(Research in Finance) Andrew H. Chen-Research in Finance, Volume 21 - Emerald Group Publishing Limited (2005) PDF

CONTENTS
LIST OF CONTRIBUTORS vii
INTRODUCTION ix
TIME DIVERSIFICATION AND STOCHASTIC DOMINANCE

Charles W. Hodges, Haim Levy and James A. Yoder 1
MINORITY EQUITY INVESTMENTS AND INTER-FIRM

COLLABORATIONS
Su Han Chan, John W. Kensinger, Arthur J. Keown
and John D. Martin 17
SIZE AND BOOK-TO-MARKET EFFECTS IN THE RETURNS

ON INFORMATION TECHNOLOGY STOCKS
Quang-Ngoc Nguyen, Thomas A. Fetherston
and Jonathan A. Batten 45
IMPLIED VOLATILITIES AND AUDITOR REPUTATION: THE

ANDERSEN CASE
Jonathan M. Godbey and James W. Mahar 93
SECONDARY EQUITY OFFERINGS: THE CASE OF

INSTALLMENTS RECEIPTS
Narat Charupat and C. Sherman Cheung 113
A NEW APPROACH TO TESTING PPP: EVIDENCE FROM

THE YEN
T. J. Brailsford, J. H. W. Penm and R. D. Terrell 135
v
vi
CORRELATION AMONG STOCK MARKETS UNDER

DIFFERENT EXCHANGE RATE SYSTEMS
Paul Sarmas 155
MULTIPLE BANKING AS A COMMITMENT NOT TO

RESCUE
Paul Povel 175
OPPORTUNITY COST AND PRUDENTIALITY: AN

ANALYSIS OF COLLATERAL DECISIONS IN BILATERAL
AND MULTILATERAL SETTINGS
Herbert L. Baer, Virginia G. France and James T. Moser 201
COLLATERALIZATION AND THE NUMBER OF LENDERS

IN PRIVATE DEBT CONTRACTS: AN EMPIRICAL
ANALYSIS
Gordon S. Roberts and Nadeem A. Siddiqi 229
AN EXAMINATION OF THE EFFICIENCY OF SINGLE VS.

MULTIPLE COMMON BOND CREDIT UNIONS
James D. Tripp, Peppi M. Kenny and Don T. Johnson 253
LIST OF CONTRIBUTORS
Herbert L. Baer Formerly at Policy Research Department, The

World Bank, Washington, DC, USA
Jonathan A. Batten College of Business Administration, Seoul
National University, Seoul, Korea and Graduate
School of Management, Macquarie University,
Sydney, Australia
T. J. Brailsford UQ Business School, The University of
Queensland, Brisbane, Australia
Su Han Chan California State University, Fullerton, Fullerton,
CA, USA
Narat Charupat School of Business, McMaster University,
Hamilton, Canada
C. Sherman Cheung School of Business, McMaster University,
Hamilton, Canada
Thomas A. Fetherston School of Business, University of Alabama,
Birmingham, Birmingham, AL, USA
Virginia G. France Department of Finance, University of Illinois at
Urbana-Champaign, Champaign Urbana, IL, USA
Jonathan M. Godbey Auburn University, Auburn, AL, USA
Charles W. Hodges Department of Accounting and Finance,
State University of West Georgia, Carrollton, GA,
USA
Don T. Johnson Western Illinois University, Macomb, IL, USA
Peppi M. Kenny Western Illinois University, Macomb, IL, USA
John W. Kensinger College of Business Administration, University of
North Texas, Denton, TX, USA
vii
viii
Arthur J. Keown R. B. Pamplin College of Business, VPI & SU,

Blacksburg, VA, USA
Haim Levy Department of Finance, Hebrew University,
Jerusalem, Israel
James W. Mahar Department of Finance, St. Bonaventure
University, St. Bonaventure, NY, USA
John D. Martin Hankamer School of Business, Baylor University,
Waco, TX, USA
James T. Moser Department of Economics & Finance, Louisiana
Tech University, Ruston, LA, USA
Quang-Ngoc Nguyen Jonathan A. Batten College of Business
Administration, Seoul National University, Seoul,
Korea and Graduate School of Management,
Macquarie University, Sydney, Australia
J. H. W. Penm Faculty of Economics and Commerce,
The Australian National University, Canberra,
Australia
Paul Povel Carlson School of Management, University of
Minnesota, Minneapolis, MN, USA
Gordon S. Roberts Schulich School of Business, York University,
Toronto, Canada
Paul Sarmas College of Business Administration, California
State Polytechnic University, Pomona, Pomona,
CA, USA
Nadeem A. Siddiqi BearingPoint Inc., Kentwood, MI, USA
R. D. Terrell National Graduate School of Management,
The Australian National University, Canberra,
Australia
James D. Tripp Western Illinois University, Macomb, IL, USA
James A. Yoder Department of Accounting and Finance, State
University of West Georgia, Carrollton, GA, USA
TIME DIVERSIFICATION AND
STOCHASTIC DOMINANCE
Charles W. Hodges, Haim Levy and James A. Yoder
ABSTRACT
We use stochastic dominance to test whether investors should prefer riskier
securities as the investment horizon lengthens. Simulated return distributions
for stocks, bonds, and U.S. Treasury bills are generated for holding periods
of one to 20 years and stochastic dominance tests are run to establish
preferences among the alternative portfolios. With independent returns, we
find no evidence that high-risk securities (stocks) dominate low-risk securities
(bonds) as the investment horizon lengthens. Under the assumption that
security returns are correlated across time, we find that common stocks
dominate corporate bonds and U.S. Treasury bills for sufficiently long
investment horizons.
1. INTRODUCTION
The issue of time diversification has generated considerable controversy.
Theoreticians, most notably Merton and Samuelson, reason that, if markets are
efficient and security returns are independent and identically distributed, then
lengthening the investment horizon should not reduce risk. Thus, an investor’s
optimal mix of securities should be independent of the planned holding period.1
Many market professionals, however, recommend that the proportion of an
investor’s holdings of high-risk securities such as equities should increase with
Research in Finance
Research in Finance, Volume 21, 1–15
Copyright © 2004 by Elsevier Ltd.
All rights of reproduction in any form reserved
ISSN: 0196-3821/doi:10.1016/S0196-3821(04)21001-2
1
2 CHARLES W. HODGES ET AL.
the length of the desired investment horizon. Lee (1990) and Reichenstein and
Dorsett (1995) offer support for this position by arguing that time diversification
will hold true if stock returns are mean reverting across time. In this case, returns
have negative autocorrelation so that volatility is reduced, because a positive or
negative price movement tends to be followed by a price movement in the
opposite direction. While numerous empirical studies employing a mean-variance
framework have been conducted on the efficacy of time diversification, they have
yielded conflicting results and have failed to resolve the issue.2
We use stochastic dominance (SD) to analyze the issue of time diversification.
Data from Ibbotson Associates 2001 Yearbook is used to generate simulated return
distributions for portfolios of small stocks, common stocks, long-term corporate
bonds, and U.S. Treasury bills for various holding periods. For each holding
period, we apply stochastic dominance tests to determine whether preferences
can be established among the equity and fixed-income portfolios. Our analysis is
first conducted under the assumption that returns are independent across time and
then under the assumption that security returns are correlated across time so that
potential mean reversion in equity returns is captured.
We employ stochastic dominance because it is a more general, analytical
technique than mean-variance analysis. Stochastic dominance examines the
entire distribution of returns and, thus, considers all the moments of the
return distribution. Furthermore, SD does not require any specific distributional
assumption such as normality. Empirical studies of time diversification based on
mean-variance analysis ignore higher moments, such as skewness and kurtosis, and
are not consistent with maximizing expected utility unless the return distributions
are normal. Albrecht (1998) demonstrates the pitfalls of applying mean-variance-
based performance measures to non-normal security return distributions.
In this paper, the stochastic dominance results clearly show that the issue of time
diversification is ultimately a question of autocorrelation in security returns. When
returns are independent across time, no dominance exists among the stock and bond
portfolios, even for long holding periods. This is consistent with the assumptions
and analyses by Merton and Samuelson. When autocorrelation across security
returns is captured, however, stock portfolios dominate the bond portfolios for suffi-
ciently long holding periods. This is consistent with the practitioner view that stock
returns are mean reverting so that lengthening the investment horizon reduces risk.
2. STOCHASTIC DOMINANCE
Stochastic dominance has important advantages over the mean-variance frame-
work that underlies other empirical time diversification studies. It is theoretically
Time Diversification and Stochastic Dominance 3
unimpeachable, requires no distributional assumptions, and takes all the moments

of the return distributions into account. Furthermore, SD requires only very general
assumptions about investor behavior.
In discriminating between the performance of the stock and bond portfolios,
stochastic dominance provides an effective method for placing investment choices
into mutually exclusive sets: an efficient set containing desirable investment
alternatives and an inefficient set containing the undesirable ones. Return
distributions of alternative portfolios are compared against one another to
determine which would be preferred (i.e. in the efficient set versus not in the
efficient set). The preference criteria is that the investor prefers more return per
dollar invested to less, is risk averse, and prefers positive skewness (potential for
great gain with limited downside risk) in a return distribution.
The basic principle underlying stochastic dominance is quite straightforward.3
Consider the two arbitrary return distributions shown in Fig. 1 and the
corresponding cumulative distributions in Fig. 2. In this example, G might
correspond to the return distribution of corporate bonds and F might correspond
to the return distribution of common stocks. Assuming that investors prefer
more return to less (first derivative of the utility function is positive), an investor
wanting to maximize expected utility would prefer return distribution F, which
lies to the right of distribution G. With distribution F, the chance of earning a
higher return is always greater than with G regardless of whether the investor likes
or dislikes risk. Formally, investment F dominates G for all utility functions if,
and only if, F(r) ≤ G(r) for all r (with at least one strict inequality), where F and
G are cumulative distributions. This constitutes first degree stochastic dominance
(FSD).
Fig. 1. Return Distributions.

Fig. 2. Cumulative Return Distributions.
In the case where F(r) lies entirely to the right of G(r), preference is readily
apparent. When two cumulative distributions cross, other factors than return
(e.g. risk aversion) must be considered in order to establish dominance. Further
assuming that investors are risk averse (second derivative of the utility function is
negative), second degree stochastic dominance (SSD) can be applied. Formally, F
dominates G for all risk-averters if and only if:
r
[G(t) − F(t)] dt ≥ 0 (1)
−∞
for all values of r, and the inequality is strict for at least one value of r. G(t) is
the cumulative distribution associated with G. F(t) is the cumulative distribution
associated with F. When F(t) lies to the right of G(t), the integral of G(t) − F(t) is
positive.
Figure 3 shows that, when the conditions of Eq. (1) are met, F(t) lies far enough
to the right of G(t) so that investment F would be preferred to investment G. This
is because the expected utility gained from the positive area to the left of R0 is
more than the decrease in expected utility lost between R0 and R1 .
Analysis of third degree stochastic dominance (TSD) is similar except that
it assumes that investors’ absolute risk aversion decreases, which implies that
investors prefer positive skewness. Formally, the TSD rule asserts that F dominates
G if, and only if:
r v
[G(t) − F(t)] dt dv ≥ 0 (2)
−∞ −∞
Fig. 3. Cumulative Return Distributions – F Preferred to G with Risk Aversion.
for all r and the expected values are:

E F (r) ≥ E G (r)
with at least one strict inequality.
Equation (2) can be interpreted as another measure of how far to the right
the cumulative distribution of investment F is relative to that of investment G. The
inside integral is merely Eq. (1). If this function is less than zero at any point, second
degree stochastic dominance is violated. Third degree stochastic dominance, on
the other hand, balances the areas where the function (i.e. the inside integral) is
positive against the areas where the function is negative. Hence, there may not be
dominance by SSD, even though a dominance by TSD prevails.
3. METHODOLOGY AND RESULTS

We now analyze the stock and bond return distributions under the assumption that
the returns are independent and then repeat the analysis allowing for autocorrelation
in the returns.
3.1. Independent Returns
We use simulation to generate sample return distributions for portfolios of small

stocks, common stocks, long-term corporate bonds, and U.S. Treasury bills for
holding periods of one to 20 years.4 Annual returns for each portfolio from 1926
through 2000 are collected from the Ibbotson Associates data. For a specified
holding period of n years and a given portfolio, we randomly select (with
replacement) n returns out of the 75 sample returns. We then compute the holding-
period return (HPR) by:
n

HPRn = (1 + R i ) − 1 (3)
i=1
where: HPRn = return for holding period of n years. Ri = ith return observation.
n = number of years in holding period.
This procedure is repeated 250 times. Thus, we generate sample return
distributions for each portfolio for annual holding periods of one to 20 years.
Note that this procedure is consistent with an efficient market because it generates
independent returns.
To illustrate, consider the case of a five-year holding period and the small-
stock portfolio. Five annual returns are selected at random from the 1926 through
2000 historical returns and the five-year holding-period return is computed. This
is repeated until a sample distribution of 250 five-year holding-period returns is
obtained.
Table 1. Mean Returns (Independent Returns).

Holding Period Small Common Corporate U.S.
(Years) Stocks Stocks Bonds T-Bills
1 0.1865 0.1396 0.0648 0.0384

2 0.4054 0.2843 0.1278 0.0785
3 0.5815 0.4185 0.1941 0.1218
4 0.8796 0.6009 0.2583 0.1676
5 1.2431 0.8002 0.3331 0.2134
6 1.8269 1.0992 0.4115 0.2609
7 2.2701 1.3921 0.4995 0.3076
8 2.7225 1.6361 0.5898 0.3593
9 3.5033 2.0382 0.6727 0.4123
10 3.8916 2.3164 0.7651 0.4705
11 5.1126 2.8870 0.8718 0.5284
12 6.2349 3.3905 0.9705 0.5863
13 7.6834 4.0060 1.0974 0.6504
14 9.1426 4.5933 1.2477 0.7159
15 9.9900 5.1838 1.3798 0.7828
16 12.7609 6.0751 1.5451 0.8558
17 14.0832 6.9254 1.6772 0.9282
18 16.3932 8.0484 1.8342 1.0045
19 19.2132 9.0971 2.0083 1.0790
20 22.3442 10.3712 2.1812 1.1584
Table 2. Standard Deviation (Independent Returns).

1 0.3415 0.2033 0.0896 0.0314

2 0.6124 0.3345 0.1249 0.0469
3 0.7956 0.4242 0.1467 0.0604
4 1.2163 0.5871 0.1970 0.0735
5 1.5884 0.7082 0.2279 0.0903
6 2.6733 1.0491 0.2724 0.1057
7 3.3439 1.2709 0.3218 0.1083
8 4.0323 1.4675 0.3602 0.1281
9 5.5973 2.0399 0.4054 0.1339
10 5.2954 2.0537 0.4472 0.1486
11 7.1563 2.6574 0.4706 0.1556
12 10.6047 3.3833 0.5106 0.1798
13 11.0493 3.7438 0.5969 0.1836
14 14.0722 4.5057 0.6794 0.1929
15 13.7674 4.8308 0.7560 0.2115
16 19.5160 5.6614 0.9234 0.2422
17 25.6209 6.8283 0.9225 0.2426
18 31.3195 9.1775 1.0202 0.2637
19 28.5125 8.6184 1.1158 0.2757
20 37.1234 10.1819 1.2254 0.2884
Descriptive statistics are computed for each sample return distribution. Tables 1
through 3 list the portfolio means, standard deviations, and skewnesses,
respectively. Table 1 shows that the mean return for all portfolios increases with the
length of the holding period. The mean return for Treasury bills, for example, grows
from 3.84% for a one-year holding period to 115.84% for a 20-year holding period.
The corresponding mean returns for small stocks are 18.656% and 2234.42%.
One might be tempted to conclude that a long-term (20-year investment horizon)
investor should invest in small stocks rather than Treasury bills since the expected
return is larger.
Risk, however, also increases with the length of the holding period. Table 2
shows that the standard deviation of returns for Treasury bills increases from
3.14% for a one-year holding period to 28.84% for a 20-year holding period. The
corresponding standard deviations for small stocks are 34.151% and 3,712.34%.
Risk, as measured by the standard deviation, grows much more rapidly with the
length of the holding period for small stocks than for Treasury bills.
Skewness coefficients given in Table 3 are generally positive and are inconsistent
with normally distributed random variables.5 Indeed, the Kolomogorov-D statistics
Table 3. Skewness (Independent Returns).

1 0.5975 −0.4003 1.4384 0.7029

2 1.1294 0.1592 1.0429 0.6224
3 1.2845 0.5065 0.9985 0.5907
4 2.3293 1.0858 1.1416 0.6260
5 2.2293 0.7606 1.1480 0.5535
6 3.2641 1.5186 1.1102 0.4931
7 3.7829 1.3195 1.4880 0.5776
8 3.2042 1.3104 1.4324 0.4032
9 3.8482 2.3178 1.5495 0.3976
10 2.9023 1.4604 1.5989 0.1392
11 3.2793 1.9264 1.1260 0.0740
12 4.9828 2.4399 0.9811 0.4420
13 3.1690 1.8461 1.2939 0.3889
14 3.8513 2.3279 1.2871 0.3007
15 2.9038 2.1656 1.3420 0.4111
16 2.9743 1.9926 1.6586 0.7835
17 6.0588 2.9756 1.1446 0.3990
18 6.7478 4.3287 1.4233 0.5260
19 2.7802 1.9004 1.2256 0.4745
20 4.5362 2.2607 1.5443 0.3502
reject normality at the 1% significance level for all the stock and bond portfolios
for all holding periods. Thus, preferences established strictly on the basis of a
mean-variance analysis would not be valid.
Stochastic dominance tests are run to determine whether preferences can be
established among the portfolios for each holding period. The SD algorithms
for discrete distributions used here are discussed in Levy (1992).6 The results
of the 20 tests are summarized in Table 4 which shows membership in the TSD
efficient set.7 As the results are the same for all holding periods, we only show the
results in five-year increments. The SSD and FSD results are identical to the TSD
results.
Table 4 shows no evidence that time diversification results in preferences among
the portfolios for risk-averse investors who prefer positive skewness. The efficient
set for each holding period includes all four portfolios. This means that an investor
with a 20-year investment horizon could rationally select any of the stock, bond,
or T-bill portfolios. Thus, with independent returns, time diversification fails and
Merton and Samuelson are correct.
Table 4. Stochastic Dominance Results with Independent Returns.a

1 Yes Yes Yes Yes

5 Yes Yes Yes Yes
10 Yes Yes Yes Yes
15 Yes Yes Yes Yes
20 Yes Yes Yes Yes
Note: Yes = Membership in the TSD Efficient Set.

a Results for all other holding periods are identical. FSD and SSD results are identical to the TSD
results.
3.2. Autocorrelated Returns
We now repeat the analysis allowing for autocorrelation across time in the security
returns. Mean reversion (negative autocorrelation) in long-run stock returns and
mean aversion (positive autocorrelation) in fixed-income securities have been
Table 5. Mean Returns Under Autocorrelation.

2 0.4115 0.2890 0.1397 0.0804

3 0.6035 0.4144 0.2094 0.1254
4 0.9260 0.5883 0.2928 0.1714
5 1.2131 0.7646 0.3829 0.2206
6 1.5646 0.9903 0.4813 0.2740
7 1.8780 1.2020 0.5797 0.3308
8 2.2722 1.4561 0.7024 0.3914
9 2.7121 1.7393 0.8114 0.4569
10 3.5052 2.1560 0.9526 0.5291
11 4.2175 2.5344 1.0765 0.6021
12 5.1086 2.9976 1.2282 0.6871
13 6.3226 3.5319 1.3557 0.7770
14 7.3999 4.1100 1.5001 0.8653
15 8.5451 4.7234 1.6490 0.9617
16 9.7845 5.3502 1.8023 1.0657
17 11.4878 6.1242 1.9380 1.1678
18 13.7256 7.0687 2.0996 1.2774
19 15.9648 7.9136 2.2472 1.3985
20 18.4647 8.9042 2.3592 1.5165
reported by various researchers.8 This autocorrelation may induce preferences

among the alternative stock and bond portfolios.
Sample return distribution for small stocks, common stocks, long-term corporate
bonds, and U.S. Treasury bills are constructed as follows. For a given holding
period of n-years, we randomly select an observation (year) from the first
75−(n−1) observations in the sample. The selected return and the next n−1
consecutive returns are then used to compute the n-year holding-period return
using equation 1. Since holding-period returns are now computed from consecutive
returns, autocorrelation across time in the portfolio returns is now captured.
Eliminating the last n−1 observations before selecting the random observation
guarantees that the n-year holding-period return can always be computed. This
procedure is repeated 250 times.
To illustrate, consider a five-year holding period. A return is randomly selected
from the first 71 (1926 through 1996) observations and the holding-period return
is computed using the next four consecutive returns. This process is repeated until
a sample distribution of 250 five-year returns is obtained. Note that if the 71st
observation (1996) is selected, then the five-year holding period can be computed
from observations 71 through 75 (1996 through 2000 returns).
Table 6. Standard Deviations Under Autocorrelation.

2 0.6034 0.3432 0.1395 0.0675

3 0.8904 0.3912 0.1818 0.1071
4 1.3331 0.4963 0.2404 0.1464
5 1.5134 0.6138 0.3082 0.1899
6 1.6694 0.7209 0.3884 0.2384
7 1.7181 0.8417 0.4737 0.2891
8 2.1213 1.0415 0.5804 0.3433
9 2.2990 1.1547 0.6620 0.4024
10 2.8453 1.4077 0.8211 0.4673
11 3.3260 1.6876 0.9444 0.5315
12 3.6656 2.0122 1.1133 0.6091
13 5.0003 2.3484 1.1756 0.6889
14 5.3469 2.7683 1.3667 0.7669
15 5.5767 3.2706 1.4700 0.8489
16 5.3380 3.5738 1.6305 0.9359
17 6.3668 4.0850 1.7634 1.0213
18 8.5244 4.7779 1.9361 1.1117
19 9.6500 5.1602 2.0700 1.2135
20 10.5444 5.7670 2.1265 1.3112
Table 7. Skewness Under Autocorrelation.

2 0.6411 −0.0591 1.1567 0.8057

3 1.7276 −0.1697 1.0532 0.7869
4 2.3218 0.3019 1.3166 0.7675
5 1.5077 0.1419 1.5876 0.7262
6 1.2361 0.1563 1.4143 0.6951
7 1.0215 0.2983 1.2562 0.6350
8 1.4028 0.2872 1.2864 0.6024
9 1.7647 0.2584 1.1687 0.5690
10 1.6208 0.2706 1.3682 0.5260
11 1.6202 0.2465 1.3391 0.4982
12 0.8436 0.2718 1.4232 0.4755
13 1.6289 0.2798 1.1737 0.4356
14 1.2038 0.3750 1.3644 0.4140
15 0.9520 0.5382 1.1949 0.3793
16 0.3263 0.5120 1.1597 0.3411
17 0.6433 0.6806 1.1334 0.3200
18 1.0471 0.7746 1.1079 0.3024
19 0.7733 0.6770 1.1268 0.2846
20 0.6426 0.8457 1.0407 0.2683
Descriptive statistics are computed for each sample return distribution. Tables 5,
6, and 7 give the portfolio means, standard deviations, and skewnesses, respec-
tively. Note that the tables start with a two-year holding period since autocorrelation
across time requires at least two holding periods. Tables 5 and 6 show that expected
returns and standard deviations increase with the holding period for all portfolios.
Table 7 shows that the skewness coefficients remain positive, indicating
that the return distributions are not normal. Mean-variance analysis remains
inappropriate.
A comparison of corresponding standard deviations in Table 2 with those in
Table 6 indicates that the portfolio returns are autocorrelated across time. The
two equity portfolios exhibit negative autocorrealtion or mean-reverting behavior.
Under mean-reversion, a positive return tends to be followed by a negative return
so that volatility is dampened. The standard deviations for the small-stock and
common-stock portfolios are all less than those for corresponding independent
returns for holding periods of five years or longer. For example, the standard
deviation for the common-stock portfolio for a 20-year investment horizon is only
576.70% in Table 6 (autocorrelated returns) compared to 1018.19% in Table 2
(independent returns).
The corporate bond and Treasury bill portfolios, however, exhibit positive
autocorrelation or mean aversion. Under mean aversion, a positive return tends to
be followed by another positive return so that volatility is enhanced. The standard
deviation of the corporate bond and Treasury bill portfolios in Table 6 are greater
than those for corresponding independent returns given in Table 2 for all holding
periods. For example, the standard deviation for the corporate bond portfolio for
a 20-year investment horizon is 212.65% in Table 6 (autocorrelated returns) but
only 122.54% in Table 2 (independent returns).
These results suggest that time diversification may hold with autocorrelated
returns. Mean reversion in the stock portfolios combined with mean aversion in
the fixed-income portfolios imply that the relative risk of the equity portfolios
compared to the fixed-income portfolios should decline as the holding period
lengthens. Thus, equities would be preferred to fixed-income securities for
sufficiently long holding periods as investment professionals claim.
Stochastic dominance tests are run to determine if preferences can be established
when autocorrelation in the return generating process is captured. The results given
in Table 8 show membership in the TSD efficient set. As predicted, preferences
Table 8. Stochastic Dominance Results Under Autocorrelation.a

2 Yes Yes Yes Yes

3 Yes Yes Yes Yes
4 Yes Yes Yes Yes
5 Yes Yes Yes Yes
6 Yes Yes Yes Yes
7 Yes Yes Yes Yes
8 Yes Yes Yes No
9 Yes Yes Yes No
10 Yes Yes Yes No
11 Yes Yes Yes No
12 Yes Yes Yes No
13 Yes Yes Yes No
14 Yes Yes Yes No
15 Yes Yes Yes No
16 Yes Yes No No
17 Yes Yes No No
18 No Yes No No
19 No Yes No No
20 No Yes No No
Note: Yes = Membership in the TSD Efficient Set.

a SSD results are identical.
for the equity portfolios are now established. Common stocks and small stocks
dominate Treasury bills for holding periods of eight years or longer and corporate
bonds for holding periods of 16 years or longer. Common stocks are preferred
over small stocks for holding periods of 18 years or longer. The SSD results are
identical to the TSD results.
4. CONCLUDING COMMENTS
The results in this paper reconcile the views of theoreticians and market
professionals. With independent security returns, no dominance exists between
the equity and fixed-income portfolios. All portfolios are in the efficient set for all
holding periods. This is consistent with the arguments of Merton and Samuelson
that time diversification does not reduce risk and that asset allocation should not
be influenced by the intended holding period.
However, when autocorrelation in security returns is considered, we find that
the stock portfolios dominate the fixed-income portfolios given sufficiently long
holding periods. Mean reversion in stock returns combined with mean aversion in
bond returns makes equities more attractive as the holding period lengthens. This is
consistent with the practitioner view that the proportion of equity holdings should
increase with the intended holding period. Thus, the issue of time diversification
is ultimately a debate about the degree to which security returns are autocorrelated
across time.
NOTES
1. See Merton and Samuelson (1974) and Samuelson (1990, 1994).

2. See, for example, Lloyd and Haney (1980), Lloyd and Modani (1983), McEnally
(1985), Butler and Domian (1991), Levy and Gunthorpe (1993), Gunthorpe and Levy
(1994), Thaler and Williamson (1994), Thorley (1995), Asness (1996), Ferguson and
Simaan (1996), Hodges, Taylor and Yoder (1997), Olsen and Khaki (1998), Hansson and
Persson (2000) and Siegel (2002).
3. For a review of stochastic dominance rules, see Levy (1992).
4. This sampling approach is similar to the bootstrapping methodology introduced by
Efron (1979).
5. Kurtosis coefficients are generally positive and are inconsistent with normally
distributed random variables. Kurtosis results are available from the authors upon
request.
6. Computer programs for stochastic dominance with discrete distributions are given in
Levy and Sarnat (1984).
7. Similar results are obtained by Hodges and Yoder (1996).
8. See, for example, Fama and French (1988), Poterba and Summers (1988), Reichenstein
and Dorsett (1995), Lewellen (2001), and Strong and Taylor (2001).
REFERENCES
Albrecht, T. (1998). The mean-variance framework and long horizons. Financial Analysts Journal, 54,
44–49.
Asness, C. S. (1996). Why not 100% equities? Journal of Portfolio Management, 22, 29–34.
Butler, K. C., & Domian, D. L. (1991). Risk, diversification, and the investment horizon. Journal of
Portfolio Management, 17, 41–47.
Efron, B. (1979). Bootstrap methods: Another look at the jack-knife. Annals of Statistics, 7, 1–12.
Fama, E., & French, K. (1988). Permanent and temporary components of stock prices. Journal of
Political Economy, 96, 246–273.
Ferguson, R., & Simaan, Y. (1996). Portfolio composition and the investment horizon revisted. Journal
of Portfolio Management, 22, 62–67.
Gunthorpe, D., & Levy, H. (1994). Portfolio composition and the investment horizon. Financial Analysts
Journal, 50, 51–55.
Hansson, B., & Persson, M. (2000). Time diversification and estimation risk. Financial Analysts
Journal, 56, 55–62.
Hodges, C., Taylor, W., & Yoder, J. A. (1997). Stocks, bonds, the Sharpe ratio, and the investment
horizon. Financial Analysts Journal, 53, 74–80.
Hodges, C., & Yoder, J. A. (1996). Time diversification and security preferences: A stochastic
dominance approach. Review of Quantitative Finance and Accounting, 7, 289–298.
Ibbotson, R. (2001). Stocks, bonds, bills, and inflation: 2001 yearbook. Chicago, IL: Ibbotson.
Lee, W. (1990). Diversification and time: Do investment horizons matter? Journal of Portfolio
Management, 16, 21–26.
Levy, H. (1992). Stochastic dominance and expected utility: Survey and analysis. Management Science,
38, 555–593.
Levy, H., & Gunthorpe, D. (1993). Optimal investment proportions in senior securities and equities
under alternative holding periods. Journal of Portfolio Management, 19, 30–36.
Levy, H., & Sarnat, M. (1984). Portfolio investment selection: Theory and practice. Englewood Cliffs,
NJ: Prentice-Hall.
Lewellen, J. (2001). Temporary movements in stock prices. Working Paper, MIT.
Lloyd, W. P., & Haney, R. L., Jr. (1980). Time diversification: Surest route to lower risk. Journal of
Lloyd, W. P., & Modani, N. K. (1983). Stocks, bonds, bills, and time diversification. Journal of Portfolio
McEnally, R. W. (1985). Time diversification: Surest route to lower risk? Journal of Portfolio
Merton, R. C., & Samuelson, P. A. (1974). Fallacy of the Log-Normal approximation to
optimal portfolio decision-making over many periods. Journal of Financial Economics, 1,
67–95.
Olsen, R. A., & Khaki, M. (1998). Risk, rationality, and time diversification. Financial Analysts Journal,
54, 58–63.
Poterba, J., & Summers, L. (1988). Mean reversion in stock prices: Evidence and implications. Journal
of Financial Economics, 22, 27–59.
Reichenstein, W., & Dorsett, D. (1995). Time diversification revisted. Charlottesville, VA: Research
Foundation of the Institute of Chartered Financial Analysts.
Samuelson, P. A. (1990). Asset allocation could be dangerous to your health. Journal of Portfolio
Samuelson, P. A. (1994). The long-term case for equities. Journal of Portfolio Management, 21, 15–26.
Siegel, J. J. (2002). Stocks for the long run. New York, NY: McGraw-Hill.
Strong, N., & Taylor, N. (2001). Time diversification: Empirical tests. Journal of Business Finance and
Accounting, 28, 263–302.
Thaler, R. H., & Williamson, J. P. (1994). College and university endowment funds: Why not 100%
equities? Journal of Portfolio Management, 23, 27–37.
Thorley, S. R. (1995). The time-diversification controversy. Financial Analysts Journal, 51, 68–76.
MINORITY EQUITY INVESTMENTS
AND INTER-FIRM COLLABORATIONS
Su Han Chan, John W. Kensinger, Arthur J. Keown

and John D. Martin
ABSTRACT
We examine the benefits for firms participating in collaborations funded via

minority equity placements. Selling firms, on average, realize significant
increases in share value – strongly correlated with the size of the equity
stake, its beta, and the relatedness of the two firms (by industry). Shares of
purchasing firms, though, show neutral responses on average (but positive
response for R&D intensive alliances). Further, purchasing firms have better
financial performance than their industry peers in the years surrounding the
announcement (suggesting, unlike joint ventures, that poor performance is not
their motivation). Selling firms, however, may be motivated by poor operating
performance.
1. INTRODUCTION
The use of interfirm collaborations or business alliances has grown dramatically

over the last decade as firms have sought to improve performance by pooling
corporate resources to engage in joint design, production, marketing, or distribution
of products and services. Interfirm collaborative agreements can manifest them-
selves in one of several organizational forms including complex contracts, strategic
Research in Finance
ISSN: 0196-3821/doi:10.1016/S0196-3821(04)21002-4
17
18 SU HAN CHAN ET AL.
alliances, joint ventures, and mergers. We focus on strategic alliances which are an
intermediate form of collaboration. Specifically, we study strategic alliances that
involve a minority equity investment by one of the partnering firms (the purchaser)
in the other (the seller).1 These agreements are particularly interesting because the
minority equity stake not only provides an infusion of cash to the seller but gives
the purchaser partial control over the seller’s resources without creating a new
entity as in a joint venture. In addition, the partial ownership interest provides a
mechanism for bonding the partnering firms to the alliance agreement.
We examine investor reactions to the announcement of corporate purchases
of minority equity stakes associated with the formation of strategic alliances.
Despite the prevalence of minority equity investments among partners to interfirm
collaborations, especially in high-tech industries, there is little empirical evidence
as to their valuation consequences. Previous studies of interfirm equity investments
(see Choi, 1991; Mikkelson & Ruback, 1985) or block shareholdings (see Bethel,
Liebeskind & Opler, 1998; Hertzel & Smith, 1993; Wruck, 1989) have not focused
on the use of minority equity stakes as a bonding mechanism for a strategic
alliance.2
Prior research on interfirm collaboration has examined joint ventures (Johnson
& Houston, 2000; McConnell & Nantell, 1985), non-equity strategic alliances
(Chan, Kensinger, Keown & Martin, 1997), and equity alliances (Allen & Phillips,
2000). The results of the joint venture and alliance studies suggest that interfirm
collaborations, in general, benefit the partnering firms.
Our study is most closely related to a recent study by Allen and Phillips (2000)
who investigate long-term block ownership by corporations and performance
changes in firms with corporate block owners. Using a sample of 150 corporate
equity purchases that accompany explicit product market relationships and 252
equity stakes that are not associated with business relationships during the
1980–1991 period, Allen and Phillips examine three potential reasons for block
ownership: accruing the benefits of product market relationships, the alleviation of
financing constraints, and improved board monitoring by corporate owners. They
find that the largest increases in the sellers’ stock prices, investment, and operating
profitability occur when the equity ownership is accompanied by alliances, joint
ventures, and other product market relationships between the purchasing and
selling firms. Furthermore, their findings were strongest in industries with high
R&D expenditures. The study concludes that block ownership has significant
benefits in product market relationships.
We provide additional evidence as to the sources of value creation in inter-firm
collaborations supported by equity participation. Our study focuses exclusively on
minority equity alliances and adds to the findings of the Allen and Phillips (2000)
study in the following ways.
Minority Equity Investments and Inter-Firm Collaborations 19
First, we examine firms that announce strategic alliances or collaborative

agreements contemporaneously with a minority equity participation.3 Our sample
permits us to examine not only the value effects of these minority-equity
participation cum alliance announcements to the participating firms but also events
immediately following the announcement to see how these relationships evolve
over time.
Furthermore, we exclude joint ventures since they have been studied extensively
elsewhere. Joint ventures also differ fundamentally from minority equity alliances
in that all joint venture partners generally contribute funding or other resources to
the joint project that forms the basis for the venture, whereas in a minority equity
alliance one partner buys and the other sells equity. The equity investment aligns
the interests of the partnering firms and thus serves as a bonding mechanism for
the partners to the alliance.
The clear distinction of buyer and seller firm in a minority-equity relationship
permits us to separately test the value effect of the alliance on the value of the
purchasing and selling firms. Similar to Allen and Phillips’s finding for their
subsample of equity purchases where alliances or joint ventures are formed, we
find that selling firms in minority equity alliances during the 1981–1992 period
realize significant increase in share value while purchasing firms garner neutral
share price responses. In addition we document that the majority of the alliances
are strengthened or extended in the period up to three years following the initial
announcement of the tie-up.
Second, we conduct multiple regression tests to provide additional insight into
the determinants of wealth gains for the purchasing and selling firms that announce
minority equity tie-ups. Allen and Phillips (2000) do not provide such evidence;
they focus on the determinants of investment and operating performance changes
following corporate block equity purchases. Our analysis reveals that purchasing
firms that are more R&D intensive seem to reap higher returns from a minority
equity alliance than their less R&D intensive counterparts. For selling firms, we
find that increases in their share values are strongly correlated with the size of
the equity stake, their riskiness, and the relatedness of the industries to which the
partnering firms belong. These results are consistent with the hypotheses discussed
in Section 2 of the paper.
Finally, we document the yearly operating performance of both purchasing
firms and selling firms in the six-year period surrounding the announcement of
the minority equity purchase to see whether the firms’ operating performance
motivates them to enter into such arrangements and how the firms perform
following the formation of the alliance. Our findings suggest that purchasing
firms have cash flows to invest while selling firms may be motivated to enter
into such alliances because of poor operating performance. The selling firms show
significant improvement in their operating performance in the third year following

the alliance.
The paper is organized as follows: Section 2 elaborates on the sources
of value creation from collaboration involving minority equity investments.
Section 3 describes the data collection procedures and the resulting test sample
as well as the test methodology. Section 4 discusses the findings, and Section 5
concludes.
2. MINORITY EQUITY STAKES AND INTERFIRM

COLLABORATIONS
Collaboration using minority equity investments is an essential strategy especially
for firms in high tech industries such as biotechnology and computer hardware
and software. By banding together the allied firms hope to benefit from reduced
investment risks, improved access to new technologies and product markets, and
the accelerated development of new products. Furthermore, cash starved firms
might gain access to capital under more favorable terms than would be available
in the capital market. Where the buyer possesses specific knowledge relevant to
the seller’s business, it may possess skill in evaluating the seller’s investment
opportunities that is superior to investors in the public capital market and thus
be willing to invest under terms more favorable to the seller than would be the
case with a public equity offering.4 Furthermore, the specialized knowledge of
the purchaser combined with the bonding effects of common ownership may help
deter opportunistic behavior by the selling firm.5 This, in turn, reduces the cost of
raising capital compared to the public capital market (resulting from asymmetric
information).
2.1. Pooling Resources and Lowering Transactions Costs
In an equity alliance one or more of the participating firms purchases common

stock from an allied firm. The purchase provides more than a cash infusion to
the selling firm; it also bonds the firms’ joint efforts by providing the purchasing
firm with access to information concerning the seller’s cost structure and often a
voice in management. (See Kensinger and Martin (1991), for a further discussion
of minority equity stakes as a means for financing network organizations.)
Harrigan (1985) argues that a minority investment by a larger firm may be the
only way for the purchasing firms’ managers to obtain access to the assets and
capabilities of smaller, knowledge-intensive firms. A case in point is provided by

Hewlett-Packard which purchased an equity stake in Conductus Inc., a closely-
held Silicon Valley startup that produces superconducting chips. Conductus
invented new equipment for forming brittle superconducting ceramics into the
thin films needed to make micro-chips. A Hewlett-Packard official was quoted as
saying, “Research in superconductors is better done by small companies such as
Conductus. It’s a very valuable relationship for us.”
Another case in point to illustrate the role minority equity stakes plays and the
potential benefits to firms in this mode of collaboration is the relationship between
IBM and Hogan Systems, Inc. (a leading developer and provider of integrated
banking applications software and services). In September 1990, IBM took a
5% equity stake in Hogan Systems, Inc. to strengthen and cement a marketing
relationship that began in 1986. At the same time, it also amended and extended
agreements that granted IBM exclusive rights to license, market and provide
support for Hogan’s banking software product in North America. In addition, IBM
will also provide funding to Hogan for the joint development of some projects. In
return, Hogan will receive 50% of revenues from IBM’s sales of its products.
This pooling of resources appears to benefit both parties because Hogan was
having problems selling its software to the larger banks and IBM needed an
integrated banking software package for its banking hardware. Psychologically,
the alliance gives the impression that IBM, the leading hardware company has just
identified Hogan as the leading bank software company. This helps bolster Hogan’s
reputation. The agreement also gives Hogan access to IBM’s deep pockets, a unique
competition over its competitors, and access to certain areas of IBM’s technology
and planning. On the downside, however, Hogan will be put in a precarious position
without a distribution capability but there is a clause in the contract that provides
Hogan with a certain amount of protection. The benefit that IBM gets is that
Hogan’s banking software package will help it sell its mainframe computers to
financial institutions.
These examples illustrate how the allied firms benefit from synergy in
production, distribution, and sourcing of inputs. Note, however, that the purchasing
firm would gain even more if it purchases the seller’s shares at a discounted price.
However, Wu (2001) shows that private placements to strategic partners (including
suppliers and customers) are at a premium as opposed to private equity sales to
informed investors (such as managers or directors, block shareholders, venture
capitalists, and institutional investors) that are generally at a discount. Therefore the
benefits to the purchasing firm, if any, will come mainly from future collaborative
opportunities. For deeper discussion of the role of knowledge, see Jensen and
Meckling (1991).
2.2. Mitigating Adverse Selection and Information Problems
Wruck (1989) shows that a private sale of equity results in a shift to more
concentrated holdings by non-managers and this generally increases firm value
if the blockholder uses his votes to see that corporate resources are managed more
efficiently or if the existence of the block increases the probability of a value-
increasing takeover. In the case of the equity alliances we study, the purchasing
partner generally does not intend to take over control or restructure the selling firm.
Therefore, any shift in assessment of the value of the seller is more likely to come
from a more efficient use of corporate resources. Also, the purchase of a minority
equity stake to form an alliance by an informed purchaser is an endorsement of the
selling firm’s management and its future prospects (e.g. the value of the seller’s
option to develop and market its new technology). Therefore, a reassessment of
the seller’s value could also be a result of the revelation of new information to the
market concerning the seller’s future prospects (which the market is incapable of
evaluating without this signal).
Following the above reasoning, we expect that the formation of an equity
alliance will create value for both the purchasing and selling firms. This value
arises from a lowering of transaction costs through pooling of resources while
outsourcing non-core functions and taking advantage of the capabilities of the
allied firms. It also arises from the mitigation of adverse selection and information
asymmetry problems regarding the seller’s future prospects. However, we expect
that the value creation will vary across the partnering firms in predictable ways. For
example, Mody (1993) suggests that the flexibility inherent in alliances facilitates
experimentation with new combinations of participants in the pursuit of new
technologies or marketing strategies. Therefore, we expect firms that are highly
R&D intensive or that have high growth opportunities will gain more value from
experimentation as well as from the pooling of resources. Also we expect firms
facing higher risks in their operations to benefit more from the pooling of resources
with a partner firm in an alliance. Given information asymmetry problems, we
expect more value added for alliances involving firms in related industries (see
also Balakrishnan & Koza, 1993).
2.3. Controlling Opportunistic Behavior Via Minority Equity Ownership
On the downside, alliances carry risks not incurred in the fully integrated
corporation. These include a variety of uncertainties arising out of dependence
on another party for resources or services. (See Klein, Crawford and Alchian
(1978), for a discussion of this general class of problems.) Furthermore, such
partnerships are inherently less stable than fully integrated (merged) firms because
the contracts used to form them cover a limited scope of activities, and residual
rights to profits are sometimes purposefully left ambiguous. According to Murray
and Siehl (1989) these agreements constitute “complex contracts” for which it is
difficult to identify a priori the potential outcomes of the relationship, the factors
causing those outcomes, the appropriate responses to the outcomes, or which party
is responsible for each response. Due to this ambiguity, a relationship of mutual
interest (bonding) replaces legalistic enforcement mechanisms. Typically, the bond
takes one of two basic forms: an equity investment by one of the partnering firms
(a minority equity stake) or hostage assets.6
For example, Mody (1993) points out that alliances have weak incentives to
curb opportunistic behavior on the part of the partnering firms when compared
to the fully integrated firm that arises out of a merger. Equity ownership in an
equity alliance helps create an ownership structure that safeguards investments
(and other claims to future income) while aligning the incentives of the partners
for cooperative behavior (Pisano, 1989). Opportunism by an equity partner
is penalized through reductions in the value of its equity investment. Thus a
higher minority equity stake ensures greater bonding of the firm’s efforts and
reduces the possibility of opportunstic behavior as it helps align the incentives of
the partners.
Furthermore, in the event of termination, equity ownership protects the acquiring
partner by providing an enforceable mechanism for dissolving the alliance. That
is, the proportionate ownership interest of the acquiring partner provides a ready
basis for determining exactly what is due to each partner to the agreement should
the alliance be terminated. We hypothesize, therefore, that the larger the size of
the minority equity stake relative to the total shares the seller has outstanding the
lower is the risk of opportunism and the greater the potential value created by the
alliance.
Pisano (1989) argues that R&D is inherently more costly to govern through
contracts and is subject to incentive losses when internalized completely. For R&D
activities partial equity ownership is preferred to pure contractual governance.
Based on this argument, we would expect equity alliances involving R&D to be
more prevalent than those involving non-R&D activities.
In summary, because minority equity investments benefit the allied firms from
a lowering of transaction costs through the pooling of resources, we expect stock
market gains for both the purchasing and the selling firm. However, the wealth
gains will vary across the partnering firms in predictable ways. We expect more
gains to accrue to firms that are highly R&D intensive, that have high growth
opportunities, and that face high risks in their operations. Also given information
asymmetry problems experienced by selling firms, we expect more value added
for selling firms that form alliances with purchasing firms from related industries.
We also hypothesize that the larger the size of the minority equity stake relative
to the total shares the seller has outstanding, the lower is the risk of opportunism
and the greater the potential value created by the alliance.
3. DATA AND METHODOLOGY

3.1. Sample Description
To obtain the sample of firms entering into alliances involving the purchase of
minority equity stakes, we searched both the Lexis/Nexis database (including
the Business Wire, PR Newswire, Southwest Newswire Reuters and United Press
International) and the Dow Jones News Retrieval Service database (including the
Dow Jones News Wire and the Wall Street Journal) for the 1981–1992 period. We
combine various keywords including “alliance,” “minority-equity” and “minority
stake” with the different agreement types (licensing, marketing, distribution,
supply, production, manufacturing, development, research, and technology).
Unlike some previous studies of interfirm equity investments (see Choi, 1991;
Madden, 1981; Mikkelson & Ruback, 1985), we did not use the SEC Schedule
13D filings to identify our initial set of firms. This was done for the following
reasons: (1) equity alliances often involve small, non-publicly traded seller firms
(13D filings are required only when 5% or more of a publicly traded firm’s stock
is acquired); (2) 13D filings do not give information as to the purpose of the
investment; and (3) previous studies have noted that the 13D filings are often
preceded by press releases or WSJ reports that reveal plans to acquire a minority
stake in another firm.7
We found 93 announcements of minority-equity investments that meet the
following sample selection criteria: (1) the purpose of the buy-in is to form an
alliance to pool knowledge and resources to achieve a common objective; (2)
the equity stake creates a minority position (does not exceed 50% of the selling
firm’s outstanding shares); (3) at least one of the partners in the alliance has
stock return data available in the CRSP daily returns file for NYSE, AMEX, or
NASDAQ firms during the period of analysis;8 and (4) at the time of the buy-in, the
purchasing firm gave no indication of an intention to acquire the seller firm. The
following announcement by Hewlett-Packard is an example of an announcement
of a minority-equity investment:
Hewlett-Packard Co. announced that it had entered into a letter of intent with Santa Barbara
Laboratories Inc. providing for a multi-faceted relationship between the two companies, as part
of which Hewlett-Packard will acquire a minority-equity position in Santa Barbara Laboratories
for an undisclosed amount of cash. . . . The company said the proposed relationship includes
arrangements for technical collaboration (Dow Jones News Wire, 9/24/84).
Of the firms involved in the minority-equity alliances, stock return data was
available in the CRSP file for 85 purchasing firms and 29 selling firms (two of
these announcements involve multiple firms purchasing the same seller firm).9 The
minority equity stake (as reported by the purchasing/selling firm in the news article)
averages 15%, ranging from 3% up to 32% of the selling firm’s outstanding shares
(59 of the announcements reported precise figures on the size of the minority-
equity position, while the others simply reported that a minority-equity position
was taken). Mikkelson and Ruback (1985) in their study of the value effects
of accumulation of large blocks report an average equity position of 37% for
13D filings associated with an outstanding takeover proposal and 9.8% where the
equity purchase is for investment purposes. In Wruck’s (1989) study of private
sales of equity securities she reports that after taking into account the purchasers’
stockholdings prior to the sale, the stake averages 26%, and ranges from about 2
to 84%. The higher stakes reported by Wruck may reflect the fact that her sample
includes cases where the purchaser intends to gain control of the selling firm,
obtain the right to elect or nominate directors, or help impede a takeover of the
selling firm by another firm. Allen and Phillips (2000) report that the average size
of the equity stakes in their sample was 20% (median 14%) which is very similar
to our sample. Thus, it appears that the average equity stake in minority equity
alliances lies somewhere between that associated with a takeover proposal and a
passive investment.
Panel A of Table 1 presents the sample observations classified by the year of
their announcement. While the sample spans twelve years, approximately 45% of
the sample is concentrated in 1989 and 1990. The concentration of minority equity
alliance announcements in these two years is consistent with a report indicating
that acquisitions and minority equity investments rose in 1989 as the U.S. high-
technology industry restructured (Electronic Business, July 9, 1990, pp. 42–44).
According to the same report the number of foreign firms acquiring equity interests
in U.S. electronics firms also rose in that year. In our sample we find four incidences
of buy-ins by foreign firms, one in 1988, two in 1989 and one in 1990. For the
foreign purchasing firms, the primary motive is to tap an existing infrastructure
and to establish an international working partnership.10
Panel B of Table 1 classifies the alliances by the purpose of the agreement. We
find that 26% of the announcements (24 cases) involve development or research
alliances followed by 23 cases of marketing alliances (25%). There are 25 cases
(27%) that involve combinations of the four types of agreement and 9 cases
(10%) that did not specify the purpose for forming the equity alliance. The above
Table 1. Distribution of Announcements of Interfirm Alliances Involving

Minority Equity Investments.
Panel A: Annual Distribution of Announcements
Year of Number of % of
Announcement Announcements Total
1981 2 2.2
1982 4 4.3
1983 5 5.4
1984 4 4.3
1985 3 3.2
1986 10 10.8
1987 6 6.5
1988 8 8.6
1989 22 23.7
1990 20 21.5
1991 2 2.2
1992 7 7.5
Total 93 100.0
Panel B: Distribution of Announcements by Type of Cooperative Agreement
Type of Cooperative Agreement Number of By Purchasing By Selling

Announcements Firms Firms
I. Licensing 4 3 1
II. Marketing 23 19 7
III. Development or Research 24 24 7
IV. Technology 8 8 2
V. Combination of II and III 14 12 9
VI. Other combinations of I–IV 11 11 1
VII. Not Specified 9 8 2
Total 93 85 29
Note: The table reports the distribution by year and by type of cooperative agreement for the 93 alliance
announcements involving the purchase/sale of minority equity stakes by NYSE, AMEX, and
NASDAQ firms in the 1981–1992 period. The sample of announcements (identified by searching
the Lexis-Nexis database) reported below include only cases where at least one of the partnering
firms in the alliance is publicly traded with return data available on the CRSP tape. Panel
A reports the number of announcements in each year during the 1981–1992 period. Panel B
classifies the announcements by the type of cooperative agreement as stated in the news article
containing the announcement of the alliance. It also reports the number of announcements under
each type made by purchasing firms and selling firms.
distribution of equity alliances by purpose of the agreement is consistent with

our conjecture that equity alliances encompassing R&D activities would be more
prevalent relative to the other types of activities (consistent with Chan et al., 1990).
Table 2 classifies the observations by industry affiliation based on SIC codes. The
table also groups the purchasing and selling firms into high-technology and low-
technology groups. This grouping is based on both the SIC code and information
in CorpTech (a database containing information about high-tech manufacturing,
development and service companies). About 86% of the purchasing firms and
93% of the selling firms belong to the high-technology industries. Furthermore, the
minority-equity buy-ins are most prevalent in the computer and software industries
(45 purchasing firms and 13 seller firms). The largest group of low-technology firms
Table 2. Industry Affiliation of Purchasing and Selling Firms.

Industry Purchasing Selling
Firms Firms
Panel A: High-technology group

Aircraft engine and parts 1
Computer programming, systems design and software 5 9
Computer equipment and industrial machinery 40 4
Medical or biotech 2
Pharmaceuticals 6 2
Chemical and thermoplastics 4 1
Instruments 5 1
Semi-conductors and electrical equipment 5 7
Telecommunications 7 1
High-tech sub-total 73 27
Panel B: Low-technology group
Air transportation 1
Farm machinery and equipment 3
Motor vehicle and parts 5 1
Paper and allied products 1
Rubber and miscellaneous plastic products services 2 1
Low-tech sub-total 12 2
Grand total 85 29
Note: The table classifies the purchasing and selling firms in our sample into a high-technology group or
a low-technology group. We classify a firm as being in the high-technology group if its operations
are primarily hitech as described in CorpTech, a database containing information about high-
tech manufacturing, development and service companies in the U.S. Both the Purchasing and
selling firms reported below are publicly-traded firms with return data available on the CRSP
tape.
(5 purchasing firms and 1 seller) comes from the motor vehicle and parts industry.
We also find that of the 93 announcements of minority-equity purchase, 38 (41%)
involve combinations between firms in related industries (the same three-digit SIC
industry).11
Most of the buy-ins involve larger firms acquiring the shares of a smaller firm.
Furthermore, roughly 69% of the purchasing firms bought into private firms. There
are only 23 cases where both purchasing and selling firms are publicly-traded with
data available on the CRSP tapes. (Note that there are 29 cases if we include
6 cases in which one of the partnering firms is a foreign-listed firm.) Note that
those sellers that are publicly-traded are relatively young in terms of the number
of years since listing. The average as well as the median age since listing is about
6 years.
Table 3 presents sample statistics to show the basic characteristics of both
the purchasing and selling firms that collaborate using a minority equity stake.
The purchasing firms are much larger, about 85 times, than the selling firms (as
measured by the market value of equity 21 days prior to the announcement). It
Table 3. Profile of Purchasing and Selling Firms.

Summary Statistics Market Value RD/Sales BV/MV Leverage Variance
of Equity ($M)
Purchasing firms
Mean 21,369 0.08 0.58 0.49 0.07%
Standard deviation 0.04 0.37 0.16 0.19%
Median 0.07 0.55 0.46 0.03%
Sample size 85 76 83 84 84
Sellers firms
Mean 252 0.95 0.40 0.50 0.16%
Standard deviation 2.32 0.28 0.26 0.11%
Median 0.17 0.38 0.48 0.12%
Sample size 29 23 27 28 29
Note: The table provides summary statistics for the Purchasing and selling firms in our sample. The
market value of equity is the market value of the firm’s common stock 21 trading days before the
initial announcement of the equity alliance. RD/Sales is the ratio of research and development
expenditures divided by firm sales in the year prior to the announcement of the minority equity
investment. BV/MV is the ratio of the book value of the firm’s assets to their market value
(measured as the sum of the market value of the firm’s equity plus the book value of its liabilities).
Leverage is the ratio of the total firm liabilities to total assets. Variance is the variance in daily
stock returns over the two years preceding the announcement of the minority equity investment.
Except for the stock returns which are obtained from CRSP tapes, all the other variables are taken
from the Compustat tapes in the year prior to the announcement. The number of observations
for each variable varies depending on the availability of data from the Compustat tapes.
is noteworthy that the average size of the selling firm sample ($252 million) is
similar to that of Wruck’s (1989) private equity sales sample ($234 million).
Purchasing firms’ R&D/sales ratio averages 8% with a median of 7% while
selling firms’ R&D/sales ratio is much higher, averaging 95% with a median of
17%. The t-statistic for the difference in the mean R&D/sales between the two
samples is significant (t-statistic = 1.80). The high average R&D/sales ratio for
the seller sample is driven mainly by four high-tech selling firms for which the
R&D/sales ratios exceed 100%. Even when comparing the purchasing and selling
samples using the median R&D/sales ratio, the selling firms are still at least twice
as R&D intensive as the purchasing firms.
Although both purchasing and selling firms exhibit growth opportunities
measured using the book-to-market value (BV/MV) ratio, the selling firms exhibit
a significantly lower mean BV/MV ratio than the purchasing firms (t-statistic of
difference in mean BV/MV = 2.67). (BV/MV is the ratio of the book value of
the firm’s assets to their market value, where market value is equal to the sum
of the market value of the firm’s equity plus the book value of its liabilities.) This
suggests that the selling firms in general exhibit higher growth opportunities than
the purchasing firms. The lower BV/MV ratio for the selling firms could also
suggest that the fraction of market value of the selling firms’ assets attributable to
tangible assets is lower than that for the purchasing firms.
The above observations about the differences in size, R&D intensity and growth
opportunities between the purchasing and selling firms are consistent with the
argument that the larger and relatively less R&D intensive firms with lower
growth prospects use minority equity alliances to gain access to the capabilities
of smaller, knowledge-intensive firms with attractive growth prospects. Although
both purchasing and selling firms exhibit similar levels of financial leverage,
the selling firms exhibit significantly higher stock return volatility (as measured
by the variance in daily stock returns) than their purchasing firm counterparts
(t-statistic of difference in mean variance = 3.09). The higher variance in the
selling firms’ stock returns could be largely due to the higher risk of their
operations.
We also examine the ownership concentration in our sample firms using the
percentage of insider ownership information from CDA/Spectrum. CDA/Spectrum
defines an insider as an officer, director or beneficial owner (holder of 10% or
more) of a company’s stock. We measure the total direct insider ownership as a
percentage of the number of shares outstanding for the quarter just prior to the
announcement of the equity alliance. The mean (median) insider holdings for the
purchasing firms and sellers are 14.6% (5%) and 17.5% (17.5%), respectively.
These data are consistent with evidence reported in other studies (Morck, Shleifer
& Vishny, 1988; Wruck, 1989) that ownership is more concentrated in smaller
firms. The ownership concentration, however, is not significantly different between

the purchasing firm and selling samples (t-statistic of difference = 0.98).
3.2. Outcomes of the Minority-Equity Relationships
The purchasing firms in our sample use a minority-equity buy-in as a means to enter
into a collaborative relationship with a selling firm. Hence we expect the outcomes
of this form of interfirm investment to differ from those buy-in situations where
the purchasing firm is considering a takeover or making a passive investment. In
the study of interfirm investment by Mikkelson and Ruback (1985) they report (see
their Table 1) that for those investments where the purchasing firm is considering
a takeover, about 49% of the selling firms were taken over by the purchasing firm
within three years of the initial buy-in while for passive investments only about
8% of the selling firms were taken over by the purchasing firm. Also for passive
investments, about 44% ended in a targeted repurchase or a sale of shares to a third
party.
It is also interesting to see how these relationships evolve when compared to
non-equity strategic alliances and joint ventures. In Chan et al.’s (1997) study of
non-equity strategic alliances, they find that in the eight year period following
the announcement of the alliances, only a small percentage (around 6%) of the
alliances resulted in early termination of the agreement. The majority of them
(about 83%) involve the signing of new agreements to strengthen or expand the
relationship. Only a small percentage of the alliances (around 6%) evolved into a
more permanent form of relationship, such as a joint venture or a merger. This is
in contrast to joint ventures in which roughly 80% end in a takeover by one of the
partners (see Bleeke & Ernst, 1995). Since minority equity alliances lie somewhere
in between non-equity alliances and joint ventures in terms of the integration of the
allied firms, we would expect to find the outcomes of the minority equity alliances
falling somewhere between the other two modes of interfirm relationships.
To track the evolution of the minority-equity alliances we examine the Wall Street
Journal (WSJ) index and news articles from the wire services for any related news
reports related to these investments. We track these news reports starting from the
date of the announcement of the minority-equity relationship and terminating three
full years after the announcement. Table 4 summarizes the results of this analysis.
We find a total of 41 outcome events reported for our sample firms in the WSJ
and newswire services within three years. More than 70% of them relates to the
strengthening or extension of the initial collorative agreement (categories A–E).12
There are nine instances where the purchasing firm increases stake in the sellers,
fifteen incidences of formation of new pact, two incidences where the partnering
Table 4. Outcomes of Interfirm Alliances Involving Minority Equity

Investment.
Outcome Within Three Years After the Initial Announcement Number of
Occurrences
Alliance strengthened or extended

A. Purchasing firm increases stake in the selling firm 9
B. Formation of new pact, expansion or strengthening of alliance 15
D. Formation of joint venture 2
E. Purchasing firm agrees to buy selling firm 4
30
Alliance reduced in scope or terminated
F. Involvement in a legal dispute 2
G. Purchasing firm reduces stake in the seller or loosen ties with selling firm 2
H. Purchasing firm sells its entire stake in the seller (targeted repurchase or 3
sale of shares to a third party)
I. Termination of partnership 5
11
Total number of announced outcomes within 3 years 41
Note: This table reports the outcomes of the interfirm alliances in our sample. The table only reports
the outcomes involving both purchasing and selling firms that are announced in either the
Dow Jones News Retrieval Service or the Lexis/Nexis database. The outcomes span the period
between the initial announcement of the alliance and the end of the third full year following the
initial announcement made during the 1981–1992 period.
firms decide to form a joint venture, and four instances (about 10% of the total
announced outcomes) where a purchasing firm agrees to acquire the selling firm.
(In two of the acquisition cases the purchasing firm agreed to acquire the selling
firm a year after purchasing a minority equity stake.) The later finding is in contrast
to that for interfirm investments with a takeover intention in which about 49% of
the selling firms were taken over (see Mikkelson & Ruback, 1985) or to that for
joint ventures in which roughly 80% end in a takeover by one of the partners (see
Bleeke & Ernst, 1995).
Table 4 also shows a total of eleven outcomes (about 27% of the total announced
outcomes) in which the alliance agreement was reduced in scope or terminated
(categories F–I). There are two cases in which the partnering firms were involved
in legal disputes over violation by one of the partners of some key elements of the
partnering agreement. There are two cases where the purchasing firm reduces its
stake in the seller and three cases (about 7% of total announced outcomes) where
a purchasing firm sells its entire equity interest in the selling firm either back to
the selling firm or to a third party.13 This contrasts with passive investments in
which about 44% ended in a targeted repurchase or a sale of shares to a third party
as reported by Mikkelson and Ruback (1985). Finally, there are five cases in the
sample where the alliance was dissolved because one of the partnering firms was
acquired by a third party.
In summary, the evidence in Table 4 indicates that for many firms the formation
of an alliance represents the beginning of a dynamic relationship between the
partnering firms. This evidence provides some limited support for Mody’s (1993)
conjecture that alliances can be used as a learning experiment. Changes in the
relationship may also be the result of adaptation to evolving market conditions as
well.
3.3. Methodology for Assessing Abnormal Stock Performance
A standard event study procedure is used to measure the immediate stock

price response associated with the announcement of minority-equity alliances.
Specifically, we use a methodology similar to Dodd and Warner (1983). Abnormal
performance for each firm is estimated using the market model and daily stock
returns obtained from the CRSP files. Defining the announcement day as day 0,
the estimation period for the market model estimate begins on day −170 and ends
with day −21. Abnormal stock returns are estimated as the prediction errors from
the market model. We calculate abnormal returns for the event days −20 through
+10 and average them across firms for each of the 31 event days. Significance
tests are based on a standardized test statistic constructed to determine whether
the mean abnormal return is significantly different than zero (see Dodd & Warner,
1983, for a detailed description of the test statistics and their calculation).
4. RESULTS
4.1. Gains to Purchasing and Selling Firms in a Minority-Equity Alliance
Table 5 presents the abnormal returns and the distribution of wealth gains between
firms participating in alliances involving a minority-equity commitment. For the
sample of 85 acquiring firms, 55% experience positive abnormal returns on the
announcement day; however, the average abnormal return is an insignificant
0.07%. The two-day results are similar. This evidence indicates that on average
the purchasing firms neither pay too much nor too little.14
For the sample of 29 selling firms, 76% of these firms experience positive
abnormal returns on the announcement day and the average abnormal return
Table 5. Average and Cumulative Abnormal Returns for Purchasing and

Selling Firms.
Statistics Purchasing Firms Selling Firms
(N = 85) (N = 29)
Pre-announcement period (−20, −1)

Cumulative average abnormal return 0.44% (0.04) 4.69% (1.13)
(Z-statistic)
Day 0 average abnormal return (Z-statistic 0.07% (0.56; 55%) 5.42% (7.84; 76%)*
and % positive)
Two-day (Day 0, +1) cumulative abnormal 0.17% (0.24; 54%) 8.15% (8.24; 83%)*
return (Z-statistic and % positive)
Post-announcement period (2, 10)
Cumulative average abnormal return 0.31% (1.03) −2.87% (1.86)∗∗
(Z-statistic)
Mean market value of equity ($M) 21,369 252
Mean change in wealth ($M) 119.36 20.04
Note: The table reports the announcement day as well as pre- and post-announcement returns for both
the purchasing and selling firms forming an alliance involving the purchase/sale of a minority
equity stake. The market value of equity as well as the dollar change in equity value is also
reported for both purchasing and selling firms. The market value of equity is the market value of
the firm’s common stock 21 trading days before the initial announcement of the equity alliance.
The mean dollar change in wealth is computed by first multiplying each participating firm’s
market value of equity by its two-day (0, +1) cumulative abnormal return and then averaging
the product across firms in the Purchasing firm or selling firm sample.
∗ Significant at the 5% level for a two-tailed test.
is 5.42% (Z-statistic = 7.84). The average abnormal return over the two-day
announcement period is a statistically significant 8.15% with a Z-statistic of 8.24.
This evidence indicates that significant value is created for the stockholders of the
selling firms when a minority equity buy-in to form an alliance is announced.
For both the purchasing and selling firms, we find that the cumulative
average abnormal return (CAAR) in the pre-announcement period (−20, −1) is
insignificant. In the post-announcement period (2, 10) the CAAR is insignificant
for the purchasing firms, but is a significant −2.87% for the selling firms. This
suggests that the market may have overreacted to the news of the minority equity
alliance for the selling firms when it was announced. However, considering that
the CAAR on days 0 and +1 is a significant 8.15%, the selling firms still garner a
net gain.
The finding that the gains accrue to selling firms but not to purchasing firms
is consistent with prior studies of mergers and acquisitions. The result is also in
line with findings for joint-ventures and non-equity strategic alliances in which the
smaller strategic partner (analogous to the selling firm) reaps a higher abnormal
return compared to the larger partner (analogous to the purchasing firm). Similarly,
in the subset of their sample that involved the announced formation of alliances
and joint ventures, Allen and Phillips (2000) find that the stock prices of selling
firms respond positively while purchasing firms do not. For their entire sample
of block equity purchases they find that the cumulative average excess stock
returns to sellers over a 21-day (−10, +10) interval is a significant 6.9%. The
response is greater (9.1%) for the subset of their sample where the equity purchase
is accompanied by the announced formation of a product market relationship in
the form of joint ventures or alliances. Similarly, we find that for the same 21-
day period the selling firms involved in minority equity alliances the cumulative
average excess stock returns is a significant 8.35%.
In their study of accumulations of 5% or more of another company’s shares
Mikkelson and Ruback (1985) find that, in general, the share prices of both the
acquiring and selling firms increase in response to the initial disclosure of the
investment position. In contrast, we find that where the motive of the equity buy-in
is to form an alliance, only the stock prices of the selling firms respond favorably.
It is possible that our inability to detect a significant abnormal return for the acquirer
is due to the large size differential between the purchasing and the selling firms in
our sample. Table 5 shows that purchasing firms are significantly larger (85 times
the market capitalization of) than the selling firms. To investigate this possibility
we investigate dollar returns corresponding to the announcement period abnormal
returns. To obtain a measure of the dollar change in wealth for the purchasing and
the selling firms, we first multiply each participating firm’s market value of equity
by its two-day cumulative abnormal return and then average the product across
firms in the sub-samples. We find that the mean dollar change in wealth (based on
two-day abnormal returns) for the sellers is $20.04 million while that experienced
by the purchasing firms is $119.36 million. These dollar returns suggest that on
average the formation of a minority equity alliance benefits both the purchasing
and selling firms.
The average excess stock returns for selling firms in our sample are higher
than that obtained by Mikkelson and Ruback (1985) in their study of interfirm
investment and by Wruck (1989) who examines private equity sales. Mikkelson
and Ruback report a two-day excess return of 3.24% while Wruck reports an excess
return of 4.5%.15 One possible explanation for this result is that the alliances
we study are expected to generate improvements in the utilization of corporate
resources or to mitigate information asymmetry problems such that they create
greater value than the sale of equity to a passive investor.
The purchasing firm results are different from those of Mikkelson and Ruback
(1985). They find statistically significant two-day abnormal returns (1.27%,
see their Table 4, p. 535) for purchasing firms while our sample produces an
insignificant two-day abnormal return (0.33%).16 This result may be due to
differences in sample composition for the two studies. Interfirm equity purchases
can also be made as a part of a passive investment strategy wherein the acquiring
firm does not intend to become involved in the business of the selling firm and only
hopes to reap its proportionate share of the selling firm’s returns. The Mikkelson
and Ruback sample may include alliances formed with both active and passive
investment motives whereas we study only those cases in which the acquirer
plans to engage in an active investment strategy involving the selling firm. It is
also noteworthy that the majority of selling firms in equity alliances are small
entreprenuerial firms, and consequently the relative size of the purchasing firm
compared to the seller may be much larger in our sample than in Mikkelson and
Ruback’s sample. Thus the lack of significance in the returns to the purchasing
firms in our study may simply be due to the relative size of the event’s valuation
impact relative to the size of the firm.
We test for the possibility of a wealth transfer between the purchasing and
selling firms by estimating the correlation between the two-day announcement
period abnormal returns for the subsample of 23 equity alliances for which both
the purchasing and selling firms are in our sample. If sellers were benefiting at the
expense of purchasing firms or vice versa, then we would expect this correlation
to be negative and significant. The resulting coefficient was −0.03 and was not
statistically different from zero. Consequently, we find no evidence of wealth
transfers between purchasing and selling firms.
4.2. Determinants of Value Effects
We conduct multiple regression tests to explain cross-sectional differences in

abnormal returns for purchasing and selling firms using variables related to
attributes that may lead to greater value creation. The explanatory variables include
the technology status of the firm (HITECH = 1 for firms in high tech industries
and 0 otherwise), R&D intensity (R&D/sales), growth opportunities as measured
by the ratio of book to market value equity (BV/MV), riskiness as measured by
the variance in daily stock returns over the two years prior to the formation of the
alliance (Variance), the proportionate control that changes hands in the transaction
as measured by the percentage of equity that the purchasing firm intends to purchase
from the selling firm (Stake), and the relatedness of the industry affiliation of
the partnering firms (Related Industry = 1 if the allied firms are from a similar
industry, and 0 otherwise). The rationales for including these explanatory variables
were discussed earlier in Section 2.
We include the logarithm of the market value of firm equity (on event day −21) in
the regression to control for the influence of firm size (Firm Size). We also construct
an indicator variable that equals 1 if the alliance agreement involves research
and/or development to test whether there is a difference in share price response
between alliances where the stated purpose of the agreement involves research
and/or development versus non-R&D types of agreements (such as marketing
agreements). Since we do not find the type of agreement to be a differentiating
factor we do not report this regression result.
Panels A and B of Table 6 present the estimated coefficients from the cross-
sectional regressions involving the two-day cumulative abnormal return for the
purchasing and selling samples, respectively. Panel A shows that the valuation
effect for the purchasing firms is positively correlated with their R&D intensity.
In other words, purchasing firms that are more R&D intensive seem to reap higher
returns from forming a minority equity alliance than their less R&D intensive
counterparts.
Panel B shows that the abnormal returns for the selling firms have significant
positive relationship with the Variance, Stake and Related Industry variables. A
positive coefficient for the Variance variable indicates that selling firms with more
volatile stock returns enjoy higher abnormal returns from forming the alliance. The
positive and significant coefficient for the Stake variable suggests that the larger the
minority stake purchased by the purchasing firm, the higher the abnormal returns
for the selling firms. This result supports the hypothesis that the larger the size of
the minority equity stake relative to the seller’s outstanding shares, the lower is the
risk of opportunism and consequently the greater the potential value that is created
by the alliance. Finally, selling firms that tie-up with purchasing firms that are in
a related industry are associated with significantly higher abnormal returns. This
result conforms to the hypothesis that purchasing firms from a related industry
are better able to evaluate the selling firm’s assets-in-place and its growth options
thus mitigating the information asymmetry problems regarding the selling firms.
Although the BV/MV variable has a negative coefficient (suggesting that firms
with higher (lower) growth opportunities tend to get a higher (lower) payback
from forming an alliance), it is not significant.
4.3. Operating Performance of Firms that Form Equity Alliances
In this section we examine the operating performance of both purchasing and

selling firms that engage in minority equity alliances. We compare the relative
operating performance of these firms to their peers in the two years preceding
and three years following the announcement of the formation of an alliance to
Minority Equity Investments and Inter-Firm Collaborations
Table 6. Determinants of Abnormal Returns for Purchasing and Selling Firms Engaged in Alliances Involving
Minority-Equity Investments in the 1981–1992 Period.
Dependent Intercept Firm HITECH R&D BV/MV Variance STAKE Related R2 Adjusted R2 F-value
Variable Size Intensity Industry
Panel A: Results for publicly-traded purchasing firms (N = 46)

CAR(0,1) −0.058 0.003 −0.012 0.262 0.011 −2.54 0.001 −0.001 17.8% 2.6% 1.17
(−1.64)** (1.306) (−1.104) (1.87)** (0.76) (−1.57) (1.07) (−0.136)
Panel B: Results for publicly-traded selling firms (N = 21)
CAR(0,1) −0.046 0.001 −0.002 −0.100 −1.102 49.995 0.006 0.062 68.5% 51.48% 4.03
(−0.60) (0.06) (−0.063) (−1.74) (−1.72) (3.52)* (3.09)* (2.42)*
Note: This table reports the cross-sectional regression results obtained by regressing the two-day (0, +1) cumulative abnormal return for investing (and
sellers) firms on selected variables representing firm size, technology status of firm (HITECH), R&D intensity, growth opportunities (BV/MV),
riskiness of firm (variance), percentage of equity stake in the selling firm (STAKE), and industry relatedness (RELATED INDUSTRY). Firm
size is measured by the logarithm of the market value of equity 21 days prior to the announcement of the alliance. Hitech is a dummy variable
that takes a value of one if the firm is classified as a high tech firm. R&D intensity is the ratio of the firm’s R&D expenditures to its net
sales. BV/MV measures a firm’s growth options and is computed as the ratio of book value of the firm’s assets to the market value of the
firm’s assets (measured as the sum of the market value of the firm’s equity plus the book value of its liabilities). Variance, a measure of the
riskiness of the firm, is the variance in daily stock returns (from CRSP tapes) over the two years preceding the announcement of the minority
equity investment. Stake is the percentage of equity of the selling firm reported in the news article that the investing firm says it will purchase.
RELATED INDUSTRY is a dummy variable that takes a value of one if the investing and selling firm are from the same 3-digit SIC industry.
R&D intensity and BV/MV are computed using data from the Compustat tapes the year prior to the announcement of the formation of the
equity alliance. There are 46 observations from the investing firm sample and 21 observations from the seller sample that have complete data
on all the variables used in the regression.
∗∗ Significant at the 10% level for a two-tailed test.
37
gain insight into the circumstances under which managers choose to engage in
such alliances. There are two issues addressed in this analysis. First, do firms that
enter into alliances do so because they are performing poorly or simply because
the purchasing firms have the cash flow to invest? Mohanram and Nanda (1996)
observe that firms entering into joint ventures tend to be experiencing deteriorating
performance. Second, does the operating performance of alliance partners improve
or deteriorate following the equity tie-up?
Data on operating performance spanning the six-year period surrounding (two
years before to three years after) the announcement of the minority equity alliance
is obtained from the COMPUSTAT. We measure operating performance using
operating cash flow return on assets (OCF-ROA) which is computed by dividing the
earnings before interest, taxes, depreciation, amortization, and extraordinary items
(COMPUSTAT mnemonic EBITDA) by the book value of total assets. To adjust for
industry influences we calculate industry-adjusted performance measures as the
difference between the announcing firm value in a given year and the median value
for all firms on COMPUSTAT with the same four-digit SIC code.17 The number
of observations in each year varies depending on the availability of COMPUSTAT
data.
Panels A and B of Table 7 report the operating performance results for the
set of purchasing and selling firms. These results indicate that firms that acquire
equity shares in an alliance partner, on average, have higher operating cash flow
to assets than their industry peers over the entire six year period beginning two
years before the announcement of the formation of an alliance. This suggests that
purchasing firms that engage in minority equity alliances tend to be cash rich
firms and therefore their participation in such alliances is not motivated by poor
performance. Furthermore, Panel A shows that when firm performance relative
to its industry peers is analyzed for three years following the formation of the
alliance, the purchasing firm’s relative cash flow position is reduced significantly
(although it is still above the industry) when compared to its cash flow position in
the year preceding the equity participation. This evidence suggests that purchase
of a minority equity stake is a way for the cash-rich firms to spend their cash.
Selling firms, on the other hand, show negative operating cash flow-to-asset
position in most of the years surrounding the announcement of the minority equity
alliance. Furthermore, the selling firms perform significantly worse than their
industry peers in the two years leading up to the equity tie-up announcement.
Unlike the purchasing firms, however, the operating cash flow-to-asset position of
the sellers show significant improvement in the third year following the equity tie-
up when compared to the year preceding the alliance announcement. The above
results suggest that poor performance may be a motivation for selling firms to
enter into interfirm collaborations involving minority equity participation by the
Minority Equity Investments and Inter-Firm Collaborations
Table 7. Operating Performance of Purchasing and Selling Firms in Minority Equity Alliances Surrounding Minority
Equity Purchase Announcements.
Year Relative to Minority Equity Purchase Announcement
−2 −1 0 1 2 3
Panel A: Operating performance of purchasing firms

Number of observations 77 82 82 83 82 81
Average of firm-level OCF-ROA (%) 18.7* 18.0* 17.8* 16.6* 15.8* 15.6*
Industry-adjusted OCF-ROA
Average (%) 8.0* 7.8* 8.0* 7.5* 7.7* 6.9*
% Positive 83 80 82 82 83 81
Change in industry-adjusted OCF-ROA Year −2 vs. Year −1: −0.40% Year −1 vs. Year +3: −1.2%**
Panel B: Operating performance of selling firms
Number of observations 26 29 27 24 23 20
Average of firm-level OCF-ROA −5.0 −4.0 −9.6 3.6 −1.5 3.6
Industry-adjusted OCF-ROA
Average (%) −12.1* −9.7* −16.2 −3.0 −8.2 −2.7
% Positive 46 45 41 46 43 50
Change in industry-adjusted OCF-ROA Year −2 vs. Year −1: 2.20% Year −1 vs. Year +3: 5.7%**
Note: Performance is measured using operating cash flow return on asset (OCF-ROA). OCF-ROA is computed using the earnings before interest, taxes,
depreciation, amortization, and extraordinary items (EBITDA) divided by total assets. The industry-adjusted performance measure equals the
difference between the announcing firm’s measure in a given year and the median value for all firms on COMPUSTAT with the same four-digit
SIC code. The number of observations in each year varies depending on the availability of COMPUSTAT data.
∗∗ Significant at the 10% level for a two-tailed test.
39
partnering firm. In addition, the alliance and equity stake seem to help improve
the operating performance of the selling firms.
In summary, the above results suggest that, unlike joint venture agreements,
poor performance is not a motivation for purchasing firms to enter into interfirm
collaborative agreements involving a minority equity stake. For selling firms,
however, poor operating performance may be a motivation for them to enter
into minority equity alliances. In addition, the post-performance results for both
the purchasing firms and selling firms seem to be in line with the evidence we
present on the market’s response to the announcement of minority equity tie-ups
by purchasing and target firms.
5. CONCLUSIONS
Minority-equity investment in a strategic partner’s common stock is frequently
used to finance an alliance and to bond the partnering firms to the agreement.
The partner that sells the equity stake gets an immediate cash infusion, while the
purchasing partner gets a percentage of the seller’s profits plus access to the seller’s
skills and technology as prescribed in the alliance agreement.
Based on 93 announcements of companies making minority-equity
commitments in strategic partners who indicate no intention to seek control of
the partnering firm, we find that the average stock price response is neutral for
the acquiring firms, but positive and significant for the selling firms. Further,
we find no evidence that the gains to the selling firms are a result of a wealth
transfer from the purchasing firms. The finding of significant positive wealth gain
for the selling firms is consistent with Mikkelson and Ruback’s (1985) study of
interfirm equity investments, Wruck’s (1989) study of private equity sales, and
Allen and Phillips’ (2000) study of corporate equity ownership associated with
product market relationships. Our cross-sectional analysis reveals that the two-
day announcement abnormal returns for the selling firms are positively related to
the amount of equity invested by the purchasing firms, their riskiness as reflected in
the variability of their stock returns, and the relatedness of the industry to which the
purchasing firm belongs. The analysis also reveals that purchasing firms that are
more R&D intensive seem to reap higher returns from a minority equity alliance
than their less R&D intensive counterparts.
The evidence from this study suggests that the pooling of resources in a minority
equity alliance benefits the selling firms. The change in the selling firm’s value
could be derived from enhanced financing flexibility, a lowering of transactions
costs, and a more efficient use of corporate resources under the minority equity
arrangement. In addition, the purchase of minority equity stake by a sophisticated
purchasing firm solely for the purpose of forming a strategic alliance can constitute
an endorsement of the selling firm’s management and future prospects. Therefore,
the positive valuation observed for the selling firm’s stock could also be due to
this revelation of new information to the market concerning the seller’s future
prospects.
Within three years of the announced formation of an alliance the majority of the
agreements continue unchanged. Of those that do change, the largest number are
expanded or strengthened. The result is an increase in the equity stake, formation
of a joint venture or merger of the partnering firms. A very small percentage of
the alliances terminate within three years of their formation. These results provide
evidence that firms engaging in minority equity alliances often take advantage of
the flexibility afforded in alliances to alter their relationships, possibly to adapt to
evolving market conditions.
Finally, firms that acquire minority stakes in alliance partner firms tend to
have better operating performance than their peers during the period two years
before through three years following the announcement of the agreement. The
cash flow position of the purchasing firms, however, deteriorates somewhat in
the three years following the announcement, compared with their own prior cash
flow position (although still better than industry peers). Selling firms, on the other
hand, tend to perform worse than their industry peers in the two years before
the announcement of their engaging in the formation of an equity alliance. Their
performance in the three years following the announcement, however, improves
significantly when compared to their pre-alliance performance. This evidence on
the selling firms differs from that observed for non-equity alliances and suggests
that poor performance may be a motivator for selling firms to participate in alliances
that involves equity participation by a cash rich partner.
NOTES
1. See Chan et al. (1997) for an analysis of non-equity strategic alliances.

2. Mikkelson and Ruback (1985) and Choi (1991) observe positive stock market
responses to the announcement of interfirm equity investments, while Wruck (1989) and
Hertzel and Smith (1993) observe positive abnormal returns to announcement of private
sales of equity mainly to institutions. Wruck notes that the change in firm value depends on
the level of ownership concentration after the sale of equity and the purchaser’s current or
anticipated future relationship with the firm. Also Bethel et al. (1998) show the effectiveness
of activist block owners in helping restructure selling firms and in improving their operating
performance. Note that although these studies distinguish between toehold investments (in
anticipation of a takeover) and all other investment motives, they do not delineate between
passive investment motives and the strategic (or collaborative) motives that we examine in
this paper.
3. Allen and Phillips (2000) examine a sub-sample of block equity purchases that are
accompanied by alliances and joint ventures formed within a two-year interval either prior
to or subsequent to the purchase.
4. Morck et al. (1988) find that smaller firms find information asymmetry problems to
be particularly acute, which would make minority equity alliances particularly attractive
for these firms. For more about public versus private equity offerings, see Wu (2000).
5. For more discussion on this issue, see Parkhe (1993) and Chan et al. (1997).
6. See Pisano (1989) for a discussion of the role of equity financing in reducing the
incentives for opportunism by partners in an equity alliance. Williamson (1985, p. 191)
describes a double hostage system for bonding an alliance “. . . reciprocity involves the sale
of specialized product to B conditioned on the procurement of specialized product from
B. The argument here is that reciprocity can serve to equalize the exposure of the parties,
thereby reducing the incentive of the purchaser or seller to defect from the exchange –
leaving the supplier to redeploy specialized assets at greatly reduced value.”
7. Madden (1981) found that the Wall Street Journal date preceded the Insider’s
Chronicle date (that reports Schedule 13D filings) by an average of 37 days. The average
delay between the filing date of the Schedule 13-D and the date of the Insider’s Chronicle
was 23 days.
8. Minority equity purchases have been used extensively to forge alliances between large
publicly-held firms and smaller privately-held firms and also between domestic U.S. and
foreign firms. Therefore, we include publicly-traded firms (with returns data available on
the CRSP) regardless of whether they form minority equity alliances with a publicly traded
firm, a private firm or a foreign firm.
9. The number of purchasing (selling) firms in the final sample excludes four (two)
publicly-traded foreign firms that do not have stock return data in the CRSP file.
10. For example, in 1989, the Ascom Group of Switzerland (the largest communications
and services automation equipment company in Switzerland) made a $6 million direct
investment in New York based Comverse Technology Inc. (one of the leading international
suppliers of computerized message management systems) for 17 million shares, giving it
an ownership position of 13.7% (Business Wire, August 17, 1989). The investment was
made as part of a multi-faceted strategic alliance formed by the two companies to distribute
message management systems to organizations that provide subscriber services in selected
markets of Western Europe. Under the pact, Ascom will also market Comverse’s defense
and related communications processing systems in Switzerland, West Germany and other
markets.
11. For those firms that are not publicly-traded, we read the news articles describing the
minority equity alliance to find out the industry in which those firms are operating.
12. Since we rely on press releases for information we recognize the possibility that
firms may simply not announce changes in the status of their alliances following the initial
announcement.
13. For example, in October, 1993 IBM sold its entire equity interest in Hogan Systems
Inc. back to Hogan in a private market transaction. IBM has bought approximately 5% of
Hogan’s system’s stock since September 1990 to strengthen a strategic partnership that has
existed between the two companies since 1986. The sale was profitable for IBM due to the
strong increase in price of Hogan’s stock. For Hogan, the sale was potentially profitable
because it hopes to make a better return from the repurchase of those shares than from
holding cash. A Hogan senior vice president said it is not clear yet if Hogan and IBM will
extend the current exclusive marketing agreement after it expires in 1995 (Business Wire,
10/4/93).
14. We also compare the excess returns for firms that purchase equity stakes in private
firms with those that purchase equity stakes in public firms and find no significant difference.
15. The results for the Mikkelson and Ruback (1985) study reflect the abnormal returns
for a sample of 106 firms that sold shares to other firms who stated that their purchase
objective was for investment purposes. Similar results for selling firms that were the target
of a takeover attempt were much higher (7.74%).
16. As in Mikkelson and Ruback, we also evaluated the abnormal returns of the
purchasing firms based on whether they are frequent (appearing in our sample six or more
times) or infrequent purchasers of minority equity. As for the full sample, the abnormal
returns for both sub-samples are insignificant.
17. Allen and Phillips (2000) use a similar technique to examine operating income
changes surrounding corporate equity purchases. See also Barber and Lyon (1996).
ACKNOWLEDGMENTS
We thank Marshall Bloom, Andrew Chen, David Denis, Diane Denis, Ning Gong,
Mike Impson, Paul Laux, Ray Miles, Ko Wang as well as participants at the
Texas Finance Symposium and the Financial Management Association Meeting
for comments on previous versions of this paper. We are also grateful to Yajat
Bindal and Jaafar Salwani for their research assistance.
REFERENCES
Allen, J. W., & Phillips, G. M. (2000). Corporate equity ownership and product market relationships.
Journal of Finance, 25, 2791–2815.
Balakrishnan, S., & Koza, M. P. (1993). Information asymmetry, adverse selection and joint ventures.
Journal of Economic Behavior and Organization, 20, 99–117.
Barber, B., & Lyon, J. (1996). Detecting abnormal operating performance: The empirical power and
specification of test-statistics. Journal of Financial Economics, 3, 359–400.
Bethel, J. E., Liebeskind, J. P., & Opler, T. (1998). Block share purchases and corporate performance.
Chan, S. H., Keown, A. J., Kensinger, J. W., & Martin, J. D. (1997). Do strategic alliances create value?
Journal of Financial Economics, 46, 199–221.
Chan, S. H., Martin, J. D., & Kensinger, J. W. (1990). Corporate research and development expenditures
and share value. Journal of Financial Economics, 26, 255–276.
Choi, D. (1991). Toehold acquisitions, shareholder wealth, and the market for corporate control. Journal
of Financial and Quantitative Analysis, 26, 391–407.
Dodd, P., & Warner, J. B. (1983). On corporate governance. Journal of Financial Economics, 11,
401–438.
Harrigan, K. R. (1985). Strategies for joint ventures. Lexington, MA: Lexington Books.
Hertzel, M., & Smith, R. L. (1993). Market discounts and shareholder gains for placing equity privately.
Jensen, M. C., & Meckling, W. (1991). Specific and general knowledge, and organizational structure.
In: L. Werin & H. Wijkander (Eds), Main Currents in Contract Economics. Oxford: Blackwell.
Johnson, S., & Houston, M. B. (2000). A reexamination of the motives and gains in joint ventures.
Journal of Financial and Quantitative Analysis, 35, 67–85.
Kensinger, J. W., & Martin, J. D. (1991). Financing network organizations. Journal of Applied
Corporate Finance, 4, 66–76.
Klein, B., Crawford, R., & Alchian, A. (1978). Vertical integration, appropriable rents, and the
competitive contracting process. Journal of Law and Economics, 21, 297–326.
Madden, G. P. (1981). Potential corporate takeovers and market efficiency: A note. Journal of Finance,
36, 1191–1197.
McConnell, J., & Nantell, T. (1985). Common stock returns and corporate combinations: The case of
joint ventures. Journal of Finance, 40, 519–536.
Mikkelson, W. H., & Ruback, R. S. (1985). An empirical analysis of the interfirm equity investment
process. Journal of Financial Economics, 523–553.
Mody, A. (1993). Learning through alliances. Journal of Economic Behavior and Organization, 20,
151–170.
Mohanram, P., & Nanda, A. (1996). When do joint ventures create value? Academy of Management
Proceedings, 36–40.
Morck, R., Shleifer, A., & Vishney, R. (1988). Management ownership and market valuation: An
empirical analysis. Journal of Financial Economics, 20, 293–315.
Murray, A. I., & Siehl, C. (1989). Joint venture and other alliances: Creating a successful cooperative
linkage. Morristown, NJ: Financial Executives Institute.
Parkhe, A. (1993). Strategic alliance structuring: A game theoretic and transaction cost examination
of interfirm cooperation. Academy of Management Journal, 36, 794–829.
Pisano, G. (1989). Using equity participation to support exchange: Evidence from the biotechnology
industry. Journal of Law, Economics, and Organization, 5, 109–126.
Williamson, O. (1985). The economic institutions of capitalism. New York, NY: Free Press.
Wruck, K. H. (1989). Equity ownership concentration and firm value: Evidence from private equity
financings. Journal of Financial Economics, 23, 3–28.
Wu, Y. (2000). The choice between public and private equity offerings. Unpublished Working Paper,
University of Chicago.
SIZE AND BOOK-TO-MARKET
EFFECTS IN THE RETURNS
ON INFORMATION TECHNOLOGY
STOCKS
Quang-Ngoc Nguyen, Thomas A. Fetherston

and Jonathan A. Batten
ABSTRACT
This paper explores the relationship between size, book-to-market, beta, and
expected stock returns in the U.S. Information Technology sector over the
July 1990–June 2001 period. Two models, the multivariate model and the
three-factor model, are employed to test these relationships. The risk-return
tests confirm the relationship between size, book-to-market, beta and stock
returns in IT stocks is different from that in other non-financial stocks.
However, the sub-period results (the periods before and after the technology
crash in April 2000) show that the nature of the relationship between stock
returns, size, book-to-market, and market factors, or the magnitude of the
size, book-to-market, and market premiums, is on average unchanged for
both sub-periods. This result suggests the technology stock crash in April
2000 was not a correction of stock prices.
Research in Finance
ISSN: 0196-3821/doi:10.1016/S0196-3821(04)21003-6
45
46 QUANG-NGOC NGUYEN ET AL.
1. INTRODUCTION
One of the most important debates among financial economists, as well as financial
practitioners, in recent years concerns the valuation of information technology
(IT) stocks. The valuation of IT stocks is not an easy task as many of these firms
do not have, or have only a short history of stock performance. Even for those
few firms that do have a long history of performance, the prediction of future
earnings is extremely difficult due to unique features of the sector. Future earnings
of firms in the IT sector depend on the opportunity, the company, the management,
or the future outcome of the research and development activity, which are as
difficult to estimate as predicting future earnings themselves (Trueman, Wong
& Zhang, 2001). Empirical studies investigating IT stock valuation have used
options (Franklin, 2000) or web traffic (for internet stocks, Rajgopal, Kotha &
Venkatachalam, 2000), while Sadorsky (2003) finds that the conditional volatilities
of oil prices, the term premium, and the consumer price index each have a signifi-
cant impact on the conditional volatility of technology stock prices. An implicit fea-
ture of these studies is the assumption that IT stocks are “special” and consequently
traditional valuation models – such as the discounted cash flow model – may argued
to be inappropriate. However, no single paper has empirically explained why IT
stocks are special, except for those that cite the abnormal price-earnings ratios
(P/E) associated with these stocks as evidence for their “specialty” (Shiller, 2000).
The objective of this paper is to investigate why IT stocks are special by utilizing
a comparative perspective which tests the relation between size, book-to-market
equity, beta, and stock returns and then compares the magnitude of this relationship
with that found in earlier studies (Fama & French, 1992, 1993). The paper also
tests the difference in the risk-return relationship for IT stocks before and after
the technology stock crash in April 2000. Investigating IT stocks in a comparative
perspective can lead to evidence that would help financial academics as well as
practitioners differentiate IT stocks from other stocks, as well as providing further
insight into questions concerning valuation.
There is one obvious benefit arising from the distinction between IT stocks and
“old economy” stocks.1 It is logical that a new valuation model for IT stocks should
be developed if IT stocks are radically different from “old economy’ stocks. It is
equally reasonable a priori that the extant discounted cash flows model should
be sufficient to produce accurate measurements of IT stock prices, if there is
no distinction between IT stocks and “old economy” stocks. In the case of no
distinction, looking for a new model for the valuation of IT stocks would be
pointless.
The evidence on the relation between size, book-to-market equity, beta, and
stock returns in the IT sector is investigated in this paper using two alternative
Size and Book-to-Market Effects in the Returns on Information Technology Stocks 47
risk-return testing models. In the first model, or the multivariate model, direct
measurements of firm size, book-to-market ratio, and market betas are applied. The
model is then used in the Fama and MacBeth (1973) cross-sectional regressions
to test the significance of the coefficients on the size, book-to-market, and
beta variables. This is the simplest approach and was widely used prior to the
introduction of the three-factor model by Fama and French (1993). In the three-
factor model, firm size, book-to-market ratio, and market beta are not applied
directly. In contrast, these variables are represented by the returns on the three
portfolios formed in a way that closely mimics firm size, book-to-market ratio,
and market beta. The time-series regressions are applied to the three-factor model.
To address any biases introduced by the impact of the technology crash in April
2000 on the behavior of stock prices, this analysis is carefully divided into two
sub-periods: one before the crash and one after the crash. Thus the paper provides
an insight into how IT stocks behave in the lead-up to the crash, after the crash as
well as how they behave over the whole sample period, from July 1990 to June
2001.
The paper improves the regression approaches used by Fama and French (1992,
1993). This paper employs the Fama and MacBeth (1973) procedure to estimate
the coefficient slopes on the size, book-to-market, and beta variables. Another
improvement, is that the paper uses the Seemingly Unrelated Regressions (SUR)
methodology to estimate the coefficients for the three-factor model. This procedure
runs time-series regressions for all portfolios as a whole, and gives estimators of the
coefficients after adjusting for cross-correlation between the portfolios’ residuals.
Individually the results are robust. However the two models, the multivariate
model and the three-factor model, produce different results. In tests using the
multivariate model, the book-to-market ratio is the only significant factor that
explains variation in stock returns. However, the book-to-market effect is only
significant at the 10% level, in contrast to Fama and French (1992) who find that
the effect is very robust (at the 5% level) for all non-financial stocks listed on the
NYSE, AMEX, and NASDAQ over the 1963–1991 period. Another difference is
that this paper finds no relationship between size and stock returns. Nevertheless,
the paper does discover that beta has weak explanatory power, which is similar to
their earlier finding.
In tests using the three-factor model, all the returns on the portfolios that are
designed to mimic the size, book-to-market, and market (beta) factors are strongly
related to returns on size – book-to-market sorted portfolios. This evidence is in
favor of the three-factor model. The different results produced by the multivariate
model and the three-factor model suggest that the two models might not be
comparable. The excess returns on the size – book-to-market portfolios and the
market portfolios before the technology crash in April 2000 (the pre-crash excess
returns) are negative while the excess returns on those portfolios after the crash (the
post-crash excess returns) are indistinguishable from zero. However, this paper
finds that the relation between size, book-to-market, beta, and portfolio returns
(using the three-factor model) is similar for the periods before and after the crash.
The remainder of this paper is organized as follows. Section 2 describes the
differing empirical evidence on the risk-return relation as well as summarizes
alternative explanations for the size and book-to-market anomaly. Specific
hypotheses are then proposed. The data, its source, and the measurements of
variables for the risk-return tests are presented in Section 3. Section 4 discusses
the multivariate model and the three-factor model as well as the reasons for their
uses. Section 5 provides the regression results. Section 6 concludes the paper and
suggests the direction for future research.
2. LITERATURE REVIEW
2.1. Background
Sharpe (1964) and Lintner (1965) suggest a positive linear relation between
individual security risk and its expected return. This relationship is expressed
by the well-known Capital Asset Pricing Model (CAPM) where:
E(R i ) = E(R f ) + [E(R m ) − E(R f )]␤i (1)
The symbols in Eq. (1) are defined as follows: E(R i ) is the expected return on
security i for the period and is equal to the change in the price of the security, plus
any dividends, interest, or other distributions, divided by the price of the security
at the start of the period; E(R m ) is the expected return on the market portfolio of all
securities taken together; E(R f ) is the return on the riskless security for the period;
␤i is the contribution of security i to the risk of the market portfolio and is often
called the “systematic risk” of security i. is defined algebraically by:
Cov(R i , R m )
␤i = (2)
Var(R m )
Nevertheless, the applicability of the CAPM has been questioned with a
number of studies in the asset-pricing literature demonstrating a weak relationship
between beta and stock returns. Many findings show that market ␤s of stocks
are not sufficient to explain stocks returns (e.g. Banz, 1981), or more bluntly
the relationship between market ␤s and average stock returns is flat (e.g. Fama
& French, 1992). Additional factors, such as firm size, book-to-market equity,
have been introduced to accommodate actual stock return behavior. The failure
of the CAPM to explain empirical evidence, or the deviation from the CAPM is
considered as an “anomaly” in the risk-return relationship and is attributed to the
following two anomalies.
2.1.1. Firm Size Anomaly

Banz (1981) discovered the firm size anomaly in a sample of common stocks
listed on the NYSE between 1926 and 1975. After forming stocks into 25 size-
beta portfolios containing similar numbers of securities he ran month-by-month
cross-sectional regressions of portfolio returns on their betas and market values
and found that in addition to portfolio ␤, the market proportion of a security was
statistically significantly related to the security’s expected return. This relationship
is negative: stocks of firms with large market values have smaller returns than stocks
of small firms of equal market betas.
The negative relationship between firm size and stock return was confirmed by
Keim (1983) who controlled for possible bias in ␤ estimates, using three different
estimators of beta: the OLS betas, Scholes-William (1977) betas and Dimson
(1979) betas. Fama and French (1992) also demonstrated a negative relationship
between stock returns and firm size. The robustness of the size effect and the
absence of a relationship between ␤s and average returns was also true for the
extended 1941–1990 period (Fama & French, 1992). Interestingly, Fama and
French claim that if Black, Jensen and Scholes (1972) and Fama and MacBeth
(1973) had controlled for the size effect, the relationship between ␤s and average
returns would have disappeared.
2.1.2. Book-to-Market Anomaly

Along with the size effect, Fama and French (1992) also find that book-to-market
equity (ratio of book equity (BE) to market equity (ME)) captures variation in
the cross-section of average returns. Firms with high ratios of book value to the
market value of common equity have higher average returns than firms with low
book-to-market ratios; a result consistent with earlier findings Rosenberg, Reid and
Lanstein (1985). Barber and Lyon (1997a) supplement Fama and French’s (1992)
with supporting empirical evidence for stocks in the financial sector. Empirical
evidence for the extended period, the 1940–1962 period, also reveals a strong
relationship between book-to-market ratio and stock returns (Davis, 1994).
2.2. Explanations for These Anomalies
The apparent violations of the CAPM have inspired financial academics to find
plausible explanations for these violations. The reasons for the failure of the CAPM
could be classified into three categories: risk-based explanations, psychology-

based explanations, and characteristics-based explanations.
2.2.1. Risk-Based Explanations

The risk-based category generally focuses on the possible errors in tests of the
CAPM, or the missing risk factors that are not captured by market betas.According
to Fama and French (1992), one of the possibilities for the failure of ␤ in predicting
average returns is that other explanatory variables (i.e. size, book-to-market equity,
leverage) are correlated with true ␤s, and this obscures the relationship between
average returns and measured ␤s. Another is that there is a positive relationship
between ␤ and average return, but the relationship is obscured by noise in the ␤
estimates. However, Chan and Chen (1988) argue that the observed size effect is
just a risk effect in disguise. They posit firm size is a good instrument for risk,
therefore if ␤s are measured in such a way that captures the size variable, ␤ itself
is sufficient to explain average returns. In other words, Chan and Chen believe the
SLB model works if ␤s are estimated more precisely. Chan and Chen propose that
the size-sorted portfolios’ ␤s follow a stationary distribution:
␤˜ it = ␤¯ i + ␪˜ t (␤¯ i − ␤)
¯ + ␩˜ it (3)
where ␤¯ i is the time-series mean of security i’s betas; ␤¯ is the cross-sectional
mean of ␤¯ i s; ␪˜ t has zero mean and ␩˜ it is a random noise independent of all
other parameters. With ␤s being estimated under this stochastic assumption, Chan
and Chen find the size premium, in tests including ␤s and firm size variables,
is not statistically different from zero. In contrast, they observe a strong positive
relationship between their adjusted ␤s and average returns. The data they use are
for the 1954–1983 period.
This argument was rejected by Jegadeesh (1992), who shows that Chan and
Chen’s (1988) result is obscured by a high correlation between ␤s and the size
factors. Thus, it is not surprising that their ␤s explain most of the cross-sectional
differences in average returns even if the average returns are related to firm size.
When the test portfolios are constructed so that the market betas have low cross-
sectional correlation with firm size (i.e. ln(size)), Jegadeesh (1992) finds the size
effect still exists.
Another approach argues that the relationship between size and average return
proxies for a more fundamental relationship between expected returns and
economic risk factors. The empirical evidence in Fama and French’s (1995) paper
further supports the risk-based explanations for the size and book-to-market effects.
Fama and French (1995) find that BE/ME is related to earnings. Low BE/ME is
typical of firms with high average returns on capital, whereas high BE/ME is
typical of firms that are relatively distressed. Size is also related to profitability.
Controlling for BE/ME, small stocks tend to have lower earnings on book equity
than do big stocks.
A third risk based explanation concerns the presence of “data-snooping
statistics” as used by Aldous (1989, p. 252) to describe the situation “where you
have a family of test statistics T(a) whose null distribution is known for fixed a,
but where you use the test statistic T = T(a) for some a chosen using the data.”
Lo and MacKinlay (1990b) argue that if the choice of a is based on the data, then
the sampling distribution of the resulting test statistic is generally not the same as
the null distribution with a fixed a; hence, the actual size of the test may differ
substantially from its nominal value under the null.
Other possible biases related to the data-snooping bias are also suggested
by many financial academics. Black (1993) and MacKinlay (1995) say that the
deviations from the CAPM are sample-specific results that are unlikely to be
observed in various sets of data. However, other studies do not support the data-
snooping bias, sample-specific bias, and the selection bias explanations for the
deviations from the CAPM. Chan, Hamao and Lakonishok (1991), Capaul, Rowley
and Sharpe (1993), and Fama and French (1998) document strong relationships
between average return and BE/ME in markets outside the U.S. Davis (1994) finds
that the relationship between average return and BE/ME observed in recent U.S.
returns extends back to 1941. Even the relationship exists for data back to 1926
with the pre-1963 value premium is close to that observed for the subsequent period
in earlier work.
2.2.2. Psychology-Based Explanations

This explanation proposes that the value premium is due to investor overreaction
to firm performance. Investors overreact to performance and assign irrationally
low values to weak firms and irrationally high values to strong firms. When the
overreaction is corrected, weak firms have high stock returns and strong firms
have low returns. Proponents of this view include DeBondt and Thaler (1987),
Lakonishok, Shleifer and Vishny (1994), and Haugen (1995). For instance,
according to Lakonishok, Shleifer and Vishny (1994), investors extrapolate past
performance too far into the future: stocks that have performed poorly in the
past (“value” stocks) are expected by investors to continue to perform poorly
in the future, and vice versa, stocks that have done well in the past (“growth”
or “glamour” stocks) are expected by investors to continue to do well in the
future. This leads to “value” stocks being underpriced and “glamour” stocks
being overpriced. Lakonishok, Sheleifer and Vishny (1994) conjecture that
those investors who employ value strategies, i.e. buying “value” stocks and
selling “glamour” stocks, could earn superior returns given the same level of risk
involved.
2.2.3. Characteristics-Based Explanations

Proponents of the characteristics-based explanation suggest that it is the
characteristics (i.e. high or low book-to-market firms, small or big firms) and not
risk that determines stock returns (Daniel & Titman, 1997). These authors argue
that not all high book-to-market firms are in financial distress, there are many firms
that are not in financial distress but still have high ratios of book-equity to market-
equity, such as firms in related lines of businesses, in the same industries, or from
the same region. Daniel and Titman (1997) conjecture that if these characteristics
drive expected returns, there should be firms with characteristics that do not match
their risk loadings (on SMB and HML in the three-factor model). For example,
there should be some strong firms in distressed industries. In the characteristics
model, these firms have low returns because they are strong. But they can have high
loadings on a distress risk factor if the factor is in part due to covariance of returns
within industries. Thus, the returns on these firms will be too low, given their risk
loadings. Conversely, there are distressed firms in strong industries. Because they
are distressed, they have high returns, but in terms of risk loadings they look like
strong firms. If characteristics drive prices, their returns will be too high given their
risk loadings. In short, the characteristics hypothesis says that the characteristics
(high BM versus low BM) drive stock returns. Low BM produces low stock returns,
irrespective of risk loadings. Similarly, high BE/ME stocks have high returns,
regardless of risk loadings.
Davis, Fama and French (2000) reexamine Daniel and Titman’s (1997) findings
with a similar approach but for a larger sample period and find that the results
from the characteristics model is unique to the short sample period of 1973–1993.
The characteristics model is not able to explain the relationship between BM and
average return for the extended period 1929–1997. Interestingly, Davis, Fama and
French’s (2000) result provides support for the risk model.
Given the contradictory results provided by empirical evidence, this paper
investigates three simple null hypotheses: (1) There is no relationship between
size and stock returns; (2) There is no relationship between book-to-market ratio
and stock returns; and (3) There is no relationship between beta and stock returns.
The alternative hypotheses are: (1) Size is related to stock returns; (2) Book-to-
market is related to stock returns; and (3) Beta is related to stock returns.
The multivariate model and the three-factor model will be used to test the null
hypotheses. The null hypotheses are similar to the hypotheses that the coefficients
on the size, book-to-market, and market variables (in those two models) are
indistinguishable from zero. Similarly, the alternative hypotheses are similar to
the hypotheses that the coefficients on those variables are statistically different
from zero. The coefficients on the size, book-to-market, and market variables will
be determined by running the Fama and MacBeth cross-sectional regressions on
the multivariate model and the time-series regressions on the three-factor model.
These regressions techniques will be described in the Method section.
3. DATA
3.1. Data Description
This study investigates Information Technology (IT) stocks that are listed on the
New York Stock Exchange (NYSE), the American Stock Exchange (AMEX) and
the NASDAQ from July 1990 to June 2001 as provided by the COMPUSTAT
Database. Not all stocks are chosen for the risk-return tests in the paper; only those
who have the market equity and book-to-market data as described in Section 3.2.2.
are included. In addition, only stocks that have monthly beta for every month during
a testing year (12 months) are selected for the tests conducted during that year. Due
to the fact that new firms enter the IT sector and existing firms exit the sector each
year, there are circumstances that a stock might be selected in a particular period
but might not be selected in a later period. Thus, the number of stocks satisfying
the testing conditions varies year by year.
Table 1 shows the total number of stocks available for the risk-return tests in
each year from July 1990 to 2000. The number of stocks in each portfolio during a
testing year (from July of year t to June of year t + 1) is also reported. For simplicity
the formation of portfolios is described in the next section. From Table 1, it can be
seen that the total number of stocks that satisfy the required conditions increases
nearly four fold over the 1990–2001 period, starting from 379 stocks for the testing
year July 1990–June 1991 to 1474 stocks for the testing year July 2000–June 2001.
This reflects the “technology boom” during the 1990s.
According to Kothari, Shanken and Sloan (1995), there are at least two aspects
of COMPUSTAT selection procedures that appear to impart a survivorship bias.
First, prior to 1978 COMPUSTAT routinely included historical financial statement
information for as many years as available going back to 1946 on firms added to
their database in a given year. In 1978 COMPUSTAT launched a major database
expansion project from about 2700 NYSE-AMEX and high-profile NASDAQ
companies to about 6,000 companies. Five years of annual data from 1973
to 1978 was added for most of these firms. Consider a firm in 1973, with
substantial assets but relatively poor earnings prospects, considerable uncertainty,
and correspondingly low market value. Suppose this high BM (book-to-market
ratio) firm performed poorly over the next five years, with earnings even lower than
expected and negative stock returns. If this company was not on COMPUSTAT
to begin with, it might not be added to the database in 1978, either because of
Table 1. Year-by-Year Numbers of Stocks Available for Testing from July 1990
to June 2001.
Portfolios July July July July July July July July July July July
1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000
Big-High 12 8 3 16 17 18 21 22 24 12 24
Big-BM2 25 25 41 32 44 39 50 68 59 46 77
Big-ME3 44 37 37 47 53 59 64 87 106 98 122
Big-Low 28 47 42 43 50 60 93 103 113 142 170
ME2-High 23 20 20 15 28 24 29 50 48 47 51
ME2-BM2 37 32 24 25 41 42 65 81 84 100 100
ME2-BM3 22 33 34 35 38 56 72 69 81 88 108
ME2-Low 28 24 37 40 39 42 50 54 73 73 101
ME3-High 23 35 39 35 38 50 64 76 95 107 107
ME3-BM2 25 22 26 34 39 45 51 65 81 89 119
ME3-BM3 21 18 17 26 26 29 44 56 51 70 94
ME3-Low 20 17 25 29 38 41 41 59 60 54 59
Small-High 33 32 43 58 62 71 97 107 111 133 183
Small-BM2 14 23 19 27 29 37 42 38 54 62 75
Small-BM3 11 14 18 17 29 24 29 41 40 39 48
Small-Low 13 7 5 8 19 22 26 36 29 31 36
Total 379 394 430 487 590 659 838 1012 1109 1191 1474
The table reports the number of stocks in each portfolio at the beginning of July of each year t from
July 1990 to June 2001.
At the end of June of each year t, all Information Technology stocks on the NYSE, AMEX, and
NASDAQ are grouped into 16 portfolios based on their market capitalization (ln(ME)) and book-to-
market ratio (ln(BM)). The size breakpoints are the first (25%), second (50%) and third (75%) quartiles
of ln(ME) of all information technology stocks on the NYSE, AMEX, and NASDAQ. Similarly, the
book-to-market breakpoints are the first, second, and third quartiles of ln(BM) of all information
technology stocks on these exchanges. Stocks with negative BM are excluded so that ln(BM) is
calculable. A portfolio’s ln(ME), ln(BM), and beta are the value-weighted averages of stocks’ ln(ME),
ln(BM), and beta, respectively. The weights are determined by stocks’ ln(ME).
Big, ME2, ME3, and Small represent the biggest size quartile, second largest size quartile, third
largest size quartile and smallest size quartile, respectively. High, BM2, BM3, and Low represent the
highest book-to-market quartile, second highest book-to-market quartile, third highest book-to-market
quartile, and the lowest high book-to-market quartile, respectively. Big-High means the size – book-
to-market portfolio whose stocks belong to the biggest size quartile and the highest book-to-market
quartile.
delisting or failure to meet minimum asset or market value requirements. On the

other hand, if this high BM company performed unexpectedly well over the next
five years, it could very well be included in 1978. The high ex post returns over
this period and the high initial BM ratio could give the appearance of a positive
relationship between BM and expected returns even when no such relationship
existed.
Second, even in recent years, COMPUSTAT’s procedures for inclusion of

financial data on firms favor surviving firms. Alford, Jones and Zmijewski (1994)
report that firms experiencing unfavourable economic conditions have a high
propensity to delay the filing of their financial statements to the Securities and
Exchange Commission (SEC) and the stock exchanges. Eventually some of
these firms’ stocks are delisted from the exchanges because of failure to comply
with disclosure requirements, thin trading activity, or financial distress. Financial
statement information for these firms during the distress period is less likely to
be available and included in the COMPUSTAT database. Some of the firms that
delay filing of financial statements due to financial distress subsequently improve
their performance. They then file their previously delayed financial statements and
COMPUSTAT incorporates data on these firms. Thus, COMPUSTAT’s selection
procedures may induce an upward bias in the average return on COMPUSTAT
firms, particularly the high BM firms, even the later period.
The sample period of July 1990–June 2001 is free of the first case of survivorship
bias, although the sample is not free of the second case of survivorship bias.
However, it is believed that the (second case of) survivorship bias is not serious
for the information technology stocks in this period. This is due to the fact that the
July 1990–June 2000 period is the “glorious” period for firms in the technology
sector. Firms that file for bankruptcy in this period are not in large number to make
the survivorship bias significant. In addition, firms that went into liquidation from
July 2000 to June 2001 have not yet recovered so that their company information
could be backfiled into the COMPUSTAT database.
3.2. Measurement of Variables
This section briefly describes how various variables such as firm size, book-to-
market ratio, market beta, and stock returns are measured.
3.2.1. Firm Size

A firm’s market equity (ME, denominated in millions of U.S. dollars) at the end
of June of year t is used to measure its size. December-end ME is also mentioned
in the measure of book-to-market ratio section but this quantity is solely for the
purpose of calculating the book-to-market ratios. The reason for the use of June-
end ME and December-end ME is discussed in the measure of book-to-market
ratios section. The natural logarithm of ME instead of the absolute value of ME
is used to proxy for the size variable. This is due to the fact that market value of
equity is highly skewed towards small firms, while there are very few firms that
have high market capitalization. Using ln(ME) will transform the distribution of
ME, making it more normally distributed and facilitating the portfolio selection
procedures. Most of previous authors also use ln(ME) in their tests of capital asset
pricing model. (e.g. Chan & Chen, 1998; Fama & French, 1992; Jegadeesh, 1992).
3.2.2. Book-to-Market Ratios

The book-to-market ratio (BE/ME or BM) is the ratio of BE to ME where ME is the
market value of a firm at the end of December in year t − 1, BE is the book value
of common equity of the same firm at the fiscal year-end that falls in calendar year
t − 1. For the same reason that is discussed above, the risk-return test in this paper
also uses the natural logarithm of BM (ln(BM)) as a proxy for the BM variable
(Fama & French, 1992). Stocks with negative BE are excluded so that ln(BM) is
calculable.
Fama and French (1992) also use December market equity and fiscal yearend
book equity in the BM ratio. They argue that because firms do not have the same
fiscal yearends, using ME at fiscal yearends would result in the BMs whose cross-
sectional variations in a given year are partly due to market-wide variation during
the year. For example, if there is a general fall in stock prices during the year, ratios
measured early in the year will tend to be lower than ratios measured later.
Some confusion might arise due to the use of different periods to measure the
size variable and the book-to-market variable. June-end ME of year t is used to
calculate firm size, yet December-end ME and fiscal yearend book equity in year
t − 1 are used to gauge book-to-market ratios. The reason is that the returns to be
used in the risk-return tests will be for the July of year t to June of year t + 1 period.
This creates a six-month gap between the ME and BM variables and the returns
they are used to explain. Many financial economists follow this measurement
procedure, such as Fama and French (1992) and Kothari, Shanken and Sloan
(1995). The six-month (minimum) gap between fiscal yearend and the return tests
is arbitrary. Nevertheless, this is motivated from Alford, Jones and Zmijewski’s
(1994) paper. These authors report that on average 19.8% of firms do not comply
with the Securities and Exchange Commission (SEC) requirement that firms must
file their 10-K reports with the SEC within 90 days of their fiscal yearends. In
addition, more than 40% of the December fiscal yearend firms that do comply with
the 90-day rule file on March 31, and their reports are not made public until April.
Thus, a six-month gap is expected to be sufficient for all accounting data to be made
publicly available. Table 2 presents the descriptive statistics for the size (ln(ME))
and book-to-market ratio (ln(BM)) variables for the July 1990–June 2000 period.
3.2.3. Market Beta

Annual beta is used in the multivariate model to test the relationship between stock
returns, size, book-to-market ratio, and beta. Monthly beta is used to calculate
Table 2. Descriptive Statistics for the Size – Book-to-Market Portfolios: July

1990–June 2001.
Portfolios Mean Median Standard Deviation Minimum Maximum Observations
Panel A. Portfolios’ book-to-market ratios and descriptive statistics.

Big-High −0.03 −0.18 0.35 −0.39 0.71 11
Big-BM2 −0.63 −0.72 0.37 −1.25 0.09 11
Big-ME3 −1.20 −1.24 0.40 −2.01 −0.50 11
Big-Low −2.15 −2.18 0.48 −3.26 −1.44 11
ME2-High 0.00 −0.11 0.32 −0.35 0.66 11
ME2-BM2 −0.62 −0.70 0.36 −1.21 0.08 11
ME2-BM3 −1.20 −1.26 0.41 −2.02 −0.45 11
ME2-Low −2.25 −2.12 0.44 −3.15 −1.43 11
ME3-High 0.11 0.01 0.32 −0.25 0.82 11
ME3-BM2 −0.59 −0.68 0.38 −1.18 0.15 11
ME3-BM3 −1.19 −1.23 0.40 −1.99 −0.43 11
ME3-Low −2.32 −2.38 0.37 −3.00 −1.71 11
Small-High 0.28 0.11 0.30 0.05 0.99 11
Small-BM2 −0.56 −0.60 0.36 −1.08 0.16 11
Small-BM3 −1.19 −1.27 0.39 −1.96 −0.44 11
Small-Low –2.39 −2.40 0.55 −3.44 −1.54 11
Panel B. Portfolios’ size (ME) and descriptive statistics.
Big-High 6.94 7.29 0.86 4.96 7.88 11
Big-BM2 7.19 7.33 0.60 5.86 8.05 11
Big-ME3 7.41 7.46 0.48 6.75 8.46 11
Big-Low 7.31 7.11 0.71 6.49 8.80 11
ME2-High 4.63 4.75 0.60 3.67 5.87 11
ME2-BM2 4.74 4.94 0.58 3.96 5.90 11
ME2-BM3 4.77 4.96 0.61 3.93 5.97 11
ME2-Low 4.79 4.89 0.61 3.99 6.00 11
ME3-High 3.40 3.50 0.51 2.61 4.33 11
ME3-BM2 3.50 3.62 0.50 2.75 4.43 11
ME3-BM3 3.51 3.63 0.52 2.81 4.42 11
ME3-Low 3.46 3.70 0.55 2.55 4.37 11
Small-High 2.02 2.05 0.34 1.43 2.72 11
Small-BM2 2.17 2.27 0.42 1.43 2.85 11
Small-BM3 2.17 2.17 0.45 1.40 2.92 11
Small-Low 2.16 2.27 0.45 1.41 2.86 11
The table reports the descriptive statistics for the 16 size – book-to-market portfolios from July 1990
to June 2001.
market ratio (ln(BM)). ME is common equity, which is denominated in millions of U.S. dollars and
is measured at the end of June of year t. BM is the ratio of book equity to market equity, with book
equity measured at fiscal yearend falling in year t − 1, and market equity measured at the end of
Table 2. (continued )
December of year t − 1. ln(ME) and ln(BM) are the natural logarithms of ME and BM, respectively.
The ME breakpoints are the first (25%), second (50%), and third quartiles (75%) of the stocks’ ln(ME).
Similarly, the BM breakpoints are the first, second, and third quartiles of the stocks’ ln(BM). Big,
ME2, ME3, and Small represent the biggest largest size quartile, second largest size quartile, third
quartile, and the lowest high book-to-market quartile, respectively.
Panel A shows the time-series averages of book-to-market ratios (ln(BM)) and the associated statistics
for the 16 size – book-to-market portfolios. A portfolio’s book-to-market ratio is the value-weighted
average of the ln(BM) of all stocks in the portfolio, with the weights calculated based on the stocks’
ln(ME).
Panel B shows the time-series averages of market equity (ln(ME)) and the associated statistics for
the 16 size – book-to-market portfolios. A portfolio’s market equity is the value-weighted average of
the ln(ME) of all stocks in the portfolio, with the weights calculated based on the stocks’ ln(ME).
annual beta. Annual beta from July of year t to June of year t + 1 is the arithmetic
average of monthly betas during the same period. That is:
beta(Jult ) + beta(Augt ) + · · · + beta(Jant+1 ) + · · · + beta(Junt+1 )
Betai,t =
12 (4)
where betai,t is the annual stock beta for the testing year beginning in July of year
t and ending in June of year t + 1; beta(Jult ), . . ., beta(Junt+1 ) are the monthly
stock betas for months from July of year t to June of year t + 1.
Monthly stock betas are extracted from the COMPUSTAT database.
COMPUSTAT calculates monthly stock beta by regressing monthly stock
returns of the previous 60 months since the calculation month (inclusive) on
monthly S&P 500 index returns of the same period. If less history is available,
COMPUSTAT uses a minimum of 24-month data to measure monthly beta. Thus,
firms that do not have at least 24 consecutive monthly returns will not have beta
available on the COMPUSTAT database and will not included in this paper’s
tests.2
Due to the fact that firms in the information technology sector do not usually
have a long history of stock prices, calculating annual beta using annual returns of
the past five years (at least two years) would leave very few firms with available
beta. This would result in too few observations for the tests of the risk-return
relationship and could give rise to spurious inferences. The use of monthly returns
in the calculation of annual beta though not producing very precise estimates of
annual beta, would work better than the use of annual returns given the short history
of stock prices for firms in the IT sector.
The procedure that COMPUSTAT uses to measure monthly beta is different
from that employed by Fama and French (1992). In this last paper (monthly) beta
is measured using the Scholes and Williams (1977) procedure, which accounts
for the degree of stock trading infrequency. Scholes-Williams beta estimates are
defined as
+1
B ik
␤i = (5)
(1 + 2r)
k=−1
where r is the autocorrelation of the value-weighted monthly market returns and

the B ik are the slope coefficients from three separate OLS regressions.
R it = ␣ik + B ik R m,t+k + vt , k = −1, 0, 1 (6)
Although trading frequency of stocks does affect a stocks’ beta, the effect of
trading frequency on Info-Tech stocks is believed to be small as these stocks are
highly traded. Furthermore, the use of annual beta will reduce the possible bias
that is inherent in the estimation of monthly beta. Possible bias might arise from
the systematic cross-temporal covariance in short-interval returns that are used to
estimate monthly beta (e.g. Lo & MacKinlay, 1990a; Mech, 1993) or seasonality
in returns (see, Kothari, Shanken & Sloan, 1995).
3.2.4. Returns
Annual returns from July of year t to June of year t + 1 are calculated in the
COMPUSTAT database; they are the percentage increase (or decrease) in the
closing stock prices in June of year t and June of year t + 1, adjusted for dividends
payments and the compounding effect of reinvested dividends. Similarly, monthly
returns are the percentage increase (or decrease) in the closing stock prices on
the last days of two consecutive months. Monthly returns are also adjusted for
the effect of dividend payments. One-year Treasury bill rate and three-month
Treasury bill rate will respectively be used to calculate annual returns and monthly
returns.
Table 3 reports the time-series averages of the portfolios’ excess returns (i.e.
the spreads of portfolio’s annual returns over the one-year Treasury Bill rates),
betas, book-to-market ratios (ln(BM)), and market equity (ln(ME)) for the July
1997–June 2001 period. The summary statistics for these variables are used in the
multivariate model. There seems to be no relationship between portfolios’ excess
returns and size (or ln(ME)). Across each book-to-market quartile, the excess
returns show no particular pattern when moving from big- to small-stock portfolios.
The relationship between the excess returns and ln(BM) is somewhat stronger, with
the excess returns increasing from high- to low-BM portfolios within the biggest-
size quartile. Within the second biggest-size quartile (ME2) the excess returns
decrease from high-BM to low-BM portfolios. Beta is highly positively correlated
Table 3. Time-Series Averages of Portfolios’ Excess Returns, Book-to-Market

Ratios, and Market Equity: July 1997–June 2001.
Book-to-Market Quartiles Size Quartiles
Excess Returns Beta
Big ME2 ME3 Small Big ME2 ME3 Small
High 0.51 44.38 17.98 45.33 1.26 1.19 0.96 0.65

BM2 15.92 31.20 25.75 31.46 1.35 1.26 1.02 0.69
BM3 19.13 11.34 28.74 42.05 1.49 1.32 1.15 0.55
Low 24.63 −4.22 8.79 3.72 1.56 1.28 1.25 0.87
Book-to-Market Ratio (ln(BM)) Market Equity (ln(ME))
High −0.23 −0.16 −0.07 0.12 7.34 5.13 3.76 2.20

BM2 −0.89 −0.85 −0.83 −0.82 7.53 5.25 3.85 2.40
BM3 −1.53 −1.50 −1.49 −1.50 7.92 5.30 3.87 2.41
Low −2.49 −2.60 −2.65 −2.82 8.34 5.21 3.90 2.43
The table reports the time-series averages of portfolios’ excess returns, beta, book-to-market ratio
(ln(BM)), and market equity (ln(ME)) for the July 1997–June 2001 period.
market ratio (ln(BM)). The size breakpoints are the first (25%), second (50%) and third (75%) quartiles
of ln(ME) of all information technology stocks on the NYSE, AMEX, and NASDAQ. Similarly, the
book-to-market breakpoints are the first, second, and third quartiles of ln(BM) of all information
technology stocks on these exchanges. Stocks with negative BM are excluded so that ln(BM) is
calculable. A portfolio’s ln(ME), ln(BM), and beta are the value-weighted averages of stocks’ ln(ME),
ln(BM), and beta, respectively. The weights are determined by stocks’ ln(ME).
Big, ME2, ME3, and Small represent the biggest size quartile, second largest size quartile, third
quartile, and the lowest high book-to-market quartile, respectively. Big-High means the size – book-
to-market portfolio whose stocks belong to the biggest size quartile and the highest book-to-market
quartile.
Annual stock returns for the period July of year t to June of year t + 1 are adjusted for monthly stock
price appreciation plus reinvestment of monthly dividends and the compounding effect of dividends paid
on reinvested dividends. A portfolio’ return is the value-weighted average of the returns on individual
stocks in the portfolio. A portfolio’s excess return is the difference between the portfolio’ return and
the 1-year U.S. Treasury bill rate in the same testing year. A testing year is from July of year t to June
of year t + 1.
with ln(ME). Across each book-to-market quartile, beta decreases monotonically

from big- to small-stock portfolios.
4. METHOD
This section consists of two parts. Section 4.1. describes the multivariate model that
is traditionally used in tests of the size and book-to-market effects. Section 4.2.
describes the three-factor model, proposed by Fama and French (1993).
Detailed portfolio formation procedures for each model are also discussed in
each part.
4.1. Multivariate Model
4.1.1. The Model

Previous researchers form stocks into portfolios based on stock characteristics,
such as a stocks’ market ␤, size, book equity-to-market equity ratio (BM) and then
run Fama and MacBeth (1973) cross-sectional regressions of portfolio returns on
the portfolios’ characteristics using the following model (for example, Banz, 1981;
Chan & Chen, 1988; Chui & Wei, 1998; Fama & French, 1992; Jegadeesh, 1992):
R pt − R ft = ␣0t + ␣1t ln(ME)pt + ␣2t ln(BM)pt + ␣3t ␤pt + ␧pt (7)
where R pt is the monthly return on a portfolio in month t; R ft is the monthly risk
free rate; ␣0t , ␣1t , ␣2t , and ␣3t are the intercept and the coefficients, respectively;
␤pt is the portfolio market beta in month t; ln(ME)pt is the portfolio size variable
in month t; ln(BM)pt is the portfolio book-to-market variable in month t; and ␧pt
is the residual in month t of the regression.
Model 7 or the multivariate model, as it will be called throughout this paper, will
be used to test the size and book-to-market effects in the returns on the Information
Technology stocks for the July 1997–June 2001 period. Although the choice of
the independent variables in Model 7 is arbitrary, it is motivated by the results of
previous studies. Fama and French (1992) report that ME and BM are sufficient to
capture variation in stock returns while market beta, whose effect on stock prices
is the center of the risk-return controversy, has no explanatory power vis-à-vis
stock returns. This paper assumes that the three factors in the multivariate model
are general enough and no additional factors are more significant than these
factors.
One of the weaknesses associated with the multivariate model is the problem
of multicollinearity. In other words, the possible high correlation between any
pair of the three variables would lead to spurious results and render any inference
from the results meaningless. To overcome this problem, various combinations of
these three factors will be used in regressions. Fama and French (1992) use the
Fama and MacBeth (1973) cross-sectional approach to estimate the coefficients
and residuals in the model. This paper improves the Fama and MacBeth procedure,
which use Ordinary Least Squares (OLS) method, by adopting the Newey-West
(1987) technique. The OLS methodology will give best, linear, unbiased estimates
of the coefficients of a model only when the residuals have constant variance, (i.e.
homoskedasticity), and are independent of each other (i.e. not autocorrelated).
If the residuals show patterns of heteroskedasticity (non-constant variance), the
variances, and therefore the standard deviations of the coefficients, will be
biased. Because the Newey-West (1987) procedure accounts for the problems
of hetoroskedasticity and autocorrelation, it is more efficient and produces more
accurate t-statistics than the OLS.
4.1.2. Data for the Multivariate Model

It has been discussed in Section 3 that the period to be covered for testing
will be from July 1990 to June 2001. However, this is achievable only when
all the necessary data for stocks are available. For instance, stocks without
ln(ME), ln(BM), beta, and returns data (with these variables calculated as
described in Part B of the Data section) as at end of June of year t would not be
included in the tests based on the multivariate model. (Monthly) beta data for
IT stocks in COMPUSTAT is only available from August 1996. Thus, annual
beta for IT stocks could not be calculated for testing years3 before July 1997.
Therefore, the sample period chosen for the tests based on the multivariate
model is from July 1997 to June 2001. Also, annual data rather than monthly
data are used. Kothari, Shanken and Sloan (1995) argue that annual data are
free of any seasonality effects and therefore is more likely to give a better
picture of the relationship between size, book-to-market ratio, beta, and stock
returns.
4.1.3. Portfolio Selection Procedure

In June of each year, all IT stocks that satisfy our selection criteria are sorted by
size (ln(ME)) to determine the quartile breakpoints for ME. These stocks are then
allocated to four size portfolios based on the breakpoints. We subdivide each size
portfolio into four portfolios on the basis of book-to-market equity (ln(BM)) for
individual stocks. The book-to-market breakpoints are determined based on all
available stocks in our sample, not on stocks in each size portfolio. As a result, we
obtain sixteen size – book-to-market portfolios with stocks independently sorted
into portfolios according to their ln(ME) and ln(BM).
4.1.4. Estimating Portfolios’ Size, Book-to-Market Ratio, Beta, and Returns

After assigning firms to the size – book-to-market portfolios in June, we calculate
the value-weighted annual returns on the portfolios with the weights based on the
natural logarithms of individual stocks’ market equity (ln(ME)). Algebraically, a
portfolio’s annual returns are calculated as:
n
i=1 R i × |ln(MEi )|
Rp = n (8)
i=1 |ln(MEi )|
where R p is the return on portfolio p; |ln(MEi )| is the absolute value of the natural
logarithm of stock i’s market equity; n is the number of stocks in portfolio p. We
obtain the value-weighted market equity, value-weighted book-to-market ratio,
and value-weighted market beta for each portfolio in a similar way: A portfolio’s
market equity is calculated as:
n
|ln(MEi )| × ln(MEi )
ln(MEp ) = i=1 (9)
|ln(MEi )|
A portfolio’s book-to-market ratio is calculated as:
n
|ln(MEi )| × ln(BMi )
ln(BMp ) = i=1n (10)
i=1 |ln(MEi )|
and a portfolio’s beta is calculated as:

n
|ln(MEi )| × betai
betap = i=1 n (11)
i=1 |ln(MEi )|
4.2. Three-Factor Model
4.2.1. The Model

Many financial economists suspect that the relation between size and average
return is not due to investors’ irrational behavior but is representative of a more
fundamental relationship between expected returns and economic risk factors
(Chan & Chen, 1991; Fama & French, 1992). In an attempt to produce more
evidence on the rational asset-pricing story, Fama and French (1993) propose a
so-called three-factor model. They use the Black, Jensen and Scholes’ (1972) time-
series regression approach. Monthly returns on stocks are regressed on the returns
to a market portfolio of stocks and mimicking portfolios for size and book-to-
market equity (BM). The three-factor model is expressed:
R(t) − RF(t) = a + b[RM(t) − RF(t)] + sSMB(t) + hHML(t) + e(t) (12)
where RM(t) is the value-weighted monthly returns on the market portfolio that
consists of all IT stocks; RF(t) is the risk free rate; SMB(t) is the difference between
returns on small stocks and returns on big stocks; HML(t) is the difference between
returns on high book-to-market stocks and low book-to-market stocks.
4.2.2. Why Employ a Three-Factor Model?

We employ the three-factor model in addition to the multivariate model (Model 7)
because firstly, we want to examine whether the size and book-to-market
effects is a function of the estimation technique. If the two models produce
different or conflicting results, there are three possibilities: (1) the two models
describe different aspects of the size and book-to-market effect and thus they are
incomparable; (2) either one of the model is wrong; and (3) both models are wrong
and the size and book-to-market effects might be the results of data-snooping biases
as conjectured by Lo and MacKinlay (1990b). Thus, a test of the size and book-
to-market effects is also a test of these two models.
Secondly, the multivariate model does not tell us why size and book-to-market
factors are related to stock returns. It might be that these factors proxy for common
risk factors in returns. The three-factor model goes a step towards finding new
factors that capture firm size and book-to-market equity. Because SMB and HML
factors can be used for any stock, if we can establish that SMB and HML are
related to some kind of risk,4 the question of why there is a relationship between
stock returns, firm size, and book-to-market equity would be solved.
Thirdly, the three-factor model has an advantage over the multivariate model in
that the former uses the pattern of returns, i.e. the return on small stocks minus the
return on big stocks (SMB), the return on high book-to-market stocks minus the
return on low book-to-market stocks (HML), in determining average stock returns
(Barber & Lyon, 1997a). These authors argue that the use of pattern of returns is
good because there are cases where larger firms or low book-to-market firms may
have common stock returns that more closely mimic those of small firms or high
book-to-market firms. The three-factor model allows for this possibility since the
pattern of returns, rather than the explicit measurement of size and book-to-market
equity, is used.
Finally, Fama and French (1993, 1995, 1996, 1997) find that empirical evidence
supports the three-factor model. They suggest that this model is an equilibrium-
pricing model. It would be interesting to see whether the model works well for
firms in the information technology sector.
4.2.3. Data for the Three-Factor Model

As the three-factor model does not require beta data for individual stocks, the
sample period used for the risk-return tests based on the three-factor model will
cover the full July 1990–June 2001 period as described in the Data section. Whereas
annual data is used for the multivariate model, we use monthly data for the three-
factor model. In Section 4.1. we argue that annual data is better than monthly data
because the effect of seasonal change in returns will be minimized when annual
data are used. This is true if the multivariate model is used. However, for the three-
factor model, it is not necessary to eliminate the seasonal effect. The reason is
that the explanatory variables, SMB, HML, and RM–RF are themselves returns,
in which the seasonal effect is inherent. The dependent variable is also return and
as such contains in itself the seasonal effect. When the time-series regressions of
the dependent variable on the explanatory variables are run, the seasonal effects in
two sides of the three-factor model will be canceled out. Thus, the period (month
versus annual) used to measure returns is appropriate and also has the advantage
of providing more observations, which would lead to more precise estimates of
the coefficients for the model.
4.2.4. Inputs for the Three-Factor Model

We mimic the procedure employed by Fama and French (1993). In June of each
year t from 1985 to 2000, all Information Technology stocks on the NYSE, AMEX,
and NASDAQ which have the ME and BM data at the start of each testing year
are size-sorted into two groups based on stocks’ ln(ME). The size breakpoint is
the median of all IT stocks’ ln(ME). The two groups are called small (S) and big
(B). All the selected IT stocks are also sorted into three book-to-market portfolios
based on the breakpoints for the bottom 30% (Low), middle 40% (Medium), and
top 30% (High) of the ranked values of ln(BM) for the stocks. BM is defined as
in the “measure of book-to-market equity” section. The size-sort and BM-sort are
conducted independently. Six portfolios (S/L, S/M, S/H, B/L, B/M, B/H) from the
intersections of the three BM and two ME groups are obtained. For example, the
S/L portfolio contains the stocks in the small-ME group that are also in the low-BM
group, and the B/H portfolio contains the big-ME stocks that also have high BMs.
Monthly value-weighted returns on the six portfolios are calculated from July of
year t to June of t + 1, and the portfolios are reformed in June of t + 1. Fama and
French (1993) say that the calculation of returns beginning in July to year t will
allow for the possibility that book equity of year t − 1 is not known until June of
year t.
(1) The Explanatory Variable: Size or (SMB)

The portfolio SMB (small minus big), mimics the risk factor in returns
related to size, is the difference, each month, between the simple average
of the returns on the three small-stock portfolios (S/L, S/M, and S/H)
and the simple average of the returns on the three big-stock portfolios
(B/L, B/M, and B/H):

R s/l + R s/m + R s/h R b/l + R b/m + R b/h
SMB = − (13)
3 3
Equation (13) can be expressed as:

R s/l − R b/l + R s/m − R b/m + R s/h − R b/h
SMB = (14)
3
From Eq. (14), it can be seen that SMB is the difference between the returns on
small- and big-stock portfolios with about the same weighted-average book-
to-market equity. This difference should be largely free of the influence of BM,
focusing instead on the difference return behaviors of small and big stocks.
(2) The Explanatory Variable: BM or (HML)
The portfolio HML (high minus low) mimics the risk factor in returns
related to book-to-market equity, is defined similarly. HML is the difference,
each month, between the simple average of the returns on the two high-BM
portfolios (S/H and B/H) and the average of the returns on the two low-BM
portfolios (S/L and B/L):
R s/h + R b/h R s/l + R b/l
HML = − (15)
2 2
or
(R s/h − R s/l ) + (R b/h − R b/l )
HML = (16)
2
The two components of HML are returns on high- and low-BM portfolios with
about the same weighted-average size. Thus the difference between the two
returns should be largely free of the size factor in returns, focusing instead on
the different return behaviors of high- and low-BM firms.
(3) The Explanatory Variable: Market or (RM–RF)
Finally, the proxy for the market factor in stock returns is the excess market
return, RM–RF. RM is the return on the value-weighted portfolio of the stocks
in the six size-BM portfolios. RF is the three-month bill rate. Fama and French
(1993) use the one-month Treasury bill rate to proxy for the RF. However, due
to the lack of information for the one-month Treasury bill rate, the three-month
Treasury bill rate is used.5 Nevertheless, the use of the three-month bill rate
does not result in any substantial difference in the risk-return tests’ results.
This paper will later prove this statement empirically.
(4) The Dependent Variables (the Returns on the Size–Book-to-Market Portfolios)
The excess returns on 16 portfolios, formed on size and book-to-market
equity, are used as dependent variables in the time-series regressions. The
reason for the use of the size – book-to-market portfolios is to determine

whether the mimicking portfolios SMB and HML capture common factors in
stock returns related to size and book-to-market equity.
The 16 size-BE/ME portfolios are formed much like the six size-BE/ME portfolios
discussed earlier. In June of each year t all Information Technology stocks are sorted
by size (ln(ME) and (independently) by book-to-market equity (ln(BM)). For the
size sort, ME is measured at the end of June. For the book-to-market sort, ME is
market equity at the end of December of t − 1, and BE is book common equity for
the fiscal year ending in calendar year t − 1. The first (25%), second (50%), and
third (75%) quartiles of stocks’ ln(ME) and ln(BM) are used as the breakpoints for
the allocation of stocks into four size quartiles and four book-to-market quartiles.
Sixteen portfolios are formed from the intersections of the size and BM quartiles.
Value-weighted monthly returns on the portfolios from July of year t to June of
year t + 1 are used as dependent variables in the time-series regressions.
Table 4 shows the descriptive statistics for SMB, HML, RM–RF, and the excess
returns on sixteen size – BM portfolios from July 1990 to June 2001. The time-
series mean of SMB is 1.12%, with a t-statistic of 1.63 (less than 1.98). This
suggests that over the sample period, small stocks do not outperform big stocks.
However, the returns on high book-to-market stocks do exceed those on low
book-to-market stocks over the same period. The time-series mean for HML is
a positive 1.48%, with a t-statistic of 2.37. The excess returns on the market
portfolio (RM–RF) are significant in the negative territory (a mean of −2.71%,
with a t-statistic of −3.04). Except for the portfolios in the smallest-size quartile
and the ME3-BM2 portfolio, other portfolios have mean excess returns that are
distinguishable from zero. These portfolios have negative mean excess returns
with −2.53% being the highest (the ME3-High portfolio). This fact, coupled with
a negative mean excess return on the market portfolio (RM–RF), indicates that IT
stocks on average produce negative returns over the July 1990–June 2001 period.
It is also clear that the mean excess return on each of the size – book-to-market
portfolio is very low in practical terms. For instance, the ME3-High portfolio,
which has the highest mean excess return compared with other size – BM portfolios,
produces a mean excess return of −2.53% per month, which is translated into
approximately −30.36% per year. That is not to mention the ME2-Low portfolio,
which loses nearly 51.12% (−4.26 × 12) per year. The lowest mean excess return
on a size – book-to-market portfolio reported by Fama and French (1993) was
0.32% per month or 3.84% per year. Even after accounting for slight differences
in research design, the positive 3.84% is still significantly different to the negative
−30.36%. Table 4 also reports low correlations between SMB, HML, and RM–RF.
The correlation coefficients between SMB and HML, SMB and RM–RF, and HML
Table 4. Descriptive Statistics for the Dependent and Explanatory Variables for
the Three-Factor Model: July 1990–June 2001.
Panel A. Descriptive Statistics for SMB, HML, and RM–RF
Mean St. dev Correlations

SMB HML RM–RF
SMB 1.12 7.87 1

(t-statistic) (1.63)
HML 1.48 7.17 0.34 1.00
(t-statistic) (2.37)
RM–RF −2.71 10.27 0.36 −0.13 1
(t-statistic) (−3.04)
Panel B. Descriptive Statistics for Portfolios’ Excess Returns

Means Standard Deviations
High −3.09 −2.66 −2.53 −0.04 9.81 10.65 10.06 11.92

(t-statistic) (−3.62) (−2.87) (−2.90) (−1.29)
BM2 −2.55 −2.84 0.08 −1.29 8.89 9.76 32.91 13.66
(t-statistic) (−3.30) (−3.34) (0.03) (−1.09)
BM3 −2.95 −3.07 −3.17 −1.15 9.56 11.28 13.70 15.08
(t-statistic) (−3.55) (−3.13) (−2.65) (−0.88)
Low −3.41 −4.26 −3.35 0.85 10.36 12.98 15.33 19.35
(t-statistic) (−3.78) (−3.77) (−2.51) (0.51)
The table reports the descriptive statistics for the returns that mimic size (SMB), book-to-market
(HML), and market (RM–RF) factors (panel A), and the excess returns on the 16 size – book-to-market
portfolios (panel B) for the July 1990–June 2001 period. SMB, HML, and RM–RF are used as the
explanatory variables, and the portfolios’ excess returns are used as the dependent variable for the
three-factor model.
R(t) − RF(t) = a + b[RM(t) − RF(t)] + sSMB(t) + hHML(t) + e(t)
Panel A shows the descriptive statistics for SMB, HML, and RM–RF. SMB (small minus big) is the
difference between the returns on small-stock and big-stock portfolios with about the same weighted
average book-to-market equity. HML (high minus low) is the difference between the returns on high
and low book-to-market equity portfolios with about the same weighted average size. RM is the value-
weighted monthly% return on the stocks in the 16 size-BM portfolios. RF is the three-month Treasury
bill rate, observed at the beginning of each month.
Panel B shows the descriptive statistics for the excess returns on the 16 size – book-to-market
portfolios. Three-month Treasury bill is used to calculate the excess returns. The portfolios are formed
as follows. At the end of June of each year t, all Information Technology stocks on the NYSE,
AMEX, and NASDAQ are grouped into 16 portfolios based on their market capitalization (ln(ME)) and
Table 4. (Continued )
book-to-market ratio (ln(BM)). ME is common equity, which is denominated in millions of U.S.
dollars and is measured at the end of June of year t. BM is the ratio of book equity to market
equity, with book equity measured at fiscal yearend falling in year t − 1, and market equity measured
at the end of December of year t − 1.The ME breakpoints are the first (25%), second (50%), and
third (75%) quartiles of the stocks’ ln(ME). Similarly, the BM breakpoints are the first, second,
and third quartiles of the stocks’ ln(BM). Stocks with negative BM are excluded so that ln(BM) is
calculable.
The reported means (in percentage) are the time-series averages of the monthly values of the
corresponding variables. The sample period ranges from July 1990 to June 2001.
A value with a t-statistic being less than –1.98 or greater than 1.98 is considered significant at the
5% level.
and RM–RF are 0.34, 0.36, and −0.13, respectively. All these figures are well
below 0.50.
5. RESULTS
This section contains three parts. Section 5.1. presents the empirical results of the
Fama and MacBeth (1973) regressions on the multivariate model. Section 5.2.
presents the empirical results of the time-series regressions on the three-factor
model. Section 5.3. compares the behavior of information technology stocks’ price
before and after the technology stock crash in April 2000.
5.1. Regressions Using the Multivariate Model
This section consists of two sub-sections. Section 5.1.1. discusses the results
from the FM (Fama & MacBeth) regressions on individual stocks. Section 5.1.2.
discusses the results from the FM regressions on the size – book-to-market
portfolios.
5.1.1. The FM Regressions on Individual Stocks

Table 5 reports the time-series averages of the slope from year-by-year FM
regressions of the cross-section of stock returns on size, book-to-market ratio,
and beta. The average slope provides standard FM tests for determining which
explanatory variables on average have non-zero expected premiums during
the July 1997 to June 2001 period. From Table 5, it can be seen that the
significance (at the 0.10 level) of the slope coefficient of the book-to-market
variable, ln(BM), is consistent in three out of the four models when ln(BM)
is present (Models 1, 2, and 7). When the book-to-market variable is alone in
Table 5. Results from the (Improved) Fama and MacBeth Regressions with
Stock Data: July 1997–June 2001.
Models Variables Intercept (%) Coefficients (%)
a1t ln(ME)it a2t ln(BM)it a3t ␤it
Model 1 ln(ME), ln(BM), and Beta 46.02 −4.18 5.03 5.05

(t-statistic) (1.11) (−1.25) (2.56)* (1.39)
Model 2 ln(BM) and Beta 30.75 8.42 4.48
(t-statistic) (1.04) (3.08)* (1.40)
Model 3 ln(ME) and Beta 44.53 −5.11 4.89
(t-statistic) (1.06) (−1.61) (1.32)
Model 4 ln(ME) and BM 48.51 −3.62 4.82
(t-statistic) (1.13) (−1.21) (2.22)
Model 5 Beta 21.81 4.05
(t-statistic) (0.75) (1.27)
Model 6 ln(ME) 47.12 −4.25
(t-statistic) (1.08) (−1.45)
Model 7 ln(BM) 34.65 7.51
(t-statistic) (1.07) (2.68)*
The table reports the time-series average coefficient slopes and the associated t-statistics (in brackets)
from the (improved) Fama and MacBeth regressions of stock returns on size, book-to-market ratio, and
beta for the July 1997–June 2001 period. The reported intercepts and slopes are in percentage.
R it − R ft = a 0t + a 1t ln(ME)it + a 2t ln(BM)it + a 3t ␤it + ␧it
The multivariate model above is a general model. There is a total of seven models where the explanatory
variables are combined in various ways. For instance, model 1 includes all three variables, ln(ME),
ln(BM), and beta while model 2 consists of only two variables, ln(BM) and beta. The aim is to separate
the effect of the multicollinearity problem.
ME is common equity, which is denominated in millions of U.S. dollars and is measured at the end
of June of year t. BM is the ratio of book equity to market equity, with book equity measured at fiscal
yearend falling in year t − 1, and market equity measured at the end of December of year t − 1. ln(ME)
and ln(BM) are the natural logarithms of ME and BM, respectively. Beta is the average of monthly
stock betas during the period July of year t to June of year t + 1. Monthly stock beta is calculated for
a 5-year (60-month) time period, ending in the current month. If less history is available, beta will be
calculated for as few as 24 months. Month-end closing prices (including dividends) are used in the
calculation. A testing year (for each yearly cross-sectional regression) is from July of year t to June of
year t + 1. There are four yearly cross-sectional regressions for the July 1997–June 2001 period.
∗ Significant at the 0.10 level.
the regressions, the average slope on ln(BM) is 7.51%, with a t-statistic of 2.68
(Model 7). The magnitude of the book-to-market coefficient increases to 8.42%
(t-statistic of 3.08) when beta is added (Model 2). This change might be caused
by a slight negative correlation coefficient of −0.08 (Table 7) between book-to-
market ratio and beta. In the regressions where ln(ME), ln(BM), and beta are
included (Model 1), the slope on ln(BM) decreases to 5.03%, with a t-statistic
of 2.56.
The slope coefficient on ln(BM) is indistinguishable from zero in Model 4, where
stock returns are regressed on the size and book-to-market variables. This lack of
explanatory power from ln(BM) is due to the fact that ln(BM) is highly negatively
correlated with ln(ME). The time-series average of the correlation coefficient
between ln(BM) and ln(ME) is reported at −0.37 (Table 7). Nevertheless, Model 7
provides strong evidence that the book-to-market ratio is statistically related to
variation in stock returns. The strong explanatory power of book-to-market ratio
sharply contrasts the weak relationship between size, ln(ME), beta, and stock
returns. In univariate regressions where either beta or size is included, the t-statistics
for the slopes on the size and beta variables are just −1.45 (Model 6) and 1.27
(Model 5), respectively. Models 1, 2, 3, and 4 say that size and beta have no power
to explain average returns in FM regressions that use various combinations of beta
with size and book-to-market ratio.
It should be noticed that our interpretation of the results in Table 5 might be
affected by an extremely low adjusted R-squared for each model; the values of
the adjusted-R2 never exceeds 0.05 across models and across years. Thus, the
book-to-market variable is significant in explaining stock returns but it is not the
only significant variable; some unknown variable(s) are more significant and more
powerful in explaining variation in stock returns.
5.1.2. The FM Regressions on the Size – Book-to-Market Portfolios

Table 6 reports the results from the FM regressions on portfolios formed on size
and book-to-market ratios. Similar to Table 5, Table 6 also reports a strong book-
to-market effect (higher book-to-market stocks have higher average returns) in
three out of the four models where ln(BM) is present (Models 2, 4, and 7). In the
yearly regressions of returns on book-to-market ratio alone, the t-statistic for the
coefficient on ln(BM) scores an impressive 2.77 (Model 7). The book-to-market
effect is thus robust in the 1997–2001 annual returns on NYSE, AMEX, and
NASDAQ Info-Tech stocks.
Like the weak size effect reported in Table 5, the results reported in Table 6
suggest that the size effect in the size – book-to-market portfolios’ returns is
insignificant over the July 1997 to June 2001 period. When firm size, ln(ME),
is the only explanatory variable in the regressions, the t-statistic for the average
Table 6. Results from the (Improved) Fama and MacBeth Regressions with
Portfolio Data: July 1997–June 2001.
Models Variables Intercept (%) Coefficients (%)
a1t ln(ME)pt a2t ln(BM)pt a 3t b pt
Model 1 ln(ME), ln(BM), and Beta 46.15 0.35 6.07 −17.13

(t-statistic) (1.09) (0.27) (2.15) (−1.72)
Model 2 ln(BM) and Beta 44.77 6.85 −14.15
(t-statistic) (1.07) (2.39)* (−1.07)
Model 3 ln(ME) and Beta 45.65 0.96 −25.60
(t-statistic) (1.15) (0.50) (−2.14)
Model 4 ln(ME) and BM 41.12 −2.14 7.23
(t-statistic) (1.01) (−0.90) (2.72)*
Model 5 Beta 44.26 −21.14
(t-statistic) (1.12) (−1.92)
Model 6 ln(ME) 32.92 −2.44
(t-statistic) (0.79) (−1.04)
Model 7 ln(BM) 31.56 7.42
(t-statistic) (1.05) (2.77)*
The table reports the time-series average coefficient slopes and the associated t-statistics (in brackets)
from the (improved) Fama and MacBeth regressions of stock returns on size, book-to-market ratio, and
beta for the July 1997–June 2001 period. The reported intercepts and slopes are in percentage.
R pt − R ft = a 0t + a 1t ln(ME)pt + a 2t ln(BM)pt + a 3t ␤pt + ␧pt
The multivariate model above is a general model. There is a total of seven models where the explanatory
variables are combined in various ways. For instance, model 1 includes all three variables, ln(ME),
ln(BM), and beta while model 2 consists of only two variables, ln(BM) and beta. The aim is to separate
the effect of the multicollinearity problem.
ME is common equity, which is denominated in millions of U.S. dollars and is measured at the end
of June of year t. BM is the ratio of book equity to market equity, with book equity measured at fiscal
yearend falling in year t − 1, and market equity measured at the end of December of year t − 1. ln(ME)
and ln(BM) are the natural logarithms of ME and BM, respectively. Beta is the average of monthly
stock betas during the period July of year t to June of year t + 1. Monthly stock beta is calculated for
a 5-year (60-month) time period, ending in the current month. If less history is available, beta will be
calculated for as few as 24 months. Month-end closing prices (including dividends) are used in the
calculation. A testing year (for each yearly cross-sectional regression) is from July of year t to June of
year t + 1. There are four yearly cross-sectional regressions for the July 1997–June 2001 period.
Portfolios are reformed based on stocks’ market equity, ln(ME), and book-to-market ratio, ln(BM)
at the end of June of each year t. Portfolios’ ln(ME), ln(BM), and beta are the value-weighted averages
of individual stocks’ ln(ME), ln(BM), and beta, respectively. The weights are determined by stocks’
ln(ME).
slope on ln(ME) is only −1.04 (Model 6). With the presence of other explanatory
variables, such as ln(BM) and beta, the average slopes on the firm size variable
are even smaller and have smaller t-statistics. The risk premiums on ␤ in Models
1, 2, 3, and 5 are interestingly also negative. This could indicate that portfolios
with lower ␤ earn higher returns than portfolios with higher ␤. Nevertheless, the
evidence on the negative relationship between portfolios’ ␤ and returns is not
conclusive because all the premiums on the beta variable are indistinguishable
from 0. The largest t-statistic (in absolute value) for the premium on ␤ is −2.14,
which is insignificant at the 0.10 level.
5.1.3. Comparison between the FM Regressions on Individual Stocks and the FM
Regressions on the Size – Book-to-Market Portfolios
The book-to-market effect exists regardless of whether portfolio data or stock data
are used. In addition, the magnitudes of this effect are similar in both cases. When
portfolio data are used, the slope on ln(BM) in Model 7 is 7.42%, with a t-statistic
of 2.77 (Table 6). When stock data are used, the slope on ln(BM) in Model 7 is
7.51%, with a t-statistic of 2.68 (Table 5). These results confirm the significance of
the book-to-market ratio in explaining stock returns. Although the slope on ln(ME)
and beta in Table 5 is different from the slope on ln(ME) and beta in Table 6, both
tables report that size and beta are not related to stock returns. Instead, both tables
suggest that ln(BM) is the only variable that explains variation in stock returns.
Table 7 shows that when stocks are formed into size – book-to-market portfolios,
the correlation between the size variable and the book-to-market variable is
reduced to −0.08 (−0.37 if stock data are used). However, the achieved low
correlation between size and book-to-market ratio comes at the expense of the
correlation between size and beta, which is increased to 0.75. The change in
the correlation coefficients between ln(ME), ln(BM), and beta due to alternative
uses of stock data and portfolio data is one of the reasons for the differing
slopes on ln(ME) and beta in Table 5 and Table 6. Nevertheless, the advantage of
sorting stocks into portfolios is that the return-generating process is more correctly
specified by the multivariate model, which assumes portfolio size, book-to-market
ratio, and beta are the three most important factors that explain variation in portfolio
returns. This is indicated by the fact that the values of R2 in regressions using
portfolio data are higher than the values of R2 obtained in regressions using stock
data, with the (average) values of adjusted R2 in the cross-sectional regressions
using portfolio data are as high as 0.20.
5.1.4. Comparison with Fama and French’s (1992) results
Similar to Fama and French’s (1992) paper, this paper finds no relationship between
beta and stock returns. Whether beta is used alone or in various combinations with
size and book-to-market ratio, the slopes on beta is indistinguishable from 0. This
Table 7. Correlation Matrices.

ln(ME) ln(BM) Beta
Stocks data
ln(ME) 1
ln(BM) −0.37 1
Beta 0.13 −0.08 1
Portfolios data
ln(ME) 1
ln(BM) −0.08 1
Beta 0.75 −0.24 1
The table reports the time-series averages of the correlation coefficients between ln(ME), ln(BM), and
beta for individual stocks as well as for the size – book-to-market portfolios. ME is common equity,
which is denominated in millions of U.S. dollars and is measured at the end of June of year t. BM is
the ratio of book equity to market equity, with book equity measured at fiscal yearend falling in year
t − 1, and market equity measured at the end of December of year t − 1. ln(ME) and ln(BM) are the
natural logarithms of ME and BM, respectively. Beta is the average of monthly stock betas during the
period July of year t to June of year t + 1. Monthly stock beta is calculated for a 5-year (60-month)
time period, ending in the current month. If less history is available, beta will be calculated for as few
as 24 months. Month-end closing prices (including dividends) are used in the calculation.
Portfolios are reformed at the end of June of each year t, using size and book-to-market quartile
breakpoints. The size breakpoints are the first (25%), second (50%) and third (75%) quartiles of ln(ME)
of all information technology stocks on the NYSE, AMEX, and NASDAQ. Similarly, the book-to-
market breakpoints are the first, second, and third quartiles of ln(BM) of all information technology
stocks on these exchanges. Stocks with negative BM are excluded so that ln(BM) is calculable. A
portfolio’s ln(ME), ln(BM), and beta are the value-weighted averages of stocks’ ln(ME), ln(BM), and
beta, respectively. The weights are determined by stocks’ ln(ME).
result confirms the weak explanatory power of beta reported by many financial
economists (e.g. Barber & Lyon, 1997b; Davis, 1994; Jegadeesh, 1992). However,
the results in this paper provide a different picture to those found by Fama and
French (1992). First, Fama and French (1992) find that the book-to-market effect
is robust at the 5% level of significance while the book-to-market effect in this
study is only significant at the 10% level. Secondly, the combination of size and
book-to-market ratio captures variation in stock returns, whereas this paper finds
no relationship between size and stock returns regardless of whether size is alone
or in combination with book-to-market ratio.
5.2. Regressions Using the Three-Factor Model
This section contains six sub-Sections 5.2.1., 5.2.2. and 5.2.3. reports the results
on the one-factor model, with HML, SMB, and RM–RF in turn plays the role of the
explanatory variable. 5.2.4. discusses the results on the two-factor model, in which
SMB and HML are the two explanatory variables. 5.2.5. discusses the results on the
three-factor model, in which SMB, HML, and RM–RF are the three explanatory
variables. And finally, 5.2.6. considers the use of the three-month Treasury bill
rate as an alternative to the one-month Treasury bill rate.
5.2.1. Regressions with One Explanatory Variable: HML

Inspired by a strong relationship between the book-to-market factor and stock
returns in Section 5.1, we do a preliminary test on the relationship between the
16 size – book-to-market portfolios’ returns and the HML returns, which mimic
the book-to-market factor. Panel A in Table 8 shows the results of the Seemingly
Unrelated Regression (SUR) of the size – BM portfolios’ returns on HML. Except
for the highest book-to-market quartile and the top three portfolios in the smallest
size quartile, all the portfolios in other size and book-to-market quartiles have HML
coefficients that are statistically different from zero. Even within the smallest size
quartile and the highest book-to-market quartile, the slopes on HML for the ME2-
High and Small-BM3 portfolios have t-statistics of −1.97, which is very close
to –1.98 – the critical value at the 5% level of significance. More interestingly,
the coefficient slopes on HML are related to the book-to-market factor: (with
the exception of the ME3–BM2 portfolio) within each size quartile, the slopes
decrease monotonically from higher- to lower – book-to-market quartiles. The
HML variable therefore is significant to explain stock returns.
However, the values of the intercepts and the SUR value for the adjusted R2
seem to suggest that other factors, in addition to HML, are needed to capture all
variation in stock returns. More than half of the sixteen size – book-to-market
portfolios have intercepts that are statistically different from zero. Particularly, all
portfolios in the biggest and second biggest size quartiles have intercepts that are
statistically different from zero. Moreover, the adjusted R2 recorded is only 0.16.
Thus, the one-factor model where HML is the only explanatory variable leaves
84% of variation in stock returns unexplained.
5.2.2. Regressions with One Explanatory Variable: SMB

Different from a weak size effect reported in Section 5.1., SMB proves to be a
significant factor that explains the shared variation in expected returns. Panel B
in Table 8 reports the intercepts and the coefficient slopes on SMB in the SUR
regression of the size – book-to-market portfolios’ returns on SMB alone. Only
five of the 16 size – book-to-market portfolios, four from the biggest size quartiles
and one from the second biggest size quartiles (the ME2-BM2 portfolio), produce
coefficients that are indistinguishable from zero. The remaining 11 portfolios have
SMB slopes that are either equal to or greater than 2.83 standard errors from zero.
76
Table 8. Seemingly Unrelated Regressions of Portfolios’ Excess Returns on SMB, HML, and RM–RF in the
One-Factor and Two-Factor Models: July 1990–June 2001.
Book-to-Market Quartiles Size Quartiles (Intercepts)
Panel A. SUR Results for One-Factor Model: HML Panel B. SUR Results for One-Factor Model: SMB
High −2.89 −2.29 −2.28 0.04 −3.05 −3.05 −3.15 −0.95

(t-statistic) (−3.34) (−2.46) (−2.58) (0.04) (−3.55) (−3.38) (−3.95) (−1.07)
BM2 −2.14 −2.41 −4.20 −0.86 −2.60 −3.00 −3.36 −2.36
(t-statistic) (−2.79) (−2.85) (−1.86) (−0.72) (−3.34) (−3.54) (−1.73) (−2.36)
BM3 −2.43 −2.44 −2.43 −0.62 −2.95 −3.45 −4.01 −2.22
(t-statistic) (−2.98) (−2.53) (−2.07) (−0.47) (−3.52) (−3.60) (−3.70) (−1.95)
Low −2.74 −3.36 −2.29 1.58 −3.42 −4.80 −4.43 −0.81
(t-statistic) (−3.14) (−3.10) (−1.79) (0.94) (−3.76) (−4.41) (−3.79) (−0.60)
QUANG-NGOC NGUYEN ET AL.

High −0.13 −0.25 −0.17 −0.06 −0.03 0.35 0.55 0.81
(t-statistic) (−1.13) (−1.97) (−1.40) (−0.39) (−0.28) (3.04) (5.44) (7.23)
BM2 −0.28 −0.29 2.90 −0.29 0.04 0.15 3.08 0.95
(t-statistic) (−2.64) (−2.51) (9.34) (−1.78) (0.43) (1.41) (12.55) (7.56)
BM3 −0.35 −0.43 −0.50 −0.36 0.00 0.34 0.75 0.96
(t-statistic) (−3.13) (−3.24) (−3.11) (−1.97) (0.00) (2.83) (5.51) (6.68)
Low −0.46 −0.61 −0.72 −0.49 0.00 0.48 0.96 1.49
(t-statistic) (−3.82) (−4.10) (−4.10) (−2.14) (0.02) (3.48) (6.53) (8.75)
Adjusted R2 0.16 0.30
Size and Book-to-Market Effects in the Returns on Information Technology Stocks
Book-to-Market Quartiles Size Quartiles (Intercepts)
Panel C. SUR Results for One-Factor Model: RM–RF Panel D. SUR Results for Two-Factor Model: SMB and HML
High −1.12 −0.05 −0.10 2.48 −2.89 −2.56 −2.66 −0.48

(t-statistic) (−1.96) (−0.14) (−0.28) (3.87) (−3.34) (−2.92) (−3.48) (−0.56)
BM 2 −0.45 −0.45 4.64 1.73 −2.22 −2.56 −5.59 −1.53
(t-statistic) (−1.26) (−1.35) (1.84) (2.59) (−2.91) (−3.09) (−3.53) (−1.67)
BM 3 −0.66 −0.25 0.15 1.71 −2.50 −2.74 −3.00 −1.30
(t-statistic) (−1.83) (−0.76) (0.31) (1.82) (−3.07) (−3.06) (−3.13) (−1.24)
Low −0.95 −1.02 0.22 4.06 −2.83 −3.78 −3.04 0.54
(t-statistic) (−2.33) (−2.66) (0.35) (3.01) (−3.27) (−3.92) (−3.28) (−0.46)
High 0.73 0.96 0.90 0.93 0.01 0.48 0.68 0.93
(t-statistic) (13.42) (28.36) (26.05) (15.35) (0.11) (4.13) (6.70) (8.18)
BM 2 0.77 0.88 1.68 1.11 0.15 0.27 2.47 1.18
(t-statistic) (23.02) (28.17) (7.06) (17.68) (1.44) (2.50) (11.76) (9.78)
BM 3 0.85 1.04 1.22 1.05 0.12 0.54 1.03 1.22
(t-statistic) (25.01) (33.16) (26.30) (11.84) (1.15) (4.55) (8.09) (8.74)
Low 0.91 1.19 1.32 1.18 0.16 0.76 1.34 1.86
(t-statistic) (23.45) (32.84) (21.67) (9.27) (1.43) (5.92) (10.92) (12.07)
High −0.14 −0.43 −0.43 −0.41
(t-statistic) (−1.10) (−3.38) (−3.82) (−3.26)
BM 2 −0.33 −0.39 1.97 −0.74
(t-statistic) (−2.99) (−3.26) (8.53) (−5.55)
77
78
BM 3 −0.40 −0.63 −0.89 −0.81
(t-statistic) (−3.35) (−4.83) (−6.35) (−5.32)
Low −0.52 −0.89 −1.23 −1.20
(t-statistic) (−4.10) (−6.37) (−9.06) (−7.06)
Adjusted R2 = 0.56
Adjusted R2 = 0.44
The table reports the Seemingly Unrelated Regressions (SUR) results for the following model:
Panel A. R(t) − RF(t) = a + hHML(t) + e(t)
Panel B. R(t) − RF(t) = a + sSMB(t) + e(t)
Panel C. R(t) − RF(t) = a + b[RM(t) − RF(t)]e(t)
Panel D. R(t) − RF(t) = a + sSMB(t) + hHML(t) + e(t)
At the end of June of each year t, all Information Technology stocks on the NYSE, AMEX, and NASDAQ are grouped into 16 portfolios based on their
QUANG-NGOC NGUYEN ET AL.

market capitalization (ln(ME)) and book-to-market ratio (ln(BM)). Ln(ME) and ln(BM) are the natural logarithms of ME (market equity) and BM
(book-to-market ratio), respectively. The ME breakpoints are the first, second, and third quartiles of the stocks’ ln(ME). Similarly, the BM breakpoints
are the first, second, and third quartiles of the stocks’ ln(BM). Big, ME2, ME3, and Small represent the biggest size quartile, second largest size
quartile, third largest size quartile and smallest size quartile, respectively. High, BM2, BM3, and Low represent the highest book-to-market quartile,
second highest book-to-market quartile, third highest book-to-market quartile, and the lowest high book-to-market quartile, respectively. Big-High
means the size – book-to-market portfolio whose stocks belong to the biggest size quartile and the highest book-to-market quartile.
SMB (small minus big) is the difference between the returns on small-stock and big-stock portfolios with about the same weighted average book-
to-market equity. HML (high minus low) is the difference between the returns on high and low book-to-market equity portfolios with about the same
weighted average size. RM is the value-weighted monthly% return on the stocks in the 16 size-BM portfolios. RF is the three-month Treasury bill
rate, observed at the beginning of the month.
A value with a t-statistic being greater than 1.98 or less than −1.98 is considered significant at the 5% level.
In addition, the SMB coefficient slopes increase monotonically from bigger to

smaller size quartiles within each book-to-market quartiles, except for the second
biggest book-to-market (BM2) quartile.
Although the adjusted R2 for the one-factor model where SMB is the only
explanatory variable is nearly twice as high as that of HML (0.30 for SMB
compared with 0.16 for HML), it is still far short of the ideal value of one. In
addition, the majority of the intercepts in Panel B are statistically distinguishable
from zero. Thus, SMB is significant in explaining variation in returns, but like
HML, it leaves much variation in returns that might be explained by other factors.
5.2.3. Regressions with One Explanatory Variable: RM–RF

In this section the covariance between the returns on the market portfolio with the
excess returns on the 16 size – book-to-market portfolios is considered. Panel C
in Table 8 shows that the excess return on the market portfolio of stocks, RM–RF,
captures variation in stock returns very well. The t-statistics on the slope coefficient
of RM–RF are all above seven, far greater than the critical value of approximately
1.98 at the 5% level of significance. It seems that RM–RF does not need an intercept
to explain expected stock returns. Only five out of the 16 size – book-to-market
portfolios have intercepts that are statistically different from zero (three from the
smallest size quartile and three from the lowest book-to-market quartile, the Small-
Low portfolio is the intersection between the smallest size quartile and the lowest
BM quartile). The adjusted R2 is 0.56, higher than those for SMB and HML.
However, the market still leaves much variation in stock returns that might be
explained by other factors.
5.2.4. Regressions with Two Explanatory Variables: SMB and HML

The combination of SMB and HML proves to explain stock returns very well. Only
one portfolio, Big-High, has the HML coefficient that is indistinguishable from 0.
Only four portfolios from the biggest size quartile have SMB coefficients that are
not statistically different from zero. These results are a significant improvement
from the one-factor model: when SMB is the only explanatory variable, five
portfolios have SMB coefficients that are indistinguishable from zero; and when
HML is the only explanatory variable, six portfolios have HML coefficients that are
indistinguishable from zero. However, SMB and HML together could not capture
all variation in stock returns. The adjusted R2 from the SUR regression of the size
– book-to-market portfolios’ returns on SMB and HML is only 0.44, leaving 56%
of variation in returns unexplained. That unexplained variation is captured by the
market factor, as we shall see in the next section.
5.2.5. Regressions with Three Explanatory Variables: SMB, HML, and RM–RF
Table 9 shows the SUR estimates of the intercepts and coefficients on SMB, HML,
and RM for the 16 size – book-to-market portfolios. With the exception of three
portfolios, the Small-High, Small-Low, and ME2-Low portfolios, the intercepts for
the remaining 13 size – book-to-market portfolios are indistinguishable from zero
and have statistically insignificant t-statistics ranging from −1.40 (ME3–BM2)
to 1.10 (Small–BM3). Nearly all the slope coefficients on SMB are significant
and wit only one exception have t-statistics that lie outside the [−1.98, 1.98]
interval. The one exception is the ME2-BM3 portfolio, which returns a t-statistic
of −1.84. This quantity, nevertheless, is very close to −1.98. More interestingly,
the slope on SMB appears related to size. In three of the four book-to-market
quartiles (the exception is the second highest book-to-market quartile (BM2))
the slope on SMB increases monotonically from bigger- to smaller-size quartiles.
The explanatory power of HML in stock returns is somewhat weaker than that
of SMB. Six out of the 16 size – book-to-market portfolios return t-statistics
are within the [−1.98, 1.98] interval. The weak explanatory power, which is in
contrast with a strong relationship between stock returns and HML where HML
is the only explanatory variable, is due to the interaction between SMB, HML,
and RM–RF. Table 4 reports that the correlation coefficients between HML and
SMB, and HML and RM–RF are 0.34 and −0.13, respectively. Therefore, the
apparently weak power of HML in the three-factor model does not invalidate
itself as a significant variable that captures stock returns. This is evidenced in the
one-factor model where HML is the only explanatory variable.
As further evidence for the role of HML in mimicking book-to-market ratio, the
slope on HML tends to decrease from higher- to smaller- BM portfolios. The two
portfolios that do not follow this tendency are ME2-BM2 and ME3-BM3. The
slope on RM–RF (or the market factor) are all more than six standard errors from
zero. The market factor, either in combination with the size and book-to-market
factors or standing alone, strongly captures variation in portfolio returns. The
adjusted R2 for the three-factor model scores an impressive 0.83, much higher
than those reported for the one-factor models (0.16 for HML, 0.30 for SMB, and
0.56 for RM–RF, Table 9). This is a strong indication that the three-factor model
is a good model that describes the relationship between size, book-to-market
ratio, beta, and stock returns well.
5.2.6. Does the Use of the 3-Month Treasury Bill Rate Matter?
In Section 4.2., it was argued that the use of the 3-month Treasury bill rate
(USTBL3M) instead of the 1-month Treasury bill rate (USTBL1M) would lead
to the same inference as if the 1-month rate was used. This section empirically
improves that statement. As the data for the 1-month Treasury bill rate on
Table 9. Seemingly Unrelated Regressions of Portfolios’ Excess Returns on

SMB, HML, and RM–RF in the Three-Factor Model: July 1990–June 2001.
Book-to-Market Quartiles Seemingly Unrelated Regressions (SUR) Results
Size Quartiles
Big ME 2 ME 3 Small
Intercepts
High −0.44 0.18 −0.34 1.72
(t-statistic) (−0.88) (0.50) (−0.95) (2.91)
BM 2 0.22 0.10 −1.50 0.98
(t-statistic) (0.80) (0.38) (−1.40) (1.75)
BM 3 0.18 0.14 −0.03 0.93
(t-statistic) (0.90) (0.48) (−0.06) (1.10)
Low −0.01 −0.65 −0.14 2.49
(t-statistic) (−0.05) (−2.20) (−0.37) (2.36)
SMB coefficients
High −0.51 −0.11 0.18 0.46
(t-statistic) (−7.32) (−2.09) (3.60) (5.54)
BM 2 −0.38 −0.30 1.59 0.65
(t-statistic) (−9.91) (−7.67) (10.53) (8.20)
BM 3 −0.45 −0.08 0.39 0.74
(t-statistic) (−15.66) (−1.84) (6.82) (6.13)
Low −0.44 0.09 0.72 1.44
(t-statistic) (−12.03) (2.07) (13.57) (9.71)
HML coefficients
High 0.22 −0.03 −0.08 −0.08
(t-statistic) (3.09) (−0.50) (−1.56) (−0.96)
BM 2 0.03 0.00 2.57 −0.37
(t-statistic) (0.72) (0.03) (16.43) (−4.49)
BM 3 0.00 −0.20 −0.45 −0.48
(t-statistic) (0.02) (−4.65) (−7.51) (−3.88)
Low −0.10 −0.43 −0.80 −0.91
(t-statistic) (−2.64) (−9.98) (−14.45) (−5.89)
RM–RF coefficients
High 0.89 0.99 0.84 0.80
(t-statistic) (17.39) (26.72) (22.86) (13.14)
BM 2 0.88 0.96 1.48 0.90
(t-statistic) (31.85) (34.15) (13.43) (15.82)
BM 3 0.97 1.04 1.07 0.81
(t-statistic) (46.26) (33.87) (25.76) (9.26)
Low 1.02 1.13 1.05 0.70
(t-statistic) (38.28) (37.30) (27.14) (6.53)
Adjusted R2 0.83
The table reports the Seemingly Unrelated Regressions (SUR) results for the three-factor model:
market ratio (ln(BM)). ME is common equity, which is denominated in millions of U.S. dollars and
is measured at the end of June of year t. BM is the ratio of book equity to market equity, with book
equity measured at fiscal yearend falling in year t − 1, and market equity measured at the end of
December of year t − 1. ln(ME) and ln(BM) are the natural logarithms of ME and BM, respectively.
The ME breakpoints are the first, second, and third quartiles of the stocks’ ln(ME). Similarly, the BM
breakpoints are the first, second, and third quartiles of the stocks’ ln(BM). Big, ME2, ME3, and Small
represent the biggest size quartile, second largest size quartile, third largest size quartile and smallest
size quartile, respectively. High, BM2, BM3, and Low represent the highest book-to-market quartile,
second highest book-to-market quartile, third highest book-to-market quartile, and the lowest high
book-to-market quartile, respectively.
SMB (small minus big) is the difference between the returns on small-stock and big-stock portfolios
with about the same weighted average book-to-market equity. HML (high minus low) is the difference
between the returns on high and low book-to-market equity portfolios with about the same weighted
average size. RM is the value-weighted monthly% return on the stocks in the 16 size-BM portfolios.
RF is the three-month Treasury bill rate, observed at the beginning of each month.
DataStream is not available after May 1996, the sample period used for the
comparison between the effects of these two rates is from July 1990 to May 1996.
Employing the SUR procedure, two time-series regressions of the size – book-
to-market portfolios’ excess returns on the size (SMB), book-to-market (HML),
and market factor (RM–RF) are undertaken. The size factor, SMB, and the book-to-
market factor (HML) are the same in these two regressions. Only the excess returns
on the size – book-to-market portfolios and the market portfolio are different: in
one regression, the 1-month rate is used; in the other, the 3-month rate is used.
Table 10 reports the results of the two regressions.
It is clear that the coefficients on SMB, HML, and RM–RF produced by the use
of the 1-month rate are nearly identical to those produced by the use of the 3-month
rate. If there is any difference, the difference is only 0.01%. The t-test for the mean
difference between the 1-month rate SMB coefficients and the 3-month rate SMB
coefficients returns a t-statistic of 0.42, which is not significant at the 5% level to
reject the null hypothesis of no difference. Similarly, the results of the t-tests for
the mean difference between the 1-month rate HML coefficients and the 3-month
rate HML coefficients, and between the 1-month rate RM–RF coefficients and
the 3-month rate RM–RF coefficients show no sign of difference (in a statistical
sense). Thus, the use of the 3-month Treasury bill rate in the preceding sections
would not invalidate the conclusion that the three-factor model is well specified
for information technology stocks on the NYSE, AMEX, and NASDAQ.
Table 10. SUR Results Using 1-Month and 3-Month Treasury Bill Rates:
July 1990–May 1996.
Portfolios SMB Coefficients HML Coefficients RM–RF Coefficients
USTBL1M USTBL3M USTBL1M USTBL3M USTBL1M USTBL3M
Big-High −0.39 −0.39 −0.05 −0.05 1.21 1.21

Big-BM2 −0.49 −0.49 0.15 0.15 0.80 0.81
Big-BM3 −0.37 −0.37 −0.06 −0.06 0.99 0.99
Big-Low −0.32 −0.32 −0.14 −0.14 1.01 1.01
ME2-High −0.30 −0.30 0.03 0.03 0.97 0.97
ME2-BM2 −0.33 −0.33 −0.05 −0.05 1.03 1.03
ME2-BM3 −0.03 −0.03 −0.30 −0.30 1.12 1.11
ME2-Low 0.05 0.05 −0.38 −0.38 1.10 1.09
ME3-High 0.06 0.06 −0.13 −0.13 0.98 0.98
ME3-BM2 2.25 2.25 2.52 2.52 1.26 1.25
ME3-BM3 0.12 0.12 −0.26 −0.26 1.03 1.03
ME3-Low 0.80 0.80 −0.91 −0.91 1.08 1.08
Small-High 0.16 0.16 −0.05 −0.05 0.90 0.90
Small-BM2 0.47 0.46 −0.31 −0.31 0.89 0.90
Small-BM3 0.06 0.06 −0.01 −0.02 0.73 0.74
Small-Low 1.49 1.48 −0.96 −0.96 0.57 0.56
t-Statistics* 0.42 −0.50 0.10
t-Critical two-tail** 2.13 2.13 2.13
The table reports the results for the Seemingly Unrelated Regressions of the size – book-to-market
portfolios’ excess returns on SMB, HML, and RM–RF, with the excess returns are measured in one of
the two ways: (1) using the 1-month treasury bill rate (USTBL1M), and (2) using the 3-month treasury
bill rate (USTBL3M).
SMB (small minus big) is the difference between the returns on small-stock and big-stock portfolios with
about the same weighted average book-to-market equity. HML (high minus low) is the difference be-
tween the returns on high and low book-to-market equity portfolios with about the same weighted aver-
age size. RM is the value-weighted monthly% return on the stocks in the 16 size-BM portfolios. RF is the
three-month Treasury bill rate or the one-month Treasury bill rate, depending on each of the approach.
The portfolios are formed as follows. At the end of June of each year t, all Information
Technology stocks on the NYSE, AMEX, and NASDAQ are grouped into 16 portfolios based on
their market capitalization (ln(ME)) and book-to-market ratio (ln(BM)). ME is common equity, which
is denominated in millions of U.S. dollars and is measured at the end of June of year t. BM is the ratio
of book equity to market equity, with book equity measured at fiscal yearend falling in year t − 1, and
market equity measured at the end of December of year t − 1.The ME breakpoints are the first (25%),
second (50%), and third (75%) quartiles of the stocks’ ln(ME). Similarly, the BM breakpoints are the
first, second, and third quartiles of the stocks’ ln(BM). Stocks with negative BM are excluded so that
ln(BM) is calculable.
∗ Tests for mean difference between two-matched samples.
∗∗ At the 5% level of significance.
Table 11. Pre-Crash and Post-Crash Returns.

Portfolios Mean
Pre-Crash Post-Crash
Panel A. The explanatory variables

SMB 1.31 0.57
(t-statistic) (1.88) (0.21)
HML 1.25 2.92
(t-statistic) (1.92) (1.31)
RM–RF −1.90 −8.09
(t-statistic) (−2.44) (−1.59)
Panel B. The dependent variables
Big-High −2.62 −6.81
(t-statistic) (−3.06** ) (−1.89)
Big-BM2 −1.94 −7.28
(t-statistic) (2.78** ) (−1.70)
Big-BM3 −2.32 −7.52
(t-statistic) (−3.12** ) (−1.60)
Big-Low −2.41 −10.67
(t-statistic) (−3.09** ) (−2.10* )
ME2-High −2.19 −5.32
(t-statistic) (−2.54* ) (1.09)
ME2-BM2 −2.15 −7.26
(t-statistic) (−2.66** ) (−1.80)
ME2-BM3 −2.23 −8.55
(t-statistic) (−2.66** ) (−1.47)
ME2-Low −3.27 −10.26
(t-statistic) (−3.57** ) (−1.45)
ME3-High −1.82 −6.87
(t-statistic) (−2.24* ) (−1.59)
ME3-BM2 1.09 −6.24
(t-statistic) (0.35) (−1.11)
ME3-BM3 −2.31 −8.01
(t-statistic) (−2.37* ) (−1.06)
ME3-Low −2.51 −8.02
(t-statistic) (−2.25* ) (0.97)
Small-High 0.96 −6.65
(t-statistic) (0.95) (−1.55)
Small-BM2 −0.35 −7.14
(t-statistic) (−0.34) (−1.04)
Small-BM3 −0.16 −7.35
(t-statistic) (−0.12) (−1.46)
Small-Low 2.26 −8.09
(t-statistic) (1.39) (−1.07)

The table compares the mean values of SMB, HML, RM–RF (Panel A), and excess returns on the size
– book-to-market portfolios (Panel B) before and after the technology crash in April 2000. Pre-crash
values are for the July 1990 – March 2000 period. Post-crash values are for the May 2000–June 2001
period. The month when the crash occurred (April 2000) is excluded.
with about the same weighted average book-to-market equity. HML (high minus low) is the
difference between the returns on high and low book-to-market equity portfolios with about the
same weighted average size. RM is the value-weighted monthly% return on the stocks in the 16
size-BM portfolios. RF is the three-month Treasury bill rate, observed at the beginning of each
month.
Portfolios’ excess returns are spreads of portfolios’ monthly returns over the three-month Treasury
bill rates. The size – book-to-market portfolios are formed as follows. At the end of June of each year t,
all Information Technology stocks on the NYSE, AMEX, and NASDAQ are grouped into 16 portfolios
based on their market capitalization (ln(ME)) and book-to-market ratio (ln(BM)). ME is common
equity, which is denominated in millions of U.S. dollars and is measured at the end of June of year t.
BM is the ratio of book equity to market equity, with book equity measured at fiscal yearend falling
in year t − 1, and market equity measured at the end of December of year t − 1.The ME breakpoints
are the first (25%), second (50%), and third (75%) quartiles of the stocks’ ln(ME). Similarly, the BM
breakpoints are the first, second, and third quartiles of the stocks’ ln(BM). Stocks with negative BM
are excluded so that ln(BM) is calculable.
The reported means (in percentage) are the time-series averages of the monthly values of the
corresponding variables.
∗∗ Significant at the 0.01 level.
5.3. Information Technology Stock Behavior Before and After

the Technology Crash in April 2000
This section investigates how well the three-factor model describes the size and
book-to-market effects in two sub-periods, the period before the technology stock
crash in April 2000 (the pre-crash period), and the period after the crash (the
post-crash period). Section 5.3.1. reports the descriptive statistics for the pre-crash
and after-crash variables. Section 5.3.2. presents the regressions results for each
of the period.
5.3.1. Inputs for the Three-Factor Model in the Periods Before (July 1990–March
2000) and After the Technology Crash (May 2000–June 2001)
Table 11 presents the descriptive statistics for the SMB, HML, and RM–RF
variables (Panel A) as well as for the size – book-to-market portfolios’ excess
returns (Panel B) for the July 1990–March 2000 period (pre-crash period) and
the May 2000–June 2001 period (post-crash period). The month when the
crash occurs, April, is excluded to separate the immediate effect of the crash.
Table 11 reports that most of the (dependent and independent) variables have
the post-crash means (i.e. the means for the May 2000–June 2001 period) that
are indistinguishable from zero. The post-crash means for SMB, HML, and
RM–RF are 0.57%, 2.92%, and –8.09%, with t-statistics of 0.21, 1.31, and –1.59,
respectively (Panel A). The post-crash means of the excess returns on the size –
book-to-market portfolios also have t-statistics that are less than 2 standard errors
from 0, except for the Big-Low portfolio (Panel B).
In contrast, (in Panel B) 11 out of the 16 size – book-to-market portfolios
have pre-crash mean excess returns (i.e. the mean excess returns for the July
1990–March 2000) that are statistically different from zero. The five portfolios
whose pre-crash mean excess returns are indistinguishable from zero are the
ME3-BM2 portfolio and those from the smallest – size quartile (the bottom four
portfolios). The pre-crash mean for the market excess returns (RM–RF) is also
different from zero (t-statistic of −2.44) (Panel A). Although SMB and HML
have the pre-crash means that are not different from zero at the 5% level of
significance, they are on average significant at the 10% level. The t-statistics for
the pre-crash means of these two variables, SMB and HML, are 1.88 and 1.92,
respectively.
5.3.2. Regression Results (Using the Three-Factor Model)

The regressions results for the pre-crash and post-crash periods are reported in
Table 12. In the post-crash period from May 2000 to June 2001, there are 16 cross-
sectional observations (16 portfolios) and 14 time-series observations (14 months).
Because the SUR regression is not applicable in circumstances where there are
more cross-sectional observations than time-series observations, the Ordinary
Least Squares (OLS) regression is used instead. For comparison purpose, the OLS
regression is also applied to the pre-crash period.6 The estimated pre-crash and
post-crash coefficients are then compared with each other to determine whether
there is any change in the size and book-to-market effects after the technology
crash.
Three t-tests (for matched sample) for the mean difference between the pre-
crash and post-crash SMB coefficients, between the pre-crash and post-crash HML
coefficients, and between the pre-crash and post-crash RM–RF coefficients are
conducted. The values of the t-statistics for these tests are reported in Table 12.
The mean difference tests indicate that on average the pre-crash coefficients and
their post-crash counterparts are the same. All the three t-statistics are well below
the t-critical value in magnitude. The highest (in magnitude) is the t-statistic
for the difference between the pre-crash HML coefficients and post-crash HML
coefficients, is just −0.47, while the critical value at the 5% level of significance
for a two-tail test is 2.13.
Table 12. Ordinary Least Squares (OLS) Regression Results on the

Three-Factor Model with Pre-Crash (July 1990–March 2000) and Post-Crash
Data (May 2000–June 2001).
Portfolios SMB Coefficients HML Coefficients RM–RF Coefficients
Pre-Crash Post-Crash Pre-Crash Post-Crash Pre-Crash Post-Crash
Big-High −0.48 −0.68 0.19 −0.78 0.93 0.46

Big-BM2 −0.39 −0.33 0.04 −0.15 0.87 0.83
Big-BM3 −0.42 −0.61 −0.01 −0.40 0.97 0.85
Big-Low −0.38 −0.71 −0.13 −0.39 0.99 0.95
ME2-High −0.14 0.29 −0.06 1.13 1.06 1.31
ME2-BM2 −0.28 −0.19 −0.06 0.57 1.02 1.05
ME2-BM3 −0.13 0.19 −0.17 0.40 1.04 1.25
ME2-Low 0.02 0.22 −0.37 −0.43 1.11 1.15
ME3-High 0.14 0.52 −0.11 0.64 0.91 0.95
ME3-BM2 1.70 0.67 2.65 0.56 1.30 1.12
ME3-BM3 0.29 0.67 −0.35 −0.45 1.06 1.09
ME3-Low 0.75 0.68 −0.82 −0.52 1.06 1.21
Small-High 0.48 0.58 −0.15 0.45 0.85 0.84
Small-BM2 0.54 1.13 −0.30 0.17 0.92 1.03
Small-BM3 0.86 0.71 −0.67 0.40 0.93 0.92
Small-Low 1.52 1.29 −0.99 −0.99 0.74 0.58
t-Statistics* −0.22 −0.47 −0.20
t-Critical two-tail** 2.13 2.13 2.13
The table reports the results for the Cross-Section Weighting regressions of the size – book-to-market
portfolios’ excess returns on SMB, HML, and RM–RF, using pre-crash data and post-crash data. Pre-
crash data are the data for the July 1990–March 2000 period. Post-crash data are the data for the May
2000–June 2001 period. The term “crash” implies the technology crash, which occurred in April 2000.
The model for the regressions is:
with about the same weighted average book-to-market equity. HML (high minus low) is the difference
between the returns on high and low book-to-market equity portfolios with about the same weighted
average size. RM is the value-weighted monthly% return on the stocks in the 16 size-BM portfolios.
RF is the three-month Treasury bill rate, observed at the beginning of each month.
The portfolios are formed as follows. At the end of June of each year t, all Information
Technology stocks on the NYSE, AMEX, and NASDAQ are grouped into 16 portfolios based on
their market capitalization (ln(ME)) and book-to-market ratio (ln(BM)). ME is common equity, which
is denominated in millions of U.S. dollars and is measured at the end of June of year t. BM is the ratio
of book equity to market equity, with book equity measured at fiscal yearend falling in year t − 1, and
market equity measured at the end of December of year t − 1.The ME breakpoints are the first (25%),
second (50%), and third (75%) quartiles of the stocks’ ln(ME). Similarly, the BM breakpoints are the
first, second, and third quartiles of the stocks’ ln(BM). Stocks with negative BM are excluded so that
ln(BM) is calculable.
∗ Tests for mean difference between two-matched samples.
∗∗ At the 5% level of significance.
6. CONCLUSION
This paper explores the relationship between size, book-to-market, beta, and
expected stock returns in the Information Technology sector. In particular, the
paper uses two models to test this relationship. The first model assumes stock
returns are related to the direct measure of size, book-to-market, and beta (the
multivariate model). The second model assumes stock returns are related to returns
on the three portfolios that are designed to mimic the size, book-to-market, and
beta factors (the three-factor model).
The risk-return tests using the multivariate model indicate that the book-to-
market effect is present in stock returns. However, this effect only exists at
the 10% level of significance in contrast to the more significant relationship
documented by Fama and French (1992) for all non-financial stocks on the NYSE,
AMEX, and NASDAQ over the 1963–1990 period. Thus, the book-to-market
effect for IT stocks, though present, is not robust over the period 1997–2001.
Except for discovering a weak book-to-market effect, the risk-return tests using
the multivariate model find no relationship between size, beta and stock returns.
Although Fama and French (1992) also document a weak explanatory power of
beta, they find that size is significant and that size and book-to-market ratio in
combination absorb variation in stock returns. The different results documented in
Fama and French’s (1992) paper and this paper might be due to different research
designs employed by the two papers. However, research design seems to be playing
a trivial part since the results from the FM regressions on individual stocks and on
portfolios (in this paper) both lead to similar conclusions. Thus, the relationship
between size, book-to-market, beta and stock returns in IT stocks appears to be
different from that previously observed in non-financial stocks.
The risk-return tests using the three-factor model show that expected returns
are strongly related to the size, book-to-market, and market factors. Unlike the
multivariate model where size and book-to-market factors are directly measured
from the market capitalization and book-equity to market-equity ratio of a stock,
the size and book-to-market factors in the three-factor model are represented
by the returns on the portfolios that are designed to mimic stocks’ market
capitalisation and book-equity to market-equity ratio. The results show that the
three-factor model works very well for IT stocks over the July 1990–June 2001
period as well as the sub-periods before (July 1990–March 2000) and after the
technology crash (May 2001–June 2001) in April 2000.
Why do the multivariate model and the three-factor model produce different
results? One possible explanation is that the multivariate model uses annual
data while the three-factor model uses monthly data. We argue that the use of
monthly data instead of annual data in the three-factor model is not a serious
problem that could lead to any spurious conclusions. Examining whether the
three-factor model produces different results due to the alternative uses of annual
data and monthly data could provide a basis for future research. Another possible
reason for the differing results produced by the two models is that they are
simply not comparable. SMB, HML, and RM–RF might not be good proxies for
direct measurements of size (ln(ME)), book-to-market ratio (ln(BM)), and beta
respectively. In other words, the mimicking returns for the size (SMB), book-to-
market (HML), and market factors (RM–RF), might not be good indicators of
movements in the characteristics (ln(ME), ln(BM), and beta) of stocks.
The sub-period results (the periods before and after the technology crash in
April 2000) show that the nature of the relationship between stock returns, size,
book-to-market, and market factors, or the magnitude of the size, book-to-market,
and market premiums, is on average unchanged for both of the sub-periods. The
coefficients on SMB, HML, and RM–RF for the July 1990–March 2000 (the pre-
crash coefficients) are not statistically different from those for the May 2000–June
2001 period. This is particularly interesting since this finding is inconsistent with
the view that the technology stock crash in April 2000 was a correction of stock
prices. As it has been known, the factors in the three-factor model are the “returns”
on some specially designed portfolios. It is possible that the behavior of a return on a
particular portfolio has been changed after the technology crash (in fact it has as the
statistics for SMB, HML, RM–RF, and the size – book-to-market portfolios’ excess
returns have changed from the pre-crash period to post-crash period, Table 12);
but because all the returns have changed by the same magnitude, the relationship
between these returns is unchanged. This conjecture, if true, would offer great
support for the three-factor model, since it could work over any time frame and be
used for tests of structural change (i.e. relationship change) after an event.
This paper makes three important contributions to theory as well as practice.
First, it provides more evidence on the risk-return relationship in the information
technology sector. The evidence in this paper brings a step towards finding a
satisfactory explanation for the deviation from the CAPM. Second, the paper
documents in detail the behavior of IT stock prices before and after the technology
stock crash in April 2000. This information could be used as a reference for
investment decisions. And finally, the paper opens a new direction for future
research in the asset-pricing field: looking for (the risk-return) evidence in
individual sectors rather than in markets as a whole.
NOTES
1. “Old Economy” stocks are those other than IT stocks.
2. Note that a firm might not have beta for a certain period but might have it at a different
time. We account for this by including the firm when its beta is available and excluding it
when its beta is unavailable. This is done subject to whether the firm fits into our portfolio
selection procedure (to be described in the methodology section).
3. A testing year is from July of year t to June of year t + 1.
4. It makes more sense to establish a relationship between SMB, HML, and systematic
risk than to find a relationship between size, book-to-market equity, and systematic risk.
This is due to the fact that size and book-to-market equity are company-specific factors.
5. Data on the 1-month Treasury bill rate in DataStream is not available after May 1996.
6. Note that the SUR and OLS techniques produce identical estimators but different
t-statistics for these estimators. As this section is only concerned with the differences in the
pre-crash and after-crash coefficients, the use of OLS is as significant as the use of SUR.
REFERENCES
Aldous, D. (1989). Probability approximations via the Poisson clumping heuristic. New York: Springer.
Alford, A. W., Jones, J. J., & Zmijewski, M. E. (1994). Extension and violation of the statutory SEC
form 10-K filing requirements. Journal of Accounting and Economics, 17, 229–254.
Banz, R. W. (1981). The relationship between return and market value of common stocks. Journal of
Financial Economics, 9, 3–18.
Barber, B. M., & Lyon, J. D. (1997a). Detecting long-run abnormal stock returns: The empirical power
and specification of test statistics. Journal of Financial Economics, 43, 341–372.
Barber, B. M., & Lyon, J. D. (1997b). Firm size, book-to-market ratio, and security returns: A holdout
sample of financial firms. Journal of Finance, 52, 875–883.
Black, F. (1993). Beta and return. Journal of Portfolio Management, 20, 8–18.
Black, F., Jensen, M. C., & Scholes, M. (1972). The capital asset pricing model: Some empirical tests.
In: M. Jensen (Ed.), Studies in the Theory of Capital Market. New York: Praeger.
Capaul, C., Rowley, I., & Sharpe, W. F. (1993). International value and growth stock returns. Financial
Analysts Journal, 49, 27–36.
Chan, K. C., & Chen, N. F. (1988). An unconditional asset-pricing test and the role of firm size as an
instrumental variable for risk. Journal of Finance, 43, 309–325.
Chan, L. K. C., Hamao, Y., & Lakonishok, J. (1991). Fundamentals and stock returns in Japan. Journal
of Finance, 46, 1739–1789.
Chui, A. C. W., & Wei, K. C. J. (1998). Book-to-market, firm size, and the turn-of-the-year effect:
Evidence from Pacific-Basin emerging markets. Pacific-Basin Finance Journal, 6, 275–293.
Daniel, K., & Titman, S. (1997). Evidence on the characteristics of cross sectional variation in stock
returns. Journal of Finance, 52, 1–33.
Davis, J. (1994). The cross-section of realized stock returns: The pre-COMPUSTAT evidence. Journal
of Finance, 49, 1579–1593.
Davis, J. L., Fama, E. F., & French, K. R. (2000). Characteristics, covariances, and average returns:
1929 to 1997. Journal of Finance, 55, 389–406.
DeBondt, W. F. M., & Thaler, R. H. (1987). Further evidence on investor overreaction and stock market
seasonality. Journal of Finance, 42, 557–581.
Dimson, E. (1979). Risk measurement when shares are subject to infrequent trading. Journal of
Fama, E. F., & French, K. R. (1992). The cross-section of expected stock returns. Journal of Finance,
47, 427–465.
Fama, E. F., & French, K. R. (1993). Common risk factors in the returns on stocks and bonds. Journal
Fama, E. F., & French, K. R. (1995). Size and book-to-market factors in earnings and returns. Journal
of Finance, 50, 131–155.
Fama, E. F., & French, K. R. (1996). Multifactor explanations of asset pricing anomalies. Journal of
Finance, 51, 55–84.
Fama, E. F., & French, K. R. (1997). Industry costs of equity. Journal of Financial Economics, 43,
153–193.
Fama, E. F., & MacBeth, J. D. (1973). Risk, return, and equilibrium: Empirical tests. The Journal of
Political Economy, 81, 607–636.
Franklin, C. (2000, September). Incubating real options: Can option pricing theory help investors make
sense of the business models and valuations of internet incubators? Working Paper, University
of Stirling.
Haugen, R. (1995). The new finance: The case against efficient markets. Englewood Cliffs, NJ: Prentice-
Hall.
Jegadeesh, N. (1992). Does market risk really explain the size effect. Journal of Financial and
Quantitative Analysis, 27, 337–351.
Keim, D. B. (1983). Size-related anomalies and stock return seasonality: Further empirical evidence.
Kothari, S. P., Shanken, J., & Sloan, R. (1995). Another look at the cross-section of expected stock
returns. Journal of Finance, 50, 185–224.
Lakonishok, J., Shleifer, A., & Vishny, R. W. (1994). Contrarian investment, extrapolation, and risk.
Lintner, J. (1965). The valuation of risk assets and the selection of risky investments in stock portfolios
and capital budgets. Review of Economics and Statistics, 47, 13–37.
Lo, A. W., & MacKinlay, A. C. (1990a). When are contrarian profits due to stock market overreaction?
Review of Financial Studies, 3, 175–205.
Lo, A. W., & MacKinlay, A. C. (1990b). Data-snooping biases in tests of financial asset pricing models.
Review of Financial Studies, 3, 431–467.
MacKinlay, A. C. (1995). Multifactor models do not explain deviations from the CAPM. Journal of
Mech, T. (1993). Portfolio return autocorrelation. Journal of Financial Economics, 34, 307–344.
Newey, W., & West, K. (1987). A simple, heteroskedastic and autocorrelation consistent covariance
matrix. Econometrica, 55, 703–708.
Rajgopal, S., Kotha, S., & Venkatachalam, M. (2000, January). The relevance of web traffic for internet
stock prices. Research Paper, Stanford University.
Rosenberg, B., Reid, K., & Lanstein, R. (1985). Persuasive evidence of market inefficiency. Journal of
Sadorsky, P. (2003). The macroeconomic determinants of technology stock price volatility. Review of
Scholes, M., & Williams, J. (1977). Estimating betas from nonsynchronous data. Journal of Financial
Economics, 5, 309–327.
Sharpe, W. F. (1964). Capital asset prices: A theory of market equilibrium under conditions of risk.
Trueman, B., Wong, F., & Zhang, X. (2001). Back to basics: Forecasting the revenues of internet firms.
Review of Accounting Studies, 6, 305–329.
IMPLIED VOLATILITIES
AND AUDITOR REPUTATION:
THE ANDERSEN CASE
Jonathan M. Godbey and James W. Mahar
ABSTRACT
Audits are a means of reducing the information asymmetry between managers
and investors. If the quality of the audit is in question, outside investors may
face a larger informational disadvantage. We test the hypothesis that this
informational disadvantage is manifested in the implied volatilities associated
with the equity options of the audited firms. We find that volatilities increased
for Andersen audited firms relative to firms audited by other Big Five
accounting firms. This finding is consistent with the view that auditors help
lessen the information asymmetry problem and that some of this reduction is
accomplished by auditor reputation.
1. INTRODUCTION
The reputation of a firm’s auditor is widely seen as an important determinant
in signaling audit quality to investors. In literature, the importance of auditors
and their reputation is shown by Watts and Zimmerman (1983). This importance
has subsequently been examined and discussed in both finance and accounting
literature. In the financial literature, much of the research into auditor reputation has
focused on IPO underpricing and the prestige of the firms’ auditor. The accounting
Research in Finance
ISSN: 0196-3821/doi:10.1016/S0196-3821(04)21004-8
93
94 JONATHAN M. GODBEY AND JAMES W. MAHAR
literature shows a consistent view, but from a different perspective, focusing on the
relationship between audit pricing and reputation as well as event-studies showing
market reaction as a function of auditor reputation. A common explanation for
much of this literature is that auditor reputation serves as a means of reducing
information asymmetries between investors and the firm’s management.
The Enron-related troubles that befell Arthur Andersen (Andersen) provide a
unique opportunity to test the importance of auditor reputation. Over the course
of approximately six months, Andersen went from being a highly respected firm
to being a symbol of corporate malfeasance. The importance of the decline of
Andersen on client stock prices has been examined previously by Chaney and
Philipich (C&P, 2002) who report that firms audited by Andersen experienced
statistically significant price declines “surrounding dates on which Andersen’s
audit procedures and independence were under severe scrutiny.” (p. 1221) There
are two possible explanations for the market-adjusted price decline: (1) a decrease
in expected cash flow of Andersen audited clients, that is, a change in the mean of
the cash flow distribution; and (2) an increase in risk stemming from either higher
information asymmetries due to poor audit quality that would allow mangers to
hide important information from investors, or an increased likelihood of incurring
the costs of changing auditors.
One way to gauge whether investors believe that the risk of Andersen audited
firms had increased is to examine the volatility implied by the pricing of equity
options. Merton (1973), Latane and Rendlemand (1976), and Hull and White
(1987) document that the implied volatility of a particular option represents the
anticipated risk over the life of the option. Information asymmetries have also
been shown to play an important role in the option pricing. For example Patell and
Wolfson (1981) show that the implied volatility on equity options drops following
firm’s earnings releases. As pointed out by Merton (1987) and later discussed by
Bellalah and Aboura (2002), as the amount of, or quality of, available information
decreases, the implied volatility, or risk, increases, which in turn increases the
required rate of return and decreases the stock price, raising the cost of capital.
In this paper, we extend the findings of C&P (2002) by examining how option
market participants use auditor reputation in their pricing of equity options. If
auditor reputation affects the informational asymmetry between investors and the
firm, then the events that undermined Andersen’s reputation should be associated
with an increase in the implied volatility of the options of Andersen audited clients.
If auditor reputation is unimportant, then there should be no change in the implied
volatilities relative to that of firms audited by other Big Five accounting firms.
For both the short and long time periods, our results demonstrate that the implied
volatility at S&P 500 firms audited by Andersen increased significantly relative
to firms audited by other Big Five accounting firms during the time period in
Implied Volatilities and Auditor Reputation 95
which Enron and Andersen actions became publicly known. Specific events where
the average implied volatility increased include: (1) the restatement of Enron’s
financial statements and the announcement that the SEC was investigating the case
(November 8, 2001); (2) the Congressional testimony of Andersen CEO Joseph
Berardino (December 12, 2001); (3) Andersen’s admission to the shredding of
documents while under SEC investigation (January 10, 2002); (4) the dual events of
the release of the Powers Report and Andersen’s hiring of former Federal Reserve
Chairman Paul Volcker (February 2 and 3, 2002); and (5) the admission by David
Duncan that he had ordered Enron documents to be destroyed. These findings are
consistent with a view that the quality of auditor reputation is an important factor
in reducing the information asymmetry problem.
The remainder of the paper is organized as follows: Section 2 provides a review
of existing research and an introduction to implied volatilities, Section 3 provides
a description of our data and methodology, Section 4 examines the results of the
tests, and Section 5 gives a summary and conclusion.
2. LITERATURE REVIEW
Auditor reputation is an important means of signaling the quality of a client’s
financial reporting and thereby reducing the information asymmetry problem.
This area of research was established by Watts and Zimmerman (1983), but
has been examined extensively in both financial and accounting literature. The
common thread through this auditor reputation literature is that investors use
auditor reputation as a signal of quality that reduces informational asymmetry.
In financial literature, much of the auditor reputation work has been done within
an IPO framework. Titman and Trueman (1986) find that audit quality can help
explain IPO underpricing. Similarly, Balvers, McDonald and Miller (1988), and
Beatty (1989) find less IPO underpricing when a Big 8 auditor is used. More
recent work by Clarkson and Simunic (1994) find that there is less underpricing
of Canadian IPOs when a high quality auditor is used.
The importance of auditor reputation has also been shown by looking at the
pricing of audits. A partial list of those who have examined the relationship between
auditor quality and their pricing includes Francis and Simon (1987), Beatty (1993),
Dye (1993), and Craswell, Francis and Taylor (1995) who all find that audit firms
with higher perceived reputation, price their audits higher than those with lower
perceived quality. This pricing may be partially due to the fact that larger (and more
highly perceived) auditors have more assets (deeper pockets) and can therefore
better withstand potential liabilities in the event of a lawsuit. However, Datar and
Alles (1999) show that audit firms understand the importance of their reputation
and strive to maintain their reputational capital even when legal liability is reduced
by their financial assets.
Auditor reputation has also been investigated with event-study type tests that
measure the market reaction for clients audited by the same auditor. Teoh and
Wong (1993) find that the perceived audit quality influences the market’s response
to earnings announcements. Franz and Crawford (1998) document that audit clients
that are not involved with auditor related litigation lose market value as a result
of litigation brought against their auditor which is seen as damaging the auditor’s
reputation. Moreover, C&P (2002) report that firms audited by Arthur Andersen
experienced market-adjusted stock price declines as the reputation of Andersen
fell.
Our paper extends the work of C&P (2002) by investigating the cause of the
stock price decline that they find for Andersen audited firms following the Enron-
related events. There are two possible explanations for any abnormal decline in
stock price: (1) an unexpected change in expected future cash flows available to
shareholders; or (2) an increase in risk. Since the stock price represents the present
value of the future cash flows, both the expected future cash flows (the mean) as
well as the risk of the cash flows are important. Moreover, by focusing on stock
price alone, it is impossible to determine which of these factors is present. The
implied volatility on the other hand, is not directly influenced by the mean of
the expected cash flows, but rather is the option market’s consensus forecast of
expected volatility of the underlying stock over the life of the option.
The significance of this distinction can be demonstrated by examining the
distribution of expected cash flows. Even if operational cash flows are held
constant, a decline in audit quality may reduce the expected cash flows available
to shareholders due to an increased probability of over-reported earnings or
under-reported liabilities. Both possibilities would lower the expected cash flows
available to shareholders. In addition, sufficient damage to the auditor’s reputation
would necessitate a costly change in auditors. This change reduces the cash flow
available to shareholders. Therefore, it is possible that shareholders face a reduction
in expected cash flows as well as an increase in the risk of cash flows. This
differentiation, which can only be theorized by using stock price data, can be
documented with implied volatility data. Thus, two important distinctions exist
between our analysis and that of C&P (2002). First, we examine the implied
volatility associated with equity option contracts rather than stock returns. This
allows us to determine if option traders use auditor reputation in their pricing
decisions. Second, and more importantly, we are able to partially disentangle
risk from the expected cash flow consequences of Andersen’s loss of prestige,
which gives us a better representation of the impact of the auditor’s reputation on
informational asymmetries.
The expected volatility of a given asset cannot be directly observed in the market.
However, given the price of an option and the price of the underlying asset, one may
use an iterative process to calculate the expected volatility implied by a particular
option-pricing model. This implied volatility is thus the market’s expectation of
the future volatility of the underlying asset, in this case the stock price. In this
paper, the implied volatility for each stock is calculated from at-the-money options
with at least 30 days to expiration using the Black-Scholes Option Pricing Model
(BSOPM) (Black & Scholes, 1973) adjusted for dividends (Merton, 1973). This
model is as follows.
C = (e−DT )S 0 N(d 1 ) − Xe−rT N(d 2 ) (1)
where: d1 = [ln((e−DT )S 0 /X) + (r +␴2 /2)T]/(␴ sqrt(T)); d2 = d1 − (␴ sqrt(T));
S0 = stock price at time zero; X = strike price; T = time to expiration; r = risk-
free rate; ␴ = volatility of the underlying asset; D = Annualized dividend yield;
N(x) = the cumulative probability distribution function for a variable that is
normally distributed with a mean of zero and a standard deviation of 1.0.
Merton (1973), Latane and Rendlemand (1976), and Hull and White (1987) show
that the implied volatility of a particular option represents the mean anticipated
daily volatility, that is the anticipated risk, over the life of the option. Canina and
Figlewski (1993) summarize much of the research on the informational content
of implied volatilities and document the “volatility-smile” whereby the implied
volatility increases as the time to maturity gets small. Dumas, Fleming and Whaley
(1998) confirm this finding and show that volatilities are not constant across time.
The impact of events, both scheduled and unscheduled, on implied volatilities
has been the topic of much research. Patell and Wolfson (1981) as well as Isakov
and Perignon (2001) examine changes in implied volatility around scheduled
earnings announcements and find that implied volatility drops once uncertainty
is resolved. French and Dubofsky (1986) study the effect of unexpected stock
splits on the implied volatility. Finally, Levy and Yoder (1993) examine the impact
of mergers and acquisitions on implied volatilities. Each of these finds that the
option market quickly incorporates the new information and the implied volatilities
change accordingly.

In order to examine the effect of changes in auditor reputation on the risk of
a firm, implied volatility data was collected for the eighteen months ending
August 31, 2002 from ivolatility.com for firms in the S&P 500 as of October
16, 2001. Implied volatility calculations are based on at-the-money put and call
options with at least 30 days to expiration. The 30-day constraint limits the effect
of a “volatility-smile” influencing the findings. At-the-money options are used
because they are the most liquid and therefore less likely to be influenced by stale
data problems brought about by infrequent trading. This infrequent trading could
introduce a bias if the stock, which is more frequently traded than the option,
falls in value after the last recorded trade of the option contract. This would be
a problem since the implied volatility would be computed with an option trade
that occurred before the stock fell in value. As C&P (2002) provide evidence that
Andersen audited clients fell in value, this systematic price drop could lead to
higher implied volatilities. However, by using liquid at-the-money options, the
stale trading problem is reduced. Additionally, the implied volatility is calculated
with using the average of the bid and ask quotes which further reduces any stale
pricing problem.
Each firm’s auditor information was drawn from the firms’ websites as of
October 16, 2002 when the news of Enron’s problems became known. The
breakdown of firms by auditors and sectors is provided in Table 1. For consistency
purposes, only firms audited by one of the Big Five accounting firms (Andersen,
Ernst and Young, Deloitte and Touche, KPMG, and PriceWaterhouseCoopers)
were studied. This resulted in a final sample size of 495 firms audited by a Big
Five accounting firm.
If auditor reputation is important in reducing the information asymmetry
problem, then the Enron-Andersen-related events of late 2001 and early 2002
Table 1. S&P 500 Breakdown by Auditor and by Sector.

Sector Auditor
AA DT EY KPMG PWC Other Totals
Consumer Discretionary 12 14 24 9 26 2 87
Consumer Staples 5 7 5 7 10 0 34
Energy 5 2 5 4 8 0 24
Financials 9 13 24 10 18 1 75
Health Care 5 5 13 5 16 1 45
Industrials 13 13 17 8 16 1 68
Information Technology 6 11 24 13 26 0 80
Materials 10 4 5 3 15 0 37
Telecommunications 3 1 3 2 4 0 13
Utilities 14 11 1 0 11 0 37
Total 82 81 121 61 150 5 500
Note: AA = Arthur Andersen; DT = Deloitte & Touche; EY = Ernst & Young; PWC = Price
Waterhouse Coopers.
Sectors are from Standard and Poors website. Auditor information is from firms’ websites.
should result in higher implied volatilities for Andersen audited firms. If auditor
reputation is not important in reducing information asymmetries, then the implied
volatility ratios should be the same in the pre and post period. This leads to the
null hypothesis:
H0 . The implied volatility of Andersen audited firms increases as events likely
damaging to Andersen’s reputation occur.
Unlike traditional event studies, the information release effecting Enron and
Andersen was not completed at one time. From August 14, 2001 (when CEO
Jeffrey Skilling resigned) to June 15, 2002 (when Andersen was found guilty of
obstruction of justice and ordered to exit the audit business), news that could be
seen as damaging was released at irregular intervals. The news that was most
directly tied to Andersen was concentrated after November 8, 2001, when Enron
announced it would have to restate earnings.
Within the aforementioned time period, other specific dates that could be judged
as being detrimental to Andersen’s reputation include the dates investigated by
C&P as well as other dates after C&P’s February 4, 2002 cutoff. These include
November 8, 2001 when the SEC began the investigation into Enron, November
29, 2001 when the SEC expanded its investigation to include Andersen, December
12, 2001 when Andersen CEO Joe Berardino testified in front of Congress,
January 10, 2002 when Andersen admitted to shredding documents, January
24, 2002 when Andersen partner David Duncan refused to testify in front of
Congressional hearings, and the early February release of the Powers Report as
well as Andersen’s hiring of former Federal Reserve Chairman Paul Volcker in an
attempt to improve the auditor’s tarnished reputation. Additionally we investigated
several dates that were not examined by C&P 2002. These events include March
14, 2002 when the U.S. Justice Department charged Andersen with obstruction of
justice, March 27, 2002 when Andersen’s CEO Joseph Berardino resigned, and
April 9 when David Duncan admitted that he ordered the shredding of documents
(Table 2).
As time progressed and the seriousness of the problems became known, many
Andersen clients dropped Andersen as their auditor. The magnitude of these
departures introduces further problems in the investigation of event dates. The
extreme case of this is shown at the end of the event period (the August 30th
court mandated end of Andersen’s audit practices) when there were no firms still
being audited by Andersen. However prior to this definitive end point, departures
from Andersen were so pronounced that any analysis is prone to capture firm
specific risk factors. For instance, on June 15, 2002 a Houston jury found Andersen
guilty of obstruction of justice. This event was not examined because only 2 of
the original sample of 82 firms were still with Andersen. To avoid these sample
Table 2. Andersen-Enron Timeline.

Date Event Description
August 14, 2001 Enron CEO Jeffrey Skilling resigns citing personal reasons
October 12, 2001 Arthur Andersen’s in-house lawyer emails Houston office regarding
document destruction policy
October 16, 2001 Enron reports $638 million loss and a $1.2 billion reduction in shareholder
equity
October 17, 2001 SEC requests information from Enron regarding losses
October 22, 2001 SEC begins official inquiry into Enron
October 31, 2001 Enron announces the SEC inquiry has been upgraded to a formal
investigation
November 8, 2001a Enron revises its last five years of financial statements, admitting net losses
of $586 million
Andersen receives a federal subpoena for Enron documents
November 9, 2001 Dynegy announces takeover of Enron for $8 billion
November 19, 2001 Enron again restates third quarter earnings and warns on debt repayments
November 28, 2001 Dynegy pulls out of takeover deal and Enron’s credit rating downgraded to
junk status
November 29, 2001 SEC investigation expanded to include Andersen
December 2, 2001 Enron files for Chapter 11 bankruptcy protection
December 12, 2001a Andersen CEO, Joseph Berardino, testifies before Congress and admits the
auditing firm may have violated securities laws
January 9, 2002 Justice department begins criminal investigation of Enron
January 10, 2002a Andersen acknowledges that it had destroyed documents
January 15, 2002 Andersen Partner, David Duncan, fired
January 22, 2002 Enron employee claims shredding
January 23, 2002 Kenneth Lay resigns as Enron’s CEO
FBI begins investigation into shredding
January 24, 2002 David Duncan refuses to testify in front of Congressional hearings
regarding shredding of “Enron-related documents”
January 25, 2002 Clifford Baxter, who had reportedly “clashed” with Jeff Skilling over
“accounting practices,” is found dead of an apparent suicide
February 2, 2002a Powers report released
February 3, 2002a Andersen hires former Fed Chairman, Paul Volcker
March 14, 2002 U.S. Justice Department charges Andersen with obstruction of justice for
allegedly shredding vital Enron documents
March 20, 2002 Andersen pleads not guilty to obstruction of justice charges
March 27, 2002 CEO Joseph Berardino announces his resignation
April 8, 2002 Andersen reports it will lay off 7,000 employees
April 9, 2002 Former Andersen auditor David Duncan enters a plea bargain with
prosecutors, admitting he ordered the shredding of incriminating Enron
documents
May 6, 2002 Andersen trial begins in a federal courtroom in Houston
June 15, 2002 Jurors find Andersen guilty of obstruction of justice charges and the firm
agrees to cease auditing public companies by August 31, 2002 as ordered
by the SEC
Date Event Description
June 27, 2002 SEC issues restraining order to prevent Andersen from destroying any
documents relating to its telecommunication client WorldCom
August 16, 2002 Andersen has its Texas accounting license revoked
August 27, 2002 Andersen Worldwide agrees to pay a $60 million settlement to the plaintiffs
in the Enron class-action suit and to creditors of the bankrupt Energy firm
August 30, 2002 Andersen “closes its books,” as its public auditing business is terminated
October 16, 2002 Andersen receives maximum sentence of five years probation and a fine of
$500,000 for obstruction of justice relating to Enron scandal
Sources: BBC, CBS News, Houston Chronicle, and Washington Post.

a Indicates that the event was also investigated by Chaney and Philipich (2002).
size problems we report individual events only up to April 9, 2002 when former
Andersen partner David Duncan admitted that he had ordered the destruction of
Enron related documents. The rationale for halting the analysis at this point was
this was the last major event where at least 50% of the original sample was still
with Andersen (Table 3).
Table 3. When Firms Dropped Andersen.

Date Event Description Number of
Andersen
Audited Firms
February 3, 2002 Andersen hires former Fed Chairman, Paul Volcker 82

March 14, 2002 Andersen charged with obstruction of justice 73
March 20, 2002 Andersen enters a not guilty plea 68
March 27, 2002 CEO Joseph Berardino resigns 57
April 8, 2002 Andersen reports 7,000 layoffs 51
April 9, 2002 Former Andersen Auditor David Duncan enters plea bargain 49
May 6, 2002 Trial begins in Houston 32
June 15, 2002 Andersen found guilty, ordered to cease public auditing by 3
August 31, 2002
June 27, 2002 SEC issues restraining order to prevent destruction of 2
WorldCom documents
August 16, 2002 Andersen’s Texas accounting license revoked 0
August 27, 2002 Andersen agrees to $60 million settlement in Enron 0
class-action suit
August 30, 2002 Andersen terminates its public accounting business 0
October 16, 2002 Andersen sentenced to five years probation and $500,000 0
fine
As a consequence of these multiple event dates, a single-event date methodology

may not capture the impact of the information. In addition, with so many events in
succession, it is possible that the impact of any one event may have been predicted
by market participants and already was incorporated into the price of the option.
Finally, since the events occur in close proximity to one another, there may be
important spillover effects that would prevent the impact of single events being
known with certainty. The most extreme example of this contamination is the
Saturday February 2, 2002 release of the Powers Report and the Sunday February
3, 2002 hiring of former Federal Reserve Chairman Paul Volcker. Since both events
occurred on the weekend, the market reaction to each event cannot be individually
measured, as the first day of trading for each was Monday, February 4, 2002.
To avoid these multi-news event contamination problems, we use a strategy
previously used to evaluate the market reaction to deregulation legislation. Since
deregulation legislation is not a single event, but rather a series of committee
meetings, press releases, and votes, the outcome could be forecasted, causing no
significant market reaction at the time of actual passage of the deregulation bill.
Schwert (1981) and Binder (1985) suggest looking before any of the coverage
began and then after all of it was over to assure that the total change was captured.
Thus, we begin our investigation by determining whether the combination of all the
events in sum, had a significant effect on the relative implied volatility of Andersen
clients as compared to other Big Five audited firms.
To control for market-wide fluctuations in implied volatilities, we begin the
analysis by using the VIX. The VIX is the most widely actively traded volatility
index. During the 2001–2002 time-period the VIX tracked the implied volatility
of at-the-money options for S&P 100 firms.1 Thus, using the VIX introduces a
possible bias as our sample is composed of S&P 500 firms. More importantly
however, is the problem that 14 of the firms in the sample (17%) are also in the
S&P 100 thus causing potential contamination problems.
To avoid these contamination problems, for the majority of our analysis we use
the ratio of implied volatilities at Andersen audited firms to the implied volatility
of the other S&P 500 firms audited by a Big Five accounting firm. This implied
volatility ratio, which will be henceforth called the IV ratio, is calculated on a daily
basis for each trading day from June 1, 2001 to August 30, 2002 when Andersen
officially ceased its audit operations.
If all audit firms were identical in audit quality and client make-up, then the
implied volatility ratio should be equal to one. However, since there are differences
in makeup of clientele (see Table 1), then the ratio may be different than one.
By analyzing changes in this ratio, rather than absolute levels, the impact of
auditor reputation can be examined. Thus, rather than examining the makeup of
the individual firms’ implied volatility, this paper studies the changes of the IV
ratio to determine whether the events surrounding the Enron-Andersen case had
differential effects on the implied volatilities of Andersen’s clients versus clients of
other auditors. This methodology also accounts for differences in the audit firms’
initial reputation and changes in market-wide volatility.
Copeland, Poon and Stapleton (2000) model implied volatility as a function of
firm specific factors. To control for these firm specific factors, which include the
“duration of the firm’s profits” and the firm’s leverage, we examine the changes
in the IV ratio for the portfolio of Andersen audited S&P 500 firms. Additionally
we end our individual event analysis on April 9, 2002 when the size of the sample
drops to 39 firms (approximately 50% of the original sample).
The August 14, 2001 resignation of Jeffery Skilling has frequently been cited
as the beginning of the end for Enron. It would therefore be theoretically possible
that the markets could have inferred the troubles with Andersen from this initial
announcement. As a result, we create a baseline IV ratio to compare Andersen
audited S&P 500 firms to other Big Five audited S&P 500 firms. This ratio is
computed on a daily basis by dividing the average implied volatility for Andersen
audited firms by the average implied volatility for firms audited by other Big Five
firms. The resulting ratio is found for the 20 trading days preceding August 14. This
baseline is then compared to the daily IV ratio for the 20 trading days following
the release of the Powers report on February 2, 2002.
To insure that any changes in implied volatility were not directly caused by the
events of September 11, 2001, the procedure is repeated using the ten trading day
period immediately prior to the October 16, 2001 announcement by Enron where
they first publicly acknowledged large amounts of off balance sheet financing and
a major loss. These dates were chosen because by October 1, 2001, market-wide
volatility, as measured by the VIX index, had returned to September 10th
levels.
Once the daily implied volatility ratios were computed, the next step in the
investigation was to check the IV ratio data for normalcy. This was accomplished
with the Jarque-Bera test, which showed that the data was likely normally
distributed (p-value of 0.41). The actual investigation is performed in two ways.
First t-tests are used to examine the average implied volatility ratio before and after
the event windows. Following the t-tests, a regression analysis was performed
with a dummy independent variable to capture the marginal impact of the
events.
Black (1976) shows that implied volatilities are inversely related to stock price.
As C&P (2002) report that Andersen audit clients experienced negative abnormal
returns following the reputation damaging events associated with Andersen and
their role in the Enron collapse. Consequently, a stock price variable is used
to control for any potentially confounding results stemming from the negative
relationship between the underlying stock price and implied volatility. We regress
the change in the implied volatility ratio from before each event to after each event.
That is we regress the model:
IV ratio = c + b1(Stock Price Ratio) + b2(Time Dummy) (2)
where: IV Ratio = Average Implied Volatility for Andersen audited firms divided
by Average Implied Volatility for firms audited by other Big Five audit firms; Price
Ratio = Average stock price for Andersen audited firms divided by average stock
price for firms audited by other Big Five audit firms; Time Dummy = 1 in the post
event time period.
Once the overall (both the 6 and 8 month event windows) impact of the events
has been determined, the same methodology is used to examine the impact of
the individual events that would be likely to have been interpreted as damaging
to Andersen’s reputation. Post event changes in the relative implied volatility are
examined for the days 0, +1 and days 0, +4 event windows. T-tests are performed
by comparing the daily implied volatility ratio to the average ratio of the 20
days preceding the event. As with the longer event-window tests, we inspect the
Table 4. Change in IV Ratio Over Long-Event Windows.

Panel A: 8-Month Window
Before August 14, 2001 After April 9 2002 Difference p-Value
IV ratio
Call data 0.9007 0.9439 0.0432 0.0000
Put data 0.9050 0.9473 0.0423 0.0000
Panel B: 6-Month Window
Before October 16 After February 4 Difference p-Value
IV ratio
Call data 0.8955 0.9335 0.0380 0.0000
Put data 0.9000 0.9361 0.0361 0.0000
Note: Changes in daily implied volatility ratio: The daily implied volatility ratio is computed by
dividing the average implied volatility of Andersen audited firms in the S&P 500 by the average
implied volatility of S&P 500 firms audited by other “Big Five” firms. The “Before” trading
dates in Panel A are based on the 30-day event window prior to Jeffrey Skilling’s August 14th,
2001 resignation. The “Before” dates in Panel B were selected to minimize the effects of the
September 11, 2001 attacks. The average implied volatility on S&P 500 stocks had returned to
its pre-attack levels by October 1, 2001. Therefore, the “Before” period in Panel B has only 11
trading days prior to the October 16, 2001 disclosure by Enron of a significant equity write-off
and net loss. The “After” trading dates in each Panel are composed of the 30-days following the
February release of the Powers report, which was critical of Andersen’s practices.
individual announcement effects by performing regression analysis using the same

equation as in longer (6 or 8 month) event window.
4. RESULTS AND DISCUSSION

Tests of both long event windows (the six and eight month) give strong evidence
that overall there was a significant increase in the implied volatilities of Andersen
audited clients relative to those of firms audited by other Big Five firms (p value
0.0000). This is shown in Table 5. This is consistent with the hypothesis that auditor
reputation helps to mitigate the information asymmetry problem.
Table 5. Long Event Window Regression Results.

Panel A: Calls
8-Month 6-Month
IV Ratio = constant + time dummy

Constant 0.9007 (212.95) 0.8955 (488.38)
Time dummy 0.0432 (8.14) 0.0380 (14.64)
IV Ratio = constant + price ratio + time dummy
Constant 1.7472 (3.86) 0.8460 (2.63)
Price ratio −0.8320 (−1.88) 0.0470(0.15)
Time dummy 0.0398 (7.50) 0.0401 (2.87)
Panel B: Puts
IV Ratio = constant + time dummy

Constant 0.9050 (188.07) 0.9000 (537.56)
Time dummy 0.0423 (7.01) 0.0361 (15.24)
IV Ratio = constant + price ratio + time dummy
Constant 1.8506 (3.61) 0.8536 (2.91)
Price ratio −0.9295 (−1.84) 0.0440 (0.16)
Time dummy 0.0385 (6.36) 0.0381 (2.98)
Note: The IV Ratio is the average implied volatility of Andersen audited firms divided the average
volatility of firms audited by other Big Five firms. The average price ratio is the average price
of Andersen audited firms divided the price of firms audited by other Big Five firms. The results
shown here are for all firms audited by Andersen as of October 16, 2001. The 8-month window
regression uses the twenty trading days before the August 14, 2001 resignation of Jeffrey Skilling
and the twenty trading days following the April 9, 2002 admission by David Duncan that he
had ordered documents to be destroyed. The 6-month window regression uses the ten trading
days before the October 16, 2001 Enron earnings announcement, and ten trading days following
the April 9, 2002 admission by David Duncan that he had ordered documents to be destroyed.
(t-Statistics are in parenthesis.)
Table 6. Changes in Auditor Implied Volatility Ratio Around Key Dates.

Date/Event Before (Days 0, 1) Days (0, 4)
(Days Difference Difference
(−20, −1) (p-Value) (p-Value)
August 14, 2001: Enron CEO Call data 0.8949 0.8959 0.8887
Jeffrey Skilling resigns 0.0010 −0.0062
(0.4612) (0.1700)
Put data 0.8908 0.8964 0.8878
0.0056 −0.0030
0.3011 (0.3904)
October 16, 2001: Enron reports Call data 0.8662 0.8589 0.8628
loss −0.0073 −0.0034
(0.2483) (0.3102)
Put data 0.8676 0.8656 0.8644
−0.0020 −0.0032
0.3776 0.3784
November 8, 2001: Andersen Call data 0.8615 0.8754 0.8710
receives federal subpoena and 0.0139 0.0097
Enron restates earnings (0.0260) (0.0221)
Put data 0.8620 0.8762 0.8697
0.0142 0.0077
(0.0204) (0.0450)
November 29, 2001: SEC Call data 0.8687 0.8561 0.8616
investigation is expanded to −0.0126 −0.0071
include Arthur Andersen (0.0425) (0.0961)
Put data 0.8683 0.8531 0.8574
−0.0152 −0.0109
0.0429 0.0226
December 12, 2001: Joseph Call data 0.8691 0.8926 0.9040
Berardino appears in front of 0.0235 0.0349
Congress (0.0049) (0.0000)
Put data 0.8683 0.8927 0.9027
0.0244 0.0344
0.0053 0.0000
January 10, 2002: Enron admits Call data 0.8864 0.8889 0.8862
to shredding 0.0035 −0.0002
(0.4224) (0.4879)
Put data 0.8861 0.8889 0.8866
0.0028 0.0005
0.4072 0.4757
January 22, 2002: Enron Call data 0.8833 0.8854 0.8929
employee claims shredding 0.0021 0.0065
(January 23) FBI begins (0.4269) (0000)
investigation into shredding Put data 0.8839 0.8873 0.8928
0.0034 0.0089
0.3663 0.0817
Date/Event Before (Days 0, 1) Days (0, 4)
(Days Difference Difference
(−20, −1) (p-Value) (p-Value)
January 24, 2002: David Duncan Call data 0.8831 0.8932 0.9159
refuses to testify 0.0101 0.0328
(0.1020) (0.0002)
Put data 0.8851 0.8932 0.9169
0.0081 0.0318
(0.1718) (0.0003)
February 4, 2002: Release of Call data 0.9031 0.9532 0.9470
Powers Report and Paul 0.0501 0.0439
Volcker hired (0.0068) (0.0022)
Put data 0.9000 0.9527 0.9443
0.0527 0.0443
0.0018 0.0001
March 14, 2002: Andersen Call data 0.9236 0.9105 0.9025
charged with obstruction of −0.0131 −0.0211
justice (0.1134) (0.0027)
Put data 0.9264 0.9160 0.9096
−0.0104 −0.0064
(0.1752) (0.0124)
March 27 2002: Andersen CEO Call data 0.9113 0.8862 0.8840
Joseph Berardino resigns −0.0251 −0.0273
(0.0452) (0.0027)
Put data 0.9156 0.8930 0.8907
−0.0226 −0.0249
(0.0539) (0.0037)
April 9, 2002: Former Andersen Call data 0.8940 0.9033 0.9099
auditor Duncan enters plea 0.0093 0.0159
bargain (0.1473) (0.0048)
Put data 0.8996 0.9152 0.9162
0.0156 0.0166
(0.0399) (0.0024)
Note: Implied Volatility Ratio = (Average Implied Volatility of Andersen-audited firms)/(Average

Implied Volatility of firms audited by other Big Five auditing firms).
While the long-event window results provide evidence that there was an increase
in implied volatilities at Andersen audited clients, they do not say specifically when
this increase occurred. As a result, we used the same methodology to examine
shorter event windows around the events that were hypothesized as being damaging
to Andersen’s reputation. The findings in this area, with both the t-tests as well
as the regression analysis, reaffirm that Andersen’s audit clients did experience
108
Table 7. Regressions Around Key Dates.
Panel A: Call Data
Nov 8 Dec 12 Jan 10 Jan 24 Feb 4 Mar 14 Mar 27 Apr 9
Avg IV Anderson/Avg IV other = constant + B2 (Time dummy)

Constant 0.8985 0.9077 0.9129 0.9139 0.9382 0.9411 0.9514 0.8853
(333.34) (230.81) (549.84) (264.19) (184.60) (277.33) (348.31) (268.05)
Time dummy 0.0050 0.0172 0.0022 0.0373 0.0169 −0.0134 −0.0026 0.0339
(1.32) (3.09) (0.94) (7.63) (2.35) (−2.79) (−0.66) (7.26)
Avg IV Anderson/Avg IV other = constant + B1 (Avg. Anderson price/Avg. other price) + B2 (Time dummy)
Constant 1.7751 2.9387 1.0684 1.4362 2.111 1.7032 2.4735 −0.1791
(8.91) (9.97) (3.74) (1.44) (1.48) (3.77) (3.21) (−0.14)
−0.8454 −1.9983 −0.1539 −0.5200 −1.1620 −0.7924 −1.5512
JONATHAN M. GODBEY AND JAMES W. MAHAR

Avg price ratio 1.0487
(−4.40) (−6.89) (−0.54) (−0.52) (−0.82) (−1.69) (−1.97) (0.081)
Time dummy −0.0016 0.0039 0.0014 0.0403 0.0167 −0.0125 0.0004 0.0315
(−0.53) (1.11) (0.47) (5.37) (2.29) (−2.72) (0.11) (5.69)
Panel B: Put Data
Avg IV Anderson/Avg IV other = constant + B2 (Time dummy)

Constant 0.9008 0.9092 0.9148 0.9161 0.9394 0.9446 0.9548 0.8905
(327.81) (239.42) (584.64) (253.90) (190.91) (284.03) (362.06) (289.40)
Time dummy 0.0028 0.0179 0.0025 0.0352 0.0154 −0.0117 −0.0002 0.0348
(0.73) (3.33) (1.13) (6.89) (2.21) (−2.48) (−0.06) (8.00)
Avg IV Anderson/Avg IV other = constant + B1 (Avg. Anderson price/Avg. other price) + B2 (Time dummy)
Constant 1.7605 2.8065 0.8413 1.4218 1.9199 1.5411 1.8856 0.1921
(8.32) (9.08) (3.17) (1.37) (1.38) (3.38) (2.37) (0.15)
Avg price ratio −0.8291 −1.8667 0.0728 −0.5034 −0.9714 −0.6204 −0.9486 0.6881
(−4.06) (−6.14) (0.27) (−0.49) (−0.70) (−1.31) (−1.17) (0.56)
Time dummy −0.0037 0.0055 0.0029 0.0380 0.0152 −0.0110 0.0016 0.333
(−1.34) (1.49) (1.07) (4.85) (2.15) (−2.37) (0.39) (6.38)
Note: The IV Ratio is the average implied volatility of Andersen audited firms divided the average volatility of firms audited by other Big Five firms. The average price
ratio is the average price of Andersen audited firms divided the price of firms audited by other Big Five firms. Ten trading days before and after each event are used
to calculate IV Ratio and Price Ratio. (t-Statistics are in parenthesis.)
statistically significant increases in their implied volatilities as a result of the

damage to Andersen’s reputation. Additionally, the results show which events
were particularly damaging to Andersen’s reputation. These findings support the
hypothesis that auditor reputation is used by market participants as a means of
inferring the quality of a firm’s financial reports and thus helps to lessen the
information asymmetry problem (Tables 4, 5, 6 and 7).
Specific events where the implied volatility ratio increases relative to the control
groups in at least one of the tests include the November 8, 2001 restatement of
Enron’s financial statements, the December 12, 2001 Congressional testimony of
Andersen CEO Joseph Berardino, the January 10, 2002 admission of shredding
documents, the January 24, 2002 refusal of David Duncan to testify, the dual events
of the weekend of February 2 and 3, 2002 with the release of the Powers Report
and the Andersen hiring of former Federal Reserve Chairman Paul Volcker, and
the April 9, 2002 plea bargain by David Duncan. All but the November 8, 2001
and December 12, 2001 events were statistically significant at the 5% level, for
both the t-tests as well as the regression analysis that controlled for changes in the
underlying stock price.
5. CONCLUSION
This paper was motivated by the opportunity to use the Enron and Andersen related
troubles to study the impact of auditor reputation on the information asymmetry
problem. Consistent with the findings of C&P (2002), who found that stock prices
of Andersen audited firms decreased over this time period, we studied how auditor
reputation impacts implied volatility at firms audited by Andersen.
Both long-term and short-term event-studies were used to examine the effects
on implied volatility, of events that were deemed as damaging to Andersen’s
reputation. The results of all of the tests yield strong evidence that the implied
volatilities at Andersen audited firms increased relative to the firms audited by
other Big Five accounting firms over the time period surrounding the events that
led to collapse of Andersen. Thus, the results of both tests are consistent with the
hypothesis that auditor reputation plays an important role in reducing information
asymmetries between investors and the audited firm.
By looking at the implied volatility of Andersen audited firms, rather than the
stock price, we are able to show that the stock price declines documented by C&P
(2002), are due, at least in part, to an increase in the risk of the cash flows. This is a
significant contribution because by looking at stock price alone, the decline could
be caused by any combination of a reduction of expected cash flows or an increase
in risk. While the implied volatility does not allow us to say anything about the
level of cash flows, we can say definitively that the market expected the risk to
increase.
NOTE
1. In September 2003, the VIX was changed to track the implied volatility of S&P 500
firms.
ACKNOWLEDGMENTS
We would like to thank David Becher, Joseph Coate, Carol Fischer, Stephan Horan,
Lance Nail, Rodney Paul, Jeffrey Peterson, Kenneth Small, and participants at
the Southern Finance Association meetings in Charleston SC. Additionally we
are indebted to the able research assistance of Patricia Dean, Jill Simme, and
Christopher Zimmer. All errors are our own.
REFERENCES
Balvers, R., McDonald, B., & Miller, R. (1988). Underpricing of new issues and the choice of auditor
as a signal of investment banker reputation. The Accounting Review, 63, 605–622.
BBC News (2002, June 17). Andersen’s fall from grace. URL: http://news.bbc.co.uk/1/hi/business/
2049237.stm.
BBC News (2002, February 4). Timeline: Enron’s rise and fall. URL: http://news.bbc.co.uk/
2/hi/business/1759599.stm.
Beatty, R. (1989). Auditor reputation and the pricing of initial public offerings. The Accounting Review,
64, 693–709.
Beatty, R. (1993). The economic determinants of auditor compensation in the initial public offerings
market. Journal of Accounting Research, 31, 294–302.
Bellalah, M., & Aboura, S. (2002). Pricing options on leveraged equity under asymmetric information
(Working Paper).
Binder, J. J. (1985). Measuring the effects of regulation with stock price data. Rand Journal of
Economics, 16, 167–183.
Black, F. (1976). Studies of stock price volatility changes. Proceedings of the American Statistical
Association, 177–181.
Black, F., & Scholes, M. (1973). The pricing of options and corporate liabilities. Journal of Political
Economy, 81, 637–654.
Canina, L., & Figlewski, S. (1993). The informational content of implied volatility. Review of Financial
Studies, 6, 659–681.
Chaney, P., & Philipich, K. (2002). Shredded reputation: The cost of audit failure. Journal of Accounting
Research, 40, 1221–1244.
Clarkson, P., & Simunic, D. (1994). The association between audit quality, retained ownership, and firm
risk in U.S. vs. Canadian IPO markets. Journal of Accounting and Economics, 17, 207–228.
Copeland, L., Poon, S., & Stapleton, R. (2000). The implied volatility of options: A test using UK data.
Journal of Business Finance and Accounting, 27, 859–885.
Craswell, A. T., Francis, J. R., & Taylor, S. L. (1995). Auditor brand name reputation and industry
specialization. Journal of Accounting and Economics, 20, 297–322.
Datar, S., & Alles, M. (1999). The formation and role of reputation and litigation in the auditor-manager
relationship. Journal of Accounting, Auditing and Finance, 14, 401–428.
Dumas, B., Fleming, J., & Whaley, R. E. (1998). Implied volatility functions: Empirical test. Journal
of Finance, 53, 2059–2106.
Dye, R. (1993). Auditor status, legal liability, and auditor wealth. Journal of Political Economy,
887–914.
Francis, J., & Simon, D. (1987). A test of audit pricing in the small-client segment of the U.S. audit
market. The Accounting Review, 62, 145–157.
Franz, D. R., & Crawford, D. (1998). The impact of litigation against an audit firm on the market value
of nonlitigating clients. Journal of Accounting, Auditing and Finance, 13, 117–134.
French, D., & Dubofsky, D. (1986). Stock splits and implied stock price volatility. Journal of Portfolio
Houston Chronicle (2002, October 22). Timeline of Andersen events. URL: http://www.chron.com/cs/
CDA/story.hts/special/enron/1452911.
Hull, J., & White, A. (1987). Pricing of option assets with stochastic volatilities. Journal of Finance,
42, 281–300.
Isakov, D., & Perignon, C. (2001). Evolution of market uncertainty around earnings announcements.
Journal of Banking and Finance, 25, 1769–1789.
Latane, H. A., & Rendlemand, R. J., Jr. (1976). Standard deviations of stock price ratios implied in
option prices. The Journal of Finance, 31, 369–381.
Levy, H., & Yoder, J. (1993). The behavior of option implied standard deviations around merger and
acquisition announcements. Financial Review, 28, 261–272.
Merton, R. (1973). Theory of rational option pricing. Bell Journal of Economics and Management
Science, 4, 141–183.
Merton, R. (1987). A simple model of capital market equilibrium with incomplete information. Journal
of Finance, 42, 483–510.
Patell, J. M., & Wolfson, M. A. (1981). The ex ante and ex post price effects of quarterly earnings
announcements reflected in option and stock prices. Journal of Accounting Research, 19,
434–458.
Schwert, G. W. (1981). Using financial data to measure the effects of regulation. Journal of Law and
Economics, 121–158.
Teoh, S., & Wong, T. (1993). Perceived auditor quality and the earnings response coefficient. The
Accounting Review, 68, 346–366.
Titman, S., & Trueman, B. (1986). Information quality and the valuation of new issues. Journal of
Accounting and Economics, 8, 159–172.
Washington Post (2002, November 14). Timeline of Enron’s collapse. URL: http://www.
washintonpost.com/ac2/wp-dyn/A25624–2002Jan10?language=printer.
Watts, R., & Zimmerman, J. (1983). Agency problems, auditing, and the theory of the firm: Some
evidence. Journal of Law and Economics, 26, 613–633.
SECONDARY EQUITY OFFERINGS:
THE CASE OF INSTALLMENTS
RECEIPTS
Narat Charupat and C. Sherman Cheung
ABSTRACT
This paper examines secondary equity offerings that were done in the
Canadian markets through “installment receipts” (IRs). Previous studies on
seasoned equity offerings tend to focus on the price reaction around the
announcement date. We extend the analysis to cover a longer period so that
the issues of liquidity effect and information asymmetry can be adequately
addressed. We also offer evidence to indicate that the use of IRs in secondary
offerings can reduce the liquidity impact in markets where market depth is
not as substantial as in the U.S.
1. INTRODUCTION
Past studies on block trading and secondary equity offerings have documented the
discounts that the sellers have to concede in order to execute the transactions.1
(See, for example, Holthausen et al. (1990) in the context of block trading, and
Hudson et al. (1993) in the context of secondary offerings.) While the two types of
transactions appear to belong to two distinct literatures (i.e. block trading belongs
to the microstructure literature whereas secondary issues are in the domain of
seasoned security offerings), there are, however, two common dominant themes –
information asymmetry and liquidity effect. Information asymmetry arises because
Research in Finance
© 2004 Published by Elsevier Ltd.
ISSN: 0196-3821/doi:10.1016/S0196-3821(04)21005-X
113
114 NARAT CHARUPAT AND C. SHERMAN CHEUNG
block sellers (in the case of block trades) and major shareholders (in the case of
secondary offerings) are traders with possibly superior information about the shares
that they sell. Liquidity effect arises if the demand for a security is not perfectly
elastic. Both factors can explain a decline in price as a result of the two types
of transactions. The major difference between the two factors is that the price
impact of the liquidity effect is temporary, whereas the price effect of information
asymmetry is permanent (Scholes, 1972).
In this paper, we study secondary equity offerings that were done through
“installment receipts” (IRs) in the Canadian markets. An IR is a security that
evidences the purchase of an underlying security on an installment basis. Typically,
buyers are allowed to pay for the underlying securities in two or three installment
payments. The payments are spread over a period of one to two years (more on
this in Section 2). Large secondary equity offerings in Canada are commonly done
through this method, instead of on a fully-paid basis.
We are interested in several issues in this paper. First, we would like to know
the role of an IR arrangement in secondary equity offerings. While IRs generally
cannot be issued on a public-market basis in the U.S. due to Securities & Exchange
Commission’s margin rules, they are quite common in Commonwealth countries
such as England, Australia, New Zealand and Canada.2 Yet there has been very
little theoretical and empirical research on this instrument. Second, to explain the
role of IRs, we trace the price behavior of stocks and, whenever possible, IRs
from the period before the announcements of the secondary offering through the
announcement dates and the issue dates, to the period after the issue dates. Previous
studies on seasoned equity offerings tend to focus on the pricing behavior around
the announcement dates and ignore the short-term pricing behavior around the
issue dates.3 As a result, the liquidity effect around the issue dates has not been
adequately addressed in the literature. We therefore offer a more comprehensive
empirical analysis of secondary offerings. Third, we will try to explain the cross-
sectional behavior around three critical dates – the announcement date, the issue
date and the post-issue period. In particular, we are interested in the pricing of the
issues and the short-term excess return from buying them.
The paper is organized as follows. In the next section, we discuss the institutional
features of IRs and the rationale for using them. Section 3 provides details of
our sample. Section 4 documents the price pattern of the stocks from the period
prior to the announcements of the offerings to post-issue period. It also reports
the magnitude of the underpricing of IRs and the short-term excess return from
buying them. Cross-section results are reported in Section 5. Section 6 examines
the post-issue performance again to address the liquidity effect. Section 7 provides
supporting evidence for the benefits of using IRs in secondary offerings. Finally,
Section 8 concludes the paper.
Secondary Equity Offerings 115
2. INSTITUTIONAL FEATURES OF AND RATIONALE

FOR INSTALLMENT RECEIPTS
In Canada, IRs cannot be used for initial or seasoned equity offerings involving
new shares as the laws prohibit companies from issuing new shares that are not
fully paid for. As a result, all of our discussions here concern secondary offerings
involving no new shares.
In a typical IR arrangement, investors are allowed to pay for the underlying
shares in two or three installments. The first payment is required at the closing
of the offering, after which the buyers will receive IR certificates. Then, after the
final payment is made, the receipt-holders will exchange the IR certificates for
common shares of the underlying stock. In the meantime, IRs entitle the holders to
the same rights and benefits (e.g. same dividends, voting right, etc.) as if they were
holding the underlying stock. This is possible because the issuer (i.e. the seller of
the shares) has to place the underlying shares with a custodian who will then pass
on the rights and benefits from the shares to the IR buyers.
The initial sale of IRs to the public is handled by underwriters who are
responsible for issuing a prospectus and filing it with the appropriate regulatory
body. The prospectus specifies, among other things, the number of installment
payments, and the timing and the dollar value of each payment.4 Generally, the
underwriters underwrite only the first installment payment. Therefore, they will
no longer be involved once the deal is closed. If the buyers fail to make future
installment payments, the issuer bears the default risk.
If a buyer defaults, the issuer will arrange to sell the underlying shares which
have been set aside for that buyer. The proceeds will be offset against the buyer’s
obligation. Any shortfall remains the buyer’s responsibility. Since the most likely
scenario for default to occur is when the price of the underlying share substantially
declines, the magnitude of the default risks depends on the stock’s volatility and
the terms of the installment payments set by the issuers. Hence, the issuers have
some control over the magnitude of the risks.5 Some issuers may choose to sell
their IR receivables to financial institutions, in which case the default risks should
be reflected in the sale prices.
After the closing of the offering, IRs are listed on stock exchanges and thus can
be easily traded by their holders. By selling their IRs in the market, the holders
pass to the buyers the obligation to make future installment payments.
As described, IRs are in effect an opportunity for investors to buy the underlying
shares on borrowed funds, where the issuers assume the role of lenders. The
fact that the issuers choose to do so and bear the associated default risks suggest
that there are offsetting benefits from using IRs. One commonly-cited benefit is
that their use allows the issuers to enlarge the set of potential buyers to include
those who want to participate in the offers but cannot do so due to borrowing
constraints. These potential buyers include mutual funds and pension funds, which
commonly are prohibited from borrowing by their charters, and small investors,
who generally find it costly or impossible to obtain leverage. Under this rationale,
IRs are not redundant securities as their use will increase the demand for the
offers.6 The higher the embedded leverage, the more are the potential buyers IRs
will attract.
The increase in demand can help reduce the underpricing of seasoned/secondary
offerings caused by liquidity effects when securities are not perfectly substitutes
for one another. If securities’ demand is not perfectly elastic, investors will require
extra compensation in the form of underpricing for having to hold more of a
particular security than initially desired.7 The increase in demand helps to reduce
the underpricing by raising the markets’ aggregate risk-bearing ability (which is
an increasing function of the number of traders in the markets).8
A closely parallel literature on liquidity-induced underpricing is that of block
trades. A block trader performs similar functions to those of an underwriter in a
public offering; i.e. to facilitate the trade by locating counterparties. It has been
observed that seller-initiated block trades are generally transacted at prices below
both pre- and post-block prices. This suggests that the sellers have to make price
concessions to the buyers. More importantly, the fact that post-issue prices are
higher than transacted prices suggests that these concessions arise due to liquidity
reasons (e.g. to compensate for search cost).9
Our sample contains all secondary offerings of common shares that were done
through installment receipts on the Toronto Stock Exchange (TSE) or the Montreal
Exchange (ME) between 1990 and 1999. In total, there are 30 issues involving
shares of 29 companies.10
Table 1 provides summary statistics for our sample. Total issue values range
from C$ 125 million to C$ 1.8 billion, with a mean of C$ 667 million and a
median of C$ 606 million.11 In terms of the percentage of the outstanding shares,
these offerings account for between 6.67 and 83.26%, with a mean of 38.91% and
a median of 37.32%. These numbers suggest that the issuers in our sample are
major shareholders of their companies. The ratios between the number of shares
offered and the average daily trading volume are also reported. They range from
22 to 9,939, with a mean of 1,031 and a median of 610. Therefore, the offerings in
our sample represent a substantial supply of equity, both in terms of dollar values
and trading impact that the markets have to absorb.
Table 1. Sample’s Summary Statistics.

Lowest Highest Mean Median
Total issue value 125 1,800 667 606

(in C$ million)a
1st installment values 64 759 297 280
(in C$ million)b
No. of shares offered as 6.67% 83.26% 38.91% 37.32%
% of no. of shares
outstandingc
No. of shares offered as 22 9,939 1,031 610
multiples of average daily
trading volumesd
Leverage Offerede 37.41% 66.64% 51.82% 48.92%
a Issue values are calculated by multiplying the number of underlying shares offered by the sum of
the required installment amounts, without taking into account the payments’ present values. Also, the
figures are based on dollars at the time of issues.
b 1st installment values are calculated by multiplying the number of the underlying shares offered by
the amounts of the 1st installments. The figures are based on dollars at the time of issues.
c The issue values and the number of offered shares do not include shares subsequently issued (if any)
through overallotment options.

d The average daily trading volumes are calculated as the average of daily volumes during a 50-day
period (from day −150 to day −101) prior to the announcement dates.
e Leverage offered is calculated by dividing the present value of future installment payments by the
sum of: (i) the initial payment; and (ii) the present value of future installment payments.
Table 1 also reports the issue values based only on their first installment
payments, which measures the immediate impact of the offerings in dollar value’s
terms. By using IRs, the issuers are able to reduce the immediate dollar impact by
an average of over 50%. In addition, the last row of Table 1 reports the degree of
leverage offered by the issuers. The degree of leverage for each issue is the ratio
between: (i) the present value of the issue’s future installment payments; and (ii)
the sum of that present value and the issue’s initial payment. In other words, it is
the percentage of the “full” price that investors do not have to pay up front.
In terms of issuers, six of the issues in our sample (20%) were sold by individual
shareholders, twenty one (70%) by corporate shareholders, and four (10%) by
either the provincial or federal governments in the privatization process. The
breakdown of the issues according to the years in which they were done is given
in Table 2.
Details of each secondary issue such as its size, its issuer(s), the number of
shares offered, the number of shares outstanding and the terms of the installment
payments, were obtained from the TSE Monthly Review and, when available, the
Table 2. A Breakdown of IR Issues by Year.

Year No. of Issues
1990 0
1991 1
1992 6
1993 6
1994 2
1995 4
1996 2
1997 7
1998 2
1999 0
Total 30
issues’ prospectuses. Closing prices of IRs and their underlying shares, together
with market returns, as proxied by the TSE 300 value-weighted index, were
collected from the TSE Summary Information Database, Canada Stockwatch and
TCE Research Services. We used the Canadian Business and Current Affairs
Database to identify the announcement dates (AD) of the offerings. Since some
announcements may have occurred before the opening or after the end of the
trading on the announcement day, our announcement period consists of AD and
AD + 1. The listing dates (LD), on which IRs started aftermarket trading, were
readily available from the TSE Monthly Review.
The first step in our analysis is to document the price behavior of IRs and
their underlying stocks. Our analysis covers the period from 150 days before the
announcements to 10 days after issue dates. Previous studies tend to focus on
the pricing behavior around the announcement and ignore the behavior around the
issue date. If the liquidity effect is present, its effect will be temporary and price
recovery will occur after issue date. It is therefore important to examine the price
pattern around the issue date.
To examine the price pattern of the underlying shares, we calculate the shares’
raw and excess holding-period returns using closing prices for the following seven
periods. The first three periods are AD − 150 to AD − 101, AD − 100 to AD −
51; and AD − 50 to AD − 1. These three periods show the stock price movements
prior to the announcements. The fourth period is the announcement period (AD and
AD + 1). The fifth period is the “interim” period, which is from AD + 2 to LD − 1.
The sixth period is the listing dates of the receipts, LD, while the last period is the
“post-issuance” period, consisting of 10 trading days after LD. Returns for these
periods are then used to determine whether offering announcements are likely to be
made after share prices have substantially risen, whether the announcements create
excess returns on the stocks, and whether the price movements reverse themselves
afterwards.
Excess holding-period returns on security i over any interval j, XRij , are
computed using the “market-adjusted” procedure (see Brown & Warner, 1985),
whereby excess returns are the difference between raw returns and market returns.
That is,
XRij = R ij − M j ,
where Rij is security i’s raw returns over interval j and Mj is market returns over
the same interval.
To measure the magnitude of underpricing of the offerings, we compare IRs’
offer prices to their closing prices on the first trading days. This “offer-to-close”
measure represents the return that one would get if one buys an IR at its offer price
and sells it at the closing price of the first day on which it is listed. This measure
has been used in studies of seasoned/secondary offerings such as Smith (1977) and
Loderer et al. (1991), and particularly in studies of initial public offerings (IPO).
Our reason for using this measure is that each issue of IRs is an offering of a new
security since the issue has not been traded before. Hence, there is no prior closing
price to which to compare.
Note that because IRs are leveraged instruments, our measure of underpricing
will also reflect the embedded leverage effect. Therefore, this measure is
appropriate if one does not plan to hold the IRs until the next installment payment
date. An alternative measure is one that adjusts for the leverage effect. This measure
– which we term “adjusted offer-to-close return” – represents the return that one
would get if, in addition to acquiring an IR, one also makes a risk-free investment
in an amount equal to the present value of its future installment payments. In other
words, adjusted returns approximate returns in case the offerings were done on a
fully-paid basis.
In addition to offer-to-close returns, we also calculate unadjusted and adjusted
raw and excess holding-period returns from holding IRs during the period from
LD + 1 to LD + 10.
4. TIME SERIES PRICE BEHAVIOR

4.1. Underlying Stocks’ Returns
Table 3 presents raw and excess returns on the underlying shares and their
associated p-values. In the three 50-day periods prior to the announcements,
Table 3. Raw and Excess Returns (in Percentages) on the Underlying Shares
around the Announcements.
Interval of Trading Days Raw Returns Excess Returns
Average p-Value % Average p-Value %

(%) Negative (%) Negative
AD − 150 to AD − 101 4.53 0.0268 33 3.48 0.0778 43

AD − 101 to AD − 51 7.62 0.0349 30 4.78 0.1462 43
AD − 50 to AD − 1 10.12 0.0003 23 7.46 0.0031 30
AD to AD + 1 −2.49 0.0092 73 −2.67 0.0054 77
AD + 2 to LD − 1 0.84 0.2838 33 0.36 0.6486 47
LD −0.71 0.3864 67 −0.71 0.3803 67
LD + 1 to LD + 10 2.57 0.0032 37 2.18 0.0126 33
Note: AD is the announcement date. LD is the installment receipt’s listing date.
the average raw and excess returns are all positive. As we get closer to the
announcement dates, both the magnitude of average returns and the proportion of
sample with positive returns increase. Note that the all pre-announcement positive
raw returns are significant at least at the 5% level and there is a clear pattern that
offering announcements are made after sustained substantial run-ups in prices.
The announcements have a significant negative impact on prices. The average
raw and excess returns during the announcement period (AD and AD + 1) are
−2.49 and −2.67% respectively, both of which are statistically significant at the
1% level. More than three quarters of the sample have negative excess returns.
In the interim period (AD + 2 to LD − 1), the average excess return is small
and positive, but not significantly different from zero. Note that this result can be
affected by a selection bias since issuers that experience a substantial decline in
their stock prices during this period may delay or cancel their offerings. (See, for
example, Mikkelson & Partch, 1986.)
On the listing dates of the receipts, the underlying stocks have insignificant
negative excess returns. On the other hand, during the 10-day period after the
issuance, we find strong evidence of price recovery, the magnitude of which closely
matches that of the average price drop on the announcement dates. Altogether, from
the announcement period up to the post-issuance periods (i.e. AD to LD + 10),
the average excess return (not shown) is −0.75% and not significantly different
from zero.
The observed pattern of positive returns prior to the announcement dates
followed by negative returns at the announcements is consistent with the findings
of previous studies of secondary offerings. For example, Korajczyk et al.
report a 10.63% cumulative excess return over a period of 100 days prior to
the announcements, and a negative excess return of −2.49% for the two-day
announcement period. Other studies with similar findings include Asquith and
Mullins (1986) and Hudson et al. (1993).12
Secondary share sales after significant price increases are commonly explained
by information asymmetry. As argued by Myers and Majluf (1984), existing
shareholders know more about their firms’ true values than new investors do. If
they recognize that the shares are overvalued, they will attempt to take advantage
of it. Therefore, their decisions to sell their holdings are construed as a signal that
the stocks are overvalued, which implies that the observed price drops is caused
by a change in the market’s perception of the shares’ value.
Nevertheless, the observed price decline at announcements does not necessarily
imply a change in the market’ valuation of the shares. It is possible that at
announcements, investors anticipate that the offered shares will have to be
underpriced when issued to accommodate the liquidity shock. Therefore, these
investors will sell their holdings of the shares at the announcements in the hope
that they can buy them back at a lower price from the offering. This kind of liquidity
effect is well documented in Scholes’ (1972) study.
As a result, we cannot say for certain what causes the observed price behavior
around the announcement period. In actuality, it is likely that both the information
asymmetry and the anticipation of a liquidity shock contribute to the results. The
price behavior after the announcements provides a clue as to the relative importance
of the two factors. If the price drops at announcement are due to information,
then the drops should be permanent and not reverse themselves since the market’s
valuation of the shares has changed (Scholes, 1972). On the other hand, if the drops
are due to anticipation of a liquidity shock, then they will be temporary. Our findings
of the post-issue date recovery suggest that the majority of the announcement effect
in our sample is caused by the liquidity effect.
4.2. Underpricing of IRs
Table 4 reports the raw offer-to-close returns on IRs, both unadjusted and adjusted
for the leverage effect. The average unadjusted offer-to-close return is 6.08% and
Table 4. Raw Unadjusted and Adjusted Offer-to-Close Returns from IRs.

Types of Return Mean (%) Median (%) Lowest (%) Highest (%) % Negative
Unadjusted 6.08 (0.001) 4.98 −16.25 22.69 27

Adjusted 2.79 (0.002) 2.67 −8.40 10.79 27
Note: p-Values are in parentheses.

is significant at the 1% level. Although the returns range from −16.25 to 22.69%,
only slightly over a quarter of our sample are negative and 13% of the sample are
zero. These figures suggest that, on average, investors will make a sizable profit if
they buy IRs at the offering and sell them at the end of the IRs’ first trading day.
However, that high return is achieved partly with the added risk from leverage.
Adjusted for the leverage effect, the average return reduces to 2.79%, which is
also statistically significant.
To put our results in perspective, we compare them to the findings of previous
studies. We are aware of only two studies that document offer-to-close returns
from seasoned and/or secondary offerings, both of which are conducted on fully-
paid issues in the U.S. markets. First, Smith (1977) reports a small but significant
average offer-to-close return of 0.82%, based on a sample of seasoned offerings
of firms listed on the New York Stock Exchange (NYSE) and the American
Stock Exchange (Amex) during the period from 1971 to 1975. Secondly, Loderer
et al. (1991) find significant average returns of 0.80% for Amex stocks and 1.94%
for NASDAQ stocks for the period from 1980 to 1984. However, they do not find
significant returns for NYSE stocks. In comparison to these two studies, our sample
exhibits higher offer-to-close returns, both adjusted and unadjusted.13 It appears,
therefore, that the offerings in our sample are more deeply underpriced than in
these previous studies. If, as our results in the previous subsection suggests, the
observed underpricing is due to liquidity shocks, then the deeper discounts may
be the result of the lack of depth in the Canadian markets compared to the U.S.
markets.
5. CROSS-SECTIONAL RESULTS
5.1. Announcement Effects
We investigate the possible factors that can help explain the observed price decline
at announcements. We do so by regressing the (excess) announcement returns of all
issues in our sample on: (i) issue size; (ii) the risk of the offered shares; (iii) stock
(excess) returns prior to the announcements, and a liquidity measure. Specifically,
the regression equation is
Ranni = a 0 + a 1 ISSUE SIZEi + a 2 VOLi + a 3 PRERETi

+ a 4 MULTIPLEi + e i . (1)
Ranni is the excess return on stock i during the announcement period. Proxies
for ISSUE SIZEi include FRACTIONi , which is the number of shares offered
divided by the number of shares outstanding for stock i; and VALUEi , which is
the dollar value of issue i, as calculated in Table 1. However, as will be discussed
later in Section 6, the two proxies are correlated. Therefore, they will not appear
in the same regression. We use VOLi , which is the standard deviation of stock
i’s returns during the period from AD − 100 to AD − 51, as our proxy for the
shares’ risk. PRERETi is the excess return on stock i over a period of 100 trading
days prior to the announcement.14 MULTIPLEi , is the ratio between the number
of shares offered and the average daily trading volume over the period from AD
− 150 to AD − 101. Strictly speaking, MULTIPLEi is a proxy for issue size.
However, since it is not correlated with the other two proxies, we treat it as an
independent liquidity measure.
Table 5 presents the results of our regressions. Each column represents
the resulting coefficients of the above regression equation. Each of the above
independent variables was individually run first and then all independent variables
were run together. None of the variables has significant explaining power. The
lack of a significant relationship between PRERETi and announcement period’s
returns is consistent with several previous studies of secondary offerings.15 It
also suggests that the offerings in our sample were, on average, not information-
related. To see this, note that according to information reasons, large price run-
ups before the offerings suggest that the issuers are attempting to sell overvalued
stocks. Hence, the higher the price run-ups, the more negative the impact of the
offering announcements will be. On the other hand, if the sales are not information-
related, no significant relationship should be observed, which is what our results
show. Note also that what we find here is consistent with the argument that we
make in Section 4.1 that the observed post-issue price recovery suggests that the
announcement effect is mainly due to liquidity reasons.
The coefficients for both proxies for issue size – FRACTIONi and VALUEi−
are insignificant, which is consistent with previous studies of seasoned/secondary
offerings.16 The coefficient of VOLi is negative, as predicted by both the
information and the liquidity reasons.17 The lack of significance is similar to the
results in Masulis and Korwar (1986), who calculate variances based on the 60-
day period prior to the announcements, but is in contrast to the results in Hudson
et al. (1993), who calculate variances based on the 250-day period (from day −500
to day −251) prior to the announcements. To make sure that our results are not
influenced by the choice of periods over which return deviation is measured, we
also ran the same regressions with deviation over AD − 150 to AD − 101. All of
the results are not significant. Hence, the lack of significance does not depend on
the choice of periods.
In sum, our regressions of the announcement effect on various independent
variables provide some support for the liquidity explanation, but none for the
124
Table 5. Regression Estimates of Announcement Period’s Excess Returns on Various Independent Variables.
Variablea (1) (2) (3) (4) (5) (6) (7)
Intercept −0.043 (0.024) −0.049 (0.017) −0.029 (0.008) −0.019 (0.227) −0.027 (0.006) −0.027 (0.444) −0.041 (0.095)
NARAT CHARUPAT AND C. SHERMAN CHEUNG

FRACTIONi 0.042 (0.345) 0.036 (0.621)
VALUEi 0.000 (0.196) 0.000 (0.250)
MULTIPLEi 0.000 (0.202) 0.000 (0.793) 0.000 (0.314)
VOLI −0.405 (0.609) −1.011 (0.393) −0.633 (0.495)
PRERETi 0.004 (0.923) 0.013 (0.817) 0.008 (0.880)
Note: p-Values are in parentheses. They are calculated based on standard errors that are computed using the heteroskedasticty consistent estimator of
White (1980).
a VOL is the standard deviation of stock returns during the period from AD − 100 to AD − 51. PRERET is the shares’ excess returns during the
i i
100-day period prior to the announcements. FRACTIONi is the number of shares offered divided by the number of shares outstanding. MULTIPLEi
is the ratio between the number of shares offered and the average daily trading volume over the period from AD − 150 to AD − 101. VALUEi is the
dollar value of issue i.
information reason. We also run the above regressions using raw Ranni and
PRERETi . The results lead to the same conclusion and therefore are not produced
here.
5.2. Underpricing of IRs
Next, we use a regression analysis to ascertain the liquidity effect by testing for
the relationship between the offer-to-close returns on IRs and the proxies for
issue size, liquidity and the risk of the offered shares (i.e. FRACTIONi , VALUEi ,
MULTIPLEi , and VOLi ). As mentioned in Section 2, underpricing is a form of
compensation to investors who have to hold more of a certain security than initially
desired. The larger the size of the offer and/or the riskier the offer, the greater is
the required compensation. As a result, the magnitude of underpricing should be
positively related to the proxies for issue size and risk.
Table 6 reports the results of the regressions. In Panel A, the dependent
variable is unadjusted returns. Again, each of the above independent variables was
individually run first and then all independent variables were run together. First,
we examine the regression results with each independent variable individually
run. Of the two proxies for issue size, VALUEi is significant at the 10% level.
The proxy for liquidity, MULTIPLEi , is significant at the 1% level. However, the
two coefficients have opposite signs to each other. The proxy for risk, VOLi , is
significant at the 10% level, but its negative sign contradicts the predictions of the
liquidity hypothesis (see Note 17). The general regression involving all variables
indicates that none of the variables is statistically significant.
To make sure that the above results are not caused by the leverage effect, we also
report in Panel B the results of the same regressions but with adjusted returns as
the dependent variable. While MULTIPLEi and VOLi continue to be significant
at the 1 and 10% levels respectively, VALUE no longer has explanation power
when regressions were run with only one independent variable. Again, the general
regression involving all variables indicates that none of the variables is statistically
significant.
6. IRS’ RETURNS FOLLOWING THE LISTING DATES

The previous section shows that investors can earn a sizable return if they subscribe
to IRs and sell them on the listing dates. In this section, we look at returns from
holdings IRs during the first 10 days of their trading. That is, we want to know the
additional return that the investors will get if they do not sell their holdings on the
listing dates but instead sell them 10 trading days later.
126
Table 6. Regression Estimates of Offer-to-Close Returns on Various Independent Variables.
Variablea (1) (2) (3) (4) (5) (6)
Panel A
Intercept 0.046 (0.202) 0.013 (0.710) 0.063 (0.001) 0.113 (0.003) 0.085 (0.190) 0.072 (0.254)
FRACTIONi 0.023 (0.773) 0.071 (0.519)
VALUEi 0.000 (0.098) 0.000 (0.231)
NARAT CHARUPAT AND C. SHERMAN CHEUNG

MULTIPLEi −0.000 (0.010) −0.000 (0.712) −0.000 (0.924)
VOLI −3.320 (0.079) −3.040 (0.302) −2.667 (0.353)
Panel B
Intercept 0.028 (0.149) 0.018 (0.338) 0.032 (0.001) 0.057 (0.003) 0.049 (0.141) 0.053 (0.122)
FRACTIONi −0.000 (0.997) 0.020 (0.721)
VALUEi 0.000 (0.435) 0.000 (0.798)
MULTIPLEi −0.000 (0.011) −0.000 (0.811) −0.000 (0.963)
VOLi −1.680 (0.091) −1.575 (0.307) −1.592 (0.320)
Note: p-Values are in parentheses. They are calculated based on standard errors that are computed using the heteroskedasticty consistent estimator of
White (1980).
a VOL is the standard deviation of stock returns during the period from AD − 100 to AD − 51. FRACTION is the number of shares offered divided
i i
by the number of shares outstanding. MULTIPLEi is the ratio between the number of shares offered and the average daily trading volume over the
period from AD − 150 to AD − 101. VALUEi is the dollar value of issue i.
Table 7. Raw and Excess Returns from IRs and Underlying Stocks for the
10-Day Holding Period.
Types of Return Mean Median Lowest Highest %
(%) (%) (%) (%) Negative
Raw returns
Unadjusted IR return 5.25 (0.008) 2.38 −12.50 28.57 33
Adjusted IR return 2.51 (0.006) 0.85 −4.17 14.16 33
Underlying stock return 2.57 (0.003) 0.28 −3.42 13.08 37
Excess returns
Unadjusted IR return 4.86 (0.010) 3.07 −8.51 28.81 40
Adjusted IR return 2.12 (0.023) 2.00 −5.57 15.07 37
Underlying stock return 2.18 (0.013) 1.88 −4.07 13.98 33
Note: p-Values are in parentheses.
Table 7 report raw and excess returns on IRs, on both unadjusted and adjusted
bases, for the 10-day holding period (not annualized). The raw unadjusted returns
have a mean of 5.25%, which is significant at the 1% level. This number is
slightly lower than the average offer-to-close return reported in Table 4, while the
percentage of the sample that have negative returns is slightly higher. Combining
the two periods, investors can, on average, earn over 10% (unadjusted) if they buy
an IR from the offering and sell it 10 days after it is listed.
Adjusted for the leverage effect, the average excess adjusted return on IRs over
the 10-day holding period is 2.12%, which is significant at the 5% level. This
figure is approximately the same as the average excess return on the underlying
shares over the same period, which, as reported in Table 3, is 2.18%. This should
come as no surprise since the adjusted return on an IR is derived from removing
the leverage effect and this should result in obtaining the underlying stock. Also,
arbitrage between the two instruments should ensure that the adjusted return on
an IR and the return on the underlying stock are the same. Hence, both of these
two positive excess returns can be regarded as an evidence of price recovery. This
evidence provides strong support for the liquidity effect.
To provide further support for the liquidity effect, we also test for the relationship
between the 10-day excess returns on the two instruments and proxies for issue
size and liquidity as follows:
(IR)RLD10i = b 0 + b 1 ISSUE SIZEi + b 2 MULTIPLEi + e i . (2)
(IR)RLD10i denotes the 10-day excess (IR) stock return on (IR) stock i. Adjusted
IR returns are not used here because they are equivalent to stock returns as explained
earlier. The results are reported in Tables 8 and 9. The coefficient of MULTIPLEi
Table 8. Regression Estimates of 10-day Excess Unadjusted IR Returns on

Proxies for Issue Size.
Variablea (1) (2) (3) (4) (5)
Intercept −0.026 (0.477) 0.055 (0.121) −0.300 (0.129) 0.033 (0.381) −0.015 (0.683)
FRACTIONI 0.191 (0.055) 0.135 (0.213)
VALUEI 0.000 (0.816) 0.000 (0.912)
MULTIPLEI 0.000 (0.033) 0.000 (0.037) 0.000 (0.381)
Note: p-Values are in parentheses. They are calculated based on standard errors that are computed
using the heteroskedasticty consistent estimator of White (1980).
a FRACTION is the number of shares offered divided by the number of shares outstanding.
i
MULTIPLEi is the ratio between the number of shares offered and the average daily trading volume
over the period from AD − 150 to AD − 101. VALUEi is the dollar value of issue i.
Table 9. Regression Estimates of 10-day Excess Stock Returns on Proxies for

Issue Size.
Variablea (1) (2) (3) (4) (5)
Intercept −0.125 (0.539) 0.251 (0.114) 0.012 (0.207) 0.013 (0.441) −0.005 (0.810)
FRACTIONI 0.088 (0.077) 0.050 (0.336)
VALUEI 0.000 (0.777) 0.000 (0.906)
MULTIPLEI 0.000 (0.004) 0.000 (0.004) 0.000 (0.123)
Note: p-Values are in parentheses. They are calculated based on standard errors that are computed
i
is positive and significant. It appears that large issues produce stronger recovery.
In other words, not only do we have price recovery, but those that potentially have
largest liquidity impact also produce the greatest recovery.
7. THE BENEFITS OF USING IRS

In this section, we test for the benefits of using IRs to facilitate the offerings. As
mentioned in Section 2, the rationale for using IRs is that their use will increase the
demand for the offers and, as a result, reduce the magnitude of the price concessions
that the issuers have to make. The larger the issue size, the more demand IRs need
to help attract.
The attractiveness of IRs depends primarily on the degree of leverage they
offer.18 Therefore, if the conjectured benefits exist, we should observe a positive
Table 10. Regression Estimates of Leverage Offered on Proxies for Issue Size
and Liquidity.
Variablea (1) (2) (3)
Intercept 0.443 (0.000) 0.460 (0.000) 0.525 (0.000)

FRACTIONI 0.193 (0.023)
VALUEI 0.000 (0.010)
MULTIPLEI −0.000 (0.273)
Notes: p-Values are in parentheses. They are calculated based on standard errors that are computed
i
relationship between the degree of offered leverage and proxies for issue size and
liquidity.
The regression estimates of leverage offerred on proxies for issue size and
liquidity are reported in Table 10. Both proxies for issue size – VALUEi and
FRACTIONi – are positively related to the levels of leverage offered to the levels
of leverage offered and are significant at the 1% or the 5% levels respectively.19
In other words, firms with large issues tend to offer more leverage by having a
small first installment payment. This supports our conjecture that issuers use IRs
to reduce the price concessions that they have to make.
8. CONCLUSIONS
This paper offers a comprehensive review of secondary offerings using IR’s.

Despite the common usage of IRs in Commonwealth countries, there has been
very little theoretical and empirical research on this instrument. While we not only
confirm empirical findings in other studies on regular seasoned equity offerings
such as abnormal price increases prior to the issue and then price declines upon
announcements, but also provide additional details on the underpricing of IRs
and their subsequent price recovery. This points to the similarity of secondary
offerings and block trades. The price decline tends to be temporary due to a less
than perfectly elastic demand function for stocks. The evidence is also consistent
with price reactions associated with the inclusion of new stocks into the S&P
500.20
We also examine the rationale for using IRs. The evidence here indicates that
firms with large issues tend to reduce the amount of the first installment, This will
increase demand and reduce the immediate liquidity impact in dollar terms when
the issues hit the market. The IR arrangement can therefore be a very useful device
to minimize the liquidity impact for markets that lack the same depth as the U.S.
markets.
NOTES
1. In this paper, the general term “seasoned equity offerings” refers to the sale of either
new or existing shares to the public, while the term “secondary equity offerings” refers to the
sale of existing shares to the public by major shareholders. Hence, in our context, “secondary
equity offerings” share the same characteristics as block trades since they involve trades in
large amounts of shares without changing the number of shares outstanding.
2. For example, in the 1980s, the British government heavily used IRs to facilitate their
privatization of state-owned enterprises. More recently, in 1998, Ameritech Corp. sold its
(substantial) holding in Telecom New Zealand through an IR plan in the New Zealand
market. In Canada, while IRs have been around for decades, it was not until the 1990s that
they became popular among Canadian investors.
3. The exception is the paper by Hudson et al. (1993).
4. Usually, the investors can choose to pay future installments early and receive the
underlying shares right away. However, that would be irrational as long as interest rate is
greater than zero and IRs can be traded in the market. An exception to this is when the buyers
buy IRs as a part of an arbitrage transaction (i.e. together with a short position in the underling
shares), in which case early payments may be made if the cost of maintaining the short-sale
position is greater than interest that can be earned on future installment payments.
5. To the best of our knowledge, there have been two cases of default in the Canadian
market. To avoid default, some issuers chose instead to reduce the amount of future
installment payments.
6. While IRs are riskier than the underlying shares due to the leverage effect, investors
can, as mentioned earlier, in most cases choose to pay the remaining installment payments
immediately and obtain the underlying shares. Alternatively, investors can easily unlever
their IRs by holding long risk-free assets. Therefore, the additional risk from leverage should
not discourage buyers. Note also that the attractiveness of IRs will depend on the implicit
leverage cost, which, in turn, depends on the terms of the installment payments and the
(unobservable) offer price if the issue were sold on a fully-paid basis.
7. Examples of this type of explanation include the price-pressure hypothesis discussed
in Scholes (1972), and the “price-of-immediacy” model proposed by Grossman and Miller
(1988).
8. Underpricing of secondary offerings can also result from information asymmetry
among the issuers, underwriters and different types of traders provided that the prices
observed in the pre-issue markets are not perfectly revealing. However, as we report in
Section 4, our results do not support this hypothesis.
9. See, for example, Holthausen et al. (1990) and Keim and Madhavan (1996).
10. One company in our sample, Hudson Bay Co., had two offerings in the sample
period.
11. These figures were calculated by multiplying the number of underlying shares offered
by the sum of the required installment amounts, without taking into account the payments’
present values. This method is the same as that used in the prospectuses. Note that the
figures are based on dollars at the time of issues. That is, no adjustment for inflation was
made to them.
12. Asquith and Mullins (1986) find a 9.60% excess return during the period from 160
days to 10 days prior to the announcement dates, followed by a –2.0% excess return during
the announcement period. Hudson et al. (1993) report a 4.28% cumulative abnormal return
during the 60-day period before the announcements, and a −2.65% excess return during
the announcement period.
13. As there has not been a study of offer-to-close returns from seasoned/secondary
offerings in the Canadian markets, we compare our results to the findings from the Canadian
initial public offerings (IPO) markets. Jog and Srivastava (1995) report an average offer-
to-close return of 5.40% between 1971 and 1992. Using a shorter data period (from
1984 to 1987), Clarkson and Merkley (1994) find an average return of 6.44%. These two
results are similar in magnitude to our unadjusted returns, but are higher than our adjusted
returns.
14. The alternative period of 150 trading days before the announcements was also used,
and yielded similar results.
15. Hudson et al. (1993) find no significant relationship between price drops and pre-
announcement returns, while Asquith and Mullins (1986) and Korajczyk et al. (1990) find
a very weak relationship. The choices of periods over which the pre-announcement returns
are measured appear to influence their results.
16. Scholes (1972), Asquith and Mullins (1986) and Korajczyk et al. (1990) find no
significant relationship between issue size and announcement period’s return. Hess and
Bhagat (1986), Masulis and Korwar (1986) and Barclay and Litzenberger (1988) report
similar findings for a sample of industrial issues.
17. If the negative returns on the announcement days are due to anticipation of a liquidity
shock, a mean-variance analysis such as that in Grossman and Miller (1988) can be used to
show that the magnitude of underpricing is positively related to the variance of the stock.
A similar positive relation exists if underpricing is due to information asymmetry. This is
because return volatility can reflect the degree of information asymmetry between existing
and potential shareholders.
18. A somewhat analogous example is the amount of required margins on futures
contracts. It has been shown that the level of margins is inversely related to the demand
for futures contracts (See, for example, Hartzmark (1986), Fishe and Goldberg (1986) and
Addrangi and Chatrath (1999)). One explanation for it is that margins are committed funds
which reduce traders’ flexibility in case profitable investment opportunities subsequently
exist. Unless the traders unwind their futures position and get back the margin deposits, they
will have to borrow to take advantage of the profitable opportunities. Hence, the flexibility
cost depends on the leverage cost and constraint that the traders face. To apply this logic to
our context, IRs can be thought of as futures contracts where the first installment payment
is the initial margin, while fully-paid offers are contracts where the full price is the initial
margin. Therefore, investing in IRs involves lower flexibility cost.
19. Different versions of leverage such as the ratio of future payments (without
adjustment for the present value factor) to the sum of all payments and the ratio of future
payments over the initial payment. The conclusions are the same.
20. See, for example, Shleifer (1986).
ACKNOWLEDGMENTS
Both authors gratefully acknowledge the financial support from the Social Sciences
and Humanities Research Council of Canada. Please address all correspondence
to C. Sherman Cheung at scheung@mcmaster.ca.
REFERENCES
Addrangi, B., & Chatrath, A. (1999). Margin requirements and futures activity: Evidence from the
soybean and corn markets. Journal of Futures Markets, 19, 433–455.
Asquith, P., & Mullins, D. (1986). Equity issues and offering dilution. Journal of Financial Economics,
15, 61–89.
Barclay, M., & Litzenberger, R. (1988). Announcement effects of new equity issues and the use of
intraday price data. Journal of Financial Economics, 21, 71–99.
Clarkson, P. M., & Merkley, J. (1994). Ex ante uncertainty and the underpricing of initial public
offerings: Further Canadian evidence. Canadian Journal of Administrative Sciences, 11,
54–67.
Fishe, R. P. H., & Goldberg, L. G. (1986). The effects of margins on trading in futures markets. Journal
of Futures Markets, 6, 261–271.
Grossman, S., & Miller, M. (1988). Liquidity and market structure. Journal of Finance, 43, 617–633.
Hartzmark, M. L. (1986). The effects of changing margin levels on futures market activity, the
composition of traders in the market, and price performance. Journal of Business, 59,
S147–S181.
Hess, A. C., & Bhagat, S. (1986). Size effects of seasoned stock issues: Empirical evidence. Journal
of Business, 59, 567–584.
Holthausen, R., Leftwich, R., & Mayers, D. (1990). Large-block transactions, the speed of response,
and temporary and permanent stock-price effects. Journal of Financial Economics, 26, 71–95.
Hudson, C. D., Jensen, M. R. H., & Pugh, W. N. (1993). Information versus price-pressure effects:
Evidence from secondary offerings. Journal of Financial Research, 16, 193–207.
Jog, V., & Srivastava, A. (1995). Underpricing in Canadian IPOs 1971–1992 – an update. Fineco, 4,
81–89.
Keim, D. B., & Madhavan, A. (1996). The upstairs market for large-block transactions: Analysis and
measurement of price effects. Review of Financial Studies, 9, 1–36.
Korajczyk, R. A., Lucas, D., & McDonald, R. L. (1990). Understanding stock price behavior around
the time of equity issues. In: R. G. Hubbard (Ed.), Asymmetric Information, Corporate Finance,
and Investment (pp. 257–277). Chicago: University of Chicago Press.
Loderer, C., Sheehan, D. P., & Kadlec, G. P. (1991). The pricing of equity offerings. Journal of Financial
Masulis, R., & Korwar, A. (1986). Seasoned equity offerings: An empirical investigation. Journal of
Mikkelson, W., & Partch, M. (1986). Valuation effects of security offerings and the issuance process.
Myers, S., & Majluf, N. (1984). Corporate financing and investment decisions when firms have
information that investors do not have. Journal of Financial Economics, 13, 187–221.
Scholes, M. (1972). The market for securities: Substitution versus price pressure and the effects of
information on share prices. Journal of Business, 45, 179–211.
Shleifer, A. (1986). Do demand curves for stocks slope down. Journal of Finance, 41, 579–590.
Smith, C. (1977). Alternative methods for raising capital: Rights versus underwritten offerings. Journal
White, H. (1980). A heteroskedasticity-consistent covariance matrix estimator and a direct test for
heteroskedasticity. Econometrica, 48, 817–838.
A NEW APPROACH TO TESTING PPP:
EVIDENCE FROM THE YEN
T. J. Brailsford, J. H. W. Penm and R. D. Terrell
ABSTRACT
Conventional methods to test for long-term PPP based on the theory of
cointegration are typically undertaken in the framework of vector error
correction models (VECM). The standard approach in the use of VECMs
is to employ a model of full-order, which assumes nonzero entries in all the
coefficient matrices. But, the use of full-order VECM models may lead to
incorrect inferences if zero entries are required in the coefficient matrices.
Specifically, if we wish to test for indirect causality, instantaneous causality, or
Granger non-causality, and employ “overparameterised” full-order VECM
models that ignore entries assigned a priori to be zero, then the power of
statistical inference is weakened and the resultant specifications can produce
different conclusions concerning the cointegrating relationships among the
variables. In this paper, an approach is presented that incorporates zero
entries in the VECM analysis. This approach is a more straightforward and
effective means of testing for causality and cointegrating relations. The paper
extends prior work on PPP through an investigation of causality between the
U.S. Dollar and the Japanese Yen. The results demonstrate the inconsistencies
that can arise in the area and show that bi-directional feedback exists between
prices, interest rates and the exchange rate such that adjustment mechanisms
are complete within the context of PPP.
Research in Finance
ISSN: 0196-3821/doi:10.1016/S0196-3821(04)21006-1
135
136 T. J. BRAILSFORD ET AL.
1. INTRODUCTION
The purchasing power parity (PPP) hypothesis implies that, over the long-run,
changes in the exchange rate between the currencies of two countries reflect
changes in the ratio of those countries’ price levels. PPP has important implications
for an explanation of exchange rate movements and for exchange rate forecasting
in the long-term. In general, most studies tend to reject PPP in the short-run while
in the long-run there is both evidence supporting and rejecting PPP.1
One factor contributing to these mixed results relates to the procedure used in
testing. In recent years the theory of cointegration has been widely applied to
tests of PPP. If the individual variables involved in the PPP relationship, namely
the nominal exchange rate and the ratio of domestic to foreign prices, are non-
stationary, but a specific linear combination of them is stationary, then PPP is
claimed to hold. A major criticism of the classical residual-based tests for non-
stationarity, such as the augmented Dickey and Fuller test, is that they lack power to
distinguish between unit root and near unit root stationary processes, and therefore
they have a tendency to accept the null hypothesis of non-stationarity (DeJong
et al., 1989; Hakkio, 1984). This has prompted the use of tests that employ the
null hypothesis of stationarity (Fisher & Park, 1991; Kwiatkowski et al., 1992).
However tests for PPP are sensitive to the null hypothesis employed. That is, PPP
could be rejected under a null hypothesis embodying the presence of a unit root,
but accepted in a test with a null hypothesis of stationarity.
In the context of a vector time-series framework, which indicates the dynamic
relationships among the relevant financial variables, conventional methods to test
for PPP are undertaken in full-order vector error correction modelling (VECM).
However standard full-order VECM models are based on nonzero elements in
all their coefficient matrices. As the number of elements to be estimated in
these possibly over-parameterised models grows with the square of the number
of variables, the degrees of freedom will be heavily reduced. In addition, in
applications of VECM models to financial market data, a priori assumptions of
zero entries may be required. Therefore the use of full-order VECM models may
lead to incorrect inferences. Specifically, if we wish to test for indirect causality, or
Granger non-causality which crucially depends on the positions of zero entries in
the coefficient matrices, and we ignore the entries assigned a priori to be zero and
employ “overparameterised” full-order VECM models, the power of our statistical
inferences is weakened. Further, if the underlying true VECM and the associated
cointegrating and loading vectors contain zero entries, the resultant specifications
can produce conclusions concerning the cointegrating relationships among the
variables which would be different to the conclusions arising where the a priori
zeros are ignored, and give rise to potentially different conclusions concerning PPP.
A New Approach to Testing PPP: Evidence from the Yen 137
If the structure is patterned then an approach that uses a zero-non-zero (ZNZ)

VECM is a more straightforward and effective means of testing for Granger
causality, Granger non-causality, indirect causality and instantaneous causality.
Further, one deficiency encountered in empirical research using cointegration
theory is to provide satisfactory financial and economic interpretation for estimated
cointegrating vectors. As demonstrated by Wickens (1996) it is important to intro-
duce a priori information, perhaps to produce ZNZ patterns. To explicitly address
this issue Penm et al. (1997) present a search algorithm in conjunction with model
selection criteria to identify the optimal specification of a ZNZ patterned VECM for
an I(1) system. However that paper did not fully discuss the existence of zero entries
in causality and cointegration theory. This paper demonstrates an approach that
incorporates zero entries in the VECM. The study re-examines PPP in the context
of a three-variable system that includes prices, interest rates and the exchange rate.
The model is tested using a variety of data relating to the Japanese Yen.
The contribution of the paper is three-fold. First, the approach used in the study
demonstrates how a ZNZ patterned VECM can be used in the context of PPP and
cointegration. Second, the paper demonstrates that instantaneous causality should
not be ignored in tests of PPP (e.g. as in Cheng, 1999), particularly with low
frequency data. Third, the paper shows that PPP does hold within the context of
the Yen in a model that incorporates interest rate effects, but this is only over the
period of an open Japanese economy.
The paper is structured as follows. Section 2 reviews VECM modelling for an
I(1) system and discusses overall causality detection. Section 3 revisits PPP and
briefly reviews prior evidence. The issue of instantaneous causality is examined
in Section 4. Section 5 presents a re-examination of the relationship between the
U.S. Dollar and Japanese Yen. Concluding remarks are presented in Section 6.
2. VECM MODELLING FOR AN I(1) SYSTEM

Begin by considering the general vector autoregressive (VAR) model of the form:
q

y(t) + B ␶ y(t − ␶) = ␧(t), (1)
τ=1
where ␧(t) is an s × 1 independently and identically distributed vector random

process with E{␧(t)} = 0 and:
E{␧(t)␧ (t − ␶)} = V, ␶ = 0,
E{␧(t)␧ (t − ␶)} = 0, ␶ > 0.
V is the residual variance-covariance matrix, and B␶ , ␶ = 1, 2 . . ., q are s × s

parameter matrices,

q
B q (L) = I + B␶L␶,
␶=1
where L denotes the lag operator, and Ly(t) = y(t − 1). It is assumed that the roots
of |B q (L)| = 0 lie outside or on the unit circle to ensure that y(t) can contain I(1)
variables.
Of note, y(t) is integrated of order d, I(d), if it contains at least one element which
must be differenced d times before it becomes I(0). Further, y(t) is cointegrated
with the cointegrating vector, ␤, of order g, if ␤ y(t) is integrated of order (d − g),
where y(t) has to contain at least two I(d) variables.2
Under this I(1) assumption:
B q (L) = B q (1)L + (I − L)B q−1 (L)
Following Engle and Yoo (1991), the equivalent VECM for (1) can be expressed
as:
B q (1)y(t − 1) + B q−1 (L) y(t) = ␧(t), (2)
where y(t) contains variables of two types, namely I(0) and I(1) and = (I − L).
Eq. (2) can be rewritten as:
B ∗ y(t − 1) + B q−1 (L) y(t) = ␧(t), (3)
where B ∗ = B q (1) and B ∗ y(t − 1) is stationary and the first term in (3) is the error
correction term. The term B q−1 (L) y(t) is the vector autoregressive part of the
VECM.
Because y(t) is cointegrated of order 1, the long-term impact matrix, B∗ , must
be singular. As a result, B ∗ = ␣␤ and ␤ y(t − 1) is stationary, where the rank of
B∗ is r, and ␣ and ␤ are matrices of dimensions s × r and r × 2s respectively.
The columns of ␤ are the cointegrating vectors and the rows of ␣ are the loading
vectors.
Model development is more convenient using VECMs, rather than the equivalent
VARs, if the systems under study include integrated time series. Engle and Granger
(1987) note that, for I(1) systems, the VARs in first difference will be mis-specified
and the VARs in levels will ignore important constraints on the coefficient matrices.
Although these constraints may be satisfied asymptotically, efficiency gains and
improvements in forecasts are likely to result by imposing them. The analogous
conclusion applies to I(1) systems, such as those typically encountered in tests
of PPP. Comparisons of forecasting performance of the VECMs versus VARs for
cointegrated systems have been reported in studies such as Engle and Yoo (1987)
and LeSage (1990). The results of these studies indicate that, while in the short-run
there may be gains in using unrestricted VAR models, the VECMs produce long-
run forecasts with smaller errors when the variables used in the models satisfy the
test for cointegration.
Further to these developments, we consider a hypothesis where every (i, j)-
th element, for specified i and j, is zero in all coefficient matrices in a VAR.
If this hypothesis is framed in the VAR expressed by (1), these zero entries
will also hold in the error-correction terms and in the vector autoregressive part
of the equivalent VECM, say (2). A discussion of this property is provided in
Appendix 1. Analogously we can achieve a result that if all (i, j)-th coefficient
elements in the error-correction terms and all (i, j)-th coefficient elements in the
vector autoregressive part of the VECM are zeros, then every (i, j)-th entry is zero
for all coefficient matrices in a VAR.
The implications of the above outcome are straightforward. If yj does not
Granger-cause yi , then any (i, j)-th entry must be zero for all coefficient matrices in
the VAR. Also all (i, j)-th coefficient elements in the equivalent VECM are zeros.
In a similar way, we can demonstrate that if yj does Granger-cause yi, then
the (i, j)-th element of Bq (L) in (1) is nonzero. Also, at least a single (i, j)-the
coefficient element is nonzero in Bq (1) or Bq-1 (L) in the equivalent VECM. Of
note, an indirect causality from yj to yi through ym indicates yj causing yi but
only through ym . Hence, yj Granger-causes ym , ym Granger-causes yi , and yj does
not Granger-cause yi directly. We can easily demonstrate that the VAR in (1) has
nonzero (m, j)-th and (i, m)-th elements and a zero (i, j)-th element in Bq (L). This
indirect causality can also be shown in the equivalent VECM, which has at least a
single nonzero (m, j)-th element and a single nonzero (i, m)-th elements in Bq (1)
and Bq-1 (L). Also all the (i, j)-the elements in the equivalent VECM are zeros.
The above discussion indicates that Granger causality, Granger non-causality
and indirect causality detected from both the ZNZ patterned VECM and its
equivalent ZNZ patterned VAR are identical. Since the use of the VECM is more
convenient, it is obvious the ZNZ patterned VECM is a more straightforward and
effective means of testing for the Granger causal relations. The same benefits will
be present if the ZNZ patterned VECM is used to analyse cointegrating relations.
3. PURCHASING POWER PARITY TESTING

The PPP theory states that exchange rates between currencies are in equilibrium
when their purchasing power is the same in each of the two countries. Formally
the PPP condition can be expressed as follows:

Pt
= Et ,
P ∗t
where Pt denotes domestic price level, P ∗t denotes foreign price level, and Et
denotes units of domestic currency per unit of foreign currency.
The classical way of testing for PPP is to regress the nominal exchange rate,
ln(E t ), against the ratio of domestic to foreign prices, ln(P t /P ∗t ). Standard Wald
statistics are then calculated to test whether the coefficient estimates are consistent
with the restrictions embodied in the PPP hypothesis. This approach usually yields
results that do not accept the PPP hypothesis (e.g. Cumby & Obstfeld, 1984;
Frenkel, 1981; Roll, 1979). A major problem with this approach is that the time-
series properties of nominal exchange rate and prices are not specifically taken
into consideration. If nominal exchange rates and the ratio of domestic to foreign
prices are integrated series, and they usually are, then this test could be biased
toward rejecting the null hypothesis of PPP.
To overcome this problem, cointegration has been widely utilised to test
for PPP. Following Engle and Granger (1987), if ln(E t ), and ln(P t /P ∗t ) are
characterised as integrated of order 1 and if there is a long-term cointegrating
relationship between them, where ␤ denotes the cointegrating vector and ␤ X t =
␤ (ln(E t ), ln(P t /P ∗t )) = ␧t with ␧t a variable which is stationary, then PPP is
claimed to hold.3
Various studies have used the methods of unit root tests and cointegration tests
for PPP by examining the bilateral exchange rates for a number of countries. A
general outcome of these studies is that long-term PPP appears not to hold, when
tests based on short- or medium-length time-series are used (e.g. Mishkin, 1984;
Piggot & Sweeney, 1985; Roll, 1979). However, when longer time samples are
used, PPP generally holds (e.g. Abauf & Jorion, 1990; Froot & Rogoff, 1994;
Lothian & Taylor, 1996). The inconsistency in results can be attributed to the
fact that statistical tests become less powerful in small samples. Another result
is that high-frequency data (e.g. monthly data) does not generally lead to support
of PPP in the long-term (e.g. Corbae & Ouliaris, 1988; McNown & Wallace,
1989; Taylor, 1988). However when researchers shift to low-frequency data
and use cointegration techniques to test PPP, the evidence usually supports the
long-term convergence of real exchange rates toward PPP (e.g. Edison, 1987;
Kim, 1990).
Recently Cheng (1999) has included the interest ratio variable to conduct PPP
testing and causality detection. If the foreign interest rate falls below the domestic
interest rate as a result of an increase in the domestic rate, the domestic currency
may appreciate against the foreign currency.4 Cheng’s analysis concerns PPP
between the USA and Japan using annual data over the period 1951–1994 and
he finds evidence supporting PPP in the long run.
Cheng (1999) uses a VECM to examine the relationship between the U.S. Dollar
and the Japanese Yen and the following one-sided Granger causality test:
M
N

(1 − L)y t = ␣0 + ␣m (1 − L)y t−m + ␤n (1 − L)x t−n
m=1 n=1
Q

+ ␩k (1 − L)z t−k + ␯t (4)
k=1
Note that neither ␤0 nor ␩0 is included in the specification (4). Since a VECM
is equivalent to a VAR model with unit roots, and (4) is valid only when no
instantaneous causality exists among the variables (Geweke, 1982; Hatanaka,
1982), instantaneous causality is implicitly neglected in (4). Moreover, when using
low frequency data, such as annual time-series as in Cheng (1999), the impact of
time aggregation suggests that instantaneous relations should not be ignored.
4. INSTANTANEOUS CAUSALITY INVESTIGATION
The use of VECM models in financial time-series is versatile. The models can
be employed as a means of detecting Granger causality, Granger non-causality
and indirect causality. In this section, the focus is on instantaneous causality.
Specifically, the section demonstrates that the one-sided Granger causality test
proposed in Cheng (1999) is valid only when no instantaneous causality exists
among the variables.
To begin with, consider a bivariate VECM of (2) and define
q−1 11
y1 (t) Bi Bi12 i
y(t) = , B q−1
(L) = L,
y2 (t) Bi21 Bi22
i=1

␧1 (t) ␣1
␧(t) = ,␣= ,
␧2 (t) ␣2
and ␤ y(t − 1) = e(t − 1). Thus we have

q−1
y1 (t) B11 B12 y1 (t − i) ␣1 ␧1 (t)
i i
+ + e(t − 1) = ,
y2 (t) Bi21 Bi22 y2 (t − i) ␣2 ␧2 (t)
i=1 (5)
where

 v11
 c
, ␶ = 0,
E{␧(t)␧ (t − ␶)} = c v22


0, ␶ > 0.
v11 0 and v22 0.
Both y 1 (t) and y 2 (t) are mutually instantaneously caused if and only if c
= 0.
Also both y 1 (t) and y 2 (t) are not instantaneously caused if and only if c = 0
(Penm & Terrell, 1984).5
1 −cv−1 22
Premultiply (5) by the matrix , then we have
0 1
q−1
−cv−1 y1 (t) B11 − cv−1 B21 Bi12 − cv−1
22 Bi
22
1 i 22 i
22 +
0 1 y2 (t) B 21
i Bi22
i=1

y1 (t − i) ␦ ␧a (t)
+ e(t − 1) = ,
y2 (t − i) ␣2 ␧2 (t)
where ␦ = ␣1 − cv−1 −1
22 ␣2 and ␧a = ␧1 − cv22 ␧2 .
Now the first equation of this new system becomes
q−1
q−1

y 1 (t) + ␥i y 1 (t − i) + ␩j y 2 (t − j) + ␦e(t − 1) = ␧a (t), (6)
i=1 j=0
where ␩0 = −cv−1 −1 21
22 . Thus ␩0 = 0, if and only if c = 0. Also ␥i = B i − cv22 B i
11
−1 22
and ␩j = B 12j − cv22 B j .
If ␩0
= 0, (6) is different from the one-sided Granger causality test. Only if
␩0 = 0, is (6) equivalent to the one-sided Granger causality test proposed in Cheng
(1999). Thus this one-sided test is valid only when no instantaneous causality exists
between y 1 (t) and y 2 (t). The same result is obtained in a trivariate system. In
Appendix B we summarise the method of instantaneous causality detection with
all cases for a trivariate system.
5. EMPIRICAL RESULTS
In this section, the PPP hypothesis is re-examined using the ZNZ patterned VECM
modelling proposed earlier. The data are drawn from the U.S.-Japan relation
and replicate Cheng’s (1999) test. Following Cheng (1999), the following three
variables are studied contemporaneously in a stochastic vector system.
(i) Japanese Yen to U.S. Dollar: exchange rate (E) per U.S. Dollar.
(ii) Japanese CPI to U.S. CPI: ratio of price levels (P).
(iii) (1 + U.S. discount rate)/(1 + Japanese discount rate): interest rate ratio (IR).6
While our interest is in assessing the presence or absence of the instantaneous
causality in the proposed E-P-IR system, using the ZNZ patterned VECM
modelling, we also use the cointegrating vectors detected to assist with the answer
to other questions, such as whether the PPP hypothesis in the long-term is supported
by the system selected.
For comparison purposes the following three data sets relating to the above three
variables are obtained (from DataStream™ ).
(i) Annual data over the period 1951–1994.
(ii) Semi-annual data over the period 1974–2000.
(iii) Quarterly data over the period 1974–2000.
The first data set is identical to that of Cheng (1999). For both the second and the
third data sets, the samples begin in 1974. This point is often chosen as a cut-off
as it coincides with the end of the Bretton Woods system and the first world oil
shock.7 For the post-Bretton Woods era, the number of annual observations for the
E-P-IR system is insufficient to conduct the analysis. Moreover, given the earlier
evidence that shows that tests of PPP are sensitive to the sampling interval, semi-
annual and quarterly data are employed. These samples allow an assessment of
long-term PPP and short-term responses.
The variables are log transformed such that y 1 (t) = log(E), y 2 (t) = log(P), and
y 3 (t) = log(IR). Unit root tests indicate that all transformed series are I(1). We
then conduct the search procedures proposed in Penm et al. (1997) to obtain the
optimal ZNZ patterned VAR models.
In the course of selecting the optimal lag order (p) for the autoregressive part
of the VECM system, the principle used by Chen and Wu (2000) to enhance the
procedure is adopted. That is, we examine whiteness for the residual vectors from
the VECM chosen by the Akaike Information Criterion (AIC). If the residual vector
process proves to be non-white, we sequentially increase p to p + 1, and check the
resultant residual vector process until the process is a vector white noise process.
The optimal ZNZ patterned VECM and the optimal ␣ and ␤ are then selected by
using the Schwarz criterion (SC). The search results for each data set are presented
in Tables 1–3 respectively.
The estimated residual variance-covariance matrices and the selected patterns
of the cointegrating vector produce some interesting results. For the annual
Table 1. Annual Sampling – The VECM for the Relationship, Linking

Exchange Rates, Consumer Price Indices and Interest Rates between Japan and
the USA Selected by SC Using the GLS Procedure.
Non-zero (i, j)-th Entries in Estimated Coefficient Matrices, B(␶) and B ∗
␶ i, j Entry i, j Entry (S.E.)

[Standard Error (S.E.)]
1 2,2 −0.7635 3,3 −0.3998

(0.110) (0.111)
3 3,3 −0.5714
(0.123)
4 1,2 0.3064 1,3 −0.2313
(0.154) (0.106)
2,2 0.2159 3,1 −0.8082
(0.111) (0.164)
5 1,1 −0.3000 1,2 −0.3792
(0.150) (0.156)
6 3,3 −0.2665
(0.104)
7 2,1 −0.1812
(0.128)
8 2,2 −0.1417
(0.101)
B∗ 3,1 0.1439 3,2 −0.1333
(0.043) (0.031)
3,3 0.4439
(0.073)
 
6.0042E-03 0 −1.9727E-04
The type of V̂ selected:  0 5.0680E-04 0 
−1.9727E-04 0 4.8548E-05
Residual Analysisa Existing Lags

0 1 2 3 4 5
Normalised value of SC 1 1.014 1.033 1.048 1.065 1.082

Long-term Cointegrating Relationship Identified: log(E) = 0.9256 log(P) − 3.0841 log(IR)
Granger Causal Patternb Recognised:

Note: Variables: y 1 (t) = log(E), y 2 (t) = log(P), y 3 (t) = log(IR). Sample Period: 1951–1994
q−1
VECM: B ∗ y(t − 1) + y(t) + ␶=1 B(␶) y(t − ␶) = ␧(t).
a The residual analysis confirms the residuals have white noise characteristics. For simplicity, the values
of SC for q > 5 are not presented, but can be supplied to readers upon request.
b x Granger-causes y only and not instantaneously: (Notation: x y). feedback, not
instantaneously: (Notation: x y). instantaneous causality only: (Notation: x y).
data set, the non-diagonal V̂ shown in Table 1 indicates the existence of the
instantaneous causality in the system. This outcome could result from the effect of
time aggregation on instantaneous causality in low frequency data. The presence
of causality in the system is prima facie consistent with PPP and the results of
Cheng (1999). Further, the positive relation between log(E) and log(P) is consistent
with PPP in that an increase in relative price levels in Japan is associated with a
depreciation of the Yen. However, a closer inspection of the cointegrating vector
reveals an opposite sign between log(E) and log(IR) indicating that, ceteris paribus,
an increase in IR leads to an appreciation in the Yen. This latter result is inconsistent
with theory which asserts an opposite relation.8
This result could be due to a number of factors. First, since the number of the
observation vectors is only 44, the poor parameter estimates in the cointegrating
vector may be caused by the small sample size. Second, over the time period
studied, the Japanese economy has undergone major change. In the early part
of the sample, the Japanese economy was subject to considerable controls. For
instance, heavy regulation and macroeconomic controls were prevalent during the
1950s and 1960s which included a fixed foreign exchange rate, restrictions on
foreign investment, subsidized central lending and a targeted industrial policy. It
was not until 1964, when Japan joined the OECD, that there was a relaxation of
foreign investment controls. In contrast, the latter part of the sample represents
a period when the Japanese economy developed into a major and open global
economy.9 The results may simply reflect the outcome of combining data from
different economic frameworks. Further disaggregation is not possible due to the
small number of observations. Hence, while there is some evidence from the annual
data that PPP holds, such a conclusion must be interpreted with caution given the
findings in relation to the interest rate effect.
For the semi-annual data set, the non-existence of instantaneous causality is
detected from the diagonal V̂ . This outcome indicates that the possible effect on
instantaneous causality through time aggregation does not arise in the semi-annual
data. In relation to PPP, there is evidence of causality in the system. Further, the
positive relation between log(E) and log(P), and the positive relation between
log(E) and log(IR) shown in Table 2 indicates that an increase in IR or an increase
Table 2. Semi-Annual Sampling – The VECM for the Relationship, Linking

Non-zero (i, j)-th Entries in Estimated Coefficient Matrices, B(␶) and B ∗ :
␶ i, j Entry (S.E.) I, j Entry (S.E.) i, j Entry (S.E.)
1 1,1 0.2935 2,1 1.5094 2,2 0.2641

(0.131) (0.321) (0.101)
2,3 −1.5557 3,1 0.6581 3,3 0.5017
(0.216) (0.137) (0.131)
2 1,3 −0.2143 2,1 1.2662 2,2 0.1452
(0.097) (0.323) (0.091)
2,3 −1.4196 3,1 0.4816
(0.215) (0.167)
3 1,3 −0.2344 2,1 1.0817 2,2 0.1918
(0.096) (0.312) (0.087)
2,3 −1.2674 3,2 −0.1358
(0.213) (0.054)
4 2,1 0.7376 2,3 −1.2130
(0.277) (0.222)
5 2,1 0.7659 2,3 −1.0656 3,1 −0.2304
(0.267) (0.219) (0.151)
6 1,1 0.2710 1,3 −0.2069 2,1 0.6782
(0.124) (0.095) (0.260)
2,3 −0.7863
(0.205)
7 1,3 −0.2610 2,1 0.7484 2,3 −0.7076
(0.093) (0.258) (0.209)
3,2 −0.1101
(0.054)
8 1,2 0.1225 2,1 0.9418 2,2 0.1266
(0.050) (0.254) (0.091)
2,3 −0.8395 3,1 −0.2828
(0.220) (0.134)
9 2,1 0.5438 2,2 0.1673 2,3 −0.5905
(0.237) (0.098) (0.203)
3,1 −0.3607
(0.149)
10 2,2 0.1854 2,3 −0.5359 3,1 −0.4603
(0.109) (0.183) (0.150)
11 2,2 0.2804 2,3 −0.4266 3,1 −0.3685
(0.113) (0.167) (0.148)
3,3 0.3242
(0.104)
12 3,3 0.3628
(0.111)
14 3,2 −0.3252
(0.062)
B∗ 2,1 −2.0284 2,2 0.2774 2,3 1.8547
(0.263) (0.039) (0.175)
 
6.3903E-03 0 0
0 0 4.5149E-05
Residual Analysis Existing Lags

0 1 2 3 4 5
Normalised value of SC 1. 1.008 1.020 1.033 1.042 1.043

Long-term Cointegrating Relationship Identified: log(E) = 0.1367log(P) + 0.9144log(IR)
Granger Causal Pattern Recognised:
Note: Variables: y 1 (t) = log(E), y 2 (t) = log(P), y 3 (t) = log(IR)

q−1
Sample Period: 1974 to 2000; VECM: B ∗ y(t − 1) + y(t) + ␶=1 B(␶) y(t − ␶) = ␧(t).
in P leads to a depreciation of the Yen. This result is now consistent with theory.
For instance, when the price level in Japan is increasing, the Yen depreciates in
order to retain PPP. Further, when the relative interest rate in Japan increases, there
is an associated appreciation of the Yen.
In reference to the Granger causal relations among the variables, feedback
relations exist between the pair of log(E) and log(P), the pair of log(P) and
log(IR), and the pair of log(E) and log(IR). Hence, the feedback within the
system is complete and shocks to any one of the variables will be processed
through the system. In addition, Fig. 1 shows that all roots detected lie outside
the unit circle (> 1). This latter finding shows that the ZNZ patterned VECM
can increase the modelling power. To check the adequacy of the model fit, the
results in Table 2 support the hypothesis that the residual vector is a white noise
process.
We now turn to the quarterly data set. The non-existence of instantaneous
causality is also detected from the diagonal V̂ in Table 3. Thus the effect of the time
aggregation on instantaneous causality can be ignored in high(er) frequency data.
The presence of causality and the positive relations between log(E) and log(P),
and log(E) and log(IR) shown in Table 3 are again consistent with theory. Hence,
we conclude that both the semi-annual and quarterly data analyses support the PPP
Table 3. Quarterly Sampling – The VECM for the Relationship, Linking

Non-zero (i, j)-th Entries in Estimated Coefficient Matrices, B(␶) and B ∗ :
␶ i, j Entry (S.E.) i, j Entry (S.E.) i, j Entry (S.E.)
1 3,1 0.2561 3,3 −0.2394

(0.102) (0.091)
2 2,3 −0.5383 3,1 0.3439
(0.115) (0.102)
3 2,2 0.1812
(0.084)
4 2,2 −0.1886
(0.082)
5 1,1 0.2297 1,3 −0.2214 2,3 −0.3539
(0.098) (0.085) (0.111)
3,2 −0.1028
(0.054)
7 2,3 −0.2054
(0.108)
8 2,2 −0.2566
(0.086)
9 2,1 0.2740 2,3 −0.2867
(0.122) (0.110)
10 1,3 −0.2356 2,2 −0.2069 2,3 −0.3198
(0.087) (0.082) (0.107)
11 2,2 0.2486
(0.074)
13 2,1 0.2807 2,3 −0.3183 3,3 −0.1692
(0.116) (0.099) (0.086)
14 2,2 0.1921
(0.086)
B∗ 2,1 −0.3913 2,2 0.0607 2,3 0.3515
(0.057) (0.010) (0.046)
 
3.2208E-03 0 0
0 0 3.1288E-05
Residual Analysis Existing Lags

0 1 2 3 4 5
Normalised value of SC 1. 1.007 1.014 1.021 1.028 1.036

Long-term Cointegrating Relationship Identified: log(E) = 0.1551 log(P) + 0.8983 log(IR)
Granger Causal Pattern Recognised:
Note: Variables: y 1 (t) = log(E), y 2 (t) = log(P), y 3 (t) = log(IR)

q−1
Sample Period: 1974–2000; VECM: B ∗ y(t − 1) + y(t) + ␶=1 B(␶) y(t − ␶) = ␧(t).
Fig. 1. Histogram of Roots for Semi-Annual Sample. Note: Minimum: 1.002 Median:
1.078 Maximum: 1.895.
Fig. 2. Histogram of Roots for Quarterly Sample. Note: Minimum: 1.015 Median: 1.148
Maximum:1.555.
hypothesis in the long run. In examining the Granger causal relations, although
there is no direct Granger causation from log(P) to log(E), there is however indirect
causation from log(P) to log(E) via log(IR). In addition direct Granger causation
exists from log(E) to log(P), and feedback relations exist between the pair of
log(P) and log(IR), and the pair of log(E) and log(IR). We therefore conclude that
Granger causal relation (directly or indirectly) exists between log(E) and log(P)
in both semi-annual and quarterly data samples. Since all detected roots shown
in Fig. 2 lie outside the unit circle, the results give no support to the hypothesis
of instability. The possibility of a structural shift, which is reflected in unstable
roots, is not a problem. Again, to check the adequacy of the model fit, the results
in Table 3 support the hypothesis that the residual vector is a white noise process.
6. CONCLUSION
In this paper we have demonstrated a new approach that involves ZNZ patterned
VECM modelling to examine PPP, and associated causality and cointegration.
Three contributions have been made. First, the paper demonstrates an approach
that allows for zero entries in the VECM and shows how it can be applied to a three
variable system that includes prices, interest rates and the exchange rate. Second,
the paper shows that the one-sided Granger causality test proposed in Cheng (1999)
is valid only when no instantaneous causality exists among the variables. Since
time aggregation can contribute to instantaneous causality in low frequency data
(such as annual data), instantaneous causality cannot be ignored. Indeed, in a test
using annual data between Japan and the USA, evidence is found of instantaneous
causality. Hence, caution is required when interpreting prior evidence. Third, prior
studies have reported inconsistent results in relation to tests of PPP. In this paper,
the sampling interval and the sample period is varied. Support for PPP is strongest
when semi-annual data are employed. The results indicate that bi-directional
feedback exists between prices, interest rates and the exchange rate and hence
sheds light on the adjustment mechanisms through which PPP is achieved.
NOTES
1. A good survey of the literature can be found in MacDonald (1995).
2. In this paper, we consider only the case d = 1, although the procedure can be generally
applied to models where d > 1.
3. For this case, McFarland et al. (1994) propose that the necessary condition for PPP
exists in the long-term. The necessary and sufficient condition means that these two variables
are cointegrated and the cointegrating vector is ␤ = (1, −1).
4. Cheng (1999) found no evidence of causality between productivity and the terms of
trade against the exchange rate, and thus excluded these variables from the model.
5. Since the VECM of (2) is equivalent to the VAR of (1), thus instantaneous causality
exists between y 1 (t) and y 2 (t) if and only if c
= 0. Also both y 1 (t) and y 2 (t) are not
instantaneously caused if and only if c = 0.
6. Consistent with the Fisher equation, the interest rate ratio is expressed in this form
and is numerically closer than the simple ratio of percentage rates.
7. In time series analysis economic and financial systems are evolving, and although
episodic structural shifts may occur in the test period, model evolution in vector systems
can be handled through time update and order update methods to investigate the structure
changes. These methods utilise each incoming observation to update the model structure
and the model parameters.
8. Cheng (1999) does not report the signs on the coefficients.
9. For a history of the Japanese post-war economy, see Komiya et al. (1988).
REFERENCES
Abauf, N., & Jorion, P. (1990). Purchasing power parity in the long run. Journal of Finance, 45,
157–174.
Chen, S. L., & Wu, J. L. (2000). A re-examination of purchasing power parity in Japan and Taiwan.
Journal of Macroeconomics, 22(2), 271–284.
Cheng, B. S. (1999). Beyond the purchasing power parity: Testing for cointegration and causality
between exchange rates, prices, and interest rates. Journal of International Money and Finance,
18, 911–924.
Corbae, D., & Ouliaris, S. (1988). Cointegration and tests of purchasing power parity. The Review of
Economics and Statistics, 70, 508–511.
Cumby, R., & Obstfeld, M. (1984). International interest rate and price level linkages under flexible
exchange rates, a review of recent evidence. In: Bilson & R. Marston (Eds), Exchange Rate
Theory and Practice. Chicago: University of Chicago Press.
DeJong, D. N., Nankervis, J. C., Savin, N. E., & Whiteman, C. H. (1989). Integration versus trend
stationarity in macroeconomic time-series. Working paper 89–99, University of Iowa.
Edison, H. J. (1987). Purchasing power parity in the long-term: A test of the dollar/pound exchange
rate (1890–1978). Journal of Money, Credit and Banking, 19, 376–387.
Engle, R. F., & Granger, C. W. J. (1987). Cointegration and error correction representation, estimation
and testing. Econometrica, 55, 251–276.
Engle, R. F., & Yoo, B. S. (1987). Forecasting and testing in co-integrated system. Journal of
Econometrics, 35, 143–159.
Fisher, E. O. N., & Park, J. Y. (1991). Testing purchasing power parity under the null hypothesis of
cointegration. The Economic Journal, 101, 1476–1484.
Frenkel, J. A. (1981). The collapse of purchasing power parities during the 1970s. European Economic
Review, 16, 145–166.
Froot, K. A., & Rogoff, K. (1994). Perspectives on PPP and long-run real exchange rates. NBER
working paper 4952.
Geweke, J. (1982). Measurement of linear dependence and feedback between multiple time-series.
Journal of the American Statistical Association, 77, 304–313.
Hakkio, C. (1984). A re-examination of purchasing power parity. Journal of International Economics,

17, 265–277.
Hatanaka, M. (1982). The causality and the exogeneity tests in the simultaneous equations. Osaka
University: Mimeo.
Kim, Y. (1990). Purchasing power parity in the long run: A cointegration approach. Journal of Money,
Credit and Banking, 22(4), 491–503.
Komiya, R., Okuno, M., & Suzumura, K. (Eds) (1988). Industrial policy of Japan. Tokyo: Academic
Press Japan.
Kwiatkowski, D., Phillips, P. C. B., Schmidt, P., & Shin, Y. (1992). Testing the null hypothesis of
stationarity against the alternative of a unit root. Journal of Econometrics, 54, 159–178.
LeSage, J. P. (1990). A comparison of the forecasting ability of ECM and VAR models. Review of
Economics and Statistics, 72, 664–671.
Lothian, J. R., & Taylor, M. P. (1996). Real exchange rate behavior: The recent float from the perspective
of the past two centuries. Journal of Political Economy, 104(3), 488–509.
MacDonald, R. (1995). Long-run exchange rate modeling: A survey of the recent evidence. Staff Papers,
International Monetary Fund, 42, 437–489.
McFarland, J. W., McMahon, P. C., & Ngama, Y. (1994). Forward exchange rates and expectations
during the 1920s: A re-examination of the evidence. Journal of International Money and
Finance, 13(6), 627–636.
McNown, R., & Wallace, M. S. (1989). National price levels, purchasing power parity, and
cointegration: A test for four high inflation economies. Journal of International Money and
Finance, 8, 533–545.
Mishkin, F. (1984). Are real interest rates equal across countries? An empirical investigation of
international parity conditions. Journal of Finance, 39, 1345–1358.
Penm, J. H., Penm, J. H. W., & Terrell, R. D. (1997). The selection of zero-non-zero patterned
cointegrating vectors in error-correcting modelling. Econometric Reviews, 16, 281–304.
Penm, J. H. W., & Terrell, R. D. (1984). Multivariate subset autoregressive modelling with zero
constraints for detecting causality. Journal of Econometrics, 3, 311–330.
Piggot, C., & Sweeney, R. (1985). In: S. Arndt, R. Sweeney & T. Willett (Eds), Exchange Rates, Trade
and the U.S. Economy. Washington, DC: American Enterprise Institute.
Roll, R. (1979). Violations of purchasing power parity and their implications for efficient international
commodity market. In: M. Marshall & S. Giorgio (Eds), International Finance and Trade.
Cambridge, MA: Ballinger.
Taylor, M. P. (1988). An empirical examination of long-term purchasing power parity using
cointegration techniques. Applied Economics, 20, 1369–1381.
Wickens, M. R. (1996). Interpreting cointegrating vectors and common stochastic trends. Journal of
Econometrics, 74, 255–271.
APPENDIX A
ZERO ENTRIES IN A VAR AND ITS EQUIVALENT
VECM FOR AN I(1) SYSTEM.
In an I(1) system, the VECM of (2) can be expressed as:
B q (1)y(t − 1) + B q−1 (L) y(t) = ␧(t).
Further, we have the following relations:
B k (L) = B k (1)L + B k−1 (L)(I − L), k = q and q − 1. (A.1)
If the (i, j)-th entries of B k (L), B k (1),and B k−1 (L)are b ij (L), b ij (1), and c ij (L)
respectively, we have
b ij (L) = b ij (1)L + c ij (L)(1 − L). (A.2)
Now we define c ij (L)by c ij (L) = c 0 + c 1 L + · · · + c k−1 L k−1 , thus
c ij (L)(1 − L) = c 0 + (c 1 − c 0 )L + · · · + (c k−1 − c k−2 )L k−1 − c k−1 L k (A.3)
If b ij (L) = 0, then b ij (1) will also be zero. From (A.2), we have that c ij (L)(1 − L) =
0. (A.3) produces c 0 = 0, (c 1 − c 0 ) = 0, . . ., c k−1 − c k−2 = 0, c k−1 = 0 which
leads to c i = 0, i = 0, 1, . . . , k − 1, and therefore c ij (L) = 0.
At this point, if the (i, j)-th entry of B q (L)is zero, then the (i, j)-th elements of
both B q (1) and B q−1 (L) are zeros.
Therefore it can be concluded that if yj does not Granger-cause yi , then any
(i, j)-th element must be zero for all coefficient matrices in the VAR. Also all (i, j)-th
coefficient elements in the error-correction terms and in the vector autoregressive
part of the equivalent VECM, will also be zeros.
Further, from (A.1) if the (i, j)-th element of B q (L) is nonzero, then at least the
(i, j)-th element is nonzero in B q (1) or B q−1 (L).
Thus, we have just demonstrated that if yj does Granger-cause yi , then the
(i, j)-th element of B q (L) in the VAR is nonzero. In addition at least a single (i, j)-th
coefficient element is nonzero in B q (1) or B q−1 (L) of the equivalent VECM.
APPENDIX B
DETECTING INSTANTANEOUS CAUSALITY IN A
TRIVARIATE SYSTEM
 
v11 e f
 
In a trivariate system of (2), with V =  e v22 h , we have the following
f h v33
eight different patterns of instantaneous causality.
(1) y 1 (t), y 2 (t)and y 3 (t) are mutually instantaneous directly caused if and
only if e
= 0, f
= 0 and h
= 0.
(2) y 1 (t) and y 2 (t) are instantaneously indirectly caused via y 3 (t) if and only
if e = 0, f
= 0 and h
= 0.
if f = 0, h
= 0 and e
= 0.
if h = 0, e
= 0 and f
= 0.
(5) y 1 (t) does not cause instantaneously y 2 (t) and y 3 (t) if and only if
e = f = 0, and h
= 0.
h = e = 0, and f
= 0.
f = h = 0, and e
= 0.
(8) No instantaneous causality exists among y 1 (t), y 2 (t) and y 3 (t) if and
only if e = f = h = 0.
The pattern of instantaneous causal relations can be detected by selecting the case
that minimises the model selection criterion.
CORRELATION AMONG STOCK
MARKETS UNDER DIFFERENT
EXCHANGE RATE SYSTEMS
Paul Sarmas
ABSTRACT
This study investigates the linkage between the Hong Kong stock market
and Singapore stock market and the U.S. stock market during the pre-
and post-East Asia Financial Crisis in 1997 and 1998. It uses multivariate
regression models to study the impact of Hong Kong’s fixed exchange rate
system and Singapore’s free-floating exchange rate system on their respective
stock markets. The results indicate that the exchange rate is not a significant
determinant of linkage between the U.S. and the two Asian stock markets, but
the evidence suggests that stronger post-crisis relationships between the U.S.
and the two Asian stock markets. The evidence also supports a stronger short-
run relationship between the U.S. and Hong Kong stock markets relative to
that between the U.S. and Singapore stock markets.
1. INTRODUCTION
Since the collapse of the Bretton Woods fixed exchange rate system in the early
1970s, most countries have switched from a fixed exchange rate system to a floating
exchange rate system. However, some countries insist on keeping their fixed
exchange rate systems in order to achieve economic stability and development.
Research in Finance
ISSN: 0196-3821/doi:10.1016/S0196-3821(04)21007-3
155
156 PAUL SARMAS
Hong Kong and Singapore are dubbed “twin cities” and share many similar
economic features. For example, both emerged under British governance, and
both are small and open metropolitan economies with well-functioning domestic
private sectors and highly rated public services. The two governments had enjoyed
healthy budget surpluses and accumulated large official foreign reserves before
the Asian economic crisis in 1997–1998. Both economies successfully pursued
an export-oriented strategy and upgraded their industrial structures toward higher
value-added activities. Each of them has become a regional financial center and the
operational headquarters of many multinational corporations in the Pacific Asia
region.
Remarkable contrasts, however, exist between the two economies’ monetary
systems and exchange rate regimes.1 Hong Kong has had a currency board regime
since 1983, with the HK dollar pegged to the U.S. dollar at a fixed rate. For
more than a half century up to the 1970s, Singapore had a classical sterling-based
currency board system. Since then, it has evolved a managed floating system that
maintains the value of the Singapore dollar against a trade-weighted basket of
currencies within an undisclosed band.
This study investigates the relationships between daily returns of two Asian
countries’ stock markets and the U.S. stock market under fixed and flexible
exchange rate systems. The motivation for this study has come about as a result of
developing a strategic plan for a diversified portfolio investment in the prosperous
Asian financial markets. The study proposes to analyze the effect of daily changes
of exchange rates and the U.S. stock market on the daily returns of Hong Kong and
Singapore stock markets, and it also tries to assess the movement of the Asian stock
markets under the two different exchange rate systems. Although other traditional
economic factors are also important, it is the perception of many financial analysts
that exchange rate systems would serve as vital links between the U.S. stock market
and the two Asian countries.
2. FOREIGN EXCHANGE RATE SYSTEMS

AND ASIAN STOCK MARKETS
The rules of the fixed exchange rate regime were set under the Bretton Woods
system and were enforced by the International Monetary Fund (IMF) from 1945
(the end of the Second World War) until 1971. The main features of the Bretton
Woods system are the relatively fixed exchange rates of individual currencies
in terms of the U.S. dollar and the convertibility of the dollar into gold for
foreign official institutions.2 However, as time passed, each country’s fiscal and
monetary policies become more complex, and different unexpected economic
Correlation Among Stock Markets 157
shocks finally made it impossible to have all the currencies pegged to each other
(Award, 1998).
Some countries, including Hong Kong, adopted the a rather rigid currency board
principle, maintaining a fixed exchange regime, while still maintaining the fixed
exchange rate policy either pegged to the U.S. dollar or to other major currencies
(Tornell, 2000). Under this system, the local monetary authority guaranteed the
complete and free convertibility of the local currency against the pound sterling at a
fixed exchange rate. The local note issue had 100% backing by sterling deposits in
London. Continuous speculative attacks could smash market confidence, resulting
in a breakdown of the peg. However, even under such adverse circumstances, the
Hong Kong Monetary Authority (HKMA) has been committed to maintaining
the fixed exchange rate system adopted in October 1983, and this has proven to
be effective in maintaining the confidence during the transition of Hong Kong to
Chinese control in 1997 (Yam, 1999).
Amid the international monetary uncertainty in 1973, as with many other
nations, the Singapore government adopted a managed floating system.
Meanwhile, the authorities continued to relax their foreign exchange controls
gradually, and finally liberalized controls by 1978.3 Under the free-floating
exchange rate system, Singapore successfully withstood the test of time, even
during the Asian Crisis in 1987, and achieved stable money policy and prosperous
business development (Wu, 2000).
3. RELATIONSHIPS OF THE EXCHANGE RATES

AND THE STOCK MARKETS
The correlation among financial markets has been studied extensively, and the
overall results are at best mixed and inconclusive. For instance, Bala Arshanapalli
and John Doukas (1995) studied the presence of a common stochastic trend
between U.S. and the Asian stock market movements during the post-October
1987 period. The evidence suggests that the “cointegrating structure” that ties
these stock markets together has substantially increased since October 1987. The
results indicated that the Asian equity markets are less integrated with Japan’s
equity market than they are with the U.S. market.
In another study, Wong (1995) investigated whether a U.S.-type intra-monthly
seasonal behavior in daily stock index returns also exists in the major Asian
markets. The study revealed that, although similar day-of-the-week effects have
been documented in the U.S. and in the major Asian markets, the U.S.-type
intra-month effect on stock returns is weak and unstable over time in Singapore,
Hong Kong, Malaysia, Taiwan and Thailand. Instead, returns generated in these
158 PAUL SARMAS
markets are fairly independent of the U.S. market. The weak linkage between the
financial markets across nations would suggest that there are potential benefits in
international diversification.
Al Awad and Goodwin (1998) examined short-run and long-run dynamic
linkages among weekly real interest rates for G-10 countries using a variety of
time-series tests. The authors focused on the time-series properties of nominal
interest rates, ex-ante expected rates of inflation, and real interest rates. In-sample
and out-of-sample Granger causality tests were also conducted to evaluate lead/lag
relationships among real interest rates. The results provided strong support for well-
integrated markets, particularly in the long run. Their study indicated that the U.S.
financial markets might have a certain impact on Hong Kong’s and Singapore’s
money markets.
Numerous empirical studies have focused on the relationship between foreign
exchange rates and stock markets. However, the validity of their findings seems
to be sensitive to underlying exchange rate regime and economic conditions. For
instance, Bahmani-Oskooee and Sohrabian (1992) found that there was a two-way
relationship between the changes in levels of exchange rate and the fluctuation of
a stock market. Their empirical results show that there is a dual causal relationship
between stock price and the effective exchange rate of the dollar in the short
run. However, the long-run relationship between the two variables failed to show
statistical significance. Yet another study by Mok (1993) used an ARIMA approach
and the Granger causality test to explore the causality of daily interest rates,
exchange rates and stock prices in Hong Kong for period 1986 to 1991. The result
of that study concluded that the interest rate and exchange rate information was
efficiently incorporated in the stock market prices, both at the daily market close
and at the opening. The author also pointed out that there is a weak bi-directional
causality between stock prices and the exchange rate.
Following the same premise, Ying Wu (2000) used an error correction model to
explore the asymmetric effects of four different exchange rates on Singapore stock
prices, and the effects’ sensitivity to economic instability. Both the Singapore
currency appreciation against the U.S. dollar and Malaysian ringgit and the
depreciation against the Japanese yen and Indonesian rupiah led to a long-run
increase in stock prices for most selected periods of the 1990s. However, the
effect associated with the U.S. dollar exchange rate has a sign reversal between
the 1997–1998 crisis periods and the 1999–2000 recovery periods. Ying Wu’s
study indicated that the influence of exchange rates on stock prices increased in
chronological order in the 1990s. So the positive relationship between the stock
market and the exchange rate tends to receive more attention in recent years.
In a more recent study, Phylaktis and Ravazzolo (2001) studied the long run
and short-run dynamics between stock price and exchange rate, and channels
through which great shocks impacted on these markets. They used cointegration
methodology and multivariate Granger causality tests, and applied the model to a
group of Pacific Basin countries over the period from 1980 to 1998. The evidence
suggested that stock and foreign exchange markets are positively related, and the
U.S. stock market acts as a conduit for these links, where these links seemed not
to be affected by foreign exchange restrictions. Thus, it showed once again that
an inconsistent relationship exists between stock markets and exchange rates, no
matter whether the exchange rate system is fixed or flexible.
4. RESEARCH DESIGN AND METHODOLOGY
Despite the recommendation by Hung and Cheung (1995) to use weekly instead
of daily data to avoid the interference of synchronous trading, this study uses daily
data to capture the effects of market shocks. Also, the results don’t show any
significant improvement when weekly data rather than daily data is used in the test
procedure. This research studies and compares the effect of the daily returns on the
U.S. stock market on Hong Kong and Singapore stock markets under the different
exchange rates system before and after the Asian Financial Crisis in 1997–1998.
It was the intent of this study to provide a more recent and specific picture by
contrasting the interrelationship between the Hong Kong stock market and the
U.S. stock market and between the Singapore stock market and the U.S. stock
market from the perspective of different exchange rate systems.
This empirical study was structured to address a number of specific issues.
First, the stationarity of all the time series variables is tested, because only the
stationary data can be employed in a regression test. The second task performed
was to examine collinearity of the explanatory variables: Hong Kong, Singapore,
and the U.S.4 A collinearity test will show the correlation among the independent
variables of three countries’ data in the regression model. Only the variables with no
collinearity were selected for regression analysis. Third, multiple regression tests
were used to determine the relationship between the dependent variable (daily
return of Hong Kong or Singapore stock market) and independent variables (U.S.
stock market return and the exchange rate changes before 1997 and after 1998).
Finally, application of the Granger causality tests was expected to determine the
pair-wise short-run relationship between those variables.
The structural design of this article can be characterized as an inferential study. It
uses multivariate analysis to assess the statistical significance of various predictors
about a single dependent variable (Zikmund, 2000). The independent variables
are: (1) the daily return of the U.S. stock market; and (2) the daily changes of
the exchange rate for the Hong Kong dollar against the U.S. dollar or the daily
160 PAUL SARMAS
changes of the exchange rate for the Singapore dollar against the U.S. dollar. The
dependent variables are: the daily return of the Hong Kong stock market or the
daily return of the Singapore stock market.
Many time series analyses have been based on the assumption that the underlying
data series are stationary. The test of stationarity is often referred to as a unit root
test. The Dickey-Fuller test is known as one of the early approaches to test unit
roots and the degree of integration. If time series data has no unit root, then it
is said to be “integrated of order zero” or I(0). If data has one or more than one
unit roots, that means it is not stationary. It has to be differenced5 once or more to
become stationary, expressed by I(1) and I(t) (Watsham & Parramore, 1997). The
unit root tests will be applied to HSI (the daily change of the Hang Seng Index of
Hong Kong), HGX (the daily change of the exchange rate for Hong Kong dollar
against the U.S. dollar), SP500 (the daily change of the Standard and Poors 500
Index of the U.S.), STI (the daily change of the Straits Times Index of Singapore),
SGX (the daily change of the exchange rate for the Singapore dollar against the
U.S. dollar).
For a time series, it is appropriate to formulate the following regression equation
for the Augmented Dickey-Fuller (ADF) test. The test will be conducted separately
for each time series data and for pre- and post-Asia Financial Crisis periods.
n

(1 − L)Y t = ␣0 + ␤Y t−1 + Y i (1 − L)Y t−i + e (1)
i=1
where: Y is the series of daily variables being tested, L is the lag operator, ␣0 is an
estimated constant, ␤ is regression coefficients, n is the number of lags differences
for E in approximately white noise,6 and e is the errors or residuals.
The null hypothesis to be tested for unit roots is that the series are non-stationary
in the data levels. For the Dickety-Fuller (ADF) test, if the absolute value of
calculated t-statistic is larger than the critical value provided by the ADF test,
the null hypothesis of unit roots (not stationary) is rejected, and the time series
variables are stationary.
Next, the natural log of the variables will be used to express relations that deal
with proportional changes in the variables for the Multiple Regression Analysis.
In this format, the coefficient of the independent variables will be the percentage
changes in the dependent variables per 1% change in each independent variable.
In addition, the variance inflationary factor (VIF) test will be conducted to test
the collinearity of the independent variables in order to set up better regression
models. If a set of explanatory variables is uncorrelated, then optimal VIF will be
equal to 1.7
The following multiple regression model will be tested separately for each
country’s historical data and for the pre- and post-Asia Financial Crisis period.
Y = ␣0 + ␤0 X + ␤1 SP500 + e i (2)
where: Y is the percentage change in the daily stock index in Hong Kong or
Singapore, ␣0 = an estimated constant, ␤0 and ␤1 are regression coefficients, X is
the percentage change of daily exchange rate, SP500 is the percentage change in
the daily S&P 500 index, and ei is the errors or residuals.
Based on the observed U.S., Hong Kong and Singapore daily stock indices and
foreign exchange rates, the regression models are as follows:
lnHSI = ␣1 + ␤1 HKX + ␤2 lnSP500 + e

(3)
lnSTI = ␣2 + ␤3 SGX + ␤4 lnSP500 + e
where: lnHSI is the logarithm of daily Hang Seng Index measured by the percentage
changes in the daily closed Hong Kong Stock Market Index, lnSTI is the logarithm
of daily Straits Times Index measured by the percentage changes in the daily closed
Singapore Stock Market Index, ␣1 , ␣2 are intercept, an estimated constant, ␤1 , ␤4
are the coefficients of the percentage change in the Hong Kong dollar or Singapore
dollar’s exchange rates and the daily return of the S&P 500, HKX is the percentage
change in HK$ daily exchange rate against US$, SGX is the percentage change in
SG$ daily exchange rate against US$, lnSP500 is the logarithm of S&P 500 index
proxied for percentage change in the daily closed U.S. stock market, and e is the
error or residual term.
In addition to examining the correlation of variables for the exchange rate and the
stock market, this study investigates the short-run dynamics by Granger causality
tests. Because correlation does not necessarily imply causation in any meaningful
sense of that world, Granger causality tests have been used frequently to investigate
short run relationships among two or more variables. A high degree of causality
from one variable to another indicates that the two variables are integrated and
that changes in one variable will tend to cause changes in the other. Thus, the
Granger approach to the question of whether X causes Y is to see how much of the
current Y can be explained by the past value of Y and then to see whether adding
lagged values of X can improve the explanation. Y is said to be Granger-caused by
X if X helps to predict Y, or equivalently if the coefficients on the lagged Xs are
statistically significant.
The causality issue between stock markets and the exchange rate is determined
with the following equation,
Y t = ␣ + ␤i Y t−1 + . . . + ␤j Y t−n + e t (4)

162 PAUL SARMAS
where ␣ is a 3 by 1 vector of parameters representing intercept terms; ␤i are 3 by

3 coefficient matrices of parameters with polynomials in the lag operator L; and
et are a 3 by 1 series of independent random vectors, white noise vector with zero
mean and finite covariance matrix.
The above method will have the expanded version of the Vector Autoregressive
(VAR)8 for Hong Kong or Singapore as follows:
       
HSI A10  A11 (L)
 A12 (L) A13 (L) HSIt−1  
       
 HKX  =  20  +  A21 (L)
A A22 (L) A23 (L) ×  HKXt−1 

 

SP A30 A31 (L) A32 (L) A33 (L) SPt−1
 
eHSI
 
+  eHKX  (5)
eSP
       
STI A10  A11 (L)
 A12 (L) A13 (L) STIt−1  
       
 SGX =  A20  +  A21 (L) A22 (L) A23 (L) ×  SGXt−1 

 

SP A30 A31 (L) A32 (L) A33 (L) SPt−1
 
eSTI
 
+  eSGX  (6)
eSP
where Aio are the parameters representing intercept terms and Aij are the
polynomials in the lag operator L. The lag structure will be arbitrarily tested from
2 to12 lags.
The following six simple hypotheses will be tested to find the pair-wise short-run
relationship (Granger Causality) between the variables for Hong Kong,
H0 : HKX does not cause HSI, H1 : HKX does cause HSI. (7)
H0 : HSI does not cause HKX, H1 : HSI does cause HKX. (8)
H0 : SP does not cause HSI, H1 : SP does cause HSI. (9)
H0 : HSI does not cause SP, H1 : HSI does cause SP. (10)
H0 : SP does not cause HKX, H1 : SP does cause HKX. (11)
H0 : HKX does not cause SP, H1 : HKX does cause SP. (12)
The following six simple hypotheses will be tested to find the pair-wise short-run
relationship (Granger Causality) between the variables for Singapore,
H0 : SGX does not cause STI, H1 : SGX does cause STI. (13)
H0 : STI does not cause SGX, H1 : STI does cause SGX. (14)
H0 : SP does not cause STI, H1 : SP does cause STI. (15)
H0 : STI does not cause SP, H1 : STI does cause SP. (16)
H0 : SP does not cause SGX, H1 : SP does cause SGX. (17)
H0 : SGX does not cause SP, H1 : SGX does cause SP. (18)
The null hypotheses will be rejected if the calculated F-statistics are significant.
It is normal to say that there exists a causation that one variable Granger causes
the other.
5. THE DATA SOURCES

This research primarily involves analyzing and interpreting historical daily data
without considering other factors such as interest rate, inflation rate, money supply,
GDP, and the unemployment rate. The reason for this limitation is that most of
those other data are available on a monthly basis rather than a daily basis, and they
are of exposed to collinearity problems as independent variables. The daily returns
of each stock market consist of the Hang Seng Index (HSI) for the Hong Kong
stock market, the Straits Times Index (STI) for the Singapore stock market, and the
S&P 500 index for the U.S. stock market. The daily foreign exchange rates include
the fixed-linked exchange rate for the Hong Kong dollar against the U.S. dollar
(HKX) and the free-floating exchange rate for the Singapore dollar against the U.S.
dollar (SGX). Due to the 1997–1998 Asian Financial Crisis, the movement of the
stock markets in Hong Kong and Singapore was considered as very erratic during
that period. In order to avoid the potential influence of the 1997–1998 crises, the
data are divided into two sets: the Pre-Asian Financial Crisis Period from January
1995 to December 1996 and the Post-Asian Financial Crisis Period from January
1999 to December 2000.
The data are gathered from the Hong Kong Monetary Authority (HKMA),
Monetary Authority of Singapore (MAS), and U.S. Federal Reserve Bank in
St. Louis. The daily data size is 456 for the pre-crisis period from January 1995 to
December 1996. The daily data size is 464 for the post-crisis period from January
1999 to December 2000.
164 PAUL SARMAS
Table 1. Unit Root Test.

HGX HSI SGX STI SP
Pre-crisis unit root test

ADF Test statistic −9.075 −10.025 −9.247 −10.464 −9.768
1% Critical value −3.457 −3.457 −3.457 −3.457 −3.457
5% Critical value −2.868 −2.868 −2.868 −2.868 −2.868
10% Critical value −2.571 −2.571 −2.571 −2.571 −2.571
Post-crisis unit root test
ADF Test statistic −11.576 −10.073 −8.953 −9.475 −10.848
1% Critical value −3.457 −3.457 −3.457 −3.457 −3.457
5% Critical value −2.878 −2.878 −2.878 −2.878 −2.878
10% Critical value −2.571 −2.571 −2.571 −2.571 −2.571
6. EMPIRICAL RESULTS
The unit root test was performed to check the stationarity of all the variables. The
unit root test results, as shown in Table 1, indicate that all the absolute values of the
calculated ADF test statistics are larger than the absolute value of the MacKinnon
critical values at the 1, 5 and 10% significance level, and thus, the null hypothesis
of a unit root will be rejected. This suggests that all the time series variables for
the regression model are stationary.
The variance inflationary factor (VIF) was used to measure collinearity of the
variable data. The VIF statistics and variance analysis as shown in Tables 2–5
are in the acceptable range from 1.00 to 1.80 for the pre- and post-crisis period.
Therefore, there is no reason to suspect any colliearity for the set of explanatory
variables, and the appropriate regression analysis can be performed.
The multiple regression estimates, as show in Tables 2 and 3, provide the
following relationship between Hang Seng Index (HSI) and daily changes in
exchange rate (HKX) and the U.S. stock market (SP500 ). The multiple regression
analysis of the data, shown in Tables 4 and 5, indicate the strength of the linear
relationship between Singapore’s STI Index and the foreign exchange rate (SGX)
and the U.S. stock market index (S&P500 ). The results can be summarized as shown
below:
lnHSIpre = 36.23 − 16.62 HKX + 1.09SP500

lnHSIpost = −52.84 + 23.65 HKX + 1.92SP500
(19)
lnSTIPre = 6.44 + 0.084 ln SGX + 0.191 ln SP500
lnSTIpost = −4.03 − 3.19 ln SGX + 1.85 ln SP500
Table 2. Regression and Variance Analysis.

lnHSI = ␣1 + ␤1 HKX + ␤2 SP500 + e
Section A – Regression Estimation
Predictor Coef. S.E. Coef. T P VIF
Constant 36.236 5.371 6.751 0.000

HGX −16.617 2.624 −6.343 0.000 1.000
S&P 1.09124 0.01139 95.772 0.000 1.000
S = 0.03169
R2 = 0.96
Adjusted R2 = 0.96
Section B – Variance Analysis
Source DF SS MS F P
Regression 2 9.0897 4.5449 4826.94 0.0000

Residual error 454 0.4276 0.0009
Total 456 9.5173
Note: Pre-Crisis: January 1995 to December 1996.

lnHSI = ␣1 + ␤1 HKX + ␤2 SP500 + e
Constant −52.837 3.542 −14.912 0.000

HKX 23.651 1.916 12.351 0.000 1.800
S&P 1.923 0.0844 22.783 0.000 1.800
S = 0.07317
R2 = 0.81
Adjusted R2 = 0.81
Source DF SS MS F P
Regression 2 10.3872 5.1946 970.53 0.0000

Total 464 12.8604
Note: Post-Crisis: January 1999 to December 2000.

166 PAUL SARMAS

lnSTI = ␣2 + ␤3 SGX + ␤4 S&P500 + e
Constant 6.4372 0.1731 37.18 0

SGX 0.0844 0.233 3.6 0.717 1.2
S&P 0.19129 0.01958 9.77 0 1.2
S = 0.04913
R2 = 0.79
Adjusted R2 = 0.79
Source DF SS MS F P
Regression 2 0.2651 0.13249 54.49 0.0000

Residual error 454 1.1038 0.00243
Total 456 1.3689
Note: Pre-Crisis: January 1995 to December 1996.

lnSTI = ␣2 + ␤3 SGX + ␤4 SP500 + e
Constant −4.0330 0.5487 −7.361 0.000

SGX −3.1913 0.2814 −11.382 0.000 1.100
S&P 1.8455 0.07797 23.671 0.000 1.100
S = 0.08920
R2 = 0.56
Adjusted R2 = 0.56
Source DF SS MS F P
Regression 2 4.7421 2.3706 297.91 0.0000

Total 464 8.4185
Note: Post-Crisis: January 1999 to December 2000.

In the case of Hong Kong, the estimated coefficients of the foreign exchange
rate (−16.62) for the pre-crisis period, and in the case of Singapore, the estimate
coefficient of the foreign exchange rate (−3.19) for the post crisis period may
appear somewhat controversial. The inverse relationship between changes in HSI
and changes in exchange rate in the pre-crisis period, and between changes in
STI and changes in exchange rate during the post-crisis period, can be explained
by the inflation factor. A higher rate of inflation in the home country forces the
domestic currency to lose its value, and hence, investors will demand a higher risk
premium and a higher rate of return, which causes stock prices to fall. Therefore,
depreciating the Hong Kong dollar (or the Singapore dollar) represents higher
inflationary pressure in Hong Kong (or Singapore), which could explain each
country’s sluggish economy and depressed stock market.
However, in the pre-crisis period of the Singapore case and the post-crisis
period of the Hong Kong case, there exists a positive correlation, and this agrees
with the previous findings of positive relationship. In the macro economic view,
depreciating the Hong Kong (or Singapore) dollar will increase exports of the
corresponding country, which usually stimulates their export-oriented economy,
and results in a booming Singapore (or Hong Kong) stock market. In this scenario,
lack of statistical significance is consistent with findings in similar studies involving
different currencies and stock markets.
The correlation coefficients for the U.S. S&P stock index are quite different from
the coefficients of the exchange rate. These two variables resulted in a significant
positive correlation between the Asian Stock markets and the U.S. stock market.
In addition, the movements of the U.S., Hong Kong, and Singapore stock markets
grew stronger during post-crisis period.
The standard errors of the regression, which describes the dispersion of data
points above and below the regression line, indicate that there would little variance
between the predicted values and the actual values. The residual analysis reveals
that the errors are normally distributed, as they are linearly related to HSI and STI,
and the linear assumption holds for this model.
The R2 and adjusted R2 are high enough to conclude that a large portion of
changes in Hang Seng Index can be explained by the Hong Kong dollar’s exchange
rate movements and the S&P 500 daily return during both sub-periods. However,
both statistics indicate a somewhat weaker relationship between the explanatory
variable and the Straits Times Index (Singapore’s stock market index).
The F-statistic is used to determine the significance or existence of the regression
line. At a 5% level of significance, the output F-statistics for both of the Hang Seng
Index (HSI) and the Straits Times Index (STI) exceed the critical value on the F
distribution (with 2 and 453 degree of freedom) which is 3.07. In addition, the
P-value of the F-test for both regression estimates during the pre- and post-crisis
168
Table 6. Granger Causality Test for Hong Kong.
Null Hypothesis Lags: 2 Lags: 4 Lags: 6 Lags: 8 Lags: 10 Lags: 12
F-Stat. Prob. F-Stat. Prob. F-Stat. Prob. F-Stat. Prob. F-Stat. Prob. F-Stat. Prob.
Pre-crisis for HSI, HKX and SP500

HIS not cause HKX 0.038 0.963 0.731 0.571 1.704 0.118 1.210 0.291 1.006 0.437 1.058 0.395
HKX not cause HIS 0.028 0.972 1.289 0.273 1.523 0.169 1.105 0.359 1.291 0.233 1.119 0.343
SP not cause HKX 0.282 0.754 0.649 0.628 0.873 0.515 0.559 0.811 1.182 0.301 1.057 0.395
HKX not cause SP 0.638 0.529 0.452 0.771 0.627 0.708 0.412 0.914 0.442 0.925 0.620 0.825
SP not cause HIS 66.926 0.000 34.117 0.000 22.961 0.000 16.952 0.000 13.605 0.000 12.161 0.000
HSI not cause SP 0.024 0.977 0.902 0.463 1.298 0.257 0.812 0.592 0.844 0.587 0.821 0.629
Post-crisis for HSI, HKX and SP500
HSI not cause HKX 1.604 0.202 2.279 0.060 0.500 0.808 0.721 0.673 0.526 0.872 0.683 0.768
HKX not cause HIS 0.927 0.397 0.730 0.572 3.052 0.006 2.297 0.020 2.097 0.024 2.313 0.007
SP not cause HKE 70.261 0.000 38.083 0.000 0.336 0.918 0.253 0.980 0.328 0.973 0.365 0.975
HKX not cause SP 0.918 0.400 1.716 0.145 0.596 0.734 0.754 0.643 0.659 0.763 0.796 0.655
SP not cause HIS 0.020 0.981 0.212 0.932 24.171 0.000 19.320 0.000 15.673 0.000 13.175 0.000
HSI not cause SP 0.944 0.390 0.741 0.564 1.069 0.380 0.895 0.520 1.072 0.382 1.031 0.419
PAUL SARMAS
Correlation Among Stock Markets
Table 7. Granger Causality Test for Singapore.
Null Hypothesis Lags: 2 Lags: 4 Lags: 6 Lags: 8 Lags: 10 Lags: 12
F-Stat. Prob. F-Stat. Prob. F-Stat. Prob. F-Stat. Prob. F-Stat. Prob. F-Stat. Prob.
Pre-crisis for STI, SGX and SP500

SP not cause SGX 0.697 0.499 0.443 0.777 0.691 0.657 0.789 0.613 0.671 0.752 1.663 0.073
SGX not cause SP 0.999 0.369 0.824 0.510 0.885 0.506 1.174 0.313 1.091 0.368 1.624 0.082
STI not cause SGX 1.566 0.210 1.039 0.386 1.308 0.252 1.639 0.112 2.318 0.012 2.743 0.001
SGX not cause STI 1.861 0.157 1.248 0.290 2.193 0.043 1.718 0.092 0.638 0.782 0.513 0.907
STI not cause SP 0.552 0.576 1.005 0.404 1.486 0.181 1.288 0.248 1.033 0.415 0.843 0.606
SP not cause STI 20.779 0.000 10.422 0.000 7.116 0.000 5.218 0.000 4.476 0.000 3.860 0.000
Post-crisis for STI, SGX and SP500
SP not cause SGX 0.966 0.381 1.106 0.353 0.998 0.426 1.444 0.176 1.381 0.186 1.834 0.041
SGX not cause SP 0.738 0.479 0.558 0.693 0.807 0.565 0.641 0.743 0.678 0.745 0.571 0.866
STI not cause SGX 0.260 0.771 0.765 0.548 1.399 0.213 0.957 0.469 1.143 0.329 1.025 0.424
SGX not cause STI 2.510 0.082 1.777 0.132 1.536 0.165 1.558 0.135 1.286 0.236 1.093 0.364
STI not cause SP 1.615 0.200 0.896 0.466 1.448 0.195 0.858 0.552 1.222 0.274 1.360 0.182
SP not cause STI 34.470 0.000 18.789 0.000 12.574 0.000 9.906 0.000 8.389 0.000 6.809 0.000
169
170 PAUL SARMAS
periods are 0.000 and less than critical value at 5% level. This indicates that at least
one of the explanatory variables is related to daily changes in HSI or STI, and the
regression lines are significant for the regression model to predict the movements
in these indices.
The results of the Granger causality test for the Hang Seng Index are illustrated
in Table 6. During the pre-crisis period, given the arbitrary 2, 4, 6, 8, 10 and 12 lags,
the calculated larger F-statistics and zero probabilities reject the null hypothesis
that the daily return of the S&P 500 index does not Granger-cause the daily return of
the Hang Seng Index. This depicts a unidirectional causality from the daily return
of the U.S. equity market to the daily return of the Hong Kong equity market.
In the post-crisis period, the unidirectional causality starts to appear in the lag
interval of 6, 8, 10 and 12, which reveal the postponed but stronger influence of
the daily change of the U.S. stock market on that of the Hong Kong stock market
after the crisis. Lack of unidirectional or bi-directional causality is evident between
the exchange rate changes and the two equities markets. However, there exists a
unidirectional causality from the S&P 500 daily return to the exchange rate daily
return given 2 and 4 lags period.
The results of the Granger causality test for the Straight Times Index are
summarized in Table 7, which shows no unidirectional or bi-directional causality
between the daily changes of the exchange rate and the two equity markets, either
pre-crisis or post-crisis given the arbitrary 2, 4, 6, 8, 10 and 12 lags. As Hong Kong,
the calculated larger F-statistics and zero probabilities reject the null hypothesis
that the daily return of the S&P 500 index does not Granger cause the daily
return of the Straits Times Index. This indicates a unidirectional causality from
the daily return of the U.S. equity market to the daily return of the Singapore
equity market, either before or after the crisis. In the post-crisis period, there is
a stronger unidirectional causality from the U.S. equity market to the Singapore
equity market, given the same lag interval of 2, 4, 6, 8, 10 and 12. This supports
the multiple regression test result.
7. SUMMARY AND CONCLUSION

Previous studies hypothesized that there exists a certain kind of relationship
between exchange rate and stock markets, but the relationship between exchange
rate fluctuations and the changes of the stock market indices is inconsistent. The
literature also notes that there exists a positive relationship among the international
stock markets. This study focused on the relationships between daily returns of
the two selected Asian markets plus the U.S. stock market and their respective
exchange rates. The main objective was to examine whether these links are different
under the different foreign exchange rate systems pre- and post-Asian Financial
Crisis in the mid 1997 and 1998.
The initial step involved examining this relationship by applying a Multiple
Regression analysis, which tests for the correlation between these countries’ stock
markets and their exchange rates. Then, the Granger Causality test was performed
to study the further interactions between the various markets under the different
exchange rate systems.
The empirical results show that different exchange rates systems have no signifi-
cant role in determining the linkage between the two Asian countries stock markets
and the U.S. stock market. The multiple regression analysis indicates that the
relationship between the stock markets and the exchange rate were inconsistent and
conflicting both before and after the crisis. The exchange rate had a positive effect
on the Hong Kong stock market in the pre-crisis period, but a negative effect in the
post-crisis period. The exchange rate had a small inverse effect on the Singapore
stock market in the pre-crisis period, but positive effect in the post-crisis period.
The further investigation by the Granger Causality test indicates no short-run re-
lationship for the Hong Kong and Singapore exchange rates and their equity market
or the U.S. stock market before the crisis. Nevertheless, after the crisis, there existed
a short-run linkage for the daily changes of the Hong Kong exchange rate and the
daily return of S&P 500. This also supports the inconsistent relationship between
the exchange rates and the equity markets. Both positive and negative effects may
be explained with the inflationary disturbance and the macro economic view.
The complex relationship between stock prices and foreign exchanges can be
supplemented by the Fisher effect and financial theories. However, the impact
of change in the fixed exchange rate on the stock market was too great when
interpreted from the regression output. It is impractical to see the dramatic effect on
the stock market after the change of the exchange rate because consequent dynamic
market adjustments eliminate the results as predicted by the efficient markets theory
and the equilibrium theory. Additionally, Hong Kong’s fixed-linked exchange rate
system limits such change. Therefore, the inference of the exchange rate effect on
the equity market is invalid because the system is shown to be incomplete without
considering other important factors.
The empirical investigation also implies that both of the Hong Kong and
Singapore stock markets were stable and positively correlated with the U.S. stock
market in the periods of pre- and post-crisis. This reflects the strong effect of the
U.S. stock market on the movements of the Asian stock markets. The two tests
also reveal that the movements of the Hong Kong and Singapore stock markets
were closer to those of the U.S. before and after East Asia Financial Crisis.
The causality test reveals only the short-run unidirectional causality from the
U.S. equity market to the two Asian equity markets. In Singapore, there were
172 PAUL SARMAS
consistent effects of the daily U.S. stock market fluctuations on the daily changes
of the Singapore market, either pre-crisis or post-crisis. In contrast, in Hong Kong,
the daily S&P 500 return had an immediate effect on the daily Hang Seng return
before the crisis. However, after the crisis, this effect was delayed until the influence
of the daily exchange rate changes given more lag periods. All these effects support
the previous findings that the linkage was increasing in a chronological order
between the Asian equity markets and the U.S. equity market in the 1990s. The
study further reveals that there were different degrees of the linkage between the
two Asian countries and the U.S. stock markets, and the East Asian Crisis appears
to differ in the degree of that linkage.
NOTES
1. For a detailed discussion refer to Zheng (2001).
2. For further study, please refer to Eiteman (1997) and Shapiro (2003).
3. Please refer to Monetary Authority of Singapore web site (www.mas.gov.sg).
4. There are evidence that both the Hong Kong and the Singapore financial markets
are influenced by the stock market movements in the U.S., the UK and Japan. For further
examination refer to the article by Chan, Cup and Pan (1992). However, this study focuses
on the correlation between the U.S. stock market and those of Hong Kong and Singapore,
in order to account for the effects of the two exchange rate systems: Hong Kong Dollar’s
fixed rate system versus Singapore Dollar’s floating rate system.
5. Differencing is the process of calculating the change in the value of a variable in
successive time periods. Integration refers to the degree of differencing that a datum requires
for it to be transformed into a stationary series (Solnik, Bourcelle & Le Fur, 1996).
6. The equation assumes that a time series is “white noise”, which means the variables
have zero mean, a constant variance, and zero correlation between successive observations
(Leady & Ormrod, 2003).
7. A more conservative criterion would employ alternatives to least-square regression if
the maximum VIF exceeds 5 (Levine, Berenson & Stephan, 1999).
8. Vestor Autoregressive (VAR) is a system in which every equation has the same right
hand variables, and those variables include lagged values of all the endogenous variables.
(Levine, Berenson & Stephan, 1999).
REFERENCES
Arshanapalli, B., & Doukas, J. (1995). Pre and post-October 1987 stock market linkages between U.S.
& Asian markets. Pacific-Basin Finance Journal, 3, 1–14.
Awad, M. A., & Goodwin, B. K. (1998). Dynamic linkages among real interest rates in international
capital markets. Journal of International Money and Finance, 453–519.
Bahmani-Oskooee, M., & Sohrabian, A. (1992). Stock prices and the effective exchange rate of the
Dollar. Applied Economics, 459–464.
Chan, K. C., Cup, B. E., & Pan, M. S. (1992). An empirical analysis of stock prices in major Asian
markets and the United States. The Financial Review, 27, 289–308.
Eiteman, D. K. (1997). Multinational business finance (8th ed.). Reading, MA: Addison-Wesley.
Federal Reserve Bank of St. Louis (FRBS). Internet address at URL: www.stls.frb.org/fred/data/
exchange.html.
Hong Kong Monetary Authority Monthly Bulletin. Internet address at URL: www.info.gov.hk/
hkma/eng/statistics.
Hung, B., & Cheung, Y. L. (1995). Interdependence of Asian emerging equity markets. Journal of
Economic Dynamics and Control, 4, 243–259.
Levine, D. M., Berenson, M. L., & Stephan, D. (1999). Statistics for managers, using Microsoft Excel,
Phstat Excel ‘98 Add-In (2nd ed.). New York: Prentice Hall.
Leady, P. D., & Ormrod, J. E. (2003). Practical research, planning and design (8th ed.). Upper Saddle
River, NJ: Prentice Hall PTR.
Mok, H. M. K. (1993). Causality of interest rate, exchange rate and stock prices at stock market open
and close in Hong Kong. Asia Pacific Journal of Management, 123–143.
Monetary Authority of Singapore. Internet address at URL: www.mas.gov.sg.
Phylaktis, K., & Ravazzolo, F. (2001). Stock price and exchange rate dynamics. Department of Banking
and Finance Working Papers. City University Business School, 321–244.
Shapiro, A. C. (2003). Foundations of multinational financial management (4th ed.). New York: John
Wiley and Sons, Inc.
Solnik, B., Bourcelle, C., & Le Fur, Y. (1996). International market correlation and volatility. Financial
Analysts Journal, 1734–1745.
Wong, K. A. (1995). Is there an intra-month effect on stock returns in developing stock markets?
Applied Financial Economics, 5, 285–289.
Wu, Y. (2000). Stock prices and exchange rates in a VEC model – the case of Singapore in the 1990s.
Journal of Economics and Finance, 36–48.
Zheng, W. Y. S. (2001). Hong Kong and Singapore equity market’s U.S. correlation under the different
exchange rate system. MBA Thesis. Pomona, CA: California State Polytechnic University.
Zikmund, W. G. (2000). Business research methods (6th ed.). New York: Dryden Press, Harcourt
College.
MULTIPLE BANKING AS A
COMMITMENT NOT TO RESCUE
Paul Povel
ABSTRACT
We show why investors may prefer not to be a firm’s unique lender, even if
they are in a strong bargaining position. Some firms need additional funds
after a first investment: providing additional funds is rational after the first
investment is sunk, but together the two investments are unprofitable. A unique
lender will always provide additional funds and make losses. Two creditors
can commit not always to provide funds: inefficient negotiations over debt
forgiveness may end with a project’s liquidation, which is harmful ex post,
but helpful ex ante, if it keeps entrepreneurs with nonpromising projects from
initially requesting funds.
1. INTRODUCTION
This paper analyzes a bank’s incentives to forgive debt and refinance a distressed
firm. We compare the decision of a unique lender with that of two banks, which
have jointly provided a loan to the firm. We show that banks may prefer such co-
financing, even if they enjoy a strong bargaining position relative to the firm. The
main difference between single and multiple banking lies in the negotiations that
are necessary, if the firm cannot repay its debt but it could profitably be refinanced.
Suppose that refinancing is profitable, once an initial investment is sunk, but
that ex ante it is not. Some firms will need refinancing, others not, and the
Research in Finance
ISSN: 0196-3821/doi:10.1016/S0196-3821(04)21008-5
175
176 PAUL POVEL
creditors would like to finance the latter, only. The entrepreneurs of the respective
firms, however, who are informed about their prospective financial needs, are only
interested in receiving a loan, irrespective of whether it will be performing well or
badly. If the creditors could commit not to refinance a firm, the entrepreneurs with
ex ante unprofitable firms would prefer to be inactive, instead of being forced to
liquidate their firm prematurely. A single lender cannot credibly commit to being
tough, as it is always sequentially rational to refinance a distressed firm, once the
initial loan is sunk. We argue that introducing multiplicity on the side of the lenders
can make such a commitment possible. Even if they agree on the need to rescue
the firm, two lenders will have to bargain about the distribution of the overall
loss. Asymmetric information between the banks is the cause of inefficiencies in
the rescue decision: with positive probability the firm is not refinanced, and it is
liquidated, instead.
There is a large literature now, which analyzes the effects of single or multiple
lending on the decisions of a firm. One strand of the literature analyzes the effects
that the structure of the creditors’ claims has on the possibilities to reorganize
a distressed firm. Gertner and Scharfstein (1991) and Detragiache (1994) for
instance assume that bonds are held by atomistic investors and therefore cannot
be renegotiated. They analyze the effects of different bankruptcy regimes on the
possibilities to reorganize a distressed firm.
These effects can be used strategically by a firm, i.e. different financial structures
can be used to achieve different goals. Several papers have asked the question why
a firm may prefer to have one or many creditors. The difference between the
market-based financial system in the U.S. and the bank-based system in Germany
and Japan are striking, and an analysis of the relative advantages of the two systems
is an important research program.
A frequently stated advantage of the “main bank” financial system in Germany
and Japan is that distressed firms are rescued more frequently (see e.g. Hoshi et al.,
1990, for the case of Japan, and Edwards & Fischer, 1994, for the case of Germany).
Some theoretical papers have analyzed the conditions under which “main bank”
finance is more efficient than a system with multiple lenders (see e.g. Dewatripont
& Maskin, 1995; Fischer, 1990; von Thadden, 1995). As Edwards and Fischer
(1990) conclude, however, these models are not compatible with the empirical
evidence for the German case. While in the models at most one “main bank” can
emerge, in reality a German firm has more than one “Hausbank.” The question to
analyze is thus why we may observe more than one nonatomistic lender. Several
answers are possible.
First, one could argue that banks are risk averse and want to spread out their
risk exposure by sharing risks with their competitors. This is certainly true, but not
a very satisfying explanation from a theoretical point of view. Banks are usually
Multiple Banking as a Commitment Not to Rescue 177
thought of as “large,” compared with the size of the average firm. They should
therefore be able to diversify away most of their risks, as was modeled in Diamond
(1984). This makes them de facto risk neutral, and they should not suffer from risk
exposure. After all, it is the banks’ business to deal with risks and to allocate them
optimally, and not to avoid risks. Additionally, it would be interesting to know
whether there is more behind multiple banking than mere risk-sharing.
Second, a bank may lack the funds to finance a project. Dewatripont and
Maskin (1995) suggested that such smallness could be a solution to the soft
budget constraint problem in centralized economies. Inability to finance a project
exclusively may be a real problem when firms are very large. However, even in cases
when the firms are very small, compared with their banks, we find multiplicity. As
before, there is a need for additional explanations.
Third, firms may want to have many banks because this protects them from
being exploited by too strong a partner, as was suggested in von Thadden (1992).
This third rationale for multiple banking implies that neither the banks nor the
firms enjoy exceptionally strong bargaining positions in their relationship. This
contrasts with the general perception that in bank-dominated financial systems,
banks are in a stronger position. Many situations can occur in which a firm has
to rely on its bank or banks and in which the bank can cheaply “punish” earlier
unfriendly behaviour.
Finally, some authors analyze the use of multiple claimants, holding different
types of securities, in solving agency problems: the investors may have poor
incentives either to really monitor their debtor, or to make proper use of their
information (e.g. to liquidate a firm). See e.g. Diamond (1993), Berglöf and von
Thadden (1995), Dewatripont and Tirole (1994), Rajan and Winton (1995), and
Repullo and Suarez (1995).
The present paper offers a rationale for multiplicity, which complements the
explanations above. We argue that multiplicity is requested by the banks, who use
it as a commitment device for eventual renegotiations of the lending contracts.
The inefficiencies that arise in rescue negotiations (the banks have to determine
their respective degrees of debt forgiveness) are a threat for entrepreneurs with bad
projects. If the inefficiencies are sufficiently strong, this allows the banks to deter
nonprofitable projects, and to finance high quality ones, only.
The idea that multiplicity can serve as a commitment device was first stated in
Hellwig (1991). Dewatripont and Maskin (1995) analyze the role of “multiple
lending” in hardening the “soft budget constraint” of a firm. In their model,
however, multiplicity is a credible commitment not to rescue only because of
the assumption that lenders are “small,” and cannot provide both an initial and a
refinancing loan. Bolton and Scharfstein (1996) analyze a renegotiation problem
that is similar in spirit to ours. In their model, too, multiplicity is used as a
178 PAUL POVEL
commitment to be inefficient in renegotiations, with the result that high quality

firms borrow from two creditors, while low quality firms prefer to borrow from a
single creditor. Our model differs from theirs in several aspects. First, we work in
a complete contracting environment. There is no variable in this model, which is
“observable but not verifiable.” In Bolton and Scharfstein (1996), the entrepreneur
can hide the returns of the project, and claim that the returns had been low. An
optimal contract “punishes” him by threatening to liquidate the assets that are
still valuable to him. In our model, the banks want to keep away nonprofitable
projects, i.e. projects with a low probability of being successful. Second, we model
the renegotiation process explicitly, and base it on observations from a financial
system with “main banks.” Bolton and Scharfstein (1996) use the Nash Bargaining
Solution and the Shapley Value, instead, to model bargaining outcomes.
Other related work includes Yosha (1995) and Bhattacharya and Chiesa (1995),
who analyze the strategic use of single or multiple lending as a commitment device
with respect to nonfinancial decisions. More precisely, they study the relative
advantages of public or bank lending, if the two regimes have different effects
on how sensitive information can leak to a firm’s competitors. They thus provide
more and richer explanations for multilateral lending, which add new aspects to
the purely financial models.
A second contribution of this paper is the development of a new model of
inefficient bargaining, which has realistic features. We model the negotiations
between the banks as a war of attrition. As soon as the banks have been informed
that the firm must be refinanced, negotiations start. In these negotiations, each
of the two banks tries to convince its opponent to write down the larger fraction
of its claims. A rescue is only possible if one of the banks gives in: it frees the
way to a rescue of the firm by accepting its opponent’s rescue plan. The reason
why the banks eventually give in is that a rescue may become impossible, and
the firm has to be liquidated. Each bank has a privately known valuation for the
business relationship with the firm, which it loses if the latter is liquidated. The
impossibility to rescue can arise at any time, as soon as the parties have started
to bargain, and the longer the rescue is delayed, the more likely it becomes that
the banks are forced to liquidate the firm. If a bank has a high valuation at risk, it
has strong incentives to accept its opponent’s plan, only to ensure that the firm is
rescued. As the opponent could have an even higher valuation, however, it also has
an incentive to hold out for a while. This tradeoff determines the banks’ strategies
in the war of attrition.
Admati and Perry (1991), Fernandez and Glazer (1991), and Abreu and Gul
(2000) are other papers, in which two parties must come to an agreement in time
consuming negotiations. We could have used variants of these models, instead
of the war of attrition, to capture the inefficiencies of the renegotiation process.
The models in the three papers, however, are somewhat technical, too, and do not
generate more elegant results than our model. We believe, therefore, that the war
of attrition is a good compromise between the requirements for the analysis and
the tractability of the results.
The rest of the paper is structured as follows: In Section 2, the projects and the
entrepreneurs are introduced, and the difficulties of a single bank are discussed.
The model is extended in Section 3, where two banks finance a firm, and renegotiate
if it must be refinanced. These renegotiations are modeled as a war of attrition.
Section 4 solves this model to find the equilibrium of the renegotiation stage, as
well as that of the whole game. Section 5 presents some empirical evidence, and
discusses implications and extensions of the model. Section 6 concludes. Proofs
are in the Appendix A.
2. THE MODEL WITH ONE BANK
There is a large number of entrepreneurs who can start one project each. Each
entrepreneur privately knows the type of project that he can start, either “good”
or “bad.” The proportion of entrepreneurs with “good” projects, ␥, is common
knowledge.
The timing of a project is the following. In the first period, an investment I must
be sunk. In period 2 the project types become publicly observable. Payoffs are
earned in the third (the last) period. A “good” project earns R > I, while a “bad”
project earns zero. Both project types can be liquidated, which earns r, where
0 ≤ r < R. A “bad” project can be “rescued” in period 2: if an additional amount
J is invested, a payoff R̄ is earned, instead of zero.
Assumption 1. It is profitable to rescue a “bad” project in period 2, as
R̄ − J > r. However, it is not profitable to finance a “bad” project ex ante:
R̄ − J − I < 0. Neither should a random sample of projects be financed:
␥(R − I) + (1 − ␥)(R̄ − J − I) < 0.
The entrepreneurs’ payoffs depend on whether their projects were started and
completed. If a project was not started, the entrepreneur earns zero utility. If the
project was started, and either completed successfully (if “good”) or rescued (if
“bad”), his utility is M > 0. If a project was started and then liquidated, this causes
harm to the entrepreneur, and his payoff is −m (where m > 0).
The entrepreneurs have no wealth of their own, and need outside finance to start
their projects. We assume that a project cannot be separated from its entrepreneur.
“Good” projects cannot be continued without him, and “bad” projects cannot
be rescued – both types would have to be liquidated. The entrepreneurs are
180 PAUL POVEL
protected by limited liability. No punishment can be used legally to influence

the entrepreneurs’ decisions, except for the liquidation of the project, which gives
them negative utility.
As we assume that it is not profitable to finance a cross section of projects,
an investor must find a way to separate the “good” from the “bad” projects.
Ideally, only the former would be financed. A bank could propose a contract which
specifies that “bad” projects are liquidated in period 2. It would like to commit
never to refinance, as this would prevent the entrepreneurs with “bad” projects
from applying for initial loans I. Unfortunately, as one can easily verify, such
a threat is not credible. Entrepreneurs with both “good” and “bad” projects will
apply for I, as those with “good” projects have nothing to fear, and those with
“bad” projects know that there will be a rescue. As a result, the single bank faces
a random sample of projects, and it has to reject all loan requests. Due to a lack
of commitment no project is undertaken, even though there would be valuable
investment opportunities.
3. THE MODEL WITH TWO BANKS

The lack of a commitment possibility in the case of a single bank can be overcome
(at least partially) by having more than one creditor for each project. If each of two
banks provides, say, half of the initial loan, both have some rights over the returns of
the firm at t = 3. If the entrepreneur asks for the additional loan J, a part of the total
investment will have to be written off. The banks will bargain over how much each
should forgive. If this bargaining is sufficiently inefficient, and the consequences
of this inefficiency cause harm to the entrepreneur, the underinvestment problem
can be solved.
It will be shown below, that two banks can commit to rescue with a probability
which is strictly smaller than one. There is a critical value for this probability,
which we denote by q̄. It is determined by the entrepreneurs’ utility functions:
q̄M − (1 − q̄)m = 0. (1)
If an entrepreneur’s “bad” project is rescued with probability q̄ and liquidated

with probability (1 − q̄), his expected payoff is exactly zero. He is thus indifferent
between applying for a loan, and being inactive (which earns a sure payoff zero). If
the rescue probability is strictly below q̄, he prefers not to apply for the loan. In this
case, only the entrepreneurs with “good” projects apply for funding. Therefore,
if the banks can credibly commit not to refinance with a probability larger than
(1 − q̄), multiple banking strictly dominates bilateral lending relationships.
The model with two banks incorporates some observations about private
workouts and bankruptcy negotiations that are reported in the business press, in
empirical and descriptive papers (e.g. Edwards & Fischer, 1994; Fischer, 1990),
in studies on the banking system and insolvency procedures in Germany, and
in the large literature on the reform of the bankruptcy laws in Germany. These
observations, or “stylized facts,” are:
(1) Banks seem to have a strong bargaining position.
(2) The parties involved try to keep the negotiations secret.
(3) The banks want to terminate the negotiations quickly.
(4) It is likely that customers and suppliers are lost if they hear that there are rescue
negotiations.
(5) Whether to rescue or not is rarely subject to dispute.
(6) The parties rather bargain about who is to sacrifice how much.
We have used these observations to construct a model of debt renegotiations,
such that it captures important elements of an existing financial system, and it
generates results which can again be compared with reality. To do so, we must
expand the model with a single bank, by adding some assumptions. Two comments
will be helpful before this is done. First, all additional assumptions could have been
added to the model with a single bank, without changing any of the results. This has
not been done, as it would have complicated the exposition unnecessarily. Second,
we will make assumptions that are much more restrictive than is necessary to
generate the results. Again, this is done to simplify the notation. Where assumptions
are “extreme,” we mention this fact, and discuss weaker alternatives.
We model the renegotiation process between the two banks of a firm as a war of
attrition. Each of the two banks tries to convince its opponent to carry the burden
of refinancing. An outside observer of the negotiations will find that no progress
is being made for a while: the banks fail to come to an agreement on how to split
the overall loss R̄ − I − J, if there should be a rescue. The negotiations can end in
two different ways. Either one of the banks gives in, i.e. it accepts the rescue plan
of its opponent. Or fate turns against the firm: a rescue becomes impossible for
exogenous reasons, and it must be liquidated. In the latter case, each bank incurs a
loss (additional to the financial loss). The size of this loss is privately known by the
respective bank. In equilibrium, the higher it is, the more a bank fears liquidation,
and the less it is willing to reject its opponent’s rescue plan.
We now introduce the extensions of the single banking model, incorporating the
observations listed above. The equilibrium of the war of attrition will be analyzed
in Section 4.
The first observation above states that banks are the main players in rescue
negotiations. This is captured by assuming that they are the only bargaining parties,
182 PAUL POVEL
and by assuming that the courts strictly enforce Absolute Priority Rules. These rules
specify that no party may receive any of the returns of the firm, if the banks have
neither been repaid in full, nor have agreed to such a payment.
Observation 2 describes how the banks want to keep the negotiations secret. It
is helpful in achieving this goal to conclude an agreement as quickly as possible
(see Observation 3). The reason for this wish for secrecy lies in the bankruptcy
laws, which in most countries favour the banks (France is a notable exception). The
assets of the firm usually are used as collateral for the loans from the banks, and
absolute priority rules enforce the need to repay these claims first. The customers
and suppliers are the parties who typically do badly in bankruptcy. Similar to a
bank run, they have every incentive to request what they are owed, as soon as they
discover the firm’s problems, and not to engage in any new trades (except possibly
on a cash-only basis). We model this sensitivity of a rescue to the cooperation of
these parties as a heavily reduced form of Observation 4.
Assumption 2. At any time during the rescue negotiations, the public can
discover that there are such negotiations going on. This happens by the time t
with probability F(t). If the negotiations have been discovered, a rescue becomes
immediately impossible, and the project must be liquidated.
Assumption 2 is stronger than is necessary for the results. Nevertheless, it is

not unrealistic. Firms whose assets consist almost exclusively of human capital
are an example. If the competitors of an advertising company find out that it is
in difficulties, they will try to hire its best employees on the spot. Robbed of its
most valuable “assets,” the distressed company is not worth rescuing anymore,
and must be liquidated. For this reason, a formal insolvency in this industry can
end after a couple of hours. Furthermore, there is anecdotal evidence from the
U.K., which indicates that secrecy may be a crucial requirement for a successful
rescue. The Bank of England assists in the rescue of distressed large companies,
by coordinating the parties’ efforts as soon as possible. It is not uncommon that
in the negotiation meetings the parties have to use coded names to identify the
distressed firms, even if everybody is informed about the real ones. Secrecy may
also be relevant if without it potential customers are lost; for example, airlines may
not able to get any more advance bookings if their customers fear being stranded
abroad in case of the airline’s bankruptcy filing.
We have to make some technical assumptions, in order to make the model
tractable:
Assumption 3. The “discovery technology” F of the public has a mass point

with measure ␲ > 0 at t = 2, and a density f with support (2, ␶], where ␶ < ∞.
The mass point at t = 2 is necessary for the uniqueness of the equilibrium

strategies.1 These are determined by two differential equations, the solution of
which is not unique without a socalled boundary condition. The mass point leads to
a static lottery over rescue and liquidation at t = 2, which gives us such a boundary
condition. This lottery is a logic extension of the dynamic war of attrition game to
a discrete pre-stage, and is therefore used in the model: as will be shown below, the
dynamic war of attrition is the limiting case of a discrete time game, if the length
of a time unit becomes infinitesimal.
Assumption 3 further restricts the support of f to a finite interval. The reason for
this is that the results would be difficult to interpret if ␶ = ∞ (it would be possible
that the banks bargain endlessly). It is by no means a necessary assumption.
Furthermore, one can easily imagine why the firm’s distress should be discovered
in finite time. For example, there may be legal obligations to make the distress
publicly known if certain contingencies arise.
Observation 5 states that the negotiating parties normally agree that the firm
should be rescued (if they start to negotiate). This is captured by the complete
information about the costs and returns of a rescue, and by the assumption that a
rescue is profitable (Assumption 1). Not everything is common knowledge between
the negotiating parties, however.
Assumption 4. After signing the initial loan contract, each bank B i develops
a privately known valuation i for the business relations with the firm. The
bank loses i if the firm is liquidated. The valuations are independently and
identically distributed, with a common probability density function g (g is
strictly positive on its support + , continuous and differentiable; denote the
cumulative distribution function by G).
There are many possible interpretations for the loss of i if the firm is liquidated.
For instance, it may be an estimate of future profits from dealing with the firm.
Alternatively, the bank may incur costs or lose profits because the liquidation of
its debtor damages its public image or leads to tighter supervision by the banking
regulator. Finally, i may parametrize agency problems within the bank. A bank
manager’s career prospects may be worsened, if “his” firm must be liquidated.
Similarly, the bank manager and the entrepreneur may have become good friends.
In both cases, the decision making unit in the bank would lose something if the
firm is liquidated, and would prefer to rescue it.
The banks’ willingness to assist a distressed debtor is frequently underlined
in studies of the German financial system (see e.g. Schneider-Lenné, 1992). It is
questioned in Fischer (1990). His evidence, however, is based on interviews with
insolvency practitioners, and can therefore be assumed to be biased to the banks’
disadvantage. In their analysis of private workouts in the U.S., Gilson et al. (1990)
184 PAUL POVEL
conclude that restructuring is the more likely, the more debt is owed to banks. This
may be caused by the banks’ superior skills and capabilities in attempting to rescue
a firm, but it may also signal that banks are more willing to rescue a firm than other
creditors. In the model this willingness to rescue is captured by the valuation i .
Assumption 4 and the next assumption jointly capture Observation 6, that the
banks bargain about who has to bear how much of the loss. The set of outcomes
that the banks can achieve is restricted to simplify the analysis, that is how the net
surplus s (the returns R̄ minus the cost J and the opportunity cost r) from rescuing
can be split (it is positive because of Assumption 1).
Assumption 5. The banks fight for the whole surplus s := (R̄ − r − J). No
offer to share the surplus is made or accepted. If one bank gives in it receives its
share r i in the liquidation value r of the firm from the other bank, where r 1 and
r 2 are specified in the initial contract. The winning bank is committed to rescue
the firm immediately, but may keep the returns for itself.
As before (in Assumption 2), the formulation of Assumption 5 is much stronger
than necessary. A sharing rule saying that the gross surplus R̄ − J can only be
shared in proportions ␣ and (1 − ␣), where ␣ = (1/2), would be sufficient. This
would lead to significant complications of the analysis, however, which are not
rewarded by the additional insight that one gets.
This completes the introduction of the model with two banks. As one can easily
see, the assumptions that have been added in this section could also have been
introduced in the single bank model, without changing anything. The loss of a
valuation i if the firm is liquidated would make a single lender even more willing
to refinance a “bad” project. This rescue happens already without the valuation,
however.
In the model with one bank, a strategy for the bank consisted of a financing and
a refinancing decision. In the case with two banks it is slightly more complicated.
We first consider the part of the strategy which is used in the rescue negotiations.
If a firm needs refinancing, the sequence of events is the following. First, the banks
decide whether they want to give in immediately. If none of the banks has given in,
the negotiations are discovered with probability ␲, and the firm must be liquidated.
With probability (1 − ␲) the continuous time war of attrition starts. We assume
that if both banks give in simultaneously, each “wins” with probability 1/2.
A strategy is a function T i : + → [2, ␶], which determines for each moment of
time whether a bank B i with valuation i should give in or not. It will be shown
in Section 4, that if the equilibrium strategy tells this bank to stop at time T i (i ),
it will stop at every later time, as well. Thus, we will define T i as determining the
first time at which a bank plans to stop. This includes the static lottery which is
played because of the mass point in F at t = 2.
One may wonder why the banks cannot renegotiate the lending contract, after
it has been signed. Both are fully aware of the inefficiency that will arise, if the
contract is renegotiated using the war of attrition. Why cannot one bank (or a third
bank) take over all debt for a flat price? Suppose B 1 would make such an offer to
B 2 . B 2 would claim to have a valuation 2 = 0 and not to fear the war of attrition,
in order to increase the takeover price. B 1 would claim to have the same valuation,
to decrease the price. None of the two has any incentive to admit having a positive
valuation, until a rescue is really needed. In this case, however, the war of attrition
will start. The time that passes by is the only credible information about one’s
valuation, as talk is “cheap,” and neither before nor during the war of attrition the
parties can renegotiate more efficiently. Even a bank with valuation i = ∞ would
wait until a rescue is necessary, as it might be that the opponent gives in. Nothing
is lost by waiting until t = 2, at which time both banks can prevent a liquidation
with probability one by giving in.
4. EQUILIBRIUM STRATEGIES
The first step in solving the renegotiation game is to determine which types
would want to start the war of attrition, and which types would prefer to give
in immediately, in order to secure the rescue of the firm. If no bank gives in
immediately, the negotiations are discovered with probability ␲ (the mass point in
F), and the firm is liquidated. With probability (1 − ␲) the continuous time war of
attrition starts.
A bank with a very high valuation at stake will not want to gamble for the
surplus s, and stop immediately. We must determine which is the lowest valuation,
for which this is still true. Denote this cut-off value of bank B i with ␭i . If its
valuation is i > ␭i , it should strictly prefer to give in immediately, while if
it is i < ␭i , it should want to start the war of attrition, and plan to stop later
than t = 2.
We define ␭i as the valuation with which a bank B i is indifferent between giving
in immediately, and starting the war of attrition, if it is sure that the opponent will
either give in immediately (with probability 1 − G(␭2 )), or will start the war of
attrition without giving in (with probability G(␭2 )).
Consider the bank with valuation i = ␭i − ␧, where ␧ > 0. Given the definition
of ␭i , there must be a ␦ > 0, such that it will strictly prefer to start the negotiations,
if the probability that the opponent gives in immediately, as soon as the negotiations
have started, is ␦. Thus, a bank with a valuation below ␭i has an incentive to hold
out for a strictly positive amount of time. A bank with a valuation higher than ␭i ,
however, strictly prefers to give in immediately.
186 PAUL POVEL
Lemma 1. The cut-off values ␭1 and ␭2 are defined implicitly by

s 1 − G(␭2 ) s 1 − G(␭1 )
␭1 = and ␭2 = . (2)
2␲ G(␭2 ) 2␲ G(␭1 )
Since ␭i is continuous and monotonic in ␭j , a symmetric solution exists. It can
happen that there are multiple solutions, since the two equations in Lemma 1 must
be solved simultaneously. We assume that the banks play the symmetric solution
in this case, and denote the common cut-off value with ␭.2
As was mentioned before, the war of attrition is only one of many possible ways
to model negotiations with inefficient delays. The model could have been slightly
simplified by assuming that the support of G is bounded (see Assumption 4).
Suppose it was common knowledge that the highest value i that a bank can
attribute to its business relationship with a firm is A < ∞, because a bank’s line
manager cannot “bet the ranch.” The results would be qualitatively the same,
except that we would have ␭ = A. Our formulation allows for banks that give in
immediately with a certain probability, depending on the parameters of the model.
If both banks decided to stay in the game, the war of attrition starts. A strategy
T i in this war of attrition specifies the earliest instant at which a bank wishes
to stop, given the realisation of its potential loss, i . Lemma 2 derives some
characteristics that equilibrium strategies must have. In Proposition 1 we will show
that these necessary conditions are also sufficient conditions for the existence of a
unique equilibrium, together with the boundary conditions that are determined in
Lemma 1.
Lemma 2. Let T 1 and T 2 be equilibrium strategies of the game defined above.
Then the strategy T i is (i) strictly decreasing in the liquidation loss i , (ii)
continuous, (iii) differentiable, and (iv) bank B i stops at ␶ if and only if i = 0.
In equilibrium it will never be the case that the bank with the higher loss level
will decide to stay in longer than its opponent. The threat of the public’s discovery
must have strictly more weight in a bank’s reasoning the higher i is, while the
gain from winning, the surplus s, is constant. Only a bank with zero liquidation
loss will wait until ␶, and it will not want to stop earlier than ␶. A bank with strictly
positive loss level will either stop immediately at t = 2 (if it has costs i ≥ ␭) or
at some moment after t = 2 but earlier than ␶.
At every moment, both players update their beliefs about the opponent’s
valuation. Since T i is continuous and strictly decreasing, each i is mapped
one-to-one with a stopping time T i (i ). T i can be inverted to yield a function
L i : [2, ␶] → + . At each instant t i there is a valuation L i (t i ) with which a
bank would plan to stop. As time passes by, a player’s expectation about the
maximal valuation that his opponent could possibly have decreases. Bygones are
not “bygones” in this game: every second that passes by signals information about
a bank’s valuation, and is relevant for the present and future decisions of the
opponent.
As was mentioned above, the finiteness of ␶ is not a necessary condition for the
tractability of the model. If the function f had an infinite support, then Lemma 2.(iv)
would state that banks with zero liquidation loss never stop, and banks with strictly
positive loss levels plan to stop at some finite time.
L i , the inverse of the strategy function T i , is the lowest cost level that would make
bank B i want to stop at time t. It will be helpful for characterising the equilibrium
strategies in the following. These are determined by finding for each moment t 1
a valuation L 1 (t 1 ), such that B 1 is exactly indifferent between stopping at t 1 , and
waiting for a small amount of time , and giving in then (the derivation is similar
to that of the cut-off values ␭i ).
If the bank gives in at time t 1 , its payoff is r 1 for sure. We require this payoff to
be equal to the expected payoff, if it decides to wait until t 1 + :

G(L 2 (t 1 + )) (F(t 1 + ) − F(t 1 )) (1 − F(t 1 + ))
r1 = (r 1 − L 1 (t 1 )) + r1
G(L 2 (t 1 )) 1 − F(t 1 ) 1 − F(t 1 )

G(L 2 (t 1 )) − G(L 2 (t 1 + )) F(t 1 + ) − F(t 1 )
+ (r 1 − L 1 (t 1 ))
G(L 2 (t 1 )) 1 − F(t 1 )

1 − F(t 1 + )
+ (R̄ − r 2 − J) . (3)
1 − F(t 1 )
The second expected payoff (on the right-hand side of Eq. (3)) has four components.
The opponent may have a low valuation, and plan to give in later than t 1 + .
By this time, the negotiations may have been discovered, and the firm must be
liquidated. The bank receives its share r 1 of the liquidation value r, but loses
L 1 (t 1 ). If the negotiations are not discovered, it will give in at time t 1 + , which
earns r 1 . On the other hand, the opponent may plan to give in between t 1 and
t 1 + . As before, the negotiations may be discovered, or they may not. In the
latter case, the firm is rescued. The bank pockets the surplus R̄ − J, and pays r 2
to the opponent. We abstract from the possibility that both may give in at t 1 +
simultaneously, as the probability that this happens is negligible.
Equation (3) can be simplified by rearranging, subtracting r 1 on both sides, and
by substituting s for (R̄ − r 1 − r 2 − J). A division of both sides by leads to

G(L 2 (t 1 )) − G(L 2 (t 1 + )) 1 − F(t 1 + )
s
G(L 2 (t 1 )) 1 − F(t 1 )

F(t 1 + ) − F(t 1 )
=− L 1 (t 1 ). (4)
(1 − F(t 1 ))
188 PAUL POVEL
Since the strategies are differentiable everywhere it is possible to take the limit as
goes to zero. The same procedure can be repeated for the second bank, and we
get a system of two differential equations:

G(L 2 (t 1 )) f(t 1 ) L 1 (t 1 )
L 2 (t 1 ) = − , (5)
g(L 2 (t 1 )) 1 − F(t 1 ) s

G(L 1 (t 2 )) f(t 2 ) L 2 (t 2 )
L 1 (t 1 ) = − . (6)
g(L 1 (t 2 )) 1 − F(t 2 ) s
Given the strategy of the opponent, Eq. (5) determines the optimal response of
bank B 1 , if it has loss level L 1 (t 1 ) = 1 (the two are equivalent, if the equilibrium
strategy tells bank B i with cost level i to stop at time t i ) and bank B 2 plays strategy
L 2 (·). If Eq. (5) were an inequality, B 1 would either want to wait longer than t 1
(if <), or it would want to have stopped earlier (if >).
Since by Assumption 2 the probability density function g is strictly positive on
−1 : [0, 1] →
+ , G has an inverse function G + . Equations (5) and (6) can be
integrated, and this leads to the following reaction function for bank B i :
ti
f(t) L i (t)
L j (t i ) = G −1 G(␭)exp − dt . (7)
2 1 − F(t) s
Equation (7) implicitly describes the strategy of bank B j that makes bank B i
exactly indifferent between stopping at t i and stopping at t i + (where is a
small amount of time), given its cost level L i (t i ). The analogous can be done
to derive the strategy of the other bank. The solution to these two equations
will give us the equilibrium strategies for the banks. We will continue with the
differential equations (5) and (6), and show that there is a unique equilibrium. The
reaction functions will be helpful in Section 5, where we present some comparative
statics.
With the help of the differential equations and the boundary conditions it is
now possible to describe the equilibrium strategies of the players for the whole
renegotiation game.
Proposition 1. The renegotiation game has a unique symmetric Bayesian
equilibrium, which is implicitly described by the system of differential equations
(5) and (6), and the boundary conditions T 1 (␭) = T 2 (␭) = 0. The equilibrium
strategy for bank B i is to stop at time t if and only if i ≥ L i (t), where L i (t) is
determined in Eq. (7).
We can now find the equilibrium strategies for the whole game with two banks,
including the financing decision. Whether an entrepreneur with a bad project
applies for a loan in the first period depends on the probability with which his
project is rescued in the second period. In Eq. (1) we determined an upper bound
q̄ to this probability, such that “bad” projects are not financed.
Proposition 2. If the probability of non-rescue due to bargaining delays is high
enough,
␭
M
2 F(T 1 (1 ))G(1 )g(1 ) d1 ≥ , (8)
0 M+m
the entrepreneurs will apply for the initial loan if and only if the project is of
the “good” type.
Proposition 2 is the main result of the paper. There are cases in which a financial
system with multiple banking performs strictly better than one with single bank
lending. If the condition in Eq. (8) is met, the banks prefer to require co-financing
by a second bank to being a single lender.
5. EMPIRICAL IMPLICATIONS
The main result of the paper is that banks might want to syndicate a loan to a firm,
if they fear to find themselves in a harmfully weak bargaining position if the firm
has to be refinanced. The loan is shared for strategic reasons, and the banks propose
to share even if they have all bargaining power. There can be other reasons for why
loans are syndicated, however, like (see Section 1) risk aversion, the sheer size
of the loan, or because the strong competition on the lenders’ side. These reasons
complement each other, and it is not clear which one was the most important if a
loan has been shared.
There is some empirical work on this question for the U.S. and for Germany. For
the U.S., Gilson et al. (1990) have analyzed the performance of private workouts.
One of their results is that debt restructurings are more likely if the number of
lenders is small, which could support the result above. For the case of Germany,
Fischer (1990) and Edwards and Fischer (1994) report that all but the very small
firms have several “main banks,” which could be interpreted as supporting the
conclusions in this paper.
Interesting evidence is reported in Armendariz (1999). She analyzes the
performance of several development banks, i.e. the default rates of their loans.
Some of these banks require that projects are co-financed by commercial banks,
while others usually are the unique providers of capital. The former enjoy
considerably less arrears in the repayment of their loans. Her interpretation of these
facts is that the requirement of co-financing hardens the Soft Budget Constraint of
development projects, exactly what the results above suggest.
190 PAUL POVEL
A similar observation can be made if firms grow: suppose that for a small firm
R̄s − J s > I s , while for a larger firm R̄l − J l < I l . Then a “main bank” could
require that a growing firm finds a second main lender, for instance by committing
to finance only a fraction of a major investment. Similarly, a bank could require
co-financing if fixed costs of rescuing a firm are higher than the net surplus s for
small firms, but lower for larger firms.
We now analyze other implications of the model. The equilibrium strategies of
all parties are unique, and therefore we can analyze the effects of varying some of
the parameters of the model.
Proposition 3. A higher expected value of the firm R̄, a lower liquidation value
r and a lower additional loan J lead to later concessions. This in turn implies
that the liquidation of a “bad” firm becomes more likely.
The intuition behind Proposition 3 is clear: if the prize is increased, and the
expected costs of fighting remain unchanged, the banks have an incentive to fight
longer. The implications for rescue negotiations are surprising, however. Of two
otherwise identical candidates for a rescue, the one with a higher post-rescue
return R̄, i.e. the more profitable, is more likely to be liquidated. Similarly, the one
with a lower liquidation value is more likely to be liquidated. This seems to be
counterintuitive, as usually we would expect a valuable rescue to be undertaken.
The result follows from two modeling assumptions. First, the negotiations are
inefficient, as the “cake” that is to be split can disappear at any time. Second,
the banks’ valuations for the surplus from a rescue and for the rescue itself are
independent. Suppose that s depends on the number of employees of the firm, and
that the banks’ public relations suffer if they cause unemployment by not assisting
a distressed debtor (they lose i ). In this case we would expect a bank to be more
willing to rescue if the firm is larger.
A “valuable” firm could therefore be rescued for different reasons, either because
a rescue is profitable (large s), or because failing to rescue would cause indirect
costs (large i ). The second reason is an incentive problem that is similar to the one
underlying our assumption: once a project has been financed, its investors have
too strong incentives to refinance it (see Mitchell, 1993, or Aghion et al., 1999, on
the problems that this can cause for banking regulation).
The result should hold, however, in situations in which the valuations i
are small, compared with the surplus from a rescue, s. One could analyze the
refinancing decisions of foreign banks, that care less about their public image
outside their home country. Similarly, one could analyze these decisions in sectors,
regions, or during time periods, in which unemployment and bankruptcies are not
considered as being major problems.
A further implication of Proposition 3 concerns the allocation of the assets of a

distressed firm. Many bankruptcy procedures are court-led, and contain rules that
are meant to protect the interests of all parties. This may make it difficult to use
the assets in the most efficient way, as for instance their quick sale to the highest
bidder. The liquidation value of a firm is therefore lower than necessary if a formal
procedure is started, with the consequence that a rescue becomes less likely.
The variables R and I (the return of a “good” project and the initial investment)
have no effect on the strategies, because of the simplified structure of the model.
As was suggested above, we could allow a bank’s valuation i to depend on the
size of its stake in the firm. The larger the loan, the more the bank is exposed to
public scrutiny, and the more it will therefore be willing to cover up “mistakes”
by rescuing the firm.
Similarly, the relative shares r i in the liquidation value r play no role. The reason
for this is that the bank receives a payment of at least r i whatever the outcome of
the negotiations. We could easily change the sharing rule such that r i plays a role
in the banks’ renegotiation strategies. For instance, a sharing rule could require
that the bank that gives in receives a share ␣ < 1/2 of the surplus.
Next, consider a variation in the public discovery technology, the density
function f. Suppose that ␲ remains constant, and that f is changed to f 1 such
that the hazard rate is higher (the term f/(1 − F) on the RHS of Eq. (7)). Assume
that this makes the second discovery technology is superior, i.e. it becomes more
skewed to the left. The RHS of Eq. (7) becomes more negative, and in order to
restore the equilibrium L 2 must become steeper and L 1 must decrease.
Proposition 4. Assume that early discovery becomes more likely, such that the
hazard rate of the discovery technology f/(1 − F) increases. Then the banks
tend to give in earlier.
Rescue negotiations can become more difficult to hide, if the disclosure
requirements for banks or firms are tightened. The introduction of a new business
paper in a region can have a similar effect. The effect of a change in the discovery
technology by varying ␲ is similar: an increase in ␲ leads to a reduced stopping time
for all types (see Lemma 1). Unfortunately the effect on the likelihood of liquidation
is not easy to specify for the general case, as two effects are opposed: the banks
stop earlier but discovery becomes more likely. This would be interesting, as one
could derive implications for disclosure rules of stock markets, or for the benefits
of having a more transparent economy. Consider the following change, however:
Proposition 5. Suppose that the support of f is rescheduled such that f 1 (t) =
f(␣ · t), where ␣ < 1. Then the banks tend to stop earlier, but the probability of
liquidation is unchanged.
192 PAUL POVEL
Suppose that the speed of all information channels is increased symmetrically.

In this case the moment of sure discovery ␶ has an effect on the stopping time
of a bank with cost i = 0, but not on the relative stopping times of the other
types (as it does not appear in the derivations). In this case, the improvement of the
discovery technology had no material effect. Thus, stricter disclosure requirements
can be neutral, and therefore (depending on the parameters) welfare reducing or
improving.
Similarly, we can analyze changes in the distribution of types. Here the “hazard
rate” is somewhat complicated, as the types are revealed “backwards,” i.e. the
first types that reveal themselves by stopping are those with high costs i . The
“reversed” hazard rate is thus g()/G(). We encounter the same difficulties as
in Proposition 4, as we can determine (using the equilibrium conditions (5) and
(6)) the effect on the banks’ strategies, but not the effect on the probability of
liquidation.
Proposition 6. Assume that the probability of being low is higher, such that
the “reverse” hazard rate of the type distribution g/G increases. Then the banks
tend to give in earlier.
This seems to be a surprising result, as one would expect “tougher” banks to
hold out longer. However, the result states that a bank with type will stop earlier.
This is intuitive, as it must be more pessimistic about its strength relative to other
types. The overall effect cannot be determined without making assumptions on the
functional forms of f and g.
Negotiation costs can easily be introduced to the model. They have been omitted
for simplicity, but can be expected to have an effect on rescue negotiations.
Examples for such costs are the need to set up a management team which analyzes
the firm’s state and the rescue plans (i.e. the opportunity costs of sending bank
managers to attend negotiations), legal costs (the costs of hiring lawyers), or the
material costs of planning and negotiating (expenses for business consultants and
industry experts, travel expenses).
Proposition 7. Assume that each bank incurs a continuous cost c per unit of
time dt, while the negotiations take place. Then the banks tend to stop earlier
than in the case of no costs, and rescues are more likely.
Even though this type of bargaining costs reduces the net surplus from a rescue,
this material loss has no effect on the banks’ decisions. At each instant, the
past costs are sunk, and “bygones are bygones.” However, c has an effect on
the decision whether to wait another infinitesimal amount of time. It decreases
the expected payoff from waiting, and therefore the banks stop earlier with
higher costs. Thus, while the already incurred costs have no effect, the costs
that have to be incurred if the negotiations continue are relevant for the decision
to stop.
Finally, the entrepreneurs’ utility functions are relevant. As m, the utility loss
that an entrepreneur incurs if his project is liquidated, increases, funds become
available for more parameter settings. Thus, there is a use in this model for the
stigma that is attached to a business failure. While we do not want to suggest
that this is a good way of solving incentive problems, we can conclude from the
model that the financing patterns of two regions or industries should be different
if bankruptcy is “not a big deal” in one of them, while it has strong negative
connotations in the other.
6. CONCLUSIONS
This paper studies the difference between single and multiple banking. It
concentrates on renegotiation problems, which are shown to be solved better in the
case of multiple banking. We assume that entrepreneurs ask banks for loans, such
that they can start projects. These may be of a “good” or “bad” type, where the type
of a project can be observed by the respective entrepreneur, only. “Bad” projects
need refinancing at an intermediate stage, which makes them nonprofitable from
an ex ante perspective. However, once the initial loan is lost, refinancing is better
than the only alternative, liquidation.
A single bank cannot commit not to refinance a bad project, which would
keep entrepreneurs with “bad” projects from applying for a loan. Two banks,
however, can commit not to refinance with some probability. The reason for
this are inefficiencies in the negotiations between the banks, when they have
to agree on their respective degree of debt forgiveness. If the probability of
liquidation is sufficiently high, entrepreneurs with “bad” projects do not ask for a
loan at all.
We model the negotiations as a war of attrition. Each of the two banks incurs
a privately known loss, if the firm is liquidated, and therefore would like to have
it refinanced. Additionally, refinancing is profitable, once the initial loan is sunk.
The banks have to agree on how to split the costs and revenues, if they refinance
the firm. These negotiations take time, and the longer they last, the more likely it
becomes that a rescue becomes impossible (for exogenous reasons). In order to
prevent this, the banks plan to “give in” after a while, i.e. to let the opponent pocket
the gain from rescuing, only to make sure that the firm is refinanced. There is a
unique equilibrium in this game: the higher the potential loss, the earlier a bank
decides to give in. The negotiations can last for a while, if both banks’ potential
losses are low, and therefore the firm is liquidated with positive probability.
194 PAUL POVEL
The model is designed to isolate the advantage of multiplicity for the lenders.
We thus abstract from many aspects which are relevant for the choice between
bilateral and multilateral finance, as well as for reorganisation procedures. One of
these is the tradeoff between single and multiple banking. Bolton and Scharfstein
(1996) analyze a case where either single or multiple lending may be optimal,
and also derive results for voting rules, as well as for the optimal use of assets as
collateral. Similarly, the effects of different bankruptcy laws need further analysis.
In the model the two banks decide to share the highest priority rank. It would be
interesting to analyze a model in which their claims have different ranks. A further
topic for future analysis is whether and how a distressed firm is rescued, if the
banks do not enjoy the highest priority rank.
NOTES
1. A simple alternative to the mass point assumption will be discussed below, see
Lemma 1.
2. A sufficient condition for uniqueness can be found by inverting ␭2 (␭1 ), and requiring
that the slope of this inverse is never equal to the slope of ␭1 (␭2 ). It is, however, difficult to
interpret:
2
g() g(G −1 [s/(2␲ + s)]) 2␲
=
∀ ∈ + .
[G()]2 [G(G −1 [s/(2␲ + s)])]2 s
REFERENCES
Abreu, D., & Gul, F. (2000). Bargaining and reputation. Econometrica, 68, 85–117.
Admati, A., & Perry, M. (1991). Joint projects without commitment. Review of Economic Studies, 58,
259–276.
Aghion, P., Bolton, P., & Fries, S. (1999). Optimal design of bank bailouts: The case of transition
economies. Journal of Institutional and Theoretical Economics, 155, 51–70.
Armendariz, B. (1999). Development banking. Journal of Development Economics, 58, 83–100.
Berglöf, E., & von Thadden, E.-L. (1995). Short-term versus long-term interests: Capital structure with
multiple investors. Quarterly Journal of Economics, 109, 1055–1084.
Bhattacharya, S., & Chiesa, G. (1995). Proprietary information, financial intermediation, and research
incentives. Journal of Financial Intermediation, 4, 328–357.
Birckhoff, G., & Rota, G.-C. (1978). Ordinary differential equations. New York: Wiley.
Bolton, P., & Scharfstein, D. (1996). Optimal debt structure with multiple creditors. Journal of Political
Economy, 104, 1–25.
Dewatripont, M., & Maskin, E. (1995). Credit and efficiency in centralized and decentralized
economies. Review of Economic Studies, 62, 541–555.
Dewatripont, M., & Tirole, J. (1994). A theory of debt and equity: Diversity of securities and
manager–shareholder congruence. Quarterly Journal of Economics, 109, 1028–1054.
Diamond, D. (1984). Financial intermediation and delegated monitoring. Review of Economic Studies,
51, 393–414.
Diamond, D. (1993). Seniority and maturity of debt contracts. Journal of Financial Economics, 33,
341–368.
Edwards, J., & Fischer, K. (1994). Banks, finance and investments in Germany. Cambridge: Cambridge
University Press.
Fernandez, R., & Glazer, J. (1991). Striking for a bargain between two completely informed agents.
American Economic Review, 81, 240–252.
Fischer, K. (1990). Hausbankbeziehungen als Instrument der Bindung zwischen Banken und
Unternehmen. Unpublished doctoral dissertation, University of Bonn.
Gertner, R., & Scharfstein, D. (1991). A theory of workouts and the effects of reorganization law.
Gilson, S., John, K., & Lang, L. (1990). Troubled debt restructurings. An empirical study of private
reorganization of firms in default. Journal of Financial Economics, 27, 315–353.
Hellwig, M. (1991). Banking, financial intermediation and corporate finance. In: A. Giovannini &
C. Mayer (Eds), European Financial Integration (pp. 35–63). Cambridge: Cambridge
University Press.
Hoshi, T., Kashyap, A., & Scharfstein, D. (1990). The role of banks in reducing the costs of financial
distress in Japan. Journal of Financial Economics, 27, 67–88.
Mitchell, J. (1993). Creditor passivity and bankruptcy: Implications for economic reform. In: C. Mayer
& X. Vives (Eds), Capital Markets and Financial Intermediation (pp. 197–224). Cambridge:
Cambridge University Press.
Rajan, R., & Winton, A. (1995). Covenants and collateral as incentives to monitor. Journal of Finance,
50, 1113–1146.
Repullo, R., & Suarez, J. (1995). Monitoring, liquidation, and security design. Review of Financial
Studies, 11, 163–187.
Schneider-Lenné, E. (1992). Corporate control in Germany. Oxford Review of Economic Policy, 8,
11–23.
von Thadden, E.-L. (1992). The commitment of finance, duplicated monitoring, and the investment
horizon. Working Paper, University of Basel.
von Thadden, E.-L. (1995). Long-term contracts, short-term investment and monitoring. Review of
Economic Studies, 62, 557–575.
Yosha, O. (1995). Information disclosure costs and the choice of financing source. Journal of Financial
Intermediation, 4, 3–20.
196 PAUL POVEL
APPENDIX A: PROOFS
Proof of Lemma 1
(A.1) compares the respective payoffs for bank B 1 with valuation ␭1 , given ␭2 :

R̄ − J − r
G(␭2 )r 1 + (1 − G(␭2 )) + r1
2

= (1 − G(␭2 ))(R̄ − r 2 − J) + G(␭2 ) (1 − ␲)(R̄ − r 2 − J) + ␲(r 1 − ␭1 ) .
(A.1)
The left-hand side of Eq. (A.1) is the expected payoff if bank B 1 gives in
immediately. With probability G(␭2 ) the opponent has a low valuation and does not
give in. The firm is rescued, and the bank receives r 1 . With probability 1 − G(␭2 )
the opponent gives in, as well, and the net surplus is shared (in expected terms).
The right-hand side of (A.1) is the payoff if the bank gives in as soon as the war of
attrition has started. With probability 1 − G(␭2 ) the opponent has a high valuation
and will give in immediately. The bank rescues, pockets the surplus R̄ − J, and
pays r 2 to the opponent. With probability G(␭2 ) the war of attrition starts. It
is discovered with probability ␲, and the firm is liquidated. With probability
(1 − ␲), the game could continue, but by definition the bank plans to stop,
which earns r 1 .
Some simplifications of (A.1) and of an analogous equation for bank B 2 lead to
the two equations in Lemma 1. There is always an interior solution for the cut-off
levels: If ␭i goes to zero, then ␭j (␭i ) goes to infinity, while if ␭i goes to infinity it
goes to zero.
Proof of Lemma 2
(i) We first show that T i is nonincreasing, and then that it is strictly decreasing. By
utility-maximisation it must be the case that
V1 (t 1 , T 2 (·), 1 ) ≥ V1 (t 1 , T 2 (·), 1 ) ∀t 1 , ∀t 1 = T 1 (1 ) (A.2)
and
V1 (t 1 , T 2 (·), 1 ) ≥ V1 (t 1 , T 2 (·), 1 ) ∀t 1 , ∀t 1 = T 1 ( 1 ), (A.3)

where Vi (t i , T j (·), i ) is the expected payoff of bank B i with cost level i , if it stops
at t i , and bank B j plays strategy T j (·):
Vi (t i , T j (·), i ) = Pr{T j (j ) ≥ t i } (F(t i )(r i − i ) + (1 − F(t i ))r i )

+ F(T j (j )) r i − i − (R̄ − r j − J)
{j |T j (j )<t i }

+ (R̄ − r j − J) g(j ) dj . (A.4)
The payoff of a bank depends on the chosen stopping time t i , the opponent’s
strategy T j and the (privately known) loss i of losing the firm. With probability
Pr{T j (j ) ≥ t i } the opponent plans to stop later than t i . If the public discovered
the negotiations (this happens with probability F(t i )), the payoff is (r i − i ). If the
secret was kept well, the bank receives r i from bank B j who rescues the firm. The
second term of Eq. (A.4)is the equivalent if the opponent plans to stop earlier. Here
the bank receives (R̄ − J) if the firm can be rescued and pays r j to the opponent.
We can rewrite the two inequalities Eqs (A.2) and (A.3) using Eq. (A.4).
Subtracting the RHS of Eq. (A.3) from the LHS of Eq. (A.2), and the LHS of
Eq. (A.3) from the RHS of Eq. (A.2), we get

Pr{T 2 (2 ) ≥ t 1 }F(t 1 )( 1 − 1 ) + F(T 2 (2 ))( 1 − 1 )g(2 ) d2
{2 |T 2 (2 )<t 1 }

≥ Pr{T 2 (2 ) ≥ t 1 }F(t 1 )( 1 − 1 ) + F(T 2 (2 ))( 1 − 1 )g(2 ) d2
{2 |T 2 (2 )<t 1 }
or, rearranging,
[(1 − Pr{T 2 (2 ) < t 1 })(F(t 1 ) − F(t 1 ))]( 1 − 1 )

≥ F(t 1 )g(2 ) d2 − F(T 2 (2 ))g(2 ) d2
{2 |t 1 <T 2 (2 )<t 1 } {2 |t 1 <T 2 (2 )<t 1 }
× ( 1 − 1 ).
If t 1 > t 1 , the following holds:
(1 − Pr{T 2 (2 ) < t 1 })(F(t 1 ) − F(t 1 ))

≥0≥ [F(t 1 ) − F(T 2 (2 ))]g(2 ) d2 ,
{2 |t 1 <T 2 (2 )<t 1 }
198 PAUL POVEL
and it must be the case that 1 ≥ 1 . On the other hand, if t 1 > t 1 :

(1 − Pr{T 2 (2 ) < t 1 })(F(t 1 ) − F(t 1 ))
≥ Pr{t 1 < T 2 (2 ) < t 1 }(F(t 1 ) − F(t 1 ))

≥ [F(t 1 ) − F(T 2 (2 ))]g(2 ) d2 ,
{2 |t 1 <T 2 (2 )<t 1 }
and it must be the case that 1 ≥ 1 . Thus for all t 1 , t 1 , in equilibrium ( 1 − 1 ) ·

(t 1 − t 1 ) ≤ 0, i.e. the strategies are nonincreasing in the liquidation loss.
Assume that T 1 is not strictly decreasing, i.e. there are a , b > a , such that
for all ∈ [a , b ], T 1 () = ␪. Then there is an ␧ > 0 such that all types 2 with
T 2 (2 ) ∈ (␪ − ␧, ␪] prefer to wait until ␪ and stop then, if the opponent did not
stop. Then the types ∈ [a , b ] could gain by stopping at ␪ − ␧ instead of ␪: The
probability of winning is not affected, but the risk of losing is diminished.
(ii) Assume that T 1 is discontinuous at . Then there are t a , t b > t a such that
a type never stops at any t ∈ (t 1 , t 2 ). A type 2 with T 2 (2 ) ∈ [t a , t b ) would
thus wait only until t a , and stop if the opponent did not stop. This implies that no
one stops at any t ∈ (t a , t b ). But then there are types ˆ 1 and an ␧ > 0 such that
T 1 (ˆ 1 ) ∈ [t b , t b + ␧], who prefer stopping at some t ∈ (t a , t b ).
(iii) Assume that T i is not differentiable at .
(a) Let the left-hand derivative be higher than the right-hand derivative (T i is
flatter to the left of ). Then there is an ␧ > 0 such that no type j stops
at any t ∈ (T i () − ␧, T i ()]. It pays to wait longer since after the point of
discontinuity it becomes relatively likely that the opponent stops. This holds
since both f and g are continuous and differentiable.
(b) Let the right-hand derivative be higher. Then there is an ␧ > 0 such that no
type j stops at t ∈ [T i (), T i () + ␧). It pays to stop earlier since it becomes
more likely that B i stops immediately after T i ().
In both cases T j is not continuous, contradicting (ii) above.
(iv) If type 1 = 0 stops at ␪ < ␶, all types with higher loss level stop earlier
because of Lemma 2(i). Then in equilibrium no type of the other bank should stop
later than ␪. There is an ␧ > 0 such that types 1 ∈ (0, ␧] find it profitable to wait
until ␪ and wait for the opponent to stop.
Proof of Proposition 1
The differential equations are Lipschitz-continuous on [2, ␶] which implies that a

solution exists and is unique (see e.g. Birckhoff & Rota, 1978, Chap. 6). At each
t ∈ [2, ␶], the (expected) payoffs from stopping or non-stopping can be compared,
as was done in deriving Eq. (5). Since the strategies are strictly decreasing, at
t < T 1 (1 ), i.e. if L 1 (t) > 1 , the payoff to bank B i with loss level 1 will be
higher if it waits. The opposite holds for L 1 (t) < 1 . For all t ≥ t 1 , type L 1 (t 1 ) can
only decrease his payoff by waiting, and will stop whenever possible.
The players constantly update their beliefs using Bayes’ Rule. If a player stops
at the wrong time (this is the only deviation that is possible) the opponent will
have no difficulties in updating his beliefs: If a player stops too early, the game is
over and beliefs are not relevant anymore. If a player waits too long, the strategy
tells him to stop immediately: Type i stops at any time t if t > T i (i ). Again, the
opponent can update his beliefs without problems.
Follows directly from the Assumptions and Proposition 1.
The reaction curves L i (see Eq. (7)) are shifted outward, if s is increased. The
indirect effect via the cut-off value ␭ goes in the same direction: ␭i (␭j ) is shifted
outward, as well (see Lemma 1).
Proof of Propositions 4, 5 and 6
As Proposition 3: analyze the equilibrium conditions Eqs (5) and (6), and the
indirect effect via the cutoff value ␭ in Lemma 1.
Equation (7) is changed to

t1
−1 f(t) L 1 (t) c
L 2 (t 1 ) = G G(␭)exp − + dt . (A.5)
2 1 − F(t) s s
A comparison of Eqs (A.5) with (7) shows that all types will want to stop earlier,
including zero-cost types.
OPPORTUNITY COST AND
PRUDENTIALITY: AN ANALYSIS
OF COLLATERAL DECISIONS
IN BILATERAL AND
MULTILATERAL SETTINGS
Herbert L. Baer, Virginia G. France and James T. Moser
ABSTRACT
This paper develops a model that explains how the creation of a futures
clearinghouse allows traders to reduce default and economize on margin.
We contrast the collateral necessary between bilateral partners with that
required when multilateral netting occurs. Optimal margin levels balance the
deadweight costs of default against the opportunity costs of holding additional
margin. Once created, it may be optimal for the clearinghouse to monitor
the financial condition of its members. If undertaken, monitoring will reduce
the amount of margin required but need not affect the probability of default.
Once created, it becomes optimal for the clearinghouse membership to expel
defaulting members. This reduces the probability of default. Our empirical
tests suggest that the opportunity cost of margin plays an important role in
clearinghouse behavior particularly their determination of margin amounts.
The relationship between volatility and margins suggests that participants
face an upward-sloping opportunity cost of margin. This appears to dominate
the effects that monitoring and expulsion might have on margin setting.
Research in Finance
© 2004 Published by Elsevier Ltd.
ISSN: 0196-3821/doi:10.1016/S0196-3821(04)21009-7
201
202 HERBERT L. BAER ET AL.
The field of competition . . . consists of all the individuals who are willing and able to recontract
about the articles under consideration.
F. Y. Edgeworth
1. INTRODUCTION AND LITERATURE REVIEW
Contract terms that specify posting of collateral effect a limit on the exposure to loss
from counterparty nonperformance. This outcome is obtained in two ways. First,
seizure of deposited collateral inherently reduces losses resulting from contract
nonperformance. Second, posting of collateral encourages contract performance.1
This paper studies decision making for instances where collateral requirements are
costly. We show that cost-avoidance within a competitive market reliably predicts
certain outcomes.
1.1. Introduction
Brennan (1986) employs the Theory of Efficient Contract Design to model

the price limits employed by futures exchanges. The theory explicitly
recognizes the influence of cost-minimization incentives for the contract
terms developed by exchanges. In that setting, he shows that price limits
lessen contract nonperformance. Improvements in contract performance enable
reductions in required collateral. Those reductions, in turn, lower contracting
costs.
This paper applies the Theory of Efficient Contract Design first to model
the determination of collateral required from contracting counterparties. We
then extend this result to highlight how reducing credit risk affects contract
terms. Futures exchanges specialize in providing facilities for designing and
exchanging contracts. Their interest in successful trading of contracts makes
these organizations a logical source for empirical examination of the hypotheses
suggested by our model.
Our empirical analysis discriminates among competing representations of an
exchange’s margin decision by examining margin coverage ratios – required
margin divided by a forward-looking measure of volatility. In the first, the marginal
opportunity cost of margin requirements is constant. The second allows for
increasing costs for margin funds.
Both time series and cross section evidence is developed. Our model predicts
margin adjustments when margin coverage is too high (and too expensive) or too
Opportunity Cost and Prudentiality: An analysis of Collateral Decisions 203
low (and too risky).2 Time series evidence concludes that coverage ratios increase
(decrease) when coverage ratios are lower (higher) than their unconditional means.
Examining a cross section of margin-coverage ratios, we find that the opportunity
cost of margin deposits significantly influences the level of magin required from
exchange members. Our regressions suggest a negative relationship between
economy-wide shifts in the opportunity cost of margin deposits and levels of margin
coverage. In addition, we find a negative relationship between margin coverage
and participant-specific shifts in participants’ borrowing needs as proxied by levels
of implied standard deviation. These suggest that opportunity costs are important
and that market participants face upward-sloping schedules of opportunity costs
for their margin deposits.
1.2. Literature Review
We advance the theory of margins by explicitly incorporating the cost of margin

deposits into the margin-setting decision. This establishes a tradeoff between these
costs and prudential concerns. Further examination of this tradeoff gives insight
into the effect of clearinghouse activities on its margin-setting decisions.
Craine (1997) models the clearinghouse as a profit-maximizing entity and
describes margin in terms of options to default. He contends that, since the
clearinghouse charges no default premia, it must keep premium values at or close
to zero. Our model, by contrast, argues that the values of default premia and
counterparty risk offset across agents. Fenn and Kupiec (1993) also implicitly
assume that the clearinghouse is an independent cost-minimizing entity. Their
clearinghouse minimizes the sum of margin costs, settlement costs, and costs
incurred when deficits arise in clearinghouse accounts. The clearinghouse sets
the probability of a deficit equal to the ratio of opportunity costs per settlement
period to the marginal cost of an account deficit. As volatility increases, increasing
settlement frequency lowers this sum, and the margin-to-volatility ratio declines.
We model the clearinghouse as a club that minimizes the joint costs of
its members. In our formulation, the clearinghouse need not make a profit,
nor even recoup any deadweight losses incurred by its membership because
members willingly subsidize the clearinghouse to avoid the greater cost of a
bilateral arrangement.3 Neither Craine nor Fenn and Kupiec provide motivation
for clearinghouse development.
Our model results for expulsion and the value added by clearing activities match
those of Bhasin and Brown (1997). They model the value of exchange seats as
stemming from trading activity. Their model analyzes intra day default incentives.
They show that the values of exchange seats secure under-margined positions held
during the day. Their model complements ours by explaining the dynamics of
collateralizing against default during the trading day.
We develop the monitoring activities of exchanges as part of their intermediating
role. In this respect, the model resembles the delegated monitoring described by
Diamond (1984), and the modeling of risk management and financial guarantees
by Merton and Bodie (1992) and Hsieh (1993). We extend the earlier work by
explicitly incorporating certain institutional features of clearinghouses. These
features include expulsion from the clearinghouse, clearinghouse monitoring of
members’ financial condition and the possibility that members face increasing
costs for external funds. These have very different effects on optimal margin
setting and the probability of default. Our model also parallels Gorton’s (1985)
modeling of bank clearinghouses, particularly concerning expulsion and risk
mutualization.
2. DETERMINING COLLATERAL FOR

BILATERAL CONTRACTS
We first model the setting of collateral requirements in a bilateral marketplace.
Two parties j and h negotiate contracts for their own accounts but are unable
to compel contract performance. Contract nonperformance, that is, failure to
fulfill contractual terms always entails deadweight loss.4 These deadweight losses
include costs of recontracting, higher borrowing costs that arise from liquidity
problems, and costs arising from financial distress. Recovery against losses is
limited to collateral deposited by the defaulting counterparty. We assume that a
default-free trustee holds collateral deposits. Counterparties posting collateral with
the trustee bond their contract performance.
There are two periods. In the opening period, the two parties enter a contract
with each other. The motivation for trading is exogenous to our model; however,
our model does imply that cost reductions increase whatever benefits trading
provides. Let N(j, h) denote the number of contracts outstanding between j and h.
If N(j, h) is positive, j holds a long position in the contract. If N(j, h) is negative,
then j’s position in the contract is short. Contra-positions are held by h, so that
N(h, j) = −N(j, h). The contract settles at the contract’s next-period price or, at
contract expiration at the value of the underlying good. The distribution of this
price has a zero mean and finite standard deviation of ␴.5
Collateral posted by j with h is denoted M(j, h), and that posted by h with j is
denoted M(h, j). Collateral are cash deposits held in interest-bearing accounts.
Interest on deposits is paid to its respective depositor.6 At the end of period
two, the contract is settled. If x is positive and less than M(h, j), x is transferred
from the account of the short to the account of the long. Thus the short now has
M(h, j) – x; the long now has M(j, h) + x. If x is negative and |x| is less than
M(j, h) then x is transferred from the long to the short.
After contracts are settled, traders immediately restore their collateral-account
balances to M(j, h) and M(h, j), by either depositing cash when they are on the
losing side or withdrawing excess balances on gains. These account adjustments
restore equal values for the default options and counterparty risk of both
counterparties.
Figlewski (1984) shows that contract counterparties implicitly give each other
options to default. In the simplest case, contract default occurs whenever losses
exceed margin-account balances. Thus, if x is positive and greater than M(h, j),
the short rationally defaults on the contract and the long takes possession of the
margin assets M(h, j). Similarly, if x is negative and |x| is greater than M(j, h), the
long rationally defaults and the short takes possession of the margin assets M(j, h).
Remaining losses include recontracting costs, higher borrowing costs arising from
liquidity problems, and costs of financial distress. These are deadweight losses.
Agent j’s expectation of these losses is:
∞
D(j, h) = ␣N (x − M(h, j)f(x, ␴) dx) (1)
M(h,j)
where N is the net number of contracts j has open with h; i.e. the absolute value
after summing across N(j, h).
Agents jointly minimize the cost of contracting as would be obtained in a
perfectly competitive market. This is realistic provided agents freely choose among
a large number of counterparties each willing to minimize joint contracting costs.
Post-trade bargaining problems are included in the cost-recovery factor subsumed
in ␣.
Proposition I. Collateral amounts are optimal when the default probability

equals the ratio of the opportunity cost for posting additional collateral to the
deadweight loss rate.
Contracting entails three costs: the total opportunity cost of margin deposits
I(j); counterparty risk, that is, the expected difference between the promised
and the actual payment when h defaults on j, L(j, h); and expected deadweight
losses incurred when h defaults on j, D(j, h). Offsetting these costs, each party
also receives an option to default O(j, h). The two parties jointly minimize the
following costs:
I(j) + I(h) Opportunity Costs
D(j, h) + D(h, j) Deadweight Losses
(2)
L(j, h) + L(h, j) Counterparty Risk
O(j, h) + O(h, j) Default Options
Default involves both a loss to one party and a corresponding gain to the other.
Alternately stated, one party’s default option is another party’s counterparty
risk, that is, that L(j, h) = O(h, j). In our representative-agent framework, the
expression for joint contracting costs reduces to:
I(j) + I(h) + D(j, h) + D(h, j) (3)
The first order conditions for minimization of (3) with respect to M(j, h) and
M(h, j) are:
i i
[1 − F(M ∗ (j, h), ␴)] = and F(M ∗ (j, h), ␴) = (4)
␣ ␣
where i is the opportunity cost for an additional dollar of margin. Equation (4)
implies that counterparties optimally collateralize when their default probabilities
equal their ratios of opportunity cost of additional margin to their deadweight
loss rate. The higher this ratio, the lower is the optimal collateral level. Nonzero
collateral requirements are optimal when i/␣ < 1. Should i/␣ exceed unity,
counterparties set margin at zero, the losing trader always defaults, and contracts
are unenforceable.
Since the objective function is linear in the number of contracts, the opportunity
cost of additional margin is constant. This implies that the collateral per unit of
exposure is independent of the aggregate level of exposure, and that collateral
amounts can be set on a per-contract basis. Further, if the distribution of price
changes is symmetric, equal collateral amounts are required for both long and
short positions.
Proposition II. when the distribution of price changes is uniquely invertible
and the opportunity cost of collateral deposits is constant, optimal coverage
ratios do not vary with volatility.
When a unique inverse exists for the price-change distribution depicted in
Eq. (4), the coverage ratio giving the level of collateral to exposure to price
changes is:

M∗ −1 i
=F ,1 (5)
␴ ␣
Inspection of (5) confirms that margin increases proportionately with ␴.
This implies that optimal coverage ratios do not vary with volatility. For instance,
Proposition II holds if the distribution of price changes is normal.
3. DETERMINING COLLATERAL FOR A

MULTILATERAL CLEARINGHOUSE
This section models a clearinghouse acting solely as a netting facility.7 We
establish the benefits from clearinghouse arrangements within a framework that
regards the clearinghouse as a club of its members, not as a separate, for-profit
agency. We then model the margin-level choices made by clearinghouses as
balancing the deadweight losses from counterparty defaults against opportunity
costs incurred by posting margin. Although depositing interest-bearing assets
can fulfill margin requirements, we argue that opportunity costs are nonzero
when the firm’s marginal borrowing cost exceeds the return on its marginable
assets.8
3.1. The Intermediary Role of a Clearinghouse
Clearinghouses intermediate by substituting themselves as contract counterparties.

This achieves certain economies by reducing both deadweight default costs and
the opportunity costs of holding assets in margin accounts. We represent the
clearinghouse acts as a club, that is, a voluntary organization that furthers the joint
interests of its members by internalizing certain shared costs. This approach is more
general than the standard profit-maximizing framework in that the model can be
modified to represent the heterogeneous interests generally found within exchange
organizations. Clearinghouse members minimize their joint contracting costs
by committing to rules that allocate contract-default losses among themselves.
Many loss-sharing rules are consistent with this objective function. Importantly,
the model is consistent with the industry practice of paying for losses from a
clearinghouse guarantee fund, in effect sharing losses pro rata among clearing
members.9
Proposition III. If members have the same i and ␣, the clearinghouse sets
margin at the same level as though contracts were cleared and settled bilaterally.

Let party j’s open interest nh=1 N(j, h) be denoted by N(j). If we assume that
f, the distribution of price changes, is symmetric then the clearinghouse chooses
M(j, h) to minimize joint contracting costs of:

n ∞
|N(j)| iM(j) + ␣ (x − M(j))f(x, ␴) dx (6)
j=1 M(j)
In our representative framework, i and ␣ are the same for all members and the
solution to this problem is given by Eq. (4). Thus, per contract, margin will be
the same whether contracts are cleared and settled bilaterally by counterparty
pairs or multilaterally through a clearinghouse. We argue this representative-agent
framework serves as a useful starting point in the following sense. Exchanges
screen members for their financial ability to fulfill contracts. That effort tends to
reduce member heterogeneity in their credit risk dimensions. Later sections of the
paper consider the monitoring efforts undertaken by exchanges as they attempt to
retain low levels of heterogeneity.
Because a clearinghouse will set the same margin rate that these agents willingly
negotiate between themselves, analyzing the benefits derived from forming a
clearinghouse is straightforward. The essential benefit of the clearinghouse is that
it permits its members to economize on margin while also reducing expected
deadweight losses. Clearinghouses economize on margins and deadweight loss
because, for the same set of contracts, participants’ net positions are less risky.
Consequently, total margin deposits required by the clearinghouse are smaller
than totals required for a comparable set of bilateral transactions. In addition, pro
rata expected deadweight losses are also smaller.
Proposition IV. Total margin deposits posted by each member will be the same
or lower under a clearinghouse system than under a system of bilateral collateral
deposits.
Under a clearinghouse system, j posts margin against the net of his position
with the rest of the market, that is M |N(j)|. In effect, a multilateral clearinghouse
secures the losing positions of a potential defaulter with its winning positions. That
is, members are prevented from “cherry picking” among their contracts realizing
gains while defaulting on those contracts where losses have been realized. The
margin posted by each member will be the same or lower under a clearinghouse
system.
Proposition V. Under an appropriate loss-sharing arrangement, total expected
deadweight losses are lower under a clearinghouse system than under a system
of bilateral collateral deposits.
Similarly, no counterparty’s expected deadweight loss is greater under a
clearinghouse system and for some expected deadweight losses will be smaller.10
In a bilateral system, j’s expected loss from counterparty

default is proportional to
the number of their open contracts; that is nh=1 |N(j, h)|. Member j benefits by
joining a multilateral clearinghouse that keeps default-loss exposure proportional
to net open contracts; that is, proportional to |N(j)|. Other loss-sharing arrangements
can also achieve this result.
3.2. Increasing Opportunity Cost of Funds
The cost of funds function may be increasing as the amount of required margin
increases. Thus, an increase in margins drives up the marginal cost of funds.
Proposition VI. Where the marginal cost of funds is increasing, the optimal
coverage ratios decrease with volatility when prices are normally distributed
and the opportunity cost of margin assets is increasing.
If marginal costs of margin are increasing in M then:
i = ␳(M); ␳ (M) > 0 (7)
and the clearinghouse sets margin to meet the condition:

␳ (M)
= [1 − F(M, ␴)] (8)
␣
An increase in ␴ now causes the clearinghouse to increase margin less than
proportionately with ␴. As the standard deviation increases, the clearinghouse
increases margin levels to keep the probability of default constant. However,
doing so drives up the members’ marginal financing costs. The members of the
clearinghouse therefore choose to bear greater deadweight losses to economize
on their financing costs. Thus, coverage ratios should decrease with volatility.
Note that, even if their cost functions are identical, individuals holding different
numbers of contracts may have different marginal costs of funds. In addition, unlike
the agents of the previous section, the slope and level of member cost functions
may differ. This will result in disagreement among members as to appropriate
margin levels, though each will have only one preferred margin level. The club
literature suggests that in decisions made by diverse interests, majority rule reflects
median voter preferences provided individuals have single-peaked preferences.11
For such preference structures, relevant marginal costs are those of median voting
members. Severe disagreement about appropriate margin levels will lead some
traders to avoid becoming a member of the clearinghouse, of these some may seek
economies through other arrangements.
4. TESTS OF THE MODEL

The modeling of the previous sections suggests three hypotheses for clearinghouse
determination of required margin. The first of these is a positive relationship
between margin levels and risks stemming from the contracts. Baer, France
and Moser (1995b) provide supporting evidence for this result. The second
is that clearinghouses incorporate the cost of maintaining margin balances
into their margin-level decisions. Thus, we can expect a negative correlation
between margin levels and the opportunity costs of clearinghouse members.
Our remaining hypothesis relates to coverage ratios, predicting that coverage
ratios are invariant to risk levels when members have constant costs while
increasing costs imply a negative correlation between coverage ratio and risk
level.
4.1. Description of the Sample Data Set
Margin data are from the clearing organizations for eighteen contracts trading
on the following futures exchanges: the Chicago Board of Trade, the Chicago
Mercantile Exchange, the Coffee, Sugar and Cocoa Exchange, the Commodity
Exchange, and the New York Mercantile Exchange. These eighteen contracts are
the most heavily traded contracts having options on the underlying futures contract.
During the sample period, with the exception of contracts listed by the New
York Mercantile Exchange, exchange affiliation was the basis for determining
margin requirements. The speculative positions of non-clearing members are
assessed the highest levels of margin.12 The initial margin requirement for clearing
members is usually the same as the initial margin amount for the hedge positions of
non-clearing members. Finally, the maintenance margin requirements of clearing
members are the same as their initial requirements. Thus, our assumption that
periodic settlement restores the account to the level M gives a lower bound for a
clearing member’s margin account. Members must always have at least the amount
of the current initial margin, and may choose to allow excess balances to remain
in the account.
Table 1 summarizes our sample. Listed under each exchange are the contracts
trading on that exchange. The start date is the first date used in the sample; generally,
this is the beginning of options trading on the respective futures contracts. In each
case, the sample extends through June 1991. Sample dates are the last Thursday of
every contract month. The number of available observations ranges from 29 for the
Treasury bond and Deutschemark contracts to 15 for the Heating Oil contract. We
report mean margin levels for positions classed as initial nonmember speculative
Table 1. Margins and Implied Volatilities.

Contract Sample Number ISD Sample Means
Start Observations
Speculative Margin Member Margin
Date
Margin Coverage Margin Coverage
Chicago board of trade

Corn 3/85 26 0.21 520.58 5.10 353.85 3.38
Soybeans 12/84 26 0.16 1396.38 5.61 1067.31 4.20
Treasury bond 3/84 29 0.11 2618.97 5.32 2120.69 4.27
Wheat 3/87 16 0.21 725.31 4.38 543.75 3.24
Chicago mercantile exchange
British Pound 3/85 26 0.12 2197.23 5.44 1938.46 5.02
Deutschemark 3/84 29 0.12 1864.17 5.45 1689.66 5.01
Eurodollar 3/85 17 0.01 925.00 7.06 823.53 6.07
Japanese Yen 6/86 21 0.10 2069.67 4.90 1788.10 4.24
Live Cattle 12/84 23 0.14 756.78 4.02 619.57 3.29
Swiss Franc 3/85 26 0.12 2111.38 4.81 1875.00 4.25
S&P 500 3/84 26 0.17 11134.62 10.17 4865.38 4.56
Coffee, sugar and cocoa exchange
Coffee 12/87 15 0.30 2733.33 5.25 1366.67 2.62
Sugar 3/85 26 0.53 1209.62 5.46 604.81 2.73
Commodity exchange
Copper 6/86 20 0.30 1734.50 4.81 1355.00 3.66
Gold 3/84 28 0.16 1692.46 5.34 1253.57 3.95
Silver 12/84 27 0.24 2004.52 5.55 1585.00 4.10
New York mercantile exchange
Crude oil 12/86 19 0.36 2284.21 7.53 2284.21 7.53
Heating oil 9/87 15 0.37 2293.33 6.79 2293.33 6.79
Note: Table reports summary statistics for a sample of futures contracts. Start date is the first sample
date. Mean margin is the average of initial speculative or initial member margin required on the
sample dates. Mean ISD is the average implied standard deviation for options trading on the
sample dates. Margin coverage is respective mean level of margin divided by the dollar volatility
of the contract.
and for clearing members (or nonmember hedgers) on the above-indicated sample
dates.
For each sample date, we impute volatilities for the respective contracts. The
needed data are from various issues of the Wall Street Journal. These data are: prices
for call options expiring in the next delivery month at each strike price traded on that
date, futures settlement prices for corresponding delivery months, and Treasury
bill rates with maturities most closely matching the time until expiration of the
option contracts. The Barone-Adesi and Whaley (1987) model was used to impute
volatilities for each of the option contracts. For each contract, we calculate a time
series of representative implied standard deviations (ISDs) on each sample date
using a Taylor-series approximation based on iterated regressions as described by
Whaley (1982). The table reports mean ISDs. These range from a low of 0.01 for
the Eurodollar contract to 0.53 for the sugar contract.13
Margin coverage ratios divide the respective margin amounts by dollar-price
volatility. Dollar-price volatility is the product of the ISD and the notional value
of the contract – futures prices times number of deliverable units – after adjusting
the annualized volatility for the length of the holding period. This gives a market-
based forecast of holding period volatility. Dividing initial speculative and member
margin requirements by their respective dollar volatilities gives coverage ratios.
Columns 6 and 8 list mean coverage ratios for the member and nonmember
categories. The peak of the frequency distribution for mean nonmember speculative
margin coverage ratios is about five. This implies that margin levels most often
cover five times the expected single-day price deviation. Comparison of the means
of nonmember speculative and member margin requirements reveals that clearing
members’ margin is about 80% of that for speculative positions. The exception is
the New York Mercantile Exchange where they are equal.
Notably, the coverage ratio for the S&P 500 contract well exceeds the typical
level obtained for nonmember speculative positions, averaging 10.17 during the
sample period. In contrast to the coverage obtained by nonmember margin levels,
the S&P 500 members’ margin, generally around four, is within the range obtained
for other contracts. The difference between coverage ratios for the S&P and other
nonmember speculative margins probably reflects considerations unique to the
sample period. Market breaks in 1987 and 1989 increased debate over the need for
regulatory intervention in the determination of stock-index margin requirements
may have resulted in higher margins than the clearinghouse would have set for
purely prudential reasons.
The contrast between margin for the S&P contract and the others is more
noticeable on recognizing that during part of this period, the S&P 500 contract
settled twice per day. Other contracts settled only once per day throughout the
period. Since we calculate coverage ratios with daily standard deviations, the
coverage ratio for the S&P 500 should be smaller, not larger. Other things
equal Fenn and Kupiec’s (1993) analysis suggests coverage ratios should be
approximately half as large.
Assuming price changes are normally distributed, the coverage ratios for
clearing members imply that the probability of a price change exceeding required
margin from one settlement period to the next is much less than 1%. The “excess”
of coverage suggests that actual distributions are kurtotic, a result that is consistent
with the findings of Kofman (1993).
4.2. Time-Series Evidence
We use daily data for four of the eighteen contracts to gain further insight into the
margin-setting process. These contracts are: Deutschemark, S&P 500, Soybean
and Treasury Bond. Daily implied volatilities, computed as previously described,
were matched with required margin levels on these dates and margin coverage
ratios computed. We examine the time series of these quantities.
The first test considers whether the coverage ratio for a contract tends to revert
to its long-run, unconditional mean. Denoting coverage ratios CRt , our model
implies that shocks to these ratios initiate intervention that restores them to desired
levels. Specifically, prudential concerns dictate that coverage ratios regarded as too
small should lead to increasing margin coverage and opportunity cost concerns
dictate that excessively large ratios lead to reducing margin coverage. Our model
equilibrates these pressures, and predicts a mean-reverting time series of coverage
ratios.
We employ the augmented Dickey-Fuller (ADF) procedure to consider this
hypothesis. Changes in coverage ratios are regressed on the first lag of their levels
and lags of changes in the coverage ratio. The specification is:
K

CRi,t = ␣i,0 + ␣i,1 CRi,t−1 + ␣i,1+j CRi,t−j + u i,t (9)
j=1
The number of lags – K – is determined by comparing Akaike’s Information

Criterion (AIC) at various lag lengths, choosing the lag length that obtains the
largest AIC value.
The test examines the coefficient on the lag level, employing Fuller’s (1976)
critical values: −1.95 at the 5% level and −2.58 at the 1% level. Table 2 reports
the results of these tests. Coefficient t statistics below these critical values suggest
mean reversion in the series. We find evidence of mean reversion at the 1% level
or better in every case.
Evidence of mean reversion in coverage ratios can be the result of mean reversion
in volatilities. Although substantial research finds evidence that the volatility of
returns on financial assets is nonstationary,14 in our sample volatility appears to
be stationary. We examine the possibility that the mean reversion in coverage
ratios is caused by mean reversion in volatility by comparing the mean half lives
from coverage-ratio shocks to the half lives for volatility shocks. From Eq. (9),
the mean half-life from a coverage-ratio shock is given by 1 − log(2)/log(␣i,1 ).
Substituting for the CR variables in (9) with the respective volatilities – these
are the denominators of the coverage ratios – gives the half-life for a shock to
volatility. Halflives are computed for two standard errors above and below the
Table 2. Estimates of Margin Coverage Adjustment.

K
CRt = ␣0 + ␣1 CRt –1 + i=1 ␣1+i CRt –i + ut
Initial Member Margin

Contract ␣1 t(␣1 )
Deutschemark−0.004579 −3.52
S&P 500 −0.004704 −2.88
Soybean −0.012160 −4.04
−0.017178 −6.84
Treasury Bond

CRt = ␣0 + 4l=1 ␣l1 Q l CRt –1 + K
i=1 ␣4+i CRt –i + u t
Quartile Rankings of Margin Coverage at Time t − 1
Contract Lowest Quartile Second Quartile Third Quartile Highest Quartile

␣1 t(␣1 ) ␣1 t(␣1 ) ␣1 t(␣1 ) ␣1 t(␣1 )
Deutschemark −0.0135 −3.18 −0.0084 −2.58 −0.0097 −3.43 −0.0090 −3.78

S&P 500 −0.0438 −4.47 −0.0417 −5.25 −0.0239 −4.44 −0.0233 −5.35
Soybean −0.0277 −2.12 −0.0265 −2.88 −0.0200 −2.91 −0.0180 −3.49
Treasury Bond −0.0408 −5.96 −0.0389 −6.48 −0.0356 −6.55 −0.0321 −6.83
Note: Table reports results for the two time series specifications listed below. CRt is the time-t ratio
of initial member margin to the option-implied volatility stated in dollars. Ql is the coverage
quartile for margin coverage during the sample period. K, the number of lagged changes in
coverage ratio included in the specification, is determined by AIC. Critical values are from
Fuller (1976): −1.95 at the 5% level and −2.58 at the 1% level. Lower values of t are indicative
of reversion to the mean; i.e. rejects the null of no mean reversion.
coefficient estimates from the volatility specification and for the coverage-ratio
specification. In no case do these ranges overlap. Hence, we reject the alternative
hypothesis that adjustments to coverage ratios stem from volatility reverting to its
long-run mean rather than from exchange action.15 We can be confident at the 5%
level that exchanges actively adjust margin levels in response to coverage-ratio
shocks.
We extend these tests to detect if reversion to the mean is more rapid when
coverage ratios are above or below their long-run averages. The prudential
hypothesis of previous authors such as Gay, Hunter and Kolb (1986) predicts that
clearinghouses respond to low coverage ratios by raising margin requirements.
Previous models of prudentiality do not predict clearinghouse response to shocks
resulting in excess margin coverage. In contrast, the model of this paper predicts
that a high cost of margin coverage induces clearinghouses to lower margin
coverage with the provision that they meet prudentiality objectives. The ADF test
is modified to test for differential slopes on the lagged level of the coverage ratio.
Quartiles are determined for the sample of coverage ratios and indicator variables,
denoted Ql , used to classify observations into quartiles.
Lagged coverage ratios are interacted with quartile-indicator variables to
measure differential responses by the clearinghouses. This specification is:

4 K

CRt = ␣0 + ␣l1 Q l CRt−1 + ␣i CRt –i + u t (10)
l=1 j=1
We report these results in the lower panel of Table 2. Most coefficients differ
reliably from zero. The exception is the speculative margin requirement of the
soybean contract where response to low coverage ratios has the correct sign
but is not statistically significant. However, in every case, coefficients on the
highest quartile classification differ reliably from zero. This is consistent with
clearinghouse policies that lower margin requirements when margin coverage
ratios exceed their long-run averages and implies an internalization of the costs of
high margins born by the exchange membership. Fenn and Kupiec (1993) predict
a similar cost internalization.
Comparing the coefficients on the low and high coverage quartiles adds evidence
for the presence of tradeoffs between prudentiality and margin costs. Coefficients
that are larger (in absolute value) imply quicker responses to shocks to the coverage
ratio. In every case, the coefficients on the low-coverage quartiles are larger in
absolute value than those on the high-coverage quartiles. This implies that these
clearinghouses respond more quickly to surety lost when coverage ratios decline
than to the increase in costs borne by clearinghouse members when coverage ratios
rise.16
4.3. Pooled Cross-Section Time Series Analysis
The opportunity cost of margin is the difference between the cost of financing an
additional dollar of margin assets and the return on those assets. If margin deposits
were non-interest-bearing cash, movements in firms’ short-term borrowing costs
would suitably measure changes in the opportunity cost of margin. However,
members most often fulfill margin requirements with securities or standby letters
of credit. For securities, the appropriate measure of opportunity cost is the risk-
adjusted difference between the yield on the margin assets and an additional dollar
of credit. During our sample period, the sampled exchanges accepted government
and agency-debt securities as margin, Treasury bills being the most widely posted
form of margin.17
Ideally, we need a time series on the spread between the risk-adjusted borrowing
costs of market participants and rates on Treasury bills. Lacking this, we adopt
a proxy for the cost of borrowing. The proxy should capture economy-wide
shifts in the borrowing costs. In addition, borrowing-cost changes should be
correlated across borrowers even though members may be at different points on
their respective credit supply curves.
Commercial banks are a significant source of credit to futures market
participants. Consequently, the prime rate measures credit costs. When the
prime rate rises, firms with prime-based loan agreements experience a change in
borrowing costs. To isolate credit premia we use differences between the prime
rate and the Treasury bill rate to represent changes in the opportunity cost of
margin.
Our model implies a negative correlation between coverage ratios and volatility
levels when members face inelastic supplies of external finance. Holding constant
the coverage ratio, open interest, and the other assets of clearing members, a
volatility increase implies higher margin deposits and greater reliance on external
finance. With an upward-sloping supply of external funds, a higher margin
requirement raises the opportunity cost for margin deposits. A clearinghouse which
is optimizing as our model suggests will respond to this higher opportunity cost by
reducing its coverage ratios. Thus, a constant cost for borrowing implies a positive
correlation between volatility and individual member borrowing costs. Hence, our
model implies a negative correlation between volatility and coverage ratio.
This discussion suggests the following test specification:
CRit = ␣0i + ␣1 R t + ␣i2 ISDit + ␮it (11)
where i denotes the ith contract, Rt is a proxy variable capturing variation in the
opportunity cost of borrowing from economy-wide changes in bank financing,
and ISDit is the implied standard deviation for the particular contract. Including
these implied standard deviations incorporates both intertemporal and cross-
sectional differences in market participants’ opportunity cost that might result from
differences in the demand for credit to finance margin positions. The increasing
opportunity cost model implies the following coefficient restrictions: ␣1 ≤ 0 and
␣i2 ≤ 0.
Table 3 presents estimation results for Eq. (11). The first three columns of data
in Table 3 present the results from OLS estimation of Eq. (11) for the eighteen
contracts included in our sample. With the exception of the British pound, gold
and silver, the coefficients on ISD are negative and differ reliably from zero. The
coefficient on the interest-rate-spread variable differs significantly from zero only
for the contracts on the British pound and wheat. From these results we conclude
that after controlling for the opportunity costs imposed by margin deposits, the
Table 3. Estimates for the Opportunity Cost Specifications.

Unrestricted Intercept Opportunity ISD Within-Exchange ␣ij1 = ␣ij1 ∃j
Coefficients Cost Coefficient Restrictions:
␣ij2 = ␣ij2 ∃j
Intercept Opportunity ISD
Cost
British Pound 14.35 −3.57 3.36 11.23 −0.91 −27.71

(1.89) (0.76) (10.32) (0.38) (0.08) (0.77)
Cattle 7.47 −0.32 −17.33 10.50
(0.91) (0.28) (2.41) (0.27)
Deutschemark 8.83 0.18 −32.72 11.06
(2.33) (0.82) (8.70) (0.40)
Eurodollar 13.56 1.07 −2807.12 10.62
(4.78) (1.64) (291.75) (0.84)
Japanese Yen 10.77 −0.38 −51.34 9.89
(0.62) (0.24) (5.20) (0.24)
Swiss Franc 9.65 −0.00 −40.41 10.54
(1.86) (0.61) (8.79) (0.32)
S&P 500 7.86 4.25 −51.23 17.38
(7.14) (2.41) (18.45) (1.00)
Corn 9.26 −0.80 −9.66 9.35 −0.19 −3.47
(2.56) (0.93) (2.52) (0.47) (0.16) (0.34)
Soy Bean 10.57 −0.20 −27.38 6.50
(2.47) (0.96) (6.99) (0.57)
Treasury Bond 7.60 0.21 −25.22 6.25
(0.80) (0.28) (4.28) (0.44)
Wheat 12.18 −2.18 −9.12 5.86
(1.77) (0.65) (2.82) (0.49)
Heating Oil 8.60 −0.19 −3.81 11.25 −0.88 −6.01
(2.51) (0.92) (0.85) (1.18) (0.42) (0.39)
Crude Oil 9.89 −0.41 −3.35 8.71
(1.84) (0.74) (0.79) (1.14)
Coffee 10.47 −1.22 −6.71 9.31 −1.02 −4.62
(2.46) (0.83) (1.56) (1.20) (0.42) (0.44)
Sugar 8.05 −0.40 −3.01 10.51
(2.32) (0.86) (0.71) (1.19)
Gold 7.90 −0.65 −6.14 7.45 −0.50 −5.70
(1.56) (0.52) (5.97) (0.74) (0.28) (0.86)
Unrestricted Intercept Opportunity ISD Within-Exchange ␣ij1 = ␣ij1 ∃j
Coefficients Cost Coefficient Restrictions:
␣ij2 = ␣ij2 ∃j
Intercept Opportunity ISD
Cost
Silver 8.80 −1.15 −1.30 8.15

(2.39) (1.05) (6.87) (0.80)
Copper 7.95 −0.57 −5.32 7.89
(1.63) (0.67) (1.81) (0.77)
Note: Standard errors in parentheses. Table reports results for two methods of estimating the following
time-series and cross-sectional specification: CRijt = ␣ij0 + ␣ij1 Rt + ␣ij2 ISDit + ␮ijt . Variable
definitions are: CRijt is the coverage ratio for contract i traded at exchange j on date t, Rt
is the excess of the prime rate over the 3-month treasury bill rate, and ISDit is the implied
standard deviation for options trading on contract i at date t. Unrestricted coefficients are from
separate OLS regressions for each contract. The within-exchange restricted coefficients are
from estimating those equations as a system of seemingly unrelated regressions. Regressions
include equality restrictions on the two right-hand side variables. Within the five exchanges, the
coefficients on R are restricted to equality and the coefficients on ISD are restricted to equality.
See Table 1 for contracts traded on the five exchanges.
coefficients on ISD generally (17 of 18 contracts) support the hypothesis of

increasing cost. Thus, volatility increases lead to lower coverage; that is, margin
coverage increases but by a smaller percentage than the accompanying volatility
increase.
We also estimate Eq. (11) as a system of 18 equations using the method
of iterated seemingly unrelated regressions. To represent better the fact that
exchange decisions on margin requirements reflect the opportunity costs and risk
tolerances of their memberships irrespective of their preferred trading venues, we
impose coefficient restrictions. Specifically, we require equal coefficients on the
opportunity costs for all contracts trading within a single exchange. Likewise,
we impose the restriction that the ISD coefficients on contracts trading within an
exchange also be equal. The effect of these restrictions is to represent exchange
members as, at the margin, indifferent as to the margin requirements for the various
contracts they may trade on the exchange.
Columns 4 through 7 of Table 3 report the SUR results. With the exception of the
opportunity cost at COMEX, all coefficients are negative and differ significantly
from zero at usual confidence levels. This result concurs with the increasing-cost
conclusion obtained from the OLS coefficients on ISD. The significantly negative
coefficients on opportunity cost bolster this conclusion. This implies that with
risk exposure held constant, an increase in the opportunity cost of margin

deposits prompts exchanges to choose lower margin and, therefore, less risk
coverage.
We also compare the margin coverages for gross- and net-margin exchanges.
The two gross-margin exchanges – the CME and the NYMEX – require clearing
members to post with their respective clearinghouses, the full amount of margin due
on all the positions carried by each clearing member. In contrast, the net-margin
exchanges – the CBOT, the CSCE and the COMEX – allow clearing members
to post margin on the net of their long and short positions. The ability to net
contracts can substantially lessen aggregate margins deposited with the exchange.
Hence, increasing costs implies that the coefficients on opportunity cost and on
ISD for gross-margin exchanges should be larger in absolute value than those for
net-margin exchanges.
Four F tests generally support this view. Comparing ISD results, the coefficient
for the CME is substantially larger in absolute value than for the CBOT. Likewise,
the equal-weighted average of ISD coefficients for the CME and the NYMEX is
larger than the equal-weighted average of ISD coefficients for the CBOT, CSCE
and the COMEX. The F tests for these comparisons conclude that the differences
are significant. Comparing opportunity cost results, the coefficient for the CME
is substantially larger in absolute value than that for the CBOT. This difference
is statistically significant. Comparing the weighted-average coefficients for the
CME and NYMEX with those for the CBOT, CSCE and COMEX also obtains
a larger effect, but the comparison is not statistically significant at the usual
levels.
5. EXTENSIONS: OTHER CLEARINGHOUSE RISK

CONTROL MECHANISMS
The preceding sections model a clearinghouse using margin deposits to
manage default losses and provide evidence supporting the model. This simple
clearinghouse need not monitor the financial condition of its participants, link
margin deposits to the riskiness of its participants, expel nonperforming members
or otherwise seek to control risk. Yet, these ancillary activities are likely sources for
additional economies. Exploring these issues adds to our understanding of present-
day institutional arrangements. This section begins by examining clearinghouse
policy toward defaulting members. We then model the monitoring activities of
the clearinghouse with respect to the value of membership and to the financial
condition of its members.
5.1. The Threat of Expulsion
Because clearinghouses reduce deadweight losses from opportunistic default and

allow participants to economize on margin costs, membership is valuable. When
traders expect to trade in more than one period, the threat of expulsion reduces
opportunistic default. Verifiable membership value reduces the amounts of required
margin because the expulsion threat induces contract performance beyond that
obtainable from margin deposits. A member’s presence may also be beneficial
to other members. Such joint benefits raise questions about the credibility of an
expulsion threat. This subsection lays out conditions for a credible expulsion policy.
Let C denote the capitalized value derived by members from continued
membership. This value has two sources: expectations of lower loss rates resulting
from contract default and lower margin requirements. We denote gains accruing
to member d as C(d, d) and the aggregate of benefits accruing to other members
as C(d, CH). A short contract position rationally responds to an expulsion threat
by performing if contract-performance costs are less than membership value:

N(d, j)
|x − M| < C(d, d) (12)

j=1
A similar condition applies to long positions. A clearinghouse rationally expels its

defaulting members if default costs, including both the contractual shortfall and
the deadweight loss, exceed the value losses other members incur by expelling d,
that is, when:

(1 + ␣)
N(d, j)
|x − M| > C(d, CH) (13)

j=1
Combining these conditions, a credible expulsion threat exists and a potential

defaulter rationally performs on the contract when

C(d, CH)

<
N(d, j)
|x − M| < C(d, d) (14)

(1 + ␣)

j=1
As the membership of a clearinghouse increases, the cost to the group as a whole of

expelling any one member will decline, but the cost to an individual of losing clear-
ing privileges will, if anything, increase. When members suffer virtually no loss
from refusing to trade with d, then C(d, CH) = 0. If C(d, CH) = 0 and C(d, d) > 0,
then an expulsion threat is always credible. Moreover, it is Pareto improving for
clearinghouse members to pre-commit to expelling any member defaulting on
a contract because the threat of membership loss reduces opportunistic default

without raising margin levels.18
The expulsion threat alters the clearinghouse’s basic maximizing problem in
that |x| > M is the optimal default rule only when the value of future clearing
privileges is ignored. Equation (14) gives a more general default rule for the short
position. This has three implications. First, due to the threat of expulsion, firms
default in fewer states of the world. Second, the value of clearing membership
C perfectly substitutes for margin deposits M in preventing default but third, the
value of membership imperfectly substitutes for margin deposits when default
occurs. This is because increases in required margin increase the amount received
in default states but the benefits of membership are not transferable.19
Proposition VII. When the clearinghouse can make a credible expulsion threat,
the optimal margin coverage ratio increases as volatility increases.
When C is identical across individuals, a credible expulsion threat increases the

lower limit of integration in Eq. (6) from M to M + C/|N*(j)|, where N* is the
median trader’s exposure. The clearinghouse now minimizes:
n
∞

|N(j)| iM(j) + ␣
x − M(j) f(x, ␴)dx (15)
j=1 M(j)+ |N(j)|
C
The first order condition for minimization of (15) with respect to M(j, h) and
M(h, j) is:

C C C i
1 − F M ∗∗ + ,␴ + f M ∗∗ + ,␴ =
|N(j)| |N(j)| |N(j)| ␣
Whenever clearinghouse membership is valuable, then C > 0 and the final term
on the hand side of (16) is strictly positive. This means that a policy of expelling a
defaulting member reduces the probability of default F(·) to less than i/␣, the level
prevailing absent expulsion. Since a credible expulsion threat acts as a substitute
for costly margin, the clearinghouse will choose a higher level of protection than
it otherwise would.
Differentiating Eq. (16) with respect to M** and C/|N(j)|, implies that when
margin is taken dM** /d(C/|N(j)|) < 0. Thus, increasing the value placed on
membership decreases the amount of margin required to obtain a given level of
safety. Intuitively, as membership and the volume cleared expands over time, the
threat of expulsion becomes more serious because clearing privileges are more
valuable, and the clearinghouse can place greater reliance on the expulsion threat
as a deterrent to default, thus decreasing the optimal level of margin.
If the threat of expulsion is an important part of the guarantee system, how would
coverage rations change when volatility changes? On the one hand, if we hold the
value of clearing C constant, then optimal margin coverage ratio M** /␴ increases
as volatility increases. Thus, margin must increase more rapidly than volatility to
supply the same level of protection. This result contrasts sharply with our analytic
results for constant or increasing costs of margin deposits.
On the other hand, it may be that the value of clearing C changes systematically
with changes in volatility. The most likely conjecture is that the value of clearing
increases as volatility increases, because increased volatility is likely to be
associated with increased volume of trading. This could reverse the result cited in
the previous paragraph: if clearing becomes more valuable in volatile markets, it
is possible that optimal margin could drop.
The above suggests the benefits of clearinghouse creation can go beyond margin
economies and avoiding deadweight default costs. Creation of a clearinghouse
assures contract performance at levels beyond that obtained by margin deposits.
Further, it can be the case that reliance on membership value is more cost effective
than relying simply on margin deposits.
5.2. The Clearinghouse as Monitor
Relaxing the assumption that only collateral can be attached in case of default, we
allow counterparties to grant senior claims on unencumbered assets k(j).20 Each
party knows its own k(j), however we assume counterparties incur an examination
cost to learn k(j). This cost is denoted e. Traders choose to be monitored when the
savings from these senior claims against k(j) exceed examination costs.
Examination, at most, saves the firm the opportunity cost of holding k(j), that
is, the maximum savings is ik(j). If the quantity ik(j) is less than e, then inspection
does not pay. However, failure of this condition is not sufficient for inspection
to occur. If the optimal margin M*|N(j)| without inspection is less than k(j), the
opportunity cost savings from granting a senior claim against k(j) is iM*|N(j)|.21
Proposition VIII. When the clearinghouse acts as monitor to verify the

existence of unencumbered assets, coverage ratios increase as volatility
increases.
Inspecting a member firm reveals one of two conditions. Members post no margin
when unencumbered assets k(j) exceed M* |N(j)|. Alternately, if k(j) is less than
M* |N(j)|, then the clearinghouse’s problem is of the same form as Eq. (6) with
M + k(j)/|N(j)| substituted for M.
If k(j) is less than M* |N(j)|, the optimal margin rule is:

∗∗∗ k(j) i
M + = F −1 ,s (16)
|N(j )| ␣
Because k(j) is less than M* |N(j)|, parties still post margin. Thus, for a constant
opportunity cost of margin, the optimal default probability is identical to the no-
examination case because firms substitute claims against unencumbered assets for
more costly forms of margin. Unlike the case where expulsion acts as a deterrent
to default, if default occurs, the other assets can be seized, making them a perfect
substitute for margin deposits.
However, when opportunity costs are increasing in total required margin,
examination decreases the optimal default rate. When other assets are substituted
for margin, the cost of additional margin decreases, and the optimal level of
protection can thus be increased.
Equation (16) implies that with monitoring, coverage ratios increase as volatility
increases, as unencumbered assets k(j) decrease, and as the number of open
contracts increases. The result that the optimal coverage ratio declines as firms
become less able to substitute unencumbered assets for collateral deposits differs
from the predictions of the constant opportunity cost model (Eq. (4)) and the
increasing opportunity cost model (Eq. (8)).
Since firm asset holdings are dissimilar, reliance on unencumbered assets
entails scrutiny of member positions, with margin setting on a member-by-
member basis. Instead, margin requirements are uniform across the memberships
of organized clearinghouses. This uniformity arises for several reasons. First,
payment delays may be the principal cause of deadweight losses for members of
the clearinghouse. The presence of unencumbered-but-illiquid assets may not be
useful for time-critical settlement requirements. Second, timely verification of the
existence of k may significantly raise clearinghouse costs, making is uneconomical
to monitor at all. Third, netting may reduce each party’s net exposure to such
low levels that intensive monitoring can not be cost-effective. In any event, the
uniformity of margins across members suggests that if clearinghouses do engage
in extensive monitoring, it must be for purposes other than controlling risk between
members.
The prediction of a positive correlation between volatility and the coverage ratio
also contrasts sharply with the independence of the coverage ratio and volatility.
Thus, this result is distinct from the simple netting model of Eq. (5) and the negative
correlation generated by the increasing opportunity cost of funds model of Eq. (10).
6. SUMMARY
We incorporate the cost of external funds and the deadweight losses associated with
counterparty default into collateral decisions. Given their scale of activities, it is not
surprising that clearinghouses internalize these costs into their margin decisions.
Thus, clearinghouse pursuit of prudentiality through margin is constrained by
the costs that members incur by carrying these balances. When margin is set
without regard to additional information about the condition of the clearinghouse
members, the coverage ratio is either uncorrelated or negatively correlated with
volatility.
The time series of coverage ratios supports the conclusion that clearinghouse
determination of margin incorporates prudential concerns. Our empirical results
demonstrate that clearinghouses respond to high levels of margin by adjusting
coverage ratios downward. This behavior is not consistent with prudentiality alone,
but is consistent with exchanges optimizing across the costs of placing margin
deposits and the costs from incurring deadweight losses.
Our pooled-regression results show that futures clearinghouses set margin in a
cost-minimizing fashion, balancing the risk of loss against the greater opportunity
costs associated with higher margins. Those results suggest that at least part of
these opportunity costs arise because market participants have imperfect access to
capital markets for their general financing.
We extend our model to capture other institutional features. We expect positive
correlations between coverage ratios and volatility when clearinghouses actively
monitor their members for risk-management purposes. Our emphasis on the
foundations of the clearinghouse makes clear that membership is valuable. Because
membership is valuable, expelling defaulting members is credible and effective for
the clearinghouse. This means that members will perform on their contracts even
when price moves exceed the value of margin on deposit.
NOTES
1. The term collateral is used for all instances where a deposit is required to reduce
exposure to credit risk. The term “margin” when used here is the special case applying
when collateral deposits are uniform within broad categories of counterparties.
2. Baer, France and Moser (1995b) re-examine some previous tests of prudentiality.
3. For-profit exchanges operating in a competitive market can also be expected to cost
minimize. Hence, our model also applies to for-profit organizations.
4. For example, courts may be unable to force the transfer of collateral quickly enough
to avoid deadweight default costs and will involve significant legal costs.
5. Generalization to a multi-contract exchange results in a relation between the loss on

a portfolio of contracts and the sum of margin deposits. The results depend on the extent of
members’ diversification. See Baer, France and Moser (1993).
6. Most margin deposits at U.S. exchanges are in interest-bearing forms though marked-
to-market gains or losses (variation margin) require cash payments. For example, a
U.S. Treasury Bill deposited as margin would be returned to the depositor when the
account is closed. This arrangement effectively gives the depositor interest on deposit.
The London Clearinghouse pays interest on cash deposits. Our model covers both
cases. If cash is deposited, the opportunity cost is driven by the levels of market
rates. Clearinghouses also allow limited use of standby letters of credit (SLOCs) for
margin.
7. A futures clearinghouse also allows its members to exploit a variety of other scale
economies accessible only by acting as a group. Centralization simplifies record keeping
since members need only keep track of their positions with the clearinghouse. Credit
monitoring and control is simplified, since members’ financial standing need only be
assessed once by the clearinghouse, rather than by every counterparty. Economies of scope
exist between record keeping and credit control, since knowledge of a member’s net position
is needed to assess exposure. Finally, precommiting to binding arbitration lowers cost
because disputes are no longer a matter for bilateral bargaining. See Baer, France, and
Moser (1995a).
8. Calomiris and Hubbard (1995), Fazzari, Hubbard and Petersen (1988); and Hubbard
and Kashyap (1992) provide evidence that nonfinancial firms behave as if financing growth
through external financing is relatively expensive. Baer and McElravey (1994) report similar
results for U.S. banking corporations.
9. The use of margin as collateral, the netting, and the attendant loss-sharing rules
effectively redefine the legal priority of claims. We assume the ability of a clearinghouse
to take possession of margin assets in the event of default is not obstructed by law. When
we say a “pre-agreed rule,” we assume that the priority of claims in the event of default is
clear. Historically, the rise of clearinghouses resulted in a clarification and streamlining of
bankruptcy law as it applied to futures claims.
10. Certain loss-sharing rules can undo this result by allocating a disproportionate
share of losses to an individual member. Since loss-sharing rules are agreed upon in
advance and since clearinghouse membership is voluntary, it can be shown that such
rules will not be adopted. Futures exchanges generally use a common fund to pay for
defaults.
11. See Laffont (1988, pp. 51–53), or Cornes and Sandler (1986). Exchanges usually set
margins, not on the basis of a direct vote, but by a committee designed to be representative
of the membership.
12. Margin amounts collected when these accounts are opened are called initial margin.
Should deposited amounts fall below a specified maintenance level, the margin balance
must be restored to the current initial level. Maintenance margin requirements in U.S. stock
markets differ. In stock markets, should a deficiency occur, margin must be restored to the
maintenance level.
13. Implied standard deviations for short-term interest rate contracts are generally
expressed in terms of yield variation. For consistency with our other contracts, they are
reported here in terms of variation of rates of return.
14. See Bollerslev, Chou, Jayaraman and Kroner (1991).
15. Another possibility not considered here is that the margin responses of exchanges
induce subsequent changes in volatility. The evidence does not generally support this claim,
however for an alternative view, see Hardouvelis (1988).
16. An F test indicates that the difference between the coefficients on the high and low
quartiles of the S&P and Deutschemark contracts is significant at better than the 95% level.
17. Other clearinghouses, for instance the Options Clearing Corporation, have long
accepted equity as margin. Futures clearinghouses are increasingly adopting this practice.
18. A minimal number of clearing members may be necessary to insure that none are
too valuable to expel.
19. This presumes that the value of an exchange seat reflects trading rather than clearing
privileges, as suggested by Bhasin and Brown (1997). If clearing memberships were a
separately traded asset, their value would reflect the value of C(d, d) for the marginal
member, and that amount could be recouped by the other members in a default. The model
would then closely resemble that in the next section, where counterparties grant senior
claims on assets.
20. Relaxing this assumption implicitly assumes that courts are effective in seizing
collateral and that the speed of payment is not an issue. If payment delay is the principal
reason that default imposes a deadweight loss on the membership, then the existence of
unencumbered assets may be irrelevant.
21. We assume that the inspection process includes assessing the probable value of k
in the default state. Clearinghouses “haircut” non-cash assets by valuing them at less than
current market value.
ACKNOWLEDGMENTS
Much of this paper was completed while France was visiting at the Chicago
Mercantile Exchange and the University of Chicago. Susanne Malek and Jan
Napoli provided valuable research assistance. The authors thank John Conley,
Ramon DeGennaro, Mark Flannery, Gary Koppenhaver, Todd Petzel, Will
Roberds, Jerry Roberts, an anonymous referee, and seminar participants at the
University of Illinois at Urbana-Champaign, the Federal Reserve Bank of Chicago,
the University of Tennessee, and the Chicago Risk Management Workshop.
Opinions expressed are entirely those of the authors and do not reflect concurrence
by the Federal Reserve, the Chicago Mercantile Exchange, or the World Bank.
REFERENCES
Baer, H. L., France, V. G., & Moser, J. T. (1993). Opportunity cost and prudentiality: A representative-
agent model of futures clearinghouse behavior. Faculty Working Paper #93–0142, University
of Illinois.
Baer, H. L., France, V. G., & Moser, J. T. (1995a, Spring). What does a clearinghouse do? Derivatives
Quarterly, 1(3), 39–46.
Baer, H. L., France, V. G., & Moser, J. T. (1995b). Determination of collateral deposits by bilateral
parties and clearinghouses. In: Proceedings of the 31st Annual Conference on Bank Structure
and Competition (pp. 525–546). Chicago: Federal Reserve Bank of Chicago.
Baer, H. L., & McElravey, J. (1994, January). Capital adequacy and the growth of U.S. banks. In:
C. Stone & A. Zissu (Eds), Global Risk Based Capital Regulations: Management and Funding
Strategies. Homewood, IL: Dow-Jones-Irwin.
Barone-Adesi, G., & Whaley, R. E. (1987). Efficient analytic approximation of American option values.
Bhasin, V., & Brown, D. P. (1997, January). The organization of competitive financial exchanges.
Working Paper, Indiana University. Journal of Financial Markets (Forthcoming).
Bollerslev, T., Chou, R. Y., Jayaraman, N., & Kroner, K. F. (1991). ARCH modeling in finance: A
selective review of the theory and empirical evidence with suggestions for future research.
Journal of Econometrics, 52, 5–59.
Brennan, M. J. (1986). A theory of price limits in futures markets. Journal of Financial Economics,
16, 49–56.
Calomiris, C. W., & Hubbard, R. G. (1995, October). Internal finance and investment: Evidence from
the undistributed profits tax of 1936–1937. Journal of Business, 68(4), 443–482.
Cornes, R., & Sandler, T. (1986). The theory of externalities, public goods, and club goods. Cambridge:
Craine, R. (1997). Valuing the futures market performance guarantee. Macroeconomic Dynamics, 1(4),
701–719.
Diamond, D. (1984). Financial intermediation and delegated monitoring. Review of Economic Studies,
51, 393–414.
Fazzari, S. M., Hubbard, R. G., & Peterson, B. C. (1988). Financing constraints and corporate
investment. Brookings Papers on Economic Activity, 141–195.
Fenn, G., & Kupiec, P. (1993). Prudential margin policy in a future-style settlement system. Journal
of Futures Markets, 13, 389–408.
Figlewski, S. (1984). Margins and market integrity: Margin setting for stock index futures and options.
Journal of Futures Markets, 4(Fall), 385–416.
Fuller, W. A. (1976). Introduction to statistical time series. New York: Wiley.
Gay, G. D., Hunter, W. C., & Kolb, R. W. (1986). A comparative analysis of futures contract margins.
Journal of Futures Markets, 6(Summer), 307–324.
Gorton, G. (1985). Clearinghouses and the origin of central banking in the U.S. Journal of Economic
History, 45(June), 277–283.
Hardouvelis, G. A. (1988, Summer). Margin requirements and stock market volatility. Federal Reserve
Bank of New York Quarterly Review, 13(2), 80–89.
Hsieh, D. A. (1993). Implication of non-linear dynamics for financial risk management. Journal of
Financial and Quantitative Analysis, 28(March), 41–64.
Hubbard, R. G., & Kashyap, A. K. (1992). Internal net worth and the investment process: An application
to U.S. agriculture. Journal of Political Economy, 100(June), 506–534.
Kofman, P. (1993). Optimizing futures margins with distribution tails. Advances in Futures and Options
Research, 6, 263–278.
Laffont, J.-J. (1988). Fundamentals of public economics. Cambridge, MA: MIT Press.
Merton, R. C., & Bodie, Z. (1992). On the management of financial guarantees. Financial Management,
21(Winter), 87–109.
Whaley, R. E. (1982). Valuation of American call options on dividend-paying stocks: Empirical tests.
COLLATERALIZATION AND THE
NUMBER OF LENDERS IN PRIVATE
DEBT CONTRACTS: AN EMPIRICAL
ANALYSIS
Gordon S. Roberts and Nadeem A. Siddiqi
ABSTRACT
Using the Dealscan database of large, U.S. corporate loans, we examine the
determinants of the number of bank relationships and the presence or absence
of collateral. Consistent with prior studies, we find that important explanatory
variables are firm quality, desire for financial flexibility, the probability of
financial distress, growth opportunities and firm size. Higher quality firms
as well as firms with a stronger desire for financial flexibility are less likely
to collateralize and borrow from more lenders. Larger firms as well as those
with lower probabilities of financial distress and greater growth opportunities
prefer multiple lenders.
1. INTRODUCTION
Why do some firms choose to borrow from a single bank while others engage in
multiple banking relationships? What factors determine whether a firm secures its
loans and how do the two decisions interact? A number of recent empirical studies
analyze multiple banking relationships in Europe: Detragiache et al. (2000) and
Research in Finance
ISSN: 0196-3821/doi:10.1016/S0196-3821(04)21010-3
229
230 GORDON S. ROBERTS AND NADEEM A. SIDDIQI
D’Auria et al. (1999) – Italy, Degryse and Ongena (2001) – Norway, Farinha and
Santos (2002) – Portugal, and Harhoff and Korting (1998) – Germany. Studies on
U.S. banking relationships have focused on small firms (Petersen & Rajan, 1994)
and the role of multiple bank relationships in mitigating hold-up problems and
creating financing flexibility (Houston & James, 1996, 2001).
The present paper examines the determinants of bank relationships and collateral
for large U.S. loans in the Loan Pricing Corporation’s DealScan database for
the period 1990–1999. Consistent with prior studies, we find that important
explanatory variables are firm quality, desire for financial flexibility, the probability
of financial distress, growth opportunities and firm size. Higher quality firms
as well as firms with a stronger desire for financial flexibility are less likely to
collateralize and borrow from more lenders. Larger firms as well as those with
lower probabilities of financial distress and greater growth opportunities prefer
multiple lenders. Although there is a common set of explanatory variables, tests
for interdependence effects suggest that the number of bank relationships and
whether to offer collateral are independent decisions.
Section 2 reviews relevant literature. Section 3 presents the models and variables
used. The data sample is presented in Section 4, which is followed by a discussion
of our simultaneous modeling technique analysis in Section 5. Section 6 presents
our statistical analysis and Section 7 concludes paper.
2. PRIOR LITERATURE
Prior research indicates that firm quality, the probability of financial distress,
growth opportunities, the desire for managerial flexibility, and firm size are relevant
determinants of both the number of lenders and whether loans are collateralized.
We discuss each decision in turn. We can summarize the testable hypotheses from
these studies with respect to the collateralization and the number of creditors in
private debt contracts as shown in Table 1.
2.1. Number of Lenders
We expect firms with greater growth opportunities to borrow from multiple lenders
for moral hazard reasons (Repullo & Suarez, 1998) as well as to limit the single
lender’s bargaining power and to mitigate any information monopoly (Houston &
James, 1996; Rajan, 1992). A similar argument applies to larger firms. Detragiache,
Garella and Guiso (2000) present a model in which borrowing from multiple banks
can ensure a more stable supply of credit. Avoiding rationing by a single lender
Collateralization and the Number of Lenders in Private Debt Contracts 231
Table 1. Summary of Hypotheses.

Collateralization Number of Creditors
Growth − Gilson and Warner (2000) + Rajan (1992), Houston and James
(1996), Repullo and Suarez (1998)
Probability of + Gorton and Kahn (1993), Rajan − Berlin and Mester (1992),
distress and Winton (1995) Chemmanur and Fulghieri (1994)
+ Degryse and Ongena (2001)
Firm quality + Bolton and Scharfstein (1996) + Berlin and Loeys (1988), Houston
and James (1996), Rajan (1992),
Bolton and Scharfstein (1996)
− Rajan and Winton (1995) − Detragiache et al. (2000), Degryse
and Ongena (2001)
Firm size + Repullo and Suarez (1998),
Detragiache et al. (2000)
Loan purpose + Gilson and Warner (2000) + Gilson and Warner (2000)
Note: The table summarizes the hypotheses and findings of previous studies on collateralization
and number of creditors in private debt contracts with respect to the five firm characteristics
mentioned. A “+” in front of the study indicates a positive hypothesized relationship between
the debt contract and the firm characteristic, and conversely a “–” in front of the study indicates
a negative relationship. Collateralization and Number of Creditors are binary variables, with “1”
indicating presence of collateral, or multiple creditors, and “0” indicating absence of collateral,
or single creditor, respectively.
reduces the risk that a profitable project will have to be prematurely liquidated.
If the adverse selection problem is very severe because outside banks have less
information than the borrower’s inside bank and suspect the firm to be a bad risk,
it may pay for the firm to establish multiple relationships from the beginning.
Hence, they theorize that larger and less profitable firms are less likely to have
single banking relationships.
Degryse and Ongena (2001) examine how the number of bank relationships
affects firm performance for a sample of Norwegian firms. Firms with valuable
proprietary information may use fewer creditors in order to prevent information
leakage. At the same time, financially distressed firms facing credit rationing at
their main bank may be forced to seek additional, more costly financing elsewhere.
Farhina and Santos (2002) obtain a similar result for Portuguese firms. A positive
association between the probability of financial distress and the number of bank
relationships is also present in Germany according to Harhoff and Korting (1998).
Firms that are looking to increase managerial discretion and flexibility, as
indicated by the loan purpose, will also borrow from multiple creditors (Gilson &
Warner, 2000). In contrast, firms with greater probability of financial distress, or
higher credit risk firms find the option to renegotiate more valuable and hence
may prefer to borrow from fewer lenders (Berlin & Mester, 1992). Further,
Chemmanur and Fulghieri (1994) find that banks are able to use reputation as a
commitment device to promise firms credibly that they will devote more resources
toward evaluating them and thereby make better renegotiation-versus-liquidation
decisions in the event of financial distress.
2.2. Collateralization
High growth firms, as well as those specifically looking to increase managerial

discretion and flexibility, as indicated by the loan purpose, are expected to put
up less collateral. Rajan and Winton (1995) analyze the role of covenants and
collateral in private debt contracts as incentives for lenders to monitor. They find
that, in order to balance monitoring against efficient liquidation, the private lender
should only take collateral in a bad state, signaling the information to the public.
Their model also predicts that the collateralization of private debt will be correlated
with financial distress possibilities of the firm. Gorton and Kahn (1993) reach a
similar conclusion: collateral and seniority are crucial since they allow the bank
to threaten the borrower and liquidate inefficient projects.
Gilson and Warner (2000) find in their empirical study that bank loans are
more often secured than junk bonds. They explain this by noting that security
reduces managerial discretion since firms cannot easily sell assets that are pledged
as collateral. They theorize that firms looking for the “flexibility to grow” would
prefer less collateralization, as well as multiple creditors, to free managers from
the immediate obligation of paying off debt, and increase their freedom to set
corporate policies.
Pozzolo (2002) examines the relationship between secured lending and
borrowers’ riskiness. He develops a theoretical model predicting that banks require
guarantees on loans to riskier borrowers. Using information on bank loans to a large
sample of Italian non-financial firms, he finds empirical support for the predictions
of his model. In contrast, John, Lynch and Puri (2002) develop and test a theoretical
model based on agency problems between managers and claimholders suggesting
that secured debt is not issued by riskier borrowers. Bolton and Scharfstein (1996)
also hold that higher quality firms offer more collateral to bond against strategic
defaults. They identify lack of liquidity as a cause of default positing that high-
quality firms borrow from multiple lenders. As a result, their work suggests that
firm decisions on the number of lenders and collateral are jointly determined.
Summarized in Table 1, our brief review of prior studies identifies growth,
probability of financial distress, firm quality, firm size and loan purpose as factors
driving choices of collateralization and the number of creditors. As the table
demonstrates, prior research fails to reach a consensus on the impacts of these

variables.
3. VARIABLES AND MODEL

The different firm characteristics reviewed in the previous section are proxied as
follows.
Firm quality We proxy firm quality by Standard and Poor’s bond rating;
the ratings of AAA, AA through C are translated to an
ordinal scale ranging from 2 to 21, as recorded by
COMPUSTAT. Blume et al. (1998) point out that rating
agencies, such as Standard and Poor’s, employ both
publicly available information, such as accounting
statements, and non-public information, such as confidential
interviews with management, to assign quality ratings as a
measure of the “creditworthiness” of a corporation with
respect to a particular debt security. Their survey of prior
studies concludes that quality ratings do have some
informational content and that these ratings contain
information beyond what is publicly available. Their own
empirical study also concludes that bond rating standards
have become more stringent over time.
Probability of This is proxied by the Z score value, calculated using
distress Altman’s (1968) model. The Z score model utilizes multiple
discriminant analysis to obtain a linear function of variables
that is able to discriminate membership between two
populations, bankrupt and non-bankrupt. The Z score
indicates how close the firm is to being classified as being in
one of the groups, i.e., the score is a measure of the
probability of being in financial distress.
Growth The standard market to book ratio is used to proxy for
opportunities growth opportunities.
Loan purpose Following Carey et al. (1998) and Dennis et al. (2000), we
place into groups the 16 stated uses of the debt by the
borrowers, as recorded by the Loan Pricing Corporation, to
indicate the purpose of the loan. We utilize Carey et al.’s
(1998) four groups and categorize loans as being made for
the following purposes: (1) general corporate purposes; (2)
debt repayment/consolidation; (3) takeover/acquisition; and

(4) other.
Firm size The size of the firm is approximated by the natural
logarithm of sales.
Aside from the above firm characteristics, we also control for the effects of
leverage in our regressions. The importance of controlling for leverage is
demonstrated by Lang, Ofek and Stulz (1996), who find that future growth and
investments are negatively related to leverage. Johnson (1998) also found that
leverage was statistically and economically significantly higher for firms with
bank debt due to the fact that bank debt attenuates negative effects on leverage of
potential asset substitution problems. Hence, omitting this variable would cause
our model to be misspecified. Leverage is measured by the book value of debt
divided by the market value of equity plus the book value of debt. We control for
the type of private debt issued since this may directly affect the need for collateral
as well as the number of creditors. The facility type is also categorized into one
of four groups as follows: (1) short-term revolver; (2) long-term revolver; (3)
term loan; and (4) other (including leases, bridge loans, notes, standby letters of
credit, etc.).
For the collateral regression, we also control for the amount of fixed assets
available to put up as collateral, as well as the type of lead lender, whether bank
or finance company. Carey et al. (1998) and Houston and James (2001) describe
finance companies as “asset-based” lenders and banks as “cash flow” lenders. In
making a loan, an asset-based lender emphasizes collateral as a source of ultimate
repayment whereas a cash flow lender relies more heavily on projected cash flow
from operations. The amount of fixed assets is measured by the ratio of fixed
assets to total assets, while the lender type is a binary variable, with 1 for a finance
company as lead lender and 0 for a bank as lead lender. A detailed formulation of
all the regression variables is in Appendix A.
We utilize a simultaneous model capable of assessing the interdependencies
between the choice of the number of creditors as well as the amount of collateral
offered hypothesized by Bolton and Scharfstein (1996). Building on the framework
of Dennis et al. (2000), we model our regressions as follows:
Collateral = ␣1 Creditors + ␤ 1 X 1 + e 1 (1)

Creditors = ␣2 Collateral + ␤ 2X2 + e2 (2)
where the ␣’s are coefficients of the interdependence effects. This permits a
direct test for the significance of interdependence effects. Both Collateral and
Creditors are binary variables, with 1 indicating presence of collateral or multiple
creditors, and 0 the absence of collateral or a single creditor, respectively. X1 and

X2 contain the firm characteristic proxies, discussed above, that are expected to
influence the decision to issue collateral for the loan and the number of creditors,
respectively.
4. DATA
We started construction of our sample set from the DealScan 5.6 database compiled
by the Loan Pricing Corporation (LPC). In this area of the literature, Shockley and
Thakor (1997) were the first to utilize the LPC database to empirically analyze
the fee structure of bank loan commitment contracts.1 The LPC database provides
detailed market information on commercial loans and private placements made
to publicly held U.S. companies that are required to file such information with
the Securities and Exchange Commission (SEC). The database also includes deal
information obtained directly from banks, which is later confirmed after the deal is
recorded with the SEC. The data include details such as the name of the borrower,
the names of all lenders party to the loan contract at origination, the type, purpose,
maturity, amount, material restrictions and contract date of the loan as well as other
details. An illustration of relevant sample data from LPC’s DealScan database is
presented in Table 2.
Table 2. Sample Data from Loan Pricing Corporation’s DealScan Database.

Borrower Name Amgen Inc. Beverly Enterprises
Ticker AMGN BEV

Deal purpose Debt repayment/consolidation General corporate purposes
Deal active date 23-Jun-95 01-Nov-94
Deal size $ 223,000,000 $ 375,000,000
Facility type Standby LC Revolver < 1Yr.
Facility size $ 73,000,000 $ 375,000,000
Facility active date 23-Jun-95 01-Nov-94
Facility maturity days 05-Dec-97 .
Seniority Senior No
Secured No No
Bid option No No
Materially restricted Yes No
Lenders and share Swiss Bank Corp (58.9%), J. P. Morgan & Co. (100%)
ABNAMRO (41.1%)
Note: The database contains over 200 searchable fields. Only some of the relevant fields are illustrated
in this sample. Each deal (or package) may consist of one or more facilities, with each facility
separately listed.
In constructing our sample, we eliminated those loan records for which

data required to fully characterize the borrower were unavailable on the
COMPUSTAT database. We use the previous year’s COMPUSTAT data to avoid
a simultaneity bias. Furthermore, we retain only those loan records for which
a bond rating was assigned. These criteria result in a total of 218 loans. Hence,
all 218 records are complete with respect to all the information needed for the
study. There are no records that have any missing information, since they are
filtered along the way. A summary of the sample characteristics is presented in
Table 3.
Since we restrict it to firms with bond ratings only, our sample represents
medium to large firms with access to both the public and private debt markets.
This bond rating “filter” should not bias our sample too much since the LPC
dataset is compiled primarily from firms’ SEC filings, and smaller firms would
have been excluded from this process to begin with.2 Furthermore, larger firms
would have more bargaining power with lenders and thus would be better able
to affect the structure of private debt contracts to suit their needs than smaller
sized firms (Rajan, 1992). This structure, chosen by the borrower rather than that
dictated by the lender, is precisely what we wish to study.
Table 3. Summary Statistics of Sample Set.

Variable Mean*** Median Std. Dev.
BONDRAT 11.10 11.00 3.32

CUMPROF 0.1446 0.1211 0.2265
DEBTASST 0.3415 0.3249 0.1611
EBITSALE 0.1562 0.1317 0.1376
FACMATUR 1405.37 1826.00 707.54
FACSIZE 533.25 250.00 851.97
FACTA 0.1520 0.0947 0.1685
FACTD 0.6058 0.3372 1.2294
FATA 0.4424 0.4024 0.2262
FIRMVAL 6801 2483 13016
LTDTA 0.2881 0.2594 0.1561
MKTBOOK 1.5301 1.4127 0.4982
MKTLEV 0.3333 0.2787 0.1920
OITA 0.0909 0.0843 0.0591
SALES 6327 2282 10804
TOTASST 6119 2286 10570
ZSCORE 4.3250 3.1884 4.5719
Note: All dollar figures are in millions of dollars, except the share price. The complete definition of
the variables is found in Appendix A.
∗∗∗ All mean values were found to be statistically significantly different from 0 at the 1% level.
5. MODEL ESTIMATION
Following Dennis et al. (2000), we apply Mallar’s (1977) two-stage estimation
procedure for simultaneous equation models with limited dependent variables to
our model.3 We note that Eqs (1) and (2) both involve discrete binary choices and
hence use a simultaneous probit model. In the first stage, a reduced form model
for each of the two endogenous variables is estimated:
Collateral = 1 X + ␧1 (3)
Creditors = 2 X + ␧2 (4)
where X is the set of all exogenous variables in X1 and X2 . Since

both Collateral

and Creditors are dichotomous variables, we can only estimate 1 /␴1 and 2 /␴2 ,
where ␴21 = Var(␧1 ) and ␴22 = Var(␧2 ). Thus, writing Collateral∗ as Collateral/␴1
and Creditors∗ as Creditors/␴2 we get the estimable structural functions.
␴2 ␤ 1 e1
Collateral∗ = ␣1 , Creditors∗ + X1 + (5)
␴1 ␴1 ␴1
␴1 ␤ 2 e2
Creditors∗ = ␣2 , Collateral∗ + X2 + (6)
␴2 ␴2 ␴2
We first estimate the reduced forms by probit maximum likelihood. Then we
substitute the predicted values of Collateral∗ and Creditors∗ into Eqs (5) and (6)
and estimate the structural equations by the probit maximum likelihood method.
The asymptotic covariance matrices are then derived as per Maddala (1983). Define

␴2 ␤1
␥1 = ␣1 ,
␴1 ␴1

␴2 ␤ 2
␥ 1 = ␣2 ,
␴1 ␴2
and let
␾1 ␾2
a1 = , a2 =
1 (1 − 1 ) 2 (1 − 2 )
∗
2 X
A 1 = ␾1 a 1 , A 2 = ␾2 a 2 , Z =
X
N
1
W1 = A 1 ZZ
N
1
N
1
W2 = A 2 XX
N
1
N
1 ␴2
W3 = A 1 ␣1 ZX
N ␴1
1
N

1
W4 = a 1 a 2 E[(y 1 − 1 )(y 2 − 2 )]XZ
N
1
Then the covariance of N 1/2 (␥ˆ 1 − ␥01 )(where ␥01 is the true value of ␥1 and ␥ˆ 1
is the two-stage estimator) is
W −1 −1 −1 −1 −1
1 [W 1 − W 3 W 2 W 4 − W 4 W 2 W 3 + W 3 W 2 W 3 ]W 1 (7)
The covariance of ␥ˆ 2 is similarly derived as follows.
␾1 ␾2
a1 = , a2 =
1 (1 − 1 ) 2 (1 − 2 )
∗
1 X
A 1 = ␾1 a 1 , A 2 = ␾2 a 2 , Z =
X
N
1
W5 = A 2 ZZ
N
1
N

1
W6 = A 1 XX
N
1
N

1 ␴1
W7 = A 2 ␣2 ZX
N ␴2
1
N
1
W8 = a 1 a 2 E[(y 1 − 1 )(y 2 − 2 )]XZ
N
1
Then the covariance of N 1/2 (␥ˆ 2 − ␥02 ) (where ␥ˆ 02 is the true value of ␥2 and ␥ˆ 2
is the two-stage estimator) is
W −1 −1 −1 −1 −1
5 [W 5 − W 7 W 6 W 8 − W 8 W 6 W 7 + W 7 W 6 W 7 ]W 5 (8)
Following Carey et al. (1998), the fit of the probit model is estimated by
McKelvey and Zavoina’s (1975) pseudo-R2 measure, calculated as in Greene
(1999).
let E[y∗ |y] = yf = ␤ X + ␭

(n − 1)var(yf) (9)
pseudo − R2 =
n + (n − 1)var(yf)
where ␭ is the inverse Mill’s ratio, and y is either the Collateral or Creditors binary
variable for the two models. This 2-stage simultaneous model approach allows us to
jointly analyze the two effects that may be missed by the single-equation approach.4
6. ANALYSIS
6.1. Univariate Analysis
To help in making comparisons, we divide the general statistics from Table 3

into two sub-tables. Table 4 compares firms offering collateral against those not
offering collateral.5
We note that firms not offering collateral are of better quality (BONDRAT),
consistent with Rajan and Winton (1995): in order to provide the private lender with
Table 4. Summary Statistics of Sample Set Separated by Collateralization.

Variable Collateral (N = 42) No Collateral (N = 176)
Mean** Median Std. Dev. Mean** Median Std. Dev.
BONDRAT 14.74 15.00 1.86 10.23 10.00 2.99

CUMPROF 0.0110 0.0413 0.1596 0.1764 0.1489 0.2287
DEBTASST 0.4704 0.4902 0.1657 0.3107 0.2874 0.1443
EBITSALE 0.0700 0.0913 0.2001 0.1768 0.1387 0.1090
FACMATUR 1582 1676 682 1364 1826 709
FACSIZE 224.42 100.00 352.62 606.95 300.00 917.92
FACTA 0.2160 0.1580 0.2524 0.1367 0.0880 0.1381
FACTD 0.5365 0.3272 0.7053 0.6224 0.3372 1.3253
FATA 0.4308 0.4310 0.2075 0.4452 0.3964 0.2309
FIRMVAL 1081 732 1095 8165 4962 14144
LTDTA 0.4215 0.4818 0.1732 0.2563 0.2344 0.1337
MKTBOOK 1.3500 1.2630 0.3968 1.5731 1.4249 0.5111
MKTLEV 0.4673 0.4783 0.1709 0.3013 0.2592 0.1831
OITA 0.0534 0.0681 0.0627 0.0999 0.0937 0.0547
SALES 1770 1025 2530 7415 3551 11707
TOTASST 1167 830 1530 7300 3863 11433
ZSCORE 2.6377 2.4402 1.1501 4.7277 3.5069 4.9762
∗∗ All mean values were found to be statistically significantly different from 0 at the 5% level, except
for RELEPS and CUMPROF, which were not significantly different from 0.
sufficient incentive to monitor and force liquidation when efficient, higher quality
firms decrease collateral. Our finding that lower quality firms are more likely to
offer collateral supports Pozzolo (2002) for Italian firms while contradicting both
Bolton and Scharfstein (1996) and John, Lynch and Puri (2002). Furthermore, those
offering collateral are not as profitable and have higher probabilities of financial
distress in the near future (OITA, ZSCORE). This is also in accordance with
the hypotheses of Rajan and Winton (1995) and Gorton and Kahn (1993), that
firms which expect to face higher probabilities of financial distress provide more
collateral to give lenders greater control as well as the ability to threaten liquidation
of inefficient projects. Finally, we find that firms borrowing from multiple lenders
have greater growth potential (MKTBOOK) than those borrowing from single
lenders. This supports the theory of Gilson and Warner (2000) that high growth
firms put up less collateral to increase managerial discretion and flexibility.
In looking at Table 5, which compares firms using multiple creditors with those
using single creditors, we find that firms using a single lender have statistically
significantly higher leverage (DEBTASST, MKTLEV) in accordance with the
findings of Detraigaiche et al. (2000), Degryse and Ongena (2001), and Harhoff and
Table 5. Summary Statistics of Sample Set Separated by Number of Lenders.

Variable Multiple Lenders (N = 182) Single Lenders (N = 36)
Mean*** Median Std. Dev. Mean*** Median Std. Dev.
BONDRAT 11.17 11.00 3.25 10.75 11.00 3.68

CUMPROF 0.1425 0.1204 0.2313 0.1552 0.1226 0.2026
DEBTASST 0.3313 0.3232 0.1584 0.3929 0.3786 0.1670
EBITTA 0.1438 0.1401 0.0643 0.1344 0.1225 0.0481
FACMATUR 1460.38 1826.00 694.92 1110.91 1096.00 711.93
FACSIZE 552.75 250.00 891.11 434.65 137.50 617.93
FACTA 0.1609 0.1073 0.1713 0.1072 0.0630 0.1481
FACTD 0.6611 0.3736 1.3185 0.3263 0.1586 0.5338
FATA 0.4361 0.3904 0.2266 0.4742 0.4592 0.2245
FIRMVAL 6379 2483 11699 8933 2748 18357
LTDTA 0.2819 0.2510 0.1546 0.3195 0.2980 0.1617
MKTBOOK 1.5590 1.4200 0.5142 1.3839 1.2985 0.3812
MKTLEV 0.3156 0.2676 0.1813 0.4228 0.4306 0.2206
OITA 0.0929 0.0885 0.0619 0.0811 0.0710 0.0413
SALES 6213 2396 10192 6903 2094 13643
TOTASST 5753 2286 9681 7966 2849 14278
ZSCORE 4.5863 3.3300 4.8500 3.0042 2.3063 2.4082
∗∗∗ All mean values were found to be statistically significantly different from 0 at the 1% level.
Korting (1998) that companies maintaining single relationships tend to be more

leveraged. However, we find firms using multiple lenders to be more profitable
with better future prospects than firms using a single lender (OITA, ZSCORE).
This is in contrast to the findings of Degryse and Ongena (2001), Detraigaiche
et al. (2000), Farinha and Santos (2002) and Harhoff and Korting (1998), but in
accordance with the hypotheses of Berlin and Mester (1992) and Chemmanur and
Fulghieri (1994), that firms with greater probability of financial distress, or higher
credit risk firms find the option to renegotiate more valuable and hence borrow from
fewer lenders. This could perhaps be explained by the differences in the financial
systems of the United States and more concentrated markets in Italy, Norway and
Portugal.
In highly concentrated markets, profitable firms may not be able to maintain
multiple banking relationships in a healthy, competitive manner, even if they so
desired. In Norway, for example, two banks have 90% of the market, and firms may
not feel that establishing multiple relationships will adequately mitigate hold-up
problems, especially at the cost of giving up a strong, healthy relationship with a
single lender. However, in the U.S., due to the large diffused market, profitable firms
are able to shop around and maintain multiple relationships easily. Furthermore, it
is the profitable firms that would be most worried about restricting relationships to
only one lender, to limit the lender’s bargaining power and mitigate any information
monopolies (Houston & James, 1996; Rajan, 1992).
We also find that firms using multiple creditors seem to have greater growth
potential (MKTBOOK). This is in line with the predictions of Repullo and Suarez
(1998) and Rajan (1992) and Houston and James (1996) that such firms borrow
from multiple creditors for moral hazard reasons as well as to limit the single
lender’s bargaining power and mitigate any information monopoly. Firms with
multiple creditors are also borrowing more, whether on an absolute basis or on a
relative basis (FACSIZE, FACTA), and are issuing debt with longer maturities.
6.2. Regression Analysis
The results of the simultaneous regressions on the collateralization and the number
of creditors in private debt are presented in Table 6.6 With pseudo-R2 of 70% and
44% respectively, both of which are significant at the 1% level, the regressions fit
reasonably well.
We find that the two decisions are not interrelated since the coefficients of the
interdependence effects are insignificant. The quality and loan purpose proxies
are significant for the collateralization regression in column (1). The positive
coefficient on the quality proxy indicates that firms of lower quality put up more
Table 6. Primary Regression Results.

Characteristic Proxy Collateralization (1) Multiple Creditors (2)
Coefficient p-Value Coefficient p-Value
Intercept −6.8057 0.0003*** 1.5185 0.6083

Interdependence −0.6382 0.4380 0.5809 0.1259
Firm quality BONDRAT 0.4325 0.0009*** −0.0102 0.9423
Leverage MKTLEV −1.5554 0.5327 −3.0098 0.0094***
Probability of distress ZSCORE 0.0091 0.8880 0.0488 0.5302
Growth potential MKTBOOK −0.1365 0.8231 0.0602 0.8833
Loan purpose PURPGRP 0.3041 0.0489** −0.1472 0.3947
Loan type FACGRP 0.4410 0.0665* −0.1614 0.5222
Lead lender FINCOT −1.2397 0.1881
Fixed assets FATA 0.7705 0.2596
Firm size LGSALE 0.2603 0.0309**
Pseudo R2 0.7031 0.0000*** 0.4375 0.0009***
N 218 218
Note: Results shown are those from the second stage of a 2-stage simultaneous equations model, using
corrected asymptotic variances. The variables are as defined in Appendix A. Column 1 is a probit
model where 1 = Collateral, and 0 = No Collateral. Column 2 is also a probit model where
1 = Multiple Creditors, and 0 = Single Creditor.
∗ Coefficients significant at the 10% level.
∗∗ Coefficients significant at the 5% level.
∗∗∗ Coefficients significant at the 1% level.
collateral and firms of higher quality put up less collateral for private debt.7 This
is in support of Rajan and Winton’s (1995) theory that private lenders only take
collateral in the bad state as incentives to monitor and force efficient liquidation.
In contrast, our result here does not support the argument by Bolton and
Scharfstein (1996) and John, Lynch and Puri (2002) that higher quality borrowers
minimize strategic defaults by increasing collateralization. The significant loan
purpose proxy is in support of Gilson and Warner’s (1999) finding that firms that
are looking for managerial freedom and “flexibility” put up less collateral. The
loan purposes labeled “1” and “2” are for general purposes and those labeled “3”
and “4” are those for special purposes like takeovers and acquisitions. Hence,
managers are looking for flexibility, freedom and discretion in everyday affairs,
while offering collateral only in special, risky projects like takeovers. We also find
that collateralization depends on the type of debt issued. Collateral is offered for
term loans, but not for revolvers. Neither the growth nor the probability of distress
proxies are significant in the collateralization regression.
In column (2) of the probit regression analyzing multiple creditors, the leverage
and size proxies are significant. The positive sign on the size proxy is in accordance
with the hypothesis of Repullo and Suarez (1998) that larger firms would prefer
multiple lenders since single lenders would be unable to enforce a credible threat
of liquidation. It is also in support of Detragiache et al.’s (2001) hypothesis that
larger firms maintain multiple creditors to avoid liquidity risk due to the lender and
to ensure a stable supply of credit. The negative sign on the leverage proxy implies
that more highly leveraged firms prefer single creditors to multiple creditors, in
support of the findings of Degryse and Ongena (2001). They state that “highly
leveraged firms choose a single relationship since firms judged fit to carry a lot of
debt do not face credit rationing by their single bank.” Hence this finding is quite
robust to regime changes, and validated in both the Norwegian and U.S. samples.
This is also in support of Johnson (1998) who found that leverage was
statistically and economically significantly higher for firms with bank debt due
to the fact that bank debt attenuates negative effects on leverage of potential asset
substitution problems. However it is inconsistent with Houston and James’ (1996)
finding of a negative relationship between leverage and the proportion of bank
financing. They explain their results as consistent with either economies of scale
in issuing public claims or the notion that bank monitoring is a public good and
firms with more leverage may use the same level of bank debt.
6.3. Alternative Specifications and Robustness Checks
Degryse and Ongena (2001) and Detragiache et al. (2002) find that profitability
impacts the number of bank relationships. Since our analysis focuses on
incremental decisions that may be based on relationship or transaction lending
(Boot & Thakor, 2000), specifically adjusting for profitability should not have any
impact on the robustness of our results. To test this, we run a set of simultaneous
regressions on collateralization and the number of creditors, but adding a fitted
profitability variable (ratio of operating income to total assets) in the second stage
of the model to analyze this effect.8 The results are presented in Table 7.
As expected, our base results are materially unchanged. We re-ran the
regressions using alternative measures for profitability, namely the EBIT to total
assets ratio, and Altman’s (1993) cumulative profitability measure, calculated as
retained earnings divided by total assets. Again, we did not find any difference in
results.
We also evaluated alternative specifications of the base model and did not
find any major difference in results. Since the correlation matrix (Table B.1 in
Appendix B) indicates that some variables are highly and significantly correlated
with others (for example market leverage), we drop each one of the variables
from the model in turn to test for robustness of the results. Two sample results
Table 7. Comparative Regression Results.

Intercept −6.1223 0.0249** 1.5834 0.5957

(Fitted) Profitability OITA −3.5702 0.7123 0.0970 0.9821
Leverage MKTLEV −1.7544 0.5099 −3.0195 0.0125**
Growth potential MKTBOOK −0.1835 0.7709 0.0617 0.8824
Loan type FACGRP 0.3992 0.1301 −0.1627 0.5199
Pseudo R2 0.7033 0.0000*** 0.4378 0. 0016***
N 218 218
corrected asymptotic variances, with a fitted profitability variable inserted in the second stage.
The variables are as defined in Appendix A. Column 1 is a probit model where 1 = Collateral,
and 0 = No Collateral. Column 2 is also a probit model where 1 = Multiple Creditors, and 0 =
Single Creditor.
are included in the appendix (Tables C.1 and C.2, with market leverage and the
market to book ratio dropped, respectively) that show that even when these highly
correlated variables are dropped, the results do not change materially.
7. CONCLUSIONS
This paper empirically assesses theoretical predictions on two aspects of debt
contracts, the number of creditors and use of collateral, employing a current
database of private U.S. corporate debt transactions. We simultaneously model
the two aspects of debt contracts, since there may exist trade-off effects not
captured in non-simultaneous models. Further, we focus on these two aspects
of debt contracts from the borrower’s strategic perspective, and not from a lender’s
pricing perspective.
We do not find support for Bolton and Scharfstein’s (1996) theory that the two
decisions are made jointly to minimize strategic defaults and the loss of value
in liquidity defaults. Rather, we find that the two decisions are not interrelated.
Firms of lower quality put up more collateral and firms of higher quality put up
less collateral possibly because private lenders only take collateral in the bad state
as incentives to monitor and force efficient liquidation (Rajan & Winton, 1995).
Furthermore, those offering collateral also give lenders greater control as well as
the ability to threaten liquidation of inefficient projects (Gorton & Kahn, 1993;
Rajan & Winton, 1995). We also find that firms that are looking for managerial
freedom and “flexibility” put up less collateral (Gilson & Warner, 2000).
Large U.S. firms in our sample with greater probability of financial distress,
or higher credit risk firms find the option to renegotiate more valuable and hence
borrow from fewer lenders (Berlin & Mester, 1992; Chemmanur & Fulghieri,
1994). Due to moral hazard reasons as well as to limit the single lender’s bargaining
power and mitigate any information monopoly, firms with greater growth oppor-
tunities borrow from multiple creditors (Rajan, 1992; Repullo & Suarez, 1998).
Larger firms prefer multiple creditors for the same moral hazard reasons (Repullo &
Suarez, 1998), or to avoid liquidity risk due to the lender and to ensure a stable
supply of credit (Detragiache et al., 2001). Finally, we find that more leveraged
firms choose a single lender, perhaps due to the fact that they do not face credit
rationing from their relationship lender (Degryse & Ongena, 2001), or due to
the attenuation of the negative effects on leverage of potential asset substitution
problems (Johnson, 1998). Overall, larger, less leveraged firms choose multiple
lenders and smaller, more leveraged firms choose a single lender.
We also discover differences and similarities in the behavior of U.S. and
European firms, perhaps due to the different financial systems in which they
operate. More leveraged firms in both environments prefer single lenders to
multiple creditors. However, profitable European firms operating in a highly
concentrated financial market maintain single relationships, while large, profitable
U.S. firms working in a large and diffused financial market, maintain multiple
relationships to mitigate lenders’ informational monopolies (Houston & James,
1996; Rajan, 1992).
Viewed broadly, our results show that the variables identified in prior studies,
firm quality, the probability of financial distress, growth opportunities, the desire
for managerial flexibility, and firm size, play major roles in explaining two
important features of loan contracts in the U.S.
NOTES
1. Other users of the LPC database include Carey, Post and Sharpe (1998), who use the
database to empirically establish the basic difference between bank and non-bank lenders
and borrowers. Angbazo, Mei and Saunders (1998) use the database to analyze credit spreads
on highly leveraged transaction loans. Hubbard et al. (1999) use it in analyzing bank effects
in borrower’s cost of funds. Strahan (1999) uses it to analyze the pricing of borrower risk.
Dennis et al. (2000) use it to analyze the determinants of contract terms in bank revolving
credit agreements.
2. Strahan (1999) compared the DealScan firms with those found in the COMPUSTAT
universe and found that the median DealScan firm was not only larger (by a factor of four)
but also more profitable (almost double) than the median COMPUSTAT firm.
3. See Chapter 8 of Maddala (1983) for more details on the two-stage estimation
procedure.
4. We appreciate the help of William Greene in resolving some econometric issues with
the model.
5. Statistical significance for ratio differences were calculated using the non-parametric
Wilcoxon sum rank test since ratios are not normally distributed and other tests would not
be valid.
6. As suggested by Gujarati (1995) and Greene (1997), since some of the variables are
significant in the regression, and the R2 is not very high (less than 75%), we conclude that
correlation amongst regressors is not a problem. The full correlation matrix is presented in
Appendix B along with some alternate regressions in Appendix C.
7. Recall that COMPUSTAT records S&P bond ratings in an “inverse” fashion: 2 for
AAA and 21 for C.
8. Using the method explained in Section 5, the profitability variable is estimated in the
first stage as an independent variable. The predicted value is then inserted into the second
stage as a fixed value to estimate the rest of the variables.
ACKNOWLEDGMENTS
We thank James Darroch, Vasil Mihov, Kamphol Panyagometh, Pauline Shum,
John Smithin, and participants at the 2000 FMA Annual Meeting for comments
that helped to improve the paper. We would also like to acknowledge the help
of William Greene in resolving econometric issues and Sumon Mazumdar and
Yuxing Yan in accessing the DealScan database.
REFERENCES
Altman, E. I. (1968). Financial ratios. Discriminant analysis and the prediction of corporate bankruptcy.
Journal of Finance.
Altman, E. I. (1993). Corporate financial distress and bankruptcy: A complete guide to predicting and
avoiding distress and profiting from bankruptcy. New York: Wiley.
Angbazo, A., Mei, J., & Saunders, A. (1998). Credit spreads in the market for highly leveraged
transaction loans. Journal of Banking and Finance, 22, 1249–1282.
Berlin, M., & Mester, L. (1992). Debt covenants and renegotiation. Journal of Financial Intermediation,
2, 95–133.
Blume, M. E., Lim, F., & Mackinlay, A. C. (1998). The declining credit quality of U.S. corporate debt:
Myth or reality? Journal of Finance, 53, 1389–1412.
Bolton, M., & Scharfstein, D. (1996). Optimal debt structure and the number of creditors. Journal of
Political Economy, 104(1), 1–25.
Boot, A. W., & Thakor, A. V. (2000). Can relationship banking survive. Journal of Finance, 55,
679–713.
Carey, M., Post, M., & Sharpe, S. A. (1998). Does corporate lending by banks and finance companies
differ? Evidence on specialization in private debt contracting. Journal of Finance, 53, 3.
Chemmanur, T. J., & Fulghieri, P. (1994). Reputation, renegotiation and the choice between bank loans
and publicly traded debt. Review of Financial Studies, 7, 475–506.
D’Auria, C., Foglia, A., & Reedtz, P. M. (1999). Bank interest rates and credit relationships in Italy.
Journal of Banking and Finance, 23, 1067–1093.
Degryse, H., & Ongena, S. (2001). Bank relationships and firm profitability. Financial Management,
30(1), 9–34.
Dennis, S., Nandy, D., & Sharpe, I. G. (2000). The determinants of contract terms in bank revolving
credit agreements. Journal of Financial and Quantitative Analysis, 35, 87–110.
Detragiache, E., Garella, P. G., & Guiso, L. (2000). Multiple versus single banking relationships:
Theory and evidence. Journal of Finance, 55, 1133–1161.
Farinha, L., & Santos, J. A. C. (2002). Switching from single to multiple bank lending relationships:
Determinants and implications. Journal of Financial Intermediation, 11, 124–151.
Gilson, S. C., & Warner, J. B. (2000). Private versus public debt: Evidence from firms that replace bank
loans with junk bonds. Working Paper. Harvard Business School.
Gorton, G., & Kahn, J. (1993). The design of bank loan contracts, collateral, and renegotiation. Working
Paper. National Bureau of Economic Research.
Greene, W. H. (1999). Econometric analysis. Upper Saddle River, NJ: Prentice-Hall.
Gujarati, D. N. (1995). Basic econometrics. McGraw-Hill.
Harhoff, D., & Korting, T. (1998). How many creditors does it take to Tango? Berlin:
Wissenschaftszentrum.
Houston, J., & James, C. (1996). Bank information monopolies and the mix of private and public debt
claims. Journal of Finance, 51, 1863–1890.
Houston, J., & James, C. (2001). Do relationships have limits? Banking relationships, financial
constraints, and investment. Journal of Business, 74, 347–374.
Hubbard, R. G., Kuttner, K. N., & Palia, D. N. (1999). Are there “bank effects” in borrowers’ cost
of funds?: Evidence from a matched sample of borrowers and banks. Working Paper. Federal
Reserve Bank of New York.
John, K., Lynch, A. W., & Puri, M. (2002). Credit ratings, collateral and loan characteristics:
Implications for yield. Journal of Business.
Johnson, S. A. (1998). Effect of bank debt on optimal capital structure. Financial Management, 27(1),
47–56.
Lang, L., Ofek, E., & Stulz, R. (1996). Leverage, investment and firm growth. Journal of Financial
Maddala, G. S. (1983). Limited dependent and qualitative variables in economics. New York, NY:
Mallar, C. D. (1977). The estimation of simultaneous probability models. Econometrica, 45(7),
1717–1722.
McKelvey, R., & Zavoina, W. (1975). A statistical model for the analysis of ordinal level dependent
variables. Journal of Mathematical Sociology, 4, 103–120.
Petersen, M., & Rajan, R. (1994). The benefits of lending relationships: Evidence from small business
data. Journal of Finance, 49, 3–37.
Pozzolo, A. F. (2002). Secured lending and borrowers’ riskiness. Working Paper. Banca d’ Italia.
Rajan, R. (1992). Insiders and outsiders: The choice between informed and arm’s-length debt. Journal
of Finance, 47, 1367–1400.
Rajan, R., & Winton, A. (1995). Covenants and collateral as incentives to monitor. Journal of Finance,
50, 1113–1146.
Repullo, R., & Suarez, J. (1998). Monitoring, liquidation, and security design. Review of Financial
Studies, 11(1), 163–187.
Shockley, R. L., & Thakor, A. V. (1997). Bank loan commitment contracts: Data, theory, and tests.
Journal of Money, Credit, and Banking, 29(4), 517–534.
Strahan, P. E. (1999). Borrower risk and the price and nonprice terms of bank loans. Working Paper.
Federal Reserve Bank of New York.
APPENDIX A
DETAILED FORMULATION OF VARIABLES
Variable Definition (and Data Source)
BONDRAT Standard and Poor’s bond rating; the ratings of AAA, AA
through C are translated to an ordinal scale ranging from 2
to 21 by COMPUSTAT.
CUMPROF Cumulative Profitability = Retained Earnings/Total Assets
(COMPUSTAT)
DEBTASST Total Debt/Total Assets (COMPUSTAT)
EBITSALE EBIT/Total Sales (COMPUSTAT)
EBITTA EBIT/Total Assets (COMPUSTAT)
FACMATUR Loan Maturity in Days (LPC)
FACSIZE Loan Size (LPC)
FACTA Loan Size/Total Assets (COMPUSTAT; LPC)
FACTD Loan Size/Total Debt (COMPUSTAT; LPC)
Loans are categorized into one of four groups depending on
their stated purpose as follows.
PURPGRP (1) General corporate purposes.
(2) Debt repayment/consolidation.
(3) Take over/acquisition.
(4) Other.
The Loan (or facility) type is also categorized into one of
four groups as follows.
FACGRP (1) Short term revolver.
(2) Long term revolver.
(3) Term loan.

(4) Other.
FATA Fixed Assets/Total Assets (COMPUSTAT)
FIRMVAL Firm Value = Market Value of Equity + Total Debt
(COMPUSTAT)
LTDTA Long Term Debt/Total Assets (COMPUSTAT)
MKTBOOK (Market Value of Equity + Total Debt)/(Common Equity +
Total Debt) (COMPUSTAT)
MKTLEV Market Leverage = Total Debt/(Total Debt + Market Value
of Equity) (COMPUSTAT)
OISALE Operating Income/Total Sales (COMPUSTAT)
OITA Operating Income/Total Assets (COMPUSTAT)
RELEPS Earnings Per Share/Total Assets (COMPUSTAT)
SALES Total Sales (COMPUSTAT)
SHAREPR Share Price (COMPUSTAT)
TOTASSET Total Assets COMPUSTAT)
ZSCORE Altman’s (1968, 1993) Z-Score = 3.3 × (EBIT/Total
Assets) + (Sales/Total Assets) + 1.4 × (Retained
Earnings/Total Assets) + 1.2 × (Working Capital/Total
Assets) + 0.6 × (Market Value of Equity/Total Debt)
(COMPUSTAT)
250
APPENDIX B
CORRELATION MATRIX FOR REGRESSION VARIABLES
Table B.1. Correlation Amongst Independent Variables.
MKTLEV MKTBOOK ZSCORE FACGRP PURPGRP FATA LGSALE BONDRAT
MKTLEV 1.0000 −0.6535 −0.5666 0.0231 0.1265 0.1830 −0.1648 0.5183

0.0001*** 0.0001*** 0.7346 0.0623* 0.0067*** 0.0148** 0.0001***
MKTBOOK −0.6535 1.0000 0.5631 0.0396 −0.1659 −0.1612 −0.0033 −0.3187
GORDON S. ROBERTS AND NADEEM A. SIDDIQI

0.0001*** 0.0001*** 0.5610 0.0142** 0.0172*** 0.9618 0.0001***
ZSCORE −0.5666 0.5631 1.0000 0.0501 −0.0912 −0.3397 0.0957 −0.2037
0.0001*** 0.0001*** 0.4615 0.1798 0.0001*** 0.1589 0.0025***
FACGRP 0.0231 0.0396 0.0501 1.0000 –0.0234 0.0665 –0.1010 0.1065
0.7346 0.5610 0.4615 0.7311 0.3285 0.1373 0.1169
PURPGRP 0.1265 −0.1659 −0.0912 −0.0234 1.0000 −0.0869 −0.2159 0.2049
0.0623* 0.0142** 0.1798 0.7311 0.2012 0.0013*** 0.0024***
FATA 0.1830 −0.1612 −0.3397 0.0665 −0.0869 1.0000 −0.0998 0.0287
0.0067*** 0.0172** 0.0001*** 0.3285 0.2012 0.1417 0.6740
LGSALE −0.1648 −0.0033 0.0957 −0.1010 −0.2159 −0.0998 1.0000 −0.6074
0.0148* 0.9618 0.1589 0.1373 0.0013*** 0.1417 0.0001***
BONDRAT 0.5183 −0.3187 −0.2037 0.1065 0.2049 0.0287 −0.6074 1.0000
0.0001*** 0.0001*** 0.0025*** 0.1169 0.0024*** 0.6740 0.0001***
Note: The numbers below the correlation coefficients are p-values indicating significance of the correlation.
APPENDIX C
ROBUSTNESS CHECKS FOR REGRESSION RESULTS
Table C.1. Alternative Regression Results.

Intercept −7.8320 0.0001*** 0.1346 0.9615

Leverage MKTLEV
Growth potential MKTBOOK 0.1519 0.7896 0.5475 0.1464
Loan type FACGRP 0.4676 0.0785* −0.1161 0.6318
Pseudo R2 0.7061 0.0000*** 0.4746 0.0118***
N 218 218
corrected asymptotic variances, with the market leverage variable dropped. The variables are as
defined in Appendix A. Column 1 is a probit model where 1 = Collateral, and 0 = No Collateral.
Column 2 is also a probit model where 1 = Multiple Creditors, and 0 = Single Creditor.
Table C.2. Alternative Regression Results.

Intercept −7.0433 0.0000*** 1.7029 0.5605

Leverage MKTLEV −1.2455 0.5413 −3.0691 0.0038***
Growth potential MKTBOOK
Loan type FACGRP 0.4315 0.0605* −0.1612 0.5247
Pseudo R2 0.7001 0.0000*** 0.4371 0.0004***
N 218 218
corrected asymptotic variances, with the growth potential variable dropped. The variables are as
defined in Appendix A. Column 1 is a probit model where 1 = Collateral, and 0 = No Collateral.
Column 2 is also a probit model where 1 = Multiple Creditors, and 0 = Single Creditor.
AN EXAMINATION OF THE
EFFICIENCY OF SINGLE VS.
MULTIPLE COMMON BOND
CREDIT UNIONS
James D. Tripp, Peppi M. Kenny and Don T. Johnson
ABSTRACT
As of 1982, federal credit unions were allowed to add select employee groups
and thus create institutions with multiple-group common bonds. We examine
the efficiency of single bond and multiple bond federal-chartered credit
unions by using data envelopment analysis (DEA), a non-parametric, linear
programming methodology. Results indicate that multiple bond credit unions
have better pure technical efficiency than single bond credit unions. However,
single bond credit unions appear to be more scale efficient than the multiple
bond credit unions. Our results also indicate that members of multiple bond
credit unions may derive greater wealth gains than members of single bond
credit unions.
1. INTRODUCTION
In 1981, approximately 500 federal-chartered credit unions closed their doors
due to failure or liquidation. The U.S. economy was suffering through one of
its worst recessions since the early 1930s. The number of business failures rose
Research in Finance
ISSN: 0196-3821/doi:10.1016/S0196-3821(04)21011-5
253
254 JAMES D. TRIPP ET AL.
dramatically, particularly in those industries referred to as “smokestack.” Many of

the occupational-based credit unions that served these smokestack industries found
themselves in critical situations with their futures jeopardized. It was at this point, in
April 1982, that the National Credit Union Administration (NCUA) implemented
a change in policy to try to reduce the number of failures that was occurring.
That change was to allow federal credit unions to add select employee groups
(SEGs) to their membership, and thereby create occupational-based credit unions
with multiple-group common bonds. Federal credit unions could then expand their
existing charters to serve different membership groups as long as each group had
its own common bond (Johnson, 1999).
This ruling by the NCUA had an immediate impact on the number of credit union
failures. In 1982, 272 credit unions failed or were liquidated and that number fell
to 90 in 1983. This policy change allowed occupational credit unions to diversify
their membership base and no longer be dependent on a single employer.
By 1997, over 3,600 federal credit unions had changed their charters and
added at least one SEG. The commercial banking industry vigorously opposed
this common bond expansion, fighting it all the way to the U.S. Supreme Court
and winning (National Credit Union Administration vs. First National Bank and
Trust). However, in 1998, Congress passed the Credit Union Membership Access
Act that effectively endorsed this looser form of organizational common bond for
credit unions. A December 1999 survey by the Credit Union National Association
(CUNA) reported that over 40% of all credit unions now served multiple groups
(Johnson, 1999).
The NCUA’s concern regarding the lack of diversification in the customer base of
occupational credit unions mirrored similar concerns that were raised regarding the
savings and loan (S&L) industry in the early 1980s. S&L regulators were concerned
about the inability of savings institutions to diversify their asset portfolios due to
the restrictions on the types of loans that could be made. This created a high
degree of concentration risk for the S&L industry. Similarly, the NCUA felt that
occupational credit unions serving only a single employer had very high levels of
concentration risk.
This notion of greater concentration risk is supported by an empirical study
done by Kohers (1986) and who reported that occupational credit unions serving
sponsors operating in unstable business-cycle environments experienced higher
loan delinquency rates and held higher levels of liquidity. A more recent study
by Frame, Karels and McClatchey (2002) also found evidence of benefits from
diversification in credit union membership. While these studies address the
reduction in risk that accrues to credit unions that expand their membership base,
little research has been undertaken to evaluate the impact of having a multiple
common bond on credit union efficiency.
An Examination of the Efficiency of Single vs. Multiple Common Bond 255
A study by Emmons and Schmid (1999), reported that similar preferences of

members in multiple-group credit unions might allow those credit unions the
opportunity to be more efficient. Additionally, they show graphically that average
operating costs decrease as the number of members increases. However, regression
results indicated a positive relation between multiple membership groups and
the ratio of total operating expenses to total assets, indicating inefficiencies
may exist.
If inefficiencies do exist in multiple bond credit unions, how does this affect
credit union members? In the profit-driven banking industry, bank shareholders
should be financially rewarded as their banks operate more efficiently. However,
in not-for-profit credit unions, members/owners may not be rewarded as efficiency
increases. Hinson and Juras (2002) posit that wealth gains may accrue to members
of credit unions that have narrower spreads between loan rates and deposit
rates.
The purpose of our paper is to examine technical efficiency, both pure technical
and scale, of single bond and multiple bond credit unions over a five-year period
(1998–2002). We utilize data envelopment analysis (DEA) a non-parametric,
linear programming methodology to expose any inefficiencies that might exist.
In addition, we utilize a method similar to Hinson and Juras (2002) to evaluate
whether the type of membership (single or multiple bond) has any impact on wealth
gains that accrue to members.
2. RELEVANT LITERATURE
Over the past twenty years, many researchers have studied the issue of efficiency
in financial institutions. The bulk of this research has focused on the commercial
banking industry. Studies focusing on the banking industry generally conclude
that inefficiencies found are primarily due to the inefficient utilization of resources
(poor management) rather that scale inefficiencies related to inappropriate size (see
Aly et al., 1990; Elysasiani & Mehdian, 1990; Ferrier & Lovell, 1990; Grabowski
et al., 1994; Miller & Noulas, 1996; Rangan et al., 1988).
The efficiency of one type of financial institution, credit unions, has been less
frequently considered. In addition to the Emmons and Schmid (1999) research,
a general study of credit union efficiency was undertaken by Fried, Lovell and
Vanden Eeckaut (1993). They reported an average of 20% productive inefficiency
in the U.S. credit unions they examined. Other credit union studies have focused
on efficiency issues in the Australian market. Worthington (1999, 2000) reported
overall technical inefficiency at 46% in one study and 30% in another. However, it
should be noted that there are substantial differences between U.S. and Australian
credit unions. For example, Australian credit unions pay a federal income tax while
U.S. credit unions do not. In order to compete effectively with other credit unions
and other financial institutions, it is imperative for credit unions to be as efficient
as possible.
3. DATA
Data were collected for the years 1998–2002 from the National Credit Union
Administration’s on-line database (NCUA, 2003) that contains financial and
statistical reports filed by U.S. credit unions. For our study, we chose occupational
credit unions, both single bond (Type 13) and multiple bond (Type 43), whose
primary sponsors operated in the machine manufacturing industry. As mentioned
earlier, credit unions with sponsors operating in the more cyclical business sectors
face more concentration risk and therefore would be more likely to diversify by
adding SEGs to the existing membership. We restricted our sample to credit unions
operating in the machine-manufacturing sector to limit the impact of economic
variability across industries on our results. The resulting data set included 36
single bond credit unions and 58 multiple bond.
In order to test for efficiency, it must first be determined what specifically do
credit unions produce, or what is their output. Two approaches used in previous
studies regarding financial institutions were considered: the intermediation
approach and the production approach. Under the intermediation approach,
financial institutions are considered producers of services that collect deposits and
purchase funds that are intermediated into loans and other assets. Additionally,
deposits are viewed as inputs, along with labor and capital, while measures of
output are defined as the financial institution’s various dollar volumes of earning
assets. Finally, under the intermediation approach, total cost is defined as total
operating costs and interest expense.
Alternatively, under the production approach, financial institutions are viewed
as producers of services related to individual deposit and loan accounts. Both
capital and labor are used to produce these account services. Output is defined as
the number of deposit and loan accounts and total cost is defined as total operating
costs and does not include interest expense. Also, according to Humphrey
(1985), some control factor for differences in the average sizes of accounts across
different-sized financial institutions must be incorporated under this approach.
According to Clark (1988), reasonable arguments have been made for both
approaches to modeling a financial institution’s behavior. Clark also noted that the
empirical results do not appear to be sensitive to the method used in defining costs
and outputs.
Table 1. Descriptive Means of Credit Uniona Variables, 1998–2002.

Single Bond Multiple Bond
Total loans $1,442,771 $18,129,815

Total investments $780,713 $9,335,428
Total assets $2,311,683 $28,875,096
Total shares and deposits $1,917,534 $24,686,060
Total interest expense $64,593 $861,687
Total non-interest expense $72,461 $842,899
Total income $166,766 $2,155,058
a Thesample consists of occupational credit unions whose primary sponsors were in the machine
manufacturing industry.
For purposes of this paper, the intermediation approach is utilized. Using this
method, input and output variables comparable to those used by Miller and
Noulas (1996) in their study of bank performance were selected. The input variables
(in dollars) are:
(1) Total shares and deposits (TSD)
(2) Total interest expense (TIE = interest on borrowed money + dividends on
shares + interest on deposits)
(3) Total non-interest expense (TNIE = total operating expenses less provision
for loan losses).
The outputs variables (in dollars) are:
(1) Total loans (TL = unsecured credit card loans + all other unsecured loans +
new vehicle loans + used vehicle loans + total all other loans to members +
total first mortgage real estate loans + total other real estate loans)
(2) Total investments (TIN)
(3) Total income (TI = interest on loans – interest refunded + income from
investments + income from trading securities + fee income + other operating
income).
Sample statistics are displayed in Table 1.
4. METHODOLOGY
These input and output variables are utilized in a data envelopment analysis model
(DEA) in order to examine efficiency of multiple common bond and single bond
credit unions. The use of DEA for purposes such as this was first introduced by
Charnes, Cooper and Rhodes (CCR) in 1978. DEA models seek to determine
which decision making units (DMUs) under consideration are efficient given the
use of m inputs for generating s outputs. The efficient DMUs out of the set of n
DMUj (where j = 1, . . ., n) will lie on the envelopment surface. Inefficient DMUs
do not lie on the surface. In developing the envelopment surface through solving
mathematical programming models, each DMUj is examined relative to the other
DMUs in the sample. Consequently, the measured efficiency is sample specific.
The first model utilized examines an input-oriented data envelopment frontier
with variable returns to scale as originally developed by Banker, Charnes, and
Cooper (BCC model, 1984). The model uses n DMUj (j = 1, 2, . . ., n) observations
with m inputs represented by xij (i = 1, 2, . . ., m) and s outputs represented by yrj
(r = 1, 2, . . ., s). With variable returns to scale the model can be formulated as
follows (Zhu, 2003):
m s

min␪−␧ s−i + s+
r
i=1 r=1
subject to
n

␭j x ij + s −
i = ␪x io , i = 1, 2, . . . , m;
j=1
n

␭j y rj − s +
r = y ro , r = 1, 2, . . . , s;
j=1
␭j ≥ 0, j = 1, 2, . . . , n
and
n

␭j = 1
j=1
with s − +
i representing the input slack, s r representing the output slack, ??, is a
non-archimedean constant, and λj being nonnegative scalars. This BCC model
provides a measure of pure technical efficiency.
For the constant returns to scale input-oriented data envelopment model as
originally developed
by Charnes, Cooper and Rhodes (CCR model, 1978), the last
condition above, nj=1 ␭j = 1, is dropped. The CCR model provides an overall
technical efficiency measure.
The desired output from the models is an efficiency score. The efficient target
with variable or constant returns to scale is given by (Zhu, 2003):
∗
x̂io = ␪∗ x io − s −
i i = 1, 2, . . . , m
∗
ŷro = y ro + s +
r r = 1, 2, . . . , s
with efficiency scores represented by ␪∗ . In order to apply the above, a two-

stage process of calculations is necessary. First optimal ␪∗ is found through the
maximization of a reduction in inputs. Second, optimizing the slack variables is
considered to move DMUs onto the efficient frontier.
In utilizing the model, efficiency scores represented by ␪∗ are generated for each
DMU. The DMUs with ␪∗ = 1 are those that fall on the efficient frontier. All other
DMUs could improve efficiency by decreasing input levels.
Overall technical efficiency, as given by the CCR model, can be separated into
two components: pure technical efficiency and scale efficiency. By adding the
convexity constraint in the BCC model, variable returns to scale are permitted and
consequently a measure of pure technical efficiency is produced from the BCC
model (Worthington, 1999). A measure of scale efficiency can then be generated by
dividing overall technical efficiency by pure technical efficiency (CCR result/BCC
result) (Mosheim, 2004). Graphical depiction of these efficiency components can
be found in Miller and Noulas (1996).
In order to test for the impact of membership type (single or multiple bond) on
wealth gains to credit union members, we used a technique similar to Hinson and
Juras (2002). These authors posit that federally tax-exempt credit unions’ profits
belong to members/owners and that credit union management should strive to offer
lower loan rates and higher deposit rates to their memberships. While Hinson and
Juras utilize net interest margin as a proxy for wealth gains, we focus on the yield
spread between the share yield (dividends on shares divided by total shares) and
loan yield (interest received on loans less rebates divided by total loans) as our
measure.
5. RESULTS
The causes of inefficiency for a DMU can stem from two sources. The DMU’s
inefficiency may be a result of inefficient operations (as measured by pure technical
efficiency) or be attributed to disadvantageous conditions (as measured by scale
efficiency) related to the environment in which it operates (Cooper et al., 2000).
The overall technical efficiency (OTE), pure technical efficiency (PTE), and
scale efficiency (SE) scores for the 36 single bond (Type 13) and 52 multiple bond
(Type 43) credit unions are displayed in Table 2.
The OTE scores for both the single bond and the multiple bond credit unions
showed relatively high levels of global efficiency under the assumption of constant-
returns-to-scale. The OTE scores for single bond credit unions ranged from a low
Table 2. DEA Mean Efficiency Scores.

Year Bond Pure Overall Scale
Technical Technical Efficiencya
Efficiencya,b Efficiencya,c
2002 Multiple 0.958335 0.896216 0.935713

Single 0.934366* 0.899950 0.964157**
2001 Multiple 0.962876 0.952796 0.989622
Single 0.941904 0.929627* 0.987044
2000 Multiple 0.956571 0.924107 0.965759
Single 0.921376** 0.904476 0.981117**
1999 Multiple 0.951139 0.877354 0.923102
Single 0.944458 0.895255 0.948185*
1998 Multiple 0.968690 0.948237 0.979117
Single 0.960163 0.950947 0.990202**
a Efficiency scores are sample specific to the credit unions within the sample. A higher score indicates
greater efficiency; the credit unions that are most efficient have a score of 1.0.
b A measure of operational efficiency.
c A measure of global efficiency.
∗ Significant at 10%.
∗∗ Significant at 5%.
of 0.895255 in 1999 to a high of 0.950947 in 1998 and averaged 0.916051 for the
five years studied. The multiple bond credit unions’ OTE scores ranged from a low
of 0.877354 in 1999 to a high of 0.952796 in 2001 and averaged 0.919742. These
results indicate that the inefficiency was approximately 8% for both types and
is smaller than reported in other U.S. credit union studies. However, sample and
methodological differences when utilizing DEA make it very difficult to compare
our results to other credit unions studies.
The PTE scores reported in Table 2 provide efficiency measures using a local
measure of scale under the assumption of variable returns to scale. The findings
show that the multiple bond credit unions were more efficient on a pure technical
basis in all five years examined. These differences were significant at the 0.10 level
in 2002 and the 0.05 level in 2000. This finding is indicative of greater managerial
efficiency in the multiple bond credit unions.
However, the scale efficiency (SE) scores displayed in Table 2 highlight an
efficiency advantage for the single bond credit unions. As reported earlier in
Table 1, the average single bond credit union was substantially smaller than
the average multiple bond credit union over the 1998–2002 time period (total
assets of $2.3 million vs. $28.9 million). According to Cooper et al. (2000), this
Table 3. Yield and Yield Spread Means.

Year Type of Bond Loan Yielda (%) Share Yieldb (%) Yield Spreadc (%)
2002 Multiple 8.3501 2.3346 6.0155

Single 10.0782*** 2.3968 7.6814***
2001 Multiple 9.0058 3.3466 5.6590
Single 9.9805** 3.5817 6.3988*
2000 Multiple 8.5101 3.6695 4.8406
Single 9.3825** 4.0961*** 5.2864
1999 Multiple 8.8000 3.5686 5.2314
Single 9.9233** 3.9639** 5.9595
1998 Multiple 9.1589 3.6891 5.4699
Single 9.9720** 3.9270 6.0450
a Interestreceived on loans, less rebates, divided by total loans.
b Dividends on shares divided by total shares.
c Loan Yield minus Share Yield.
∗ Significant at 10%.
∗∗ Significant at 5%.
∗∗∗ Significant at 1%.
may indicate that these smaller credit unions are operating under advantageous
conditions relative to larger multi-bond credit unions. SE results in Table 2 reveal
that single bond credit unions had significantly higher efficiency scores in four of
the five years studied.
Table 3 presents the results regarding differences in member wealth gains
between single bond and multiple bond credit unions over the 1998–2002 time-
period. The yield spread is lower in all five years for the multiple bond credit unions
and significantly so in 2001 (0.10 level) and 2002 (0.01 level). Multiple bond credit
unions had significantly lower loan rates in all five years studied while the single
bond credit unions provide higher share rates. However, as interest rates in the
economy trended lower in 2001 and 2002, there was no significant difference
in the share rates of the two types of credit unions. This finding indicates that
multiple bond credit unions may be improving their members’ financial welfare
by providing better interest rates than single bond credit unions.
6. CONCLUSIONS
Prior credit union research has shown that the creation of multiple bond credit
unions has reduced the amount of concentration risk in the industry. The results of
our study reveal additional benefits of this diversification. As measured by the pure
technical efficiency scores, multiple bond credit unions appear to offer superior
managerial expertise. These larger multiple bond credit unions are quite likely
managed by individuals with higher levels of education and greater experience
resulting in greater operational efficiency. In addition, these multiple bond credit
unions appear to pass along the financial gains of this managerial efficiency to
their members by offering superior rates over their single bond counterparts. The
only negative finding regarding multiple bond credit unions concerns their scale
efficiency. Our results indicate that the larger multi-bond credit unions are less
efficient on a size basis than the smaller single bond credit unions.
From a regulatory perspective, the results of our study should be encouraging
since both single and multiple bond credit unions appear to be operating at relatively
high levels of overall efficiency. Additionally, the finding that multiple bond credit
unions may offer better managerial efficiency and wealth gains to their members
might justify regulators promoting further expansion in this area.
REFERENCES
Aly, H. Y., Grabowski, R., Pasurka, C., & Rangan, N. (1990). Technical, scale and allocative efficiencies
in U.S. banking: An empirical investigation. Review of Economics and Statistics, 72, 211–218.
Banker, R. D., Charnes, A., & Cooper, W. W. (1984). Some models for estimating technical and scale
inefficiencies in data envelopment analysis. Management Science, 30(9), 1078–1092.
Charnes, A., Cooper, W. W., & Rhodes, E. (1978). Measuring efficiency of decisions making units.
European Journal of Operational Research, 2, 429–444.
Clark, J. A. (1988). Economies of scale and scope at depository financial institutions: A review of the
literature. Economic Review, Federal Reserve Bank of Kansas City, September/October, 16–23.
Cooper, W. C., Seiford, L. M., & Tone, K. (2000). Data envelopment analysis: A comprehensive text
with models, applications, references and DEA-solver software. Boston: Kluwer.
Elysasiani, E., & Mehdian, S. M. (1990). A nonparametric approach to measurement of efficiency and
technological change: The case of large U.S. commercial banks. Journal of Financial Services
Research, 4, 157–168.
Emmons, W. R., & Schmid, F. A. (1999). Credit unions and the common bond. The Federal Reserve
Bank of Saint Louis Review, 81(September/October), 41–64.
Ferrier, G. D., & Lovell, C. A. K. (1990). Measuring cost efficiency in banking: Econometric and linear
programming evidence. Journal of Econometrics, 46(October/November), 229–245.
Frame, W. S., Karels, G. V., & McClatchey, C. (2002). The effect of the common bond and membership
expansion on credit union risk. Financial Review, 37(November), 613–636.
Fried, H. O., Lovell, C. A. K., & Vanden Eeckaut, P. (1993). Evaluating the performance of U.S. credit
unions. Journal of Banking & Finance, 17, 251–265.
Grabowski, R., Rangan, N., & Rezvanian, R. (1994). The effect of deregulation on the efficiency of
U.S. banking firms. Journal of Economics and Business, 46, 39–54.
Hinson, Y. L., & Juras, P. E. (2002). Performance evaluation of credit unions: Reaping the benefit of
tax status. Journal of Managerial Issues, 14, 145–161.
Humphrey, D. B. (1985). Cost and scale economies in bank intermediation. In: R. Aspinwall & R.
Eisenbeis (Eds), Handbook for Bank Strategy. New York: Wiley.
Johnson, E. (1999). Multiple-group memberships, http://www.cuna.org/data/newsnow.
Kohers, T. (1986). Parent organization’s activities as a cause for differential credit union operating
characteristics. Southwest Journal of Business and Economics, 4(Fall), 11–21.
Miller, S. M., & Noulas, A. G. (1996). The technical efficiency of large bank production. Journal of
Banking & Finance, 20, 495–509.
Mosheim, R. (2004). The efficiency of credit unions in Hawaii and Puerto Rico.
http://www.ceterisparibus.uprm.edu/articulos/vol3/articulo6.htm.
NCUA (2003). Call report data files under FOIA. http://www.ncua.gov/data/FOIA/foia.html,
Alexandria: NCUA.
Rangan, N., Grabowski, R., Aly, H. Y., & Pasurka, C. (1988). The technical efficiency of U.S. banks.
Economic Letters, 28, 169–175.
Worthington, A. C. (1999). Measuring technical efficiency in Australian credit unions. The Manchester
School, 67(March), 231–248.
Worthington, A. C. (2000). Cost efficiency in Australian non-bank financial institutions: A non-
parametric approach. Accounting and Finance, 40, 75–97.
Zhu, J. (2003). Quantitative models for performance evaluation and benchmarking: Data envelopment
analysis with spreadsheets and DEA Excel solver. Boston: Kluwer.

(Research in Finance) Andrew H. Chen-Research in Finance, Volume 21 - Emerald Group Publishing Limited (2005) PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

(Research in Finance) Andrew H. Chen-Research in Finance, Volume 21 - Emerald Group Publishing Limited (2005) PDF

Uploaded by

Copyright:

Available Formats

CONTENTS

LIST OF CONTRIBUTORS vii

TIME DIVERSIFICATION AND STOCHASTIC DOMINANCE

MINORITY EQUITY INVESTMENTS AND INTER-FIRM

SIZE AND BOOK-TO-MARKET EFFECTS IN THE RETURNS

IMPLIED VOLATILITIES AND AUDITOR REPUTATION: THE

SECONDARY EQUITY OFFERINGS: THE CASE OF

A NEW APPROACH TO TESTING PPP: EVIDENCE FROM

CORRELATION AMONG STOCK MARKETS UNDER

MULTIPLE BANKING AS A COMMITMENT NOT TO

OPPORTUNITY COST AND PRUDENTIALITY: AN

COLLATERALIZATION AND THE NUMBER OF LENDERS

AN EXAMINATION OF THE EFFICIENCY OF SINGLE VS.

Herbert L. Baer Formerly at Policy Research Department, The

Arthur J. Keown R. B. Pamplin College of Business, VPI & SU,

Charles W. Hodges, Haim Levy and James A. Yoder

unimpeachable, requires no distributional assumptions, and takes all the moments

Fig. 1. Return Distributions.

Fig. 2. Cumulative Return Distributions.

Fig. 3. Cumulative Return Distributions – F Preferred to G with Risk Aversion.

for all r and the expected values are:

3. METHODOLOGY AND RESULTS

3.1. Independent Returns

We use simulation to generate sample return distributions for portfolios of small

Table 1. Mean Returns (Independent Returns).

1 0.1865 0.1396 0.0648 0.0384

Table 2. Standard Deviation (Independent Returns).

1 0.3415 0.2033 0.0896 0.0314

Table 3. Skewness (Independent Returns).

1 0.5975 −0.4003 1.4384 0.7029

Table 4. Stochastic Dominance Results with Independent Returns.a

1 Yes Yes Yes Yes

Note: Yes = Membership in the TSD Efficient Set.

3.2. Autocorrelated Returns

Table 5. Mean Returns Under Autocorrelation.

2 0.4115 0.2890 0.1397 0.0804

reported by various researchers.8 This autocorrelation may induce preferences

Table 6. Standard Deviations Under Autocorrelation.

2 0.6034 0.3432 0.1395 0.0675

Table 7. Skewness Under Autocorrelation.

2 0.6411 −0.0591 1.1567 0.8057

Table 8. Stochastic Dominance Results Under Autocorrelation.a

2 Yes Yes Yes Yes

Note: Yes = Membership in the TSD Efficient Set.

1. See Merton and Samuelson (1974) and Samuelson (1990, 1994).

Su Han Chan, John W. Kensinger, Arthur J. Keown

We examine the beneﬁts for ﬁrms participating in collaborations funded via

The use of interfirm collaborations or business alliances has grown dramatically

First, we examine firms that announce strategic alliances or collaborative

significant improvement in their operating performance in the third year following

2. MINORITY EQUITY STAKES AND INTERFIRM

2.1. Pooling Resources and Lowering Transactions Costs

In an equity alliance one or more of the participating firms purchases common

capabilities of smaller, knowledge-intensive firms. A case in point is provided by

2.2. Mitigating Adverse Selection and Information Problems

2.3. Controlling Opportunistic Behavior Via Minority Equity Ownership

3. DATA AND METHODOLOGY

Table 1. Distribution of Announcements of Interfirm Alliances Involving

Type of Cooperative Agreement Number of By Purchasing By Selling

distribution of equity alliances by purpose of the agreement is consistent with

Table 2. Industry Affiliation of Purchasing and Selling Firms.

Panel A: High-technology group

Table 3. Profile of Purchasing and Selling Firms.