Professional Documents
Culture Documents
Olmo Insider Trading
Olmo Insider Trading
1. Introduction
The detection of insider trading is generally considered to be essential to maintaining the
integrity of the financial system and consequently its detection is given a high priority by the
Securities Exchange Commission (SEC) in the US and the Financial Services Authority
(FSA) in the UK. Recent research of Grgoire and Huang (2009) has analyzed some of the
consequences that insider trading and inside information have on the cost of issuing new
equity. In particular, their model indicates that such information can cause the market to
demand a higher premium over the risk-freerate of interest on newly issued equity
1
. As such
financial regulators haveincentive to detect such trading and where appropriate prosecute this
form of market abuse.
There are several alternative approaches in the literature to detect insider trading. From a
statistical point of view, Dubow and Monteiro (2006) develop a measure of market
cleanliness based on detecting abnormal stock returns prior to the release of an announcement
of price sensitive information. The authors implement an extended capital asset pricing model
to capture the dynamics of risky returns and use a definition of abnormal returns as the
residuals of the corresponding time series regression. To detect insider trading, they use
bootstrap techniques to approximate the finite sample distribution of the sequence of
abnormal returns before an unscheduled announcement and compare this distribution against
the magnitude of four-day and two-day cumulative returns taken four days before and one
day after the announcement to see if these observations are in the tails of the bootstrap
1
In a similar veinVo (2008) has shown thatwhen information of this nature is not disclosed
to financial markets in a timely fashion the price of equity is positively correlated with this
information.
distribution. This method is further refined by Monteiro et al (2007) to allow for serial
correlation and conditional heteroscedasticity in the data.
Other methods are employed by regulators, for example, the Korea Exchange employs a
market stock price monitoring model combined with a period stock price monitoring model
to detect abnormal transactions. The market stock price model develops two linear
regressions: one for stock price and the other for trading volume. Both models are used to
detect when prices/volumes deviate from some normal trading range. The period stock price
monitoring model detects variations in cumulative average returns and consists of three
models: a stock price base model, a trading volume base model and a concentration ratio base
model. A more sophisticated model of insider trading is developed by Park and Lee (2010)
who use their model to characterize the time series of stock price returns. To identify insider
transactions using a time series, they assume that information exposed from insider trading at
time t can be determined by a particular mixed strategy AR(1) process which they use to
establish that the return series must follow an ARMA(1,1) process. They develop three
criteria for detecting insider trading and conduct two validation tests.
In this paper, we argue that a natural methodology to detect possible insider trading is to look
at unexpected changes in the idiosyncratic component of capital asset pricing models. In
particular, we claim that the occurrence of insider trading leading to abnormal price
movements will be potentially reflected in sudden shifts in the mean of asset pricing
equilibrium models. Thus, we identify insider trading with a structural change in the intercept
of an extended asset pricing model that, as for the previous authors, includes lags of the
idiosyncratic return and of returns on the market portfolio. However, in contrast to the work
of Dubow and Monteiro (2006) and Monteiro et al (2007) we take a novel view on the
4
statistical detection of insider trading. Using their theoretical extended capital asset pricing
model (ECAPM) for pricing the risk premium on a risky asset, we develop a powerful new
statistical test called the consistent timing structural break test (CTSB) which is designed to
detect insider trading prior to the announcement of price sensitive information and,
importantly, we are able to place an approximate timing on the insider trading.
The main advantage of our proposed test is that it can detect possible insider trading in the
run up to the release of price sensitive information compared to the traditional CUSUM test
which proves to be poor at picking up potential insider trading in the lead up to a price
sensitive announcement. Also, compared to the bootstrap technique previously discussed, our
method is more straightforward and less data-dependent since the asymptotic theory of the
test is well known, implying that critical values can be tabulated. As such, by means of
simple critical values it can alert regulators of potential cases of insider trading. The
application of our method to a rich data set covering 370 price sensitive announcements
relating to FTSE 350 companies shows that there is evidence of abnormal returns and
potential insider trading for 92 companies comprising the index. As a by product, we also
show that the majority of suspected insider trading takes place in the 25 days prior to the
release of market sensitive information.
The paper is structured as follows. In Section 2 we discuss how to identify the occurrence of
abnormal returns using an asset pricing model in equilibrium. In Section 3 we develop our
novel test statistic for a structural break in the intercept and relate it to potential insider
trading detection for an extended capital asset pricing model (ECAPM). Section 4 shows via
a Monte Carlo simulation that the statistical power of our CTSB test statistic is significantly
greater than that of the traditional CUSUM test. Section 5 compares the performance of our
5
test at picking up potential insider trading cases against a CUSUM test using a confidential
data set relating to the release of 370 price sensitive announcements on FTSE 350 companies
supplied by the FSA. Finally, Section 6 concludes.
2. Detecting insider trading via abnormal returns in equilibrium pricing models
Insider trading can be detected directly by looking at unusual trading volumes in the equities
or derivative markets or alternatively by looking for unusual share price movements prior to a
price sensitive announcement or a combination of the two. In their study, Dubow and
Monteiro (2006) use the pricing approach and examine two kinds of announcements, trading
statements by company issuers and public takeover announcements by companies to which
takeover code applies. The methodology developed by Dubow and Monteiro (2006) and
Monteiro et al (2007) for detecting informed price movements and insider trading defines
abnormal stock returns as:
1
[ ] (1)
it it t it it
AR R E R c
= =
where AR
it
is the abnormal returns, R
it
refers to returns on stock i at time t and
1
[ ]
t it
E R
is the
expected return at time t conditional on information up to time t-1. The expected return can
be modelled using time series or cross-section methods. We follow the literature on asset
pricing in equilibrium and describe the dynamics of the expected return via an Extended
Capital Asset Pricing Model (ECAPM) similar in spirit to the above authors model but based
on excess returns given by equation (2):
* * * *
1 1 2 1 3 1
[ ] (2)
t it mt mt it
E R R R R | | |
= + +
6
where
*
it
R denotes returns in excess of the risk-free asset ,
t
R
mt
R refers to the market return at
time t, and
1
| ,
2
| and
3
| are the slope parameters corresponding to the different risk factors.
The use of lagged variables in the model acts as a filter for the presence of serial dependence
in the data.
We argue that a natural methodology to detect possible insider trading is to look for a
positive/negative shift in the mean of the abnormal return sequence caused by a change in the
intercept of the above ECAPM. Theoretically, under a normal functioning of the market and
the standard assumptions on market efficiency the risk premium on a risky asset can be
modelled by the standard CAPM. We assume an extended version of it given by equation (2).
Thus, if there is a positive (negative) price sensitive information that is only revealed to a
reduced group of market participants the price of the stock is bound to increase (decrease) by
a smaller amount than it would be in the case that the information was publicly available.
Figure 1 illustrates the difference for a positive price shock.
Figure 1 The effects of insider trading
Price
Quantity
D
1
D
2
D
3
Q
1
S
1
P
1
P
2
P
3
Note that unless the company raises more equity its supply curve is fixed at S
1
. On the other
hand, the demand curve is downward sloping. The D
1
curve is the demand curve of
uninformed investors. The equilibrium price determined by P
1
corresponds to no insider
trading and defines the return on that asset. In equilibrium the return on the asset can be
expressed by equation (2). If there is a small group of informed traders that decide to trade on
the asset then the new demand curve shifts to the right to D
2
implying a new equilibrium
price P
2
which is higher than P
1
but smaller than the equilibrium price P
3
that would prevail if
the inside information were fully publicly available corresponding to demand curve D
3
. The
return implied by the difference between P
2
and the initial price P
1
, denoted R
ins
, is the sum of
the return produced by trade from uninformed investors, denoted R
uni
plus an extra
quantityo due to private information and given by the difference between P
2
and P
1
. In
equilibrium, the risk premium on the asset required by uninformed investors is only affected
by the correlation with the market portfolio. Mathematically,
* * * * * *
, , 1 , 1 1 2 1 1 3 1 , 1
[ ] [ ] [ ] [ ] (3)
ins t uni t t uni t t mt t mt t uni t
R R and E R E R E R E R o | | |
= + = + +
This result implies that the observed risk premium on the asset affected by insider trading is
* * * * *
1 , 1 , 1 1 2 1 1 3 1 , 1
[ ] [ ] [ ] [ ] [ ] (4)
t ins t t uni t t mt t mt t uni t
E R E R E R E R E R o o | | |
= + = + + +
in contrast to the no insider trading case, where the risk premium is
* * * * *
1 , 1 , 1 1 2 1 1 3 1 , 1
[ ] [ ] [ ] [ ] [ ]. (5)
t ins t t uni t t mt t mt t uni t
E R E R E R E R E R | | |
= = + +
The existence of potential insider trading or abnormal price movements can therefore be
detected by unexpected changes in the asset pricing formula in equilibrium. A similar
technique is widely used in the mutual fund industry to uncover assets and portfolios
outperforming the market. The difference in our case is that we aim to detect changes in the
value of the intercept and occurrence of excess returns prior to the announcement of price
8
and which after applying a suitable supremum functional provides a consistent estimator of
the timing of the change. Further, this new test can accommodate the presence of weight
functions that can be tuned to have more power against specific alternatives, as for example
in early/late detection. We begin by illustrating the U-statistic type process considered by
Gombay et al (1996). These authors consider the following setting:
Let {Y
1
, Y
2
, ... , Y
T
} for T 2; 3;...., be a sequence of independent and identically distributed
observations. The interest is in testing for the presence of at most one change in variance at a
distinct but unknown time in the process
*
, 1 * (8)
, *
t
t
t
t t
Y
t t T
oc
o c
+ s s
=
+ < s
where is the mean of the process, o and o
*
are positive constants and the errors
t
c are
independent and identically distributed, with
2
[ ] 0, [ ] 1
t t
E E c c = = ,
4
[ ] ,
t
E c < and
*
t is the
timing of the change in the intercept. Assuming thato =
*
o then the no change in variance
null hypothesis can be formulated as:
H
0
: * t T >
versus the at-most-one change in variance alternative:
H
A
: 1 * . t T s <
To test the null hypothesis Gombay et al (1996) use the change in mean framework to
develop a statistic suited to testing for at most one change in the variance. Their process,
reproduced below, compares two estimators of the variance before and after the change:
11
[ ]
(1) 1/ 2 2 2
1 1
1 1
( ) (1 ) ( ) ( ) 0 1 (9)
T T
T t t
t t T
M T Y Y
T T T
t
t
t t t t
t
= = +
= s <
`
)
where
(1)
T
M is a U-statistic type process, t is the fraction of the sample where the change in
the variance of the process {Y
t
} occurs and [.] denotes the integer part. One estimator is
fashioned from the first Tt observations and then compared to the estimator constructed
from the last ( 1) T T t + observations. After some simple algebra, the above process can be
re-expressed as:
[ ]
(1) 1/ 2 2 2
1 1
( ) ( ) ( ) 0 1 (10)
T T
T t t
t t
M T Y Y
t
t t t
= =
= s <
`
)
Gombay et al (1996) substitute the sample mean
_
T Y
of the series
t
Y for ,
the population
mean, to arrive at:
[ ]
_ _
(1)
1/ 2 2 2
1 1
( ) ( ) ( ) 0 1 (11)
T T
T T
t T t
t t
M T Y Y Y Y
t
t t t
= =
= s <
`
)
This methodology can be extended to a mean process, (X
t
), that depends on a vector of
explanatory variables X
t
. Let Y
t
be
*
* *
( ) , 1 (12)
( ) ,
t t
t
t t
X t t
Y
X t t T
oc
o c
+ s s
=
+ < <
Further, Gombay et al (1996) explore the use of weight functions to improve the statistical
power of related tests to detect changes in the parameters produced at specific subsamples of
the evaluation period. In particular, they study the following family of weight functions to the
process captured in equation (10):
( , ) {( (1 )) , 0 1/ 2} (13) q
v
t v t t v = s s
12
=
+ + < s
where
*
Y=R ,
t it
* * *
t 1 1
X (R R R ) '
mt mt it
= ,
t
defined as in (7),
(1) (2)
o o = and the error term
satisfies conditions detailed after equation (7).
The null and alternative hypotheses are:
O
H : * t T >
versus the at-most-one change in intercept alternative
A
H : 1 * t T s <
The task here is to construct a test to detect such deviations and which is robust to the
presence of conditional heteroscedasticity in the data. The U-statistic type process proposed
by Olmo and Pouliot (2008) is:
( )
[ ]
( 2) 1/ 2 ' '
1 1
( ) ( ) (15)
T T
T t T t T
t t
M T Y X Y X
t
t | t |
= =
=
`
)
Using this process a test statistic for H
0
and a consistent estimator of * t under the alternative
hypothesis can be fashioned. In this context, we are concerned with how large this process
13
can be for 0<t <1. If there is in fact a change in the intercept, the value of the supremum of
the above process should be large and leads to the following test statistic for a one time
change in the intercept, denoted now as a consistent timing structural break test (CTSB):
( 2)
0 1
( )
sup . (16)
( , )
T
M
q
t
t
t v
< <
This statistic makes use of the family of weight functions introduced by Gombay et al (1996)
to increase the power of the test against early and late detection. It does, however, depend on
unknown slope parameters of the ECAPM which must be estimated in order for the test to be
implemented. When these slope parameters | are replaced with a consistent estimator
| ,
we arrive at the following process
( 2)
0 1
( )
sup . (17)
( , )
T M
q
t
t
t v
< <
This substitution does not affect the asymptotic distribution which still is
0 1
| ( ) |
sup
( , )
B
q
t
t
t v
< <
with
( )
( , )
B
q
t
t v
a weighted Brownian bridge process characterized by the Brownian bridge ( ) B t .
Further, following Gombay et al (1996) and Antoch et al (1995), a consistent estimator of t*
is
*
, t defined by:
( 2) ( 2)
*
1
Theorems 1 and 2 of Antoch et al (1995) detail the consistency and the limiting distribution
of the above estimator using the weight function q(t/T ,). The critical values of the CTSB
test are obtained from the supremum of the weighted Brownian bridge process introduced
above. Gombay et al (1996) detail the limiting distribution for v = 0.5, the asymptotic
distribution has an extreme value distribution from which critical values can be generated by
most standard statistical packages; for v values between 0 and 0.5 see Olmo and Pouliot
(2008, 2011).
The test statistic set out in equation (16) can be made robust to changes in the slope
parameter by calculating the statistic using the available data and then estimating
*
t as in
equation (18). This
*
t and then
estimating (15) again but using data corresponding to t =
*
( )
2
1
(19)
r
r
t
t K
W w
o
= +
=
where
( ) r
W is the sum of the first r recursive residuals
2
t
w , and o
is the OLS estimate of the
standard deviation. The test statistic is:
1
( )
1
max
1
1 2
1
(20)
k r T
r
W
T K
CUSUM test
r K
T K
+ < s
=
+
The null hypothesis of parameter constancy is rejected whenever this test statistic exceeds
some critical value obtained from the distribution of
0 1
sup | ( ) | B
t
t
< <
, where ( ) B t is a
Brownian bridge. Tabulated values of this distribution can be obtained from Orasch and
Pouliot (2004) (cf. Table I). Here, K, refers to the number of parameters in the linear
regression model under estimation. The estimator of the timing of the break
*
t is detailed by
the following formula:
1
( ) ( )
1 1 *
min :
max
1 1
1 2 1 2
1 1
(21)
k r T
r r
W W
T K T K
CUSUM t r
r K r K
T K T K
+ < s
= =
`
+ +
)
The comparison of the standard CUSUM against the test given by equation (17) is done via a
Monte-Carlo simulation for the following model:
2
For more information on recursive residuals we refer the reader to Brown et al (1975).
16
(1) '
, 1 * (22)
(2) '
, *
X t t
t t
Y
t
X t t T
t t
o | oc
o | oc
+ + s s
=
+ + < s
where is constant and the error terms satisfy the conditions given after equation (7).
For the Monte Carlo exercise considered here, the number of slope parameters K was set to 1
in equation (21). As a result of this restriction, the linear regression model considered for this
exercise consisted of an intercept and one slope parameter | . Both the
parameters | and
(1)
o were set to 1 while
(2)
o was set to values 1.25, 1.5, 1.75, 2. Table 1
details the simulated nominal coverage probability of both the CTSB and CUSUM test of
Brown et al (1975) under the null hypothesis of no change in the intercept. As the
significance level was set at the usual 5% level the simulated nominal coverage should
approximate this value. From the table, we observe that both the CTSB and CUSUM
statistics perform well in terms of nominal coverage.
Table 1 Nominal coverage
T=75 T=100 T=125
CTSB 0.087 0.079 0.067
CUSUM 0.050 0.029 0.040
The second part of our Monte Carlo experiment consists of comparing the simulated power of
the two tests to detect a one-time change in the intercept for different percentage changes.
This comparison allows a realistic assessment of the ability of the CTSB statistic to detect a
one-time change in the intercept and in particular its ability to pick up the timing of the
change in the intercept. Table 2 tabulates the empirical power for the model set out in
17
equation (22) using the same changes in intercept and sample sizes discussed above. The
change in the intercept varies from 25% to 100% in the simulation, while the fraction of the
sample before the break occurs has been set at 5% (early detection), 50% (middle detection)
and 90% (late detection). When it comes to early detection, the CTSB performs better than
the CUSUM particularly as the magnitude of the break becomes larger. When it comes to
middle detection, the CTSB is generally speaking much better than the CUSUM test. Further,
the estimate of * t , where
*
* / , t T t =
based on the CUSUM statistic ranged from a minimum
of 0.5 (25% change in intercept and T = 75) to a maximum of 0.923 (100% change in
intercept and T = 125). This emphasises the inconsistency of the CUSUM estimator for the
change fraction,t , which should be near 0.5 for a one-time change in intercept that occurs in
the middle of the sample. We see that the CTSB estimator of * t had smaller variability and
was much closer to 0.5, especially for moderate to larger sample sizes; the estimator was
0.491 when T=125 and there was a 100% change in intercept. For a change late in the
sample, the CTSB has higher empirical power than the CUSUM test across sample sizes and
models explored under the alternative hypothesis. In particular, for a 100% change in
intercept and T = 125 the power of the CTSB statistic is 0.574 and that of the CUSUM test is
0.077. Also, the estimator of * t based on the CTSB statistic estimates the break fraction to
be 0.766 which is much closer to the true value of t (0.9) than the CUSUM that estimated
the break fraction to be 0.476. The latter result is consistent with the lack of power of the
CUSUM method.
Overall, this small simulation experiment clearly shows the outperformance of the CTSB
method to detect structural breaks in the intercept of linear regression models compared to the
standard CUSUM technique. The following section illustrates these results with a financial
application to detect insider trading from return data on companies trading on FTSE 350.
18
5. Empirical application to FTSE 350 companies return series
We now use our CTSB estimator to detect potential insider trading activity. Our confidential
data set consists of 370 price sensitive announcements relating to FTSE 350 companies
supplied by the Financial Services Authority with 251 return observations per company per
announcement and standardization of the announcement on the 250th day
3
. We only have
information concerning the timing of the announcement but not on the nature of this
announcement or the name of the company under study. The number of announcements (370)
implies that for some companies there is more than one announcement.
Table 3 shows equation (6) parameter estimates for five of the 370 series in the available
sample. The results show that the ECAPM model performs well at explaining the return
series. The residuals of the different series are well behaved. Unreported results show that the
coefficients corresponding to the conditional volatility process in equation (7) are not
statistically significant. The results for the rest of series under study are similar to those
reported in Table 3.
3
We are very grateful to the FSA for supplying us with this dataset. The dataset covers 370
price sensitive announcements on FTSE 350 companies relating to the period 2000 to 2006.
In order to maintain complete confidentiality, the series does not name any of the companies,
the nature of the price sensitive announcement which could be due to a reporting of profits, a
takeover bid, a profit warning or some other price sensitive announcement.
19
rejections and see their statistical reliability we have also computed 95% confidence intervals
using the bootstrap method described in Antoch et al (1995). The results suggest that from
these 38 potential cases there are 4 cases in which the confidence interval indicates that the
rejection is right before the announcement. For the remaining 34 cases the uncertainty in the
abnormal returns makes the precise timing of the rejection inconclusive
4
. Table 4 summarizes
these results for the 4 significant cases.
Figure 2 The CUSUM versus CTSB for detecting insider trading, N=250
4
Bootstrap confidence intervals cannot be computed for the CUSUM of Brown et al since
the method does not satisfy the assumptions in Antoch et al (1995). This represents another
advantage of using statistical tests based on equation (15).
22
Figure 3 The empirical effects of different values of the tuning parameter, N=250
Some of these timings, especially those corresponding to early detection of potential insider
trading, can be due to considering the entire trading year for our analysis
5
. To check the
robustness of our results for smaller sample sizes we have repeated our method for different
sample sizes corresponding to six and three months before the announcement. The exercise is
more challenging due to the lower number of observations in the sample. Nevertheless, as
shown in our results reported in Figures 4 and 5 for sample sizes 125 (6 months of the trading
year) and 63 (3 months of the trading year) the CTSB method is still able to detect structural
breaks for relatively small sample sizes.
5
Since time of the announcement can be any day during the trading year, the early detections
in the 250 day trading year are unlikely to be January effect observations. This is further
confirmed in our 6 month and 3 month tests, where we have removed the first 125 and 187
observations and still find significant cases of early detection.
24
The results of these two experiments confirm our previous findings on the existence for some
companies of potential insider trading right before the occurrence of relevant announcements.
For the CTSB test, both sample sizes of 125 and 63 observations yield results similar to those
reported for the sample size of 250. We observe a slight overall decrease in the number of
abnormal returns as the sample size decreases. Interestingly, for the change in the sample size
from 250 to 125 this decrease occurs in late detection while for sample sizes 125 to 63 the fall
is in early detection. The decrease in detection of potential insider trading cases can be due to
a fall in statistical power of the test as the sample size decreases or a true decay due to there
being less actual breaks when the data set is decreased. Concerning the CUSUM test, for a
sample size of 125, the test still shows some power for early detection but continues to fail at
late detection. For the sample size of 63, the CUSUM test became infeasible due to the lack
of meaningful observations necessary to invert the relevant matrix to compute the recursive
residuals.
Figure 4 The CUSUM versus CTSB for detecting insider trading, N=125
25
volume
6
. Finally, the GARCH structure that we have adopted enables us to distinguish cases
of potential insider trading even in periods of changing market volatility.
We have shown that the presence of potential insider trading can be detected by running
structural break tests for the intercept of an extended capital asset pricing model. Our
procedure which requires a change in the intercept of the regression model yields a number of
possible abnormal price movements meriting investigation for insider trading. More
importantly, our test statistic enables us to check for potential insider trading over an
extended trading range rather than be limited to say an evaluation of trading volumes in the
five days run up to an announcement. Our test statistic based upon a U-statistic type process
has a considerable advantage over the CUSUM test since it has more power to detect changes
that occur over the entire evaluation period. Furthermore, the CTSB has far more statistical
power and accuracy in detecting the timing of a structural break than the CUSUM test.
Another advantage of using the CTSB test is that it proves to be quite robust and effective
with smaller sample sizes.
Further research could include the detection of timing for more than one break in the
intercept during the evaluation period and also combining breaks in the price data with
information from trading volumes as well as information from changes in the volatility
process.
6
Insider traders may well use derivatives markets such as call and put options so as to
maximise the leverage on their trades. Such derivatives activity will affect the price of the
underlying shares. It is for this reason that looking at volumes of trades on a share is an
unreliable guide to insider trading activity and why regulators increasing look at abnormal
price movements.
28
References
Andrews, D. W. K., 1993. Tests for parameter instability and structural change with unknown
change point. Econometrica 61, 821-859.
Andrews, D. W. K., Ploberger, W., 1994. Optimal tests when a nuisance parameter is present
only under the alternative. Econometrica 62, 1383-1414.
Antoch, J., Huskov, M., Veraverbeke, N., 1995. Change-point problem and bootstrap.
Journal of Non-parametric Statistics 5, 123-144.
Brown, R., Durbin, J., Evans, J., 1975. Techniques for testing the constancy of regression
coefficients over time. Journal of the Royal Statistical Society Series B, 37, 149-192.
Chow, G.C., 1960. Tests of quality between sets of coefficients in two linear regression
models. Econometrica 28, 591-605.
Dubow, B., Monteiro, M., 2006. Measuring market cleanliness. Financial Services Authority
Occasional Paper 23.
Dufour, J.M., 1988. Recursive stability analysis of linear regression relationships. Journal of
Econometrics 19, 31-76.
Ferger, D., 2001. Analysis of change-point estimators under the null hypothesis. Bernoulli 7,
487-506.
Gombay E., Horvth L., Huskov, M., 1996. Estimators and tests for change in variances.
Statistics & Decisions 14, 145-159.
Gregoire, P., Huang, H., 2009. Informed trading, noise trading and the cost of equity.
International Review of Economics & Finance 17, 13-32.
Maddala, G. S., 1999. Unit Roots, Cointegration and Structural Change. Cambridge
University Press, New York.
McCabe, B. P. M., Harrison, M. J., 1980. Testing the constancy of regression relationships
over time using least squares residuals. Applied Statistics 29, 142-148.
Monteiro, M., Zaman, Q., Leitterstorf, S., 2007. Updated measure of market cleanliness.
Financial Services Authority Occasional Paper 25.
Olmo, J., Pouliot, W., 2008. U-Statistics type tests for structural breaks in linear regression
models. Department of Economics, City University London Discussion Paper Series, 08/15.
Olmo, J., Pouliot, W., 2011. Early detection techniques for market risk failure. Studies in
Nonlinear Dynamics & Econometrics, forthcoming.
Orasch, M., Pouliot, W., 2004. Tabulating sup-norm used in change-point analysis. Journal of
Statistical Computation and Simulation 74, 249-276.
29
Park, S., Lee, J., 2010. Detecting insider trading: The theory and validation in Korea
Exchange. Journal of Banking and Finance 34, 2110-2120.
Vo, M.T., 2008. Strategic trading when some investors receive information before others.
International Review of Economics and Finance 17, 319-332.
Worsley, K. J., 1983. The power of likelihood ratio and cumulative sum tests for a change in
a binomial probability. Biometrika 70, 455-464.