Time-Variation in Expected Returns Author(s): Jennifer Conrad and Gautam Kaul Source: The Journal of Business, Vol.

61, No. 4 (Oct., 1988), pp. 409-425 Published by: The University of Chicago Press Stable URL: http://www.jstor.org/stable/2352789 . Accessed: 10/08/2011 01:02
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at . http://www.jstor.org/page/info/about/policies/terms.jsp JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.

The University of Chicago Press is collaborating with JSTOR to digitize, preserve and extend access to The Journal of Business.

http://www.jstor.org

Jennifer Conrad
University of North Carolina

Gautam Kaul
University of Michigan

Time-Variation in Expected Returns*

I. Introduction Theoretical models of asset pricing put few, if any, restrictionson the behavior of expected returnsover time. However, in the implementation of most tests of marketefficiency and/ora particular marketequilibriummodel, expected returns are assumed to remain constant over some period of time. Given market efficiency, such an assumption is unrealistic in light of recent evidence that security returnscan (to some extent) be predicted(see, e.g., Fama and Schwert 1977; Fama 1981; Keim and Stambaugh 1986; Fama and French 1987, 1988;and Kaul 1987). In this article, we attempt to characterize the stochastic nature of expected returns. Specifically, we assume market efficiency and test whether expected returns are constant relative to a particular alternative hypothesis. Based on the findings of recent empirical papers, we
* We are grateful to Victor Bernard, Michael Bradley, Mark Flannery, Mustafa Gultekin, Stanley Kon, Albert Madansky,GregoryNiehaus, Jay Ritter,SarabjeetSeth, Nejat Seyhun, and especially Eugene Fama for helpful comments and suggestions. We would like to thank the finance workshopparticipants the Universityof Michiganand the at Universityof North Carolina.We also appreciatethe invaluable researchassistance of M. Nimalendran,the secretarial assistanceof DianneHaftandPattiLamparter, the editoand rial suggestionsof KristinKaul. Partialfinancialsupportfor this articlewas providedby the School of Business Administration,Universityof Michigan.

This article characterizes the stochastic behaviorof expected returnson common stock. We assume market efficiency and postulate an autoregressive process for conditional expected returns.We use weekly returnsof 10 size-basedportfolios over the 1962-85 period and find that (1) the variationthrough time in expected returns is well characterized by a stationary autoregresfirst-order sive process; (2) the extractedexpected returns explain a substantialproportion(up to 26%)of the variance in realizedreturns,and the magnitudeof this proportionhas a monotonic (inverse)relation with size; (3) the degree of variationin expected returnsalso changes systematically over time; and (4) the forecasts subsumethe in information other potentialpredictorvariables.

0021-9398/88/6104-0003$01 .50

(Journal of Business, 1988,vol. 61, no. 4) ? 1988by The Universityof Chicago.All rightsreserved. 409

410

Journal of Business

hypothesize that expected returns follow an autoregressiveprocess. We use weekly returns of 10 size-based portfolios over the 1962-85 biases through period and attemptto eliminatemarketmicrostructure careful sample selection. We use a Kalmanfilter technique proposed by Ansley (1980)to extract expected returnsand findthat constancy is stronglyrejectedfor all portfolios. Movementsin expected returnsare well characterizedby a parsimonious (stationary)first-orderautoregressive model. The most strikingaspect of our results is that the variationthrough time in short-horizonexpected returnsis a relatively large fraction of returnvariances. Moreover, there is a monotonicrelationbetween the size rankings of the portfolios and the relative time-variationin expected returns: variationin expected returns explains 26% of return variance for the smallest portfolio, and this proportiondrops systematically to 1%for the largest portfolio. The significanttime-variation, and its relationto size, is found duringeach of the 5-year subperiods, even when we separately allow for a Januarydummy. There is also strong evidence of systematic changes in the relative variationin expected returns across subperiods. Finally, to gauge the economic content of the forecasts, we test whether they are informativewith respect to other relevant ex ante information. Among others, we consider two predeterminedvariables-the treasury bill rate and the lagged return on an equalweighted market portfolio. Although forecasts based on the autoregressive model rely solely on each portfolio's past returns, they subsume the informationin such variables. Section II describes the model; the empiricalanalysis is presented in Section III. Section IV contains a brief summary. II. The Model The choice of a particularmodel for the time-varyingbehavior of expected returns is, by nature, somewhat arbitrary.Ideally, an equilibrium model should specify both the stochastic process for, and the underlyingeconomic determinantsof, expected returns.However, existing asset-pricingtheories do not specify any particulara priori restrictions on the variationthrough time in expected returns. In fact, more recent models of market equilibrium(e.g., Merton 1973;Lucas 1978; Breeden 1979; Cox, Ingersoll, and Ross 1985) do not rule out even negative expected returns. Underthese circumstances,it becomes essential to empiricallycharacterize the stochastic nature of expected returns. Such an exercise will help us understandthe behaviorof expected returnsover time and across different assets and perhaps also help identify the economic determinantsof security returns. The basic approachin this article is

Time-Variation

411

one of stochastic-parameter estimation, which permits both the identificationand estimation of the expected returnprocess.' We model movements in expected returns as a first-orderautoregressive process. The choice of an autoregressive process is based largely on recent empiricalevidence of the predictabilityof stock returns. A majorityof the predeterminedvariablesthat have significant correlationswith realized stock returns are themselves characterized by highly autocorrelatedbehavior. For example, all three of the (ex ante) predictive variablesused by Keim and Stambaugh(1986)exhibit while higherorderautolarge and positive first-order autocorrelations, correlations decay across longer lags. Similarly, Fama and French (1987) use dividend yield to forecast stock returns, and this variable also exhibits a similar autocorrelationstructure (see also Campbell 1987;and Ferson, Kandel, and Stambaugh1987). Since movements in expected returnspresumablyreflect variation throughtime in such forecast variables, they themselves may be characterizedby an autoregressiveprocess. We, therefore,write our model as
Rt=

Et- 1 (Rt) +

Et,

(1)

and
Et -I(Rt) =4Et-2(Rt-1)
+ Ut-1,

(2)

where Rt = realizedreturnon a particularsecurity over period t 1 to t;

Et-j(Rt j+ 1) = expected return for a security over period t - j to t - j
Et -

+ I as of period t - j; independently and identically distributed(i.i.d.) N(0,

cr2);

Ut - i.i.d. N(0, cr-2);and

+ '

1.

We use a Kalman filter technique proposed by Ansley (1980) to extract the expected returns. Before discussing the estimationprocedure, we present a simple representationof the model. If expected returns are representedby the process in equation (2), we can write realized returns as
Rt = Rti
+
Et E6tl

Ut-19

(3)

1. OhIsonand Rosenberg(1982)employ such an approachto estimatethe stochastic common stock index. behaviorof the systematicrisk of the equal-weighted

412

Journalof Business

by which, in turn, implies that realizedreturnscan be characterized an ARMA (1,1) process of the form
Rt = 4)Rt+ at - Oat-1. (4)

In the invertedform we can then write the conditionalexpected return as

E,-I

(4)t

)=

O)Rt-I

0(
+

O)R2

+ 0(4-

(5)
-

0)Rt-3

Invertibility of the process requires that the sum E1. OO(4)

0) con-

verge and, hence, that Io0< 1. In the pure, moving average form, we can write
Et-1(Rt)=

(4) - 0)at-1 + 4(4+ 4)2(4) _

)at-2
....

(6)

O)at3 +

For the process to be stationary,the sum of weights, I'o 4)( _ 0), < must converge, and hence 14)1 1 (see Box and Jenkins 1970). Therefore, conditional expected returns are an exponentially weightedsum of past returns,where the weights add up to less than 1.0 (eq. [5]). Conversely, conditionalexpected returnscan be expressed as a weighted sum of all past shocks, where the weights given to past shocks decline exponentially (eq. [6]). In other words, if expected
returns follow a stationary process, a shock at t 1, at, has a

progressivelysmallereffect on future expected returns.2 A special case of this model is one in which the autoregressiveparameter,4),is constrainedto be equal to 1.0. Under this specification, randomwalk process, which expected returnsfollow a (nonstationary) leads to realized returnsfollowing a randomwalk-plus-noiseprocess. This, in turn, implies that a shock at t - 1, at-1, has two parts, a permanentand a temporaryone. We can interpret(1 - 0)at, l as the permanentcontributionof a shock to realizedreturnsin the sense that it affects all futureexpected returnsby this amount.Correspondingly, 1 Oat_ can be viewed as the temporarycontribution.We estimate both the constrainedand unconstrainedversions of the model.
2. Rosenberg(1973)first developed such a convergentparameter model, as opposed to the random-walk model, in which expected returnshave no tendency to converge. Using his notation,we can rewriteour model in the followingform:
R. = Et-I(Rt) + Et-I(R,)
Et,

= (1 - 8)R + BEt-2(Rt-1)

+ U,-1,

where8 is the convergenceparameter, Kis the population"norm"towardwhichthe and expected returnprocess converges. We can then rewriterealizedreturnsas R.= (1 - 8)R + B~R,1 + t - b6t-l + Ut-1, which is an ARMA (1,1) process.

Time-Variation

413

III. A.

The Evidence Data Description

We use the Centerfor Researchin SecurityPrices (CRSP)daily master file to calculate weekly portfolioreturns.The choice of a weekly sampling interval is largely a compromise between the relatively few monthlyobservationsand the potentialbiases associatedwith nontrading, the bid-ask effect, and so on, in daily data. We use portfolio returns, rather than individual security returns, because it is much more difficultto extract the (expected return)signalfrom noisy weekly returnsof a single security (see Lo and MacKinlay 1987). At the end of each year, stocks are sortedinto 10portfoliosbased on marketvalue (numberof sharesoutstandingtimes price per share).For each week (Wednesdayclose-Wednesday close) of the followingyear, 1-week simple returnsare calculatedfor securities that actuallytraded on both Wednesdays. For the July 1962-December 1962 period we form portfolios based on marketvalues as of December 1962. Weekly holding-periodsecurity returnsin each portfolio are equally weighted to form 10 series of portfolio returns, which are then continuously compounded.We also constructan equal-weighted"marketportfolio" returnusing all our sample securities. Hence, we have a total of 1,226 weekly, continuously compoundedreturnsof 10 size-based portfolios and one marketportfolio from July 1962to December 1985.
B. Autocorrelations

1. The WeeklyEvidence. Table 1 shows the summarystatistics for the weekly portfolio returns for the 1962-85 period. The first-order autocorrelationsare large and significant,and the higher-orderautocorrelations(though significant)decay across longer lags. The returns on the equal-weightedmarketportfolio(EWMR)exhibit similarpersistence in autocorrelations.Fama (1965) and Lo and MacKinlay(1987) also find positive autocorrelationsin short-horizonreturns. (We replicate all our results using excess returns, i.e., returnsin excess of the weekly risk-freereturn.The results are virtuallyidentical.) The autocorrelationstructuredisplays a consistent patternas we go fromthe smallestportfolio(R1) to the largest(R10):the magnitudeand persistence of the autocorrelationsdecline monotonically. However, autocorrelationsremain significantfor all but the largest higher-order portfolio (which exhibits only first-orderautocorrelation). Finally, the first-orderautocorrelationsof weekly changes in reautocorreturns, AR, are all significantlynegative, while higher-order lations are close to zero. This behavior of sample autocorrelationsis consistent with slowly moving expected returns.
2. Market Microstructure Biases. The positive autocorrelations in

portfolio returns may also be consistent with the presence of market

414 TABLE 1

Journal of Business Summary Statistics of Weekly Returns of 10 Equal-weighted Portfolios of New York and American Stock Exchange Common Stocks, Formed by Decile Rankings of Market Value of Equity Outstanding at the End of the Previous Year, 1962-85 (1,226 weeks) Al .41 .35 .31 .30 .29 .27 .24 .22 .18 .09 .28 -.35
- .38
P2

Variable (x)

P3

P4

P5

P6

s(x)*

RI R2 R3 R4 R5 R6 R7 R8 R9
RIO

EWMRt AR1t
AR2

.24 .20 .17 .15 .13 .12 .11 .09 .07 - .00 .14 -.08
-.07

.16 .14 .12 .10 .08 .07 .06 .07 .07 .04 .10 -.01
-.01

.10 .10 .08 .05 .05 .04 .04 .03 .02 .00 .06 .02
.04

.01 .01 .00 .00 .00 - .01 - .00 - .00 -.02 - .02 - .00 -.10
-.10

.04 .04 .03 .01 .00 - .00 - .01 - .01 -.03 - .03 .01 .07
.07

.7231 .4650 .3949 .3515 .2896 .2921 .2747 .2466 .2370 .1919 .3358 .0031
- .0009

3.158 2.736 2.638 2.528 2.481 2.395 2.256 2.153 2.076 1.978 2.295 3.435
3.110

AR3 AR4 ARS AR6 AR7 AR8 AR9 AR 10

AEWMR

-.40 -.39 -.39 -.40 -.41 -.42 -.43 -.45

- .40

-.07 -.07 -.07 -.07 -.06 -.06 -.07 -.08

- .07

-.01 -.00 -.01 -.01 -.01 .01 .03 .05

.00

.03 -.00 .01 .01 .01 -.00 -.00 -.01

.01

-.08 -.04 -.04 -.04 -.02 -.01 -.02 -.00

- .05

.06 .02 .00 -.00 -.01 -.02 -.03 -.05

.01

-.0026 -.0012 -.0025 -.0019 -.0016 -.0007 -.0007 -.0002

- .0009

3.098 2.996 2.958 2.893 2.777 2.692 2.658 2.673

2.758

NOTE.-R1-RIO are the continuously compounded weekly returns on 10 size-based portfolios in ascending order from smallest to largest. Values x and s(x) are the sample mean and standard deviation of the variable, and p, is the sample autocorrelation at lag t. Under the hypothesis that the true autocorrelations are zero, SEs of the estimated autocorrelations are about .03. RI is the smallest portfolio; RIO the largest. * The returns are rates of return per week in decimal fraction units x 102. t EWMR is the equal-weighted market portfolio return. t The operator A denotes first differences.

microstructure biases caused by infrequentor nonsynchronoustrading and/or bid-ask effects (see, e.g., Fisher 1966; Scholes and Williams 1977;and Cohen et al. 1983). in The basic intuitionfor positive autocorrelation portfolioreturnsis as follows: an infrequentlytraded (or small firm) security's observed returnat time t may contain a component of the price adjustmentto news released at t - 1, which will tend to be positively related to the observed returnsof frequentlytraded(largefirm)securitiesat t - 1. In other words, good (bad) news is incorporatedimmediately in frequently traded firms' returns and with a lag in the returns of infrequently traded firms, thus inducing artificialserial correlationin the
returns on a portfolio of securities.

However, our samplingprocedureminimizesthe possibility of such structureof biases, and the evidence indicates that the autocorrelation returnsis unlikely to have been caused by infrequenttradingin small-

Time-Variation

415

firm stocks. First, we use weekly portfolio returns and exclude all securitiesthat do not trade on the adjacentWednesdaysof a particular week. Hence, any remainingnonsynchronoustradingbias would be restrictedto differences in tradingintervals of less than one day, and, relative to weekly returns, such a bias should be small. Moreover, we constructweekly portfolio returnsby first calculatingweekly security returns and then forming equally weighted portfolios. This method minimizesthe bias in some earlierstudies that use arithmeticaverages of returnswithin the review period. See Blume and Stambaugh(1983) and Roll (1983)for a discussion of this issue. Second, the magnitudeof the first-orderautocorrelations(see table 1) is unlikely to have been caused by microstructurebiases. Lo and MacKinlay (1987) model the nontradingphenomenon as a binomial process and show that even if (on average)50%of stocks on the New York and AmericanStock Exchanges do not trade each day, the theoof retical first-orderweekly autocorrelation portfolioreturnswould be about 17%.Furthermore,Working(1960)shows that averagingtempoin autocorrelation the average/ rallyordereddata can induce first-order autocorrelationapproaches a maxindex, but the magnitudeof this imum of 25%as the numberof items in the average/indexapproaches infinity.Our evidence indicates that most of the portfolioreturnshave first-orderserial correlations greater than such extreme theoretical are autocorrelations not peculiarto values. Hence, the high first-order just infrequentlytraded (small) firms. Third, table 1 also shows persistence in the serial correlation at longer lags for a majorityof the portfolios. Such a pattern in weekly returns, which should exist if our model for expected returns is a of robustrepresentation the true process, is againunlikelyto have been biases. Finally, our analysis(see Sec. caused by marketmicrostructure of the extracted expected returns suggests that III) of the properties such biases are unlikely causes of the observed autocorrelationstructure in weekly returns.
C. The Model Estimates

procemaximum-likelihood We use the Kalmanfilterand a Marquardt dure to estimate the parametersof our model. The first-orderautoregressive parameter,A, is allowed to vary across portfolios.The results are reportedin table 2. A stationaryautoregressiveprocess for expected returnsappearsto be well specified. The estimates of 4 are significantlydifferentfrom both 0.0 and 1.0 for all portfolios, and the residualsbehave like white noise. Moreover, there is a monotonic relationbetween the estimates of 4 and size: the magnitude of + declines systematically from the smallest- to the largest-sized portfolios. We also estimate the model with a dummy variable for the first week of January;the results are

416 TABLE 2

Journal of Business Weekly Estimates of the Parameters of the Model in Which Expected Returns Follow a Stationary AR(1) Process
R=

Et1(R,)

,,

(1) (2)

E-,_(Rt) = 4E,2(R,-l) where

+ Ut-1,

4? 1
With January Dummy (D)
()

Overall Period: July 1962-December 1985 (N = 1,226) Without January Dummy Portfolio 1 2 3 4 5 6 7 8 9 10
*

s(e)t .02872 .02549 .02501 .02409 .02373 .02304 .02187 .02101 .02042 .01971

Ait

s(E)

P .004 .003 .003 .004 .002 .001 - .000 - .001 - .003 .001

.589 (.054) .584 (.061) .555 (.071) .489 (.080) .430 (.086) .403 (.094) .394 (.104) .217 (.028) .180 (.028) .087 (.028)

- .001 - .002 - .003 - .004 - .006 - .006 - .006 - .010 - .007 .001

.649 (.037) .617 (.046) .601 (.056) .566 (.064) .523 (.072) .498 (.081) .489 (.091) .463 (.105) .409 (.132) .086 (.028)

.068 (.005) .047 (.005) .035 (.005) .029 (.005) .023 (.005) .017 (.005) .014 (.004) .010 (.004) .007 (.004) .001 (.004)

.02707 .02462 .02452 .02376 .02352 .02292 .02179 .02095 .02039 .01971

NOTE.-Portfoliosize rangesfrom smallest size (1) to largestsize (10). Numbersin parentheses realizedreturnfor week t; below estimatedcoefficientsare SEs. R. = continuouslycompounded u, I) E,1j(R,1+ = expectedreturnover week t - j to t - j + 1as of week t - i; 't - N(O o2); - N(O,
u2).

* 4 = the estimatedautoregressive in parameter eq. (2). error. t s(e) = residualstandard = residualautocorrelation lag 1. Under the hypothesisthat the true autocorrelations are at t 01 are zero, the SEs of the residualautocorrelations about .03. and ? 8 = estimatedcoefficienton the dummyvariable(D), whereD = 1 for firstweek in January, D = 0 V other weeks.

shown in table 2. The dummy variable is significantfor the smaller eight market-valueportfolios, which indicates that seasonality is not phenomenon. Moreover, the estimates of 4)maintain just a small-firm their systematic relationwith size. Evidence of a stationaryexpected-returnprocess is consistent with the results of Fama and French (1987), who find positive autocorrelations in expected returns,which are documentedin theirregressionsof long-horizonreturnson dividend/priceratios. The patternof the coefficients in these regressionsfor differentholdingperiods is suggestive of a mean-revertingexpected-return process. However, the time-

Time-Variation

417

variationin short-horizonexpected returnsis quite differentfrom the slowly decaying variationin long-horizonexpected returnsfound by Fama and French. Specifically, our estimates of + (which are always less than 0.65) indicate rapidly decaying time-variationin expected returns:the effects of an expected-returnshock are largely dissipated after a month. The conclusion that expected returnsfollow a stationaryautoregressive process is supportedwhen we estimate the model in which 4 is restrictedto be equal to 1.0 (or, equivalently,expected returnsfollow a random walk). The results, not reported, show strong evidence of movements in expected returnsacross all portfolios and a systematic relationbetween the magnitudeof such variationand the size rankings of the portfolios. However, the constrained specification is clearly misspecified. First, the residuals from the model exhibit significant first-orderautocorrelationsfor all portfolios. Second, although the extracted expected returns from the constrained model contain significant information about subsequentrealizedreturns,they have no marginalexplanatorypower in regressionsof realized returnson forecasts obtainedfrom both the constrainedand unconstrainedmodels. Brown, Kleidon, and Marsh(1983), hereafterBKM, also use a Kalman filter technique to analyze the variationthroughtime in the size effect. Specifically, they test for nonstationarity the excess returns in of 10 size-based portfolios, where excess returnsare definedas deviations from returnspredictedby the Sharpe-Lintner version of the capital asset-pricingmodel (CAPM).Theirresultsindicatethatthe assumption of constancy of excess returns is most seriously violated for the smallest and largest portfolios. Our objective is quite differentfrom the BKM study since we are concerned more generally in the stochastic process for time-varying expected returns.Unlike BKM, our approachhas the advantageof not imposing, and thereforenot requiringthe estimationof, any particular asset-pricingmodel with.its attendantassumptions. We only require the assumptionof market efficiency, which enables us to extract expected returnsby a linear projectionof realized returnson past information(in particular,past portfolioreturns).Consequently,we do not entertainthe notion of excess returns.
D. Statistical Properties of Weekly Expected-Return Forecasts

Table 3 shows regressions of weekly portfolio returns,Rt, on the extractedexpected returns(with the Januarydummy)fromthe stationary autoregressivemodel, ERAR, 1. The criteriafor a good extractedexpected-return series are: (1) conditionalunbiasedness,that is, an intercept close to zero and a slope coefficient close to 1.0, and (2) serially uncorrelatedresiduals. The extracted expected returnsare condition-

418 TABLE 3

Journal of Business Estimates of Regressions of Weekly Realized Portfolio Returns on Extracted Expected Returns, July 1962-December 1985
R, = ot +

PERAR,-1

Portfolio 1 2 3 4 5 6 7 8 9 10 -.000003 (.00088) - .000001 (.00083) .000009 (.00084) .000001 (.00084) .000005 (.00080) .000004 (.00081) .000007 (.00081) .000004 (.00077) .000010 (.00078) - .000000 (.00104) 1.000 (.086) .998 (.099) .994 (.112) .997 (.116) .995 (.115) .995 (.125) .994 (.140) .995 (.152) .992 (.184) 1.000 (.421)

R2* .265 .191 .136 .117 .101 .084 .067 .053 .035 .007

Olt .004 .004 .005 .005 .003 .002 .001 .000 - .002 .001

below the estimatedregressioncoefficientsare SEs basedon NOTE.-Thenumbersin parentheses correction.Portfoliosare listed fromsmallestsize (1) to White's(1980)consistentheteroskedasticity largestsize (10). for realizedreturnfor week t; ERARtl = expectedreturn week t compounded R. = continuously AR(1) as of week t - 1 extractedfrom the model, in which expectedreturnsfollow a (stationary) term. dummy);a, = randomdisturbance process (withJanuary * l = (adjusted) coefficientof determination. at are t 01 = residualautocorrelation lag 1. Under the hypothesisthat the true autocorrelations are zero, the SEs of the residualautocorrelations about .03.

ally unbiased for all portfolios: the slope coefficients are all close to 1.0, while the intercepts are close to zero.3 Moreover, the regression residualsbehave like white noise. The most striking aspect of the regressions in table 2 is the large proportionof variance of short-horizonreturnsexplainedby variation through time in expected returns. The (adjusted)R2 values of up to 26% are much larger than those typically found in tests that use monthly data. Moreover, there is a monotonic relation between the
3. The heteroskedasticitytest of White (1980) produces chi-squarestatistics well above conventionalsignificancelevels for most portfolioregressions.We thereforereport the more conservative standarderrorsbased on the heteroskedasticity-consistent errorsin the extractedexpectedreturns,ERARt 1, method.We also have measurement since they are estimates. However, most methods for the computationof corrected problem)assume homoskedasticerrors(see, errors(for the errors-in-variables standard e.g., Murphyand Topel 1985). Since heteroskedasticityis potentiallya more serious problemin stock returnregressions,we choose to reportthe alreadyconservativehetstandarderrors. eroskedasticity-consistent

Time-Variation TABLE 4 Estimates of Variances of Realized Returns, Expected Returns, and Unexpected Returns on 10 Equal-weighted Portfolios Rt* .099751 .074866 .069596 .063921 .061530 .057340 .050903 .046336 .043099 .039138 ERAR,- 1 .026500 .014374 .009615 .007562 .006300 .004887 .003498 .002502 .001574 .000298 URARt .073291 .060592 .060136 .056435 .055320 .052519 .047464 .043871 .041560 .038847

419

Portfolio 1 2 3 4 5 6 7 8 9 10

NOTE.-R. = continuously compounded realized return for week t; 1 ERARt- = estimatedexpectedreturnfor week t as of week t - 1 extracted fromthe modelin whichexpectedreturnsfollow a (stationary) AR(1)process (withJanuary dummy);URARt = unexpectedreturn. * The estimatedvariancesare reportedin decimalfractionunits x 102.

size rankings of the portfolios and the relative time-variationin expected returns: variation in expected returns explains 26% of return variance for the smallest portfolio, and this proportiondrops systematically to about 1%for the largest portfolio. Hence, not only is the variance of realized returnsfor small firms systematicallyhigherthan for large firms (see table 1), but the relative variance of expected returns is higher as well. This, in turn, enables the signal extraction technique to detect the significantvariationin the expected returnsof (especially) the smallerfirms. The variabilityof movements in,expected returns across portfolios can perhaps be better gauged by comparingthe absolute variances of the expected-returncomponentsof differentportfolioreturns.In table 4 we present the estimated variances of realized, expected, and unexpected returns for all 10 portfolios. The absolute variationin the expected returns of the smallest portfolio is about 90 times the corresponding variation of the largest portfolio. Moreover, the estimated varianceof expected returnsdeclines systematicallyas the size of the portfolio increases. Finally, to analyze the extent of variationin expected returnsover differenttime periods, we reestimate the model over five subperiods (results not reported). The significantaspect of the results is the evidence of systematic changes in the relative variationin expected returns over time, across portfolios. Typically, the proportionof variance in realizedreturnsexplainedby the extractedforecasts duringthe seventies is about twice the proportionexplained in the sixties and eighties. This result can serve as a useful guidein identifyingthe important economic determinantsof security returns.

420 TABLE 5

Journal of Business Estimates of Regressions of Weekly Realized Portfolio Returns on Predetermined Variables and Extracted Expected Returns, July 1962-December 1985

R. = oa + PI RF,-1 + P2 EWMR,-1 + P3 ERARt1

Portfolio 1: & .0107 (.0018) .0054 (.0008) .0012 (.0016) .0086 (.0016) .0032 (.0008) .0019 (.0015) .0075 (.0015) .0026 (.0007) .0020 (.0016) .0071 (.0015) .0023 (.0007) .0021 (.0016) .0062 (.0014) .0018 (.0007) .0019 (.0015) .0063 (.0014) .0020 (.0007) .0017 (.0015) .0056 (.0013) .0019 (.0006) .0015 (.0014) .0050 (.0013)

+ Ti Pit .002 .159 .273 .004 .132 .195 .003 .108 .140 .004 .096 .119 .003 .082 .101 .004 .067 .084 .41 .02 - .06 .35 - .02 -.05 .31 - .04 -.04 .29 - .03 -.02 .28 - .01 -.00 .26 .00 .01 .24 .00 .01 .21

1
- 3.263 (1.478) ... - .666 (1.309) - 3.708 (1.319) ... - 1.382 (1.222) - 3.335 (1.298)
...

2 ... .550 (.052) .166 (.059) ... .434 (.045) .121 (.060) ... .378 (.044) .119 (.067) ... .342 (.042) .097 (.066) ... .310 (.041) .040 (.075) ... .270 (.039) -.027 (.085) ... .231 (.037) -.007 (.075) ...

P3 ... ... .849 (.108) ... ... .825 (.139) ... ... .769 (.176) ... ... .775 (.190) ... ... .883 (.217) ...
...

(i) (ii) (iii)

2:

(i) (ii)
(iii) 3:

(i)
(ii) (iii) 4:

- 1.371 (1.234) -3.362 (1.294) ... -1.553 (1.250) -3.148 (1.287)

...

(i)
(ii) (iii) 5:

(i)
(ii) (iii) 6:

- 1.551 (1.257) - 3.178 (1.270) ... -1.640 (1.247) -2.631 (1.192) ... -1.440 (1.164) -2.370 (1.150)

(i) (ii)
(iii) 7:

1.063 (.272) ... ... 1.005 (.287)

...

(i)
(ii) (iii) 8:

.003 .055 .067 .002

(i)

Time-Variation TABLE 5 (Continued)

i
322* 2

421

Portfolio & PO~~fO~~iO & (ii) (iii) 9: (i) (ii) (iii) 10: (i) (ii) (iii)
.0046 .0018

2*

Pi1t

ii

...

.191

...

.041

.01

(.0006) .0010 (.0014) (.0012)

.0019

-1.331 (1.131)
-2.086

(.035) -.060 (.082)

... .143

1.235 (.360)
... ...

.053

.01

.002 .024

.18 .03

(1.122)
...

(.0006) .0005 (.0014)

.0039

-1.291 (1.107)
-1.836

(.033) -.098 (.074)

... .067

1.498 (.412)
... ...

.036

.01 .08 .02 .00

.001 .005

(.0012)
.0017

(1.131)
...

(.0005) .0018 (.0018)

- 1.583 (1.123)

(.031) .002 (.060)

.947 (.825)

.007

below the estimatedregressioncoefficientsare SEs basedon NOTE.-Thenumbersin parentheses correction.Portfoliosare listedfromsmallestsize (1) to White's(1980)consistentheteroskedasticity rate for realizedreturn week t; RF, - = the risk-free compounded largestsize (10).R, = continuously calculatedas the continuouslycompoundedreturnon a 1-weektreasurybill for week t, knownat marketportfolio returnon the equal-weighted week t - 1; EWMR, l = continuouslycompounded fromthe for for week t - 1; ERAR,_I = estimatedexpectedreturn week t as of week t - 1 extracted AR(1) process (with Januarydummy); , = model in which expected returnsfollow a (stationary) term. randomdisturbance * K2 = (adjusted) coefficientof determination. are at t 01 = residualautocorrelation lag 1. Under the hypothesisthat the true autocorrelations are zero, the SEs of the residualautocorrelations about .03.

E.

The Information Content of the Extracted Expected Returns

We now consider the informationcontent of the expected returnsextracted using the stationary model, ERAR, 1, with respect to other relevant ex ante information. It is by no means necessary that the expected returns, conditioned solely on each portfolio's past returns, should also incorporate other information.However, we can get an idea of the economic content of the forecasts by testing whether the in information other relevantex ante variablesis alreadyimpoundedin them. The choice of the predeterminedvariablesis largelydictatedby the findings in other papers and data availabilityconsiderations. Among other variables, we consider the nominal risk-free rate (Fama and and the lagged returnon the equalSchwert 1977;and Shanken 1987)4
4. Since new treasurybills are introducedevery Thursdayin the Wall Street Journal quotations,the estimated risk-freerate series is for an 8-day instrument.We assume skip-daysettlementand continuouslycompoundthe returns.The priceused to calculate the return is an average of the bid-and-askprices derived from the quoted bid/ask discountrates.

422

Journal of Business

weightedmarketportfolio(Gibbonsand Ferson 1985).Moreover,since the Januaryeffect is statistically significant,we use the extracted expected returnswhich include the Januarydummy. Regressions (i) and (ii) in table 5 show estimates of regressions of realized portfolio returns on predeterminedvariables for the overall period. The risk-freerate, RF, -1, is significantlynegatively related to the returnsof all portfolios (see regressions [i]). There is residualautocorrelationin all the regressions,which impliesthat the standarderrors are biased. However, if the informationset incorporatedin expected returnscontains elements (other than RF,_ 1) that are autocorrelated, then we would expect such autocorrelatedresiduals. (and The laggedreturnon the market,EWMR, 1,is also significantly positively) related to realized returnsof all the portfolios (see regressions [ii]). These regressions have lower standarderrors than regressions (i), and the residuals behave like white noise. Estimates of regressions, in which all of the predictorvariablesand the extracted expected returnsfrom the stationarymodel, ERAR,_1, are includedsimultaneously,are also shown in table 5 (see regressions [iii]). The expected returnforecasts are informativewith respect to the two ex ante variables. Specifically, the forecasts contain all the informationin the risk-freerate that is pertinentto expected returns;RF,_ I has no marginalexplanatorypower in any of the portfolioregressions. The lagged returnon the market, EWMR, again has no marginal -l, explanatorypower in the larger portfolio (4-10) regressions and has only some explanatorypower in the regressionsof portfolios 1-3. The fact that the explanatory power of the lagged market return drops substantiallydue to the inclusion of the extracted expected returns suggests that the autoregressive model cannot merely be capturing autocorrelationin portfolio returns due to infrequenttradingin small stocks. In fact, the remainingexplanatorypower of EWMR,-I in the small-sizedportfolio regressions may be a reasonableestimate of the infrequenttradingeffect (see simulationresults of Lo and MacKinlay [1987]). More important, the lagged market return does not significantly alter the explanatorypower of the extracted forecasts. We also use other ex ante informationto predictreturns(resultsnot reported). Based on the findings of Campbell(1987), we use various term-structurevariables-for example, lagged weekly returns on 1-, 3-, 6-, and 12-monthtreasury bills (in excess of the weekly risk-free rate). Following Shanken (1987), we use a measure of interest rate volatility in an attemptto capture shifts in the investmentopportunity set (see Merton 1973).Finally, we use the laggedreturnon the smallest portfolio. However, these variables either do not have systematically significantrelationswith portfolio returnsor are renderedinsignificant
in regressions that also include ERAR, -l.

Hence, forecasts extractedfrom the stationaryautoregressivemodel (whichrely solely on each portfolio'spast returns)tend to subsumethe

Time-Variation

423

in information other potentialpredictorvariables.This result is important because it supports the robustness of our parsimonious,autoregressive model for expected returns. Regressions (iii) in table 5 also about indicatethat the predictorvariablesdo containsome information expected returnsbecause the standarderrorsof the coefficients of the extractedexpected returnsare largerthan in table 3 (especiallyfor the large-firm portfolios).
IV. Summary and Conclusions

In this article, we attempt to characterizethe stochastic behavior of expected returns on common stock. We assume market efficiency, and, based on recent empiricalevidence, we postulate an autoregressive process for conditionalexpected returns. We use weekly returnsof 10 size-based portfolios over the 1962-85 period and extract expected returnsfor all portfolios. In implementing our signal-extractionmethodology, we attempt to eliminate market microstructurebiases through careful sample selection. Our major findings are: (1) the time-variationin expected returns is well characterized by a parsimonious (stationary) first-order autoregressive expected returns model; (2) the variationthroughtime in short-horizon is a relatively large fraction of returnvariance, and this fractionhas a monotonic (inverse) relation with the size rankingsof the portfolios; (3) there is strongevidence of systematic changes in the relativevariation in expected returnsacross subperiods;and (4) althoughthe foremodel rely solely on each portfolio'spast recasts based on a stationary variables. predictor in turns,they subsumethe information otherpotential characterizes the nature of exThis article, therefore, explicitly pected return movements. Our findingsreveal significantvariationin weekly expected returns, and this variation has systematic patterns both over time and across portfolios. Furthermore,the fact that the expected returnstend to incorporateother relevanteconomic information (includingthe effects of the laggedmarketreturn)suggeststhat the signal extractedfrom past returnsis not due to marketmicrostructure biases. However, there are several unexploredissues. First, futureresearch is needed to determine the underlyingeconomic determinantsof the time-variationin expected returns. This will help us understandthe natureof the systematic differences in the degree of relative variation in expected returns both across assets and over time.5 Second, the rapidlydecaying time-variationin short-horizonexpected returnsapon 5. Ourresultsindicatethatthe expected riskpremium the marketfollows a stationary process. Preliminaryresults also show that systematic risk measures (i.e., betas marketportfolio)changeslowly over time (see computedrelativeto the equal-weighted also Ohlsonand Rosenberg 1982).Hence, it appearsthat expected returnson common stock vary due to changes in both risk premiumsand risk measures.

424

Journal of Business

pears to be quite differentfrom (and unrelatedto) the slowly moving long-horizonexpected returns.An understanding the basic determiof nants of expected returnsmay also help explain these differences.
References
Ansley, C. F. 1980. Signal extractionin finite series and the estimationof stochastic
regression coefficients. Proceedings of the American Statistical Association, Business

and Economics Statistics Section, pp. 251-55. An R. to Blume,M. E., andStambaugh, F. 1983.Biases in computedreturns: application the size effect. Journal of Financial Economics 12:387-404.
Box, G. E. P., and Jenkins, G. M. 1970. Time Series Analysis: Forecasting and Control.

San Francisco:Holden-Day. Breeden,D. 1979.An intertemporal and asset pricingmodelwith stochasticconsumption

investment opportunities. Journal of Financial Economics 7:265-96.

Brown,P.; Kleidon,A. W.; andMarsh,T. A. 1983.New evidenceon the natureof sizerelated anomalies in stock prices. Journal of Financial Economics 12:33-56.

Campbell,J. Y. 1987. Stock returnsand the term structure.Journalof FinancialEconomics 18:373-99.

Cohen, K.; Hawawini,G.; Maier,S.; Schwartz,R.; andWhitcomb,D. 1983.Frictionin the tradingprocess and the estimationof systematicrisk. Journalof Financial Economics 12:263-78.

modelfor Cox, J. C.; Ingersoll,J. E.; andRoss, S. A. 1985.An intertemporal equilibrium asset prices. Econometrica53:363-84. Fama, E. F. 1965.The behaviorof stock marketprices. Journalof Business 38:34-105. Fama,E. F. 1981.Stock returns,realactivity,inflation,andmoney.AmericanEconomic
Review 71:545-65.

Fama, E. F., and French, K. R. 1987. Dividend yields and expected stock returns. Workingpaper. Chicago:Universityof Chicago. Fama, E. F., and French, K. R. 1988. Permanentand temporarycomponentsof stock prices. Journal of Political Economy 96 (April):246-73. Fama, E. F., and Schwert,G. W. 1977.Asset returnsandinflation. Journalof Financial
Economics 5:115-46.

Ferson, W. E.; Kandel, S.; and Stambaugh, F. 1987.Test of asset pricingwith timeR. varyingrisk premiumsand marketbetas. Journalof Finance 42:201-20. Fisher, L. 1966. Some new stock-market indexes. Journalof Business 39:191-225. Gibbons, M. R., and Ferson, W. E. 1985. Testing asset pricingmodels with changing expectationsand an unobservablemarketportfolio.Journalof FinancialEconomics 14:217-36. Kaul, G. 1987.Stock returnsand inflation:The role of the monetarysector. Journalof
Financial Economics 18:253-76.

Keim, DonaldB., and Stambaugh, F. 1986.Predicting R. returnsin the stock and bond
markets. Journal of Financial Economics 17:357-90.

Lo, A. W., and MacKinlay, A. C. 1987. Stock marketprices do not follow random walks: Evidence from a simple specificationtest. Workingpaper. Philadelphia: University of Pennsylvania. Lucas, R. E. 1978.Asset prices in an exchangeeconomy. Econometrica46:1429-66. Merton,R. C. 1973.An intertemporal capitalasset pricingmodel.Econometrica 41:86787. Murphy,K. M., and Topel, R. H. 1985. Estimationand inference in two-step econometric models. Journal of Business and Economics Statistics 3:370-79.

Ohlson,J., and Rosenberg,B. 1982. Systematicrisk of the CRSPequal-weighted common stock index: A historyestimatedby stochastic-parameter regression.Journalof
Business 55:121-45.

Roll, R. 1983. On computingmean returnsand the small firmpremium.Journalof Financial Economics 12:371-86.

Rosenberg, R. 1973. The analysis of a cross-section of time series by stochastically

Time-Variation

425

convergent parameter regression. Annals of Economic and Social Measurement

2:399-428. Scholes, M., andWilliams,J. 1977.Estimating betasfromnonsynchronous data.Journal

of Financial Economics 5:309-27.

Shanken,J. 1987.The intertemporal capitalasset pricingmodel:An empiricalinvestigation. Workingpaper. Rochester, N.Y.: Universityof Rochester. White,H. 1980.A heteroskedasticity-consistent covariancematrixestimatoranda direct test for heteroskedasticity. Econometrica48:817-38. Working,H. 1960. Note on the correlationof first differencesof averagesin a random
chain. Econometrica 4:916-18.