loumal of Financial E∞nomics 22 (1988) 27-59.

North-Holland

MEAN REVERSION IN STOCK PRlCES
Evidence and Implications*

James M. POTERBA
M따sachuselts lnslÏl ule 0/ Technolo,양" Cambridge, MA.01l 39, USA
Nalional Bureau 0/ Economic Research, Cambridge, MA 02138, USA

Lawrence H. SUMMERS
Haroard UniversilY, Bosloπ MA 02138, USA
Nalional Bureau 0/ Economic Research, Cambridg,ι MA 02138, USA

Received August 1987 , final version received March 1988

Thi‘ paper investigates transitory components in stoc k. prices. After showing that statistical tests
have little power to detect persistent deviations between market prices and fundamental values, we
consider whether pri야s are mean-reverting, using data from the United States and 17 other
countries. Our point estimates imply positive autocorrelation in returns over short horizons and
negative autocorrelation over longer horizons, a1 though random-walk price behavior cannot be
r던 ected at ∞nventional statisticallevels. Substantial movements in required retums are needed to
account for these correlation pattems. Persistent , but tran치 tory, dispa디 ties between prices and
fundamental values could 외so explain our findings.

1. Introduction

The extent to which stock prices exhibit mean-reverting behavior is crucial
in assessing assertions such as Keynes' (1 936) that ‘ all sorts of considerations
enter into market valuation which are in 00 way relevant to the prospective
yield' (p. 152). If market and fundamental values diverge, but beyond some
range the 마 d iffl
마 하rencπc
e 야 않s are eliminated by sψpecω
e 비la
없t“뼈
iv
ve ,
fo
야“rπc
야 앓
e sκ, then stock prices
will revert to their mean. Retums must be negatively serially correlated at

.We are grateful to Changyong Rhee , lelf Zwcibel and , especially, David Cutler for excellent
rescarch assistance, to Ben Be rnanke , Fischer Black, Oli에er Blanchard , lohn Campbell , Robert
En gle, Eugene Fama, Terence Gorman, Pete Kyle , An drew Lo, Greg Mankiw, Robert Merton ,
lu \i o Rotemberg , Kenneth 피n밍eton, Mark Watson , an anonymous referee, and the editor,

繼驚
Wi1l iam Schwert, for helpful comments , and to lames Darcel 삐d Matthew Shapiro for data
assistance. Thi s research was supported by the National Science Foundation and a Batterymarch
Financial Fel10wship to the first author, and is part of the NBER Programs in Economic
Fluctuations and Fi nancial Markets. A data appendix is on file with the ICPSR in Ann Arbor ,
Michigan.

03 04- 405X/88/S3.50@1988 , Elsevier Science Publishers B. V. (North-Hol1 and)

be accorded to our null hypothesis of serially independent returns. 2 We argue that it is difficult to account for of these approaches involves using a particular function of the sample autocor- observed transit relations to test the hypothesis that all autocorrelations equal zero. presence of any Iimils on valuation errors sel by speè비 alors or real inveslmenl 3 Fama (1976) acknowledges the diffic버ty of distinguishing the random-walk model from some opporlunilies .n slock pr. In Ihe. 에 B lac 짜 k (1986 뼈야), Campbe Jl and Kyle (1986) . and S삐ler 6 (1984) discuss a v하iety of possible models for noise Irader behavior tions from random-walk beha에or. These tests nevertheless have little power. marketsjconstant-required-returns model by exarnining individual autocorre- , ‘: Section 2 evaluates alternative statistical tests for transitory price components. A vast literature dating at least to Kendall (1953) has tested the efficient- t. Huizinga (1987) pro꺼des a speclr. Fama and French Section 5 investigates whether observed patterns of mean reversion and the (1988b) use regression tests that also involve studying the serial correlation in associated movements in ex ante returns are better explained by shifts in multiperiod returns. depend- íng on our assumption about its persistence. with returns tive hypotheses . follows a random walk.S.1 Our investigation of mean reversion in stock prices is organized as follows.on. prices and fundamenlal values wilhoul negative 'serial correlalion in relums.period. Section 4 uses our variance-ratio estimates to gauge the significance of transttory príce components. Campbell and Mankiw (1987) study the importance of required returns due to changes in interest rates or market volatilitv or as transitory components in real output using parametric ARMA models. Mean ret' ers.ces 1‘~ j서 ‘a 끼 conclusion discusses some implications of our results and directions for future some frequency if ‘ erroneous’ market moves are eventually corrected. surveyed in Fama (1970) . iιινι‘ J. the return ISlochaslic speculalive bubbles. the presence of negative autocorrelation 2. Perron (1 985) and Summers (1 986) has shown that such tests have relatively ble alternative hypotheses such as the fads model suggested by Shiller (1984) little power against interesting alternatives to the null hypothesis of market and Summers (1986). incI uding those used in recent studies. . Th e point estimates imply that Recent studies use different but related tests for mean reversion. such bubbles could 001 exist allemative specifications.oll . Fama and transltory components account for more than half of the monthly return French (1 986a) and Lo and MacKi nlay (1988) ∞mpare the relative variability variance. Summers. lations in security returns. The early literature. considered by Blanchard and Walson (1982) . Polt! rba and LH. For the then present Monte Carlo evidence on each test’ s power against plausible U.외 inlerpretalion of the Ihan Ihe.enous. result of maximizing a conventional utilily function 뼈 u s잉in 매g 떼 raa와i“Ilon 매따외 n a때1 앉 e Xp 야e‘이 :Ia따때 Ilon 마 1 variance-ratio estimalor and reporls evidence that exchange rates also show long-horizon devia- 떠lu π 삐 m 바lπ1m 마 d i샤s“t빼 u 띠삐t“ion. Each byproducts of noise trading. tion over longer periods. We concI ude with a behavior of individual ∞φorate securities in the U.S. although it could also arise ‘,‘ from variation in risk factors over time. Polerba and LH. Test methods the transitory price component varies between 15% and 25% of value. 0 ’ Brien (1 987) demonslrales the presence of negative serial correlation at veπ long (up to twenly-year) horizons. even with efficiency with constant required returns. ralher MacKinlay (1988). Several recent studies using new tests monthly data for a 60-year period. we analyze monthly data on real and excess New York Stock Exchan2e alternatives to the null hypothesis of serially independent returns. (1987) notes that reasoning of this type has been used to draw concI usions about market valuation from failure to r멍 ect the absence of negative serial correlation in returns. as well as annual returns data for the 1871-1985 most powerful tests have little power against these alternatives to the random . We find that variance-ratio tests of the type used by Fama and French (1 986a) found little evide따e of patterns in secu디 ty returns and is frequently adduced and Lo and Mac Ki nlay (1988) are cI ose to the most powerful tests of the null in support of the efficient-markets hypothesis. We concI ude that a sensible balancing of for serial dependence have nonetheless r얻jected the random-walk model. however. Summers. 1 Merton ji1R 에 3 ‘‘,a*!**‘*i research.1.05 level.05. For the U.M. of returns over different horizons using variance-ratio tests.S. Even the (NYSE) returns since 1926. stationary stock-price components. Recent work by Shiller and hypothesis of market efficiency with constant required returns against plausi. The variance-ratio test exploits the fact that if the logarithm of the stock price. MethodologicaI issues involved in testing for transitory components ,2--: may signal departures from fundamental values . could creale devialions belween marke l. J. (1987) .n slock prices 29 28. 3 Type 1 and Type 11 errors s맨gests using critical values above the conventional This section begins by describing several possible tests for the presence of 0. Conversely .M. We also analyze 17 other equity markets and study the mean-reverting walk when we specify the conventional size of 0. addressing in partic비 ar the degree of presumption that should showing positive autocorrelation over short periods but negative autocorrela. We Section 3 exarnines the extent of mean reversion in stock prices. including cumulated dividends. we find the standard deviation of 2. Mean rl't' ers. In addition to the recent work of F뻐1a and French (1988b) and Lo 때d 2 Noise Iraders are inveslors whose demands for sec‘lrilies are besl trealed as eXOl1. a finding confirmed by international evidence. The results consistentlv discussion of test design when the data can only weakly differentiate alterna- suggest the presence of transitory components in stock prices . De Long el 떠.

will r멍 ect the null hypothe- transitory factors .) ~!esting the~ relationship. Our variance 2. 5 For monthly returns. We analyze the power.j). pararneters that were c1 0se to those from the data-generation process. If some of the p디ce variation is due to then positive. the estimation package 얀 the variance ratio is large in relation to the sample size. not a general rule about the relative power of the two tests. (1) can be written are feasible. the bias is -0. SWe use twelve-month retums in the denominator of the variance ratio to permit comparability 8 We tried estimating ARMA models for the pseudo.*) and a transitory (u .M.M. between the variability of retums at dilferent horizons has a long 1 Further details on the relationship between regression tests and the sample autocorrelogram tradition: see Osbome (1959) and A1exander (1961). standard ARMA 6 Kendall 뻐d Stuart (1976) show that under weak restrictions. The variance ratio places increasing positive weight on autocorrelations up to and including lag 11 . Less than 10% of the cases led to welI-estimated Monte Carlo analysis of the varian야-ratio statislic may be found in Lo and MacKinlay (1988). than the variance-ratio tes t. I. used by Fama and French (1988b). or computing likelihood-ratio tests of the null The variance ratio is closely' related to earlier tests based on estimated hypothesis of serial independence against particular parametric alternatives.069.ratio denominator is study. R denoting the total return in month t. For example.720 뼈d k . however.i mately alternatives we consider. autocorrelations. the regression coefficient is ity of returns at different horizons.2 친 (뜩딛)까 (2) rate is a lower bound on the error rates that other tests with the same size could achieve. the variance-ratio statistic is therefore ßk= 효k(kfRtk)/헐k(한_k)2 ” / l I l ‘ ·‘ n I / ‘ πús statistic applies negative weight to autocorrelations up to order 2k/3 . this bound is unlikely to be achieved . are presented in an earlier draft. the variance ratio will fall below one. We show below that returns display positive.60.) embodies both a permanent (p . the Neyman-Pearson lemma dictates that the like \i- hood-ratio test is the most powerful test of the null of serial independence against the particular alternative that generated the data. encountered noninvertibilities in the moving-average poJynomial and therefore brQke down in For exarnple. Thi s is the result of the actual properties of the . with declining positive weight thereafter. of tests for transitory components against the A second test for mean reversion.-0 returns data. with sarnple autocorrelation is -lj(T . se디al correlation as . since we do not know the precise data-generation process. In practice. Using this r야비 t, we ∞mpute 타 VR(k)). alternative hypotheses that Summers (1986) su잃,ests, where the logarithm of regresses m바 stock prices (p. Power ca/cu/ations ratios for k-period annual returns place declining weight on all autocorrela- tions up to order k.6 time-series models for returns. Mean reversion in stock pηces 31 30 I. Thi s statistic converges to unity if the horizon lengthens. Using Cochrane’ s (1 988) result that the ratio of the k-month Because returns are nearly white noise under both the null hypothesis and the return variance to k times the one-month return variance is approx. Summers. 4 We study the variabil. 、 followed by increasing positive weight up to lag k . autocorrelations at some lags will be negative and sis of serial independence more often than the variance-ratio tes t. however. in relation to the variation over a one-year 끼 period. sive and moving-average c야 lllcients of roughly equal but opposite signs. With annu허 data. Mean reversion in stock prices variance should be proportional to the return horizon. by virtue of its negative. 8 When they equal to a linear combination of sample autocorrelations.120. then negative. It rises to -0. .2. returns are uncorrelated through time.order autoregres- var( R .retums generated in our Monte Carlo with our results using annual retums data. When the horizon three-quarters of the variation in retums due to transitory factors. 1) model with first. so its Type 11 error 깨 (k) 프 1+2델 (펀)Pj. available on reques t.160 if k . this can be substantially less than u띠ty. In this case the regression test . the expected value of the }th estimation packages (i. followed by decaying where positive weights. . standard ARMA techniques often fai l. 7 Fama and French (1988b) report regression tests because k-l they r엠 ect the null hypothesis of serially independent returns more strongly R~ = L R t-i. de-meaned k-period return. with T . RATS) had dilllculty recovering this process.). Summers. The statistics reported below are A third method of detecting mean reversion involves estimating parametric corrected for small-sample bias by dividing by E[ VR(k)].e. the variance. Ahhough these data were generated by an ARMA(I . Poterba and LH. Poterba and LH. Detailéd more than a third of 외I Monte Carlo runs. weights on sample autocorrelations.

2112 In evaluating Type n error rates .05 test is 0. Retum regression 12 months 0. Parameters of retum-generating process 서 .931 0. Th e Th e first row in table 1 analyzes a test based on the first-order autocorrela- Type II error rate for a size 0.98 、 p B. 1). 48 months 0.903 -0332 generated with 8 = 0 to deterrnine the critical region for a one~sided 0.250 0.OOO sequences of 36 months 0.002 0.0.934 -0.001 r얀um va민ation due to transitory factors is deterrnined by the relative size of Variance ratio 0.~.M.n stock pr. t ·.924) when one-quarter tion coefficient.8 1.75).894 0. . return components. 84 months 0. variation is due to transitory factors (8 = 0.044 0.841 -0. Mean revers. φ -p" The next panel in table 1 considers variance-ratio tests comparing return variances for several different horizons ..1 p/) equals 1 + 2상/(1 + P1) and ~t parameters for the return-gener.210 lransltory price component have a half-life of 2.ces J.-1 + κ, (4) Each row describes the statistica1 properties of a parti띠ular test for mean reversion.--(p,+ 앙 )/9.7ω 4.927 0. but persistent.842 -0.25 and 0.815 60 months 0. Poterba and LH.167 0.8117 -0.973 0. the Type II error rate never falls below 0.05 test 96 months 0. the share of return variation due to transitory components .813 0.75.906 0.934 -0.814 0.0.733 retums (. test has rninimal power against the a Iternative hypotheses we consider. with and E/ = P/. but they still have little power to detect IO~n p.491 table 1 report Type II error rates for each test when the data are generated by the pr@ess indicated at the column head. As Shiller and Perron (1985) and Summers (1986) observe.943 -0.820 0. first-order autocorrelation coefficients . The mean value of the test statistic under the a Itemative hypothesis is also reported.844 0.941 -0.0119 nent.920 0. Summers. 24 months 0.1 p.25) ..0.863 -0.929 0.935 0.M. indexed by k . If the stationary component is an AR(l) process Table 1 Simulated Type n error rates of a1 temative tests for transitory components in security returns ".p, )[4앙 + (1 +써 '/2} /(2앙 +2p,), variarlees. Poterba and LH. Th e parameter 8 96 months 0.354 of the random-walk null against the mean-reverting alternative. Each test we analyze has s따 0.927 0. the probability of failing to reject the null 72 months 0. 1) process.98 for both cases.863 0.927 We perform Monte Carlo experiments by generating 25 .929 -0.856 -0.839 0. When one-quarter of the return truct dp. = P1" .. implying that innovations in the 36 months 0.929 0..on . . It is useful in considering the empirical res비ts below to note .884 0. we use the empiricaI distribution of the test statistic 84 months 0. φL)dp, .952 0.929 -0. and pj .l) model (1 .112 0. + ".221 0..25 p .ra~tice we draw 720 pairs of random variables.The parameIer Pl First-order determines the persistence of the transitory component. the then monthly serial correlation of the transitory comφonen t.0. associate them with (e" κ), and then persistent. follows an ARMA (1.868 -0.P/":-l denotes the innovation to the nonstationary component.98 .080 0. We consider cases where 8 때als 0. L )(1 . Summers.924 -0.2.933 -0.찌4 0.925 0.15 Test statistic If p/ and E/ are independent.25 and 8 = 0. . 32 J.771 Securities Prices (CRSP) data base. this (three-quarters) of the variation in returns is from the stationary component 9까le parameters of the ARMA(l .814 0. = p . e. the number of monthly observations in the Center for Research in 48 months 0. Th e panels of LR test 0. The variance-ratio tests are more powerful than tests based on a.. 24 months 0.P1 L) -1κ. AII tabulations are based on one set of 25α)() Monte Carlo experiments using 720 monthly returns generated by the process described at the column heading.194 0..933 0.670 denotes the share of return ‘ variance accounted for by the stationary compo.3011 hypothesis when it is false.158 We set P\ equal to 0. lO We set 낭 = 1 so that the variance of 72 months 0. Mean revers.on . with one-period return 9-{-(1+ 야 )-2앙+ (1.+ (1 .1 P/ = E . parameters set by B..141 0.700 ating process by ch∞sing P1 and 8 = 2 강/(1 + P1 + 2 앙). 9 Th is and return Type 11 Mean value Type 11 Mean value measurement error of test error of test description of returns allows us to capture in a simple way the possibility that interval rate stabsbc rate statJstJ c stock prices contain transitory. but transitory.n stock pr.05.924 1.941 (0.ces 33 componen t.914 -0. 8 = 0. components.9 years.867 720 returns.941 -0.(l + 9L) 씨 are (i. We assume that p . “ I / .941 -0.ana O.250 60 months 0. ð and P1 deterrnine 0. and the share of autocorrelation 0. Both underlying processes are ARMA(l .

showing the empirical distribution function for the 96-month variance ratio in both the homoskedastic and heteroskedastic cases.s~ o! autocorrelations. ‘@ α 4 The final panel of the table presents results on likelihood-ratio tests. For the case where three-quarters of the return variation is due to transitory o Type 1 Errar 11 꺼le likelih∞d value under each hypothesis is evaluated using Harvey’s (1981) exact maximum Fig. 2. the figure shows the associated Type 11 error rate against the altemative hypothesis.Iog( 바 1) + ω. regression-beta. cu~e pre~nts .M.67. The ~yp~ 1I erro: rate fo~ ~ac~ tes! u~der the altemativé hypo Ïhesis of 8 .for ~ pa..2 the significance level should be a decreasing function of sample size' (p. ll Although these are more power띠 than the'variance-ratio tests. but allowing for hetero- to vary through time according to this process. compared with 0. the o error rates are still high. 10 > the regression tests appear to be somewhat less powerful against our alterna- ‘Õ 8 tive hypotheses.,. yet transitory.Li kelihood R이 ia ----.929 for the most § 6 powerful regression tes t.0.925 .05 @a M level.243 + 0. 1.3.0.rtic~lar te. The logarithm of the simulated retum variance evolves from the resulting simulations are similar to those in table 1.2. Regression Bela @ For most of the tests described above.05 .25 case and 0. of homoskedasticity in the return-generating process.2) .95 if the Type 1 eπor rate were set at the conventional 0. Type 11 versus Type 1error rates for three altemative tests of mean reversion. aIthough ir.98 is calculated using another set ofsiIllulated o log( 0 . For variance-ratio. . P\ .85 and 0. Schwert. setting the significance level arbi- trarily at 0. Leamer (1978) echoes a point made in most statistics courses when he > LF writes that ‘ the [popular] rule of thumb. we fit a first-order autoregressive model to montbly data on the lated from 25 ,α)() replications of 72O-observation time series under the null bypothe~is of se?외Iy logarithm of volatilityP We expand our Monte Carlo experiments to aIlow 앙 independent draws irom an identical distribution: lbe bro~en. Because estimating the mean induces a small-sample bias toward negative Each curve displays the tradeoff between Type 1 and Type 11 er~or r~tes .34 I. Ii kelihood method. is . Poterba and LH.Variance Rali。 ‘。」」티 2. Summers. even under the null hypothesis of serial independence the mean Iikelihood ratios mean reversion-in-stock retums.922 in the lì = 0. price components. Fig.. Th e results are similar to those for variance ratios. the 14 variance ratio at 96 months is 0. and Stambaugh (1987)..statisti~, cal~~­ tance .7689 . To investigate its impor. Mean reversion in stock prices 35 that when the transitory component in prices has a half-life of less than three years and accounts for three-quarters of the variation in returns (lì = 0. and likelih∞d-ratio tests under the null 12 The estimated volatility process that we use for our simulations is hypothesis of serially independent.691. Critical reglons for each test are found using simulated empirical for each altemative hypothesis are above one ‘ dis ributions for the variance-ratio.75 . reJl.~ si~la~ e~pirical distribution calculated from the same number of Monte Carlo draws. Mean reversion in stock prices 1.. Oand likelihood-ratio tests with glven Type 1 error rates shown 떠ong the horizon때 싫. The monthly lions are c외C띠at혀 using 25 ,α)() replicalions of 72O-observation time series for synthetic re씨 ms.75) .. deficient in the sense that from every reasonable viewpoint 0.760 in the lì = 0. M.. 1 illustrates through time as noted in footnote 12 this. The solid curve shows the empirical distribution of the 96-month variance~ratio .. empirical distributions: Under both the nu\l and the altemative hypothesi~, t~e empi?ca~ distribu- where 와 has a nonnal distribution with mean zero and standard de찌 ation 0. ’ --Homoskedas ic Relurns The next panel in table 1 shows Type 11 error rates for the long-horizon 12 ’ ’ ----He eroskedas ic Relurns regression tests. Empiric외 distribution of 96-month variance-ratio statistic with homoskedastic 뻐d heterσ One potential shortcoming of our Monte Carlo analysis is our assumption skedastic retums.75 case. the Type 11 error rate would be 띠 between 0. homoskedastic retums. volatility data are described in French. Even the best possible tests therefore have Ii ttle power to distinguish the random-walk model of stock prices from alternatives that Variance Ralia imply highly persistent. Summers.25 case has a Type 11 error rate of 0. Th e Type 11 error calculations skedasticitv in the simulated retums. with Type 11 2 error rates of 0. 92). Fig.ression-beta. the best variance-ratio test against the lì = 0. Eua/uating statistica/ significance . Poterba and LH. . For example.

358) (0.522 0. The divergence between their fìndings and thosc of Lo and one-year retums. monthly dat a.S. Mean reversion in stock prices 37 36 J.232) (0. This finding of first positive then negative serial correlation parallels Lo and MacKi nlay’ s (1988) result that variance ratios exceed unity in their weekly data . along with the Monte Carlo shown in table 2. Furthermore. The variance-ratio statistic is defined as VR(k) .6% 0. as well as real retums measured using the Consumer Price Monle Carlo resu1ts. Values in parenthescs are Montc Carlo estimates of the standard errpr of the variance the variance-ratio test.610 0. involving rational and real returns (0‘ 150) (0. 까le variance-ratio statistics for these series are normal appro싱mation to the distribution of the variance ratio.278) (0. 3. 14Prench and Roll (1986) apply vari빼야-ratio tests to daily retums for a sample of NYSE and Eight-year returns are about four rather than eight times as variable as AMEX stocks for thc peri여 1963-1982.394) importance of transitory components.7% 0.855 0.764 1. The variance of the one-month retum on the equal- ptê디od since 1871 . Summers. Further details are available on reques t.79 times as large as the variability of twelve-month markets outside the U. significance levels in excess Data scrics dcviation month months months months months m어Iths months months of 0.797 0.010 0. We confirm the Fama and French (1988b) finding that both estimales of the standard devialion of the variance ratio. A similar conclusion applies to the value-weighted post-1926 period. so it Eq u허-wei빼 tcd 29.1 5이 (0 .u reversion is more pronounced for the equal-weighted than for the value- This section uses variance-ratio tests to analyze the importance of stationary weighted retums.05 test .278) (0. The va더ance ratios also suggest positive retum autocorrelation at horizons annual returns on the Standard and Poor’ s-Cowles stock price indices for the shorter than one year. bascd on 25.232) (0. Mean retJersion in stock prices Tablc 2 factors .809 0. especi외Iy among smaller securities.785 1.M. In Equal-weighted 29.394) onJy solution to the problem of low power is the collection of more data.108) (0.786 0. 2 depicts the attainable tradeoff between Type 1 and Type 11 errors Variance ratios for U. To justify using the conven- experiments under the null hypothesis of no scrial correlation. . as rcported in the CRS!. one would have to assign three times as great a cost to Type I as to Type 11 errors.1 7끼 (0. Monthψ NYSE returns.278) (0.(12/k). For interval. the null hypothesis of MacKi nlay (1988) is presumably due to dilferen야s in the Iwo data scts serial independence is r,에ected at the 0.835 0. tion aI 0. Poterba and LH.1. as well as for the CalculatioDS are bascd OD the monthly returns for the value.weighted and 여ual-weighted NYSE likelihood-ratio tes t.05 seem appropriate in evaluating the importance of transitory compo- Valuc-weighted 20.320) (0. monthly retums file.1 77) (0. Retum measurcmcnt intcrval An nual retum Since there is little theoretical basis for strong attachment to the nu l1 standard 24 36 48 60 72 84 96 hypothesis that stock pric강s fo lIow a random walk.487 0. They find evidence of negative serial ∞rrelation.873 0.278) (0. We analyze excess and real returns using four for both.40 significance level is appropriate if the goal is to ratio. Summers.6% 0. πle Type 11 error curve for the variance-ratio test lies portfolios.358) (0.v뼈 R k )/var( RI2 ), where RJ denotes returns over a j-period measurement between the frontiers attainable using regression and likelihood-ratio tests.320) (0. real and excess retums at long horizons show negative serial correlation.232) (0. Valuc-wei방Ited 20.1 7끼 (0. based on Treasury bill yield.320) (0. Many asset-pricing models. If some of the securities in equal-weighted NYSE indices from the CRSP data base for the 1926-1985 pe다od.565 0. whereas va디an야 ratios fa l1 below one in other studies concemed with 3. post-World War 11 monthly stock retums for 17 stock weighted index is only 0.745 0. J.358) (0. Statistical evidence on mean reversion retums and at the 0. a 0.917 0.M.575 nents in stock prices.747 0.394) plague our search for transitory components complicate investors' lives.878 0.036 0.S.1 77) (0. 1926-1985 longer horizons. but the variance ratios at long horiwns are well below unity components in stock prices. fig.400 0. 1926-1985 for the most powerfl비 variance-ratio and regression tests. suggest the presence of transitory components and time.689 0. They permit rejeclion at lower levels than would be possible using the Index (CPI) inflation rate.08 level for value-weighted excess .6% 0.232) (0.925 0.108) (0. and retums on individual firms in the U.667 0.320) (0.649 0.150) (0.358) (0. the same problems of statistical power that ex야ss returns (0.963 0.642 0.677 varying returns.1 5이 (0.989 0. Th e real returns (0.973 0. We consider excess retums with the risk-free rate measured as the \3까lese p-v외ues are calculated from Ihe empiric외 distribulion of our tesl statistic. Poterba and LH.S.α)() replications under the null hypothesis of serially independent retums.425 the next section .781 0.005 level for equal-weighted excess retums Mean .108) (0. we bring to bear as much data as possible in evaI uating the excess relurns (0. Each varian야 ratio is corrected for small-samplc bias by di피ding by the mean value from Monte Carlo minimize the sum of Type 1 and Type 11 errors. Despite the low power of our tests.353 may be difficult for speculative behavior to eliminate these components. index. 14 One potential difficulty in interpreting our finding of positive serial correla- We begin by analyzing monthly retums on both the vaI ue-weighted and tion at short horizons concems nontrading effects. for the returns implies it should be.394) irrational behavior.1 08) (0. major data sets: monthly retums on the NYSE for the period since 1926.

We Additional evidence on mean reversion can be obtained by analyzing the 15 We 외so experimented with cmde Iechniques for accounling for tIme-varying stock market behavior of equity markets outside the U. value-weighted index rise to 0.436) A second issue that arises in analyzing the post-1926 data is the sensitivity Real retums 17. Excludìng the first ten y않rs weakens the evidence for (1987).425 0. expehen@.736 at the 96-month horizon. excess returns variance ratios are not substantially changed by treatment of the early years. For the post-1936 period .1 43) (0. For real returns. Equity markets outside the United States by Merton (1987) suggest the value of examining other data as well. for example. This may make the ex post inflation rate an unreliable years of data strengthens the evidence for m C( an revers measure of expected inflation during thisperiod. Values in parentheses are Monte Carlo standard would require that one security in ten typically did not trade in a given month. Mean reuersion in slα kprices 39 the market index trade infrequently.313) (0.tima~ing ~rian~e ~a. with variance ratios of 0..'elati~~~-wiihJ~-het~~~~ie::i. 16 까1e CRSP data are the best available for analyzing recent U.913 0. Th e results for both equal-weighted real and used in studies of the serial correlation properties of stock returns.1 40) (0.806 0.3. POlerba and LH.737 0.29이 excluding these years from analyses designed to shed light on current condi. have documented the unusual behav.807 of the findings to inclusion or exclusion of the Depression years.s.436) previous studies.767 0.and equal-weighted reaJ returns and 0.265) (0. Summers. contain a great deal of information about the persistence of price shocks. pre.210) (0. but real and excess returns provide less evidence of mean reversion than the monthly post-1925 CRSP data. Es timating sample autoco. Index [see Poterba and Summers (1986)] for the period 1928-1986. Summers. 1871-1985 returns.009. Nontrading at this frequency is likely to affect only a small fraction of Each entry is a bias-adjusted variance ratio with a mean of l!nity und~! the null hyPothesis. and Stmbaugh’ s (1987) estimate of the previous month’s 16 The variance ratio for the full sample (1871-1985) period is not a simple wei빼 ted avera응e of retum volatility elfect바ly r,벼uces the weight of the early Depression ye따s, yielding variance-ratio the variance ratios for the two subperiods.441 1871-1925 (0. data. backdated to 1871 using the Cowles data as r,φorted in Wil50n and Jones (1987).358) (0. although excess returns are robust to the sample choice.398) (0.782 and 0.211) (0.211) (0.M.10.143) (0. but the low power of available statistical tests and data-mining risks stressed 3.769 0.587 they have been used in some studies of stock market volatility.922 18. returns will show positive seri aJ correla. Excess retums 1.047 0. πle Retum measurement interval first-order autocorrelation coefficient for daily returns is only 0.095) (0.0.996 0. One could argue for 1871-1985 (0. which are avaílable beginning in 1871. revised by Wilson and Jones beginning and the end.601 0. but the long-horizon variance ratios on the (1 981).731 . Table 3 tion. respectively. For the pre-1925 period .833 ior of stock price volatility during the early 1930s.095) (0. Excess retums 16.9% 0. Historical data for the United States serial correlation at long lags . display negative serial correlation at long horizons.464 0.851 for vaJ ue. The securities.825 the pattern is weaker.847 0. Polerba and LH.064. where RJ denotes the retum measured over a j~month interval. 1S Truncating the sample to exclude the last ten years before 1900. although the sham increas¢ in market voIatility in the last quarter of 1871-19115 (0. Both series show negative 3.(12jk).313) (0.398) (0. the post-1926 period excess and re외 S&P data. πlÌs s때gests that autocorrelation pat- terns in monthly returns are not likely to be due to infrequent trading.M. deviations of the variance ratio.915 0.737 0.876 0.24이 (0.612 0.929 0.179) (0. the one-month variance ratios are 0.781 tlons.and unclear. and Annual retum standard 24 36 48 60 72 84 % groupíng returns into nonoverlapping five-day periods yields a first-order Data series deviation months months m여IthS months months months months autocorrelation coefficient of .266) (0. 38 J.797 0. These data have rarely been mean reversion at long horizons. composite value-weighted index by analyzing daily returns on the Standard and Poor’ s stock index.and post-1926. such as Shiller a찌 0. respectively . We analyze returns in Canada for ~olatility i~ es.880 0.240) (0.0% 1.265) (0. We doubt this explanation of our results since we are analyzing monthly Varian야 ratios for U. Real re!ums 19. by virtue of the large movements in prices. whereas accounting for the degree of positive correlation we observe variance-r"atio statisticis defined as VR(k) .290) l987 undercuts this view. The one-month The results are presented in table 3. 2.358) (0.463 and 0. based on 25 ,1이)() replications under the null hypothesis of seri외 We also investigated the incidence of nontrading in a portfolio similar to the independence 까le un4erlying data are annu외 retums on the Standard and Poor's.S. are 0.S. var( R k )jvar( R 12 ) . it appears to result from the volatility of the CPI inflation rate in the equal-weighted excess returns.035 0.856 0.2% 0.2% 0.97 and 1. Mean reuersion in slockp끼ces J.1 40) (0.591 0. The counterargument suggesting inclusion of this period is that the 1930s.tios. Although the explanation for this phenomenon is for value. such as Officer (1973).210) (0.1 79) (0.833 and 0. however.266) (0.821 0. A number of 1871-1925 (0. The 96-month variance ratios for estimates closer to unity. The two lower rows in table 3 present results for the full 1871-1985 sample period.? b~ on ~rench, Schwert. We explored therefore consider real and excess returns based on the Standard and the robustness of our nndings by tmncating the sample penod at both the Poor's-Cowles Commission stock price indices .855 0.ti~~ ity corre~tio..

뼈찌 om--m o o 뺏때 o ””) u % % o % :。 1l ψ- 6 3 n)-) m m-o 떠“m 4% -sh‘。 4 쩌… ω-o m--m -o -) m-o (- - 、---‘, 、 m 、 .i ‘ 。。 、,, ¢ ιι 빠빼 ι 、, g l ‘ 11 -m nm 뻐 뼈떼 W 역·MrSS깅 o 3 Oq l % μm매 omm jj %”끄m α 1mw ”” n ”“ R lm ) /.1 % 0. average of variance-ratio statistics) for a particular nation and retum horizon.16끼 (0. om 8% om ”잉 omm l 1m mu ) 때 9 mm 2m 1 여 끼 m 찌처써빼 m m J ”1l 2 ) 이” Average value α7 u … 뼈 y ·in내I m (excluding U..403) (0.334) (0.317) (0.¢ ιi 뼈빼 、, g 꾀뼈 없 o mm 1 l ”” z % 233 2m ”u 36 찌없찌… l mX ) .214) (0.3 34) (0. Th e U.S.RJ denotes the real retum over a j-month measurement interval.518) (0 .1 98) (0.5 18) (0. these are computed 외lowing for a constant correlation across countries.‘ ”” 꾀죄 ”“ R 2@ ι %6 얘 ‘ ψr 뼈 %삐 어“ 9 뼈%뼈 (o m m% m ro ψ m m m m 04 N’ξ ‘‘。: 、,, 、‘, 빼빼 때.‘ 폐 뼈服펴 j (o m* % m3 세 04 ”’” 찌 、,/ 、‘, m ¢ &. For the averages.358) (0.3 96 1.822 0. Spain) -s‘ .868 0. where.5 18) (0 .504 1.254) (0.462 (capital gains o n1y) (0.156) (0./1939-1986 20.477 0.464) (0.262 1. Data under1 ying the variance ratios are real dividend-exc1 usive retums‘ calculated from share price indices in the 1ntemational Monetary Funds' ~。‘악용SRF 녹·%§‘ Jnternational Financial Statistics.014 (capital gains 0피y) (0.102) (0.1 56) (0.218) (0.752 0.581 0.575 (capital gains o n1y) (0./ ‘ -gh ---J om Mm 째써 때W j … o .regation induced by averaging closing values of the index within each month Return measurement interval s‘a·ξSa、이eSss Retum series Annual retum 1 24 36 48 60 72 84 % Country/sample standard deviation month months months months months months months months C때ada/1919-1986 20. 1n 외I cases ex∞pt those marked with an asteri양 (*) the data are monthly averages of daily or weekly values.9% 0.054 1.687 0.S.써 g I o9 oj %wq o65 o %mM ) . m“* ‘.1% 0.617 0. 215 1.518) (0 .141 1.740 0.794 (0. data are point-sampled but on1y at the end of each year.. other data ranges are noted.864 0.166) (0.336) (0..156) (0.582 0.566) 。 ‘ Bel밍um/1957-1986 17.‘ ( “… tF t o63 ( y xm ”m o”” αα α om O o o ”M2 On η써 n % ι V& 5 n“* ‘.666 0. g Table4 Vari빼야 ratios for intemational data on re외 monthly retums Each entη reports the variance-ratio statistic (or in the bottom panel.987 0.692 0. we assume it is zerò.309 1.261) (0.566) 뼈쩌 뼈씨씨찌씨 째쩌얘쩌 찌 m 鋼 뼈빼 짜찌째빼씨뼈뼈뼈뼈째 ” H뻐 %빠 “‘ 43 α ια o OMm5 nm k4 a% (‘ α O αα ”“ % 애써 r 뼈 m빼 ” 1 Km m“! ‘.403) (0 뼈4) (0.K. m4 …빼 Average value m m ‘.251 0. … --m -M ” --y ” -% 8l occ.4% 1.1 32 0.403) (0.400) (0.566) Colombia/1959-1983 21.259) (0.987 0. 째때펴M폐 ” - o “ gm m3 n . 」--1 ‘ The variance rati。 잉 defined by VR(k) .998 0.403) (0.566) Germany/1957-1986 23.K.334) (0.0% 0. 205 1. For most countries the monthly 1MF data span the period 1957:1-1986:12.1 37 1.386 0.254) (0 .724 0 .464) (0. 삐 ” r 3 죄 4 m4 %‘‘l mm (o 씨쩌 ψ m .806 0. The varian야 ratios are corrected for the time a잃.060 0. lf this correlation was estimated to be negative.055 0.464) (0. mu m2 ‘17 m 갱 j g ‘- 2 、u7 ω l m”” 、.610 1. ” u ) (excluding U.832 0.214) (0 .502 ‘,‘a‘:” (capital gains only) (0.125) (0 ./1 κ-A l1ι g -” … -7-v .1 21 1. | mw j 찌 매째 ( 、, ι h、 ---- 째씨 g 8% om W om OOnn o m” vι om O따 om o “m ’ 끼 j 뼈니뼈 찌 。 j mn R W 양h .4% 0.876 0.807 0..776 (capital gains 0피y) (0. Values in parentheses are Monte Carlo standard deviations of the variance-ratio statistics (and standard deviations for the averages in the last three rows).663 1. …¢‘,쩌 삐 ml o 베 에 m 펴생 ψ mv m ( o j” R nι mM ι* 씨 ( “ “”、 ” w tr 1m MXn (n mmm αι α ao omm4 % αμ 아야 m% om L …… B r*ij 뼈빼빼 ‘·ι 9 %40 뻐 5 ‘. 죄 낀좌 씨 ””껴써 비 써 /.3 34) (0.156) (0.381 1.711* 1.799 0.180 (capital gains on1y) (0.156) (0 254) (0.747 0.912 0. 써 I (o 에 째얘 m nι 、, t ‘ ¢ 빼빼 빼 뼈폐m펴 ‘‘。‘맞a y vS om Lm o mi oo o& o a8j T ‘Om“% Mω 5쩨 갱 N W R V& 때 여m찌 // m E o4 ‘. 3mX‘.566) Finland/1957-1986 22.) oX 8n om j om o 야 ”o lm o 1E “ 펴 l 때 W% ‘ ι m3 n‘.141) (0.743 0.464) (0 .334) (0.214) (0.718‘ 1.370) U.5 83 0..‘a ψ펴이 %ψ m m 씨 ( ( 삐 잊 삐 」 ιm”이‘m ¢ te W폐 쩌 」·ii 83 ‘ ”” a따 L n 까 ιJ m L l Lm4 % 야 m% om Lm 쩌%쩌 찌없쩌 R @m 7 찌 m (m 째 m“‘ ‘.254) (0.403) (0.807 0.463 1.214) (0.254) (0 .518) (0.223 0.206 1. ψ F ( OJ” “ι mF mF m ρ 이 -뻐뼈 뼈 .301) (0.(12/k)*v ar( R )/v뼈 R12 ) .‘ 84 야잉 o “ J % ”“ % om o 0 ω / m3 m ( o3 m % 떠 ι* ψ까ψ m m 0 、1/ t m 、,, 、‘l 빼빼 왜펴 t g 빼뼈 m o 1 ”” o4 mm1 o33 3”여 1om l ”” X) -m 찌없 % ”“ % o mm l 0 S「 Em 찌얘씨 까 a mm m“4 “% m3 mm ” ‘m M 뼈-낀 m →이깅‘ m ( 4M ‘내/ 객 、, m 、‘, M ti ¢ ……펴m g “ 깨뺏빼 o8 mw 떠죄 o6 % o8 % oI mm 찌.8% 0.214) (0.440) ‘:‘‘n Austria/1957-1986 21.

~~ue of ~e first:()rder autocÓrrelation to b~c-1/il'=':l')-~h~~ mean for N observations. We suspect that yield-inclusive data . make it dimcult to relation to the overall market or a risk-free asse t. would affect our results in only minor ways.ces J.. The British data are monthly returns. me monihly stock lndex data fIom the IFS are time averages of daiIy or weekly 18 The standard errors for the cross-country averages aJlow for correlation between the varian야 index v~~~~.샤/앙 The variance of the sample Ins!ead.25 as the variance ratios and these variance 떼ios have constant variance 야, then the expected vaJ ue of the number of observaIlOns in the average becomes large.585 . /ndividual firm data 0.ces 43 the period since 1919.on . This is a biased sample. Firms that went bankrupt or began trading during the sample period returns is 0. Although we expect transitory components to be less likely in v Fun 찌I띠 d’녕sll “’ nternat/On 삐1 the relative prices of individual stocks than in the market as a whole. some yle 리l삐 ds’ so the reported returns correspond to capital gains alone.545. variance ratios that exceed unity. The point estimates s앤gest r멍 ect the null hypothesis of serial independence for any individual country. To assess the previous work has suggested that individual stock returns may show negative importanc건 of this omission . we also obtain Financial 깐mes-Actuaries Share Price Index. an outlier because of the Th e Canadian data consist of monthly capital gains on the Toronto Stock unusual pattern of hyperinflation followed by deflation that it expe디enced Exchange. In Germany . the 96-month variance ratio is and short-selling $82 of the aggregate marke t. 18 0. 42 J. ev aJ uating 티 VR(k)) we use 0. available from (1985)]. we examine properties 0 . omission.653 when we exclude Spain. along Evidence of positive seriaI correlation at short horizons is also peπasive. Replacing the expected For the countries with time aggregated data we therefore m여파 our sm aJ l-sample bias correction.R mt where R mt is the value-weighted NYSE retum. for example.25 -lj(T-1) 까le reported variance ratios have been bias.832. For example. ln %me cases. the first difference of a time-a:eraged r~dom ~. Table 5 reports the mean values of the individual-firm variance ratios. The first two rows of table 4 a more precise estimate of the long-horizon variance ratio. each with the same variance ø. For Canada. and associated large standard errors. For Britain. in France it is 0. Mean revers. . generalized to allow adjusted by dividing by the re잉lting expected vaJue for different sampling variances for different variance ratios on the basis of our Monte Carlo standard errors from table 4.~0) showed tha t. sample variance with the actu외 V외ue, we estimate 'T as 1 . DeBondt and Th aler exclusive CRSP and British stock market returns.718 times the value that would be predicted on the basis of the 12-month variance. Summers. By averaging across many countries . We use our estimate of 'T to evaJuate this expression. Arbitrageurs should be better at trading in individual seαuities to correct The variance ratios for the 15 other stock markets are calculated from mispricing than at taking positions in the entire market to offset persistent monthly returns based on stock pric야e indices in the International Mo 아 one하t때arπ misvaluations. aggregated and 0.794. show only minor differences as a result of dividend no missing return information between 1926 and 1985. of ~~_,:g. independent firms enables us to r멍 ect the null hypothesis that all of the price Aver variation arises from nonstationary factors. shows 17. and the increased precision gained by studying retums for many finding of substantial transitory price components. Beca use the retums for different firms are not independent. Poterba and LH. the comparable value is 0. we also on data starting in 1957. This Wlll bias our estimated vahance ratios.462. with a first-order autocorrelation α>efficient of 0.'T).11 We ∞mpute va디 ance ratios using both real and excess returns for these 82 Table 4 presents the va디ance ratios for individual countries. Most of the countries display negative serial correla. Th e 96-month efficiency gain is attenuated because the results for different countries are not variance ratio for the Canadian data is 0. the 96-month variance ratio for real value-weighted weighted toward large firrns that have been traded actively over the entire CRSP retums inclusive of dividends is 0. That is. Summers. Both markets also display statistically signi6cant positive se다al correla- tion at lags of less than 12 months.x ïiut constant cross-correlation' 'T. and many are substantially below one. returns we use.575 and that for dividend-exclusive period.4. although the show that both markets display mean reversion at long horizons. sample variance of the varian야-ratio statist때 is 티셔).Onlv with standard errors that take account of cross-firm correlation. in Britain since 1939.앙 (1 . Only three of the fifteen countries have 96-month of the time series R i / . Mean revers.438.on in stock pr. which reports variance-ratio calculations using the residuals from market-model equations estimated for each firm (assuming a constant ß for the entire period) . although superior to the are necessarily exclud려. The suggest some long-horizon mean reversion for individual stock prices in short data samplg.Jk ratios for different ∞뻐tries. inclusive of dividends . Wo~I?ng (1. We examine the 82 firms in the CRSP monthly master file that have the authors on request. that 12% of the eight-year variance in excess retums is due to stationary The similarity of the results across nations nevertheless supports our earlier factors . The last row. while for the British data it is independen t.M. examine the retums on portfolios formed by buying one dollar of each firm tion at long horizons. Poterba and LH. The results one ∞untq, Colombia. based typically firms. If 왜I nations have a constant pairwise correlation 'T between their would exhibit positive seri aJ correlation. and in 15 other nations for a shorter postwar period. se다al correlation over some horizons [Lehmann (1987). The results..n stock pr. on the during our sample period.M. the one-month variance is 3.th~ expected. has a one-month variance ratio greater ihan unity. we reestimated the variance ratios for dividend. is 야 [1 + (N -1) 'TJ/ N.

the results are stronger than an띠y individual fin뼈 1 . We do not follow thlS strategy because our 깅u。a염 〉gS어g깅씌운g < 형 쩌 Q ”。,‘애』 ”E3E 펴 6nding of positive autocorrelation over short intervals implies that the AR(1) 냥해 SDeci6cation of the transitory component is inappmpdate and because of our ”며에』U〉〈 엉엉끽여〉 a엉띠 ·aEiiu ug g R -합 ( ( diHiculties in estimating the ARMA(1 . g g 따 . 1) models implied by this approach ∞∞。 m 8gg> N *a g S a S F F {. Summers. {. Furthermore. but nevertheless allows us to focus on its standard 영쩍 ‘。P뼈@ip〉@ deviation and the fraction of the one-period return variance that can be 얘 attributed to it.an reve~~i~n in the relati~e ”경ili””ιi g 경aka £·컴김 ‘g ‘ ‘ retums on the wo indlCes.:e~ ~ove~en!s in th~ permanen샤omponent of the ” 효ip-”a【@〉 Q t 야펴 겁8 ”u two indices. ). ‘‘ @ {. ). Assuming similar-s:. 。 〕 PSIg딩 ” ”i 4u 엄 < not by a. Mean reuersion in stock prices less evidence of serial correlation than the results that subtract the market P . Summary m m @ @ 검 aa깅잉 ‘。-ggQ 써 U겁-엄。t ” 킹』얘갱엉잉 S@성”그。릅겁g ).S. Poterba and LH.S a Ng 。) guEua whole.M. nlese Ieturns show positive serial correlatlOn a all lags . B .ν 엄 < 」 ιi‘,생김 m g In recent years. One possible approach would involve calibrating models of the class m。 캠 { 。 --。 랙 ] pugg> 。 Considered in the 6rst section. . ) j 。 。 。 。 • --) 엎% • Sau Our point estimates generally suggesl that over long horizons return van- 짧찌 ance increases less than proportionally with time.Immnmlulκc야ha없s they would be if the various dFta sets were independent.5. < 。) 。 ·∞ZSE8‘ 디il gS 。 of the results rgiect tile null hypothesis of se때 independence at the 0 . 。 。 {gAB。u ”-@μ < ”【gE%-- NYSE stocks shows more mean reversion than the value-weighted portf이io. ). g ι . Instead. 깅음껴〉경g』。i (-。。 g(( ( m 。。 @ a {. ).。。깅씌Eg < the traIisitory mmponent. each of the difrerenl types of data we analyze ”g ” ($ {( *m [( 경디 ∞잉@ ∞∞∞。 nrOvides evidence of departure from serial independence in stock returns. 〔。 〔。 。 --) {. . S 。 隨폈 。 .) ) @ (] .。 B . sg (a ∞ ∞。 m g( m g 훌훌 3.ttszr짧껴S옆억ne?않i$5i램e£ESS앓lXJxe앓앓채atr짧 넓 q. The U. mean reversion is more pronounced in smaller foreign equíty U필 픔쩌 a ·”。a얘 uis쩌』‘ g” markets than in the U . ).M. i ·m∞얻 l잉Ngi “ i@ a level that m매 be appropriate giveu our previous discussion of síze versus ”EEE 경a잉디iS잊딩석킹에”B power tradeoRs.。) *。 @ . a). Where ”-g동 U뼈에』잉〉어야디경S경 C 뚫“끓엠뻐얀야행원 ” ( ( 검 。R∞ 。 ∞ ∞。 ( @ m m m。。 transitorv factors accounted for three-fouIlhs of the variatIOn in returns. . {. . ‘ 싫뾰짧면션 (*. { N @ R 。。 。 다( 。 ”g m S 。 B μ ‘。 ] 애。앓 ”g 깅‘성 s 4. and in many cases 때ey 협 imply more mean reversion than our examples in the last sectlon.1 5 level. Poterba and LH. 엎 ι침 。。깅쩌강‘ 잉。jSPμ 띠m〉 Z 검--S M 경C% 。다 ‘。 ” 원 냈협 ug i』RP겸%양{ I9We conjectured that the grealer mean reverslOIlln the equal-weighted than the value-weighted 빽 gi&μi s 앓. ·g {. 。 ’”EaZ { 운aken t02ether. These results suggest that transitory factors account for a smaller share ∞ 잉∞ g m ∞ 잉 i쩌잉 § mN a [ ∞. . this conjecture can be κsted by analyzing the degree of me.. R @ N [ @ 。。 R @ F .S.. {. 딛 { . contrary to 핸웰 성엌 Raug %uQt 쉰엌닙 되융따 g애김U” our conJecture.''-~~i~t~d portfolio. ) 염 m。 prices. F R @ .g딩μi ”””Q야띠 양 a8 잉염 륨환 ggEag} 【융。a -μ ~~.。 。. Mean reuersion in stock prices 44 J. data before 1925 show greater eVIdence ”될 m R ( g( gis 。。 @ F N。。) R。 m ∞ g m ) . The substantive importance of transitory components in stock prices S엉a ga얘강』딩 ‘。‘ 8』 협쇄 에【PC。 디애김삐@ This section assesses the substantive importance of mean reversion in stock ”。a어』 딩Ua3u ( μ (N m g * S m m m (N *륭 . 45 J. Summers. we use an approach that does not require us to specify a pro야ss for i。 )(< < 。) . {. T11ere is some tendency for more mean reversIon ln less bro때-bas맺 and {』퍼읍。딩 {검디。-eg”si녕 N∞ E U ””agaauu‘ -m 쉴읍 ( sophisticated equity markets.앞?nm:짧i앓bt템뜸 맏 *.。 F @ of the variance in relative returns for individual stocks than for the market as a . gga얘』&BER§ (홍 딩【 며그디잉g 』용aaμ3경〉 ( 얻얼 return. The 여ual-wei뱅ted portfolio 약 * { Ra‘”“‘ () . 。 of mean reversion than the post-1926 data. Many ).

‘.7% 0.(.712 37. data.15 12.i [ VR ( T )( 1 .' = T(J.30 12 . and PT is lhe 7:period auiocone.0'.6'.1112 11% -0.386 20. ca\c u13ted from to 아ry com 띠 1(ψponent may re바배flec더t fads . the transitory componenl accounts to argue for assuming that transitory components should die out rapidly.15 21. I'o(erhu alld 1.안 /GÏI ’‘ 1 - _2 . Given data on the va때Ice of returns over p%톨 O 아) 16. Meall rellt'rsion in s(o("/< priæs J.M. the variance of T-period returns is ‘ . A transitory component that is almost as persistent as a Table 6 presents estimales of the standard deviation of the transitorv random walk . !l9<’ lwo horlZOIIS T and Tf and assumplions about PT and pr. ‘ ‘ Each en ry reports he standard deviation of the transitory component of prices.30 34.0. For various degrees of serial C?rrelalion ai other hong ns. Sufficien tI y persistent transitory components will be unable to f“m때때l띠d메mgs are mse떼ive to our choice of P%: we 때Orl values of 0 .VR ( T')( 1 .500 fJW. and VR(T) for lhe 7:period variance :atio in rckill(1Il lo one-period returns. a pair of equalions 21. Results for value-weighted returns also suggesl 메 -‘사 제 -기 끼 -」 rji---. 0. 46 J. . 0.70 ,fo야 “I대P I2' ‘“thc twe 뼈e리Ive-mon o빼 ‘0이nt 때 n1띠i“빼 뻐h 띠 a뻐비 ultoco componen t.3% 0. Summer.PT ) T ’ (7a) depending on our serial correI ation assumption . measured al tíon may bc 믿gl annu때 rates (G. monthly data 깐nd the trans띠ry component follows a stationary process.7% 0.) .month) autocorrelation in the transitory price componen l.앙/셰 Value-weighted excess retums (J'. (7b) ’ U.8% 0. --. it is difficult P% = 0..2 is the variance 이 innovations to the permanent price ωmponent, 강 is the variance of 빠 Siamnaq componem .‘‘ t historical periods show patterns of variance-ratio decline similar to those in -샤 7ll -、4 -- .:. Using (J~ for the va따ncc of one-period relurns. and it has a slandard devia- tion of between 14% and 37%. .5% 0.(.\'. Summers. Po(erba and LH.S. More persistent transitory components are less able to accounl for declining variance ratios at long horizons. In cither case.7% 0.373 19.PT )(J. increasing the transilory component’ s persistence requires in- due 10 따 trr띠a꾀nstl …‘ l‘0 아 or 1πry cαcon 째)m매lψpone히야nls creasing the weight on the transitory component in relation to Ihe permanenl v써J펴lωue정so시fOι’ u 떠35’ and 0.6'.speculation-induced de야 씨l뻐at“ions of prices v eqs.li and account for low long-horizon variance ratios. l ‘ ‘ 터 -I -‘ -o -T -.2~‘ 센 h lhF form (6) can k solved to yield estimates of q2 and 강. will be unable to explain very much long-horizon ‘ompo따nl in slock p따es for the value-weighted and equal-weighted NY삶 mean reverS lO n. (7 a) and (7b) to match the observed pattem of variances in long.and short-horizon retums from fur빼mental v뻐es . and the share of the return varialion variance ratio.4% 0. pnce component is a way of characterizing the part of stock price movements that 때IIIOl he explained by changing expectaIions about future cash Rows. (6) 11% -0. Pl" ) T . . Since other nations and 세 n--R v T‘•• --------- T T o--u= ’ .0. return variation.2 and oj are given hy n3 = o . Pl2 . We begin by postulat- component raises both its standard deviation and its contribution 10 the relurn llIg ilull UI lS Serially uncorrelated al lhe horizon of 96 monlhs. 0.Pr ) T ._2 G. describing 따 stochastic properties of the stationarv assumptions about the 12-month (96.v 1'. we do not present parallel calculations for them.11% 0.1% 0. transitory componen l.. The permanent component evolves as a ‘ random walk Permanent and transitory retum componenκ U.0 Pl2 .. As one would expect. a substantial . 0.-/øÏl .0.4αl 21. This decomposi.(.. variance. _2 _2 1 . The diffeπnt cases of Pl2 (11%) correspond to different r바 urns. even if they account for all of the 0.00 9.S.M.70 Ulven our assumptions.687 35. 0.or it may be a consequence of changes in re폐lired ‘ Th e variance-ratio estimates that underlie this table are drawn from he entires for 96-month variance ratios for e"cess retums in table 2. Equ외-weighted ‘ exces‘ re ums lation 따 the statio때y co때onen t. 0. Many pairs of variance ratios and assumptions about the serial C\lrrelation Table 6 indicates that increasing the assumed persistence of the transilory properlícs of u.9116 1'% = 0. Mean revt' rsion in s(ock prices 47 We treat thc logarithm of the stock price as the sum of a permanent and a Table 6 trans Jtory c찌ponen t.0.35 PI2 . nations with 96-month variance ratios lower than those for the U. For lhe equal-welglued POIifollo.2 + 2(1 . have Jarger transitory components. we can then eSImaIe the .657 21.5 '1.456 where (J. though smaller.30. 11.664 0.369 12. for example. ωuld he analyzed hy using (7a)-(7b). as well as the share of retum variation due to transitory factors . porIfolk1s over the permd l926-I985 for various values of Pl1· assuming Which cases in table 6 are most relevant? As an a priori matter. estimates of o.(1 .554 fJW.Pr ) T’ l for between 43% and 99% of the variance in equal-weighted monthly returns ‘ (1 . To rationalize a given long-horizon vanance of thc t때sJtory component.S... (J.

Poterba a'ld LH. ing ret 써 s model ala the fads model would not hold in this case- . 값 a R(l 끼 e얹s잃않s 잃 1) Pπproc웠 a s post 뻐tu late혀d in Summers (1986). mo 짧 . Thi s has the 찌 rtue 01 띠비 with P96 = 0 . 퍼e assume for simplicity that the horizon.23 ex post returns (R . respectively. Campbell (1987) finds that appro피 mately 11 % of the variation ln ex야 ss retums can be explained on the basis of lagged information derived from the term Lt:. With geometric decay. th 닮 negative serial correlation. positive autocorrelation.on .igi않3많 없앓 피잉월f structure. (6) demonstrates . it does not suffer 1 + . 27 at 60 months . in stock h뼈av%e transitory tmd generate to account for the 벼 components orices. implying that the elements in the table correspond- ing to P12 = 0.0. that PT not be very negative over periods of transitory component follows a? up to 96 months. followed by 빼 er reduce share prices This will induce a holding for other large values of P12' Thi s is because actual variance ratios decline returns follow an AR(1) returns. although it is inconsistent with the observation that actual returns If required returns snow stationary component of -0. Keim 뼈d Stambaugh (1986) find that between 8% 뻐d 13% of the th1S 며 말닝 lin <l lf variation in retums for a portfolio of stocks in the bottom quintile of the NYSE can be predicted cormlatlOIl in dlRRent Stocks may be attributable lo a COInFIOn ‘ using lagged information. papers . exacI parallel belween lhe ilme-vary- whICh however fraction of retums over longer horizons. the써1 innovation that raises .:dgta댐.274 at 84 months . and inlerpret렌 v using various information sets. about the future bath of requi 때 returns (~,), reft 맺 future dividends. is satisfied only for cases where P12 is large. imposing P12 = 0 . it must become negative to account for the obseπed pattem. Larger autocorrelations at short horizons do not necessitate such pattems. 21 If changes in r여 uired returns and profits are positively ∞ πelated , then the Any stochastic process for the transitory price component can be mapped into Wlll understate the varlance m a stochastic process for ex ante returns . 20 orthogonal to innovatio where L a serially uncorrelated innovation that is 면 s revisions 10 expecte<!.744 at 36 months.: Previous claims that there are fads in stock prices have typically suggested t half-lives of several years .ces J. Summers . l+g Insofar as the evidence in the last section and in Fama and French (1 988b) R. 앉 an& returns I1랴ded to rationalize ’ mean reversion in stock prices. lt is interest rates examples in which pro6ts and The central issue is whether variations in ex ante retums are better explained DoSsible to construct theoretical ZI·e negatively related. Several recem .: Transitory components in stock prices imply variation in ex ante returns. and Ross (l985) Silldy lhe pricing of assets and assuming constant required retums is controversial.gsSRZsc파찮e띔댐찮 넓34:었 짧k3깅S뚫a$$sb잖. A reasonable restriction.n stock pr.0 . Campbell Kyle (l986). Shiller’ s con cI usion that market retums are too volatile to be reconciled with valuation models wiih lime-varylIlg 22Lucas (1978) and Cox.r) is persuasive in su잃esting the presence of transitory components in stock prices. as in Campbell (1986). Since our analysis does not rely on the present-value . Mean revers. Summers.M. including Black (l986) . The average dividend 센터 d and S. (η+1 .48 J. 뻐dS비빠 21 Several recent studies have cχ)nsidered the extent to which equity retums can be predicted possible lilt1ueIm 때 ers (1984) .r) + rl ' (8) relation between stock prices and expected future dividends.:~ P1 (I + 홍) from some of the problems that have been highlighted in the volatility-test debate.70 and P96 = 0 .) are given by this requires declining values of PT' If PT starts small. see West (1988) for a survey of recent work.. Fama 빼 French (l 986b)fac or.。3.-R 르 1+. and show poShive.1 73 .f noise DeLong et 떠. when P12 = 0. The source of the transitory component in stock prices dividend growth rate are d and 흥, respectively.$C. reQUired returns.3 5 yields an implied autocorrelation for the tractabiliiv. The appendix shows that when required between long and longer horizons. this section’ s results confirm S비ller’s (1 981) conclusion that models assuming constant ex ante retums cannot account for all of the variance in (1 +. One other consideration supports large values for P12' For given values of 쉰 and 강, eq.M. this s뼈gests a half-life of two years. and as eq. have dlscussed the oF 짜 ow that the ne security prices 뼈d r여 uired retums. In contrast. IngersolI.1 68 and .. but the eI삐 20.tt 밟&훌말쁨 i 짧 :cg 삶證positive nC짧felllrn iJt2:낌 찮The때꽤?따‘ S can be accounted for in this way. Similar resu1ts obtain period loss. the implied required returns will values of P36 and P60 are 0 .70 are most relevan t.:)-1(1 +홍 )2 stock market retums. For example. (1987) . A much smaller share of the variation in retums to larger companies as support for the time-varying retums view of mean reversion. and any pattem for ex ante returns asslimntion that F and tt are orthogonal can be represented by describing the associated transitory price componen t. = d + 용· . rationalizing process . (6) permits us to calculate PT over any th 멀i뭔l3z g. Poterba and LH.:-P1(1+ 용) (η. -1. in steady state. Fama and French (1988a) find that lagged dividend yields can predict a much higher ret1likes the exnected excess to be . Mean reversíon ín stκk príces 49 gx.l:Z생$임C:앓짧꾀랩obseπ fa짜 acc다tOl아띤 expected returns that 뼈 歐빼 r.

4% 5. wilh :1 di‘ l‘knJ 、 icl,1 of 4.'tulll‘ Ill'l'J hl h\._.: .'rl>“ ‘”’ d 1. SlOn in prices.li.ln 11\ L. 1926-1985 period. Even larger amounts of required return varialion are needed components (.'.)2(1 +.111 in sl “ 'l príιc‘ Uur assumption that required returns are given hy η -r= (1 -p L)' l~, .'rn illlpli.‘ F‘'.:)-1(1 + 흐 f/[l +r-PI(l + 치 J..'n "f r. implied by a variely and market volatilities‘ tìnd no evidence that stock pri l'cs h e\ p to f 'rel'asl ."!'I.L)(R . (K).. These estimates of the standard deviation of re4uircd relurns are ( 10) large in relation to the mean of ('‘ post excess relurns and imply that if ‘’‘ t1I l(‘’ {(1 + J)(l + 미).:)2 persisten t.l . ~~’ Pro 뼈 ‘0 1 o얀{> O. lhe standard deviation of ex ante returns mllSI be negative serial correlation .. th.'aku l:llions arc cωi퍼li싸l싸ihr:띠‘a싸 i The first-order autocova때ce of the expression 011 the 때lI -hand side of … tlhc 1n6-19 !1. .!1‘F 7.if we poSullate that the slandard deviain1Il Of Ihe The second diftìculty in explaining the ohservcd correlation pallerns with transítory price component is 20% ‘ then even when required return shocks models of time-varying returns arises from our tinding of positive fo l\owed hy have a half-life of 2. nlc‘C ‘a씨‘Cfi,찌‘a앤g" "、X I,.\'1 fl'lurn “‘’ c th l'n :1“ \\h‘It ‘:lhl" “ f ". white noise . It shows the standard deviation reqliired returns . 떠re히tu 바rn떠s fo l1 ow 때 a n ARιMA( l.' vnl‘Il ilil~ pallcrn‘ ’‘ providcd hy Frcnch.N'l marized in eq . f 'n..' ’t" ‘XI.' ‘’‘ l’‘. aIId Wrobleski (1977) prove lhal an alllocorrelogram with zero tntries heWIld order k implies an MA(k) proces~ !bConlrary cvidcncc suggesting Ihal ‘ lock rClurns J" prcJicl flllur .' 7 ”혀k negative correlation between shocks to cash tlo\\'s and re배 ired returns. We use estimaies of the variance prediction .9 vears 4.‘ I ‘’“IIr‘'('t".'‘l as rcturns. ('r’‘·η ‘l positive correlation bet\\’ een bond and stock returns suggests either posit Ï\'e or T:lhl.'r\' pnιc palt .PI(1 + gW(I .“1\llII n £ 프 -~'+I(1 + . but a l1 higher-order autocovariances equal ~ero. 1 ‘ ---AInky.' 씨’l1 l' tint. 9% pcr ycar.luld thc stanJarJ dc\'Ìa ti.ld i‘ K.<I rcturn‘')' Our ,'alιulati‘、ns "lIIpl‘、、 th . l) proc ’lCe야ss: if C1 = O.5'f … … … “ 띠rn ιf,‘l、πlfr tlhli i‘ (인 ís nonzero.an rationalize the second but not Ihe tirst of Ihcse ohservations ‘ It is instru .P.' fι‘ t Ih:1I “ ilh AR(I) <<'q \l ir. generate the same time-series process for ohseπed returns as fads of Varil)US sues. For example.M. 9 vears 6 . they are also large in r e\ ation to traditional views of Substantial variability in required returns is needed to explaín mean rever. … corporale equ es was 0.7‘r. thaI Ihe required reIurn variaI1야 corresponding 10 a given fad variance is to explai l1 the same size price Cads when required rcturn shod.(I~ 25.. S"1/I'"‘’r ‘.t'-l' (9) . of fad models.0’} 1‘’ ‘’‘i‘ The si띠ple model of stationary and no찌 ationary price compol1 ents sum.. 1.O'i.s arc Icss C1_2 = [1 + . 4 \lcars 7. then re'turns are i Slan‘iard Jc\'iali"ll "f trall서 tl'~ 'llIUpl)n 'nl .P.' ‘.5 pcri미‘띠 and arπc hascd "…n cq ‘(씨‘ H0씨)) in ‘th .9't 1O. . hcading.:n.22 ror Ihe 1959-1979 pcriod and 0. ""an . -R) 르 흐 + t . (5) also yields an ARMA(1 ‘ 1) representation for returns . market efficiency .f‘II ,~ Ih. 1O. 1f> Although thcy show that slod.fl‘”’ ”’ ”‘κ-k pr. 20. P‘’I .2~‘ 12. the sign is :ounter 10 Ihc hcory 's ‘ ‘ an average dividend growlh rate of 44%. .l't !I.‘'t".8% per annum. Eadl cntry answcr‘ th l' qU l'stion: 'If hoth r.~" Anwunt nf 、.9 years. This 1.' hl g. On the other hand ‘ if the transitory componcnts are vicwed as a ratío at 96 months to calibrate the degree of mean reversion.'’”‘ flπ‘t이?끼tur pcn‘.unl “ r 1tl.‘S\l IIICJ t" f,'\l‘、“ an AR( 1) that yidds a pr‘.\f‘’“" n'I','r. Campbel\ and Shiller (1987).'quircJ r. defining autorcgrl'‘si,'n~ “ 'ith halr. S"''' ‘’r ‘.1979-::'ï 9iõ period. ohSl'f\'l'J rcturns is ιonS1‘ tcnt ‘ ith a priιc faJ ‘ ith a stanJarJ ‘k、 l‘lh,ll\ 끼’‘) g’‘. reflection of mispricing.1 h、 n“ Thc (l-p .''f‘u.In Ihe appeFdix we show.셔 1 + (1 + ,ï)까 (1 + g)~ 낀 relurns are never negative they must frequently exceed 20%. P >ter l>a ‘”’‘I L lI.'n in th .:‘ procc‘ s for ‘’‘ ('.11.(1 + J){. l ‘\.sl rcturn‘ prncl'Ss is givcn h、 cq.6'{. Silllilarl、 thc priι" f:lJ i‘ a.'I,. 1.1‘r allows us io calculale the varialion in required reIurns that is needed lo 2.~1‘”’ ”’ ‘“’‘ J. what ’‘. “、r illlpli디 tl\1 악 is nccdcJ t" gl' ncratc a giwn ‘ 1Zl' tr‘lIl Slt . We calibrate Ihe calculations using tl1e average excess RIllrn future movements in discou l1 t rales ‘ as they should if stol'k price movcmcnls (89% per year) on the NYSE equal-weighted share price index over the are caused by tluctuations in these factors. The dividend yield on lhese shares averages 45%. ~(l ‘ J .l'n,.Spivey.:-s. anJ th . Sc hwerl. .\1.'an f""\'l'r‘ Î4.tive to consider what type of cxpecled retu fIl s hehavior is necessilry !4Campbell (l 9R7) eslimales Ihallhe correlalion belw~엉ncxcess relurns on long-tcrni honds anJ to account for both observations. 1i、 cs ind ’“at l'J in thl' ro“ lII argin. implying prices do forecast consumplion tluctuations."!'{.1~ 11 .. measured on an annual basis.'quirúl r.파iation in r,'q버 rcd rcturns nl'l'dl'd t" ‘l‘:ùlunl f~\r nh.'turns ‘mJ priζ" f‘IJ‘ f"ll‘”‘ ’ir‘’-‘、rJ,'r enables us to reι-rite (8) . using data 011 rcal intcrcsl rates of required excess reIurns.36 for Ihe morc~,:~~'.( 1/I J. 삐d Slamhallgh (1째 7). It is diftkult to think of risk factors that could accounl for SUdl varialion in Table 7 reports calculations based on (10) ..'CS‘ Iilc (5) for ‘. Models with first-order aUlorcgrcssive transilory 5. Half-lirc lS.' ‘1nh". '.

An aitemative possibility is because increases in expected dividends. 568 + 0. POlerba and LH. R . while the second process exhibits positive to stock retum data in an efforl to extract more information about the autocorrelation at 떠1 lags._. then negative.75 . It may also 27‘ Two exarnples of required retum pr,α:esses are twelfth-order moving-average processes wi th justify some institutions' practice of spending on the basis of a weighted the following coefficients: 1. emerges from data on equal-weighted and value-weighted NYSE retums over (1. This could lead to positive autocorrelation at short horizons age representations show declining coefficients. ‘. may be less risky than it appears to be when the variance of single-period lation of ex post retums over short horizons...75.5 . but the results vanance m retums.844 • IDI V. an investor’ s horizon will autocorrelogram of ex post retums reftects the dynamics of required retums. weighted index on lagged values of IDI κ.IDI V. 0.75 .061 • IDIκ. is shown below.5 . POlerba a.5 . 째 6.75 . 5.75 . is measured in percentage points and retums.75 average of their past endowment vnlues. 3. autocorrelation in ex post retums will also generate this pattem in ex post ing six lagged values.5 . Ö. (1. 4. 380) . volatility such as French. and is corroborated by data from other nations and time periods.. which would raise share prices. πe autocorrelograrn of the former process displays Although the temptation to apply more sophisticated statistical techniques positive. A representative equation. -0. 1. 0. If stock price movements contain quent returns. and a set of in equilibrium must equal 1 . Th e presence of have identitì.nding of signitì. Mea. 0. mag띠tude and structure of transit‘ π components is ever present .이) that follow a moving-average process similar monthly dummy variables.1. they su잃,est a negative but Our tì. Th e notion that noise trading impulses intensify and then decline standard errors are given in parentheses: comports with qualitative discussions of fads . a time trend . retums is extrapolated using the random-walk model. First. such as those considered by DeBondt and Thaler (1985). assume that the required retum of sophisticated monthly ‘ dividend innovations' (IDI κ) for the 1926-1985 pe디od as the .. 52 J. We explored this possibility by forming of noise traders. Summers. 3. Although individual data sets do not consistently permit 띠ec­ R 2 = 0. but further work is clearly necessary to evaluate this conjecture..cant relationship between dividend innovations and subse. If the investor’ s relative risk aversion is Some required-retum processes could generate positive. Mean reversion i. 80th processes generate positive. . Our point estimates The coefficients on lagged values of IDIV should be positive if required suggest that transitory price components account for a substantial part of the retums and prospective dividends are positively correlated . Samuelson (1 988) dem- The se∞nd potential explanation for positive se디al correlation is that the onstrates that in the presence of mean reversion. -0. greater (less) than unity.cance levels.380) ‘ (1. 5. 1. rather than current market value.75 . tion of the random-walk hypothesis at high signitì. would that movements in required retums are due to changes in the equity demands be fo l1owed by higher ex ante retums. in slock prices J. This would suggest that positive dividend news is followed bv large transitory components. 377) . and Stambaugh (1987) or Poterba and shocks to required retums and to prospective dividends may be positively Summers (1986) suggest that shocks are persistent but that their moving-aver- correlated. . 3.ed all show increasing coefficients in some part of their moving transitory price components also s맹gests the desirability of investment average representationY We are unaware of evidence suggesting that obserι strategies. Studies of observed retums at short lags.d LH. 2. 0. For example. ~ . 2. slock prices 53 There are two potential explanations for the positive autocorrelation in able proxies for r어uired retums display such stochastic properties. -0. R . contrary to our maintained specitì. If anything. provide no support for this view. aUlocorrel~tion in ex pOSI relums. then negative. we doubt . 2.04이 (1. traders is equal to a + ß피, where S is the fraction of the outstanding common residuals from a regression of real dividends (on the value-weighted NYSE stock that these investors must hold. (0. 1927:7-1985:12. Equity demands of noise traders (which portfolio) on twelve lagged values of real dividends .cant transitory p디ce components has potentially im- statistically insignitì.75 . -1. The required-retum processes with this feature that we equities than he would with serially independent retums. influence his portfolio decisions.0 .cation.1 09 • IDI V. 0. followed by negative.nancial practice.. ∞rrelation in required rcturns. This conclusion . as his horizon lengthens he will invest more (less) in return autocorrelation.. 4. Summers.m v 빼 mWI i l l t --’ J i v • l u ·‘ -- / n * 1 ‘ Our results suggest that stock retums show positive serial correlation over 、 、 short periods and negative correlation over longer intervals. a pattern that should be reftected in negative autocorre.‘ .5. involving the purchase of securities that have recently declined in value. the various data sets together strengthen the case against its validity. 0~75 , 0.037 . We then regressed real retums on the value. Conclusions 빼 ‘‘ M .769 . then for long-horiwn investors the stock market lower required retums. 4. 374) the 1926-1985 period . reversio. an이 1. then negative.M.= 1. includ. Schwert. to one of those for required retums that generate positive. portant implications for tì.5 . 1.M.

-D (l + 홍) ∞ . T'J .+ 잠뚫 [η+j. .I •• -. 서 패 니 0 ’ . supported 홀鐵. \ /-:0'.d+D. t PI l l E ‘.2) to obtain Appendix . D. as well as theories of changing 디 sk factors .leidon’ s (1986) results .g) as P.:) +건.E ,( 화 j lO'T (r. 용vJ}l} [g'+i- where 끼 +i denotes the required real return in period t + i . such as data on fundamental values..Igt+l.(1 + r) U.:.흥) j~O The price of a common stock. l 'TJ J} nents in stock prices. R" is given by available as of period t. fraction of relurn varialion on Ihe basis of observable news aboul Culure cash Hows or discounl rales..-P.M.. proxies for noise trading such as the net purchases by odd-lot traders .2) We suggest in the paper’ s final section that noise trading. Summers.:)/(i . + 펄E,{흘ßj[g,+j-gl} Evaluating such theories is lik e\ y to require information other than stock returns. provides a plausible explanation for the transitory compo. or the level of participation in investment clubs.+I: J. ‘ l + (A .. g . (A.r) = Pl(η-1 . We lineaTÏze inside the expectation operator in μ . 28 Pursuing this wi\l involve constructing and testing theories of noise trading.홍) -.. we can simplify the second term in (A. As the debate ap. 니 1 r 까 . Only by comparing models based on the presence of case of noise traders with models based on changing 디 sk factors can we judge whether financial markets are efficient in the sense of rationally valuing assets. Summers.S) R ,= 수 τ Cutler. Mean reversion - that a great deal can be learned in this way. r) + ~I' (A.M. 28 P. ‘‘ • +a D. J. 7 . In the special implied by stock options.4) --시 + ∞ r Z .:)(.l) l l l . and indicators of risk factors such as ex ante volatilities where ß= (1 + 흥)/(1 + r). . and g. Polerba and LH.D. is that such Monte Carlo analysis of much of the more elaborate work on stock-price volatility would reveal poor statistical + properhes.} designates expectations formed using information Now recall that the holding period return .~'~"\I'"ð'=\E,{ Eßj(r.rl 1 ∞ over volatility tests has illustrated .홍 )(I+.홍1) by K.. sophisticated statistical results are often very sensitive to maintained assumptions that are ilifficult to evaluate.종 (1+.{ . POlerba and LH. that could account for the characteristic stock return autocorrelogram documented here..:] Derivation 0/ ex post return process when required returns are AR(l) P.:.:-Pl(l+g))' r-g-'\j~o'- .‘ jl l ‘ 껴 (1 +흥 )(η-.3) as precluding the generation of excess profits.흥 l} (A . Poterba. as well (η . trading by investors whose demand for shares is deterrnined by factors other than their D.+i is the real dividend growth rate between pe디 ods t + i and t + i + 1. turnover. 르 Eß 해 Iη - (1+. We have validated the statistical procedures in this paper by applying them to pseudo data conforrning to the random-walk model. PI' equals ”n ’ + ‘l 필 E,{흘 ßi(g'+i . Our suspicion.)(.. Even the broad characteristics of in slock prices -톨.+j-r]} l expected return .. We denote D.(1 + . and Summers (1 988) doαlment the difficulty of explaining a significant →-1. is the dividend paid in period t . (A. Mean reversion in slock prices 55 the data exarnined in this paper cannot be estimated precisely.\J. D (1 + 흥) ~ f ∞ ,. R 응 E,(롤(뽑r. and E ..

며).Pl (1 + 흩rr(η+1 -. 1986. Price movements in speculative markets: Trends or random walks . We find them by equating the variances and first-order autocovari- R . and rational expectations. ~ D.Pl l!. P'+l-P'+1 ( Pt+ l + D. . 1) process to be the same as (A.홍 )(1 +. equals 상/(1. Jouma1 of Financia\ Economics 18. (A.β (l-PI L )-l테 + ι , Alexander.샤1 + (1 + d)2) 0.r) _ß-l(η . Crises in the economic and fin뻐cial structure (Lexington B∞ks, Lexington . and Mark Watson.-R) 프 E. This ~ields an ARMA(l .. 10hn Y. . Bubbles. .M.12) 회 -냉 1 -n-+ i」 l --페 -때 + ---‘ - --메 --! 7 --T rv ’ l l l I . ed. 1) representation of retums. -1 + 까 . Quarterly 10umal (1-P 1 L)(κ. and Wi1liam Wrobleski . 1987.ll) we find 디 (1 +. Poterba and L H.-gj' -1 + t . 56 J.l 4) x [(η+l.Pl (1 + 흥)]H, +l = ξ , we can multiply Black. in: P. 르 R+ 부화수-.M. 1+. MA) 295-315.) ance of the right-hand sides of (A. W.6) We now explore the parallel between the time-varying retums model and the fad model ..g} (A. Craig.t_ ’ . and the last expression exploits the fact thatj1 .찌 르호 -1\1낯냐 . (8) in the tex t. 1982. . "'-1' (A.:. Since (1 + . 1977. .Pl)2(1 + r)1 7 (A .9). 785-803 Campbell .9) of Economics 102. Bond and stock retums in a simple exchange model .g)[l +. Fischer.p . ..7) 얀 (1 + d)(l + 야). Mean reversion in stock prices r、」 J.P1L) so that Blanchard. Now defining {ß(l + 흩)/[1 + .1이 yields For this ARMA(l . -R 프 p'(r. Sidney S. .1 3) that the variance of required returns -. 10uma\ of Finance 41 . P,(r.-Pl(l+ 용 )]2(1 . Summers.Pl(l + 용)) ? l+.F) = &I processes. 121-134. .9) and (A . 373-400. 1961 .8) β(1 +흥) πÚS leads immediately to (10) in the tex t.Pl (I + g) References X[ (I -PIL)-I~'+l. which postulates that retums evolve according to where R+ I = (1 + a)R and R = F(1 + a)/(1 + F)· Substituting (A. Noise . two restrictions must be + 흥) Dt+ l(l satisfied.헨τ웅I 짚 ßJE, {g'+J .12) to eliminate 상 from (A. 파 {(1 +d)(l+ 야). On the structure of moving averagc and t + 1. Industri a1 Management Review 2. Poterba and LH.Pl)2 、 .L x (1 + d) . 529-543. Using (A.13) (A.1 0): [1 + (1 +d)2] 랜 + (1 + 며 ) 상 =2상 + (1 + 야 ) 앙, IJ 0。 (A. Olivier 1.. Stock retums and the term structure .4) into (A. g} ” 、 m + f ----. through by (1 . Rz (R-R (A. ”+ (1 + d) 랜 + Pl상=상+ Pl앙.(1 + 흥)2 corresponding to a given fad variance is R.) ∞ (1. Mean reversion in stock prices 57 It can be linearized around 껴 and P +l as follows: .+l+J . 10hn Y. ‘ι ‘ (A.ll) + 닮찮침 ßJEt+ l {g . this is eq.셰 + (1 + d)2]} (1 + g)2 따 (A.r)] + t .PlL)(1. Recall that thξvariance of the fad . Using this and the definition of ~" we find from (A . Summers. 10uma1 of Econometrics 6. Campbell . 1986. R.)/(1 + 흥)'" n . Al1en Spivey. Wachte1. 7-26 where L reRects changes in expected future dividend growth rates between t Ansley. crashes.-R 르- .- This can be rewritten as 0.6) (l-PI L )(R .

The Samuelson. and J. Are ou‘P띠 fluctuations transÌtory?. 639-655 Journal of Financial Economics. MA). 916-918. Asset prices in 때 exchange economy. aIId LaWIfIl@ H.I 58 • . Summers. Jqhn H l988. . 1986. Note on the corre1a tion of tìrst diπerences of averages in a random French ‘ Kennelh R. fads. shiller. Mean rer'ersion in stock pr. 1960. Cutler. 1981 . . 1988a. ‘ Hu~nga. Cambridge. I986b. David C . 1973 . 2395 (NBER. Operations Research 7. Quarlerly Merton. and Robert I. Officer. . 1987. Cambridge. Joumal of Po\i tical Economy 96 . American Economic Revíew 76. 1142-115 1. William Schwerl . 246-273. prices. James M. Dombusch. and Robert Waldman. Mimeo. 삐dRφert Stambaugh. Poterba and LH. The variabi \i ty of lhe market factor of Ihc Ncw York Stock Ex change. Chicago. Joumal of Financial Economics 19. On estimating Ihe expecled rctum on Ihe markel. and Lawrence H. Stock retum variances: Th e arrival of information chain. French. 1981 . PnnCeton. Gregory 'Mankiw. 793-805. Campben. John . Longrun risk tolerance when equily relurns are mean regressing: economlc con않quences of noise traders. ~uture d. 857-880 Economics 8. Fama. Washington. Eugene F. Stock pri야s and social dynamics . CUller for Research in Security Prices working paper no. 3-30 Harvey. (Board of Cox . (Massachusells lnslilule of pap~r no. Economics Lellers 18 . Roberl . Fischer . 363-384. . I986a. Do stock prices move too much to be juslítìed by subsequent changes in of Finance 25 . John Y. 1987. 323-361. Macroeconomícs and tìnance: Essays in C@hrane. Mímeõ. Fama. honor of Franco Modigliani (MIT Press. Joumal of Financial Eιonomics 17. Eugene F and Kennelh R French. 뻐d S~cp~e. John c. Summers. . . Summers. and stock p다ce volatility: A parlial evalualíon. Chicago. Summers. 1970. Does t11e slock markcl raliona \l y reflecl fundamenlal values? sl@k RIllI1ls .. forthcoming Osbomc. 1984. Economelri :a 66. 5-26. Perron. Robert C. and N. An empirical investigation'of he lon . Lawrence Summers. Andrew. and Ihe reaclion of Iraders.M. Joumal of Fa~a, Eu~en_e _F. 1429-1445.Iohn Y. Janles M. lournal oI P이itic허 Economy. Kenncth . 1987. Frcnch. A comparison of annual common slock relurns Fama: Eug~ne F. . Jonathan Ingcrsoll. 145-173 DeBondt. Robert R. MA). Jack W. 381-386. 434-453. Eugene F. Campbell. 1980. 591-601. 1987 . model of asset prices. 1976. Testing the random walk hypolhesis: Power versus Chicago. Kyle. DeLong.- J. . MA) 93-124 forthcoming O'Brien . Efficient capít외 markets: A re씨ew of thcory and empirical work .n~e~h_R. Permanent and lemporxy componenls of sl@k 2.' c:n: ín: R. . 1988. James M.. Iohn Y ald Albert s. New Hav~. whal moves slock prices?‘ Joumal of Business 46. Expected stock retums and stoçk market volatility. D@S the stock market overreact. IL) West . . Robert J. Econometrica 28. and P. 1978. M. Mi meo. Frenc~,. 239-258. pnces. Bubbles. Time series mode\s (Philip AII삐, Oxford). and _~nneth R. Smart money? I1Olse lradlng. market retums . I78 (University of Chicago. IL) frequency of observation.. Cambridge . Robert 1. Journal Shiller. Holbrook .M. poieIba. 1988. Working. FIencl1. Robcrt c. Centr for Research in Secllr1Iy PriC£s working paper no 200 (UIliverslty or Joumal of Financc 41 . . Bradford. Shiller. 1988 . NI) ‘ Lucas. The dividend-price ralio and expectations of Merlon . American Economic Review 71.~ ~~s~, 1985 . (Yale Universiiy. Wilson . Werner and Richard Tbaler. Unpubllshed manuscripl (Princelon University. this issuc. l986. and Kennelh R. eds . 383-417 dividends? . An intertemporal general equilibrium Govemors of the Federal Rcserve . Foundations of tìnancc (B앓ic Books . DC). Bossons. Di vidend yi리ds and expectcd stock retums . and stock pnce 훌훌 • J. Robert J. Joumalof Portfo\io Management. Poterba al/ d LH. Nation외 Bureau of Eco nomic Rese하ch working Pseudoparadoxes and vindication of ‘ business man’ s risk’.ces Campbell. l985. NY) Shiller. Journal of Business 60. Lawrence H. . 1987.F. Paul A.421-436 Fama . On Ihe slale of the efficienl market hypolhcsis ín tìnancial economícs. How big is the random walk in GNP?. Journal of Finance Poterba. S.iv~de~~s ,~~ discounl factors . and Richard RolI. Kennelh R. 1986. 1959. 1987 . New York. Brookings Papcrs on Economic Activily Fama. 1985 . Common faclors m lhe senal correlation of Summers.an? _K~. Econometrica 53. and Charles Joncs .t ~88b' ,Permancnt and tcmporary components of stock 1871-1925 wílh 1926-1985 . Testing for Iransitory elements ín stock prices. Technology. Financc 43 . Andrei Shleifer. Mea l/ re!'ersio꺼 in sfock prices 59 behavlOr. J. The persislence of volalilily and slock 40. 1986. Brownian motion in thc slock markel.Eueene F. . 457-498.. 1987. 1987 . Journal of Financial Joumal of Economics 102 .