<<

Journalof Economic Perspectives—Volume 17,Number 1—Winter 2003— Pages 83– 104

FromEfŽ cient Markets Theoryto BehavioralFinance

Robert J.Shiller

cademicŽ nance has evolveda longway from the days when theefŽ cient marketstheory was widelyconsidered to be provedbeyond doubt. Behav- A ioralŽ nance—that is, Žnance froma broadersocial science perspective includingpsychology and sociology—isnow oneof the most vital research pro- grams,and itstands insharp contradiction tomuch ofefŽ cient markets theory. TheefŽ cient markets theory reached its height of dominance inacademic circlesaround the1970s. At that time,the revolution in economictheory was initsŽ rstblush ofenthusiasm, afreshnew idea that occupied thecenter of . Theidea that speculativeasset pricessuch as stock prices alwaysincorporate the best information about fundamental valuesand that prices change onlybecause of good, sensibleinformation meshed very well with theoret- icaltrends of thetime. Prominent Ž nance modelsof the 1970s related speculative asset pricesto economicfundamentals, using rationalexpectations to tie together Žnance and theentire economy in one elegant theory. For example, Robert Mertonpublished “ An IntertemporalCapital Asset Pricing Model” in 1973,which showed how togeneralize the capital asset pricingmodel to a comprehensive intertemporalgeneral equilibrium model. Robert Lucas published“ AssetPrices in an ExchangeEconomy” in 1978, which showed that ina rationalexpectations generalequilibrium, rational asset pricesmay have aforecastableelement that is relatedto the forecastability of consumption. DouglasBreeden published his theoryof “ consumption betas”in 1979, where a stock’s beta(which measuresthe sensitivityof itsreturn compared to some ) was determinedby thecorrelation y RobertJ. Shiller isthe Stanley B.ResorProfessor of andalso afŽ liated with the CowlesFoundation and the International Center for Finance, Yale University,New Haven, Connecticut.He isa ResearchAssociate at the National Bureau ofEconomic Research, Cambridge,Massachusetts. His e-mail addressis [email protected] . 84Journal of Economic Perspectives

of the stock’sreturnwith per capita consumption. Thesewere exciting theoretical advances at thetime. In 1973,the Ž rstedition of Burton Malkiel ’sacclaimedbook, ARandomWalk Down Wall Street, appeared, which conveyedthis excitementto a wideraudience. In thedecade of the 1970s, I was agraduatestudent writinga Ph.D. dissertation on rationalexpectations models and an assistant and associate professor,and Iwas mostlycaught up inthe excitement of the time. One could easilywish that these modelswere true descriptions of the world around us, forit would then bea wonderfuladvance forour profession. Wewouldhave powerfultools to study and toquantify the Ž nancial worldaround us. Wishfulthinking can dominatemuch ofthework of aprofessionfor a decade, but not indeŽ nitely.The 1970s already saw thebeginnings of some disquiet over thesemodels and atendencyto push themsomewhat aside in favor of a more eclecticway of thinkingabout Ž nancial marketsand theeconomy. Browsingtoday again through Ž nance journals fromthe 1970s, one sees some beginnings of reportsof anomalies that didn ’tseemlikely to square with the ef Ž cientmarkets theory,even if they were not presentedas signi Ž cant evidenceagainst thetheory. Forexample, ’s1970article, “EfŽ cientCapital Markets: A Reviewof EmpiricalWork, ” whilehighly enthusiastic initsconclusions formarket ef Ž ciency, did reportsome anomalies like slight serial dependencies in stockmarket returns, though withthe tone of pointing out how smallthe anomalies were.

The1980s andExcess Volatility

Frommy perspective,the 1980s were a timeof importantacademic discussion ofthe consistency ofthe ef Ž cientmarkets model for the aggregate stock market witheconometric evidence about thetime series properties of prices,dividends and earnings. Of particularconcern was whetherthese stocks show excessvolatility relativeto what wouldbe predicted by the ef Ž cientmarkets model. Theanomalies that had beendiscovered might be considered at worstsmall departuresfrom the fundamental truth ofmarket ef Ž ciency,but ifmost of the volatilityin thestock marketwas unexplained,it wouldcall into question the basic underpinnings ofthe entire ef Ž cientmarkets theory. The anomaly representedby thenotion ofexcess volatility seems to be much moretroubling for ef Ž ciency marketstheory than someother Ž nancial anomalies, such as theJanuary effector theday-of-the-week effect. 1 Thevolatility anomaly ismuch deeperthan those representedby pricestickiness or tatonnement oreven by exchange-rate overshoot- ing.The evidence regarding excess volatility seems, to some observers at least,to implythat changes inpricesoccur forno fundamental reason at all,that theyoccur because ofsuch things as “sunspots” or “animalspirits ” orjust mass .

The efŽ cientmarkets model can bestated as assertingthat theprice Pt of a

1 Agooddiscussion of the majoranomalies, and the evidencefor them, isinSiegel (2002). RobertJ. Shiller 85

share (orof a portfolioof shares representingan index)equals the mathematical expectation,conditional on allinformation available at thetime, of the present P*t ofactual subsequent dividendsaccruing tothat share (orportfolio of shares). P*t isnot known at time t and has tobe forecasted. Ef Ž cientmarkets say that priceequals the optimal forecast of it. Differentforms of the ef Ž cientmarkets model differ in the choice of the discount ratein the present value, but thegeneral ef Ž cientmarkets model can be writtenjust as Pt 5 Et P*t, where Et refersto mathematical expectation conditional on publicinformation available at time t.This equationasserts that any surprising movementsin the stock marketmust have at theirorigin some new information about thefundamental value P*t. It followsfrom the ef Ž cientmarkets model that P*t 5 Pt 1 Ut, where Ut is a forecasterror. The forecast error Ut must beuncorrelated with any information variableavailable at time t,otherwisethe forecast would not beoptimal; it would not betaking into account allinformation. Since the price Pt itselfis information at time t, Pt and Ut must beuncorrelatedwith each other.Since the variance of the sum oftwo uncorrelated variables is the sum oftheir variances, itfollows that the variance of P*t must equalthe variance of Pt plus thevariance of Ut, and hence, sincethe variance of Ut cannot benegative, that thevariance of P*t must begreater than orequal to that of Pt. Thus, thefundamental principleof optimal forecasting is that theforecast must beless variable than thevariable forecasted. Any forecasterwhose forecast consistentlyvaries through timemore than thevariable forecasted is making a seriouserror, because then high forecastswould themselves tend toindicate forecastpositive errors, and lowforecasts indicate negative errors. The maximum possiblevariance of the forecast is thevariance of the variable forecasted, and this can occur onlyif the forecaster has perfectforesight and theforecasts correlate perfectlywith the variable forecasted. If onecomputes foreach yearsince 1871 the present value subsequent tothat yearof the real dividends paid on theStandard &Poor ’sCompositeStock Price Index, discounted bya constant realdiscount rateequal to the geometric average realreturn 1871 –2002on thesame Standard &PoorIndex, one Ž nds that the presentvalue, if plottedthrough time,behaves remarkably like a stabletrend. 2 In contrast, theStandard &Poor ’sCompositeStock Price Index gyrateswildly up and down around this trend.Figure 1 illustratesthese patterns. How, then, can wetakeit as receiveddoctrine that, according tothe simplest efŽ cientmarkets theory, the stock pricerepresents the optimal forecast of this presentvalue, the price responding only to objectiveinformation about it?I argued inShiller (1981), as didalso StephenLeRoy and Richard Porter(1981), that the stabilityof the present value through timesuggests that thereis excessvolatility in

2 ` ( t2 t) Thepresent value, constant discountrate, iscomputed for each year t as p*const, t 5 ¥t5 t1 1 r Dt , where r isa constant discountfactor, and Dt isthe realdividend at time t.An assumptionwas made aboutreal dividends after 2002.See note to Figure 1. 86Journal of Economic Perspectives

Figure 1 RealStock Prices and PresentValues of Subsequent Real Dividends (annual data)

10000

Real Stock Price (S&P 500)

1000 PDV, Rates e c i

r PDV, Constant Discount Rate P

100 PDV, Consumption

10 1860 1880 1900 1920 1940 1960 1980 2000 2020

Notes: Theheaviest line is the Standard &Poor500 Index for January ofyearshown. Theless-heavy lineis the presentvalue for each year of subsequentreal dividends accruing to the indexdiscounted bythe geometric-averagereal return for the entiresample, 6.61 percent. Dividends after 2002were assumedequal to the 2002dividend times 1.25 (to correctfor recent lower dividend payout) and growingat the geometric-averagehistorical growth rate for dividends, 1.11 percent. The thin lineis the presentvalue for each year of subsequentreal dividends discounted by one-year interest rates plusa riskpremium equal to the geometricaverage real return on the marketminus the geometric averagereal one-year . Thedashed line is the presentvalue for each year of subsequent realdividends discounted by marginal rates ofsubstitution inconsumptionfor a representative individualwith acoef Ž cientof relativerisk aversion of 3whoconsumes the real per capita nondurableand serviceconsumption from the U.S.NationalIncome and ProductAccounts. Realvalues were computed from nominal values by dividing by the consumerprice index (CPI-U since1913, linked to the Warrenand Pearsonproducer price index before 1913) and rescalingto January 2003 5 100.Some of the verylatest observationsof underlyingseries were estimated based ondata availableas ofthis writing; forexample, the consumerprice index for January 2003was estimatedbased on data fromprevious months. Sourcedata areavailable on http://www.econ.yale. edu/ shiller ,and the further descriptionsof someof the data arein Shiller (1989). Seealso footnotes1, 5and 6. theaggregate stock market,relative to the present value implied by the ef Ž cient marketsmodel. Our worklaunched aremarkableamount ofcontroversy, from which Iwillrecall here just afewhighlights. Theprincipal issue regarding our original work on excessvolatility was in regardto thinking about thestationarity of dividends and stockprices. My own workin the early 1980s had followeda traditionin the Ž nance literatureof From EfŽcientMarkets Theory to Behavioral Finance 87

assuming that dividends  uctuated around aknown trend. 3 However,one might also argue,as do Marsh and Merton(1986), that dividendsneed not stay closeto atrendand that evenif earnings followed a trend, share issuance orrepurchase could makedividends depart from a trendinde Ž nitely.In addition, ifbusiness managersuse dividends to provide a smoothed  owof payouts fromtheir busi- nesses, then thestock prices might be expected to shift more rapidly than divi- dends. Marsh and Mertonargued that such dividendsmoothing could makestock pricesunstationary insuch awaythat in Ž nitesamples prices appear morevolatile than thepresent values. Thus, thechallenge became how toconstruct atestfor expected volatility that modeleddividends and stock pricesin a moregeneral way. As such testswere developed,they tended to con Ž rmthe overall hypothesis that stock priceshad morevolatility than an ef Ž cientmarkets hypothesis could explain.For example, West(1988) derived an inequalitythat thevariance of innovations (that is, surprises) instock pricesmust beless than orequal to the variance of the innovations in the forecastedpresent value of dividends based on asubset ofinformation available to themarket. This inequalityis quitegeneral: it holds evenwhen dividendsand stock priceshave in Ž nitevariances so longas thevariance of the innovation (the unexpectedchange) inthese is Ž nite.Using long-term annual data onstock prices, Westfound that thevariance of innovations instock prices was fourto 20 times its theoreticalupper bound. 4 John Campbelland I(1988)recast the time series model interms of a cointegratedmodel of real prices and realdividends, whilealso relaxingother assumptions about thetime series, and again found evidenceof excessvolatility. 5 Campbell(1991) provided a variancedecomposition for stock returnsthat indicatedthat mostof the variability of the aggregate stock market conveyedinformation about futurereturns, rather than about futuredividends. Anothercontested issue regarding the early work on excessvolatility ques- tionedthe assumption oftheearly work that theef Ž cientmarkets model was best conveyedthrough an expectedpresent value model in which thereal discount rate isconstant through time.The assumption ofa constant discount rateover time can onlybe considereda Ž rststep, forthe theory suggests more relationships.

3 It shouldbe pointed out that dividendpayouts as afraction ofearnings have showna gradual downtrendover the periodsince 1871 and that dividendpayouts have increasinglybeen substituted by sharerepurchases. share repurchases reached approximately 1 percentof sharesoutstanding bythe late1990s. However, share repurchases do not invalidatethe theoreticalmodel that stockprices should equalthe presentvalue of dividends. See Cole, Helwege and Laster(1996). 4 In moretechnical terms, this argumentis over whether dividends could be viewed as astationary series. Thediscussion was often phrasedin termsof the “unit root” propertyof the timeseries, where a unit rootrefers to notion that whena variableis regressed on its ownlags, the characteristicequation of the differenceequation has arooton the unit circle.West (1988) can beviewedas away ofaddressing the unit rootissue. In our1988 paper, Campbell and Ihandlednonstationarity byusing a vectorautore- gressivemodel including the logdividend-price ratio and the changein log dividends as elements. 5 Barskyand DeLong(1993), however,later showed that if oneassumes that realdividends must be twice differencedto induce stationarity (so that dividendsare even more unstationary inthe sensethat dividend growthrates ,not just levels,are unstationary), then the ef Ž cientmarkets model looks rather moreconsistent with the data. 88Journal of Economic Perspectives

One such efŽ cientmarkets model makes the discount ratecorrespond to interestrates. The line in Figure 1 labeled “PDV,InterestRates ” illustratesthis concept.6 However,allowing time-varying interest rates in the present value for- muladoes littleto support theef Ž cientmarkets model. The actual priceis still morevolatile than thepresent value, especially for the latest half century.More- over,what changes through timethere are in the present value bear little resem- blanceto thechanges through timein thestock prices. Note for example that the presentvalue is extremely high throughout thedepression years of the1930s, not lowas was theactual stock market.The present value is high then becausereal interestrates were at extremelows after 1933, into the early 1950s, and sincereal dividendsreally did not fallmuch after1929. After 1929, real Standard &Poor ’s dividendsfell to around 1925levels for just afewyears, 1933 –1935and 1938,but, contraryto popular impressions,were generally higher in the1930s than theywere inthe1920s. 7 Analternativeapproach tothepossibility of varying real discount rateslooks at theintertemporal marginal rate of substitution forconsumption, which isshown in Figure1 withthe line labeled “PDV,Consumption. ”8 Themodels of ef Ž cient Ž nancial marketsfrom the 1970s like Merton (1973), Lucas (1978)and Breeden (1979)concluded that stock pricesare the expected present value of future dividendsdiscounted using marginalrates of substitution ofconsumption, and in thesemodels the equations forstock returnswere derived in the context of amodel maximizingthe of consumption. Grossman and Shiller(1981) produced a plotof that presentvalue since 1881, using Standard &Poordividend data and using aggregateconsumption data tocompute the marginal rates of substitution as discount factors, and this plotis updated here,and this iswhat isshown inFigure 1. We

6 Thepresent value, discounted by interestrates, isa plotfor each year t of

2002 t 2002 * 5 ~ 1 1 f! 1 ~ 1 1 f! * P r,t O 1/ 1 rt1j Dt 1/ 1 rt1j P const,2003. t5t j50 j5t

Seenote to Figure 1. 7 Campbelland I(1989) recastthe argumentin terms of a vectorautoregressive model of real stock prices,real interest rates and realdividends, in which each of these variables was regressedon lagsof itself and lagsof the othervariables. We found that the dividend-priceratio not onlyshows excess volatility, butshows very little correlation with the dividenddivided by the forecastof the presentvalue offuturedividends. 8 Thepresent value, consumption discounted, is a plotfor each year t of

2002 * 5 ~ !3 1 ~ !3 * P c,t O Ct /Ct Dt Ct/C2003 P const,2003, t5t11

where Ct is real per capita realconsumption at time t.This expressionis inspired by Lucas(1978) and derivedin Grossman and Shiller(1981) assuminga coef Ž cientof relative aversion of 3. Seenote to Figure 1. RobertJ. Shiller 89

found, as can also beseen here in Figure1, that thepresent value of dividendsas discounted inthis modelhad onlya tenuous relationto actual stock prices,and did not appear volatileenough tojustify the price movements unless wepushed the coefŽ cientof relativerisk aversion to ridiculously high levels,higher than thevalue ofthree that was used forthe plot. Grossman and Shiller(1981) stressed that therewere some similarities be- tweenthe present value and theactual realprice, notably thepresent value peaks in1929and bottomsout in1933, close to the actual peakand troughof the market. Butthe present value does this becauseconsumption peakedin 1929 and then dropped verysharply, bottomingout in1933,and thepresent value takes account ofthis, as ifpeople had perfectforesight of the coming depression. But in fact it appears veryunlikely that peoplesaw this outcomein 1929,and iftheydid not, then the efŽ cientmodel does not predictthat theactual realprice should have tracked thepresent value over this period. Actually,the consumption discount model,while it may show somecomove- mentsat timeswith actual stock prices,does not workwell because itdoes not justify thevolatility of stock prices.I showed (1982)that thetheoretical model implies a lowerbound on thevolatility of the marginal rate of substitution, abound which is withthe U.S. data much higherthan could beobserved unless riskaversion were implausiblyhigh. Hansen and Jagannathan latergeneralized this lowerbound and elaboratedon itsimplications, and today theapparent violationof this “Hansen- Jagannathan lowerbound ” isregarded as an importantanomaly in Ž nance.9 Somevery recent research has emphasizedthat, eventhough theaggregate stock marketappears tobewildly inef Ž cient,individual stock pricesdo show some correspondenceto ef Ž cientmarkets theory. That is, whilethe present value model forthe aggregate stock marketseems unsupported bythe data, thereis some evidencethat cross-sectional variationsin stock pricesrelative to accounting measures show somerelation to the present value model. Paul Samuelsonsome years ago positedthat thestock marketis “micro efŽ cientbut macroinef Ž cient,” since there isconsiderablepredictable variation across Ž rmsin theirpredictable future paths ofdividends but littlepredictable variation in aggregate dividends. Hence, Sam- uelsonasserted, movements among individualstocks makemore sense than do movementsin the market as awhole.There is now evidenceto back up this assertion. Vuolteenaho(2002) showed, using vector-autoregressivemethods, that the ratioof book-to-market-valueof U.S. Ž rmsexplainsa substantial fractionof changes in future Ž rms’ earnings. Cohen, Polkand Vuolteenaho(2002) concluded that 75 to80 percent of the variation across Ž rmsin their book-to-market ratios can be explainedin terms of futurevariation in pro Ž ts. Jung and Shiller(2002) show that, cross-sectionally,for U.S. stocks that have beencontinually traded since 1926, the price-dividendratio is a strongforecaster of the present value of future dividend

9 See,for example, John Cochrane ’s(2001) book AssetPricing ,whichsurveys this literature.Much of the olderliterature is summarized in my 1989book MarketVolatility . 90Journal of Economic Perspectives

changes. So, dividend-priceratios on individualstocks do serveas forecastsof long-termfuture changes intheir future dividends, as ef Ž cientmarkets assert. This does not meanthat thereare not substantial bubblesin individual stock prices,but that thepredictable variation across Ž rmsin dividends has oftenbeen so largeas tolargely swamp out theeffect of the bubbles. A lotof this predictable variationacross Ž rmstakes the form of Ž rms’ payingzero dividends for many years and investorscorrectly perceiving that eventuallydividends will be coming, and of Ž rmsin very bad shape withinvestors correctly perceiving they will not bepaying substantial dividendsmuch longer.When itcomes to individual stocks, such predictablevariations, and theireffects on price,are often far larger than the bubblecomponent ofstock prices. Thereis a clearsense that thelevel of volatility of the overall stock market cannot bewell explained with any variantof theef Ž cientmarkets model in which stock pricesare formed by looking at thepresent discounted valueof future returns.There are many ways totinkerwith the discount ratesin thepresent value formulas,and somedaysomeone may Ž nd some deŽ nitionof discount ratesthat produces apresentvalue series that “Ž ts” theactual pricebetter than any ofthe seriesshown inFigure1. 10 Butit isunlikely that theywill do so convincingly,given thefailure of ourefforts to date to capture the volatility of stock prices.To justify thevolatility in termsof such changes inthe discount rates,one will have toargue that investorsalso had agreatdeal of information about changes inthe factors in uencingthese future discount rates. Afterall the efforts to defend the ef Ž cientmarkets theory, there is still every reason tothink that, whilemarkets are not totallycrazy, theycontain quitesub- stantial noise, so substantial that itdominates the movements in the aggregate market.The ef Ž cientmarkets model, for the aggregate stock market,has stillnever beensupported byany study effectivelylinking stock market  uctuations with subsequent fundamentals. Bythe end ofthe 1980s, the restless minds ofmany academicresearchers had turnedto other theories.

TheBlossoming of Behavioral Finance

In the1990s, a lotof thefocus ofacademic discussion shiftedaway fromthese econometricanalyses oftime series on prices,dividends and earningstoward developingmodels of human psychologyas itrelatesto Ž nancial markets.The Ž eld ofbehavioral Ž nance developed.Researchers had seentoo many anomalies,too

10 Otherfactors areconsidered by McGrattan and Prescott(2001), whoemphasize tax ratechanges, and Siegel(2002), whoconsiders not onlytax ratechanges but also changes in the volatilityof the economy, changesin the in  ation rate, and changesin transactions costs. Neitherof thesestudies shows a “Ž t” betweenpresent value and pricesover the longsample, however. Notably, the factors they usedo not gothrough sudden changes at the timeof the stockmarket booms and crashessurrounding 1929 and 2000. From EfŽcientMarkets Theory to Behavioral Finance 91

littleinspiration that ourtheoretical models captured important  uctuations. An extensivebody ofempirical work, summarized in Campbell, Lo and MacKinlay ’s 1996 book TheEconometrics of FinancialMarkets ,laidthe foundation fora revolution in Ž nance. Richard Thalerand Istartedour National Bureau of Economic Research conferenceseries on behavioral Ž nance in1991,extending workshops that Thaler had organizedat theRussell Sage Foundation afewyears earlier. 11 Many other workshops and seminarson behavioral Ž nance followed.There is so much going on in the Ž eldthat itis impossible to summarize in a short space. Here,I will illustratethe progress of behavioral Ž nance withtwo salient examples from recent research:feedback models and obstaclesto smartmoney. For overall surveys of the Ž eldof behavioral Ž nance, theinterested reader might begin with ’s Beyond Greedand Fear: Understanding Behavioral Finance and the Psychology ofInvesting (2000)or Andrei Shleifer ’s InefŽcientMarkets (2000).There are also somenew books ofcollected papers inbehavioral Ž nance, includinga three-volumeset, BehavioralFinance ,editedby Hersh Shefrin(2001), and Advancesin Behavioral Finance II,editedby Richard H.Thaler(2003).

FeedbackModels One ofthe oldest theories about Ž nancial markets,expressed long ago in newspapers and magazinesrather than scholarlyjournals, is, iftranslated into academicwords, aprice-to-pricefeedback theory. When speculativeprices go up, creatingsuccesses forsome investors, this mayattract publicattention, promote word-of-mouth enthusiasm, and heightenexpectations for further price increases. Thetalk attracts attentionto “new era” theoriesand “popular models ” that justify theprice increases. 1 2 This process inturn increasesinvestor demand and thus generatesanother round ofprice increases. If thefeedback is not interrupted,it mayproduce after many rounds aspeculative “bubble,” inwhich high expectations forfurther price increases support veryhigh currentprices. The high pricesare ultimatelynot sustainable, sincethey are high onlybecause ofexpectations of furtherprice increases, and so thebubble eventually bursts, and pricescome falling down. Thefeedback that propelledthe bubble carries the seeds of its own destruc- tion, and so theend ofthe bubble may be unrelated to news stories about fundamentals. Thesame feedback may also producea negativebubble, downward pricemovements propelling further downward pricemovements, promoting word- of-mouth pessimism,until the market reaches an unsustainably lowlevel. Such afeedbacktheory is veryold. Aslongago as 1841,Charles MacKay inhis

11 Fora list ofour programs since 1991, with linksto authors ’ websites,see http://cowles.econ. yale.edu/beh Ž n . 12 Descriptionsof new era theories attending variousspeculative bubbles are described in my book (2000). Popularmodels that accompaniedthe stockmarket crash of1987, the realestate bubbles peakingaround 1990 and variousinitial public offering booms are discussed in my paperin this journal (1990). 92Journal of Economic Perspectives

in uentialbook Memoirsof Extraordinary Popular Delusions describedthe famous tulipmaniain Holland inthe1630s, a speculativebubble in tulip  owerbulbs, with wordsthat suggestfeedback and theultimate results of the feedback (pp. 118 –119):

Many individualsgrew suddenly rich. Agoldenbait hung temptinglyout beforethe people, and oneafter another, theyrushed tothe tulip marts, like  iesaround ahoney-pot ....Atlast, however,the more prudent beganto see that this follycould not last forever.Rich peopleno longerbought the  owers tokeep them in their gardens, but tosell them again at centper cent pro Ž t. It was seenthat somebodymust losefearfully in the end. As this conviction spread, pricesfell, and neverrose again. 13

Thefeedback theory seems to be even much olderthan this. Noteof such feedback, and therole of word-of-mouth communications inpromotingit, was infact made at thetime of the tulipmania itself. One anonymous observerpublishing in 1637 (theyear of the peak of thetulipmania) gives a Ž ctionalaccount ofa conversation betweentwo people, Gaergoedt and Waermondt, that illustratesthis author ’s impressionof the word-of-mouth communications ofthat time:

Gaergoedt: “Youcan hardlymake a returnof 10% with the that you investin youroccupation [as aweaver],but withthe tulip trade, you can make returnsof 10%, 100%, yes, even 1000%. Waermondt: “ ....Buttell me, should Ibelieveyou? ” Gaergoedt: “Iwilltell you again, what Ijust said. ” Waermondt: “ButI fearthat, sinceI wouldonly start now, it ’stoo late,because now thetulips are very expensive, and Ifearthat I ’llbehit with the spit rod, beforetasting the roast. ” Gaergoedt: “It’snevertoo late to make a pro Ž t, you makemoney while sleeping. I’vebeen away fromhome for four or Ž vedays, and Icamehome just lastnight, but now Iknowthat thetulips I have have increasedin value bythree or four thousand guilder;where do you have pro Ž ts likethat from othergoods? ” Waermondt: “Iam perplexedwhen Ihearyou talkinglike that, Idon ’t know what todo; has anybody becomerich with this trade? ” Gaergoedt: “What kindof questionis this?Look at allthe gardeners that used towear white-gray out Ž ts, and now they ’rewearing new clothes. Many weav- ers,that used towearpatched up clothes, that theyhad ahard timeputting

13 Garberquestions MacKay ’sfacts aboutthe tulipmaniain his 1990article in this journaland inhis book FamousFirst Bubbles .Forexample, the crash was not absolutely Ž nal; Garberdocuments very high tulipprices in 1643.The actual courseof the bubbleis ambiguous,as allcontracts weresuspended by the states ofHollandin 1637 just after the peak,and noprice data areavailable from that date. RobertJ. Shiller 93

on, now wearthe glitteriest clothes. Yes,many who tradein tulipsare riding ahorse, have acarriageor a wagon, and duringwinter, an icecarriage, . ... ”14

Casual observationsover the years since then areplentiful evidence that such talk, provokinga senseof relative futility of one ’sday-to-day workand envyof the Ž nancial successes ofothers, and includingsome vacuous answerto doubts that the pricerise may be over, is effective in overcoming rational doubts among some substantial numberof people and tends tobring successive rounds ofthem into the market. In my book IrrationalExuberance ,published(with some luck) at thevery peak ofthe stock marketbubble in March 2000,I arguedthat verymuch thesame feedback,transmitted by word-of-mouth as wellas themedia, was at workin producing thebubble we were seeing then. Ifurtherargued that thenatural self-limitingbehavior of bubbles, and thepossibility of downward feedbackafter the bubblewas over,suggested a dangerous outlookfor stocks inthe future. One mightwell also presumethat such simplefeedback, if it operates so dramaticallyin events like the tulip bubble or the stock marketboom until 2000, ought oftento recur at asmallerscale and toplay an importantif lesser role in morenormal day-to-day movementsin speculative prices. Feedback models, in the formof difference equations, can ofcourse produce complicated dynamics. The feedbackmay be an essentialsource of much ofthe apparently inexplicable “randomness” that wesee in Ž nancial marketprices. Butthe feedback theory is very hard to Ž nd expressedin Ž nance oreconomics textbooks,even today. Sincethe theory has appearedmostly in popular discourse, and not inthetextbooks, one might well infer that ithas longbeen discredited by solidacademic research. In fact, academicresearch has untilrecently hardly addressed thefeedback model. Thepresence of such feedbackis supported bysome experimental evidence. Andreassen and Kraus (1988)found that when peopleare shown real historicalstock pricesin sequence(and which theyknew were real stock prices)and invitedto trade in a simulatedmarket that displays theseprices, they tended to behaveas ifthey extrapolate past pricechanges when theprices appear toexhibit atrendrelative to period-to-period variability. Smith, Suchanek and Williams (1988)were able to create experimental markets that generatedbubbles that are consistent withfeedback trading. Marimon,Spear and Sunder(1993) showed experimentsin which repeatingbubbles were generated if subjects wereprecon- ditionedby past experienceto form expectations of bubbles. Thepresence of such feedbackis also supported byresearch in , which shows that human judgments ofthe probability of futureevents show systematicbiases. Forexample, psychologists Tverskyand Kahneman have shown that judgments tend tobe made using arepresentativenessheuristic,

14 Anonymous(1637). BjornTuypens translated this passage. 94Journal of Economic Perspectives

wherebypeople try to predict by seekingthe closest match topast patterns, without attentionto the observed probability of matching thepattern. Forexample, when asked toguess the occupations ofpeople whose personality and interestsare describedto them, subjects tendedto guess the occupation that seemedto match thedescription as closelyas possible,without regard to the rarity of the occupation. Rationalsubjects wouldhave chosen humdrum and unexceptionaloccupations morebecause morepeople are in these occupations. (Kahneman and Tversky, 1974).By the same principle, people may tend tomatch stock pricepatterns into salientcategories such as dramaticand persistentprice trends, thus leadingto feedbackdynamics, evenif these categories may be rarely seen in fundamental underlyingfactors. Daniel,Hirschleifer and Subramanyam (1999)have shown that thepsycho- logicalprinciple of “biasedself-attribution ” can also promotefeedback. Biased self-attribution,identi Ž edbypsychologist Daryl Bem (1965), is a patternof human behaviorwhereby individuals attribute events that con Ž rmthe validity of their actions totheir own high abilityand attributeevents that discon Ž rmtheir actions tobad luckor sabotage. Upon readingthe above passage fromthe time of the tulipmania, oneeasily imagines that Gaergoedtis basking in self-esteem and relishingthe telling of the story. Many readerstoday can probablyeasily recall similarconversations, and similarego-involvement by thespreaders of theword, in the1990s. Such human interactions,the essential cause ofspeculative bubbles, appear torecur across centuriesand across countries:they re  ectfundamental parametersof human behavior. Thereis also evidencesupportive of feedback from natural experiments,which maybe more convincing than thelab experiments when theyoccur inreal time, withreal money, with real social networks and associated interpersonalsupport and ,with real and visceralenvy of friends ’ investmentsuccesses, and with communications-mediapresence. Ponzi schemesmay be thought ofas represent- ingsuch natural experiments.A Ponzi scheme(or pyramid scheme or money circulationscheme) involves a super Ž ciallyplausible but unveri Ž ablestory about how moneyis made for investors and thefraudulent creation of high returnsfor initialinvestors by givingthem the money invested by subsequent investors.Initial investorresponse to the scheme tends tobe weak,but as therounds ofhigh returns generatesexcitement, the story becomes increasingly believable and enticingto investors.These schemes are often very successful ingenerating extraordinary enthusiasms among someinvestors. We have seensome spectacular Ponzi schemes recentlyin countries that do not have effectiveregulation and surveillanceto preventthem. A numberof Ponzi schemes in Albania 1996 –1997were so largethat totalliabilities reached half ayear ’sGDP;theircollapse brought on aperiodof anarchy and civilwar in which 2000people were killed ( Jarvis,1999). Real world stock-marketspeculative bubbles, Iarguedin my 2000 book IrrationalExuberance , resemblePonzi schemes in thesense that some “new era” storybecomes attached tothe bubble and acquiresincreasing plausibility and investorenthusiasm as the marketcontinues toachieve high returns.Given the obvious success ofPonzi From EfŽcientMarkets Theory to Behavioral Finance 95

schemeswhen theyare not stopped bythe law, we would need a good reason to thinkthat analogous phenomena ofspeculative bubbles are not also likely. Thestock marketboom that endedin early2000 is another relevantepisode. Accordingto my surveydata, now expressedin theform of stock marketcon Ž dence indexesproduced bythe Yale School of Management and availableat http:// icf.som.yale.edu/con Ž dence.index , the conŽ denceof individualinvestors that the stock marketwill go up inthe next year, and willrebound fromany drop, rose dramatically1989 –2000.As inthe tulipmania centuries before, there was afocusing ofpublic attention and talkon thespeculative market and aproliferationof wishful-thinkingtheories about a “new era” that wouldpropel the stock market on acoursethat, whileuneven, isrelentlessly upward, theoriesthat werespread by wordof mouth as wellas themedia. It iswidely thought that thereis a problemwith the feedback theories: the theorieswould seem to imply that speculativeprice changes arestrongly serially correlatedthrough time,that pricesshow strongmomentum, continuing uniformly inonedirection day afterday. This seemsinconsistent withthe evidence that stock pricesare approximately a random walk. Butsimple feedback models do not implystrong serial correlation, as Istressed inShiller(1990). There, I presenteda modelof the demand fora speculativeasset as equalinga distributedlag withexponentially declining weights on past price changes through time(the distributed lag representing feedback distributed over time),plus otherfactors that affectdemand. Themodel asserts that peoplereact graduallyto price changes overmonths oryears, not just toyesterday ’s price change. Ahistoryof priceincreases over the last year may encourage buying today evenif yesterday ’spricechange was down. Also, themodel recognizes that thereare othershocks, besidesfeedback, in  uencingprice. In such amodel,a disturbance insome demand factorother than feedback can incertaincases beampli Ž ed, at leastfor a time,because itchanges theprice and thus affectsfuture prices through thedistributed lag. 15 However,unless we knowsomething about theother factors that drivedemand, such adistributedlag modeldoes not implyanything at allabout theserial correlation properties of speculativeprice changes. Thefeedback model does not implythat thereis much serialcorrelation in day-to-day stockprice changes, sincethe noise in the other factors feedsdirectly into short-run changes, and theeffect on today ’s price of laggedother factors operatesat alowfrequency that isessentially unrelated to day-to-day changes and has effectsthat can beobserved only from its cumulative effectafter a longperiod of time. Thus, theapproximate random walkcharacter of stockprices is not evidence

15 t 2 g ( t2 t) Thefeedback model is pt 5 c 2 ` e dpt 1 pt, 0 , c , 1, 0 , g. Here, pt isprice at time t, and pt isthe combinedeffect ofotherfactors ondemand.It followsthat pt 5 pt 1 (c/(1 2 c))(pt 2 t 2 ( g / ( 1 2 c) ) ( t2 t ) pt), where pt 5 (g/(1 2 c)) 2 ` e ptdt isaweightedaverage of lagged p.SeeShiller (1990, p.60). Sucha modeldoes not implythat pricebehaves smoothly through time: price can look muchlike a randomwalk if, forexample, pt isa randomwalk. 96Journal of Economic Perspectives

against feedback.Moreover, even if feedback did imply some momentum, we can also notethat therandom walkcharacter of stock pricesis reallynot fullysupported bythe evidence anyway, and that infact therehas beenmore than alittlemomen- tumto stock prices.Jegadeesh and Titman(1993) found that winningstocks, stocks that showed exceptionallyhigh six-month returns,beat losing stocks, stocks that showed exceptionallylow six-month returns,by 12 percentover the following year. In contrast, overlonger periods of timethis momentumseems to reverse itself. De Bondt and Thaler(1985) Ž nd that overthe period 1926 to 1982, stocks represented on theCenter for Research in Security Prices data setof the whosereturns had beenin the top decileacross Ž rmsover three years (thus, “winner” stocks) tendedto show negativecumulative returns in the succeeding threeyears. They also found that “loser” stocks whosereturns had beenin the bottomdecile over the prior three years tended to show positivereturns over the succeedingthree years. Thus, thereis a tendencyfor stock pricesto continue inthe samedirection over intervals of six months toa year,but toreversethemselves over longerintervals. Campbell, Lo and Mackinlay(1996) document this fact carefully. 16 Apatternlike this iscertainlyconsistent withsome combination offeedbackeffects and otherdemand factors drivingthe stock marketlargely independently of fundamentals.

SmartMoney vs. Ordinary Investors Theoreticalmodels of ef Ž cient Ž nancial marketsthat representeveryone as rationaloptimizers can beno morethan metaphors forthe world around us. Nothingcould bemore absurd than toclaim that everyone knows how tosolve complexstochastic optimizationmodels. For these theoretical models to have any relevanceto thestock market,it must somehowbe the case that asmallerelement of “smart money” or the “marginaltrader ” can offsetthe foolishness ofmany investorsand makethe markets ef Ž cient. The efŽ cientmarkets theory, as itis commonly expressed, asserts that when irrationaloptimists buy astock, smartmoney sells, and when irrationalpessimists sella stock, smartmoney buys, therebyeliminating the effect of the irrational traderson marketprice. But Ž nance theorydoes not necessarilyimply that smart moneysucceeds infullyoffsetting the impact of ordinaryinvestors. In recentyears, researchin behavioral Ž nance has shed someimportant light on theimplications ofthe presence of these two classes ofinvestors for theory and also on some characteristicsof the people in thetwo classes. Froma theoreticalpoint ofview, it isfar from clear that smartmoney has the powerto drive market prices to fundamental values.For example, in one model withboth feedbacktraders and smartmoney, the smart money tended to amplify, ratherthan diminish,the effect of feedback traders, by buying in ahead ofthe

16 Grinblattand Han (2001) have arguedthat this tendencyof stockprices to show momentum for a whileand then reversethemselves might be relatedto the phenomenonthat investorstend tohold on tolosers and sellwinners (Statman and Shefrin, 1985;Odean, 1998). RobertJ. Shiller 97

feedbacktraders in anticipation ofthe price increases they will cause (DeLong, Shleifer,Summers and Waldman, 1990b).In arelatedmodel, rational, expected- utility-maximizingsmart money never chooses tooffset all of the effects of irrational investorsbecause they are rationally concerned about therisk generated by the irrationalinvestors and do not want toassume therisk that theircompletely offsettingthese other investors would entail (De Long, Shleifer,Summers and Waldman, 1990b). 17 Often, speculativebubbles appear tobe common toinvestments of a certain “style,” and thebubbles may not includemany otherinvestments. For example, the stock marketbubble that peakedin the year 2000 was strongestin tech stocks or Nasdaq stocks. Barberisand Shleifer(2002) present a modelin which feedback traders’ demand forinvestments within a particularstyle is relatedto a distributed lagon past returnsof that styleclass. Bytheir budget constraint, when feedback tradersare enticed by one style, they must moveout ofcompeting styles. The smart moneyare rational utility maximizers. Barberis and Shleiferpresent a numerical implementationof their model and Ž nd that smartmoney did not fullyoffset the effectsof the feedback traders. Style classes gothrough periodsof boom and bust ampliŽ edby the feedback. Goetzmannand Massa (1999)provided some direct evidence that itisreason- ableto suppose that thereare two distinct classes ofinvestors: feedback traders who followtrends and thesmart money who movethe other way. FidelityInvestments providedthem with two years of daily account informationfor 91,000 investors in aStandard and Poor ’s500index fund. Goetzmannand Massa wereable to sort theseinvestors into two groups based on how theyreact to daily price changes. Therewere both momentuminvestors, who habituallybought moreafter prices wererising, and contrarian investors,or smart money, who habituallysold after priceswere rising. Individual investorstended to stay as oneor the other, rarely shiftedbetween the two categories. Recentresearch has focused on an importantobstacle to smart money ’s offsettingthe effects of irrational investors. The smart money can alwaysbuy the stock, but ifthe smart money no longerowns thestock and Ž nds it difŽ cultto short thestock, then thesmart money may be unable tosell the stock. Somestocks could bein asituation wherezealots have bought intoa stockso much that onlyzealots own shares, and tradeis only among zealots,and so thezealots alone determine the priceof the stock. Thesmart money who knowthat thestock is priced ridiculously high maywell use up allthe easily available shortable shares and then willbe standing on thesidelines, unable toshort moreshares and pro Ž t from their knowledge.Miller (1977) pointed out this  aw inthe argument for market ef Ž - ciency,and his paperhas beendiscussed eversince. It seemsincontrovertible that insome cases stocks have beenheld primarily by zealotsand that short sellershave found itverydif Ž cultto short. One exampleis the

17 Shleiferand Summers(1990) presenta nicesummary of these themes in this journal. 98Journal of Economic Perspectives

3Comsale of Palmnear the peak of thestock marketbubble (Lamont and Thaler, 2001).In March 2000,3Com, a pro Ž tableprovider of network systems and services, sold tothe general public via an initialpublic offering 5 percentof its subsidiary Palm, amakerof handheld computers. 3Comannounced at thesame time that the restof Palm would follow later. The price that these Ž rstPalm shares obtainedin themarket was so high, when comparedwith the price of the 3Com shares, that if onesubtracts theimplied value of the remaining 95 percent of Palm from the 3Commarket value, one Ž nds that thenon-Palm part of3Com had a negative value. Sincethe worst possible price for 3Com after the Palm sale was completedwould bezero, there was thus astrongincentive for investors to short Palmand buy 3Com. But, theinterest cost ofborrowing Palm shares reached35 percent by July 2000, puttinga damperon theadvantage toexploitingthe mispricing. 18 Evenan investor who knewfor certain that thePalm shares wouldfall substantially may have been unable tomake a pro Ž tfromthis knowledge.The zealots had won withPalm and had controlover its price, for the time being. ThePalm example is an unusual anomaly. Shortingstocks onlyrarely becomes so costly. Butthe example proves the principle. The question is: How importantare obstacles tosmart money ’ssellingin causing stocks todeviate from fundamental value? Of course, inreality, the distinction between zealots and smartmoney is not alwayssharp. Instead, thereare sometimes all gradations inbetween, especially sincethe objective evidence about thefundamental valueof individual stocks is alwayssomewhat ambiguous. If sellingshort isdif Ž cult, anumberof individual stocks could becomeoverpriced. It wouldalso appear possiblethat major segments ofthe stock market,say theNasdaq in1999,or even the entire stock market,could windup ownedby, ifnot zealots,at leastrelatively optimistic people. Short-sale constraints could bea fatal  aw inthe basic ef Ž cientmarkets theory. Theproblem with evaluating Miller ’s(1977)theory that alackof short selling can cause Ž nancial anomalieslike overpricing and bubblesis that therehas been littleor no data on which stocks aredif Ž cultto short. Thereare long time series data serieson “short interest, ” which isthe total number of shares that areshorted. Figlewski(1981) found that high levelsof short interestfor individual stocks predictslow subsequent returnsfor them, a directionthat wouldbe predicted by Miller’stheory.But the predictability was weak.On theother hand, differencesin short interestacross stocks do not have an unambiguous connection withdif Ž culty ofshorting. Stocksdiffer from each otherin terms of the fraction of shares that are inaccounts that areshortable. Differences across stocks inshort interestcan also re ectdifferent demand forshorting for hedging needs. Thus, thereis a signi Ž cant

18 Put optionprices on Palmalso began to re  ectthe negativeopinions and becameso expensive that the usualrelation between options prices and stockprice, the so-called “put-callparity, ” failedto hold. Onemust rememberthat optionsmarkets are derivative markets that clearseparately from stock markets,and overpricedputs have nodirect impact on the supplyand demandfor stock unless arbitrageurscan exploitthe overpricingby shortingthe stock. From EfŽcientMarkets Theory to Behavioral Finance 99

errors-in-variablesproblem when using short interestas an indicatorof the cost of shorting. Somerecent papers have sought todetectthe presence of barriers that might limitshort salesindirectly by observing the differences of opinion that can have an impacton priceif thereis adif Ž cultyshorting stocks. Without observingbarriers to shortingstocks directly,we can stillinfer that when differencesof opinion arehigh about astock, itis more likely that short-salerestrictions will be binding for that stock, and thus that themore pessimistic investors will not preventthe stock from becomingoverpriced and hencesubject tolower subsequent returns. Scherbina(2000) measured differences of opinion bycalculating the disper- sion ofanalysts ’ earningsforecasts. Shefound that stocks witha high dispersionof analysts’ forecastshad lowersubsequent returns,and she linkedthe low returns to theresolution of the uncertainty. Chen, Hong and Stein(2000) measured differ- enceof opinion bya breadth ofownership measure derived from a database on mutualfund portfolios.The breadth variablefor each quarteris the ratio of the numberof mutual funds that hold alongposition in thestock tothe total number ofmutual funds forthat quarter.They Ž nd that Ž rmsin the top decileby breadth ofownership outperformed those inthebottom decile by 4.95percent per annum afteradjusting forvarious other factors. What wewould really like to have totest the importance of short sales restrictionson stock pricingis someevidence on thecost ofshorting. If those stocks that have becomevery costly to short tend tohave poorsubsequent returns,then wewill have moredirect con Ž rmationof Miller ’s(1977)theory. There is surpris- inglylittle available information about thecost ofshortingindividual stocks. Such data have not beenavailable for economic research until recently. A numberof recentunpublished papers have assembleddata on thecost ofshorting individual stocks, but thesepapers have assembleddata forno morethan ayeararound 2000. Recently,Jones and Lamont (2001)discovered an oldsource of data on the cost ofshorting stocks. In the1920s and 1930sin theUnited States, there used to be a “loan crowd” on the  oorof theNew York Stock Exchange, whereone could lendor borrow shares, and theinterest rates at which shares wereloaned were reportedin the WallStreet Journal .Jones and Lamont assembledtime series of the interestrates charged on loans ofstocks from1926 to 1933, eight years of data on an averageof 80 actively-traded stocks. Theyfound that, aftercontrolling for size, overthis periodthe stocks that weremore expensive to short tendedto be more highlypriced (in terms of market-to-book ratios), consistent withthe Miller (1977) theory.Moreover, they found that themore expensive-to-short stocks had lower subsequent returnson average,again consistent withthe Miller theory. Of course, theirdata span onlyeight years from a remoteperiod in history, and so their relevanceto today ’smarketsmight be questioned. Why has therenot beenmore data onthecost ofshorting? Why didthe loan crowdon theNew York Stock Exchange disappear and theloan ratesin the Wall StreetJournal withit? Perhaps afterthe crash of1929 the widespread hostility to short sellers(who werewidely held responsible for the crash) forcedthe market to go 100Journal of Economic Perspectives

underground. Jones and Lamont (2001)document aconsistent patternof political opposition toshort sellersafter 1929 and point out that J.EdgarHoover, the head ofthe Federal Bureau of Investigation, was quotedas sayingthat hewould inves- tigatea conspiracy tokeep stock priceslow. By 1933, the rates shown on theloan listbecome all zeros, and the WallStreet Journal stopped publishingthe loan listin 1934. Fortunately,this longdrought ofdata on thecost ofshorting stocks maybe over,and stocks should becomeeasier to short. In 2002,a consortiumof Ž nancial institutionsestablished an electronicmarket for borrowing and lendingstocks onlinevia a new Ž rm,EquiLend, LLC. Thenew securities lending platform at http://www.equilend.com exceeded$11 billion in transactions inits Ž rst two weeks,and dailyavailability posting exceed$1 trillion. Butthe true cost ofshorting stocks isprobablymuch higherthan theexplicit interestcost ofborrowing the shares, becauseof thepsychological cost that inhibits short selling.Most investors,even some very smart investors, have probablynever evenconsidered shorting shares. Shortingshares iswidelyreputed to involve some substantial risksand nuisances. Forexample, the short-seller always stands therisk that theultimate owner of the shares willwant tosell the shares, at which timethe short-selleris forced to return the shares. This detailmay be little more than a nuisance, forthe short sellercan likelyborrow them again fromanother lender,but it may Ž gurelargely in potential short-sellers ’ minds. Amoreimportant consideration that mayweigh on short sellers ’ minds is the unlimitedloss potentialthat short salesentail. When an investorbuys astock, the potentialloss isno greaterthan theoriginal . But when an investor shorts astock, thepotential losses can greatlyexceed the original investment. An investorcan alwaysterminate these losses by covering the shorts, but this action typicallybrings considerable psychological anguish. Decidingto cover one ’s shorts and getout ofa short positionafter losses is psychologically dif Ž cult, giventhe evidenceon thepain ofregret. Kahneman and Tversky ’sprospecttheory (1979) suggeststhat individualsare far more upset bylosses than theyare pleased by equivalentgains; infact, individualsare so upset bylosses that theywill even take greatrisks with the hope ofavoiding any lossesat all.The effects of this pain of regrethave beenshown toresult in atendencyof investors in stocks toavoidselling losers,but thesame pain ofregret ought tocause short sellersto want toavoid coveringtheir shorts inalosingsituation. Peopleprefer to avoid putting themselves insituations that mightconfront themwith psychologically dif Ž cultdecisions in the future. Thestock marketthat wehave today alwayslimits the liability of investors.As Moss (2002)has documented, theidea that allpublicly traded stocks should have limitedliability for their investors was theresult of experimenting with different kinds ofstockholder liability in the in the early nineteenth century and thediscovery of the psychological attractiveness of limited liability stocks. The debatesin theearly nineteenth century were concerned withthe balancing ofthe agencycosts oflimited liability, which encouragesbusinesses totake greater , RobertJ. Shiller 101

against thebene Ž ts interms of peace of mind to investors. Various alternatives were consideredor experimented with, including unlimited liability, unlimited propor- tionalliability (where individual investors in a company arelimited to their pro- portionateshare ofthe company ’slossesaccording totheir share inthe company), and doubleliability (where individual investors are accountable forthe capital subscribedonce again). Byaround 1830,it was apparent fromexperiments in New Yorkand surrounding states that investorsfound itveryappealing that theycould put moneydown tobuy astock today, and fromthat day forwardface no further lossesbeyond what theyalready put down. It allowedthem, once having purchased astock, toconcentratetheir emotions on thesmall probability of the stock doing extremelywell, rather on thesmall probability that someonewould come after themfor more money. People have always beenvery attracted to lottery tickets, and theinvention of limited liability, Moss concludes, turnedstock investmentspsycho- logicallyinto something a lotlike lottery tickets. By the same theory, then, investors will not Ž nd shortingstocks veryattractive. Remarkablyfew shares arein fact sold short. Accordingto New York Stock Exchangedata, from1977 to 2000 year-end short interestranged from 0.14 percent to1.91 percent of all shares. Accordingto Dechow,Hutton, Muelbroekand Stone (2001),less than 2percentof all stocks had short interestgreater than 5percentof shares outstanding 1976 –1983.Given the obviously large difference of opinion about and differenceof public attention to different stocks, itis hard tosee how such asmallamount ofshort sellingcould offsetthe effect on stock priceof the extrademand ofinvestors who developan irrational Ž xationon certainstocks.

Conclusion

Thecollaboration between Ž nance and othersocial sciences that has become known as behavioral Ž nance has ledto aprofound deepeningof our knowledge of Ž nancial markets.In judging theimpact of behavioral Ž nance todate, itisimpor- tant toapply theright standards. Of course, wedo not expectsuch researchto providea methodto make a lotof moneyoff of Ž nancial marketinef Ž ciency very fast and reliably.We should not expectmarket ef Ž ciencyto be so egregiouslywrong that immediatepro Ž tsshould becontinually available. But market ef Ž ciencycan be egregiouslywrong in othersenses. Forexample, ef Ž cientmarkets theory may lead todrastically incorrect interpretations of events such as major stock market bubbles. In his reviewof the literature on behavioral Ž nance, EugeneFama (1998) found faultfor two basic reasons. The Ž rstwas that theanomalies that were discoveredtended to appear tobe as oftenunderreaction by investors as overre- action. Thesecond was that theanomalies tended to disappear, eitheras time passed oras methodologyof the studies improved. His Ž rstcriticism re  ects an incorrectview of the psychological underpinnings ofbehavioral Ž nance. Since thereis no fundamental psychologicalprinciple that peopletend alwaysto over- 102Journal of Economic Perspectives

reactor always tounderreact, it isno surprisethat researchon Ž nancial anomalies does not revealsuch aprincipleeither. His second criticismis also weak.It isthe natureof scholarly research, at thefrontier, in all disciplines, that initialclaims of importantdiscoveries are often knocked down bylater research. The most basic anomaly, ofexcess volatility, seems hardly to have beenknocked down, and itis in fact graphicallyreinforced by the experience of the past fewyears in the stock marketsof the world. Moreover, the mere fact that anomaliessometimes disappear orswitch signs withtime is no evidencethat themarkets are fully rational. That is also what wewould expect to see happen evenin highly irrational markets. (It wouldseem peculiar to argue that irrationalmarkets should displayregular and lastingpatterns!) Even the basic relationsuggested by market inef Ž ciency,that stocks whoseprice is bid up byinvestors will tend togo back down later,and stocks that areunderpriced by investors will tend togo up later,is not arelationthat can beeasily tested or that should hold inalltime periods. The fundamental valueof stocks ishard tomeasure, and, moreover,if speculative bubbles (either positive bubblesor negative bubbles) last a longtime, then eventhis fundamental relation maynot beobserved except in very long sample periods. In furtherresearch, it is important to bear in mind the demonstrated weak- nesses of efŽ cientmarkets theory and maintain an eclecticapproach. Whiletheo- reticalmodels of ef Ž cientmarkets have theirplace as illustrationsor characteriza- tions ofan idealworld, we cannot maintain themin their pure form as accurate descriptorsof actual markets. Indeed, wehave todistance ourselvesfrom the presumption that Ž nancial marketsalways work well and that pricechanges alwaysre  ectgenuine informa- tion. Evidencefrom behavioral Ž nance helpsus tounderstand, forexample, that therecent worldwide stock marketboom, and then crash after2000, had itsorigins inhuman foiblesand arbitraryfeedback relations and must have generateda real and substantial misallocationof resources. The challenge for is tomake this realitya betterpart oftheir models.

References

Andreassen,Paul and Stephen Kraus. 1988. “StyleInvesting. ” NBERWorking Paper No. “JudgmentalPrediction by Extrapolation. ” Un- w8039. publishedpaper, Department of Psychology, Barsky, Robertand J. BradfordDe Long. HarvardUniversity. 1993. “Why Doesthe StockMarket Fluctuate? ” Anonymous. 1637. Samen-spraecktusschen Waer- QuarterlyJournal of Economics .May, 108,pp. 291 – mondtende Gaergoedt nopende de opkomste endeon- 311. derganghvan ora. Haerlem:Adriaen Roman, Bem,Daryl J. 1965. “An ExperimentalAnalysis reprintedin EconomischHistorisch Jaarboek. 1926, ofSelf-Persuasion. ” Journalof Experimental Social 12:20– 43, pp. 28–29. Psychology. 1, pp.199–218. Barberis,Nicholas and Andrei Shleifer. 2000. Breeden,Douglas T. 1979. “An Intertemporal From EfŽcientMarkets Theory to Behavioral Finance 103

Asset PricingModel with Stochastic Consump- kets: II.” Journalof Finance. December,46:5, tionand Investment Opportunities. ” Journal of pp. 1575–617. FinancialEconomics .7:2, pp. 265 –96. Fama,Eugene F. 1998. “Market EfŽ ciency, Campbell,John Y. 1991. “AVarianceDecom- Long-TermReturns, and BehavioralFinance. ” positionfor Stock Returns. ” EconomicJournal. Journalof .September,49:3, March, 101:405,pp. 157 –79. pp. 283–306. Campbell,John Y. andRobert J. Shiller. 1988. Figlewski,Stephen. 1981. “TheInformational “StockPrices, Earnings, and ExpectedDivi- Effects ofRestrictionson Short Sales: Some Em- dends.” Journalof Finance .July,43:3, pp. 661 –76. piricalEvidence. ” Journalof Financial and Quan- Campbell,John Y. andRobert J. Shiller. 1989. titativeAnalysis .November,16:4, pp. 463 –76. “TheDividend-Price Ratio and Expectationsof Figlewski,Stephen and Gwendolyn P. Webb. FutureDividends and DiscountFactors. ” Review 1993. “Options, ShortSales, and MarketCom- ofFinancial Studies .1:3, pp. 195 –228. pleteness.” Journalof Finance .June,48:2, pp. Campbell,John Y., AndrewW. Loand A. 761–77. CraigMacKinlay. 1996. TheEconometrics of Finan- Garber,Peter M. 1990. “FamousFirst Bub- cialMarkets. Princeton: bles.” Journalof Economic Perspectives. Spring,4:2, Press. pp. 35–54. Chen,Joseph, Harrison Hong and Jeremy C. Garber,Peter M. 2000. FamousFirst Bubbles . Stein. 2000. “Breadthof Ownership and StockRe- Cambridge:MIT Press. turns.” Unpublishedpaper, . Goetzmann,William N. andMassimo Massa. Cochrane,John H. 2001. AssetPricing . Prince- 1999. “DailyMomentum and ContrarianBehav- ton: PrincetonUniversity Press. iorof IndexFund Investors. ” Unpublishedpa- Cohen,Randolph, Christopher Polk and per,Yale University. TuomoVuolteenaho . 2002. “TheValue Spread. ” Grinblatt,Mark and Bing Han. 2001. “The Unpublishedpaper, Harvard Business School. DispositionEffect and Momentum. ” Unpub- Forthcoming, Journalof Finance . lishedpaper, Anderson School, UCLA. Cole,Kevin, Jean Helwege and David Laster . Grossman,Sanford J. andRobert J. Shiller. 1996. “StockMarket Valuation Indicators: Is 1981. “TheDeterminants of the Variabilityof This TimeDifferent? ” FinancialAnalysts Journal. StockMarket Prices. ” AmericanEconomic Review . May/June,pp. 56 –64. May, 71:2, pp. 222 –27. Daniel,Kent, David Hirshleifer and Avanidhar Hansen,Lars P. andRavi Jagannathan. 1991. Subramanyam. 1998. “InvestorPsychology and “Implicationsof SecurityMarket Data forMod- SecurityMarket Under- and Overreactions. ” elsof Dynamic Economies. ” Journalof Political Journalof Finance .December,53:6, pp. 1839 – 885. Economy.April,99:2, pp. 225 –62. DeBondt,Werner F. M.andRichard H. Tha- Jarvis,Chris. 1999. “TheRise and Fallof the ler. 1985.“Doesthe StockMarket Overreact? ” PyramidSchemes in Albania. ” International Journalof Finance .July,40:3, pp. 793 – 805. MonetaryFund WorkingPaper No. 99-98. DeLong, J. Bradford,Andrei Shleifer, Law- Jegadeesh,Narasimhan and Sheridan Titman. renceH. Summersand Robert J. Waldmann. 1993. “Returns toBuying Winners and Selling 1990a. “NoiseTrader Risk in Financial Markets. ” Losers:Implications for Stock Market Ef Ž - Journalof Political Economy .August, 98:4, pp. ciency.” Journalof Finance .March, 48:1, pp. 703–38. 65–91. DeLong, J. Bradford,Andrei Shleifer, Law- Jones,Charles M. andOwen A. Lamont. renceH. Summersand Robert J. Waldmann. 2001.“Short-SaleConstraints and StockRe- 1990b. “PositiveFeedback Investment Strategies turns.” NBERWorkingPaper No. 8494. and DestabilizingRational Speculation. ” Journal Jung,Jeeman and Robert J. Shiller. 2002. of Finance.June,45:2, pp. 379 –95. “OneSimple Test ofSamuelson ’sDictumfor the Dechow,Patricia M., Amy P.Hutton,Lisa StockMarket. ” NBERWorking Paper No. Muelbroekand Richard G. Stone. 2001. “Short- w9348. Selling,Fundamental Analysis and StockRe- Kahneman,Daniel and . 1979. turns.” Journalof Financial Economics . Forthcom- “ProspectTheory: An Analysis of DecisionUn- ing. der Risk.” .March, 47:2, pp. 263 –92. Fama,Eugene F. 1970. “EfŽ cientCapital Mar- Lamont,Owen A. andRichard H. Thaler. kets: AReviewof Empirical Work. ” Journal of 2001. “Canthe MarketAdd and Subtract?Mis- Finance.May, 25:2, pp. 383 – 417. pricingin Stock Market Carve-Outs. ” National Fama,Eugene F. 1991. “EfŽ cientCapital Mar- Bureauof Economic Research Working Paper 104Journal of Economic Perspectives

No.8302, 2000. Forthcoming, Journalof Political Shefrin,Hersh. 2001. BehavioralFinance (The Economy. InternationalLibrary of Critical Writings in Financial LeRoy,Stephen F. andRichard D. Porter. Economics), VolumesI throughIV. Cheltenham, 1981. “ThePresent-Value Relation: Tests Based U.K.: EdwardElgar. onImpliedVariance Bounds. ” Econometrica . May, Shiller,Robert J. 1982. “Consumption,Asset 49:3, pp. 97–113. Marketsand MacroeconomicFluctuations. ” Lucas,Robert E. 1978. “Asset Pricesin an Carnegie-RochesterConference Series on Public Policy. ExchangeEconomy. ” Econometrica . November, Autumn, 17, pp. 203 –38. 46:6, pp. 1429 – 445. Shiller,Robert J. 1989. MarketVolatility . Cam- MacKay,Charles. 1996. “Memoirsof Extraor- bridge,Mass.: MIT Press. dinaryPopular Delusions, ” 1841, in Extraordinary Shiller,Robert J. 1990a. “SpeculativePrices PopularDelusions and the Madness of Crowds and and PopularModels. ” Journalof Economic Perspec- Confusio´ndeConfusiones .Martin Fridson,ed. New tives.Spring,4:2, pp. 55 –65. York:John Wiley. Shiller,Robert J. 1990b. “MarketVolatility Malkiel,Burton G. 1973. ARandomWalk Down and InvestorBehavior. ” AmericanEconomic Re- Wall Street. NewYork: W. W.Norton. view. May, 80:2, pp. 58 – 62. Marimon,Ramon, Stephen E. Spearand Shiller,Robert J. 2000a. IrrationalExuberance . Shyam Sunder. 1993. “ExpectationallyDriven Princeton,N.J.: PrincetonUniversity Press. MarketVolatility: An ExperimentalStudy. ” Jour- Shiller,Robert J. 2000b. “MeasuringBubble nalof Economic Theory .October,61:1, pp. 74 –103. Expectationsand InvestorCon Ž dence.” Journal Marsh,Terry A. andRobert C. Merton. 1986. ofPsychology andFinancial Markets . 1:1, pp. “DividendVariability and VarianceBounds Tests 49 –60. forthe Rationality ofStockMarket Prices. ” Amer- Shiller,Robert J. 2002. “Bubbles,Human icanEconomic Review .June,76:3, pp. 483 –98. Judgment,and ExpertOpinion. ” FinancialAna- McGrattan, EllenR. andEdward C. Prescott. lysts Journal . 58, pp. 18 –26. 2001. “Taxes, Regulation,and Asset Prices. ” Un- Shleifer,Andrei. 2000. InefŽcientMarkets . Ox- publishedpaper, Federal Reserve Bank ofMin- ford: OxfordUniversity Press. neapolis. Shielfer,Andrei and Lawrence Summers. Merton,Robert C. 1973. “An Intertemporal 1990. “TheNoise-Trader Approach to Finance. ” CapitalAsset PricingModel. ” Econometrica . Sep- Journalof Economic Perspectives .Spring,4:2 pp. tember,41:5, pp. 867 – 87. 19 –33. Miller,Edward M. 1977. “Risk, Uncertainty Siegel,Jeremy J. 2002. Stocks forthe Long Run. and Divergenceof Opinion. ” Journalof Finance . ThirdEdition. New York: McGraw-Hill. September,32:4, pp. 1151 –168. Smith,Vernon, Gerry Suchanek and Arlington Moss,David A. 2002. WhenAll ElseFails: Gov- Williams. 1988. “Bubbles,Crashes, and Endoge- ernmentas theUltimate Risk Manager .Cambridge, nousExpectations in Experimental Spot Asset Mass.: HarvardUniversity Press. Markets.” Econometrica .September,56:5, pp. Odean,Terrance. 1998. “AreInvestors Reluc- 1119–153. tant toRealize their Losses? ” Journalof Finance . Statman, Meirand Hersh Shefrin. 1985. “The October,53:5, pp. 1775 –798. Dispositionto Sell Winners Too Early and Ride Samuelson,Paul A . 1998. “SummingUp on LosersToo Long: Theory and Evidence. ” Journal BusinessCycles: Opening Address, ” in Beyond of Finance.July,40:3, pp. 777 –90. Shocks:What Causes Business Cycles .JeffreyC. Fu- Thaler,Richard . 2003. Advancesin Behavioral hrerand Scott Schuh, eds.Boston: FederalRe- Finance II.NewYork: Russell Sage. serveBank ofBoston, pp. 33 –36. Tversky,Amos and . 1974. Scherbina,Anna. 2000. “StockPrices and Dif- “Judgmentunder Uncertainty: Heuristicsand ferencesin Opinion: Empirical Evidence that .” Science.185:4157,pp. 1124 –31. Prices Re ectOptimism. ” Unpublishedpaper, Vuolteenaho,Tuomo . “What DrivesFirm- NorthwesternUniversity. LevelStock Returns? ” Journalof Finance. Febru- Shefrin,Hersh. 2000. BeyondGreed and Fear: ary, 57:1, pp. 233 – 64. UnderstandingBehavioral Finance and the Psychol- West,Kenneth D. 1988. “DividendInnova- ogy ofInvesting .Boston, Mass.: HarvardBusiness tions and StockPrice Volatility. ” Econometrica . SchoolPress. January, 56:1, pp. 36 –71.