Eugene Fama - Mercados Eficientes

SESSION TOPIC: STOCK MARKET PRICE BEHAVIOR
SESSION CHAIRMAN: BURTON G. MALKIEL
EFFICIENT CAPITAL MARKETS: A REVIEW OF

THEORY AND EMPIRICAL WORK*
EUGENE F. FAMA**
I. INTRODUCTION
THE PRIMARYROLE of the capital market is allocation of ownership of the
economy'scapitalstock.In generalterms,theideal is a marketin whichprices
provideaccuratesignalsforresourceallocation: that is, a marketin which
firmscan make production-investment decisions,and investorscan choose
amongthe securitiesthat representownershipof firms'activitiesunderthe
assumptionthat securitypricesat any time "fullyreflect"all available in-
formation. A marketin whichpricesalways"fullyreflect"availableinforma-
tionis called"efficient."
This paper reviewsthe theoreticaland empiricalliteratureon the efficient
marketsmodel.Aftera discussionof the theory,empiricalwork concerned
withthe adjustmentof securitypricesto threerelevantinformation subsets
is considered.First,weak formtests,in whichthe information set is just
historicalprices,are discussed.Then semi-strong formtests,in whichthecon-
cernis whetherpricesefficiently adjust to otherinformation thatis obviously
publiclyavailable (e.g., announcements of annual earnings,stocksplits,etc.)
are considered.Finally,strongformtestsconcernedwithwhethergivenin-
vestorsor groupshave monopolistic access to any information relevantfor
price formation are reviewed.'We shall concludethat,with but a few ex-
ceptions,theefficient marketsmodelstandsup well.
Though we proceed fromtheoryto empiricalwork,to keep the proper
historicalperspectivewe shouldnoteto a largeextentthe empiricalworkin
thisarea precededthedevelopment of thetheory.The theoryis presentedfirst
here in orderto moreeasily judge whichof the empiricalresultsare most
relevantfromtheviewpointof thetheory.The empiricalworkitself,however,
willthenbe reviewedin moreor less historicalsequence.
Finally,the perceptivereaderwill surelyrecognizeinstancesin thispaper
whererelevantstudiesare not specifically discussed.In such cases my apol-
ogies should be taken for granted. The area is so bountifulthat some such
injusticesare unavoidable. But the primarygoal here will have been ac-
complishedif a coherentpicture of the main lines of the work on efficient
an
marketsis presented,along with accuratepictureof the currentstate of
thearts.
* Researchon thisprojectwas supported by a grantfromtheNationalScienceFoundation.I
am indebtedto ArthurLaffer,RobertAliber,Ray Ball, MichaelJensen,JamesLorie, Merton
Miller,CharlesNelson,RichardRoll,WilliamTaylor,and Ross Wattsfortheirhelpfulcomments.
** University of Chicago-JointSessionwiththe EconometricSociety.
betweenweak and strongformtestswas firstsuggested
1. The distinction by HarryRoberts.
383
384 The Journalof Finance
II. THE THEORY OF EFFICIENT MARKETS
A. ExpectedReturnor "Fair Game" Models

The definitionalstatement thatin an efficientmarketprices"fullyreflect"
availableinformation is so generalthatit has no empiricallytestableimplica-
tions.To make the modeltestable,the processof price formation mustbe
specifiedin moredetail. In essencewe mustdefinesomewhatmore exactly
whatis meantby the term"fullyreflect."
One possibilitywouldbe to posit that equilibriumprices (or expectedre-
turns) on securitiesare generatedas in the "two parameter"Sharpe [40]-
Lintner[24, 25] world.In general,however,the theoreticalmodelsand es-
peciallythe empiricaltests of capital marketefficiency have not been this
specific.Most of the availableworkis based onlyon the assumptionthatthe
conditionsof marketequilibriumcan (somehow)be statedin termsof ex-
pectedreturns.In generalterms,like the two parametermodelsuch theories
wouldpositthatconditionalon somerelevantinformation set,the equilibrium
expectedreturnon a securityis a function of its "risk."And different
theories
woulddiffer primarily in how "risk"is defined.
All membersof the class of such "expectedreturntheories"can, however,
be describednotationally as follows:
E(gj,t+,I|@t) =[I + E(r-,t+1|0t) ]pjtl 1
whereE is theexpectedvalue operator;pit is thepriceof securityj at timet;

cash income
of any intermediate
pj,t+iis its priceat t + 1 (withreinvestment
fromthe security); ri,t+iis the one-periodpercentagereturn(pi,t+l- pjt)/
pjt; (Dtis a generalsymbolforwhateverset of information is assumedto be
"fully reflected"in the price at t; and the tildes indicate that pj,t+i and r,t+i
are randomvariablesat t.
The value of theequilibriumexpectedreturnE(rj,t+llijt) projectedon the
basis of theinformation iJtwouldbe determined fromtheparticularexpected
returntheoryat hand. The conditionalexpectationnotationof (1) is meant
to imply,however,thatwhateverexpectedreturnmodelis assumedto apply,
the information in 1t is fullyutilizedin determining equilibriumexpected
returns. in theformation
And thisis the sensein which1t is "fullyreflected"
of the price pjt.
But we shouldnoterightoffthat,simpleas it is, the assumptionthat the
conditionsof marketequilibriumcan be statedin termsof expectedreturns
elevatesthe purelymathematicalconceptof expectedvalue to a status not
necessarily implied by the general notion of market efficiency.The expected
value is just one of manypossible summarymeasuresof a distribution of
returns, and marketefficiencyper se (i.e., thegeneralnotionthatprices"fully
reflect"availableinformation) does notimbueit withany specialimportance.
Thus, theresultsof testsbased on thisassumptiondependto some extenton
its validityas wellas on theefficiency of themarket.But somesuch assump-
tion is the unavoidableprice one mustpay to give the theoryof efficient
marketsempiricalcontent.
The assumptionsthat the conditionsof marketequilibriumcan be stated
Capital Markets
Efficient 385
in termsof expectedreturnsand thatequilibrium expectedreturnsare formed
on thebasis of (and thus"fullyreflect")the information set (Dthave a major
empiricalimplication-theyrule out the possibilityof tradingsystemsbased
onlyon information in (Dtthathave expectedprofitsor returnsin excess of
equilibriumexpectedprofitsor returns.Thus let
Xj,t+l - Pj,t+l- E(pj,t+1I4Dt). (2)
Then
E (:j',t+lJ4t) =?0 (3)
which,by definition,
says thatthesequence{xjt} is a "fairgame"withrespect
to theinformationsequence{@t}. Or, equivalently,let
zjt+l =rj,t+l - E(rj t+lt), (4)
then
E(Zjt+i141t) y, (5)
so thatthesequence{zjt} is also a "fairgame"withrespectto theinformation
sequence {41}.
In economic terms,xJ,t+i is the excess market value of security j at time
t + 1: it is the differencebetweenthe observedpriceand the expectedvalue
of the pricethatwas projectedat t on the basis of the information (Dt.And
similarly,zj,t+l is the returnat t + 1 in excess of the equilibriumexpected
returnprojectedat t. Let
a(1(t) [al(QDt), a2(2Dt), . . ., a1((Dt)]
be any tradingsystembased on 1?t whichtellstheinvestortheamountsaj ((It)

of fundsavailableat t thatare to be investedin each of then availablesecu-
rities.The totalexcessmarketvalue at t + 1 thatwillbe generatedby such a
systemis
n
Vt+ Ejja((Dt)
j=1
[rj,t+l-E(rj,t+llt)],
which,fromthe "fairgame"propertyof (5) has expectation,
Z
n
E (Vt+lIDt) cj((<Dt)E(!j,t+1[(Dt) = 0.
j=l
The expectedreturnor "fair game" efficient marketsmodel2has other
importanttestableimplications,but theseare bettersaved forthe later dis-
cussionof theempiricalwork.Now we turnto twospecialcases of themodel,
thesubmartingale and the randomwalk,that (as we shall see later) play an
importantrolein the empiricalliterature.
2. Though we shall sometimesreferto the model summarizedby (1) as the "fair game" model,
keep in mind that the "fair game" propertiesof the model are implicationsof the assumptionsthat
(i) the conditionsof marketequilibriumcan be stated in termsof expected returns,and (ii) the
information(Pt is fullyutilized by the market in formingequilibriumexpected returnsand thus
currentprices.
The role of "fair game" models in the theory of efficientmarkets was first recognized and
studied rigorouslyby Mandelbrot r27] and Samuelson [38]. Their work will be discussedin more
detail later.
B. Tke Submartingale Model
Supposewe assumein (1) thatforall t and (Dt
E("',t+1Ilt) > Pit, orequivalently, E(i,t+1iDt) > 0. (6)
This is a statement thatthepricesequence{pit} forsecurityj followsa sub-
martingalewith respectto the information sequence Ol?t},whichis to say
nothingmorethanthat the expected value of nextperiod'sprice,as projected
on thebasis oftheinformation (Dt,is equal to or greaterthanthecurrent price.
If (6) holds as an equality(so thatexpectedreturns and price changesare
zero), thenthepricesequencefollowsa martingale.
A submartingale in priceshas one important empiricalimplication.Consider
theset of "one securityand cash" mechanicaltradingrulesby whichwe mean
systemsthatconcentrate on individualsecuritiesand thatdefinetheconditions
underwhichthe investorwouldhold a givensecurity,sell it short,or simply
hold cash at any timet. Then the assumptionof (6) that expectedreturns
conditionalon (Dt are non-negative directlyimpliesthat such tradingrules
based onlyon theinformation in Ct cannothave greaterexpectedprofitsthan
a policyof alwaysbuylng-and-holding the securityduringthe futureperiodin
question.Tests of such rules will be an importantpart of the empirical
evidenceon theefficient marketsmodel.8
C. The Random Walk Model

In the earlytreatments marketsmodel,the statementthat
of the efficient
the currentprice of a security"fully reflects"available information was
assumedto implythatsuccessivepricechanges (or moreusually,successive
one-periodreturns)are independent. In addition,it was usuallyassumedthat
successivechanges(or returns)are identicallydistributed. Togetherthe two
hypotheses the randomwalk model.Formally,themodelsays
constitute
= f(rj,t+?),
f(rj,t+?ItDt) (7)
whichis the usual statementthat the conditionaland marginalprobability
of an independentrandomn
distributions variable are identical.In addition,
thedensityfunction f mustbe thesame forall t.4
3. Note thattheexpectedprofitability and cash" tradingsystems

of "one security buy-
vis-'a-vis
and-holdis not ruledout by the generalexpectedreturnor "fairgame"efficient marketsmodel.
The latterrulesout systemswithexpectedprofits in excessof equilibrium expectedreturns, but
sincein principleit allowsequilibriunmexpectedreturns to be negative,holdingcash (whichalways
hag zero actual and thusexpectedreturn)may have higherexpectedreturnthan holdingsome
security.
And negativeequilibriumn expectedreturnsforsomesecurities are quite possible.For example,
in theSharpe[40]-Lintner[24, 25] model(whichis in turna naturalextension of the portfolio
modelsof Markowitz[30] and Tobin [43]) theequilibrium expectedreturnon a security depends
on theextentto whichthe dispersion in thesecurity'sreturndistributionig relatedto dispersion
in the returnson all othersecurities. A security whosereturns on averagemove oppositeto the
generalmarketis particularly valuable in reducingdispersionof portfolioreturns,and so its
equilibriumexpectedreturn maywellbe negative.
4. The terminology is loose.Priceswill onlyfollowa randomwalk if pricechangesare inde-
pendent,identically distributed; and even thenwe shouldsay "randomwalk with drift"since
expectedpricechangescan be non-zero.If one-periodreturnsare independent, identicallydis.
tributed,priceswillnot followa randomwalk sincethe distribution of pricechangeswill depend
CapitalMarkets
Efficient 387
Expression(7) of coursesays muchmorethanthegeneralexpectedreturn
modelsummarizedby (1). For example,if we restrict(1) by assumingthat
theexpectedreturnon securityj is constantovertime,thenwe have
E(,j,t+1|f't)= E(rj,t+?). (8)
This says thatthe meanof the distribution of rj,t+lis independent of thein-
formation available at t, t, whereasthe randomwalk modelof (7) in addi-
tionsays thatthe entiredistribution is independent of CF.5
We argue later that it is best to regardthe randomwalk model as an
extensionof the generalexpectedreturnor "fair game" efficient markets
modelin the sense of makinga moredetailedstatementabout the economic
environment. The "fairgame" modeljust says thatthe conditionsof market
equilibrium can be statedin termsof expectedreturns, and thusit says little
about thedetailsof thestochasticprocessgenerating returns.A randomwalk
arises withinthe contextof such a modelwhen the environment is (fortu-
itously)such thattheevolutionof investortastesand the processgenerating
new information combineto produceequilibriain whichreturndistributions
repeatthemselves throughtime.
Thus it is not surprising thatempiricaltestsof the "randomwalk" model
thatare in facttestsof "fairgame" propertiesare morestrongly in support
of themodelthantestsof theadditional(and, fromtheviewpointof expected
returnmarketefficiency, superfluous)pure independence assumption.(But it
is perhapsequally surprising that,as we shall soon see, the evidenceagainst
theindependence of returnsovertimeis as weak as it is.)
D. MarketConditionsConsistentwithEfficiency
Before turningto the empiricalwork,however,a few words about the
marketconditionsthatmighthelp or hinderefficient adjustmentof pricesto
information are in order.First,it is easy to determine sufficientconditionsfor
capitalmarketefficiency. For example,considera marketin which (i) there
are no transactions costsin tradingsecurities,(ii) all availableinformation is
costlesslyavailable to all marketparticipants, and (iii) all agree on the im-
plicationsof currentinformation for the currentprice and distributions of
futurepricesof each security.In sucha market,thecurrentpriceof a security
obviously"fullyreflects"all availableinformation.
But a frictionless marketin whichall information is freelyavailable and
investorsagreeon its implications is, of course,notdescriptive of marketsmet
in practice.Fortunately, theseconditionsare sufficient formarketefficiency,
but not necessary.For example,as long as transactorstake account of all
on the pricelevel.But thoughrigorousterminology is usuallydesirable,our loose use of terms
shouldnot cause confusion;and our usage followsthat of the efficient
marketsliterature.
Note also thatin the randomwalk literature,
theinformation set (t in (7) is usuallyassumed
to include only the past returnhistory,rj,t,rj t-1 . . .
5. The randomwalk modeldoes not say, however,that past information is of no value in

assessingdistributions
of futurereturns.Indeed since returndistributions
are assumedto be
stationarythroughtime,past returnsare the bestsourceof suchinformation.
The randomwalk
modeldoessay,however, thatthesequence(or theorder)of thepast returns
is of no consequence
of futurereturns.
in assessingdistributions
388 The Journal
of Finance
available information,
even large transactionscosts that inhibitthe flowof
transactionsdo notin themselvesimplythatwhentransactions do take place,
priceswillnot "fullyreflect"available information.Similarly(and speaking,
as above, somewhatloosely),the marketmay be efficient if "sufficient
num-
bers" of investorshave readyaccess to available information. And disagree-
mentamonginvestorsabouttheimplications of giveninformation does notin
itselfimplymarketinefficiencyunlessthereare investorswho can consistently
make betterevaluationsof available information thanare implicitin market
prices.
But thoughtransactions costs,information
thatis not freelyavailableto all
investors,and disagreement amonginvestorsabout the implicationsof given
information are notnecessarilysourcesof marketinefficiency, theyare poten-
tial sources.And all threeexistto someextentin real worldmarkets.Measur-
ing theireffectson theprocessof priceformation is, of course,themajorgoal
of empiricalworkin thisarea.
III. THE EVIDENCE

All the empiricalresearchon the theoryof efficient marketshas been con-
cerned with whetherprices "fully reflect"particularsubsets of available
information. Historically,the empiricalworkevolvedmoreor less as follows.
The initialstudieswereconcernedwithwhatwe call weak formtestsin which
theinformation subsetof interestis just past price (or return)histories.Most
of theresultsherecomefromtherandomwalkliterature. Whenextensivetests
seemedto supportthe efficiency hypothesis at thislevel,attentionwas turned
to semi-strongformtestsin whichtheconcernis thespeed of priceadjustment
to otherobviouslypubliclyavailable information(e.g., announcements of
stock splits,annual reports,new securityissues,etc.). Finally,strongform
testsin whichthe concernis whetherany investoror groups (e.g., manage-
mentsof mutualfunds)have monopolistic access to any information relevant
forthe formation of priceshave recentlyappeared.We reviewthe empirical
researchin moreor less thishistoricalsequence.
First, however,we should note that what we have called the efficient
marketsmodel in the discussionsof earliersectionsis the hypothesisthat
securitypricesat any pointin time"fullyreflect"all available information.
Thoughwe shall arguethatthemodelstandsup ratherwell to the data, it is
obviouslyan extremenull hypothesis.And, like any otherextremenull hy-
posthesis,we do not expectit to be literallytrue.The categorization of the
testsintoweak,semi-strong, and strongformwill servetheusefulpurposeof
allowingus to pinpointthelevelofinformation at whichthehypothesis breaks
down.And we shall contendthatthereis no important evidenceagainstthe
hypothesisin the weak and semi-strong formtests (i.e., prices seem to effi-
to
cientlyadjust obviouslypublicly available information), and onlylimited
evidenceagainst the hypothesisin the strongformtests (i.e., monopolistic
access to informationaboutpricesdoes notseemto be a prevalentphenomenon
in the investment community).
Efficient
CapitalMarkets 389
A. Weak Form Tests of theEfficient MarketsModel
1. RandomWalks and Fair Games: A Little HistoricalBackground
marketscan be con-
As notedearlier,all of the empiricalworkon efficient
sideredwithinthe contextof the generalexpectedreturnor "fair game"
model,and muchof the evidencebears directlyon the special submartingale
expectedreturnmodelof (6). Indeed,in the early literature, discussionsof
the efficientmarketsmodelwerephrasedin terms of the even morespecial
randomwalkmodel,thoughwe shallarguethatmostof theearlyauthorswere
in factconcernedwithmoregeneralversionsof the "fairgame" model.
Someof theconfusionin theearlyrandomwalkwritings is understandable.
Researchon securitypricesdid not beginwiththe development of a theory
of price formation whichwas thensubjectedto empiricaltests.Rather,the
impetusforthe development of a theorycame fromthe accumulationof ev-
idencein the middle 1950's and early 1960's that the behaviorof common
stockand otherspeculativepricescould be well approximated by a random
walk. Faced withthe evidence,economistsfeltcompelledto offersome ratio-
nalization.What resultedwas a theoryof efficientmarketsstatedin termsof
randomwalks,but usuallyimplyingsomemoregeneral"fairgame" model.
It was notuntiltheworkof Samuelson[38] and Mandelbrot[27] in 1965
and 1966 thatthe role of "fair game" expectedreturnmodelsin the theory
marketsand the relationships
of efficient betweenthesemodelsand thetheory
of randomwalks wererigorously studied.6And thesepapers came somewhat
afterthemajorempiricalworkon randomwalks. In the earlierwork,"theo-
retical"discussions,thoughusuallyintuitivelyappealing,werealwayslacking
in rigorand ofteneithervague or ad hoc. In short,until the Mandelbrot-
Samuelsonmodelsappeared,thereexisteda large body of empiricalresults
in searchof a rigoroustheory.
Thus, thoughhis contributions wereignoredforsixtyyears,the firststate-
mentand testof the randomwalk modelwas thatof Bachelier[3] in 1900.
But his "fundamental principle"forthe behaviorof priceswas thatspecula-
tionshouldbe a "fairgame"; in particular,the expectedprofitsto thespecu-
lator should be zero. With the benefitof the moderntheoryof stochastic
processes,we knownowthattheprocessimpliedby thisfundamental principle
is a martingale.
AfterBachelier,researchon thebehaviorof securitypriceslaggeduntilthe
6. Basing theiranalyses on futurescontractsin commoditymarkets,Mandelbrot and Samuelson

show that if the price of such a contractat time t is the expected value at t (given information
t) of the spot price at the terminationof the contract,then the futuresprice will follow a
martingalewith respectto the informationsequence {jt); that is, the expected price change from
period to period will be zero, and the price changes will be a "fair game." If the equilibriumex-
pected returnis not assumed to be zero, our more general "fair game" model, summarizedby (1),
is obtained.
But though the Mandelbrot-Samuelsonapproach certainly illuminates the process of price
formationin commoditymarkets,we have seen that "fair game" expected returnmodels can be
derived in much simplerfashion.In particular,(1) is just a formalizationof the assumptionsthat
the conditions of market equilibriumcan be stated in terms of expected returns and that the
informationt is used in formingmarketprices at t.
comingof the computer.In 1953 Kendall [21] examinedthe behaviorof
weeklychangesin nineteenindicesof Britishindustrialshare pricesand in
spot prices for cotton (New York) and wheat (Chicago). Afterextensive
he suggests,in quite graphicterms:
analysisof serial correlations,
Theserieslookslikea wandering one,almostas ifoncea weektheDemonofChance
drewa random number from population
a symetrical andaddedit
of fixeddispersion
to thecurrentpriceto determinethenextweek'sprice[21,p. 13].
Kendall's conclusionhad in fact been suggestedearlierby Working[47],
thoughhis suggestion lackedtheforceprovidedby Kendall'sempiricalresults.
And theimplications of theconclusionforstockmarketresearchand financial
analysiswere laterunderlinedby Roberts[36].
But thesuggestion by Kendall,Working,and Robertsthatseriesof specula-
tivepricesmaybe well describedby randomwalkswas based on observation.
None of theseauthorsattemptedto providemucheconomicrationaleforthe
hypothesis,and,indeed,Kendall feltthateconomists wouldgenerallyrejectit.
Osborne [33] suggestedmarketconditions,similar to those assumed by
Bachelier,thatwouldlead to a randomwalk. But in his model,independence
of successivepricechangesderivesfromthe assumptionthatthe decisionsof
investorsin an individualsecurityare independentfrom transactionto
transaction-which is littlein theway of an economicmodel.
Whenevereconomists(priorto Mandelbrotand Samuelson) triedto pro-
vide economicjustificationfor the randomwalk, their argumentsusually
implieda "fairgame." For example,Alexander[8, p. 200] states:
If onewereto startoutwiththeassumption thata stockor commodity is
speculation
a "fairgame"withequal expectation withan
of gainor loss or, moreaccurately,
expectationof zerogain,onewouldbe wellon thewayto picturing thebehavior of
pricesas a random
speculative walk.
There is an awarenessherethatthe "fairgame" assumptionis not sufficient
to lead to a randomwalk, but Alexandernever expands on the comment.
Similarly, Cootner[8, p. 232] states:
If anysubstantial groupof buyersthought pricesweretoo low,theirbuyingwould
forceup theprices.The reversewouldbe trueforsellers. due
Exceptforappreciation
to earningsretention,theconditional of tomorrow's
expectation price,giventoday's
price,is today'sprice.
In sucha world, theonlypricechanges thatwouldoccurarethosethatresultfrom
newinformation. to be non-ran-
Sincethereis no reasonto expectthatinformation
domin appearance, pricechangesof a stockshouldbe random
theperiod-to-period
movements, statistically
independentofoneanother.
Thoughsomewhatimprecise, thelast sentenceof thefirstparagraphseemsto
pointto a "fairgame" modelratherthana randomwalk.' In thislight,the
secondparagraphcan be viewedas an attemptto describeenvironmentalcon-
ditionsthatwouldreducea "fairgame" to a randomwalk. But the specifica-
tionimposedon theinformationgenerating forthispur-
processis insufficient
pose; one would, for example,also have to say somethingabout investor
conditioning
7. The appropriate would be "Giventhe sequenceof historicalprices."
statement
CapitalMarkets
Efficient 391
tastes. Finally,lest I be accused of criticizingotherstoo severelyfor am-
biguity,lack of rigorand incorrectconclusions,
By contrast,the stock markettraderhas a much more practicalcriterionfor
judgingwhat constitutesimportantdependencein successiveprice changes.For his
purposesthe randomwalk modelis valid as long as knowledgeof the past behavior
of theseriesof pricechangescannotbe used to increaseexpectedgains.More specif-
ically,the independence assumptionis an adequate descriptionof realityas long as
theactual degreeof dependencein theseriesof pricechangesis not sufficient
to allow
the past historyof the seriesto be used to predictthe futurein a way whichmakes
expectedprofitsgreaterthan they would be under a naive buy-andhold model
[10, p 35].
We knownow, of course,that this last conditionhardlyrequiresa random
walk. It willin factbe metby thesubmartingalemodelof (6).
But one shouldnot be too hard on the theoreticalefforts
of the earlyem-
piricalrandomwalk literature.The arguments wereusuallyappealing;where
theyfellshortwas in awarenessof developments in the theoryof stochastic
processes.Moreover,we shall now see thatmostof the empiricalevidencein
the randomwalk literaturecan easilybe interpreted
as testsof moregeneral
expectedreturnor "fairgame" models.8
2. Tests of MarketEfficiency
in theRandomWalk Literature
As discussed earlier,"fair game" models imply the "impossibility"of
varioussortsof tradingsystems.Someof therandomwalk literature has been
concernedwithtestingtheprofitability of suchsystems.More of theliterature
has, however,been concernedwithtestsof serial covariancesof returns.We
shall now show that,like a randomwalk, the serial covariancesof a "fair
game" are zero, so thatthesetestsare also relevantforthe expectedreturn
models.
If Xt is a "fair game,"its unconditionalexpectationis zero and its serial
covariancecan be written in generalformas:
E (it+r iit) xtE (it+rIxt) f(xt)dxt,

xt
wheref indicatesa densityfunction.

But if Xt is a "fairgame,"
E (5Et+ lxt) = 0.
8. Our briefhistoricalreviewis meant only to provide perspective,and it is, of course,somewhat

incomplete.For example, we have ignored the importantcontributionsto the early random walk
literaturein studies of warrants and other options by Sprenkle, Kruizenga, Boness, and others.
Much of this early work on options is summarizedin [8].
9. More generally,if the sequence {xj is a fair game with respectto the informationsequence
{(Dt}, (i.e., E(Xt+1?It) = 0 for aH Pt); then xt is a fair game with respect to any Vt that is a
subset of (t (i.e., E(xt+? I t) = 0 for all 't). To show this,let (P = (Vt, V"t). Then, using
Stieltjes integralsand the symbol F to denote cumulative distinctionfunctions,the conditional
expectation
E(xt+ll,t) = f xt+dF(xt+i t1e, = f[f xt+dF(xt+1I4t)

, % ] dF(O)-
bt Xtt+ (Pt Xt.+1
From this it followsthat forall lags, the serial covariancesbetweenlagged
valuesofa "fairgame"variableare zero.Thus, observations of a "fairgame"
variableare linearlyindependent.10
But the "fair game" model does not necessarilyimply that the serial
covariancesof one-periodreturnsare zero. In the weak formtests of this
modelthe "fairgame" variableis
zj,t -rj,t- E(r-j,tIrj,t_j,
rj,t-2, . . .). (Cf. fn.9) (9)
But thecovariancebetween,forexample,ritand rj,t+iis
E( rFj,t+j-E(r'j,t+j)] [r-jt-E(r'jt)])
- [rjt-E(rjt)] [E(rj,t+lIrjt)-E(rj,t?i)]f(rjt)drjt,
rjt
and (9) does not implythat E(rj,t+?irjt) E(ij,t+1): In the "fair game"
marketsmodel,the deviationof the returnfort + 1 fromits condi-
efficient
tionalexpectationis a "fair game" variable,but the conditionalexpectation
itselfcan dependon thereturnobservedfort.1'
In the randomwalk literature, this problemis not recognized,since it is
assumed that the expected return(and indeed the entiredistribution of
returns)is stationarythroughtime.In practice,thisimpliesestimating serial
covariancesby takingcross productsof deviationsof observedreturnsfrom
theoverallsamplemeanreturn.It is somewhatfortuitous, then,thatthispro-
cedure,whichrepresents a rathergrossapproximation fromthe viewpointof
the generalexpectedreturnefficientmarketsmodel,does not seemto greatly
affecttheresultsof the covariancetests,at least forcommonstocks.'2
But the integralin bracketsis just E(xt?iI |t) which by the "fair game" assumptionis 0, so that
E(xt?+l 't) = 0 forall Vt C t.
10. But thoughzero serial covariancesare consistentwith a "fair game," they do not implysuch
a process. A "fair game" also rules out many types of non linear dependence. Lhus using argu-
mentssimilarto those above, it can be shown that if x is a "fair game," E(xtxt+l . . . xt+r) = 0
for all -r,which is not implied by E(Xtxt+T) = 0 for all T. For example, considera three-period
case where x must be either? 1. Suppose the process is xt+2 = sign (xtxt+?), i.e.,
xt Xt+l i Xt+2
? + e +
- ?
If probabilitiesare uniformlydistributedacross events,
E(xt?21xt+l) = E(xt+2Ixt) .= E(xt+llxt) = E(xt+2) = E(xt+?) = E(xt) = 0,

so that all pairwise serial covariances are zero. But 'the process is not a "fair game," since
E(Xt?2lXt+?, xt) & 0, and knowledgeof (xt+i, Xt) can be used as the basis of a simple "system"
with positive expected profit.
11. For example,suppose the level of one-periodreturnsfollows a martingaleso that
E(fijt+1?rjt, rj,t_1 ... ) = rjt.
Then covariances between successive returnswill be nonzero (though in this special case first
differences
of returnswill be uncorrelated).
12. The reason is probably that for stocks, changes in equilibrium expected returns for the
EfficientCapital Markets 393
TABLE 1 (from [10])
forOne-, Four-, Nine-,and Sixteen-Day
First-orderSerial CorrelationCoefficients
Changesin Loge Price
Differencing
Interval(Days)
Stock One Four Nine Sixteen
Allied Chemical .017 .029 -.091 -.118
Alcoa .118* .095 -.112 -.044
AmericanCan -.087* -.124* -.060 .031
A. T. & T. -.039 -.010 -.009 -.003
AmericanTobacco .111* -.175* .033 .007
Anaconda .067* -.068 -.125 .202
BethlehemSteel .013 -.122 -.148 .112
Chrysler .012 .060 -.026 .040
Du Pont .013 .069 -.043 -.055
Eastman Kodak .025 -o.006 -.053 -.023
GeneralElectric .011 .020 -.004 .000
GeneralFoods .061* -.005 -.140 -.098
General Motors -.004 -.128* .009 -.028
Goodyear -.123* .001 -.037 .033
InternationalHarvester -.017 -.068 -.244* .116
InternationalNickel .096* .038 .124 .041
InternationalPaper .046 .060 -.004 -.010
JohnsManville .006 -.068 -.002 .002
Owens Illinois -.021 -.006 .003 -.022
Procter& Gamble .099* -.006 .098 .076
Sears .097* -.070 -.113 .041
StandardOil (Calif.) .025 -.143* -.046 .040
StandardOil (N.J.) .008 -.109 -.082 -.121
Swift& Co. -.004 -.072 .118 -.197
Texaco .094* -.o53 -.047 -.178
Union Carbide .107* .049 -.101 .124
United Aircraft .014 -.190* -.192* -.040
U.S. Steel .040 -.006 -.056 .236*
Westinghouse -.02 7 -.097 -.137 .067
Woolworth .028 -.033 -.112 .040
* Coefficient
is twiceits computedstandarderror.
For example,Table 1 (taken from[10]) showsthe serialcorrelations be-

tweensuccessivechangesin the naturallog of price for each of the thirty
stocksof theDow JonesIndustrialAverage,fortimeperiodsthatvaryslightly
fromstockto stock,but usuallyrunfromabout theend of 1957 to September
26, 1962.The serialcorrelations
of successivechangesin loge priceare shown
fordifferencing
intervalsof one,four,nine,and sixteendays.13
commondifferencing intervalsof a day, a week, or a month,are trivial relativeto other sources of
variation in returns.Later, when we consider Roll's work [37], we shall see that this is not true
for one week returnson U.S. GovernmentTreasury Bills.
13. The use of changesin loge price as the measure of returnis common in the random walk
literature.It can be justifiedin several ways. But for currentpurposes,it is sufficient
to note that
for price changesless than fifteenper cent,the changein loge price is approximatelythe percentage
price change or one-periodreturn.And for differencing intervalsshorterthan one month,returns
in excess of fifteenper cent are unusual. Thus [10] reportsthat for the data of Table 1, tests
carried out on percentage or one-period returnsyielded results essentiallyidentical to the tests
based on changesin loge price.
The resultsin Table 1 are typicalof thosereportedby othersfortestsbased
on serial covariances.(Cf. Kendall [21], Moore [31], Alexander[1], and
the resultsof Grangerand Morgenstern[17] and Godfrey,Grangerand
Morgenstern [16] obtainedby meansof spectralanalysis.) Specifically, there
is no evidenceof substantiallineardependencebetweenlaggedprice changes
or returns. In absolutetermsthemeasuredserialcorrelations are alwaysclose
to zero.
Lookinghard,though,one can probablyfindevidenceof statistically "sig-
nificant"lineardependencein Table 1 (and again thisis trueof resultsre-
portedby others).For the daily returnselevenof the serial correlations are
morethantwicetheircomputedstandarderrors,and twenty-two out of thirty
are positive.On theotherhand,twenty-one and twenty-four of thecoefficients
forthefourand nineday differences are negative.But withsamplesof thesize
underlying Table 1 (N- 1200-1700observations per stockon a dailybasis)
statistically deviationsfromzero covarianceare not necessarily
"significant"
a basis forrejectingthe efficient marketsmodel.For the resultsin Table 1,
the standarderrorsof the serial correlationswere approximatedas (1/
(N-i) )'/2,whichforthe daily data impliesthat a correlationas small as .06
is morethantwiceits standarderror.But a coefficient thissize impliesthata
linearrelationship withthe laggedpricechangecan be used to explainabout
.36% of the variationin the currentprice change,whichis probablyinsig-
nificant froman economicviewpoint. In particular,
it is unlikelythatthesmall
absolutelevels of serialcorrelation thatare alwaysobservedcan be used as
thebasis of substantially profitabletradingsystems.'4
It is, of course,difficultto judge what degreeof serial correlationwould
implythe existenceof tradingrules withsubstantialexpectedprofits.(And
indeedwe shall soonhave to be a littlemorepreciseaboutwhatis impliedby
"substantial"profits.)Moreover,zero serialcovariancesare consistent witha
"fairgame" model,but as notedearlier(fn. 10), thereare typesof nonlinear
dependencethatimplythe existenceof profitable tradingsystems,and yet do
not implynonzeroserial covariances.Thus, formanyreasonsit is desirable
to directlytesttheprofitability of varioustradingrules.
The firstmajorevidenceon tradingruleswas Alexander's[1, 2]. He testsa
varietyof systems,but the most thoroughly examinedcan be decribedas
follows:If the price of a securitymovesup at least y%7,buy and hold the
securityuntilits price movesdownat least y%' froma subsequenthigh,at
whichtimesimultaneously sell and go short.The shortpositionis maintained
untilthe price rises at least y%oabove a subsequentlow, at whichtimeone
coverstheshortpositionand buys.Moves less thany% in eitherdirection are
14. Giventhe evidenceof Kendall [21], Mandelbrot[28], Fama [10] and othersthat large
pricechangesoccurmuchmorefrequently thanwouldbe expectedif the generating processwere
Gaussian,the expression(1/(N-1))'/2 understatesthe samplingdispersion of the serialcorrelation
coefficient,
and thusleads to an overstatement of significance
levels.In addition,the fact that
sampleserialcorrelationsare predominantly of one signor the otheris not in itselfevidenceof
lineardependence.
If, as theworkof King [23] and Blume[7] indicates, thereis a marketfactor
whosebehavioraffects the returnson all securities, the samplebehaviorof this marketfactor
maylead to a predominance of signsof one typein theserialcorrelationsforindividual securities,
even thoughthe populationserial correlations forboth the marketfactorand the returnson
individual are zero.For a moreextensive
securities analysisof theseissuessee [10].
Capital Markets
Efficient 395
ignored.Such a systemis calleda y% filter.It is obviouslya "one securityand
cash" tradingrule,so that the resultsit producesare relevantforthe sub-
martingale expectedreturnmodelof (6).
Afterextensivetestsusingdaily data on price indicesfrom1897 to 1959
and filtersfromone to fiftyper cent, and aftercorrectingsome incorrect
presumptions in theinitialresultsof [1] (see fn.25), in his finalpaperon the
subject,Alexanderconcludes:
In fact,at thispointI shouldadviseanyreaderwhois interested onlyin practical
results,
and whois nota floortraderand so mustpaycommissions, to turnto other
sources
onhowtobeatbuyandhold.The restofthisarticleis devotedprincipally to
a theoreticalconsiderationof whether the observedresultsare consistent witha
random walkhypothesis [8], p. 351).
Later in the paper Alexanderconcludesthat thereis some evidencein his
resultsagainstthe independence assumptionof the randomwalk model.But
marketefficiency does not requirea randomwalk,and fromthe viewpointof
thesubmartingale modelof (6), theconclusionthatthefilters cannotbeat buy-
and-holdis supportforthe efficient marketshypothesis.Furthersupportis
providedby Fama and Blume [13] who comparetheprofitability of various
filtersto buy-and-hold forthe individualstocksof the Dow-JonesIndustrial
Average.(The data are thoseunderlying Table 1.)
But again, lookinghard one can findevidencein the filtertests of both
Alexanderand Fama-Blumethat is inconsistent withthe submartingale ef-
ficientmarketsmodel,if thatmodelis interpreted in a strictsense.In partic-
ular,the resultsforverysmallfilters(1 per centin Alexander'stestsand .5,
1.0, and 1.5 per centin the testsof Fama-Blume)indicatethatit is possible
to devisetradingschemesbased on veryshort-term (preferablyintra-daybut
at most daily) price swingsthat will on average outperform buy-and-hold.
The averageprofitson individualtransactions fromsuch schemesare minis-
cule, but theygeneratetransactionsso frequently that over longerperiods
and ignoringcommissionsthey outperform buy-and-holdby a substantial
margin.These resultsare evidenceof persistenceor positivedependencein
veryshort-term price movements. And, interestingly,this is consistentwith
the evidencefor slightpositivelinear dependencein successivedaily price
changesproducedby theserialcorrelations.15
15. Though strictlyspeaking, such tests of pure independence are not directly relevant for
expected returnmodels, it is interestingthat the conclusionthat very short-termswings in prices
persistslightlylonger than would be expected under the martingalehypothesisis also supported
by the resultsof non-parametricruns testsapplied to the daily data of Table 1. (See [10], Tables
12-15.) For the daily price changes,the actual numberof runs of price changes of the same sign
is less than the expectednumberfor 26 out of 30 stocks.Moreover,of the eightstocksfor which the
actual numberof runs is more than two standarderrorsless than the expectednumber,five of the
same stocks have positive daily, firstorder serial correlationsin Table 1 that are more than
twice theirstandard errors.But in both cases the statistical"significance"of the resultsis largely
a reflectionof the large sample sizes. Just as the serial correlationsare small in absolute terms
(the average is .026), the differences
between the expected and actual number of runs on average
are only three per cent of the total expected number.
On the other hand, it is also interestingthat the runs tests do not support the suggestionof
slight negative dependencein four and nine day changes that appeared in the serial correlations.
In the runs tests such negative dependencewould appear as a tendencyfor the actual number of
runs to exceed the expectednumber.In fact, for the four and nine day price changes,for 17 and
But whenone takesaccountof even theminimum tradingcoststhatwould
be generatedby small filters,theiradvantageover buy-and-hold disappears.
For example,evena floortrader(i.e., a personwho ownsa seat) on theNew
York StockExchangemustpay clearinghouse feeson his tradesthatamount
to about .1 per cent per turnaround transaction(i.e., sales plus purchase).
Fama-Blumeshow that because small filtersproducesuch frequenttrades,
theseminimumtradingcosts are sufficient to wipe out theiradvantageover
buy-and-hold.
Thus thefiltertests,like theserialcorrelations,
produceempirically notice-
able departuresfromthe strictimplicationsof the efficient marketsmodel.
theymighthave,froman economic
But,in spiteof any statisticalsignificance
viewpointthe departuresare so small that it seems hardlyjustifiedto use
themto declarethe marketinefficient.
3. OtherTests of Independencein the Random Walk Literature
It is probablybest to regardthe randomwalk modelas a special case of
themoregeneralexpectedreturnmodelin thesenseof makinga moredetailed
specificationof theeconomicenvironment. That is, thebasic modelof market
equilibriumis the "fair game" expectedreturnmodel,with a randomwalk
arisingwhenadditionalenvironmental conditionsare such that distributions
of one-periodreturnsrepeat themselvesthroughtime.From this viewpoint
violationsof thepureindependence assumption of therandomwalk modelare
to be expected.But whenjudged relativeto the benchmarkprovidedby the
randomwalk model,theseviolationscan provideinsightsinto the natureof
the marketenvironment.
For example,one departurefromthepure independence assumptionof the
randomwalk modelhas been notedby Osborne[34], Fama ([10], Table 17
and Figure8), and others.In particular,largedailypricechangestendto be
followedby large daily changes.The signsof the successorchangesare ap-
parentlyrandom,however,whichindicatesthat the phenomenonrepresents
a denialoftherandomwalkmodelbut notof themarketefficiency hypothesis.
Nevertheless, it is interesting
to speculatewhythe phenomenon mightarise.
It may be that whenimportantnew information comes into the marketit
cannot always be immediatelyevaluated precisely.Thus, sometimesthe
initialpricewill overadjustto the information,and othertimesit will under-
adjust.But sincetheevidenceindicatesthatthepricechangeson days follow-
ingtheinitiallargechangeare randomin sign,theinitiallargechangeat least
represents an unbiasedadjustment to theultimatepriceeffects of theinforma-
tion,and tlhisis sufficient
forthe expectedreturnefficient marketsmodel.
Niederhoffer and Osborne [32] documenttwo departuresfromcomplete
randomnessin commonstockprice changesfromtransactionto transaction.
First,theirdata indicatethat reversals(pairs of consecutiveprice changes
of oppositesign) are fromtwoto threetimesas likelyas continuations (pairs
of consecutiveprice changesof the same sign). Second, a continuationis
18 of the 30 stocksin Table 1 the actual numberof runs is less than the expected number.Indeed,
runs testsin generalshow no consistentevidence of dependencefor alnydifferencing intervallonger
than a day, whichseems especiallypertinentin light of the commentsin footnote14.
Capital Markets
Efficient 397
morefrequentaftera precedingcontinuation
slightly than aftera reversal.
Thatis,let (+I++) indicatetheoccurrenceof a positivepricechange,given
twopreceding positivechanges.Then the events(+?++) and (-I---)
areslightly
morefrequent
than(+1+-) or ( _|+).1B
Niederhoffer and Osborneofferexplanations forthesephenomenabased on
themarketstructure of theNew York Stock Exchange (N.Y.S.E.). In par-
thereare threemajortypesof ordersthatan investormightplace in
ticular,
a givenstock: (a) buy limit (buy at a specifiedprice or lower), (b) sell
limit(sell at a specifiedpriceor higher),and (c) buy or sell at market(at
thelowestsellingor highestbuyingprice of anotherinvestor).A book of
unexecuted limitordersin a givenstockis keptby thespecialistin thatstock
onthefloorof the exchange.Unexecutedsell limitordersare, of course,at
higher prices than unexecutedbuy limit orders.On both exchanges,the
smallestnon-zeropricechangeallowedis Y8 point.
Supposenowthatthereis morethanone unexecutedsell limitorderat the
lowestpriceof any such order.A transactionat thisprice (initiatedby an
orderto buy at market'7)can onlybe followedeitherby a transaction at the
sameprice(if thenextmarketorderis to buy) or by a transaction at a lower
price(if the nextmarketorderis to sell). Consecutiveprice increasescan
usuallyonly occurwhenconsecutivemarketordersto buy exhaustthe sell
limitordersat a givenprice.'8In short,the excessivetendencytowardre-
versalforconsecutivenon-zeropricechangescould resultfrombunchingof
unexecuted buy and sell limitorders.
The tendency fortheevents(+ ++) and (- -) to occurslightly more
frequently than(+?+-) and (-I-+) requiresa moreinvolvedexplanation
which we shall not attemptto reproducein fullhere.In brief,Niederhoffer
and Osbornecontendthat the higherfrequencyof (+|++) relativeto
(+I+-) arises froma tendencyforlimitorders"to be concentrated at in-
tegers(26, 43), halves (26X2, 43'2), quarters and odd eighthsin descending
orderof preference."'9 The frequencyof the event (+I++), whichusually
requiresthatsell limitordersbe exhaustedat at least twoconsecutivelyhigher
prices(the last of whichis relativelymore frequently at an odd eighth),
moreheavilyreflects the absenceof sell limitordersat odd eighthsthanthe
event(+?+-), whichusuallyimpliesthatsell limitordersat onlyone price
havebeen exhaustedand so moreor less reflectsthe averagebunchingof
limit
ordersat all eighths.
But thoughNiederhoffer and Osbornepresentconvincing evidenceof sta-
16.On a transaction to transactionbasis,positiveand negativepricechangesare aboutequally

Thus,underthe assumption
likely. thatpricechangesare random,any pair of non-zerochanges
shouldbe as likelyas any other,and likewisefortriplets of consecutive
non-zerochanges.
17.A buy limitorderfora priceequal to or greaterthanthe lowestavailablesell limitprice
an orderto buyat market,
iseffectively and is treatedas suchby thebroker.
18.The exception is whenthereis a gap of morethan IX betweenthe highestunexecuted buy
limitand the lowestunexecuted sell limitorder,so that marketorders(and new limitorders)
canbe crossedat intermediate prices.
19.Theirempiricaldocumentation for this claim is a few samplesof specialists'books for
selected
days,plusthe observation [34] thatactualtradingprices,at leastforvolatilehighpriced
seemto be concentrated
stocks, at integers,halves,quartersand odd eighths in descendingorder.
tisticallysignificantdeparturesfromindependencein price changes from
transaction to transaction,
and thoughtheiranalysisof theirfindingspresents
interestinginsightsintotheprocessof marketmakingon themajorexchanges,
the types of dependenceuncovereddo not implymarketinefficiency. The
best documented sourceof dependence,the tendencytowardexcessiverever-
sals in pairs of non-zeroprice changes,seems to be a directresultof the
abilityof investorsto place limitordersas well as ordersat market,and this
negativedependencein itselfdoes notimplytheexistenceof profitable trading
rules. Similarly,the apparenttendencyfor observedtransactions(and, by
implication, limitorders)to be concentrated at integers,
halves,even eighths
and odd eighthsin descendingorder is an interesting fact about investor
behavior,but in itselfis not a basis on whichto concludethatthe marketis
inefficient.20
The Niederhoffer-Osborne analysisof marketmakingdoes, however,point
clearlyto theexistenceof marketinefficiency, but withrespectto strongform
testsof the efficient marketsmodel.In particular,the list of unexecutedbuy
and sell limitordersin the specialist'sbook is important information about
the likelyfuturebehaviorof prices,and thisinformation is onlyavailable to
the specialist.When the specialistis asked fora quote, he gives the prices
and can give the quantitiesof the highestbuy limit and lowest sell limit
orderson his book,but he is preventedby law fromdivulgingthebook's full
contents.The interested readercan easilyimaginesituationswherethe struc-
tureof limitordersin the book could be used as the basis of a profitable
tradingrule.2'But therecordseemsto speak foritself:
It shouldnotbe assumedthatthesetransactions undertaken by thespecialist,
andin
whichhe is involvedas buyeror sellerin 24 per centof all marketvolume,are
necessarilya burdento him.Typically, thespecialist
sellsabovehislastpurchase on
83 per centof all his sales,and buysbelowhis last sale on 81 per centof all his
purchases ( [32], p. 908).
Thus it seemsthatthespecialisthas monopolypoweroveran importantblock
of information,and, not unexpectedly,
uses his monopolyto turna profit.
Andthis,ofcourse,is evidenceof marketinefficiency
in thestrongformsense.
The importanteconomicquestion,of course,is whetherthe marketmaking
20. Niederhofferand Osborneoffer littleto refutethisconclusion.For example([32], p. 914):
Although thespecificpropertiesreported in thisstudyhave a significance froma statistical point
of view,the readermay well ask whetheror not theyare helpfulin a practicalsense.Certain
tradingrulesemergeas a resultof ouranalysis.One is thatlimitand stop ordersshouldbe placed
at odd eights,preferablyat Y8 forsell ordersand at /8 forbuy orders.Another is to buywhena
stockadvancesthrough a barrierand to sell whenit sinksthrough a barrier.
The first"tradingrule"tellstheinvestor to resisthis innateinclination
to place ordersat integers,
but ratherto placesell ordersI/8below an integerand buy ordersI/8above.Successful execution
of the ordersis thenmorelikely,sincethe congestion of ordersthatoccurat integers is avoided.
But the costof thissuccessis apparent.The second"tradingrule"seemsno morepromising, if
indeedit can evenbe translated intoa concrete foraction.
prescription
21. See, forexample,([32], p. 908). But it is unlikelythatanyonebut thespecialistcouldearn
substantialprofitsfromknowledge of the structure of unexecutedlimitorderson the book. The
specialistmakestradingprofitsby engagingin manytransactions, each of whichhas a small
averageprofit;but forany othertrader,including thosewithseatson theexchange, theseprofits
wouldbe eatenup by commissions to thespecialist.
Capital Markets
Efficient 399
moreeconomicallyby some non-
functionof the specialistcould be fulfilled
monopolisticmechanism.22
4. DistributionalEvidence
At this date the weightof the empiricalevidenceis such that economists
wouldgenerallyagree thatwhateverdependenceexistsin seriesof historical
returnscannotbe used to make profitable predictionsof the future.Indeed,
forreturnsthatcoverperiodsof a day or longer,thereis littlein the evidence
thatwould cause rejectionof the strongerrandomwalk model,at least as a
goodfirstapproximation.
Rather,the last burningissue of the randomwalk literaturehas centered
on the natureof the distribution of price changes (which,we should note
immediately, is an important issue forthe efficientmarketshypothesissince
thenatureof thedistribution affectsboththetypesof statisticaltoolsrelevant
fortestingthe hypothesisand the interpretation of any resultsobtained).A
model implyingnormallydistributedprice changes was firstproposedby
Bachelier[3], who assumedthatprice changesfromtransactionto transac-
tion are independent,identicallydistributedrandomvariables with finite
variances.If transactions are fairlyuniformly spread across time,and if the
numberof transactions per day,week,or monthis verylarge,thentheCentral
LimitTheoremleads us to expectthat theseprice changeswill have normal
or Gaussian distributions.
Osborne [33], Moore [31], and Kendall [21] all thoughttheirempirical
evidencesupportedthenormality hypothesis,but all observedhightails (i.e.,
higherproportions of largeobservations)in theirdata distributions vis-a-vis
whatwouldbe expectedif the distributions were normal.Drawing these
on
findingsand someempiricalworkofhis own, Mandelbrot[28] thensuggested
thatthese departuresfromnormalitycould be explainedby a moregeneral
formof the Bacheliermodel.In particular,if one does not assume that dis-
tributionsof price changesfromtransactionto transactionnecessarilyhave
finitevariances,thenthe limitingdistributions forprice changesover longer
differencingintervalscouldbe any memberof thestable class, whichincludes
the normalas a special case. Non-normalstable distributions have higher
tailsthanthenormal,and so can accountforthisempirically observedfeature
of distributionsof price changes.Afterextensivetesting(involvingthe data
fromthe stocksin Table 1), Fama [10] concludesthat non-normalstable
distributionsare a betterdescription of distributionsof dailyreturnson com-
mon stocksthan the normal.This conclusionis also supportedby the em-
piricalworkof Blume [7] on commonstocks,and it has been extendedto
U.S. Government TreasuryBills by Roll [37].
Economistshave,however,been reluctantto accepttheseresults,2" primar-
22. With moderncomputers, it is hard to believethat a more competitiveand economical
systemwould not be feasible.It does not seemtechnologically to replacethe entire
impossible
floorof theN.Y.S.E. witha computer, fedby manyremoteconsoles, thatkeptall thebooksnow
keptby thespecialists,thatcouldeasilymaketheentirebook on any stockavailableto anybody
(so that interested
individualscould then competeto "make a market"in a stock) and that
carriedout transactionsautomatically.
23. Somehave suggestedthatthelong-tailed empirical
distributions
mightresultfromprocesses
ily because of the wealthof statisticaltechniquesavailable for dealingwith
normalvariablesand the relativepaucityof such techniquesfornon-normal
stable variables.But perhapsthe biggestcontribution of Mandelbrot'swork
has been to stimulateresearchon stable distributions and estimationpro-
ceduresto be appliedto stablevariables.(See, forexample,Wise [46], Fama
and Roll [15], and Blattbergand Sargent[6], amongothers.)The advance
of statisticalsophistication(and the importanceof examiningdistributional
assumptionsin testingthe efficientmarketsmodel) is well illustratedin Roll
[37], as compared,forexample,withtheearlyempiricalworkof Mandelbrot
[28] and Fama [10].
5. "Fair Game" Models in the TreasuryBill Market
Roll's workis novel in otherrespectsas well. Comingafterthe efficient
marketsmodelsof Mandelbrot[27] and Samuelson[38], it is the firstweak
formempiricalworkthat is consciouslyin the "fair game" ratherthan the
randomwalk tradition.
Moreimportant, as we saw earlier,the"fairgame"properties
of thegeneral
expectedreturnmodelsapplyto
zjt= rjt- E(fjtjDt_j). (10)
For data on commonstocks,testsof "fair game" (and randomwalk) pro-
pertiesseem to go well whenthe conditionalexpectedreturnis estimatedas
theaveragereturnforthesampleof data at hand.Apparently thevariationin
commonstockreturnsabout theirexpectedvalues is so large relativeto any
changesin the expectedvalues thatthe lattercan safelybe ignored.But, as
Roll demonstrates,thisresultdoes not hold forTreasuryBills. Thus, to test
the "fair game" model on TreasuryBills requiresexpliciteconomictheory
fortheevolutionof expectedreturnsthroughtime.
Roll uses threeexistingtheoriesof thetermstructure(thepureexpectations
hypothesisof Lutz [26] and two marketsegmentation hypotheses,one of
whichis the familiar"liquiditypreference"hypothesisof Hicks-[18] and
Kessel [22]) forthispurpose.24In his modelsrntis therateobservedfromthe
termstructure at periodt forone week loans to commenceat t + j - 1, and
can be thoughtof as a "futures"rate. Thus rj+i,t-i is likewisethe rate on
thatare mixtures of normaldistributionswithdifferent variances.
Press[35], forexample, suggests
a Poissonmixture of normalsin whichtheresulting of pricechangeshave longtails
distributions
but finitevariances.On the otherhand,Mandelbrot and Taylor[29] showthatothermixtures of
normalscan stilllead to non-normal stabledistributions of pricechangesforfinitedifferencing
intervals.
If, as Press'modelwouldimply,distributions of pricechangesare long-tailed but have finite
variances,thendistributions of pricechangesover longerand longerdifferencing intervalsshould
be progressively closerto the normal.No such convergence to normality was observedin [101
(thoughadmittedly the techniquesused weresomewhatrough). Rather,exceptfor originand
scale, the distributionsfor longerdifferencing intervalsseem to have the same "high-tailed"
characteristics forshorter
as distributins differencing whichis as wouldbe expectedif the
intervals,
distributionsare non-normal stable.
24. As notedearlyin our discussions, all availabletestsof marketefficiency are implicitly
also
testsof expectedreturn modelsof marketequilibrium. But Roll formulates the economic
explicitly
modelsunderlying his estimatesof expectedreturns, and emphasizesthat he is simultaneously
testingeconomicmodelsof the termstructure as well as marketefficiency.
Capital Markets
Efficient 401
one week loans to commenceat t + j -1, but observedin thiscase at t - 1.
Litis theso-called"liquiditypremium"in rjt;thatis
Similarly,
rjt E((ro,t+j_iIIt) + Ljt.
In words,the one-week"futures"rateforperiodt + j - 1 observedfromthe
termstructure at t of the "spot" ratefort + j -1 plus
at t is theexpectation
a "liquiditypremium"(whichcould,however,be positiveor negative).
In all threetheoriesof the termstructureconsideredby Roll, the condi-
tionalexpectationrequiredin (10) is of the form
- rj+?,tl +
E(r"j,tPt_1) E(LjtJI~t-L)- Lj+L,t-..
The threetheoriesdifferonly in the values assignedto the "liquiditypre-
miums."For example,in the"liquiditypreference" hypothesis,investorsmust
alwaysbe paid a positivepremiumforbearinginterestrate uncertainty, so
thatthe Lit are alwayspositive.By contrast,in the "pure expectations"hy-
pothesis,all liquiditypremiumsare assumedto be zero,so that
i( tJOt-:tL)- rj+L,t -L.
Afterextensivetesting, Roll concludes(i) thatthetwomarketsegmentation
hypothesesfitthe data betterthan the pure expectationshypothesis, with
perhapsa slightadvantageforthe "liquiditypreference" hypothesis,and (ii)
thatas faras his testsare concerned,themarketforTreasuryBills is effcient.
that when the best fittingtermstructuremodel is
Indeed, it is interesting
used to estimatetheconditionalexpected"futures"ratein (10), the resulting
variablezjt seemsto be seriallyindependent!It is also interestingthatif he
simplyassumedthathis data distributions were normal,Roll's resultswould
notbe so strongly marketsmodel.In thiscase taking
in supportof theefficient
accountof the observedhightails of the data distributions substantiallyaf-
fectedthe interpretation of the results.25
6. Tests of a MultipleSecurityExpected ReturnModel
Though the weak formtests supportthe "fair game" efficient markets
model,all of the evidenceexaminedso far consistsof what we mightcall
"single securitytests." That is, the price or returnhistoriesof individual
securitiesare examinedforevidenceof dependencethatmightbe used as the
basis of a tradingsystemforthat security.We have not discussedtestsof
whethersecuritiesare "appropriately priced"vis-a-visone another.
But to judgewhetherdifferences betweenaveragereturnsare "appropriate"
an economictheoryof equilibriumexpectedreturnsis required.At the mo-
ment,theonlyfullydevelopedtheoryis thatof Sharpe [40] and Lintner[24,
25. The importanceof distributionalassumptionsis also illustratedin Alexander'swork on trad-
ing rules. In his initial tests of filtersystems[1], Alexanderassumed that purchases could always
be executed exactly (rather than at least) y% above lows and sales exactly y% below highs.
Mandelbrot [281 pointed out, however, that though this assumptionwould do little harm with
normallydistributedprice changes (since price series are then essentiallycontinuous), with non-
normal stable distributionsit would introducesubstantialpositive bias into the filterprofits(since
with such distributionsprice series will show many discontinuities). In his later tests [2],
Alexanderdoes indeed find that taking account of the discontinuities(i.e., the presence of large
of the filters.
price changes) in his data substantiallylowers the profitability
402 The Journal
of Finance
25] referredto earlier.In this model (which is a directoutgrowthof the
mean-standard deviationportfoliomodels of investorequilibriumof Mar-
kowitz[30] and Tobin [43]), theexpectedreturnon securityj fromtimet to
[
t+ 1 is
E(f,t+1j1t) = rf,t+l + E(Fm,t+lfDt) -rf,t+l co]
C(j,to,y rm,t+4lDt)
(11)
whererf,t+1is thereturnfromt to t + 1 on an asset thatis risklessin money
terms; rm,t+1is the returnon the "marketportfolio"m (a portfolioof all
investment assetswitheach weightedin proportion to the totalmarketvalue
of all its outstanding
units); 02(rm,t+110t) is thevarianceof the returnon m;
cov (rij,t+i, rm,t+:Lit)is the covariance between the returnson j and m; and
the appearanceof lIt indicatesthat the various expectedreturns,variance
and covariance,could in principledependon 'Dt. ThoughSharpeand Lintner
derive(11) as a one-periodmodel,the resultis givena multiperiodjustifica-
tion and interpretation
in [11]. The modelhas also been extendedin (12)
to thecase wherethe one-periodreturnscould have stable distributions
with
infinitevariances.
In words,(11) says thattheexpectedone-periodreturnon a securityis the
one-period risklessrateof interestrf,t+1plus a "riskpremium"thatis propor-
tional to cov(rij,t+i, rm,t+ilDt)/6(rmt+11100.In the Sharpe-Lintnermodel
each investorholds some combinationof the risklessasset and the market
portfolio, so that,givena mean-standard deviationframework, the riskof an
individualasset can be measuredby its contribution to the standarddeviation
of the returnon the marketportfolio.This contributionis in fact cov
(rj,t+i, rm,t+l r(t)/I(imst+it) *26 The factor
[E(r-m,t+,ifDt)- rf,t+1]/0(rm,t+1j@I t),

whichis the same forall securities,is thenregardedas the marketprice of
risk.
Publishedempiricaltestsof theSharpe-Lintner modelare notyetavailable,
thoughmuchworkis in progress.There is some publishedwork,however,
which,thoughnot directedat the Sharpe-Lintner model,is at least consistent
withsomeof its implications. The statedgoal of thisworkhas been to deter-
minethe extentto whichthe returnson a givensecurityare relatedto the
returnson othersecurities.It started (again) with Kendall's [21] finding
thatthoughcommonstockpricechangesdo notseemto be seriallycorrelated,
thereis a high degreeof cross-correlationbetweenthe simultaneousreturns
of different
securities.This line of attackwas continuedby King [23] who
(usingfactoranalysisof a sampleof monthly returnson sixtyN.Y.S.E. stocks
fortheperiod1926-60) foundthaton averageabout 50% of thevarianceof
an individualstock's returnscould be accountedfor by a "marketfactor"
whichaffects thereturnson all stocks,with"industryfactors"accountingfor
at mostan additional10%'oof the variance.
26. That is,
coy (rjt+i rm Crm =
,t+ilt)/o ,t+iI1,t) (Yrm,t+iI,Dd.
CapitalMarkets
Efficient 403
For our purposes,however,the workof Fama, Fisher,Jensen,and Roll
[14] (henceforthFFJR) and the more extensivework of Blume [7] on
monthly returndata is morerelevant.Theytestthefollowing "marketmodel,"
originally suggestedby Markowitz[30]:
r,t+i = aj + ijrM,t+1+ ij,t+ (12)
wherera,t+1 is the rateof returnon securityj formontht, rm,t+i is the cor-
respondingreturnon a marketindexM, aj and ij are parametersthat can
vary fromsecurityto security,and uj,t+l is a randomdisturbance. The tests
of FFJR and subsequently thoseof Blumeindicatethat (12) is well specified
as a linearregressionmodel in that (i) the estimatedparametersaj and ij
remainfairlyconstantover longperiodsof time (e.g., the entirepost-World
War II periodin thecase of Blume), (ii) rM,t+1and theestimatedufj,t+,, are
close to seriallyindependent, and (iii) the uj,t+i seem to be independent of
rM,t+1.
Thus theobservedpropertiesof the"marketmodel"are consistent withthe
expectedreturnefficient marketsmodel,and, in addition,the "marketmodel"
tellsus something about theprocessgeneratingexpectedreturnsfromsecurity
to security.In particular,
E(r- t+c) = aj + PjE(riM,t+1). (13)
The questionnow is to what extent(13) is consistentwith the Sharpe-

Lintnerexpectedreturnmodel summarizedby (11). Rearranging(11) we
obtain
E(r-j,t+1J|t) aj((Dt)+ (3j((Dt)E(rim,t+i1|Dt), (14)

where,notingthatthe risklessrate rf,t+1is itselfpart of the information
set
t, we have
aj(@Dt) rf,t+l[ P-j (Dt)], (15)
and
Pj( D) =cov (r'jja,~rmt+11(Dt) ( 16)
Withsome simplifying assumptions,(14) can be reducedto (13). In partic-

ular, if the covarianceand variancethat determineWj(Ct) in (16) are the
same forall t and Dt,thenPjf(Dt)in (16) corresponds to Pj in (12) and (13),
and the least squares estimateof Pj in (12) is in fact just the ratio of the
samplevalues of the covarianceand variancein (16). If we also assumethat
rf,t+1is thesame forall t, and thatthebehaviorof thereturnson themarket
portfoliom are closelyapproximated by the returnson some representative
indexM, we willhave comea longway towardequating(13) and (11). In-
deed,theonlymissinglinkis whetherin the estimatedparametersof (12)
ajrf(I S) (17)
Neither FFJR nor Blume attack this question directly,thoughsome of
Blume's evidenceis at least promising.In particular,the magnitudesof the
estimated`j are roughlyconsistent with (17) in the sense thatthe estimates
are alwaysclose to zero (as theyshouldbe withmonthlyreturndata).27
In a sense, though,in establishingthe apparentempiricalvalidityof the
"marketmodel"of (12), both too muchand too littlehave been shownvis-
a-vis theSharpe-Lintner expectedreturnmodelof (11). We knowthatduring
thepost-World War II periodone-month interestrateson risklessassets (e.g.,
government bills withone monthto maturity)have not been constant.Thus,
if expectedsecurityreturnswere generatedby a versionof the "market
model"thatis fullyconsistent withthe Sharpe-Lintner model,we would,ac-
cordingto (15), expectto observesome non-stationarity in the estimatesof
aj. On a monthly basis, however,variationthroughtimein one-periodriskless
interestratesis probablytrivialrelativeto variationin otherfactorsaffecting
monthlycommonstock returns,so that more powerfulstatisticalmethods
wouldbe necessaryto studythe effects of changesin therisklessrate.
In any case, since the workof FFJR and Blume on the "marketmodel"
was not concernedwithrelatingthismodelto the Sharpe-Lintner model,we
can onlysay thattheresultsforthe formerare somewhatconsistent withthe
implications of the latter.But the resultsforthe "marketmodel" are, after
all, just a statisticaldescriptionof thereturngenerating process,and theyare
probablysomewhatconsistentwith other models of equilibriumexpected
returns.Thus theonlyway to generatestrongempiricalconclusionsabout the
Sharpe-Lintner modelis to testit directly.On theotherhand,any alternative
modelof equilibrium expectedreturnsmustbe somewhatconsistent withthe
"marketmodel,' giventhe evidencein its support.
B. Tests of MartingaleModels of the Semi-strong
Form
In general,semi-strong
formtestsof efficientmarketsmodelsare concerned
with whethercurrentprices "fullyreflect"all obviouslypubliclyavailable
information. Each individualtest,however,is concernedwiththe adjustment
of securityprices to one kind of information generatingevent (e.g., stock
splits,announcements of financialreportsby firms,
new securityissues,etc.).
Thus each testonlybringssupporting evidenceforthe model,withthe idea
thatby accumulating such evidencethe validityof the modelwillbe "estab-
lished."
In fact,however,thoughthe availableevidenceis in supportof theefficient
marketsmodel,it is limitedto a few major typesof information generating
events.The initialmajorworkis apparently thestudyof stocksplitsby Fama,
27. With least squares applied to monthlyreturndata, the estimateof (X in (12) is
aj = rj,t - jrm,t,
where the bars indicate sample mean returns.But, in fact, Blume applies the marketmodel to the
wealth relativesRjt = 1 + rjt and RMt = 1 + rmt.This yields preciselythe same estimateof ,1 as
least squares applied to (12), but the interceptis now
a'J=Rjt- 3jRMt = 1 + rJt-3j(1 + rMt) = 1- pj + aj

Thus what Blume in fact finds is that for almost all securities, j'j + 3j 1, which implies that
ctj is close to 0.
Capital Markets
Efficient 405
Fisher,Jensen,and Roll (FFJR) [14], and all the subsequentstudiessum-
marizedhere are adaptationsand extensionsof the techniquesdevelopedin
FFJR. Thus, thispaper will firstbe reviewedin some detail,and thenthe
otherstudieswill be considered.
1. Splits and the Adjustmentof Stock Prices to New Information
Since theonlyapparentresultof a stocksplitis to multiplythenumberof
sharesper shareholder withoutincreasing claimsto real assets,splitsin them-
selves are not necessarilysourcesof new information. The presumption of
FFJR is that splitsmay oftenbe associatedwith the appearanceof more
fundamentally important information.
The idea is to examinesecurityreturns
aroundsplitdates to see firstif thereis any "unusual"behavior,and, if so,
to what extentit can be accountedforby relationships betweensplitsand
othermorefundamental variables.
The approachof FFJR to theproblemreliesheavilyon the"marketmodel"
of (12). In thismodelif a stocksplitis associatedwithabnormalbehavior,
thiswouldbe reflected in the estimatedregressionresidualsforthe months
surrounding thesplit.For a givensplit,definemonth0 as themonthin which
theeffectivedate of a splitoccurs,month1 as themonthimmediately follow-
ingthe splitmonth,month-1 as the monthpreceding,etc. Now definethe
averageresidualoverall splitsecuritiesformonthm (whereforeach security
mis measuredrelativeto thesplitmonth)as
N
u
N'1
wherefUjmis thesampleregression
residualforsecurityj in monthm and N is
thenumberof splits.Next,definethecumulativeaverageresidualUm as
m
Um i Uk.
k=-29
The averageresidualum can be interpreted as the average deviation(in

monthm relativeto splitmonths)of the returnsof split stocksfromtheir
normal relationshipswiththemarket.Similarly, Um can be interpreted
as the
cumulative deviation(frommonth-29 tomonthm). Finally,defineu+, u;, U+
and Um as the averageand cumulativeaverageresidualsforsplitsfollowed
by"increased"(+) and "decreased"(-) dividends.An "increase"is a case
wherethe percentagechangein dividendson the splitsharein theyearafter
thesplitis greaterthanthe percentagechangeforthe N.Y.S.E. as a whole,
whilea "decrease"is a case ofrelativedividenddecline.
The essenceof theresultsof FFJR are thensummarized in Figure1, which
showsthe cumulativeaverage residualsUr U+ and U- for -29 ` m
30. The sampleincludesall 940 stocksplitson the N.Y.S.E. from1927-59,
wherethe exchangewas at least fivenew sharesforfourold, and wherethe
securitywas listedforat least twelvemonthsbeforeand afterthe split.
For all threedividendcategoriesthe cumulativeaverageresidualsrise in
the29 monthspriorto thesplit,and in factthe averageresiduals(not shown
here) are uniformly
positive.This cannotbe attributed
to thesplitting
process,
sincein onlyabouttenper centof thecases is thetimebetweentheannounce-
mentand effectivedates of a splitgreaterthan fourmonths.Rather,it seems
thatfirmstendto splittheirsharesduring"abnormally"good times-thatis,
duringperiodswhenthepricesof theirshareshave increasedmorethanwould
U
m
o.44 , , , , , ' ' ' ' '
0.33 -
0.22
0.11
o t
-29 25-20_15-10_50 5 10 15 20 25 30
Monthrelative to split--m
FIGURE la
Cumulative average residuals-all splits.
u + u-
m m
10.44 , , , , ,T
o.~~~~~~~~~
44
0.33 ~0.33
0.22 .
0.22
- . .
0.11 0.11 _.
o:i
0 ~-29 --20 -loO0 1 0 3
-29 25-2 15-10 _50 5 10 15 20 25 30 -25 -15 5 5 1015 20523
Monthrelative to split--m Monthrelative to Split--m
FIGURE lb FiGuRE lc
Cumulative average residuals for dividend Cumulativeaverage residualsfor dividend
"increases." "decreases."
CapitalMarkets
Efficient 407
be impliedby theirnormalrelationships with generalmarketprices,which
itselfprobablyreflectsa sharp improvement, relativeto the market,in the
earningsprospectsof thesefirmssometimeduringtheyearsimmediately pre-
cedinga split.28
Afterthesplitmonththereis almostno further movement in Un, thecumu-
lativeaverageresidualforall splits.This is striking, since 71.5 per cent (672
out of 940) of all splitsexperienced greaterpercentagedividendincreasesin
theyear afterthe splitthanthe averageforall securitieson the N.Y.S.E. In
lightof this,FFJR suggestthatwhena splitis announcedthe marketinter-
prets this (and correctlyso) as a signal thatthe company'sdirectors are
probablyconfident thatfutureearningswillbe sufficient to maintaindividend
paymentsat a higherlevel. Thus the largepriceincreasesin the monthsim-
mediatelyprecedinga splitmay be due to an alterationin expectations con-
cerningthe futureearningpotentialof thefirm,ratherthan to any intrinsic
effectsof thesplititself.
If thishypothesis is correct,returnbehaviorsubsequentto splitsshouldbe
substantiallydifferent forthe cases wherethe dividendincreasematerializes
thanforthe cases whereit does not. FFJR arguethatin factthe differences
are in the directionsthat would be predicted.The fact that the cumulative
averageresidualsforthe "increased"dividends(Figure lb) driftupwardbut
onlyslightlyin the year afterthesplitis consistentwiththe hypothesisthat
whenthesplitis declared,thereis a priceadjustment in anticipation of future
dividendincreases.But thebehaviorof theresidualsforstocksplitsassociated
with "decreased" dividendsofferseven strongerevidencefor the splithy-
pothesis.The cumulativeaverageresidualsforthesestocks(Figure lc) risein
the fewmonthsbeforethe split,but thenfall dramatically in the fewmonths
afterthe split when the anticipateddividendincrease is not forthcoming.
When a year has passed afterthe split,the cumulativeaverageresidualhas
fallento aboutwhereit was fivemonthspriorto thesplit,whichis about the
earliesttimereliableinformation about a splitis likelyto reachthe market.
Thus by the timeit becomesclear that the anticipateddividendincreaseis
not forthcoming, the apparenteffectsof the split seem to have been wiped
away,and the stock'sreturnshave revertedto theirnormalrelationship with
marketreturns.
Finally,and mostimportant, althoughthebehaviorof post-splitreturnswill
be verydifferent dependingon whetheror notdividend"increases"occur,and
in spite of the fact that a large majorityof split securitiesdo experience
dividend"increases,"when all splits are examinedtogether(Figure la),
subsequentto thesplitthereis no netmovement up or downin thecumulative
28. It is importantto note, however,that as FFJR indicate,the persistentupward driftof the
cumulativeaverage residualsin the monthsprecedingthe split is not a phenomenonthat could be
used to increaseexpectedtradingprofits.The reason is that the behavior of the average residuals
is not representativeof the behavior of the residuals for individual securities.In months prior to
the split, successivesample residualsfor individual securitiesseem to be independent.But in most
cases, there are a few months in which the residuals are abnormally large and positive. The
monthsof large residualsdifferfromsecurityto security,however,and these differences in timing
explain why the signs of the average residualsare uniformlypositive for many months preceding
the split.
averageresiduals.Thus, apparentlythe marketmakesunbiasedforecasts of
the implicationsof a split forfuturedividends,and theseforecastsare fully
in thepricesof thesecurityby theend of thesplitmonth.Aftercon-
reflected
siderablymoredata analysisthancan be summarized here,FFJR conclude
thattheirresultslend considerablesupportto the conclusionthatthestock
marketis efficient,at leastwithrespectto its abiliyto adjust to theinforma-
tionimplicitin a split.
2. OtherStudiesof PublicAnnouncements
Variantsof the methodof residualanalysisdevelopedin [14] have been
used by othersto studytheeffects of differentkindsof publicannouncements,
and all of thesealso supportthe efficientmarketshypothesis.
Thus usingdata on 261 majorfirmsfortheperiod1946-66,Ball and Brown
[4] applythemethodto studytheeffects of annualearningsannouncements.
They use theresidualsfroma timeseriesregression of theannualearnings of
a firmon theaverageearnings of all theirfirmsto classifythe firm'searnings
fora givenyearas having"increased"or "decreased"relativeto themarket.
Residualsfromregressions of monthlycommonstockreturnson an indexof
returns(i.e., themarket model of (12)) are thenused to computecumulative
average returnresidualsseparatelyfor the earningsthat "increased,"and
thosethat"decreased."The cumulativeaveragereturnresidualsrisethrough-
out the year in advance of the announcement for the earnings"increased"
category,and fall forthe earnings"decreased"category.29 Ball and Brown
[4, p. 175] concludethatin factno morethanabout tento fifteen percent of
theinformation in theannualearningsannouncement has notbeenanticipated
by themonthof the announcement.
On themacrolevel,Waud [45] has used themethodof residualanalysisto
examinethe effectsof announcements of discountrate changesby Federal
ReserveBanks. In thiscase the residualsare essentiallyjust the deviations
of the dailyreturnson the Standardand Poor's 500 Index fromthe average
daily return.He findsevidenceof a statistically "announcement
significant
effect"on stockreturnsforthe firsttradingday following an announcement,
but the magnitudeof the adjustmentis small,neverexceeding.5%. More
interesting fromthe viewpointof the efficient marketshypothesis is his con-
clusionthat,if anything, the marketanticipatesthe announcements (or in-
formation is somehowleaked in advance). This conclusionis based on the
non-random patternsof the signs of average returnresidualson the days
immediately precedingthe announcement.
Furtherevidencein supportof the efficient marketshypothesisis pro-
vided in the workof Scholes [39] on large secondaryofferings of common
stock (ie., largeunderwritten sales of existingcommonstocksby individuals
and institutions) and on newissuesof stock.He findsthaton averagesecon-
daryissuesare associatedwitha declineof betweenone and twoper centin
thecumulativeaverageresidualreturnsforthecorresponding commonstocks.
Since the magnitudeof the price adjustmentis unrelatedto the size ofthe
28 is againrelevanthere.
of footnote
29. But thecomment
CapitalMarkets
Efficient 409
issue,Scholesconcludesthatthe adjustmentis not due to "sellingpressure"
(as is commonly believed),but ratherresultsfromnegativeinformation im-
in
plicit thefactthatsomebody trying is to sell a largeblock of a firm'sstock.
Moreover,he presentsevidencethatthe value of the information in a secon-
darydependsto someextenton the vendor;somewhat as would be expected,
by far the largestnegativecumulativeaverage residualsoccur where the
vendoris the corporationitselfor one of its officers, withinvestment com-
paniesa distantsecond.But theidentity of the vendoris notgenerallyknown
at the timeof the secondary,and corporateinsidersneed only reporttheir
transactions in theirowncompany'sstockto theS.E.C. withinsix days aftera
sale. By thistimethemarketon averagehas fullyadjustedto theinformation
in the secondary,as indicatedby the fact thatthe averageresidualsbehave
randomlythereafter.
Note, however,thatthoughthisis evidencethatpricesadjust efficiently to
public information, it is also evidencethatcorporateinsidersat least some-
timeshave importantinformation about theirfirmthat is not yet publicly
known.Thus Scholes' evidenceforsecondarydistributions providessupport
forthe efficient marketsmodelin the semi-strong formsense,but also some
strong-form evidenceagainstthe model.
Thoughhis resultshere are onlypreliminary, Scholes also reportson an
applicationof the methodof residualanalysisto a sampleof 696 new issues
of commonstockduringtheperiod1926-66.As in the FFJR studyof splits,
thecumulativeaverageresidualsrisein themonthsprecedingthenewsecurity
offering(suggestingthat new issues tend to come after favorablerecent
events)30but behaverandomlyin the monthsfollowing the offering (indicat-
ingthatwhateverinformation is containedin thenewissueis on averagefully
reflected in thepriceof themonthof the offering).
In short,the available semi-strong formevidenceon the effectof various
sortsof publicannouncements on commonstockreturnsis all consistent with
the efficient marketsmodel.The strongpointof the evidence,however,is its
consistencyratherthan its quantity;in fact, few different typesof public
information have been examined,thoughthose treatedare among the ob-
viouslymostimportant. Moreover,as we shall now see, the amountof semi-
strong form evidence is voluminouscomparedto the strongformtests that
are available.
C. StrongForm Tests of the Efficient MarketsModels

The strongformtests of the efficient marketsmodel are concernedwith
whetherall available informationis fullyreflected
in pricesin the sense that
no individualhas higherexpectedtradingprofitsthanothersbecause he has
monopolistic access to someinformation.We wouldnot,of course,expectthis
modelto be an exact descriptionof reality,and indeed,theprecedingdiscus-
sions have alreadyindicatedthe existenceof contradictory evidence.In par-
ticular,Niederhoffer and Osborne [32] have pointedout thatspecialistson
theN.Y.S.E. apparentlyuse theirmonopolistic access to informationconcern-
30. Footnote 28 is again relevanthere.
ing unfilledlimitordersto generatemonopolyprofits,and Scholes' evidence
[39] indicatesthatofficersof corporationssometimes have monopolisticaccess
to information abouttheirfirms.
Sincewe alreadyhave enoughevidenceto determine thatthe modelis not
strictlyvalid,we can nowturnto otherinteresting how
questions.Specifically,
far down throughthe investment community do deviationsfromthe model
permeate?Does it pay forthe averageinvestor(or the averageeconomist)
to expendresourcessearchingout littleknowninformation? Are such ac-
tivitiesevengenerally profitableforvariousgroupsof market"professionals"?
More generally,who are the people in the investment community that have
access to "special information"?
Thoughthis is a fascinating problem,only one grouphas been studiedin
any depth-the managements of open end mutualfunds.Severalstudiesare
available (e.g., Sharpe [41, 42] and Treynor[44]), but the most thorough
are Jensen's[19, 20], and our comments willbe limitedto his work.We shall
firstpresentthe theoretical modelunderlying his tests,and thengo on to his
empiricalresults.
1. TheoreticalFramework
In studyingtheperformance of mutualfundsthe majorgoals are to deter-
mine (a) whetherin generalfundmanagersseem to have access to special
information whichallowsthemto generate"abnormal"expectedreturns, and
(b) whethersome fundsare betterat uncoveringsuch special information
thanothers.Since thecriterion will simplybe the abilityof fundsto produce
higherreturnsthan some normwith no attemptto determinewhat is re-
sponsibleforthe high returns,the "special information" that leads to high
performance could be eitherkeenerinsightinto the implications of publicly
available informationthanis implicitin marketpricesor monopolistic access
to specificinformation.Thus the testsof theperformance of themutualfund
industryare not strictlystrongformtestsof the efficient marketsmodel.
The major theoretical(and practical) problemin usingthe mutualfund
industryto test the efficientmarketsmodel is developinga "norm"against
whichperformance can be judged.The normmustrepresent the resultsof an
investment policybased on the assumptionthatpricesfullyreflectall avail-
able information. And if one believesthatinvestorsare generallyriskaverse
and so on averagemustbe compensatedforany risksundertaken, thenone
has theproblemof finding appropriatedefinitionsof riskand evaluatingeach
fundrelativeto a normwithits chosenlevelof risk.
Jensenuses the Sharpe [40]-Lintner[24, 25] model of equilibriumex-
pectedreturnsdiscussedabove to derivea normconsistentwiththesegoals.
From (14)-(16), in thismodeltheexpectedreturnon an asset or portfolioj
fromt to t + 1 is
E(r'j,t?l 10t) rf,t+l [1 - (Dt))] + E (rm,t+1ikt)Pj(3t), (18)
wherethevarioussymbolsare as definedin SectionIII. A. 6. But (18) is an
ex ante relationship,
and to evaluateperformance
an ex post normis needed.
CapitalMarkets
Efficient 411
One way the lattercan be obtainedis to substitutethe realizedreturnon the
marketportfolioforthe expectedreturnin (18) withthe result3'
E(rij,t+,(Dt, rm,t+,) rf,t+l [1 - 3j((Dt)] + rm,t+,(j(QDt). (9)
Geometrically,(19) says that withinthe contextof the Sharpe-Lintner
model,theexpectedreturnon j (giveninformation (Dtand thereturnrm,t?lon
of its risk
themarketportfolio)is a linearfunction
p(34Q>) - COV (irj,t+1, rm
as indicatedin Figure2. Assumingthatthevalue of (3j(Dt) is somehowknown,
or can be reliablyestimated,if j is a mutualfund,its ex post performance
fromt to t + 1 mightnowbe evaluatedby plottingits combination of realized
returnrj,t+land riskin Figure2. If (as forthepointa) thecombination falls
above theexpectedreturnline (or, as it is morecommonly called,the "market
line"), it has donebetterthanwouldbe expectedgivenits level of risk,while
if (as forthepointb) it fallsbelowtheline it has doneworse.
r
E(ib,t?1lit,rn ,t+1)- ------------------------------------
b,t+- - --------------
m,t+IL
%+1 t rm,t+1-
a,tR
o- - - .
? Paia( t) PM(1t PO t) 04t)
FIGURE 2
Performance
EvaluationGraph
the marketline shows the combinations
Alternatively, of returnand risk
providedby portfoliosthat are simplemixturesof the risklessasset and the
marketportfoliom. The returnsand risksforsuch portfolios(call themc)
are
r =,t+l arf,t+l + (1 -0rM,t+1
(z1)) cv
0 (" ,t+1, rm,
(rcst t
"m,t+1! t) co
l|a
cov ( ?( )mtat
(1a
) r'm, + l, rm, )t
(D)=coy mt+1IlIPt+l| a
(3~QD~)
O2(irm,t+ 11t) -adrm,t+ij(Dt)1-
accordingto
hereis thatthe returnr; t--l is generated
31. The assumption
rj,t+l = ri,t+,[l - j(Dt)] + rm,t+j0j((Dt) + Uj,t+ls

and
t+IIrm,t+,)= 0 forall rm,t+.
E(UJ,
wherea is the proportionof portfoliofundsinvestedin the risklessasset.
Thus,when1 > a > 0 we obtainthecombinations of returnand riskalongthe
marketline fromrf,t+?to m in Figure 2, whilewhena < 0 (and underthe
assumptionthat investorscan borrowat the same rate that theylend) we
obtainthe combinationsof returnand risk along the extensionof the line
throughm. In this interpretation,
the marketline representsthe resultsof
a naive investment
strategy,whichthe investorwho thinkspricesreflectall
availableinformation
mightfollow.The performance of a mutualfundis then
measuredrelativeto thisnaive strategy.
2. EmpiricalResults
Jensenuses thisrisk-return framework to evaluatetheperformance of 115
mutual fundsover the ten year period 1955-64. He argues at lengthfor
measuringreturnas the nominalten year rate withcontinuouscompounding
(i.e., the naturallog of the ratioof terminalwealthafterten yearsto initial
wealth) and for using historicaldata on nominalone-yearrates with con-
tinuouscompounding to estimaterisk.The Standardand Poor Index of 500
majorcommonstocksis used as theproxyforthemarketportfolio.
The generalquestionto be answeredis whethermutualfundmanagements
have any special insightsor information whichallows themto earn returns
above thenorm.But Jensenattacksthequestionon severallevels.First,can
the fundsin generaldo well enough to compensateinvestorsfor loading
charges,management fees,and othercosts thatmightbe avoided by simply
choosingthe combination of the risklessasset f and the marketportfoliom
withrisklevel comparableto thatof the fund'sactual portfolio?The answer
seemsto be an emphaticno. As faras net returnsto investorsare concerned,
in 89 out of 115 cases, the fund'srisk-return combinationfor the ten year
periodis belowthemarketline fortheperiod,and theaverageoverall funds
of the deviationsof tenyear returnsfromthemarkettimeis -14.6%o. That
is, on averagethe consumer'swealthafterten yearsof holdingmutualfunds
is aboutfifteen per centless thanifhe held thecorresponding portfolios
along
the marketline.
But theloadingchargethatan investor pays in buyingintoa fundis usually
a pure salesman'scommissionthatthe funditselfnevergets to invest.Thus
one mightask whether,ignoringloading charges (i.e., assumingno such
chargeswere paid by the investor),in generalfundmanagements can earn
returnssufficiently above the normto coverall otherexpensesthatare pre-
sumablymore directlyrelated to the managementof the fund portfolios.
Again,the answerseemsto be no. Even whenloadingchargesare ignoredin
computingreturns,the risk-return combinations for 72 out of 115 fundsare
belowthemarketline,and theaveragedeviationof tenyear returnsfromthe
marketline is -8.9%.
Finally,as a somewhatstrongertest of the efficient marketsmodel,one
would like to knowif, ignoringall expenses,fundmanagements in general
showedany abilityto pick securitiesthat outperformed the norm.Unfortu-
nately,this questioncannotbe answeredwithprecisionforindividualfunds
since,curiously,data on brokeragecommissions are not publishedregularly.
Capital Markets
Efficient 413
But Jensensuggeststhe available evidenceindicatesthat the answerto the
questionis again probablynegative.Specifically, addingback all otherpub-
lishedexpensesof fundsto theirreturns,the risk-return combinations for58
out of 115 fundswerebelowthemarketline,and the averagedeviationof ten
year returnfromthe line was -2.5%o. But part of thisresultis due to the
absence of a correctionfor brokeragecommissions.Estimatingthese com-
missionsfromaverageportfolioturnoverrates forall fundsfor the period
1953-58,and addingthemback to returnsforall fundsincreasesthe average
deviationfromthe marketline from-2.5%o to .09%o,whichstill is not in-
dicativeof the existenceof specialinformation amongmutualfundmanagers.
But thoughmutualfundmanagersin generaldo notseemto have access to
information not alreadyfullyreflected in prices,perhapsthereare individual
fundsthatconsistently do betterthanthenorm,and so provideat least some
strongformevidenceagainstthe efficient marketsmodel.If thereare such
funds,however,they escape Jensen'ssearch. For example,for individual
funds,returnsabove thenormin one subperioddo not seem to be associated
withperformance above thenormin othersubperiods.And regardlessof how
returnsare measured(i.e., netor grossofloadingchargesand otherexpenses),
thenumberof fundswithlargepositivedeviationsof returnsfromthemarket
line of Figure 2 is less than the numberthatwould be expectedby chance
with115 fundsundertheassumptionthatfundmanagements have no special
talentsin predicting returtis.32
Jensenargues that thoughhis resultsapply to only one segmentof the
investment community, theyare nevertheless striking evidencein favorof the
efficientmarketsmodel:
Although theseresultscertainly do notimplythatthestrong formofthemartingale
hypothesisholdsforall investors and forall time,theyprovidestrongevidencein
support of thathypothesis.One mustrealizethattheseanalystsare extremely well
endowed. Moreover, theyoperatein thesecurities markets everydayandhavewide-
ranging contactsand associations in boththe businessand financial communities.
Thus,thefactthattheyareapparently unableto forecast returnsaccurately enough
to recovertheirresearchandtransactions costsis a strikingpieceofevidence in favor
of the strongformof themartingale hypothesis-at leastas faras the extensive
subsetofinformation availableto theseanalystsis concerned [20, p. 170].
IV. SUMMARY AND CONCLUSIONS
The preceding(ratherlengthy)analysiscan be summarizedas follows.In
marketsis concernedwithwhetherprices
generalterms,thetheoryof efficient
The theoryonlyhas
at anypointin time"fullyreflect"availableinformation.
empiricalcontent,however,withinthe contextof a more specificmodel of
32. On the other hand, thereis some suggestionin Scholes' [39] work on secondaryissues that
mutual fundsmay occassionallyhave access to "special information."Aftercorporateinsiders,the
next largest negative price changes occur when the secondary seller is an investmentcompany
(includingmutual funds), though on average the price changes are much smaller (i.e., closer to 0)
than when the seller is a corporateinsider.
Moreover, Jensen'sevidence itself,though not indicative of the existenceof special information
among mutual fund managers,is not sufficiently precise to conclude that such informationnever
exists. This strongerconclusion would require exact data on unavoidable expenses (including
brokeragecommissions) of portfoliomanagementincurredby funds.
414 of Ftnance
The Journal
marketequilibrium,that is, a model that specifiesthe nature of market
equilibriumwhenprices "fullyreflect"available information. We have seen
thatall of theavailableempiricalliterature is implicitly based on
or explicitly
the assumptionthat the conditionsof marketequilibriumcan be stated in
termsof expectedreturns.This assumptionis thebasis of theexpectedreturn
or "fairgame" efficient marketsmodels.
The empiricalwork itselfcan be dividedinto threecategoriesdepending
on thenatureof theinformation subsetof interest.Strong-form testsare con-
cernedwithwhetherindividualinvestorsor groupshave monopolistic access
to any information relevantforprice formation. One would not expectsuch
an extrememodelto be an exact description of the world,and it is probably
best viewedas a benchmark againstwhichthe importance of deviationsfrom
marketefficiency can be judged.In the less restrictive semi-strong-form tests
the information subset of interestincludesall obviouslypubliclyavailable
information, while in the weak formtests the information subset is just
historicalpriceor returnsequences.
Weak formtestsof the efficient marketmodel are the most voluminous,
and it seems fair to say that the resultsare stronglyin support.Though
statisticallysignificantevidencefor dependencein successiveprice changes
or returnshas been found,some of this is consistentwiththe "fair game"
modeland the restdoes not appear to be sufficient to declarethe marketin-
Indeed,at leastforpricechangesor returnscoveringa day or longer,
efficient.
thereisn'tmuchevidenceagainstthe"fairgame" model'smoreambitiousoff-
spring,the randomwalk.
Thus, thereis consistentevidenceof positive dependencein day-to-day
price changesand returnson commonstocks,and the dependenceis of a
formthatcan be used as the basis of marginally profitabletradingrules. In
Fama's data [10] the dependenceshows up as serial correlationsthat are
consistently positivebutalso consistentlycloseto zero,and as a slighttendency
forobservednumbersof runsofpositiveand negativepricechangesto be less
thanthenumbersthatwouldbe expectedfroma purelyrandomprocess.More
important, thedependencealso showsup in thefiltertestsof Alexander[1, 2]
and thoseof Fama and Blume [13] as a tendencyforvery small filtersto
produceprofitsin excessof buy-and-hold. But any systems(like the filters)
that attemptto turnshort-term dependenceinto tradingprofitsof necessity
generateso manytransactions thattheirexpectedprofitswould be absorbed
by even theminimum commissions (securityhandlingfees) thatfloortraders
on majorexchangesmustpay. Thus, usinga less thancompletely strictinter-
pretationof marketefficiency, this positive dependencedoes not seem of
sufficientimportance to warrantrejectionof the efficient marketsmodel.
Evidence in contradiction of the "fair game" efficient marketsmodel for
price changesor returnscoveringperiodslongerthan a single day is more
difficultto find.Cootner[9], and Moore [31] reportpreponderantly negative
(but againsmall) serialcorrelations in weeklycommonstockreturns, and this
resultappears also in the fourday returnsanalyzedby Fama [10]. But it
does notappear in runstestsof [10], where,if anything, thereis some slight
indicationof positivedependence,but actually not much evidenceof any
Capital Markets
Efficient 415
dependenceat all. In any case, thereis no indicationthatwhateverdependence
existsin weeklyreturnscan be used as thebasis of profitable tradingrules.
Otherexistingevidenceof dependencein returnsprovidesinteresting in-
sightsinto the processof price formation in the stockmarket,but it is not
relevantfor testingthe efficient marketsmodel. For example,Fama [10]
showsthatlargedailypricechangestendto be folowedby largechanges,but
of unpredictablesign. This suggeststhat importantinformation cannotbe
completelyevaluatedimmediately, but that the initialfirstday's adjustment
of pricesto theinformation is unbiased,whichis sufficient forthe martingale
model.More interesting and important, however,is the Niederhoffer-Osborne
[32] findingof a tendencytowardexcessivereversalsin commonstockprice
changesfromtransaction to transaction. They explainthisas a logical result
of the mechanismwherebyordersto buy and sell at marketare matched
against existinglimitorderson the books of the specialist.Given the way
this tendencytowardexcessivereversalsarises,however,thereseems to be
no way it can be used as thebasis of a profitable tradingrule.As theyrightly
claim,theirresultsare a strongrefutation of the theoryof randomwalks,at
least as appliedto pricechangesfromtransaction to transaction, but theydo
not constituterefutation of the economically morerelevant"fair game" effi-
cientmarketsmodel.
Semi-strong formtests,in whichprices are assumed to fullyreflectall
obviouslypubliclyavailable information, have also supportedthe efficient
marketshypothesis. Thus Fama, Fisher,Jensen,and Roll [14] findthatthe
information in stocksplitsconcerning the firm'sfuturedividendpaymentsis
on averagefullyreflected in thepriceof a splitshareat the timeof the split.
Ball and Brown[4] and Scholes[39] cometo similarconclusions withrespect
to the information containedin (i) annual earningannouncements by firms
and (ii) newissuesand largeblocksecondaryissuesof commonstock.Though
onlya few different typesof information generatingeventsare represented
here,theyare amongthe moreimportant, and the resultsare probablyin-
dicativeof whatcan be expectedin futurestudies.
As notedearlier,thestrong-form marketsmodel,in whichpricesare
efficient
assumedto fullyreflectall availableinformation, is probablybest viewedas
a benchmarkagainstwhichdeviationsfrommarketefficiency (interpretedin
its strictestsense) can be judged.Two such deviationshave in factbeen ob-
served.First,Niederhoffer and Osborne [32] point out that specialistson
major securityexchangeshave monopolistic access to information on unex-
ecutedlimitordersand theyuse thisinformation to generatetradingprofits.
This raises the questionof whetherthe "marketmaking"functionof the
specialist (if indeed this is a meaningfuleconomicfunction)could not as
effectivelybe carriedout by some other-mechanism that did not implymon-
opolisticaccess to information. Second, Scholes [39] findsthat,not unex-
pectedly,corporateinsidersoftenhave monopolisticaccess to information
about theirfirms.
At themoment, however,corporateinsidersand specialistsare theonlytwo
groups whose monopolistic access to information has been documented. There
is no evidence that deviations from the strong form of the efficientmarkets
416 of Finance
The Journal
modelpermeatedownany further through theinvestment community. For the
purposesof mostinvestorstheefficient marketsmodelseemsa good first(and
second) approximation to reality.
In short,theevidencein supportof theefficient marketsmodelis extensive,
and (somewhatuniquely in economics) contradictory evidence is sparse.
Nevertheless,we certainlydo notwantto leave the impression thatall issues
are closed.The old saw, "muchremainsto be done,"is relevanthereas else-
where.Indeed,as is oftenthe case in successfulscientificresearch,now that
we knowwe'vebeenin thepast,we are able to pose and (hopefully)to answer
an evenmoreinteresting set of questionsforthe future.In thiscase themost
pressingfieldof futureendeavoris the development and testingof modelsof
marketequilibriumunderuncertainty. When the processgenerating equilib-
riumexpectedreturnsis betterunderstood(and assumingthatsomeexpected
returnmodelturnsout to be relevant),we willhave a moresubstantialframe-
workformoresophisticated intersecuritytestsof marketefficiency.
REFERENCES
1. Sidney S. Alexander."Price Movements in Speculative Markets: Trends or Random Walks."
IndustrialManagement Review,2 (May 1961), 7-26. Also reprintedin [8], 199-218.
2. . "Price Movements in Speculative Markets: Trends or Random Walks. No. 2,"
in [8], 338-72.
3. Louis Bachelier. TheIoriede la Speculation (Paris: Gauthier-Villars,1900), and reprintedin
English in [8], 17-78.
4. Ray Ball and Phillip Brown. "An Empirical Evaluation of Accounting Income Numbers."
Journalof Accounting Research,6 (Autumn, 1968), 159-78.
5. William Beaver. "The InformationContent of Annual Earnings Announcements."Empirical
Researchin Accounting:SelectedStudies,1968,supplement to Vol. 7 of the Journalof
Accounting Research,67-92.
6. Robert Blattberg and Thomas Sargent. "Regression with Non-Gaussian Disturbances: Some
Sampling Results," forthcomingin Econometrica.
7. Marshall Blume. "The Assessment of Portfolio Performance." Unpublished Ph.D. thesis,
Universityof Chicago, 1968. A paper summarizingmuch of this work will appear in the
April, 1970,Journalof Business.
8. Paul Cootner (ed.). The Random Characterof Stock MarketPrices. Cambridge: M.I.T.,
1964.
9. . "Stock Prices: Random vs. Systematic Changes." Industrial Management
Review,3 (Spring 1962), 24-45. Also reprintedin [8], 231-52.
10. Eugene F. Fama. "The Behavior of Stock Market Prices." Journal of Business, 38 (January,
1965), 34-105.
11. . "Multiperiod Consumption-Investment Decisions." AmericanEconomic Review,
(March, 1970).
12. . "Risk, Return and Equilibrium." Report No. 6831, University of Chicago,
Center for Math. Studies in Business and Economics,June, 1968.
13. and Marshall Blume. "Filter Rules and Stock Market Trading Profits."Journal
of Business,39 (Special Supplement,January,1966), 226-41.
14. , Lawrence Fisher,Michael Jensenand Richard Roll. "The Adjustmentof Stodc
Prices to New Information."InternationalEconomic Review, X (February, 1969), 1-21.
15. and Richard Roll. "Some Propertiesof SymmetricStable Distributions."Journal
of the American Statistical
Association,63 (September, 1968), 817-36.
16. Michael D. Godfrey,C. W. J. Grangerand 0. Morgenstern."The Random Walk Hypothesis
of Stock Market Behavior." Kyklos, 17 (1964), 1-30.
17. C. W. J. Grangerand 0. Morgenstern."Spectral Analysisof New York Stock Market Prices,"
Kyklos,16 (1963), 1-27. Also reprintedin [8], 162-88.
18. John R. Hicks. Value and Capital. Oxford: The Clarendon Press, 1946.
19. Michael Jensen. "The Performanceof Mutual Funds in the Period 1945-64," Journal of
Finance,23 (May, 1968), 389-416.
Capital Markets
Efficient 417
20. . "Risk, the Pricing of Capital Assets, and the Evaluation of InvestmentPort-
folios,"Journal of Business,42 (April, 1969), 167-247.
21. Maurice G. Kendall. "The Analysis of Economic Time-Series,Part I: Prices," Journal of the
Society,96 (Part I, 1953), 11-25.
Royal Statistical
22. Ruben A. Kessel. "The Cyclical Behavior of the Term Structureof InterestRates," National
Bureau of Economic Research Occasional Paper No. 91. New York: Columbia University
Press, 1965.
23. Benjamin F. King. "Market and Industry Factors in Stock Price Behavior," Journal of
Business, 39 (Special SupplementJanuary, 1966), 139-90.
24. John Lintner. "Security Prices, Risk, and Maximal Gains from Diversification,"Journal of
Finance, 20 (December, 1965), 587-615.
25. . "The Valuation of Risk Assets and the Selection of Risky Investmentsin Stock
Portfoliosand Capital Budgets," Review of Economics and Statistics,47 (February, 1965),
13-37.
26. Fredrich A. Lutz. "The Structure of Interest Rates," Quarterly Journal of Economics, 40
(1940-41).
27. Benoit Mandelbrot. "Forecasts of Future Prices, Unbiased Markets, and Martingale Models,"
Journal of Business,39 (Special Supplement,January, 1966), 242-55.
28. . "The Variation of Certain SpeculativePrices." Journal of Business,36 (October,
1963), 394-419.
29. and Howard M. Taylor. "On the Distribution of Stock Price Differences."
OperationsResearch, 15 (November-December,1967), 1057-62.
of Investment.New York:
Diversification
30. Harry Markowitz.PortfolioSelection:Efficient
John Wiley & Sons, 1959.
31. Arnold Moore. "A Statistical Analysis of Common Stock Prices. Unpublished Ph.D. thesis,
Graduate School of Business,Universityof Chicago, 1962.
32. Victor Niederhofferand M. F. M. Osborne. "Market Making and Reversal on the Stock
Exchange." Journal of the AmericanStatistical Association,61 (December, 1966), 897-916.
33. M. F. M. Osborne."Brownian Motion in the Stock Market," OperationsResearch, 7 (March-
April, 1959), 145-73. Also reprintedin [8], 100-28.
34. . "Periodic Structure in the Brownian Motion of Stock Prices." Operations
Research, 10 (May-June, 1962), 345-79. Also reprintedin [8], 262-96.
35. S. James Press. "A compound Events Model for Security Prices." Journal of Business, 40
(July, 1968), 317-35.
36. Harry V. Roberts. "Stock Market 'Patterns' and Financial Analysis: Methodological Sug-
gestions."Journal of Finance, 14 (March, 1959), 1-10.
37. Richard Roll. "The EfficientMarket Model Applied to U.S. TreasuryBill Rates." Unpublished
Ph.D. thesis,Graduate School of Business,Universityof Chicago, 1968.
38. Paul A. Samuelson. "Proof That ProperlyAnticipatedPrices Fluctuate Randomly." Industrial
Management Review, 6 (Spring, 1965), 41-9.
39. Myron Scholes. "A Test of the CompetitiveHypothesis: The Market for New Issues and
SecondaryOfferings." UnpublishedPH.D. thesis,Graduate School of Business,Universityof
Chicago, 1969.
40. William F. Sharpe. "Capital Asset Prices: A Theory of Market Equilibriumunder Conditions
of Risk." Journal of Finance, 19 (September,1964), 425-42.
41. . "Mutual Fund Performance."Journal of Business, 39 (Special Supplement,
January, 1966), 119-38.
42. . "Risk Aversion in the Stock Market." Journal of Finance, 20 (September,
1965), 416-22.
43. James Tobin. "Liquidity Preferenceas Behavior Towards Risk," Review of Economic Studies,
25 (February, 1958), 65-85.
44. Jack L. Treynor. "How to Rate Management of Investment Funds." Harvard Business
Review, 43 (January-February,1965), 63-75.
45. Roger N. Waud. "Public Interpretationof Discount Rate Changes: Evidence on the 'An-
nouncementEffect."' forthcomingin Econometrica.
46. John Wise. "Linear Estimators for Linear Regression Systems Having Infinite Variances."
Unpublished paper presentedat the Berkeley-StanfordMathematical Economics Seminar,
October, 1963.
47. Holbrook Working. "A Random DifferenceSeries for Use in the Analysis of Time Series."
Journalof theAmerican
Statistical 29 (March,1934), 11-24.
Association,

Eugene Fama - Mercados Eficientes

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Eugene Fama - Mercados Eficientes

Uploaded by

Copyright:

Available Formats

SESSION TOPIC: STOCK MARKET PRICE BEHAVIOR

SESSION CHAIRMAN: BURTON G. MALKIEL

EFFICIENT CAPITAL MARKETS: A REVIEW OF

A. ExpectedReturnor "Fair Game" Models

whereE is theexpectedvalue operator;pit is thepriceof securityj at timet;

be any tradingsystembased on 1?t whichtellstheinvestortheamountsaj ((It)

which,fromthe "fairgame"propertyof (5) has expectation,

C. The Random Walk Model

3. Note thattheexpectedprofitability and cash" tradingsystems

5. The randomwalk modeldoes not say, however,that past information is of no value in

III. THE EVIDENCE

6. Basing theiranalyses on futurescontractsin commoditymarkets,Mandelbrot and Samuelson

E (it+r iit) xtE (it+rIxt) f(xt)dxt,

wheref indicatesa densityfunction.

8. Our briefhistoricalreviewis meant only to provide perspective,and it is, of course,somewhat

E(xt+ll,t) = f xt+dF(xt+i t1e, = f[f xt+dF(xt+1I4t)

If probabilitiesare uniformlydistributedacross events,

E(xt?21xt+l) = E(xt+2Ixt) .= E(xt+llxt) = E(xt+2) = E(xt+?) = E(xt) = 0,

For example,Table 1 (taken from[10]) showsthe serialcorrelations be-

16.On a transaction to transactionbasis,positiveand negativepricechangesare aboutequally

[E(r-m,t+,ifDt)- rf,t+1]/0(rm,t+1j@I t),

The questionnow is to what extent(13) is consistentwith the Sharpe-

E(r-j,t+1J|t) aj((Dt)+ (3j((Dt)E(rim,t+i1|Dt), (14)

Withsome simplifying assumptions,(14) can be reducedto (13). In partic-

a'J=Rjt- 3jRMt = 1 + rJt-3j(1 + rMt) = 1- pj + aj

The averageresidualum can be interpreted as the average deviation(in

C. StrongForm Tests of the Efficient MarketsModels

E(ib,t?1lit,rn ,t+1)- ------------------------------------

rj,t+l = ri,t+,[l - j(Dt)] + rm,t+j0j((Dt) + Uj,t+ls

You might also like