Risk, Return, and Equilibrium: Empirical Tests

Author(s): Eugene F. Fama and James D. MacBeth

Source: Journal of Political Economy, Vol. 81, No. 3 (May - Jun., 1973), pp. 607-636
Published by: University of Chicago Press
Stable URL: http://www.jstor.org/stable/1831028
Accessed: 08-02-2016 11:52 UTC

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at http://www.jstor.org/page/
info/about/policies/terms.jsp
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content
in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship.
For more information about JSTOR, please contact support@jstor.org.

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

http://www.jstor.org

This content downloaded from 152.118.24.10 on Mon, 08 Feb 2016 11:52:52 UTC
All use subject to JSTOR Terms and Conditions

Risk, Return,and Equilibrium:

Empirical Tests

Eugene F. Fama and James D. MacBeth

Universityof Chicago

This paper tests the relationshipbetweenaverage returnand risk for

New York Stock Exchange commonstocks. The theoreticalbasis of
the testsis the "two-parameter"
portfoliomodel and modelsof market
equilibriumderivedfromthe two-parameter
portfoliomodel. We cannot reject the hypothesisof these models that the pricingof common
stocks reflectsthe attemptsof risk-averseinvestorsto hold portfolios
thatare "efficient"
in termsof expectedvalue and dispersionof return.
Moreover,the observed"fair game" propertiesof the coefficients
and
residualsof the risk-return
regressionsare consistentwithan "efficient
capital market"-that is, a market where prices of securitiesfully
reflectavailable information.

I.

Theoretical Background

In the two-parameterportfoliomodel of Tobin (1958), Markowitz (1959),

and Fama (1965b), the capital market is assumed to be perfect in the
sense that investors are price takers and there are neither transactions
costs nor informationcosts. Distributions of one-period percentage returns
on all assets and portfolios are assumed to be normal or to conform to
some other two-parametermemberof the symmetricstable class. Investors
are assumed to be risk averse and to behave as if they choose among
portfolios on the basis of maximum expected utility. A perfect capital
market, investor risk aversion, and two-parameter return distributions
imply the important "efficientset theorem": The optimal portfolio for
any investormust be efficientin the sense that no other portfoliowith the
same or higherexpected returnhas lower dispersionof return.1
Received August 24, 1971. Final versionreceived for publication September2, 1972.
Research supported by a grant from the National Science Foundation. The commentsof ProfessorsF. Black, L. Fisher,N. Gonedes, M. Jensen,M. Miller, R. Officer,
H. Roberts, R. Roll, and M. Scholes are gratefullyacknowledged.A special note of
thanks is due to Black, Jensen,and Officer.
1 Althoughthe choice of dispersionparameter is arbitrary,the standard deviation
607

This content downloaded from 152.118.24.10 on Mon, 08 Feb 2016 11:52:52 UTC
All use subject to JSTOR Terms and Conditions

6o8

JOURNAL

OF POLITICAL

ECONOMY

In the portfolio model the investor looks at individual assets only in

termsof their contributionsto the expected value and dispersion, or risk,
of his portfolioreturn.With normal returndistributionsthe risk of portfolio p is measured by the standard deviation, cY(Rp), of its return,Rp,2
and the risk of an asset for an investor who holds p is the contributionof
the asset to cY(Rp). If xip is the proportionof portfoliofunds invested in
asset i, ij - cov(Ri, Rj) is the covariance between the returnson assets i
and j, and N is the number of assets, then

i a;= ILi
Ej
N

-N

(Rp

Ntv

cov(Rs Rp)

Thus, the contribution

of asset i to o(Rp)-that
is,
the
cov (Ri,
Exjpaij/6(Rpu)
Rp)/
R) risk of asset i in
a p==to
R) is =proportional
ip
the portfolio
aA
N

Note that since the weights xj, vary fromportfolio to portfolio,the risk
of an asset is differentfor differentportfolios.
For an individual investor the relationshipbetween the risk of an asset
and its expected returnis implied by the fact that the investor's optimal
portfoliois efficient.Thus, if he chooses the portfoliom, the fact that m
is efficientmeans that the weightsxim,i -1, 2, .. ., N, maximize expected
portfolioreturn

Z Xir E(Ri),
N

E(Rm)

i==

subject to the constraints

is common when returndistributionsare assumed to be normal, whereas an interfractile range is usually suggested when returns are generated from some other
symmetricstable distribution.
It is well known that the mean-standard deviation version of the two-parameter
portfoliomodel can be derived from the assumption that investors have quadratic
utilityfunctions.But the problems with this approach are also well known. In any
case, the empiricalevidence of Fama (1965a), Blume (1970), Roll (1970), K. Miller
(1971), and Officer(1971) provides support for the "distribution"approach to the
model. For a discussionof the issues and a detailed treatmentof the two-parameter
model, see Fama and Miller (1972, chaps. 6-8).
We also concentrateon the special case of the two-parametermodel obtained with
the assumptionof normallydistributedreturns.As shown in Fama (1971) or Fama
and Miller (1972, chap. 7), the importanttestable implicationsof the general symmetricstable model are the same as those of the normal model.
2 Tildes (~) are used to denote random variables. And the one-period percentage
returnis most oftenreferredto just as the return.

This content downloaded from 152.118.24.10 on Mon, 08 Feb 2016 11:52:52 UTC
All use subject to JSTOR Terms and Conditions

RISK,

RETURN,

AND EQUILIBRIUM

609
N

a(Rp)

=(Rm)

and

Xim - 1.

i==1

Lagrangian methods can then be used to show that the weights xjm must
be chosen in such a way that for any asset i in m
N

E(Ri)

- E(Rm)- SmL X1mij

]n

(1)

where Sm is the rate of change of E(R')

with respect to a change in
o(Rp) at the point on the efficientset correspondingto portfolio m. If
there are nonnegativityconstraintson the weights (that is, if short selling
is prohibited), then (1) only holds for assets i such that xim> 0.
Although equation (1) is just a condition on the weights xjm that is required for portfolioefficiency,
it can be interpretedas the relationshipbetween the risk of asset i in portfoliom and the expected returnon the asset.
The equation says that the differencebetween the expected returnon the
asset and the expected returnon the portfoliois proportionalto the difference between the risk of the asset and the risk of the portfolio.The proportionalityfactor is Snl, the slope of the efficientset at the point corresponding to the portfoliom. And the risk of the asset is its contributionto
total portfoliorisk, 6( R,)II.

Testable

Implications

Suppose now that we posit a market of risk-averse investors who make

portfoliodecisions period by period according to the two-parametermodel.3
We are concerned with determining what this implies for observable
propertiesof securityand portfolioreturns.We consider two categories of
implications. First, there are conditions on expected returns that are implied by the fact that in a two-parameterworld investors hold efficient
portfolios.Second, there are conditions on the behavior of returnsthrough
time that are implied by the assumption of the two-parametermodel that
the capital market is perfect or frictionlessin the sense that there are
neithertransactionscosts nor informationcosts.
A.

Expected Returns

The implications of the two-parametermodel for expected returnsderive

fromthe efficiencycondition or expected return-riskrelationship of equation ( 1). First, it is convenient to rewrite ( 1) as
3A multiperiodversion of the two-parametermodel is in Fama (1970a) or Fama
and Miller (1972, chap. 8).

This content downloaded from 152.118.24.10 on Mon, 08 Feb 2016 11:52:52 UTC
All use subject to JSTOR Terms and Conditions

6io

JOURNAL

ECONOMY

- Sn,0(Rm)] + Sm (Rm)f3i,

[E(Rm)

E(Ri)

OF POLITICAL

(2)

where
N

~~

~xjmaij

=Zj

cov(Ri, Rn,)

2(RG )

~o

cov(Ri, Rm)/y(Rm)

(Rm)

0(Pa)

The parameter Pican be interpretedas the risk of asset i in the portfolio

m, measured relative to o(R,,), the total risk of m. The interceptin (2),
E (RO) =-:E(Rsn) -Sm a (Rm),

(4)

is the expected return on a security whose return is uncorrelated with

Rm-that is, a zero-0 security. Since (
0 implies that a security contributesnothingto y(Rt,,),it is appropriate to say that it is risklessin this
portfolio. It is well to note from (3), however, that since ximG.i.
Xim
0 does not imply that security
o2(RI) is just one of the N termsin Pi, Hi
i has zero variance of return.
From (4), it follows that

Sm~

E(Rnb)

(R)

E(RO)

'

(5)

so that (2) can be rewritten

E(Ri)

E(R)

+ [E(Ril)

E(R,)](3.

(6)

In words, the expected returnon securityi is E(R0,), the expected return

on a security that is riskless in the portfoliom, plus a risk premium that
is Pitimes the differencebetween E(R,,.) and E(R)).
Equation (6) has three testable implications: (Cl) The relationship
between the expected returnon a securityand its risk in any efficientportfolio m is linear. (C2) Piis a complete measure of the risk of securityi in
the efficientportfoliom; no other measure of the risk of i appears in (6).
(C3) In a marketof risk-averseinvestors,higher risk should be associated
with higher expected return; that is, E(Real ) - E(R0) > 0.
The importance of condition C3 is obvious. The importance of C1 and
C2 should become clear as the discussion proceeds. At this point sufficeit
to say that if C1 and C2 do not hold, market returnsdo not reflectthe
attempts of investors to hold efficientportfolios: Some assets are systematically underpriced or overpriced relative to what is implied by the
expected return-riskor efficiencyequation (6).
B.

Market Equilibrium and the Efficiencyof the Market Portfolio

To test conditions C l-C3 we must identify some efficientportfolio m.

This in turn requires specification of the characteristic of market equi-

This content downloaded from 152.118.24.10 on Mon, 08 Feb 2016 11:52:52 UTC
All use subject to JSTOR Terms and Conditions

RISK,

RETURN,

6i i

AND EQUILIBRIUM

libriumwhen investorsmake portfoliodecisionsaccordingto the twoparameter

model.
Assumeagain that the capital marketis perfect.In addition,suppose
that fromthe information
available withoutcost all investorsderivethe
same and correctassessmentof the distribution
of the futurevalue of any
asset or portfolio-an assumptionusually called "homogeneousexpectations." Finally,assume that shortsellingof all assets is allowed. Then
Black (1972) has shown that in a marketequilibrium,the so-called
marketportfolio,
definedby the weights
Xtm

totalmarketvalue ofall unitsof asset i

totalmarketvalue of all assets

is alwaysefficient.
Since it containsall assets in positiveamounts,the marketportfoliois
a convenient
reference
pointfortestingthe expectedreturn-risk
conditions
C 1-C3 of the two-parameter
model. And the homogeneous-expectations
assumptionimplies a correspondence
between ex ante assessmentsof
returndistributions
and distributions
of ex post returnsthat is also requiredformeaningful
testsof thesethreehypotheses.
C. A StochasticModel forReturns
Equation (6) is in termsof expectedreturns.But its implications
mustbe
testedwith data on period-by-period
securityand portfolioreturns.We
wishto choosea modelof period-by-period
returnsthat allows us to use
observedaverage returnsto test the expected-return
conditionsC1-C3,
but one thatis nevertheless
as generalas possible.We suggestthe following stochasticgeneralization
of (6):
Rit = YOt+

Alto

7'2/j2

'3tSi

Nit.

(7)

The subscriptt refersto periodt, so thatRit is the one-periodpercent-

age returnon securityi fromt -

1 to t. Equation (7) allows yet and yit

to varystochastically
fromperiodto period.The hypothesisof condition
C3 is that the expectedvalue of the riskpremium'it, whichis the slope
[E(Rn11)- E(R<t)] in (6), is positive-that is, E(71t) -E(Rmt) -

E(R(t)

> 0.

The variableP12 is includedin (7) to test linearity.The hypothesisof

conditionC1 is E(y.2t)
.0 althoughyet is also allowedto varystochasticallyfromperiodto period.Similarstatements
applyto the terminvolving
sj in (7), whichis meantto be somemeasureof the riskof securityi that
is not deterministically
relatedto Pi. The hypothesisof conditionC2 is
E(7 t) - 0, but y3tcan varystochastically
throughtime.
The disturbance is assumedto have zeromeanand to be independent
of all othervariablesin (7). If all portfolioreturndistributions
are to be

This content downloaded from 152.118.24.10 on Mon, 08 Feb 2016 11:52:52 UTC
All use subject to JSTOR Terms and Conditions

612

JOURNAL

OF POLITICAL

ECONOMY

normal(or symmetric
stable), thenthe variablesNt ?'t,
ht, Y2t and y3t
'y
musthave a multivariate
normal(or symmetric
stable) distribution.
D.

Capital Market Efficiency:The Behaviorof ReturnsthroughTime

C1-C3 are conditionson expectedreturnsand riskthat are impliedby

the two-parameter
model. But the model,and especiallythe underlying
assumption
of a perfectmarket,impliesa capitalmarketthatis efficient
in
thesensethatpricesat everypointin timefullyreflectavailableinformation.This use of the wordefficient
is, of course,not to be confusedwith
portfolioefficiency.
The terminology,
if a bit unfortunate,
is at least
standard.
Marketefficiency
in combination
withconditionCI requiresthatscrutiny
of the time seriesof the stochasticnonlinearity
coefficient
'2t does not
lead to nonzeroestimatesof expectedfuturevalues of y2t. Formally,Y2t
mustbe a fairgame. In practicalterms,althoughnonlinearities
are observedex post,because Y2tis a fairgame,it is always appropriateforthe
investorto act ex ante under the presumptionthat the two-parameter
model,as summarized
by (6), is valid. That is, in his portfoliodecisions
he alwaysassumesthat thereis a linearrelationship
betweenthe riskof
a securityand its expectedreturn.Likewise,marketefficiency
in the twoparametermodelrequiresthat the non-( riskcoefficient
y t and the time
seriesof returndisturbancesnitare fairgames.And the fair-game
hypothesis also applies to the time series of lit - [E(Rmt)

E(Rot)],

the

difference
betweenthe riskpremiumforperiodt and its expectedvalue.
In the terminology
of Fama (1970b), theseare "weak-form"propositionsabout capital marketefficiency
fora marketwhereexpectedreturns
are generated
by thetwo-parameter
model.The propositions
are weaksince
theyare onlyconcernedwithwhetherpricesfullyreflectany information
in the timeseriesof past returns."Strong-form"
testswouldbe concerned
withthe speed-of-adjustment
of pricesto all available information.
E.

MarketEquilibriumwithRisklessBorrowingand Lending

We have as yet presentedno hypothesisabout y t in (7). In the general

two-parameter
model,given E(' t) E(73t)
E("Qit) 0 then, from
(6), E(y)t) is just E(Rt), the expectedreturnon any zero-dsecurity.
And marketefficiency
requiresthat t- E(R(t) be a fairgame.
But if we add to the modelas presentedthus far the assumptionthat
thereis unrestricted
risklessborrowing
and lendingat the knownrateR1t,
thenone has themarketsettingof theoriginaltwo-parameter
"capitalasset
pricingmodel"of Sharpe (1964) and Lintner(1965). In thisworld,since
-

At

0, E(`

a fairgame.

t) - Rft. And market efficiencyrequires that

$1t -

Rft be

This content downloaded from 152.118.24.10 on Mon, 08 Feb 2016 11:52:52 UTC
All use subject to JSTOR Terms and Conditions

RISK,

RETURN,

613

AND EQUILIBRIUM

It is well to emphasizethatto refutethe proposition

thatE(7Yot)- Rft
is only to refutea specifictwo-parameter
model of marketequilibrium.
Our view is that tests of conditionsC1-C3 are more fundamental.
We
regardC1-C3 as the generalexpectedreturnimplicationsof the twoparametermodelin the sense that theyare the implicationsof the fact
thatin thetwo-parameter
portfolio
modelinvestorsholdefficient
portfolios,
and theyare consistentwith any two-parameter
model of marketequilibriumin whichthe marketportfoliois efficient.
F.i The Hypotheses
To summarize,
giventhe stochasticgeneralization
of (2) and (6) that is
providedby (7), the testableimplicationsof the two-parameter
model
forexpectedreturnsare:
C1 (linearity)-E('2t)-- 0.
C2 (no systematic
of non-0risk)-E(y:jt) - 0.
effects
C3 (positive expected return-risk
tradeoff) E('1t) -

E(Rot) > 0.

Sharpe-Lintner
(S-L) Hypothesis-E(7 t)

E(R'1,)

Rft.

Finally,capitalmarketefficiency
in a two-parameter
worldrequires
ME

(market efficiency) the stochastic coefficientsY2't, Y.-It,7t

[E(R,,,t) - E(R(t) j, 7f,- E(Rot), and the disturbancesIVt are fair

games.4
III.

Previous

Work5

The earliesttests of the two-parameter

model were done by Douglas
(1969), whoseresultsseemto refuteconditionC2. In annualand quarterly
returndata, thereseemto be measuresof risk,in additionto (, that contributesystematically
to observedaveragereturns.These results,if valid,
are inconsistent
withthehypothesis
thatinvestorsattemptto hold efficient
portfolios.
Assumingthat the marketportfoliois efficient,
premiumsare
paid forrisksthat do not contributeto the riskof an efficient
portfolio.
Miller and Scholes (1972) take issue both with Douglas's statistical
techniquesand withhis use of annual and quarterlydata. Using different
methodsand simulations,
theyshow that Douglas's negativeresultscould
be expectedevenif conditionC2 holds.ConditionC2 is testedbelowwith
extensivemonthlydata, and this avoids almostall of the problemsdiscussedby Millerand Scholes.
If Y'Y.fand Y are fair games, then E(y.,,) = E(
-t) -0. Thus, C1 and C2 are
implied by ME. Keeping the expected return conditions separate, however, better
emphasizesthe economic basis of the various hypotheses.
5 A comprehensivesurvey of empirical and theoreticalwork on the two-parameter
model is in Jensen (1972).

This content downloaded from 152.118.24.10 on Mon, 08 Feb 2016 11:52:52 UTC
All use subject to JSTOR Terms and Conditions

JOURNAL

6I4

OF POLITICAL

ECONOMY

model is
Much of the available empiricalworkon the two-parameter
thatE(70ot) - Rft.The testsof
concernedwithtestingtheS-L hypothesis
Friendand Blume (1970) and thoseof Black, Jensen,and Scholes (1972)
indicatethat,at least in the periodsince 1940, on averageyot is systematicallygreaterthanRft.The resultsbelow supportthis conclusion.
of the linearitycondiIn theempiricalliterature
to date,theimportance
tion C1 has been largelyoverlooked.Assumingthat the marketportfolio
m is efficient,
if E('2t) in (7) is positive,the pricesof high-( securities
are on averagetoo low-their expectedreturnsare too high-relativeto
thoseof low-0securities,
whilethe reverseholdsif E(2't) is negative.In
in the capital marketreflectsthe
short,if the processof price formation
portfolios,
thenthe linearrelationattemptsof investorsto hold efficient
ship of (6) betweenexpectedreturnand riskmusthold.
model has
Finally,the previousempiricalworkon the two-parameter
not been concernedwithtestsof marketefficiency.
IV.

Methodology

The data forthis studyare monthlypercentagereturns(includingdividends and capital gains, with the appropriateadjustmentsfor capital
changessuch as splitsand stockdividends)forall commonstockstraded
on theNew YorkStockExchangeduringtheperiodJanuary1926 through
June1968. The data are fromthe CenterforResearchin SecurityPrices
of the University
of Chicago.
A. GeneralApproach
model immediatelypresentsan unavoidable
Testing the two-parameter
conditionor expected
"errors-in-the-variables"
problem: The efficiency
return-risk
equation (6) is in termsof true values of the relativerisk
measurePi,but in empiricaltestsestimates,Hi,mustbe used. In thispaper
cov (Ri,

Rm)

wherecov(Ri, R,,,) and a'(Rm) are estimatesof cov(Ri, R,,) and 62(Rm)
obtainedfrommonthlyreturns,and wherethe proxychosen forRmtis
"Fisher'sArithmetic
Index," an equally weightedaverageof the returns
on all stockslistedon the New York Stock Exchangein montht. The
of this index are analyzedin Fisher (1966).
properties
Blume (1970) showsthat forany portfoliop, definedby the weights
- 1, 2, . . . , NN
xip,i
g

I p-

cov(Rp,Rr)
0*'(Rn)

cov(Ri, Rm)

(R,)
~i=1

This content downloaded from 152.118.24.10 on Mon, 08 Feb 2016 11:52:52 UTC
All use subject to JSTOR Terms and Conditions

RISK,

RETURN,

AND EQUILIBRIUM

615

If the errorsin the 1 are substantially

less than perfectly
positivelycorrelated,the Rt'sof portfolioscan be muchmorepreciseestimatesof true
P's thanthe 03'sforindividualsecurities.
To reduce the loss of information
in the risk-return
tests caused by
usingportfoliosratherthan individualsecurities,a wide rangeof values
of portfolioit'sis obtainedby forming
portfolioson the basis of ranked
values of jb forindividualsecurities.But such a procedure,naYvelyexecuted could resultin a seriousregression
phenomenon.
In a cross section
of V, highobservedj tendto be above the corresponding
truefPjand low
observed^Pitendto be below the truePi. Formingportfolioson the basis
of ranked/j thuscauses bunchingof positiveand negativesamplingerrors
withinportfolios.
The resultis thata largeportfolio
jUpwouldtendto overstatethe true PP,whilea low Pp wouldtendto be an underestimate.
The regression
phenomenon
can be avoidedto a largeextentby forming
portfolios
fromranked,j computedfromdata forone timeperiodbut then
usinga subsequentperiodto obtain the 'P for these portfoliosthat are
used to test the two-parameter
model.With freshdata, withina portfolio
A
errorsin the individualsecurityP3 are to a large extentrandomacross
securities,
so thatin a portfolio%ptheeffects
of theregression
phenomenon
are, it is hoped,minimized.6
B.

Details

The specificsof theapproachare as follows.Let N be the totalnumberof

securitiesto be allocated to portfoliosand let int(N/20) be the largest
integerequal to or less thanN120. Using the first4 years (1926-29) of
monthlyreturndata, 20 portfoliosare formedon the basis of ranked i
for individualsecurities.The middle 18 portfolioseach has int(N/20)
If N is even,the firstand last portfolios
securities.
each has int(N/20) +
-')[N - 20 int(N/20)] securities.The last (highest^3) portfoliogets an
additionalsecurityif N is odd.
The following5 years (1930-34) of data are thenused to recompute
theHi,and theseare averagedacrosssecuritieswithinportfoliosto obtain
tests.The subscriptt is added to
20 initialportfolio
Aptforthe risk-return
indicatethat each montht of the followingfouryears (1935-38) these
as simpleaveragesof individualsecurityPi, thus ad1pt are recomputed
justingthe portfoliojpt monthby monthto allow fordelistingof securities.The component
Adforsecuritiesare themselvesupdatedyearly-that
6 The errors-in-the-variables
problem and the technique of using portfolios to
solve it were firstpointed out by Blume (1970). The portfolioapproach is also used
by Friend and Blume (1970) and Black, Jensen,and Scholes (1972). The regression
phenomenonthat arises in risk-returntests was first recognized by Blume (1970)
and then by Black, Jensen,and Scholes (1972), who offera solution to the problem
that is similarin spiritto ours.

This content downloaded from 152.118.24.10 on Mon, 08 Feb 2016 11:52:52 UTC
All use subject to JSTOR Terms and Conditions

6i6

JOURNAL

OF POLITICAL

ECONOMY

is, they are recomputedfrommonthlyreturnsfor 1930 through1935,

1936, or 1937.

As a measureof the non-0riskof securityi we use s('tj), the standard

deviationof theleast-squaresresiduals'it fromthe so-calledmarketmodel
Rit = ai + Pi Rmt+ Zjt.

(8)

The standarddeviations(ti.) is a measureof non-0riskin the following

sense. One view of risk,antitheticto that of portfoliotheory,says that
the risk of a securityis measuredby the total dispersionof its return
distribution.
Givena marketdominatedby riskaverters,thismodelwould
predictthata security'sexpectedreturnis relatedto its total returndispersionratherthanjust to thecontribution
of thesecurityto thedispersion
in thereturnon an efficient
portfolio.7
If Bi Gocov (Ri, R,) /a2(Rf), then
in (8) cov( Eii,Rm)

0, and

(Ri) -- i32o2(Rm) +

E2(i)

+ 2Picov(Rm,A).

(9)

Thus, from(9), one can say that s(ti) is an estimateof thatpart of the
dispersion
of thedistribution
of thereturnon securityi thatis notdirectly
relatedto P3i.
The month-by-month
returnson the 20 portfolios,
withequal weighting
of individualsecuritieseach month,are also computedfor the 4-year
period 1935-38. For each montht of this period,the followingcrosssectionalregression-theempiricalanalog of equation (7)-is run:
Rpt = cot +

fit Apt-l

72t 2

?3tSpstl-(hi)

Pt

(10)

p - 1,2,... ,20.

The independent
variableAp t-l is the averageof the A*forsecuritiesin
portfoliop discussedabove; i92pt_ is the averageof the squared values
of thesefi,(and is thussomewhatmislabeled); and Tpjt_-(?i.) is likewise
the averageof s(,X) forsecuritiesin portfoliop. The s(^E) are computed
fromdata forthesameperiodas thecomponentPiof kt-1, and like these
ti, they are updated annually.

The regression
equation (10) is (7) averagedacross the securitiesin a
withestimates(plt-1, 32p't-1, and spjt_('u) used as explanatory
portfolio,
variables,and with least-squaresestimatesof the stochasticcoefficients
time seriesof month-byYbt, Yet, ?2t, and y3t. The results from (10)-the
monthvalues of the regressioncoefficients
and at for the
t2t,
9ot, ?11t,
4-yearperiod1935-38-are the inputsforour testsof the two-parameter
modelforthisperiod.To get resultsforotherperiods,the stepsdescribed
7 For those accustomed to the portfolio viewpoint, this alternative model may
seem so naive that it should be classifiedas a straw man. But it is the model of risk
and returnimplied by the "liquidity preference"and "market segmentation"theories
of the termstructureof interestrates and by the Keynesian "normal backwardation"
theoryof commodityfuturesmarkets.For a discussionof the issues with respect to
these markets,see Roll (1970) and K. Miller (1971).

This content downloaded from 152.118.24.10 on Mon, 08 Feb 2016 11:52:52 UTC
All use subject to JSTOR Terms and Conditions

RISK,

RETURN,

617

AND EQUILIBRIUM

above are repeated.That is, 7 yearsof data are used to formportfolios;

the next 5 years are used to computeinitial values of the independent
variablesin (10); and then the risk-return
regressionsof (10) are fit
monthby monthforthe following
4-yearperiod.
The nine different
portfolioformationperiods (all except the first7
yearsin length),initial5-yearestimationperiods,and testingperiods(all
but thelast 4 yearsin length)are shownin table 1. The choiceof 4-year
testingperiodsis a balance of computationcosts against the desire to
reformportfoliosfrequently.The choice of 7-yearportfolioformation
periodsand 5-8-yearperiods for estimatingthe independentvariables
in the risk-return
reflectsa desireto balregressions
Ppst-l and sp,tI (f)
ance the statisticalpowerobtainedwitha large samplefroma stationary
processagainstpotentialproblemscaused by any nonconstancy
of the Pit.
The choices here are in line with the resultsof Gonedes (1973). His
resultsalso led us to requirethat to be includedin a portfolioa security
availablein thefirstmonthof a testingperiodmustalso have data forall
5 yearsof theprecedingestimationperiodand forat least 4 yearsof the
portfolioformation
period.The total numberof securitiesavailablein the
firstmonthof each testingperiodand the numberof securitiesmeeting
the data requirement
are shownin table 1.
C.

Some Observations on the Approach

Table 2 showsthe values of the 20 portfoliosfkvt-1and theirstandard

errorss(0ptj)

for four of the nine 5-year estimationperiods. Also shown

are: r(RA,R, )2, the coefficient

of determination
betweenRpt and Rest;
s(Rp), thesamplestandarddeviationof Rp; and s(Ep), thestandarddeviation of the portfolioresidualsfromthe marketmodel of (8), not to be
confused with spvt1 (h), the average for individual securities,which is also
E ) are the independent variables in the risk
shown. The jktand -S

returnregressions
of (10) forthe firstmonthof the 4-yeartestingperiods
following
the fourestimation
periodsshown.
Underthe assumptionsthat fora givensecuritythe disturbancesEjt in
(8) are seriallyindependent,
of Rmt,and identicallydistribindependent
uted throughtime,the standarderrorof NI is
(

N/fl

(Rm)

wheren is the numberof monthsused to computePi. Likewise,

/
(it t1 )

-"-

Thus, the factthatin table 2, s(Ep) is generallyon the orderof one-third

to one-seventh
impliesthats(Nj-1)
is one-third
to one-seventh
sp'tE(fi,)

This content downloaded from 152.118.24.10 on Mon, 08 Feb 2016 11:52:52 UTC
All use subject to JSTOR Terms and Conditions

6i8

JOURNAL OF POLITICAL ECONOMY

TABLE 1
PORTFOLIO

FORMATION,

AND TESTING

ESTIMATION,

PERIODS

PERIODS

Portfolio formationperiod ...

Initial estimationperiod.
Testing period .1935-38
No. of securitiesavailable
No. of securitiesmeeting
data requirement
.435

1926-29
1930-34

1927-33
1934-38
1939-42

1931-37
1938-42
1943-46

1935-41
1942-46
1947-50

1939-45
1946-50
1951-54

710

779

804

908

1,011

576

607

704

751

s(f,). Estimates of ( for portfoliosare indeed more precise than those for
individual securities.
Nevertheless,it is interestingto note that if the disturbances 3jt in (8)
were independent from security to security, the relative increase in the
precision of the AP obtained by using portfolios rather than individual
securities would be about the same for all portfolios. We argue in the
Appendix,however,that the results from (10) imply that the fit in (8) are
interdependent,and the interdependenceis strongest among high-0 securities and among low-e securities. This is evident in table 2: The ratios
s('p) /pt-1 ('i) are always highest at the extremesof the pt-I range and
lowest for Ppet_1 close to 1.0. But it is importantto emphasize that since
these ratios are generally less than .33, interdependenceamong the fit of
differentsecuritiesdoes not destroy the value of using portfoliosto reduce
the dispersion of the errorsin estimated 0's.
Finally, all the tests of the two-parametermodel are predictive in the

sensethattheexplanatory
variables(,t-1

and Spt-1('(i)

in (10) are com-

puted fromdata fora period prior to the month of the returns,the Rpt, on
which the regressionis run. Although we are interestedin testing the twoparameter model as a positive theory that is, examining the extent to
which it is helpfulin describingactual returndata-the model was initially
developed by Markowitz (1959) as a normativetheory-that is, as a model
to help people make betterdecisions. As a normativetheorythe model only
has content if there is some relationshipbetween future returnsand estimates of risk that can be made on the basis of currentinformation.
Now that the predictive nature of the tests has been emphasized, to
simplify the notation, the explanatory variables in (10) are henceforth
referredto as fp (3p2, and s('~).

V.

Results

model are in
The major tests of the implicationsof the two-parameter
table 3. Results are presented for 10 periods: the overall period 1935-

This content downloaded from 152.118.24.10 on Mon, 08 Feb 2016 11:52:52 UTC
All use subject to JSTOR Terms and Conditions

RISK,

RETURN,

6I9

AND EQUILIBRIUM

TABLE 1 (Continued)
PERIODS

Portfolio formationperiod ...

Initial estimationperiod ......
Testing period ..............
No. of securitiesavailable ... .
No. of securitiesmeeting
data requirement..........

1943-49
1950-54
1955-58

1947-53
1954-58
1959-62

1951-57
1958-62
1963-66

1955-61
1962-66
1967-68

1,053

1,065

1,162

1,261

802

856

858

845

and six
6/68; three long subperiods, 1935-45, 1946-55, and 1956-6/68;
subperiods which, except for the first and last, cover 5 years each. This
choice of subperiods reflects the desire to keep separate the pre- and postWVorldWar II periods. Results are presented for four different versions of
the risk-return regression equation

(10):

Panel D is based on (10)

itself,

but in panels A-C, one or more of the variables in (10) is suppressed.

For each period and model, the table shows: ^j, the average of the monthby-monthregressioncoefficientestimates, Yjt; s(9j), the standard deviation of the monthly estimates; and r' and s(r2), the mean and standard
deviation of the month-by-monthcoefficientsof determination,rt2,which
are adjusted for degrees of freedom.The table also shows the first-order
serial correlationsof the various monthly Vjt computed either about the
sample mean of jt [in which case the serial correlations are labeled
p,1(I) ] or about an assumed mean of zero [in which case they are labeled
p(,(j)]. Finally, t-statisticsfor testingthe hypothesisthat \j = 0 are presented. These t-statisticsare
s (% ) / 'n
where n is the number of months in the period, which is also the number
of estimates ?jt used to compute %jand s(9j).
In interpretingthese t-statisticsone should keep in mind the evidence
of Fama (1965a) and Blume (1970) which suggests that distributionsof
common stock returns are "thick-tailed" relative to the normal distribution and probably conformbetter to nonnormalsymmetricstable distributions than to the normal. From Fama and Babiak (1968), this evidence
means that when one interpretslarge t-statisticsunder the assumption that
the underlyingvariables are normal, the probability or significancelevels
obtained are likely to be overestimates.But it is important to note that,
with the exceptionof condition C3 (positive expected return-risktradeoff),
upward-biased probability levels lead to biases toward rejection of the
hypothesesof the two-parametermodel. Thus, if these hypotheses cannot

This content downloaded from 152.118.24.10 on Mon, 08 Feb 2016 11:52:52 UTC
All use subject to JSTOR Terms and Conditions

JOURNAL

620

STATISTICS

Statistic

FOR FOUR

ECONOMY

TABLE
SAMPLE

OF POLITICAL

SELECTED

ESTIMATION

PERIODS

10

Portfolios for Estimation Period 1934-38

r(Rp, Rm)2 .......

s (R p) ............
S (p
.............

.508
.027.027027
.861
.709
.058
.040
.022
.022

.651
.025
.921
.072
.020

.674
.023
.936
.074
.019

.695
.028
.912
.077
.023

.792
.026
.941
.087
.021

.921
.032
.932
.101
.026

.942
.029
.946
.103
.024

.970 1.005
.034
.02 7
.958
.933
.106
.109
.028
.022

.......... .
p-t _1()
S ('tp) I' t- 1 (,i~) ..

085
.259

.083
.241

.078
.244

.090
.256

.095
.221

.109
.238

.106
.226

.111
.252

.097
.227

p't-l

S(

...
t 1)

........
........

.322

.075
.293

Portfolios for Estimation Period 1942-46

ftp.t-1 .467
s (pt-1)
.........

.045
.645
.035
.021

r(Rp, Rm)2 .......

s (R p)...........

step) .............
.
SP - 1 ()
S(2p) /Sp t 1

055
..

.382

.537
.041
.745
.037
.019

.593
.044
.753
.041
.020

.628
.037
.829
.041
.017

.707
.02 7
.919
.044
.013

.721
.032
.898
.046
.015

.770
.035
.889
.049
.016

.792
.035
.898
.050
.016

.805
.028
.934
.050
.013

.894
.040
.896
.057
.018

.055
.345

.063
.317

.058
.293

.058
.224

.063
.238

.064
.250

.064
.250

.062
.210

.069
.261

Portfolios for Estimation Period 1950-54

s

P t_ 1
t_)

............
........

r(Rp, Rm)2 .......

s (R p)............

S(fp) .............
PsItl1(d.).*---*

S (' p)Ip

~t-

(i)

.418
.042
.629
.019
.012
.040
.300

.590
.047
.723
.025
.013
.044
.295

.694
.045
.798
.028
.013
.046
.283

.751
.037
.872
.029
.010
.048
.208

.777
.038
.878
.030
.010
.051
.196

.784
.035
.895
.030
.010
.051
.196

.929
.050
.856
.036
.014
.052
.269

.950
.038
.913
.036
.011
.053
.208

.996 1.014
.029
.035
.954
.933
.038
.037
.008
.010
.054
.057
.140
.185

Portfolios for Estimation Period 1958-62

t
S _(Pt-

...........
........

1)

r(Ry, Rm)2 .......

s (R ,) ............
S (e) .............
p.t- 1 ()
S (I'p) /.p, t

.............
..

(i)

.626
043
.783
.030
.014
.049
.286

.635
.048
.745
.031
.016
.052
.308

.719
.039
.851
.033
.013
.056
.232

.801
.046
.835
.037
.015
.059
.254

.817
.047
.838
.038
.015
.064
.234

.860
.033
.920
.038
.011
.061
.180

.920
.037
.913
.041
.012
.070
.171

.950
.038
.915
.042
.012
.069
.174

.975
.032
.939
.043
.011
.068
.162

.995
.037
.925
.044
.012
.064
.188

underthe assumptionof norare interpreted

be rejectedwhent-statistics
mality,the hypothesesare on even firmergroundwhen one takes into
accountthe thicktails of empiricalreturndistributions.
on monthly
Furtherjustification
to testhypotheses
forusingt-statistics
commonstockreturnsis in theworkof Officer
(1971). Undertheassumpstable,he estiof monthlyreturnsare symmetric
tion that distributions
exponent
matesthat in the post-WorldWar II periodthe characteristic

This content downloaded from 152.118.24.10 on Mon, 08 Feb 2016 11:52:52 UTC
All use subject to JSTOR Terms and Conditions

RISK,

AND

RETURN,

62 1

EQUILIBRIUM

TABLE

Statistic

11

12

2 (Continued)

13

14

15

16

17

18

19

20

Portfolios for Estimation Period 1934-38

1.046 1.122 1.181 1.192 1.196 1.295 1.335 1.396 1.445 1.458
.053
.039
.032
.028
.028
.032
.053
.031
.035
.029
.927
.922 .958
.967
.969
.966
.966
959
.956 .951
.160
.156
.122
.140
.144
.154
.128
.128
.129
.113
.043
.032
023
.029
.026
.026 .043
.026
.023
.024

.............

SPPt- I

t 1) ........

s(p

r(Rp1 Rm)2 ....... ,.

s(Rp)............

s (iUp) ..............
1
P. t_ .........
s (p
p -1)

.094
.245

.124

.210

.120
.242

.122
.188

.132
.182

.125
.208

.129
.202

.158
.272

.145
.221

.170

.253

Portfolios for Estimation Period 1942-46

.........
S(P. t - I) .......
r(Rp, Rm)2 .......
SP. t I

.............

sagp)

sp,

TP,t-- 1 (f

........
t _1(t

.031

.942
.059
.014

s(Rp) ............

S (2p

.949

.073
. .

.192

.952 1.010 1.038 1.254 1.312 1.316 1.473 1.631 1.661

.077
.083
.041
.084
.040
.030
.034
.039
.036
.867 .887
.839
.951
.945
.923
.917
.954
.958
.106
.081
.097
.105
.063
.064
.077
.081
.060
.036
.019
.039
.038
.018
.014
.016
.018
.016
.074
.216

.085
.212

.077
.182

.096
.167

.083
.217

.086
.221

.134
.291

.117
.325

.122
.295

Portfolios for Estimation Period 1950-54

............

S(

)
-1

.......... .

r (R p Rol) 2 .......
s (R p)..
..........

s(1p) .............
Tp t- I (hi) ........
SpSs) t-l~g)

..

1.117 1.123 1.131 1.134 1.186 1.235 1.295 1.324 1.478 1.527
.086
.058
.046
.027
.045
039
.037
.049
.044
.033
.841
.917
.933
.934
.919
.952
.944
.915
.934
.968
.060
.049
.050
.056
.043
.042
.044
.047
.041
.042
.024
.013
.013
.016
.012
.010
.014
.011
.007
.009
.088
.076
.068
.066
.066
.064
.064
.065
.060
.057
.167
.182
.123
.150 .156 .219 .200 .192 .210 .273
Portfolios for Estimation Period 1958-62

OptsI

............

s ( [p tI ) .........
r(Ray R1") 2 .......

s (R p)............

s(2p) .............
P-t - I ('h) ............
S (e'P) /SXt l(i)

1.013 1.019 1.037 1.048 1.069 1.081 1.092 1.098 1.269 1.388
.065
.048
.038
.045
.045
.038
.031
.036
.033
.036
.886
.922
.93 1 .907
.910
.934
.945
.936
.922
.948
.063
.048
.049
.049
.056
.046
.046
.047
.045
.045
.02 1
.013
.015
.015
.016
.012
.011
.012
.013
.010
.078
.070
.068
.069
.066
.062
.070
.072
.076
.067
.228
.269
.180
.197
.220
.179
.177
.171
.188
.152

for these distributionsis about 1.8 (as compared with a value of 2.0 for a
normal distribution). From Fama and Roll (1968), for values of the characteristic exponent so close to 2.0 stable nonnormal distributions differ
noticeably from the normal only in their extreme tails-that is, beyond
the .05 and .95 fractiles.Thus, as long as one is not concerned with precise estimates of probabilitylevels (always a somewhat meaningless activity), interpretingt-statisticsin the usual way does not lead to serious errors.

This content downloaded from 152.118.24.10 on Mon, 08 Feb 2016 11:52:52 UTC
All use subject to JSTOR Terms and Conditions

C,

C0

s0

In

C)

--

to 0),1oC)

C- CO

CDO0

C)w ))

o2~~~~~~~~~~~oC

!..

...

0_

) t0

00

K:

.....

....

C'- . --t1......
-l In
...In

e b'

1t'

...

00)0)
t/

--1

14 t-)0

... -0

..0)0C0 )h s t.

n ,t 00In

0n

...............-

.o ..........,

00

O~~~~~~~~~~~~~~~~~~~~1
C~lr
CC

-C0

0~
6 ~~~~~~~~~

<>

tLtL

CN

C D oc

in

t_ t.-

Ct.r-

z
<:0

7-1

0)

X~~~~~~~~~~~C
C:
CH CCC

>~

00

C.

In

CD

~C

.o

o,-

C,] t- In
C C) e

-e

) C)

oo

0 0
0)). C

0.000)0,)
0)C00

-0

C0)

_t.00000000

CD C

00

CDOCD00

n00 00

00

000000

0-00-/\3100
'o
C

0) 0

t_ 0 r-)

0)
)

00
0

t
0C 00)0
000

CDV-0
0-I

~
0

D-

00140

"

-0-t C

0 000

g
)t

00
0)

0)

.........
..........

o.....

~C - ~C) ~C

000

000000

IC
00000
0.00

a, !
00

o0

0 ........
h

0~0)0--

0D

00
C

Q-

0.........000

. ... ......

..0

....

0-

000t--/

!C,I -)00!

00 t- I
-0,

.00).0
*I0.00
000

_
000

.0

000000.

000

In 00
1ne

0)
0
0

~~ ~ ~~~~~~C ~~~~~~~~0t
-

tt

0000000

00

C)C'-)

0)0)000)~~~~~~~~~~~~)
0
C0

-t-. C)t

I t , '
0)00~~~~~~~CDI

0)00D )00

_ )

-..

00

<o

[a.;~

00

1t

00)0)

t0

I
I~~~~~~~~~~~~~~~~~~~~~~In
)n
r

)00

CDo CD

Ch I

~~~~.

<0

<:0)

)-.-

z +-

gi

tL 1/

c)

000-0

oo0
C00

)
,t t-D
-C)C)

"

"
C

C) "-t

-0-0
0-0-00"
C

00 00-0
0-0-0-0
000000
0),n
).I
- '0-000)
00

)0
o

00-0

, a, o111, o

o)

R
-

>I)

0>I

o0)
.t0t0
0

..

00
t10
ooCt-0)
oJ-n 0
-0

10
,I- -'n

00
.0
.0.0

1-0

.o

0)
.0
C)
0

t-_0
I-n.

In

.1

00
0)

'

0to00
I-')o.0

00
10

. '1

This content downloaded from 152.118.24.10 on Mon, 08 Feb 2016 11:52:52 UTC
All use subject to JSTOR Terms and Conditions

-lk

))./o~

cr0

0)0)0)

In

> C7

00

O7

or

o
Co.
0

.00
C!

0)-0

<Z"?

-\

) r)00 -

-I

-. oo
--.)
_-

__ord

I00000)

CD

In
4 c 4 co C4
-

1)

00

n"-I

1o~

00

1111

-.0

C ll C
0)O
00C
0

I
O
C

\0

00 0

Ch O O \0
-0000

0)
C,0
_
0
-00-31

:)0

0--0
o

000--0

0-

00

00

.~

00

db-000

0.

l
t 4o ,0O C

.- 0

IC,
0 0

000
00
0-0

00
_0

00000 .

000000
>
00
000--0

>

-)

CD

ON

C0

C>_r Ct

000

000000

0000

0000000

0
0

_ ey O e s
1
0

C0
.C

n CD
C

0)00000
0 000

.0 5
t0

CD

000

u0 e C)t

0
00
0

t
C, )

00 0)0In)NC0
0)
C4
r- I In In ruX
0

..

r~~~t t~

00t
CD0

--

tr)00

ON

00

~ ~

ON

ON

t_

en)0-

..

tri

00 t-o

C_

0-0
D
r00
-00000

00
-

- 0

C>

0
0

0 - 000
0
,0-t-r

CD In

_ t)

C0

000

C0

aC>

00

0)00

00

00000

C,

tC

CD
C,

o00
C)t C

000 "00000
00
, 00
00
CD000
0000-00
Cr

,)- * )t
o t- -) -.)-. o
-0-0C00
00)0)00

~~~~~~~0
C>

..

000000

-000
C

ll
0-0-0)n00o
00io

0r- 000000

00

C 'n0
tr)

In C] 00 Ct0 0. O0
.
0r0)
-0000
ll,

000
-t-

-0
0
00

0-0

,
-0

oo)C0
000

O
_

C0

Oo rooo

4(<

-.o

~~~~~~~~~~~~~~~ll

...........

00
I

- 00000

K2

-)0

,00-- -0_-

000

(,A,

00

t0 0-0,-

0
D000

C-4

r)-0

0o000

C0

00

00--

<b

0
-11

11

0000

CO.

C,~rl

T00

0CD

,I 0

CDo

CD

o C KN~
r _I

1
.

OA

co 1-c

KfIU I~~ N~mx

-

000-.

r)0CD

C0o
)

'O

))).-~
cCE
C~e

00

10
.

- C

00-0
0000

o00

o~o

l-

1<b'.

0)0

CD

,d-

00

Cxco

K,-.-

0)

o--.-o0~~~~~~~t -

0000

Co O

0)

~~~~~~I

0'

t-

C,

00
-

>

\0
rC>

00
0

t-t

aC

000000

)000000
00-00..00
t0

This content downloaded from 152.118.24.10 on Mon, 08 Feb 2016 11:52:52 UTC
All use subject to JSTOR Terms and Conditions

624

JOURNAL

OF POLITICAL

ECONOMY

Inferences
based on approximate
normality
are on even safergroundif
one assumes,againin linewiththe resultsof Officer(1971), thatalthough
distributions
witha -- 1.8,
theyare wellapproximated
by stablenonnormal
ofmonthly
distributions
returns
in facthave finitevariancesand convergebut veryslowly-towardthenormalas one takessumsor averagesof individual returns.Then the distributions
of the means of month-by-month
regression
coefficients
fromthe risk-return
modelare likelyto be close to
normalsinceeach meanis based on coefficients
formanymonths.
A. Tests of the Major Hypothesesof the Two-Parameter
Model
ConsiderfirstconditionC2 of the two-parameter
model,whichsays that
no measureof risk, in addition to I, systematicallyaffectsexpected
returns.This hypothesisis not rejectedby the resultsin panels Cand D
of table 3. The values of t(y3) are small,and the signsof the t(y3) are
randomly
positiveand negative.
Likewise,the resultsin panels B and D of table 3 do not rejectcondition C1 of the two-parameter
bemodel,whichsays that the relationship
tweenexpectedreturnand ( is linear.In panel B, the value of t(y2) for
the overall period 1935-6/68 is only -.29. In the 5-year subperiods,
t(r2) for 1951-55 is approximately
-2.7, but forsubperiodsthat do not
cover1951-55,thevalues of t(^2) are muchcloserto zero.
So far,then,the two-parameter
modelseemsto be standingup well to
thedata. All is fornaught,however,if the criticalconditionC3 is rejected.
That is, we are not happy withthe model unless thereis on averagea
positivetradeoff
betweenriskand return.This seems to be the case. For
theoverallperiod1935-6/68,t(71) is largeforall models.Except forthe
period1956-60,thevalues of t(^y,) are also systematically
positivein the
subperiods,
but not so systematically
large.
The small t-statisticsfor subperiodsreflectthe substantialmonth-tomonthvariabilityof the parametersof the risk-return
regressions.For
example,in the one-variableregressions
summarizedin panel A, for the
period 1935-40, 9- .0109. In otherwords,for this period the average
incremental
returnperunitof ( was almost1.1 percentper month,so that
on average,bearingriskhad substantialrewards.Nevertheless,
because of
thevariability
of ^it-in thisperiods(51) is 11.6 percentper month(!)t(?1) is only.79. It takesthe statisticalpowerof the largesampleforthe
overallperiodbeforevalues of 9i that are large in practicaltermsalso
yieldlarget-values.
But at least withthe sampleof the overallperiodt(Q,) achievesvalues
of theconclusionthaton averagethereis a statistically
supportive
observable positiverelationship
betweenreturnand risk.This is not thecase with
respectto t(92) and t(Y3). Even, or indeed especially,for the overall
period,theset-statistics
are close to zero.

This content downloaded from 152.118.24.10 on Mon, 08 Feb 2016 11:52:52 UTC
All use subject to JSTOR Terms and Conditions

RISK,

RETURN,

AND EQUILIBRIUM

62

The behavior through time of t 9u, and 9:t is also consistent with
The serialcorrelations
ME thatthe capitalmarketis efficient.
hypothesis
pr,,(91), pO(Q)), and p,,(':j), are always low in terms of explanatory power

The proportionof
and generallylow in termsof statisticalsignificance.
serial correlationis estimated
the varianceof 'Yt explainedby first-order
under
by p(7j)2 whichin all cases is small.As forstatisticalsignificance,
is zero, the standarddeviathat the trueserialcorrelation
the hypothesis
tion of the sample coefficientcan be approximated by a(p)

1/ n. For

.05, whileforthe 10- and 5-year

theoverallperiod,a(^) is approximately
subperiodsG(,) is approximately.09 and .13, respectively.Thus, the
values of pjlr(?),

and po(q:j) in table 3 are generally statistically

pf,("y't),

close to zero. The exceptionsinvolveprimarilyperiodsthat include the

1935-40subperiod,and the resultsfortheseperiodsare not independent.8
of the portfolioresiduals,ivt,
To conservespace, the serialcorrelations
negativevalues predominate.
are not shown.In theseserial correlations,
of the9's, thoseof the?'s are close to zero.
But like theserialcorrelations
of the "'s and -l's have been computed,and
serialcorrelations
Higher-order
large.
thesealso are neversystematically
In short,one cannotrejectthe hypothesisthat the pricingof securities
model forexpected
is in line withthe implicationsof the two-parameter
pricingmodel,the behaviorof returns
And givena two-parameter
returns.
capitalmarket.
withan efficient
timeis consistent
through
A

B.

The Behaviorof the Market

periods
Some perspectiveon the behaviorof the marketduringdifferent
and
in
the
of
the
risk-return
coefficients
and on the interpretation
Vit
Yot
can be obtainedfromtable4. For thevariousperiodsof table3,
regressions
table 4 showsthe samplemeans (and withsomeexceptions),the standard
A
about means that are assumed to be zero
serial correlationsof 9.> and
provide a test of the fair game propertyof an efficientmarket,given that expected
)
returnsare generatedby the two-parametermodel-that is, given E(t.,t) =E(y
with respectto the
0. Likewise, P)()ty - Rft) provides a test of market efficiency
behavior of 9" through time, given the validity of the Sharpe-Lintnerhypothesis
computing
(about which we have as yet said nothing). But, at least for Dot and A
the serial correlationsabout sample means produces essentiallythe same results.
To test the marketefficiency
[E(R,1t) - E(1R%)) I, the sample
hypothesison '-t
mean of the ylt is used to estimateE(Rm) - E(Rlf~), thus implicitlyassumingthat
the expected risk premiumis constant. That this is a reasonable approximation [in
the sense that the pAd
(%) are small], probably reflectsthe fact that variation in
variation in l
E(R111 - E(Rf,1) is trivial relative to the month-by-month
Finally, it is well to note that in termsof the implicationsof the serial correlations
for making good portfoliodecisions-and thus for judging whethermarket efficiency
is a workable representationof reality-the fact that the-serial correlationsare low
in termsof explanatorypower is more importantthan whetheror not they are low
in termsof statisticalsignificance.

MThe

This content downloaded from 152.118.24.10 on Mon, 08 Feb 2016 11:52:52 UTC
All use subject to JSTOR Terms and Conditions

626

JOURNAL OF POLITICAL ECONOMY

TABLE 4
THE

BEHAVIOR

OF THE MARKET

STATISTIC*
PERIOD

Rm
.0143

Rm-Rt

Rt

RR_-Rtt
s(Rm)

s(Rm)

s(Rrn)

s(Rm)

.0130

.0085

.0061

.0013

.2 136

.1388

.061

.066

1935-45 .......
1946-55 .
.
1956-6/68

.0197
0112
0121

.0195
.0103
.0095

.0163
.0027
.0062

.0039
.0087
.0060

.0002
.0009
.0026

.2207
.2378
.2387

.1844
.0614
.1560

.089
.043
.040

.098
.041
.044

1935-40 ....... .
1941-45 ....... .
1946-50 ....... .
1951-55 .......
1956-60 .......
1961-6/68

0132
0274
0077
.0148
.0090
.0141

.0132
.0272
.0070
.0136
.0070
.0111

.0024
.0056
.0050
.0123
.0148
.0001

.0001
.0002
.0007
.0012
.0020
.0030

.1221
.4715
.1351
.4174
.2080
.2567

.1009
.3963
.0564
.0735
-.1755
.3294

.108
.058
.052
.033
.034
.043

.116
.069
.047
.035
.034
.048

1935-6/68

.....

.0109
.0229
.0029
.0024
--.7.0059
.0143

* Since s(R,) is so small relative to

which is not shown, is essentially the same
s(R.n) s(Rrn as s(Rl).
The standard deviations of (Rm - R)/s(R?,t)
also not shown, can be
and 9l/s(R.),
obtained directly from s(RnL - R,) s(A ) and s(Rm). Finally, the t-statistics for (Rm, - Rt)s(Rm,,)
and 91/S(Rrl) are identical with those for R,-, -Ra and

deviations, t-statisticsfor sample means, and first-orderserial correlations

for the month-by-month
values of the followingvariables and coefficients:
the market returnR,,,t; the riskless rate of interest Rft, taken to be the
yield on 1-month Treasury bills; Rmt- Rft; (Rmt- Rft)ls(Rn);
got
and "yt, repeated frompanel A of table 3; and Alt/s(R,,,). The t-statistics
on sample means are computed in the same way as those in table 3.
If the two-parametermodel is valid, then in equation (7), E(7st)
E(R0t), where E(Rot) is the expected return on any zero- security or
portfolio.Likewise, the expected risk premiumper unit of is E(Rmt)
E(Rot)
E(ffit). In fact, for the one-variable regressions of panel A,
table 3, that is,

Rpt =yt

Vt PP + %~tl

we have, period by period,

_-Rmt
R tyot-

12)

This condition is obtained by averaging (11) over p and making use of

the least-squares constraint
E

i~pt

0.5

Moreover, the least-squares estimate 9,t can always be interpretedas the

returnfor month t on a zero-[i portfolio,where the weights given to each
'There is some degree of approximationin (12). The averages over p of Rptand
N are R,,,tand 1.0, respectively,only if every securityin the marketis in some portfolio. With our methodology (see table 1) this is never true. But the degree of
approximationturnsout to be small: The average of the Rptis always close to Rm1t
and the average jp is always close to 1.0.

This content downloaded from 152.118.24.10 on Mon, 08 Feb 2016 11:52:52 UTC
All use subject to JSTOR Terms and Conditions

RISK,

RETURN,

AND EQUILIBRIUM

62

TABLE 4 (Continued)
STATISTIC*

s(@(,)

s(R,)

.038

.0012

.052
.026
.030

.0001
.0004
.0009

.064
.034
.031
.019
.020
.034

.0001
.0001
.0003
.0004
.0007
.0008

t(&O)

Pgt)Ps(r

t(R.-Rf)

t(al")

4.71

4.28

2.57

3.24

-.01

-.01

.02

.14

.98

2.56
2.84
3.72

2.54
2.60
2.92

1.92
.70
1.73

.86
3.71
2.45

-.07
.09
.14

-.07
.09
.14

--.03
.07
.15

.10
.10
.2 5

.88
.94
.92

1.04
3.68
1.15
3.51
2.07
3.08

1.04
3.65
1.05
3.22
1.60
2.44

.79
2.55
.48
.53
-1.37
2.81

.32
1.27
1.2 7
5.06
5.68
.03

--.13
.14
.09
-.02
.12
.13

-.13
.14
.09
-.01
.13
.13

-.09
.15
.04
.08
.18
.09

.07
.21
.18
-.07
.13
.21

.72
.83
.97
.89
.80
.93

t(R.)

pm(R,,)

pM(R,,,-Rr)

pm('1

of the 20 portfoliosto formthis zero-3jportfolioare the least-squares

weightsthatare appliedto theRptin computingot.10
In the Sharpe-Lintnertwo-parameter
model of market equilibrium
E(fy-ot)

- E(Rot) Rft and E(7

t)

E(Rmnt)- E(R0t)

E(Rmt)

Rft.In theperiod1935-40 and in themostrecentperiod1961-6/68,yetis

close to Rm- Rf and the t-statistics
forthe two averagesare similar.In
otherperiods,and especiallyin theperiod1951-60,YJis substantially
less
thanRfl- Rf. This is a consequenceof the factthatfortheseperiodsyo
is noticeablygreaterthan Rf. In economicterms,the tradeoff
of average
returnfor risk betweencommonstocks and short-term
bonds has been
moreconsistently
large throughtimethan the tradeoff
of averagereturn
forriskamongcommonstocks.Testingwhetherthe differences
between
R,,- Rf and 9yare statistically
large,however,is equivalentto testing
the S-L hypothesis
Rft,whichwe preferto take up afterexamE(y)t)
iningfurther
thestochasticprocessgenerating
monthlyreturns.
Finally,althoughthedifferences
betweenvaluesof R - Rf fordifferent
periodsor betweenvaluesof yj are neverstatistically
large,thereis a hint
in table 4 thataverage-risk
premiumsdeclinedfromthe pre- to the postXVorldWar II periods.These are average risk premiumsper unit of j,
however,whichare not of primeinterestto the investor.In makinghis
portfoliodecision,the investoris more concernedwith the tradeoffof
expectedportfolioreturnfor dispersionof return-thatis, the slope of
model this slope is
the efficient
set of portfolios.In the Sharpe-Lintner
I0 That 9o is the returnon a zero-8 portfoliocan be shown to follow from the
in the cross-sectionalrisk-returnregresunbiasednessof the least-squarescoefficients
sions. If one makes the Gauss-Markov assumptionsthat the underlyingdisturbances
'1pt of (11) have zero means, are uncorrelatedacross p, and have the same variance
for all p, then it follows almost directlyfrom the Gauss-Markov Theorem that the
least-squaresestimate Y is also the returnfor month t on the minimumvariance
zero-8 portfoliothat can be constructedfrom the 20 portfolioUP.

This content downloaded from 152.118.24.10 on Mon, 08 Feb 2016 11:52:52 UTC
All use subject to JSTOR Terms and Conditions

628

JOURNAL

~~~~~
~~I

OF POLITICAL

ECONOMY

always [E(Rmt) -Rft] 1s(Rmt), and in the moregeneralmodelof Black

(1972), it is [E(Rmt)

E(R(t)]1a(Rmt)

at the point on the efficientset

corresponding
to themarketportfolio
m. In table4, especiallyforthe three
longsubperiods,
dividingR,, - Rf and ^y,by s(R,,) seemsto yield estimated risk premiumsthat are moreconstantthroughtime.This results
fromthe factthatany declinesin 7, or Rm-Rf are matchedby a quite
noticeabledownwardshiftin s(Rm) fromthe early to the later periods
(cf. Blume [1970] or Officer[1971]).
C. Errorsand True Variationin theCoefficients
Vjt
Each cross-sectional
regression
coefficient
^jt in (10) has twocomponents:
the true?it and the estimationerror, -jt
7t. A naturalquestion
Ajt is: To whatextentis the variationin 9jt throughtimedue to variationin
In additionto providingimportant
yjt and to whatextentis it due to Aft?
information
about theprecisionof thecoefficient
estimatesused to testthe
two-parameter
model, the answerto this question can be used to test
hypothesesabout the stochasticprocessgeneratingreturns.For example,
althoughwe cannotrejectthe hypothesisthat E(7 t)
0, does including
the terminvolvingj%2 in (10) help in explainingthe month-by-month
behaviorof returns?That is, can we rejectthe hypothesisthat forall t,
0? Likewise,can we reject the hypothesisthat month-by-month
Y2t
timein 9ot due entirelyto bi)t and
'It - 0? And is the variationthrough
to variationin Rft?
The answersto thesequestionsare in table 5. For the modelsand time
periodsof table 3, table 5 showsforeach 7j: S2(j ^), the samplevariance
_

of the month-by-month9jt; s2(j),

the average of the month-by-month

values of S2'(t),
wheres('jt) is the standarderrorof ?jt fromthe crosssectionalrisk-return
regressionof (10) for montht; S2(1)
S2(j)
s2(j);
and theF-statisticF -2 (sIj) IS2(+),
whichis relevantfortesting
the hypothesis,
of F has n - 1 df where
s2('1(). The numerator
S2(^j)
n is thenumberof monthsin the sampleperiod; and the denominator
has
in
model."
n(20 - K) df,whereK is the numberof coefficients
the
9j

11The standard error of t/jt' S( 't), is proportionalto the standard error of the
risk-return
for month t, which has 20 - K df. And n values of S2('1t)
residuals,
are averaged to get s2(-,),
so that the latter has n(20-K)
df. Note that if the
underlyingreturndisturbancesapt of (10) are independentacross p and have identical
normal distributionsfor all p, then "it is the sample mean of a normal distribution
and s2("jt)
is proportionalto the sample variance of the same normal distribution.
If the process is also assumed to be stationary through time, it then follows that
S2(9jt) and s2($'t) are independent,as required by the F-test. Finally, in the Fstatisticsof table 5, the values of n are 60 or larger,so that, since K is from 2 to 4,
n(20 - K) >3 960. From Mood and Graybill (1963), some upper percentage points
of the F-distributionare:

This content downloaded from 152.118.24.10 on Mon, 08 Feb 2016 11:52:52 UTC
All use subject to JSTOR Terms and Conditions

RISK,

RETURN,

AND EQUILIBRIUM

629

One clear-cutresultin table 5 is that thereis a substantialdeclinein

thereliability
of thecoefficients
y,,tand y9lt-thatis, a substantialincrease

in s'())

and s2($ )-when

and/or

jp2

are included in the risk-

p('hj)

returnregressions.
The variablej%2 is obviouslycollinearwithjce and, as
can be seen fromtable 2,
likewise increases with -. Frompanels B
and C of table 5, the collinearitywith ^,p is strongerfor Ap2 than for
Sp(ti)

In spiteof the loss in precisionthat arisesfrommulticollinearity,

however,theF-statisticsfor y,,(the coefficient
of j2) and 9;: [the coefficient
of -Sp(,)] are generallylargeforthe modelsof panels B and C of table 5,
and forthe modelof panel D whichincludesboth variables.From the Fstatisticsin panel D, it seems that,except for the period 1935-45, the
variationthroughtimeof
is statistically
morenoticeablethan that of
but
there
are
periods(1941-45, 1956-60) when the values of F for
'tY,
both cot and t are large.
The F-statistics for ^It -

ylt + b t also indicate that 7 t has substantial variationthroughtime.This is not surprising,

however,since it is
alwaysdirectlyrelatedto R,,W.For example,fromequation (12), forthe

one-variable model of panel A, )i

t-

iRf -

Finally,the F-statisticsfor AOt 'ot + rot are also in generallarge.

And the month-by-month
variationin ,yt cannot be accountedfor by
variationin Rft.The varianceof Rftis so smallrelativeto S2(,t),
S (y(t),
and s2($,}t) thatdoingthe F-testsin termsof )t- Rftproducesresults
almostidenticalwiththosefor'%'t.
Rejectionof thehypothesis
that7ot-Rft0 does not implyrejection
of the S-L hypothesis-tobe testednext-that E(,fft) Rft.Likewise,
to findthatmonth-by-month
y2t # 0 and 'Yfitz 0 does not implyrejection
of hypothesesC1 and C2 of the two-parameter
model.These hypotheses,
whichwe are unable to rejecton the basis of the resultsin table 3, say
that E(y.t)

0 and E(Q:it)

0.

What we have foundin table 5 is that thereare variablesin addition

to fpthat systematically
affectperiod-by-period
returns.Some of these
omittedvariablesare apparentlyrelatedto Spj and 9 (-f). But the latter
are almostsurelyproxies,since thereis no economicrationalefor their
presencein our stochasticrisk-return
model.

F
60
60
120
120

(120) ............
.1.29
()
.
(120)
(oo) ..............

1.35
1.26
1.19

.
....F

1.47
1.39
1.35
1.25

F.
1.58
1.48
1.43
1.31

F,)
1.73
1.60
1.53
1.38

This content downloaded from 152.118.24.10 on Mon, 08 Feb 2016 11:52:52 UTC
All use subject to JSTOR Terms and Conditions

F,()q
1.83
1.69
1.61
1.43

630

JOURNAL

OF

POLITICAL

ECONOMY

TABLE 5
COMPONENTS

PERIOD

Panel A:
1935-6/68

S2(.y)

S2(90)

OF THE X ARIANCES OF THE 9jt

S2(o)

s2(a)

s2(51)

.00105

.00142

.00037

3.84

.00401

.00436

.00035

12.46

1935-45 .
1946-55
1956-6/68

.00182
.00057
.0007 7

.00273
.00066
.00090

.00091
.00009
.00013

3.00
7.33
6.92

.00863
.00163
.00181

.00950
.00171
.00193

.00087
.00008
.00012

10.92
21.38
16.08

1935-40 .....
1941-45 .....
1946-50 .....
1951-55 .....
1956-60 .....
1961-6/68

.00265
.00086
.00086
.00027
.00032
.00100

.00404
.00118
.00094
.00036
.00041
.00114

.00139
.00032
.00008
.00009
.00009
.00014

2.91
3.69
11.75
4.00
4.56
8.14

.01212
.00452
.00216
.00113
.00104
.00217

.01347
.00481
.00224
.00121
.00112
.00231

.00135
.00029
.00008
.00008
.00008
.00014

9.98
16.59
28.00
15.12
21.50
16.50

Panel B:
1935-6/68

...

.00092

.00267

.00175

1.52

.00564

.01403

.00839

1.67

.0005
7
00053
.00155

.003 77
.00112
.00294

.00320
.00059
.00139

1.18
1.90
2.12

.00372
.00651
.00667

.01941
.00897
.01338

.01569
.00245
.00671

1.24
3.66
1.99

1935-40
..
00018
1941-45 .....
.00101
1946-50 .....
.00084
1951-55 .....
.00024
1956-60 .....
.00037
.00232
1961-6/68

.00476
.00254
.00136
.00090
.00087
.00431

.00458
.00153
.00052
.00066
.00050
.00199

1.04
1.66
2.62
1.36
1.74
2.16

.00374
.00389
.00862
.00447
.00289
.00928

.02555
.01225
.01071
.00729
.005 17
.01894

.02181
.00836
.00209
.00282
.00228
.00966

1.17
1.46
5.12
2.58
2.2 7
1.96

.00192

.00266

.00075

3.55

.00285

.00428

.00142

3.01

00394
00083
00100

.00533
.00101
.00164

.00139
.00018
.00063

3.83
5.61
2.60

.00433
.00261
.00178

.00717
.00310
.00270

.00283
.00050
.00092

2.52
6.20
2.93

.00473
.00307
.00103
.00061
.00079
.00109

.00669
.00377
.00117
.00083
.00134
.00177

.00196
.00070
.00014
.00022
.00055
.00068

3.41
5.38
8.36
3.77
2.44
2.60

.00732
.00085
.00386
.00140
.00106
.00212

.01094
.002 74
.00439
.00188
.00204
.00300

.00362
.00189
.00053
.00047
.00098
.00088

3.02
1.45
8.28
4.00
2.08
3.41

...

1935-45 .
1946-55 .
1956-6/68

Panel C:
1935-6/68

...

1935-45 .
1946-55 .
1956-6/68 .
1935-40 .....
1941-45 .....
1946-50 .....
1951-55 .....
.
1956-60 .
1961-6/68
Panel D:
1935-6/68
1935-45 .
1946-55 .
1956-6/68
1935-40 .....
1941-45 .....
1946-50 .....
1951-55 .....
1956-60 .....
1961-6/68

D.

...

.00150

.00566

.00406

1.39

.00608

.01521

.00913

1.66

.00233
.00013
.00194

.01065
.00176
.00420

.00832
.00163
.00226

1.28
1.08
1.86

.00402
.00647
.00763

.02118
.00916
.01485

.01716
.00269
.00722

1.23
3.41
2.06

. .00157
.00340
.00023
.00006
.00092
.00260

.01263
.00843
.00220
.00136
.00239
.00539

.01106
.00503
.00197
.00130
.00147
.00279

1.14
1.68
1.12
1.05
1.62
1.93

.00457
.00365
.00858
.00442
.00328
.01060

.02910
.01196
.01119
.00719
.00602
.02081

.02453
.00832
.00261
.00277
.002 74
.01021

1.19
1.44
4.29
2.60
2.20
2.04

Tests of the S-L Hypothesis

In the Sharpe-Lintner two-parametermodel of market equilibrium one

has, in addition to conditions C1-C3, the hypothesis that E('tt)
Rft.
The work of Friend and Blume (1970) and Black, Jensen, and Scholes
(1972) suggests that the S-L hypothesis is not upheld by the data. At
least in the post-World War II period, estimates of E('Y/,t) seem to be
significantlygreater than Rft.
Each of the fourmodels of table 3 can be used to test the S-L hypothe-

This content downloaded from 152.118.24.10 on Mon, 08 Feb 2016 11:52:52 UTC
All use subject to JSTOR Terms and Conditions

RISK,

AND EQUILIBRIUM

RETURN,

631I

TABLE 5 (Continued)

~ ~

S2(y.,)

..

...

PERIOD

s2

s2(9A )

S2(-3

s2(93

S2(

Panel A:

19 456-50
19356-60

. .
...

....

19561-6/68

...

...

...

1935-6/6 .....
.002.038.017
1945-45 ..
0071
.05
..0077
1946-55 ..
0063
.012.004.229..
1956-6/68 ...
.002.028.016
0041
19356-40..
.054..0 .056
.00066

1951-55.....00058
1956-6068....00033
1961-6/6 ....00018
19315-66

...

1946-5 5 .
1956-6/6

..

.0006
.005
.00033

1935-408

.008

1935-450..

1941-456

50 ...
1951-55 .
1956-606

1961-6/6

..
.

....

1941-45 ..
1946-50 ....
195 1-55.....00039
1956-60.....00037
1961-6/68 ...00202

...

...

16

...

...

..

.45

.. .

17

...

...
...
...

...

...

..

..

...

..

..
..

..
..
..

...

.001120
.000783

.00062
.00050

1.294
1.66

...

.00410

.005227

1.81

...

...

.341

.753

.41.

18

.3
.16
.304

87
30
.968

33
.20

.66.

27
1.80
14

.553

.282

1.96

.3136

1.871

.000 2..62
.0003

2.78
19

.006.

.05

...

1..66

.027 ..20 ...

040

1
.
.753.840

...

. ..

..

.0006

.
..

2.78

...
. ..
...

00162.
00083

...
...

2.62.

...

.. .
....

...
.

.006 .030.121 032

19365-66 ....
1935-455 . ..
...
.002.004.9
1946-556060.04
.0030
19561-6/68 ..0134
1935-408

...

...

...

193546

.
...

.0020

.0010
.0012
.083

. .
...

.00037

.002

.....037

. .
...

.003103

1941-45603268057

194566-50

. .
...

..

..

...
.

..

.118

.2175

...

.....

..
.

3.4

1.18839
.2547
.437

.622

1.355

.140

1.722

..

.276
.734

1.79
1.85

.617

1.417

.276

.864

.588

1.47

.0
18

.321
.028
.374

1.001
1.383
1.1252

.6136
.7355
.7517

1.63
1.08
1.50

...

008
.006.

341

...

...

.007623
.00515
.001480

.00886

1.821

.1206

.8682

.562

1.21

.00353
.00096

1.46
1.870

.7920
.023

1.395
.3483

.675
.325

2.07
1.07

.00116
.00103
.00440

.00077
.00066
.00238

1.51
1.56
1.85

.038
.712
.163

.424
1.654
.78 7

.386
.941
.624

1.10
1.76
1.26

sis.12 The most efficienttests, however, are provided by the one-variable

12 The least-squaresintercepts
%~)tin the four cross-sectionalrisk-returnregressions
can always be interpretedas returnsfor month t on zero-A_
portfolios(n. 10). For the
three-variablemodel of panel D, table 3, the unbiasedness of the least-squares cocan be shown to imply that in computing7,t negative and positive weights
efficients
are assigned to the 20 portfoliosin such a way that the resultingportfoliohas not
2 and of the 20 _S(~) Analogous
only zero-j~ but also zero averages of the 20
statementsapply to the two-variable models of panels B and C.
Black, Jensen,and Scholes test the AS-L hypothesiswith a time series of monthly
returnson a "minimumvariance zero-O portfolio"which they derive directly.It turns

This content downloaded from 152.118.24.10 on Mon, 08 Feb 2016 11:52:52 UTC
All use subject to JSTOR Terms and Conditions

632

JOURNAL

OF POLITICAL

ECONOMY

model of panel A, since the values of s(^9,) for this model [which are
nearly identical with the values of s(,, - Rif)] are substantially smaller
than those forother models. Except for the most recent period 1961-6/68,
the values of 9y,, Rf in panel A are all positive and generallygreater than
0.4 percent per month. The value of t(9o - Rf) for the overall period
1935-6/68 is 2.55, and the t-statisticsfor the subperiods 1946-55, 195155, and 1956-60 are likewise large. Thus, the results in panel A, table 3,
support the negative conclusions of Friend and Blume (1970) and Black,
Jensen, and Scholes (1972) with respect to the S-L hypothesis.
The S-L hypothesis seems to do somewhat better in the two-variable
quadratic model of panel B, table 3 and especially in the three-variable
model of panel D. The values of t(, - Rf) are substantially closer to
zero for these models than for the model of panel A. This is due to values
of 0 - Rf that are closer to zero, but it also reflectsthe fact that s(90)
is substantially higher for the models of panels B and D than for the
model of panel A.
But the effectsof jp2 and sP(^) on tests of the S-L hypothesis are in
fact not at all so clear-cut. Consider the model
Rit -

'Ot +

'1tfi + y2t(1 -

(i)2

3tSi +

(13)

Nit-

Equations (7) and (13) are equivalent representationsof the stochastic

process generating returns, with y

1t -

2` t and yot -

y'ot +

Y2t.

Moreover, if the steps used to obtain the regressionequation (10) from

the stochastic model (7) are applied to (13), we get the regressionequation,
Rpt -_*ot +

2'*Jp

22t(l

jp)2

y3tSP(Ei)

%pt,

(14)

jp2 in (10) is the average of ^P-2 for securities i in portfolio

is the average of (1 2. The values of the estimates
and
are
identical
in
and
(14); in addition, ^it - Q'lt - 22t
art
(10)
y3t
and yet ot + %.t But although the regression equations (10) and
(14) are statistically indistinguishable,tests of the hypothesis E('ot)

where,just as

p, (1

)2

out, however,that this portfoliois constructedunder what amountsto the assumptions

of the Gauss-Markov Theorem on the underlyingdisturbancesof the one-variable
risk-returnregression(11). With these assumptions the least-squares estimate yot,
obtained from the cross-sectionalrisk-returnregressionof (11) for month t, is precisely the returnfor month t on the minimumvariance zero4j portfoliothat can be
constructedfromthe 20 portfolioOP. Thus, the tests of the S-L hypothesisin panel A
of table 3 are conceptuallythe same as those of Black, Jensen,and Scholes.
If one makes the assumptions of the Gauss-Markov Theorem on the underlying
disturbancesof the models of panels B-D of table 3, the regressioninterceptsfor these
models can likewise be interpretedas returnson minimum-variancezero-13portfolios.
These portfoliosthen differin termsof whetheror not they also constrainthe averages
of the 20 > 2 and of the 20 S (1i) to be zero. Given the collinearityof 0pI Ag2, and
Sp('d), however, the assumptionsof the Gauss-Markov Theorem cannot apply to all
four of the models.

This content downloaded from 152.118.24.10 on Mon, 08 Feb 2016 11:52:52 UTC
All use subject to JSTOR Terms and Conditions

RISK,

RETURN,

633

AND EQUILIBRIUM

Rft from (10) do not yield the same results as tests of the hypothesis
E(^yt) - Rft from (14). In panel D of table 3,
-Rf is never statistically very differentfromzero, whereas in tests (not shown) from (14), the
resultsare similar to those of panel A, table 3. That is, y -Rf is systematically positive for all periods but 1961-6/68 and statistically very
differentfromzero for the overall period 1935-6/68 and for the 1946-55,
1951-55, and 1956-60 subperiods.
Thus, tests of the S-L hypothesis from our three-variablemodels are
ambiguous. Perhaps the ambiguity could be resolved and more efficient
tests of the hypothesiscould be obtained if the omitted variables for which
are almost surely proxies were identified.As indisp('i), hp. or (1-Up)2
cated above, however, at the moment the most efficienttests of the S-L
hypothesisare provided by the one-variable model of panel A, table 3, and
the results for that model support the negative conclusions of others.
Given that the S-L hypothesisis not supported by the data, tests of the
market efficiencyhypothesisthat ot- E(Rot) is a fair game are difficult
since we no longer have a specific hypothesis about E(R,,t). And using
the mean of the ant as an estimate of E(Rt) does not work as well in this
case as it does for the market efficiencytests on Yit. One should note,
however, that although the serial correlationspm(^o)
in table 4 are often
large relative to estimates of theirstandard errors,they are small in terms
of the proportionof the time series variance of Q(,tthat they explain, and
the latter is the more important criterion for judging whether market
efficiencyis a workable representationof reality (see n. 8).

VI.

Conclusions

In sum our results support the importanttestable implicationsof the twoparameter model. Given that the market portfolio is efficient or, more
specifically,given that our proxy for the market portfoliois at least approximatelyefficient we cannot reject the hypothesisthat average returns
on New York Stock Exchange common stocks reflectthe attempts of riskaverse investors to hold efficientportfolios. Specifically,on average there
seems to be a positive tradeoffbetween return and risk, with risk measured from the portfolioviewpoint. In addition, although there are "stochastic nonlinearities" from period to period, we cannot reject the hypothesis that on average their effectsare zero and unpredictably different
from zero from one period to the next. Thus, we cannot reject the hypothesis that in making a portfolio decision, an investor should assume
that the relationshipbetween a security's portfolio risk and its expected
returnis linear, as implied by the two-parametermodel. We also cannot
reject the hypothesisof the two-parametermodel that no measure of risk,
in addition to portfoliorisk,systematicallyaffectsaverage returns.Finally,
the observed fair game properties of the coefficientsand residuals of the

This content downloaded from 152.118.24.10 on Mon, 08 Feb 2016 11:52:52 UTC
All use subject to JSTOR Terms and Conditions

634

JOURNAL

OF POLITICAL

ECONOMY

risk-returnregressions are consistent with an efficientcapital market

that is, a market where prices of securities fully reflectavailable information.
Appendix
Some Related Issues
Al.

MarketModels and Tests of MarketEfficiency

The time series of regressioncoefficients

from (10) are, of course,the inputs
for the tests of the two-parameter
can also be
model. But these coefficients
usefulin tests of capital marketefficiency-that
is, tests of the speed of price
adjustmentto different
typesof new information.
Since the workof Fama et al.
(1969), suchtestshave commonlybeen based on the "one-factormarketmodel":
Rit= -a+

(15)

Rmt +'it

In this regressionequation,the terminvolvingR,,t is assumed to capture the

effectsof market-widefactors.The effectson returnsof events specificto
companyi, like a stock split or a changein earnings,are then studiedthrough
the residualsfit.
But giventhatthereis period-to-period
variationin iOt. YAt, and %t in (10)
that is above and beyond pure samplingerror,then these coefficients
can be
interpreted
as marketfactors,(in addition to R,,t) that influencethe returns
on all securities.To see this, substitute(12) into (11) to obtain the "twofactormarketmodel":
R Pt

1_

Up

Rmt+

16

pt.

In like fashion,from equation (10) itself we easily obtain the "four-factor

marketmodel":
Rpt-

(t( 1

Upp)

-pR1(Ipt
+

(%

(2p p -

j')

9'3t
[Spfi)

p(i]

apt,

(17)
where2 and s
are theaveragesoverp of thej%' and thesp('i).
Comparingequations (15-17) it is clear that the residuals tit from the
one-factormarketmodel containvariationin the marketfactors%,t 9t, and
9:t
Thus, if one is interestedin the effecton a security'sreturnof an event
specificto thegivencompany,this effectcan probablybe studiedmoreprecisely
fromthe residualsof the two- or even the four-factor
marketmodels of (16)
and ('17) than from the one-factormodel of (15). This has in fact already
been done in a study of changes in accountingtechniquesby Ball (1972), in
a study of insider tradingby Jaffe( 1972), and in a study of mergersby
Mandelker(1972).
Ball, Jaffe,and Mandelker use the two-factorrather than the four-factor
marketmodel, and there is probably some basis for this. First, one can see
fromtable 5 that because of the collinearityof (3p, (3p', and Sp(j;), thecoeicient estimates ,, and %/
t have much smaller standard errors in the twofactor model. Second, we have computedresidual variances for each of our
20 portfoliosfor various timeperiods fromthe time series of ipt and apt from
(15), ('16), and (173. The decline in residual variance that is obtained in

This content downloaded from 152.118.24.10 on Mon, 08 Feb 2016 11:52:52 UTC
All use subject to JSTOR Terms and Conditions

RISK,

RETURN,

AND EQUILIBRIUM

635

goingfrom(15) to (16) is as predicted:That is, the declineis noticeableover

moreor less the entirerangeof Upand it is proportionalto (1 - jp)2. On the
other hand, in going from the two- to the four-factormodel, reductionsin
residualvarianceare generallynoticeableonly in the portfolioswith the lowest
and highestj%,and the reductionsfor these two portfoliosare generallysmall.
Moreover,includingSp(tj) as an explanatoryvariable in additionto p and j 2
neverresultsin a noticeablereductionin residualvariances.
A2. MultifactorModels and Errorsin the j
If the return-generating
process is a multifactormarketmodel, then the usual
estimatesof Pi fromthe one-factormodel of (15) are not most efficient.
For
example, if the return-generating
process is the population analog of (16),
moreefficient
estimatesof Picould in principlebe obtainedfroma constrained
regressionapplied to
+ Cut
Rit -ot =Pi(Rmt -ot)
But this approach requires the time series of the true yot. All we have are
estimates9n, themselvesobtained fromestimatesof (p fromthe one-factor
model of (15).
It can also be shown that with a multifactorreturn-generating
process the
errorsin the computedfromthe one-factormarketmodel of (8) and (15)
are correlatedacross securitiesand portfolios.This resultsfromthe fact that if
the true process is a multifactormodel, the disturbancesof the one-factor
model are correlatedacross securitiesand portfolios.Moreover, the interdependenceof the errorsin the is higherthe fartherthe true (3's are from
1.0. This was already noted in the discussionof table 2 wherewe found that
the relativereductionin the standarderrorsof the 3's obtainedby usingportfoliosratherthan individualsecuritiesis lower the fartherfA1p
is from 1.0.
Interdependence
of the errorsin the A~ also complicatesthe formalanalysis
of the effectsof errors-in-the-variables
on propertiesof the estimatedcoefficients (the 7jt) in the risk-return
regressionsof (10). This topic is considered
in detail in an appendixto an earlierversionof this paper that can be made
available to the readeron request.

References
Ball, R. "Changes in AccountingTechniquesand Stock Prices." Ph.D. dissertation,Universityof Chicago, 1972.
Black, F. "Capital Market EquilibriumwithRestrictedBorrowing."J. Bus. 45
(July 1972): 444-55.
Black, F.; Jensen,M.; Scholes, M. "The Capital Asset PricingModel: Some
EmpiricalResults." In Studies in the Theoryof Capital Markets,edited by
Michael Jensen.New York: Praeger,1972.
Blume,M. E. "PortfolioTheory: A Step towardIts Practical Application."J.
Bus. 43 (April 1970): 152-73.
Douglas, G. W. "Risk in the Equity Markets: An Empirical Appraisal of
MarketEfficiency."
Yale Econ. Essays 9 (Spring1969): 3-45.
Fama, E. F. "The Behaviorof Stock MarketPrices." J. Bus. 38 (January1965):
34-105. (a)
. "PortfolioAnalysisin a Stable Paretian Market." ManagementSci. 11
(January1965): 404-19. (b)

This content downloaded from 152.118.24.10 on Mon, 08 Feb 2016 11:52:52 UTC
All use subject to JSTOR Terms and Conditions

636

JOURNAL

OF POLITICAL

ECONOMY

- "MultiperiodConsumption-Investment
Decisions." A.E.R. 60 (March
1970): 163-74. (a)
. "Efficient
Capital Markets: A Review of Theoryand EmpiricalWork."
J. Finance 25 (May 1970): 383-417. (b)
. "Risk, Returnand Equilibrium."J.P.E. 79 (January/February
1971):
30-55.
Fama, E. F., and Babiak, H. "Dividend Policy: An Empirical Analysis." J.
AmericanStatis. Assoc. 48 (December 1968): 1132-61.
Fama, E. F.; Fisher,L.; Jensen,M.; Roll, R. "The Adjustmentof Stock Prices
to New Information."Internat.Econ. Rev. 10 (February 1969): 1-21.
Fama, E. F., and Miller,M. The Theoryof Finance. New York: Holt, Rinehart
& Winston,1972.
Fama, E. F., and Roll, R. "Some Propertiesof SymmetricStable Distributions."
J. AmericanStatis.Assoc. 48 (September1968): 817-36.
Fisher,L. "Some New Stock MarketIndexes." J. Bus. 39 (January1966): 191225.
Friend,I., and Blume, M. "Measurementof PortfolioPerformanceunder Uncertainty."A.E.R. 60 (September1970): 561-75.
Gonedes,N. J. "Evidence on the InformationContentof AccountingNumbers:
Accounting-basedand Market-based Estimates of Systematic Risk." J.
Financial and QuantitativeAnalysis(1973): in press.
Jaffe,
J. "SecurityRegulation,Special Information,
and InsiderTrading."Ph.D.
dissertation,
Universityof Chicago, 1972.
Jensen,M. "The Foundationsand CurrentState of Capital Market Theory."
Bell J. Econ. and ManagementSci. 3 (Autumn1972): 357-98.
Lintner,J. "The Valuation of Risk Assets and the Selection of Risky Investmentsin Stock Portfoliosand Capital Budgets." Rev. Econ. and Statis. 47
(February 1965): 13-37.
Mandelker,G. "Returnsto StockholdersfromMergers."Ph.D. proposal,Universityof Chicago, 1972.
Markowitz,H. Portfolio Selection: EfficientDiversificationof Investments.
New York: Wiley, 1959.
Miller,K. D. "FuturesTradingand InvestorReturns:An Investigationof CommodityMarket Risk Premiums."Ph.D. dissertation,Universityof Chicago,
1971.
Miller,M., and Scholes,M. "Rates of Returnin Relationto Risk: A Re-Examinationof Some Recent Findings."In Studies in the Theoryof Capital Markets,editedby Michael Jensen.New York: Praeger,1972.
Mood, A. M., and Graybill,F. A. Introductionto the Theory of Statistics.
New York: McGraw-Hill,1963.
Officer,
R. R. "A Time Series Examinationof the Market Factor of the New
York Stock Exchange." Ph.D..dissertation,Universityof Chicago, 1971.
Roll, R. The Behaviorof InterestRates. New York: Basic, 1970.
Sharpe,W. F. "Capital Asset Prices: A Theory of Market Equilibriumunder
Conditionsof Risk." J. Finance 19 (September1964): 425-42.
Tobin, J. "LiquidityPreferenceas Behavior towardsRisk." Rev. Econ. Studies
25 (February1958): 65-86.

This content downloaded from 152.118.24.10 on Mon, 08 Feb 2016 11:52:52 UTC
All use subject to JSTOR Terms and Conditions