A_test_of_the_efficiency_of_a_given_portfolio_Gibbons_M_R__Ross_S_A_and_Shanken_JA_test_of_the_efficiency_of_a_given_portfolio_Gibbons_M_R__Ross_S_A_and_Shanken_JA_test_of_the_efficiency_of_a_given_portfolio_Gibbons_M_R__Ross_S_A_and_Shanken_JA_test_of_the_efficiency_of_a_given_portfolio_Gibbons_M_R__Ross_S_A_and_Shanken_JA_test_of_the_efficiency_of_a_given_portfolio_Gibbons_M_R__Ross_S_A_and_Shanken_JA_test_of_the_efficiency_of_a_given_portfolio_Gibbons_M_R__Ross_S_A_and_Shanken_JA_test_of_the_efficiency_of_a_given_portfolio_Gibbons_M_R__Ross_S_A_and_Shanken_JA_test_of_the_efficiency_of_a_given_portfolio_Gibbons_M_R__Ross_S_A_and_Shanken_JA_test_of_the_efficiency_of_a_given_portfolio_Gibbons_M_R__Ross_S_A_and_Shanken_JA_test_of_the_efficiency_of_a_given_portfolio_Gibbons_M_R__Ross_S_A_and_Shanken_J

© All Rights Reserved

34 views

A_test_of_the_efficiency_of_a_given_portfolio_Gibbons_M_R__Ross_S_A_and_Shanken_JA_test_of_the_efficiency_of_a_given_portfolio_Gibbons_M_R__Ross_S_A_and_Shanken_JA_test_of_the_efficiency_of_a_given_portfolio_Gibbons_M_R__Ross_S_A_and_Shanken_JA_test_of_the_efficiency_of_a_given_portfolio_Gibbons_M_R__Ross_S_A_and_Shanken_JA_test_of_the_efficiency_of_a_given_portfolio_Gibbons_M_R__Ross_S_A_and_Shanken_JA_test_of_the_efficiency_of_a_given_portfolio_Gibbons_M_R__Ross_S_A_and_Shanken_JA_test_of_the_efficiency_of_a_given_portfolio_Gibbons_M_R__Ross_S_A_and_Shanken_JA_test_of_the_efficiency_of_a_given_portfolio_Gibbons_M_R__Ross_S_A_and_Shanken_JA_test_of_the_efficiency_of_a_given_portfolio_Gibbons_M_R__Ross_S_A_and_Shanken_JA_test_of_the_efficiency_of_a_given_portfolio_Gibbons_M_R__Ross_S_A_and_Shanken_J

© All Rights Reserved

- Risk & Return
- Market Efficiency: Finance week 09
- mat133
- Multinational Business Finance by Shapiro Chapter 1 Solutions
- Practice Problems for FINN3222_Midterm_Fall2011
- Agriculture In India
- Appendix_W_-_Fama_French_Model.pdf
- Question Bank.docx for Mba
- Credit Ratings and The Cross-Section of Stock Returns
- Wind Forecasting Using Artificial Neural Networks
- Investment Management
- MCQ's for Midterm Test -7th April 2012
- 3. Finance - Ijfm - Empirical Testing of Capm in - Seethanaik
- Report
- Assignment 2
- FM 8th Edition Chapter 12 - Risk and Return
- Artigo 2 - r2 Baixo
- 318072
- Assuring Not-For-Profit Hospital cost of capital.pdf
- 4_Standard Valuation Models (1)

You are on page 1of 33

Source: Econometrica, Vol. 57, No. 5 (Sep., 1989), pp. 1121-1152

Published by: The Econometric Society

Stable URL: http://www.jstor.org/stable/1913625 .

Accessed: 26/11/2013 04:41

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .

http://www.jstor.org/page/info/about/policies/terms.jsp

.

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of

content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms

of scholarship. For more information about JSTOR, please contact support@jstor.org.

The Econometric Society is collaborating with JSTOR to digitize, preserve and extend access to Econometrica.

http://www.jstor.org

All use subject to JSTOR Terms and Conditions

BY MICHAEL R. GIBBONS, STEPHEN A. Ross, AND JAY SHANKEN1

A test for the ex ante efficiency of a given portfolio of assets is analyzed. The relevant

statistic has a tractable small sample distribution. Its power function is derived and used to

study the sensitivity of the test to the portfolio choice and to the number of assets used to

determine the ex post mean-variance efficient frontier.

Several intuitive interpretations of the test are provided, including a simple mean-standard deviation geometric explanation. A univariate test, equivalent to our multivariate-based

method, is derived, and it suggests some useful diagnostic tools which may explain why the

null hypothesis is rejected.

Empirical examples suggest that the multivariate approach can lead to more appropriate

conclusions than those based on traditional inference which relies on a set of dependent

univariate statistics.

KEYWORDS:

Asset pricing, CAPM, multivariate test, portfolio efficiency.

1. INTRODUCTION

The modern theory of finance has always been rooted in empirical analysis.

The mean-variance capital asset pricing model (CAPM) developed by Sharpe

(1964) and Lintner (1965) has been studied and tested in more papers than can

possibly be attributed here. This is only natural; the quality and quantity of

financial data, especially stock market price series, are the envy of other fields in

economics.

The theory is generally expressed in terms of its first-order conditions on the

risk premium. Expected returns on assets are linearly related to the regression

coefficients, or betas, of the asset returns on some index of market returns. In

other words, risk premiums in equilibrium depend on betas. The standard tests of

the CAPM are based on regression techniques with various adaptations. For

some notable examples, see Black, Jensen, and Scholes (1972) and Fama and

MacBeth (1973). Usually, cross-sectional regressions are run of asset returns on

estimated beta coefficients, and estimates of the slope are reported. Often the

data are grouped to reduce measurement errors, and sometimes the estimation is

done at a sequence of time points to create a time series of estimates from which

the precision of the overall average can be determined.

Roll (1977, 1978), among others, has raised serious doubts whether these

procedures are, in fact, tests of the CAPM. Insofar as proxies are used for the

market portfolio, the Sharpe-Lintner theory is not being tested. Furthermore, as

Roll emphasizes, the regression tests are probably of quite low power, and

1

We are grateful to Ted Anderson, Fischer Black, Douglas Breeden, Michael Brennan, Gary

Chamberlain, Dave Jobson, Allan Kleidon, Bruce Lehmann, Paul Pfleiderer, Richard Roll, and two

anonymous referees as well as the seminar participants at Duke University, Harvard University,

Indiana University, Stanford University, University of California at San Diego, University of Illinois

at Urbana, and Yale University for helpful comments. We appreciate the research assistance of Ajay

Dravid, Jung-Jin Lee, and Tong-sheng Sun. Financial support was provided in part by the National

Science Foundation and the Stanford Program in Finance. This paper supersedes an earlier paper

with the same title by Stephen Ross.

1121

All use subject to JSTOR Terms and Conditions

1122

grouping may lower the power further.These objections leave the empirical

testing of the CAPM in an odd state of limbo. If the proxy is not a valid

surrogate,then as tests of the CAPM the existing empiricalinvestigationsare

somewhatbeside the point.2 On the other hand, if the proxy is valid, then the

small sample distributionand powerof the tests are unknown.

This is unfortunateand indicativeof a missedopportunity.The CAPMis one

of many financial theorieswhich suggest quite specific hypothesescouched in

terms of observables.The rich data availablefor testing these hypothesesare an

incentive to develop tests which are explicitlydirectedat them. In this paperwe

develop a canonicalexampleof such a test usingmultivariatestatisticalmethods.

The problem we consideris the central one addressedin tests of the CAPM.

Since the theory is equivalent to the assertion that the market portfolio is

mean-varianceefficient, we wish to test whether any particularportfolio is

ex ante mean-varianceefficient.

While the paper is organizedinto seven sections, it also can be viewed as

consisting of three parts. The first part (Sections 2 through 4) considers a

multivariatestatisticfor testingmean-varianceefficiencyand examinesthe properties of such a test. The second part (Sections 5 and 6) studies the relation

between this multivariatetest and alternativeapproachesbased on a set of

univariatestatistics. The third part (Sections7 and 8) concludes the paper by

extending the frameworkto related hypothesesand providing suggestionsfor

future research.A moredetailedsummaryof each section follows.

In Section2 we recalla necessaryconditionfor the efficiencyof some portfolio.

We use this implicationas a null hypothesisthat can be tested using a statistic

which has a tractablefinitesampledistributionunderboth the null and alternate

hypotheses. In addition,we relate this statistic to three alternativeapproaches

which are based on asymptoticapproximations.In the third section the multivariate test is given a geometricinterpretationin the mean-standarddeviation

space of portfolio theory.The method and geometryare then applied to a data

set from one of the classic empiricalpapersin modernfinance;we reaffirmand

complementthe findingsof Black,Jensen,and Scholes(1972).The fourthsection

turns to issues relatingto the powerof the test. Here we considerthe sensitivity

of the test to the choice of the portfoliowhich is examinedfor efficiencyand the

effect of the numberof assets used to determinethe ex post efficientfrontier.A

new data base is analyzed in this section, and we demonstratethat one's

conclusionsregardingthe efficiencyof a givenindexcan be alteredby the type of

assets used to constructthe ex post frontier.

The fifth section attemptsto contrastactualempiricalresultswhen the multivariate method is used versus informalinferencebased on a set of dependent

univariatestatistics.Herewe provideexampleswherethe multivariatetest rejects

even thoughnone of the univariatestatisticsseem to be significant.We also have

2Recent work by Kandeland Stambaugh(1987) and Shanken(1987b)do considertests of the

CAPM conditionalon an assumptionabout the correlationbetweenthe proxy and the true market

portfolio.

All use subject to JSTOR Terms and Conditions

PORTFOLIO EFFICIENCY

1123

the reverse situation where there are a seemingly large number of "significant"

univariate statistics; yet, the multivariate test fails to reject at the traditional

levels of significance. In this section we also introduce another data set which

allows us to re-examine the size-effect anomaly. Section 6 develops an alternative

interpretation of the multivariate test. The statistic is equivalent to the usual

calculation for a t statistic on an intercept term in a univariate simple regression

model, with the ex post efficient portfolio used as the dependent variable and the

portfolio whose ex ante efficiency is under examination as the explanatory

variable. This section also develops some useful diagnostics for explaining why

the null hypothesis may not be consistent with the data. Most of the empirical

work in this section focuses on the size effect only in the month of January.

Section 7 extends the analysis to a case where one wishes to investigate the

potential efficiency of some linear combination of a set of portfolios, where the

weights in the combination are not specified. This turns out to be a minor

adaptation of the work in Section 2.

2. TEST STATISTIC FOR JUDGING THE EFFICIENCY OF A GIVEN PORTFOLIO

We assume throughout that there is a given riskless rate of interest, Rft, for

each time period. Excess returns are computed by subtracting Rft from the total

rates of return. Consider the following multivariate linear regression:

(1)

'it=aip+

flipppt + iit

Vi = 1,. .., N,

where ri the excess return on asset i in period t; Fpp- the excess return on the

portfolio whose efficiency is being tested; and 9it- the disturbance term for asset

i in period t. The disturbances are assumed to be jointly normally distributed

each period with mean zero and nonsingular covariance matrix 2, conditional on

the excess returns for portfolio p. We also assume independence of the disturbances over time. In order that 2 be nonsingular, Fp and the N left-hand side

assets must be linearly independent.

If a particular portfolio is mean-variance efficient (i.e., it minimizes variance

for a given level of expected return), then the following first-ordercondition must

be satisfied for the given N assets:

(2)

(Fit) = fiPeQPt).

Thus, combining the first-order condition in (2) with the distributional assumption given by (1) yields the following parameter restriction, which is stated in the

form of a null hypothesis:

(3)

Ho: aip = 0

li

=19,...,9N.

Testing the above null hypothesis is essentially the same proposal as in the

work by Black, Jensen, and Scholes (1972), except that they replace ip, by a

portfolio which they call the market portfolio and refer to their test as a test of

the CAPM. In addition, they do not report the joint significance of the estimated

All use subject to JSTOR Terms and Conditions

1124

values for a?1p across all N equations; instead, they report N univariate t

statisticsbased on each equation.

Given the normalityassumption,the null hypothesisin (3) can be tested using

"Hotelling's T2 test," a multivariategeneralizationof the univariatet-test (e.g.,

see Malinvaud(1980, page 230)). A brief derivationof the equivalentF test is

included for completenessand as a meansof introducingsome notationthat will

be needed later. If we estimatethe multivariatesystemof (1) using ordinaryleast

squaresfor each individualequation,the estimatedinterceptshave a multivariate

normal distribution,conditionalon rp,(Vt= 1,..., T), with

(4)

/T(+k)ap

- NtT(

a sT;

p)Op;

P);

p -plsp; rp- samplemean of r and sp samplevarianceof rp, without an

adjustmentfor degreesof freedom.Furthermore,ap and T are independentwith

(T - 2)2 having a Wishartdistributionwith parameters(T - 2) and 2. These

facts imply (see Morrison (1976, page 131)) that (T(T - N - 1)/N(T - 2))WJK

has

a noncentral F distributionwith degreesof freedom N and (T - N - 1), where

(5)

+ 2

on the maximumlikelihoodestimateof 2 will be denotedas W.)The noncentrality parameter,A, is givenby

X [T/(1+2

(6)

)]a - lap.

Under the null hypothesisthat ap equalszero, A= 0, and we have a centralF

distribution.More generally,the distributionunder the alternativeprovides a

way to study the powerof the test; morewill be said about this in a later section.

It is also interestingto note that underthe null hypothesisthe Wustatistichas a

central F distributionunconditionally,for the parametersof this centralF do not

depend on rin any way. However,we do not know the unconditionaldistributionofn

or Wuunderthe alternate,for the conditionaldistributiondependson

the samplevalues of -p,through0.

Generally, the normalityassumptionhas been viewed as providinga "good

workingapproximation"to the distributionof monthlystock returns(see Fama

(1976, Chapter1) for a summaryof the relevantempinrcalwork).Thereis some

evidence, however,that the true distributionsare slightlyleptokurticrelativeto

the normaldistribution.Whiledeparturesfrom normalityof the disturbancesin

(1) will affect the small-sampledistributionof the test statistic, simulation

evidence by MacKinlay(1985) suggeststhat the F test is fairly robust to such

misspecifications.4This is important,since the applicationof standardasymptotic tests to the efficiencyproblem can result in faulty inferences,given the

sample sizes often used in financialempiricalwork.

4We assumethat N is less thanor equalto T - 2 so that I is nonsingular.

Tests for normalityof the residualsof the size and industryportfolios,whichare used below,do

revealexcesskurtosisand someskewnessas well.Theseresultsareavailableon requestto the authors.

All use subject to JSTOR Terms and Conditions

1125

PORTFOLIO EFFICIENCY

TABLE 1

GIVEN PORTFOLIO. THE W STATISTIC IS DISTRIBUTED AS A TRANSFORM OF A CENTRAL F

DISTRIBUTION IN FINITE SAMPLES. THE WALD TEST, THE LIKELIHOOD RATIO TEST (LRT),

AND THE LAGRANGE MULTIPLIER TEST (LMT) ARE MONOTONE TRANSFORMS OF W, AND

EACH IS DISTRIBUTED AS CHI-SQUARE WITH N DEGREES OF FREEDOM AS T APPROACHES

INFINITY.

10

20

40

58

10

20

40

58

118

10

20

40

58

118

238

10

20

40

58

10

20

40

58

118

10

20

40

58

118

238

60

60

60

60

120

120

120

120

120

240

240

240

240

240

240

60

60

60

60

120

120

120

120

120

240

240

240

240

240

240

P-Value Using

Exact Distribution

of W

.05

.05

.05

.05

.05

.05

.05

.05

.05

.05

.05

.05

.05

.05

.05

.10

.10

.10

.10

.10

.10

.10

.10

.10

.10

.10

.10

.10

.10

.10

P-Values Using

Asymptotic Approximations

Wald

LRT

LMT

.008

.000

.000

.000

.023

.005

.000

.000

.000

.035

.109

.003

.000

.000

.000

.025

.000

.000

.000

.056

.017

.000

.000

.000

.076

.048

.009

.001

.000

.000

.027

.007

.000

.000

.038

.023

.003

.000

.000

.044

.035

.017

.006

.000

.000

.061

.019

.000

.000

.081

.053

.010

.000

.000

.090

.075

.041

.018

.000

.000

.071

.094

.173

.403

.060

.070

.094

.122

.431

.055

.059

.069

.079

.123

.451

.122

.146

.216

.404

.111

.122

.147

.175

.432

.106

.111

.122

.133

.178

.452

Note: N is the number of assets used together with portfolio p to construct the ex post frontier, and T is the

number of time series observations.

multipliertests, each of whichis asymptoticallydistributedas chi-squarewith N

degrees of freedom as T -- oo.S Since the small-sample distribution of W is

the information in the first three columns of Table I (i.e., N, T, and the

hypothetical p-value). The implied asymptoticp-values given in the last three

5 Jobsonand Korkie(1982)also discussthesethreetests usinga simulation.Theyapproximatethe

distributionof the likelihoodratio test with an F distributionbased on Rao's (1951)work.In -their

1985 paperthey recognizethat a smallsampledistributionis availableunderthe null hypothesis.

All use subject to JSTOR Terms and Conditions

1126

columns are then obtained using the fact that each test statistic is a monotonic

function of W.6

Consistent with the results of Berndt and Savin (1977), the p-values are always

lowest for the Wald test and highest for the Lagrange multiplier test with the

likelihood ratio test in between. Clearly, the asymptotic approximation becomes

worse as the number of assets, N, approaches the number of time series

observations, T. Shanken (1985) reaches similar conclusions based on an approximation when the riskless asset is not observable.

3. A GEOMETRIC INTERPRETATION

So far, the primary motivation for the W statistic has been its well-known

distributional properties. For rigorous statistical inference such results are an

absolute necessity. Just as important, though, is the development of a measure

which allows one to examine the economic significance of departures from the

null hypothesis. Fortunately, our test has a nice geometric interpretation.

It is shown in the Appendix that:

W=

(7)

?*2

-12

=2_

where &* is the ex post price of risk (i.e., the maximum excess sample mean

return per unit of sample standard deviation) and Opp~~~~~

is the ratio of ex post

average excess return on portfolio p to its standard deviation (i.e., p -p/sp).

Note that 4 cannot be less than one since 0* is the slope of the ex post frontier

based on all assets used in the test (including portfolio p).

The curve in Figure la represents the (ex post) minimum-variance frontier of

the risky assets. When a riskless investment is available, the frontier is a straight

line emanating from the origin and tangent to the curve at m. 0* is the slope of

the tangent line whereas 0p is the slope of the line through p.

An examination of (7) suggests that 42 should be close to one under the null

hypothesis. When 0* is sufficiently greater than 0 , the return per unit of risk for

portfolio p is much lower than the ex post frontier tradeoff, and we will reject the

hypothesis that portfolio p is ex ante mean-variance efficient. In Figure la 4 is

just the distance along the ex post frontier up to any given risk level, a, divided

by the similar distance along the line from the origin through p.

The reader may wonder why the test is based on the square of the slopes as

opposed to the actual slopes. The reason is straightforward. Our null hypothesis

only represents a necessary condition for ex ante efficiency. This condition is

satisfied even if portfolio p is on the negative sloping portion of the minimumvariance frontier for all assets (including the risk-free security). Thus, only the

6

The relations are LRT= T ln(1 + W) and LMT= TW/(1 + W). Shanken (1985) has discussed

this result for the case where the riskless asset does not exist. A proof of the result in the case where

the riskless asset does exist is available upon request to the authors. Bemdt and Savin (1977) discuss

similar relationships among alternative asymptotic tests in a more general setting.

All use subject to JSTOR Terms and Conditions

1127

PORTFOLIO EFFICIENCY

p~~~~~~~~~~~

X

Standard

Deviation

of Excess

Return

is

I + 02,

V1 +*

--

2.4

2.2

2/

1.2]

4)

0.80.6-] 0

1 2-

81

0.4

0.2-

0I

0

Standard

lb.)

Deviation

of Excess

10

Return

portfolios and tihe CRSP Equal-Weighted Index

using monthly data, 1931-1965. Point p represents the CRSP Equal-Weighted Index.

All use subject to JSTOR Terms and Conditions

1128

1.7

1.6

1.5

1.4

1.3

1.2

0.9

*i

0.7XOB

*

0.8

N

0.5

0.40.3-

0

0

Standard

Deviation

6

of Excess

Return

portfolios and the CRSP Value-Weighted 1n(lex

using monthly data, 1926-1982. Point p represents the CRSP Value-Weighted Index.

1.21.1

0.9

0.8

*

x

0.?

0.6

0.5

0.40.30.20.

0

2

Standard

4

Deviation

of Excess

Return

portfolios and the CRSP Value-Weighted Index

using monthly data, 1926-1982. Point p represents the CRSP Value-Weighted Index.

FIGURE1.-Continued.

All use subject to JSTOR Terms and Conditions

1129

PORTFOLIO EFFICIENCY

TABLE II

SUMMARY

ALL SIMPLE EXCESS RETURNS ARE NOMINAL AND IN PERCENTAGE FORM, AND THE

(T=420).

CRSP EQUAL-WEIGHTEDINDEX IS PORTFOLIOp. THE FOLLOWINGPARAMETERESTIMATES

10 AND Vt =1.

ARE FOR THE REGRESSION MODEL: Pit = alp + Pip Fpt + 9,t Vi = 1.

420,

WHERE R2 IS THE COEFFICIENT OF DETERMINATION FOR EQUATION i.

Portfolio

Number

&,P

1

2

3

4

5

6

7

8

9

10

-0.19

-0.19

-0.06

-0.09

-0.06

0.05

0.03

0.12

0.14

0.22

S(&,P)

0.17

0.10

0.09

0.07

0.07

0.07

0.07

0.07

0.08

0.09

/3P

1.54

1.37

1.24

1.17

1.06

0.92

0.86

0.74

0.63

0.51

s(Ap)

0.02

0.01

0.01

0.01

0.01

0.01

0.01

0.01

0.01

0.01

R2

0.94

0.97

0.98

0.98

0.98

0.97

0.97

0.96

0.92

0.85

NOTE: For this sample period Op and 0* are 0.166 and 0.227, respectively. These imply a value for Wu equal to

0.023, which has a p-value of 0.476. Under the hypothesis that the CRSP Equal-Weighted Index is efficient, '( WTu)

is 0.024 and SD( W,,) is 0 01 1.

absolutevalue of the slope is relevantfor our null hypothesis,and our test is then

based on the squaredvalues.

Figure lb is based on a data set that is very similarto the one used by Black,

Jensen, and Scholes (1972) (hereafter,BJS).7 Using monthly returns on 10

beta-sortedportfolios from January,1931 throughDecember,1965, 0* = 0.227

while the CRSP Equal-WeightedIndex, which is portfoliop, has 0p = 0.166. To

judge whether these two slopes are statistically different, we can calculate

_ 1), which is 0.02333.Basedon the resultsin Section2, we can use a central

(#2

F distribution with degrees of freedom 10 and 409 to judge the statistical

significanceof this differencein slopes.The resultingF statisticis 0.96, whichhas

a p-value of 0.48. Our multivariatetest confirmsthe conclusionreachedby BJS

for their overall time period in that the ex ante efficiencyof the CRSP EqualWeightedIndex cannotbe rejected;equivalently,if this Indexis takenas the true

market portfolio, then the Sharpe-Lintnerversion of the CAPM cannot be

rejected.Table II providessome summarystatisticson the beta-sortedportfolios

that were used for Figure lb. Table II, when comparedwith Table II in BJS,

verifiesthat our data base is very similarto the one used by BJS.

BJS provide variousscatterplots of averagereturnsversusestimatedbetas to

judge the fit of the datato the expectedlinearrelationif the CRSPEqual-Weighted

7 WhileBJSreliedon the datafromthe Centerfor Researchin SecurityPrices(hereafter,CRSP)at

the Universityof Chicago,it is not possibleto replicatetheirdata.The CRSPtapes are continually

revised to reflect data errors,and one would need the same versionof the CRSP file to perfectly

duplicatea data base. For example,we wereable to findmorefirmsper yearthanreportedin Table1

of BJSbecauseof correctionsto the data base.Also we reliedon Ibbotsonand Sinquefield(1979)for

the returnof US TreasuryBills as the risklessrate. This latter data base was not used by BJS.

However,we followedthe groupingprocedureoutlinedin BJSin formingthe 10 portfoliosthat were

used in constructingFigure1 and TableII.

All use subject to JSTOR Terms and Conditions

1130

scatter plots, for they summarize the multivariate test in a manner familiar to

financial economists. The advantage of the scatter plots in BJS is that they may

provide some information as to which asset or which set of assets is least

consistent with the hypothesis that the index is efficient; figures like Figure lb

really do not provide such information. On the other hand, the scatter plots in

BJS can be difficult to interpret due to heteroscedasticity across the different

portfolios as well as contemporaneous cross-sectional dependence. Section 6 will

suggest some other types of diagnostic information based on the multivariate

framework.

To understand further the behavior of our measure of efficiency, 4?,its small

sample distribution given in Section 2 is helpful. Since a linear transform of 42

has a central F distribution with degrees of freedom N and (T - N - 1), we can

use the first two moments of the central F to calculate:

(8)

-1) = [ T-N-3

and

(9)

SD(P2_1)

=[T-N-3

T-N-5

The first moment for 42 only exists if T > N + 3 while the second moment for 42

only exists if T> N + 5. These last two equations for the moments can be

applied to the BJS data set for 1931-1965 where N = 10 and T= 420, so

1) and the standard deviation of p2 are 0.024 and 0.011, respectively. As

6,(+2the realized value of 42 _ 1 is less than its expectation, it is not surprising that

the ex ante efficiency of the Equal-Weighted Index cannot be rejected for this

time period.

This measure, 4, is a new variant of the geometry developed to examine

portfolio performance. In past procedures the efficient frontier has been taken as

given, and a distance such as mb in Figure la has been used as a measure of p's

performance. Note that mb is simply the return differential of the ex post

optimal portfolio over p, computed at the sample standard deviation of the ex

post optimal portfolio. Another suggestion has been to use the difference in their

slopes * - p as a measure of p's relative performance. How the true ex ante

frontier is to be known is unclear, and if the ex post frontier is used, then we face

the statistical problem of this paper.

4. THE POWEROF THE MULTIVARIATETEST FOR EFFICIENCY

The empirical illustration in the previous section fails to reject the ex ante

efficiency of the Equal-Weighted Index when using 10 beta-sorted portfolios as in

BJS.8 Such a result may occur because the null hypothesis is in fact true, or it

8 We have also examined our data base using the same subperiods as in BJS. When we aggregate

the results of the multivariate test across these four subperiods, we can reject ex ante efficiency at

usual levels of significance. This confirms the conclusions reached by BJS.

All use subject to JSTOR Terms and Conditions

PORTFOLIO EFFICIENCY

1131

may be due to the use of a test which is not powerful enough to detect

economically importantdeviationsfrom efficiencyof the Index. Questions of

power for various types of test statistics have been a long standing concern

among financialeconomists(e.g., see Roll (1977),amongothers).This sectionwill

focus on the powerof the multivariatetest.

From Section 2 we know that underboth the null and alternatehypothesesa

simple transformof W, or 42, has an F distributionwith degreesof freedomN

and T - N - 1. The F distributionis noncentralwith the noncentralityparameter

given by equation (6); underthe null hypothesisthe noncentralityparameteris

zero. It deserves emphasis that the F distributionunder the alternativeis

conditional on the returns of portfolio p since the noncentralityparameter

depends on O2. Thus, we will be studyingthe power functionconditionalon a

value for 0, not the unconditionalpowerfunction.

The probabilityof rejectinga false null hypothesisincreasesas the noncentrality parameterincreases,holding constant the numeratorand denominatordegrees of freedom(Johnsonand Kotz (1970,page 193)). Studyingthe factorsthat

affect the noncentralityparameter,X, will give some guidanceabout the powerof

the multivariatetest. From equation(6) we can see that A is a weightedsum of

squareddeviations about the point ap =0. The weightingmatrixis the inverse

of the covariancematrixof the ordinaryleast squaresestimatorsfor ap. Thus,

estimated departuresfrom the null are weightedaccordingto the variabilityof

the estimatorand the cross-sectionaldependenceamongthe estimators.

The noncentralityparametercan also be given an intuitiveeconomicinterpretation. The derivationof equation (23) in the Appendix would hold for the

populationcounterpartsof the sampleestimates,so it is also truethat a' -ap=

9*2

a2

X = [ T/(I

Not surprisingly,the powerof the test will increaseas the ex ante inefficiencyof

portfolio p increasesas measuredin termsof the slope of the relevantopportunity sets. If p2 increases,the precisionof the estimatorfor ap declines, so the

power of the test decreases.

Figure2 summarizeshow the powerof the test is affectedby 0* and Op.When

the proportion of potential efficiency (i.e., Op/O*)is equal to one, the null

hypothesis is true. As this proportionapproacheszero, the given portfolio is

becomingless efficient.Figure2 is based on values for the significancelevel, N,

and T that are commonfor existingempiricalworkon asset pricingmodels;we

have used N = 10 or 20 and T= 60 or 120 and a five percentsignificancelevel.

Empiricalwork on the CRSPIndexesreportsestimatesof Opbetween0 and 0.4.

We have used this range to guide our selection of a grid for OPand 0*. In

addition,Figure2 is basedon the assumptionthat p= Opto eliminateone of the

parametersthat affect A; this assumptionsuggeststhat our calculationsof power

of the underlyingpopulation.

are for situationswherethe sampleis representative

Even within the range of parametersthat we consider, the probabilityof

rejecting the null hypothesis ranges from five percent to nearly 100 percent

depending on the differencebetween the two relevantmeasuresof slope. For

All use subject to JSTOR Terms and Conditions

1132

C

4

~~~0

01

2a.) N

10 and rr

OTENTIAL EFFICIENC'

60.

0~

ONI0

NT

2b.) N

PRP

ENIA

7F

ION OF p0TrNTIAL

FIGURE2.-Sensitivity of the power of the test to the choice of the index. Each figure is based on a

different combination of the number of assets (N) and the number of time series observations (T). In

all cases a critical level of five percent is used.

and if N = 20 and T = 60, then the probabilityof rejectinga false null hypothesis

ranges from ten percent(for 9* = .3) to 98 percent(for 9* = 1.0).

Given the data bases that are available,an empiricistis alwaysfaced with the

question of the appropriatesizes for N and T. For example,with the CRSP

monthly file we have a data base which extendsback to 1926 for every firm on

the New York Stock Exchange.This would permitthe empiricistto use around

700 time seriesobservations(i.e., T) and well over 2000 firms(i.e., N). However,

the actual N used may be restrictedby T to keep estimates of covariance

matrices nonsingular,and the actual T used is constrainedby concerns over

All use subject to JSTOR Terms and Conditions

PORTFOLIO EFFICIENCY

2c.) N 1 arid

2d.) N

20 and T

1133

OF PO

PROPORTION

01

4

120.PROORrION

FOTENTIAL EFFICIENC-y

120.

FIGURE2.-Continued.

monthly observations and N is between 10 and 20. While these numbers for N

and T are common, we are not aware of any formal attempts to study the

appropriate values to select. We will now examine this issue in the context of the

specific hypothesis of ex ante efficiency. While the analysis is focused on an

admittedly special case, our hope is that it may shed some light on other cases as

well.

To get more intuition about the impact of N on equation (6), consider a case in

which S has a constant value down the diagonal and a constant (but different)

value for all off-diagonal elements. Since S represents the contemporaneous

covariances across assets after the "market effect" has been removed, sucha

All use subject to JSTOR Terms and Conditions

1134

structure includes the Sharpe (1963) diagonal model as a special case when the

off-diagonal terms are zero. The more general case where the off-diagonal terms

are constant but are nonzero is motivated by the work of Elton and Gruber

(1973) and Elton, Gruber, and Urich (1978).9 Under this structure we can

parameterize T as:

(10)

= (1 -P)

2IN + P2

'I,

where p the correlation between eit and ij,; 2 the variance of eit; IA, an

identity matrix of dimension N; and l I- a 1 x N vector of l's. The inverse10of

this patterned matrix is (Graybill (1983)):

(lP)2

[IN

1 + (N-l)p LNLNb

(11)

= T/(1+

(1p)W2

) [N2

[N

p

(1-p)+Np

(a'a )/N. One could view ju as a measure of the

"average" misspecification across assets while A2 indicates the noncentral dispersion of the departures from the null hypothesis across assets.

When N is relatively large and p is not equal to zero, equation (11) implies:

(12)

=

X/N_

A)y2

2) =

(1(

)VAR(ap),

where VAR (ap) denotes the cross-sectional variance of the elements of ap. Thus,

X is approximately proportional to N and T."1Alternatively, if either p = 0 or

Al = 0, then X is exactly proportional to N and T.12 Unfortunately, this is still

not adequate to determine the impact of changing N and T, for these two

parameters affect not only the noncentrality parameter but also the degrees of

freedom.

We have evaluated the power of the multivariate test for various combinations

of X, N, and T.13These numerical results provide some guidance on the proper

9Strictly speaking, the Sharpe diagonal model allows for heteroscedasticity in the disturbances of

the market model equations; our formulation assumes homoscedasticity. Also, the constant correlation model of Elton and Gruber is usually applied to the correlation matrix for total returns; we are

assuming constant correlation after the market effect has been removed.

10Necessary and sufficient conditions for this inverse to exist are that p # 1 and p # (1 - N) ';

see Graybill (1983, page 190-191). In addition, the matrix should be positive definite; this would

require that p> - l/(N - 1).

"1In general, since p < 1, X/N is less than or equal to the right side of (12) when p ? 0.

12 If the Equal-Weighted Index is portfolio p, then we would expect

Al to approach zero as N

becomes large.

13

These numerical calculations require evaluation of a noncentral F distribution and an inverse of

a central F distribution. The latter calculation relied on the MDFI subroutine provided by IMSL. The

former calculations are based on a subroutine written by J. M. Bremner (1978), and a driver program

written by R. Bohrer and T. Yancey of the University of Illinois at Champaign-Urbana. Each

subroutine was checked by verifying its output with the published tables reported in Tang (1938) and

Titu (1967).

All use subject to JSTOR Terms and Conditions

1135

PORTFOLIO EFFICIENCY

10 1

0 .8-

0.7

0.6

\SA=O.1T

,-

;+

0-4

0.3

A

\

0.002NT

0.00002NT

02

0.2

0.6

0.4

0.8

Ratio of N Divided by T

3a.) T

60.

0.9

0.8

,

0.7-

;:

0.6-

0.5-

0.4-

0.3-

0.2

0.1

iA

A- 0.00002NT

A

I

,

I

0]

0.2

0.4

.

0.6

0.002NT

O.B

Ratio of N Divided by T

3b.) T

120.

FIGURE3.-Sensitivity of the powerof the test to the choiceof the numberof assets(N) gives a

fixed number of time series observations (T).

We selected this

provides various values for the constant of proportionality.14

constant of proportionalitybased on equation(11) when p = 0. In this case, the

+

constantis 2/[2(1

)]. We then replacedI2 and o2 with the cross-sectional

and 6i2,respectively,from an actualdata set. We also know that

averagesof

Op is 0.166 for the CRSP Equal-WeightedIndex (1931-1965) and 0.109 for the

14

of the

MacKinlay (1987) studies the power of the test using alternativeparameterizations

noncentralityparameter.

All use subject to JSTOR Terms and Conditions

1136

0.9I.

A=O.1T

0.7

.6

0.002NT

0.5

0.4

0.3

0.2

0

o. 0

A

0.2

0.002N

0.4

0.6

0.8

Ratio of N Divided by T

3c.) T

240.

FIGURE 3.-Continued.

values for the constant of proportionality.The constant is 0.004 using the

beta-sorted portfolios, and it is 0.002 using a set of industry portfolios. For

size-sorted portfolios the constant is 0.004 using all months and 0.763 for

monthlydata only usingJanuary.(The detailson how the industryportfoliosand

size-sortedportfolioswerecreatedwill be providedlater in this paper.)

In Figure 3, we look at cases wherethe constantis 0.00002 and 0.002, which

are small relativeto the abovecalculations.For purposesof comparison,Figure3

also includes a case where X is not affectedby N; instead we set X=.1T. This

representsa situationwherean investigatorhas one asset that violates the null

hypothesis, and all the remainingassets that are added are consistentwith the

efficiencyof some givenportfoliop. WhileFigure3 is based on specificvaluesfor

the constant of proportionality,the generalpatternthat is observedis consistent

with a wide rangeof choicesthat we triedbut did not reporthere.15

For a fixed numberof time series observations,Figure 3 demonstratesthat

there may be an importantdecisionto be made by the empiricist.Even though

the noncentralityparameterincreasesas N increases,it is not necessarilyappropriate to choose the maximumN possible.Given our particularparameterization

of the problem,it appearsthat N shouldbe roughlya third to one half of T, or

when five years of monthlydata are used, 20 to 30 assets may be appropriate.

When the constantis very low, the poweris so small for all possiblevaluesof N

15We were not able to evaluatethe noncentralF for very high values of A, so we have little

is high.If the

knowledgeabout the shapeof the powerfunctionwhenthe constantof proportionality

constant is large enough, it is conceivablethat a corner solution of setting N = T - 2 may be

appropriate.

All use subject to JSTOR Terms and Conditions

PORTFOLIO EFFICIENCY

1137

is proportionalto T and not affectedby N, clearlysetting N = 1 is the preferred

strategy.In this case addingsecuritiesdoes not providemore informationabout

departuresfrom the null hypothesis;however,additionalsecuritiesincreasethe

numberof unknownparametersto be estimated.It deservesemphasisthat these

conclusionsabout the properchoiceof N may not be appropriatefor all possible

situationsand models.

The choices of N and T are not the only decisions facing the empiricistin

designingthe econometricanalysis.Since N must alwaysbe less than T (unless

highly structuredcovariancematricesare entertained),the empiricistmust also

decide how to select the assetsto maximizethe powerof the test. Given N and T

we wish to maximizethe quadraticform a-T'ap, or equivalently9*; however,

these parametersare unobservable.A common approachis to use beta-sorted

portfolios. While dispersion in betas is useful in decreasingthe asymptotic

standarderror in estimatesof the expectedreturnon the zero-betaasset (Gibbons (1980) and Shanken(1982)), such sorting need not maximizedepartures

from the null hypothesisas measuredby X.16

Empiricalexamplespresentedbelowillustratethe effectthat differentasset sets

can have on the outcome of the test. First, we consider a set of 12 industry

portfolios.17 An industrygroupingseems reasonableon economic groundsand

also capturessome of the importantcorrelationsamong stocks.To measurethe

returnfrom a "buy-and-hold"investmentstrategy,the relativemarketvalues of

the securitiesare used to weightthe returns.Almost everymonthlyreturnon the

CRSP tape from 1926 through1982is included,whichshouldminimizeproblems

with survivorshipbias.18 Table III provides some summarystatistics on the

industryportfolios.

The multivariateF statisticrejectsthe hypothesisof ex ante efficiencyat about

the one percentsignificancelevel. The relevantF statisticis 2.13 with degreesof

freedom 12 and 671; its p-value is 0.013.19To complementthese numerical

results, Figure Ic, whichis similarto Figuresla and lb, providesa geometrical

summary.

16 In fact, for a given set of N securities,the multivariate

test is invariantto how we groupthese

assets into N portfolios;we could form N portfoliosso that they have verylittle dispersionin their

beta valueswith no impacton the power.This followsfromthe well-knownresultin the multivariate

statisticsliteraturethat our test is invariantto lineartransformations

of the data (Anderson(1984,

pages 321-323)). Of course,the selectionof the originalsubsetof assetsto be analyzedis important

even thoughthe way theyare aggregatedinto portfoliosis not (giventhat the numberof portfoliosis

the same as the numberof originalassets).

17 For the detailsof the database, see Breeden,Gibbons,and Litzenberger

(1987),who developed

these data for tests of the consumption-based

asset pricingmodel. The industrygroupingclosely

follows a classificationused by Sharpe(1982).

18

However,all firmswith a SIC numberof 39 (i.e., miscellaneousmanufacturing

industries)are

excluded to avoid any possible problemswith a singular covariancematrix when the CRSP

Value-WeightedIndex is used as portfoliop.

19 Whilenot

reportedhere,we also analyzedthis data set acrossvarioussubperiods.Basedon five

year subperiods,the p-valuefor the F statisticis less thanfivepercentin 7 out of 11 cases,is less than

10 percentin 9 out of 11 cases,andrejectswhenaggregatedacrossthe subperiods.Thus,the rejection

of the overallperiodis confirmedby the subperiodsas well.

All use subject to JSTOR Terms and Conditions

1138

To understandthis low p-value, consider the fact that for this time period

= 0.109 while the slope of the opportunityset using the ex post optimal

portfolio, 9*, is more than doublewith a value of 0.224. With these numberswe

can calculate 42 as 1.038.For N= 12 and T= 684, e(42) iS 1.018 with SD( 2)

of 0.007. Thus, the realizedvalue of 42 is nearlythreestandarddeviationsfrom

its expectedvalue if the CRSPValue-WeightedIndex is trulyex ante efficient.

Perhapsof greaterinterestis the fact that the multivariatetest rejectsthe null

hypothesisat the one percentlevel even thoughall 12 univariatet statisticsfail to

reject at even the five percentlevel. The next sectionbuilds on such contrasting

results by analyzingwhy univariatetest may be difficultto summarizeacross

differentassets.

0

Table II suggests that high beta portfolios earn too little and low beta

portfolios too much if the Equal-WeightedIndex is presumedto be efficient;

similarevidencewas used by BJS to garnersupportfor the zero-betaversionof

the CAPM.Yet, this patternis difficultto interpret.The uppertriangularportion

of Table IV providesthe samplecorrelationmatrixof the marketmodel residuals

based on the regressionsthat are summarizedin Table II. A very distinctive

patternemergesin that the residualsof portfolioswith similarbetas are positively

correlated while those of portfolios with very different betas are negatively

matrixfor ap in equation(4), it is

correlated.Based on the variance-covariance

clear that the estimatorsfor aip will have the samepatternof correlation.Thus,it

is difficultto infer whetherthe observedpatternin estimatedvalues of a 's is

TABLE III

SUMMARY STATISTICSON INDUSTRY-SORTEDPORTFOLIOSBASED ON MONTHLY DATA,

1926-82 (T = 684). ALL SIMPLEEXCESSRETURNSARE NOMINALAND IN PERCENTAGEFORM,

AND THE CRSP VALUE-WEIGHTEDINDEX IS PORTFOLIOp. THE FOLLOWINGPARAMETER

ESTIMATES ARE FOR THE REGRESSION MODEL: ,,= a?ip +/ApFpt+t

Vi =1,.12

AND

Vt= 1.684,

WHEREI IS THE COEFFICIENT

OF DETERMINATION

FOR EQUATIONi.

IndustryPortfolio

a,

Petroleum

Financial

Consumer Durables

Basic Industries

Food and Tobacco

Construction

Capital Goods

Transportation

Utilities

Trade and Textiles

Services

Recreation

0.17

-0.05

0.03

0.00

0.12

-0.17

0.10

-0.17

0.05

0.00

0.43

-0.03

s(&,p)

0.14

0.09

0.09

0.00

0.07

0.12

0.08

0.14

0.09

0.00

0.37

0.13

P,

0.93

1.19

1.29

1.09

0.76

1.20

1.08

1.20

0.74

0.94

0.80

1.22

s(fi

0.02

0.02

0.02

0.01

0.01

0.02

0.01

0.02

0.02

0.02

0.06

0.02

R2

0.69

0.89

0.90

0.94

0.83

0.85

0.91

0.78

0.76

0.77

0.19

0.78

NOTE: For this sample period 6, and 6* are 0.109 and 0.224, respectively. These imply a value for Wu equal to

0.038, which has a p-value of 0.013. Under the hypothesis that the CRSP Value-Weighted Index is efficient, 9'( W")

is 0.018 and SD(W") is 0.007.

All use subject to JSTOR Terms and Conditions

1139

PORTFOLIO EFFICIENCY

TABLE IV

EXCESS RETURNS.

THE UPPER TRIANGULAR PORTION OF THE TABLE IS BASED ON 10 BETA-SORTED PORTFOLIOS

FOR THE DEPENDENT VARIABLES AND THE CRSP EQUAL-WEIGHTED INDEX FOR PORTFOLIO

p. ALL MONTHLY DATA FROM 1931-65 (T = 420) ARE USED. TABLE II SUMMARIZES THE

OTHER PARAMETER ESTIMATES FOR THIS REGRESSION MODEL. THE LOWER TRIANGULAR

PORTION OF THE TABLE IS BASED ON 10 SIZE-SORTED PORTFOLIOS FOR THE DEPENDENT

INDEX FOR PORTFOLIO p. ALL MONTHLY

VARIABLES AND THE CRSP VALUE-WEIGHTED

DATA FROM 1926-82 (T = 684) ARE USED. TABLE V SUMMARIZES THE OTHER PARAMETER

ESTIMATES FOR THIS REGRESSION MODEL.

.52

.62

.72

.68

.66

.70

.66

.63

.61

.63

.41

.52

.39

.39

.28

.35

-.54 -.59

.39

.38

.03

.01

.08

-.32

-.16

-.06

-.06

.75

.70

.68

.57

.51

.21

-.68

.72

.65

.55

.46

.27

-.66

Portfolio Number:

6

.72

.62

.50

.26

-.68

-.51

-.35

-.25

-.07

.21

.67

.59

.36

-.68

10

-.64

-.50

-.37

-.11

.25

.34

-.60

-.52

-.43

-.13

.12

.36

.43

-.64

-.53

-.46

-.32

.03

.26

.46

.49

-.50

-.55

-.49

-.22

- .16

.06

.23

.27

.56

.52

.27

-.61

.30

-.66

-.38

NOTE: For the upper triangular portion of the table, portfolio 1 consists of firms with the highest values for

historical estimates of beta while portfolio 10 contains the firms with the lowest values. For the lower triangular

portion of the table, portfolio 1 is a value-weighted portfolio of firms whose market capitalization is in the lowest

decile of the NYSE while portfolio 10 contains firms in the highest decile.

parameters.

Otherexamplesfromempiricalworkin financialeconomicscould also be cited

whereunivariatetests are difficultto interpret.Sincethe workof Banz(1981)and

Reinganum(1981), the "size effect"has receiveda great deal of attention.(For

more informationabout this researchsee Schwert(1983), who summarizesthe

existing evidence and also providesa useful bibliography.)While most of the

researchin this area now focuses on returnsin January,we begin by looking at

the originalevidencewhichdid not distinguishbetweenJanuaryand non-January

returns.

We have created a data base of monthly stock returnsusing the CRSP file.

Firms were sorted into 10 portfoliosbased on the relativemarketvalue of their

total equity outstanding.In otherwords,we rankedfirmsby theirmarketvalues

in December, 1925 (say), and we then formed 10 portfolios where the first

portfolio contains all those firmsin the lowest decile of firm size and the tenth

portfolio consists of companiesin the highestdecileof firmsize on the New York

Stock Exchange.Each of the ten portfoliosis value-weighted,and the firmsare

not resortedby their marketvalues for five years.Thus, the returnson these 10

portfolios from January,1926 through December, 1930 representthe returns

from a buy-and-holdstrategywithoutany rebalancingfor five years;this portfolio formationwas adoptedto representa low transactioncost investmentstrat-

All use subject to JSTOR Terms and Conditions

1140

TABLE V

(T = 684). ALL SIMPLE EXCESS RETURNS ARE NOMINAL AND IN PERCENTAGE FORM, AND THE

CRSP VALUE-WEIGHTED INDEX IS PORTFOLIO p. THE FOLLOWING PARAMETER ESTIMATES

10 AND Vt =1.

684,

Portfolio

Number

/,p

1

2

3

4

5

6

7

8

9

10

0.28

0.34

0.25

0.18

0.19

0.18

0.08

0.08

0.00

-0.01

s(6,p)

0.24

0.18

0.14

0.13

0.10

0.09

0.07

0.06

0.00

0.05

f3,lip

1.59

1.45

1.40

1.36

1.27

1.25

1.18

1.17

1.16

0.94

s(ip)

0.04

0.03

0.02

0.02

0.02

0.02

0.01

0.01

0.01

0.00

RI

0.68

0.78

0.84

0.85

0.89

0.91

0.93

0.95

0.96

0.98

NOTE: Portfolio 1 is a value-weighted portfolio of firms whose market capitalization is in the lowest decile of the

NYSE while portfolio 10 contains firms in the highest decile. For this sample period # and #* are 0.109 and 0.172,

respectively. These imply a value for Wu equal to 0.017, which has a p-value of 0.301. Jnder the hypothesis that the

CRSP Value-Weighted Index is efficient, 6(W") is 0.015 and SD(PWV)is 0.007.

time period.

Given the existing evidence on the size effect, some readers may find it

somewhat surprisingthat, in the overall period from 1926 through 1982, the

multivariatetest fails to rejectefficiencyof the CRSP Value-WeightedIndex at

the usual levels of significance.The firstrow of TableVI reportsthe statisticand

its correspondingp-value; Figure Id providesa geometricalinterpretationfor

this overallperiod.20

The correlationmatrix of the marketmodel residualsof the size portfolios

exhibits a distinctivepattern.The lower triangularportionof Table IV provides

this informationbased on the overallperiod. However,the patternis identical

acrosseveryten year subperiodreportedin TableVI, and a similarpatternis also

described by Brown, Kleidon, and Marsh(1983, page 47) and Hubermanand

Kandel (1985b).The correlationis positiveand high amongthe low decile firms.

The correlationdeclinesas one comparesportfoliosfrom very differentdeciles.

Even more striking is the fact that the highest decile portfolio has negative

sample correlationwith all other decile portfolios.(In some of the subperiods,

this negative correlationoccurredfor the ninth decile as well.) Thus, if we

observe that the lowest decile performswell (i.e., estimated alphas that are

positive),we would then expectthat the highestdecilewould do poorly(and vice

versa). This is the case, for example,in the period 1946-1955, wherefive out of

20

The subperiod results in Table VI are consistent with the conclusions of Brown, Kleidon, and

Marsh (1983) who find the size effect is not constant across all subperiods.

All use subject to JSTOR Terms and Conditions

1141

PORTFOLIO EFFICIENCY

TABLE VI

p) RELATIVETO 10 SIZE-SORTEDPORTFOLIOS.ALL SIMPLEEXCESSRETURNSARE NOMINAL

AND IN PERCENTAGEFORM. OVERALL PERIOD IS BASED ON ALL MONTHLYDATA FROM

1926-82.

THE FOLLOWING MODEL IS ESTIMATEDAND TESTED: r, alp+ Aprpt

+ 9,t

Vi = 1.

Time

Period

W.

(P-Value)

(T)

p

1926-1982

(684)

1926-1935

(120)

1936-1945

(120)

1946-1955

(120)

1956-1965

(120)

1966-1975

(120)

1976-1982

(84)

0.109

0.172

0.065

0.354

0.146

0.286

0.308

0.469

0.216

0.604

-0.019

0.093

0.408

0.538

0.018

(0.301)

0.119

(0.227)

0.059

(0.765)

0.113

(0.264)

0.302

(0.001)

0.164

(0.065)

0.275

(0.039)

Numberof

?t(ap)

I > 1.96

1

0

0

5

1

0

9

NOTE:#P is the ratio of the sample averageexcess returnon the CRSP Value-WeightedIndex divided by its

sample standarddeviation, and #* is the maximumvalue possible of the ratio of the sample averageexcess return

divided by the sample standarddeviation. W_ (#*2 - )/(1 + 2), and it is distributedas a transformof a central

F distributionwith degreesof freedom10 and T - 11 under the null hypothesis. W should converge to zero as T

approachesinfinity if the CRSP Value-WeightedIndex is ex ante efficient.By convertingthe p-values for the W.

statistics to an implied realizationfor a standardizednormalrandomvariable,the results across the 6 subperiods

can be summarizedby summingup the 6 independentand standardizednormalsand dividingby the squareroot of

6 as suggested in Shanken(1985). This quantityis 2.87 which implies a rejectionacross the subperiodsat the usual

levels of significance.

ten portfolioshave significantalphas(at the five percentlevel),but the multivariate test cannot rejectthe efficiencyof the Value-WeightedIndex.

Even though summarizingthe results of univariatetests can be difficult,

applied empiricalwork continuesto report such statistics.This is only natural,

for univariatetests are moreintuitive(perhapsbecausethey are used more) and

seem to give more diagnosticinformationabout the natureof the departurefrom

the null hypothesiswhenit is rejected.Partof the goal of this paperis to provide

some intuitionbehindmultivariatetests. Section3 has alreadydone this to some

extent by demonstratingthat the multivariatetest can be viewed as a particular

measurementin mean-standarddeviation space of portfolio theory. The next

section shows that the multivariatetest is equivalentto a "t test"on the intercept

in a particularregressionwhichshouldbe intuitive.A way to generatediagnostic

informationabout the natureof the departuresfrom the null hypothesisis also

provided.

6. ANOTHER INTERPRETATIONOF THE TEST STATISTIC,W

The hypothesisthat aoip=0 Vi is violatedif and only if some linearcombination of the a's is zero; i.e., if and only if some portfolio of the N assets has a

All use subject to JSTOR Terms and Conditions

1142

With this in mind, it is interestingto considerthe portfolio which, in a given

sample, maximizesthe squareof the usual t statisticfor the intercept.It is well

known in the literatureon multivariatestatistics that this maximumvalue is

Hotelling'sT2 statistic,our TW,. In this sectionwe focus on the compositionof

the maximizingportfolio, a, and its economicinterpretation.

Thus, let Frat a'F2, where F2, is an N x 1 vector with typical element ?,

Vi-1,..., N. Let a&be the N x 1 vectorof regressioninterceptestimates.Then

aa a a'&t and VAR( a) = (1 + 0 )a'>a/T

by (4) above.21Therefore,

f

(A'&)2AY

[&A]2

(13)

t2 =

[S(^)]=

AA

adopt the normalizationthat a'&t= c, where c is any constant differentfrom

zero. With this normalization,T(a'&t)2 and (1 + 0p2) in (13) are fixed given the

sample. Hence, maximizingt2 is equivalentto the followingminimizationproblem:

a'Ta

min:

a

subjectto:

a'ap =c.

the solution is:

a

ap

2_

a

1

+ 02

p

distributionof ta is not Student t, for portfolio a was formed after examining

the data.

The derivation of t2 suggests some additional information to summarize

empiricalwork on ex ante efficiency.Given the actual value of a based on the

sample, one will know the particularlinear combination which led to the

rejectionof the null hypothesis.If the null hypothesisis rejected,then a may give

us some constructiveinformationabout how to createa bettermodel.

Portfolio a has an economicbasis as well. When this portfolio is combined

properlywith portfolio p, the combinationturns out to be ex post efficient.In

21

Since we are workingwith returnsin excessof the risklessrate, t' a need not equal 1, for the

risklessasset will be held (long or short)so that all wealthis invested.

All use subject to JSTOR Terms and Conditions

PORTFOLIO EFFICIENCY

1143

(14)

c so that the samplemeans of rpt,rat, and Ft*are all equal. The equivalenceof

these three means requiresthat:

ap

aCf

a2rp

- Ip

r2

For the remainderof the paper, we will refer to portfolio a as the "active"

portfolio. In many applicationsof our methodology, portfolio p will be a

"passive"portfolio, i.e., a buy-and-holdinvestmentstrategy.Whileour methods

are applicable to situationswhere portfolio p is not passive, certainlyin its

applicationto tests of the CAPM,portfoliop will be passive.In such a setting

portfolio a is naturallyinterpretedas an activeportfolio,for it representsa way

to improve the efficiency of portfolio p. The terminologyof "active" and

"passive"has been used by Treynorand Black(1973), amongothers.

To establishthis relationbetweenthe ex post efficientportfolioand portfolios

a and p, we firstrecallthe equationfor the weightsof an efficientportfolio, w*.

Using equation(22) in the Appendixand setting m in that equationto rp:

(15)

w*

V can be parameterizedas:

V- X

2tS

Appendix)on the last expressionand substitutingthis into equation(15) for V-1,

algebra.

equation(14) can be derivedafter some tedious,but straightforward,

The previous paragraphshave established that the square of the usual t

statistic for the estimatedintercept, &a, equals the T2 test statistic, TW,. A

similarresultcan be establishedas well for the ex post efficientportfoliowith the

same sample mean as portfoliop, i.e., the estimatedintercept,a*, from regressing * on rpt has a squaredt statistic,t*2, whichis identicalto t 2. Sincewe will

not use this resultin what follows,we only note the fact here withoutproof.22

To illustratethe usefulnessof the active portfoliointerpretation,we returnto

the example of Section 5 where the size effect (acrossall months)is examined.

The second columnof Table V is roughlyconsistentwith the findingsof Brown,

22 The two key facts used in the proof are that a* =

kaa and SD(9,*) = kSD(9a,), where 9,* is the

disturbancein the regressionof F,* on Fp,;these equalitieshold for the estimatesas well as the

parameters.Since t* is essentiallya ratio of &* and SD(E^*),k cancels.The two key facts can be

establishedby workingwith the momentsof Ft*basedon equation(14).

All use subject to JSTOR Terms and Conditions

1144

TABLE VII

INFORMATION

THE

AREBASED

DESCRIPTIVE

ABOUT ACTIVE

a. THESESTATISTICS

PORTFOLIO,

PORTFOLIOS

USINGMONTHLY

1926-82 (T = 684). ALL SIMPLE

ON SIZE-SORTED

RETURNS,

EXCESS RETURNSARE NOMINALAND IN PERCENTAGE

FORM, AND THE CRSP

INDEXIS PORTFOLIO

VALUE-WEIGHTED

p.

Monthly Returns

in all Months

(T = 684)

&a

t2

ta

k

a,

a2

a3

a4

a.

a6

0.05

11.97

7.56

-0.02

0.04

0.04

0.00

0.06

0.08

Monthly Returns

in only January

(T-= 57)

1.36

71.57

0.87

Monthly Returns

Excluding January

(T = 627)

0.05

11.09

7.95

0.09

-0.04

0.07

0.08

0.07

-0.01

0.03

-0.02

0.03

-0.04

0.12

0.06

0.01

0.10

0.10

0.04

0.03

0.09

aRF

0.04

0.60

0.12

-0.52

-0.16

1.25

0.01

0.63

0.09

Ea,

1.00

1.00

1.00

a7

a8

a9

al0

NOTE: Portfolio 1 is a value-weighted portfolio of firms whose market capitalization is in the lowest decile of the

NYSE while portfolio 10 contains firms in the highest decile. The portion of wealth invested in the riskless asset is

denoted by aRF.

Kleidon, and Marsh in that the estimated alphas are approximately monotonic in

the decile size rankings. However, such a result does not imply that an optimal

portfolio should give large weight to small firms. As Dybvig and Ross (1985)

point out, alphas only indicate the direction of investment for marginal improvements in a portfolio. The portfolio that is globally optimal may have a very

different weighting scheme than is suggested by the alphas. A comparison of

Tables V and VII verifies this.

For example, the portion of the active portfolio invested in the portfolio of the

smallest firms (i.e., a,) has a sign which is opposite that of its estimated alpha.

Furthermore, the active portfolio suggests spreading one's investment fairly

evenly across the portfolios in the bottom 9 deciles and then investing a rather

large proportion in the portfolio of large firms, not small firms. Table VII also

reports a '2, and k for the overall period. Note that as k is much greater than

one (k = 7.56), the ex post efficient portfolio has a huge short position in the

value-weighted index. Since this index is dominated by the largest firms, the net

large firm position in the efficient portfolio is therefore actually negative. It is

interesting that ex post efficiency is achieved by avoiding (i.e., shorting) large

firms rather than aggressively investing in small firms.

The reader should keep in mind that Tables IV through VI and the second

column of Table VII have examined the size effect across all months. Based on

just these results, the size effect seems to be less important than perhaps

originally thought. However, if the data are sorted by January returns versus

All use subject to JSTOR Terms and Conditions

1145

PORTFOLIO EFFICIENCY

TABLE VIII

(T= 57). ALL SIMPLE EXCESS RETURNS ARE NOMINAL AND IN PERCENTAGE FORM, AND THE

CRSP VALUE-WEIGHTED INDEX IS PORTFOLIO p. THE FOLLOWING PARAMETER ESTIMATES

ARE FOR THE REGRESSIONMODEL: Fit= aip +3,,iprpt + ,it Vi=.

10, WHERE R2 IS THE

COEFFICIENT OF DETERMINATION FOR EQUATION i.

Portfolio

Number

a,p

s(&,p)

3,,p

1

2

3

4

5

6

7

8

9

10

6.12

4.60

3.43

2.88

1.79

1.79

0.85

0.80

0.31

-0.52

0.87

0.67

0.47

0.49

0.34

0.30

0.20

0.23

0.17

0.10

1.67

1.52

1.44

1.44

1.19

1.21

1.16

1.17

1.05

0.94

s(I,P)

0.18

0.14

0.10

0.10

0.07

0.06

0.04

0.05

0.03

0.02

RI

0.62

0.69

0.80

0.80

0.85

0.88

0.94

0.92

0.95

0.98

NOTE:

Portfolio 1 is a value-weighted portfolio of firms whose market capitalization is in the lowest decile of the

NYSE while portfolio 10 contains firms in the highest decile. For this sample period #P and (* are 0.259 and 1.197,

respectively. These imply a value for WI,equal to 1.256, which has a p-value of 0.000. Under the hypothesis that the

CRSP Value-Weighted Index is efficient, 4(W ,) is 0.219 and SL(W,,) is 0.111.

size effect-at least for the monthof January.TableVIII summarizesthe sample

characteristicsof our 10 size-sortedportfolioswhenusing only returnsin January

from 1926 through1982.A comparisonof TablesV and VIII revealsthat the size

effect is much more pronouncedin January than in other months; this is

consistent with the work by Keim (1983). This impressionfrom the univariate

statisticsis confirmedby the multivariatetest of ex ante efficiency,for the F test

is 5.99 with a p-value of zero to three decimal places. In contrast,the F test

based on all months excludingJanuaryis 1.09 with a p-value of 0.36. The

weights of the activeportfolio, a, are presentedin the last two columnsof Table

VII for Januaryversusnon-Januarymonths.As in the firstcolumnof Table VII,

the active portfolio is not dominatedby small firms.For the month of January,

one's investment should be evenly spread (roughlyspeaking)across the eight

portfoliosin the bottom deciles(or smallerfirms);however,firmsin the top two

deciles (or largerfirms)shouldbe shorted.23Resultsfor non-Januarymonthsare

similarto those based on all monthlydata.

The evidence in Table VII suggests that the optimal active portfolio is not

dominatedby smallfirmsevenin the monthof January-at least based on the ex

post sample moments.Nevertheless,in the marketplacewe see the development

of mutual funds which specializein holding the equities of just small firms.24

23

The active portfolio for the month of January involves a rather large position in the riskless asset

equals 1.25). This investment in the riskless security is necessary to maintain a sample mean

return on the active portfolio equal to that of the CRSP Value-Weighted Index.

24

Examples of such funds include The Small Company Portfolio of Dimensional Fund Advisors

and the Extended Market Fund of Wells Fargo Investment Advisors.

(CRF

All use subject to JSTOR Terms and Conditions

1146

indexes like the S&P 500 (or the CRSP Value-Weighted Index) with a portfolio

of small firms. We now turn to an examination of the ex ante efficiency of such a

linear combination in the next section. A multivariate statistical test of such an

investment strategy turns out to be a simple extension of the test developed in

Section 2.

7. TESTING THE EFFICIENCY OF A PORTFOLIO OF L ASSETS

restrictions on the joint distribution of excess returns similar to those considered

earlier. Specifically, if P = E 1xj/j, (where ELlxj = 1) and if Fr is efficient,

then

L

(16)

(r it)=

E Sii

j=1

(Fir),

L

(17)

'Pit=

io + E 8ijryt + r1it

vi

N.

j=1

(We will assume that the stochastic characteristics of nit are the same as those of

git in equation (1).) Conversely, (16) implies that some portfolio of the given L

portfolios is on the minimum variance frontier (Jobson and Korkie (1982)). Thus,

a necessary condition for the efficiency of a linear combination (

r2t,..., rLF)

with respect to the total set of N + L risky assets is:

(18)

Ho:

3io= 0

Vi = 1,..., N.

The above null hypothesis follows when the parameter restriction given by (16) is

imposed on (17).

In this case, [T/N][(T- N - L)/(T- L - 1)](1 + Q2-Irp) l 2o'- 8 has a

noncentral F distribution with degrees of freedom N and (T - N - L), where

is a vector of sample means for rpt (i1t, 2t,... Lt), Q is the sample variancecovariance matrix for rpt, So has a typical element Sio, and So is the least squares

estimator for 83 based on the N regression equations in (17) above. Further, the

noncentrality parameter is given by [T/(1 + I-2

80)T-1 So. Under the null

hypothesis (18), the noncentrality parameter is 0.

For an application of the methodology developed in this section, we return to

the results based on the size-sorted portfolios using returns only during the

month of January. In the previous section, we found that we could reject the ex

ante efficiency of the CRSP Value-Weighted Index. It could be that there exists a

linear combination of the lowest decile portfolio and the Value-Weighted Index

which is efficient. To consider such a case, we set L = 2 and N = 9. (Since

portfolio 1 has become a regressor in a system like (17), we can no longer use it

as a dependent variable.) The F statistic to test hypothesis (18) is 1.09 with a

All use subject to JSTOR Terms and Conditions

PORTFOLIO EFFICIENCY

1147

levels of significance.Of course,this inferenceignoresthe obvious pre-testbias.

Throughoutthis paper we have assumedthat there is an observableriskless

rate of return,in which case the efficientfrontieris simply a line in mean-standard deviation space. Suppose,now, that we wish to determinewhethera set of

L + 1 portfolios (L ? 1) spans the minimum-variancefrontier determinedby

these portfoliosand the N otherassets.The N + L + 1 asset returnsare assumed

to be linearlyindependent.If we observethe returnon the "zero-beta"portfolio

(which in practice we do not), this spanninghypothesis(with L = 1) naturally

arisesin the context of the zero-betaversionof the CAPMdue to Black(1972).25

To formulate the test for spanning for any L> 1, consider the system of

regressionequations,

L+1

(19)

(19

311,1

Rk=10

= sio + E, sji

Rit

jt

= 1,...,5N,

Vii=,***,N

it

j=1

where kit denotes total returns, not excess returns. Hubermanand Kandel

(1985a) observe that the spanning hypothesis is equivalent to the following

restrictions:

(20)

Vi=

3io=0

... ,N

and

L+1

(21)

E

j=1

sij=l

Vi1

N.

of the returnson portfolio L + 1, we derive (17). Thus, the problemof testing

(20) in the context of (17) is identicalto that of testing (18) in the risklesscase

above. All we have learnedabout testingthe risklessasset case is equallyrelevant

to the spanningproblem,providedthat "excess returns"are interpretedappropriately. Perhapsmost importantly,the exact distributionof our test statisticis

known under both the null and alternativehypotheses,permittingevaluationof

the power of the test. Note that this test of spanningimposes (21) and then

assesses whether the interceptsin the resultingregressionmodel are equal to

zero.26 In contrast,Hubermanand Kandel(1985a)proposea joint F test of (20)

25

Moregenerally,supposethe L + 1st portfoliois uncorrelatedwith each of the first L portfolios

and has minimumvarianceamong all such orthogonalportfolios.A simple generalizationof the

for all i. It then follows

argumentin Fama (1976, page 373) establishesthat 8,L+l=1:=

(details are availableon request)from the resultsof Hubermanand Kandel(1985a)that the L + 1

frontierif and only if some combinationof the first L

portfolios span the minimum-variance

portfoliosis on the frontier.Thus,a test of the latterhypothesiscan be conductedas in this section

orthogonalportfoliois observable.

providedthat the minimum-variance

26 An intermediateapproachwouldbe to firsttest (21) directlyand then,providedthe null is not

rejected,proceedto test (20).Onceagain,the test of (21) is an F test, and the exactdistributionunder

the alternativemay be determinedalong the lines of our earlieranalysis.Of course,this test statistic

does requirethat we observethe returnon the L + 1 spanningportfolios.

All use subject to JSTOR Terms and Conditions

1148

statistic has not been studied under the alternative.

8. SUMMARYAND FUTURE RESEARCH

While this paper focuses on a particular hypothesis from modern finance, this

apparently narrow view is adopted to gain better insight about a broad class of

financial models which have a very similar structure to the one that we examine.

The null hypothesis of this paper is a central hypothesis common to all risk-based

asset pricing theories.27The nature of financial data and theories suggests the use

of multivariate statistical methods which are not necessarily intuitive. We have

attempted to provide some insight into how such tests function and to explain

why they may provide different answers relative to univariate tests that are

applied in an informal manner. In addition, we have studied the power of our

suggested statistic and have isolated factors which will change the power of the

test. There are at least two natural extensions of this work, and we now discuss

each in turn.

First, the multivariate test considered here requires that the number of assets

under study always be less than the number of time series observations. This

restriction is imposed so that the sample variance-covariance matrix remains

nonsingular. A test statistic which could handle situations with a large number of

assets would be interesting.28

Second, we have not been very careful to specify the information set on which

the various moments are conditioned. Gibbons and Ferson (1985), Grossman and

Shiller (1982), and Hansen and Singleton (1982, 1983) have emphasized the

importance of this issue for empirical work on positive models of asset pricing.

Our methods provide a test of the ex ante unconditional efficiency of some

portfolio-that is, when the opportunity set is constructed from the unconditional moments, not the conditional moments. When the riskless rate is changing

(as it is in all of our data sets), then our methods provide a test of the conditional

efficiency of some portfolio given the riskless rate. Of course, such an interpretation presumes that our implicit model for conditional moments given the riskless

rate is correct. Ferson, Kandel, and Stambaugh (1987) and Shanken (1987a)

provide more detailed analysis of testing conditional mean-variance efficiency.29

27 If there is no riskless asset, then the null hypothesis becomes nonlinear in the parameters, for the

intercept term is proportional to (1 - fl,p). Gibbons (1982) has explored this hypothesis using

statistics which only have asymptotic justification. These statistics have been given an elegant

geometric interpretation by Kandel (1984). While we still do not have a complete characterization of

the small sample theory, Shanken (1985, 1986) has provided some useful bounds for the finite sample

behavior of these tests.

2X See Affleck-Graves and McDonald (1988) for some preliminary work on this problem.

29

As Hansen and Richard (1987) emphasize, efficiency relative to a given information set need not

imply efficiency relative to a subset. This implication does hold given some additional (and admittedly

restrictive) assumptions, however. Let the information set, I, include the riskless rate, and let p be

efficient, given I. Assume betas conditional on I are constant and 9(Pi, Irp, I) is linear in rp, It

r, Irp, Rf,) = lprpt, where Rf, is

follows that (,t Ip, I) = iprpt, and by iterated expectations

the riskless rate. Thus, p is on the minimum-variance frontier, given Rf,, and the methods of this

paper are applicable.

All use subject to JSTOR Terms and Conditions

PORTFOLIO EFFICIENCY

1149

School of Organization and Management, Yale University, New Haven, CT,

U.S.A.,

and

Business

Simon School of

Administration, University of Rochester, Rochester,

NY, U.S.A.

ManuscriptreceivedApril, 1986; final revision receivedNovember,1988.

APPENDIX

DERIVATION OF EQUATION (7)

To understand the derivation of (7), first consider the basic portfolio problem:

min:

w'Vw

subject to w'r = m, a mean constraint, where w the vector of N + 1 portfolio weights; V the

variance-covariance matrix of N + 1 assets; and r the vector of N + 1 sample mean excess returns.

Without loss of generality, we assume that p itself is the first component of our excess return vector.

Thus, r' = (r, i) where -5 is a column vector of mean excess returns on the original N assets. The

first-order conditions for this problem are:

(22)

qV-

and

m

2

mean

m2

standard deviation

w'Vw

m2

Fr t

rV

= rV-lT

=

0*2

(23)

'

=*2

02

where in contrast to the rest of the paper S is now the maximum likelihood estimator. The last

equality follows from rewriting the elements of V in terms of sp, 4p,and S and then finding Vusing the formula for a partitioned inverse. These steps lead to:

(24)

i[s

_ ;1/.

I]

2-

lr=

(-2/s2)

[(

)-

)] .

All use subject to JSTOR Terms and Conditions

1150

- fpp and since the first term on the left-hand side of the above equation is 0*2 and the

Since

=2

first term on the right-hand side is 9p,we can rewrite the last equation as:

0*2

- 02

? &p

p p

- "&

p

or

ap

2Cp

ap &=

*2

_ p2

Thus,

W=

0*2 _2

I P

-2

REFERENCES

AFFLECK-GRAVES,J., AND B. McDONALD (1988): "Multivariate Tests of Asset Pricing: The Compar-

ANDERSON, T. W. (1984): An Introductionto Multivariate Statistical Analysis, Second Edition. New

BANZ, R. W. (1981): "The Relationship between Return and Market Value of Common Stock,"

BERNDT, E. R., AND N. E. SAVIN (1977): "Conflict among Criteria for Testing Hypotheses in the

BLACK, F. (1972): "Capital Market Equilibrium with Restricted Borrowing," The Journal of Business,

45, 444-454.

BLACK, F., M. C. JENSEN, AND M. SCHOLES (1972): "The Capital Asset Pricing Model: Some

Empirical Findings," in Studies in the Theoryof Capital Markets, ed. by M. C. Jensen. New York:

Praeger.

BREEDEN, D. T., M. R. GIBBONS, AND R. H. LITZENBERGER(1987): "Empirical Tests of the

Consumption-Oriented CAPM," Research Paper #879, Graduate School of Business, Stanford

University, Stanford, CA.

BREMNER,J. M. (1978): "Mixtures of Beta Distributions, Algorithm AS 123," Applied Statistics, 27,

104-109.

BROWN, P., A. W. KLEIDON, AND T. A. MARSH (1983): "New Evidence on the Nature of Size-Related

Anomalies in Stock Prices," Journal of Financial Economics, 12, 33-56.

DYBVIG, P. H., AND S. A. Ross (1985): "The Analytics of Performance Measurement Using a

Security Market Line," Journal of Finance, 40, 401-416.

ELTON, E. J., AND M. J. GRUBER (1973): "Estimating the Dependence Structure of Share PricesImplications for Portfolio Selection," Journal of Finance, 28, 1203-1232.

ELTON, E. J., M. J. GRUBER, AND T. J. URICH (1978): "Are Betas Best?" Journal of Finance, 33,

1375-1384.

FAMA, E. F. (1976): Foundationsof Finance. New York: Basic Books.

FAMA, E. F., AND J. D. MAcBETH (1973): "Risk, Return, and Equilibrium: Empirical Tests," Journal

of Political Economy, 81, 607-636.

FERSON, W., S. KANDEL, AND R. STAMBAUGH(1987): "Tests of Asset Pricing with Time-Varying

Expected Risk Premiums and Market Betas," Journal of Finance, 42, 201-220.

GIBBONS, M. R. (1980): "Estimating the Parameters of the Capital Asset Pricing ModelA Minimum Expected Loss Approach," Research Paper #565, Graduate School of Business,

Stanford University, Stanford, CA.

(1982): "Multivariate Tests of Financial Models: A New Approach," Journal of Financial

Economics, 10, 3-28.

GIBBONS, M. R., AND W. FERSON(1985): "Testing Asset Pricing Models with Changing Expectations

and an Unobservable Market Portfolio," Journal of Financial Economics, 14, 217-236.

GRAYBILL, F. A. (1983): Matrices with Applications in Statistics, Second Edition. Belmont, CA:

Wadsworth.

All use subject to JSTOR Terms and Conditions

PORTFOLIO EFFICIENCY

1151

GROSSMAN,

Economies with Non-Traded Assets and Heterogeneous Information," Journal of Financial Economics, 10, 195-210.

HANSEN, L., AND S. RICHARD (1987): "The Role of Conditioning Information in Deducing Testable

Restrictions Implied by Dynamic Asset Pricing Models," Econometrica, 55, 587-614.

HANSEN, L., AND K. SINGLETON(1982): "Generalized Instrumental Variables Estimation of Nonlinear Rational Expectations Models," Econometrica,50, 1269-1286.

(1983): "Stochastic Consumption, Risk Aversion, and the Temporal Behavior of Stock Market

Returns," Journal of Political Economy, 91, 249-265.

HUBERMAN, G., AND S. KANDEL (1985A): "Likelihood Ratio Tests of Asset Pricing and Mutual Fund

Separation," Graduate School of Business, University of Chicago, Chicago, IL.

(1985b): "A Size Based Stock Returns Model," Graduate School of Business, University of

Chicago, Chicago, IL.

IBBOTSON, R., AND R. SINQUEFIELD(1979): Stocks, Bonds, Bills and Inflation: Historical Returns

(1926-1978). Charlottesville, VA: Financial Analysts Research Foundation.

JOBSON, J. D., AND B. KORKIE (1982): "Potential Performance and Tests of Portfolio Efficiency,"

Journal of Financial Economics, 10, 433-466.

(1985): "Some Tests of Linear Asset Pricing with Multivariate Normality," Canadian Journal

of Administrative Sciences, 2, 114-138.

JOHNSON, N. L., AND S. KOTZ(1970): Continuous Univariate Distributions, Volume 2. New York:

John Wiley.

KANDEL, S. (1984): "The Likelihood Ratio Test Statistic of Mean-Variance Efficiency without a

Riskless Asset," Journal of Financial Economics, 13, 575-592.

KANDEL, S., AND R. STAMBAUGH(1987): "On Correlations and Inferences about Mean-Variance

Efficiency," Journal of Financial Economics, 18, 61-90.

KEIM, D. B. (1983): "Size-Related Anomalies and Stock Return Seasonality: Further Empirical

Evidence," Journal of Financial Economics, 12, 13-32.

LINTNER, J. (1965): "The Valuation of Risk Assets and the Selection of Risky Investment in Stock

Portfolios and Capital Budgets," Review of Economics and Statistics, 47, 13-37.

MAcBETH, J. (1975): "Tests of Two Parameter Models of Capital Market Equilibrium," Ph.D.

Dissertation, Graduate School of Business, University of Chicago, Chicago, IL.

MACKINLAY, A. C. (1985): "An Analysis of Multivariate Financial Tests," Ph.D. Dissertation,

Graduate School of Business, University of Chicago, Chicago, IL.

(1987): "On Multivariate Tests of the CAPM," Journal of Financial Economics, 18, 341-372.

MALINVAUD, E. (1980): Statistical Methods of Econometrics, Third Edition. Amsterdam: North

Holland Publishing.

MERTON, R. (1973): "An Intertemporal Capital Asset Pricing Model," Econometrica, 41, 867-887.

MORRISON, D. F. (1976): MultivariateStatistical Methods, Second Edition. New York: McGraw-Hill.

RAo, C. R. (1951): "An Asymptotic Expansion of the Distribution of Wilks Criterion," Bulletin of the

Institute of International Statistics, 33, 33-38.

REINGANUM, M. R. (1981): "Misspecification of Capital Asset Pricing: Empirical Anomalies Based

on Earnings Yields and Market Values," Journal of Financial Economics, 9, 19-46.

ROLL, R. (1977): "A Critique of the Asset Pricing Theory's Tests-Part 1: On Past and Potential

Testability of the Theory," Journal of Financial Economics, 4, 129-176.

(1978): "Ambiguity When Performance is Measured by the Securities Market Line, Journal

of Finance, 33, 1051-1069.

SCHWERT, G. W. (1983): "Size and Stock Returns, and Other Empirical Regularities," Journal of

Financial Economics, 12, 3-12.

SHANKEN, J. (1982): "An Analysis of the Traditional Risk-Return Model," Ph.D. Dissertation,

Graduate School of Industrial Administration, Carnegie-Mellon University, Pittsburgh, PA.

(1985): "Multivariate Tests of the Zero-Beta CAPM," Journal of Financial Economics, 14,

327-348.

(1986): "Testing Portfolio Efficiency when the Zero-Beta Rate Is Unknown: A Note," Journal

of Finance, 41, 269-276.

(1987a): "The Intertemporal Capital Asset Pricing Model: An Empirical Investigation,"

Working Paper, University of Rochester.

(1987b): "Multivariate Proxies and Asset Pricing Relations: Living with the Roll Critique,"

Journal of Financial Economics, 18, 91-110.

SHARPE, W. F. (1963): "A Simplified Model of Portfolio Analysis," Management Science, 9, 277-293.

All use subject to JSTOR Terms and Conditions

1152

(1964): "Capital Asset Prices: A Theory of Market Equilibrium under Conditions of Risk,"

Journal of Finance, 19, 425-442.

(1982): "Factors in New York Stock Exchange Security Returns, 1931-79," Journal of

Portfolio Management, 8, 5-19.

SILVEY,S. D. (1975): Statistical Inference. London: Chapman and Hall.

R. F. (1982): "On the Exclusion of Assets from Tests of the Two-Parameter Model:

STAMBAUGH,

A Sensitivity Analysis," Journal of Financial Economics, 10, 237-268.

TANG, P. C. (1938): "The Power Function of the Analysis of Variance Tests with Tables and

Illustrations of Their Use," Statistical Research Memoirs, 2, 126-150.

TIKU, M. L. (1967): "Tables of the Power of the F-Test," Journal of the American Statistical

Association, 62, 525-539.

TREYNOR,J. L., AND F. BLACK(1973): "How to Use Security Analysis to Improve Portfolio

Selection," Journal of Business, 46, 66-86.

All use subject to JSTOR Terms and Conditions

- Risk & ReturnUploaded byajeet090
- Market Efficiency: Finance week 09Uploaded byIRPS
- mat133Uploaded byBrian Griffin
- Multinational Business Finance by Shapiro Chapter 1 SolutionsUploaded byEdward Yang
- Practice Problems for FINN3222_Midterm_Fall2011Uploaded byangelbear2577
- Agriculture In IndiaUploaded byAjay Sahu
- Appendix_W_-_Fama_French_Model.pdfUploaded byjamilkhann
- Question Bank.docx for MbaUploaded byneeshusharma
- Credit Ratings and The Cross-Section of Stock ReturnsUploaded bymirando93
- Wind Forecasting Using Artificial Neural NetworksUploaded byMumawaddah Rizqiyatul Hayyah
- Investment ManagementUploaded byRonak Koticha
- MCQ's for Midterm Test -7th April 2012Uploaded byNalin Indika Kumara
- 3. Finance - Ijfm - Empirical Testing of Capm in - SeethanaikUploaded byiaset123
- ReportUploaded byJesa Singhbal
- Assignment 2Uploaded bymarc
- FM 8th Edition Chapter 12 - Risk and ReturnUploaded byKa Io Chao
- Artigo 2 - r2 BaixoUploaded byCatarina Silva
- 318072Uploaded byLuisMendiola
- Assuring Not-For-Profit Hospital cost of capital.pdfUploaded byAnonymous Sl6O21UT
- 4_Standard Valuation Models (1)Uploaded bychirag
- Lecture 11 With Lecture NotesUploaded byAlex
- Cost of CapitalUploaded byHazem Ali
- Cost of CapitalUploaded byAkshatra Pareek
- Cost of CapitalUploaded byAkshatra Pareek
- Chap 008Uploaded bysarahjohnson
- Lecture 10 Beta2 Intro Stat of PortUploaded bydrlebedovich
- Internet SecurityUploaded byKhoa Nguyen
- Nike Inc Case4Uploaded byRegine Marie Talucod
- 8 25 2009Uploaded bybentleyinvestmentgroup
- SyllbussUploaded byJuan Miguel Gaddi

- What Your Employees Really ThinkUploaded byAzeriboye
- Asimov AlgebraUploaded byChichiMiguel
- Study Guide for the Risk Managm Abdulla AlkuwaitiUploaded byAzeriboye
- Valentine Ch20 Best Practices Equity SampleUploaded byharpoon11
- Study Guide for the Risk Managm Abdulla AlkuwaitiUploaded byAbdullah Alrimawi
- Excel 2010 ShortcutsUploaded bywhackz
- 2015 Level I OnDemandVideo SS2 R6 HandoutUploaded byAzeriboye
- F5 Progress Test Question 1Uploaded byAzeriboye
- Cyber AttackUploaded byAzeriboye
- Resource 1385Arshad Resource 1385Arshad Resource 1385Arshad Resource 1385Arshad Resource 1385Arshad Resource 1385Arshad Resource 1385Arshad Resource 1385Arshad Resource 1385Arshad Resource 1385Arshad Resource 1385Arshad Resource 1385Arshad Resource 1385Arshad Resource 1385Arshad Resource 1385Arshad Resource 1385Arshad Resource 1385Arshad Resource 1385Arshad Resource 1385ArshadUploaded byAzeriboye
- Calculation+of+Security+Market+IndicesUploaded byAzeriboye
- Distinguished Lecture by Jörg AsmussenUploaded byAzeriboye
- Us Crisis Management Corporate GuidelinesUploaded byAzeriboye
- Phi27106 Fm i Xxxiv1Uploaded byAzeriboye
- Direct FinanceUploaded byAzeriboye
- Strategic FrameworkUploaded byAzeriboye
- IAS 7IAS 7IAS 7IAS 7IAS 7IAS 7IAS 7IAS 7IAS 7IAS 7IAS 7IAS 7IAS 7IAS 7Uploaded byAzeriboye
- IAS 2Uploaded byAzeriboye
- 360 Degree Feedback sampleUploaded byAzeriboye
- Calculus I (E-book Az)Calculus I (E-book Az)Calculus I (E-book Az)Calculus I (E-book Az)Calculus I (E-book Az)Calculus I (E-book Az)Calculus I (E-book Az)Calculus I (E-book Az)Calculus I (E-book Az)Calculus I (E-book Az)Calculus I (E-book Az)Calculus I (E-book Az)Calculus I (E-book Az)Calculus I (E-book Az)Calculus I (E-book Az)Calculus I (E-book Az)Calculus I (E-book Az)Calculus I (E-book Az)Calculus I (E-book Az)Calculus I (E-book Az)Calculus I (E-book Az)Calculus I (E-book Az)Calculus I (E-book Az)Calculus I (E-book Az)Uploaded byAzeriboye
- Math NotesMath NotesMath NotesMath NotesMath NotesMath NotesMath NotesMath NotesMath NotesMath NotesMath NotesMath NotesMath NotesMath NotesMath NotesMath NotesMath NotesMath NotesUploaded byAzeriboye
- F1 - Answers - May 2011Uploaded byAzeriboye
- Equilar 2013 Ceo Pay Stategies ReportUploaded byAzeriboye
- f1 Country x Tax May 2011Uploaded byYahan Samarajeewa
- efa-paper-101213Uploaded byAzeriboye
- CIMA f1 f2 Syllabus ChangeCIMA f1 f2 syllabus changesUploaded byAzeriboye
- calculusUploaded byOladeleIfeoluwaOlayode
- C1Uploaded bykhurramsundrani

- ESL PaperUploaded byaini71
- Key Instruction WordsUploaded bydinhchungxd
- Basic Queuing TheoryUploaded byStefan Ristevski
- Actual Post Test Development, Validation and Acceptability of Guidebook in Writing Investigatory ProjectsUploaded byCiv Nortub
- multicultural dance webquestUploaded byapi-398534676
- CV Theodorus Garry PG 178115163Uploaded byTheodorus Garry Putra Gana
- Predictive Modeling Sample ProgramUploaded bysiva5256
- Patient Safety in EMS Full ReportUploaded byAlvi Rahmi
- APBHP MethodologyUploaded byMarvin Pulao
- Module 1 Doing Scientific InvestigationUploaded byLJ Valdez
- Quality of Life in Persons with Intellectual Disabilities and Mental Health Problems: An Explorative StudyUploaded byRaquelMaia
- THE DETERMINATION OF WAX ESTERS IN LUBE OIL BY INFRARED SpectroscopyUploaded byYousry ElToukhee
- Acceptance CriteriaUploaded byNitish Kumar
- CAR DLL wk 1Uploaded byjoana torculas
- cer graphic organizer lesson planUploaded byapi-324337073
- Assignment for Marketing -Uploaded byTharanga Kariyawasam
- Applied Predictive Modeling - Max KuhnUploaded byDavid
- engg-s-2012Uploaded bybhavesh512
- apcoUploaded byArya Ramesh
- EDU 5170LessonPlanUploaded byCody Kober
- The Origins and Practice of Participatory Rural.pdfUploaded byandres bautista
- Colin Gardner - Meta-Interpretation and Hypertext Fiction A Critical Response.pdfUploaded byKlro
- Ginger for Nausea and Vomiting in Pregnancy .17Uploaded byMarnia Sulfiana
- NICE Guidelines for Dental RecallUploaded bynaomi
- 2014-2020 Programming Period_ex_ante Assessment MethodologyUploaded byfionarml4419
- Force Field AnalysisUploaded byCanumay Jbc Jennely
- 30-1-108Uploaded byamin_zuhdi
- CV SELLYUploaded bySelly Purnama Nduth
- Pios Goi Mhrd as May 25 2017Uploaded byAnonymous n4UccmT6q
- Proprioceptive Processing Difﬁculties Among Children With Autism Spectrum Disorders and Developmental DisabilitiesUploaded byLuli Arrieta