This article was downloaded by: [Michigan State University]

On: 17 February 2015, At: 19:53

Publisher: Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House,
37-41 Mortimer Street, London W1T 3JH, UK

Journal of the American Statistical Association

Publication details, including instructions for authors and subscription information:
http://www.tandfonline.com/loi/uasa20

Estimation for Markowitz Efficient Portfolios

a a
J. D. Jobson & B. Korkie
a
Department of Finance and Management Science , University of Alberta , Edmonton , T6G
2G1 , Canada
Published online: 12 Mar 2012.

To cite this article: J. D. Jobson & B. Korkie (1980) Estimation for Markowitz Efficient Portfolios, Journal of the American
Statistical Association, 75:371, 544-554

To link to this article: http://dx.doi.org/10.1080/01621459.1980.10477507

PLEASE SCROLL DOWN FOR ARTICLE

Taylor & Francis makes every effort to ensure the accuracy of all the information (the Content) contained in the
publications on our platform. However, Taylor & Francis, our agents, and our licensors make no representations
or warranties whatsoever as to the accuracy, completeness, or suitability for any purpose of the Content. Any
opinions and views expressed in this publication are the opinions and views of the authors, and are not the
views of or endorsed by Taylor & Francis. The accuracy of the Content should not be relied upon and should be
independently verified with primary sources of information. Taylor and Francis shall not be liable for any losses,
actions, claims, proceedings, demands, costs, expenses, damages, and other liabilities whatsoever or howsoever
caused arising directly or indirectly in connection with, in relation to or arising out of the use of the Content.

This article may be used for research, teaching, and private study purposes. Any substantial or systematic
reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in any
form to anyone is expressly forbidden. Terms & Conditions of access and use can be found at http://
www.tandfonline.com/page/terms-and-conditions
 Estimation for Markowitz Efficient Portfolios
J.D. JOBSON and B. KORKIE*

Given a set of N assets a portfolio is determined by a set of weights where X is a n N X 1 vector representing the unknown
. xr=l
zi, i = 1, 2, . ., N ; zi = 1 indicating the proportion of the proportions invested in risky positive-variance assets ;
value of the portfolio devoted to each asset. A Markowitz efficient
portfolio is the vector of weights X, that minimizes the variance Z is the N X N covariance matrix of risky assets with
urn2of the total return from the portfolio, subject t o the condition rank T = N , the number of assets;
that the portfolio mean premium return p, has a certain value. p = (p* - eE,) is the N X 1 vector of asset mean-
The estimators for the N X 1 vector X ,, the return premium p,,,,
and the variable urn2require estimators for the mean premium return premiums, where p* is the mean-return vector,
return vector I and for the covariance matrix 2 . The expectations, e is the unit vector, and E , is the fixed return on the
variances, and asymptotic distributions of the estimators of X , riskless asset x ;and
p,, and urn2are derived under the assumption that returns are
normally distributed. The use of these sampling properties for p, is the desired mean premium return = (mean
statistical inference is also discussed. The derived results are also portfolio return -E,).
Downloaded by [Michigan State University] at 19:54 17 February 2015

compared with results obtained from a Monte Carlo simulation

for a population of 20 stocks and several sample sizes.
Solution of (1.1) gives the investment proportions or
weight vector X,, mean premium return p,, and vari-
KEY WORDS : Asymptotic distributions; Sampling distributions; ance urn2of the optimal portfolio m. (See Merton 1972
Taylor series approximations ; Monte Carlo simulation ; Distribu-
tions of ratios; Fieller-Creasy problem. and Roll 1977 for a summary.)
Xm = Ap/b = Fm/b 1

1. INTRODUCTION
pm =pX, = a / b , (1.2)
The theory of portfolio analysis involves the deter- am2= X,ZX, = a/b2 ,
mination of sets of assets that are efficient in a risk-
return space. Efficient portfolios are those combinations where A is the inverse of 8 , F, = A g is the nonstandard-
of assets that have maximum return for a given level of ized weight vector, a = pAp, and b = eAp.
risk or, alternatively, minimum risk for a given level of The portfolio m, for given E,, is a unique Markowitz
return. I n the Markowitz (1959) formulation of port- portfolio, which could be combined with investment in
folio analysis, the measures of return and risk are the the riskless asset z to produce portfolios that are termed
mean and variance of the portfolios returns. The objects Sharpe (1964) efficient. The solutions (1.2) to this
of choice are, therefore, the mean and variance because optimization problem do not depend on the existence
investors possess quadratic preference functions for re- of a zero-variance asset, with return E,. I n the absence
turn, or, alternatively, the distributions of asset returns of a zero-variance asset, E, can be thought of as the
are completely specified by their first two moments. mean return from a positive-variance portfolio, whose
The number of assets available to investors is nearly return is orthogonal to the return on the unique portfolio
limitless. The available assets range from riskless zero- m. I n the parlance of financial economics, E , is referred
variance securities (federal government bills and zero to as the return on the zero beta portfolio (see Black
coupon bonds held to maturity) to various other financial 1972 for a discussion).
securities and real assets. Most financial research has A major problem, which belies the implementation of
been restricted to financial assets and, more often, this normative theory of portfolio analysis, is the forma-
common stocks. tion of rational expectations regarding the mean-return
An efficient portfolio (allowing unrestricted short sales premium vector p and the covariance matrix that is
of assets) is determined by minimizing portfolio variance, appropriate for the investors holding period. I n this
subject to a mean portfolio premium return and the article, we assume that the returns from the N stocks
additional constraint that investment proportions in are stationary random time series, which are distributed
risky assets sum to one. The Lagrangian is as a multivariate normal with mean p and covariance
matrix 8 . Although the assumption of multivariate
minL = XzX - X1{Xp - p,) - X2{Xe - 1 ) , (1.1) normality is suspect for daily and weekly returns, the
X
distribution does not seem to be significantly different
* J.D. Jobson and B. Korkie are Associate Professors, Depart- from a normal for monthly returns. (See Fama 1976
ment of Finance and Management Science, University of Alberta, for a review of the evidence on normality and station-
Edmonton, Canada T6G 2G1. The authors are indebted to S.
Beveridge for his comments on inference for ratios of random
~~~

variables. I n addition the authors acknowledge the helpful com- @ Journal of the American Statistical Association
ments of T. Daniel, G.A. Whitmore, S. Tinic, P. Talwar, and September 1980, Volume 75, Number 371
a referee. Applications Section
544
 Jobson and Korkie: Markowitz Portfolio Estimation 545

arity.) Given that the time series of monthly returns is The columns of A will be denoted by A,, j = 1, . . ., N .
stationary, sample estimates of the mean-return vector The elements o f f , F,, X,, p, W, 8 , and A will be denoted
and covariance matrix may be obtained from past re- by ?,, F m , , XmL,p t , W,,, u,,, and A,,, respectivcly, i, j =
turns of the N assets under consideration. 1, 2, . . ., N . In subsequent sections exact cxpressions
Using a simulation approach for three stocks, Frank- are dcrivcd for thc expectations and variances of the
furter, Phillips, and Seagle (1971) conclude that, since estimators of the ratio components a, b, and Fm. I n
sampling error is so large, portfolios selected according addition, approximate exprcssions are obtained for the
to the Markowitz criterion are likely not more efficient expectations and variances of the estimators p,, & m 2 ,
than an equally weighted portfolio. For the two-asset and the elements of x,. Finally, the asymptotic dis-
case, Dickenson (1974a, b) indicates poor reliability of tributions of the estimators arc derived.
the estimators of the proportions vector and variance of
the global minimum-risk portfolio. I n addition, for the
two-asset case he provides preliminary indication of the 2.1 Expectations and Variances for 8, b, and F,,
bias in estimating the weight vector of Markowitz The estimators 6 and 8, are unbiased, since f and W
portfolios. Other research by Barry (1974, 1975, 1978) are unbiased and statistically indepcndent. The ex-
and Klein and Bawa (1976,1977) incorporated estimation pectation of ci is given by
risk in the problem with Bayesian procedures and
N N
investigated the effect of this risk on final portfolio
Downloaded by [Michigan State University] at 19:54 17 February 2015

choice.
E[ci] = E [ C C FtF7W;j]
I n this article we are concerned with the sampling N N
N
distributions and asymptotic properties for estimators
of the weight vector, mean, and variance of Markowitz i i T
efficient portfolios for arbitrary numbers of assets. The Thcrefore, ci is biased and thc bias is given by ( N / T ) ,
following section of the article derives the means, vari- independent of p and A.
ances, covariances, and asymptotic distributions of the It is shown in Marx and Hocking (1977) that the co-
estimators. The third section gives the results of a variance between the (i, j ) t h and ( k , 4)th element of V-'
Monte Carlo experiment. Section 4 discusses the effects can be written as
of changes in p, 8 , N , and E,, and Section 5 discusses
inference for the unknown parameters.

2. EXPECTATIONS, VARIANCES, AND ASYMPTOTIC

DISTRIBUTIONS OF THE ESTIMATORS wherc

Assume a random sample of T return premium ob- 2

kl = _____
servations on each of N stocks that is denoted by the (T - N - 1) (T - N - 4)
.
N X 1 vector rt, t = 1, 2, . . , T . Now assume that rt is and
multivariate normal with mean vector p and covariance
matrix Z, and denote the sample mean vector and co- (T - N - 2 )
variance matrix by
ke - --.
T
( T - N - l ) (T - N - 4)
f = (l/T) 1 rt Expressions for the variances of ci and b and the co-
t= 1
and variance matrix for 8, may therefore be derived. After
T
some simplification, this leads to the following formulas :
V = [1/(T - I)] C (rt - f ) (rt - f)' .
t= 1
c ( T - 2)(T - N - 2)
T o estimate X,, F,, p,, crmz, a, and b, replace and V ( b ) = --
T ( T - N - 1)(T - N - 4)
A = 8 - l in (1.2) by estimators based on f and V. It is
U C ( T- N - 2)
shown in Anderson (1958, p. 53) that f is an unbiased
estimator of p. Marx and Hocking (1977) have shown
+-(T - N - l ) ( T - N - 4)
that W = [ ( T - N - 2)/(T - l)]V-I is an unbiased
estimator of A. The sample estimators, of the popula-
tion parameters identified in (1.2), are therefore
+---( T - N b2(T -N)
- 1) (T - N - 4)
> (2.2)

ci = PWf , - 2)(T - N - 2)
Xjk(T
cov(P,jP,h) =
b ?'We ,
= T ( T - N - 1)(T - N - 4)
pm = d/b , -N)
6
,
2 = ci/b2 ,
(2.1)
+ (T -Fm,Fmk(T
N - 1) (T - N - 4)
P, Wf , - N - 2)
fi,
=

= Wr/8 . + (.T -aXjk(T

N - 1)(T - N - 4)
* (2.3)
 546 Journal of the American Statistical Association, September 1980

4a(T - 2) 2a2( T )
V(6)= +
T ( T - N - 4) T ( T - N - 4)

+ 2N(T N)
-
T 2 ( T - N - 4)
(2.4)
+-[-
b2
1 2 cov(6, b )
b
--
12V(6) oov(6,
b3 -I 6)

cov(6, 6) =
2b(T - 2)
T ( T - N - 4)
+ ( T - N -4>
2ab
(2.5)

- ~ ~=, -
C O V ( F b)
A t j e ( T - 2 ) ( T - N - 2)
T ( T - N - 1)(T- N - 4 )
- V[Sm2]=
[a+ ( N / T ) I 2 4V(6)
b4
__
[ b2 +--I
___ GS[V(6)I2
b4

2 cov(d, 6) -42V(6) cov(6, b)

+ ---___( T - N)bFm,
(T - N - l)(T - N -4>
+----
b4
-

lOV(d)V(6)
b b3
lGCcov(6, &)I2
1
+-( T a(A'je)(T - N
- N -
- 2)
1) ( T - N - 4) '
b2 + b2 1
+o(T-~) .
- 2Fmj(T - 2)
COV(F,,, 6) = -+- ( T - N - 4)
2aFm,
(2.6) Finally, the approximate expectations and variances of
Downloaded by [Michigan State University] at 19:54 17 February 2015

7
T(T - N - 4) X,, arc
where c = e'Ae.
E[rTrrn,]=--1 V(6) 3[V(b)I2]
Fm.[l+--+
For the remainder of the article these moments shall be b b2 b4
referred to as exact moments.
The higher moments for 6, b, and the elements of F,
can be approximated by assuming that the vector
[ d , $, F',] is multivariate normal. (We shall see in
Section 3 that this assumption is reasonable.) With this +0( T - 9
assumption, all other moments can be expressed as a
function of the moments given earlier. In Section 2.2
these moments are required to obtain approximations to
1
mean and variances of p,, km2,and the mean vector and
covariance matrix for X,.
+-
""fb2
C O V ( P~m, j ) SV(6) C O V (F~m, j )
-_____--
b
3V(Pmi)v(6) 5[cov(6, 27,j)]2
b3 I
2 2 Expectations and Variances for

Because Pm, km2,and fi,

A,Gmz, and x,
are ratios involving ti, $, and
1
+b;[ v(pmj)+ b2
--+
b2
~o(27-3)
1 .
8,, Taylor series expansions may be employed to ap-
proximate the moments of p,, am2,and fi, in terms of For the remainder of the article these moments of Em,
the moments of 6, b, and 8,. To obtain higher-order 6,2, and X, obtained earlier shall be referred to as
moments of 6, 6, and 8,, normality is assumed. The approximate moments.
resulting approximate expectations and variances are Correction for the bias in d produces new estimators
given by for p, and am2given by

[ d - ( N / T ) ] / b and [6 - (N/T)]/b2,
respectively. The expectation and variance for these

+-[-
1 -COV(ci,
b b
6) -3V(6) cov(6, 6 )
b3
]+ estimators can be found, from the expressions given for
0 ( P 3 )f i m and km2, by dropping all terms in ( N / T ) . The expres-
sion for p m is then similar in structure to Xmi,with
[6 - ( N / T ) ]in place of F m i .

-cov(6, 6) SV(6) cov(6, 6) 2.3 Derivation of Asymptotic Distributions

+2[a+ ( N I T ) ] ______
b2 [ b ' b3 1 Applying a theorem given in Rao (1973, p. 387), it
can be shown that the asymptotic distributions of the
estimators given by (2.1) are normal and asymptotically
unbiased. The variances and covariances for the
+ o ( T - ~ ). asymptotic distribution of [a, b, Ft,,,l, as T + a , are
 Jobson and Korkie: Markowitz Portfolio Estimation 547

given by 3. RESULTS OF A SAMPLING EXPERIMENT

4a 2d
A V(d) =-+-
T (T-N-4)
=Va1+Va2
A Monte Carlo simulation was conducted to compare
the sample properties of the estimators in (2.1) with
b2(T -N ) the corresponding theoretical properties obtained in
Section 2 .
(T - N - 1)(T -N -%
ac (T - N - 2)
'- + (T-N-1) _-
(T-N-4)
]=Vbl+Vb2
3.1 The Design

An asset population of size 20 was chosen by random

A cov(Prni,Pmk) selection from the population of common stocks, con-
="+[ (T--N) ( F m j ) (Fmk) tinuously listed on the New York Stock Exchange dur-
ing the period December 1949 to December 1975. A
T (T-N-l)(T-N-4)
total of 313 monthly return observations were obtained
akjk (T - N 2)
+-(T-N-1) -
(T-N-4)
2b 2ab
1 =Clik+C2jk (2.7) for each stock from the Centre for Research in Security
Prices (CRSP) financial data tape, where
. . ., 20
A cov(6, 8) =-+ j = 1, 2,
Downloaded by [Michigan State University] at 19:54 17 February 2015

=Cabl+Cab2
T (T-N-4) t = 1, 2, . . ., 313
A cov(6, P,J and
=".[ (T -N )Fmib
T~~ = the monthly effective return premium from stock
T (T - N - 1)(T - N -4)
j in month t ,
+ ( Ta(&e) (T-N-2)
--N 1)( T - N -4) 1
-
=Cbjl+Cbj2
pj1=
djl =
the closing price of j a t month end t , and
the divided per share paid in month t .
The expressions for the asymptotic variances, for p m The return was used to compute a population mean-
and dm2, and the asymptotic covariances of the elements return vector p and covariance matrix X of returhs from
of X, also can be written as the 20 stocks. These population parameters are shown in
Table 1, where the mean returns range from .50 to 1.82
and variances range from 23 to 178. The efficient set
parameters a and b and the nonstandardized weight
vector F, were computed, which then permitted the
calculation of the mean p, and variance urn2,as well as
the proportions vector X, for the Markowitz portfolio.
Initially the value of E , was arbitrarily set to zero. At a
later stage some results for the case E , = .6 percent are
also presented.
The population parameters were then used in the com-
putation of the theoretical moments defined in Section 2.
Finally, the simulation experiment employed the popula-
tion mean vector and covariance matrix to generate
sample observations of the parameters. For a fixed
number of stocks N = 20 and sample size T, ranging
from 60 to 1,000 return observations, 100 multivariate
normal random deviates with mean p and covariance Z
were generated. All computations in the simulation,
An examination of the expressions for the asymptotic including the inverse operation, were performed in
variances and covariances given shows that, unlike the double-precision FORTRAN. The multivariate normal
expressions in Section 2.2, terms of O(T-2) have been deviates were generated from subroutine GGNRM of
excluded. We shall see in Section 3 that inclusion of the the IMSL Subroutine Library.
terms of O(T-2) provides much greater accuracy for The sample estimates of the efficient set parameters
small T. a, b, the Markowitz portfolio mean return pm and
I n later sections all the moments derived in this sec- variance am2, the portfolio weight vectors X, and F,
tion-exact, approximate, and asymptotic-will be re- were computed by equations (2.1). The mean values and
ferred to as theoretical moments. Detailed derivations of variances of these statistics over the 100 trials werc
the expressions obtained throughout Section 2 are calculated and compared with their theoretical counter-
available from the authors on request. parts.
 548 Journal of the American Statistical Association, September 1980

1. Population Parameters for Simulation Experiment

Variable in %

Mean .50 .SO 1.10 1.74 1.82 1.11 .91 1.18 1.35 1.07 1.16 1.23 .81 1.18 .88 1.20 .72 1.16 .92 1.25

Covariance 53.64
matrix 6.60 29.84
19.84 16.68 82.88
34.14 20.66 48.01 178.07
6.32 6.48 18.66 27.54 118.09
5.76 11.83 21.04 19.28 26.29 57.07
16.92 8.43 22.16 32.35 23.88 20.24 52.05
15.26 9.26 16.21 26.87 12.41 11.69 15.29 48.25
9.57 10.80 16.26 18.29 14.16 15.23 12.10 9.69 29.80
10.12 11.22 18.94 22.39 23.13 16.27 17.74 9.37 11.21 35.12
10.33 9.59 21.59 21.75 31.03 13.72 17.95 8.59 13.05 22.58 47.64
18.89 8.76 27.01 41.73 13.08 19.29 21.39 14.42 13.83 12.96 16.56 65.62
8.45 13.50 8.80 17.34 5.40 7.78 9.58 9.86 7.27 7.92 5.97 7.92 23.51
14.58 14.06 23.31 42.93 20.36 21.53 26.36 11.34 16.74 17.62 19.75 23.10 12.03 51.20
14.64 16.47 17.40 26.27 9.88 11.33 16.24 13.33 11.42 10.70 9.26 11.55 14.27 16.42 28.72
14.34 8.76 22.27 30.30 14.33 13.21 15.17 16.96 8.21 12.57 13.46 25.84 8.54 14.72 12.20 56.03
27.85 14.93 36.71 65.96 17.07 13.45 25.64 32.08 15.70 16.21 20.46 35.76 15.22 26.17 19.87 32.30 109.46
41.43 47.56 12.32 24.80 21.66 20.61 21.51 18.76 26.41 14.23 25.60 24.34 24.46 50.78 131.75
Downloaded by [Michigan State University] at 19:54 17 February 2015

25.08 16.74 20.20

11.77 22.82 21.44 20.67 13.46 16.84 15.50 14.34 14.80 14.23 13.32 16.96 15.80 20.42 22.36 13.14 18.58 26.97 44.66
16.92 10.26 27.74 43.93 18.45 18.08 25.27 15.75 10.72 14.74 17.69 23.65 9.72 20.93 13.76 14.48 32.30 29.18 16.09 58.69

3.2 Comparison of Monte Carlo and Theoretical Moments tained for each statistic. The estimators of a, b and the
The means for 6 - N / T , 6,and the elements of 8, elements of 8, showed probabilities much larger than .10
are in general comparable to the theoretical means for for all sample sizes, while the estimators of p, and the
all values of T . In addition, their exact variances are elements of X, were well behaved for sample sizes of 300
comparable to the simulation variances for all sample or more. In the case of the estimator of urn2,the null
sizes. Comparison of the simulation variances of 6, 6, hypothesis of normality is only marginally acceptable a t
and the elements of 8, to their asymptotic counterparts, sample size 1,000 ( p = .13). For this estimator both the
however, demonstrated that sample sizes of 300 or more skewness and kurtosis were relatively large.
are required before the simulation and asymptotic For the ratio of two normally distributed random
variances are comparable. variables W = X/Y,Hayya, Armstrong, and Gressis
Tables 2a and b compare the theoretical means and (1975) conclude that W is approximately normal if
variances of p,, am2,and X,, with the simulation means ~ P X Y5~ .5, U Y / ~ Y< .19, and U X / ~ X > .09, where p x ,
and variances. For sample sizes of a t least 300, the p y , ux2, u y 2 ,and P X Y are the means, variances, and cor-

approximate means and variances of the estimators relation, respectively, of the joint distribution of X and
p,, am2,and s,,, j = 1, 2, . . ., 20, are comparable to Y.For our study the values of the coefficients of varia-
tion and correlations were computed for the terms 6, 6,
their simulation means and variances and, in general,
are much closer to the simulation values than the b2, and P m l , j = 1, 2, . . ., N , of the ratio estimators
asymptotic values. I n some cases the asymptotic values x,.
am,c+,~,and For the numerators ci and P,,, j = 1, 2,
are not comparable to the simulation values at T = 1,000. . . ., N , the coefficients of variation obtained were well
For small sample sizes the differences between the above the critical value of .09. For the denominator b,
approximate means and variances and their simulation the coefficients of variation for sample sizes 60 through
counterparts indicate a lack of convergence of the Taylor 1,000 were .61, .43, .23, .17, and .12, respectively, while
series expressions for the moments. It can be shown that, for b2 the corresponding coefficients were .98, .76, .44, 3 4 ,
when T > 1/b2, the Taylor series expressions for the and .24. The correlations between 6 and 6 are .56, .63, .75,
moments converge. I n our simulation experiment b = .084 .78, and .81 for the respective sample sizes 60, 100, 300,
and hence T > 140 is required for convergence. This 500, and 1,000, while the corresponding correlations
result is consistent with the findings in Tables 2a and b. between 6 and h2 are .09, .11, .13, .13, and .14. The cor-
I n conclusion, comparisons between approximate and relations between 6 and P,,, j = 1, 2, . . ., N , ranged
simulation moments can only be made at sample sizes from -.18 to .32 and did not vary with sample size.
in excess of 140 for our example. Our earlier finding that the ratios X,,,j = 1, 2, . . .,
N , and p, are normally distributed for sample sizes of
3.3 Normal Goodness of Fit 300 or more is consistent with the values of the coefficients
Normal goodness-of-fit tests were also performed for of variation obtained. By sample size 300, the coefficient
each statistic by using the Kolmogorov-Smirnov test. of variation 6 has declined to .23 and seems to be
The mean and variance used in the tests were computed sufficiently close to the required .19 for normality. The
from the sample data. The probability of a larger correlation coefficient requirement given by Hayya,
deviation under the assumption of normality was ob- Armstrong, and Gressis (1975) does not seem to be im-
 Jobson and Korkie: Markowitz Portfolio Estimation 549

2a. Comparison of Simulation and Approximate I n the case of am2,the coefficient of variation for b2
Means for Xm, finl, and Gm2 was still above .19 a t sample size 1,000, which seems to
be consistent with our findings that 8m2 could not be
Means at Various Sample Sizes
Param- Parameter
assumed to be normally distributed below samples of
eter Value 60 700 300 500 1,000 size 1,000. At a sample size of 1,000, the coefficient of
variation of b2 was .24, which appears to be sufficiently
-.070 SIM - 1.865 -.408 -.070 -.070 - ,072
close to the critical .19, so that the null hypothesis of
APP -.194 -.115 -.080 -.075 - ,073
normality discussed before could not be rejected at
,116 SIM - 1.489 ,042 ,110 ,111 ,121
APP .074 ,101 ,113 ,114 ,115 significance levels below .13.
-.017 SIM ,095 -.042 -.008 -.012 - ,006
4. T H E EFFECTS OF CHANGES IN p,
APP - ,023 -.019 -.017 -.017 - .017
z,N, AND ,
,016 SIM ,457 ,079 ,009 ,015 ,017
APP ,071 .036 ,020 ,018 ,017 From Sections 2 and 3, the parameters a, b, c, and F,,
,090 SIM ,356 ,135 .071 .076 ,079 together with T and N , determine the sampling proper-
APP ,145 .110 ,094 ,092 ,091 ties of the estimators Pm, am2,and Em.Because a, b, c,
-.032 SIM - ,406 -.193 -.056 -.046 - ,035 and F, are functions of p and z, it is important to
APP - ,069 -.045 -.035 -.034 - ,033 determine how the parameters are affected by changes in
Downloaded by [Michigan State University] at 19:54 17 February 2015

-.071 SIM ,858 -.083 -.084 -.091 - ,070 p, 8 , N , and E,.

APP -.130 -.092 -.075 -.073 - ,072 For simplicity, assume that all the diagonal elements
,178 SIM -.182 ,262 ,178 ,180 .171 of B are equal to a and all the off-diagonal elements are
APP .243 ,201 ,183 ,181 ,179 equal to p. The mat,rix A = 8-1 can be written as
.379 SIM ,821 ,531 ,399 ,384 ,375
APP 518 ,428 ,389 ,384 ,381
,077 SIM ,374 ,171 ,121 ,096 ,083
APP .064 ,072 .076 ,077 ,077
where J is a matrix of unities (See Graybill 1969, p. 172).
.046 SIM - ,074 ,074 ,028 ,051 ,060 Hence,
APP ,037 ,043 ,045 ,045 .045
,086 SIM ,314 ,125 ,088 ,091 ,083
APP ,115 ,096 ,088 ,087 ,087
,193 SIM 1.536 ,347 ,210 ,209 .194 i = 1, 2, . . ., N (4.1)
APP ,122 ,167 .187 ,190 ,191
,021 SIM - ,027 -.042 .028 ,011 ,012 X.. -
-P
APP ,044 ,029 ,023 ,022 .021 23 -
(a + ( N - l>P>
.059 SIM
APP
- ,200
,040
-.lo7
,052
,047
,057
.066
,058
,057
,058
i # j i,j=l,2, ..., N .
,131 SIM - ,340 ,158 .136 .130 ,127 The parameters a, b, c, and Fmi may therefore be
APP ,173 ,146 .134 .133 ,132 written a s
-.157 SIM ,023 -.193 -.149 -.142 -.147
APP - ,230 -.183 -.162 -.160 -.158
-.016 SIM -.426 ,003 -.019 -.023 - ,022
APP - .003 -.011 -.015 -.015 - ,016

-.155 SIM -.221 -.152 -.171 -.173 -.157

APP -.174 -.162 -.157 -.156 -.155
,127 SIM 1.394 ,295 ,133 ,136 ,129
APP ,177 .145 ,131 ,129 ,128
1.38 SIM 12.98 5.61 2.15 1.83 1.61
APP 8.07 4.28 2.19 1.85 1.61
16.33 SIM 1.8 x lo4 413.87 26.87 21.54 18.75
APP 212.66 74.18 27.75 22.68 19.35 . i = 1 , 2,..., N ,
where
portant, in that the correlation between 6 and b exceeded N N

the .5 limit for all sample sizes, even though normality p = C p ; / N and V ( p ) = (C p i 2 / N ) - p2
i= 1 i= 1
of p, was not rejected for sample sizes exceeding 300.
In addition, the correlations between b and the Pmi, are the mean and variance, respectively, of the elements
j = 1, 2, . . ., 20, were below .5 in absolute value, and of p.
the ratios Xmi did not appear to approach normality at Examination of these expressions reveals that, as N
sample sizes lower than those for Pm. increases, b, c , and F m j approach p/p, l / P , and ( p i - ji)/
 550 Journal of the American Statistical Association, September 1980

(a - p), respectively. Thus, if the number of stocks N Table 2b. (Continued)

is large, b grows as p grows relative to p, while c in-
creases as p decreases. In the case of F m l ,as N increases, Variances at Various Sample Sizes
Param-
F,, grows as (pf - p) increases relative to (a - p). eter 60 100 300 500 1,000
I n the case of a, as N increases a increases without
limit. For fixed N , a varies directly with V ( p ) ,p, and p SIM 6.579 ,271 ,018 .007 ,005
ASY ,069 ,040 ,013 ,008 .004
and inversely with a. Therefore, larger values of a, 6, c, APP .206 ,075 ,016 .009 ,004
and F,, result if decreases. Increases in p bring about 158.47
SIM 1.466 ,042 ,021 ,009
increases in a and b, while increases in pJ relative to p ASY ,178 ,104 ,034 ,020 ,010
will cause F,, to incrcase. APP ,519 ,190 ,041 ,023 ,011
I n Section 3, it was shown that the magnitudr of b SIM 1.050 1.000 ,021 ,012 ,005
influenced the sample size required for the convergence ASY ,103 ,060 ,020 ,012 ,006
APP ,322 ,115 ,024 ,013 ,006
of the Taylor series approximations for the moments of
fin, dm2, and x,. Therefore, increases in p and/or dc- SIM
ASY
15.410
,225
,946
.131
,052
,043
,025
,026
,011
,013
creases in are beneficial in reducing the sample sizr APP ,699 ,250 ,053 ,029 ,013
required to obtain good approximations to the moments.
SIM 16.492 ,213 ,016 ,008 ,004
I n Section 1, the mean vector of return prrmiums p ASY .070 ,041 .013 .008 ,004
was defined as p* - E,e, where E , is a constant and p*
Downloaded by [Michigan State University] at 19:54 17 February 2015

APP ,202 ,074 ,016 ,009 ,004

is a vector of mean returns. Increases in E,, therefore, ,109 ,014 ,007 ,003
SIM 5.934
bring about decreases in p and hence corresponding ASY ,048 ,028 ,009 ,005 ,003
decreases in the parameters a and b. In the simulation APP ,127 ,048 ,011 ,006 ,003
experiment of Section 3, E , was set at .O. For the samr SIM 11.922 ,066 ,006 ,003 .002
p* and I; and samplc sizc T = 300, a n additional simula- ASY ,031 ,018 .006 ,004 .002
APP ,097 ,035 ,007 ,004 ,002
26. Comparison of Simulation, Asymptotic, and SIM 1.353 ,547 ,032 .019 ,008
Approximate Variances for X,, ,!&, and emz ASY
APP
,137
,402
,080
,147
,026
,032
,017
,018
.008
,008
Variances at Various Sample Sizes SIM 182.204 1.240 .019 ,008 ,005
Param- ASY ,076 ,044 ,014 ,009 ,004
eter 60 100 300 500 7,000 APP ,222 ,082 ,018 .010 ,005
SIM 260.052 2.579 ,022 ,010 ,004 SIM 4244.400 76.000 ,140 ,065 ,026
ASY ,077 ,045 ,015 ,009 ,004 ASY ,397 ,236 ,078 ,047 .023
APP ,252 ,089 ,018 ,010 ,005 APP 24.014 3.098 .1 68 ,072 ,028
SIM 208.183 1.123 ,040 ,024 ,012 SIM 3.2 x 1Olo 5.7 x 106 99.73 23.70 7.40
ASY .184 ,107 .035 .021 .010 ASY 70.16 40.86 13.34 7.98 3.98
APP ,561 ,202 ,043 ,024 ,011 APP 28,162.1 3,226.00 97.54 29.295 8.49
SIM 2.541 .751 ,011 ,005 ,003
ASY ,053 ,031 ,010 ,006 ,003 tion was performed for E , = .60. The sampling properties
APP ,165 ,059 ,012 ,007 ,003 of the estimators of a, b, and the elements of F, were
SIM 12.222 ,077 ,007 ,004 ,002 comparable to the asymptotic properties. For the ratios
ASY ,027 ,016 .005 ,003 ,002 p,, am2and the elements of X,, however, the simulation
APP ,087 .031 ,006 ,003 ,002
moments far exceeded the asymptotic moments.
SIM 6.867 ,239 ,007 .005 ,002 The decrease in p by .60 caused b to decrease from
ASY .033 ,019 ,006 ,004 ,002
APP ,097 ,035 ,008 ,004 ,002 .084 to .029 and a to decrease from .116 to .043. The
sample size required for convergence of the Taylor series
SIM 20.051 ,868 ,018 .009 ,005
ASY ,074 ,043 ,014 .008 ,004 approximations increased from 140 to l / b 2 = 1,190. Thus,
APP ,232 ,083 ,017 .010 ,004 the lack of agreement between the asymptotic moments
SIM 98.681 ,460 ,022 ,011 ,007 and simulation moments is easily accounted for by the
ASY ,096 ,056 ,018 .011 ,005 lack of convergence of the Taylor series expansions.
APP ,297 .lo6 .022 .012 ,006 Expressions 4.1 and 4.2 may be used t o determine the
SIM 27.066 1.040 .021 .010 ,005 effects of changes in p and I; on the variances and
ASY ,081 ,047 .015 .009 ,005 coefficients of variation. The coefficient of variation for b
APP ,225 ,083 ,018 .010 ,005
can be written approximately as
SIM 21.501 .848 ,037 ,017 .007
ASY .141 ,082 .027 .016 ,008 1 c ac
APP ,301 .1 24 .030 .017 ,008
SIM 2.728 ,248 ,034 .013 .008
ASY .138 .080 ,026 ,016 ,008
APP ,424 ,152 ,032 ,018 ,008
SIM 3.804 ,889 ,021 .011 ,005
ASY ,101 ,059 ,019 .011 ,006
APP ,311 .111 ,024 ,013 ,006
 Jobson and Korkie: Markowitz Portfolio Estimation 55 1

Increases in 9, or decreases in ,ri cause increases in For each of the 100 samples of Section 3, 2 values were
CV(b). As mentioned in Section 3, a n increase in CV(b) obtained by computing the quantities
increases the sample size T required for the distributions
of the ratios p, and x,,, j = 1, 2, . . ., N ,to approach Z, = [(ci - N/T)- p m b ] /
normality. We conclude that a n increase in E , of .60 [~(a) + pmzV(b> - 2 p m COV(&, b)]: ,
(hence a decrease in p) results in a marked departure ZU2 = [(ci - N/T)- um2b2]/
from normality for the ratios g,,, j = 1, 2, . . ., 20, fim
[V(ci) 4- um4V(b2)- 2um2cov(Ci, bz)]: (5.2)
and urnz,at T = 300.
~

I n addition to this simulation experiment the authors zx, = [ P m , - xrnlbl/

have repeated some of these analyses for other popula- +

[ v ( P ~ , > Xm12V(b)- 2 X m , cOv(Prn,, b)]' ,
tions with as many as 500 trials. The results of these
simulations are quite comparable to our results in this using the true values of p,, urn2,and X,,, j = 1 , 2 , . . . ,20.
article. When ( l / b 2 ) > T,the simulation moments for Estimates of the variances and covariances were made by
the estimators of p m and urn2can be extremely large replacing a, 6, c, F,,, and A,, by 6, b, 8, P m , , and W,,,
relative to the approximate moments, and the magnitudcs j = 1, 2 , . . ., 20, respectively, in the expressions (2.2)
of the simulation moments are unstable from experiment through (2.6). As in Section 2, V ( b ) and cov(ci, 6)z were
to experiment. obtained by assuming that ci and b are normally dis-
tributed. The means and variances of the Z values were
Downloaded by [Michigan State University] at 19:54 17 February 2015

computed ovm the 100 trials and K - S normal goodness-

5. STATISTICAL INFERENCE FOR pm, urn2,A N D X,, of-fit tests were performed.
It was shown in Sections 2 and 3 that the sampling Table 3 shows the means and variances of the various
distributions of the estimators ci, b, and P,,, j = 1, 2, Z values. The means and variances of 2, and ZU2 were
. . ., N are approximately normal and that closed-form omitted for sample size 60 and 100 because of several
expressions for the means and variances exist. The means extremely large sample values. Note that, for Zx,,
and variances for the ratio estimators fi,, am2,and x,,j = 1, 2 , . . ., 20, the means and variances are ap-
however, are not obtainable in closed form, and for proximately 0 and 1, respectively. For sample sizes 60
small samples the Taylor series approximations to the and 100, thc variances of the Zx,, j = 1, 2 , . . ., 20
moments do not converge. I n addition, the sampling tended to be less than 1. I n the case of 2, and ZU2,the
distribution5 of these ratio estimators differed from means were negative for all sample sizes. The variances
normality for small sample sizes. of 2, were larger than 1.0 a t sample size 300 and 500.
I n the case of bivariate normal random variables X The normal goodncss-of-fit tests (excluding 2, and Z,?
and Y , the distribution of the ratio W = Y / X was first for T = 60, 100) showed that the null hypothesis of

studied by Geary (1930) and Fieller (1932). Later normality could not be rejected exccpt a t very large
articles by Fieller (1954), Crcasy (1954), Marsaglia significance levels.
(1965), and Hinkley (1969) have also discussed various It was suggested by Fiellcr (1932) that confidence
aspects of the distribution of W , as well as infercnce intervals for W in (5.1) can be obtained from a confidence
procedures for W . interval for Z by solving a quadratic equation in W .
This approach has been used in practice by Fuller and
It was shown by Hinkley (1969) that the transformcd
variable Martin (1961) and Beveridgc (1975), for example. From
our results, we conclude that the transformcd variables
+
Z = ( X - W Y ) / ( U X ' W%y2 - 2 W ~ x y ) : (5.1) Z X , , j = 1, 2, . . ., 20, Z,, and Zuz given in ( 5 . 2 ) may be
used to make inferences about X,?, = 1, 2 , . . ., 20, p,,
approaches a standard normal random variable, as the and urn2for samples of size 300 or more. I n the case of
coefficient of variation of Y , u y / p y , approaches zero. the X,,, j = 1, 2, . . ., 20, the Z X , may also be used for
Monte Carlo simulation studies by Hayya, Armstrong, samples as small as 60.
and Gressis (1975) have shown that, if u y / p y < .39 I t may be of intcwst to tcst whether a given portfolio
and U X / ~ X> .05, the distribution of Z is approximately with the proportions vector X, is significantly different
normal. I n their simulation, Hayya, Armstrong, and from any Markowitz c>fficicmt portfolio, as opposed to
Gressis used the true parameters C T ~ U, X , and u x y to the efficient portfolio obtained from a specified Ez. If
compute 2. X, is a Markowitz efficient portfolio other than the
From Section 3, the coefficients of variation for the global minimum portfolio, then there exists a n E z such
P,,, j = 1, 2 , . . ., 20 and ci are well above the required that X, = h ( p * - e E z ) / e ' h ( p * - e E z ) . Defining b* =
.05 for all sample sizes. The coefficient of variation of b e'hp*, F,* = A p * and recalling that c = e ' h e and p*
is below the required .39 for samples of size 300 or more, is the mean-return vcctor dcfinc.d in Section 1, we may
and for bz the coefficient of variation is below .39 for write (b*X, - F,*) = E z ( c X , - A e ) , and, hence, for
samples of size 500 or more. I n addition, a t sample each j , j = 1, 2 , . . ., N , we have (b*X,, - F,,*) =
sizes 100 for b and 300 for b2, the coefficient of variation Ez(cX,, - A',e). For a given X,, confidence intervals
is only slightly above .39, a t .43 and .44, respchvely. for Ez may be obtained in a manner similar to (5.2).
 552 Journal of the American Statistical Association, September 1980

3. Sample Moments of Z Distributions + cov(Fm,*,W,e) - X,, cov(3, F,,)

- x,, cov(b*, ~ , e ) ]
Param- Mo-
Sample Size
+ E z ~ [ X , , ~ V (-~ *2X,,
) cov(c*,W,e) + V ( W , e ) ] .
eter ment 60 100 300 500 1,000
Expressions for V(b*), cov(b*, F,,), and V(F,,*),
Mean - .057 -.193 ,074 .069 ,013 are obtained from the expressions for V ( b ) ,cov(i), F,,),
var ,951 1.164 1.183 1.149 1.059 and V(P,,) in Section 2.1 by replacing b, F,, and a by
Mean -.134 ,009 ,012 -.OOO ,067 b*, Fm,* and a* = p*hp*. The remaining expressions
var ,840 ,927 ,927 1.092 1.114
arc dffincd by
Mean - .040 ,076 .116 .061 ,193
var 1.054 1.072 ,935 ,815 1.124 2c2
V [ L ] = ______
Mean ,144 ,090 -.147 -.058 -.026 ( T - N - 4)
var .685 .916 1.131 1.118 1.131
2 (A,e)
Mean - ,092 -.131 - ,256 -.267 -.273 cov(L, W t j e ) =
var ,771 1.ooo 359 1.196 1.144 (T - N - 4)
Mean ,073 - ,095 -.154 -.122 -.021 (T - N ) ( L % , ~ ) ~
var ,848 ,948 1.024 1.051 1.095 v[w.e] = ---_________
( T - N - 1 ) ( T - N - 4)
Mean ,078 ,013 -.065 -.162 ,036
Downloaded by [Michigan State University] at 19:54 17 February 2015

var ,953 ,916 .876 ,922 1.225 + -___( T - N - 2)cXjj

Mean - ,075 -.132 - ,042 -.016 -.110 (T - N - 1 ) (T - N - 4)
var ,961 ,861 1.121 1.056 .939
2b*c
Mean .128 .170 .044 ,013 -.064 cov[6*, t ] =
var ,965 ,932 1.127 ,986 ,894 (T - N - 4)
Mean ,114 ,124 ,223 ,104 ,039 (T - N)F,,*(A,e)
var ,666 ,718 ,843 ,805 1.065 cov[P,,*, wtje]= --
(T - N - l ) ( T - N - 4)
Mean - ,083 -.lo7 -.lo3 ,071 ,203
var
Mean
,791
- ,027
,947
,009
.81 1
- ,007
,929
,051
,933
-.040
+-( T - N - 2)b*Xjj
( T - N - 1)(T - N - 4)
var 1.001 ,914 1.031 ,919 1.185
( T - N)b*(Atje)
Mean ,148 ,184 ,102 ,126 ,013 cov[6*, ~ ~= e ] ~

var ,703 ,992 1.001 ,938 ,873 ( T - N - 1 ) ( T - N - 4)

( T - N - 2)cF,,*
Mean
var
- ,000
301
- ,040
,754
.019
,948
-.114
,993
-.128
,803 +-
(T-N- l)(T-N-ij
Mean - ,065 - ,089 - ,047 ,034 -.026
var 1.017 1.127 ,960 ,919 ,842 - 2cFmj*
COV[E, F,,*] = -
Mean - .049 ,025 ,016 -.017 -.078 ( T - N - l)(T-N-4)
var ,762 ,697 1.014 ,988 ,995
2 ( T - N - 2)b*(hf,e)
Mean
var
- ,056
,974
,092
1.129
.148
1.283
,248
1.166
,237
1.141
+-
(T - N - 1)(T - N - 4)
Mean - ,072 - .036 - ,091 -.133 -.168 Estimators for the components of V [ D , ] can be obtained
var ,974 1.172 ,866 ,883 ,990
by replacing p* and A by f* and W in the definitions of
Mean - ,015 - ,092 - ,058 -.128 -.014 a*, b*, c, and F,*.
var ,836 ,966 ,964 1.128 1.037
The magnitudes of the interval estimators of Ez may
Mean - ,061 ,059 .016 ,066 .025 now be judged for reasonablcness. In theory O<Ez
var ,779 ,918 1.076 ,916 1.037
< b*/c, wherc b*/r is the mean return for the global
Mean - - -.321 -.248 -.lo0 minimum-variance portfolio. A confidence interval for
var - - 1.459 1.338 1.031
po = b*/c can be obtained from the sample information
Mean - - - ,363 -.376 -.311 by employing the estimator po = i)*/3. I t can be shown
var - - 1.066 1.076 ,970
that E[po] = P O and
Define the Z variable Z E , = D,/[V(D,)]:, where V[po] = p02V(c*)/c2- 2po cov(9*, c*)/c2
D, = (b*X,, - F,,*) - Ez(c*X,, - W,e),j = 1,2, . . ., V(b*)/cZ O(T--2) , + +
N , and b*, 8,*, and 2 are obtained by replacing v* and
h by the sampIe quantities f * and W,where ?* is the and, therefore, we define 2, = (no- pO)/[V(pO)]f.
sample mean-return vector. I t is easily shown that Estimating V(p0) by replacing I* and A by f* and W
in the expressions for V(c*),cov(&*,c*), and V(b*),we can
E[D,] = 0 and that
obtain a confidence region for po. If pa, is the upper limit
v [ D ~ ]= X , , Z V ( ~ *-
) 2xp1C O V ( ~P*m,, ) obtained from this interval than the acceptable range for
+ V(P,,*) - ~ E Z ~ [ XC, O , ~V ( ~2)* , EZ is (0, PO,).
 Jobson and Korkie: Markowitz Portfolio Estimation 553

For each component of X,, X,,, j = 1, 2 , . . ., N the The distributions of p, and X,,,j = 1, 2, . . ., N , in
procedure shown before gives a n interval estimate for the simulation were found t o be close to normal a t sample
Ez. If a single estimator of E z is desired, two conventional sizes of 300 or more. For dm2the normality assumption
estimators that may be used arc the ratio estimator was barely acceptable a t sample size 1,000. The effect
N
on these distributions of the coefficient of variation of
I?,, = C ( 6 * ~ , , - 8,,*)/(tx,,- W?e) the denominators of these ratios was shown to be con-
3= 1 sistent with the findings of Hayya, Armstrong, and
and the regression estimator Gressis (1973). The correlations between the numerator
and denominator of these ratios were found to be
N unimportant.
gZ2= [C ( 6 * ~ , , - P,,*)(ZX,, - ~ , e ) ] / The applicability of transformed variables of the form
j= 1
2 = ( X - WY)/(ux2 - W 2 a y 2- 2Waxy): for the ratio
N
W = X / Y was examined. Sample estimates for the
(ex,,
[C
j=1
- W3e)21.
variances and covariances were used, and the distribu-
tions of 2 were studied for each of p,, urn2,and the
Both these estimators, however, are biased and for elements of X,. The 2 values corresponding to X, were
inference purposes suffer from the same shortcomings as reasonably well behaved for all sample sizes, while for
Downloaded by [Michigan State University] at 19:54 17 February 2015

the estimators of p,,, urn2,and X, discussed in Section 3 . p, and am2the 2 values were applicable for samples of
size 300 or more.
6. SUMMARY AND CONCLUDING REMARKS The impact on the sampling properties of changes in
the mean premium-return vector p, the covariance matrix
Under the assumption of the normality of asset returns, 8 , and the portfolio size N were examined. For simplicity
this article has examined the sampling properties of the it was assumed that the diagonal and off-diagonal
conventional estimators for the parameters of a n efficient elements of X have magnitude O( and p, respectively, and
portfolio. The parameters estimated are the efficient set that the mean of the elements of the vector p is given by
constants a and b, the standardized and nonstandardized p. For large N , the magnitude of b and the coefficient of
weight vectors X, and F,, and the mean and variance variation of b were found to be dependent on the ratio
p, and urn2of a n efficient portfolio. Because many of the p / p . Thus, the sample size T required for both the
derived sampling properties are approximate, a sampling convergence of the expressions for the approximate
experiment was used to determine the applicability of moments and for the normality of the ratio estimators
the derived properties a t various sample sizes. decreases as the ratio p / p increases. Therefore, if the
The exact moments of the estimators of a, b, and the covariances among the assets are relatively small but the
elements of F,, were derived. The estimators of a, 6 , mean returns are relatively large the sampling properties
and the elements of F, were found t o be comparable to of the estimators are improved. The effect of a n increase
a normal distribution for all sample sizes. The estimators in the risk-free rate E, is equivalent to a decrease in p.
of b and F, were shown to be unbiased, while the esti- We may therefore conclude that the estimators are not
mator of a was shown to have bias ( N / T ) .The asymptotic applicable for populations of stocks in which p is small
variances for 6, b, and 8, differed substantially from their relative to the off-diagonal elements of 8 .
exact counterparts a t sample size less than 300. The applicability of the conclusions of the Monte
Assuming normality for the distributions of 6, b, and Carlo study to the financial market depends in part on
the elements of 8,, Taylor series approximations were the realism of the mean-return and covariance param-
obtained from the means and variances of the ratios eters. Because these parameters were computed over 313
p,, dn2, and a,,, j = 1, 2, . . . , N . The theoretical means months or 26 years, temporal changes in market condi-
and variances and simulation means and variances were tions would cause elements of the covariance matrix,
compared. For sample sizes of a t least 300 the approxi- including p, t o be unrealistically large. I n addition, sample
mate means and variances were found to be comparable outcomes such that E , > b / c are included in our results.
to the simulation values. The convergence of the expres- Merton (1972) has shown that such results are not
sions for the means and variances of p,, dm2, and a,,, economically realistic. These sample outcomes tend t o
j = 1 , 2 , . . ., N , were shown to depend on the magnitude reduce the values of the elements of the mean-return
of T relative to l/b2. For the X,,,j = 1, 2, . . ., N , the vector and therefore p. Sample size requirements men-
asymptotic means and variances were comparable to the tioned throughout the study therefore would in general
simulation means and variances for sample sizes 500 and be less stringent.
1,000. For p m and dm2,the asymptotic means were com- From this study we conclude that the estimators p,,
parable to the simulation counterparts a t sample size dm2,and x, do not lend themselves to making inferences
500 and 1,000. The asymptotic variances of p, approxi- in small samples. The key t o improving the small-
mated the simulation variance at sample size 1,000, sample properties of these estimators lies in improving
while for am2the asymptotic variance was still not com- the estimators of p and A. Other studies by the authors
parable a t T = 1,000. (1979, 1980) have shown that the use of James-Stein-
 554 Journial of the American Statistical Association, September 1980

type estimators of p and A can bring about substantial Geary, R.C. (1930), The Frequency Distribution of the Quotient
improvements to the estimators of nm,
km2, and x,. of Two Normal Variables, Journal of the Royal Statistical Society,
43, 442-446.
Graybill, Franklin A. (1969), Introduction to Matrices W i t h A p -
[Received October 1978. Revised October 1979.1 plications in Statistics, Belmont, Calif. : Wadsworth Publishing Co.
Hayya, Jack, Armstrong, Donald, and Gressis, Nicholas (1975),
A Note on the Ratio of Two Normally Distributed Variables,
REFERENCES Management Science, 21, 1338-1341.
Anderson, T.W. (1958), An lntroduction to Multivariate Analysis, Hinkley, D.V. (1969), On the Ratio of Two Correlated Normal
New York: John Wiley & Sons. Random Variables, Biometrika, 56, 635-639.
Barry, C.B. (1974), Portfolio Analysis Under Uncertain Means, Jobson, J.D., and Korkie, B. (1980), Improved Estimation and
Variances and Covariances, Journal of Finance, 29, 515-522. Selection Rules for Markowitz Portfolios, paper presented at
~- (1975), Specification Uncertainty in Portfolio Analysis, the June 1980, Western Finance Association Meetings, Sm Diego.
Proceedings of the American Statistical Association, Business and Jobson, J.D., Korkie, B., and Ratti, V. (1979), Improved Esti-
Economic Statistics Section, 223-237. mation for Markowitz Portfolios Using James-Stein Type Esti-
(1978), Effects of Uncertain and Non-Stationary Parame- mators, Proceedings of the American Statistical Association,
ters Upon Capital Market Equilibrium Conditions, Journal of Business and Economic Statistics Section, 279-284.
Financial and Quantitative Analysis, 13, 419-433. Klein, R.W., and Bawa, V.S. (1976), The Effect of Estimation
Beveridge, Stephen (1975), The Dynamic Properties of the St. Risk on Optimal Portfolio Choice, Journal of Financial Eco-
Louis Model, unpublished PhD dissertation, University of nomics, 3, 215-231.
Chicago, Graduate School of Business. -~ (1977), The Effect of Limited Information and Estimation
Black, Fisher (1972), Capital Market Equilibrium With Restricted Risk on Optimal Portfolio Diversification, Journal of Financial
Borrowing, Journal of Business, 45, 444-454. Economics, 5, 89-1 11.
Downloaded by [Michigan State University] at 19:54 17 February 2015

Creasy, Monica A. (1954), Limits for the Ratio of Means, Journal Markowitz, Harry, M. (1959), Portfolio Selection: Eficient Diver-
of the Royal Statistical Society, Ser. B, 16, 186-192. sification of Investment, New York: John Wiley & Sons.
Dickenson, J.P. (1974a), The Reliability of Estimation Procedures Marsaglia, George (1965), Ratios of Normal Variables and Ratios
in Portfolio Analysis, Journal of Financial and Quantitative of Sums of Uniform Variables, Journal of the American Statis-
Analysis, 9, 447-462. tical Association, 60, 193-204.
(1974b), Some Statistical Aspects of Portfolio Analysis, Marx, D.L., and Hocking, R.R. (1977), Moments of Certain
The Stalistician, 23, 5-16. Functions of Elements in the Inverse Wishart Matrix, paper
Fama, Eugene F. (1976), Foundation of Finance, New York: presented at the Annual Meeting of the American Statistical
Basic Books. Association, Chicago.
Fieller, E.C. (1932), The Distribution of the Index in a Normal Merton, Robert C. (1972), An Analytic Derivation of the Effi-
Bivariate Population, Biometrika, 24, 428-440. cient Portfolio Frontier, Journal of Financial and Quantitative
___ (1954), Some Problems in Interval Estimation, Journal Analysis, 7, 1851-1872.
of the Royal Statistical Society, Ser. B, 16, 175-185. Rao, C. Radhakrishna (1973), Linear Statistical Inference and Its
Frankfurter, George M., Phillips, Herbert E., and Seagle, John P. Applications, New York: John Wiley & Sons.
(1971), Portfolio Selection: The Effects of Uncertain Means, Roll, Richard (1977), A Critique of the Asset Pricing Theorys
Variances and Covariances, Journal of Financial and Quanti- Tests, Part I: On Past and Potential Testability of the Theory,
tative Analysis, 6, 1251-1262. Journal of Financial Economics, 4, 129-176.
Fuller, Wayne A,, and Martin, James E. (1961), The Effects of Sharpe, William F. (1964), Capital Asset Prices: A Theory of
Autocorrelated Errors on the Statistical Estimation of Distributed Market Equilibrium Under Conditions of Risk, Journal of
Lag Models, Journal of Farm Economics, 43, 71-82. Finance, 19, 425-442.