Hammer 1992
Hammer 1992
Hammer 1992
JERRY A. HAMMER
HERBERT E. PHILLIPS
The purpose of the single-index model is to obtain solutions to the general portfolio selection
problem which are equivalent to those obtained by the full variance-covariance portfolio
selection model, but to do so at lower cost. The model’s computational advantage hinges on
the assumption that the error terms are cross-sectionally independent. Although it is generally
understood that this assumption is not strictly correct, the implications have never been
demonstrated. This paper shows that the single-index model’s distributional assumption is
substantially violated in sample applications, and that the model’s solutions are markedly
different from those produced by the full variance-covariance model as a direct result.
I. INTRODUCTION
The selection of Harry Markowitz, William Sharpe and Merton Miller for the 1990 Nobel
Prize in economics marks the acceptance of modem finance as an essential part of the
academic literature. For students of finance, this recognition, though long overdue, is also a
challenge. Modem theory is based on rigorous argument and the application of scientific
method. In this regard especially it differs from previous methodologies that involved little
more than ad hoc rules and exceptions to rules. The rapid and significant advances in the
theory of finance that have taken place in recent years were possible because, in the main,
they were developed within a common theoretical framework that is quite general. This
interlocking of the models and theories of finance, however, is a two-edged sword. Errors
and misconceptions that may long have been overlooked or ignored can be interlaced through
a broad spectrum of models in which the findings and conclusions of previous work are
treated as axiomatic. This paper deals with one such issue.
The Markowitz (1952) mean-variance utility maxim is the theoretical basis and motivation
for much of modem finance. Implementation of the full variance-covariance portfolio selec-
Jerry A. Hammer l Assistant Professor of Finance, Department of Finance, School of Business and Management,
Temple University, Philadelphia, PA 19122; Herbert E. Phillips l Professor of Finance. Deaartment of Finance.
&h&l of Business and Management. Temple University. Philadelphia. PA 19122.
InterrmtIod Review of Fbmocial AdysIs, Volume 1, Number 1,1992, pages 39-50. ISSN: 10574219
Copyright 0 1992 by JAI Press, Inc. All rlgbts of reproduction In any form reserved.
39
40 INTERNATIONAL REVIEW OF FINANCIAL ANALYSIS / Vol. 1, No. 1
tion model, however, involves iterative inversion of an nth order matrix-which is costly in
terms of computer execution time and memory. In order to diagonalize the variance-
covariance matrix, which is key to the single-index model’s computational efficiency, Sharpe
(1963, p. 281) invoked the assumption that the residuals are cross-sectionally independent.
King (1968) questions the applicability of this assumption, as do Blume (1971) and many
others. Markowitz and Perold (1981) take an alternate path by considering the use of factors
and scenarios to revise the prior on a covariance matrix. Bayesian methodology does not
resolve the issues that arise in sample-based applications, however. Vasicek (1973, p. 1239)
makes this point clear.
Bayesian theory deals with the distribution of parameters given the available information,
while sample theory deals with the properties of sample statistics given the value of the
parameters.
Relatedly, the present paper shows that cross-sectionally correlated residuals in sample data
affect the security composition of the single-index model’s solution set, and thus alter the
very nature of the portfolio efficiency that is at the heart of the Markowitz theory that many
other theoretical developments build upon.
Complications arising from cross-sectionally correlated error terms are not limited to the
portfolio selection framework. Black, Jensen and Scholes (1972) and Roll (1977) consider
the implications of cross-sectionally distributed beta statistics in tests of the capital asset
pricing model. The inability of the traditional CAPM to satisfactorily explain capital asset
prices has led to the development and study of the arbitrage pricing theory by Ross (1976),
Roll and Ross (1980) and others. Brown (1989) points out, however, that purely empirical
attempts to determine the number of factors in equity returns are likely to lead to false
conclusions.
Rosenberg (1974, p. 271), moreover, contrasts a multiple factor model with a variant form
of the diagonal model and concludes that, although the problem of correlated residuals
remains, the magnitude of the problem is reduced. Even this very limited conclusion,
however, would not follow so easily if the factor relationships required more than a purely
empirical justification. Koopmans’s (1947, p. 191) point in this regard deserves reference:
“What is relevant can only be determined with the help of some notions as to the generation
of.. . fluctuations.” In seeking plausible tests of positive theory, therefore, we should first
resolve the micro issues.
Frankfurter and Lamoureux (1990), for example, argue that the performance of the single-
index model “does not hinge on the . . . assumption of cross-sectionally independent error
terms.” This would seem to imply that the securities actually selected by the ex ante model,
and the weights that correspond to them, are likewise unaffected. The present paper shows,
by contrast, that the single-index model’s assumption of cross-sectionally independent error
terms is seriously violated in sample applications. As a result, the single-index model
disregards an important component of statistical covariation, and, by doing so, invokes
different stock substitution and selection criteria than the full variance-covariance model
does. The empirical results show that the security composition of the single-index model’s
solutions are very different from those obtained by the full variance-covariance model, and
leads to fundamentally different stock selection strategies.
The Single-Index Model 41
It is generally assumed that Sharpe’s (1963) single-index portfolio selection model [hence-
forth SIM] and the full variance-covariance model [henceforth COV] are equivalent, and
produce equivalent solution sets. Elton and Gruber (1977), for example, use the models
interchangeably in one paper, and exploit the assumed equivalence in another (Elton, Gruber
and Padberg 1976). This issue of normative equivalence is obscured in the literature by a
faulty definition. Cohen and Pogue (1967), for example, were among the first to show that
sample-based SIM and COV solution sets tend to plot in close proximity in EV space; on this
basis, they conclude that the models are equivalent. Markowitz (1987, p. 25) formally states
that “two models are equivalent if they have the same set of efficient EV combinations.” This
paper shows that this notion of equivalence may be misleading.
The SIM and COV solution sets do plot in close proximity in EV space. This overlapping
does not necessarily mean that the models are equivalent, however. SIM disregards an
important component of statistical covariation that COV takes fully into account; thus, the
models invoke different substitution criteria when branching from one comer solution to the
next. An analytic justification for these conclusions is presented in Section III. Empirical
results presented in Section IV show that SIM’s assumption of cross-sectionally independent
error terms is substantially violated and that the security composition of the SIM and COV
solution sets differ as a direct result.
subject to budget and boundary constraints. In the expression, A is the coefficient of risk
aversion, and pp and a,* are portfolio mean and variance parameters that need to be estimat-
ed. 1 To facilitate applications, Sharpe (1963) replaces COV’s input functions with an altema-
tive set based on a regression structure:
where Fiilis a rate of return, Ai and Gi are regression coefficients, rm, is the rate of return on a
broad stock market index and eit is a sample residual.
In order to diagonalize the variance-covariance matrix, which is key to SIM’s computa-
tional efficiency, Sharpe (1963, p. 284) assumes that the residuals are cross-sectionally
independent:
cov(f?i,i;j> =
1O
Var( eii)
vi+j
Vj=j
This is not a trivial assumption; to see why and in what sense, we relax the assumption.
42 INTERNATIONAL REVIEW OF FINANCIAL ANALYSIS / Vol. 1, No. 1
one obtains
This result differs from Sharpe’s (1963, p. 282), where i # j, in just one respect: he assumes
that the cross-sectional residual covariance term is zero and, consistent with Equation (3)
suppresses it.
To illustrate the effect of this assumption on the derivation of SIM’s portfolio variance
estimator, we substitute from Equation (5) into the standard COV portfolio variance esti-
mator:
(6)
i=l j-1
which yields:
n n
i#j
where COV Var( fP) is defined by Equation (6) and SIM Var( fP) is substituted for the first two
terms of Equation (7). Equation (8) makes it clearer that the algebraic equivalence of the SIM
and COV portfolio variance estimators depends critically on the assumption that the cross-
The Single-Index Model 43
sectional covariance terms, Cov(ci, cj), are zero.2 This is an empirical question, to which we
now turn.
Data for this study were obtained from Standard & Poor’s PDE Compustat Tape. Firms
included in Standard and Poor’s 500 Industrials [henceforth S&P] as of December 1989 were
excluded from the sample. Dividend-adjusted rates of return were calculated for the remain-
ing 1,653 firms for which there were no missing observations during the period December
1979 to December 1989.
Rates of return on these stocks were regressed on the S&P. Two regressions were fitted for
each stock: one using 60 monthly observations covering the period from January 1985 to
December 1989, and a second based on the 120 monthly observations covering the period
from January 1980 to December 1989. A short- and a long-time series are used to account for
Table 1
Cross-Sectional Residual Covariances: Values Corresponding to CDF Percentilesa
Cumulative Cumulative
Sample Size Sample Size
density: density:
percentile 60 120 percentile 60 120
CUMULATIVE
PERCENT
100 I
60
60 ]
120 MONTHLY
40 ; 60 MONTHLY OBSERVATIONS
OBSERVATIONS
\
/
20 1
\
1
I ‘/
,’ ,’
_*
,’ ,..,’
-
I- --- -T- -
differences in methodology which are common in empirical work. Two series are used to
demonstrate that degrees of freedom is not a major issue determining the implications of
SIM’s assumption of cross-sectionally independent error terms. The residuals were saved,
and 1,365,378 cross-sectional residual covariance statistics were calculated for each time
series. Cumulative density functions for these covariance statistics were then generated.
A cumulative density function for each time series is shown in Table 1, and the central
portions of the CDF’s are plotted in Figure 1. The table and figure show that SIM’s central
assumption, Cov(ci,Zj) = 0, Vi + j, is substantially violated in each case. The figure shows
that the CDF’s are sharply skewed to the right in general, and more so for the longer time
series. The table shows that 64% of the residual covariance terms are positive when the
shorter time series is used, and 72% are positive when estimation is based on the longer time
series.3 It follows from Equation (8), therefore, that SIM understates the average security’s
contribution to portfolio risk in the sample.
A random sample of 800 of the 1,653 stocks was selected for the contrast.4 SIM and COV
solution sets were derived, and the resulting efficient frontiers are shown in Figure 2.
Solution sets based on the shorter time series are presented in Figure 2A, and the solution sets
for the longer time series are shown in Figure 2B.
An intuitive reading of Equation (8) might suggest that the relative position of the SIM and
COV frontiers~is unique. Since the security composition of corresponding SIM and COV
The Single-Index Model 45
RETURh
x 100
COV FRONTIER
I
0
J---Y; *
i
_,I -- .-
! x
/
I ,/ ‘- A
3 7 / ..
/
I ,’
+* \
! ,/ SIM FRONTIER
I / i
* -I *
i ,/:
\ ADJUSTEO
I ! SIM PORTFOLIOS
3 4 /,”
i/
j \A
I
1 ’
j-----_r- ~___ _ r---- _ .?.__ _ ~~.~~
_ _--r----. .T .~_~~~_~~__~___
WRI’ANCE
0 20 40 60 SO 100 120 140 x 10000
RETURN
x 100
--G-*
A--
..- /
,/’ m
\
* COV FRONTIER
,_;/ G+c
I ;G
3 1 \
ADJUSTED
1; SIM PORTFOLIOS
1 /
/ , , i ___-T_ _... i WdANCE
0 20 40 SO 30 100 120 x 10000
FIGURE 2. SIM and COV efficient frontiers: (a) 60 monthly observations, (b) 120 monthly observa-
tions.
46 INTERNATIONAL REVIEW OF FINANCIAL ANALYSIS / Vol. 1, No. 1
portfolios may be different, this uniqueness does not follow. Figure 2 shows that the SIM and
COV frontiers plot in close proximity in EV space, but there is no necessary ordering-the
relative location of the frontiers varies from sample to sample. The tendency of the SIM and
COV ex ante frontiers to plot in close proximity is well-known, but is misleading. In fact, the
SIM and COV solution sets represent markedly different stock selection strategies.
The SIM efficient frontiers are derived according to SIM variable definitions, and the COV
solution sets are obtained according to COV criteria. Once a set of security weights is
established by either model, however, the same weights can be substituted into the other
model’s portfolio mean and variance estimator functions. The resulting adjusted portfolio
means and variances may then be plotted on the same graph for the purpose of direct
contrast.
Notice that points marked by asterisks in Figure 2 lie internal to the COV feasible region
for both samples. These points are obtained by adjusting each SIM portfolio for the cross-
sectional residual covariances in Equation (8) that SIM disregards. The figure shows that
these adjusted mean and variance combinations plot below and to the right of the correspond-
ing COV frontiers. The optimizing SIM portfolio solution weights, therefore, are different
from the COV solutions. The nature and significance of the differences is now examined.
Column 1 of Table 2 reports selected levels of portfolio variance for the SIM and COV
algorithms.5 Columns 2 and 3 report the number of stocks contained in the SIM and COV
solutions at each level of risk, and Column 4 reports the number contained in both the SIM
and COV solutions.
Columns 2 and 3 show that SIM tends more and more toward superfluous diversification
as the level of portfolio risk is reduced. Column 4 shows that the security composition of the
SIM and COV solution sets are substantially different. Figure 2, by contrast, shows that the
SIM and COV frontiers plot in close proximity in EV space, notwithstanding the differences
observed in the table. This result can now be explained.
The COV and SIM portfolio expectation estimators,
SIM E(Fp,)
= i Xihi+ (i Xiii]E(rm)
i=l i=l
are algebraically equivalent if E(r,)is a sample mean. Both models see the same feasible
limits of portfolio expectation, therefore, but not necessarily corresponding to the same
levels of portfolio risk or risk-return tradeoffs. This equality is tantamount to a boundary
condition that forces SIM frontiers to mirror COV frontiers.
Disregarding the cross-sectional covariance terms in Equation (7) causes SIM to substitute
different stock combinations than COV as it branches from one comer portfolio to the next in
its search for optimal risk-return tradeoffs; this is demonstrated in Columns 2, 3, and 4 of the
The Single-Index Model 47
Table 2
Components of Portfolio Variance and Portfolio Size for Selected Levels of Risk”
‘The portfolio variance and SIM error statistics are scaled by 104.
table, and by contrast of Blocks A and B. SIM tends toward superfluous diversification
relative to COV as risk is reduced in both Blocks A and B. Recall from Table 1 and Figure 1,
however, that the residual covariance CDF’s are more highly skewed to the right for the
longer time series than the shorter one. Perhaps this is a reflection of sample differences,
such as the 1979 to 1982 experience and the period of adjustment that followed. Contrasting
48 INTERNATIONAL REVIEW OF FINANCIAL ANALYSIS / Vol. 1, No. 1
Blocks A and B of Table 2, we see that COV compensates for greater cross-sectional
covariation by bringing more stocks into solution. SIM, which assumes that the cross-
sectional effects are zero, has no reason to compensate and brings fewer stocks into the more
highly diversified portfolios in Block B than in Block A.
Panels A and B of the table show that SIM’s tendency toward superfluous diversification is
not substantially affected by the length of the time series. This relative invariance of the
optimization algorithm’s behavior to the length of the time series is strong evidence that
SIM’s solutions are dominated by a variable that is not sensitive to degrees of freedom.
Notice that a residual variance statistic, Var(ci), in the context of regression, is not much
affgcted by degrees of freedom, but the standard errors of the estimator functionsf(hi) and
f( B i) are affected.
A major conclusion supported by these results is that the SIM and COV portfolio optimiza-
tion algorithms lead to different stock selections and selection strategies. SIM’s tendency
toward superfluous diversification is not systematically related to degrees of freedom, but
may be sensitive to methodological differences that affect average residual variation. This
conclusion is consistent with Frankfurter’s (1976) finding that the degree of superfluous
diversification is affected by goodness of fit of the index proxy.
Portfolio variance for selected SIM and COV portfolios are shown in Column 1 of Table 2.
The cross-sectional residual covariance terms that the SIM portfolio variance estimator
disregards are shown in Column 5; these terms provide a measure of mis-specification error.
The ratio of this residual covariance error to the reported SIM portfolio variance,
is shown in Column 6. The data in Column 6 show that this error ratio varies from 20.7% to
117.1% for the shorter time series, and from 8.6% to 86.0% for the longer time series. These
results lead to two important conclusions.
The portfolios marked by asterisks in Figure 2 are obtained by adjusting the corresponding
SIM portfolios for the cross-sectional residual covariances [see Equation (8)] that SIM
disregards. Because of the scale of the graphs, however, one may be given the mistaken
impression that the adjustments are relatively minor. Column 6 of the table, by contrast,
shows that the adjustments are as high as 117% of reported SIM variance for the shorter time
series and 86% of reported SIM variance for the longer one. The table shows that these cross-
sectional error adjustments are not uniform over the range of solutions for either time series,
but become larger and larger, both in absolute terms and as a proportion of reported SIM
variance, as diversification increases.
Figure 2 also shows that these adjusted portfolio mean and variance combinations are
interior to the COV feasible regions, and are thus dominated by points that lie on the
The Single-index Model 49
corresponding COV efficient frontiers. Notice that the variances shown in Column 1 of the
table correspond to both SIM and COV portfolios. Since the SIM error ratios measure
the proportionate adjustments to the SIM variances, they also provide an indication of the
degree of mis-specification error in SIM’s portfolio variance estimates. The data in Column 6
indicate that the impact of mis-specification error is not trivial at any level of estimated risk,
and increases as the model branches to lower and lower levels of reported variance.
V. CONCLUSIONS
Various papers have attempted to contrast the equivalence of the single-index and full
variance-covariance models. It has been found elsewhere, as in this paper, that the efficient
frontiers for both models plot in close proximity in mean-variance space. This result has led
to the mistaken impression that their solution sets are equivalent. For the solution sets to be
“equivalent” in the sense of selecting the same or similar securities, however, it is necessary
that all sample cross-sectional covariance terms be equal to zero. This condition is not even
approximated in standard applications.
The single-index portfolio selection model disregards an important component of statisti-
cal covariation that the full variance-covariance model takes fully into account and exploits.
Thus, the models invoke different stock selection and risk reduction criteria. The experimen-
tal results illustrate that the single-index model’s solutions are different from those obtained
by the full variance-covariance model.
The solution sets differ not only with regard to which stocks are selected for inclusion, but
also in terms of how many stocks are selected. The single-index model’s tendency toward
superfluous diversification is not systematically related to degrees of freedom, but is sensi-
tive to methodological differences that affect the residual variance statistics. The empirical
results show that the model’s ex ante risk estimates are seriously biased and that the SIM
covariance ratio and superfluous diversification increase roughly in unison as the optimiza-
tion algorithm branches to lower and lower levels of portfolio risk.
ACKNOWLEDGMENTS
This research has benefitted from the comments and suggestions of George M. Frankfurter.
We wish to thank him particularly for his encouragement.
Notes
1. Markowitz (1987) suggests that these inputs may be obtained on the basis of expert judgment.
Sharpe (1963) expresses a similar view. Many, if not most, applications are based strictly on sample
data, and not expert judgment.
2. For the solution sets to be identical, each cross- sectional residual covariance term must be zero; it
would not be sufficient that they merely average out.
3. The residual covariance terms in Table 1 have not been formally tested for significance, although
the data clearly indicate that the hypothesis that the residuals are cross-sectionally independent (in
parameter space) would be rejected for at least some i f j. It is important to note, however, that even if
50 INTERNATIONAL REVIEW OF FINANCIAL ANALYSIS / Vol. 1, No. 1
none of the 1,365,378 cross-sectional covariance statistics are significantly different than zero in the
sense of a conventional 1% or 5% significance test, it would have no bearing on the difference or
similarity of the SIM and COV solution sets. This issue is determined by cross-sectional covariance
sample statistics, and not parameters.
4. The amount of virtual memory needed to run 1,653 stocks with COV exceeds the 14 mg
maximum available to the authors.
5. Portfolio variances were chosen as close as possible to the values shown and within 3% of the
tabulated values. Risk levels for which SIM and COV portfolio variance statistics, as defined by
Equation (8), do not fall within a 3% bound are not reported in Table 2.
References
Black, Fischer, Jensen, Michael C., and Scholes, Myron. 1972. “The Capital Asset Pricing Model:
Some Empirical Tests.” In Studies in the Theory ofcapital Markets, edited by Michael C. Jensen.
New York: Praeger Publishers.
Blume, Marshall E. 1971. “On the Assessment of Risk.” Journal of Finance 26: l-10.
Brown, Stephen J. 1989. “The Number of Factors in Security Returns.” Journal of Finance 44: 1247-
1262.
Cohen, Kalman J., and Pogue, Jerry A. 1967. “An Empirical Evaluation of Alternative Portfolio-
Selection Models.” Journal of Business 40: 166- 193.
Elton, E.J., and Gruber, M.J. 1977. “Risk Reduction and Portfolio Size: An Analytical Solution.”
Journal of Business 50: 415-437.
Elton, E.J., Gruber, M.J., and Padberg, M.W. 1976. “Simple Criteria for Optimal Portfolio Selection.”
Journal ofFinance 31: 1341-1357.
Frankfurter, George M. 1976. “The Effect of ‘Market Indexes’ on the Ex Post Performance of the
Sharpe Portfolio Selection Model.” Journal of Finance 3 1: 949-955.
Frankfurter, George M., and Lamoureux, Christopher G. 1990. “Insignificant Betas and the Efficacy of
the Sharpe Diagonal Model for Portfolio Selection.” Decision Sciences 21: 853-861.
King, Benjamin F. 1968. “Market and Industry Factors in Stock Price Behavior.” Journal of Business
41: 139-170.
Koopmans, Tjalling C. 1947. “Measurement Without Theory.” Review of Economics Statistics 29.
Reprinted in Readings in Business Cycles, edited by Robert A. Gordon and Lawrence R. Klein,
American Economic Association. Homewood, IL: Richard D. Irwin.
Markowitz, Harry M. 1952. “Portfolio Selection.” Journal of Finance 7: 77-91.
-. 1987. Mean-Variance Analysis in Portjolio Choice and Capital Markets. Oxford: Basil Black-
well.
Markowitz, Harry M., and Perold, Andre F. 1981. “Portfolio Analysis with Factors and Scenarios.”
Journal of Finance 36: 871-878.
Roll, Richard. 1977. “A Critique of the Asset Pricing Theory’s Tests: On Past and Potential Testability
of the Theory.” Journal of Financial Economics 4: 129-176.
Roll, Richard, and Ross, Stephen A. 1980. “An Empirical Investigation of the Arbitrage Pricing
Theory.” Journal of Finance 35: 1073- 1104.
Rosenberg, Barr. 1974. “Extra-Market Components of Covariance in Security Returns.” Journal of
Financial and Quantitative Analysis 9: 263-274.
Ross, Stephen A. 1976. “The Arbitrage Theory of Capital Asset Pricing.” Journal of Economic Theory
13: 341-360.
Sharpe, William F. 1963. “A Simplified Model of Portfolio Analysis.” k4unagement Science 9: 277-
293.