Journal ofPortfolio Management, 19, 6-11 (1 993) 249

18
The Effect of Errors in Means,
Variances, and Covariances on Optimal
Portfolio Choice*
Vijay K. Chopra and William T. Ziemba
The Kelly Capital Growth Investment Criterion Downloaded from www.worldscientific.com

Good mean forecasts are critical to the mean-variance framework

There is considerable literature on the strengths and limitations of mean-

variance analysis. The basic theory and extensions of MV analysis are dis-
by KAINAN UNIVERSITY on 02/14/17. For personal use only.

cussed in Markowitz [1987] and Ziemba & Vickson [1975]. Bawa, Brown &
Klein [1979] and Michaud [1989] review some of its problems.
MV optimization is very sensitive to errors in the estimates of the inputs.
Chopra [1993] shows that small changes in the input parameters can result in
large changes in composition of the optimal portfolio. Best & Grauer [1991]
present some empirical and theoretical results on the sensitivity of optimal
portfolios to changes in means. This article examines the relative impact of
estimation errors in means, variances, and covariances.
Kallberg & Ziemba [1984] examine the question of mis-specification in
normally distributed portfolio selection problems. They discuss three areas
of misspecification: the investor's utility function, the mean vector, and the
covariance matrix of the return distribution.
They find that utility functions with similar levels of Arrow-Pratt absolute
risk aversion result in similar optimal portfolios irrespective of the functional
form of the utilityl; Thus, mis-specification of the utility function is not a
major concern because several different utility functions (quadratic, nega-
tive exponential, logarithmic, power) result in similar portfolio allocations
for similar levels of risk aversion.
Misspecification of the parameters of the return distribution, however, does
make a significant difference. Specifically, errors in means are at least ten
times as important as errors in variances and covariances.
We show that it is important to distinguish between errors in variances
and covariances. The relative impact of errors in means, variances, and co-
variances also depends on the investor's risk tolerance. For a risk tolerance of
-Reprinted, with permission, from Journal 0/ Portfolio Management, 1993. Copyright
1993 Institutional Investor Journals.
IFor an investor with utility function U and wealth W, the Arrow- Pratt absolute risk
aversion is ARA == -U"(W)jU'(W) . Friend and Blume 11975] show that investor behavior
is consistent with decreasing ARA; that is, as investors' wealth increases, their aversion to
a given risk decreases.

53
 250 V K. Chopra and W T Ziemba

54 Chopra and Ziemba

50, errors in means are about eleven times as important as errors in variances,
a result similar to that of Kallberg & Ziemba. 2 Errors in variances are about
twice as important as errors in covariances.
At higher risk tolerances, errors in means are even more important relative
to errors in variances and covariances. At lower risk tolerances, the relative
impact of errors in means, variances, and covariances is closer. Even though
errors in means are more important than those in variances and covariances,
the difference in importance diminishes with a decline in risk tolerance.
The Kelly Capital Growth Investment Criterion Downloaded from www.worldscientific.com

These results have an implication for allocation of resources according to

the MV framework. The primary emphasis should be on obtaining superior
estimates of means, followed by good estimates of variances. Estimates of
by KAINAN UNIVERSITY on 02/14/17. For personal use only.

covariances are the least important in terms of their influence on the optimal
portfolio.

Theory
For a utility function U and gross returns r - i (or return relatives) for assets
i = 1,2, .. . , N, an investor's optimal portfolio is the solution to:
N
maximize Z(x) = E[U(Wo ~]ri)xi)]
i=l
N
such that Xi> 0, L= 1,
i=1

where Z(x) is the investor's expected utility of wealth, Wo is the investor's

initial wealth, the returns ri have a distribution F(r), and Xi are the portfolio
weights that sum to one.
Assuming a negative exponential utility function U(W) = - exp( -aW)
and a joint normal distribution of returns, the expected utility maximization
problem is equivalent to the MV-optimization problem:
N 1 N N
maximize Z(x) = L E[ri]Xj - t L L xixjE[Uij]
i=1 i=lj=1
N
such that Xi > 0, LXi = 1,
i=l

2The risk tolerance reflects the investor's desired trade-off between extra return and
extra risk (variance). It is the inverse slope of the investor's indifference curve in mean-
variance space. The greater the risk tolerance, the more risk an investor is willing to
take for a little extra return. Under fairly general input assumptions, a risk tolerance of 50
describes the typical portfolio allocations of large US pensions funds and other institutional
investors. Risk tolerances of 25 and 75 characterize extremely conservative and aggressive
investors, respectively.
 The Effect ofErrors in Means, Variances, and Covariances on Optimal Portfolio Choice 251

The Effect of Errors 55

EXHIBIT 1:
List of Ten Randomly Chosen DJIA Securities
1. Aluminum Co. of America
2. American Express Co.
3. Boeing Co.
4. Chevron Co.
5. Coca Cola Co.
The Kelly Capital Growth Investment Criterion Downloaded from www.worldscientific.com

6. E.!. Du Pont De Nemours & Co.

7. Minnesota Mining and Manufacturing Co.
8. Procter & Gamble Co.
9. Sears, Roebuck & Co.
by KAINAN UNIVERSITY on 02/14/17. For personal use only.

10. United Technologies Co.

where E[TiJ is the expected return for asset i, t is the risk tolerance of the
investor, and E[OijJ is the covariance between the returns on assets i and j.3
A natural question arises: How much worse off is the investor if the distri-
bution of returns is estimated with an error? This is an important considera-
tion because the future distribution of returns is unknown. Investors rely on
limited data to estimate the parameters of the distribution, and estimation
errors are unavoidable. Our investigation assumes that the distribution of
returns is stationary over the sample period. If it is time-varying or non-
stationary, the estimated parameters will be erroneous.
To measure how close one portfolio is to another, we compare the cash
equivalent (CE) values of the two portfolios. The cash equivalent of a risky
portfolio is the certain amount of cash that provides the same utility as the
risky portfolio, that is, U(CE) = Z(x) or CE = U- 1 [Z(x)] where, as defined
before, Z(x) is the expected utility ofthe risky portfolio.4 The cash equivalent
is an appropriate measure because it takes into account the investor's risk
tolerance and the inherent uncertainty in returns, and it is independent of
utility units. For a risk-free portfolio, the cash equivalent is equal to the
certain return.
Given a set of asset parameters and the investors risk tolerance, a MY-
optimal portfolio has the largest CE value of any portfolio of those assets. The
3 Although the exponential utility function is convenient for deriving the MV problem
with normally distributed returns, the MV framework is consistent with expected utility
maximization for any concave utility function, assuming normality.
4For negative exponential utility, Freund [19561 shows that the expected utility of port-
folio x is Z(x) = 1 - exp( -aE[x] + (a 2 /2)Var[x]), where E[X] and Var[xl are the expected
return and variance of the portfolio. The cash equivalent is eEl: = (lla) log(l - Z(x)).
If returns are assumed to have a multivariate normal distribution, this is also the cash
equivalent of an MV-optimal portfolio. See Dexter, Yu & Ziemba [1980] for more details.
 252 V K. Chopra and W T. Ziemba

56 Chopra and Ziemba

percentage cash equivalent loss (CEL) from holding an arbitrary portfolio, x

instead of an optimal portfolio 0 is
CEL = CEo - CEx
CEo
where CEo and CEx are the cash equivalents of portfolio 0 and portfolio x
respectively.
The Kelly Capital Growth Investment Criterion Downloaded from www.worldscientific.com

Data and Methodology

The data consist of monthly observations from January 1980 through Decem-
by KAINAN UNIVERSITY on 02/14/17. For personal use only.

ber 1989 on ten randomly selected Dow Jones Industrial Average (DJIA) se-
curities. We use the Center for Research in Security Prices (CRSP) database,
having deleted one security (Allied-Signal, Inc.) because of lack of data prior
to 1985. Each of the remaining twenty-nine securities had an equal probabil-
ity of being chosen. The securities are listed in Exhibit 1.
MV optimization requires as inputs forecasts for: mean returns, variances,
and covariances. We computed historical means (Ti), variances (aii), and
covariances (aij), and assumed that these are the 'true' values of these pa-
rameters. Thus, we assumed that E[ri] = Ti, E[aii] = aii, and E[aij] = aij' A
base optimal portfolio allocation is computed on the basis of these parameters
for a risk tolerance of 50 (equivalent to the parameter a = 0.04).
Our results are independent of the source of the inputs. Whether we use
historical inputs or those based on a complete forecasting scheme, the results
continue to hold as long as the inputs have errors.
Exhibit 2 gives the input parameters and the optimal base portfolio re-
sulting from these inputs. To examine the influence of errors in parameter
estimates, we change the true parameters slightly and compute the resulting
optimal portfolio. This portfolio will be suboptimal for the investor because
it is not based on the true input parameters.
Next we compute the cash equivalent values of the base portfolio and the
new optimal portfolio. The percentage cash equivalent loss from holding the
suboptimal portfolio instead of the true optimal portfolio measures the impact
of errors in input parameters on investor utility.
To evaluate the impact of errors in means, we replaced the assumed true
mean Ti for asset i by the approximation Ti(l + kzi ) where Zi has a standard
normal distribution. The parameter k is varied from 0.05 through 0.20 in
steps of 0.05 to examine the impact of errors of different sizes. Larger values
of k represent larger errors in the estimates. The variances and covariances
are left unchanged in this case to isolate the influence of errors in means.
The percentage cash equivalent loss from holding a portfolio that is optimal
for approximate means Ti (1 + kzi ) but is suboptimal for the true means r, is
 The Effect of Errors in Means, Variances, and Covariances on Optimal Portfolio Choice 253

The Effect of Errors 57

EXHIBIT 2:
Inputs to the Optimization and the Resulting Optimal Portfolio for a
Risk Tolerance of 50 (January 198G-December 1989)
Alcoa Amex Boeing Chev. Coke Du Pont MMM P&G Sears UTech
Means ('Yo
per month) 1.5617 1.9477 1.907 1.5801 2.1643 1.6010 1.4892 1.6248 1.4075 1.1537
Std. Dev. ('Yo
per month) 8.8308 8.4585 10.040 8.6215 5.988 6 .8767 5.8162 5.6385 8.0047 8.212
The Kelly Capital Growth Investment Criterion Downloaded from www.worldscientific.com

Correlations
Alcoa 1.0000
by KAINAN UNIVERSITY on 02/14/17. For personal use only.

Amex 0.3660 1.0000

Boeing 0.3457 0 .5379 1.0000
Chev. 0.1606 0.2165 0.2218 1.0000
Coke 0.2279 0.4986 0.4283 0 .0569 1.0000
Du Pont 0.5133 0.5823 0.4051 0.3609 0.3619 1.0000
MMM 0.5203 0.5569 0.4492 0.2325 0.4811 0.6167 1.0000
P&G 0.2176 0.4760 0.3867 0.2289 0.5952 0.4996 0.6037 1.0000
Sears 0.3267 0.6517 0 .4883 0.1726 0 .4378 0.58Il 0.5671 0.5012 1.0000
UTech 0.5101 0 .5853 0.6569 0.3814 0.4368 0.5644 0.6032 0 .4772 0.6039 1.0000

Optimal Port .
Weights 0 .0350 0.0082 0 .0 0.1626 0.7940 0.0 0.0 0.0 0.00 0 .00

then computed. This procedure is repeated with a new set of Z values for a
total of 100 iterations for each value of k.
To investigate the impact of errors in variances each variance forecast O'ii
was replaced by O'ii(l + kZj ). To isolate the influence of variance errors, the
means and covariances are left unchanged.
Finally, the influence of errors in covariances is examined by replacing
each covariance O'ij (i # j) by O'ij + kZ ij where Zij has a standard normal
distribution, while retaining the original means and variances. The procedure
is repeated 100 times for each value of k, each time with a new set of Z values,
and the cash equivalent loss computed. The entire procedure is repeated for
risk tolerances of 25 and 75 to examine how the results vary with investors'
risk tolerance.

Results
Exhibit 3 shows the mean, minimum, and maximum cash equivalent loss over
the 100 iterations for a risk tolerance of 50. Exhibit 4 plots the average eEL
 254 V K. Chopra and W T. Ziemba

58 Chopra and Ziemba

EXHIBIT 3:
Cash Equivalent Loss (CEL) for Errors of Different Sizes
k (size of Parameter Mean Min. Max.
error) with Error CEL CEL CEL
0.05 Means 0.66 0.01 5.05
0.05 Variances 0.05 0.00 0.34
0.05 Covariances 0.02 0.00 0.25
The Kelly Capital Growth Investment Criterion Downloaded from www.worldscientific.com

0.10 Means 2.45 0.01 15.61

0.05 Variances 0.22 0.00 1.39
0.10 Covariances 0.11 0.00 0.66
by KAINAN UNIVERSITY on 02/14/17. For personal use only.

0.15 Means 5.12 0.15 24.35

0.15 Variances 0.55 0.00 3.35
0.15 Covariances 0.27 0.00 1.11

0.20 Means 10.16 0.17 3609

0.20 Variances 0.90 0.01 4.16
0.20 Covariances 0.47 0.00 1.94

as a function of k. The CEL for errors in means is approximately eleven

times that for errors in variances and over twenty times that for errors in
covariances. Thus, it is important to distinguish between errors in variances
and errors in covariances.:; For example, for k = 0.10, the CEL is 2.45 for
errors in means, 0.22 for errors in variances, and 0.11 for errors in covariances.
Our results on the relative importance of errors in means and variances
are similar to those of Kallberg & Ziemba [1984J. They find that errors in
means are approximately ten times as important as errors in variances and
covariances considered together (they do not distinguish between variances
and covariances).
Our results show that for a risk tolerance of 50 the importance of errors
in covariances is only half as much as previously believed. Furthermore, the
relative importance of errors in means, variances, and covariances depends
upon the investor's risk tolerance.
Exhibit 5 shows the average ratio (averaged over errors of different sizes, k)
of the CELs for errors in means, variances, and covariances. An investor with
a high risk tolerance focuses on raising the expected return of the portfolio
5The result for covariances also applies to correlation coefficients, as the correlations
differ from the covariances only by a scale factor equal to the product of two standard
deviations.
 The Effect a/Errors in Means, Variances, and Covariances on Optimal Portfolio Choice 255

The Effect of Errors 59

% Cash
Equivalant lo..
11,---------------------------~----_,
Means
10
9
8
7
6
5
4
The Kelly Capital Growth Investment Criterion Downloaded from www.worldscientific.com

3
2
___ .... __ .... . V.ri.~ce•
... ~_.:_-::::: .... _................ Covlnances
O~~--~-*~~~~~~----~----~
by KAINAN UNIVERSITY on 02/14/17. For personal use only.

o 0.05 0.10 0.15 0.20

Magnituda of error (kl

EXHIBIT 4
Mean percentage cash equivalent loss due to errors in inputs

EXHIBIT 5
Average Ratio of CELs for Errors in Means, Variances, and Covariances
Risk Errors in Means Errors in Means Errors in Variances
Tolerance versus Variances versus Covariances versus Covariances
25 3.22 5.38 1.67
50 1.98 22.50 2.05
75 21.42 56.84 2.68

and discounts the variance more relative to the expected return. To this
investor, errors in expected returns are considerably more important than
errors in variances and covariances. For an investor with a risk tolerance of
75, the average CEL for errors in means is over twenty-one times that for
errors in variances and over fifty-six times that for errors in covariances.
Minimizing the variance of the portfolio is more important to an investor
with a low risk tolerance than raising the expected return. To this investor,
errors in means are somewhat less important than errors in variances and
covariances. For an investor with a risk tolerance of 25, the average CEL for
errors in expected returns is about three times that for errors in variances
and about five times that for errors in covariances.
Most large institutional investors have a risk tolerance in the 40 to 60 range.
Over that range, there is considerable difference in the relative importance
of errors in means, variances, and covariances. Irrespective of the level of
risk tolerance, errors in means are the most important, followed by errors
 256 V K. Chopra and W T. Ziemba

60 Chopra and Ziemba

in variances. Errors in covariances are the least important in terms of their

influence on portfolio optimality.

Implications and Conclusions

Investors have limited resources available to spend on obtaining estimates of
necessarily unknowable future parameters of risk and reward. This analysis
The Kelly Capital Growth Investment Criterion Downloaded from www.worldscientific.com

indicates that the bulk of these resources should be spent on obtaining the
best estimates of expected returns of the asset classes under consideration.
Sometimes, investors using the MV framework to allocate wealth among
individual stocks set all the expected returns to zero (or a non-zero constant).
by KAINAN UNIVERSITY on 02/14/17. For personal use only.

This can lead to a better portfolio allocation because it is often very difficult
to obtain good forecasts for expected returns. Using forecasts that do not
accurately reflect the relative expected returns of different securities can sub-
stantially degrade MV performance.
In some cases it may be preferable to set all forecasts equal. 6 The opti-
mization then focuses on minimizing portfolio variance and does not suffer
from the error-in-means problem. In such cases it is important to have good
estimates of variances and covariances for the securities, as MV optimizes
only with respect to these characteristics.
Of course, if investors truly believe that they have superior estimates of
the means, they should use them. In this case it may be acceptable to use
historical values for variances and covariances.
For investors with moderate to high risk tolerance, the cash equivalent loss
for errors in means is an order of magnitude greater than that for errors in
variances or covariances. As variances and covariances do not much influence
the optimal MV allocation (relative to the means), investors with moderate-
to-high risk tolerance need not expend considerable resources to obtain better
estimates of these parameters.

References
Bawa, Vijay S., Stephen J. Brown and Roger W. Klein (1979). 'Estimation Risk and
Optimal Portfolio Choice.' Studies in Bayesian Econometrics, Bell Laboratories
Series. North Holland.
6This approach is in the spirit of Stein estimation and is discussed in Chopra, Hensel, and
'Thrner [1993] . As a practical matter, it should be used for assets that belong to the same
asset class. e.g., equity indexes of different countries or stocks within a country. It would
be inappropriate to apply it to financial instruments with very different characteristics; for
example, stocks and T-bills.
 The Effect ofErrors in Means, Variances, and Covariances on Optimal Portfolio Choice 257

The Effect of Errors 61

Best, Michael J. and Robert R. Grauer (1991). 'On the Sensitivity of Means-
Variance-Efficient Portfolios to Changes in Asset Means: Some Analytical and
Computational Results.' Review of Financial Studies 4, No.2, 315-342.
Chopra, Vijay K. (1991). 'Mean-Variance Revisited: Near-Optimal Portfolios and
Sensitivity to Input Variations.' Russell Research Commentary.
Chopra, Vijay K., Chris R. Hensel and Andrew L. Thrner (1993). 'Massaging
Mean-Variance Inputs: Returns from Alternative Global Investment Strategies
in the 1980s.' Management Science, (July): 845-855.
The Kelly Capital Growth Investment Criterion Downloaded from www.worldscientific.com

Dexter, Albert S., Johnny N.W. Yu and William T. Ziemba (1980). 'Portfolio
Selection in a Lognormal Market when the Investor has a Power Utility Func-
tion: Computational Results.' In Proceedings of the International Conference
on Stochastic Programming, M.A.H. Dempster (ed.), Academic Press, 507-523.
by KAINAN UNIVERSITY on 02/14/17. For personal use only.

Freund, Robert A. (1956). 'The Introduction of Risk into a Programming Model.'

Econometrica 24253-263.
Freund, L. and M. Blume (1975). 'The Demand for Risky Assets.' The American
Economic Review, December, 900-922.
Kallberg, Jarl G. and William T. Ziemba (1984). 'Mis-specification in Portfolio
Selection Problems'. In Risk and Capital, G. Bamberg and K. Spremann (eds.),
Lecture Notes in Econometrics and Mathematical Systems. Springer-Verlag.
Klein, Roger W, and Vijay S. Bawa (1976). 'The Effect of Estimation Risk on
Optimal Portfolio Choice. J. of Financial Economics 3 (June), 215-231.
Markowitz, Harry M. (1987) . Mean- Variance Analysis in Portfolio Choice and
Capital Markets. Basil Blackwell.
Michaud. Richard O. (1989) . 'The Markowitz Optimization Enigma: is 'Opti-
mized' Optimal?' Financial Analysts Journal 45 (January-February), 31-42.
Ziemba, William T. and Raymond G. Vickson, eds. (1975) Stochastic Optimization
Models in Finance. Academic Press.