The opportunity cost of E-V optimal portfolios.
Tompkins, Daniel L.
ABSTRACT
This paper examines empirical common stock data sets to investigate
the opportunity costs of mean--variance optimal portfolios as compared
to an investor's direct expected utility maximization. We
investigate whether the opportunity cost can be attributed to (1) the
investor's utility function, (2) the number of securities included
in the data sets, (3) the number of observation used to determine the
optimal portfolios, or (4) the investor's preferences for skewness and kurtosis. We have found, that the E-V model provides a very low cost
approximation to the actual optimal portfolio. Eighty-one percent of the
mean-variance optimal portfolios all the cases have an opportunity cost
that equal zero.
INTRODUCTION
Investing has two stages. The first is to determine which assets to
invest in and the second is to determine how much should be put in each
asset. Markowitz' (1952) developed the Mean Variance model
(hereafter, E-V) portfolio optimization model to solve the second stage.
Using this model, investors maximize their portfolio by finding the
efficient frontier--selecting securities with the highest return for the
given amount of risk or the lowest risk for the given amount of return.
However, many argue that those who use the E-V model might not be
maximizing their utility. Some argue that for E-V to hold one of two
conditions must exist: (1) the individual must have a quadratic utility
function, or (2) security returns must be normally distributed. The
first condition is argued to be invalid because the quadratic utility
function possesses undesirable properties, such as increasing absolute
risk aversion. The second condition doesn't hold since securities
aren't normally distributed.
Though these conditions may not hold, existing evidence shows that
the E-V model accurately approximates an investor's optimal
portfolio. Evidence from Levy and Markowitz (1979); Pulley (1981 and
1982); Kroll, Levy, and Markowitz (1984); and Tew, Reid, and Witt (1991)
shows that E-V performs well, even when the investor's utility
function differs from the quadratic. To test how closely the results of
the E-V model approximated the direct expected utility maximization
model (EUM) results Levy and Markowitz used the correlation coefficient;
Pulley developed the Pulley Score; and Kroll, Levy, and Markowitz
developed the KLM Index. However, it has been shown that the Pulley
Score is not invariant to a linear transformation of the utility
function. Also, the Kroll, Levy, and Markowitz Index is not linearly
invariant to a transformation of the naive portfolio.
Another method used to find the goodness of fit for the E-V
approximation is Simaan's (1986) Opportunity Cost Approach. This
method can be explained as follows: suppose an investor has a choice
between the free use of the E-V approximation to the optimal portfolio
or the use of a costly direct EUM optimal portfolio. Let E(r) equal the
return on the EUM portfolio and E(r') equal the E-V
portfolio's expected return. Then it follows that the investor will
be indifferent between the choice if:
E (r - O) = E (r') (1)
where O equals the opportunity cost of the E-V solution. Tew, Reid,
and Witt (1989) simulated farm data to find the opportunity cost of an
E-V approximation. They found that for 76% of their samples, the
opportunity cost equaled zero. In only six percent of the cases did the
costs exceed $0.00025 and the highest opportunity cost amounted to
$0.04175. They concluded that an E-V approximation adequately
represented the actual situation.
The purpose of this research is to provide information on the
opportunity cost of E-V optimal common stock portfolios. By examining
empirical data, the research investigates whether the opportunity cost
can be attributed to (1) the investor's utility function, (2) the
number of securities included in the data sets, (3) the number of
observations used to determine the optimal portfolios, or (4) the
investor's preferences for skewness and kurtosis.
DATA AND METHODOLOGY
As previously mentioned, this paper examines the opportunity cost
of the E-V approximation to the direct EUM optimal solution. Thus, for
each data set we need to find the EUM optimal portfolio and the E-V
approximate solution for each utility function. Direct expected utility
maximization will be computed with the use of non-linear mathematical
programming. The general form of the model to maximize EU:
= SUM (U (SUM [r.sub.jt] [X.sub.j])) (2) subject to,
X, a > b
X > = 0
Sum [X.sub.j] = 1.0
Where [r.sub.jt] is the daily return of asset j for observation t;
[X.sub.j] is the percentage of available capital devoted to asset j; n
is the number of securities from which a portfolio is selected;
U(Sum([r.sub.jt][X.sub.j])) is the value of the utility function for the
portfolio of assets for observation t where U is the specific utility
function; a is a (q * n) matrix of constraint coefficients. This
constraint, a budget constraint, requires that all available funds are
spent. Thus, the program is put in the return form. Also, b is the
q-dimensional column vector of constraint resource levels; x is the
n-dimensional column vector of [X.sub.j,], the weight for each asset in
the portfolio. Finally, 0 is a n-dimensional column vector of zeros.
Thus the user specifies the right-hand side of the constraint
matrix (-1.0), the return matrix, the upper bounds (maximum investment
allowed for investment in security j), the lower bounds (the minimum
amount to be invested in each security), and the initial starting
position.
We will compare the optimal portfolios obtained through direct
utility maximization with the optimal portfolios obtained through an E-V
approximation. The algorithm used to determine the E-V optimal portfolio
involved several steps. First, it used a linear programming subroutine to solve for the top half of the efficient frontier. The second step
involves finding the minimum variance for each expected value. Thus, it
traced out the E-V efficient frontier by each combination of E and
minimized variance. During this search, the program calculated the
expected utility of each of the utility functions to see if it had
reached a maximum along the frontier. The solution from this second
search was determined to be the optimal E-V utility (denoted as E*U).
In both the direct EUM and the E-V approximation we will find the
optimal solutions for several different utility functions. We use the
same utility functions as the papers by Levy and Markowitz (1979),
Pulley (1981), and Kroll, Levy, and Markowitz (1979). They are:
Logarithmic: U(TR) = LN(TR) (3)
Power: U(TR) = [W.sup-a] (4)
Negative Exponential: U(TR) = [-e.sup.-aW] (5)
where "TR" equals the terminal total return value and
"a" equals the risk aversion parameter. We will calculate the
risk aversion parameter over the same range as used in the previous
studies. As given in Pulley (1981) these ranges for the Power and
Negative Exponential functions are:
Power: a = .1, .5, .9
Negative Exponential: a = 1, 2, 3, 4, 5
The objective function for the programming will be from a
Taylor's series expansion. To represent investors who consider the
mean, variance, skewness, and kurtosis, four forms of the Taylor's
series expansion will be used. The first one represents investors who
only consider the security's expected return:
F(x) = f(x)/0! +(f'(x)/1!) (x-[x.sub.0]) (6)
The second form represents those who consider the expected return
and variance:
F(x) = f(x)/0! + (f'(x)/1!) (x-[x.sub.0]) + (f"(x)/2!)
[(x-[x.sub.0]).sub.2] (7)
The third form adds consideration of skewness:
F(x) = f(x)/0! + (f'(x)/1!) (x-x0) + (f"(x)/2!) (x-x0)2 +
(f'"(x)/3!)([(x-[x.sub.0]).sub.3] (8)
Finally, the last form considers expected return, variance,
skewness, and kurtosis:
F(x) = f(x)/0! + (f'(x)/1!) (x-[x.sub.0])+ (f"(x)/2!)
[(x-[x.sub.0]).sup.2] + (f;'"(x)/3!)[(x-[x.sub.0]).sup.3] +
f""(x)/4!)[(x-[x.sub.0]).sup.4] (9)
Thus, after the user inputs the relative risk aversion level (a)
and the initial weights, the procedure checks to find out what the value
of the objective function is (it seeks to maximize the investor's
utility). Then, through an iterative process, it changes the weight of
each security. With each iteration, it calculates what the value of the
objective function is, as it tries to find the weights that provide the
highest utility to the investor.
In collecting the data we had several considerations. We wanted to
examine the issue empirically.. We also wanted to examine whether the
number of securities or observations, or the investor's preferences
for skewness or kurtosis provides for differences in the opportunity
cost. The empirical data sets are taken from the returns of those firms
that make up the Standard & Poor's 500 Stock Index. We chose
these securities for several reasons: (1) many large institutional
investors are restricted to investing in these securities because of
size or other constraints such as the Prudent Investor rule, and (2)
Since these are among the largest securities (in terms of market value),
they are among the most liquid securities. The time period for the data
was chosen randomly. We gathered daily returns, including dividends,
from Nov. 1987 to Dec. 1988.
To examine if the number of securities makes a difference in the
opportunity cost of the E-V approximation, we vary the number of firms
in the data sets. Though they have not explicitly explored this issue,
other studies such as Burgess and Bey (1988), have also varied the
number of firms included in the data set. We randomly selected firms for
portfolios of size ten (an amateur portfolio), twenty, thirty, and forty
firms (a small professional portfolio). Since as many as three
portfolios are selected randomly for each size, the portfolios are not
necessarily mutually exclusive. We also examine whether the number of
observations will matter. To examine this issue, we vary the number of
observations for each firm in the portfolio. We will use data sets with
100, 200, and 300 observations. We have chosen these observations for
the following reasons. First, as Lau and Wingerter (1989) point out, a
problem with calculating skewness and kurtosis is the large sampling
errors. With less than a few hundred observations, Lau and Wingerter
say, skewness and kurtosis may have misleading and erratic behavior.
And, as Ederington (1988) mentions, using a small sample of annual
returns may not include any observations from the tails of the
distribution. However, by expanding a data set of annual returns (if
enough data would be available), Ederington adds there may be a problem
in that the distribution has changed from the original one. Thus to get
large enough samples from within the same distribution, we have chosen
to use daily data. Table 1 shows the number of firms and observations
that are used in each of the data sets.
The measurement method to determine the quality of the E-V solution
is the opportunity cost approach. For a constant relative risk aversion
function, we can directly solve for the opportunity cost. For other
utility functions we employ a search routine to find the opportunity
cost. The search routine involves specifying the returns for the data
set, the E-V and EU portfolio weights for each asset, the utility
function, risk aversion level, and initial value for the opportunity
cost. The routine calculates and compares the value of EU(R-C) and
EU(R'). Then it adjusts the value of C in small increments until
the difference between EU(R-C) and EU(R') becomes negligible.
RESULTS
The results show that generally E-V provides a very good
approximation of the direct EUM optimal portfolio. Table 2 provides the
opportunity costs of the E-V approximation for the data sets. Eighty-one
percent of the portfolios have an opportunity cost of zero.
Next, we broke the results down to examine whether the
investor's utility function would account for any differences in
the opportunity costs. As seen in Table 3, the utility function does not
make a significant difference in the opportunity cost of the E-V
portfolio. Though the results for the negative exponential utility function with a risk parameter of 1, have the lowest percent of cases
with an opportunity cost of zero (fifty-nine percent), it still has a
high percentage with an opportunity cost of $0.0000025 or less
(eighty-two percent). And all of the other utility functions have higher
percentages of cases with an opportunity cost of $0.0000025 or less (and
greater percentage of case with a zero costs).
Since one criticism of E-V is that it assumes the investor only
considers a security's mean and variance, we wanted to see if
inclusion of skewness and kurtosis would result in a larger opportunity
cost for the E-V optimal portfolio. Table 4 summarizes the opportunity
cost results as a function of the number of moments included in the
utility function. The results show that the E-V approximations with the
fewest cases with a zero opportunity cost are the utility functions that
only include the first moment. For the first moment investors, seventy-f
our percent of the cases have an opportunity cost of zero. However,
eighty-eight percent of the cases for the first moment category have an
opportunity cost of $0.0000025 or less per dollar invested, similar to
the other categories.
With ten percent of the cases having an opportunity cost more than
$0.00025 per dollar invested or more, the utility functions that contain
the first three moments have the most cases in the highest cost
category. However, the other ninety percent of the cases have
opportunity costs of $0.0000025 or less, similar to the other
categories. The E-V approximation performs best for the utility
functions that include all four moments. In this category, eighty-three
percent of the cases have an opportunity cost of zero with an additional
nine percent having an opportunity cost of $0.000025 or less.
Table 5 breaks down the results by the number of securities
included in the data sets. This table shows there is little difference
among the ten, twenty, and forty securities sets. These categories
respectively have ninety-six, ninety-five, and one hundred percent of
their cases with an opportunity cost of $0.000025 or less. The twenty
securities category has ninety-two percent of its cases with an
opportunity cost of zero. The forty securities category is close behind
with ninety percent of the cases with an opportunity cost of zero. The
ten security data set has a total of seventy-nine percent of its cases
with an opportunity cost of zero. Though most of the thirty security
cases (sixty-six percent) have an opportunity cost of zero, twenty-nine
percent of the cases have a cost between zero and $0.00025 per dollar
invested. In general, the results show that the E-V model results in a
low cost approximation of the investors direct expected utility
maximization portfolio at a low cost.
Finally, we examined the empirical results of the opportunity cost
as a function of the number of observations. Table 6 summarizes this
information. The 200 and 300 observation categories each provide E-V
approximations with very low opportunity cost. Respectively,
Ninety-seven and ninety-five percent of the cases have an opportunity
cost of $0.0000025 or less. Eighty-three percent of the cases for the
200-observation category have an opportunity cost of zero, with just two
cases having a cost of $0.00025 or more. For the 300-observation
category, ninety-three percent of the cases have a cost of zero, with a
single case that has an opportunity cost of $0.00025 or more. The
100-observation category has a majority of cases that have an
opportunity cost of zero, sixty-nine percent; however, a greater
percentage of cases, sixteen percent, have an opportunity cost of
$0.00025 per dollar invested or higher.
CONCLUSIONS
We have used the opportunity cost approach to examine common stock
portfolios to evaluate how closely the E-V model approximates the direct
EUM optimal portfolio. We find that the E-V model provides a very low
cost approximation to the actual optimal portfolio. Eighty-four percent
of all the cases have an opportunity cost that equal zero.
These results are similar to the results reported by Tew, Reid, and
Witt, as well as the other studies using different comparison methods.
Thus we conclude that investors can use the E-V model and closely
approximate their actual optimal portfolio.
REFERENCES
Beedles, W.L. (1979). On the Asymmetry of Market Returns, Journal
of Finance and Quantitative Analysis, 14 (Sept), 653-660.
Burgess, R. C. & Bey, R. P. (1988). Optimal Portfolios:
Markowitz Full Covariance Versus Simple Selection Rules, Journal of
Financial Research, 11 (Summer), 153 164.
Ederington, L. H. (1986). Mean-Variance as an Approximation to
Expected Utility Maximization, Washington University working paper no
86-5.
Fama, E. (1968). Risk, Return and Equilibrium: Some Clarifying
Comments, Journal of Finance, (March) 29-40.
Francis, J. C. (1975). Skewness and Investor's Decision,
Journal of Finance and Quantitative Analysis,(March) 163-172.
Gelles, G. M. & Mithcell, D. W. (1999). Broadly Decreasing Risk
Aversion, Management Science, 45 (October), 1432-1439.
Grauer, R. (1986). Normality, Solvency, and Portfolio Choice,
Journal of Finance and Quantitative Analysis, 21 (September), 265-270.
Kroll, Y., Levy, H. & Markowitz, H. (1984). Mean-Variance
Versus Direct Utility Maximization, Journal of Finance, 39 (March),
47-61.
Lau, H.-S. & Wingender, J. R. (1989). The Analytics of
Intervaling Effect on Skewness and Kurtosis of Stock Returns, The
Financial Review, 24 (May), 215-234.
Levy, H. & Markowitz, H. (1979). Approximating Expected Utility
by a Function of Mean and Variance, American Economic Review, 69 (June),
308-317.
Loistl, O. (1976). The Erroneous Approximation of Expected Utility
by Means of a Taylor Series Expansion: Analytic and Computational
Results, American Economic Review, 66 (Dec), 904-910.
Markowitz, H. (1952). Portfolio Selection, Journal of Finance, 7
(1), 77-91.
Markowitz, H. (1959). Portfolio Selection: Efficient
Diversification of Investments, New Haven, CT: Yale University Press.
Markowitz, H. (1987). Mean-Variance Analysis in Portfolio Choice
and Capital Markets, New York: Basil Blackwell.
Meyer, J. (1987). Three Moment Decision Models and Expected Utility
Maximization, American Economic Review, (June), 421-430.
Pulley, L. B. (1981). A General, Mean-Variance Approximation to
Expected Utility for Short Holding Periods, Journal of Financial and
Quantitative Analysis, 16 (Sept.). 361-373.
Pulley, L. B. (1983). Mean-Variance Approximations to Expected
Logarithmic Utility, Operations Research, 31 (July-August), 686-696.
Quirk, J. P. & Saposnik, R. (1962). Admissibility and
Measurable Utility Functions, Review of Economic Studies, 29, 140-146.
Reid, D. W. & Tew, B. V. (1986). Mean-Variance versus Direct
Utility Maximization: A Comment, Journal of Finance, 41 (Dec),
1177-1179.
Reilly, F. K. (1988). Investment Analysis and Portfolio Management,
Chicago: The Dryden Press.
Rubinstein, M. (1976). The Strong Case for the Generalized
Logarithmic Utility Function as the Premier Model of Financial Markets,
Journal of Finance, 31 (May), 555-571.
Samuelson, P. (1967). A General Proof that Diversification Pays,
Journal of Finance and Quantitative Analysis, 2 (1),. 1-13.
Samuelson, P. (1970). The Fundamental Approximation Theorem of
Portfolio Analysis in Terms of Means, Variances, and Higher Moments, The
Review of Economic Studies, 37,. 537-542.
Simkowitz, M.A. & Beedles, W. L. (1978). Diversification in a
Three Moment World, Journal of Finance and Quantitative Analysis, 13
(Dec.). 927-941.
Simaan, Y. (1986). Portfolio Selection and Capital Asset Pricing
for a Class of Non-Spherical Distributions of Asset Returns, Ph.D.
Dissertation, Baruch College, City University of New York.
Statman, (1999). Foreign Stocks in Behavioral Portfolios, Financial
Analysts Journal, 55 (March-April),. 12-16.
Tew, B. V. & Reid, D. H. (1987). More Evidence on the Expected
Value-Variance Analysis Versus Direct Utility Maximization, Journal of
Financial Research, 10 (Fall). 249-257.
Tew, B. V. & Witt, C. A. (1991). Opportunity Cost of a
Mean-Variance Efficient Choice, Financial Review, 26 (Feb), 31-44.
Tsiang, S.C. (1972). The rationale of the Mean-Standard Deviation
Analysis, Skewness Preference, and the Demand for Money, American
Economic Review, 62,. 354-371.
Witt, C. A. (1988). Agriculture Diversification Under Risk:
Pragmatic Use of E-V Portfolio Analysis, Ph.D. Dissertation, University
of Kentucky
Zopounidis, C., Doumpos, M. & Zanakis, S. (1998). Stock
Evaluation Using a Preference Disaggregation Methodology, Decision
Sciences, 30 (Spring), 313-336.
Daniel L. Tompkins, Niagara University
TABLE 1:
Number of Firms and Observations
Data Set number of firms in portfolio number of observations
1 10 100
2 10 200
3 10 300
4 20 100
5 20 200
6 20 300
7 30 100
8 30 200
9 30 300
10 40 100
11 40 200
12 40 300
Table 2:
Summary of Opportunity Cost Results
Opportunity Cost
0.00000 to 0.0000025 to greater
0.00000 0.0000025 0.00025 than 0.00025 Totals
321 (81%) 34 (8%) 6 (2%) 35 (9%) 396 (100%)
Opportunity Cost is per dollar invested.
Table 3:
Summary of Opportunity Cost Results As a Function of Investor's
Utility Function
Opportunity Cost
Utility Function
Negative 0.00000 0.00000 to 0.0000025 to
Exponential 0.0000025 0.00025
1 26 (59%) 10 (23%) 3 (7%)
2 30 (68%) 7 (16%) 3 (7%)
3 30 (68%) 9 (20%) 0 (0%)
4 39 (89%) 1 (2%) 0 (0%)
5 41 (93%) 0 (0%) 0 (0%)
Power
.1 40 (91%) 0 (0%) 0 (0%)
.5 40 (91%) 0 (0%) 0 (0%)
.9 40 (91%) 0 (0%) 0 (0%)
Logarithmic 35 (80%) 7 (16%) 0 (0%)
Totals 321 (81%) 34 (8%) 6 (2%)
Opportunity Cost
Utility Function
Negative greater than Totals
Exponential 0.00025
1 5 (11%) 44 (100%)
2 4 (9%) 44 (100%)
3 5 (11%) 44 (100%)
4 4 (9%) 44 (100%)
5 3 (7%) 44 (100%)
Power
.1 4 (9%) 44 (100%)
.5 4 (9%) 44 (100%)
.9 4 (9%) 44 (100%)
Logarithmic 2 (4%) 44 (100%)
Totals 35 (9%) 396 (100%)
Opportunity Cost is per dollar invested.
Table 4:
Summary of Opportunity Cost Results As a Function of Number of
Moments Included in Utility Function
Opportunity Cost
Number of 0.00000 0.00000 to 0.0000025 to
Moments 0.0000025 0.00025
1 73 (74%) 14 (14%) 5 (5%)
2 81 (82%) 8 (8%) 1 (1%)
3 83 (84%) 6 (6%) 0 (0%)
4 84 (85%) 6 (6%) 0 (0%)
Totals 321 (81%) 34 (8%) 6 (2%)
Opportunity Cost
Number of greater than Totals
Moments 0.00025
1 7 (7%) 99 (100%)
2 9 (9%) 99 (100%)
3 10 (10%) 99 (100%)
4 10 (9%) 99 (100%)
Totals 35 (9%) 396 (100%)
Opportunity Cost is per dollar invested.
Table 5:
Summary of Opportunity Cost Results As a Function of Number of
Securities Included in Data Set
Opportunity Cost
Number of 0.00000 0.00000 to 0.0000025 to
Securities 0.0000025 0.00025
10 85 (79%) 19 (17%) 0 (0%)
20 100 (92%) 3 (3%) 5 (5%)
30 71 (66%) 5 (4%) 1 (1%)
40 65 (90%) 7 (10%) 0 (0%)
Totals 321 (81%) 34 (8%) 6 (2%)
Opportunity Cost
Number of greater than Totals
Securities 0.00025
10 4 (4%) 108 (100%)
20 0 (0%) 108 (100%)
30 31 (29%) 108 (100%)
40 0 (0%) 72 (100%)
Totals 35 (9%) 396 (100%)
Opportunity Cost is per dollar invested.
Table 6:
Summary of Opportunity Cost Results As a Function of Number of
Observations
Opportunity Cost
Number of 0.00000 0.00000 to 0.0000025 to
Observations 0.0000025 0.00025
100 100 (69%) 12 (6%) 0 (0%)
200 120 (83%) 20 (14%) 2 (1.5%)
300 101 (93%) 2 (2%) 4 (4%)
Totals 321 (81%) 34 (8%) 6 (2%)
Opportunity Cost
Number of greater than Totals
Observations 0.00025
100 32 (16%) 144 (100%)
200 2 (1.5%) 144 (100%)
300 1 (1%) 108 (100%)
Totals 35 (9%) 396 (100%)
Opportunity Cost is per dollar invested.