Portfolio Optimization in A Mean-Semivariance Fram
Portfolio Optimization in A Mean-Semivariance Fram
Portfolio Optimization in A Mean-Semivariance Fram
Vigdis Boasson
AUTHORS Emil Boasson
Zhao Zhou
Vigdis Boasson, Emil Boasson and Zhao Zhou (2011). Portfolio optimization in a
ARTICLE INFO mean-semivariance framework. Investment Management and Financial
Innovations, 8(3)
0 0 0
businessperspectives.org
Investment Management and Financial Innovations, Volume 8, Issue 3, 2011
only the returns below a certain threshold. This thre- 1. Downside risk measurements
shold captures the risk perspectives from investors
1.1. Markowitz’s mean-variance framework. Mar-
to investors. Unlike standard deviation, downside
kowitz’s mean-variance approach, which leads to
risk accommodates different views of risk. Suppose
optimal investment decision, has two important
an investor is concerned only about losing the initial
limitations. Firstly, it assumes that the distribution
wealth, that threshold would be zero, and the proba-
of investment returns is jointly elliptically distri-
bility of losing the principal would be viewed as
buted, i.e., a symmetric bell-shaped distribution.
risky. Suppose an investor’s minimum required rate
However, if the underlying return data is not nor-
of return is 10 percent, any return below ten percent
mally distributed, the variance is likely to give mis-
would be considered risky. In the case of institu-
leading results. A number of studies have demon-
tional investors, this threshold can also be certain
strated that investment returns are not normally dis-
peer performance benchmark.
tributed (Fama & Roll, 1968; and Jansen & de
Semivariance, a special case of downside risk mea- Vries, 1991). In the real financial world, security
surement, is defined as the weighted sum of square returns tend to be asymmetrically distributed, e.g.,
deviations from certain threshold considering only approximately lognormal distribution. The skewed
those values below the expected value of returns distribution of investment returns makes the va-
(Ballestero, 2005). Semivariance also receives early riance as an inefficient risk measure, because va-
support from Markowitz (1959, 1970), who pre- riance treats the favorable upside dispersion of in-
sented it as an alternative risk measure. Since inves- vestment return over the mean value of return as a
tors are more concerned about downside risk than part of risk and penalizes it as much as the unfa-
about overall volatility, measuring risk by semiva- vorable downside deviation from the mean re-
riance, instead of variance, produces better portfo- turns. If the returns are not normally distributed,
lios Markowitz (1959). The study of mean-semiva- investors using variance or standard deviation to
riance efficient frontier has been focused on the nu- measure risk are likely to reach wrong asset allo-
merical calculation of the frontier and comparison cation decisions. Skewness and kurtosis in real
between mean-semivariance and mean-variance fron- return data with non-normal distributions can
tiers. To our knowledge, there has been a paucity of cause variance or standard deviation to underes-
research in empirically testing the advantages of the timate risk. Secondly, the mean-variance ap-
semivariance model over the variance model. proach ignores the investor’s risk aversion. Because
In this paper, we intend to fill this gap by empirical- the variance can only measure the dispersion of re-
ly constructing efficient portfolios and efficient fron- turns distribution around a mean, it cannot be custo-
tiers using the mean-semivariance model as well as mized for individual investors’ aversion.
the mean-variance model so as to compare the dif- Moreover, the real-world implementation of Marko-
ferences between the two models’ efficient frontiers witz’ mean-variance optimal portfolio construction has
and asset allocations. While acknowledging the many pitfalls. The optimal portfolio constructed in a
usefulness of other downside risk measurements mean-variance framework may not lead to an optimal
such as Variance-at-Risk (VaR) and Conditional portfolio that optimizes expected returns while mini-
Variance-at-Risk (VarCVaR), we do not intend to mizing risk. Mitchaud (1989) indicates that these
include VaR or CVaR into our portfolio optimiza- real world portfolio optimizers are essentially “error
tion framework for the reasons discussed in the lite- maximizers” because “optimizers” tend to treat the
rature review. We focus instead on asset allocations inputs as if they were exact quantities, while in reality
and portfolio optimization in a mean-semivariance they can only be estimated with error. The optimal
framework. Our selection of investment assets are portfolios constructed based on this framework tend
based on the investment portfolios in the insurance to suggest large bets on stocks with large estimation
industry, the industry which is most concerned with error in expected returns, often leading to poor-out-
the downside risk in security investment. of-sample performance. In fact, Markowitz (1970)
The remainder of this paper is organized as follows. himself realized the drawback of variance. He showed
Section 1 outlines the literature on downside risk mea- that both the downside risk measurement and the
surements. Section 2 describes the mean-semivariance variance measurement can produce the same cor-
model; section 3 presents the data and empirical rect results when return distributions are normal.
asset allocation using mean-semivariance frame- However, in situations where return distributions are
work, and discusses empirical results; and the last not normal, the downside risk measurement is more
section concludes this research. likely to produce a better solution.
59
Investment Management and Financial Innovations, Volume 8, Issue 3, 2011
1.2. Alternative risk measures. Roy (1952) was the days and a level of probability of loss exceeding the
first to discuss the downside risk measure in in- VaR limit. Such a short-term horizon is not suitable
vestment literature. He asserted that it is reasonable for asset allocations and portfolio optimization for
for investors to reduce the possibility of disaster as long-term investors. Moreover, the greater the time
much as possible, and perceived the downside devia- horizon is the less precise is the efficient VaR fron-
tion as the “safety-first” rule, which measures the tier (Campbell, Huisman, & Koedijk, 2001). Camp-
investment risk by the probability of investment bell et al. (2001) also shows that the optimal portfo-
value falling below certain target or disaster level. lios with VaR constraints are sensitive to the confi-
He adds a criterion to Markowitz’s efficient fron- dence level selected. In addition, VaR ignores ex-
tier which selects the efficient portfolio with the treme events below the specified quantile. Indeed,
lowest probability to fall short of a given target val- the mean-VaR optimization does not necessarily
ue. Specifically, given an expected return, r, and improve upon the portfolio optimization in a mean-
standard deviation, s, investors tend to choose the variance framework (Alexander and Baptista, 2002).
portfolio that has the lowest possibility of falling In fact, when switching from mean-variance frame-
below the disaster level, d. That is, they will try to work to mean-VaR framework, we may end up with
maximize the reward-to-variability ratio, (r-d)/s. more volatile portfolios. Many researchers also sug-
Roy’s main contribution í the concept that investors gest that VaR is not a coherent risk measure since
will prefer the principal of safety first when faced the subadditivity property is not satisfied. In other
with uncertainty í is instructive to the later evolu- words, if we combine several securities into a port-
tion of the downside risk measurement research. Fish- folio, the combined VaR at certain confidence level
burn (1977) uses a utility function model to incorpo- may not necessarily result in a lower VaR than the
rate downside risk based on risk aversion level and a
sum of the VaRs of individual securities. In fact
target return. Bawa (1978) extended Roy’s work
we may end up with a higher combined VaR than
from 1st order to a more generalized nth order “safety-
the sum of the VaRs. In addition, VaR is also
first” rule, and showed that the nth order “safety-
criticized for its incompleteness in risk measure-
first “rule is computationally feasible. Tse et al.
ment since it cannot provide any information
(1993) further discussed the “safety-first” rule in a
about the magnitude of losses once the VaR limit
dynamic structure.
is exceeded. The drawbacks of VaR can be miti-
In the literature of risk management, several risk gated by CVaR which is a conditional expectation
measures are proposed for measuring downside that gives the expected loss beyond the VaR. For
risks. The most prevalent risk measures debated in example, while VaR at 99% measures the maximum
the risk management literature are Value-at-Risk loss in 99% of cases, CVaR at 99% measures the
(VaR) and Conditional Value-at-Risk (CVaR). average loss in the 1% of the worst cases. That is,
CVaR is the conditional expectation of losses that CVaR measures not only the magnitude but also the
exceed the VaR level and is proposed for direct and likelihood of losses. However, for portfolio optimi-
asymmetric control over the distribution of residual zation, CVaR model requires either an assumption
errors, and for constraining one of its tail means not
about the return distribution or a substantial amount
to exceed some pre-specified value (see Jorion,
of return observations below the target return which
1997; Trinidade, Uryasev, Shapiro & Zrazhevsky,
could pose a significant problem of real world data
2007; Rachev, Stoyanov & Fabozzi, 2011). In
other words, both VaR and CVaR are downside limitations for empirical research. For example, for
risk measures where the objective is to maximize a sample of 100 real world observations, a CVaR at
expected return given a VaR or a CVaR, respec- 99% will be based on 1 observation only, namely,
tively. Campbell, Huisman, & Koedijk (2001) 1% (1 out of 100 observations). Although we can
incorporates VaR as a shortfall constraint into the address this limitation using simulation techniques
portfolio selection decision by maximizing ex- to generate larger samples, this is often a practical
pected return subject to the constraint that the disadvantage of this model.
expected maximum loss should meet the VaR 1.3. Semivariance as a downside risk measure-
limits set by the risk manager. They measure risk ment. The two downside risk measurements sug-
in terms of the VaR over and above the risk free
gested by Markowitz are of particular interest in
rate on the initial wealth.
finance: 1) the semivariance below the mean value;
However, both VaR and CVaR require an investor and 2) the semivariance below the target return.
to specify a probability level of cumulative losses Though he theoretically preferred semivariance, Mar-
(Wiener, 1998). The problem with VaR or CVaR is kowitz still insisted on using variance as the risk
that the investor has to specify a holding period, measure simply because of the difficulty in compu-
typically a very short-term horizon from 1 to 10 ting lower semivariance. The major difficulty lies in
60
Investment Management and Financial Innovations, Volume 8, Issue 3, 2011
gauging the co-movement, or correlation of Lower Foo & Eng (2000)’s downside risk optimization
Partial Moments (LPM), which is the most impor- model extended the former work of Harlow & Rao
tant provision for investment diversification. (1989) by incorporating the model with downside
covariance of correlated asset returns. But their
Complication in calculation has not prevented
work is still complicated and computationally bur-
scholars from pursuing the study and research on
dened. Hogan & Warren (1974) introduced the con-
the downside risk measurement. Based on the con-
cept of downside risk, Fishburn (1977) and Har- cept of co-lower-partial-variance, which measures
low and Rao (1989) introduced a generalized form of risky asset and market portfolio. Bawa & Linden-
lower partial moments (LPM) and developed the berg (1977) further developed this co-lower-partial-
“ (D t ) ” model, in which t represents the target variance measure to an n-degree framework called
generalized asymmetric co-LPM.
return of investment or disaster level as proposed by
Roy (1952), and D denotes the investor’s risk aver- Ballestero (2005) developed a semivariance matrix
sion. The higher the value of D , the greater is the using a strictly mathematical derivation while rely-
investor’s risk aversion. For risk-neutral investors, ing on the validity of Sharpe’s beta regression equa-
D = 1; for risk seeking investors, D < 1; and for risk tion. This approach has greatly eased the computa-
averse investors, D > 1. Fishburn (1977) also intro- tion complexity, allowing one to obtain the compu-
duced the Mean-Lower Partial Moment model tational results of the semivariance model from the
(MLPM-model). Note that when D = 2 and t equals traditional mean-variance model.
to the mean value of the investment return, and the
Since the 1990s, researchers have started to apply the
“ (D t ) ” model is specified as a LPM model. Har-
downside risk measures to their empirical research.
low (1991) employed LPM as a downside risk Sortino & Meer (1991) introduced the downside dev-
measure in portfolio selections. He defined LPM as: iation, i.e., below-target deviation, and the reward-
¦ Pp W R p , where Pp is the pro-
T n
LPM n Rp x
to-downside variability ratio as the tools for the
measurement of downside risk. Balzer (1994)
bability that return, Rp occurs. He explained that the
discussed the skewness existed in asset returns,
type of “moment,” n, specified in the LPM equation and the issues of applying downside variance.
captures an investor’s preferences. For n = 0, the Bookstaber & Clarke (1985) and Merriken (1994)
risk measure becomes a 0th-order moment (LPM0) tried consecutively to prove how the semivariance
which measures the probability of falling below the could be applied in the downside risk manage-
target rate. However, for n = 1, LPM1 becomes the ment of different hedging policies using stock
expected deviation of returns below the target. For n options and interest rate swaps.
= 2, LPM2 is analogous to variance, in that it is a
probability weighting of squared deviations. Thus, It is not difficult to see that the academic downside
LPM2 can be referred to as a target semivariance. risk research has emphasized in general asset classes,
Harlow (1991) further explained that many popular e.g., the optimization of individual stocks, and the
notions of risk are special cases of the generalized performance evaluation of various mutual funds.
LPMn measure. For example, with n = 0 and a target However, few researchers have applied this theory
rate = 0%, LPM0 is simply the probability of a loss. into other industries such as insurance industry
For n = 2 and a target rate = mean return, LPM2 which also needs the portfolio theory to manage
becomes the traditional semivariance. Overall, asset allocation efficiently. The insurance indus-
LPM1 (target shortfall) and LPM2 (target semiva- try, which has large amount of fund to invest,
riance) provide an intuitive set of risk definitions needs further discussion in the downside risk op-
that are more useful than traditional approaches timization theorem.
(Harlow, 1991).
2. Portfolio optimization using mean-
Although theoretically and intuitively sound, LPM semivariance framework
has its own limit, because it creates much more
complexity in computation than does the variance The underlying principle for semivariance model is
measurement. Because Harlow and Rao (1989) the same as the variance model, in that investors
failed to consider the correlation of individual are willing to bring downside risk as low as possible
asset returns, i.e., co-movement between individ- while keeping the rate of return above a certain level.
ual asset returns that fall below the target return, The definition of semivariance below the mean value
and because the co-movement is very important can be expressed by the following formula:
for risk diversification, their results can only be n ~ n
limited to those assets whose returns are perfectly SV () ¦[(¦ rjtZ j ¦ rjZ j )2 p(t )] , (1)
or highly correlated. j 1 j 1
61
Investment Management and Financial Innovations, Volume 8, Issue 3, 2011
~
where for each observation rj satisfying where
n
~ ~
n ~
¦r Z d ¦r Z
jt j
n
j j . (2)
T ¦T
j 1
j Z j. (9)
j 1 j 1
Based on (7), we can rewrite the equation of SV (>) as:
Similarly, semivariance above the mean value can
be expressed as: n
SV (!) ¦[(T (rM rM )¦ E jZ j )2 p(t )] . (10)
n ~ n
¦[(¦ r Z ¦ r Z )
j 1
SV (!) jt j j j
2
p(t )] (3)
j 1 j 1 When the level of diversification goes to infinity,
with we can prove that:
n ~ n ¦
lim SV ( ! ) ( E E Z Z ) SV ( ~r ! r ). (11) Ȃ Ȃ
¦r Z ! ¦r Z
j h j h
L ov
jt j j j , (4) j ,h
j 1 j 1
The definition of SV(>) and SV(<) implies that
where p (t ) denotes the probability. If we assign t n n
the same probability for all observations, then we
have p (t ) = 1/T.
SV () SV (!) V ¦[¦rjtZj ¦rjZj ]2 p(t) . (12)
t 1 j 1 j 1
To further derive the simplified formula for semiva- Hence, we can get the expression of SV(<) by sub-
riance model, we have to make an important as- tracting SV(>) from V:
sumption, which is the validity of Sharpe (1964) lim SV ( ) V lim SV ( ! ) , (13)
beta regression equation: Lof Lof
~ r0 H~j .
rj D j E j ~ (5) lim SV ()
Lov
¦[V
j ,h
jh E j E h SV (~
rȂ ! rȂ )]Z jZh . (14)
62
Investment Management and Financial Innovations, Volume 8, Issue 3, 2011
category of ETFs for the period from August 2002 to the trading day. This provides a broad benchmark to
December 2007. The rationale for selecting this period measure the return of a particular asset type and
is that this period does not have extreme downside enables us to approach the real market return for a
movements nor upside movements as we have wit- certain type of investments as a whole.
nessed in the period from 2008 to 2010.
3.1. Application of Markowitz’s mean-variance
The rationale for using ETFs rather than individual efficient frontier. The descriptive statistics of our
stocks and bonds is that ETF is an index fund that sample are reported in Table 1 and the optimal port-
attempts to track a basket of stocks, bonds or certain folio asset allocation results constructed out of the
indexes but can be traded like a stock on the market, Markowitz mean-variance efficient frontier model are
where it experiences price fluctuations throughout reported in Table 2.
Table 1. Summary statistics of the monthly returns on investment assets
Asset types N Mean return Median return Std. dev.
Government bond 53 0.333% 0.360% 1.792%
Muni bond 53 0.554% 1.066% 3.544%
High yield bond 53 0.662% 1.103% 5.845%
Real estate 53 1.979% 3.111% 4.437%
MBS 53 0.584% 1.027% 3.383%
Investment grade bond 53 0.510% 0.713% 1.816%
Large cap stocks 53 1.037% 1.252% 3.171%
Value-weighted market returns 53 1.049% 1.282% 3.251%
63
Investment Management and Financial Innovations, Volume 8, Issue 3, 2011
3.2. Application of mean-semivariance efficient derive the beta coefficients of various assets. We
frontier. In this section, we illustrate our computa- calculate the variance and covariance matrix of dif-
tion of the semivariance matrix. The first step is to ferent asset types. Results are shown in Table 3.
Table 3. Covariance matrix of assets
High yield Investment grade Large cap
Government bond Muni bond Real estate MBS
bond bond stock
Government bond 0.031% 0.040% 0.048% 0.037% 0.039% 0.026% -0.010%
Muni bond 0.040% 0.123% 0.074% 0.070% 0.063% 0.039% -0.010%
High yield bond 0.048% 0.074% 0.335% 0.107% 0.077% 0.047% 0.002%
Real estate 0.037% 0.070% 0.107% 0.193% 0.091% 0.026% 0.029%
MBS 0.039% 0.063% 0.077% 0.091% 0.112% 0.033% 0.022%
Investment grade bond 0.026% 0.039% 0.047% 0.026% 0.033% 0.032% -0.007%
Large cap stock -0.010% -0.010% 0.002% 0.029% 0.022% -0.007% 0.099%
Std. dev 1.775% 3.511% 5.789% 4.395% 3.351% 1.799% 3.141%
Skewness -0.782 -1.374 -0.001 -2.147 -1.133 -0.774 -0.521
E(M) 1.049%
Var(M) 0.104%
Cov(j,M) -0.012% -0.010% -0.003% 0.027% 0.020% -0.008% 0.101%
Beta -0.111 -0.098 -0.027 0.262 0.197 -0.072 0.972
V(>) 0.046%
T ~
¦max(RMt EM ,0)2
The second step is to compute the market mean and
variance values as follows:
VM (!) t 1 . (19)
1 T T
EM ¦vwretdt ,
Tt1
(15)
The results of equations (15)-(19) are shown in Table 3.
The third step is to calculate the market portfolio’s The result of the semivariance matrix is presented in
semivariance above the mean return: Table 4.
Table 4. Downside semivariance matrix V (<)
Government High yield Investment Large cap
Muni bond Real estate MBS
bond bond grade bond stock
Government bond 0.031% 0.039% 0.048% 0.038% 0.040% 0.026% -0.005%
Muni bond 0.039% 0.123% 0.073% 0.071% 0.064% 0.039% -0.006%
High yield bond 0.048% 0.073% 0.335% 0.107% 0.078% 0.047% 0.003%
Real estate 0.038% 0.071% 0.107% 0.190% 0.089% 0.027% 0.018%
MBS 0.040% 0.064% 0.078% 0.089% 0.111% 0.033% 0.014%
Investment grade bond 0.026% 0.039% 0.047% 0.027% 0.033% 0.032% -0.003%
Large cap stock -0.005% -0.006% 0.003% 0.018% 0.014% -0.003% 0.055%
Std. dev. of semivariance 1.759% 3.504% 5.789% 4.359% 3.324% 1.792% 2.349%
Std. dev of variance 1.775% 3.511% 5.789% 4.395% 3.351% 1.799% 3.141%
64
Investment Management and Financial Innovations, Volume 8, Issue 3, 2011
The optimal asset allocation results constructed of the variance investor. For example, when the
from the semivariance model are reported in Ta- risk free rate c is 0.7614%, a semivariance inves-
ble 5. Comparing the results of the traditional tor will have a portfolio return of 0.931% as com-
variance model (see Table 2) and the semiva- pared to a variance investor’s return of 0.931%;
riance model (see Table 5), we show that the in- while at the same time the semivariance investor
vestor can generate almost the same portfolio insurers will have a lower downside risk of
returns as the traditional variance investor while 1.891%, as compared to 1.910% for the variance
lowering the downside risk below the level of that investor.
Table 5. Results of the mean-semivariance efficient frontier
Portfolio
Portfolio Govern- High yeld Investment Large cap
c expected Muni bond Real estate MBS
semi-dev. ment bond bond grade bond stocks
returns
0.744 1.879 0.930 0.000 0.000 0.000 25.000 0.000 65.000 10.000
0.753 1.879 0.930 0.000 0.000 0.000 25.000 0.000 65.000 10.000
0.761 1.891 0.931 0.000 0.000 0.724 25.000 0.000 64.276 10.000
0.770 1.904 0.932 0.000 0.000 1.443 25.000 0.000 63.557 10.000
0.779 1.919 0.933 0.000 0.000 2.256 25.000 0.000 62.744 10.000
0.788 1.937 0.935 0.000 0.000 3.183 25.000 0.000 61.817 10.000
0.797 1.972 0.937 0.000 0.000 4.193 25.000 1.457 59.350 10.000
0.805 2.019 0.941 0.000 0.000 5.377 25.000 3.444 56.179 10.000
0.814 2.080 0.945 0.000 0.000 6.799 25.000 5.831 52.370 10.000
0.823 2.159 0.949 0.000 0.000 8.539 25.000 8.751 47.710 10.000
0.832 2.267 0.955 0.000 0.000 10.744 25.000 12.473 41.784 10.000
0.841 2.340 0.959 0.000 0.000 13.262 25.000 12.500 39.238 10.000
0.849 2.463 0.965 0.000 2.238 16.537 25.000 12.500 33.726 10.000
0.858 2.643 0.973 0.000 5.811 20.870 25.000 12.500 25.819 10.000
0.867 2.911 0.985 0.000 10.731 26.838 25.000 12.500 14.931 10.000
0.876 3.311 1.000 0.000 17.326 35.174 25.000 12.500 0.000 10.000
0.885 3.409 1.004 0.000 14.069 38.431 25.000 12.500 0.000 10.000
0.893 3.538 1.008 0.000 10.035 42.465 25.000 12.500 0.000 10.000
0.902 3.716 1.014 0.000 4.908 47.592 25.000 12.500 0.000 10.000
0.911 3.898 1.019 0.000 0.000 52.500 25.000 12.500 0.000 10.000
0.920 3.898 1.019 0.000 0.000 52.500 25.000 12.500 0.000 10.000
0.929 3.898 1.019 0.000 0.000 52.500 25.000 12.500 0.000 10.000
0.937 4.156 1.024 0.000 0.000 59.666 25.000 5.334 0.000 10.000
0.946 4.360 1.029 0.000 0.000 65.000 25.000 0.000 0.000 10.000
0.955 4.360 1.029 0.000 0.000 65.000 25.000 0.000 0.000 10.000
0.964 4.360 1.029 0.000 0.000 65.000 25.000 0.000 0.000 10.000
0.973 4.360 1.029 0.000 0.000 65.000 25.000 0.000 0.000 10.000
0.981 4.360 1.029 0.000 0.000 65.000 25.000 0.000 0.000 10.000
0.990 4.360 1.029 0.000 0.000 65.000 25.000 0.000 0.000 10.000
0.999 4.360 1.029 0.000 0.000 65.000 25.000 0.000 0.000 10.000
Figure 1 shows the graph of the two efficient indicate that the real differences between the va-
frontiers, in which the semivariance efficient riance-oriented investors and the semivariance-
frontier is moving outward to the left of the va- oriented investors are negligible. If the optimizing
riance efficient frontier. That is, the efficient fron- process is fed with more assets instead of just
tier based on mean-semivariance framework has seven ETF index funds, the ultimate difference
higher return-risk tradeoff than the efficient fron- between the two frontiers might be much more
tier based on mean-variance framework. Although obvious. Moreover, the inter-correlations between
the improvement from the mean-variance efficient asset classes are much higher than those of indi-
frontier to the mean-semivariance efficient fron- vidual securities, and the skewness may be re-
tier is not substantial, this does not necessarily duced through diversification.
65
Investment Management and Financial Innovations, Volume 8, Issue 3, 2011
In terms of asset weight allocation, we can observe The differences in the asset allocation strategy sug-
some similarities as well as differences between the gested by the two models may be partly explained by
Markowitz mean-variance model and the mean- each asset’s standard deviation and semi-deviation of
semivariance model. As risk-free rate goes higher, different asset returns. Municipal bonds and high
the real estate and common stock investments al- yield bonds both have relatively higher standard
ways stay at their highest investment constraints, deviation and standard semivariance than those of
25% and 10% respectively, while at the same time, MBS and investment grade corporate bonds. The
each model allocates a zero percentage to govern- difference between the standard semivariance and
ment bonds (see Table 2 and Table 5). For the mu- the standard deviation of the municipal bonds and
nicipal bond allocation, neither model allocates any high yield bonds are smaller than that between the
MBS and investment grade corporate bonds. So the
capital until risk-free rate goes higher. However, the
semivariance model suggests a smaller asset alloca-
semivariance model allocates slightly less than the
tion in these two assets than what the variance mod-
variance model when the risk-free rate is below
el suggests.
0.867%. When the risk-free rate is above 0.867%,
the semivariance model allocates more than the The relative difference in the asset allocation in the
variance model, but both models suggest dropping portfolio has practical economical implications. The
this asset from the portfolio after the risk-free rate mean-semivariance model allows investors to select a
goes above 0.911%. For the high yield bond alloca- portfolio that can achieve an expected portfolio return
tion, as the risk-free rate goes up, the optimal asset that matches and even exceeds what the mean-variance
allocation in this investment category goes up, but model can offer while keeping the risk level at the
the semivariance model always indicates to maintain minimum variance and the semivariance level. These
a lower investment ratio in this investment category results are consistent with Ballestero’s findings.
than does the variance model. For the Mortgage- Conclusion
Backed Securities (MBS), the investment alloca-
In general, Markowitz’s mean-variance model has
tions suggested by both models are all increasing
been one of the most commonly used methods in
initially, but dropping dramatically to a zero percen-
real-world portfolio management and asset alloca-
tage after the risk-free rate rises above a level of
tion. However, it has received criticisms for its strict
0.946%. Interestingly, for this investment catego- mathematical assumption that the returns of the
ry, the semivariance model suggests a much higher assets in the portfolio are normally distributed.
allocation to the MBS than does the variance model. When portfolio is composed of assets with skewed
For the investment grade corporate bond invest- returns, the results of the mean-variance model will
ment, things get a little twisted. The allocation be ineffective. Moreover, variance which is used as
suggested by the semivariance model starts out a risk measurement in the mean-variance framework
lower than that of the variance model, but the treats both the returns above and below the mean
allocation of the semivariance model soon ex- return equally. Thus, variance as a risk measurement
ceeds that of the variance model as the risk-free tends to give a misleading estimation for the down-
rate increases. In the end, both models allocate a side risk which investors weigh more heavily than
zero percentage to this asset category. the upside volatility of security returns.
66
Investment Management and Financial Innovations, Volume 8, Issue 3, 2011
Research on the downside risk, including the semi- correlation with each other than with other assets.
variance below the mean value or certain target By constructing a portfolio as suggested by the
value, has undergone for many years. The difficul- mean-semivariance model, investors can simulta-
ties in the calculation of the downside risk measure neously minimize the downside risk and achieve
have made the downside risk optimization not as similar or better expected returns. In contrast, the
popular as the mean-variance optimization algo- mean-variance model tends to lead to a portfolio that
rithm. However, the introduction of a strictly de- has limited upside gain and higher downside risk.
rived method has simplified the process: the sym-
Our findings of the semivariance model have prac-
metric co-semivariance matrix is derived from the
tical implications for both individual investors and
empirical validity of Sharpe’s beta regression equa-
institutional investors for asset allocation and port-
tion. Few prior studies have devoted to empirically
folio optimization while managing their downside
testing these two models and compare their efficient
risk exposure. It is especially important for insurance
frontiers. In this paper, we empirically test these
and banking sectors, the industries that have a higher
two models using real world data and compare the
risk aversion for downside risk. Insurance companies
two efficient frontiers in investment portfolio and
and commercial banks in the United States, for ex-
asset allocation.
ample, are required by regulators to maintain a cer-
The efficient frontiers comparison seems to indicate tain level of capital, which is determined by the level
that the mean-semivariance framework could pro- of the risk in their invested assets. While insurance
vide clearer indications in terms of asset allocations companies and banks seek to reduce the required
that lead to an optimal portfolio that not only capital to a minimum level, they are very concerned
matches if not exceeds those expected returns from about minimizing the downside risk while maintain-
the traditional mean-variance framework, but also ing a certain level of return on investment. The
lowers downside risk. In other words, the results mean-semivariance model could be very instrumen-
indicate that the mean-semivariance model could tal to these companies for their risk management.
provide investors with asset allocation strategy that Moreover, the mean-semivariance model allows
minimizes asset allocation not only in those assets portfolio managers to have a clear definition of
with higher variance but also in those assets with risk that combines the objectives and constraints
higher semivariance in particular and with higher of the entire investment portfolio.
References
1. Alexander, G. and A.M. Baptista (2002). Economic Implications of Using a Mean-VaR Model for Portfolio Selec-
tion: A Comparison with Mean-Variance Analysis, Journal of Economic Dynamics & Control, 26, pp. 1159-1193.
2. Ballestero, E. (2005). Mean-Semivariance Efficient Frontier: A Downside Risk Model for Portfolio Selection,
Applied Mathematical Finance, 12 (1), pp. 1-15.
3. Balzer, L.A. (1994). Measuring Investment Risk: A Review, Journal of Investing, 3 (3), pp. 47-58.
4. Bawa, V.S. (1978). Safety-First, Stochastic Dominance, and Optimal Portfolio Choice, The Journal of Financial
and Quantitative Analysis, 13 (2), pp. 255-271.
5. Bawa, V.S. & Lindenberg, E.B. (1977). Capital market equilibrium in a mean-lower partial moment framework,
Journal of Financial Economics, 5 (2), pp. 189-200.
6. Bookstaber, R. & Clarke, R. (1985). Problems in Evaluating the Performance of Portfolios with Options, Financial
Analysts Journal, 41 (1), pp. 48-62.
7. Campbell, R., Huisman, R., & Koedijk, K. (2001). Optimal portfolio selection in a Value-at-Risk framework,
Journal of Banking & Finance, 25, pp. 1789-1804.
8. Fama, E.F., Roll, R. (1968). Some properties of symmetric stable distributions, Journal of the American Statistical
Association, 63, pp. 817-846.
9. Jansen, D., de Vries, C. (1991). On the frequency of large stock returns: Putting booms and busts into perspective,
The Review of Economics and Statistics, 73, pp. 18-24.
10. Jorion, P. (1997). Value at Risk: The New Benchmark for Controlling Derivatives Risk, McGraw-Hill, New York.
11. Fishburn, P.C. (1977). Mean-Risk Analysis with Risk Associated with Below-Target Returns, The American Eco-
nomic Review, 67 (2), pp. 116-126.
12. Harlow (1991), Asset Allocation in a Downside-Rils Framework, Financial Analysts Journal, 40.
13. Harlow, W.V. & Rao, R.K.S. (1989). Asset Pricing in a Generalized Mean-Lower Partial Moment Framework:
Theory and Evidence, The Journal of Financial and Quantitative Analysis, 24 (3), pp. 285-311.
14. Hogan, W.W. & Warren, J.M. (1972). Computation of the Efficient Boundary in the ES Portfolio Selection Model,
The Journal of Financial and Quantitative Analysis, 7 (4), pp. 1881-1896.
15. Hogan, W.W. & Warren, J.M. (1974). Toward the Development of an Equilibrium Capital-Market Model Based
on Semivariance, The Journal of Financial and Quantitative Analysis, 9 (1), pp. 1-11.
16. Ibbotson, R.G. & Sinquefield, R.A. (1976). Stocks, Bonds, Bills, and Inflation: Year-by-Year Historical Returns
(1926-1974), The Journal of Business, 49 (1), pp. 11-47.
67
Investment Management and Financial Innovations, Volume 8, Issue 3, 2011
17. Markowitz, H. (1952). Portfolio Selection, The Journal of Finance, 7 (1), pp. 77-91.
18. Markowitz, H. (1987). Mean-variance Analysis in Portfolio Choice and Capital Markets. Basil Blackwell, New York.
19. Markowitz, H.M. (1959). Portfolio selection: efficient diversification of investments, John Wiley & Sons, New
York, p. 344.
20. Markowitz (1970). Portfolio Selection: Efficient Diversification of Investments, Yale University Press.
21. Merriken, H.E. (1994). Analytical Approaches to Limit Downside Risk: Semivariance and the Need for Liquidity,
Journal of Investing, pp. 65-62.
22. Michaud, R. (1989). The Markowitz Optimization Enigma: Is ‘Optimized’ Optimal?” Financial Analysts Journal,
45 (1) (January/February), pp. 31-42.
23. Rachev, S.T., Stoyanov, Stoyan V., Fabozzi, Frank J. (2011). A Probability Metrics Approach to Financial Risk
Measures. Chichester, West Sussex, U.K., Malden, MA: Wiley-Blackwell.
24. Roy, A.D. (1952). Safety First and the Holding of Assets, Econometrica, 20 (3), pp. 431-449.
25. Sharpe, W.F. (1964). Capital Asset Prices: A Theory of Market Equilibrium under Conditions of Risk, The Jour-
nal of Finance, 19 (3), pp. 425-442.
26. Sortino, F.A. & van der Meer, R. (1991). Downside Risk, Journal of Portfolio Management, 17 (4), 27-31.
27. Swisher, P. & Kasten, G.W. (2005). Post-Modern Portfolio Theory, Journal of Financial Planning, 18 (9), p. 74.
28. Foo, T.S. & Eng, S.O. (2000). Asset allocation in a downside risk framework, Journal of Real Estate Portfolio
Management, 6 (3), p. 213.
29. Trindade, A.A., Urasey, S., Shapiro, A. & Zrazhevsky, G. (2007). Financial prediction with constrained tail risk,
Journal of Baking & Finance, 31 (11), p. 3524.
30. Tse, K.S.M., Uppal, J. & White, M.A. (1993). Downside risk and investment choice, The Financial Review, 28
(4), p. 585.
31. Wiener, Z. (1998). Value-at-Risk (VaR), Mathematica in Education and Research, 7, pp. 39-45.
68