Professional Documents
Culture Documents
To cite this article: Wolfgang Bessler, Heiko Opfer & Dominik Wolff (2014): Multi-asset portfolio optimization
and out-of-sample performance: an evaluation of Black–Litterman, mean-variance, and naïve diversification
approaches, The European Journal of Finance, DOI: 10.1080/1351847X.2014.953699
Taylor & Francis makes every effort to ensure the accuracy of all the information (the “Content”)
contained in the publications on our platform. However, Taylor & Francis, our agents, and our
licensors make no representations or warranties whatsoever as to the accuracy, completeness, or
suitability for any purpose of the Content. Any opinions and views expressed in this publication
are the opinions and views of the authors, and are not the views of or endorsed by Taylor &
Francis. The accuracy of the Content should not be relied upon and should be independently
verified with primary sources of information. Taylor and Francis shall not be liable for any
losses, actions, claims, proceedings, demands, costs, expenses, damages, and other liabilities
whatsoever or howsoever caused arising directly or indirectly in connection with, in relation to or
arising out of the use of the Content.
This article may be used for research, teaching, and private study purposes. Any substantial
or systematic reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or
distribution in any form to anyone is expressly forbidden. Terms & Conditions of access and use
can be found at http://www.tandfonline.com/page/terms-and-conditions
The European Journal of Finance, 2014
http://dx.doi.org/10.1080/1351847X.2014.953699
The Black–Litterman model aims to enhance asset allocation decisions by overcoming the problems of
mean-variance portfolio optimization. We propose a sample-based version of the Black–Litterman model
Downloaded by [FU Berlin] at 01:39 15 May 2015
and implement it on a multi-asset portfolio consisting of global stocks, bonds, and commodity indices,
covering the period from January 1993 to December 2011. We test its out-of-sample performance relative
to other asset allocation models and find that Black–Litterman optimized portfolios significantly outper-
form naïve-diversified portfolios (1/N rule and strategic weights), and consistently perform better than
mean-variance, Bayes–Stein, and minimum-variance strategies in terms of out-of-sample Sharpe ratios,
even after controlling for different levels of risk aversion, investment constraints, and transaction costs.
The BL model generates portfolios with lower risk, less extreme asset allocations, and higher diversifica-
tion across asset classes. Sensitivity analyses indicate that these advantages are due to more stable mixed
return estimates that incorporate the reliability of return predictions, smaller estimation errors, and lower
turnover.
1. Introduction
The traditional mean-variance (MV) portfolio optimization (Markowitz 1952) has played a
prominent role in modern investment theory and has been widely discussed and tested in the
literature. In theory, it provides the investor with the optimal asset allocation if the portfolio
variance and return are the only relevant parameters and future asset returns and the covariance
matrix are given. However, when applied in an asset management setting, estimation errors in
input parameters (Jobson and Korkie 1981a; Michaud 1989), corner solutions (Broadie 1993),
and high transaction costs, resulting from extreme portfolio reallocations (Best and Grauer 1991),
often result in a poor out-of-sample portfolio performance.
Asset managers frequently try to cope with these shortcomings by implementing constraints
on the portfolio weights and turnover. In the literature, several variations and extensions of MV
are discussed that attempt to overcome its limitations. These suggestions range from imposing
portfolio constraints (Frost and Savarino 1988; Jagannathan and Ma 2003; Behr, Guettler, and
Miebs 2013) to the use of factor models (Chan, Karceski, and Lakonishok 1999), and Bayesian
methods for estimating the MV input parameters (Jorion 1985, 1986; Pastor 2000; Pastor and
Stambaugh 2000). In a recent study, DeMiguel, Garlappi, and Uppal (2009) analyze whether
MV strategies and various extensions suggested in the literature are able to outperform a naïve-
diversified 1/N portfolio across a range of different stock data sets. They report that none of
the different MV strategies consistently outperforms a naïve equally weighted portfolio (1/N) in
an out-of-sample setting. However, Kirby and Ostdiek (2012) question these results, suggesting
that the results are mainly due to their research design and their data sets. They show that by
focusing on the tangential portfolio, DeMiguel, Garlappi, and Uppal (2009) place the MV model
at an inherent disadvantage in terms of turnover and estimation risk. Nevertheless, Kirby and
Ostdiek (2012) also find that the MV approach hardly outperforms the 1/N strategy at statistically
significant levels when transaction costs are included.
In contrast to the professional asset management industry, the academic literature has paid lit-
tle attention to the performance of the Black and Litterman (1992) model, which was derived
as an alternative approach to portfolio optimization more than 20 years ago. Since then the
Black–Litterman (BL) model has experienced an increasing attention among quantitative port-
folio managers (Satchell and Scowcroft 2000; Jones, Lim, and Zangari 2007). In the literature,
Downloaded by [FU Berlin] at 01:39 15 May 2015
only a few authors analyze the rationale of the BL model, provide examples for applying the
methodology, or propose extensions of the model (Lee 2000; Satchell and Scowcroft 2000; Dro-
betz 2001; Herold 2005; Idzorek 2005; Meucci 2006; Chiarawongse et al. 2012; Bessler and
Wolff 2013). So far, there exists only little empirical evidence for the out-of-sample performance
of the BL model.
Our study contributes to the literature by empirically testing the BL model performance out-
of-sample. The main objectives are first to analyze whether the BL model is able to alleviate
the typical shortcomings of MV optimization, and second to evaluate its performance relative
to alternative asset allocation strategies in an out-of-sample setting with realistic investment
constraints. We propose a sample-based approach of the BL model that allows us to compute
its out-of-sample performance based on historic data sets. For the period from January 1993 to
December 2011, we compute the out-of-sample performance of BL, MV, and Bayes–Stein (BS),
as one of the most prominent extensions of MV, minimum variance (MinVar), as well as for two
naïve diversification approaches and evaluate the performance of each strategy. Furthermore, we
propose variations of the BL model by incorporating different reference portfolios and finally
analyze expansionary and recessionary sub-periods separately.
Relative to other studies on asset allocation strategies such as DeMiguel, Garlappi, and Uppal
(2009), Kirby and Ostdiek (2012), Behr, Guettler, and Miebs (2013), Murtazashvili and Voz-
lyublennaia (2013), our study extends the literature with respect to the following three aspects.
First, and most importantly, we include the BL model in the analysis. Second, we analyze all opti-
mization strategies for three different investor types with maximum desired portfolio volatilities
of 5%, 10% and 15%, while earlier studies usually focus only on the tangential portfolio. Third,
we implement portfolio optimization strategies based on multi-asset portfolios. Earlier studies
cited above focus solely on stock portfolios. Our empirical analysis covers the asset classes
developed and emerging stocks, government and corporate bonds, as well as commodities. In
particular, we expect that the additional asset classes, government bonds and commodities, offer
high diversification potentials, due to their historically low correlations with stocks. Given that
government bonds usually gain in value during stock market downturns, asset allocation models
that shift wealth from stocks to bonds during these periods should benefit relative to naïve strate-
gies such as the 1/N approach. Therefore, it is important to analyze how different asset allocation
strategies performed in a multi-asset context over the last two decades.
The European Journal of Finance 3
Our empirical results offer new insights into several dimensions. We find that for a multi-asset
universe all portfolio optimization strategies outperform relative to naïve equally weighted (1/N)
or strategically weighted portfolios. Yet the outperformance is insignificant for most MV and BS
portfolios, which is consistent with the results of Kirby and Ostdiek (2012) and DeMiguel, Gar-
lappi, and Uppal (2009). However, the BL portfolios exhibit significantly higher Sharpe ratios
than naïve strategies and consistently outperform MV. Additionally, BL portfolios have lower
portfolio turnover and offer more diversification across asset classes relative to MV portfolios.
Our sub-period analysis suggests that the BL model outperforms MV and naïve-diversified port-
folios particularly in recessionary periods. Sensitivity analyses indicate that the superiority of the
BL model results from more stable mixed return estimates that include the reliability of return
estimates (‘views’) and lead to a lower portfolio turnover. A variety of sensitivity analyses indi-
cates that our results are robust to different estimation windows, optimization constraints, and
to restrictions of the optimized portfolio weights. Our results also hold for different levels of
transaction costs and when transaction costs are included in the optimization function as well as
for larger and different sets of asset universes.
Downloaded by [FU Berlin] at 01:39 15 May 2015
The remainder of this study is organized as follows. In Section 2, we discuss the asset alloca-
tion models employed, the related literature as well as the employed performance measures. We
provide the data and some descriptive statistics in Section 3 and present and discuss our empiri-
cal results in Section 4. Section 5 includes a description and discussion of the robustness checks.
Section 6 concludes.
2. Methodology
We analyze the out-of-sample performance of different optimization techniques and naïve diver-
sification rules. The optimization models employed include three alternative implementations
of the Black–Litterman (BL) model, the mean-variance (MV) and minimum-variance (MinVar)
portfolios, as well as the Bayes–Stein (BS) approach as one of the most prominent extensions of
MV. The naïve diversification rules analyzed are equally weighted portfolios (1/N) and strate-
gically weighted portfolios (st.w.). Table 1 summarizes all asset allocation models used in this
study. For the period from January 1993 to December 2011, we calculate monthly optimized
portfolios at the first trading day of each month, based on the different optimization approaches.
In line with DeMiguel, Garlappi, and Uppal (2009) and Daskalaki and Skiadopoulos (2011),
we employ a rolling sample approach. This means that at any point in time t (month), we
use data available up to and including time t (k observations) to compute optimized portfolio
weights. These weights are then used to compute the out-of-sample realized portfolio return over
the period [t, t + 1]. We repeat this process by moving the sample period one month forward
and computing the optimized weights for the next month. Rolling estimation windows offer the
advantage of being more responsive to structural breaks than expanding estimation windows. To
ensure the robustness of our results, we analyze different estimation window lengths. All opti-
mization models include realistic investment constraints. First, we include a budget restriction
according to Equation (1) which ensures that portfolio weights sum to one:
N
ωi = 1, (1)
i=1
where ωi is the portfolio weight of asset i and N is the number of assets in the portfolio. Second,
we exclude short selling since many institutional investors are restricted to long positions only
4 W. Bessler et al.
Notes: This table lists the various asset allocation models that we include in this study. The last column
of the table gives the abbreviation used to refer to the strategy in the tables in which the performance of
Downloaded by [FU Berlin] at 01:39 15 May 2015
(Equation (2)):
ωi ≥ 0, with i = 1, . . . , N. (2)
Third, we limit the maximum portfolio volatility in order to distinguish between different
investor types in terms of their maximum desired portfolio risk (Equation (3)):
ω ω ≤ σ̂P , (3)
where ω is the vector of portfolio weights, is the covariance matrix of the asset returns, and
σ̂P is a predefined portfolio volatility constraint. The volatility constraint represents an upper
volatility bound, rather than a target volatility for the optimized portfolio. Such a volatility bound
provides the advantage that the asset allocation model may shift large fractions of wealth to low-
risk government bonds during stock market downturns, thereby preventing losses, whereas for
a specific target volatility (e.g. 10% p.a.) substantial fractions of wealth have to be invested in
stocks even if stock return estimates are negative. In the remainder of this section, we briefly
describe the methodology of the various asset allocation strategies implemented in this study.
Notes: This table provides the strategic weights for the three analyzed investor types: conservative, moderate, and aggres-
sive, which are used to compute naïve-diversified portfolios and implied return estimates. Within the asset class stocks,
emerging market stocks obtain a strategic weight of 25%, while developed market stocks obtain a strategic weight of
75%. Accordingly, in the asset class bonds, high-yield corporate bonds have a strategic weight of 25% and government
bonds 75%. We assume that the three investor types prefer a maximum expected portfolio volatility of 5%, 10% and
15%, respectively, which is compatible with the historic volatilities of the benchmark portfolios.
Downloaded by [FU Berlin] at 01:39 15 May 2015
determine strategic weights we rely on the results of earlier studies and on discussions with prac-
titioners. Analyzing multi-asset portfolios including US stocks, bonds, and commodities for the
period from 1974 to 1997, Anson (1999) suggests that a moderate investor should allocate about
15% to commodities. Usually the risk of commodities is similar to equities (Bodie and Rosansky
1980; Gorton and Rouwenhorst 2006) and substantially larger than for bonds. Therefore, we set
the strategic allocation to commodities and stocks lower for conservative investors and higher
for aggressive investors. For the conservative, moderate, and aggressive investor clienteles, the
strategic weights for commodities are 5%, 15%, and 25% and for stocks 15%, 40%, and 65%,
respectively. Accordingly, the strategic weights for bonds are 80% for the conservative investor,
which might be, for instance, reasonable for pension funds, 45% for the moderate investor, and
10% for the aggressive investor type. Table 2 summarizes the strategic weights for the different
investor types. The maximum desired portfolio volatilities are set to 5%, 10%, and 15% for the
conservative, moderate, and aggressive investor clienteles, respectively, which is in line with
the historic volatilities of the strategically weighted portfolios for these investor types before
the evaluation period from 1988 to 1992. We employ these volatility bounds as constraints for
the optimization strategies. As for the optimized portfolios, both naïve-diversified portfolios are
rebalanced at every first trading day of each month. Rebalancing, however, only maintains the
naïve 1/N or strategic weight for each asset. As robustness check, we also compute equally and
strategically weighted buy-and-hold benchmark portfolios that do not include any rebalancing.
Our findings also hold for the buy-and-hold benchmark portfolios, and the performance differs
only marginally between the portfolios with and without rebalancing. To be parsimonious, we
only report the results with rebalancing.2
does not require any return estimates, which are usually subject to large estimation errors. The
effects of estimation errors in the covariance matrix are about 10 times smaller than the effects
of estimation errors in returns (Chopra and Ziemba 1993). In line with all other optimization
strategies, we include realistic investment constraints, specifically a budget restriction according
to Equation (1) and disallow short selling according to Equation (2).
returns or mean-variance preferences, focusing only on the mean and variance of returns and
ignoring higher moments. Although this seems to be a critical assumption, Landsman and Nešle-
hová (2008) show that it is sufficient that returns are elliptically symmetrically distributed so
that all investor preferences are equivalent to mean-variance preferences. Therefore, it is rea-
sonable to rely on the mean-variance framework for portfolio optimization even if asset returns
are non-normal, but are symmetric. To implement the mean-variance strategy, we employ the
sample mean μ̂ and the sample covariance matrix ˆ as described above. We include a budget
restriction according to Equation (1), disallow short selling according to Equation (2), and limit
the maximum allowed portfolio volatility (Equation (3)). The risk aversion coefficient is initially
set to 2 and variations are analyzed as robustness check.
The BS approach estimates the covariance matrix as = ((K − 1)/(K − N − 2)), ˆ where
ˆ
μ̂ is the expected return of the minimum-variance portfolio, is the usual unbiased sample
min
covariance matrix, N is the number of assets included, and K is the sample size (Jorion 1986). In
The European Journal of Finance 7
addition to shrinking the return estimates, the Bayes–Stein approach also adjusts the covariance
matrix that we implement as described in Jorion (1986). We employ the same optimization pro-
cedure, optimization constraints, and estimation windows as for the MV approach and analyze
variations of all input parameters as a robustness check.
be the market or any other benchmark portfolio, if they are able to provide reliable estimates of
future returns. The BL model uses implied returns as a prior that we compute based on the asset
weights of the reference portfolio. The implied excess returns used in the original Black and
Litterman (1992) approach are derived using reverse optimization, assuming that the observable
market or benchmark weights ω∗ are the result of a mean-variance optimization. Implied returns
used in the BL model are computed as
=δ ω∗ , (8)
e
where e is the vector of implied asset excess returns, is the covariance matrix, δ is the
investor’s risk aversion coefficient, and ω∗ is the vector of observable market or benchmark
weights.
The BL framework combines the vector of implied returns with the investor’s ‘views’
expressed in the vector Q, incorporating the reliability of each ‘view’ quantified in a matrix . To
derive the combined return estimates, the original Black and Litterman (1992) paper references
Theil’s mixed estimation model (1971). Figure 1 illustrates the procedure of the BL approach.
The combined return estimates are written as
in which P is a binary matrix containing the information for which asset a subjective return
estimate exists, and τ is a factor that measures the reliability of implied return estimates. The
combined return estimate is a matrix-weighted average of implied returns and ‘views’ with
respect to the correlation structure (Lee 2000). The weighting factors are the uncertainty mea-
sures of implied returns τ and subjective return estimates , which we discuss in the following
section.
In the BL approach, the posterior covariance matrix is (Satchell and Scowcroft 2000)
= +[(τ )−1 + P −1 P]−1 . (10)
BL
After computing combined return estimates and the posterior covariance matrix, a traditional
risk-return optimization is conducted maximizing the investor’s utility function presented in
Equation (5).
8 W. Bessler et al.
Downloaded by [FU Berlin] at 01:39 15 May 2015
estimates (He and Litterman 1999; Lee 2000; Drobetz 2001; Idzorek 2005) and suggest confi-
dence intervals of the return estimates as a measure of uncertainty (Black and Litterman 1992).
Since the portfolio performance critically depends on the exogenously assumed estimates, these
approaches are hardly capable to evaluate the performance of the BL model in comparison
to MV and naïve-diversified benchmark portfolios. Moreover, mutually estimating returns and
their respective confidence intervals might be a challenging task for analysts and might hinder
a successful implementation of the BL model. Therefore, we employ sample means as subjec-
tive return estimates based on rolling estimation windows as in the other optimization models
(MV and BS). However, in contrast to MV and BS, the BL model additionally incorporates the
reliability of these estimates that are expressed in the matrix .
For a portfolio of N assets, is a N × N diagonal matrix with the reliability measures for
the assets on its diagonal. A low reliability in all ‘views’ results in optimized portfolio weights
close to the reference portfolio. Meucci (2010) proposes to assume simply an overall level of
confidence in the ‘views’ that is constant over time by setting as
Downloaded by [FU Berlin] at 01:39 15 May 2015
1
= P P . (11)
c
In this approach, it is assumed that the reliability of ‘views’ is proportional to the reliability
of implied returns with the proportionality factor 1/c. As a result no additional and time-varying
information on the reliability of ‘views’ is included. We suggest that the out-of-sample perfor-
mance of the BL model is superior if reasonable and time-varying information on the reliability
of ‘views’ is included. Therefore, we measure the reliability of each ‘view’ i based on historic
forecast errors εi . For this, we employ the same rolling estimation windows as for return esti-
mates and estimate the uncertainty of return estimates as the variance of historic forecast errors.
We analyze different estimation window lengths as robustness check. The idea is that in uncer-
tain market environments when the last months return estimates differ strongly from the realized
returns, the reliability for the subsequent return estimate is likely to be lower, resulting in portfo-
lios closer to the reference portfolio. In contrast, in stable market conditions when the last months
return forecasts are close to the realized returns, the return estimate for the next period should be
more reliable and the optimized portfolio may depart more from the reference portfolio.
We analyze the contribution of this historic reliability measure by comparing the out-of-sample
portfolio performance of our approach with the assumption of an overall time-invariant level of
confidence in the ‘views’ as in Meucci (2010) by substituting according to Equation (11).
R̄Net,i − R̄f
ŜRi = . (12)
σ̂Net,i
Using Sharpe ratios as performance measure is often criticized, because asset return distri-
butions are usually non-normal. However, Meyer (1987) shows that the general LS condition
10 W. Bessler et al.
(location and scale) of returns is sufficient to rely on the μ-σ framework. Therefore, relying
on Sharpe ratios is reasonable even for asymmetric return distributions and fat tails as long as
the LS condition holds. We use the two-sample statistic to test if the difference in Sharpe ratios
of two portfolios is significant (Opdyke 2007). In contrast to earlier Sharpe ratio tests (Jobson
and Korkie 1981b; Lo 2002), this test is appropriate under very general conditions – stationary
and ergodic returns. Most important for our analysis, the test permits autocorrelated and non-
normal distributed returns and allows for a likely high correlation between the portfolio returns
of different strategies.3
As alternative performance measure, we compute the Omega ratio (Shadwick and Keating
2002) also referred to as gain-loss ratio. It measures the average gains to average losses in which
we define gains as returns above the risk-free rate and losses as returns below the risk-free rate.
Hence, investments with a larger Omega measure are preferable. Formally, the omega measure is
(1/T) Tt=1 max(0, rt,i − rf )
Omegai = , (13)
(1/T) Tt=1 max(0, rf − rt,i )
Downloaded by [FU Berlin] at 01:39 15 May 2015
where ri,t is the return of asset i at time t and rf is the risk-free rate. The advantage of the Omega
measure is that it does not require any assumption on the distribution of returns.
As alternative risk measures besides volatility, we compute the maximum drawdown (MDD)
as proposed by Grossman and Zhou (1993). The MDD reflects the maximum accumulated loss
that an investor may suffer during the entire investment period if she buys the portfolio at a high
price and subsequently sells at the lowest price. As the Omega measure, it does not require any
assumption on the return distribution. We compute the percentage maximum drawdown (MDD)
of strategy i as
Pi,t − Pi,t∗
MDDi = Maxi,t∗∈(0,T) Maxi,t∈(0,t∗) , (14)
Pi,t
where Pi,t is the price of portfolio i at time t, when the portfolio is bought, and Pi,t∗ is the price
of portfolio i at time t*, when the portfolio is sold.
Furthermore, we compute the portfolio turnover, in line with Daskalaki and Skiadopoulos
(2011) and DeMiguel, Garlappi, and Uppal (2009), which quantifies the extent of trading required
to implement a certain strategy. The portfolio turnover PTi of strategy i is the average absolute
change of the portfolio weights ω over the T rebalancing points in time and across the N assets:
1
T N
PTi = (|ωi,j,t+1 − ωi,j,t+ |) (15)
T t=1 j=1
in which ωi,j,t is the weight of asset j at time t under strategy i; ωi,j,t+ is the portfolio weight before
rebalancing at t + 1; and ωi,j,t+1 is the desired portfolio weight at t + 1, after rebalancing. ωi,j,t
is usually different from ωi,j,t+ due to changes in asset prices between t and t + 1.
We account for trading costs by assuming proportional transaction costs of 30 basis points of
the transaction volume in the base case and compute all performance measures after transaction
costs. As a robustness check, we analyze the impact of different transaction costs.
3. Data
To construct multi-asset portfolios we include global stocks, bonds, and commodity indices in
the investment universe. We use the MSCI World and MSCI Emerging Markets stock indices
The European Journal of Finance 11
to cover both developed and emerging markets. Emerging markets usually provide higher stock
returns than developed markets because they have a higher exposure to additional risk factors
such as illiquidity or institutional and political conditions (Iqbal, Brooks, and Galagedera 2010).
Chiou, Lee, and Chang (2009) find that international diversification is beneficial for US investors,
reducing portfolio volatility, and improving risk-adjusted returns.
Government bond returns typically have low or even negative correlations with stock returns
and often prices of government bonds increase during stock market downturns. To ensure their
role as a low-risk investment we employ US government bonds, thereby eliminating default and
currency risks. To represent US government bond investments we rely on the Bank of Amer-
ica/Merill Lynch US Government Bond Index (all maturities). In addition, we include the Bank
of America/Merill Lynch US High Yield 100 Bond Index to add exposure to corporate default
risk. This index is expected to provide higher returns and higher risks compared to government
bonds, but lower returns and lower risks compared to stock indices.
The S&P GSCI Light Energy Index represents a diversified commodity investment, providing
investors with an exposure to a wide range of commodity price changes. While the main deter-
Downloaded by [FU Berlin] at 01:39 15 May 2015
minants of the S&P GSCI Index are energy prices, the S&P GSCI Light Energy Index offers
more balance across different commodity classes. At the end of our sample period, it reflects
the price developments on the future markets for energy (37.4%), agricultural products (31.2%),
and livestock (10%), as well as industry (14%) and precious metals (7.4%). Commodities should
have low correlations with the traditional asset classes such as stocks and bonds, because their
prices depend on different risk factors such as weather, geographical conditions, and supply
constraints. Moreover, as several studies document a positive correlation between commod-
ity returns and future inflation, investing in commodities is often viewed as a hedge against
inflation (Bodie and Rosansky 1980; Erb and Harvey 2006; Gorton and Rouwenhorst 2006).
Several studies also find that by including commodities, the efficient frontier of stock-bond port-
folios improved (for instance, Satyanarayan and Varangis 1996; Abanomey and Mathur 1999;
Anson 1999; Jensen, Johnson, and Mercer 2000). More recent evidence suggests that diversifi-
cation benefits of commodities are regime-dependent (Cheung and Miu 2010), but that portfolio
benefits of commodities in out-of-sample optimized portfolios are ambiguous (Daskalaki and
Skiadopoulos 2011; You and Daigler 2013). However, given that investors still perceive com-
modities as an important asset class in portfolio optimization, we include commodities in our
analysis.
We obtain monthly total return index data for the period from January 1988 to December
2011 from Thomson Reuters Datastream. All data are denominated in US dollar. In line with
DeMiguel, Garlappi, and Uppal (2009), we use the three months US T-Bill rate as the risk-free
rate. Because we need several years of historical data to set up the first optimized portfolio the
evaluation period ranges from January 1993 to December 2011. Table 3 provides descriptive
statistics of the monthly asset returns for the full evaluation period.
The table shows similar annualized mean returns for stock and bond indices ranging from
6.16% to 8.00% p.a. For the entire period, the average return of the commodity index is slightly
lower than the average risk-free rate of 3.12%, resulting in a negative Sharpe ratio. The US
Government Bond Index generates the highest Sharpe ratio of 0.655. The maximum drawdowns
(MDD) of the assets reveal that the maximum loss an investor could have suffered during the
evaluation period by investing in stocks was between 55.16% and 63.04% of the invested cap-
ital. This figure was 59.95% for commodities, 27.21% for the US High Yield Bond Index, and
5.29% for the US Government Bond Index. The Jarque–Bera statistic is significant for all asset
12 W. Bessler et al.
Notes: This table provides sample moments, Sharpe ratios, maximum drawdown, and Jarque–Bera statistics
of the five asset classes considered in the empirical analysis. The period covers the months from January 1993
to December 2011. ‘Mean return p.a.’ denotes annualized time-series mean of monthly returns, while ‘SD
p.a.’ denotes the associated annualized standard deviation. ‘Skewness’ and ‘Kurtosis’ represent the third and
Downloaded by [FU Berlin] at 01:39 15 May 2015
fourth moment of the return distribution. ‘Sharpe ratio’ shows the annualized Sharpe ratios of the respective
asset classes using the average 1993–2011 risk-free interest rate of 3.12% per year. MDD shows the maximum
drawdown of the respective asset class during the period from January 1993 to December 2011 and ‘JB’ is the
Jarque–Bera statistic for testing normality of returns.
*Statistically significant at the 1% level.
classes. Hence, the assumption of normal distributed returns has to be rejected. However, as men-
tioned above, Meyer (1987) shows that the general LS condition (location and scale) of returns
is sufficient to apply the mean-variance framework.
Table 4 presents evidence on potential diversification effects in terms of pair-wise correlation
coefficients. Over the entire period the diversification benefits when investing only in stocks
are limited. The correlation between the MSCI World and MSCI Emerging Markets is highly
significant and larger than 0.8, indicating a strong co-movement of developed and emerging
stock markets. While the US-High-Yield-Bond Index and commodities offer a slightly better
diversification effect with correlation coefficients ranging between 0.35 and 0.65, the highest
diversification potential during our sample period is provided by investing in the US Government
Bond Index, which is reflected in negative correlation coefficients and positive Sharpe ratios.
Consequently, we expect to find significant portfolio benefits by applying the BL, BS, and MV
frameworks on a multi-asset portfolio, including bonds and commodities, rather than on a stock-
only portfolio.
4. Empirical results
4.1 Results for the full sample
Table 5 summarizes the out-of-sample performance of the different asset allocation strategies for
the three investor types ‘conservative’, ‘moderate’, and ‘aggressive’ for the base case over the
full evaluation period from January 1993 to December 2011. In the base case, we use 36 months
rolling estimation windows for the covariance matrix and 12 months for return estimates. We
use shorter estimation windows for return estimates as we expect the correlation structure and
variances to be more stable. In robustness checks, we provide the results for different estimation
windows (5.2). The strategically weighted portfolio (st.w.) is computed according to the asset
weights in Table 2. Within the asset class ‘stocks’, the strategic weights are specified to be 25%
for emerging markets and 75% for developed markets. Accordingly, for the asset class ‘bonds’,
The European Journal of Finance 13
MSCI World 1
MSCI Emerging Markets 0.811* 1
US Gov. Bond Index − 0.182* − 0.221* 1
US High Yield Bond Index 0.652* 0.619* − 0.069 1
S&P GSCI Light Energy 0.470* 0.463* − 0.128** 0.353* 1
Notes: This table provides the correlation matrix for the asset classes considered in the analysis over the period from
January 1993 to December 2011.
*Statistically significant at the 1% level.
**Statistically significant at the 5% level.
the strategic weights are set to 75% for government bonds and to 25% for high-yield corpo-
Downloaded by [FU Berlin] at 01:39 15 May 2015
rate bonds. In the equal-weighted (1/N) portfolio, all five asset classes obtain an equal portfolio
weight of 20%.
The performance evaluation in Table 5 reveals that the three BL approaches, with different
reference portfolios, yield larger Sharpe ratios and Omega measures than MV, BS, MinVar, and
both naïve-diversified portfolios for all investor types. The Sharpe ratios of the BL portfolios are
significantly larger (5% level) than that of the naïve-diversified 1/N strategy. In contrast, the MV
strategy does not result in a significant outperformance relative to the 1/N strategy.
Our results suggest that the impact of the reference portfolio on the performance of the BL
approach is rather low. For all analyzed reference portfolios, the BL optimization leads to a
consistently superior performance relative to MV, BS, and both naïve-diversified portfolios. This
result holds for all investor types. We also observe that for the conservative investor type, the BL-
MinVar approach performs slightly better than the other BL approaches. For aggressive investors,
however, the BL model based on strategic weights as reference portfolio performs marginally
better.
Our risk measures volatility and maximum drawdown (MDD) indicate a consistently lower
risk for all BL optimized portfolios in comparison to MV and BS, independent of the investor
type. At first, it seems surprising that the ex post realized volatilities differ from the ex ante
determined volatility constraints of 5%, 10%, and 15%, but the explanation is rather straightfor-
ward. On the one hand, the sample volatility estimates include estimation errors that may lead to
a discrepancy between ex ante estimated volatility and ex post realized volatility. On the other
hand, the optimization framework with volatility constraint as described in Equation (5) does
not necessarily favor the asset allocation with the largest possible volatility that is close to the
volatility constraint. The optimized portfolio is rather determined by the trade-off between risk
and return. For instance, in recessionary periods with negative expected stock returns, the opti-
mization framework is likely to allocate a large fraction of the portfolio to low-risk government
bonds, independent of the volatility constraint. Therefore, the average volatility over the entire
evaluation period is below the volatility constraint for all optimization models.
Table 5 also presents the average yearly portfolio turnover as an indicator for the magnitude
of trading volume and transaction costs required to implement a certain strategy. However, all
performance measures already include transaction costs. The results reveal that for all investor
types the BL approach requires a lower portfolio turnover and, therefore, has lower transaction
costs relative to the MV approach. ‘The BL-st.w.’ approach, for instance, yields an average yearly
14
Table 5. Empirical results for the full sample (evaluation period 1993–2011).
W. Bessler et al.
BL-st.w. BL-1/N BL-MinVar BS MV Min Var st.w. 1/N st.w. B&H 1/N B&H
Net Omega measure 1.96 1.93 1.98 1.78 1.69 1.70 1.66 1.24 1.67 1.25
Net MDD (%) 5.99 6.07 6.17 7.30 9.41 7.45 13.99 40.59 14.19 40.50
VaR99% (%) 2.77 3.04 2.87 3.33 3.92 2.74 3.16 9.72 9.70 12.33
Avrg. number of assets 3.32 3.79 3.20 2.91 2.69 3.71 5.00 5.00 5.00 5.00
Avrg. turnover p.a. 2.13 2.06 2.11 1.96 2.67 0.77 0.18 0.29 0.00 0.00
Investor type: moderate
Net mean return p.a. (%) 9.58 9.51 9.28 8.27 8.21 – 6.03 6.08 6.39 6.25
Volatility p.a. (%) 8.65 8.87 8.05 9.40 10.37 – 9.16 11.08 9.52 11.08
Net Sharpe ratio 0.75∗∗##†† 0.72†† 0.76†† 0.55 0.49 – 0.32 0.27 0.34 0.28
Net Omega measure 1.77 1.73 1.81 1.48 1.46 – 1.27 1.24 1.31 1.25
Net MDD (%) 9.46 11.02 9.21 14.35 16.18 – 35.10 40.59 35.30 40.50
VaR99% (%) 6.36 6.91 5.52 6.87 7.53 – 7.65 9.72 − 9.70 − 12.33
Avrg. number of assets 3.26 3.48 2.86 2.33 2.16 – 5.00 5.00 5.00 5.00
Avrg. turnover p.a. 3.32 3.29 3.28 3.34 4.44 – 0.27 0.29 0.00 0.00
Investor type: aggressive
Net mean return p.a. (%) 11.72##†† 11.32##†† 10.37##†† 9.46 10.35 – 5.81 6.08 6.27 6.25
Volatility p.a. (%) 11.68 11.30 9.85 12.38 13.62 – 14.27 11.08 14.74 11.08
Net Sharpe ratio 0.74#†† 0.73##†† 0.74##†† 0.51 0.53 – 0.19 0.27 0.21 0.28
Net Omega measure 1.79 1.78 1.82 1.51 1.53 – 1.15 1.24 1.18 1.25
Net MDD (%) 14.39 11.67 12.10 22.23 24.59 – 50.99 40.59 52.04 40.50
VaR99% (%) 9.43 7.86 6.73 10.26 10.77 – 12.38 9.72 − 9.70 12.33
Avrg. number of assets 3.12 3.21 2.63 1.86 1.72 – 5.00 5.00 5.00 5.00
Avrg. turnover p.a. 3.62 3.83 3.89 4.76 4.61 – 0.25 0.29 0.00 0.00
Notes: This table reports the portfolio performance measures for the full sample from 1993 to 2011 in the base case. In all optimized portfolios, the maximum expected
volatility is constrained to 5%, 10%, and 15% p.a. for the conservative, moderate, and aggressive investor type, respectively. All portfolios are rebalanced at the first trading
day of every month. */**, # /## , † /†† indicate a significantly higher Sharpe ratio compared to mean-variance, the strategically weighted benchmark, and the 1/N benchmark at
the significance levels of 1% and 5%, respectively.
The European Journal of Finance 15
turnover equal to 3.32 times the portfolio volume for the moderate investor, while the yearly
turnover for MV equals 4.44 times the portfolio volume. The explanation for this result is that
the combined return estimates used in the BL approach are more stable over time than the sample
means used in MV, which is indicated by a lower variance of return estimates (unreported).
Additionally, we find that the average number of assets in the optimized portfolio, as an indi-
cator for the magnitude of diversification across asset classes, is higher for BL than for MV and
BS portfolios. Consequently, BL portfolios offer more diversification across asset classes and
have less extreme allocations. A possible explanation is that the combined return estimates used
in the BL approach are less heterogeneous. Figure 2 presents the optimized portfolio weights for
BL and MV optimization for the three investor types during the 1993–2011 period. In line with
the turnover and diversification measures, the figure reveals less extreme portfolio reallocations
and a stronger benchmark orientation of the BL optimized portfolios relative to MV.
Our findings for the multi-asset data set suggest a slight outperformance of the sample-based
MV approach relative to naïve-diversified strategies. In line with Kirby and Ostdiek (2012), the
level of outperformance (after transaction costs) of MV is insignificant. However, our findings
Downloaded by [FU Berlin] at 01:39 15 May 2015
differ from that of DeMiguel, Garlappi, and Uppal (2009) and Murtazashvili and Vozlyublennaia
(2013) who conclude that none of the variations of MV is able to outperform a naïve 1/N
strategy. One explanation is the difference in the employed data set. While both earlier stud-
ies analyze stock-only portfolios, we additionally include government bonds, corporate bonds,
and commodities that should result in broader diversification and might enhance the portfolio
optimization benefits. Particularly, by including government bonds, asset allocation models may
outperform naïve strategies if they actively shift wealth from stocks to bonds during stock mar-
ket downturns and vice versa. In fact, Figure 2 shows that for both, BL and MV, and across all
investor types, the optimized portfolios are shifted completely to government bonds during the
stock market downturns between 2000 and 2003 (end of the new economy period) and between
2008 and 2009 (the financial crisis).
However, while the multi-asset approach might explain why MV performs superior in our
analysis relative to the DeMiguel, Garlappi, and Uppal (2009) study, additional analyses are
required to explain the reasons for the performance differences between BL, BS, and MV
approaches. To obtain additional insights, we analyze the return estimates used in the opti-
mization approaches. We find that for all assets the forecast error of estimated to subsequently
realized returns, measured as mean absolute error (MAE), is consistently lower for mixed return
estimates employed in the BL approach than for return estimates employed in the BS and MV
approaches. Additionally, we compute the coefficients of determination (R2 ) as the squared cor-
relation between forecasted and subsequently realized returns and use it as an indicator for the
ex post performance. In line with the results for the MAE, we find that the coefficients of deter-
mination of forecasted returns to subsequently realized returns are larger for BL mixed return
estimates than for BS shrinked return estimates and for sample mean returns employed in MV.
The analyses of forecast errors are available from the authors upon request.
monetary cycle is the first change of the short-term interest rate by the central bank that runs
counter to the previous trend (Jensen and Mercer 2003). The stock market signal is determined
as the intersection of the 24-months moving average of the MSCI World with the actual index
from either below (expansionary state) or above (recessionary state), indicating a change in the
business cycle. For the transition from one state to another, it is required that both instruments,
the monetary policy and the stock market, provide a consistent signal. In Figure 3, we present
the definition of sub-periods derived from the joint monetary policy and stock market signals,
where shaded areas denote recessionary periods.
The first sub-period ranges from January 1993 to January 2001 and includes a number of
events such as the Asian crisis, the Russian default, and the buildup of the new economy bubble.
The European Journal of Finance 17
Downloaded by [FU Berlin] at 01:39 15 May 2015
This period consists of 97 months and can be characterized as ‘expanding’ with increasing values
in developed stock markets and relatively high interest rates with T-Bills yielding, on average,
4.77% p.a. The second sub-period between February 2001 and June 2004 covers the end of the
new economy period and the subsequent rebound of international stock markets. It includes 41
months and is characterized as ‘recessionary’ with bearish stock markets and an average risk-
free rate of 1.85% p.a. The third sub-period ranges from July 2004 to February 2008 covering 44
months, and containing bullish stock markets and high interest rates. The average risk-free rate
in this period is 3.67% p.a. The final sub-period from March 2008 to December 2011 includes
46 months and incorporates the financial crisis that led to substantial declines in the values of
equities and alternative asset classes, such as commodities and hedge funds. The average risk-
free rate in the fourth period was only 0.37% p.a.
The performance of the out-of-sample optimized portfolios for the four sub-periods is sum-
marized in Table 6 for the moderate investor.4 We find that BL optimized portfolios outperform
BS and MV in all four sub-periods and perform superior than naïve-diversified portfolios in
three of four sub-periods. For both recessionary sub-periods, we observe substantially higher
Sharpe ratios for the BL strategies relative to MV. Naturally, due to the shorter time series of sub-
periods, the power of the significance test is lower relative to the full period so that differences
in Sharpe ratios are not significant at the 5% level anymore. In comparison to Bayes–Stein and
both naïve-diversified portfolios, all BL approaches achieve a substantially higher performance
in both recessionary periods.
In contrast, for both expansionary periods we find a relatively smaller outperformance of BL
in comparison to MV and BS. In the third sub-sample, which covers the bullish stock mar-
kets between July 2004 and February 2008, the naïve-diversified portfolios outperform all other
optimization strategies (BL, BS, MV, and MinVar).
For all sub-periods, the BL optimized portfolios are less risky than BS, MV, and both naïve-
diversified portfolios. This is indicated by a lower maximum drawdown and is consistent with
18 W. Bessler et al.
Net MDD (%) 6.52 7.08 4.88 9.43 10.67 3.40 14.19 17.37
Avrg. number of assets 2.76 3.10 2.32 2.10 1.90 4.56 5.00 5.00
Avrg. turnover p.a. 2.67 3.12 1.95 4.03 4.42 0.62 0.30 0.31
Avrg. risk-free rate 1.85% 1.85% 1.85% 1.85% 1.85% 1.85% 1.85% 1.85%
Obs. 41 41 41 41 41 41 41 41
July 2004–February 2008
Net mean return p.a. (%) 13.81 14.05 13.36 10.71 11.97 6.60 10.51 12.44
Volatility p.a. (%) 9.41 9.54 9.38 9.42 11.35 2.80 5.08 6.66
Net Sharpe ratio 1.08 1.09 1.03 0.75 0.73 1.05 1.35 1.32
Net MDD (%) 6.63 6.69 7.25 8.98 10.48 6.65 34.14 39.57
Avrg. number of assets 3.84 3.84 3.59 2.66 2.59 3.95 5.00 5.00
Avrg. turnover p.a. 3.45 3.24 4.08 4.24 4.20 0.97 0.22 0.25
Avrg. risk-free rate 3.67% 3.67% 3.67% 3.67% 3.67% 3.67% 3.67% 3.67%
Obs. 44 44 44 44 44 44 44 44
March 2008–December 2011
Net mean return p.a. (%) 5.84 6.21 5.06 3.17 3.00 4.31 0.14 − 0.43
Volatility p.a. (%) 8.63 8.73 8.28 8.79 9.75 5.08 15.04 18.10
Net Sharpe ratio 0.63 0.67 0.57 0.32 0.27 0.78 − 0.01 − 0.04
Net MDD (%) 9.34 11.02 6.90 14.35 16.18 7.45 35.10 40.59
Avrg. number of assets 3.81 4.06 3.00 2.21 2.13 2.77 5.00 5.00
Avrg. turnover p.a. 2.02 1.97 2.10 2.10 2.57 0.59 0.39 0.38
Avrg. risk-free rate 0.37% 0.37% 0.37% 0.37% 0.37% 0.37% 0.37% 0.37%
Obs. 47 47 47 47 47 47 47 47
Notes: This table reports portfolio performance measures for the four sub-periods from 1993 to 2011 for the moderate
investor type. In all optimized portfolios, the maximum expected volatility is constrained to 10% p.a. All portfolios
are rebalanced at the first trading day of every month. */** Indicate a significantly higher Sharpe ratio compared to
mean-variance at the significance levels of 1% and 5%, respectively.
the analysis of the full sample. Furthermore, the analysis reveals consistently lower portfolio
turnovers and, hence, lower transaction costs for BL compared to MV for all sub-periods. Addi-
tionally, for all sub-periods BL portfolios are more diversified across asset classes than MV and
BS portfolios as indicated by a higher average number of assets in the optimized portfolios. Over-
all, our results suggest that the BL model outperforms MV, BS, and naïve-diversified portfolios
particularly in recessionary periods. The explanation is that the BL model reacts more quickly to
changes in the economic cycle and adjusts the asset allocation more rapidly.
The European Journal of Finance 19
5. Robustness checks
We perform a variety of robustness checks and sensitivity analyses to confirm the robustness of
our results. We vary the constraints on optimized portfolio weights and relax the short selling
constraint (Section 5.1). We analyze variations in all input parameters (Section 5.3) and estima-
tion windows (Section 5.2) and modify the level of transaction costs and, alternatively, include
transaction costs in the optimization function (Section 5.4). Finally, we alter the investment uni-
verse and employ different sets of assets (Section 5.5). We primarily report the results for the
BL-st.w. strategy because the results for the other reference portfolios used in the BL approach
(st.w., 1/N, MinVar) are qualitatively similar. The same applies for the three investor types in that
we focus on the moderate investor type, as the results for the others are similar. All unreported
results are available upon request.
Conservative
Net Sharpe ratio 0.95† 0.86† 0.91† 0.77†† 0.73†† 0.74†† 0.65 0.27
Net Omega measure 2.02 1.88 1.97 1.77 1.71 1.73 1.66 1.24
Net MDD (%) 7.86 8.46 6.49 9.76 10.83 7.67 13.99 40.59
Avrg. number of assets 4.42 4.20 3.91 3.48 3.51 3.69 5.00 5.00
Avrg. turnover p.a. 1.11 1.00 0.99 1.36 1.53 0.51 0.18 0.29
Moderate
Net Sharpe ratio 0.64##† 0.65##† 0.62##† 0.52†† 0.49 – 0.32 0.27
Net Omega measure 1.60 1.62 1.62 1.48 1.44 – 1.27 1.24
Net MDD (%) 21.84 20.46 19.33 25.02 25.99 – 35.10 40.59
Avrg. number of assets 4.45 4.48 4.31 3.91 3.88 – 5.00 5.00
Avrg. turnover p.a. 1.44 1.32 1.42 1.80 1.87 – 0.27 0.29
Aggressive
Net Sharpe ratio 0.40## 0.41## 0.44∗∗## 0.25 0.31 – 0.19 0.27
Downloaded by [FU Berlin] at 01:39 15 May 2015
Net Omega measure 1.36 1.37 1.41 1.22 1.27 – 1.15 1.24
Net MDD (%) 40.37 39.87 38.61 43.31 43.33 – 50.99 40.59
Avrg. number of assets 4.21 4.36 4.52 4.28 4.15 – 5.00 5.00
Avrg. turnover p.a. 1.49 1.44 1.46 1.89 1.64 – 0.25 0.29
Notes: This table reports the portfolio performance measures for the full sample from 1993 to 2011. In all optimized
portfolios, the asset weights are restricted to a maximum absolute deviation of 15% from the respective benchmark
weight reported in Table 2. The maximum expected volatility is constrained to 5%, 10%, and 15% p.a. for the
conservative, moderate, and aggressive investor type, respectively. All portfolios are rebalanced at the first trading
day of each month. */**, # /## , † /†† indicate a significantly higher Sharpe ratio compared to mean-variance, the
strategically weighted benchmark, and the 1/N benchmark at the significance levels of 1% and 5%, respectively.
as well as the portfolio turnover. The BL model performs best for risk aversion coefficients
between 2 and 4, which is in line with the assumed moderate investor type.
Avrg. number of 3.64 2.91 2.69 3.45 2.59 2.41 3.26 2.33 2.16 3.00 1.86 1.71 2.76 1.52 1.35 5.00 5.00
assets
Avrg. turnover 2.00 1.96 2.67 2.80 2.78 3.79 3.32 3.34 4.44 3.91 3.80 4.61 4.07 3.61 4.36 0.27 0.29
p.a.
0.5 1 2 4 8 Benchmark
Net Sharpe ratio 0.61 0.54 0.47 0.68 0.54 0.48 0.75∗∗##†† 0.55 0.49 0.78∗∗##†† 0.53 0.50 0.69##†† 0.58 0.48 0.32 0.27
Net MDD 11.62% 14.35% 16.18% 10.35% 14.35% 16.18% 9.46% 14.35% 16.18% 9.94% 14.35% 16.18% 18.59% 10.93% 15.81% 35.10% 40.59%
Avrg. number of 2.79 2.31 2.17 2.95 2.31 2.16 3.26 2.33 2.16 3.93 2.35 2.19 4.53 2.42 2.25 5.00 5.00
assets
Estimation 12 24 36 48 60 Benchmark
window
number of
months BL-st.w. BS MV BL-st.w. BS MV BL-st.w. BS MV BL-st.w. BS MV BL-st.w. BS MV st.w. 1/N
Net Sharpe ratio 0.63 0.66 0.56 0.73∗∗##†† 0.58 0.48 0.75∗∗##†† 0.55 0.49 0.75∗∗##†† 0.42 0.46 0.73†† 0.44 0.47 0.32 0.27
Net MDD 16.13% 10.18% 18.28% 14.58% 14.04% 15.73% 9.46% 14.35% 16.18% 8.97% 17.55% 16.99% 8.76% 16.82% 18.26% 35.10% 40.59%
Avrg. number of 3.04 2.37 1.90 3.14 2.16 2.09 3.26 2.33 2.16 3.38 2.38 2.24 3.41 2.50 2.29 5.00 5.00
assets
Avrg. turnover 4.34 3.29 4.63 3.56 2.86 4.33 3.32 3.34 4.44 3.48 3.51 4.60 3.27 3.31 4.59 0.27 0.29
p.a.
(Continued).
21
22
W. Bessler et al.
Downloaded by [FU Berlin] at 01:39 15 May 2015
Table 8. Continued.
5.00% 7.50% 10% 15% 20% Benchmark
Maximum
volatility p.a. BL-st.w. BS MV BL-st.w. BS MV BL-st.w. BS MV BL-st.w. BS MV BL-st.w. BS MV st.w. 1/N
Net Sharpe ratio 0.45 0.26 0.23 0.57 0.53 0.32 0.75∗∗##†† 0.55 0.49 0.56 0.43 0.33 0.48 0.44 0.34 0.32 0.27
Net MDD 14.48% 19.59% 22.69% 15.65% 15.69% 24.93% 9.46% 14.35% 16.18% 17.45% 19.44% 20.58% 28.95% 22.70% 30.87% 35.10% 40.59%
Avrg. number of 2.97 2.10 2.01 3.10 2.26 2.15 3.26 2.33 2.16 3.38 2.24 2.13 3.69 2.17 2.03 5.00 5.00
assets
Avrg. turnover 12.30 14.44 15.29 4.76 5.68 6.41 3.32 3.34 4.44 2.92 2.45 4.07 1.81 2.64 2.80 0.27 0.29
p.a.
Notes: The table shows sensitivity analysis for different optimization constraints, risk aversion coefficients, and estimation windows for the moderate investor type and the full
sample from 1993 to 2011. ∗ /∗∗ , # /## , † /†† indicate a significantly higher Sharpe ratio compared to mean-variance, the strategically weighted benchmark, and the 1/N benchmark
at the significance levels of 1% and 5%, respectively.
The European Journal of Finance 23
Relative to BS, the BL approach performs consistently superior with one exception: for a 12-
months estimation window for the covariance matrix, the BS approach achieves a slightly higher
Sharpe ratio than BL. However, further analysis suggests that this result is due to inflating the
covariance matrix in the BS model, which is equivalent to reducing the volatility constraint to
6.74%.6 With a volatility constraint of 6.74%, the BL and MV approaches yield a Sharpe ratio of
0.83 and 0.56 (unreported), respectively. Hence, when comparing equally constrained portfolios,
the BL approach outperforms BS and MV, which confirms our base case results.
dence in implied returns. For values of τ close to zero, the unconstrained optimized BL portfolio
converges to the benchmark portfolio. In contrast, for very large values of τ , the optimized BL
portfolio converges to the MV optimized portfolio. Panel A of Table 9 shows that for the com-
monly used values for τ between 0.025 and 1.00 (Black and Litterman 1992; He and Litterman
1999; Drobetz 2001; Idzorek 2005), our results are robust. This holds for the out-of-sample
Sharpe ratios as well as for the portfolio risk (MDD), the portfolio diversification, and the port-
folio turnover. Furthermore, we observe that the BL portfolio’s deviation from the benchmark
declines with lower values of τ , which results in declining tracking errors. This is consistent with
the interpretation of τ as an uncertainty measure of implied returns and confirms its function to
control the desired deviation from the reference portfolio.
Second, we vary the estimation window for , which quantifies the reliability of ‘views’ based
on the historic estimation errors of return forecasts. Panel B of Table 9 confirms that our results
are robust for estimation windows of between 6 and 36 months. Third, we assume an overall
time-invariant level of confidence in the ‘views’ (Meucci 2010) by substituting according
to Equation (11). In the approach no additional time-varying information on the reliability of
‘views’ is included. Panel C of Table 9 presents the results for different overall time-invariant
levels of confidence, measured by the scalar c. We find that for all confidence levels (scalars c
between 0.1 and 100), the outperformance of the BL model relative to MV vanishes. Therefore, it
seems that including time-varying information on the reliability of return estimates is an essential
factor for the superior performance of the BL model relative to MV. This leads, on the one hand,
to an investment close to the reference portfolio when market conditions are uncertain and, on
the other hand, to larger deviations from this reference portfolio when markets are more stable
and return forecasts are more reliable.
W. Bessler et al.
BS MV BL-st.w.: varying parameter τ st.w.
Avrg. number of assets 2.33 2.16 2.72 2.86 3.06 3.26 3.87 2.34 5.00
Avrg. turnover p.a. 3.34 4.44 4.01 3.88 3.56 3.32 2.79 4.45 0.27
Tracking error (%) 2.17 2.50 2.34 2.20 2.12 2.03 1.68 2.40 –
Panel B: Variation of estimation window for uncertainty measure of views ()
Net Sharpe ratio 0.55 0.49 0.52 0.68 0.75∗∗##†† 0.73†† 0.76##†† 0.74†† 0.32
Net MDD (%) 14.35 16.18 13.01 13.66 9.46 10.53 10.64 11.08 35.10
Avrg. number of assets 2.33 2.16 2.89 3.15 3.26 3.22 3.26 3.30 5.00
Avrg. turnover p.a. 3.34 4.44 6.28 4.28 3.32 3.10 2.98 2.98 0.27
Tracking error (%) 2.17 2.50 2.17 2.18 2.03 2.07 2.09 2.07 –
Panel C: Variation of assumed overall level of confidence in views (c)
Net Sharpe ratio 0.55 0.49 0.46 0.52 0.51 0.52 0.51 0.49 0.32
Net MDD (%) 14.35 16.18 29.13 15.68 15.60 15.74 15.88 16.09 35.10
Avrg. number of assets 2.33 2.16 4.86 2.67 2.44 2.33 2.24 2.17 5.00
Avrg. turnover p.a. 3.34 4.44 1.14 4.07 4.38 4.44 4.47 4.44 0.27
Tracking error (%) 2.17 2.50 0.51 2.02 2.23 2.33 2.37 2.42 –
Notes: The table shows sensitivity analyses for setting different model parameters in the BL model for the moderate investor type and the full sample from 1993 to 2011. The
strategically weighted benchmark is the reference portfolio for BL and is used to compute tracking errors. */**, # /## , † /†† indicate a significantly higher Sharpe ratio compared
to mean-variance, the strategically weighted benchmark, and the 1/N benchmark at the significance levels of 1% and 5%, respectively.
The European Journal of Finance 25
Notes: The table shows a sensitivity analysis for different levels of transaction costs for the moderate investor type
and the full sample from 1993 to 2011. */**, # /## , † /†† indicate a significantly higher Sharpe ratio compared to mean-
variance, the strategically weighted benchmark, and the 1/N benchmark at the significance levels of 1% and 5%,
respectively.
Downloaded by [FU Berlin] at 01:39 15 May 2015
Alternatively, we include variable transaction costs in the optimization function. The opti-
mization problem with transaction costs is
δ
max U = ω μ −
ϕ − ω ω, (16)
ω 2
where
is a vector that contains the changes in portfolio weights required to rebalance the
portfolio at the monthly rebalancing dates, and ϕ is the vector of variable transaction costs of
the assets. The results in Table 11 confirm our findings for the base case. This holds for the
out-of-sample Sharpe ratios as well as for the portfolio risk (MDD), the portfolio diversification,
and the portfolio turnover. As expected, including transaction costs in the optimization function
substantially reduces portfolio turnover for all optimization strategies and it slightly enhances
performance after transaction costs for most strategies.
Conservative
Net Sharpe ratio 0.96† 0.86†† 0.92†† 0.82†† 0.73 0.64 0.65 0.27
Net Omega measure 2.03 1.90 1.99 1.85 1.72 1.61 1.66 1.24
Net MDD (%) 6.50 6.69 6.09 7.35 7.80 8.35 13.99 40.59
Avrg. number of assets 4.39 4.28 3.73 2.86 2.72 3.81 5.00 5.00
Avrg. turnover p.a. 0.73 0.82 0.82 1.58 2.16 0.21 0.18 0.29
Moderate
Net Sharpe ratio 0.76##† 0.83##†† 0.85##†† 0.50 0.50 – 0.32 0.27
Net Omega measure 1.78 1.87 1.92 1.47 1.47 – 1.27 1.24
Net MDD (%) 13.89 8.45 7.56 13.75 16.18 – 35.10 40.59
Avrg. number of assets 4.12 3.85 3.42 2.39 2.21 – 5.00 5.00
Avrg. turnover p.a. 1.27 1.34 1.34 3.25 3.81 – 0.27 0.29
Downloaded by [FU Berlin] at 01:39 15 May 2015
Aggressive
Net Sharpe ratio 0.72#†† 0.73##†† 0.77##†† 0.52 0.53 – 0.19 0.27
Net Omega measure 1.77 1.78 1.86 1.52 1.53 – 1.15 1.24
Net MDD (%) 17.98 14.72 11.00 22.23 24.59 – 50.99 40.59
Avrg. number of assets 3.28 3.30 2.91 1.97 1.75 – 5.00 5.00
Avrg. turnover p.a. 1.71 1.77 1.87 3.67 4.10 – 0.25 0.29
Notes: This table reports the portfolio performance measures for the full sample from 1993 to 2011 when considering
transaction costs in the optimization function according to Equation (16). The variable transaction costs are set to 30 bp
for all assets. The maximum expected volatility is constrained to 5%, 10%, and 15% p.a. for the conservative, moder-
ate, and aggressive investor type, respectively. */**, # /## , † /†† indicate a significantly higher Sharpe ratio compared to
mean-variance, the strategically weighted benchmark, and the 1/N benchmark at the significance levels of 1% and 5%,
respectively.
includes industry indices for US stocks (US-industries) such as Oil and Gas, Basic Materi-
als, Industrials, Consumer Goods, Health Care, Telecom, Utilities, Financials, Technology, and
Consumer Services for the 1993–2011 period (Datastream indices).
Panel A of Table 12 presents the performance measures (Sharpe ratios) for the two alter-
native multi-asset universes for a moderate investor. The conclusions for the conservative and
aggressive investors are similar. The results confirm the base case in that the BL portfolios are
consistently superior relative to the MV, BS, and naïve approaches. The MV and BS portfolios
also perform slightly better than naïve-diversified portfolios. However, the outperformance of
BS and MV is lower in magnitude and insignificant.
Panel B of Table 12 contains the results for stock-only asset universes.7 Because the risk of
stock-only portfolios is substantially larger than for mixed portfolios, we compute the optimiza-
tion models for the aggressive investor with volatility threshold of 15% p.a. As in our base
case results, the BL model performs consistently better than MV and BS. However, similar
to DeMiguel, Garlappi, and Uppal (2009) and Murtazashvili and Vozlyublennaia (2013), we
find that for stock-only portfolios none of the optimization models is able to outperform naïve-
diversified portfolios such as the 1/N strategy. The likely explanation for this finding is that
in stock-only portfolios the diversification effects and the benefits of reallocating the portfolio
over time are lower and are outweighed by estimation errors (as discussed by Murtazashvili
and Vozlyublennaia 2013) and transaction costs. We conclude that in order to outperform naïve-
diversified portfolios, optimization models require more precise return estimates than provided
The European Journal of Finance 27
Notes: This table reports the portfolio performance (Sharpe ratios) for different asset universes. The column number of
assets shows the number of assets considered in the asset universe, where s is the number of stock indices, b is the number
of bond indices, and c is the number of commodity indices. */**, # /## , † /†† indicate a significantly higher Sharpe ratio
compared to mean-variance, the strategically weighted benchmark, and the 1/N benchmark at the significance levels of
1% and 5%, respectively.
Downloaded by [FU Berlin] at 01:39 15 May 2015
by sample moments. Alternatively, quantitative investors need an asset universe that includes
not only stocks, but also a broader set of assets, providing larger diversification effects and larger
benefits from dynamically reallocating the portfolio, thereby enabling optimization strategies to
provide superior results relative to naïve diversification strategies.
6. Conclusion
We implement a sample-based version of the BL model and analyze its out-of-sample per-
formance relative to BS, MV, minimum-variance, and naïve-diversified portfolios based on
multi-asset rolling sample optimizations for the period from January 1993 to December 2011.
To ensure the comparability of all optimization approaches, we use the same sample moments in
all approaches. Our empirical results contribute to the literature in several dimensions. We find
that the BL model generates a consistently higher out-of-sample performance (Sharpe ratios and
Omega measures) relative to MV, BS, and minimum-variance optimized portfolios. In compar-
ison to naïve diversification strategies with equal or strategic portfolio weights, the BL model
significantly performs better in almost all analyzed cases even after transaction costs.
Furthermore, BL portfolios are less risky as indicated by lower volatility and ‘maximum draw-
down’ (MDD) measures, and are more diversified across asset classes, including a larger number
of assets compared to MV and BS optimized portfolios. Sensitivity analyses suggest that the out-
of-sample outperformance of the BL model compared to MV is due to incorporating additional
information on the reliability of return estimates, resulting in more stable and more accurate
return estimates and consequently in a lower portfolio turnover.
We separate the full sample (1993–2011) into four sub-periods based on the monetary cycle
and stock market signals and find that the BL model outperforms MV and BS in all four sub-
periods, but the outperformance is larger in magnitude during recessionary periods. In line with
the results for the full sample period, we observe the additional benefits of the BL strategy relative
to MV in all sub-periods, such as lower risk (maximum drawdown and volatility), lower portfolio
turnover, and broader portfolio diversification.
The robustness checks confirm that our results are insensitive to restricting optimized portfolio
weights, to short selling restrictions, to all input parameter variations, and to different levels of
28 W. Bessler et al.
transaction costs. This holds even when transaction costs are included in the optimization func-
tion. Finally, our results hold for larger and different sets of assets. For stock-only portfolios,
the BL approaches consistently perform better than MV and BS optimization. However, when
investing solely in stocks, we find that none of the optimization models is able to outperform
consistently naïve-diversified portfolios such as the 1/N strategy. For these stock portfolios, the
diversification effects are lower, and the estimation errors and transaction costs seem to out-
weigh the benefits of reallocating the portfolio over time. Hence, to outperform naïve-diversified
portfolios, optimization models require more precise return estimates than provided by sample
moments or an asset universe that includes not only stocks, but also a broader set of assets
offering larger diversification effects and higher benefits of reallocating the portfolio over time.
Both enable optimization strategies to demonstrate their strengths vis-à-vis naïve diversification
strategies.
Overall, we provide empirical evidence that the BL model consistently performs superior rel-
ative to MV and BS, and significantly outperforms naïve strategies for all analyzed data sets,
sub-periods, and robustness checks. Consequently, quantitative portfolio managers may be able
Downloaded by [FU Berlin] at 01:39 15 May 2015
to improve their asset allocation decisions when employing the BL approach rather than the MV
or BS framework.
Acknowledgements
We are grateful to Chris Adcock (editor) and two anonymous referees for helpful advice and comments that significantly
improved the quality of this research. The authors also thank Lawrence Kryzanowski, Paul Söderlind, Michael Frömmel,
as well as participants at the European Financial Management Symposium 2012, the 19th Annual Meeting of the German
Finance Association 2012, the International Annual Conference of the German OR Society 2012, and the Verein fuer
Socialpolitik Annual Congress 2012 for helpful comments and suggestions.
Notes
1. Value-weighted portfolios weight constituents proportional to their relative market weights and price-weighted
portfolios allocate the fraction of each constituent proportional to its actual market price.
2. Moreover, benchmark portfolios with rebalancing are more common in institutional asset management and we expect
the performance of the portfolios with rebalancing to be less sensitive to the evaluation period compared to the
buy-and-hold portfolio because the portfolio composition of the buy-and-hold portfolio varies over time based on
the relative performance of the assets during the evaluation period.
3. See Opdyke (2007) for a detailed description of the Sharpe ratio test.
4. The results for the other investor types are qualitatively similar so that we only report the performance measures
for the moderate investor. We use the same continuous sample to compute optimized portfolios and divide the
resulting return time series into four sub-samples, thereby avoiding the problems of rolling sample estimations on a
discontinuous sample.
5. It seems surprising that restricting asset weights results in higher Sharpe ratios for conservative investors, but to
an inferior performance for moderate and aggressive investors. The explanation is that the restriction of the asset
weights for the aggressive (moderate) investor results in a maximum portfolio weight of government bonds below
22.5% (50%). Therefore, in periods of stock market downturns, the asset allocation algorithm cannot fully reallocate
the portfolio into the safe asset, thereby loosing performance relative to the unrestricted case.
6. In the BS approach the sample covariance matrix is inflated by the factor (T − 1)/(T − N − 2), where T is the
sample size and N is the number of assets. While for larger observation windows the covariance matrix inflation
in the BS approach plays only a minor role, it gets pronounced for a shorter estimation window of only 12 months
√ factor increases to 2.2 (T = 12; N = 5). In this case, the expected portfolio volatility increases by
as the inflation
the factor 2.2 = 1.48. Because the volatility constraint requires the expected portfolio volatility to be below 10%,
inflating
√ the covariance matrix by a factor of 2.2 is equal to setting the volatility constraint to a level of 6.74%
(10%/ 2.2) instead of 10%.
The European Journal of Finance 29
7. For the stock-only case, we do not compute strategically weighted portfolios as strategic weights are only used to
determine the fraction of stocks relative to bonds and commodities and we do not have any reason to set different
strategic weights for different industries or countries on an ex ante basis.
References
Abanomey, W. S., and I. Mathur. 1999. “The Hedging Benefits of Commodity Futures in International Portfolio
Diversification.” The Journal of Alternative Investments 2 (3): 51–62.
Allaj, E. 2013. “The Black-Litterman Model: A Consistent Estimation of the Parameter Tau.” Financial Markets and
Portfolio Management 27 (2): 217–251.
Anson, M. J. P. 1999. “Maximizing Utility with Commodity Futures Diversification.” The Journal of Portfolio
Management 25 (4): 86–94.
Behr, P., A. Guettler, and F. Miebs. 2013. “On Portfolio Optimization: Imposing the Right Constraints.” Journal of
Banking & Finance 37 (4): 1232–1242.
Benartzi, S., and R. H. Thaler. 2001. “Naive Diversification Strategies in Defined Contribution Saving Plans.” American
Downloaded by [FU Berlin] at 01:39 15 May 2015
Idzorek, T. M. 2005. A Step-by-Step Guide Through the Black-Litterman Model, Incorporating User Specified Confidence
Levels. Chicago: Ibbotson Associates, 1–32.
Iqbal, J., R. Brooks, and D. U. A. Galagedera. 2010. “Testing Conditional Asset Pricing Models: An Emerging Market
Perspective.” Journal of International Money and Finance 29 (5): 897–918.
Jagannathan, R., and T. Ma. 2003. “Risk Reduction in Large Portfolios: Why Imposing the Wrong Constraint Helps.”
The Journal of Finance 58 (4): 1651–1684.
James, W., and C. Stein. 1961. “Estimation with Quadratic Loss.” Proceedings of the 4th Berkeley Symposium on
Probability and Statistics 1: 361–379.
Jensen, G. R., R. R. Johnson, and Jeffrey M. Mercer. 2000. “Efficient Use of Commodity Futures in Diversified
Portfolios.” Journal of Futures Markets 20 (5): 489–506.
Jensen, G. R., and J. M. Mercer. 2003. “New Evidence on Optimal Asset Allocation.” The Financial Review 38 (3):
435–454.
Jobson, J. D., and R. M. Korkie. 1981a. “Putting Markowitz Theory to Work.” The Journal of Portfolio Management 7
(4): 70–74.
Jobson, J. D., and B. M. Korkie. 1981b. “Performance Hypothesis Testing with the Sharpe and Treynor Measures.” The
Journal of Finance 36 (4): 889–908.
Jones, R. C., T. Lim, and P. J. Zangari. 2007. “The Black-Litterman Model for Structured Equity Portfolios.” The Journal
Downloaded by [FU Berlin] at 01:39 15 May 2015