Copula-Based Black Litterman Portfolio Optimizaion

European Journal of Operational Research 297 (2022) 1055–1070
Contents lists available at ScienceDirect
European Journal of Operational Research

journal homepage: www.elsevier.com/locate/ejor
Innovative Applications of O.R.
Copula-based Black–Litterman portfolio optimization

Maziar Sahamkhadam a, Andreas Stephan a,b,∗, Ralf Östermark c
a
Linnaeus University, Växjö, Sweden
b
Jönköping International Business School, Jönköping, Sweden
c
Åbo Akademi, Turku, Finland
a r t i c l e i n f o a b s t r a c t
Article history: We extend the Black-Litterman (BL) approach to incorporate tail dependency in portfolio optimization
Received 22 September 2020 and estimate the posterior joint distribution of returns using vine copulas. Our novel copula-based BL
Accepted 7 June 2021
(CBL) model leads to flexibility in modeling returns symmetric and asymmetric multivariate distribution
Available online 12 June 2021
from a range of copula families. Based on a sample of the Eurostoxx 50 constituents (also for S&P 100 as
Keywords: robustness check), we evaluate the performance of the suggested CBL approach and portfolio optimization
Finance technique using out-of-sample back-testing. Our empirical analysis and robustness check indicate better
Portfolio optimization performance for the CBL portfolios in terms of lower tail risk and higher risk-adjusted returns, compared
Black–Litterman framework to the benchmark strategies.
Truncated regular vine copula
Tail constraints © 2021 The Authors. Published by Elsevier B.V.
Conditional value-at-risk This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/)
1. Introduction One of the challenges faced by fund managers is to combine

specific views on underlying assets with allocation techniques.
The mean-variance approach suggested by Markowitz In particular, when using a risk model representing the market
(1952) plays an integral role in modern asset allocation and equilibrium, fund managers might be interested in applying some
has been vastly studied in the finance literature. Using statistical quantifiable views or opinions in constructing diversified portfo-
measures of mean and variance, Markowitz developed the modern lios. The Black–Litterman (BL) model is a widely used approach to
portfolio theory that provides investors with the optimal asset incorporate investors’ views into Markowitz-type allocation prob-
allocation considering their risk-return trade-off. Although the lems (Black & Litterman, 1991; 1992). The BL framework belongs
mean-variance optimization has been widely used in practice, it is to the Bayesian and shrinkage approaches that are also used to re-
sensitive to input parameters causing a number of shortcomings duce the sensitivity of the Markowitz’s mean-variance optimiza-
including extreme weights, corner solutions, and highly chang- tion concerning input parameters (Kolm et al., 2014). In the BL
ing weights over time (Bera & Park, 2008; Best & Grauer, 1991; model expected returns are estimated based on the market equilib-
Chopra & Ziemba, 1993; Grauer & Shen, 20 0 0; Green & Hollifield, rium and the Bayesian method is used to obtain the posterior dis-
1992; Jobson & Korkie, 1980; Michaud, 1989). Several approaches tribution of the mean and covariance matrix, which include both
e.g., Bayesian framework, robust portfolio optimization, higher subjective relative and absolute views of returns (He & Litterman,
moments, tail risk optimization and weight constraints have been 1999).
developed to alleviate the effects of estimation error (see Kolm, Although the BL approach, in its original form, improves the
Tütüncü, & Fabozzi, 2014, and reference therein). Other extensions Markowitz’s portfolio optimization by reducing estimation errors,
include constrained portfolio optimization (Clarke, De Silva, & it enforces normality and constant conditional distribution for as-
Thorley, 2002), transaction costs (e.g., De Roon, Nijman, & Werker, set returns. As regards to the normality assumption, several studies
2001), risk-based constraints (Levy & Levy, 2014), multi-period have extended the original BL model to capture non-normal dis-
portfolios (Cui, Gao, Shi, & Zhu, 2019; Merton, 1969; Östermark, tributions (Fabozzi, Focardi, & Kolm, 2006; Giacometti, Bertocchi,
1991) and new risk measures (e.g., Ahmadi-Javid & Fallah-Tafti, Rachev, & Fabozzi, 2007). For the constant distribution assumption,
2019; Fulga, 2016; Huang, Zhu, Fabozzi, & Fukushima, 2010; volatility modeling, both univariate and multivariate, has been in-
Rockafellar & Uryasev, 20 0 0). corporated into the BL approach (Beach & Orlov, 2007; Deng, 2018;
Harris, Stoja, & Tan, 2017; Palomba, 2008). Although these stud-
∗
ies extend the BL framework to non-Gaussian distributions, mostly
Corresponding author at: Jönköping International Business School, Jönköping,
Sweden.
Student-t, their suggested models fail to capture asymmetries in
E-mail address: [email protected] (A. Stephan). the return distribution. For this reason, copula pooling techniques
https://doi.org/10.1016/j.ejor.2021.06.015
0377-2217/© 2021 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/)
M. Sahamkhadam, A. Stephan and R. Östermark European Journal of Operational Research 297 (2022) 1055–1070
have been developed, where an implied (mixture) distribution is empirical results including out-of-sample portfolio back-testing.
obtained by blending investors’ views with market-based returns In Section 5, we present robustness checks. Finally, concluding
(Meucci, 20 06; 20 08). As discussed in Palczewski & Palczewski remarks are provided in Section 6.
(2019), a disadvantage of the distribution pooling methods is a lack
of statistical interpretation. Kolm & Ritter (2017) provide a gen- 2. Methodology
eralized BL approach allowing views on asset pricing models’ pa-
rameters, e.g., risk premia in the Arbitrage Pricing Theory. Using In the CBL model, the posterior distribution of returns is esti-
Verbal Decision Analysis, Silva, Pinheiro, & Poggi (2017) address mated using (i) the prior distribution, (ii) investors’ views, and (iii)
the difficulty of quantifying subjective views and suggest an alter- the dependency structure between assets. To estimate the prior
native approach to incorporate investor’s profile and perspective. distribution, an equilibrium model is used to obtain the prior mean
Pang & Karan (2018) suggest a closed-form solution for the clas- (also known as equilibrium returns). Furthermore, a prior covari-
sical BL portfolio optimization problem using conditional value-at- ance matrix is required to estimate the equilibrium returns. There-
risk (CVaR) as the risk measure. fore, copula models are applied to obtain both the prior covari-
In this paper, we extend the BL approach to incorporate both ance and the dependency structure, which is one of the novel-
symmetric and asymmetric dependence in return distribution by ties of the current paper. Using the directional views obtained
using copula modeling. The copula models can capture symmetric from the VECM/VAR (vector autoregressive) model, we estimate
or asymmetric tail dependence, which is an important aspect of the CBL posterior mean and covariance matrix. We first describe
financial risk modeling and portfolio optimization based on down- the CBL approach using the CAPM-based and risk-adjusted equi-
side risk (Embrechts, Höing, & Juri, 20 03; Patton, 20 04). Several librium models. Thereafter, we present the truncated Rvine copula
studies have shown the importance and advantages of asymmetric model and portfolio optimization methods. Finally, we discuss the
dependence modeling in asset allocation using Archimedean copu- steps involved in estimating the copula-based optimal portfolios.
las (see Al Janabi, Hernandez, Berger, & Nguyen, 2017; Boubaker &
Sghaier, 2013; Low, Alcock, Faff, & Brailsford, 2013; Okimoto, 2014). 2.1. BL approach
Furthermore, vine copulas are more flexible for estimating a pair-
wise dependency structure (Czado, 2019; Joe, 2014). Black & Litterman (1991, 1992) show how to incorporate in-
We combine the regular vine (Rvine) copula model with the BL vestors’ views when constructing portfolio strategies. While the BL
approach. This leads to a generalized version of the BL approach approach imposes consistency between the expected returns and
in which the covariance matrix is estimated from the joint distri- market equilibrium, it allows investors’ views to cause deviations
bution obtained from copula modeling. In our vine copula-based from the market equilibrium. The BL approach also allows us to
BL (CBL) approach, directional investors’ views are constructed by include as many views as investors require. Based on a Bayesian
applying cointegration rank tests and the vector error correction framework, the BL model combines investors’ views with expected
model (VECM) to asset prices. We evaluate the CBL approach in a returns that are consistent with the market equilibrium (known as
portfolio optimization setting. To further reduce the sensitivity of equilibrium returns). In this model, inputs include a prior distri-
our CBL approach to input parameters, we impose tail constraints bution estimated from the equilibrium model, investors’ views and
in reward/risk maximization as suggested in Alexander, Baptista, & their confidence levels, and a scaling parameter indicating the de-
Yan (2007). An empirical analysis of CBL portfolio optimization re- viation from the market equilibrium.
veals that there are gains from incorporating copula modeling into Let d be the total number of assets, rt = (r1t , r2t , ..., rdt ), r jt =
the BL approach to maximize the investor’s utility function, par- p jt −p j,t−1
[ p j,t−1 ] − r f t be the excess returns computed based on the ob-
ticularly for the Min CVaR and Max Sharpe ratio (SR) portfolios.
Most of the tail constrained optimal portfolios result in lower risk served adjusted prices p jt and risk-free rate of return r f t at time
(both in terms of volatility and downside risk). In general, we find t, with a d × d covariance matrix and a d × 1 vector of ex-
that the CBL models outperform the benchmark portfolios for most pected returns μ such that rt = μ + t , t ∼ D (0, ), where D (., . )
out-of-sample measures. is a location-scale distribution. Denoting π as equilibrium returns
This study makes the following contributions. First, we show (prior mean) and imposing the scaling parameter τ , the prior dis-
how to incorporate the vine copula models into the BL approach. tribution for μ is defined as:
The estimation of tail dependency for the posterior conditional dis- μprior ∼ D (π, τ ), (1)
tribution of returns enhances the CBL approach with the flexibil-
ity to model downside risk. This leads to a generalized version where τ is the prior covariance for μ. Let q be a vector of K
of the BL approach in which the covariance matrix is estimated views with a K × K matrix, , representing the confidence in in-
from the joint distribution obtained from copula modeling. Sec- vestors’ views. Given a K × d pick matrix, P, the posterior distribu-
ond, we contribute to the literature on the BL approach with non- tion for μ is:
Gaussian equilibrium returns by modeling the multivariate distri- μposterior |q; ∼ D (μBL , μBL ),
bution from asymmetric copula models. Third, we extend the risk-
μBL = [(τ )−1 + P −1 P]−1 [(τ )−1 π + P −1 q], (2)
adjusted approach in Giacometti et al. (2007) to a copula-based
equilibrium model. Our novel CBL model leads to (i) more flexi-
μBL = [(τ )−1 + P −1 P]−1 ,
bility in modeling returns symmetric and asymmetric multivariate where μBL and μBL denote the posterior mean and covariance for
distribution from a range of copula families, (ii) a simulation-based μ.1 To construct confidence in investors’ views, we follow Meucci
approach that can be applied not only to downside risk modeling (2009, 2010) and set = κ1 PP , where κ ∈ (0, ∞ ) is the gen-
but also for portfolio optimization, and (iii) scenario-based opti- eral confidence level. The pick matrix P is used to represent the
mization and stochastic programming. K views on d returns. Each row in P includes weights for the re-
The remainder of the paper is structured as follows. turns, expressed as a views vector.2
Section 2 presents the methodology including the CBL approach,
the capital asset pricing model (CAPM)-based and risk-adjusted 1
See Meucci (2010) and He & Litterman (1999) for more details on the derivation
equilibrium, investors’ views, truncated vine copulas, and optimal of the posterior distribution.
portfolios with tail constraints. The data set used for the empirical 2
We set κ = 1; hence, the confidence in investors’ views is a diagonal matrix
investigation is described in Section 3. Section 4 provides the containing conditional variances.
1056
Rewriting the posterior covariance as BL = μBL + , the pos- the market capitalization portfolio does not account for an asset’s
terior distribution for r is given as: tail risk. Instead, we suggest using the CVaR-adjusted equilibrium
rBL |q; ∼ D (μBL , BL ), model suggested by Giacometti et al. (2007) with the weights for
(3) a portfolio constructed from copula models. In this case, the risk-
BL = (1 + τ ) − τ 2 P (τ PP + )−1 P.
adjusted equilibrium returns become:
2.2. CAPM-based equilibrium δ ẃ
π= CVaRα √
− E ( r ) , (7)
2 ẃ ẃ
In the original BL model, a CAPM-based market equilibrium is
where ẃ is a vector of the optimal weights from the mean-CVaR
used to estimate the expectations for implied returns. Considering
optimization for the Rvine copula model. CVaRα is the α −level
the investor’s general utility function maximization as:
CVaR, defined in Eq. (22), and E (r ) are the expected returns for
maximize w π − 2δ w w, (4) the optimal portfolio.
w
and applying reverse optimization, given the optimal portfolio 2.4. Investors’ Views
weights w, we can obtain the equilibrium returns as:
π = δ w, (5) The original BL model allows incorporating subjective views in
the portfolio optimization, and therefore, provides an efficient ap-
where δ is the risk aversion coefficient, which can be estimated proach that does not require the utilization of historical data. This
as the ratio of the market risk premium (expected excess returns) could be considered as one of the main motivations for apply-
to market variance. Assuming the CAPM holds, setting w to the ing the BL approach. However, the complex nature of financial re-
market capitalization weights leads to a CAPM-based equilibrium turns and advancements in econometric time-series modeling, in-
model. In this conventional approach, the historical covariance ma- spired several studies to extend the BL approach and to apply data-
trix is used to create the equilibrium returns. Furthermore, no ac- generating processes for constructing investors’ views. For instance,
tual optimization is conducted at this stage; instead, the weight Zhou (2009) suggests a Bayesian learning approach to include his-
vector w is directly based on market capitalization. torical data, market equilibrium, and investors’ views. Geyer & Lu-
Considering the BL model in Eq. (3), we make three main as- civjanská (2016) use predictive regressions, with price-to-earnings
sumptions: (1) normally distributed returns, (2) a constant condi- and dividend-to-price ratios as regressors, to construct investors’
tional distribution, and (3) no tail dependency (either symmetric views. Beach & Orlov (2007) suggest using the GARCH process in
or asymmetric). Relaxing the normality assumption is addressed by defining views vector and its uncertainty, while Deng (2018) incor-
Meucci (20 06, 20 08). Regarding the constant conditional distribu- porates investors’ views using the VECM model augmented with
tion assumption, Harris et al. (2017) use the time-varying CAPM DCC. All these studies aim at improving the BL approach by in-
suggested by Bollerslev, Engle, & Wooldridge (1988), which allows corporating objective views that consider the distributional prop-
modeling the time-varying conditional mean and covariance ma- erties of financial data. The VECM, developed in the seminal work
trix for the equilibrium returns. of Engle & Granger (1987), takes advantage of the cointegration
In the dynamic conditional correlation (DCC) based BL approach relations between asset prices in a multivariate setting and has
suggested by Harris et al. (2017), the marginal distribution is con- been applied to both the investigation of the interdependence of
sidered as normal or Student-t, which cannot capture an asymmet- financial assets (e.g., Östermark, 2001) and the forecasting of stock
ric tail distribution. Therefore, we suggest using copula modeling prices (e.g., Cheung, Cheung, & Wan, 2009; Kuo, 2016).
to estimate returns’ joint distribution. As mentioned above, cop- To construct investors’ views, we follow Pfaff (2016) and use the
ula modeling can also be used to estimate the covariance matrix VECM to create expectations of price changes. The VECM allows
between assets. To do so, first, GARCH models can be used to es- modeling possible cointegration between asset prices that might
timate the univariate conditional volatilities and marginal distribu- lead to better estimation. Let pt = ( p1t , p2t , ..., pdt ) be the observed
tions: asset prices at time t. Then, the VAR process is given by:
r jt = μ
j + jt pt = A1 pt−1 + · · · + Aρ pt−ρ + et . (8)
jt = h jt z jt
(6) The corresponding VECM, with a transitory form, is given as:
z jt ≈ (iid ), ∀ j ∈ {1, 2, ..., d}

h jt = ω j + α j 2j,t−1 + β j h j,t−1 , ∀ j ∈ {1, 2, ..., d}. pt = αβ pt−1 + 1 pt−1 + · · · + ρ −1 pt−ρ +1 + et ,
(9)
i = − (Ai+1 + . . . + Aρ ), i = 1, . . . , ρ − 1,
Then, the pseudo observations from the inverse of the marginal
distributions are added into the copula model. Having estimated where i represents the cumulative transitory effects, αβ is of
the copula parameters, asset returns rˆ can be simulated from the reduced rank in the case of cointegration, and α and β are the
joint distribution.3 We use these simulations to create a copula- d × k loading and long-run coefficient matrices with k cointegra-
based prior covariance matrix = cov(rˆ ). Inserting this covariance tion rank, respectively (see Pfaff, 2008). The VECM can be es-
matrix into Eq. (5), we obtain a copula-based equilibrium model. timated using maximum likelihood estimation as suggested by
Johansen (1995). To find the best VAR lag order ρ , we use the
2.3. Risk-adjusted equilibrium model Schwartz information criterion (see supplementary material, Fig.
S1). To select an optimal cointegration relation k, we use the max-
In the original BL approach, equilibrium returns are modeled imum likelihood cointegration test (see Section 3).
using the reverse optimization of an investor’s utility function. In Converting the VECM into a VAR process, we obtain the one-
this utility function, portfolio variance is considered as risk mea- step ahead price trajectories with 10% confidence intervals ( pˆLj,t+1
sure. Giacometti et al. (2007) suggest an adjustment that incorpo- and pÛj,t+1 ). Finally, comparing the last adjusted price p jt with the
rates tail risk measures, both VaR and CVaR, to estimate equilib- forecasts, the return views qt = (q1t , q2t , ..., qdt ) are computed as:
rium returns. Furthermore, market capitalization weights are con-
⎧ pˆL −p
sidered as an optimal portfolio when the CAPM holds. However, ⎪
⎨ p jt ,
j,t+1 jt
p jt < pˆLj,t+1
q jt = 0, pˆLj,t+1 < p jt < pÛj,t+1 (10)
3
For more information and the steps involved in GARCH-copula models, see ⎪
⎩ pÛj,t+1 −p jt ,
Sahamkhadam, Stephan, & Östermark (2018). p jt
pÛj,t+1 < p jt .
1057
Since we estimate the views vector qt using Eq. (10) with K = d, sequence of i = 1, 2, . . . , d − 1 linked trees, the decomposition can
the pick matrix is set to identity P = Id such that each row in P be presented in a graphical PCC, known as the regular vine. Let
captures the expected return of one asset.4 e ∈ Ei be the edge between two nodes. ne and ke represent a pair-
copula cne ,ke ;De conditioned on De , with the copula parameter(s)
2.5. CBL posterior distributions ne ,ke |De . Let uDe = {ui |i ∈ De } be the variables in the conditioning
set De , and Cne |De is the conditional distribution of Une |U De . When
In our CBL approach, we suggest estimating returns’ posterior the number of trees increases, the conditioning set De also grows
distribution using a prior covariance matrix from the copula and it is common to consider only the dependence of cne ,ke ;De on
model. To do so, the copula-based posterior mean μ ˆ CBL and covari- the indices in De , ignoring the impact of uDe . This is the so-called
ance ˆ CBL are estimated using Eqs. (2) and (3). Given the residuals simplifying assumption (see Acar, Genest, & NešLehová, 2012; Haff
ηˆ simulated from copula modeling, the CBL posterior distribution et al., 2013). Several studies have investigated the quality of this
is: assumption. For instance, Haff, Aas, & Frigessi (2010) show that the
simplified PCC provides an acceptable approximation of the depen-
rˆCBL = μ ˆ CBL ] 21 η
ˆ CBL + [ ˆ, (11)
dence structure. Killiches, Kraus, & Czado (2017) also conclude that
where ηˆ ∼ i.i.d. is the vector of the covariance-standardized resid- the simplified PCC results in a smooth fit and is preferred over the
uals, which can be simulated from a joint distribution estimated non-simplified vine copulas for practical applications with high-
with the copula density function. In the CBL approach, both the dimensional data.
posterior covariance matrix and the simulated residuals preserve The copula density for a simplified Rvine copula is:
the dependency structure between assets.
d−1
2.6. Rvine copula c ( u| ) = cne ,ke ;De Cne |De (une |uDe ), Cke |De (uke |uDe )| ne ,ke |De ,
i=1 e∈Ei
Sklar (1959) showed how to obtain the d-dimensional copula C (15)

which for the distribution function F with marginal distributions
with the corresponding log-likelihood function:
F1 , ..., Fd is:
∀z ∈ d
: F (z1 , z2 , . . . , zd ) = C (F1 (z1 ), F2 (z2 ), . . . , Fd (zd ) ) d d−1
L( |u ) = ln[cne ,ke ;De (Cne |De (u j,ne |u j,De ), Cke |De (u j,ke |u j,De )| ne ,ke |De )].
= C (u1 , u2 , . . . , ud ), (12) j=1 i=1 e∈Ei
where z j = Fj−1 u j , u j ∼ U [0, 1]d , ∀ j ∈ {1, 2, . . . , d}. If all the (16)

marginals are continuous, then C is unique; if C is a d-dimensional Although vine copulas exhibit flexibility in estimating tail de-
copula, then F is a multivariate distribution function with the pendency, there is a tradeoff between higher flexibility and in-
marginals F1 , F2 , ..., Fd and therefore: creased computational load in high-dimensional settings. For this
reason, truncated and simplified vine structures that allow for
C (u1 , u2 , ..., ud ) = F (F1−1 (u1 ), F2−1 (u2 ), ..., Fd−1 (ud ))=F (z1 , z2 , ..., zd ).
high-dimensional dependence estimation have been developed
(13) (see e.g., Brechmann, Czado, & Aas, 2012; Brechmann & Joe, 2015;
Heinen, Valdesogo et al., 2009; Kurowicka, 2011). In particular,
Let be the parameter set in the copula multivariate distribu-
Brechmann et al. (2012) show how to truncate the Rvine structure.
tion function C (u1 , u2 , . . . , ud | ) and f j the derivative of the uni-
This truncation is applied to the number of trees in the vine by
variate marginal distribution Fj . Therefore, the density function for
setting an independence copula at each edge from a specific tree
the d-dimensional joint distribution is given as:
I ∈ {1, 2, . . . , d − 1} to the final tree. The I-level truncated Rvine has
∂ dC (F1 (z1 ), F2 (z2 ), . . . , Fd (zd )| ) a density of:
f ( z1 , z2 , . . . , zd ) =
∂ z1 , ∂ z2 , . . . , ∂ zd I
d cTruncated (u ) = cne ,ke |De Cne |De (une |uDe ), Cke |De uke |uDe | ne ,ke |De .
= c (F1 (z1 ), F2 (z2 ), . . . , Fd (zd )| ) × fj zj . (14) i=1 e∈Ei
j=1
(17)
In Eqs. (12)–(14), only one copula function C is used to con-
Nagler, Bumann, & Czado (2019) suggest a modified version of
struct the joint distribution, which means only one copula family
the Bayesian Information Criteria (BIC) for vine copulas and show
is considered for the whole set of marginal uniforms u1 , u2 , . . . , ud .
advantages in its performance for portfolio downside risk forecast-
Drawing upon the idea of Joe (1996) to decompose the joint den-
ing. This selection criteria can be used to identify both the trunca-
sity function to several pair-copula densities, Bedford & Cooke
tion level I and the copula families in a mixed vine copula struc-
(2001) and Bedford, Cooke et al. (2002) derive a graphical rep-
ture. Following Nagler et al. (2019), the mBICV is defined as:
resentation of the pair-copula construction (PCC) in the form of
nested trees including regular (Rvine), drawable, and canonical
structures. Aas, Czado, Frigessi, & Bakken (2009) show how to de-
d−1
mBICV( ˆ ) = −2L ( ˆ ) + ˆιlnd − 2 ˆ i lnϕ0i + d − i − ˆ i ln 1 − ϕ0i ,
rive the latter two from Rvine structures. More properties and sta- i=1
tistical inference of vine copulas are developed in Joe (2014) and (18)
Czado (2019).
For a d-dimensional set of continuous random variables, there where î represents the number of independent copulas in the
exist d (d − 1 )/2 pair-copulas, and the copula density c can be de- i:th tree, ˆι is the number non-zero parameters in the set ˆ re-
ˆ
composed into a product of these pair-copulas’ densities. Using a spectively. L ( ˆ ) = dj=1 c u j is the log-likelihood and ϕ0i is the
probability of not having an independence copula in tree i.5
4
Regarding Eq. (10), the investor’s wealth or market evolution (e.g., bull/bear
conditions) requiring different investment strategies in different times and contin-
5
gent on wealth as a dynamic concept could be incorporated (see Östermark, 2017). See Nagler et al. (2019) for the derivation of the mBICV.
1058
2.7. Portfolio Optimization Methods ˆ t , rˆt ) − l]+ = max{[f(w

where [f(w ˆ t , rˆt ) − l], 0}, and:
According to Markowitz (1952), for a d-dimensional portfo- VaRα (w

ˆ t ) = argmin Fα (w
ˆ t , l ), CVaRα (w
ˆ t ) = Fα w
ˆ t , VaRα (w
ˆ t) .
l∈
lio with asset returns rˆt = (rˆ1t , rˆ2t , ..., rˆdt ), asset weights w ˆt =
(wˆ 1t , wˆ 2t , ..., wˆ dt ), a d × d positive-definite covariance matrix ˆ t, (24)
and a d × 1 vector of asset means μ ˆ t = (μ
ˆ 1t , μ
ˆ 2t , ..., μ
ˆ dt ) at To approximate Fα (w ˆ t , l ) and express it as piecewise linear
time (out-of-sample iteration) t, the portfolio’s expected return w.r.t. l, simulations can be drawn from the probability distribution
ˆ
and variance are w ˆt μ
ˆ t and w ˆt ˆ t , respectively. Markowitz
tw of rˆt according to p(rˆt ), such that rˆt = {rˆ mt , m = 1, ..., M}. In this
(1952) suggests that investors should maximize the mean-variance case, we have:
utility function. Another possibility is to maximize the Sharpe ra-
tio (SR) which is defined as a ratio of the expected portfolio re- 1
M +
Fˆα (w
ˆ t, l) = l + f (w
ˆ t , rˆ mt ) − l . (25)
turn to the standard deviation (c.f., Sharpe, 1966; 1994). SR max- M (1 − α )
m=1
imization aims to find the optimal portfolio with the highest SR
from Markowitz’s mean-variance efficient frontier (see Stoyanov, Rockafellar & Uryasev (20 0 0) suggest using the approximation
Rachev, & Fabozzi, 2007, for a discussion on reward/risk portfolios). presented in Eq. (25) to minimize CVaR. This can be achieved

Both reward (portfolio expected return) and risk (portfolio vari- by setting f(w ˆ t , rˆ mt ) = −w ˆ t rˆ mt and l = VaRα (w ˆ t ) and introduc-
ance) measures are positive homogeneous. Furthermore, the re- ing an auxiliary variable υ t = ( υ 1t , υ 2t , ..., υ Mt ) such that υmt =

ward and risk measures are concave and convex functions, respec- −wˆ t rˆ mt − VaRα (w ˆ t ) + . Therefore, the CVaR minimization prob-
tively. Therefore, using the fractional programming suggested by lem can be solved using linear programming:
Charnes & Cooper (1962) and Dinkelbach (1967), Max SR optimiza-

M
tion reduces to convex linearly constrained quadratic programming minimize VaRα + 1
M ( 1 −α )
υmt portfolio risk
ˆ t ,VaRα ,υt
w m=1
and can be posed as: subject to

ˆ t rˆ mt + VaRα + υmt ≥ 0,
w
∀m ∈ {1, 2, .., M}

minimize ˜ t
w ˆ tw
˜t portfolio risk ˆt 1=1
w full investment
˜ t ,ν
w
0≤w ˆ jt ≤ 1, ∀ j ∈ {1, 2, .., d} long positions only
subject to ˜ t ˆ t ≥
w μ μ0 portfolio return υmt ≥ 0,
˜ t 1 =
w ν full investment
0≤w ˜ jt ≤ ν, ∀ j ∈ {1, 2, .., d} long positions only (26)
ν > 0,
The stable tail-adjusted return ratio, known as the STARR or
(19) ˆ t μ
w ˆt
mean/CVaR ratio, is given by CVaRα (w ˆt)
, and can, similar to SR, also
where μ0 is the target portfolio return, ν is an auxiliary scaling be maximized by applying fractional programming. In this case,
˜ t is a vector of the unconstrained weights;
variable (scalar) and w the fractional objective function is reduced to a linear program-
˜
w ming formulation:
the final optimal weights are obtained as w ˆ jt = νjt . An important
assumption in this optimization problem is the limitation of posi-

M
tive reward and risk measures. minimize VaRα + 1
M (1−α )
υmt portfolio risk
˜ t ,VaRα ,υt ,ν
w m=1
Downside risk is another risk measure that has been consid-
subject to ˜ t ˆ t ≥
w μ μ0 portfolio return
ered in many studies of optimal portfolio allocation. For instance,
˜ t rˆ mt +
w VaRα + υmt ≥ 0,
in combination with the classical Markowitz mean-variance port- ∀m ∈ {1, 2, .., M}
folio optimization, VaR has been employed as the risk measure ˜ t 1 = ν
w full investment
leading to an optimal mean-VaR portfolio (see e.g., Baixauli-Soler, 0≤w ˜ jt ≤ ν, ∀ j ∈ {1, 2, .., d} long positions only
Alfaro-Cid, & Fernandez-Blanco, 2011; Consigli, 2002; Lwin, Qu, & υmt ≥ 0
MacCarthy, 2017). ν > 0.
Rockafellar & Uryasev (20 0 0) propose linearizing CVaR and us- (27)
ing linear programming to solve the minimum CVaR optimization
problem. Let f(wˆ t , rˆt ) be the loss function from the simulated out- As the CVaR problem can be expressed in linear terms, it is pos-
of-sample returns, α ∈ (0, 1 ) be the confidence level, and p(rˆt ) be sible to add it as a constraint into the portfolio optimization prob-
the probability density of rˆt . The probability that f(w ˆ t , rˆt ) does not lem.
exceed a threshold value l ∈ is: As shown in Alexander et al. (2007), the CVaR constraint better
reduces large losses in the mean-variance framework, compared to
(w
ˆ t, l) = p(rˆt )d (rˆt ), (20) the VaR constraint. The choice of equality or inequality constraint
f (w
ˆ t ,rˆt )≤l
depends on the extent to which a risk-neutral investor’s prefer-
where (w ˆ t , l ) is assumed to be non-decreasing and continuous ence differs from that of a risk-averse investor. In other words,
w.r.t. l; therefore, it is used to derive α −level VaR as: assuming that the risk-neutral investor seeks to maximize the re-
ward/risk ratio, the CVaR constraint limits downside risk to a value
ˆ t ) = min{l |(w
VaRα (w ˆ t , l ) ≥ α}. (21) close to the risk-averse Min CVaR optimization. This difference can
be imposed on the optimization problem by introducing an upper
CVaR is an alternative when the objective is to minimize losses
boundary for downside risk, which could be considered as a con-
beyond the VaR threshold:
fidence level for CVaR estimated from the Min CVaR optimization.

1 This boundary provides the CVaR level with flexibility when maxi-
CVaRα (w
ˆ t) = f (w
ˆ t , rˆt ) p(rˆt )drˆt , (22)
1−α f (w
ˆ t ,rˆt )≥VaRα (w
ˆt)
mizing the reward/risk ratio. In the first step, we estimate the Min
CVaR and Max SR (or STARR) portfolio weights and obtain the min-
where P[f(w ˆ t )] = 1 − α and (w
ˆ t , rˆt ) ≥ VaRα (w ˆ t , l ) = α. imum level of CVaR for the corresponding portfolios. Then, we take
Rockafellar & Uryasev (20 0 0) suggest using a convex and con- the average of the CVaR values and use it as the upper boundary.6
tinuously differentiable function Fα (w ˆ t , l ) as:

1 6
We use this method to obtain a plausible upper boundary for CVaR. Any other
Fα ( w
ˆ t, l) = l + [f ( w
ˆ t , rˆt ) − l]+ p(rˆt )drˆt , (23)
1−α rˆt ∈ d
logical values can be imposed as long as there is a solution for the optimization.
1059
In this case, the constrained Max SR portfolio optimization is given Step 2: Estimate the return’s posterior distribution. Insert the es-
as: timated parameters {, π , q, P, , τ } from Step 1 into Eqs.
(2) and (3) and obtain the posterior mean μ ˆ CBL and co-
minimize ˜ t
w ˆ tw
˜t portfolio risk
w variance ˆ CBL . Use the M simulated standardized residuals
subject to ˜ t μ
w ˆ t ≥ μ0 portfolio return ηˆ from the chosen distribution in Step 1.b, as well as the

M
VaRα + 1
M (1−α )
υmt posterior mean and covariance in Eq. (11), and generate
m=1 the M CBL-return vectors rˆt = {rˆ mt , m = 1, ..., M}.
≤ ν UCVaR
Step 3: Solve the chosen portfolio problem for the optimal scaled
˜ t rˆ mt + VaRα + υmt ≥ 0,
w
∀m ∈ {1, 2, .., M} composition vector w ˆ t using the active optimization
˜ t 1 = ν
w full investment model discussed in Section 2.7. Register the performance

0≤w ˜ jt ≤ ν, ∀ j ∈ {1, 2, .., d} long positions only of the portfolio Pt = w ˆ t pt using the observed prices, with
υmt ≥ 0 comparisons to the representative benchmark portfolios
ν > 0, constructed and held over Ht .
(28) Step 4: Let t = t + t and repeat from Step 1 for t ∈ [t1 , t2 ].
where UCVaR is the maximum acceptable value of the downside

risk of the portfolio. For the constrained Max STARR optimization,
we define: 3. Data

M
minimize VaRα + 1
M (1−α )
υmt portfolio risk To test the suggested model, we focus on European stocks. Our
w m=1 data set includes daily adjusted prices for constituents of the Eu-
subject to ˜ t μ
w ˆ t ≥ μ0 portfolio return
rostoxx 50 index. The sample period runs from February 1998 to

M
VaRα + 1
M (1−α )
υmt December 2020, resulting in 5833 trading days. The data including
m=1 adjusted prices, historical constituent list, and market capitaliza-
≤ ν UCVaR
tion are obtained from Eikon Thompson Reuters’ database. We use
˜ t rˆ mt + VaRα + υmt ≥ 0,
w
3-month government bonds for Germany (until December 2006)
∀m ∈ {1, 2, .., M}
˜ t 1 = ν
w full investment and 3-month government AAA-rated spot rates for the Eurozone
0≤w ˜ jt ≤ ν, ∀ j ∈ {1, 2, .., d} long positions only (from January 2007) as the risk-free rates.7 Table S1 (see supple-
υmt ≥ 0 mentary material) reports the descriptive statistics of the stock re-
ν > 0. turns for constituents of the Eurostoxx 50 index. Except for AIB
Group (AIBG.I) and Pharol SGPS SA (PHRA.LS), all series have pos-
(29)
itive average returns, with Adyen NV (ADYEN.AS) having the high-
2.8. Steps est. The lowest volatility, 1.36%, is observed for Electrabel SA (EL-
CBt.BR.G07). The minimum and maximum returns are reported for
In this section, the steps involved in constructing the CBL- Ageas SA (AGES.BR) and Volkswagen AG (VOWG.DE), respectively.
based portfolio strategies are presented. Let Tt = Ot + Ht be the While skewness differs by series, all have positive kurtosis along
number of time points Tt in the observation interval Ot and the with the results of Jarque–Bera’s normality test, indicating that
out-of-sample holdout interval Ht at time t , Ot ∩ Ht = ∅, ∀t (see these return time series are non-normally distributed. The results
Appendix A for more details). Repeat steps 1–4 for a specified set of the ARCH test (Engle, 1982) with one lag indicate volatility clus-
of copula, equilibrium, and portfolio optimization models. tering and autocorrelation in the squared residuals for all the se-
ries, except for Volkswagen AG (VOWG.DE). The test statistics for
Step 1: Estimate the return’s prior distribution and investors’
the Ljung–Box test with 10 lags suggest the presence of serial cor-
views for Step. For this, generate the parameters
relation for most of the series.
{u, , π, q, P, , τ }: To construct portfolio strategies, we consider the historical con-
(a) u: Extract the standardized residuals z from the es-
stituent list and only include 50 stocks that are the components of
timated univariate GARCH models in Eq. (6) for the
Eurostoxx 50 at each trading day. In the BL approach, we use the
chosen set of stocks using the observed returns r over
VECM model to construct investor’s views. To investigate whether
Ot . Convert the standardized residuals z into pseudo-
this model is suitable and there is cointegration rank among the
observations u using the probability marginal distribu-
stock price series, we perform the maximum likelihood test sug-
tion.
gested by Johansen (1995). We test the null hypothesis of rank k
(b) Copula cumulative distribution function: Estimate the
against k + 1, and select the first rank when the test does not reject
conditional multivariate cumulative distribution func-
the null hypothesis. Fig. 1 plots the selected cointegration ranks
tion of the returns using the chosen copula model on
and their p-values using a rolling window of 500 days. As one can
the pseudo-observations from Step 1.a.
see for most out-of-sample intervals, there are cointegration rela-
(c) : Simulate the returns from the cumulative distribu-
tionships between stock prices. We use these selected ranks when
tion function estimated in Step (1.b) a sufficient num-
constructing investor’s views q in the CBL approach.
ber of times to compute the prior covariance matrix
As a robustness check, we also apply the suggested portfolio
. models to U.S. stocks (i.e., the constituents of the S&P 100 index).
(d) π : Apply the prior covariance matrix from Step (1.c)
The sample for the U.S. market expands from August 1998 to De-
to obtain the prior mean from the chosen equilibrium
cember 2020, resulting in 5631 trading days. We use 3-month U.S.
model in Eq. (5) or (7).
treasury bills as the risk-free rate.8
(e) q: Carry out the VECM/VAR estimation in Eqs. (8) and
(9) using the observed prices p over Ot to generate the
price trajectories and estimate the directional views q.
7
(f) Define P = Id , = κ1 PP , κ ∈ (0, ∞ ) and τ . In this AAA-rated refers to the highest rating a bond issuer can receive from credit-
rating agencies.
study we set κ = 1 and τ = 0.5 (see Section 4 for more 8
The historical constituent list and descriptive statistics of the U.S. stock returns
details on τ ). are available upon request.
1060
Fig. 1. This figure illustrates (i) selected cointegration ranks, and (ii) p-values of the cointegration rank test for the constituents of Eurostoxx 50 index over the out-of-sample
using a rolling window of 500 days.
Table 1
Out-of-sample performance of the CBL Portfolios with the CAPM equilibrium.
Portfolio Ave. Std. Sortino Sharpe CVaR VaR STARR Mean Ave. Portfolio
Strategy Return Devation Ratio Ratio Ratio /VaR Turnover Wealth
Panel A: Benchmark Portfolios

(i) EQW 0.024 1.52 0.021 0.016 6.00 4.50 0.004 0.005 0.010 193.36
(ii) Historical-based
Min CVaR 0.014 1.17 0.016 0.012 4.45 3.26 0.003 0.004 0.032 143.17
Max SR 0.010 1.45 0.009 0.007 5.65 3.88 0.002 0.003 0.126 91.53
Constrained Max SR 0.008 1.27 0.009 0.007 4.75 3.64 0.002 0.002 0.127 95.26
Max STARR 0.004 1.50 0.004 0.003 5.83 4.30 0.001 0.001 0.176 62.50
Constrained Max STARR 0.004 1.33 0.004 0.003 5.10 3.81 0.001 0.001 0.195 70.42
Panel B: BL Portfolios
Min CVaR 0.029 1.25 0.032 0.023 4.79 3.55 0.006 0.008 1.39 146.54
Max SR 0.057 1.60 0.048 0.036 6.37 4.62 0.009 0.012 1.47 492.46
Max STARR 0.055 1.58 0.048 0.035 6.13 4.21 0.009 0.013 1.65 410.77
Panel C: Mixed CBL Portfolios
Min CVaR 0.036 1.25 0.038 0.029 4.91 3.41 0.007 0.011 1.20 241.59
Max SR 0.044 1.49 0.039 0.030 6.00 4.16 0.007 0.011 1.40 281.44
Max STARR 0.017 1.68 0.011 0.010 7.29 4.19 0.002 0.004 1.47 51.39
Panel D: Student-t CBL Portfolios
Min CVaR 0.031 1.21 0.033 0.025 4.77 3.31 0.006 0.009 1.28 174.14
Max SR 0.046 1.47 0.041 0.031 5.81 4.18 0.008 0.011 1.40 306.89
Max STARR 0.023 1.64 0.015 0.014 7.05 4.09 0.003 0.006 1.59 67.17
Panel E: Clayton CBL Portfolios
Min CVaR 0.028 1.26 0.030 0.023 4.91 3.27 0.006 0.009 1.08 167.42
Max SR 0.044 1.45 0.041 0.030 5.68 4.14 0.008 0.011 1.35 292.75
Max STARR 0.012 1.74 0.008 0.007 7.44 4.58 0.002 0.003 1.34 38.85
Notes: The portfolios consist of 50 components of the Eurostoxx 50 index. The out-of-sample period runs from February 20 0 0 to
December 2020, consisting of 5333 trading days. The results are obtained by applying rolling window estimation with a training sample
size of 500 days. Panel A reports the results for the benchmark model including the EQW and historical-based portfolios. Panel B reports
the results for the BL approach. Panels C–E report the results for the CBL model with mixed, Student-t, and Clayton copula. VaR and
CVaR are estimated empirically at the 1% level. Economic measures are average turnover and final wealth for the portfolio at the end
of the sample, assuming a € 100 initial investment with the proportional transaction cost set to 1 basis point (bp). Bold values show
the five best portfolios for each measure.
4. Empirical analysis rest of the targeted sample, which yields 5333 out-of-sample it-
erations. We use the CBL model and estimate the one-step ahead
To evaluate the suggested BL model combined with copula conditional distribution. In the CBL model, we apply the truncated
modeling, we perform out-of-sample back-testing and examine the copula models in which the truncation level is selected based on
optimal portfolio strategies with a tail constraint. To do so, we the mBICV criterion. Estimating the one-step ahead conditional
use a training sample of 500 days and roll this window over the distribution and tail dependence structure, we forecast and sim-
1061
Table 2
Out-of-sample performance of the CBL portfolios with the risk-adjusted equilibrium.
Panel A: BL Portfolios
Min CVaR 0.031 1.19 0.035 0.026 4.59 3.29 0.007 0.009 1.10 196.52
Max SR 0.042 1.35 0.042 0.031 5.42 3.76 0.008 0.011 0.841 376.07
Max STARR 0.034 1.37 0.033 0.025 5.67 3.90 0.006 0.009 1.21 190.60
Panel B: Mixed CBL Portfolios
Min CVaR 0.037 1.12 0.046 0.033 4.11 2.97 0.009 0.013 1.10 289.22
Max SR 0.050 1.21 0.057 0.041 4.45 3.24 0.011 0.015 1.19 506.40
Max STARR 0.046 1.25 0.051 0.037 4.59 3.32 0.010 0.014 1.32 373.44
Panel C: Student-t CBL Portfolios
Min CVaR 0.040 1.13 0.047 0.035 4.23 2.89 0.009 0.014 1.17 314.22
Max SR 0.051 1.20 0.058 0.043 4.49 3.20 0.011 0.016 1.21 537.79
Max STARR 0.048 1.23 0.053 0.039 4.64 3.25 0.010 0.015 1.42 403.03
Panel D: Clayton CBL Portfolios
Min CVaR 0.036 1.15 0.043 0.031 4.33 3.06 0.008 0.012 1.00 280.61
Max SR 0.048 1.22 0.054 0.040 4.52 3.29 0.011 0.015 1.13 480.84
Max STARR 0.040 1.29 0.043 0.031 4.77 3.52 0.008 0.011 1.20 279.50
Notes: The portfolios consist of 50 components of the Eurostoxx 50 index. The out-of-sample period runs from February 20 0 0 to
December 2020, consisting of 5333 trading days. The results are obtained by applying rolling window estimation with a training sample
size of 500 days. Panel A reports the results for the BL approach. Panels B–D report the results for the mixed, Student-t, and Clayton
CBL portfolios. VaR and CVaR are estimated empirically at the 1% level. Economic measures are average turnover and final wealth for
the portfolio at the end of the sample, assuming a € 100 initial investment with the proportional transaction cost set to 1 basis point
(bp). Bold values show the five best portfolios for each measure.
ulate the stock returns. For each out-of-sample iteration, we ob- parison of different distributional assumptions (e.g., symmetric or
tain the optimal weights and, using the realizations of stock re- asymmetric).9
turns over the out-of-sample, we calculate the portfolio returns Regarding the tests in general, we have compared the perfor-
based on a daily rebalancing strategy. Having obtained the re- mance of our empirical models in three different settings. We have
turns for the optimal portfolio strategies based on the suggested not recognized transaction costs, since all methods use the same
forecasting models, we compare the out-of-sample descriptive one-step ahead forecasts with continuous reshuffling of the port-
statistics, risk-adjusted ratios, and economic performance of these folios over time. We assume that the different strategies do not,
portfolios. on average, impose significant differences in the true transaction
In the BL approach, the τ parameter is the confidence level for costs that would be imposed in real-world conditions. Hence, the
investors’ views. In the original BL approach, this parameter is set results are comparable.
to a value close to zero (He & Litterman, 1999). This dominates the We divide our empirical investigation into three parts. First, we
equilibrium model and places a small weight on investors’ views. apply the CBL models with CAPM equilibrium returns and con-
However, because we are using the VECM/VAR model, we expect struct the unconstrained and constrained optimal portfolios. We
to gain from modelling cointegration relations between stocks and also evaluate different copula families, including Clayton, Student-t,
possible improvements of the mean forecasts. As we also imple- and a mixed version where pair-copula families are selected from
ment reward/risk optimization, the BL conditional mean modeling Gaussian, Student-t, Clayton, Frank, Joe, and Gumbel. Second, in
will have an effect on the results of the portfolio back-testing. In Section 4.2, we present and analyze the results of the CBL mod-
this study, we suggest to set τ to 0.5, i.e., the equilibrium and the els with risk-adjusted equilibrium returns. Then, in Section 4.3, we
view models have equal weights. We investigate the appropriate- present a multi-period analysis in which the portfolio strategies
ness of this choice by performing robustness analyses on the scal- are compared based on several sub-periods.
ing parameter (see Tables S2 and S3 in the supplementary mate-
rial).
4.1. CAPM equilibrium
In addition to the equally-weighted (EQW) and historical port-
folios, where we apply the portfolio optimization techniques to
Table 1 reports the out-of-sample performance for the CBL port-
historical and observed asset returns, we define, as another bench-
folios. In the mixed version, pair-wise copulas are selected from
mark, a modified BL approach which is different from the sug-
the Gaussian, Student-t, Clayton, Frank, Gumbel, and Joe families
gested CBL model. First, we use the historical method to obtain
based on the mBICV criterion.
the prior covariance . Second, we estimate the returns’ posterior
distribution by drawing M simulations from the multivariate Gaus-
sian distribution. Since we apply the investor’s view model similar 9
For a comparison of alternative investor view modeling, we also construct the
to the CBL model, we consider the BL approach a benchmark on BL and CBL portfolios using the momentum strategy suggested by Fabozzi et al.
how the equilibrium model is estimated. This provides us a com- (2006) (see supplementary material, Table S4).
1062
Table 3
Multi-period analysis with CAPM equilibrium.
Portfolio 20 0 0–20 06 20 07–20 09 2010–2019 2020

Strategy
Ave. Std. CVaR Ave. Std. CVaR Ave. Std. CVaR Ave. Std. CVaR
Return Devation Return Devation Return Devation Return Devation

(i) EQW 0.024 1.48 5.38 -0.003 2.04 7.79 0.034 1.27 4.50 0.000 2.15 9.34
Min CVaR 0.015 1.16 4.34 -0.058 1.49 5.68 0.040 1.03 3.49 -0.039 1.48 6.96
Max SR -0.011 1.41 5.86 -0.030 1.89 7.58 0.032 1.27 4.16 0.063 1.81 8.15
Constrained Max SR -0.007 1.24 4.72 -0.044 1.63 5.76 0.031 1.13 3.77 0.040 1.63 7.31
Max STARR -0.016 1.40 5.51 -0.064 2.05 8.20 0.034 1.32 4.43 0.048 1.90 8.27
Constrained Max STARR -0.015 1.26 4.71 -0.064 1.78 7.06 0.036 1.18 3.97 0.021 1.70 7.55
Panel B: BL Portfolios
Min CVaR 0.032 1.24 4.59 0.006 1.61 6.21 0.034 1.06 3.73 0.029 1.76 7.19
Max SR 0.084 1.70 6.24 0.041 2.10 8.46 0.042 1.27 4.73 0.078 2.16 8.33
Constrained Max SR 0.062 1.50 5.54 0.040 1.94 7.91 0.041 1.17 4.27 0.074 2.01 7.97
Max STARR 0.068 1.66 6.25 0.057 2.04 7.79 0.042 1.30 4.78 0.095 2.03 7.18
Constrained Max STARR 0.053 1.50 5.72 0.050 1.93 7.23 0.040 1.22 4.31 0.091 1.93 7.20
Panel C: Mixed CBL Portfolios
Min CVaR 0.060 1.12 3.54 -0.017 1.76 7.76 0.040 1.06 3.77 -0.003 1.95 10.12
Max SR 0.092 1.50 5.28 -0.074 1.96 8.72 0.048 1.23 4.34 0.034 2.03 8.25
Constrained Max SR 0.090 1.34 4.46 -0.055 1.78 7.78 0.042 1.14 4.09 0.034 1.91 8.35
Max STARR 0.083 1.48 4.83 -0.181 2.79 17.03 0.031 1.25 4.30 0.021 2.22 9.20
Constrained Max STARR 0.072 1.34 4.33 -0.192 2.72 17.38 0.029 1.18 4.09 0.009 2.12 9.31
Panel D: Student-t CBL Portfolios
Min CVaR 0.047 1.08 3.47 0.001 1.74 7.21 0.036 1.02 3.53 -0.050 1.82 9.07
Max SR 0.096 1.48 5.08 -0.068 1.91 7.93 0.047 1.22 4.35 0.031 2.03 8.24
Max STARR 0.082 1.46 5.03 -0.140 2.68 15.79 0.034 1.23 4.28 -0.010 2.21 9.77
Constrained Max STARR 0.072 1.32 4.48 -0.141 2.62 15.77 0.035 1.15 3.94 -0.026 2.14 9.73
Panel E: Clayton CBL Portfolios
Min CVaR 0.056 1.13 3.55 -0.010 1.81 7.67 0.028 1.06 3.70 -0.041 1.79 8.91
Max SR 0.089 1.46 4.94 -0.067 1.87 7.69 0.046 1.21 4.30 0.046 2.00 8.14
Max STARR 0.078 1.49 4.70 -0.187 2.93 17.42 0.028 1.26 4.41 -0.010 2.43 9.19
Notes: The portfolios consist of 50 components of the Eurostoxx 50 index. The results are obtained by applying rolling window estimation with a training
sample size of 500 days. Panel A reports the results for the benchmark model including the EQW and historical-based portfolios. Panel B reports the results
for the BL approach. Panels C–E report the results for the CBL model with mixed, Student-t, and Clayton copula. CVaR is estimated empirically at the 1%
level. Bold values show the five best portfolios for each measure.
Regarding the benchmarks, in Panel A, the historical-based Min higher volatility and downside risk than their historical counter-
CVaR portfolio outperforms EQW in terms of volatility and down- part, improvement in average return leads to considerably higher
side risk. In particular, it results in an out-of-sample CVaR of 4.45%. Sharpe ratios (0.030 and 0.032). Similar to those in Panel B, both
Though the historical-based reward/risk maximization cannot in- the constrained Max SR and the constrained STARR maximization
crease the out-of-sample average return and the risk-adjusted ra- from the mixed CBL model reduce out-of-sample standard devia-
tios, it results in a less volatile portfolio with lower tail risk com- tion and CVaR compared to non-constrained reward/risk optimiza-
pared with the EQW. For instance, for Max SR, the out-of-sample tion.
standard deviation and CVaR are 1.45% and 5.65%, respectively. The Panels D and E in Table 1 provide the results for the CBL port-
constrained SR and STARR portfolios both reduce volatility and folios with Student-t and Clayton Rvine copulas. Comparing the
downside risk compared with the unconstrained strategies. In this Min CVaR CBL portfolios, the Clayton copula results in lower CVaR
case, constrained Max SR has a volatility of 1.27% and a CVaR of (3.27%), while the mixed copula leads to higher average return and
4.75%. economic performance. In general, there is almost no gain from
Panel B in Table 1 provides the results for the BL portfolios. All changing the copula family to either Clayton or Student-t copulas
the BL-based strategies achieve a higher average return as well as in terms of risk-adjusted ratios. However, most of the CBL portfo-
a better risk-adjusted and economic performance. For instance, the lios achieve less volatility and downside risk compared to the BL-
BL-based Max SR portfolio has an average return of 0.057% and implied portfolios.
an accumulated wealth of € 492.46 at the end of out-of-sample, In summary, these results show that (1) there are advantages to
assuming a € 100 initial investment after considering the propor- incorporating copula modeling into the BL approach to maximize
tional transaction costs. In general, the BL approach with CAPM the investor’s utility function, in particular, for reducing the port-
equilibrium improves the portfolio return, but at the cost of in- folio tail risk, and (2) most of the constrained optimal portfolios
creasing the portfolio volatility and downside risk. result in lower risk (both volatility and downside risk) compared
In Panel C of Table 1, most of the mixed CBL portfolios can to their unconstrained counterparts.
outperform both EQW and historical-based counterparts in terms
of higher average returns and risk-adjusted ratios. Comparing the
4.2. Risk-adjusted equilibrium
EQW with the CBL Min CVaR portfolios in reducing downside risk,
we see that the CBL model reduces out-of-sample CVaR to 3.41%.
As mentioned before, to estimate the risk-adjusted equilibrium
Although the Max SR and constrained Max SR portfolios result in
for the CBL models, we consider a copula-based mean-CVaR port-
1063
Table 4
Multi-period analysis with risk-adjusted equilibrium.
Portfolio 20 0 0-20 06 20 07-20 09 2010-2019 2020

Strategy
Ave. Std. CVaR Ave. Std. CVaR Ave. Std. CVaR Ave. Std. CVaR
Return Devation Return Devation Return Devation Return Devation
Panel A: BL Portfolios
Min CVaR 0.033 1.16 4.26 -0.007 1.50 6.19 0.044 1.03 3.51 -0.008 1.65 7.70
Max SR 0.064 1.38 5.48 0.019 1.89 7.56 0.039 1.08 3.74 -0.003 1.57 6.91
Constrained Max SR 0.055 1.29 5.03 0.017 1.73 6.94 0.039 1.05 3.65 -0.004 1.56 6.89
Max STARR 0.053 1.37 5.40 0.009 1.91 8.23 0.037 1.11 3.87 -0.066 1.68 6.59
Constrained Max STARR 0.044 1.30 4.97 0.016 1.82 7.93 0.040 1.08 3.82 -0.035 1.79 6.58
Panel B: Mixed CBL Portfolios
Min CVaR 0.048 1.00 2.98 0.001 1.50 5.92 0.041 1.00 3.40 0.037 1.63 7.43
Max SR 0.090 1.13 3.44 -0.026 1.66 6.64 0.043 1.04 3.57 0.057 1.58 7.16
Max STARR 0.075 1.15 3.46 -0.004 1.74 6.51 0.039 1.08 3.68 0.057 1.74 8.06
Panel C: Student-t CBL Portfolios
Min CVaR 0.050 1.00 3.00 0.010 1.59 6.47 0.041 0.985 3.38 0.036 1.55 7.19
Max SR 0.089 1.11 3.44 -0.015 1.68 6.91 0.045 1.03 3.62 0.050 1.52 6.88
Max STARR 0.073 1.15 3.78 0.010 1.73 6.63 0.043 1.06 3.65 0.039 1.58 7.22
Constrained Max STARR 0.067 1.09 3.60 0.005 1.69 6.69 0.041 1.03 3.50 0.039 1.56 7.12
Panel D: Clayton CBL Portfolios
Min CVaR 0.061 1.06 3.29 0.008 1.52 5.83 0.030 1.01 3.58 0.009 1.70 8.54
Max SR 0.087 1.15 3.57 -0.004 1.66 6.41 0.037 1.06 3.70 0.050 1.61 7.67
Max STARR 0.071 1.17 3.63 0.032 1.83 6.34 0.023 1.10 3.88 0.011 1.79 9.78
Constrained Max STARR 0.065 1.13 3.52 0.017 1.77 6.21 0.025 1.07 3.76 -0.044 1.89 10.50
Notes: The portfolios consist of 50 components of the Eurostoxx 50 index. The results are obtained by applying rolling window estimation with a training
sample size of 500 days. Panel A reports the results for the BL approach. Panels B–D report the results for the CBL model with mixed, Student-t, and Clayton
copula. CVaR is estimated empirically at the 1% level. Bold values show the five best portfolios for each measure.
folio that represents an investor’s general utility function maxi- model. In most cases, the CBL model with Student-t copula pro-
mization. Then, instead of market capitalization weights, we use an duces both Max SR and Max STARR portfolios with higher av-
asset’s weights from this mean-CVaR optimization in Eq. (7) and erage returns, risk-adjusted ratios, and final wealth. For instance,
estimate the return’s one-step ahead conditional distribution from the Max SR portfolio from the CBL model with Student-t copu-
the CBL model. las in Panel C achieves the highest average returns (0.051%) and
Table 2 presents the out-of-sample performance for the port- accumulated wealth (€ 537.79). In general, the mixed CBL portfo-
folio strategies obtained from the CBL models with risk-adjusted lios achieve lower downside risk compared to their counterparts
equilibrium returns. In Panel A, the BL-based portfolios, computed with Student-t and Clayton copulas. This indicates the advantage
using the risk-adjusted equilibrium, show lower out-of-sample of modeling the dependence structure with a mixed copula model,
volatility and downside risk compared to portfolios based on the that captures both symmetric and asymmetric tail dependence,
CAPM equilibrium (see Table 1). In Panel B, when copula is set to to reduce portfolio downside risk. Comparing results in this ta-
mixed families, the portfolios from the CBL models do outperform ble with those of Table 1, the risk-adjusted equilibrium model
both the benchmarks (Panel A, Table 1) and the BL-based strate- generally produces portfolios with lower volatility and higher
gies in terms of both minimizing portfolio volatility and downside returns.
risk as well as maximizing the risk-adjusted ratios. For instance, In summary, there are gains from using risk-adjusted equi-
the Mixed CBL-based Min CVaR portfolio leads to an out-of-sample librium returns in terms of both maximizing reward/risk ratios
CVaR of 4.11%, which is lower than the benchmarks’ and that of and minimizing portfolio downside risk. When using risk-adjusted
the CBL model with the CAPM equilibrium. Both the constrained equilibrium returns, the CVaR-constrained portfolios can achieve
and the unconstrained Max STARR portfolios from the mixed CBL lower volatility and downside risk than the unconstrained portfo-
model result in higher average returns and thus better economic lios, while providing similar reward/risk ratios. For a comparison of
performance than the same portfolios from the CBL model with the economic performance of CBL portfolios (consisting of the Eu-
CAPM Equilibrium. Applying the risk-adjusted equilibrium in the rostoxx 50 components) with the benchmarks portfolios, see Figs.
CBL approach results in both constrained and unconstrained Max S4–S6 in the supplementary material.
SR portfolios with higher reward/risk ratios, and economic perfor-
mance as well as lower volatility and downside risks than sim- 4.3. Multi-period analysis
ilar portfolios with CAPM equilibrium. In general, the results in
Tables 1 and 2 show that the portfolios obtained from the CBL To perform a multi-period analysis, we divide the out-of-sample
models with risk-adjusted equilibrium returns lead to better out- into several sub-periods. Using these holding periods, we can com-
of-sample performance. pare the short-term performance of the suggested portfolio strate-
Panels C and D in Table 2 report the results of back-testing gies. To do so, we use the first 500 out-of-sample portfolio returns
for the CBL-based portfolios with Student-t and Clayton copulas as a holding period and compute the performance measures. Then,
and risk-adjusted equilibrium returns. Similar to Table 1, there rolling this holding period, we obtain the performance measures
is a limited gain from changing the copula family in the CBL for 4833 sub-periods. We limit our multi-period analysis to the
1064
Table 5
Robustness check - weekly frequency.

(i) EQW 0.180 3.12 0.074 0.058 13.25 8.75 0.014 0.021 0.022 338.56
Min CVaR 0.150 2.71 0.068 0.055 11.92 8.01 0.013 0.019 0.068 285.09
Max SR 0.200 3.53 0.080 0.057 12.86 7.74 0.016 0.026 0.152 373.12
Max STARR 0.201 3.58 0.076 0.056 14.46 8.50 0.014 0.024 0.234 362.04
Panel B: BL Portfolios (CAPM Equilibrium)
Min CVaR 0.196 2.83 0.088 0.069 12.08 8.44 0.016 0.023 1.32 381.29
Max SR 0.214 3.65 0.081 0.059 13.76 9.93 0.016 0.022 1.66 341.06
Max STARR 0.292 3.72 0.113 0.078 13.49 9.61 0.022 0.030 1.70 695.37
Panel C: BL Portfolios (Risk-adjusted Equilibrium)
Min CVaR 0.171 2.84 0.077 0.060 12.21 8.63 0.014 0.020 1.22 303.61
Max SR 0.187 3.57 0.071 0.052 13.78 10.33 0.014 0.018 1.51 275.99
Max STARR 0.210 3.68 0.080 0.057 14.20 9.84 0.015 0.021 1.58 331.64
Panel D: Mixed CBL Portfolios (CAPM Equilibrium)
Min CVaR 0.189 2.46 0.102 0.077 9.75 6.69 0.019 0.028 1.22 395.88
Max SR 0.266 3.23 0.125 0.082 10.48 7.72 0.025 0.034 1.62 642.01
Max STARR 0.227 3.11 0.106 0.073 10.90 8.00 0.021 0.028 1.62 461.28
Panel E: Mixed CBL Portfolios (Risk-adjusted Equilibrium)
Min CVaR 0.218 2.38 0.119 0.091 9.40 6.34 0.023 0.034 1.09 532.70
Max SR 0.248 2.62 0.128 0.094 9.56 6.89 0.026 0.036 1.43 645.35
Max STARR 0.218 2.65 0.111 0.083 9.64 7.10 0.023 0.031 1.44 487.03
Notes: The portfolios consist of 50 components of the Eurostoxx 50 index. The out-of-sample period runs from December 2002 to De-
cember 2020, consisting of 940 weeks. The results are obtained by applying rolling window estimation with a training sample size of 250
weeks. Panel A reports the results for the benchmark models including the EQW and historical approach. Panel B–C report the results
for the BL approach. Panels D–E report the results for the CBL model with mixed copulas. VaR and CVaR are estimated empirically at the
1% level. Economic measures are average turnover and final wealth for the portfolio at the end of the sample, assuming a € 100 initial
investment with the proportional transaction cost set to 1 basis point (bp). Bold values show the five best portfolios for each measure.
based portfolios, with the CBL-based strategies shows the gain

from applying the CBL model to maximize the SR. For most of the
holding periods ending before 2005, the benchmarks result in neg-
ative out-of-sample SRs, while the CBL portfolios produce higher
average returns and, consequently, positive SRs. In particular, dur-
ing this period, the mixed CBL Max SR portfolio produces a higher
SR than the Student-t and Clayton copula models. As expected,
during the global financial crisis (20 07–20 09), all the portfolios
produce lower out-of-sample SRs. During this period, the worst SR
is shown for the EQW portfolio (-0.059), while the Clayton CBL-
based portfolio results in a higher minimum SR (-0.027). After the
global financial crisis and until August 2011, the EQW performs
better than all the other portfolios, with a maximum SR equal to
0.097. For most sub-periods after 2011 until 2020, the CBL portfo-
lios outperform the benchmarks. During the COVID-19 pandemic,
the minimum SR is reported for EQW (-0.044), while the Student-
Fig. 2. This figure illustrates the realized SR for the Max SR portfolio strategies
t CBL Max SR achieves the highest SR (0.087). Comparing with the
computed for each holding period consisting of 500 days. The benchmarks include
the EQW and historical-based portfolios. The CBL portfolios are obtained using the benchmarks, the copula-based CBL portfolios result in higher SRs
Student-t, Clayton, and mixed copula families. for most sub-periods (or holding periods). This indicates advantage
of incorporating asymmetric tail dependence that can be captured
constrained and unconstrained Max SR portfolios obtained from in both the Clayton and the mixed CBL models.
the CBL model with risk-adjusted equilibrium returns.10 Fig. 3 shows the result for the constrained Max SR portfolios.
Fig. 2 plots the realized SRs for the Max SR portfolio strate- Even with tail constraints, the CBL portfolios can outperform the
gies. Comparing the benchmarks including the EQW and historical- benchmarks. Similar to Fig. 2, the realized SR obtained from the
constrained portfolios peaks during 2006. The highest SR (0.176) is
reported for the Student-t CBL portfolio, while the historical-based
10
Further results are available on request. optimization produces the lowest SR (-0.096). Similar to the un-
1065
Table 6
Robustness check - monthly frequency.

(i) EQW 0.607 5.26 0.176 0.115 11.19 7.59 0.054 0.080 0.050 177.66
Min CVaR 0.540 4.13 0.217 0.131 8.17 6.21 0.066 0.087 0.090 174.82
Max SR 0.715 4.50 0.258 0.159 8.82 7.01 0.081 0.102 0.138 211.97
Max STARR 0.679 4.37 0.270 0.156 7.96 7.47 0.085 0.091 0.172 204.57
Min CVaR 0.571 5.34 0.135 0.107 12.96 8.14 0.044 0.070 1.52 165.27
Max SR 0.407 7.24 0.075 0.056 16.50 12.18 0.025 0.033 1.77 115.45
Max STARR 0.597 7.06 0.117 0.085 15.78 11.75 0.038 0.051 1.77 148.82
Min CVaR 0.583 5.30 0.146 0.110 12.59 8.44 0.046 0.069 1.50 168.50
Max SR 0.370 7.36 0.066 0.050 17.20 12.76 0.022 0.029 1.77 109.10
Max STARR 0.453 7.28 0.083 0.062 16.84 11.89 0.027 0.038 1.77 122.03
Min CVaR 0.451 5.02 0.129 0.090 10.76 8.11 0.042 0.056 1.40 146.32
Max SR 0.277 6.46 0.054 0.043 16.38 10.92 0.017 0.025 1.70 105.44
Max STARR 0.302 6.59 0.059 0.046 15.05 9.89 0.020 0.030 1.71 107.33
Min CVaR 0.754 4.44 0.254 0.170 9.58 6.20 0.079 0.122 0.998 220.86
Max SR 0.383 6.17 0.098 0.062 12.77 8.43 0.030 0.045 1.53 125.43
Max STARR 0.543 5.69 0.157 0.096 11.56 8.94 0.047 0.061 1.54 157.56
Notes: The portfolios consist of 50 components of the Eurostoxx 50 index. The out-of-sample period runs from September 2010 to Decem-
ber 2020, consisting of 123 months. The results are obtained by applying rolling window estimation with a training sample size of 150
months. Panel A reports the results for the benchmark models including the EQW and historical approach. Panel B–C report the results
for the BL approach. Panels D–E report the results for the CBL model with mixed copulas. VaR and CVaR are estimated empirically at the
5% level. Economic measures are average turnover and final wealth for the portfolio at the end of the sample, assuming a € 100 initial
investment with the proportional transaction cost set to 1 basis point (bp). Bold values show the five best portfolios for each measure.
Fig. 3. This figure illustrates the realized SR for the constrained Max SR portfo- Fig. 4. This figure illustrates the realized CVaR for the mixed CBL portfolio strate-
lio strategies computed for each holding period consisting of 500 days. The bench- gies computed for each holding period consisting of 500 days.
marks include the EQW and historical-based portfolios. The CBL portfolios are ob-
tained using the Student-t, Clayton, and mixed copula families.
For a comparison of the (constrained) Max STARR CBL portfolios,
constrained optimization, the Clayton CBL model leads to the most see Figs. S2 and S3 in the supplementary material.
sub-periods (1674) with the highest SR, followed by the mixed and Table 3 reports performance measures obtained by holding the
Student-t CBL models (1499 and 931, respectively) ). portfolios over four periods. During the 20 0 0-20 06 period, the BL
To show the improvements obtained from adding the CVaR con- portfolios (with CAPM equilibrium) achieve higher average returns
straint into the Max SR optimization, we plot the realized CVaR than the benchmarks. In most cases, the reward/risk maximization,
estimated empirically for the mixed CBL portfolios over the hold- augmented with the BL-implied returns’ posterior distribution, re-
ing periods. As shown in Fig. 4, the constrained Max SR and STARR sults in riskier optimal portfolios than those from the benchmarks.
portfolios result in lower CVaR than the unconstrained portfolios. The CBL-based optimal portfolios perform better than those from
1066
Table 7
Robustness check - stock market.

(i) EQW 0.033 1.23 0.034 0.027 5.04 3.51 0.007 0.009 0.010 364.84
Min CVaR 0.038 0.965 0.051 0.039 3.91 2.70 0.010 0.014 0.056 535.32
Max SR 0.026 1.16 0.029 0.022 4.46 3.32 0.006 0.008 0.118 248.56
Max STARR 0.027 1.26 0.028 0.021 4.72 3.69 0.006 0.007 0.204 234.07
Min CVaR 0.035 1.14 0.040 0.031 4.56 3.28 0.008 0.011 1.48 199.32
Max SR 0.055 1.33 0.056 0.041 4.95 3.64 0.011 0.015 1.37 524.82
Max STARR 0.043 1.40 0.042 0.031 5.43 3.95 0.008 0.011 1.55 253.55
Min CVaR 0.044 1.06 0.056 0.041 4.09 3.01 0.011 0.015 1.26 374.70
Max SR 0.050 1.08 0.062 0.046 4.23 3.01 0.012 0.017 0.801 633.76
Max STARR 0.040 1.21 0.045 0.033 4.51 3.44 0.009 0.012 1.25 278.36
Min CVaR 0.053 1.13 0.062 0.047 4.48 3.02 0.012 0.017 1.29 551.90
Max SR 0.055 1.23 0.059 0.045 4.79 3.54 0.012 0.016 1.29 598.15
Max STARR 0.048 1.33 0.046 0.036 5.27 3.66 0.009 0.013 1.47 350.42
Min CVaR 0.044 0.975 0.059 0.045 3.83 2.50 0.011 0.018 1.17 412.21
Max SR 0.048 0.990 0.063 0.048 3.95 2.69 0.012 0.018 1.19 483.52
Max STARR 0.045 1.03 0.056 0.043 4.15 2.66 0.011 0.017 1.35 375.41
Notes: The portfolios consist of 50 most capitalized components of the S&P 100 index. The out-of-sample period runs from July 2002 to
December 2020, consisting of 5131 days. The results are obtained by applying rolling window estimation with a training sample size of
500 days. Panel A reports the results for the benchmark models including the EQW and historical approach. Panel B–C report the results
for the BL approach. Panels D–E report the results for the CBL model with mixed copulas. VaR and CVaR are estimated empirically at
the 1% level. Economic measures are average turnover and final wealth for the portfolio at the end of the sample, assuming a $100
initial investment with the proportional transaction cost set to 1 basis point (bp). Bold values show the five best portfolios for each
measure.
the BL approach in terms of volatility and downside risk. Dur- copula families. For both frequencies, the data set runs from Febru-
ing the 20 07-20 09 (global financial crisis) and 2020 (COVID-19 ary 1998 to December 2020, resulting in 1190 weekly and 270
pandemic), both the BL and CBL approaches fail to produce less monthly observations. For the weekly (monthly) frequency, we use
volatile portfolios with lower downside risk than those based on a rolling window of 250 (150) observations to construct the port-
the historical approach. However, the BL-based (constrained) Max folio strategies.
SR and STARR portfolios lead to higher average returns during the Table 5 reports the out-of-sample performance of the port-
crisis periods. folio strategies obtained from the weekly frequency. In Panels B
Table 4 presents the results for the BL and CBL portfolios with and C, most of the BL portfolios result in higher average return
risk-adjusted equilibrium over different periods. In regard to port- and volatility than the benchmarks in Panel A. Those based on
folio risk, the CBL-based portfolios generally outperform the BL- the CAPM equilibrium generally achieve higher accumulation of
based strategies. Except for the 20 07-20 09 period, most of the wealth. The mixed CBL-based Min CVaR optimization reduces out-
CBL-based portfolios achieve higher average returns than the BL- of-sample CVaR to 9.40% compared to 11.92% of the historical ap-
based portfolios. Comparing the copula families, the Student-t proach. The Max SR and STARR portfolios also result in low CVaR
Rvine copula results in portfolios with lower downside risk during values (9.56% and 9.64%, respectively). In Panel B, both the BL-
the COVID-19 pandemic. However, during the 20 07-20 09 period, based Max STARR portfolios increase the out-of-sample average re-
the Clayton copula performs better in reducing the out-of-sample turn compared to their counterparts from the mixed copula CBL
CVaR. (Panels D and E) models. When applying the risk-adjusted equilib-
rium, the mixed CBL approach also reduces the average turnover,
compared to the BL model. For a multi-period analysis with weekly
5. Robustness checks
frequency, see supplementary material Table S5.
Table 6 presents the results for the monthly frequency. Con-
5.1. Data frequency
sidering the benchmarks in Panel A, most of the historical-based
portfolios result in lower portfolio volatility and higher average re-
As a robustness check, we compare the performance of the
turn. Both the BL and CBL approaches fail to produce portfolios
portfolios obtained from the CBL model with the different copula
with better performance measures than the historical-based. Most
families by changing the data frequency to weekly and monthly.
likely this is because the parameters in these models cannot be
We limit our analysis in this section to the CBL model with mixed
1067
estimated with a sufficient precision which is needed for an im- CBL models outperform the benchmark portfolios for most out-of-
provement of portfolio performance. However, comparing the BL- sample measures. Higher portfolio returns and risk-adjusted per-
based portfolios with those from the CBL model, we can conclude formance are obtained from the CBL models with the risk-adjusted
that the suggested CBL approach results in lower portfolio down- equilibrium compared with models with the CAPM-based equilib-
side risk and volatility for both risk minimization and reward/risk rium. Based on a multi-period analysis, we show the advantages
maximization. (e.g., higher out-of-sample SR) of incorporating the asymmetric tail
dependence captured in both the Clayton and the mixed CBL mod-
5.2. U.S. stock market els.
The CBL approach can be extended in several aspects. First,
To investigate the performance of suggested models in another other vine copula models (e.g., canonical and drawable vines) can
financial market, we apply the CBL approach to a data set consist- be used with this approach. Second, because of its ability to model
ing of the 50 most capitalized U.S. stocks that are constituents of both lower and upper tail dependency, the CBL model can also be
the S&P 100 index. Similar to the Eurostoxx 50 index, we obtain applied to other reward/risk maximization problems with different
the historical constituents of S&P 100, and from this list, we in- risk and reward measures. The CBL model can also be combined
clude those stocks with the highest market capitalization and con- with multi-criteria portfolios where several risk and reward mea-
struct several portfolio strategies.11 Similar to Section 5.1, we only sures are included in a scenario-based optimization.
include CBL-based portfolios with mixed copula families.
Table 7 reports the results for the U.S. market. All the BL and
CBL portfolios achieve higher average returns than both the EQW
and historical-based strategies. Although the risk-adjusted equilib- Acknowledgment
rium leads to portfolios with lower volatility and downside risk,
the CAPM equilibrium generally increases the portfolio average
We want to thank Claudia Czado for her helpful comments.
return. The CBL-based Min CVaR portfolio (Panel E) shows the
We are thankful to the developers of R packages “rmgarch” and
lowest out-of-sample CVaR (3.83%). Comparing to their uncon-
“rvinecopulib”. In addition, we are grateful for the suggestions of
strained counterparts, all the tail-constrained portfolios result in
seminar participants at the 13th International Conference on Com-
lower volatility and downside risk, while achieving similar risk-
putational and Financial Econometrics (CFE 2019) in London, 3rd
adjusted ratios. Overall, these results indicate the advantages of (i)
Annual Workshop on Financial Econometrics organized by Örebro
incorporating copula modeling with the BL approach, (ii) applying
University School of Business, 2nd Annual Workshop on Emerg-
risk-adjusted equilibrium, and (iii) including tail constraints in the
ing Topics in Financial Economics organized by Linkping Univer-
reward/risk maximization, in achieving less volatile portfolios with
sity, 5th Econometric Research in Finance (ERFIN 2020) organized
less downside risk. For a comparison of the economic performance
by SGH Warsaw School of Economics. The usual disclaimer applies.
of the CBL portfolios (consisting of the S&P 100 components) with
the benchmarks, see Figures S7– S9 in the supplementary material.
6. Conclusions
Appendix A. Advancement of the data window over time
In this study, we suggest and evaluate a novel CBL approach
for modeling and forecasting portfolio returns. The CBL approach The estimation sample Ot length is fixed (e.g., 500 days), and
includes estimating the tail dependency structure and covariance when changing t, it is updated with a new information set (since
matrix from truncated Rvine copulas. Using investors’ views as this is a rolling window). This is the same for the out-of-sample
well as CAPM-based and risk-adjusted equilibrium models, we ob- holdout interval Ht (e.g., for daily re-balancing Ht = 1.). Below we
tain the BL posterior mean and covariance matrix. Then, using provide more clarifications:
standardized residuals from copula modeling, we simulate one- Tt points to the starting point of the current test interval con-
step ahead asset returns. We consider three vine copula models: sisting of the observation set Ot for the estimation and the hold-
Student-t, Clayton, and mixed versions. We then use the mBICV out set Ht for performance evaluation. Example: Ot , Ht = 50 time
criterion to select both the truncation level and the copula families points each, hence Tt = Ot + Ht = 10 0. Let t1 = 1.1.20 0 0 and initial-
(in the mixed version). Additionally, we impose tail constraints in ize t = t1 before step 1 (Section 2.8). Then the final date of the first
the reward/risk maximization to reduce downside risk. We apply observation interval is t + Ot − 1 = 1.1.20 0 0 + 50 − 1 = 19.2.20 0 0
the models to the 50 components of the Eurostoxx 50 index (also and that of the first holdout interval t + Ot + Ht − 1 = 1.1.20 0 0 +
S&P 100 for robustness check) and perform out-of-sample portfolio 100 − 1 = 9.4.2000 (not separating workdays from holidays in this
back-testing. example). Iterating steps 1 - 4 through the data sample with step
The results of the portfolio back-testing support using the sug- length t gives the first date of the last observation interval, for
gested models. For the daily data frequency, we find gains from example t2 = 1.6.2019. The final date of the last holdout interval at
t = t2 = 1.6.2019 then is t + Tt − 1 = 1.6.2019 + 100 − 1 = 8.9.2019.
incorporating copula modeling into the BL approach to maximize

the investor’s utility function, particularly for the Min CVaR and
Max SR portfolios. Most of the constrained optimal portfolios result
in lower risk (both volatility and downside risk). In general, the Supplementary material
11
Due to the computational intensity, we refrained to use all the 100 stocks of Supplementary material associated with this article can be
the S&P 100 as input for the various models. found, in the online version, at 10.1016/j.ejor.2021.06.015.
1068
References Green, R. C., & Hollifield, B. (1992). When will mean-variance efficient portfolios be
well diversified? The Journal of Finance, 47(5), 1785–1809.
Aas, K., Czado, C., Frigessi, A., & Bakken, H. (2009). Pair-copula constructions of mul- Haff, I. H., Aas, K., & Frigessi, A. (2010). On the simplified pair-copula construc-
tiple dependence. Insurance: Mathematics and Economics, 44(2), 182–198. tion simply useful or too simplistic? Journal of Multivariate Analysis, 101(5),
Acar, E. F., Genest, C., & NešLehová, J. (2012). Beyond simplified pair-copula con- 1296–1310.
structions. Journal of Multivariate Analysis, 110, 74–90. Haff, I. H., et al. (2013). Parameter estimation for pair-copula constructions.
Ahmadi-Javid, A., & Fallah-Tafti, M. (2019). Portfolio optimization with entropic val- Bernoulli, 19(2), 462–491.
ue-at-risk. European Journal of Operational Research, 279(1), 225–241. Harris, R. D., Stoja, E., & Tan, L. (2017). The dynamic Black–Litterman approach to
Al Janabi, M. A., Hernandez, J. A., Berger, T., & Nguyen, D. K. (2017). Multivariate de- asset allocation. European Journal of Operational Research, 259(3), 1085–1096.
pendence and portfolio optimization algorithms under illiquid market scenarios. He, G., & Litterman, R. (1999). The intuition behind Black-Litterman model portfolios.
European Journal of Operational Research, 259(3), 1121–1131. Goldman Sachs Investment Management Research.
Alexander, G. J., Baptista, A. M., & Yan, S. (2007). Mean–variance portfolio selection Heinen, A., Valdesogo, A., et al. (2009). Asymmetric CAPM dependence for large di-
with at-risk constraints and discrete distributions. Journal of Banking & Finance, mensions: the canonical vine autoregressive model. No. Universidad Carlos III de
31(12), 3761–3781. Madrid.
Baixauli-Soler, J. S., Alfaro-Cid, E., & Fernandez-Blanco, M. O. (2011). Mean-VaR port- Huang, D., Zhu, S., Fabozzi, F. J., & Fukushima, M. (2010). Portfolio selection under
folio selection under real constraints. Computational Economics, 37(2), 113–131. distributional uncertainty: A relative robust cvar approach. European Journal of
Beach, S. L., & Orlov, A. G. (2007). An application of the Black–Litterman model Operational Research, 203(1), 185–194.
with EGARCH-M-derived views for international portfolio management. Finan- Jobson, J. D., & Korkie, B. (1980). Estimation for Markowitz efficient portfolios. Jour-
cial Markets and Portfolio Management, 21(2), 147–166. nal of the American Statistical Association, 75(371), 544–554.
Bedford, T., & Cooke, R. M. (2001). Probability density decomposition for condition- Joe, H. (1996). Families of m-variate distributions with given margins and m (m-1)/2
ally dependent random variables modeled by vines. Annals of Mathematics and bivariate dependence parameters. In L. Rüschendorf, B. Schweizer, & M. E. Taylor
Artificial Intelligence, 32(1-4), 245–268. (Eds.), Distributions with fixed marginals and related topics (pp. 120–141). Hay-
Bedford, T., Cooke, R. M., et al. (2002). Vines–a new graphical model for dependent ward.
random variables. The Annals of Statistics, 30(4), 1031–1068. Joe, H. (2014). Dependence modeling with copulas. Chapman & Hall/CRC.
Bera, A. K., & Park, S. Y. (2008). Optimal portfolio diversification using the maximum Johansen, S. (1995). Likelihood-based inference in cointegrated vector autoregressive
entropy principle. Econometric Reviews, 27(4-6), 484–512. models. Oxford University Press, Oxford..
Best, M. J., & Grauer, R. R. (1991). On the sensitivity of mean-variance-efficient port- Killiches, M., Kraus, D., & Czado, C. (2017). Examination and visualisation of the
folios to changes in asset means: some analytical and computational results. The simplifying assumption for vine copulas in three dimensions. Australian & New
Review of Financial Studies, 4(2), 315–342. Zealand Journal of Statistics, 59(1), 95–117.
Black, F., & Litterman, R. (1991). Asset allocation: combining investor views with Kolm, P., & Ritter, G. (2017). On the Bayesian interpretation of Black–Litterman. Eu-
market equilibrium. Journal of Fixed Income, 1(1), 7–18. ropean Journal of Operational Research, 258(2), 564–572.
Black, F., & Litterman, R. (1992). Global portfolio optimization. Financial Analysts Kolm, P. N., Tütüncü, R., & Fabozzi, F. J. (2014). 60 years of portfolio optimization:
Journal, 48, 28–43. Practical challenges and current trends. European Journal of Operational Research,
Bollerslev, T., Engle, R. F., & Wooldridge, J. M. (1988). A capital asset pricing model 234(2), 356–371.
with time-varying covariances. Journal of Political Economy, 96(1), 116–131. Kuo, C.-Y. (2016). Does the vector error correction model perform better than others
Boubaker, H., & Sghaier, N. (2013). Portfolio optimization in the presence of depen- in forecasting stock price? An application of residual income valuation theory.
dent financial returns with long memory: A copula based approach. Journal of Economic Modelling, 52, 772–789.
Banking & Finance, 37(2), 361–377. Kurowicka, D. (2011). Optimal truncation of vines. In D. Kurowicka, & H. Joe (Eds.),
Brechmann, E. C., Czado, C., & Aas, K. (2012). Truncated regular vines in high di- Dependence modeling: Vine copula handbook. World Scientific Publishing Co..
mensions with application to financial data. Canadian Journal of Statistics, 40(1), Levy, H., & Levy, M. (2014). The benefits of differential variance-based constraints
68–85. in portfolio optimization. European Journal of Operational Research, 234(2),
Brechmann, E. C., & Joe, H. (2015). Truncation of vine copulas using fit indices. Jour- 372–381.
nal of Multivariate Analysis, 138, 19–33. Low, R. K. Y., Alcock, J., Faff, R., & Brailsford, T. (2013). Canonical vine copulas in the
Charnes, A., & Cooper, W. W. (1962). Programming with linear fractional functionals. context of modern portfolio management: Are they worth it? Journal of Banking
Naval Research Logistics Quarterly, 9(3-4), 181–186. & Finance, 37(8), 3085–3099.
Cheung, Y.-L., Cheung, Y.-W., & Wan, A. T. (2009). A high–low model of daily stock Lwin, K. T., Qu, R., & MacCarthy, B. L. (2017). Mean-VaR portfolio optimization:
price ranges. Journal of Forecasting, 28(2), 103–119. A nonparametric approach. European Journal of Operational Research, 260(2),
Chopra, V. K., & Ziemba, W. T. (1993). The effect of errors in means, variances, and 751–766.
covariances on optimal portfolio choice. The Journal of Portfolio Management, Markowitz, H. (1952). Portfolio selection. The Journal of Finance, 7(1), 77–91.
19(2), 6–11. Merton, R. C. (1969). Lifetime portfolio selection under uncertainty: The continu-
Clarke, R., De Silva, H., & Thorley, S. (2002). Portfolio constraints and the fundamen- ous-time case. The Review of Economics and Statistics, 247–257.
tal law of active management. Financial Analysts Journal, 58(5), 48–66. Meucci, A. (2006). Beyond Black-Litterman in practice: A five-step recipe to input
Consigli, G. (2002). Tail estimation and mean–VaR portfolio selection in markets views on non-normal markets. Risk, 19, 114–119.
subject to financial instability. Journal of Banking & Finance, 26(7), 1355–1382. Meucci, A. (2008). Fully flexible views: Theory and practice. Risk, 21, 97–102.
Cui, X., Gao, J., Shi, Y., & Zhu, S. (2019). Time-consistent and self-coordination strate- Meucci, A. (2009). Risk and asset allocation. Springer Science & Business Media.
gies for multi-period mean-Conditional Value-at-Risk portfolio selection. Euro- Meucci, A. (2010). The Black-Litterman approach: Original model and extensions. The
pean Journal of Operational Research, 276(2), 781–789. Encyclopedia of Quantitative Finance. Wiley.
Czado, C. (2019). Analyzing dependent data with vine copulas. A practical guide Michaud, R. O. (1989). The Markowitz optimization enigma: Is optimized optimal?
with R. Lecture Notes in Statistics. Cham: Springer. Financial Analysts Journal, 45(1), 31–42.
De Roon, F. A., Nijman, T. E., & Werker, B. J. (2001). Testing for mean-variance span- Nagler, T., Bumann, C., & Czado, C. (2019). Model selection in sparse high-dimen-
ning with short sales constraints and transaction costs: The case of emerging sional vine copula models with an application to portfolio risk. Journal of Mul-
markets. The Journal of Finance, 56(2), 721–742. tivariate Analysis, 172, 180–192.
Deng, Q. (2018). A generalized VECM/VAR-DCC/ADCC framework and its application Okimoto, T. (2014). Asymmetric increasing trends in dependence in international
in the Black-Litterman model: Illustrated with a China portfolio. China Finance equity markets. Journal of Banking & Finance, 46, 219–232.
Review International, 8(4), 453–467. Östermark, R. (1991). Vector forecasting and dynamic portfolio selection: Empiri-
Dinkelbach, W. (1967). On nonlinear fractional programming. Management Science, cal efficiency of recursive multiperiod strategies. European Journal of Operational
13(7), 492–498. Research, 55(1), 46–56.
Embrechts, P., Höing, A., & Juri, A. (2003). Using copulae to bound the Value-at-Risk Östermark, R. (2001). Multivariate cointegration analysis of the Finnish Japanese
for functions of dependent risks. Finance and Stochastics, 7(2), 145–167. stock markets. European Journal of Operational Research, 134(3), 498–507.
Engle, R. F. (1982). Autoregressive conditional heteroscedasticity with estimates of Östermark, R. (2017). Massively parallel processing of recursive multi-period portfo-
the variance of United Kingdom inflation. Econometrica, 50, 987–1007. lio models. European Journal of Operational Research, 259(1), 344–366.
Engle, R. F., & Granger, C. W. (1987). Co-integration and error correction: represen- Palczewski, A., & Palczewski, J. (2019). Black–Litterman model for continuous distri-
tation, estimation, and testing. Econometrica: Journal of the Econometric Society, butions. European Journal of Operational Research, 273(2), 708–720.
251–276. Palomba, G. (2008). Multivariate GARCH models and the Black-Litterman approach
Fabozzi, F. J., Focardi, S. M., & Kolm, P. N. (2006). Incorporating trading strategies in for tracking error constrained portfolios: an empirical analysis. Global Business
the Black-Litterman framework. The Journal of Trading, 1(2), 28–37. and Economics Review, 10(4), 379–413.
Fulga, C. (2016). Portfolio optimization with disutility-based risk measure. European Pang, T., & Karan, C. (2018). A closed-form solution of the Black–Litterman model
Journal of Operational Research, 251(2), 541–553. with conditional value at risk. Operations Research Letters, 46(1), 103–108.
Geyer, A., & Lucivjanská, K. (2016). The Black Litterman approach and views from Patton, A. J. (2004). On the out-of-sample importance of skewness and asymmetric
predictive regressions: Theory and implementation. The Journal of Portfolio Man- dependence for asset allocation. Journal of Financial Econometrics, 2(1), 130–168.
agement, 42(4), 38–48. Pfaff, B. (2008). VAR, SVAR and SVEC models: Implementation within R package
Giacometti, R., Bertocchi, M., Rachev, S. T., & Fabozzi, F. J. (2007). Stable distributions vars. Journal of Statistical Software, 27(4), 1–32.
in the Black-Litterman approach to asset allocation. Quantitative Finance, 7(4), Pfaff, B. (2016). Financial risk modelling and portfolio optimization with R. John Wiley
423–433. & Sons.
Grauer, R. R., & Shen, F. C. (20 0 0). Do constraints improve portfolio performance? Rockafellar, R. T., & Uryasev, S. (20 0 0). Optimization of conditional value-at-risk.
Journal of Banking & Finance, 24(8), 1253–1274. Journal of Risk, 2, 21–42.
1069
Sahamkhadam, M., Stephan, A., & Östermark, R. (2018). Portfolio optimization based Sklar, M. (1959). Fonctions de répartition à n dimensions et leurs marges. Publica-
on GARCH-EVT-copula forecasting models. International Journal of Forecasting, tions de l’Institut de Statistique de l’Université de Paris, 8, 229–231.
34(3), 497–506. Stoyanov, S. V., Rachev, S. T., & Fabozzi, F. J. (2007). Optimal financial portfolios.
Sharpe, W. F. (1966). Mutual fund performance. The Journal of Business, 39(1), Applied Mathematical Finance, 14(5), 401–436.
119–138. Zhou, G. (2009). Beyond Black Litterman: Letting the data speak. The Journal of Port-
Sharpe, W. F. (1994). The sharpe ratio. Journal of Portfolio Management, 21(1), 49–58. folio Management, 36(1), 36–45.
Silva, T., Pinheiro, P. R., & Poggi, M. (2017). A more human-like portfolio optimiza-
tion approach. European Journal of Operational Research, 256(1), 252–260.
1070

Copula-Based Black Litterman Portfolio Optimizaion

Uploaded by

Copyright:

Available Formats

Copula-Based Black Litterman Portfolio Optimizaion

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Copula-Based Black Litterman Portfolio Optimizaion

Uploaded by

Copyright:

Available Formats

European Journal of Operational Research 297 (2022) 1055–1070

Contents lists available at ScienceDirect

European Journal of Operational Research

Innovative Applications of O.R.

Copula-based Black–Litterman portfolio optimization

1. Introduction One of the challenges faced by fund managers is to combine

Sklar (1959) showed how to obtain the d-dimensional copula C (15)

where z j = Fj−1 u j , u j ∼ U [0, 1]d , ∀ j ∈ {1, 2, . . . , d}. If all the (16)

2.7. Portfolio Optimization Methods ˆ t , rˆt ) − l]+ = max{[f(w

According to Markowitz (1952), for a d-dimensional portfo- VaRα (w

where UCVaR is the maximum acceptable value of the downside

Panel A: Benchmark Portfolios

Portfolio 20 0 0–20 06 20 07–20 09 2010–2019 2020

Panel A: Benchmark Portfolios

Portfolio 20 0 0-20 06 20 07-20 09 2010-2019 2020

Panel A: Benchmark Portfolios

based portfolios, with the CBL-based strategies shows the gain

Panel A: Benchmark Portfolios

Panel A: Benchmark Portfolios

incorporating copula modeling into the BL approach to maximize

You might also like