International Journal of Forecasting 34 (2018) 497–506

Contents lists available at ScienceDirect

International Journal of Forecasting

journal homepage: www.elsevier.com/locate/ijforecast

Portfolio optimization based on GARCH-EVT-Copula

forecasting models
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

article info a b s t r a c t
Keywords: This study uses GARCH-EVT-copula and ARMA-GARCH-EVT-copula models to perform
GARCH models out-of-sample forecasts and simulate one-day-ahead returns for ten stock indexes. We
Extreme value theory construct optimal portfolios based on the global minimum variance (GMV), minimum
Copula models
conditional value-at-risk (Min-CVaR) and certainty equivalence tangency (CET) criteria,
Conditional value-at-risk
and model the dependence structure between stock market returns by employing elliptical
Portfolio optimization
(Student-t and Gaussian) and Archimedean (Clayton, Frank and Gumbel) copulas. We
analyze the performances of 288 risk modeling portfolio strategies using out-of-sample
back-testing. Our main finding is that the CET portfolio, based on ARMA-GARCH-EVT-
copula forecasts, outperforms the benchmark portfolio based on historical returns. The
regression analyses show that GARCH-EVT forecasting models, which use Gaussian or
Student-t copulas, are best at reducing the portfolio risk.
© 2018 International Institute of Forecasters. Published by Elsevier B.V. All rights reserved.

1. Introduction improvement on the traditional GARCH volatility mod-

els. In general, the downside risk can be measured by
A range of different portfolio optimization methods, the Value-at-Risk (VaR) and Conditional Value-at-Risk
generally consisting of two steps (Markowitz, 1952), have (CVaR) (Gençay & Selçuk, 2004). The latter measures the
been proposed over the last few decades. The first step expected loss at a specific confidence level. GARCH models
involves forecasting the future returns of the underlying can be employed for forecasting asset returns (Crato &
assets. One of the models that has been used for this step is Ruiz, 2012), though it is recommended that forecasters
the generalized autoregressive conditional heteroscedas- take into account the fact that the financial assets (e.g. stock
ticity (GARCH) extreme value theory (EVT) copula (Longin, prices, interest rates and exchange rates) possess the fea-
1996). The second step consists of optimal portfolio alloca- tures of fat-tailed distributions (Bondt & Thaler, 1985; Da-
tion, which is achieved by defining each asset’s weight in corogna & Pictet, 1997; Harmantzis, Miao, & Chien, 2006).
EVT can be used to estimate the tail behavior of the re-
the corresponding portfolio. There are three main methods
turns of these financial assets, and the volatility has been
that are used for this step: Min-CVaR (minimizing the
shown to be an important factor in extreme value fore-
conditional Value-at-Risk), GMV (minimizing the variance)
casting models (Gençay & Selçuk, 2006; Gençay, Selçuk,
and CET (maximizing the Sharpe ratio).
& Ulugülyaci, 2003). A combination of GARCH and EVT
GARCH-EVT-copula models are used mainly for min-
models has been proposed and used in several previous
imizing the downside risk, and are considered to be an
studies (Bhattacharyya & Ritolia, 2008; Chan & Gray, 2006;
Deng, Ma, & Yang, 2011; McNeil & Frey, 2000). The basic
idea behind this combination is that EVT is suitable for the
* Corresponding author.
E-mail address: maziar.sahamkhadam@lnu.se (M. Sahamkhadam). independently and identically distributed series that can

https://doi.org/10.1016/j.ijforecast.2018.02.004
0169-2070/© 2018 International Institute of Forecasters. Published by Elsevier B.V. All rights reserved.
498 M. Sahamkhadam et al. / International Journal of Forecasting 34 (2018) 497–506

be derived from the residuals obtained from GARCH mod- each model during the financial crisis and post-crisis peri-
els. ods. Their results indicate that daily and weekly strategies
Furthermore, one of the characteristics of financial time based on GARCH-EVT-copula models outperform others
series is that movements in one series can affect those in with respect to both the annual average return and the
other series (Christoffersen, Errunza, Jacobs, & Jin, 2014). Sharpe ratio. Another study performed by Low, Alcock, Faff,
This provides researchers with a vast array of choices when and Brailsford (2013) employed a Clayton C-vine copula for
modeling the dependency structures between different testing the portfolio performance (accumulation wealth)
types of financial assets, such as stock markets, exchange based on minimizing the conditional value-at-risk. Two
rates, commodities and so forth. Of the various choices, portfolio sizes are considered, namely 12 and 3 dimensions.
copula models have shown adequate practicality, perhaps They find that the use of this model is more appropriate for
because of the technique employed in copula models. In the portfolio with more assets.
this technique, originally shown by Sklar (1959), the corre- Following Huang and Hsu (2015) and Wang, Chen et al.
lations between assets are obtained from the joint distribu- (2010), we use the GARCH-EVT-copula and ARMA-GARCH-
tion and then used in each separate marginal distribution. EVT-copula models for performing out-of-sample forecast-
The application of GARCH-EVT-copula models to risk ing and portfolio allocation. We begin by applying rolling
management has been the focus of several previous stud- window estimation and using univariate GARCH(1,1) and
ies. For instance, a study performed by Wang, Chen, Jin, and ARMA(1,1)-GARCH(1,1) models separately to obtain the
Zhou (2010) used three copula models to estimate the in- parameters for one-day-ahead forecasts, then use EVT for
vestment risk of foreign currencies. Of these copulas, both tail modeling and for obtaining the uniforms. As for the de-
the Clayton and Student-t copulas yield better estimates of pendency structure, we consider five d-dimensional cop-
the correlation between exchange rates than the Gaussian ula models, namely Student-t, Gaussian, Clayton, Frank
copula. Berger (2013) forecasts the portfolio risk (VaR) by and Gumbel. Finally, we use the parameters from ARMA-
employing a time-varying dynamic conditional correlation GARCH (or GARCH) models and the dependency structures
(DCC) copula in combination with EVT, which leads to a from copula models to simulate one-day-ahead returns.
better estimation of VaR than a static copula approach. We capture the performances of these models fully by
In addition, a semi-parametric GARCH-EVT-copula model applying them to stock markets, and also evaluate different
re-balancing frequencies.
was proposed by Koliai (2016), who performed stress tests
Our work contributes to the existing literature in several
on corresponding portfolios and found that different mod-
ways. First, the addition of CET and GMV portfolios allows
els affect stress scenarios in various different ways.
us to present novel results. As was mentioned earlier, the
In another study, based on the DCCGARCH-copula (but
GARCH-EVT-copula and ARMA-GARCH-EVT-copula mod-
without EVT), two dynamic robust portfolio optimization
els are used in the first step of portfolio optimization by
methods are considered and compared with non-robust
forecasting future returns and volatilities. In contrast to
portfolios (Han, Li, & Xia, 2017). Several more complex cop-
previous studies, we use the simulated returns from the
ula models, including the regular vine (R-vine), canonical
copula model not only in the Min-CVaR, but also for the
vine (C-vine) and drawable vine (D-vine), are examined
CET and GMV portfolios. Moreover, the aim of our study
by Zhang, Wei, Yu, Lai, and Peng (2014), who use these cop-
is to answer the question of how much of a reduction in
ulas to forecast both VaR and CVaR. The authors conclude
risk (or gain in returns) can be achieved by combining
that D-vine is better than the other vine copulas in terms of
the different forecasting models with EVT and the above-
forecasting the CVaR. Bhatti and Nguyen (2012) suggest the
mentioned portfolio optimization techniques. To the best
use of a conditional EVT and time-varying copula for mod- of our knowledge, this is the only study to do so by es-
eling the tail dependency between stock markets. Wang, timating the contribution of each part in risk modeling
Jin, and Zhou (2010) use the GARCH-EVT-copula model to using a regression analysis based on the sample of portfolio
evaluate the risk of foreign currencies. They concentrate returns obtained from back-testing. We find that almost all
on Student-t, Gaussian and Clayton copulas, and, in cor- of the risk models decrease the portfolio risk significantly.
respondence with Huang, Lee, Liang, and Lin (2009), who Overall, GARCH forecasting models in combination with
implement GARCH-copula and GJRGARCH-copula models EVT and copula models appear to be particularly suitable
for estimating the VaR of a portfolio of stock indexes, point for the CET optimization framework.
out that the Student-t provides better estimation of the The rest of the paper is organized as follows. Section 2
VaR when compared with other copula models. presents the methodology, including the ARMA-GARCH
Most studies of the GARCH-EVT-copula model have fo- model, extreme value theory, copula models and opti-
cused on its application to forecasting and the examination mization methods. The data set is described in Section 3.
of the resulting downside risk. However, some studies have Section 4 provides details of our empirical results. We
used these models not only for risk modeling, but also for describe the robustness analysis in Section 5. Finally, con-
portfolio allocation and back-testing. For instance, Huang cluding remarks are provided in Section 6.
and Hsu (2015) consider two GARCH-EVT-copula models
(based on Student-t and Gaussian copulas) for simulating 2. Methodology
future returns of stock markets, and use a rolling window
to compute the optimal weights based on the Min-CVaR The GARCH-EVT-copula approach assumes that the re-
allocation for the out-of-sample period. In addition, they turns are ergodic processes (Boltzmann, 1896) and the
also conduct four re-balancing strategies (daily, weekly, residuals are independently identically distributed (i.i.d.)
biweekly and monthly) and evaluate the performance of random variables.
 M. Sahamkhadam et al. / International Journal of Forecasting 34 (2018) 497–506 499

2.1. ARMA-GARCH forecasting model Table 1

The parameter θ and generator ψ (uj ) for Archimedean copulas.

In the ARMA-GARCH approach, the mean equation is Copula θ ψ (uj ), j = 1, 2, . . . , d

1
driven by the recursive volatility process. In line with pre- Clayton θ ∈ (0, ∞) (1 + uj )− θ
vious studies, we assume that the conditional mean µjt Frank θ ∈ (0, ∞) − log(1 − (1 − e−θ ) exp(−uj ))/θ
1
follows an ARMA(1,1) process and that the conditional Gumbel θ ∈ (1, ∞) exp(−ujθ )
variance hjt follows a GARCH(1,1) process (Kim & Jung,
2016; Pircalabu, Hvolby, Jung, & Høg, 2017; Righi, Schlen-
der, & Ceretta, 2015). The ARMA(1,1)-GARCH(1,1) can be there exists a d-dimensional copula C such that:
modeled as:
rjt = µjt + ϕj (rj,t −1 − µjt ) + θi ϵj,t −1 + ϵjt ∀z ∈ ℜd : F (z1 , z2 , . . . , zd )
⎧
⎪
= C (F1 (z1 ), F2 (z2 ), . . . , Fd (zd )) = C (u1 , u2 , . . . , ud ).
⎪
(4)
ϵjt = zjt hjt
⎪
⎨ √
(1)
⎪
⎪zjt ≈ (i.i.d.) Note that the copula parameter vector is not necessary
⎩h = ω + α ϵ 2 + β h
j j,t −1 ,
for the proof of Sklar’s theorem, and is therefore sup-
⎪
jt j i j,t −1
pressed here. The parameter vector will be introduced be-
where rjt denotes the actual returns for asset j = 1, 2, . . . , low. If all margins are continuous, then C is unique, and the
d, zjt is the standardized residuals, and the parameter re- converse is also true; i.e., if C is a d-dimensional copula and
strictions are ωj > 0, αj ≥ 0, βj ≥ 0, αj + βj < 1, F1 , F2 , . . . , Fd are distribution functions, then F is a mul-
ϕj + θj ̸= 0. On the other hand, the autoregressive part of tivariate distribution function with margins F1 , F2 , . . . , Fd .
the mean equation can be replaced by a constant, in which Note that zj = Fj−1 (uj ), uj ∈ [0, 1], ∀j and therefore, by
case the univariate GARCH(1,1) is defined as: Eq. (4),
rjt = µjt + ϵjt
⎧
⎪
⎪ C (u1 , u2 , . . . , ud ) = F (F1−1 (u1 ), F2−1 (u2 ), . . . , Fd−1 (ud ))
ϵjt = zjt hjt
⎪
⎨ √
(2) = F (z1 , z2 , . . . , zd ). (5)
⎪zjt ≈ (i.i.d.)
Using Eq. (5) and the notation F = φ for the multivariate
⎪ ∑
⎩h = ω + α ϵ 2 + β h
j j,t −1 ,
⎪
jt j j j,t −1
standard normal distribution and Fj (uj ) = φ (uj ), j =
where rjt denotes the actual returns for asset j = 1, 2, . . . , 1, 2, . . . , d, for the univariate standard normal distribution,
d, and zjt is the standardized residuals with ωj > 0, αj ≥ we can write the (standardized) Gaussian copula as fol-
0, βj ≥ 0, αj + βj < 1. Note that we formulate the lows:
two ARMA(1,1)-GARCH(1,1) and GARCH(1,1) models here
because we are interested in comparing forecasting models C Gaussian (u1 , u2 , . . . , ud )
that are based on a ARMA(1,1) term in the mean equation = φ∑ (φ −1 (u1 ), φ −1 (u2 ), . . . , φ −1 (ud ))
with those that merely consider a constant mean. = φ∑ (z1 , z2 , . . . , zd ), (6)
where φ denotes
∑
∑ the standard multivariate normal dis-
2.2. Tail behavior
tribution, and and φ −1 are the correlation matrix and
the inverse of the normal cumulative distribution function,
The tail behavior of the asset returns can be modeled
respectively.
via extreme value theory. The method is known as peak
If we consider tν ∑ as the standard
∑ multivariate t dis-
over threshold (POT), and, in combination with GARCH(1,1)
tribution with correlation matrix , then the Student-t
or ARMA(1,1)-GARCH(1,1) models, the jth marginal dis-
copula can be written analogously as:
tribution function Gj (zj ) can be obtained based on two
distributions, namely the generalized Pareto (GPD) for the C t (u1 , u2 , . . . , ud )
upper and lower tail and the Gaussian kernel for the middle
part:
= tν ∑ (tν−1 (u1 ), tν−1 (u2 ), . . . , tν−1 (ud ))
⎧ = tν ∑ (z1 , z2 , . . . , zd ), (7)
N L uL − zj − 1L
⎪ u {1 + ξ L } ξ , z j < − uL
⎪
where ν and tν−1 are degrees of freedom and the inverse
⎪
⎨ N
⎪
⎪ βL
of the t cumulative distribution function respectively. In
Gj (zj ) = φ (zj ), uL < zj < uR (3)
⎪ the case of Archimedean copulas, we can estimate the
N R uR − zj − 1R parameter θ as follows:
⎪
⎪
⎩1 − u {1 + ξ L } ξ , zj > uR ,
⎪
⎪
N βR
C Archimedean (u1 , u2 , . . . , ud )
where ξ , β , u and u denote the shape, scale, and upper
R L
= ψ (ψ −1 (u1 ), ψ −1 (u2 ), . . . , ψ −1 (ud ))
and lower thresholds, respectively.
= ψ (z1 , z2 , . . . , zd ), (8)
2.3. Copula models where ψ is the generator of the Archimedean copula and
ψ −1 is the inverse of the generator. Table 1 gives the
Sklar (1959) proved that, for any d-dimensional distri- parameter θ and the generator ψ for Archimedean copu-
bution function F with marginal distributions F1 , . . . , Fd , las (Hofert, Maechler, & McNeil, 2012).
500 M. Sahamkhadam et al. / International Journal of Forecasting 34 (2018) 497–506

2.4. Maximum likelihood estimation on Markowitz’s mean-variance portfolio characterization,

there is an enormous body of literature focusing on im-
Since the copula in Eq. (4) is a distribution function, it is proving the portfolio allocation methods. One important
a multiple integral in uj ∈ [0, 1], ∀j: extension is the maximization of the Sharpe ratio, as sug-
gested by Sharpe (1963, 1994). The Sharpe ratio (SR) is
C (u1 , u2 , . . . , ud ) defined as the expected return of the portfolio over its stan-
uj ud
∂ d C (z1 , . . . , zd )
∫ ∫
dard deviation. For a d-dimensional portfolio with asset
= ... dz1 , . . . , dzd
returns rt = (rt1 , rt2 , . . . , rtd ), j = 1, . . . , d, asset weights
∫0 uj ∫0 ud ∂ z1 , . . . , ∂ zd
wt = (wt1 , wt2 , . . . , wtd ) and a d × d covariance matrix
= ... c(z1 , . . . , zd )dz1 , . . . , dzd . (9) Σ at time t, the portfolio expected return and variance are
0 0
wTt rt and wTt Σ wt , respectively. The CET portfolio can be
The copula density in the interior (u1 , . . . , ud )T ∈ ]0, 1[d is expressed as:
defined as: wTt rt
maximize Sharpe ratio
∂ d C (u1 , . . . , ud ) wt
√
wTt Σ wt
c(u1 , . . . , ud ) = . (10)
∂ u1 , . . . , ∂ ud subject to wTt 1 = 1, full investment (15)
Introducing the copula parameter as the generic vector Ω , ∀j ∈ {1, 2, . . . , d} :
we can write the multivariate distribution function Eq. (4) wt ,j ≥ 0, long positions only.
as:
Another portfolio optimization method is that proposed
∀z ∈ ℜd : F (z1 , z2 , . . . , zd ) by Merton (1980), who showed that the variance and co-
= C (F1 (z1 ), F2 (z2 ), . . . , Fd (zd )|Ω ) variance estimates from the asset returns are reliable when
= C (u1 , u2 , . . . , ud |Ω ), (11) considering optimization. The GMV portfolio is based on
minimizing the variance of the portfolio wTt Σ wt , which is
and the density function is obtained by differentiating the ideal for risk-averse investors who are not interested in
copula as: maximizing the expected return (Bodnar, Mazur, & Okhrin,
∂ d C (u1 , . . . , ud |Ω ) 2017). In this case, the GMV can be solved as:
c(u1 , . . . , ud |Ω ) = . (12)
∂ u1 , . . . , ∂ ud minimize wTt Σ wt variance
wt
Differentiating Eq. (11) and using Eq. (12) yields Sklar’s subject to wTt 1 = 1, full investment (16)
theorem in terms of density functions: ∀j ∈ {1, 2, . . . , d} :
∂ d F (z1 , . . . , zd ) wt ,j ≥ 0, long positions only.
∀z ∈ ℜd :
∂ z1 , . . . , ∂ zd On the other hand, downside risk is another risk mea-
∂ d C (F1 (z1 ), . . . , Fd (zd )|Ω ) sure that has been considered in many studies regarding
= f (z1 , . . . , zd ) = = optimal portfolio allocation. For instance, VaR has been
∂ z1 , . . . , ∂ zd
d employed as the risk measure in combination with the
classic Markowitz’s mean-variance portfolio, leading to
∏
c(F1 (z1 ), F2 (z2 ), . . . , Fd (zd )|Ω ) fj (zj ), (13)
an optimal mean-VaR portfolio (Consigli, 2002). However,
j=1
CVaR is an alternative when the objective is to minimize
where fj denotes the derivative of the distribution function losses beyond the VaR threshold (Xu, Zhou, Jiang, Yu, & Niu,
Fj w.r.t zj ; i.e., fj is the jth density function. Using Sklar’s 2016). Based on the algorithm proposed by Rockafellar and
theorem in terms of density functions (Eq. (13)) gives the Uryasev (2000), the integral form for CVaR is:
log-likelihood function: 1
∫
CVaRβ (wt ) = f (wt , rt )p(rt )drt , (17)
l(z1 , z2 , . . . , zd ) 1−β f (wt ,rt )≥αβ (wt )
n d
∑ ( ∏ ) where f (wt , rt ) is the loss function for asset weights wt and
= log c(F1 (zi1 ), F2 (zi2 ), . . . , Fd (zid )|Ω ) fj (zij ) = p(rt ) is the probability of rt at time t. The Min-CVaR portfo-
i=1 j=1 lio can be solved by linear programming (see Rockafellar &
n [
∑ Uryasev, 2000, 2002, for more details). The use of Monte
log(c(F1 (zi1 ), F2 (zi2 ), . . . , Fd (zid )|Ω )) Carlo integration allows the CVaR to be minimized for a
i=1 given level of return (rt ):
d
1
∑ ]
+ log(fj (zij )) . (14) minimize fα (wt , β ) = α +
(wt ,α ) q(1 − β )
j=1 q
∑
× [−wTt rkt − α]+ conditional
k=1
2.5. Allocation methods
Value-at-Risk (18)
Markowitz (1952) considered two characteristics in subject to wTt 1 = 1, full investment
portfolio optimization, namely the expected return and ∀j ∈ {1, 2, . . . , d} :
the risk measure. In Markowitz’s optimal portfolio, the wt ,j ≥ 0, long positions only
risk measure is considered to be the variance. Based µ(wt ) ≤ −R,
 M. Sahamkhadam et al. / International Journal of Forecasting 34 (2018) 497–506 501

where q is the number of Monte Carlo draws, α is the VaR in order to estimate new standardized residuals for
at the β threshold value (here, 0.9), and rkt is the kth vector use in portfolio optimization at time t:
of the simulated returns at time t. From the optimization,
wt is the vector of asset weights that minimizes the CVaR Ẑ = [ẑi ] = (ẑi1 , ẑi2 , . . . , ẑid )
for a given R. = (F̂1−1 (ûi1 ), F̂2−1 (ûi2 ), . . . , F̂d−1 (ûid )),
i = 1, . . . , M . (23)
2.6. Steps
6. Substitute the estimated residual vectors Ẑ from
In this section, we begin by defining the following pa- step 5 in the estimated [ARMA]-GARCH forecasting
rameters: L = window length, t0 = starting point for the model in Eqs. (1) or (2) and generate M one-step
iterations, T = length of the out-of-sample interval, and forecasts for time point t = t0 + 1:
M = number of drawings from the multivariate return
distribution estimated for each iteration [t0 + 1, T ]. We use r̂it = (r̂ti,1 , r̂ti,2 , . . . , r̂ti,d ), i = 1, . . . , M . (24)
a rolling window estimation for the out-of-sample period,
7. Substitute the forecasts for returns in the opti- r̂it
in which we iterate steps 1 to 8 for each time point in
mization methods explained in Section 2.5 to get
[t0 + 1, T ], meaning that we reestimate all of the param-
the optimal weights for the CET, Min-CVaR and GMV
eters at each iteration. The first set of observations, needed
portfolios.
for the one-period out-of-sample prediction at t0 + 1, is
8. Use the optimal weights wt together with the real
[t0 − L + 1, t0 ], and the next set is [t0 − L + 2, t0 + 1]. We
asset returns rt to compute an
∑destimate of the one-
initialize t1 = t0 . The whole procedure can be summarized
as follows. period portfolio return Rt = j=1 wtj r tj = wT
t r t .

1. Use Eq. (1) for ARMA(1,1)-GARCH(1,1) or Eq. (2) for

3. Data
GARCH(1,1) to estimate the parameters by maxi-
mum likelihood estimation (MLE), and obtain stan-
Our data set includes daily adjusted prices of ten stock
dardized residuals:
indexes (S&P 500, FTSE 100, DAX 30, EURO STOXX, MSCI
x̂t = (x̂t1 , x̂t2 , . . . , x̂td ), x̂tj ≈ i.i.d., ∀j, World, CAC 40, OMXC 20, OMXH, OMXS 30 and TOPIX)
obtained from Thomson Reuters’ Datastream. We define
t ∈ [t0 − L + 1, t0 ]. (19)
logarithmic returns. The sample period starts in August
2. Use the estimated standardized residual vector xˆt 1996 and ends in August 2016, giving a total of 5218
from step 1 to estimate the center of the distribu- observations. Similarly to Huang and Hsu (2015) and Low
tions as a Gaussian kernel and the upper/lower tails et al. (2013), we implement a rolling window of 1260
as a GPD (Eq. (3)): observations.1 Accordingly, the out-of-sample forecasting
is performed 3958 times.
v̂tj = F̂k (x̂tj ), t ∈ [t0 − L + 1, t0 ], j ∈ [1, d], Table 2 provides descriptive statistics for the returns
v̂tj ∼ U(0, 1). (20) of stock indexes. All series have positive average returns
except for TOPIX. Only OMXS30 has a positive skewness.
3. Apply Sklar’s theorem (see Eq. (4)) and insert the The positive kurtosis seen for all series indicates that they
estimated uniforms v̂tj , j ∈ [1, d] of step 2 in the are all more peaked than a standard normal distribution.
chosen multivariate copula, then estimate the pa- Finally, the results of Jarque–Bera’s test reject the null
rameter Ω of the distribution function using MLE hypothesis of normality for all series.
(see Eq. (14)). For the elliptical copulas∑
(i.e., Gaussian
(Eq. (6)) and Student-t (Eq. (7))), Ω = , and for the 4. Results
Archimedean copulas (Eq. (8)), Ω = θ :
We employ GARCH-EVT and ARMA-GARCH-EVT models
F̂ (v̂t1 , v̂t2 , . . . , v̂td ) to define the marginal distribution for the innovations. We
= Ĉ (F̂1 (v̂t1 ), F̂2 (v̂t2 ), . . . , F̂d (v̂td )|Ω̂ ). (21) assume that the conditional distribution for residuals in
the GARCH(1,1) is Student-t. However, for the ARMA(1,1)-
4. Generate M uniform random numbers (wi1 , wi2 , . . . , GARCH(1,1), we use a normal distribution as the condi-
wid ), i = 1, . . . , M for each series j = 1, . . . , d tional distribution.2 The parameter estimation results are
and insert them into the estimated empirical mul- given in Table 3. The estimated parameters are significant
tivariate copula distribution of step 3 to obtain M and suggest that the error terms follow a t distribution.
uniforms with the estimated dependency structure: We use extreme value theory to fit GPD to the upper
and lower tails of the standardized residuals. The method
ûi = (ûi1 , ûi2 , . . . , ûid ) = Ĉ (F̂1 (wi1 ),
1 We select 1260 observations as the window length because we need
F̂2 (wi2 ), . . . , F̂d (wid )|Ω̂ ), ûij ∼ U(0, 1). (22) enough observations for each series in order to fit the semi-parametric
distribution to the residuals and obtain the lower and upper thresholds.
5. Insert the simulated uniform random numbers ûi = 2 Our intent is to examine the effects of changing the conditional
ûi1 , ûi2 , . . . , ûid , i = 1, . . . , M, of step 4 with the esti- distribution in GARCH model specifications. However, we did not test
mated dependency structure into the inverse of the for other distributions, such as the skewed normal or skewed Student-t,
estimated marginal distribution functions of step 3 because this would add too many results for the current paper.
502 M. Sahamkhadam et al. / International Journal of Forecasting 34 (2018) 497–506

Table 2
Descriptive statistics for the daily log returns of stock indexes.
Series Mean Std. dev. Skewness Kurtosis JB
S&P 500 0.023 0.012 −0.23 7.99 13 941***
FTSE 100 0.011 0.012 −0.15 5.63 6 926***
DAX 30 0.027 0.015 −0.14 4.05 3 583***
EURO STOXX 0.012 0.015 −0.09 4.22 3 876***
MSCI 0.015 0.010 −0.38 7.33 11 819***
CAC 40 0.015 0.015 −0.06 4.45 4 309***
OMXC 20 0.040 0.013 −0.28 5.13 5 786***
OMXH 0.026 0.018 −0.37 7.17 11 292***
OMXS 30 0.025 0.015 0.04 3.79 3 125***
TOPIX −0.004 0.014 −0.30 5.83 7 483***
Notes: The total number of observations is 5218 for each stock index. The mean is expressed as a percentage. JB shows the result of Jarque-Bera’s normality
test.
***
Denotes statistical significance at the 1% level.

Table 3
GARCH parameters for daily log returns of stock indexes.
Series GARCH(1,1) ARMA(1,1)-GARCH(1,1)
µ ω α1 β1 λ µ ϕ θ ω α1 β1
S&P 500 0.068 0.014 0.094 0.899 6.161 0.057 0.800 −0.837 0.018 0.095 0.892
FTSE 100 0.049 0.015 0.104 0.888 8.836 0.042 0.924 −0.950 0.017 0.103 0.887
DAX 30 0.096 0.021 0.094 0.900 8.358 0.082 0.869 −0.888 0.028 0.098 0.892
EURO STOXX 0.074 0.022 0.093 0.900 7.759 0.063 0.773 −0.807 0.027 0.093 0.896
MSCI 0.061 0.012 0.095 0.895 7.734 0.048 0.010 0.142 0.012 0.094 0.894
CAC 40 0.071 0.022 0.087 0.905 8.342 0.061 0.815 −0.852 0.026 0.089 0.900
OMXC 20 0.090 0.045 0.122 0.852 7.406 0.083 −0.296 0.340 0.048 0.105 0.863
OMXH 0.095 0.013 0.070 0.929 6.221 0.076 −0.137 0.174 0.013 0.061 0.936
OMXS 30 0.082 0.019 0.086 0.908 8.330 0.076 0.785 −0.822 0.024 0.090 0.902
TOPIX 0.040 0.040 0.089 0.891 6.966 0.037 0.115 −0.075 0.047 0.100 0.876
Note: The total number of observations used for each stock index is 5218.

Table 4
Estimates of EVT parameters for standardized residuals from both GARCH(1,1) and ARMA(1,1)-GARCH(1,1) models.
Series GARCH(1,1) ARMA(1,1)-GARCH(1,1)
Upper tail Lower tail Upper tail Lower tail
µR ξR βR µL ξL βL µR ξR βR µL ξL βL
S&P 500 1.148 −0.182 0.568 −1.325 0.059 0.592 1.154 −0.177 0.558 −1.326 0.032 0.631
FTSE 100 1.194 −0.066 0.481 −1.301 −0.095 0.699 1.185 −0.064 0.474 −1.316 −0.109 0.712
DAX 30 1.168 −0.052 0.479 −1.352 −0.019 0.608 1.170 −0.063 0.487 −1.355 −0.031 0.617
EURO STOXX 1.165 0.002 0.471 −1.329 −0.003 0.607 1.158 −0.011 0.477 −1.340 −0.008 0.613
MSCI 1.134 −0.116 0.542 −1.319 −0.017 0.644 1.165 −0.109 0.550 −1.285 −0.011 0.646
CAC 40 1.158 −0.045 0.510 −1.309 −0.036 0.641 1.160 −0.037 0.490 −1.326 −0.043 0.648
OMXC 20 1.198 −0.018 0.504 −1.291 0.045 0.573 1.210 −0.024 0.504 −1.282 0.037 0.581
OMXH 1.135 0.049 0.485 −1.262 0.087 0.598 1.148 0.055 0.483 −1.237 0.082 0.599
OMXS 30 1.160 −0.091 0.553 −1.285 −0.015 0.627 1.157 −0.099 0.547 −1.294 −0.025 0.643
TOPIX 1.190 −0.006 0.467 −1.318 0.071 0.560 1.190 −0.006 0.467 −1.318 0.071 0.560
Note: The total number of observations for each stock index is 5218.

is peak-over-threshold (POT), and the first step is to choose obtaining the marginal distribution, we use copula mod-
appropriate thresholds. To do so, we follow Wang, Chen et els to estimate the dependency structure. Following that,
al. (2010) and choose the 10th percentile of the standard- 10,000 simulated uniforms are obtained and inserted into
ized residuals.3 The estimated parameters µ (threshold), the inverse function of the marginal distribution for each
ξ (scale) and β (location) for the upper and lower tails are series. Next, one-day-ahead returns are derived and the
reported in Table 4. optimal weights are computed based on the three opti-
As was mentioned earlier, after fitting the GPD to the mization methods (Min-CVaR, CET and GMV). We repeat
upper and lower tails of the standardized residuals and the whole procedure for each day in the out-of-sample
period and perform portfolio back-testing for each model
to obtain the portfolio returns. Fig. 1 illustrates the tan-
3 We also check the mean excess function (MEF) and Hill plots. In
gency (CET) portfolio accumulation wealth with an initial
the Hill plot, the threshold is seen when the graph is horizontally stable.
investment of $100. As we can see from the figure, all of
In the case of MEF plot, a straight line with a positive slope shows the
threshold (see Gençay & Selçuk, 2004, and Karmakar & Shukla, 2015, for the forecasting models outperform the benchmark, which
the interpretation of the MEF and Hill plots). is the tangency portfolio obtained by using historical data.
 M. Sahamkhadam et al. / International Journal of Forecasting 34 (2018) 497–506 503

Fig. 1. The tangency portfolio (CET) accumulation wealth plot for an initial investment of $100. The plot is based on the results of the out-of-sample
portfolio back-testing, where the sample includes 5218 logarithmic daily returns for stock indexes. The portfolio returns are obtained based on a two-step
procedure. First, parameter estimation is performed and one-day-ahead returns are simulated using each forecasting model. Second, the simulated one-
day-ahead returns are inserted into each portfolio allocation strategy in order to obtain the optimal weights. This two-step procedure is employed based on
a rolling window estimation. The training sample size is 1260 days, which gives an out-of-sample period of 3958 iterations. The portfolio realized returns
are computed by applying the out-of-sample optimal weights to realized one-day-ahead returns. The benchmark portfolio includes tangency portfolio
based on the historical data. The portfolio back-testing is performed based on the daily rebalancing strategy.

5. Impact of risk modeling on the portfolio risk and in the regression analysis. The strategies that are con-
performance sidered in the regression analyses are GARCH forecasting
models with and without ARMA terms, the application of
We determine and evaluate the impact of model choice EVT to the modeling of extreme returns in the lower/upper
on portfolio risk by summarizing the portfolio returns gen- tail, various copula specifications for modeling the depen-
erated from each strategy, calculating the standard devi- dence, including elliptical and Archimedean copulas, and
ation (SD), first percentile (P1) and 99th percentile (P99) rebalancing the portfolio weights daily, weekly, monthly,
of the 3958 out-of-sample predicted portfolio returns. The semi-annually or annually in the backtesting procedure
first two statistical measures SD and P1 describe the port- (note that annual rebalancing is the reference category).
folio risk, with the latter describing the downside risk Each risk modeling strategy is included as a categorical
in particular, while P99 describes the positive (wanted) variable in the regression model, and the reference cat-
potential in terms of the upper tail of portfolio returns. egory is defined as the corresponding strategy based on
Each combination of the portfolio optimization methods historical returns instead of GARCH model forecasts.
and forecasting techniques reflects the performance of a We focus in particular on the ability of risk modeling
particular risk modeling strategy. Since we apply 96 fore- techniques to reduce the portfolio risk, which means either
casting techniques to each portfolio optimization method decreasing the portfolio return volatility (SD) or increasing
(Min-CVaR, GMV and CET), this gives a total of 3 × 96 = 288 the first percentile (P1) of portfolio returns. We also test the
observations for each statistical measure (SD, P1, P99). The GARCH-STD-copula model in order to evaluate the effects
regression estimates indicate the connection between the of using EVT for the transformation of the marginals.4
portfolio risk —measured by SD, P1 and P99 —and specific The main finding from the regression analysis is that
forecasting and risk modeling techniques. The statistical the Gaussian and Student-t copulas based on GARCH fore-
measure of risk in the regression model is a function of casting with or without EVT reduce the volatility (SD)
categorical variables that describe the respective modeling of the corresponding portfolio returns for the Min-CVaR
strategy. optimal portfolios. As the results of Table 5 show, these
SDi = αSD + βSD,1 D1i + βSD,2 D2i + µi models typically also increase the lower tail (P1) of the
{
P1i = αP1 + βP1,1 D1i + βP1,2 D2i + νi (25) returns, meaning that a portfolio’s downside risk is reduced
P99i = αP99 + βP99,1 D1i + βP99,2 D2i + γi , on average. However, the significant negative coefficients
indicate that these models also reduce the P99 of portfolio
where D1i and D2i (i = 1, 2, . . . , 96) denote sets of the
returns in many cases. This implies that, while the use
dummy variables for forecasting models and rebalancing
of these forecasting models can reduce the portfolio risk,
strategies. We report regression results for each portfolio
this also reduces the likelihood of high portfolio returns.
optimization approach separately, because the effects of
Thus, in terms of risk-modeling strategies, the benefits
the modeling strategies can differ across portfolio methods,
and thus, each regression model uses 96 observations (see
Table 5). 4 In this model, after employing GARCH(1,1) to obtain the standard-
The reference model for each portfolio optimization ized residuals, we simply use the probability function of the t distribution
to perform the transformation. We set the t distribution as the conditional
method is based on historical returns using a rolling win-
distribution in the GARCH(1,1) model and obtain the parameter required
dow of 1260 days. If a specific risk modeling strategy using for the transformation, which is degrees of freedom. Moreover, after sim-
GARCH model forecasts reduces the portfolio risk or im- ulating uniforms from copula models, we again use the quantile function
proves portfolio returns, we should find significant effects of the t distribution. This is the parametric method of transformation.
504 M. Sahamkhadam et al. / International Journal of Forecasting 34 (2018) 497–506

Table 5
Regression results of risk modeling strategies on statistical measures of the portfolio return distribution for the three portfolio optimization methods (Min-
CVaR, GMV and CET).
Forecasting model Optimization: Min-CVaR Optimization: GMV Optimization: CET
SD P1 P99 SD P1 P99 SD P1 P99
ARMA-GARCH-EVT- 0.0523*** −0.131 ** 0.185*** 0.0527*** −0.108 * 0.227*** −0.0926 *** 0.292*** −0.208 **
Clayton (0.0187) (0.0612) (0.0438) (0.0195) (0.0648) (0.0415) (0.0157) (0.0560) (0.0848)
GARCH-EVT-Clayton 0.00495 0.0919 0.116** 0.0121 0.107 0.196*** −0.0228 0.255*** 0.191**
(0.0187) (0.0612) (0.0438) (0.0195) (0.0648) (0.0415) (0.0157) (0.0560) (0.0848)
GARCH-STD-Clayton 0.0125 0.0676 0.129*** 0.0123 0.109* 0.185*** −0.0934 *** 0.409*** 0.107
(0.0187) (0.0612) (0.0438) (0.0195) (0.0648) (0.0415) (0.0157) (0.0560) (0.0848)
ARMA-GARCH-EVT- 0.0548*** −0.141 ** 0.232*** 0.0515** −0.132 ** 0.238*** −0.0958 *** 0.305*** −0.181 **
Frank (0.0187) (0.0612) (0.0438) (0.0195) (0.0648) (0.0415) (0.0157) (0.0560) (0.0848)
GARCH-EVT-Frank 0.0156 0.0335 0.159*** 0.0102 0.108* 0.182*** −0.0303 * 0.307*** 0.197**
(0.0187) (0.0612) (0.0438) (0.0195) (0.0648) (0.0415) (0.0157) (0.0560) (0.0848)
GARCH-STD-Frank 0.0102 0.0905 0.145*** 0.0102 0.110* 0.170*** −0.0904 *** 0.386*** 0.0806
(0.0187) (0.0612) (0.0438) (0.0195) (0.0648) (0.0415) (0.0157) (0.0560) (0.0848)
ARMA-GARCH-EVT- −0.00792 0.0839 −0.00525 −0.0104 0.0192 0.0613 −0.113 *** 0.391*** −0.253 ***
Gaussian (0.0187) (0.0612) (0.0438) (0.0195) (0.0648) (0.0415) (0.0157) (0.0560) (0.0848)
GARCH-EVT-Gaussian −0.0476 ** 0.177*** −0.0256 −0.0434 ** 0.183*** 0.00738 −0.390 *** 1.157*** −1.148 ***
(0.0187) (0.0612) (0.0438) (0.0195) (0.0648) (0.0415) (0.0157) (0.0560) (0.0848)
GARCH-STD-Gaussian −0.0507 *** 0.162** −0.0568 −0.0454 ** 0.183*** −0.0348 −0.152 *** 0.540*** −0.136
(0.0187) (0.0612) (0.0438) (0.0195) (0.0648) (0.0415) (0.0157) (0.0560) (0.0848)
ARMA-GARCH-EVT 0.0608*** −0.159 ** 0.250*** 0.0536*** −0.129 ** 0.230*** −0.103 *** 0.345*** −0.217 **
-Gumbel (0.0187) (0.0612) (0.0438) (0.0195) (0.0648) (0.0415) (0.0157) (0.0560) (0.0848)
GARCH-EVT-Gumbel 0.0216 0.0258 0.199*** 0.0119 0.107 0.181*** −0.0454 *** 0.299*** 0.0988
(0.0187) (0.0612) (0.0438) (0.0195) (0.0648) (0.0415) (0.0157) (0.0560) (0.0848)
GARCH-STD-Gumbel 0.0135 0.0864 0.162*** 0.0144 0.0626 0.183*** −0.0710 *** 0.326*** 0.111
(0.0187) (0.0612) (0.0438) (0.0195) (0.0648) (0.0415) (0.0157) (0.0560) (0.0848)
ARMA-GARCH-EVT- −0.00842 0.118* 0.000114 −0.0121 0.0343 0.0223 −0.0934 *** 0.336*** −0.153 *
Student-t (0.0187) (0.0612) (0.0438) (0.0195) (0.0648) (0.0415) (0.0157) (0.0560) (0.0848)
GARCH-EVT-Student-t −0.0484 ** 0.218*** −0.0300 −0.0437 ** 0.210*** −0.00759 0.0214 −0.0377 0.398***
(0.0187) (0.0612) (0.0438) (0.0195) (0.0648) (0.0415) (0.0157) (0.0560) (0.0848)
GARCH-STD-Student-t −0.0484 ** 0.188*** −0.0478 −0.0435 ** 0.202*** 0.00462 −0.0502 *** 0.0849 0.316***
(0.0187) (0.0612) (0.0438) (0.0195) (0.0648) (0.0415) (0.0157) (0.0560) (0.0848)
Rebalancing semi-annually −0.0294 ** 0.147*** −0.0641 ** −0.0284 ** 0.109*** −0.0766 *** 0.00296 −0.0301 0.0431
(0.0115) (0.0375) (0.0268) (0.0119) (0.0397) (0.0254) (0.00960) (0.0343) (0.0519)
Rebalancing quarterly −0.0352 *** 0.181*** −0.0926 *** −0.0351 *** 0.166*** −0.0935 *** 0.00485 −0.0236 0.218***
(0.0115) (0.0375) (0.0268) (0.0119) (0.0397) (0.0254) (0.00960) (0.0343) (0.0519)
Rebalancing monthly −0.0530 *** 0.269*** −0.175 *** −0.0538 *** 0.268*** −0.184 *** 0.00163 −0.0460 0.141***
(0.0115) (0.0375) (0.0268) (0.0119) (0.0397) (0.0254) (0.00960) (0.0343) (0.0519)
Rebalancing weekly −0.0799 *** 0.357*** −0.190 *** −0.0821 *** 0.322*** −0.223 *** −0.0169 * 0.0919*** 0.141***
(0.0115) (0.0375) (0.0268) (0.0119) (0.0397) (0.0254) (0.00960) (0.0343) (0.0519)
Rebalancing daily −0.138 *** 0.569*** −0.331 *** −0.140 *** 0.548*** −0.333 *** −0.0793 *** 0.321*** 0.136**
(0.0115) (0.0375) (0.0268) (0.0119) (0.0397) (0.0254) (0.00960) (0.0343) (0.0519)
Constant 0.995*** −3.080 *** 2.367*** 0.988*** −3.061 *** 2.319*** 1.293*** −3.852 *** 3.210***
(0.0152) (0.0496) (0.0355) (0.0158) (0.0525) (0.0336) (0.0127) (0.0454) (0.0687)
Observations 96 96 96 96 96 96 96 96 96
R2 0.795 0.835 0.832 0.776 0.811 0.842 0.939 0.918 0.879
Notes: The portfolio returns are obtained by applying rolling window estimation (L = 1260) and portfolio back-testing. The dependent variables are the
standard deviation (SD), first percentile (P1) and 99th percentile (P99) of portfolio returns. The risk modeling strategies are included as categorical variables
(dummy variables) in the regressions. The reference category (benchmark) is the respective optimal portfolio based on historical returns using a 1260-day
rolling window. The reference category for the rebalancing of portfolio weights is annually. Standard errors are shown in parentheses.
*
p < 0.1.
**
p < 0.05.
***
p < 0.01.

come at the cost of reducing the risk (opportunities) in the almost all of the risk models decrease the portfolio risk
upper tail as well, implying lower performances. Another significantly; on the other hand, though, all techniques also
finding from the regression analysis is that the main results decrease the 99th percentile of portfolio returns. Overall,
for the GMV and Min-CVaR portfolios are quite similar. GARCH forecasting models in combination with EVT and
According to our study, the gains from using Min-CVaR copula models appear to be particularly well suited for the
optimal portfolios instead of GMV portfolios, in terms of CET optimization framework. As expected, we find that a
reducing the portfolio risk (P1), are minor in most cases, as more frequent rebalancing of portfolio weights decreases
is shown by the small differences between the regression the portfolio risk for different rebalancing strategies,5 but
intercepts for the two portfolio optimization methods. To
the best of our knowledge, no previous study has provided
evidence on this issue. When it comes to CET portfolios, 5 This corresponds to the results of Huang and Hsu (2015).
 M. Sahamkhadam et al. / International Journal of Forecasting 34 (2018) 497–506 505

again also decreases the 99th percentile of portfolio re- Bhattacharyya, M., & Ritolia, G. (2008). Conditional VaR using EVT—
turns. Towards a planned margin scheme. International Review of Financial
Analysis, 17(2), 382–395.
Bhatti, M. I., & Nguyen, C. C. (2012). Diversification evidence from inter-
6. Conclusions national equity markets using extreme values and stochastic copulas.
Journal of International Financial Markets, Institutions and Money, 22(3),
622–646.
This study uses GARCH-EVT-copula and ARMA-GARCH-
Bodnar, T., Mazur, S., & Okhrin, Y. (2017). Bayesian estimation of the global
EVT-copula models to forecast and simulate the one-day- minimum variance portfolio. European Journal of Operational Research,
ahead returns of ten stock indexes. Using the forecasts, 256(1), 292–307.
we employ three portfolio optimization techniques (Min- Boltzmann, L. (1896). Vorlesungen über Gastheorie, Vol. 1. JA Barth.
CVaR, GMV and CET) to compute the optimal weights Bondt, W. F., & Thaler, R. (1985). Does the stock market overreact? The
Journal of Finance, 40(3), 793–805.
and perform portfolio back-testing for the out-of-sample Chan, K. F., & Gray, P. (2006). Using extreme value theory to measure
period based on different rebalancing strategies. Further- value-at-risk for daily electricity spot prices. International Journal of
more, we perform regression analyses to examine how, on Forecasting, 22(2), 283–300.
average, the various risk modeling strategies affect three Christoffersen, P., Errunza, V., Jacobs, K., & Jin, X. (2014). Correlation dy-
namics and international diversification benefits. International Journal
statistical measures of the portfolio return distribution: the of Forecasting, 30(3), 807–824.
return volatility and lower and upper tails of the portfolio Consigli, G. (2002). Tail estimation and mean-VaR portfolio selection in
returns. markets subject to financial instability. Journal of Banking & Finance,
In terms of wealth accumulation, we find that GARCH- 26(7), 1355–1382.
Crato, N., & Ruiz, E. (2012). Can we evaluate the predictability of financial
EVT-copula forecasting models are not able to outperform markets? International Journal of Forecasting, 28(1), 1–2.
models that are based on historical returns (our bench- Dacorogna, M. M., & Pictet, O. V. (1997). Heavy tails in high-frequency
mark) for Min-CVaR and GMV portfolios. However, for financial data. Available as SSRN 939.
CET, the use of forecasting models improves the portfolio Deng, L., Ma, C., & Yang, W. (2011). Portfolio optimization via pair copula-
GARCH-EVT-CVaR model. Systems Engineering Procedia, 2, 171–181.
wealth.
Gençay, R., & Selçuk, F. (2004). Extreme value theory and value-at-risk:
In addition, our regression analyses provide evidence Relative performance in emerging markets. International Journal of
that elliptical copulas (Gaussian and Student-t) reduce Forecasting, 20(2), 287–303.
the return volatility of Min-CVaR portfolios better than Gençay, R., & Selçuk, F. (2006). Overnight borrowing, interest rates and
extreme value theory. European Economic Review, 50(3), 547–563.
Archimedean copulas. Overall, the improvement in portfo-
Gençay, R., Selçuk, F., & Ulugülyaǧci, A. (2003). High volatility, thick
lio risk reduction from using risk modeling and forecasting tails and extreme value theory in value-at-risk estimation. Insurance:
techniques in combination with Min-CVaR and GMV port- Mathematics & Economics, 33(2), 337–356.
folios appears to be modest. As minimizing the risk already Ghalanos, A. (2015). spd: Semi-Parametric Distribution. URL https://cran.
forms part of determining the optimal weights for MinC- r-project.org/package=spd R package version 2.0-1.
Ghalanos, A. (2018). rugarch: Univariate GARCH models. URL https://cran.
VaR and GMV portfolios, combining them with risk mod- r-project.org/package=rugarch R package version 1.4-0.
eling techniques might not improve the results further. Han, Y., Li, P., & Xia, Y. (2017). Dynamic robust portfolio selection with
However, there is a strong indication that risk modeling copulas. Finance Research Letters, 21, 190–200.
strategies based on return forecasts in combination with Harmantzis, F. C., Miao, L., & Chien, Y. (2006). Empirical study of value-at-
risk and expected shortfall models with heavy tails. The Journal of Risk
the optimal tangency portfolio can reduce the portfolio risk Finance, 7(2), 117–135.
significantly. Hofert, M., Maechler, M., & McNeil, A. J. (2012). Estimators for Archimedean
Future research should try to improve the forecasting copulas in high dimensions. arXiv preprint, arXiv:1207.1708.
ability of the ARMA-GARCH models by including explana- Huang, C.-W., & Hsu, C.-P. (2015). Portfolio optimization with GARCH-
EVT-Copula-CVaR Models. Banking and Finance Review, 7(1), 19–31.
tory factors in both the mean and conditional volatility
Huang, J.-J., Lee, K.-J., Liang, H., & Lin, W.-F. (2009). Estimating value
parts. Fama–French factors and volatility indexes (VIX) are at risk of portfolio by conditional copula-GARCH method. Insurance:
natural candidates for this. Mathematics & Economics, 45(3), 315–324.
Karmakar, M., & Shukla, G. K. (2015). Managing extreme risk in some ma-
jor stock markets: An extreme value approach. International Review of
Acknowledgments Economics and Finance, 35, 1–25.
Kim, J.-M., & Jung, H. (2016). Linear time-varying regression with Copula-
Some of the computations for this study were done DCC-GARCH models for volatility. Economics Letters, 145, 262–265.
Koliai, L. (2016). Extreme risk modelling: An EVT-pair-copulas approach
using the R packages ‘‘rugarch’’ and ‘‘spd’’ (Ghalanos, 2015,
for financial stress tests. Journal of Banking & Finance, 70, 1–22.
2018). We want to thank Bernhard Pfaff for his guidance Longin, F. M. (1996). The asymptotic distribution of extreme stock market
in the use of GARCH-copula models (Pfaff, 2016). In addi- returns. Journal of Business, 69, 383–408.
tion, we are grateful to seminar participants at the 10th Low, R. K. Y., Alcock, J., Faff, R., & Brailsford, T. (2013). Canonical vine
International Conference on Computational and Financial copulas in the context of modern portfolio management: Are they
worth it? Journal of Banking & Finance, 37(8), 3085–3099.
Econometrics (CFE 2016) in Seville. Finally, we thank Scott Markowitz, H. (1952). Portfolio selection. The Journal of Finance, 7(1),
Hacker, Kristofer Månsson, Gazi Salah Uddin and Pär Sjö- 77–91.
lander for their helpful comments and suggestions on a McNeil, A. J., & Frey, R. (2000). Estimation of tail-related risk measures
previous version of this paper. The usual disclaimer applies. for heteroscedastic financial time series: an extreme value approach.
Journal of Empirical Finance, 7(3), 271–300.
Merton, R. C. (1980). On estimating the expected return on the mar-
References ket: An exploratory investigation. Journal of Financial Economics, 8(4),
323–361.
Berger, T. (2013). Forecasting value-at-risk using time varying copulas and Pfaff, B. (2016). Financial risk modelling and portfolio optimization with R.
EVT return distributions. International Economics, 133, 93–106. John Wiley & Sons.
506 M. Sahamkhadam et al. / International Journal of Forecasting 34 (2018) 497–506

Pircalabu, A., Hvolby, T., Jung, J., & Høg, E. (2017). Joint price and volumet- Maziar Sahamkhadam is a Ph.D. student at the School of Business and
ric risk in wind power trading: A copula approach. Energy Economics, Economics, Linnaeus University, Sweden. His research is in the area of
62, 139–154. portfolio optimization and risk modeling. He holds a master degree in
Righi, M. B., Schlender, S. G., & Ceretta, P. S. (2015). Pair copula construc- International Financial Analysis from Jönköping International Business
tions to determine the dependence structure of treasury bond yields. School, Sweden.
IIMB Management Review, 27(4), 216–227.
Rockafellar, R. T., & Uryasev, S. (2000). Optimization of conditional value-
at-risk. Journal of Risk, 2, 21–42.
Rockafellar, R. T., & Uryasev, S. (2002). Conditional value-at-risk for gen- Andreas Stephan studied industrial engineering, statistics and economics
eral loss distributions. Journal of Banking & Finance, 26(7), 1443–1471. at TU Berlin and received a Ph.D. in economics at the Humboldt University
Sharpe, W. F. (1963). A simplified model for portfolio analysis. Management Berlin. He previously held positions at the Social Science Research Center
Science, 9(2), 277–293. in Berlin (WZB), at the German Institute for Economic Research (DIW
Sharpe, W. F. (1994). The Sharpe ratio. The Journal of Portfolio Management, Berlin) and as an Assistant Professor of economics at the European Uni-
21(1), 49–58. versity Viadrina in Frankfurt/Oder. His main areas of interest are financial
Sklar, M. (1959). Fonctions de répartition à n dimensions et leurs marges. economics including financial engineering. Previous work includes the
Université Paris 8. estimation of implicit betas from option prices. He is affiliated with CESIS
Wang, Z.-R., Chen, X.-H., Jin, Y.-B., & Zhou, Y.-J. (2010). Estimating risk - Centre of Excellence for Science and Innovation Studies at the Royal
of foreign exchange portfolio: Using VaR and CVaR based on GARCH- Institute of Technology, Stockholm, and with CeFEO - Centre for Family
EVT-copula model. Physica A. Statistical Mechanics and its Applications, Enterprise and Ownership, Jönköping.
389(21), 4918–4928.
Wang, Z., Jin, Y., & Zhou, Y. (2010). Estimating portfolio risk using GARCH-
EVT-copula model: An empirical study on exchange rate market.
In Advances in neural network research and applications (pp. 65–72).
Springer. Ralf Östermark studied accounting and economics at Turku School of
Xu, Q., Zhou, Y., Jiang, C., Yu, K., & Niu, X. (2016). A large CVaR-based port- Buiness Economics and Åbo Akademi University (ÅAU). He has held po-
folio selection model with weight constraints. Economic Modelling, 59, sitions in business economics and accounting at ÅAU. He is currently
436–447. Professor in accounting and optimization systems at ÅAU. His main areas
Zhang, B., Wei, Y., Yu, J., Lai, X., & Peng, Z. (2014). Forecasting VaR and ES of interest are high performance computing and financial engineering.
of stock index portfolio: A Vine copula method. Physica A. Statistical Previous work includes multiperiod portfolio management systems and
Mechanics and its Applications, 416, 112–124. scalability testing of challenging numerical problems.