Muller e Righi 2018

Risk Manag
DOI 10.1057/s41283-017-0026-8
ORIGINAL ARTICLE
Numerical comparison of multivariate models

to forecasting risk measures
Fernanda Maria Müller1 • Marcelo Brutti Righi1
Ó Macmillan Publishers Ltd 2017
Abstract We evaluated the performance of multivariate models for forecasting

Value at Risk (VaR), Expected Shortfall (ES), and Expectile Value at Risk (EvaR).
We used Historical Simulation (HS), Dynamic Conditional Correlation-Generalized
Autoregressive Conditional Heteroskedastic (DCC-GARCH) and copula methods:
Regular copulas, Vine copulas, and Nested Archimedean copulas (NAC). We
assessed the performance of the models using Monte Carlo simulations, considering
different scenarios, regarding the marginal distributions, correlation, and number of
portfolio assets. Numerical results evidenced the accuracy forecasting risk measures
are associated with marginal distributions. For a data-generating process where the
marginal distribution is Gaussian, Regular and Vine copulas demonstrated better
performance. For data generated with Student’s t distribution, we verified better
performance by NAC. In addition, we identified the superiority of copula methods
over HS and DCC-GARCH, which reduces the model risk.
Keywords Risk measures DCC-GARCH Copulas Model risk Monte Carlo

simulation
Introduction
Over the last few years, interest from financial organizations, regulators, and of the
academic community in the development of better tools to measure market risk has
increased. The main factors contributing to this were the increasing number of
& Fernanda Maria Müller

[email protected]
Marcelo Brutti Righi
[email protected]
1
School of Business, Federal University of Rio Grande do Sul, Washington Luiz, 855,
Porto Alegre 90010-460, Brazil
F. M. Müller, M. B. Righi
collapses and trading activities on the financial markets. In this context, Value at
Risk (VaR) has become one of the most popular risk management tools. This
measure refers to the maximal loss expected to occur by a financial position for a
given period of time and confidence level. For a general review, one can consult the
works of Duffie and Pan (1997) and Jorion (2006). Despite its simplicity and
popularity, VaR has been criticized for presenting theoretical deficiencies as a
measure of market risk. One of the principal weaknesses is that it is not a coherent
measure of risk1 once it does not meet subadditivity. In contrast to the principle of
diversification, the VaR of a portfolio is not necessarily smaller than the sum of the
VaR for individual assets. This measure ignores potential losses that exceed the
quantile of interest.
To circumvent the limitations of VaR, Artzner et al. (1999) presented the concept
of Tail Conditional Expectation (TVaR), known in the literature as Expected
Shortfall (ES), a measure proposed by Acerbi and Tasche (2002). ES is defined as
the conditional expectation of a loss, given this loss exceeds VaR. By its definition,
ES considers the losses that exceed the quantile of interest, besides being a coherent
risk measure. Another alternative to VaR is to use expectiles, introduced by Newey
and Powell (1987). This measure is referred to as Expectile Value at Risk (EVaR)
and is more sensitive to extreme losses compared to VaR, since the tail probability
is determined by the underlying distribution. In addition, it is a coherent risk
measure, being the only example of an elicitable2 coherent risk measure beyond the
mean. Besides these measures, in the literature there are other options for
quantifying the risk, such as the Entropic, studied by Föllmer and Schied (2002),
and Shortfall Deviation Risk (SDR) proposed by Righi and Ceretta (2016).
Under the practical relevance of risk measures for financial risk management,
there is a need for reliable estimates and forecasts. One of the most commonly used
methods is the non-parametric approach, known as Historical Simulation (HS),
which is based on the empirical distribution of returns. Parametric methods are also
used, such as Generalized Autoregressive Conditional Heteroskedastic (GARCH)
models. Regarding the estimation of risk measures for portfolios, difficulties can be
observed due to the complexity of modelling the joint distribution of assets. To
solve this problem, many authors have proposed the use of copulas. A copula results
in a multivariate distribution, combining univariate marginal distributions and
dependence between variables (Nelsen 2006). They allow the aggregation of
marginal distributions of two or more variables and capture important features
found in risk management, such as asymmetries and heavy tails. The mathematical
basis of copulas was introduced by Sklar (1959). Their application in financial
practical situations has become clearer more recently. Examples of the use of
bivariate copulas to estimate risk measures are found in Mendes and de Souza
(2004), Junker and May (2005), Palaro and Hotta (2006), and Jäschke (2014).
1
A coherent risk measure is a risk measure that satisfies four axioms: translation invariance,
subadditivity, positive homogeneity, and monotonicity (Artzner et al. 1999).
2
Risk measures are called elicitable when the verification and comparison of competing estimation
procedures is possible (Ziegel 2014). See also Ziegel (2014) and Bellini and Bernardino (2017) for a
detailed discussion referent to elicitability for risk measures.
Numerical comparison of multivariate...
For the bivariate case, a wide variety of different types of copulas is available,
exhibiting flexible and complex dependence patterns. However, for the general
multivariate case, the appropriate choice of the copula family is more limited.
Standard multivariate copulas do not fit data well and lack flexibility for modelling
the dependence of several assets. Aas et al. (2009) suggested the use of Vine
copulas, originally proposed by Joe (1996), for modelling multivariate distributions.
Vine copulas are a flexible model for describing multivariate relationships, using a
cascade of bivariate copulas, so-called pair-copulas (Kurowicka 2011). Such pair-
copula constructions (PCCs) decompose a multivariate density into bivariate copula
densities, including dependence structures of unconditional bivariate distributions
and dependence structures of conditional bivariate distributions. Applications of
Vine copula to estimate risk measures are found in Aas and Berg (2009), Righi and
Ceretta (2013b), Righi and Ceretta (2013a), Brechmann and Czado (2013), Low
et al. (2013), and Righi and Ceretta (2015b).
A different structure for building higher-dimensional copula is the Nested
Archimedean copulas (NAC). The hierarchical functional form first appears in Joe
(1997). NAC are generalizations of Archimedean copulas,3 allowing asymmetries
and providing more flexibility, sharing properties such as an explicit functional
form. Similar to the Vine copulas, NAC are hierarchical and based on pair-copulas.
Applications of NAC to estimate risk measures are found in Aas and Berg (2009).
Among a vast range of estimation techniques used in the literature of risk
management, a class of studies emerges to perform a comparison among the
different estimation methods of risk measures to determine the most efficient model.
We refer to Kuester et al. (2006) as an example of study for the univariate case.
Despite this literature, many studies focus their investigation on estimation methods
for VaR. Regarding ES and EVaR, there is not enough research to point out the best
estimation method. Most studies use empirical data and focus on the univariate
perspective for the comparison of the performance of different models for
forecasting risk measures. Multivariate data are a crucial issue in risk management
literature and practical situations. An inappropriate model can lead to sub-optimal
portfolios and inaccurate risk exposure assessments. The Monte Carlo simulation is
a computational method utilized to evaluate the performance of competing models.
It allows an experiment to be replicated several times to understand their behaviour.
The numerical results obtained from many iterations can be considered represen-
tative of the real situation.4
In this context, the main objective of the current paper was to analyse the
performance of models for forecasting VaR, ES, and EVaR from the multivariate
perspective. We opted for these risk measures because they are the most used in the
literature. Regarding multivariate models, we considered the following approaches:
HS, Dynamic Conditional Correlation (DCC)-GARCH (Engle 2002), Regular
3
A bivariate distribution function C(u, v) with marginals F1 and F2 is generated by an Archimedean
copula if it can be given by Cðu; vÞ ¼ u1 ½uðF1 ðuÞÞ þ uðF2 ðvÞÞ, where u is a function u : I ! Rþ ,
continuous, decreasing, convex and such that /ð1Þ ¼ 0 (Cherubini et al. 2004).
4
We refer to the Law of Large Numbers and Central Limit Theorem (CLT) in which Monte Carlo
simulations are based for a better understanding.
copulas, Vine copulas, and NAC.5 The performance of the models was evaluated
using Monte Carlo simulations. Pseudo-random samples were generated and
different scenarios considered. In each scenario, we computed absolute bias (A.
Bias), relative bias (R. Bias), and root-mean-square error (RMSE).
To the best of our knowledge, our paper is the first to perform an evaluation with
Monte Carlo simulations for forecasting VaR, ES, and EVaR using multivariate
models. Previous studies using Monte Carlo simulations, such as Manganelli and
Engle (2001), have VaR as the first measure of interest and restrict their analysis to
the univariate case. Zhou (2012), Degiannakis et al. (2013), Tolikas (2014), and
Righi and Ceretta (2015a) compared estimators for VaR and ES; however, they
considered empirical data and focus on univariate models; the authors did not
provide an extensive focus on VaR and ES forecasting patterns as our study has
done. Regarding EVaR, Bellini and Bernardino (2017) performed a review of its
known properties, and financial meaning and compared them with VaR and ES.
Their main objective was not to investigate which model offers better performance
for computing these risk measures but to present EVaR as an alternative measure to
VaR and ES. In the work of Righi and Borenstein (2017), although they considered
VaR, ES, and EVaR, their main objective was compare risk measures regarding
performance of optimal portfolio strategies.
Another contribution of our paper is that it is the first study to perform a
comparison of HS, DCC, and copulas to estimate risk measures, in a general
multivariate context. Previous studies, such as Weiß (2013), considered bivariate
copula and analysed the forecasting of VaR and ES. In addition, they have not used
Monte Carlo simulations. As a final contribution of our paper, it was possible to
identify the model that reduces the model risk6 of the risk measures analysed.
Appropriate risk-measurement techniques are paramount in finance, once they are
used as an input into expensive decisions, such as asset allocation, hedge strategies,
and capital requirement. Risk measures estimated with models with a lower model
risk are fundamental to improving risk management and to strengthen global
financial stability.7
The remainder of this paper is structured as follows: Sect. 2 presents a
background of the risk measures used in this paper and the estimation methods we
considered. In Sect. 3 numerical procedures are described, and in Sect. 4, numerical
results are presented and discussed. Section 5 summarizes and concludes the paper.
5
This research focused on comparing the performance of traditional multivariate models, such as HS and
DCC-GARCH, with copulas. We did not use models such as multivariate quantile regression and
multivariate Extreme Value Theory.
6
Model risk refers to errors in modelling assumptions that introduce errors in risk measurement
(Glasserman and Xu 2014).
7
Revisions to the Basel II Market Risk Framework require financial institutions quantify model risk
(Basel Committee on Banking Supervision 2009).
Background
Risk measures
In this subsection, we defined VaR, ES, and EVaR. We focused on the continuous
distribution case, but generalizations to the discrete case are straightway. X
represents the random result of any asset or portfolio, where X 0 a profit and X\0
a loss, and a 2 ð0; 1Þ corresponds to the significance level. Formally, expression (1)
defines VaR:
VaRa ðXÞ ¼ inffx : FX ðxÞ ag ¼ FX1 ðaÞ; ð1Þ
where FX is the probability function of X and its inverse is FX1 . Based on formu-
lation (1), we observed VaR does not give information about the severity of losses
that occur with low probability. To outperform VaR drawback, Acerbi and Tasche
(2002) suggested the use of ES, which is defined as
ESa ðXÞ ¼ E½XjX FX1 ðaÞ: ð2Þ

This risk measure represents the expected value of a loss, given that it is beyond the
a-quantile of interest. Nowadays, ES is the most used coherent risk measure. Given
the same significance level, ES is higher than VaR. From a risk management point
of view, this can lead to higher levels of security. Despite its theoretical advantages,
ES is estimated with more uncertainty, because it is subject to errors in the esti-
mation of VaR and of the expectation of tail observations. Besides that, ES is not an
elicitable risk measure.
Another coherent possibility to measure risk is the EVaR. This risk measure is
linked to the concept of expectiles, which is a generalized quantile function
employed for obtaining VaR. Mathematically, EVaR is defined as (3)
EVaRa ðXÞ ¼ arg min E½ða 1X h ÞðX hÞ2 ; ð3Þ
h
where 1a is an indicator function with the value 1 if a is true and 0 otherwise. Bellini
et al. (2014) and Bellini and Bernardino (2017) argued EVaR is a coherent risk
measure, for a 0:5. Delbaen (2013) described the properties of EVaR as a
coherent risk measure. This measure is also elicitable. Some authors have suggested
it as an alternative to both ES and VaR (Emmer et al. 2015). Although EVaR has
good theoretical properties, in the literature it is criticized for not presenting
financial interpretation. To circumvent this obstacle, Bellini and Bernardino (2017)
discussed its interpretation in relation to acceptance set. In this context, according to
EVaR, a position is acceptable when the ratio between the expected value of the
gain and of the loss is sufficiently high. According to the authors, the measure
presents similarities to VaR and ES, and their numerical results pointed to EVaR as
a good alternative to quantify market risk.
Copulas
To facilitate the notation, we focused on the bivariate case. Extensions to the n-

dimensional case are straightforward. A function C : ½0; 12 ! ½0; 1 is a copula, if
for 0 u; v 1, and u1 u2 ; v1 v2 ; ðu1 ; v1 Þ; ðu2 ; v2 Þ 2 ½0; 12 , it respects the
following properties (Cherubini et al. 2004):
1. Cðu; 0Þ ¼ Cð0; vÞ ¼ 0, such that Cðu; 1Þ ¼ u and Cð1; vÞ ¼ v, for every (u, v) of
A B, which represents two non-empty subsets of I ¼ ½0; 1 2 R. This property
means uniformity of the margins.
2. Cðu2 ; v2 Þ Cðu2 ; v1 Þ Cðu1 ; v2 Þ þ Cðu1 ; v1 Þ 0, which indicates the function
C is called n-increasing. This property means that Pðu1 U u2 ;
v1 V v2 Þ 0 for (U, V) with distribution function C.
Given that C is a copula and F1 and F2 univariate distribution functions, C is unique

if F1 and F2 are continuous for each ðu; vÞ 2 ½0; 12 . A bivariate distribution F with
marginals F1 and F2 , is defined by
Fðu; vÞ ¼ CðF1 ðuÞ; F2 ðvÞÞ; ðu; vÞ 2 R2 : ð4Þ
For a two-dimensional distribution function F with marginal F1 and F2 , there is a

copula C, which is given by the expression:
Cðu; vÞ ¼ FðF11 ðuÞ; F21 ðvÞÞ; ð5Þ
where F11 and F21 are the inverse of the marginal distribution functions.
Vine copula
Vine copula represents a flexible and intuitive way of extending bivariate copulas to
higher dimensions (Aas et al. 2009). In the literature, C-vines, R-vines (Bedford and
Cooke 2001, 2002), and D-vines (Kurowicka 2005) have been proposed. In this
work, we focused on the D-vine estimation.8
According to Aas et al. (2009), D-vine estimation has density c given by
Y
n n1 Y
Y nj
Fðui juiþ1 ; . . .; uiþj1 Þ;
cðu1 ; . . .; un Þ ¼ fk ðuk Þ Ci;j ; ð6Þ
Fðuiþj juiþ1 ; . . .; uiþj1 Þ:
k¼1 j¼1 i¼1
where u1 ; . . .; un ¼ u 2 ½0; 1n are pseudo-observations; fk is the density function,

which is different for each asset; Cð; Þi;j is a bivariate copula density, which can be
different for each pair-copula.
8
To validate the choice of a D-vine, we compared the estimated model with its counterpart through the
test presented by Clarke (2007). We have not included the results here, but they are available upon
request.
The conditional distribution functions are computed with the following equation
(Joe 1996):
oCui ; uj juj fFðui juj Þ; Fðuj juj Þg
Fðui juÞ ¼ ; ð7Þ
oFðuj juj Þ
where i; j 2 N; i 6¼ j, Cui ; uj juj is the dependence structure of ui and uj bivariate
conditional distribution conditioned on uj (uj is the vector u excluding the
component uj ).
Nested Archimedean copulas
We have assumed the reader is familiar with bivariate Archimedean copulas. A

good review can be found in Nelsen (2006). Differently from elliptical copulas,
Archimedean copulas can capture different kinds of tail dependences; however, they
cannot capture asymmetry. A possibility is to model dependences using a
hierarchical structure composed by Archimedean copulas. This hierarchical
structure is called Nested Archimedean copulas. Details can be found in McNeil
(2008) and Hofert (2008).
For the n-dimensional case, the corresponding expression becomes
Cðu1 ; . . .; un Þ ¼ u1 1
n1;1 fun1;1 ðun Þ þ un1;1 oun2;1 fun2;1 ðun1 Þ
ð8Þ
þ un2;1 o. . .ou1
11 fu11 ðu1 Þ þ u11 ðu2 Þggg:
Here C refers to fully Nested Archimedean copulas, where all bivariate margins are
Archimedean copulas and u the corresponding generators.
DCC-GARCH
This description suggested by Engle (2002) presents a way to model the dynamic
processes of conditional volatilities and correlations simultaneously. Conditional
covariance matrix Ht can be defined as
pffiffiffiffiffiffi
Ht ¼ Dt Rt Dt ; Dt ¼ diagf ri;t g; ð9Þ
where Dt refers to the diagonal matrix of time-varying standard deviations (r of the
asset i and period t from univariate GARCH models).9 Rt is the correlation matrix
containing the conditional correlations coefficients, which can be defined as
Rt ¼ Q1
t Qt Q1
t ; ð10Þ
9
The elements of univariate GARCH were obtained through the structure given in (12). The model
P
GARCH(P,Q) can be described in the following manner: Xt ¼ lt þ rt zt ; r2t ¼ x þ Pp¼1 ap 2tp þ
PQ 2
q¼1 bq rtq , where lt represents the expectancy, zt is independent and identically distributed (i.i.d.) with
zero mean and unit variance, x is the constant, a the component ARCH, and b represents the component
GARCH. One can use other generalizations of the univariate GARCH model.
where
þ at1 0 þ bQt1 ;
Qt ¼ ð1 a þ bÞQ ð11Þ
t1
where Qt is the conditional covariance matrix of residuals with unconditional

covariance matrix Q obtained from univariate GARCH models; a and b are the non-
negative scalar parameters satisfying a þ b\1; Qt is a diagonal matrix containing
the square root of the diagonal elements of Qt .
Numerical procedures
We used Monte Carlo simulations to compare the performance of the models. The
number of Monte Carlo replications was set at 10,000. The length of analysed
samples was 1000 observations.10 All computational implementations were
conducted in the R programming language (R Core Team 2016). In each replicate,
we generated one-step-ahead forecast of risk measures.11
The generating process of the (log)-returns Xi;t , which represent the simulated
data, were drawn from AR(1)-GARCH(1,1) models. This type of specification is
commonly used to fit financial data, because it considers stylized facts of daily
returns as in Angelidis et al. (2007) and other works. The benchmark process
utilized to generate the returns is given by12
Xi;t ¼ 0:10Xi;t1 þ i;t
i;t ¼ ri;t zi;t ; zi;t i:i:d:Fð0; 1Þ; ð12Þ
r2i;t ¼ r2 ð1 0:10 0:85Þ þ 0:102i;t1 þ 0:85r2i;t ;
where, for time t and asset i, Xi;t is the return, r2i;t is the conditional variance, i;t is
the innovation in expectation, and zi;t is a white noise process with distribution
F. The unconditional volatility is 0.02.
For the analysis, eight scenarios were considered. Marginal distribution was
generated with Normal and Student’s t distribution with six degrees of freedom. The
Normal distribution was considered, because it is often utilized in stock market
analysis, and the Student’s t distribution was used to consider the heavy-tailed
behaviour of financial assets. We considered high and low association (correlation
of 0.20 and 0.80) among assets, obtained from the multivariate Normal and
Student’s t distributions. In addition, we generated portfolios with 4 and 16 assets.13
10
Studies show 1000 observations (four years of daily data) is a good sample size for daily data.
11
We focused on the 1-day-ahead forecast, because according to the literature, this is the horizon usually
used in empirical studies and simulations analysis.
12
Many works use similar values for the parameter, such as Christoffersen and Gonçalves (2005) and
Righi and Ceretta (2015a), because this data-generating process matches the daily returns obtained on the
S & P 500 Index.
13
We generated portfolios with 2n assets, where n ¼ 2; 3; 4. We have not presented the results of the
portfolio with 8 (23 ) assets, because the results are similar to those of the portfolios with 4 (22 ) and 16
To forecast the risk measures, we used HS, DCC-GARCH, and copula

methods.14 Regarding copula methods, after generating returns (considering the
eight scenarios described), we fitted the expectation l and dispersion components r
with AR(1)-GARCH(1,1) models15 to isolate the marginal behaviour. Then, we
transformed standardized residuals into pseudo-observations u 2 ½0; 1 by inversing
the distribution fitted to each of them. This procedure was needed because of the
definition of the copula functions. With this pseudo-information, we could estimate
the considered copula. Given the marginal and joint parameters already estimated,
we used the following algorithm to forecast the risk measures:
1. Obtain forecasts of the conditional mean li;tþ1 and standard deviation ri;tþ1 of
each asset through marginal models (AR(1)-GARCH(1,1)).
2. Simulate N ¼ 10; 000 samples16 ui;N with the size of 1000 for each asset
i through estimated Regular, Vine and Nested Archimedean copulas.
3. Convert each set of simulations ui;N to zi;N samples through the inversion of
their marginal probability, according to zi;N ¼ Fi1 ðui;N Þ.
4. For each asset i, determine the returns in accordance to the following
specification Xi;N;tþ1 ¼ li;N;tþ1 þ ri;tþ1 zi;N .
5. Compute portfolio returns as w0 XN , where w ¼ fw1 ; w2 ; . . .; wn g is the weights
vector and XN ¼ fX1;N;tþ1 ; X2;N;tþ1 ; . . .; Xn;N;tþ1 g represents the returns. In this
study, we assumed equally weighted portfolios.
6. Forecast VaRatþ1 ðw0 XN Þ, as the negative of a-quantile of the distribution of the
simulated portfolio returns w0 XN .
7. Forecast the ESatþ1 ðw0 XN Þ, which can be obtained by the negative mean of
portfolio returns w0 XN below the a-quantile (-VaRatþ1 ðw0 XN Þ).
8. Forecast the EVaRatþ1 ðw0 XN Þ, in accordance with Eq. (3).
This procedure is an extension of the proposed algorithm by Aas et al. (2011) to

forecast VaR and the algorithm utilized by Righi and Ceretta (2013a) to forecast ES.
It is similar to the procedure of filtered historical simulation (FHS) used for the
forecasting of risk measures. We considered a equal to 1% and 5%. These values are
the most common in the literature. In each replicate, we also performed the forecasts
of risk measures with HS and DCC-GARCH. In the approach based on HS, portfolio
returns were computed according to steps 5–8 of the algorithm. The only difference
was that the returns XN used were the returns obtained by the data-generated process
(12). For the approach based on DCC-GARCH, the difference in relation to the
algorithm presented was that, in steps 2–3, we obtained zi;t from the DCC-GARCH
Footnote 13 continued
(24 Þ assets. Portfolios with 25 or more assets are computationally complicated, due to the copulas
approach.
14
Regarding the parameters estimation, we used quasi-maximum likelihood (Q-ML) estimate.
15
The distribution used to fit the returns was the same as the scenario considered, to avoid marginal
interference (Fantazzini 2009).
16
In each replicate of Monte Carlo, we generated 10,000 samples ui;N with the size of 1000 for each asset
i. This procedure was repeated 10,000 times.
model, instead of using the copulas approach. To assess the performance of the
forecasts for all scenarios and each replicate, we computed A. Bias, R. Bias, and
RMSE from VaR, ES, and EVaR forecasts.17 Then, for each scenario, we calculated
the average values of these metrics. Similar metrics to identify the model with lower
model risk were used by Yao et al. (2006) and Telmoudi et al. (2016).
Numerical results
Based on the numerical procedures, we obtain numerical results of the performance

of VaR, ES, and EVaR for the different scenarios analysed. Tables 1, 2, 3, 4, 5, and
6 present the A. Bias, R. Bias, and RMSE for a ¼ 1% and 5% of one-step-ahead
forecasts of VaR, ES, and EVaR. Table 7 presents the summary of the best models
at each scenario. Results were analysed in the following order for a better
understanding: first, we analysed the results of the scenarios generated with Normal
distribution, second, we analysed the scenarios generated with Student’s t distribu-
tion, and finally, the behaviour of the absolute bias and relative bias was analysed.
Results for VaR, as shown in Tables 1 and 2, showed that, for scenarios with
Normal distribution, the best results were presented for Regular and Vine copulas.
We noted that the values of the criteria are close for these two types of copulas. An
interesting result was that, regardless of the value for a, number of assets, and
correlation, the best performances were presented by the Regular and Vine copulas.
Similar results were reported for estimates of ES (Tables 3, 4) and EVaR (Tables 5,
6), for the scenarios generated with Normal distribution.
Although Normal distribution is one of the most used, there is evidence that
financial data usually distance themselves from this distribution. In their seminal
work, Mandelbrot (1963) stated the Normal distribution is insufficient for modelling
financial data and their heavy-tailed behaviour (we refer to the work Cont 2001 as
an overview of stylized facts present in asset returns). It is usually observed that the
market returns display negative skewness, excess kurtosis, and structural shifts in
the distribution. Improper use of Gaussian models can lead to underestimation of
tail risk. This may lead to high losses without proper prevention of investor
portfolios (Brooks and Persand 2002; Tolikas 2014). A recent example was the
losses observed in the events of 2007 and 2008, which resulted among other factors
of the inappropriate use of Gaussian copulas. During periods of crisis, there is an
increase in the dependency between the assets, which is not properly captured by
Gaussian copulas. In financial data, use of Normal distribution can illustrate a
potential model risk, since it is an apparent model specification error (Miller et al.
2006).
Regarding scenarios generated with Student’s t distribution, we noticed that for
VaR, considering a = 1%, NAC presents lower values for the metrics considered.
This data generation process has more heavy tails and can generate more extreme
values than a Normal distribution. The implications of returns with heavy tails can
17
The method with the better performance would present the smallest value in at least two of these
metrics.
Table 1 Monte Carlo simulations results for VaR with four assets
a = 1% A. Bias R. Bias RMSE a = 5% A. Bias R. Bias RMSE
Normal distribution, low correlation

Vine copula -0.0004 -0.0146 0.0021 Vine copula -0.0001 -0.0063 0.0013
Nested copula -0.0039 -0.1463 0.0044 Nested copula -0.0021 -0.1261 0.0027
Regular copula -0.0003 -0.0124 0.0020 Regular copula -0.0001 -0.0051 0.0013
DCC-GARCH -0.0093 -0.2674 0.0119 DCC-GARCH -0.0084 -0.2848 0.0110
HS 0.0052 0.2845 0.0102 HS 0.0032 0.4099 0.0087
Normal distribution, high correlation
HS 0.0093 0.3410 0.0175 HS 0.0048 0.3866 0.0135
Student’s t distribution, low correlation
Vine copula 0.0246 0.2712 0.0376 Vine copula 0.0137 0.0798 0.0210
Nested copula 0.0089 0.0709 0.0302 Nested copula 0.0119 0.2308 0.0226
Regular copula 0.0241 0.2649 0.0409 Regular copula 0.0139 0.1392 0.0208
HS 0.0469 0.8997 0.2722 HS 0.0104 0.4464 0.0901
Student’s t distribution, high correlation
HS 0.1719 0.7657 0.4578 HS 0.0467 0.5908 0.2491
This table shows numerical results of the performance evaluation of multivariate models (Historical
Simulation (HS), Dynamic Conditional Correlation-Generalized Autoregressive Conditional
Heteroskedastic (DCC-GARCH), Regular copulas, Vine copulas, and Nested Archimedean copulas
(NAC)) for forecasting VaR of a portfolio with four assets. The results are based on 10,000 Monte Carlo
replications of length 1001 (1000 to consider the larger estimation window plus 1 for out-sample per-
formance). Data generation process of returns corresponds to AR(1)-GARCH(1,1), considering Normal
and Student’s t distribution, and low and high correlation between the assets. The performance of the
forecasts was analysed using absolute bias (A. Bias), relative bias (R. Bias), and root-mean-square error
(RMSE). The values in bold correspond to the model presented 2 or 3 with lower value criteria
be drastic to a risk manager. This is more pronounced in high quantile (Bradley and
Taqqu 2003). Moreover, Angelidis et al. (2004) showed that leptokurtic distribu-
tions, especially the Student’s t, are more appropriate than Normal for VaR
forecasting.
For the significance level of 5% and VaR forecasts, we observed NAC were
superior. In the scenario with four assets and low correlation, Vine copulas
exhibited the lowest value for criteria. When we considered a portfolio of 16 assets
and high correlation, estimates of VaR obtained with the DCC-GARCH model
demonstrated better performance, as noted in Table 2. Regarding ES in scenarios
Table 2 Monte Carlo simulations results for VaR with 16 assets


Regular copula -0.0001 -0.0035 0.0013 Regular copula 0.0002 0.0035 0.0008
HS 0.0041 0.2825 0.0073 HS 0.0028 0.6397 0.0067
Regular copula -0.0006 -0.0184 0.0023 Regular copula -0.0001 0.0056 0.0015
HS 0.0084 0.3464 0.0168 HS 0.0044 0.4377 0.0136
Nested copula 0.0048 -0.0012 0.0426 Nested copula 0.0140 0.4011 0.0249
HS 0.0239 0.6622 0.1376 HS 0.0044 0.8888 0.0864
Nested copula -0.0120 -0.1188 0.0161 Nested copula 0.0167 0.2540 0.0245
HS 0.0692 1.1725 0.1729 HS 0.0215 0.8544 0.0720
(NAC)) for forecasting VaR of a portfolio with 16 assets. The results are based on 10,000 Monte Carlo
with Student’s t distribution, only in the scenario with high association, the portfolio
with 16 assets, and a significance level of 1%, the DCC-GARCH model presents
best values for the analysed criteria (Table 4). In other cases for ES, NAC also
showed a better performance. In relation to EVaR, we realized, in most scenarios,
Nested Archimedean copulas outperformed other models. In general, we realized
NAC, among the analysed models, for the scenarios generated with Student’s
t distribution, was the best method for forecasting risk measures.
Table 3 Monte Carlo simulations results for ES with four assets


HS 0.0067 0.2845 0.0115 HS 0.0046 0.3156 0.0096
HS 0.0132 0.3747 0.0211 HS 0.0079 0.3567 0.0159
HS 0.0736 0.9866 0.4244 HS 0.0331 0.8321 0.1940
HS 0.2667 0.8294 0.6621 HS 0.1273 0.7275 0.3748
(NAC)) for forecasting ES of a portfolio with four assets. The results are based on 10,000 Monte Carlo
We observed the HS did not perform better in any of the analysed scenarios. This
result was more pronounced in scenarios with Student’s t distribution. Problems
with HS are presented in Pritsker (2006), who suggested this method responds
slowly to changes in volatility and large price movements, features present in
financial returns. Corroborating with the results, Christoffersen and Gonçalves
(2005) argued the HS model results in bad point estimates for VaR and ES and poor
confidence intervals. This result is troubling. According to the results identified by
Table 4 Monte Carlo simulations results for ES with 16 assets


HS 0.0049 0.2616 0.0081 HS 0.0037 0.3292 0.0070
HS 0.0119 0.3729 0.0199 HS 0.0071 0.3704 0.0155
HS 0.0456 0.7419 0.0869 HS 0.0226 0.6335 0.0492
HS 0.1029 1.2898 0.2441 HS 0.0523 1.0416 0.1382
(NAC)) for forecasting ES of a portfolio with 16 assets. The results are based on 10,000 Monte Carlo
Pérignon and Smith (2010), about 73% of institutions that disclose their VaR use
this procedure. Estimation errors of this method accumulate and become large.
When employing this method to compose a portfolio, the manager may have
difficulties identifying most efficient asset allocation.
We realized the absolute bias and the relative bias of the risk measures with HS
were positive. This result was stronger in scenarios with Student’s t distribution.
The HS method overestimates the market risk. This can cause higher capital
requirements that could be applied more profitably. A trader may also be required to
Table 5 Monte Carlo simulations results for EVaR with 4 assets


HS 0.0039 0.3699 0.0086 HS 0.0024 0.4759 0.0077
Regular copula -0.0001 0.0026 0.0023 Regular copula 0.0001 0.0297 0.0016
DCC-GARCH -0.0072 -0.0978 0.0135 DCC-GARCH -0.0060 0.1262 0.0125
HS 0.0063 0.4131 0.0145 HS 0.0032 0.8001 0.0123
HS 0.0350 1.0684 0.1383 HS 0.0139 1.4188 0.0687
HS 0.0569 1.2837 0.2196 HS 0.0218 1.3672 0.1118
(NAC)) for forecasting EVaR of a portfolio with four assets. The results are based on 10,000 Monte Carlo
rebalance his/her portfolio at an inopportune time. Concerning the absolute bias and
relative bias of the DCC-GARCH model, in most scenarios, they are negative,
especially in scenarios with Normal distribution. In such cases, this model
underestimates the risk, which can be dangerous and lead the institution to
bankruptcy. The portfolio manager, when using DCC-GARCH to monitor their
portfolio, will underestimate the probability of events of tail risk, being sub-optimal
for investors that have as their primary objective the minimization of risk. In this
case, the portfolio composition may not be appropriate for their risk tolerance,
Table 6 Monte Carlo simulations results for EVaR with 16 assets


Regular copula -0.0002 -0.0145 0.0011 Regular copula 0.0001 0.0002 0.0007
HS 0.0031 0.3490 0.0067 HS 0.0023 0.8799 0.0063
HS 0.0054 0.2406 0.0113 HS 0.0027 0.7221 0.0097
HS 0.0196 0.8163 0.0439 HS 0.0085 1.2112 0.0288
HS 0.0468 1.2111 0.1093 HS 0.0208 1.5755 0.0607
(NAC)) for forecasting EVaR of a portfolio with 16 assets. The results are based on 10,000 Monte Carlo
besides compromising their investment strategies. Corroborating, Hwang and

Pereira (2006) and Carnero et al. (2007) identified the estimates of the univariate
GARCH model usually have a negative bias. Given the bias identified in this
model’s estimators, Fantazzini (2009) noted that the VaR model estimated with
GARCH has a poor performance, especially for small sample sizes.
For the estimated risk measures through copulas, we noticed a pattern in the sign
of bias. Data generated with Normal distribution displayed negative bias. Regarding
scenarios generated with Student’s t distribution, we noted a positive bias for the
Table 7 Summary of the best models for risk measure estimation at each scenario
Scenario/measure VaR ES EVaR
Significance level 1%
Normal, low correlation, 4 assets Vine, Regular Vine, Regular Vine, Regular
Normal, high correlation, 4 assets Vine, Regular Vine, Regular Vine, Regular
Normal, high correlation, 16 assets Vine Vine, Regular Vine, Regular
Student, low correlation, 4 assets NAC NAC NAC
Student, high correlation, 4 assets NAC NAC NAC
Student, low correlation, 16 assets NAC NAC NAC
Student, high correlation, 16 assets NAC DCC NAC
Significance level 5%
Student, low correlation, 4 assets Vine NAC NAC
Student, high correlation, 4 assets NAC NAC NAC
Student, low correlation, 16 assets NAC NAC DCC
Student, high correlation, 16 assets DCC NAC NAC, DCC
This table provides summarized numerical results regarding performance evaluation of multivariate
models for forecasting VaR, ES, and EVaR of portfolios with 4 and 16 assets. The results are based on
10,000 Monte Carlo replications of length 1001 (1000 to consider the larger estimation window plus 1 for
out-sample performance). Data generation process of returns corresponds to AR(1)-GARCH(1,1), con-
sidering Normal and Student’s t distribution, and low and high correlation between the assets. The
multivariate models used were Historical Simulation (HS), Dynamic Conditional Correlation-Generalized
Autoregressive Conditional Heteroskedastic (DCC-GARCH), Regular copulas, Vine copula, and Nested
Archimedean copulas (NAC)
measures, except for some scenarios, where the estimates obtained by NAC
presented negative relative bias. See, for example, Table 3 with high correlation and
a = 1%. In this way, in periods of lull (normal distribution), the copulas method
results in small risk estimates, resulting in less protection for the portfolios (negative
bias), underestimating the market risk. However, in the most turbulent periods
(Student’s t distribution), copulas result in risk measures more parsimonious,
reflecting greater protection. We emphasize that, in general, the relative bias of
estimated measures with copulas is small.
We can highlight the results provided evidence of the superiority of copula
models over the Historical and DCC-GARCH methods to forecast risk measures.
The data generated in the process simulation present marginals with normal and
Student’s t distribution and linear dependencies, which should favour the DCC-
GARCH model. Some studies have discussed the superiority of copulas over the
correlation approach, because they offer more flexibility to model the dependence
between variables (Kole et al. 2007). Corroborating with our results, Chen and Tu
(2013) demonstrated, under the coverage rates of 95 and 99%, the copula-based
VaR model demonstrates performance superior to both the constant conditional

correlation-generalized autoregressive conditional heteroskedasticity model (CCC-
GARCH) (Bollerslev 1990) and the DCC-GARCH model. However, differently
from what we found, Weiß (2013), when analysing a large number of portfolios,
identified that the DCC-GARCH does not lose to any of the considered copula
(elliptical copulas, such as Gaussian and Student’s t, and Archimedean copulas,
such as Clayton, Frank, and Gumbel) models in terms of VaR and ES forecasting.
One of the possible explanations for this is that the author analysed bivariate
portfolios, and in our work, we considered portfolios with 4 and 16 assets. In
addition, we point out that using copulas will reduce the model risk,18 as also noted
by Chen and Tu (2013) for VaR.
Overall, risk measures are sensitive to the choice of model. By making use of a
different model or a different parameterization, the risk manager or regulator is
susceptible to different results. For instance, consider the determination of banks’
capital requirement, which is obtained directly by the point estimates of risk
measure multiplied by portfolio value. Using an inappropriate model, HS or DCC-
GARCH, may cause insufficient quantities of capital, which may be insufficient to
absorb losses from unexpected impacts, especially in times of crisis, or can be
calculated at a high capital requirement quantity, which could be applied to more
profitable investments. Our results can assist the portfolio manager in selecting the
most suitable model to manage their portfolio. The choice of model with lower
model risk, besides reducing financial losses, can avoid poor strategic decision-
making and damaging the reputation of the institution. Another practical implication
to use a model with lower model risk is to reduce the problem of regulatory
arbitrage and model misspecification, approached by Kellner and Rösch (2016).19
We noticed similar behaviour in the estimation methods in the two sizes of
portfolios. Regarding the a analysed, we realized, in general, there is no pattern for
the value of the relative bias. The results of Chen and Tu (2013) suggested the risk
management models should apply a smaller nominal coverage rate (95% instead of
99%) to reduce model risk. Differently from our work, the author considered only
the VaR in their analysis.
According to our results, models with good performance in estimating one of the
three measures assessed might provide accurate measurements of the VaR, ES, or
EVAR in a multivariate perspective. This result is most evident when we analyse
Table 7, which summarizes the results identified. Zhou (2012), Degiannakis et al.
(2013), and Righi and Ceretta (2015a) also identified a similar performance of
analysed estimators to measure the VaR and ES. Unlike our research, they restricted
their analysis to univariate models. This may be related with the elicitable. ES is not
elicitable, but ES can in practice be jointly elicited with VaR (Acerbi and Szekely
2014). EVaR is elicitable; however, there are no studies that examine the
18
In this paper, we considered a model presents a greater model risk if the estimate measure with this
model presents greater relative bias.
19
Institutions that present the same portfolio and use different internal models, approved by the
regulator, must hold the same or at least almost the same amount of regulatory capital, which rarely
happens, even if institutions use models that pass in backtesting. This problem is referred to as regulatory
arbitrage.
performance of models to forecast this measure. Another possible reason for the
similar performance models is the three measures are tail risk measures, although
they present conceptual differences. VaR can be represented as the quantile of the
distribution of profit and loss, while ES considers the magnitude of losses beyond
the quantile, instead of only the quantile of interest. Differently to VaR and ES,
EVaR is associated with expectiles, being more sensitive to the extreme values than
VaR. Corroborating with our results, Pfingsten et al. (2004) noted that risk measures
of the same group deliver similar risk rankings in their analysis. Differently,
Marinelli et al. (2007), to compare with a backtesting study the performance of
univariate models for VaR and ES based on stable laws and on extreme value theory
(EVT), observed distinct performance models to measure VaR and ES.
Final considerations
In this paper, we evaluated the performance of models for forecasting VaR, ES, and
EVaR from the multivariate perspective. To assess the performance of the models,
we used Monte Carlo simulations. The models considered were HS, DCC-GARCH,
Regular copulas, Vine copulas, and NAC. To evaluate the quality of forecasts, we
calculated absolute bias, relative bias, and RMSE. Different scenarios regarding
marginal distributions, correlation, and the number of assets in portfolios were
considered in the study.
The numerical results showed that performing the estimators is associated with
marginal distribution. When the marginal distribution is Gaussian, Regular and Vine
copulas performed better in predicting the risk measures considered. For data which
follow Student’s t distribution, the NAC showed the best performance. Although
Vine copulas have been used more in finance literature, our results suggest which
NAC are a good alternative to be considered to forecast risk measures in the
multivariate context. We realized the copula method offers better performance than
traditional methods, such as HS and DCC-GARCH. The better performance of the
forecasts obtained through copulas may be explained by the fact that they allow the
dependence structure to be identified and allow to capture the possible non-linear
relationships among variables.
We also noticed the HS method tends to overestimate the risk, i.e. exhibited
positive bias, besides presenting with the worst performance. Forecasts of the risk
measures obtained with copulas, in the scenarios generated with Student’s
t distribution, also exhibited positive bias. Referring to the DCC-GARCH model,
we observed the risk forecasts have a negative bias. The same was observed for the
measures obtained with copulas in scenarios with Normal distribution. Here, the
forecasts are more aggressive, underestimating the risk.
In addition, the performance of the models to forecast risk measures does not
change with respect to the size of portfolios (number of assets), the value of
correlation, and the significance level. Our analysis also demonstrated that the
performance of the models does not change according to the measure. However, we
emphasize the models adjust the data, and when choosing the measure, we should
consider the theoretical definition of each one. For future research, we suggest the
use of Vine copulas and NAC in empirical applications of financial data for risk
management. In accordance with our results, we believe they will improve risk
control and reduce losses.
Acknowledgements We gratefully acknowledge the partial financial support from Coordenação de

Aperfeiçoamento de Pessoal de Nı́vel Superior (CAPES), Brazil.
References
Aas, K., and D. Berg. 2009. Models for construction of multivariate dependence–A comparison study.
The European Journal of Finance 15 (7–8): 639–659.
Aas, K., D. Berg, and D. Kurowicka. 2011. Modeling dependence between financial returns using pair-
copula constructions. In Dependence modeling: Vine copula handbook, ed. D. Kurowicka, and H.
Joe, 305. Hackensack, NJ: World Scientific.
Aas, K., C. Czado, A. Frigessi, and H. Bakken. 2009. Pair-copula constructions of multiple dependence.
Insurance: Mathematics and economics 44 (2): 182–198.
Acerbi, C., and B. Szekely. 2014. Backtesting expected shortfall. Risk Magazine 27: 76–81.
Acerbi, C., and D. Tasche. 2002. On the coherence of expected shortfall. Journal of Banking & Finance
26 (7): 1487–1503.
Angelidis, T., A. Benos, and S. Degiannakis. 2004. The use of GARCH models in VaR estimation.
Statistical Methodology 1 (1–2): 105–128.
Angelidis, T., A. Benos, and S. Degiannakis. 2007. A robust VaR model under different time periods and
weighting schemes. Review of Quantitative Finance and Accounting 28 (2): 187–201.
Artzner, P., F. Delbaen, J.-M. Eber, and D. Heath. 1999. Coherent measures of risk. Mathematical
Finance 9 (3): 203–228.
Basel Committee on Banking Supervision. 2009. Revisions to the Basel II market risk framework. BCBS:
Bank for international settlements. Basel.
Bedford, T., and R.M. Cooke. 2001. Probability density decomposition for conditionally dependent
random variables modeled by vines. Annals of Mathematics and Artificial Intelligence 32 (1–4):
245–268.
Bedford, T., and R.M. Cooke. 2002. Vines: A new graphical model for dependent random variables.
Annals of Statistics 30 (4): 1031–1068.
Bellini, F., and E.D. Bernardino. 2017. Risk management with expectiles. The European Journal of
Finance 23 (6): 487–506.
Bellini, F., B. Klar, A. Müller, and E.R. Gianin. 2014. Generalized quantiles as risk measures. Insurance:
Mathematics and Economics 54: 41–48.
Bollerslev, T. 1990. Modelling the coherence in short-run nominal exchange rates: A multivariate
generalized ARCH model. The Review of Economics and Statistics 72 (3): 498–505.
Bradley, B.O., and M.S. Taqqu. 2003. Financial risk and heavy tails. In Handbook of heavy-tailed
distributions in finance, ed. S.T. Rachev, 35–103. Amsterdam: Elsevier.
Brechmann, E.C., and C. Czado. 2013. Risk management with high-dimensional vine copulas: An
analysis of the Euro Stoxx 50. Statistics & Risk Modeling 30 (4): 307–342.
Brooks, C., and G. Persand. 2002. Model choice and value-at-risk performance. Financial Analysts
Journal 58 (5): 87–97.
Carnero, M.A., D. Peña, and E. Ruiz. 2007. Effects of outliers on the identification and estimation of
GARCH models. Journal of Time Series Analysis 28 (4): 471–497.
Chen, Y.-H., and A.H. Tu. 2013. Estimating hedged portfolio value-at-risk using the conditional copula:
An illustration of model risk. International Review of Economics & Finance 27: 514–528.
Cherubini, U., E. Luciano, and W. Vecchiato. 2004. Copula methods in finance. New York: Wiley.
Christoffersen, P., and S. Gonçalves. 2005. Estimation risk in financial risk management. The Journal of
Risk 7 (3): 1.
Clarke, K.A. 2007. A simple distribution-free test for nonnested model selection. Political Analysis 15
(3): 347–363.
Cont, R. 2001. Empirical properties of asset returns: Stylized facts and statistical issues. Quantitative
Finance 1 (2): 223–236.
Degiannakis, S., C. Floros, and P. Dent. 2013. Forecasting value-at-risk and expected shortfall using
fractionally integrated models of conditional volatility: International evidence. International Review
of Financial Analysis 27: 21–33.
Delbaen, F. 2013. A remark on the structure of expectiles. arXiv preprint arXiv:1307.5881.
Duffie, D., and J. Pan. 1997. An overview of value at risk. The Journal of Derivatives 4 (3): 7–49.
Emmer, S., M. Kratz, and D. Tasche. 2015. What is the best risk measure in practice? A comparison of
standard measures. Journal of Risk 18 (2): 31–60.
Engle, R. 2002. Dynamic conditional correlation: A simple class of multivariate generalized
autoregressive conditional heteroskedasticity models. Journal of Business & Economic Statistics
20 (3): 339–350.
Fantazzini, D. 2009. The effects of misspecified marginals and copulas on computing the value at risk: A
Monte Carlo study. Computational Statistics & Data Analysis 53 (6): 2168–2188.
Föllmer, H., and A. Schied. 2002. Convex measures of risk and trading constraints. Finance and
Stochastics 6 (4): 429–447.
Glasserman, P., and X. Xu. 2014. Robust risk measurement and model risk. Quantitative Finance 14 (1):
29–58.
Hofert, M. 2008. Sampling archimedean copulas. Computational Statistics & Data Analysis 52 (12):
5163–5174.
Hwang, S., and P.L.V. Pereira. 2006. Small sample properties of GARCH estimates and persistence. The
European Journal of Finance 12 (6–7): 473–494.
Jäschke, S. 2014. Estimation of risk measures in energy portfolios using modern copula techniques.
Computational Statistics & Data Analysis 76: 359–376.
Joe, H. 1996. Families of m-variate distributions with given margins and m (m-1)/2 bivariate dependence
parameters. Lecture Notes-Monograph Series 28: 120–141.
Joe, H. 1997. Multivariate models and multivariate dependence concepts. Boca Raton: CRC Press.
Jorion, P. 2006. Value at risk: The new benchmark for managing financial risk, 3rd ed. New York:
McGraw-Hill.
Junker, M., and A. May. 2005. Measurement of aggregate risk with copulas. Econometrics Journal 8 (3):
428–454.
Kellner, R., and D. Rösch. 2016. Quantifying market risk with value-at-risk or expected shortfall?
Consequences for capital requirements and model risk. Journal of Economic Dynamics and Control
68: 45–63.
Kole, E., K. Koedijk, and M. Verbeek. 2007. Selecting copulas for risk management. Journal of Banking
& Finance 31 (8): 2405–2423.
Kuester, K., S. Mittnik, and M.S. Paolella. 2006. Value-at-risk prediction: A comparison of alternative
strategies. Journal of Financial Econometrics 4 (1): 53–89.
Kurowicka, D. 2005. Distribution-free continuous Bayesian belief. Modern Statistical and Mathematical
Methods in Reliability 10: 309.
Kurowicka, D. 2011. Dependence modeling: Vine copula handbook. Singapore: World Scientific.
Low, R.K.Y., J. Alcock, R. Faff, and T. Brailsford. 2013. Canonical vine copulas in the context of modern
portfolio management: Are they worth it? Journal of Banking & Finance 37 (8): 3085–3099.
Mandelbrot, B.B. 1963. The variation of certain speculative prices. Journal of Business 36: 394–419.
Manganelli, S., and R.F. Engle. 2001. Value at risk models in finance. Working paper.
Marinelli, C., S. D’addona, and S.T. Rachev. 2007. A comparison of some univariate models for value-at-
risk and expected shortfall. International Journal of Theoretical and Applied Finance 10 (06):
1043–1075.
McNeil, A.J. 2008. Sampling nested Archimedean copulas. Journal of Statistical Computation and
Simulation 78 (6): 567–581.
Mendes, B.V.M., and R.M. de Souza. 2004. Measuring financial risks with copulas. International Review
of Financial Analysis 13 (1): 27–45.
Miller, D.J., W.-H. Liu, et al. 2006. Improved estimation of portfolio value-at-risk under copula models
with mixed marginals. Journal of Futures Markets 26 (10): 997–1018.
Nelsen, R.B. 2006. An introduction to copulas, 2nd ed. Berlin: Springer.
Newey, W.K., and J.L. Powell. 1987. Asymmetric least squares estimation and testing. Econometrica 55
(4): 819–847.
Palaro, H.P., and L.K. Hotta. 2006. Using conditional copula to estimate value at risk. Journal of Data
Science 4: 93–115.
Pérignon, C., and D.R. Smith. 2010. The level and quality of value-at-risk disclosure by commercial
banks. Journal of Banking & Finance 34 (2): 362–377.
Pfingsten, A., P. Wagner, and C. Wolferink. 2004. An empirical investigation of the rank correlation
between different risk measures. The Journal of Risk 6 (4): 55.
Pritsker, M. 2006. The hidden dangers of historical simulation. Journal of Banking & Finance 30 (2):
561–582.
R Core Team. 2016. R: A language and environment for statistical computing. Vienna: R Foundation for
Statistical Computing. https://www.R-project.org/.
Righi, M.B., and D. Borenstein. 2017. A simulation comparison of risk measures for portfolio
optimization. Finance Research Letters. doi:10.1016/j.frl.2017.07.013.
Righi, M.B., and P.S. Ceretta. 2013a. Pair copula construction based expected shortfall estimation.
Economics Bulletin 33 (2): 1067–1072.
Righi, M.B., and P.S. Ceretta. 2013b. Risk prediction management and weak form market efficiency in
Eurozone financial crisis. International Review of Financial Analysis 30: 384–393.
Righi, M.B., and P.S. Ceretta. 2015a. A comparison of expected shortfall estimation models. Journal of
Economics and Business 78: 14–47.
Righi, M.B., and P.S. Ceretta. 2015b. Forecasting value at risk and expected shortfall based on serial pair-
copula constructions. Expert Systems with Applications 42 (17–18): 6380–6390.
Righi, M.B., and P.S. Ceretta. 2016. Shortfall deviation risk: An alternative for risk measurement. Journal
of Risk 19 (2): 81–116.
Sklar, A. 1959. Fonctions de répartition á n dimensions et leurs marges. lInstitut de Statistique de
LUniversité de Paris 8: 229–231.
Telmoudi, F., E.L. Ghourabi, and M. Limam. 2016. On conditional risk estimation considering model
risk. Journal of Applied Statistics 43 (8): 1386–1399.
Tolikas, K. 2014. Unexpected tails in risk measurement: Some international evidence. Journal of Banking
& Finance 40: 476–493.
Weiß, G.N.F. 2013. Copula-GARCH versus dynamic conditional correlation: An empirical study on VaR
and ES forecasting accuracy. Review of Quantitative Finance and Accounting 41 (2): 179–202.
Yao, J., Z.-F. Li, and K.W. Ng. 2006. Model risk in VaR estimation: An empirical study. International
Journal of Information Technology & Decision Making 5 (03): 503–512.
Zhou, J. 2012. Extreme risk measures for reits: A comparison among alternative methods. Applied
Financial Economics 22 (2): 113–126.
Ziegel, J.F. 2014. Coherence and elicitability. Mathematical Finance 26 (4): 901–918.

Muller e Righi 2018

Uploaded by

Copyright:

Available Formats

Muller e Righi 2018

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

Muller e Righi 2018

Uploaded by

Copyright:

Available Formats

Risk Manag

Numerical comparison of multivariate models

Fernanda Maria Müller1 • Marcelo Brutti Righi1

Ó Macmillan Publishers Ltd 2017

Abstract We evaluated the performance of multivariate models for forecasting

Keywords Risk measures DCC-GARCH Copulas Model risk Monte Carlo

& Fernanda Maria Müller

ESa ðXÞ ¼ E½XjX  FX1 ðaÞ: ð2Þ

To facilitate the notation, we focused on the bivariate case. Extensions to the n-

Given that C is a copula and F1 and F2 univariate distribution functions, C is unique

For a two-dimensional distribution function F with marginal F1 and F2 , there is a

Cðu; vÞ ¼ FðF11 ðuÞ; F21 ðvÞÞ; ð5Þ

where u1 ; . . .; un ¼ u 2 ½0; 1n are pseudo-observations; fk is the density function,

Nested Archimedean copulas

We have assumed the reader is familiar with bivariate Archimedean copulas. A

where Qt is the conditional covariance matrix of residuals with unconditional

To forecast the risk measures, we used HS, DCC-GARCH, and copula

This procedure is an extension of the proposed algorithm by Aas et al. (2011) to

Based on the numerical procedures, we obtain numerical results of the performance

Normal distribution, low correlation

Table 2 Monte Carlo simulations results for VaR with 16 assets

Normal distribution, low correlation

Table 3 Monte Carlo simulations results for ES with four assets

Normal distribution, low correlation

Table 4 Monte Carlo simulations results for ES with 16 assets

Normal distribution, low correlation

Table 5 Monte Carlo simulations results for EVaR with 4 assets

Normal distribution, low correlation

Table 6 Monte Carlo simulations results for EVaR with 16 assets

Normal distribution, low correlation

besides compromising their investment strategies. Corroborating, Hwang and

VaR model demonstrates performance superior to both the constant conditional

Acknowledgements We gratefully acknowledge the partial financial support from Coordenação de

You might also like

ESa ðXÞ ¼ E½XjX FX1 ðaÞ: ð2Þ