You are on page 1of 17

Journal of Econometrics 227 (2022) 151–167

Contents lists available at ScienceDirect

Journal of Econometrics
journal homepage: www.elsevier.com/locate/jeconom

Efficient estimation of high-dimensional dynamic covariance


by risk factor mapping: Applications for financial risk
management

Mike K.P. So a , , Thomas W.C. Chan a , Amanda M.Y. Chu b
a
Department of Information Systems, Business Statistics and Operations Management, The Hong Kong University of Science and
Technology, Clear Water Bay, Kowloon, Hong Kong
b
Department of Social Sciences, The Education University of Hong Kong, Tai Po, New Territories, Hong Kong

article info a b s t r a c t

Article history: This paper aims to explore a modified method of high-dimensional dynamic variance–
Received 25 September 2018 covariance matrix estimation via risk factor mapping, which can yield a dependence
Received in revised form 21 April 2020 estimation of asset returns within a large portfolio with high computational efficiency.
Accepted 29 April 2020
The essence of our methodology is to express the time-varying dependence of high-
Available online 3 June 2020
dimensional return variables using the co-movement concept of returns with respect to
Keywords: risk factors. A novelty of the proposed methodology is to allow mapping matrices, which
Dynamic covariance modeling govern the co-movement of returns, to be time-varying. We also consider the flexible
Dynamic mapping modeling of risk factors by a copula multivariate generalized autoregressive conditional
Multivariate GARCH heteroscedasticity (MGARCH) model. Through the proposed risk factor mapping model,
Risk contribution the number of parameters and the time complexity are functions of a small number
Tail risk of risk factors instead of the number of stocks in the portfolio, making our proposed
methodology highly scalable. We adopt Bayesian methods to estimate unknown param-
eters and various risk measures in the proposed model. The proposed risk mapping
method and financial applications are demonstrated by an empirical study of the Hong
Kong stock market. The assessment of the effectiveness of the mapping via risk measure
estimation is also discussed.
© 2020 Elsevier B.V. All rights reserved.

1. Introduction

Variance–covariance structure plays a central role in risk management and investment. Under the assumption of
elliptical distributions, different risk measures of a portfolio, such as value-at-risk (VaR) or expected shortfall (ES), can be
derived analytically, given the variance–covariance structure of the underlying assets. The classical Markowitz’s portfolio
selection problem asserts that the optimal portfolio weight vector of the portfolio can be solved using the variance–
covariance matrix of the assets. In the univariate cases, numerous volatility models have been proposed to describe
the variances of economic data; for example, the ARCH model by Engle (1982) and its generalizations in Bollerslev
(1986), Nelson (1991), Glosten et al. (1993) and Hamilton and Susmel (1994).
In recent decades, univariate volatility models have been modified in a multivariate setting. A parsimonious multi-
variate GARCH (MGARCH) model was introduced by Bollerslev (1990), in which constant correlations among different
variables were assumed. Further developments of this model were carried out by Engle and Kroner (1995a), Ding

∗ Corresponding author.
E-mail address: immkpso@ust.hk (M.K.P. So).

https://doi.org/10.1016/j.jeconom.2020.04.040
0304-4076/© 2020 Elsevier B.V. All rights reserved.
152 M.K.P. So, T.W.C. Chan and A.M.Y. Chu / Journal of Econometrics 227 (2022) 151–167

(1994), Alexander and Chibumba (1997), Klaassen (1999), Engle and Sheppard (2001), Tse and Tsui (2002), Engle (2002),
and Engle and Kelly (2012). In particular, Engle (2002) and Tse and Tsui (2002) allowed the entire correlation matrix to
follow an autoregressive process. In addition, McAleer et al. (2008) developed a generalized autoregressive conditional
correlation model for a random coefficient vector autoregressive standardized residual process. McAleer et al. (2009)
proposed a vector ARMA-asymmetric GARCH model. Audrino and Trojani (2011) and So and Yip (2012) presented a new
multivariate GARCH model by having threshold nonlinearity in the conditional volatilities and correlations of multivariate
time series. As an alternative of the traditional MGARCH modeling, Asai et al. (2006) reviewed the idea of stochastic
volatility (SV) to describe the empirical features of financial returns. Chib et al. (2009) gave a summary of the modern
development of multivariate stochastic volatility modeling. Correlation modeling with SV was discussed in Asai and
McAleer (2009), where both constant and dynamic correlation models were introduced. Further modeling features were
incorporated in an orthogonal GARCH model (Alexander, 2001), a factor GARCH (Vrontos et al., 2003) and a portfolio index
approach (Asai and McAleer, 2008), with the threshold nonlinearity property (So and Choi, 2009) and with correlation
clustering behaviors (So and Yip, 2012).
In the modeling multivariate financial time series, one big challenge is the issue of dimensionality. A typical portfolio
managed by hedge funds or financial institutions may consist of hundreds of assets. However, a direct evaluation of
the variance–covariance matrix is computationally inefficient, as its size grows quadratically with the number of assets,
except in those restrictive models. Estimating parameters in an MGARCH model, using the method of maximum likelihood,
requires the computation of inverse of the covariance matrix at each step of estimation, and such a gigantic matrix makes
this procedure computationally expensive and time-consuming. This issue is even more pronounced when one attempts
to produce a dynamic risk estimation of a large portfolio.
Research works have been conducted on MGARCH model estimation in a high dimension setting. Engle (2002) and Tse
and Tsui (2002) divided the model estimation into two steps, the volatility estimation and correlation estimation. Other
estimation approaches of MGARCH models include the equation-by-equation estimation (Francq and Zakoïan, 2016),
the weighted quasi-likelihood (Ling, 2007), the Student-t quasi-MLE (Fiorentini et al., 2003), and the variance targeting
estimation (Engle and Mezrich, 1996; Francq et al., 2014).
Another challenge in multivariate volatility modeling is the capturing of fat-tailed properties of financial returns
and the nonlinear dependence between returns. Researchers proposed various non-Gaussian distributions, such as t-
distribution or mixture distribution for modeling fat-tailed behaviors by Paolella (2015). Copula-based GARCH was
also considered to model nonlinear dependence. Jondeau and Rockinger (2006) proposed a Gaussian copula and t-
copula with skewed t as the error distribution. Lee and Long (2009) developed a copula-GARCH model with dynamic
correlation. Riccetti (2013) constructed a portfolio with macro asset allocation. Messaoud and Aloui (2015) applied a
GJR-GARCH copula to financial data. So and Yeung (2014) and Acar et al. (2019) modeled dynamic dependence using
vine-copula GARCH models. Tang et al. (2016) adopted copula-DCC-GARCH to model volatility in the S&P500 index. This
paper proposes capturing the fat-tailed properties and nonlinear dependence via a meta-t copula structure in risk factors.
The main objectives of this paper are to propose a new methodology to efficiently estimate the variance–covariance
structure of a vast number of asset returns, and the associated financial applications, through time-varying beta
coefficients linking asset returns and risk factors, as well as the use of correlated errors to account for possible return
co-movement that cannot be explained by the risk factors, and copula modeling of risk factors to capture the nonlinear
dependence of returns. We adopt the concept of risk mapping to simplify the estimation procedure. Instead of treating
each individual asset as the source of randomness, we assume all of them are driven by the same set of variables that are
known as risk factors. The main advantage of our research is the application of MGARCH modeling to a small number of
selected risk factors, which can yield an efficient estimation.
The process of mapping generally involves three stages. The first is the construction of a set of benchmark instruments
or risk factors. Benchmark instruments might include key bonds, equities, commodities, and so on. We then derive
synthetic substitutes for each instrument we hold, which are made up of positions in the core instruments. This synthetic
substitution is the actual mapping. The final stage is to estimate risk measures using the mapped instruments (i.e., the
synthetic substitutes) instead of the actual instruments we hold. The main features of the proposed high-dimensional
dynamic covariance estimation are: (1) using risk factors for risk mapping to facilitate efficient calculation; (2) enabling
time-varying beta coefficients for dynamic mapping; and (3) modeling risk factor returns using a t-copula to capturing
potential extremal dependence.
The organization of this paper is as follows. Section 2 provides a discussion of the risk factor mapping and its
formulation. The portfolio variance is decomposed into two parts: a systematic risk component representing the variance–
covariance matrix of the risk factors and the beta coefficients vector governing the connection between each individual
asset and the risk factors, and a non-systematic risk component allowing serial dependence between mapping errors
of asset returns. Section 3 presents three financial applications based on risk factor mapping. Section 4 describes the
simulation setting and shows the results of the efficiency of the Bayesian estimation to the proposed model by imitating
the real-world data. Section 5 presents an empirical study in the Hong Kong stock market with 100 listed companies
where we demonstrate the practical application of the mapping methodology in analyzing risk contribution, in searching
the minimum variance portfolio, and in value-at-risk estimation. Section 6 discusses and concludes our study.
M.K.P. So, T.W.C. Chan and A.M.Y. Chu / Journal of Econometrics 227 (2022) 151–167 153

2. Methodology

This section introduces a risk factor model for multivariate financial return series and its parameter estimation
procedures. The large covariance matrix of the financial asset returns is assumed to have a linear relationship with a small
covariance matrix of risk factor returns to reduce the number of parameters in the model. The reduction in the number
of parameters via the decomposition of the asset returns’ covariance matrix using risk factors alleviates the issue on high
model complexity and hence leading to an efficient estimation. Moreover, a copula-based meta-t distribution captures the
tail property of risk factor returns, providing extra flexibility in integrating the returns’ distribution in dynamic covariance
modeling.

2.1. Model settings

Suppose we are interested in modeling the dynamic covariance of K asset returns and there are N risk factors explaining
their covariance movement. Let Mt = [M1t , M2t , . . . , MNt ]⊤ be a collection of risk factor return random variables. Under
our risk mapping framework, we assume

Rt = at + Bt Mt + εt , (1)

where Rt = [R1t , R2t , . . . , RKt ] is the return of assets at time t, at = [a1t , a2t , . . . , aKt ] is the time-varying intercept,
⊤ ⊤

εt = [ε1t , ε2t , . . . , εKt ]⊤ is the random mapping error and Bt is a K × N matrix such that its (i, j)th entry is the time-varying
factor loading of the jth risk factor on the ith asset at time t, together with the following standard assumptions of risk
mapping:

1. The random mapping errors, εit ’s, have zero mean (i.e., E(εit ) = 0 for all i);
2. The random mapping errors are uncorrelated with the risk factors (i.e., Cov (εit , Mjt ) = 0 for all i, j).

This idea of risk mapping dates back to Sharpe (1964), who proposed a model that returns are linearly related to a risk
factor. Ross (1976) further extended this idea of ‘‘mapping’’ to involve several common ‘‘risk factors’’. Risk factor mapping
leads to a simplified covariance structure among the asset returns and avoids the need to estimate a huge number of
parameters. This also avoids the rank problem caused by closely correlated instruments, with the algorithm no longer
working or producing pathological estimates, as our risk factors can be chosen to be not too closely related to each other
by selecting an appropriate set of risk factors and mapping our instruments onto them.
The choice of risk factors depends on the nature of the portfolio. For example, asset returns of listed companies in Hong
Kong are related to market indices, such as the Hang Seng Index and the Hang Seng China Enterprise Index. For stocks
traded in U.S. stock markets, the risk factors can be the S&P500 Index and the Dow Jones Industrial Average. Benchmark
zero-coupon bonds can be an appropriate choice for a portfolio of fixed income securities.

2.2. Distribution assumptions

In traditional risk factor models similar to Eq. (1), the random errors of any two distinct assets are assumed to be
uncorrelated (i.e., Cov (εit , εjt ) for any i ̸ = j), implying that all the co-movement of asset returns is explained by the N
risk factors. In this paper, we relax the above restriction to allow εt , usually known as the non-systematic errors, to
follow a time-homogeneous ‘‘standardized’’ multivariate t-distribution tv1 (0, Σ ε ), where a p-dimensional ‘‘standardized’’
Γ ((v+p)/2)
)−(v+p)/2
multivariate t-distribution tv (µ, Σ ) has a density of 1 + (v − 2)−1 (x − µ)⊤ Σ −1 (x − µ)
(
, a
Γ (v/2)((v−2)π )p/2 |Σ |1/2
mean of µ and a covariance matrix of Σ , the ith diagonal element of the covariance matrix of εt is the variance σi2
and (i, j)th off-diagonal element is the covariance ρσi σj , for i, j = 1, . . . , K and i ̸ = j. Generally speaking, we expect weak
correlations among the non-systematic errors, such that ρ is a small number or close to zero. Furthermore, in order to
keep Σ ε positive definite, ρ has to be bounded below by −(K − 1)−1 , which is approximately 0 when K is large. Under
the risk factor model in (1), conditional on Mt and the information up to time t − 1, denoted by 𭟋t −1 , Rt has a multivariate
t-distribution such that Rt |Mt , 𭟋t −1 ∼ tv1 (at + Bt Mt , Σ ε ).
In addition, the risk factor vector Mt is assumed to follow the meta-t distribution with individual degrees of freedom
given the previous information, that is, t-distributions are assigned to each marginal with ( distinct degrees of freedom:
)
Mjt |𭟋t −1 ∼ tv2j (0, σjt2 ) for j = 1, . . . , N, but the joint of transformed risk factors M′t = Fv−2,10 Fv2,i (M1t ) · · · Fv−2,10 Fv2,i (MNt )
given 𭟋t −1 has a multivariate t density with degrees of freedom v2,0 :

M′t |𭟋t −1 ∼ tv2,0 (0, Σ t ), (2)

where Fv (M) is the distribution function of M with degree of freedom v , and Fv−1 is the quantile (inverse) function
of Fv .
154 M.K.P. So, T.W.C. Chan and A.M.Y. Chu / Journal of Econometrics 227 (2022) 151–167

In the specification in (2), we assume Σ t to be the covariance matrix of the transformed risk factors M′t and is
decomposed into Dt Γ t Dt , where Dt = diag(σ1t , σ2t , . . . , σNt ) is a diagonal matrix of standard deviation. We model the
conditional variance, square of the standard deviation, as σjt2 = ηj + αj Mj2,t −1 + βj σj2,t −1 , and the conditional correlation
matrix as

Γ t = corr(M′t |𭟋t −1 ) = (1 − θ1 − θ2 )Γ + θ1 Γ t −1 + θ2 Ψ t −1 , (3)

where Γ is a constant positive definite correlation matrix with all off diagonal entries equal to γ and Ψ t = {ψij,t }
∑N
h=1 ei,t −h+1 ej,t −h+1
is a measure of previous correlation with ψij,t = √(∑ ),
Mit
)(∑ and eit = σit
. The original form of this
N 2 N 2
h=1 ei,t −h+1 h=1 ej,t −h+1

model was proposed by Tse and Tsui (2002) but they assumed that the conditional distribution of the return vector
follows a multivariate normal distribution, which may fail to capture the heavy tail property of financial returns. The
above framework was also used by So and Tse (2009) in their investigation of tail dependence structures among different
market indices in Asia.
The conditional variance of each individual risk factor follows a GARCH(1, 1) process. To ensure the positivity of
conditional volatility, McAleer (2014) proposed restricting all ηj , αj and βj to be positive as a sufficient condition. We
further impose αj + βj < 1 for covariance stationarity. As in So and Yip (2012), for positive definiteness of the conditional
correlation matrix Γ t , we require θ1 and θ2 to be positive, θ1 + θ2 < 1 and Γ to be positive definite by having
(N − 1)−1 < γ < 1.
A novelty of the risk factor model in (1) is to allow the mapping matrix Bt to be time-varying. Compared with a
static mapping matrix, this time-varying feature likely enhances the mapping effectiveness by promptly adjusting the
relationship between Rt and Mt . Let Yt be the covariance between the risk factors Mt and individual asset returns Rt , given
previous information 𭟋t −1 . We impose an EWMA dynamic on the covariance matrix, i.e. Yt = λYt −1 + (1 − λ)Mt −1 R⊤ t −1 ,
with 0 < λ < 1. We estimate the mapping matrix Bt via an approximation of the conditional variance of Mt :

Yt = Cov (Mt , Rt |𭟋t −1 ) = v ar(Mt |𭟋t −1 )B⊤


t (4)
= v ar Fv2 Fv2,0 (M1t ) · · · Fv2 Fv2,0 (MKt ) |𭟋t −1 Bt ≈ Σ t Bt ,
−1 ′ −1 ′ ⊤ ∗ ⊤
(( ) )

where the (i, j)th entry of Σ ∗t is a quadratic approximation of the covariance of the transformed risk factors at the point
0, such that

∑2
∑ l
(Fv−2,1i Fv2,0 )(k) (0) (Fv−2,1j Fv2,0 )(l−k) (0) (
[Σ ∗t ]ij = E (Mit′ )k (Mjt′ )l−k ,
)
k! (n − k)!
l=0 k=0

Γ ((v2,i + 1)/2)Γ ((v2,j + 1)/2) Γ (v2,0 /2) v2,0
( )2
= √ [Σ t ]ij ,
Γ (v2,i /2)Γ (v2,j /2) v2,i v2,j Γ ((v2,0 + 1)/2)
where f (l) is the lth derivative of f for positive integer l given that they exist, and f (0) = f . We chose the point of expansion
of 0, which simplifies the expression by removing most of the terms because t density has a stationary point at 0 regardless
∗−1
of the degrees of freedom and the variance, and the means of the risk factors are fixed at 0. Hence, Bt ≈ Y⊤ t Σt ,
which took a similar form as the estimate of slope parameter in simple linear regression. This completes our modeling
framework.
Another novelty of the proposed model is that the total number of parameters in the model is K + 4N + 7, which is
a linear function of N (the number of risk factors) and K (the number of asset returns), and is likely dominated by K as
in many applications, the number of asset returns is substantially larger than the number of risk factors. Therefore, the
model complexity is of order O(K ). Most existing multivariate GARCH models, for example, VEC(1, 1) by Bollerslev et al.
(1988) and BEKK(1, 1, K) by Engle and Kroner (1995b), have at least quadratic order of complexity, or the computational
time for parameter estimation is of order O(K 2 ) or higher. Our model is efficient in the sense that the computational cost
for parameter estimation in linear over K , making our proposed methodology highly scalable.

2.3. Parameter estimation

The unknown parameters are (ηj , αj , βj , θ1 , θ2 , v1 , v2,j , σk , ρ, λ, γ ) for j = 1, 2, . . . , N, k = 1, 2, . . . , K . They can be


estimated by a Bayesian analysis producing posterior samples using Markov chain Monte Carlo (MCMC) methods. A similar
procedure has been adopted by Chen and So (2006) and So and Yeung (2014).
In each iteration of MCMC, all of the parameters are sampled sequentially. A random draw is first generated and the
acceptance probability is then calculated, which determines whether the next iterate is the new draw or the latest iterate.
Let ζ be one of our parameters to be sampled in the rth iteration. Generating a new draw ζ (r) according to the proposal
distribution, conditional on ζ (r −1) , involves the following:
M.K.P. So, T.W.C. Chan and A.M.Y. Chu / Journal of Econometrics 227 (2022) 151–167 155

Parameter Proposal distribution of ζ (r)


(r) (r −1)
1/v2,j , for j = 0, . . . , N TN(0,0.5) (1/v2,j , σ12/v2,j )
ηj(r) , for j = 1, . . . , N TN(0,∞) (ηj
(r −1)
,σ2)
(( ) ( ηj ))
(r) (r)
(r −1)
α σ 2
αj ραj βj σαj σβj
(α , β j ), for j = 1, . . . , N N j
,
j (r −1)
jβ ρ σ σ
αj βj αj βj σβ2j
(r −1)
γ (r) TN(−1,1) ( j γ
,σ 2
γ)
(r) (r −1)
1/v1 /v
TN(0,0.5) (1 1 ,σ 2
1/v1 )
(r) (r −1)
σk , for k = 1, . . . , K σ
TN(0,∞) ( k ,σ 2
σk )
ρ (r) ρ ,σ
TN(−(K −1)−1 ,1) ( (r −1) ρ2 )
λ(r) λ , σ
TN(0,1) ( (r −1) λ2 )
θ (r −1)
σ 2
ρθ 1 θ 2 σθ 1 σθ 2
(( ) ( ))
(r) (r) θ1
(θ1 , θ2 ) N 1 ,
θ (r −1)
2
ρ σ σ θ1 θ2 θ1 θ2 σθ22

where TNB (µ, σ 2 ) is the truncated normal distribution with a mean of µ and a variance of σ 2 restricted on the set B.
N(µ, Σ ) is bivariate normal with a mean of µ and a covariance matrix of Σ .
Those σ with the parameter being estimated as subscript control the step size in each sampling. Correlations terms
ρ were introduced to any pairs of parameters that are sampled simultaneously (i.e., those follow the bivariate normal in
our case). To obtain a good sample (i.e., a converging and well-mixed chain), a suitable step size is essential. Therefore,
MCMC adaptation (Chen and So, 2006) was applied after completing 40% of the iterations. We adjusted the step size by
the standard error of the observations after burn-in, which is the first 10% in our whole sample.
post
Then, by calculating the target density, that is, our joint posterior likelihoods LM ,R for both ζ (r) and ζ (r −1) , we accept ζ (r)
post
fζ (ζ (r −1) )LM ,R (ζ (r) )
as the new iterate with a probability of min {1, α} , where α = post and fζ is the proposal density. Otherwise,
fζ (ζ (r) )LM ,R (ζ (r −1) )
set ζ (r) to be ζ (r −1) .
post
Since those parts unrelated to ζ are canceled out in α , it is sufficient to write down the joint posterior likelihoods LM ,R ,
post
which are proportional to ζ . Therefore, we have LM ,R (ζ ) ∝ LM ,R (ζ ) × π (ζ ) according to Bayes theorem, where LM ,R is the
joint likelihood of our assets and risk factors and π is the prior distribution.
For the prior π , we adopted the following setting to provide a non-informative prior to the parameters.

Parameter Prior distribution


1/v1 and 1/v2,j , for j = 0, . . . , N Unif((0, 0.5))
ηj , for j = 1, . . . , N Inv-Gamma(α(ηj ) , β(ηj ) )
(αj , βj ), for j = 1, . . . , N Unif({(αj , βj ) : αj , βj > 0 and αj + βj < 1})
γ and λ Unif((0, 1))
σk , for k = 1, . . . , K Inv-Gamma(α(σk ) , β(σk ) )
ρ Unif((−(K − 1)−1 , 1))
(θ1 , θ2 ) Unif({(θ1 , θ2 ) : θ1 , θ2 > 0 and θ1 + θ2 < 1})

where Unif(A) indicates a (univariate or multivariate) uniform distribution on the region bounded by set A, α(·) and β(·)
are preassigned hyperparameters and Inv-Gamma(α, β ) refers to an inverse gamma with shape α and scale β , that is, if
X ∼Inv-Gamma(α, β ), the density of X is β α x−α−1 exp(−β/x)/Γ (α ) for x > 0 and 0 otherwise.
The log-likelihood of Mt and Rt can be decomposed as the sum of conditional likelihood of asset return on the risk
factors and likelihood of risk factors itself. Using the above likelihood function and the suggested prior distributions, the
MCMC sampling is carried out by repeating the simulation procedure until we obtain an adequate number of iterates for
posterior inference.

3. Risk estimation with risk mapping

We discuss how the proposed risk factor model in (1) helps assess the financial risk of portfolios. Suppose we construct
a portfolio with the K individual assets with a weight vector of ω = (ω1 , ω2 , . . . , ωK )⊤ . The portfolio return at time t is
given by Rωt = ω Rt = ω (at + Bt Mt + εt ) = ω at + ω Bt Mt + ω εt . Thus
⊤ ⊤ ⊤ ⊤ ⊤

v ar(Rωt ) = ω⊤ Bt v ar(Mt )B⊤


t ω + ω v ar(εt )ω

≈ ω⊤ Bt Σ ∗t B⊤
t ω + ω Σ εω

= σtω,M + σtω,ε , (5)


ω,M ω,ε ω,M
where σt = ω⊤ Bt Σ ∗t B⊤
t ω and σt = ω⊤ Σ ε ω. In the finance literature, the first term σt is usually called a systematic
ω,ε
risk component and the second term σt is a non-systematic risk component. Under the classical assumption of ρ = 0,
156 M.K.P. So, T.W.C. Chan and A.M.Y. Chu / Journal of Econometrics 227 (2022) 151–167

ω,ε
the non-systematic risk σt can be diversified away when K → ∞. When ρ ̸ = 0 but small, the non-systematic risk may
make a small contribution to the overall risk. The following section presents three financial applications of the risk factor
model.

3.1. Risk contribution

The main consideration in using the risk factor model in (1) is the effectiveness of capturing portfolio risk using the
risk factors. We can see how well the risk factors estimate the portfolio variance by calculating the proportion of portfolio
variance explained by the systematic risk component,

σtω,M
. (6)
σtω,M + σtω,ε
If this value is close to 1 (i.e., the unexplained part is significantly smaller than the explained), this indicates that the
factors work satisfactorily.
By using Euler’s theorem,
√ we can decompose the portfolio risk into a sum of risk contributions by each asset:
∂ v ar(Rω )
v ar(Rωt ) = Ki=1 ωi ∂ω t , from which we define the risk contribution by asset i as
√ ∑
i

∂ v ar(Rω )
ωi ∂ω t ω⊤ v ar(Rt )1i 1⊤
i ω
fit = √ i
= , (7)
ω
v ar(Rt ) ω ⊤ v ar(R )ω
t

where 1i is a column vector with the ith entry equal to 1 and 0 otherwise.
According to (1), the systematic risk component is given by Bt Mt . Using a similar approach as in (7), we can decompose
the systematic risk into a sum of risk factors by considering a hypothetical portfolio ω⊤ Bt Mt . Let ω(M) = B⊤ t ω such that
ω,M
ω(M)

Mt = ω⊤ (Bt Mt ). Then, we can decompose the systematic risk component variance σt using Euler’s theorem as
∂σ ω,M
σtω,M = = σtω,M
∑N ∑N
j=1 ω(M)j ∂ωt j=1 gjt , where
(M)j

∂σ ω,M
ω(M)j ∂ωt ω(M)

v ar(Mt )1j 1⊤
j ω(M)
.
(M)j
gjt = ω,M
= (8)
σt ω(M)

v ar(Mt )ω(M)
The above information is particularly helpful for portfolio managers because they can easily identify which risk factors
explain substantial exposure to market risk and perform hedging accordingly. For example, if the portfolio concerned
is a basket of stocks and the risk factors are equity indices, the exposure can be easily hedged by trading index futures
linked to each risk factor, with reference to the weighting (g1t , g2t , . . . , gNt ), instead of trading the individual K stocks. This
greatly simplifies the risk monitoring process and could potentially save a considerable amount of money in transaction
costs.

3.2. Minimum variance portfolio (MVP)

It is always possible to select some assets of different weights to form a portfolio. While a high expected return portfolio
is appreciated, it often comes with a high level of risk. Therefore, in order to control the risk, the standard deviation of
a portfolio, which is a common measure of risk, has to be minimized. The minimum variance portfolio is achieved by
selecting a portfolio with a weight of ω∗ as the solution of

min ω⊤ v ar(Rt )ω subject to 1⊤ ω = 1, (9)


ω

where 1 is a K -dimensional vector of 1 and ω⊤ v ar(Rt )ω is obtained in (5).


In some cases, short-selling is not allowed (i.e., the weight has to be non-negative) and so an extra constraint is added
to (9). The optimization problem becomes

min ω⊤ v ar(Rt )ω subject to 1⊤ ω = 1, and ω ≥ 0,


ω

where x ≥ a means that every entry of vector x has to be at least a.


Under these two√scenarios, the minimum risk portfolio (MVP) is formed with an expected return ω∗⊤ E(Rt ) and a
standard deviation ω∗⊤ v ar(Rt )ω∗ . If the expected return is required to stay at a certain level µ0 , one more constraint
has to be introduced and the problem becomes

min ω⊤ v ar(Rt )ω subject to 1⊤ ω = 1, ω ≥ 0, and ω⊤ E(Rt ) = µ0 .


ω

We will demonstrate in Section 5 how the risk factor model helps to dynamically determine the MVP.
M.K.P. So, T.W.C. Chan and A.M.Y. Chu / Journal of Econometrics 227 (2022) 151–167 157

3.3. Tail risk calculation

Suppose that we take at = 0. With the estimated portfolio variance, we can easily compute various risk measures, such
as value-at-risk or expected shortfall, under simple distribution assumptions. For example, the q-percentile value-at-risk
(VaR) under normal distribution is given by

v ar(Rωt ) × zq ,

VaRq = (10)
where zq is the q-percentile of a normal distribution. Under normal distribution assumptions for both Mt and εt , v ar(Rω t )
is calculated in (5) with Σ ∗t = Σ t . Notice that the variance of the portfolio can be estimated without the direct estimation
of the variance–covariance matrix of the underlying assets which has a dimension of N × N. Through Eq. (5), we only
need to evaluate the variance–covariance matrix of the risk factors Σ t and the beta coefficients matrix Bt and thus the
dimension of the system is reduced to N × K . This will greatly reduce the computational effort when N ≪ K .
Although Eq. (10) provides a quick normal approximation, this VaR estimation may underestimate risk when q is small.
Recall that Rωt = ω Bt Mt + ω εt , which is a sum of a t-distribution and a linear combination of another t-distribution
⊤ ⊤

with distinct degrees of freedom. As there is no analytical distribution for Rω t , the VaR percentile computation will hence
rely on simulation. A similar approach has been adopted by Wong and So (2003) and So and Wong (2012).
ω,ε
Conditional on Mt , we have Rω t |Mt , 𭟋t −1 ∼ tv1 (ω Bt Mt , σt

). Hence, the probability density function of Rω
t can be
expressed in an integral form
∫ +∞
fRω (x)
t
= fRωt |Mt ,𭟋t −1 (x|y)fMt |𭟋t −1 (y)dy. (11)
−∞
(l) ω(l)
Using (11), we simulate Mt from Mt |𭟋t −1 , the t-copula specified in Section 2.2, for l = 1, . . . , L. Then, we simulate Rt
ω,ε
from Rωt |Mt , 𭟋t −1 , a t-distribution with a mean of ω Bt Mt , a variance of σt

and degrees of freedom of ν1 . A one-step-
ω(l) (l) ω(l)
ahead forecast of the portfolio VaR at time t is given by the empirical q-percentile of Rt , l = 1, . . . , L. Using Mt and Rt ,
we can also estimate the extremal dependence between risk factor returns and portfolio returns. For example, to estimate
ω(l)
the extremal correlation corr(Rω t , Mjt |Mjt ≤ δ ), where δ is a downside risk level, we can use the sample correlation of Rt
(l) (l)
and Mjt based on the Monte Carlo sub-sample with Mjt ≤ δ .

4. Simulation study

We conduct a simulation study to examine the performance of the Bayesian estimation described in Section 2.3.
There are two main purposes in the simulation study. First, we would like to demonstrate the reliability of our Bayesian
estimation in cases with a large number of assets. Second, we provide numerical evidence that the computation time
required for estimation grows linearly with K , the number of assets. We generated 100 replications of four risk factors
(N = 4) with 100 individual assets (K = 100). Model parameters are specified in order to imitate empirical characteristics
of financial returns. For example, the GARCH parameters αj ’s are small and βj ’s are large with αj + βj quite close to one to
indicate high volatility persistence. In each replication, there are T = 2000 serial returns for the risk factors and assets,
corresponding roughly to a time horizon of eight years of daily observations in financial markets. We performed 100,000
MCMC iterations to estimate the parameters.
The initial values of the MCMC were calculated in several steps. First, we found the maximum quasi-likelihood
estimates of ηj , αj , and βj for each j = 1, . . . , N. Then, we fixed the above GARCH parameters at their estimates and found
the maximum quasi-likelihood estimates, based on a multivariate normal distribution, of θm and γ for m = 1, 2. After that,
assuming that the kurtosis exists, we estimated the degrees of freedom of v2,j , for j = 0, . . . , N by the method of moments.
In particular, v2,0 is based on a multivariate t distribution and the remaining degrees of freedom, v2,1 , . . . , v2,N , are based
on their univariate t distributions. Next, the regression related parameters λ and v1 were estimated using maximum
likelihood method. The ρ and σk , for k = 1, . . . , K were initialized as the corresponding sample correlations and standard
deviations. For the hyperparameters of the prior distributions, which are α(ηj ) , β(ηj ) , α(σk ) , and β(σk ) for j = 1, . . . , N and
k = 1, . . . , K , they were set to have the means of the corresponding inverse gamma prior distributions the same as the
initial value of ηj and σk , and the corresponding prior standard deviations equal to 100 times the prior means.
The true values of the model parameters and parameter estimation results in the simulation are presented in Table 1.
We report in the table the means and standard errors of the estimates, which are the averages of the posterior mean
estimates and the standard deviations of the posterior mean estimates in the 100 replications. All means of the posterior
mean estimates are close to their respective true values, indicating that the Bayesian estimation produces unbiased
estimates in the sample size of T = 2000. All standard errors are reasonably small that the Bayesian estimation is reliable.
The highest standard errors are associated with the degrees of freedom parameters, v2,0 , . . . , v2,4 , implying that there is
relatively higher uncertainty in the estimation of degrees of freedom.
One novelty of the proposed method for modeling high-dimensional dynamic variance–covariance matrix is by risk
factor mapping, where the computation time required in the Bayesian estimation is in the linear order of K , rather than K 2
or higher orders. With the fixing of N, the number of factors in the model, we expect that the time required to complete
a replication is proportional to K . To demonstrate the linear order property, we repeated the above simulation study in a
158 M.K.P. So, T.W.C. Chan and A.M.Y. Chu / Journal of Econometrics 227 (2022) 151–167

Table 1
The first column lists the true parameter values. The second and third columns are the means and standard deviations
(i.e., the standard error or SE) of the posterior mean estimates in 100 replications.
GARCH True Mean SE MGARCH True Mean SE
η1 0.000010 0.000010 0.000000 θ1 0.150 0.144 0.038
η2 0.000010 0.000010 0.000000 θ2 0.100 0.100 0.006
η3 0.000010 0.000010 0.000001 v2,0 8.000 7.952 0.804
η4 0.000010 0.000010 0.000001 v2,1 6.000 6.075 0.711
α1 0.150 0.151 0.005 v2,2 7.333 7.405 0.840
α2 0.117 0.116 0.004 v2,3 8.667 8.651 1.329
α3 0.083 0.083 0.004 v2,4 10.000 10.142 1.994
α4 0.050 0.050 0.003 γ 0.000 −0.000 0.002
β1 0.750 0.749 0.008 Others True Mean SE
β2 0.800 0.800 0.007 λ 0.990 0.990 0.000128
β3 0.850 0.850 0.007 v1 10.000 10.130 0.296
β4 0.900 0.899 0.006 ρ 0.100 0.099 0.003188
σ 0.0075 0.0075 0.000126

Fig. 1. Time taken to complete the Bayesian estimation versus the number of assets. The diagonal line is the simple linear regression line suggesting
the growth rate is linear.

single replication but with different number of assets, K = 10, 30, 50, 70, 90, 100. The results in Fig. 1 support our claim
that the computation time required to complete the estimation grows linearly with the number of assets. The simulation
results confirm that our proposed methodology is highly scalable with the computation time controlled in a linear order
of K .

5. Empirical study

In this section, we implement the risk factor model introduced in Section 2 and study the dynamic characteristics of
portfolios of 100 Hong Kong stocks.

5.1. Study settings

One hundred stocks listed in the Hong Kong Stock Exchange with the highest market capitalization as of April 2018
are used as the individual assets, where the data is collected from Bloomberg.1 Four market indices are selected as risk
factors. They are the Hang Seng Index (HSI), the Hang Seng China Enterprises Index (HSCEI), the Shanghai Stock Exchange
Composite Index (SSECI) and the Shenzhen Stock Exchange Composite Index (SZECI). The study covers a period of eight
years, from April 2010 to March 2018, with 1969 return values for each stock and risk factor. All the closing prices Pit have
been adjusted in case of any corporate action, such as dividend payout, right issues, or split. For each individual stock, we
consider their logarithm return defined by Rit = ln Pit − ln Pi,t −1 . The subject statistics are presented in Appendix.
A rolling window of 1000 days is applied to the data for dynamic parameter estimation. At a particular time point,
we use the most recent 1000 observations for parameter estimation and update the estimation window every day to re-
estimate the parameters. For each moving window, we perform 200,000 MCMC iterations and the last 120,000 iterations

1 To remove the noise from the financial returns, stocks that have been suspended from trading for more than five business days are not
considered in our study. Moreover, we only consider the trading days on which the Hong Kong market was open.
M.K.P. So, T.W.C. Chan and A.M.Y. Chu / Journal of Econometrics 227 (2022) 151–167 159

Fig. 2. Time series of the posterior means of the MGARCH parameters. The solid green line represents the means. (For interpretation of the references
to colour in this figure legend, the reader is referred to the web version of this article.)

are used for the Bayesian estimation of parameters. The one-step forecast of the portfolio variance and VaR is obtained
based on the estimated model using the methods described in Section 3. Altogether, we obtain 100 one-step-ahead
forecasts of the covariance matrix of Rt and risk measures.

5.2. Numerical results

Following the same estimation procedure used in Section 4, except the adaption is done 10 times in total when there
are 3% of new observations produced after burn-in to achieve a better convergence, the GARCH and MGARCH parameters
associated with the risk factors are estimated and reported in Table 2.
The average estimates appear to be typical where α in the GARCH models tends to be small, while β is large, as shown
in Fig. 2. Meanwhile, α + β is close to 1, which indicates the strong volatility of the persistence of the market returns.
Except for a few trading days, HSI and HSCEI from Hong Kong keep a constant η, increasing α and decreasing β throughout
the whole testing period. Nevertheless, the sum of α and β remains almost constant over time. On the other hand, SSECI
and SZECI from mainland China show a decreasing trend in η and α but an increasing β during the first half of the moving
windows. The sum of α and β fluctuates. However, in the second half of the moving window, all the GARCH parameters
and the sum become stable.
For the MGARCH model parameters presented in Fig. 3, θ1 dominates θ2 throughout the whole testing period. The
fluctuation of occurs in the first half, but they become stable in second. The sums almost equal 1 across the whole period,
but drop slightly in the last 25% of the testing period. The constant term γ maintains an increasing trend across the whole
testing period.
The average degree of freedom v2,0 in the t-copula shown in Table 2 is around 3.35, indicating quite strong extremal
dependence in the joint distribution of the four market index returns. There is a similar level in v2,1 and v2,2 in the degrees
of freedom of HSI and HSCEI, in that they have means of around 3.2 and 3.5, respectively. However, in Fig. 3, SSECI and
SZECI exhibit stable v2,3 and v2,4 at around 6 in the first half of the testing period, and dropping to around 4 in the second
half of the period.
The correlation of the non-systematic risk component εt , ρ , has a small value of 0.035, indicating that the four
risk factors in (1) can explain most of the co-movement of the 100 stock returns. With such a small ρ , the portfolio
diversification is expected to work well, especially when the portfolio size K is large. In addition, the large degree of
freedom, v1 = 11, increases from around 8.5 in the first half of the testing period to around 12 in the second half,
consistent with our observation that, in the second half of the testing period, the four market indices are more stable.

5.3. Application to risk management

5.3.1. Risk contribution


In our example, we adopt an equally weighted portfolio. As discussed in Section 3, each stock has a changing
contribution to the portfolio risk over time. Under our risk factor model, we can quantify the risk contributions of
individual stocks and the risk factors to the portfolio risk.
First of all, we calculate the contribution of systematic risk to the overall risk using Eq. (6). From the first plot of Fig. 4,
we find that the systematic risk contributes a high proportion, on average, 84.67%, the portfolio variance most of the time,
especially in 2016. The high proportion is consistent with a small estimated ρ and the use of a large K of 100, and is a
good sign of having sufficient factors to explain the portfolio risk by the systematic risk component.
160 M.K.P. So, T.W.C. Chan and A.M.Y. Chu / Journal of Econometrics 227 (2022) 151–167

Fig. 3. Time series of the posterior means of the MGARCH parameters. The solid red line represents the means.

Table 2
Averages of the posterior means (the first column) and the posterior standard deviations (the second column).
GARCH Posterior mean Posterior SD MGARCH Posterior mean Posterior SD
η1 0.000001 0.000000 θ1 0.976 0.001
η2 0.000002 0.000000 θ2 0.024 0.001
η3 0.000008 0.000002 v2,0 3.349 0.094
η4 0.000011 0.000003 v2,1 3.212 0.086
α1 0.012 0.001 v2,2 3.509 0.118
α2 0.012 0.001 v2,3 4.506 0.294
α3 0.041 0.006 v2,4 4.421 0.380
α4 0.041 0.006 γ 0.175 0.449
β1 0.981 0.001 Others Posterior mean Posterior SD
β2 0.979 0.002 λ 0.990 0.000358
β3 0.939 0.011 v1 10.709 0.503018
β4 0.951 0.009 ρ 0.033 0.002038
σ 0.017 0.000414
M.K.P. So, T.W.C. Chan and A.M.Y. Chu / Journal of Econometrics 227 (2022) 151–167 161

Fig. 4. First plot: Time series plot of the proportion of variance explained by the systematic risk component, calculated using Eq. (6). The solid
red line represents the means. Second plot: Time series of the relative risk contribution of four selected stocks. Third plot: Time series of the risk
contribution of the four risk factors to the systematic risk.

We move on to consider the risk contribution of some stocks over time. In particular, we have chosen four banks: HSBC
Holdings PLC (0005.HK), Hang Seng Bank (0011.HK), China Construction Bank Corp (0939.HK), and Industrial & Commercial
Bank of China Ltd (1398.HK). The second plot in Fig. 4 shows the time series plots of the relative risk contribution defined
by fjt × K . If the relative risk contribution is higher/lower than one, the risk contribution is higher/lower than average. We
observe that the risk contribution of HSBC Holdings PLC and Hang Seng Bank (local banks in Hong Kong) has lower risk
contributions than the two other banks, the major businesses of which are based in mainland China. An increasing trend
is observed in the risk contribution of HSBC Holdings PLC and Hang Seng Bank in mid-2015, which drops again in 2017.
162 M.K.P. So, T.W.C. Chan and A.M.Y. Chu / Journal of Econometrics 227 (2022) 151–167

Fig. 5. Returns and standard deviations under MVP in the testing period. The solid red line represents the means.

In regard to China Construction Bank Corp and Industrial & Commercial Bank of China Ltd, the relative risk contribution
starts an increasing trend in 2016 and reaches about 1.5 in early 2018.
In the third plot of Fig. 4, we plot git in (8), the contribution of the ith market index to the systematic risk component
at time t. Comparing the four indices, HSI and HSCEI contribute higher risk than SSECI and SZECI most of the time. This is
not surprising as HSI and HSCEI are composed of Hong Kong stocks. HSCEI contributes as much as 40% of the systematic
risk. The contributions of SSECI and SZECI rise to a level higher than that of HSI in 2015, when the markets in mainland
China are very volatile.

5.3.2. Minimum variance portfolio (MVP)


Besides the equally-weighted portfolio, the minimum variance strategy is also common in asset allocation. Assuming
that short-selling is allowed, a set of changing weights is produced across time. Fig. 5 gives the portfolio returns and
standard deviations of MVP. The average portfolio return and average portfolio standard deviation over time are 0.022%
(roughly 5.5% annual return) and 0.411%, respectively. We see that the portfolio risk is higher before 2017 and falls below
0.3% in late 2017. The MVP allocation based on the risk factor model successfully yields portfolios of daily variation to
within 1% in most trading days.

5.3.3. Value-at-risk (VaR)


The time series of daily VaR of an equally-weighted portfolio of the 100 stocks is estimated under Gaussian and meta
t-distribution assumptions, using the methods in Section 3.3. For meta t-distribution assumption, we compute the VaR
based on the Monte Carlo samples of the portfolio return Rω t . We compare VaR estimates against the actual portfolio
returns and count the number of days in which the portfolio return is worse than the VaR estimate. Fig. 6 plots the
returns and VaR estimates. We observe that the return series and the VaR bands are very volatile in 2015 during the
stock market turbulence in China. The VaR bands narrow toward the year 2018 as the markets gradually recover.
When considering the empirical coverage of VaR as shown in Table 3, a normal assumption performs well in regard
to the 5% and 2.5% VaR, as the empirical coverage is closer to the nominal values of 5% and 2.5%, compared to meta-t.
However, the VaR estimation under meta t-distribution performs better for extreme percentiles of 1% and 0.5%. This
makes sense because meta t-distribution is a leptokurtic distribution that places higher weight in the tail behavior of
a distribution. To understand the effectiveness of the risk factor model in estimating VaR with the meta-t and normal
distributions, we randomly generate 100 portfolios and analyze their VaR coverages. Fig. 7 presents boxplots of the ratios
of the empirical coverage of VaR over the nominal value under meta t and normal distributions. The closer the ratio is
to 1, the better the estimation of VaR is. The boxplots illustrate that the 2.5% and 5% VaR are not a good estimator with
the meta-t distributions. In addition, the 0.5% and 1% VaR are underestimated using normal distributions. However, the
normal performance regarding 2.5% and 5% VaR and the meta-t in regard to the 0.5% and 1% VaR is very satisfactory. From
the above analysis, we can conclude that the risk factor model is able to reliably estimate the VaR of a large portfolio in
various scenarios, with a suitable choice between normal and meta-t distributions.
We also estimate the extremal correlation between the equally-weighted portfolio returns and Hang Seng Index returns
(for δ = −3%) in 2015–2018 using the method in Section 3.3. Fig. 8 displays the time series plot of the extremal
correlations. These correlations stay above 0.6 in most of the time, and are relatively high in late 2016 and mid 2017.
The average extremal correlation is 0.87, which indicates that the chance for the portfolio to follow is high when the
market downturn is significant.
M.K.P. So, T.W.C. Chan and A.M.Y. Chu / Journal of Econometrics 227 (2022) 151–167 163

Fig. 6. Daily VaR estimates obtained under meta-t and normal distributions vs the actual portfolio returns.

Table 3
Empirical coverage of VaR under normal and meta-t distributions.
VaR Coverage — Normal Coverage — meta-t
5% 5.160% 7.946%
2.5% 2.580% 2.993%
1% 1.548% 1.135%
0.5% 0.929% 0.516%
164 M.K.P. So, T.W.C. Chan and A.M.Y. Chu / Journal of Econometrics 227 (2022) 151–167

Fig. 7. Boxplot of the ratio between empirical coverage and the nominal value under meta-t and normal distributions, based on 100 randomly
generated portfolios.

Fig. 8. Time series plot of the correlation of the HSI returns and the equally-weighted portfolio returns given that the HSI returns are less than
−3%.

6. Conclusion

This paper proposes an alternative approach to estimating the variance–covariance structure of financial returns with
high computational efficiency. The computational requirement in applying the proposed approach is proportional to the
number of financial returns. By selecting a small number of observable market risk factors, multivariate time series models
can be fitted to the joint risk factors, allowing us to recover the high-dimensional dependence structure of potentially
large number of financial returns as well as estimating various portfolio risk measures dynamically. We considered
the MGARCH model proposed by Tse and Tsui (2002) and generalized it to a risk mapping framework. A copula-based
meta-t distribution captures the individual and overall tail risk of the risk factors. We performed a simulation study to
demonstrate the effectiveness of Bayesian estimation and the linear order of computation time with respect to the number
of financial returns. We then applied the proposed model in an empirical study of 100 Hong Kong stocks with four risk
factors.
Results from the empirical study show that the portfolio VaR has a tendency to underestimate the risk at less extreme
tail percentiles, but performance will gradually improve when we move toward the tail. This may be related to the
probabilistic property of a t-distribution, in that it is only accurate in describing the extreme tail of financial returns.
Alternative distribution assumptions might be proposed, such that we can obtain consistent risk estimation performances
at different percentiles. For example, we currently assume that non-systemic risk within the mapping system has a
time-homogeneous distribution, but more realistic features can be imposed, such as stochastic volatility or correlation
clustering.
There are two possible extensions of our current study. On the empirical side, we need to select a good set of risk
factors or adequately explain the portfolio risk concerned. In our current study of Hong Kong stocks, only equity indices
M.K.P. So, T.W.C. Chan and A.M.Y. Chu / Journal of Econometrics 227 (2022) 151–167 165

from Hong Kong and China are considered, which may be too narrow. Other possible candidates include international
equity indices, foreign exchange levels, benchmark interest rates, and commodity prices. On the technical side, our risk
mapping framework is a robust one that can be implemented with different time series models other than the MGARCH
type model considered in this paper. After we have identified an appropriate group of risk factors, the next question
of interest is how we can optimize the time series modeling strategy, such that the risk estimation performance can be
further enhanced. Once we have clear results on the above issues, this risk mapping framework will prove to be beneficial
to any high-dimensional risk management problem faced by practitioners.

Acknowledgments

The authors would like to thank Professor Michael McAleer and two reviewers for their constructive comments. This
work was supported by the Hong Kong RGC Theme-based Research Scheme [project number T31-604/18-N].

Appendix. List of indices and Hong Kong stocks for the empirical study

The table below gives the list of Hong Kong stocks and indices from April 2010 to March 2018. They were used in both
the empirical study and the summary statistics of their daily returns.

Mean SE SKEW Kurtosis


700 HK EQUITY 0.135% 1.971% −0.140 5.379
939 HK EQUITY 0.032% 1.547% 0.047 5.046
5 HK EQUITY 0.013% 1.287% −0.434 7.366
941 HK EQUITY 0.012% 1.298% 0.050 5.723
2318 HK EQUITY 0.055% 2.057% 0.045 7.922
1398 HK EQUITY 0.026% 1.627% 0.171 6.975
883 HK EQUITY 0.009% 1.999% 0.219 6.807
2388 HK EQUITY 0.054% 1.463% −0.077 6.791
1 HK EQUITY 0.029% 1.437% 0.364 9.215
16 HK EQUITY 0.016% 1.384% −0.687 11.610
3988 HK EQUITY 0.026% 1.533% 0.153 5.971
11 HK EQUITY 0.044% 1.030% 0.223 7.942
2007 HK EQUITY 0.106% 2.673% 0.133 5.846
1928 HK EQUITY 0.077% 2.550% 0.376 8.521
3333 HK EQUITY 0.143% 3.205% 0.472 7.239
267 HK EQUITY −0.016% 1.847% 0.413 8.273
388 HK EQUITY 0.044% 1.770% 1.144 13.735
27 HK EQUITY 0.152% 2.810% 0.012 7.217
762 HK EQUITY 0.011% 1.994% 0.208 5.033
688 HK EQUITY 0.031% 2.203% 0.322 7.171
945 HK EQUITY −0.001% 1.680% −0.053 6.547
2888 HK EQUITY −0.044% 1.852% −0.375 8.394
66 HK EQUITY 0.029% 1.078% −0.389 6.620
3 HK EQUITY 0.040% 0.973% −0.129 4.883
3328 HK EQUITY 0.001% 1.712% 0.014 5.960
12 HK EQUITY 0.033% 1.562% 0.040 5.057
175 HK EQUITY 0.090% 2.948% 0.221 5.769
2 HK EQUITY 0.035% 0.807% −0.289 4.787
1109 HK EQUITY 0.035% 2.418% 0.245 5.369
386 HK EQUITY 0.034% 1.720% 0.201 5.547
2018 HK EQUITY 0.127% 2.705% 0.039 4.993
1038 HK EQUITY 0.052% 1.191% 0.009 5.627
1972 HK EQUITY 0.022% 1.508% −0.586 9.234
2628 HK EQUITY −0.022% 1.970% 0.062 5.865
2382 HK EQUITY 0.221% 3.129% 0.326 4.761
3968 HK EQUITY 0.035% 2.099% 0.681 12.557
1128 HK EQUITY 0.053% 2.682% 0.138 6.067
656 HK EQUITY 0.059% 2.321% 0.412 8.423
960 HK EQUITY 0.061% 2.300% 0.093 5.619
1093 HK EQUITY 0.083% 2.418% 0.290 6.710
6 HK EQUITY 0.036% 1.137% −0.595 16.963
166 M.K.P. So, T.W.C. Chan and A.M.Y. Chu / Journal of Econometrics 227 (2022) 151–167

Mean SE SKEW Kurtosis


1177 HK EQUITY 0.111% 2.656% −0.097 10.374
2313 HK EQUITY 0.112% 2.188% 0.033 4.950
20 HK EQUITY 0.053% 1.841% 0.080 5.765
291 HK EQUITY 0.028% 1.948% 0.415 9.089
857 HK EQUITY −0.015% 1.732% 0.196 5.004
17 HK EQUITY 0.010% 1.765% −1.066 14.204
2319 HK EQUITY 0.047% 2.147% −0.648 18.001
2020 HK EQUITY 0.076% 2.749% 0.052 6.657
1044 HK EQUITY 0.023% 1.688% 0.126 4.807
2601 HK EQUITY 0.008% 2.088% 0.204 5.939
966 HK EQUITY 0.002% 2.611% 0.290 5.939
322 HK EQUITY −0.002% 2.044% 0.208 6.515
23 HK EQUITY 0.017% 1.442% 0.077 5.965
669 HK EQUITY 0.108% 2.311% 0.226 4.632
83 HK EQUITY 0.010% 1.729% −0.005 4.841
4 HK EQUITY 0.073% 1.882% 0.556 8.315
101 HK EQUITY −0.016% 1.660% −0.022 5.275
270 HK EQUITY 0.066% 1.788% 0.212 4.258
998 HK EQUITY 0.011% 1.780% 0.091 5.657
1114 HK EQUITY 0.105% 2.981% 0.083 4.986
2688 HK EQUITY 0.066% 2.357% −0.181 5.725
151 HK EQUITY 0.011% 1.998% 0.198 4.466
1169 HK EQUITY 0.085% 2.720% −0.003 6.463
6823 HK EQUITY 0.067% 1.554% 0.156 14.064
813 HK EQUITY 0.041% 2.622% 0.336 5.358
19 HK EQUITY 0.006% 1.279% −1.059 15.421
486 HK EQUITY −0.032% 2.427% 0.348 5.469
836 HK EQUITY 0.007% 2.000% −0.029 5.412
1088 HK EQUITY −0.016% 2.066% 0.193 6.025
2328 HK EQUITY 0.042% 2.230% 0.136 5.625
3383 HK EQUITY 0.041% 2.809% −0.049 9.270
914 HK EQUITY 0.052% 2.499% 0.127 4.907
144 HK EQUITY −0.013% 2.031% 0.135 4.413
2689 HK EQUITY −0.001% 3.035% 0.308 6.815
1211 HK EQUITY 0.004% 3.139% 0.343 7.280
69 HK EQUITY 0.006% 1.606% −0.107 8.582
659 HK EQUITY 0.043% 1.737% 0.056 5.932
135 HK EQUITY −0.017% 1.957% −0.591 12.083
392 HK EQUITY −0.009% 1.874% 0.020 4.742
1988 HK EQUITY 0.028% 1.948% −0.086 6.094
551 HK EQUITY 0.027% 1.874% −0.009 9.371
683 HK EQUITY 0.004% 1.510% 0.034 6.005
817 HK EQUITY 0.040% 2.414% 0.238 5.033
1099 HK EQUITY 0.008% 2.123% 0.415 6.137
257 HK EQUITY 0.055% 2.492% 0.227 6.640
992 HK EQUITY −0.004% 2.214% −0.284 7.023
728 HK EQUITY 0.002% 1.708% 0.309 4.497
3311 HK EQUITY 0.070% 2.452% 0.279 6.780
881 HK EQUITY 0.034% 2.839% 0.087 5.685
868 HK EQUITY 0.084% 2.531% 0.421 7.007
753 HK EQUITY 0.017% 2.517% 0.049 4.649
1313 HK EQUITY 0.038% 2.495% 0.308 5.304
522 HK EQUITY 0.029% 2.229% −0.144 5.764
316 HK EQUITY 0.013% 2.278% −0.301 15.029
3377 HK EQUITY 0.008% 2.516% 0.456 6.244
14 HK EQUITY 0.044% 1.482% 0.042 5.216
425 HK EQUITY 0.061% 2.609% −0.292 7.775
1208 HK EQUITY 0.032% 3.045% 0.104 5.710
87 HK EQUITY 0.003% 1.122% −1.699 25.952
M.K.P. So, T.W.C. Chan and A.M.Y. Chu / Journal of Econometrics 227 (2022) 151–167 167

Mean SE SKEW Kurtosis


HSI Index 0.017% 1.144% −0.346 5.669
HSCEI Index −0.003% 1.475% −0.052 5.121
SSECI Index 0.000% 1.550% −0.564 11.404
SZCEI Index 0.020% 1.821% −0.671 7.229

References

Acar, E.F., Czado, C., Lysy, M., 2019. Flexible dynamic vine copula models for multivariate time series data. Econom. Stat. 12.
Alexander, C., 2001. Orthogonal GARCH. In: Mastering Risk. Financial Times-Prentice Hall, pp. 21–38.
Alexander, C., Chibumba, A., 1997. Orthogonal GARCH: An Empirical Validation in Equities, Foreign Exchange and Interest Rates. School of Mathematical
Sciences Discussion Paper, Sussex University.
Asai, M., McAleer, M., 2008. A portfolio index GARCH model. Int. J. Forecast. 24, 449–461.
Asai, M., McAleer, M., 2009. The structure of dynamic correlations in multivariate stochastic volatility models. J. Econometrics 150, 182–192.
Asai, M., McAleer, M., Yu, J., 2006. Multivariate stochastic volatility: a review. Econometric Rev. 25, 145–175.
Audrino, F., Trojani, F., 2011. A general multivariate threshold GARCH model with dynamic conditional correlations. J. Bus. Econom. Statist. 29,
138–149.
Bollerslev, T., 1986. Generalized autoregressive conditional heteroskedasticity. J. Econometrics 31, 307–327.
Bollerslev, T., 1990. Modelling the coherence in short-run nominal exchange rates: a multivariate generalized ARCH model. Rev. Econ. Stat. 72,
498–505.
Bollerslev, T., Engle, R.F., Wooldridge, J.M., 1988. A capital asset pricing model with time-varying covariances. J. Political Econ. 96, 116–131.
Chen, C.W.S., So, M.K.P., 2006. On a threshold heteroscedastic model. Int. J. Forecast. 22, 73–89.
Chib, S., Omori, Y., Asai, M., 2009. Multivariate stochastic volatility. In: Handbook of Financial Time Series. Springer.
Ding, Z., 1994. Time Series Analysis of Speculative Returns (Ph.D. thesis). University of California, San Diego, Department of Economics.
Engle, R.F., 1982. Autoregressive conditional heteroscedasticity with estimates of the variance of United Kingdom inflation. Econometrica 50, 987–1007.
Engle, R.F., 2002. Dynamic conditional correlation: A simple class of multivariate generalized autoregressive conditional heteroskedasticity models. J.
Bus. Econom. Statist. 20, 339–350.
Engle, R.F., Kelly, B., 2012. Dynamic equicorrelation. J. Bus. Econom. Statist. 30, 212–228.
Engle, R.F., Kroner, K.F., 1995a. Multivariate simultaneous generalized ARCH. Econometric Theory 11, 122–150.
Engle, R.F., Kroner, K.F., 1995b. Multivariate simultaneous generalized ARCH. Econometric Theory 11, 122–150.
Engle, R.F., Mezrich, J., 1996. Garch for groups: A round-up of recent developments in garch techniques for estimating correlation. RISK 9, 36–40.
Engle, R.F., Sheppard, K., 2001. Theoretical and Empirical Properties of Dynamic Conditional Correlation Multivariate GARCH. Technical Report. National
Bureau of Economic Research.
Fiorentini, G., Sentana, E., Calzolari, G., 2003. Maximum likelihood estimation and inference in multivariate conditionally heteroscedastic dynamic
regression models with student t innovations. J. Bus. Econom. Statist. 21, 532–546.
Francq, C., Horvath, L., Zakoian, J.M., 2014. Variance targeting estimation of multivariate GARCH models. J. Financ. Econom. 14, 353–382.
Francq, C., Zakoïan, J.M., 2016. Estimating multivariate volatility models equation by equation. J. R. Stat. Soc. Ser. B Stat. Methodol. 78, 613–635.
Glosten, L.R., Jagannathan, R., Runkle, D.E., 1993. On the relation between the expected value and the volatility of the nominal excess return on
stocks. J. Finance 48, 1779–1801.
Hamilton, J.D., Susmel, R., 1994. Autoregressive conditional heteroskedasticity and changes in regime. J. Econometrics 64, 307–333.
Jondeau, E., Rockinger, M., 2006. The copula-GARCH model of conditional dependencies: An international stock market application. J. Int. Money
Finance 25, 827–853.
Klaassen, F., 1999. Have Exchange Rates Become More Closely Tied? Evidence From a New Multivariate GARCH Model. Tilburg University, CentER
Discussion Paper Series.
Lee, T.H., Long, X., 2009. Copula-based multivariate GARCH model with uncorrelated dependent errors. J. Econometrics 150, 207–218.
Ling, S., 2007. Self-weighted and local quasi-maximum likelihood estimators for ARMA-GARCH/IGARCH models. J. Econometrics 140, 849–873.
McAleer, M., 2014. Asymmetry and leverage in conditional volatility models. Econometrics 2, 145–150.
McAleer, M., Chan, F., Hoti, S., Lieberman, O., 2008. Generalized autoregressive conditional correlation. Econometric Theory 24, 1554–1583.
McAleer, M., Hoti, S., Chan, F., 2009. Structure and asymptotic theory for multivariate asymmetric conditional volatility. Econometric Rev. 28, 422–440.
Messaoud, S.B., Aloui, C., 2015. Measuring risk of portfolio: GARCH-copula model. J. Econ. Integr. 30, 172–205.
Nelson, D.B., 1991. Conditional heteroskedasticity in asset returns: A new approach. Econometrica 59, 347–370.
Paolella, M.S., 2015. Multivariate asset return prediction with mixture models. Eur. J. Finance 21, 1214–1252.
Riccetti, L., 2013. A copula-GARCH model for macro asset allocation of a portfolio with commodities. Empir. Econom. 44, 1315–1336.
Ross, S., 1976. The arbitrage theory of capital asset pricing. J. Econom. Theory 13, 341–360.
Sharpe, W.F., 1964. Capital asset prices: A theory of market equilibrium under conditions of risk. J. Finance 19, 425–442.
So, M.K.P., Choi, C.Y., 2009. A threshold factor multivariate stochastic volatility model. J. Forecast. 28, 712–735.
So, M.K.P., Tse, A.S.L., 2009. Dynamic modeling of tail risk: Applications to China, Hong Kong and other Asian markets. Asia-Pac. Financial Mark. 16,
183–210.
So, M.K.P., Wong, C.M., 2012. Estimation of multiple period expected shortfall and median shortfall for risk management. Quant. Finance 12, 739–754.
So, M.K.P., Yeung, C.Y.T., 2014. Vine-copula GARCH model with dynamic conditional dependence. Comput. Statist. Data Anal. 76, 655–671.
So, M.K.P., Yip, I.W.H., 2012. Multivariate GARCH models with correlation clustering. J. Forecast. 31, 443–468.
Tang, J., Sriboonchitta, S., Ramos, V., Wong, W.K., 2016. Modelling dependence between tourism demand and exchange rate using the copula-based
GARCH model. Current Issues in Tour. 19, 876–894.
Tse, Y.K., Tsui, A.K.C., 2002. A multivariate generalized autoregressive conditional heteroscedasticity model with time-varying correlations. J. Bus.
Econom. Statist. 20, 351–362.
Vrontos, I.D., Dellaportas, P., Politis, D.N., 2003. A full-factor multivariate GARCH model. Econom. J. 6, 312–334.
Wong, C.M., So, M.K.P., 2003. On conditional moments of GARCH models, with applications to multiple period value at risk estimation. Statist. Sinica
13, 1015–1044.

You might also like