Measuring Systemic Risk of The Chinese Banking - 2021 - The North American Jour

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 13

North American Journal of Economics and Finance 55 (2021) 101354

Contents lists available at ScienceDirect

North American Journal of Economics and Finance


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

Measuring systemic risk of the Chinese banking industry: A


wavelet-based quantile regression approach
Qifa Xu a,b , Bei Jin a , Cuixia Jiang a ,∗
a School of Management, Hefei University of Technology, Hefei, PR China
b Key Laboratory of Process Optimization and Intelligent Decision-making of Ministry of Education, Hefei, PR China

ARTICLE INFO ABSTRACT

Keywords: In systemic risk measure, a large amount of literature has emerged, but few of them take
Systemic risk into account the multi-scale natures of financial data. Considering these natures, we develop a
Banking industry novel W-QR-CoVaR method to measure systemic risk. To be specific, the W-QR-CoVaR method
CoVaR
combines the wavelet multiresolution analysis (MRA) with the conditional value-at-risk (CoVaR)
Quantile regression
method based on the quantile regression (QR) framework. We then apply it to measure the
Wavelet analysis
systemic risk in the Chinese banking industry covering the period from September 2007 to
September 2018. Our experiment results show that the hybrid W-QR-CoVaR method performs
better than the traditional CoVaR method in terms of predictive accuracy. Furthermore, we also
explore the relation between the systemic risk contribution of each individual bank and the
bank-specific characteristics. Size and leverage appear to be the most robustness determinants.
The findings suggest that regulators should pay more attention to the banks with smaller size
and higher leverage.

1. Introduction

Systemic risk refers to the possibility of the collapse of the whole financial system due to a cascading failure which is triggered
by the distress of a single entity or cluster of entities. Since the global financial crisis of 2007–2009, the systemic risk of financial
sectors and its measuring tools have drawn more and more attention from academics, practitioners, and regulators. In recent years,
China has gradually become the second largest economy in the world and the Chinese financial system also evolves rapidly. It
is well-known that the Chinese financial system is actually a bank-based system. The Chinese banking industry has occupied a
dominant position in the domestic financial system for a long time and has expanded significantly over the past decades. At present,
the Chinese banking system includes several largest banks all over the world. The top four banks in the latest ranking are Industrial
and Commercial Bank of China, China Construction Bank, Agricultural Bank of China, and Bank of China. They are all state-owned
commercial banks and their assets account for a large proportion of the whole banking system. As a result, the banking industry is
regulated and guided by the government through the China Banking Insurance Regulatory Commission (CBIRC), which replaced the
China Banking Regulatory Commission (CBRC) in April 2018. The Chinese banking industry has gradually transitioned to a more
open system since the reform began in the early 1980s, which supports the China’s emergence into the global economic and keeps
China a tight association with the international market. Consequently, the Chinese banking system has potential impacts not only
on the domestic economic cycle, but also on the global financial stability and economic growth. To this end, the supervision of
systemic risk in the Chinese banking industry is of great significance.

∗ Correspondence to: School of Management, Hefei University of Technology, Hefei 230009, Anhui, PR China.
E-mail addresses: xuqifa@hfut.edu.cn (Q. Xu), jinbei@mail.hfut.edu.cn (B. Jin), jiangcuixia@hfut.edu.cn (C. Jiang).

https://doi.org/10.1016/j.najef.2020.101354
Received 26 July 2020; Received in revised form 2 November 2020; Accepted 23 December 2020
Available online 29 December 2020
1062-9408/© 2020 Elsevier Inc. All rights reserved.
Q. Xu et al. North American Journal of Economics and Finance 55 (2021) 101354

Accurate prediction of systemic risk can provide an effective decision-making basis for risk regulators. For the reason that when
the systemic risk increases, the regulators need to make polices to manage the financial risk, such as encouraging cash hoarding,
which will reduce the profitability of traditional banks in the way of undermining the financial transmission. Too much cash hoarding
will lead to the decline of bank profitability, while too little cash hoarding will not be able to resist financial risks. Costs vary across
different combination of cash balances and liquidity needs. According to this trade-off, the accurate prediction of systemic risk is
of great importance.
Among the literature on quantifying systemic risk, the most widely used methods are conditional value-at-risk (CoVaR) of Adrian
and Brunnermeier (2016) and Girardi and Ergün (2013), marginal expected shortfall (MES) of Acharya et al. (2017), and systemic
risk measure (SRISK) of Brownlees and Engle (2017); see Silva et al. (2017) for more details. In particular, CoVaR is preferred
for its effectiveness and simplicity. In statistics, CoVaR is a signal number evaluating the maximal potential loss of the entire
financial system conditional on institutions being under distress for a given probability. In CoVaR estimation, the commonly used
methods are quantile regression, multivariate GARCH, and copulas (see Adrian and Brunnermeier (2016), Girardi and Ergün (2013),
and Reboredo and Ugolini (2015)). In this study, we follow Adrian and Brunnermeier (2016) and adopt the two-stage quantile
regression method (abbreviated as QR-CoVaR henceforth), which is based on the framework of quantile regression (see Koenker
and Bassett (1978)) to estimate the model. This QR-CoVaR method has already been applied in systemic risk measure. For
example, López-Espinosa et al. (2012) employ this method to identify the determinants of systemic risk contribution in the sample
of 54 international banks and find that short-term wholesale funding is a key factor in triggering systemic risk episodes. Bernal
et al. (2014) assess the systemic risk contribution of different financial sectors (e.g., banking, insurance, and other financial
services industries) in the Eurozone and United States, and find that at the time of distress, other financial services and the
insurance industry contribute the most to systemic risk in the Eurozone and United States, respectively. In de Mendoncça and
da Silva (2018), they examine systemic risk in the Brazilian banking sector and conclude that the QR-CoVaR method has a good
ability to predict risk. Moreover, the QR-CoVaR method has also been extended in practice. For example, Girardi and Ergün
(2013) modify the CoVaR definition as the conditional state of institutions being at most rather than being exactly at their
risk levels, thus considering more severe distress events. Xu et al. (2019) introduce a Least Absolute Shrinkage and Selection
Operator (LASSO) method of Tibshirani (1996) into QR-CoVaR to investigate the interconnectedness and systemic risk network
of the Chinese financial institutions. Bonaccolto et al. (2019) introduce autoregressive components into QR-CoVaR and employ
a quantile-located process to capture the state in which the financial system and a conditioning company are jointly under
distress.
To date, the QR-CoVaR method has proven to be an effective tool in measuring systemic risk. Actually, it is based on the nature of
time series in time domain. However, only using time domain analysis is not adequate to understand all structures of financial time
series, especially for high–frequency components. To the best of our knowledge, we can also study the behavior of time series from
both time and frequency domains simultaneously through wavelet transform. In this way, multiresolution analysis (MRA) of Mallat
(1989) can be implemented as well so that one can analyze any time series with desired scales and time intervals. Recently, MRA
has been successfully applied to the field of risk management. For example, Huang (2011) combines MRA with the GARCH model
to investigate spillover effects across financial markets of US and Taiwan, and shows that the direction and magnitude of volatility
spillovers significantly vary with their scales. He et al. (2012) comprehensively use MRA, artificial neural network (ANN), and VaR
to evaluate the risk of metals markets. Jammazi and Reboredo (2016) combine MRA and copula to study the dependence structures
of Brent oil and MSCI world stock market at different scales, and analyze their implications for optimal portfolio selection. Teply
and Kvapilikova (2017) employ MRA to measure the systemic risk of American banking industry in time–frequency domain, and
prove that wavelet decomposition can improve the predictive ability of CoVaR. Meng and Huang (2019) employ MRA to examine
time–frequency co-movement characteristics of effective exchange rates for four Asian economics and estimate the 𝛥CoVaR of Asian
foreign exchange markets at different scales to assess their stress performance.
In this study, we temp to improve the accuracy of systemic risk measure by exploiting time–frequency information in financial
time series. To this end, we propose a novel W-QR-CoVaR method through combining the MRA with the QR-CoVaR method. The
hybrid W-QR-CoVaR method contains three main steps: (1) the wavelet decomposition of financial time series, (2) the QR-CoVaR
estimation on multi-scales, and (3) the wavelet reconstruction of QR-CoVaRs. Then we apply this method to evaluate the systemic
risk of the Chinese banking industry covering the period from September 2007 to September 2018. In this empirical application,
we compare the performance of W-QR-CoVaR and QR-CoVaR, and investigate the relation between systemic risk and risk factors.
The empirical findings and conclusions are as follows. First, the backtesting results show that our W-QR-CoVaR method has a
better performance in measuring systemic risk. Second, according to the decomposition results of system returns, we find that risk
tends to concentrate on the high–frequency component. Meanwhile, the financial crisis also induces fluctuations in higher scales
(corresponding to lower frequencies), even in the approximate component (corresponding to the lowest frequency), which can help
to warn the risk. We therefore argue that all information at different frequencies is important and should be taken into account in
systemic risk measurement. Third, size and leverage appear to be the most robust determinants of systemic risk contribution in the
Chinese banking industry. Smaller banks with higher leverage seems to contribute more systemic risk, which should be paid more
attention by regulators.
Accordingly, different from the existing literature, we employ the two-stage quantile regression method to calculate CoVaR, rather
than a DCC-GARCH method which is employed in Teply and Kvapilikova (2017). Moreover, we consider a reconstruction process
to aggregate the components of financial systemic risk at different scales to the original level, which is also obviously different

2
Q. Xu et al. North American Journal of Economics and Finance 55 (2021) 101354

from that Teply and Kvapilikova (2017) estimate the wavelet based conditional value-at-risk (WCoVaR) at each scale separately. As
for Meng and Huang (2019), they follow Teply and Kvapilikova (2017) and calculate the WCoVaR to quantify the stress performance
of Asian foreign exchange markets, whose purpose is to examine the co-movement characteristics of effective exchange rates across
different frequencies and over time. Unlike Meng and Huang (2019), our purpose is to improve the accuracy of systemic risk measure
by exploiting time–frequency information in financial time series.
The contributions of this paper to the existing literature are two-fold. First, we propose a novel W-QR-CoVaR method which
enables us to exploit the information in both time and frequency domains simultaneously for improving the accuracy of systemic
risk measure. Unlike the previous literature using the wavelet analysis in measuring financial risk on multi-scales, we further consider
a reconstruction process to aggregate the components of financial systemic risk at different scales to the original level. Second, we
apply the W-QR-CoVaR method to measure systemic risk in the Chinese banking industry. The volatile style of the system returns at
different scales helps investors better understand the opportunities, trade-offs, and costs involved with different investment horizons.
Our discussions on the determinants of systemic risk contribution of financial institutions help both investors and regulators to
identify key risk factors, supervise risk level, and prevent financial collapse.
The remainder of this paper is organized as follows. In Section 2 we introduce the MRA based on the maximum overlap discrete
wavelet transform (MODWT) approach, present the W-QR-CoVaR method, and provide a backtesting method. We present the data
and the empirical results in Section 3. Section 4 concludes.

2. Methodology

In this section, we first introduce the MODWT-based MRA approach. Then we propose the W-QR-CoVaR method. Finally, we
present a backtesting procedure for evaluating the accuracy of the systemic risk measure.

2.1. MRA based on MODWT

MRA proposed by Mallat (1989) is a method to decompose a signal into components at different scales and has been widely
used in processing financial data. In general, MRA employs wavelet transform to decompose the data. In this study, we adopt the
MODWT since it can handle data with any sample size. Actually, MODWT is a modification of the discrete wavelet transform (DWT).
Unlike DWT, MODWT has different normalizations for filters, with no downsampling by power of two. As a consequence, the data
at each decomposition level has the same length as the original signal (Zhu et al., 2014), which makes MODWT shift invariant and
ensure the time matching of signals at each decomposition level.
Actually, the MODWT is a linear filtering that transforms a series into coefficients related to variations over a set of scales.
This makes it very convenient for us to analyze high–frequency components of the data through high-pass filters and low-frequency
components through low-pass filters. Since MODWT is a modified version, its definition can be obtained directly from the DWT.
Let {ℎ𝑗,𝑙 , 𝑗 = 1, 2, … , 𝐽 ; 𝑙 = 0, 1, … , 𝐿 − 1} be the DWT wavelet filter and {𝑔𝑗,𝑙 , 𝑗 = 1, 2, … , 𝐽 ; 𝑙 = 0, 1, … , 𝐿 − 1} be the scaling filter,
where 𝐿 is the length of the filter and 𝑗 is the level of the decomposition. The 𝑗th level wavelet filter coefficients ℎ𝑗,𝑙 and the scaling
filter coefficients 𝑔𝑗,𝑙 represent the high-pass and low-pass filters respectively. For a given mother wavelet 𝜓, which is essentially
a band-pass filter, the DWT wavelet filter is actually a finite-length vector of wavelet filter 𝜓 elements. As for MODWT, the 𝑗th
ℎ 𝑔
level wavelet and scaling filters are defined by ℎ̃ 𝑗,𝑙 = √𝑗,𝑙 and 𝑔̃𝑗,𝑙 = √𝑗,𝑙 . Similarly, the 𝑗th level MODWT wavelet filter coefficients
2𝑗 2𝑗
ℎ̃ 𝑗,𝑙 and the scaling filter coefficients 𝑔̃𝑗,𝑙 represent the high-pass and low-pass filters respectively. In this case, 𝑔̃𝑗,𝑙 is related to ℎ̃ 𝑗,𝑙
through a quadrature mirror filter relation 𝑔̃𝑗,𝑙 = (−1)𝑙+1 ℎ̃ 𝑗,𝐿−1−𝑙 . To guarantee the basic properties of wavelets, the MODWT wavelet
∑ ∑𝐿−1 ̃ 2 1 ∑𝐿−1 ̃ ̃
filter ℎ̃ 𝑗,𝑙 must satisfy three conditions: 𝐿−1 ̃
𝑙=0 ℎ𝑗,𝑙 = 0, 𝑙=0 ℎ𝑗,𝑙 = 2𝑗 and 𝑙=0 ℎ𝑗,𝑙 ℎ𝑗,𝑙+2𝑛 = 0, 𝑛 ≠ 0. These conditions are sufficient
to ensure that translated wavelet filters are orthogonal to each other; see Gallegati (2008) for more details. Then, the MODWT of a
{ }
financial time series or return series 𝑹 = 𝑹𝑡 , 𝑡 = 1, 2, … , 𝑇 can be formulated as
𝐿𝑗 −1

̃ 𝑗,𝑡 =
𝒘 ℎ̃ 𝑗,𝑙 𝑹𝑡−𝑙 mod 𝑇 (1)
𝑙=0
𝐿𝑗 −1

𝒗̃ 𝑗,𝑡 = 𝑔̃𝑗,𝑙 𝑹𝑡−𝑙 mod 𝑇 (2)
𝑙=0
( )
where 𝐿𝑗 = 2𝑗 − 1 (𝐿 − 1) + 1, 𝒘
̃ 𝑗,𝑡 denotes the wavelet coefficients, and 𝒗̃ 𝑗,𝑡 denotes the scale coefficients. We rewrite Eqs. (1) and
(2) in a matrix notation as
̃ [(𝐽 +1)×𝑇 ]×𝑇 𝑹𝑇 ×1
̃ [(𝐽 +1)×𝑇 ]×1 = 𝑾
𝒘 (3)
[ ]′
where 𝒘
̃ [(𝐽 +1)×𝑇 ]×1 = 𝒘̃ ′1 , … , 𝒘 , 𝒘̃ ′𝑗 denotes the transpose of 𝒘
̃ ′𝑗 , … , 𝒘̃ ′𝐽 , 𝒗̃ ′𝐽 ̃𝑗, 𝒘̃ 𝑗 and 𝒗̃ 𝐽 are T-dimensional column vectors
𝑇 ×1 𝑇 ×1 𝑇 ×1 𝑇 ×1 𝑇 ×1 𝑇 ×1
𝑇 ×1𝑇 ×1 [ ]′
̃ ̃ ̃ ̃ ̃
which represent the 𝑗th wavelets coefficient and the 𝐽 th scale coefficient separately, and 𝑾 [(𝐽 +1)×𝑇 ]×𝑇 = 𝑾 1 , … , 𝑾 𝑗 , … , 𝑾 𝐽 , 𝑽 𝐽
𝑇 ×𝑇 𝑇 ×𝑇 𝑇 ×𝑇 𝑇 ×𝑇

3
Q. Xu et al. North American Journal of Economics and Finance 55 (2021) 101354

⎛ ℎ̃ 𝑗,0 ℎ̃ 𝑗,𝑇 −1 … ℎ̃ 𝑗,1 ⎞


ℎ̃ 𝑗,2
⎜ ⎟
⎜ ℎ̃ 𝑗,1 ℎ̃ 𝑗,0 … ℎ̃ 𝑗,3
ℎ̃ 𝑗,2 ⎟
consists of wavelet filters and scaling filters, the elements 𝑾 ̃ 𝑗 =⎜ ⋮ ⋮ ⋱ ⋮ ⋮ ⎟ and 𝑽̃ 𝐽 is expressed the similar
⎜̃ ⎟ 𝑇 ×𝑇
𝑇 ×𝑇
⎜ℎ𝑗,𝑇 −2 ℎ̃ 𝑗,𝑇 −3 … ℎ̃ 𝑗,0 ℎ̃ 𝑗,𝑇 −1 ⎟
⎜̃ ⎟
⎝ℎ𝑗,𝑇 −1 ℎ̃ 𝑗,𝑇 −2 … ℎ̃ 𝑗,1 ℎ̃ 𝑗,0 ⎠
̃ 𝑗 with each ℎ̃ 𝑗,𝑙 replaced by 𝑔̃𝐽 ,𝑙 . Then the original time series 𝑅𝑇 ×1
as 𝑾 can be recovered from a reverse exercise of Eq. (3) through:
𝑇 ×𝑇

′ ∑
𝐽
̃
𝑹𝑇 ×1 = 𝑾 ̃ [(𝐽 +1)×𝑇 ]×1 = ̃ ′𝒘 ̃ ′ ̃𝐽 (4)
[(𝐽 +1)×𝑇 ]×𝑇 𝒘 𝑾 𝑗 ̃𝑗 +𝑽𝐽 𝒗
𝑗=1 𝑇 ×𝑇 𝑇 ×1 𝑇 ×𝑇 𝑇 ×1

Then we introduce the wavelet detail 𝐷̃ 𝑗 and the 𝐽 level wavelet smoothed series 𝑆̃𝐽 , which are defined by
𝑇 ×1 𝑇 ×1

̃ = ̃ ′𝒘 (5)
𝐷𝑗 𝑾 𝑗 ̃𝑗
𝑇 ×1 𝑇 ×𝑇 𝑇 ×1

̃ = 𝑽̃ ′ 𝒗̃ 𝐽
𝑆𝐽 (6)
𝐽
𝑇 ×1 𝑇 ×𝑇 𝑇 ×1

In this way, return series can be finally reconstructed by



𝐽
𝑹𝑇 ×1 = ̃
𝐷𝑗 + 𝑆𝐽̃ (7)
𝑗=1 𝑇 ×1 𝑇 ×1

where the 𝑗th level wavelet detail 𝐷̃ 𝑗 corresponds to a frequency band with the period ranging from 2𝑗 to 2𝑗+1 . As a consequence,
the original time series can be partly or even fully recovered. This is the so called MODWT-based MRA. Using MRA we can extract
useful and more comprehensive information at different scales from the time series. It should be mentioned that the selected mother
wavelet, such as Morlet wavelet, Mexican hat wavelet, and Gaussian wavelet, determines the type of MRA. The mother wavelet is
essentially a bandpass filter and defines the shape of wavelet and scaling filters.

2.2. W-QR-CoVaR measure

For a given random variable 𝑅𝑖𝑡 which represents the return of financial institution 𝑖 at time 𝑡, the negative return is considered
as a loss. Under the confidence level of 100 × (1 − 𝑞) %, the VaR of institution 𝑖 at time 𝑡 is generally defined as the 𝑞th quantile of
the returns distribution, namely
( )
Pr 𝑅𝑖𝑡 ≤ 𝑉 𝑎𝑅𝑖𝑞,𝑡 = 𝑞 (8)

where Pr (⋅) denotes the probability. In Adrian and Brunnermeier (2016), CoVaR of the system conditional on institution 𝑖 being in
financial distress at time 𝑡 is defined by
( )
𝑠|𝑖
Pr 𝑅𝑠𝑡 ≤ 𝐶𝑜𝑉 𝑎𝑅𝑞,𝑡 |𝑅𝑖𝑡 = 𝑉 𝑎𝑅𝑖𝑞,𝑡 = 𝑞 (9)
𝑠|𝑖
where 𝑅𝑠𝑡 stands for the return of the system at time 𝑡. In other words, 𝐶𝑜𝑉 𝑎𝑅𝑞,𝑡 is essentially the VaR of the system conditional
{ }
on the event 𝑅𝑖𝑡 = 𝑉 𝑎𝑅𝑖𝑞,𝑡 , that is, the 𝑞th quantile of the system returns distribution under the institution 𝑖 being in financial
distress. Therefore, the two-stage quantile regression scheme can be used to estimate CoVaR, which is termed as QR-CoVaR. In this
study, we apply the combination of wavelet analysis and quantile regression to the estimation of CoVaR and call this hybrid method
W-QR-CoVaR. Its schematic diagram is shown in Fig. 1, which contains three main steps.
Step 1: Wavelet decomposition. For each institution 𝑖 with 𝑇 observations, we decompose the original return series 𝑅 into 𝐽
{ }
levels using MRA based on MODWT. Then we acquire 𝐽 +1 time series at different scales as 𝑅1 , … , 𝑅𝑗 , … , 𝑅𝐽 , 𝑅𝐽 +1 , which consists
of wavelet details 𝐷𝑗 = 𝑅𝑗 𝑓 𝑜𝑟 𝑗 = 1, 2, … , 𝐽 and smoothed series 𝑆𝐽 = 𝑅𝐽 +1 .
Step 2: CoVaR estimation on multi-scales. At each scale, we estimate CoVaR via the two-stage quantile regression method.
{ }
Specifically, we use the decomposition components 𝑅1 , … , 𝑅𝑗 , … , 𝑅𝐽 , 𝑅𝐽 +1 acquired at Step 1 as the dependent variable and a
set of lagged state variables as independent variables. Thus, the two-stage scheme is conducted with two linear quantile regression
models
( )
𝑄𝑢𝑎𝑛𝑡𝑞 𝑅𝑖𝑗,𝑡 = 𝛼𝑗,𝑞
𝑖 𝑖
+ 𝛾𝑗,𝑞 𝑀𝑡−1 (10)
( )
𝑠𝑦𝑠𝑡𝑒𝑚|𝑖 𝑠𝑦𝑠𝑡𝑒𝑚|𝑖 𝑖 𝑠𝑦𝑠𝑡𝑒𝑚|𝑖
𝑄𝑢𝑎𝑛𝑡𝑞 𝑅𝑠𝑦𝑠𝑡𝑒𝑚
𝑗,𝑡 = 𝛼𝑗,𝑞 + 𝛽𝑗,𝑞 𝑅𝑗,𝑡 + 𝛾𝑗,𝑞 𝑀𝑡−1 (11)
( ) ( )
where 𝑄𝑢𝑎𝑛𝑡𝑞 𝑅𝑖𝑗,𝑡 and 𝑄𝑢𝑎𝑛𝑡𝑞 𝑅𝑠𝑦𝑠𝑡𝑒𝑚
𝑗,𝑡 denote the 𝑞th conditional quantiles of 𝑅𝑖𝑗,𝑡 and 𝑅𝑠𝑦𝑠𝑡𝑒𝑚
𝑗,𝑡 , 𝑀𝑡−1 denotes the first-order lag
state variables and the lag periods are changeable in empirical applications. It should be noted that the variables in 𝑀𝑡−1 are not
interpreted as systematic risk factors, but rather as systemic state variables, which are used to shift the conditional mean and the
conditional volatility (Adrian & Brunnermeier, 2016). Consequently, it is acceptable that the variables in 𝑀𝑡−1 are not decomposed

4
Q. Xu et al. North American Journal of Economics and Finance 55 (2021) 101354

Fig. 1. Schematic representation of the proposed method.

in Eqs. (10)–(13), but are consistent with 𝑅𝑠𝑦𝑠𝑡𝑒𝑚


𝑗,𝑡 and 𝑅𝑖𝑗,𝑡 as conditioning variables. In addition, considering the signal decomposition
with wavelet analysis is a process of discretization and thus the decomposed series have discretization errors. If we decompose 𝑀𝑡−1
into 𝐽 levels and employ the decompositions in Eqs. (10)–(13), the accumulation of such errors in a set of state variables at the 𝑗th
level will lead a big error, which is also proved( in our experiments. To this end, ) we use 𝑀𝑡−1 in our systemic risk model. Based on
quantile regressions, we estimate parameters 𝛼𝑗,𝑞 𝑖 , 𝛾 𝑖 , 𝛼 𝑠𝑦𝑠𝑡𝑒𝑚|𝑖 , 𝛽 𝑠𝑦𝑠𝑡𝑒𝑚|𝑖 , 𝛾 𝑠𝑦𝑠𝑡𝑒𝑚|𝑖 and use them to obtain
𝑗,𝑞 𝑗,𝑞 𝑗,𝑞 𝑗,𝑞

𝑉 𝑎𝑅𝑖𝑗,𝑡 (𝑞) = 𝛼̂ 𝑗,𝑞


𝑖 𝑖
+ 𝛾̂𝑗,𝑞 𝑀𝑡−1 (12)
𝑠|𝑖 𝑠𝑦𝑠𝑡𝑒𝑚|𝑖 𝑠𝑦𝑠𝑡𝑒𝑚|𝑖 𝑠𝑦𝑠𝑡𝑒𝑚|𝑖
𝐶𝑜𝑉 𝑎𝑅𝑗,𝑡 (𝑞) = 𝛼̂ 𝑗,𝑞 + 𝛽̂𝑗,𝑞 𝑉 𝑎𝑅𝑖𝑗,𝑡 (𝑞) + 𝛾̂𝑗,𝑞 𝑀𝑡−1 (13)
𝑠|𝑖
where 𝑉 𝑎𝑅𝑖𝑗,𝑡 (𝑞) denotes the VaR of institution 𝑖 at scale 𝑗 under the confidence level of 100 × (1 − 𝑞) %, and (𝑞) is the risk 𝐶𝑜𝑉 𝑎𝑅𝑗,𝑡
of the system when the institution 𝑖 is in financial distress at scale 𝑗 under the confidence level of 100 × (1 − 𝑞) %.
Step 3: CoVaR reconstruction. Finally, we employ the inverse wavelet transform to aggregate CoVaRs on multi-scales into a
𝑠|𝑖
reconstruction CoVaR, namely 𝑊 -𝑄𝑅-𝐶𝑜𝑉 𝑎𝑅𝑡 (𝑞)
Furthermore, we also consider the systemic risk contribution of a particular institution 𝑖 and defined it as
𝑠|𝑖 𝑠|𝑅𝑖𝑡 =𝑉 𝑎𝑅𝑖𝑡 (𝑞) 𝑠|𝑅𝑖𝑡 =𝑉 𝑎𝑅𝑖𝑡 (50%)
𝛥𝑊 -𝑄𝑅-𝐶𝑜𝑉 𝑎𝑅𝑡 (𝑞) = 𝑊 -𝑄𝑅-𝐶𝑜𝑉 𝑎𝑅𝑡 (𝑞) − 𝑊 -𝑄𝑅-𝐶𝑜𝑉 𝑎𝑅𝑡 (50%) (14)

where 𝑉 𝑎𝑅𝑖𝑡 (50%) means the median of returns distribution of institution 𝑖 and we use 𝑅𝑖𝑡 = 𝑉 𝑎𝑅𝑖𝑡 (50%) as the normal state of
institution 𝑖. That is, the systemic risk contribution is the difference between the systemic risk of institution 𝑖 being in financial
distress and that of institution 𝑖 being in a normal state. As a result, it can be calculated by
𝑠|𝑖 𝑠|𝑖 𝑠|𝑖
𝛥𝑊 -𝑄𝑅-𝐶𝑜𝑉 𝑎𝑅𝑡 (𝑞) = 𝑊 -𝑄𝑅-𝐶𝑜𝑉 𝑎𝑅𝑡 (𝑞) − 𝑊 -𝑄𝑅-𝐶𝑜𝑉 𝑎𝑅𝑡 (50%) (15)

2.3. Backtesting method

Referring to Christoffersen (1998) and Kupiec (1995), we design a procedure to backtesting W-QR-CoVaR measure. Suppose
that there are 𝐼 institutions selected in the sample, and the number of observations for each institution 𝑖 (𝑖 = 1, 2, … , 𝐼) is 𝑇 with

5
Q. Xu et al. North American Journal of Economics and Finance 55 (2021) 101354

𝑡 = 1, 2, … , 𝑇 . For each institution 𝑖, we can compare the estimated VaRs with the losses distributed in the lower tail of 𝑅𝑖𝑡 and define
the hit sequence of violations as
{
𝑖 1 if 𝑅𝑖𝑡 = 𝑉 𝑎𝑅𝑖𝑞,𝑡
𝑍𝑡 = (16)
0 if 𝑅𝑖𝑡 ≠ 𝑉 𝑎𝑅𝑖𝑞,𝑡

In reality, the condition 𝑅𝑖𝑡 = 𝑉 𝑎𝑅𝑖𝑞,𝑡 is difficult to guarantee. As an alternative, we introduce an error 𝜀 as small as possible to
| | | |
replace 𝑅𝑖𝑡 = 𝑉 𝑎𝑅𝑖𝑞,𝑡 with |𝑅𝑖𝑡 − 𝑉 𝑎𝑅𝑖𝑞,𝑡 | < 𝜀 and replace 𝑅𝑖𝑡 ≠ 𝑉 𝑎𝑅𝑖𝑞,𝑡 with |𝑅𝑖𝑡 − 𝑉 𝑎𝑅𝑖𝑞,𝑡 | ≥ 𝜀. As a consequence, the hit sequence will
| | | |
return one when the loss of institution 𝑖 is very close to its VaR, while zero otherwise. Observations with returned one value will
be selected to form a sub-sample with the sample size assumed to be 𝑁. For this sub-sample, we construct a second hit sequence,
which compares the loss of the financial system with CoVaR and is defined as
{ 𝑠|𝑖
𝑠|𝑖 1 if 𝑅𝑠𝑡 ≤ 𝐶𝑜𝑉 𝑎𝑅𝑞,𝑡
𝑍𝑡 = 𝑠|𝑖
(17)
𝑠
0 if 𝑅𝑡 > 𝐶𝑜𝑉 𝑎𝑅𝑞,𝑡
It should be noted that the optimal 𝜀 is selected using a data-driven method. In practice, to make the backtesting process valid,
we need a relatively large sample (e.g., the sample size is equal to or greater than 30). In the process of backtesting CoVaR, it is
| |
important to consider the condition 𝑅𝑖𝑡 = 𝑉 𝑎𝑅𝑖𝑞,𝑡 with |𝑅𝑖𝑡 − 𝑉 𝑎𝑅𝑖𝑞,𝑡 | < 𝜀. In this case, a larger 𝜀 can make much more observations
| |
to meet the condition. However, when 𝜀 increases, the deviation between 𝑅𝑖𝑡 and 𝑉 𝑎𝑅𝑖𝑞,𝑡 may become too large to guarantee the
condition, which will render the backtesting less efficient. To sum up, when 𝜀 is too small, the selected observations in sub-sample
with size N is too few to backtesting, while 𝜀 is too large, the backtesting results tend to be inaccurate. To this end, an appropriate
value of 𝜀 is important in this procedure (see Xu et al. (2018)). In the experiment, we select 𝜀 by trial-and-error.
To evaluate the accuracy of W-QR-CoVaR measure, we adopt the unconditional coverage (UC) test of Kupiec (1995) and the
conditional coverage (CC) test of Christoffersen (1998), respectively. As far as we know, valid CoVaR forecasts should satisfy the
hypothesis: the probability that ex-post losses exceed the CoVaR forecasts is equal to the expected probability of failure. In the
experiments, we set the expected violation rate 𝑞 of CoVaR
( measure
) with the confidence level 100 × (1 − 𝑞) % in advance and acquire
𝑠|𝑖
the corresponding real violation rate through 𝑝 = Pr 𝑍𝑡 = 1 for both tests.
For the UC test, the null hypothesis is 𝐻0 ∶ 𝑝 = 𝑞, and we use the likelihood ratio test of Kupiec (1995) which is defined as
[ ( ) ( )]
𝐿𝑅𝑢𝑐 = −2 ln 𝑞 𝑁1 (1 − 𝑞)𝑁0 − ln 𝑝𝑁1 (1 − 𝑝)𝑁0 ∼ 𝜒 2 (1) (18)
𝑠|𝑖 𝑠|𝑖
where 𝑁0 is the number of observations such that 𝑅𝑠𝑡 > 𝐶𝑜𝑉 𝑎𝑅𝑞,𝑡 , 𝑁1 is the number of observations such that 𝑅𝑠𝑡 ≤ 𝐶𝑜𝑉 𝑎𝑅𝑞,𝑡 , and
𝑁1 = 𝑁 − 𝑁0 .
𝑠|𝑖
As regards the CC test, assuming that the second hit sequence 𝑍𝑡 is dependent over time, the null hypothesis is 𝐻0 ∶ 𝑝 = 𝑝01 =
𝑠|𝑖 𝑠|𝑖 𝑠|𝑖 𝑠|𝑖
𝑝11 , where 𝑝01 = Pr(𝑍𝑡 = 1|𝑍𝑡−1 = 0) and 𝑝11 = Pr(𝑍𝑡 = 1|𝑍𝑡−1 = 1). We employ the likelihood ratio test of Christoffersen (1998)
which is defined as:
⎡ ⎤
(1 − 𝑝)𝑁00 +𝑁10 𝑝𝑁01 +𝑁11
𝐿𝑅𝑐𝑐 = 𝐿𝑅𝑢𝑐 − 2 ln ⎢ ⎥ ∼ 𝜒 2 (2) (19)
⎢ (1 − 𝑝 )𝑁00 𝑝 01 (1 − 𝑝 10 )𝑝 11 ⎥
𝑁 𝑁 𝑁
⎣ 01 01 11 11 ⎦
𝑠|𝑖 𝑠|𝑖
where 𝑁𝑢,𝑣 (𝑢, 𝑣 = 0, 1) represent the number of periods satisfying the conditions 𝑍𝑡−1 = 𝑢 and 𝑍𝑡 = 𝑣, for example, 𝑁0,1 denotes
𝑠|𝑖 𝑠|𝑖 𝑁01 +𝑁11 𝑁01 𝑁11
the number of periods with 𝑍𝑡−1 = 0 followed by a period with 𝑍𝑡 = 1; 𝑝 = 𝑁00 +𝑁01 +𝑁10 +𝑁11
, 𝑝01 = 𝑁00 +𝑁01
and 𝑝11 = 𝑁10 +𝑁11
.
Then we define failure times (FT) and average error (AE) to evaluate the backtesting results. FT refers to the number of institutions
which reject the null hypothesis, and is defined by

𝐼
( )
𝐹𝑇 = 𝟏 𝐻0 is rejected for institution 𝑖 (20)
𝑖=1

where 1 is the indicator function, and 𝐼 is the number of selected institutions. Moreover, AE stands for the average value of all
errors of 𝐼 institutions and is defined by

1 ∑|
𝐼
𝐴𝐸 = 𝑝 − 𝑞 || (21)
𝐼 𝑖=1 | 𝑖

where 𝑝𝑖 denotes the violation rate of institution 𝑖 and 𝑞 denotes the expected violation rate. In general, the preferred model tends
to have smaller average error (AE) or fewer failure times (i.e., FT(uc) and FT(cc)), which means a more accurate prediction in
statistics.

3. Empirical analysis

In this section, we compare the performance of W-OR-CoVaR and QR-CoVaR for measuring systemic risk in the Chinese banking
industry, and study the contribution of each individual bank to systemic risk. Furthermore, we explore the relation between the
systemic risk and risk factors.

6
Q. Xu et al. North American Journal of Economics and Finance 55 (2021) 101354

Table 1
The information of selected banks in China.
Bank Abbreviation Total equity Founding date Stock code
Bank of Communications BOCOM 74 262.73 1987.03.30 601328
Industrial and Commercial Bank ICBC 356 406.26 1984.01.01 601398
China Construction Bank CCB 250 010.98 2004.09.17 601939
Bank of China BOC 294 387.79 2004.08.26 601988
Pingan Bank PAB 17 170.41 1987.12.22 000001
Shanghai Pudong Development Bank SPDB 29 352.08 1992.10.19 600000
Huaxia Bank HXB 12 822.69 1992.10.14 600015
China Minsheng Bank CMBC 43 782.42 1996.02.07 600016
China Merchants Bank CMB 25 219.85 1987.03.31 600036
China Industrial Bank CIB 20 774.19 1988.08.22 601166
China Citic Bank CITICB 48 934.80 2006.12.31 601998
Bank of Beijing BOB 21 142.98 1996.01.29 601169
Bank of Nanjing BON 8 482.21 1996.02.06 601009
Bank of Ningbo BONB 5 069.74 1997.04.10 002142

Notes: The total equity (Million RMB Yuan) is observed on Sep. 28, 2018.

3.1. Data

Considering that extreme market conditions are suitable for exploring systemic risk or its spillover effects, we select all 14 banks
listed in China before 2008. The sample period is from September 28, 2007, to September 28, 2018, including 2242 trading days,
which contains the 2008 global financial crisis and the 2015 Chinese stock market crash. The detailed information of 14 selected
banks, such as stock code and full name with the corresponding abbreviation, is shown in Table 1.
We collect the data from the Chinese Stock Market and Accounting Research Database (CSMAR) and the Genius Finance Database.
The daily return of each bank is calculated through the first difference of the logarithm closing price, namely
[ ( ) ( 𝑖 )]
𝑅𝑖𝑡 = ln 𝑃𝑡𝑖 − ln 𝑃𝑡−1 × 100 (22)
where 𝑅𝑖𝑡 and 𝑃𝑡𝑖 denote the return and closing price of bank 𝑖 at time 𝑡, respectively. The daily return of the whole system is
constructed by the weighted average of daily returns of all banks. In this study, we use the proportion of the total equity as the
weight, and define the system return as

𝐼
𝑇 𝐸𝑡𝑖
𝑅𝑠𝑡 = ∑ 𝑅𝑖
𝑖 𝑡
(23)
𝑖=1 𝑖 𝑇 𝐸𝑡

where 𝑅𝑠𝑡 denotes the return of the whole system at time 𝑡 and 𝑇 𝐸𝑡𝑖 is the total equity of bank 𝑖 at time 𝑡.
The summary statistics of daily returns are reported in Table 2. We see that all the values of kurtosis are greater than three,
which means that the returns distribution is fat-tailed. Moreover, the J-B test results also reject the hypothesis that the returns are
normally distributed.
Following Adrian and Brunnermeier (2016) and Wang et al. (2018), we choose the return and volatility of CSI300 index, the
3-month Treasury bill rate, term spread and liquidity spread as state variables shown in Table 4. They are daily observed and
summarized in Table 3. In addition, bank-specific characteristics used as risk factors are also listed in Table 4. Except for Beta,
the other five firm characteristics are quarterly observed. To ensure that all regression variables have the common frequency, we
average the weekly Beta within each quarter to obtain the quarterly Beta.

3.2. MRA results

We use the MODWT method to conduct MRA for the daily returns of the banking system and all individual banks, respectively.
In the MODWT, we adopt the mother wavelet LA(8) of the Daubechies family. For the decomposition level, we take the common
trading cycles into account and use the MRA method to acquire four decomposed time series with different cycle lengths, which are
represented by wavelet details 𝐷1 , 𝐷2 , 𝐷3 and the smooth series 𝑆3 . More specifically, 𝐷1 denotes the short-term (less than 8 days),
𝐷2 denotes the medium-term (8–32 days), 𝐷3 denotes the long term (32–64 days), and 𝑆3 denotes the cycle longer than 64 days.
We plot in Fig. 2 the MRA results of the banking system returns. Three general conclusions then emerge. First, the volatility
becomes smaller as the decrease of frequency (or increase of scale), which shows that risk tends to concentrate on the higher
frequencies. It is intuitively sound since the high–frequency information contains more noise components and the low-frequency
information contains more trend components. Second, during the two periods of the 2007–2009 global financial crisis and the
2015–2016 China’s stock market crash, the Chinese banking industry experiences significantly higher fluctuations compared to the
other years. This phenomenon occurs on all decomposed series at different scales. The increase in medium-term (𝐷2 ) and long-term
(𝐷3 ) volatility suggests that the financial crisis can generate fluctuations on a monthly or quarterly basis, which means a longer
period of financial distress. Third, compared with wavelet details, the cycle 𝑆3 (representing cycles longer than 64 days) is smoother
and has less energy. However, it still has significant fluctuations when the crisis breaks out, which could play a crucial role in early
warning and guiding the financial crisis. To sum up, both high–frequency information and low-frequency information are of great
importance for measuring systemic risk.

7
Q. Xu et al. North American Journal of Economics and Finance 55 (2021) 101354

Table 2
Summary statistics for the returns of the banking system and each bank.
System and bank Obs. Min Mean Max S.D. Skewness Kurtosis J-B test
Statistic P-value
SYSTEM 2242 −10.4973 −0.0036 9.5327 1.5892 −0.1755 10.2673 4945 0
BOCOM 2242 −10.5999 −0.0196 9.6239 1.9514 −0.1452 10.2214 4879 0
ICBC 2242 −10.5365 0.0038 9.5329 1.6128 −0.2314 11.2436 6368 0
CCB 2242 −10.6415 0.0003 9.5670 1.7820 −0.0463 9.9544 4519 0
BOC 2242 −10.5790 −0.0068 9.6812 1.6248 0.2038 11.7367 7146 0
PAB 2242 −10.5629 −0.0100 9.5699 2.1929 −0.1538 8.0104 2354 0
HXB 2242 −10.5677 −0.0093 9.5922 2.2365 −0.1241 7.1599 1622 0
CMBC 2242 −10.5368 −0.0155 9.5780 2.0269 0.0302 8.1727 2500 0
CMB 2242 −10.5278 0.0024 9.5540 2.0999 −0.0092 7.4660 1863 0
CIB 2242 −10.5585 −0.0088 9.5786 2.2417 −0.0629 7.4854 1881 0
CITICB 2242 −10.5636 −0.0253 9.6149 2.2247 0.1517 7.3779 1799 0
BOB 2242 −10.5521 −0.0140 9.5811 2.0554 0.0612 8.1008 2432 0
BON 2242 −10.5502 0.0051 9.5914 2.1968 −0.1046 7.5900 1972 0
BONB 2242 −10.5532 0.0055 9.5994 2.3390 −0.0556 6.4762 1130 0

Notes: (1) SYSTEM represents the banking system whose returns are constructed by Eq. (22). (2) The other abbreviations denote 14 banks whose full names are
listed in Table 1.

Table 3
Summary statistics for the state variables.
Variables Obs. Min Mean Max S.D. Skewness Kurtosis J-B test
Statistic P-value
CSI300-return 2242 −8.75 −0.02 9.34 1.67 −0.43 7.17 1690 0
CSI300-vol 2242 0.27 1.49 4.62 0.81 1.36 4.60 932 0
Treasury3m 2242 0.80 2.71 5.11 0.73 −0.29 3.31 41 0
Term spread 2242 −1.51 0.90 2.50 0.55 0.77 4.14 340 0
Liquidity spread 2242 −1.06 1.47 7.99 0.72 1.11 7.49 2339 0

Table 4
The detailed information of state variables and firm characteristics.
Category Variable Frequency Definition
State variables CSI300-return Daily Computed from the closing price of the China Securities Index 300
( )
𝑅𝑡 = ln 𝑃𝑡 ∕𝑃𝑡−1 × 100
CSI300-vol Daily The 5-day rolling standard deviation of daily return of the China
Securities Index 300
Treasury3m Daily The 3-month Treasury bill rate
Term spread Daily The spread between the 10-year and 3-month Treasury bill rate
Liquidity spread Daily The difference of the 3-month collateral repo rate and the 3-month
Treasury bill rate
Firm Market/book Quarterly The ratio of the market value to the book value of total equity
characteristics PE ratio Quarterly The ratio of share price to annual earnings per share
Size Quarterly The logarithm of market valued total assets
ROA Quarterly The value of net profit divided by total assets
Leverage Quarterly The value of total assets divided by total equity (in book values)
Beta Weekly The value of the Beta coefficient of each company

3.3. CoVaR backtesting

In empirical analysis, we obtain the coefficients in the two-stage quantile regression models, which are not presented here to
save limited space. The results show that the coefficients of CSI300-vol in Eqs. (12) and (13) are negative and significant at all
decomposed levels, which implies that the greater the volatility of CSI300 Index, the larger the absolute value of CoVaR of the
system will be. Moreover, the coefficients of VaR in Eq. (13) are positive and significant at all decomposed levels, which means that
the systemic risk will increase when an individual bank gets into a trouble. These results are in line with our expectations.
We apply the W-QR-CoVaR and QR-CoVaR methods to measure systemic risk in the Chinese banking industry with 𝑞 =
0.01, 0.05, 0.1 corresponding to the 99%, 95%, and 90% confidence levels. However, in the extreme condition of 𝑞 = 0.01, the number
of selected sub-sample according to the definition in Section 2.3 is too small to continue the backtesting procedure. Consequently,
we only present in Table 5 the backtesting results of 𝑊 -𝑄𝑅-𝐶𝑜𝑉 𝑎𝑅𝑠|𝑖 and 𝑄𝑅-𝐶𝑜𝑉 𝑎𝑅𝑠|𝑖 measures at 𝑞 = 0.05, 0.1 under different
values of 𝜀.
For backtesting CoVaR, the significance level used to reject the null hypothesis of the two tests is 5%. In the case of 𝑞 = 0.05,
𝜀 ranges from 0.6 to 1. From Table 5 we find that W-QR-CoVaR has smaller values of AE, FT(uc) and FT(cc) than those of QR-
CoVaR, and the results are consistent under different values of 𝜀. Moreover, the backtesting results of 𝑞 = 0.1 are similar to those

8
Q. Xu et al. North American Journal of Economics and Finance 55 (2021) 101354

Fig. 2. MRA of the banking system returns.

of 𝑞 = 0.05. In the case of different values of 𝜀, which ranges from 0.28 to 0.36, the W-QR-CoVaR method always performs better
than QR-CoVaR in terms of both AE and FT. Referring to Girardi and Ergün (2013), we provide the average test statistics with the
corresponding average 𝑝-values in Table 6, where 𝐿𝑅𝑢𝑐 denotes the test statistic for the UC test of Kupiec (1995) and 𝐿𝑅𝑐𝑐 denotes
the test statistic for the CC test of Christoffersen (1998). The larger 𝑝-value implies that we do not have sufficient evidence to reject
the null hypothesis. As shown in Table 6, W-QR-CoVaR has smaller average test statistics and larger average 𝑝-values than those of
QR-CoVaR. These results are consistent under different values of 𝜀 at both 𝑞 = 0.05 and 𝑞 = 0.1 quantiles. Therefore, we conclude
that the W-QR-CoVaR method performs well in backtesting experiments.
In addition, we depict the returns of the banking system, 𝑊 -𝑄𝑅-𝐶𝑜𝑉 𝑎𝑅𝑠|𝑖 and 𝑄𝑅-𝐶𝑜𝑉 𝑎𝑅𝑠|𝑖 for each bank with 𝑞 = 0.05 in Fig. 3
and with 𝑞 = 0.1 in Fig. 4. In each subfigure, the gray points denote the returns of the banking system, the blue line and the red line
stand for 𝑊 -𝑄𝑅-𝐶𝑜𝑉 𝑎𝑅𝑠|𝑖 and 𝑄𝑅-𝐶𝑜𝑉 𝑎𝑅𝑠|𝑖 of each bank, respectively. As shown in Figs. 3 and 4, the lines for 𝑊 -𝑄𝑅-𝐶𝑜𝑉 𝑎𝑅𝑠|𝑖 and
𝑄𝑅-𝐶𝑜𝑉 𝑎𝑅𝑠|𝑖 have a similar trend, and the blue line for 𝑊 -𝑄𝑅-𝐶𝑜𝑉 𝑎𝑅𝑠|𝑖 is mostly below the red line for 𝑄𝑅-𝐶𝑜𝑉 𝑎𝑅𝑠|𝑖 . Moreover,
when the global financial crisis or the China’s stock market crash happens, the values of the two systemic risk measures for the
Chinese banking system significantly increase. These empirical results lead us to the conclusion that the wavelet-based quantile
regression framework is sufficiently flexible to estimate CoVaR. Moreover, they also show that the W-QR-CoVaR method is an
efficient tool for risk-averse investors to control risk losses and financial institutions to implement robust risk management.

3.4. Relation between systemic risk and risk factors

Besides the predictive accuracy of measures, the pro-cyclicality is another important issue of systemic risk regulation. It is
necessary to identify the determinants of systemic risk contribution. To this end, we calculate 𝛥𝑊 -𝑄𝑅-𝐶𝑜𝑉 𝑎𝑅 to quantify the
systemic risk contribution of each individual bank, and then explore the relation between 𝛥𝑊 -𝑄𝑅-𝐶𝑜𝑉 𝑎𝑅 and observable bank-
specific characteristics. Accordingly, we employ the panel regression model with fixed effects on the panel data which contain lagged
independent variables of bank-specific characteristics and the dependent variable of quarterly-aggregated 𝛥𝑊 -𝑄𝑅-𝐶𝑜𝑉 𝑎𝑅 at both

9
Q. Xu et al. North American Journal of Economics and Finance 55 (2021) 101354

Table 5
The backtesting results of 𝑞 = 0.05 and 𝑞 = 0.1.
𝑞 𝜀 QR-CoVaR W-QR-CoVaR
AE (%) FT(uc) FT(cc) Size AE (%) FT(uc) FT(cc) Size
0.05 0.6 3.3939 6 6 155 2.5999 0 0 34
0.7 3.1814 8 3 184 2.6891 0 0 40
0.8 2.7342 6 6 219 1.8209 0 0 47
0.9 2.1005 5 3 256 1.6212 0 0 56
1 1.6862 2 2 296 1.5198 0 0 67
0.1 0.28 3.7320 5 6 138 3.0213 2 1 41
0.30 3.8592 6 7 146 3.2730 2 1 44
0.32 3.6026 6 8 160 2.7920 1 1 47
0.34 3.8121 5 8 169 2.9335 1 2 51
0.36 3.3300 5 7 177 3.1345 1 2 54

Notes: (1) AE denotes the average error of CoVaR measure defined in Eq. (21). (2) FT is the
number of banks which reject the null hypothesis that real violation rate is equal to excepted
violation rate at the significance level of 5%. (3) FT(uc) denotes the results of the UC test of
Kupiec (1995); FT(cc) denotes the results of the CC test of Christoffersen (1998). (4) Size denotes
| |
the observations that meet the condition |𝑅𝑖𝑡 − 𝑉 𝑎𝑅𝑖𝑞,𝑡 | < 𝜀.
| |

Table 6
Average test statistics and 𝑝-values for CoVaR backtesting.
𝑞 𝜀 LR test QR-CoVaR W-QR-CoVaR
Average 𝑝 values Average 𝑝 values
0.05 0.6 𝐿𝑅𝑢𝑐 4.03 0.22 0.57 0.56
𝐿𝑅𝑐𝑐 5.24 0.24 0.91 0.71
0.7 𝐿𝑅𝑢𝑐 4.10 0.21 0.70 0.52
𝐿𝑅𝑐𝑐 4.77 0.23 1.13 0.63
0.8 𝐿𝑅𝑢𝑐 3.44 0.20 0.49 0.61
𝐿𝑅𝑐𝑐 4.67 0.17 0.93 0.67
0.9 𝐿𝑅𝑢𝑐 2.50 0.25 0.46 0.61
𝐿𝑅𝑐𝑐 3.62 0.26 0.97 0.65
1 𝐿𝑅𝑢𝑐 2.00 0.31 0.57 0.61
𝐿𝑅𝑐𝑐 3.15 0.34 1.68 0.52
0.1 0.28 𝐿𝑅𝑢𝑐 2.40 0.24 0.91 0.55
𝐿𝑅𝑐𝑐 4.31 0.27 1.50 0.56
0.30 𝐿𝑅𝑢𝑐 2.75 0.22 1.09 0.51
𝐿𝑅𝑐𝑐 5.00 0.22 1.85 0.50
0.32 𝐿𝑅𝑢𝑐 2.61 0.25 1.00 0.55
𝐿𝑅𝑐𝑐 5.17 0.25 1.90 0.49
0.34 𝐿𝑅𝑢𝑐 2.98 0.20 0.93 0.51
𝐿𝑅𝑐𝑐 5.81 0.17 2.06 0.47
0.36 𝐿𝑅𝑢𝑐 2.53 0.25 1.04 0.47
𝐿𝑅𝑐𝑐 5.59 0.21 2.15 0.47

Note: 𝐿𝑅𝑢𝑐 is the statistic of Kupiec (1995) test for the unconditional coverage property and
𝐿𝑅𝑐𝑐 is the statistic of Christoffersen (1998) test for the conditional coverage property.

95% and 90% confidence levels. We obtain the quarterly Beta by averaging the weekly Beta within each quarter and the quarterly
risk measurements (i.e., 𝛥𝑊 -𝑄𝑅-𝐶𝑜𝑉 𝑎𝑅 and VaR) by averaging their daily estimated values within each quarter.
For the reason that the complete data of bank-specific characteristics are available from 2010, we use the panel data for a sample
period which spans from Q3 2010 to Q3 2018, including 34 quarters. The summary statistics for the variables in the panel data
model is presented in Table 7. We consider three different lag periods for the variables of bank-specific characteristics, namely one
quarter, one year and two years. For the convenience of exploring the relation, we change the negative sign of 𝛥𝑊 -𝑄𝑅-𝐶𝑜𝑉 𝑎𝑅 or
VaR into a positive one. The panel regression results based on three different forecast horizons are reported in Table 8, including the
estimated coefficient of each variable and its standard errors in parentheses. We can use these statistically significant coefficients to
learn the extent to which the bank-specific characteristics affect system risk contributions. For example, the regression coefficient
of leverage is 0.036 for the two-year horizon, which indicates that for every unit increment in leverage of a bank, the systemic risk
contribution at 𝑞 = 0.05 will decrease by 0.036 units.
From Table 8, the significance of size and leverage persists throughout all forecast horizons. In this regard, we infer that these two
variables appear to be the most robust determinants of systemic risk contribution. The regression coefficients of size are significantly
negative, suggesting that banks with larger size contribute smaller systemic risk. As regards leverage, the regression coefficients are
positive and statistically significant, implying that banks with higher leverage contribute more systemic risk. Furthermore, the VaR
of each bank has significantly positive effects on the systemic risk contribution at both one quarter and one year horizon, which
means that an individual bank can decrease the systemic risk through decreasing its own risk. Unlike Adrian and Brunnermeier
(2016) and Zeb and Rashid (2019), size has a negative impact on the systemic risk contribution. We believe that larger banks often

10
Q. Xu et al. North American Journal of Economics and Finance 55 (2021) 101354

Fig. 3. Time-varying 𝑊 -𝑄𝑅-𝐶𝑜𝑉 𝑎𝑅𝑠|𝑖 and 𝑄𝑅-𝐶𝑜𝑉 𝑎𝑅𝑠|𝑖 of 14 banks with 𝑞 = 0.05. In each subfigure the gray points is the returns of the banking system, the
blue line stands for 𝑊 -𝑄𝑅-𝐶𝑜𝑉 𝑎𝑅𝑠|𝑖 of each bank and the red line stands for 𝑄𝑅-𝐶𝑜𝑉 𝑎𝑅𝑠|𝑖 of each bank. (For interpretation of the references to color in this
figure legend, the reader is referred to the web version of this article.)

Table 7
Summary statistics of variables in panel regressions.
Variables Min Median Mean Max S.D.
5%𝛥𝑊 -𝑄𝑅-𝐶𝑜𝑉 𝑎𝑅 2.02 4.35 4.41 10.33 1.27
10%𝛥𝑊 -𝑄𝑅-𝐶𝑜𝑉 𝑎𝑅 1.47 3.08 3.13 7.32 0.88
5%VaR 1.48 4.20 4.23 10.97 1.47
10%VaR 1.06 3.02 3.05 7.90 1.06
Market/book 0.57 1.13 1.24 1.24 1.06
PE ratio 4.19 7.02 7.61 25.14 2.53
Size 21.59 23.79 24.06 26.60 1.50
ROA 0.16 0.69 0.69 1.40 0.30
Leverage 11.68 15.70 15.88 30.17 2.39
Beta 0.32 0.95 0.93 1.80 0.27

Notes: (1) The frequency of all variables is quarterly. (2) 𝛥𝑊 -𝑄𝑅-𝐶𝑜𝑉 𝑎𝑅 and VaR are obtained
by averaging their daily estimated values within each quarter. (3) Beta is obtained by averaging
the weekly Beta within each quarter.

have better management abilities to control the occurrence of adverse events. Consequently, regulators should pay attention to
supervise the systemic risk contribution from the banks with high leverage and small size.

4. Conclusion

In this paper, we propose a novel W-QR-CoVaR method, to measure systemic risk. It is a hybrid method that combines the
wavelet analysis with the QR-CoVaR method. Actually, different measures have different emphases on systemic risk. In Zhou et al.
(2020), who use the CoVaR, MES, and SRISK methods to examine systemic risk in the Chinese banking system, the empirical results
show that CoVaR and MES exhibit an abnormal rise during the financial crisis, whereas SRISK shows a steady increase throughout
the observation window. The SRISK measure focuses on the long-run expected capital shortfall, and is insensitive to temporary
fluctuations compared with the other two measures. When it turns to CoVaR measure, the results may differ based on different
methods. In general, GARCH method can capture the dynamic evolution of systemic risk contributions explicitly. However, it does
not consider the impact of exogenous variables. As an alternative, quantile regression method considers the influence of risk factors

11
Q. Xu et al. North American Journal of Economics and Finance 55 (2021) 101354

Fig. 4. Time-varying 𝑊 -𝑄𝑅-𝐶𝑜𝑉 𝑎𝑅𝑠|𝑖 and 𝑄𝑅-𝐶𝑜𝑉 𝑎𝑅𝑠|𝑖 of 14 banks with 𝑞 = 0.1. In each subfigure the gray points is the returns of the banking system, the
blue line stands for 𝑊 -𝑄𝑅-𝐶𝑜𝑉 𝑎𝑅𝑠|𝑖 of each bank and the red line stands for 𝑄𝑅-𝐶𝑜𝑉 𝑎𝑅𝑠|𝑖 of each bank. (For interpretation of the references to color in this
figure legend, the reader is referred to the web version of this article.)

Table 8
Panel regression of 𝛥𝑊 -𝑄𝑅-𝐶𝑜𝑉 𝑎𝑅𝑠|𝑖 on lagged values of firm characteristics.
Panel A:5%𝛥𝑊 -𝑄𝑅-𝐶𝑜𝑉 𝑎𝑅𝑠|𝑖 Panel B:10%𝛥𝑊 -𝑄𝑅-𝐶𝑜𝑉 𝑎𝑅𝑠|𝑖
1 Quarter 1 Year 2 Years 1 Quarter 1 Year 2 Years
VaR 0.316*** 0.063* 0.025 0.336*** 0.065* 0.016
(0.029) (0.036) (0.038) (0.029) (0.037) (0.041)
Market/book 0.022 −0.040 0.172* −0.035 −0.083 0.064
(0.072) (0.084) (0.087) (0.055) (0.066) (0.069)
PE ratio −0.004 0.028* 0.003 0.009 0.027** 0.008
(0.013) (0.015) (0.016) (0.010) (0.012) 0.012)
Size −0.036*** −0.087*** −0.118*** −0.040*** −0.044*** −0.083***
(0.013) (0.015) (0.016) (0.009) (0.012) (0.012)
ROA 0.607*** 0.557*** 0.232 0.227** 0.170 −0.080
(0.147) (0.175) (0.189) (0.114) (0.138) (0.151)
Leverage 0.036*** 0.039*** 0.036*** 0.021*** 0.027*** 0.015**
(0.007) (0.009) (0.009) (0.006) (0.007) (0.007)
Beta −0.093 0.104 −0.149 −0.100* 0.063 −0.174**
(0.073) (0.093) (0.100) (0.058) (0.076) (0.083)
Adjusted R2 55.04% 39.92% 39.10% 48.44% 29.00% 26.60%

Notes:
***Denote significance at the level of 1%.
**Denote significance at the level of 5%.
*Denote significance at the level of 10%.

and VaR on systemic risk, and can be employed in many cases without the dependence on data distribution. On this basis, our
proposed W-QR-CoVaR method can exploit potential information in financial time series and typical natures at different scales,
which can produce an accurate systemic risk measure.
To investigate the efficacy of the proposed W-QR-CoVaR method, we conduct an empirical application on measuring the systemic
risk of the Chinese banking industry. The empirical results show that the W-QR-CoVaR method performs better than the QR-CoVaR
method in terms of AE and FT. Based on the accurate systemic risk prediction, regulators can find the right amount of reserves

12
Q. Xu et al. North American Journal of Economics and Finance 55 (2021) 101354

and formulate corresponding regulatory rules. In addition, we investigate the impact of individual financial institutions on the
systemic risk by exploring the relation between their systemic risk contribution and characteristics. We employ the panel regression
to identify the determinants of systemic risk contribution across three different time horizons. The regression results show that size
and leverage stand out and have significant explanatory power. Specifically, banks with higher leverage and smaller size appears to
have more contributions to the systemic risk, which should be paid more attention by regulators. Our findings can help regulators
evaluate the systemic risk and formulate macro-prudential regulatory policies.
The future research can be extended in the following directions. First, different mother wavelets can be introduced into our
method. In this study, we select the LA(8) mother wavelet, which is suitable for processing this particular financial data set. In future
study, we can also try some other mother wavelets from different wavelet families and compare their performance in systemic risk
measurement for different types of data. Second, unidirectional risk spillovers can be extended to multi-directional risk spillovers of
institutions and the financial system. In this study, we primarily consider the risk of the system conditional on individual institutions
being in financial distress. In the future, we can also study the systemic risk of an individual institution conditional on the system
or other institutions being in financial distress. This extension will help to understand the true risks faced by individual financial
institutions and help them take strategies against systemic risk.

CRediT authorship contribution statement

Qifa Xu: Conceptualization, Software, Formal analysis, Resources, Data curation, Writing - original draft, Writing - review
& editing, Supervision, Project administration, Funding acquisition. Bei Jin: Methodology, Software, Investigation, Resources,
Data curation, Writing - original draft, Writing - review & editing, Visualization. Cuixia Jiang: Conceptualization, Methodology,
Validation, Investigation, Resources, Writing - review & editing, Supervision, Funding acquisition.

Acknowledgments

The authors are grateful to the Editor-in-Chief and two anonymous referees for their helpful comments and constructive guidance.
The authors also gratefully acknowledge financial support from the National Natural Science Foundation of China (71671056,
91846201), the National Statistical Science Research Projects of China (2019LD05), and the Key Research and Development Plan
Projects in Anhui Province, China (202004a05020020).

References

Acharya, V. V., Pedersen, L. H., Philippon, T., & Richardson, M. (2017). Measuring systemic risk. Review of Financial Studies, 30(1), 2–47.
Adrian, T., & Brunnermeier, M. K. (2016). CoVAr. The American Economic Review, 106(7), 1705–1741.
Bernal, O., Gnabo, J.-Y., & Guilmin, G. (2014). Assessing the contribution of banks, insurance and other financial services to systemic risk. Journal of Banking
& Finance, 47, 270–287.
Bonaccolto, G., Caporin, M., & Paterlini, S. (2019). Decomposing and backtesting a flexible specification for CoVar. Journal of Banking & Finance, 108, Article
105659.
Brownlees, C., & Engle, R. F. (2017). SRISK: A conditional capital shortfall measure of systemic risk. Review of Financial Studies, 30(1), 48–79.
Christoffersen, P. F. (1998). Evaluating interval forecasts. International Economic Review, 39(4), 841–862.
Gallegati, M. (2008). Wavelet analysis of stock returns and aggregate economic activity. Computational Statistics & Data Analysis, 52(6), 3061–3074.
Girardi, G., & Ergün, A. T. (2013). Systemic risk measurement: Multivariate GARCH estimation of CoVar. Journal of Banking & Finance, 37(8), 3169–3180.
He, K., Lai, K. K., & Yen, J. (2012). Ensemble forecasting of value at risk via multi resolution analysis based methodology in metals markets. Expert Systems with
Applications, 39(4), 4258–4267.
Huang, S.-C. (2011). Wavelet-based multi-resolution GARCH model for financial spillover effects. Mathematics and Computers in Simulation, 81(11), 2529–2539.
Jammazi, R., & Reboredo, J. C. (2016). Dependence and risk management in oil and stock markets. a wavelet-copula analysis. Energy, 107, 866–888.
Koenker, R. W., & Bassett, G. (1978). Regression quantile. Econometrica, 46(1), 33–50.
Kupiec, P. (1995). Techniques for verifying the accuracy of risk measurement models. The Journal of Derivatives, 3(2), 73–84.
López-Espinosa, G., Moreno, A., Rubia, A., & Valderrama, L. (2012). Short-term wholesale funding and systemic risk: A global CoVar approach. Journal of Banking
& Finance, 36(12), 3150–3162.
Mallat, S. G. (1989). A theory for multiresolution signal decomposition: The wavelet representation. IEEE Transactions on Pattern Analysis and Machine Intelligence,
11(7), 674–693.
de Mendoncça, H. F., & da Silva, R. B. (2018). Effect of banking and macroeconomic variables on systemic risk: An application of ▵COVAR for an emerging
economy. The North American Journal of Economics and Finance, 43, 141–157.
Meng, X., & Huang, C.-H. (2019). The time-frequency co-movement of Asian effective exchange rates: A wavelet approach with daily data. The North American
Journal of Economics and Finance, 48, 131–148.
Reboredo, J. C., & Ugolini, A. (2015). Systemic risk in European sovereign debt markets: A CoVar-copula approach. Journal of International Money and Finance,
51, 214–244.
Silva, W., Kimura, H., & Sobreiro, V. A. (2017). An analysis of the literature on systemic financial risk: A survey. Journal of Financial Stability, 28, 91–114.
Teply, P., & Kvapilikova, I. (2017). Measuring systemic risk of the US banking sector in time-frequency domain. The North American Journal of Economics and
Finance, 42, 461–472.
Tibshirani, R. (1996). Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society. Series B., 58(1), 267–288.
Wang, G.-J., Jiang, Z.-Q., Lin, M., Xie, C., & Stanley, H. E. (2018). Interconnectedness and systemic risk of China’s financial institutions. Emerging Markets Review,
35, 1–18.
Xu, Q., Chen, L., Jiang, C., & Yuan, J. (2018). Measuring systemic risk of the banking industry in China: A DCC-MIDAS-t approach. Pacific-Basin Finance Journal,
51, 13–31.
Xu, Q., Li, M., Jiang, C., & He, Y. (2019). Interconnectedness and systemic risk network of chinese financial institutions: A LASSO-CoVar approach. Physica A.
Statistical Mechanics and its Applications, 534, Article 122173.
Zeb, S., & Rashid, A. (2019). Systemic risk in financial institutions of BRICS: Measurement and identification of firm-specific determinants. Risk Management,
21(4), 243–264.
Zhou, H., Liu, W., & Wang, L. (2020). Systemic risk of China’s financial system (2007-2018): A comparison between ▵CoVar, MES and SRISK across banks,
insurance and securities firms. The Chinese Economy, 53(3), 221–245.
Zhu, L., Wang, Y., & Fan, Q. (2014). MODWT-ARMA model for time series prediction. Applied Mathematical Modelling, 38(5–6), 1859–1865.

13

You might also like