Economic Modelling xxx (xxxx) xxx–xxx

Contents lists available at ScienceDirect

Economic Modelling
journal homepage: www.elsevier.com/locate/econmod

Expected default based score for identifying systemically important banks

⁎
Yanzhen Yaoa,b, Jianping Lia,b, Xiaoqian Zhua, , Lu Weia,b
a
Institute of Policy and Management, Chinese Academy of Sciences, Beijing 100190, China
b
University of Chinese Academy of Sciences, Beijing 100049, China

A R T I C L E I N F O A BS T RAC T

JEL classification: The issue of identifying systemically important banks has gained prominence since the recent global financial
G21 crisis in 2007. However, the extant methods either neglect the adverse impact on the financial system posed by a
G28 bank or ignore the various interactions among banks. To resolve this issue, the objective of this study is to put
Keywords: forward an expected default based score (EDBS) that overcomes the drawbacks of the existing methods from the
Systemically important banks perspective of contagion risk. This indicator measures the systemic importance of a bank by calculating the
Contagion risk expected bank defaults triggered by its initial failure. In the empirical study, the expected default based score is
Systemic risk applied to identify the systemically important banks in the Chinese banking system. Both the quantitative
Bank risk
comparison with other major methods and the qualitative evaluation of the Delphi method validate the
reliability of the EDBS method. The empirical results also demonstrate that interconnectedness among banks is
an important and complementary driver of systemic importance in addition to asset size.

1. Introduction
The default of a single bank can spark widespread contagion risk in
the entire banking industry and even harm the smooth functioning of
the real economy (Molyneux et al., 2014; Pourkhanali et al., 2016;
Zhou, 2010; Fiala and Havranek, 2017). Two illustrative examples are
Continental Illinois Bank and Bear Stearns. When the Continental
Illinois Bank, the seventh largest bank in U.S. history, failed in 1984,
nearly 2300 other banks held deposits at or loaned funds to the
Continental (Kaufman, 2000). The Federal Deposit Insurance
Corporation's (FDIC's) bailout of the Continental was justified on the
grounds that its collapse would have posed a severe threat to the U.S.
banking system. The government assistance for J. P. Morgan's acquisi-
tion of Bear Stearns, which nearly failed during the subprime crisis that
started in 2007, is supported by the fact of Bear Stearns’ active
participation in the credit risk transfer market and high systemic
importance (Chan-Lau, 2010). Therefore, particularly motivated by the
severe aftermath of the global financial crisis of 2007–2009, super-
visory authorities have been eager to call for effective analytical
methods to identify banks with high systemic importance (Gravelle
and Li, 2013).
In the early years, systemically important banks (hereafter SIBs) were
simply deemed as the “too-big-to-fail” banks and were thought should
face more stringent regulation (O’Hara and Shaw, 1990). However, this
“size only” definition does not fit the increasingly complicated banking
system (Adrian and Brunnermeier, 2016; Thomson, 2010). Recently, the
viewpoint of the International Monetary Fund et al. (2009) has become
widely accepted. It defines the bank to be systemically important if its
failure would cause propagation of contagion risk through the rest of the
financial system and even to the real economy. Based on this definition,
two typical features, the risk propagation in and the adverse impact on the
banking system, should be captured when identifying SIBs. Various
interactions among banks, such as interbank lending relationships, have
become a major risk propagation channel of systemic risk. Given that a
systemically important bank fails initially, it might pose a severe adverse
impact on the entire system. However, most of the existing methods for
identifying SIBs ignore one of the two features.
The objective of this paper is to propose an expected default based
score (EDBS) from the perspective of contagion risk to identify
systemically important banks. It signifies the expected number of total
bank defaults caused by the initial failure of a particular bank. Because
the EDBS is based on contagion risk, it can not only provide an
intuitive interpretation of SIBs but also well overcome the demerits of
most existing methods. Specifically, the EDBS captures contagious
defaults among banks through a contagion mechanism. In this paper, a
sequential default algorithm proposed by Furfine (2003) is employed as
an alternative technique for the realization of the EDBS. In addition, a
critical revision is made in this algorithm for incorporating more and
wider factors and circumstances.
The remainder of this paper is organized as follows. Section 2
reviews the relevant literature. Section 3 introduces the proposed
expected default based score. The empirical analysis is presented in

⁎
Corresponding author.
E-mail addresses: yaoyanzhen14@mails.ucas.ac.cn (Y. Yao), ljp@casipm.ac.cn (J. Li), zhuxq@casipm.ac.cn (X. Zhu), weilu2014ucas@163.com (L. Wei).

http://dx.doi.org/10.1016/j.econmod.2017.04.023
Received 14 October 2016; Received in revised form 6 March 2017; Accepted 27 April 2017
0264-9993/ © 2017 Elsevier B.V. All rights reserved.

Please cite this article as: Yao, Y., Economic Modelling (2017), http://dx.doi.org/10.1016/j.econmod.2017.04.023
Y. Yao et al.
Section 4, where the proposed score is applied to identify the SIBs in
China. Section 5 presents the conclusions.
2. Literature review
Several measures of systemic importance have been put forward in
recent academic studies, which can be generally categorized into two
types of methods (Lu and Hu, 2014). One type of method is called
indicator-based measurement, which incorporates bank-level data, e.g.,
balance sheet data and the volume of transactions. The other type of
method is called market-based measurement, which uses related
market volatility data such as stock returns of different banks, to
measure the contribution of SIBs to the system risk. The main
difference between the two types of methods is their different perspec-
tives in understanding the meaning of systemic importance and the
data involved.
The widely recognized indicator-based measurement was proposed
by the Basel Committee on Banking Supervision (BCBS for short, 2011)
for identifying Global Systemically Important Banks (G-SIBs). The
BCBS (2013) further updated the indicators in this method. The
systemic importance of a bank is expressed as a final score by summing
the sub-scores in the five selected categories, size, interconnectedness,
substitutability, complexity, and cross-jurisdictional activity, with
equal weights. Afterward, Brämer and Gischer (2013) expeditiously
extended it to identify domestic systemically important banks (D-SIBs).
The extension mainly focused on the last category of indicators, “cross-
jurisdictional activity”, which is not applicable to domestic systemically
important banks (D-SIBs) because it was created to express the global
reach of a bank. Chen et al. (2014) further revised this method to be
applicable to the Chinese banking system. The indicator-based mea-
surement is believed to be capable of select SIBs quickly, transparently
and dynamically (Lu and Hu, 2014) due to its comprehensive indicator
system. However, it has several demerits. Firstly, it needs an extensive
collection of bank data (Brämer and Gischer, 2013). Secondly, it does
not capture the adverse impact on the financial system posed by a
distressed or troubled bank dynamically.
Market-based measurement primarily includes Marginal Expected
Shortfall (MES), CoVaR and Shapley Value. Acharya (2009) and
Acharya et al. (2010) use MES to measure systemic risk and extend
it to identify systemically important financial institutions. CoVaR,
conceptualized in Adrian and Brunnermeier (2016), is designed for
bilateral risk spillover (Bianconi et al., 2015; Liu, 2016). It quantifies a
bank's systemic importance as its marginal contribution to the overall
systemic risk when treating all other institutions as a whole. However,
it is difficult to be generalized to measure a group of banks’ contribu-
tion to systemic risk (Ferrari, 2010). Shapley Value, suggested in
Tarashev et al. (2009) and Drehmann and Tarashev (2011), is an
important instrument used in game theory and for calculating the
degree of systemic importance of each bank based on average
contribution to the risk of all groupings of institutions (Tarashev
et al., 2016). Intuitively it provides interpretation for systemic im-
portance, but a large computational effort is required when empirically
applied to a large financial system.
Generally, market-based measurement is more forward-looking
than indicator-based measurement on account of high-frequency
market data. However, it neglects the increasing importance of various
interactions, such as contagious defaults, in systemic risk (Kanno,
2015), thus failing to adequately capture the relationship between
interconnectedness and systemic importance in a financial system.
Moreover, some crucial data in market-based measurement are
difficult to obtain sometimes (Lu and Hu, 2014). For instance, when
applying the CoVaR method to China, one of the systematic state
variables, the Volatility Index (VIX), which captures the implied
volatility in the stock market, is not officially released and thus cannot
be obtained directly.
The literature review above shows that both types of methods have
2
Economic Modelling xxx (xxxx) xxx–xxx
drawbacks, either neglecting the adverse impact on the financial system
posed by a distressed or troubled bank or ignoring the various
interactions among banks. By definition, with the expected number
of extra bank defaults caused by an initially failed bank, our new
method can capture its adverse impact on the financial system
dynamically. It is noted that ‘dynamically’ here means our new method
is based on the spreading of contagion risk. This method can also
capture various interactions through an interbank network comprising
bilateral exposures.
3. Methodology
In this section, the proposed method, the expected default based
score (EDBS), is introduced in detail. Firstly, the basic definition of the
EDBS is described. Then, the sequential default algorithm proposed by
Furfine (2003) is used as an alternative technique for its realization.
Given that a particular bank fails or defaults, it is natural to evaluate
its adverse impact on the banking system by considering the expected
number of total extra bank failures in the system from the perspective
of contagion risk. This is defined as our method for measuring banks’
systemic importance, namely, the expected default based score (EDBS).
The value of the score represents the systemic importance of a bank.
Clearly, EDBS captures the adverse impact of a distressed bank.
Consider a banking system containing n banks; their values of
EDBS are represented by a vector (EDBS1, …, EDBSn ), where EDBSi
denotes the expected number of total bank defaults in the banking
system given that bank i fails initially due to an idiosyncratic shock.
With a specific value of EDBSi , its relative systemic importance in the
system emerges. The larger the value is, the more systemically
important it is in the banking system. Measuring relative systemic
importance is the key to identifying systemically important banks.
Because our new method considers contagion risk throughout, it
calls for a contagion mechanism to simulate the spread of default risk.
Two often-cited contagion algorithms based on bilateral exposures
include fictitious default algorithm put forward by Eisenberg and Noe
(2001) and sequential default algorithm proposed by Furfine (2003).
Between the two algorithms, the sequential default algorithm is the
most frequently used mainly for two reasons (Upper, 2011). One is that
the interbank market, comprising bilateral exposures among banks,
plays an essential role in a well-functioning integrated financial system
(de Souza et al., 2016) and acts as an important transmission channel
for contagion spreading during crises (Grilli et al., 2014; Kuzubaş et al.,
2014; Souza et al., 2015; Tabak et al., 2014; Toivanen, 2013). The other
is that it is easy to understand and use.
Although the sequential default algorithm has been extensively
used in contagion studies across many countries, such as Germany
(Upper and Worms, 2004), the United Kingdom (Wells, 2004), Holland
(Van Lelyveld and Liedorp, 2006) and Italy (Mistrulli, 2011), it has an
inherently noteworthy limitation. As illustrated in Furfine (2003), the
algorithm confines its scope to the credit contagion channel of
interbank market and thus ignores many other possible contagion
channels. Thus, its assumption is somewhat too simple and cannot
address a variety of practical situations. In the following text, the
limitation of the algorithm is explained mainly from two aspects.
Firstly, contagion can spread through a multitude of channels, other
than interbank credit exposure (Upper, 2011). For example, banks
might have correlated exposures, and an adverse economic shock may
result directly in simultaneous multiple bank defaults (Elsinger et al.,
2006). Other contagion channels also include liquidity risk from
information effects (Degryse and Nguyen, 2007). Therefore, only
considering contagion due to interbank credit exposure is a somewhat
limited perspective. Other channels should also be considered in the
original algorithm.
Secondly, single interbank linkages might not trigger any contagion
in some cases (Georg, 2013; Glasserman and Young, 2015). A
necessary condition for contagion to occur is that the volume of a
Y. Yao et al. Economic Modelling xxx (xxxx) xxx–xxx

bank's interbank loans should exceed its capital. In some countries
such as Germany, this is generally the case. Of the 3246 banks that
existed in Germany at the end of 1998, 2758 (85%) had domestic
interbank loans in excess of their capital. The ratio of domestic
interbank loans to the capital of some banks is even well above 10
(Upper and Worms, 2004). By contrast, in some countries such as
China, interbank activities account for a very small part of banking
business due to strict supervision. Even the largest two banks, the
Industrial and Commercial Bank of China (ICBC) and China
Construction Bank (CCB), have interbank loans that are not larger
than equity capital. As a consequence, interbank lending linkages
might not bring a systemic crisis.
In general, the original sequential default algorithm is not suitable
for a variety of practical situations, such as in countries with restricted
interbank markets. A revision is made to incorporate losses resulting
from non-interbank factors that include all other contagion sources in
addition to the interbank market. Next, the revised sequential default
algorithm is described and presented in detail.
The original sequential default algorithm confines the contagion to
the interbank market and is used to estimate the potential contagion
process by regarding the contagion as a sequential default process.
Initially, one bank fails due to an idiosyncratic shock. Then, this
troubled bank may default on its interbank liabilities and hence cause
other banks to default. The only and key parameter in this algorithm is
the interbank loss rate, referring to the percentage of interbank assets
that cannot be recovered because of debtor bank's bankruptcy. It is
noted that the parameter is usually assumed to be constant across all
banks and rounds in most studies; see van Lelyveld and Liedorp
(2006); Wells (2004).
Non-interbank loss is supposed to be an integration of all types of
losses except interbank loss. We introduce a non-interbank loss rate
and assume that it directly absorbs the capital of banks. Contagion does
not occur until the total loss of a bank resulting from the initial failure,
including an exposure loss and an equity capital loss, exceeds its equity
capital. The revised sequential default algorithm based on the banking
system constructed above is presented in Algorithm 1. As Algorithm 1
shows, the major difference between the original Furfine's algorithm
and the revised one is the addition of parameter β , which reflects the
non-interbank loss directly consuming the capital of banks. If β is set to
0, the revised algorithm degrades into Furfine's algorithm. Compared
with Furfine's algorithm, the revised algorithm is able to simulate a
wider scenario because the addition of β allows for more flexibility.
Both parameters α and β are constrained in the interval [0, 1].
Additionally, they are assumed to be constant across all banks and
rounds.
Algorithm 1. The revised sequential default algorithm.
Round 1: Bank i fails by assumption.
Round 2: Bank j suffers an interbank loss of αxji and a non-
interbank loss of βej .
If αxji + βej > ej , bank j fails.
Round 3: Bank suffers an interbank loss of α (xki + xkj ) and non-
values to two loss rates. Each pair of parameters constructs a
simulation scenario. Thus, to be specific, EDBSi can be written as Eq.
(1).
EDBSi: = E (kir −1), i = 1, … n ; r = 1, …, m (1)
where kir is defined as the total bank defaults given the initial failure of
bank i under the simulation scenario r , which is defined as different
(α , β ) pairs; m denotes the number of simulation scenarios. Here, one is
subtracted from kir to exclude the initial failure of a particular bank in
the calculation of its expected default based score. More specifically, if
the probability distribution of two loss rates α and β is determined,
EDBSi can be expressed as Eq. (2).
m
EDBSi = ∑r (kir −1) × p (r )
m
= ∑r (kir −1) × p (αr , βr ), i = 1, … n ; r = 1, …, m; α , β ∈ (0, 1)
(2)
where p (r ) signifies the probability of simulation scenario r , expressed
as the joint probability of interbank loss rate and non-interbank loss
rate, p (αr , βr ).
Based on the above introduction, the proposed EDBS provides a
measure of systemic importance and thus can serve as an analytical
tool to identify SIBs if it is combined with an effective category
technique or threshold value. Compared to existing methods, the
expected default based score provides a more informative measure of
systemic importance. In contrast to indicator-based measurement, it
can capture the adverse impact posed by a distressed or troubled bank
on the financial system dynamically; in contrast to market-based
measurement, it incorporates the contagious defaults by considering
contagion risk and thus captures various interaction among banks to a
great degree. Table 1 intuitively shows the comparison results of the
proposed EDBS and the two other major methods.
4. Empirical study
This section applies the proposed method to the Chinese banking
system. Firstly, the data used is described and pre-processed. Then, the
systemically important banks of China are identified and analysed by
applying the EDBS method. Moreover, a result comparison among
indicator-based measurement, CoVaR and EDBS is presented.
4.1. Data description and pre-processing
According to the annual report of the China Banking Regulatory
Commission (CBRC) in 2014, the Chinese banking system consists of
4091 banking institutions, among which only 16 are listed banks.
However, the 16 listed banks are the largest banks in China, with their
total assets amounting to nearly 61.37% of the total assets of the entire
banking system in 2014. Additionally, non-listed banks do not release
their financial reports periodically. Therefore, we choose the 16
Chinese listed banks as our experiment sample in this paper. Table 2
Table 1
Comparisons of the proposed EDBS and the two other major methods.

interbank loss of βek . Method Indicator Typical features

If α (xki + xkj ) + βek > ek , bank k fails.
Round 4: The contagion process will continue until no further Adverse impact Interactions
among banks
bank defaults.
———————————————————————— Expected default based EDBSi Considered Considered
α : interbank loss rate; score
β : non-interbank loss rate Indicator-based The sum of five Not considered Considered
measurement sub-scores
ei : equity capital of bank i ;
Market-based ΔCoVaRti (q ) Considered Not considered
xij : credit exposure of bank i to bank j . measurement
(CoVaR as an
Based on the revised algorithm, given the initial failure of a example)
particular bank, the contagion risk is simulated by assigning different

3
Y. Yao et al.
Table 2
Sixteen listed commercial banks of China and their corresponding abbreviations.
No. Bank No. Bank
1 Industrial and Commercial Bank of 9 Huaxia Bank (HXB)
China (ICBC)
2 China Construction Bank (CCB) 10 Industrial Bank (IB)
3 Bank of China (BOC) 11 China Everbright Bank
(CEB)
4 Agriculture Bank of China (ABC) 12 China Minsheng Bank
(CMB)
5 Bank of Communications (BOCOM) 13 China CITIC Bank
(CITIC)
6 China Merchants Bank (CMB) 14 Bank of Beijing (BOBJ)
7 Pingan Bank (PAB) 15 Bank of Ningbo (BONB)
8 SPD Bank (SPDB) 16 Bank of Nanjing (BONJ)
provides the 16 listed banks and their corresponding abbreviations. For
convenience, we use the abbreviation for each bank in the rest of the
paper.
The China Banking Regulatory Commission usually classifies these 16
banks into three categories, i.e., state-owned commercial banks (ICBC,
CCB, BOC, ABC, and BOCOM), joint-stock commercial banks (CITIC,
CMBC, SPDB. HXB, IB, CEB, PAB and CMB) and city commercial banks
(BOBJ, BONB and BONJ). The five state-owned banks are the largest, the
eight joint-stock commercial banks are medium-size, and the three city
commercial banks are the smallest in total assets.
The EDBS is used to identify systemically important banks in
China. In addition, two other major methods, i.e., indicator-based
measurement and CoVaR, are also applied to make a quantitative
comparison. Thus, data for the three methods need to be collected.
Data involved in the CoVaR include the growth rate of market-valued
assets of each bank and systematic static variables. Specifically, in our
experiment, five systematic state variables are chosen, including 1) the
weekly stock market returns; 2) the weekly implied stock market
volatility; 3) the weekly “liquidity spread”, defined as the difference
between the three-month interbank lending rate and the three-month
bill rate; 4) the weekly yield spread between the ten-year Treasury rate
and the three-month bill rate; 5) the weekly credit spread between AA-
rated bonds and the Treasury rate with the same maturity of ten years.
The data for the growth rate of market-value assets and five static
variables are collected from Wind, the leading providing service for
financial data in China. For the indicator-based measurement, we
employ the method of Chen et al. (2014) in this empirical study. For
specific indicators, please refer to Chen et al. (2014). All of the data can
be collected from the annual reports of the 16 listed commercial banks.
For EDBS, the data are needed for estimating the bilateral exposure
matrix in the revised sequential default algorithm described in Section
3, namely, interbank assets and liabilities. Because the financial reports
of Chinese banks provide only the sub-items of interbank assets and
liabilities, we need to sum the sub-items to obtain the aggregate
interbank assets and liabilities of these banks. Specifically, the inter-
bank assets include due from banks, call loans to banks and buying
back the sale of assets while the interbank liabilities include loans from
other banks, deposits from interbank and assets sold for repurchase.
For the EDBS method and indicator-based measurement, the sample
period is from December 31, 2007 to December 31, 2014 because in
2007 a new accounting standard for these commercial banks was
enacted in China; for CoVaR, the sample period is from August 18,
2010 to December 31, 2014 because the CEB was the last to be listed,
on August 18, 2010.
Numerous researchers have used the maximum entropy method to
map the bilateral exposure matrix (Mistrulli, 2011; Upper and Worms,
2004). This method is usually perceived as a sensible and unbiased
approach because it practically provides a distribution of interbank
loans and deposits, which is as even as possible given the aggregate
interbank loans and deposits. Please see Appendix B for more details
4
Economic Modelling xxx (xxxx) xxx–xxx
about how the maximum entropy method proceeds to estimate the
bilateral exposures matrix.
The maximum entropy method needs to be conducted based on a
complete Chinese banking system in which the total interbank assets
equal the total interbank liabilities. Thus, we merge all non-listed banks
into a single bank, namely, a hypothetical bank, to estimate the
bilateral exposure matrix. However, it is excluded in the contagion
risk simulation process for the fact that it is a mixture of small banks,
among which the interbank lending relationships are unknown and
complex. For its interbank assets and liabilities, they are derived by
performing a linear regression analysis between a bank's total assets
and aggregate interbank assets.
4.2. Empirical results
The premise of the simulation is the probability distribution over
two loss rate parameters. They are defined exogenously. The interbank
loss rate shares many similarities with the loss given default (LGD), one
of the most important parameters of credit risk in the New Basel
Accord (i.e., Basel II, 2003). Firstly, by definition, both are measured in
the event of default and denoted as the ratio of losses to exposure.
Secondly, both are constrained to the interval [0, 1]. Thirdly, most
importantly, most of the LGD as a percentage of exposure is either
relatively high or relatively low (Schuermann, 2004), which is similarly
found for the interbank loss rate in Li et al. (2013). As much literature
has found that the LGD is suited for the beta distribution as it supports
random variables in [0, 1] and is flexible with two parameters, α and β
are similarly beta distributed, hypothetically. In the following empirical
study, it is assumed that the distribution of α and β are identical.
Based on the revised sequential default algorithm, the potential
contagion process is simulated given that the 16 listed commercial
banks fail initially one by one. As analysed in Section 2, the loss rate is
well suited for a beta distribution. The beta distribution is well known
to be very flexible and can model random variables constrained in the
interval [0, 1]. In particular, when the two parameters in the beta
distribution are 1, the probability density function becomes uniform
(Ferrari, 2010). Thus, firstly, we consider the General Beta Regression
(GBR) proposed by Huang and Oosterlee (2011) for modelling the two
loss rates α and β . Please see details about the GBR distribution in
Appendix C. Then, the results of the uniform distribution are regarded
as a benchmark. Because empirical evidence on the loss given default
modelling in the Chinese banking system is lacking and the relevant
dataset is hard to obtain, we apply Wang et al.’s (2010) work, in which
the parameters of the beta distribution are 0.677 and 0.289, respec-
tively. In Wang et al. (2010), the values for the parameters of beta
distribution are obtained by using the generalized beta regression
model proposed by Huang and Oosterlee (2011). The data of loan
defaults were collected from LossMetrics, built by China Orient Asset
Management Corporation (COAMC), which is the largest database
pertaining to loan defaults among different banks in China. Robustness
checks were also performed to validate the robustness and reliability of
the results.
Using the Monte Carlo Simulation for random sampling, we
respectively obtain 1000 values for parameters α and β , thus con-
structing a million simulation scenarios (1 × 10 6 ). Based on the re-
vised sequential default algorithm, the values of total defaults,
kir , i = 1, …, 16; r = 1, …, 1000000 , are obtained. As shown in Eq.
(2), the joint probability distribution of α and β is used as a weight
for the scenarios. It is assumed that α and β are independently and
identically distributed. Finally, we derive expected default based scores
for 16 listed banks from 2007 to 2014, conditional on the beta
distribution B (0.677, 0.289) and even distribution U (0, 1) respec-
tively, as shown in Tables 3 and 4.
The score results are slightly different between Tables 3 and 4. It
can obviously be found that the ranking results during the sample
period are correspondingly identical. Take results under GBR in 2014
Y. Yao et al. Economic Modelling xxx (xxxx) xxx–xxx

Table 3
Expected default based scores and corresponding ranking order of the 16 Chinese listed banks from 2007 to 2014 (Generalized Beta Regression (GBR) for parameters α and β ).

Bank Years

2007 2008 2009 2010 2011 2012 2013 2014

Score Ranking Score Ranking Score Ranking Score Ranking Score Ranking Score Ranking Score Ranking Score Ranking

ICBC 5.010 1 5.171 2 4.848 2 5.560 2 7.572 2 7.041 2 6.421 2 1.854 2

CCB 3.552 2 4.102 4 4.676 3 4.704 3 7.350 3 5.421 3 3.390 6 1.537 4
BOC 3.814 3 5.246 1 5.531 1 6.671 1 8.125 1 7.551 1 6.133 1 1.957 1
ABC 3.075 4 3.542 5 3.198 4 3.119 5 5.145 4 4.683 4 4.333 4 1.333 6
BOCOM 2.005 5 4.544 3 3.892 5 3.219 4 5.142 5 3.901 6 5.446 3 1.429 5
CMB 1.952 6 1.200 9 2.690 7 1.638 8 1.959 11 2.230 10 2.560 8 0.868 9
PAB 0.521 14 0.581 13 0.863 13 0.752 14 1.666 14 2.003 11 1.896 12 0.567 12
SPDB 0.732 11 2.534 7 1.165 10 1.890 7 4.246 6 2.688 8 2.987 7 1.006 7
HXB 0.652 12 1.256 8 0.953 12 0.975 12 1.685 13 1.671 13 1.568 13 0.295 14
IB 1.576 7 2.718 6 2.709 6 2.581 6 4.269 7 4.330 5 3.529 5 1.625 3
CEB 1.289 9 1.068 11 1.889 8 1.567 9 2.098 9 2.519 9 2.000 11 0.737 11
CMBC 1.443 8 1.124 10 1.041 11 1.253 10 2.093 10 3.000 7 2.202 9 0.962 8
CITIC 0.890 10 0.879 12 1.584 9 1.003 11 3.452 8 1.812 12 2.178 10 0.820 10
BOBJ 0.620 13 0.514 14 0.206 16 0.887 13 1.765 12 1.535 14 1.318 14 0.361 13
BONB 0.359 15 0.198 16 0.228 14 0.423 15 0.552 16 0.315 15 0.251 15 0.211 15
BONJ 0.324 16 0.146 15 0.159 15 0.346 16 0.579 15 0.235 16 0.268 16 0.153 16

as an example; we can observe that the highest score comes from BOC,
which is 1.957. This indicates that if BOC fails initially, it will cause an
average of 1.957 extra failures in the system now that we have
abstracted one from the value of kir . Although this is not a high
systemic impact compared with the size of the system, it shows that
BOC is the most systemically important bank. The lowest score comes
from BONJ, which is only 0.153, indicating that the initial failure of
BONJ would barely cause extra defaults. Thus, its systemic importance
was the lowest among the 16 listed banks in 2014. Therefore, the
relative systemic importance of 16 listed banks can be determined in
light of the scores.
To identify systemically important banks, we attempt to classify the
16 listed commercial banks based on their scores by using clustering
techniques. Widely used clustering techniques include hierarchical
clustering analysis (Gordon, 1987) and k-means analysis (Everitt,
2001). In this study, we choose hierarchical clustering analysis because
it not only does not need to assign the number of clusters in advance
but is also effective when the number of features is larger than 3 (In
this paper, each sample bank has eight years of scores in succession as
its clustering features). The clustering for score results in Tables 3 and
4, in the form of a dendrogram, are presented in Fig. 1. Although the
dendrogram provides a complete view of the classification process, our
object is to capture two categories including systemically import banks
and non-systemically important banks.
As shown in Fig. 1, in spite of a slight difference, the dendrogram for
the results of the GBR distribution and that for the uniform distribution
both indicate that the 16 listed banks are initially assigned to their own
cluster, and then pairs of clusters are repeatedly merged until two clusters,
namely, “SIBs” cluster and “Non-SIBs” cluster, are obtained. The SIBs
include 5 banks while the Non-SIBs include 11 banks. From Fig. 1, we can
see that the five SIBs are precisely the five state-owned banks, which
always rank at the top of the list with high scores during the sample
period, as shown in Tables 3 and 4. The five state-owned banks have the
largest total assets in China, accounting for 70.77% of the 16 listed banks
and are thus labelled as the “Big Five” in China. In November 2011, The
Financial Stability Board (2011) publishes a list of global systemically
important banks (G-SIBs) based on the indicator-based methodology
developed by BCBS (2013) and updates the list every November. In 2015,
except BOCOM, all the other four banks are in the list of G-SIBs (Financial
Stability Board, 2015). BOC even has been included in the list for four
consecutive years. The evidence validates the reasonability of our cluster-
ing results to a degree.

Table 4
Expected default based scores and corresponding ranking order of the 16 Chinese listed banks from 2007 to 2014 (uniform distribution for parameters α and β ).

Bank Years

2007 2008 2009 2010 2011 2012 2013 2014

Score Ranking Score Ranking Score Ranking Score Ranking Score Ranking Score Ranking Score Ranking Score Ranking

ICBC 4.050 1 4.230 2 4.048 2 4.060 2 8.370 2 7.610 2 5.410 2 1.925 2

CCB 3.023 2 3.390 4 3.676 3 3.070 3 6.360 3 4.847 3 3.090 6 1.387 4
BOC 2.814 3 4.630 1 4.378 1 5.670 1 8.580 1 7.951 1 6.550 1 2.003 1
ABC 2.075 4 2.540 5 3.198 4 2.720 5 5.150 4 4.238 4 3.340 4 1.230 6
BOCOM 1.966 5 3.450 3 3.082 5 2.910 4 5.140 5 3.998 5 3.740 3 1.322 5
CMB 1.742 6 1.200 9 1.590 7 1.440 8 1.910 11 2.260 10 2.660 8 0.886 9
PAB 0.491 14 0.630 13 0.613 13 0.590 14 1.330 14 1.882 11 1.720 12 0.471 12
SPDB 0.632 11 1.740 7 1.275 10 1.890 7 3.670 6 2.866 8 2.900 7 0.910 7
HXB 0.562 12 1.310 8 0.753 12 0.750 12 1.490 13 1.507 13 1.290 13 0.424 14
IB 1.176 7 1.820 6 1.699 6 2.160 6 3.530 7 4.530 6 3.260 5 1.454 3
CEB 1.109 9 0.970 11 1.499 8 1.180 9 2.080 9 2.652 9 1.860 11 0.637 11
CMBC 1.234 8 1.150 10 0.941 11 1.170 10 2.040 10 4.000 7 2.200 9 0.910 8
CITIC 0.990 10 0.830 12 1.694 9 0.820 11 3.420 8 1.732 12 2.080 10 0.776 10
BOBJ 0.494 13 0.410 14 0.254 16 0.610 13 1.580 12 1.305 14 1.260 14 0.452 13
BONB 0.157 15 0.160 16 0.281 14 0.510 15 0.370 16 0.571 15 0.530 15 0.210 15
BONJ 0.179 16 0.190 15 0.259 15 0.350 16 0.560 15 0.445 16 0.490 16 0.206 16

5
Y. Yao et al. Economic Modelling xxx (xxxx) xxx–xxx

Fig. 1. Hierarchical clustering of the 16 Chinese listed banks. The left dendrogram is the results of GBR distribution, and the right dendrogram is the results of uniform distribution.
Notes: Clustering distance: SqEuclidean. Clustering method: UPGMA scheme (also known as group average). SIBs: systemically important banks. Non-SIBs: non-systemically
important banks.

In addition to the above SIBs we have identified, some other Table 5

commercial banks are increasing in their ranking orders of the scores Categories of Chinese listed banks.
during the sample period, which indicates that they are becoming
Category Systemically Non-systemically important banks
increasingly more systemically important to the banking system. important banks
Therefore, these banks should also be paid close attention to and Potential Others
classified into potential candidates of SIBs. According to the results in candidates
Tables 3 and 4, in which the ranking orders for the 16 banks are
Bank ICBC, CCB, BOC, IB SPDB, CMBC, CMB,
correspondingly identical, Fig. 2 illustrates the changes of the ranking ABC, BOCOM CITIC, CEB, HXB, BOBJ,
orders of the non-SIBs from 2007 to 2014. PAB, BONJ, BONB
From Fig. 2, it is easy to see that among the non-SIBs, only the
ranking order of IB apparently remains at the top and has a clear upward
trend while the ten other banks either remain at the bottom or fluctuate in 4.3. Results analysis
score ranking during the sample period. More specifically, it can be
observed from Tables 3 and 4 that the ranking of IB increases from 7th to In this subsection, firstly, the ranking results of the proposed EDBS,
3rd, even surpassing CCB, ABC, and BOCOM, which are SIBs, in 2014. indicator-based measurement and CoVaR are compared and analysed.
Therefore, IB is suggested to be included in the potential SIB list. IB is Then, a deeper analysis of the relationships among size, interconnect-
characterized by its active engagement in interbank business and thus edness and systemic importance is conducted.
referred to as the “King of Interbank Business” in China. This may be a
possible reason accounting for its increasing systemic importance. 4.3.1. Ranking results comparison
Based on the above analyses, SIBs have been identified from the 16 As illustrated in Section 4.1, based on Chen et al.’s (2014) work, the
Chinese listed commercial banks because they have maintained a state five categories of indicators and the corresponding multiple indicators
of high systemic importance during the past eight years, as shown in are determined. The systemic importance of a bank is expressed as the
Tables 3 and 4. Table 5 shows that the SIBs include the Big Five state- final score by summing the sub-scores of five categories of indicators
owned banks. These banks should be devoted special attention to by with equal weights. The results during the sample period, including the
supervisory authorities. Non-SIBs include the eleven other banks. value of score and ranking, are shown in Table 6. Fig. 3 presents the
Among non-SIBs, the IB is recognized as a potential candidate of SIB hierarchical clustering based on the score results in Table 6. As to the
because of its increasing systemic importance, which should also merit CoVaR method (Adrian and Brunnermeier, 2016), the difference of
attention from regulatory authorities. CoVaR at different quantiles (usually 1% and 5%) between when the

Fig. 2. Changes of ranking order of non-systemically important banks during the sample period.

6
Y. Yao et al. Economic Modelling xxx (xxxx) xxx–xxx

Table 6
Systemic importance score and corresponding ranking of the Chinese listed banks from 2007 to 2014 by using indicator-based measurement.

Bank Years

2007 2008 2009 2010 2011 2012 2013 2014

Score Ranking Score Ranking Score Ranking Score Ranking Score Ranking Score Ranking Score Ranking Score Ranking

ICBC 0.217 1 0.192 2 0.191 2 0.189 2 0.207 1 0.200 1 0.196 1 0.193 1

CCB 0.161 3 0.158 3 0.164 4 0.148 4 0.147 4 0.146 4 0.160 3 0.157 3
BOC 0.212 2 0.217 1 0.196 1 0.201 1 0.181 2 0.168 2 0.148 4 0.150 4
ABC 0.156 4 0.152 4 0.172 3 0.159 3 0.153 3 0.158 3 0.163 2 0.165 2
BOCOM 0.055 5 0.066 5 0.060 5 0.065 5 0.061 5 0.059 5 0.061 5 0.059 5
CMB 0.041 6 0.041 6 0.043 6 0.042 6 0.038 7 0.041 7 0.043 7 0.047 6
PAB 0.006 14 0.011 14 0.010 14 0.010 14 0.013 13 0.018 12 0.021 12 0.019 12
SPDB 0.021 10 0.031 7 0.025 10 0.032 8 0.034 9 0.032 9 0.032 10 0.032 10
HXB 0.020 11 0.018 12 0.015 12 0.015 12 0.015 12 0.015 13 0.016 13 0.014 14
IB 0.029 7 0.029 8 0.026 9 0.034 7 0.038 8 0.043 6 0.044 6 0.044 7
CEB 0.020 12 0.022 10 0.026 8 0.027 10 0.025 11 0.027 11 0.023 11 0.023 11
CMBC 0.023 9 0.021 11 0.021 11 0.025 11 0.028 10 0.040 8 0.034 9 0.039 8
CITIC 0.025 8 0.026 9 0.033 7 0.029 9 0.038 6 0.031 10 0.034 8 0.034 9
BOBJ 0.010 13 0.012 13 0.010 13 0.013 13 0.013 14 0.012 14 0.014 14 0.014 13
BONB 0.002 16 0.003 16 0.004 16 0.007 15 0.005 15 0.007 15 0.008 15 0.006 15
BONJ 0.003 15 0.003 15 0.004 15 0.005 16 0.005 16 0.004 16 0.004 16 0.005 16
Ratio of Big Four 74.6 71.9 72.4 69.7 68.8 67.0 66.7 66.5

Fig. 3. Hierarchical clustering of the 16 Chinese listed bank based on scores of the indicator-based measurement.

Table 7 As shown in Table 6, the first four state-owned banks’ scores are all
ΔCoVaR on the 1% and 5% quantiles and corresponding ranking of the 16 Chinese listed larger than 0.1, which is distinguished from the other banks.
banks during the sample period by using the CoVaR model.
Furthermore, they take more than 66.5% of the overall systemic
Bank importance of the system. Fig. 3 shows that the four state-owned
ΔCoVaRti (1%) ΔCoVaRti (5%)
banks are classified into the “SIBs” category, which is characterized
Average Ranking Average Ranking with high systemic importance. From Table 7, the results are largely
qualitatively similar for the two quantiles. The Big Five state-owned
ICBC −0.1140 1 −0.0486 1
banks have a higher systemic risk contribution relative to other banks
CCB −0.0652 4 −0.0390 5
BOC −0.0668 3 −0.0412 4
with respect to ranking. Thus, they are quantitatively characterized
ABC −0.0738 2 −0.0472 2 with high systemic importance. This is, to a great degree, similar to the
BOCOM −0.0600 5 −0.0435 3 results in Tables 3 and 4.
CMB −0.0472 9 −0.0359 8 With regard to the dissimilarity, two aspects are noteworthy,
PAB −0.0385 14 −0.0290 13
Firstly, BOCOM is categorized as a non-SIB in indicator-based
SPDB −0.0428 11 −0.0347 9
HXB −0.0412 13 −0.0342 10 measurement, as presented in Fig. 3. While in EDBS and CoVaR, it
IB −0.0519 6 −0.0370 7 is incorporated into the list of SIBs. This is greatly attributed to its
CEB −0.0464 10 −0.0298 12 increasing interconnectedness among other banks, which is however
CMBC −0.0514 8 −0.0379 6
always equally weighted as 20% in indicator-based measurement.
CITIC −0.0524 7 −0.0309 11
BOBJ −0.0325 15 −0.0262 14
Secondly, except the SIBs, the potential SIBs are different. In EDBS
BONB −0.0418 12 −0.0211 15 method, IB is included on the potential candidate list for its growing
BONJ 0.0000 16 0.0000 16 systemic importance in the banking system. However, as shown in
Table 3, the potential SIB is CMB for the fact that its systemic
importance ranks quite near the top, with the sixth five times and
underlying bank is in distress and when it is in the median state is the seventh three times. Similarly, in the CoVaR method, IB needs to be
usually used to capture the risk externality that the underlying poses on scrutinized.
the system, and then extended to measuring systemic importance. The Based on the above analysis, the comparison results are shown in
results, including the average ΔCoVaRti (q ) when q = 1% and q = 5%
Table 8 as follows. The list of systemically important banks is basically
during the sample period and the ranking, are shown in Table 7.

7
Y. Yao et al. Economic Modelling xxx (xxxx) xxx–xxx

Table 8
Chinese SIBs results of the proposed EDBS and two other major methods.
Method SIBs Potential SIBs
Expected default based score ICBC, CCB, BOC, ABC, IB
and BOCOM
Indicator-based measurement ICBC, CCB, BOC, and CMB
ABC
Market-based measurement (CoVaR ICBC, CCB, BOC, ABC, IB
as an example) and BOCOM
the same for the three methods, which validate the reliability of our
new method.
In addition, the EDBS method is compared to the two other types of
methods with respect to data requirements and computational complex-
ity. In data requirements, EDBS needs a fraction of balance sheet data
that are involved in the indicator-based measurement. Balance sheet
data for 16 listed banks are often easier and more available to obtain
than market data involved in CoVaR in the Chinese banking system.
Then, the computational time is used to compare the computational
complexity. Because the ranking results of the three methods are all
obtained in the same statistic software, MATLAB 2012a, the computa-
tional time needed for each method can be easily obtained: 0.0012 sec-
onds are needed for indicator-based measurement, 0.0075 seconds for
EDBS and 0.0150 seconds for CoVaR. From the comparison results in
Table 1 and Table 9, it is concluded that our EDBS method not only is
intuitive in the interpretation of a bank's systemic importance but also
performs better in various aspects.
Furthermore, we have conducted a Delphi process. We carefully
screened and invited eleven experts from People's Bank of China
(PBC), China Banking Regulatory Commission (CBRC), large commer-
cial banks, and well-recognized research institutes to participate in the
questionnaire survey and interview. All the eleven experts are in favour
of the reasonability of our method, among which nine of them agree
with our list of SIBs. The final responses are taken as a consensus on
the reliability of our new method.
easier its failure or impairment is, the greater will be the harm to the
entire banking system (BCBS, 2013). This, to a degree, confirms the
usual “too big to fail” argument.
However, it is observed that the ranking order of systemic importance
of the Big Five is not in full accordance with their respective total assets.
BOC is followed by ICBC and ranks the first with the exception that ICBC
ranks the first in 2007. Thus, the systemic importance of BOC is higher
than that of ICBC. In fact, BOC has been listed as one of the G-SIBs since
2011 while ICBC was not on the list until 2012. In the context that the
total assets of ICBC exceed BOC, it is not completely a bank's absolute size
that impacts its systemic importance. Furthermore, the difference be-
tween the ranking order of total assets and systemic importance is
becoming larger, indicating that the effect of the bank's size on its
systemic importance is weakening over time.
A growing amount of literature has revealed the importance of
interconnectedness in systemic risk contagion (Capponi and Chen,
2015; de Souza et al., 2016; Ladley, 2013). The interconnectedness of a
bank is necessarily linked to the contagion effect it might cause (Martinez-
Jaramillo et al., 2014). It could capture situations when distress in one
bank raises the likelihood of distress in others. Therefore, the relationship
between systemic importance and interconnectedness among banks is
further investigated. Interconnectedness refers to the degree of inter-
dependence that they have with each other and other financial institutions
(International Monetary Fund (IMF) et al., 2009). In this paper, total
interbank liabilities and assets of one bank are chosen as the proxy of its
interconnectedness. By checking the total interbank liabilities and assets
of IB, SPDB, and CMB in their annual reports, we find that although the
size of their assets is of a medium level, their interconnectedness vis-à-vis
other banks is quite significant, thus resulting in a relatively high systemic
importance. Thus, the interconnectedness towards other banks is defi-
nitely an important driver of systemic importance, which has also been
revealed in Drehmann and Tarashev (2013). More specifically, given a
complex network of obligations or exposure to other banks, sudden
financial distress at a bank can materially increase the likelihood of risk
contagion in the entire banking system (BCBS, 2013), and thus, its
systemic importance is large in the system.

4.3.2. Size, interconnectedness and systemic importance 5. Conclusions

To further investigate the relationship between the size of a bank
and its systemic importance, the line chart of the expected default In this paper, we propose an expected default based score to
based score in Table 3 and the total assets (as a proxy for the size of a identify SIBs from the perspective of contagion risk. The score
bank) of 16 listed banks is plotted, as presented in Fig. 4. measures the expected number of total bank defaults triggered by the
The Big Five state-owned banks have the largest total assets in the initial failure of a particular bank. Furfine's sequential default algo-
banking system, which can also be determined from the annual reports. rithm is employed as an alternative technique for the realization of
They possess the highest systemic importance as well. For the city EDBS. In addition, a critical revision is made in this algorithm for
commercial banks, the systemic importance ranking and the total incorporating more and wider factors and circumstances.
assets ranking are also basically consistent, both low. We can therefore Then, we conduct an empirical analysis of the Chinese banking system
conclude that the size of a bank is the primary determinant of systemic for the eight-year period from 2007 to 2014. By analysing the expected
importance. More specifically, if a bank's total assets size is large and default based score of 16 listed banks, the SIBs list including the Big Five
its activities comprise a large share of the entire financial system, its state-owned banks and the potential candidate list including IB are
failure or impairment is more likely to incur a system-wide risk and provided. Both the quantitative result comparison with two other existing
endanger even the real economy. The larger the bank's size is, the more methods and qualitative evaluation of the Delphi process validate the
difficult it is to replace it with other banks once it fails, and thus, the reliability of the proposed method. Further analysis reveals that inter-
connectedness among banks is an important and complementary driver of
Table 9 systemic importance in addition to asset size.
The comparisons of the three methods in computational complexity and data Several questions remain to be studied in future research. For example,
requirements.
the maximum entropy technique estimates the bilateral exposure matrix
Method Data requirements Computational based on the pre-assumption that each sample bank is connected to other
complexity banks with probability one, thus determining a complete banking network.
Therefore, there is need for a method that can generate a network closer to
Expected default based Available; small dataset Relatively low
the actual interbank lending relationships.
score
Indicator-based Available; relative small Low
measurement dataset Acknowledgements
CoVaR Available conditionally; High
large dataset This research has been supported by grants from the National
Natural Science Foundation of China (71425002, 71601178, and

8
Y. Yao et al. Economic Modelling xxx (xxxx) xxx–xxx

Fig. 4. Expected default based score and total assets for the 16 Chinese listed banks from 2007 to 2014. Note: The solid line represents the systemic importance while the dotted line
represents the total assets. The scale on the right of the y-axis changes to illustrate the systemic importance. The total assets are put in a descending order to provide a clear comparison
between the systemic importance and the size of the 16 listed banks.

71403251) and Youth Innovation Promotion Association of Chinese editor and three reviewers for their very valuable and professional
Academy of Sciences (2012137 and 2017200). We sincerely thank the comments.

Appendix A. Estimated interbank bilateral exposures of Chinese listed banks

The interbank bilateral exposures of the 16 Chinese listed banks are estimated by using the maximum entropy method in this paper. Taking the
year 2014 as an example, the whole picture of bilateral exposure matrix of 2014 is provided in Table A1. Each table element signifies the amount of

9
Y. Yao et al. Economic Modelling xxx (xxxx) xxx–xxx

Table A1
The 2014 interbank bilateral exposures of Chinese listed banks estimated by maximum entropy method. (Unit: billion CNY).

ICBC CCB BOC ABC BOCOM CMB PAB SPDB HXB IB CEB CMBC CITIC BOBJ BONB BONJ

ICBC 0.0 57.2 83.7 50.1 54.1 35.1 17.1 36.3 14.9 59.8 24.3 40.4 30.4 16.1 5.2 5.0
CCB 49.8 0.0 51.8 31.0 33.5 21.7 10.6 22.5 9.2 37.0 15.0 25.0 18.8 10.0 3.2 3.1
BOC 72.8 51.8 0.0 45.3 49.0 31.8 15.5 32.9 13.5 54.2 22.0 36.6 27.5 14.6 4.7 4.5
ABC 93.4 66.5 97.2 0.0 62.9 40.8 19.9 42.2 17.3 69.5 28.2 46.9 35.3 18.7 6.0 5.8
BOCOM 33.0 23.5 34.4 20.6 0.0 14.4 7.0 14.9 6.1 24.6 10.0 16.6 12.5 6.6 2.1 2.1
CMB 32.6 23.2 33.9 20.3 21.9 0.0 6.9 14.7 6.0 24.2 9.8 16.3 12.3 6.5 2.1 2.0
PAB 17.8 12.7 18.5 11.1 12.0 7.8 0.0 8.0 3.3 13.3 5.4 8.9 6.7 3.6 1.2 1.1
SPDB 22.4 15.9 23.3 13.9 15.1 9.8 4.8 0.0 4.1 16.6 6.8 11.2 8.5 4.5 1.4 1.4
HXB 12.5 8.9 13.0 7.8 8.4 5.5 2.7 5.6 0.0 9.3 3.8 6.3 4.7 2.5 0.8 0.8
IB 54.7 38.9 56.9 34.0 36.8 23.9 11.6 24.7 10.1 0.0 16.5 27.4 20.7 10.9 3.5 3.4
CEB 28.3 20.1 29.4 17.6 19.0 12.3 6.0 12.8 5.2 21.0 0.0 14.2 10.7 5.7 1.8 1.8
CMBC 57.8 41.1 60.1 36.0 38.9 25.2 12.3 26.1 10.7 43.0 17.4 0.0 21.8 11.6 3.7 3.6
CITIC 18.4 13.1 19.2 11.5 12.4 8.0 3.9 8.3 3.4 13.7 5.6 9.2 0.0 3.7 1.2 1.1
BOBJ 18.8 13.4 19.5 11.7 12.6 8.2 4.0 8.5 3.5 14.0 5.7 9.4 7.1 0.0 1.2 1.2
BONB 3.1 2.2 3.2 1.9 2.1 1.3 0.6 1.4 0.6 2.3 0.9 1.5 1.2 0.6 0.0 0.2
BONJ 2.5 1.8 2.6 1.6 1.7 1.1 0.5 1.1 0.5 1.9 0.8 1.3 1.0 0.5 0.2 0.0

money that the bank in the corresponding row lends to the bank in the corresponding column. It is noted that the elements in the principal diagonal
are zeros because one bank cannot lend to or borrow from itself.

Appendix B. Maximum entropy method for estimating interbank bilateral exposures

The lending relationships in the interbank market can be represented by the following N × N matrix:
⎡ x11 x12 ... x1N ⎤
X1 = ⎢⎢ x...
21 x22
...
... x2N ⎥
... ... ⎥ ,
⎣ xN1 xN 2 ... xNN ⎦

where xij represents the credit exposure of bank j vis-á-vis bank i and N is the number of banks. We cannot observe the bilateral exposures xij
N N
directly, but we do know ∑ j =1 xij = ai , ∑i =1 xij = l j are, respectively, the total amount of money bank i lends to other banks and bank j raises from
other banks. In the absence of any other assumptions about the distribution of bilateral exposures, the matrix X1 cannot be identified as N2 − 2N
unknowns have to be estimated.
The common and widely used approach is to assume that banks maximize the dispersion of their interbank exposures. It amounts to solving the
N N
following mathematical problem: min ∑i =1 ∑ j =1 xij ln xij , which is subject to the following constraints:
N N
∑ xij = ai , ∑ xij = l j , xij ≥ 0.
j =1 i =1

By constructing the corresponding Lagrange function, the bilateral exposures are given by a simple solution: xij* = ai l j , from which we obtain
matrix X1*. However, elements on the main diagonal of X1* are non-zero, which would imply that for banks that are both lender and borrower in the
market, they lend to themselves. To rule out this outcome, we, therefore, need to modify the matrix X1* by setting xij* = 0 for ∀ i = j . We then obtain
the transition matrix X 0 :
⎡ 0 x12 ... x1N ⎤
⎢ ⎥
x 0 ... x2N ⎥
X0 = ⎢ 21 .
⎢ ... ... 0 ... ⎥
⎢⎣ xN1 xN 2 ... 0 ⎥⎦

The problem is then to estimate bilateral exposures such that the matrix X * obtained by the maximum entropy method finally can be as close as
possible to the matrix X 0 . This is generally obtained by minimizing the cross-entropy between the two matrices:
N N ⎛ xˆij ⎞
min ∑ ∑ xˆij ln ⎜⎜ ⎟⎟ ,
i =1 j =1 ⎝ xij* ⎠
N N
s.t. ∑ xˆij = ai , ∑ xˆij = l j , xˆij ≥ 0, if i ≠ j ; xˆij = 0, if i = j .
j =1 i =1

This optimization problem can be solved numerically with the RAS algorithm (Censor and Zenios, 1997), and the process is summarized as the
following iterations:

ai
Step 1: (row adjustment): xiju ← xiju ρiu , where ρiu = ∑∀ j / x 0 >0 xiju
;
ij
lj
Step 2: (column adjustment): xjiu ← xjiu σ ju , where σ ju = ∑∀ i / x 0 >0 xiju
;
ij
Step 3: return to step 1.

10
Y. Yao et al. Economic Modelling xxx (xxxx) xxx–xxx

Appendix C. Generalized beta regression model

Parameters of a beta distribution

Assume that
x ∼ B (φ , γ ),
we have
x φ −1 − (1 − x )γ −1 Γ (φ + γ ) φ −1
f (x ) = = x (1 − x )γ −1,
B (φ , γ ) Γ (φ ) Γ (γ )
where Γ (⋅) denotes the Gamma function.
Next,
φ
μ = E (X ) = ,
φ+γ
φγ μ (1 − μ)
σ 2 = Var (X ) = .
(φ + γ )2 (φ + γ + 1) φ + γ + 1

Let π = φ + γ ; then
φ = μπ , γ = (1 − μ) π .
Therefore, a beta distribution can also be uniquely determined by its mean and dispersion. The beta distribution is well-known to be very
flexible, modelling variables belonging to the interval [0, 1].

Generalized beta regression model

The mean model for μ has following two components:

1) A linear predictor η
η = λς,
where ς denotes a vector of explanatory variables and λ is a vector of the corresponding regression coefficients. As a convention, the first element
of ς is set to 1 so that the first element of λ is an intercept term.
2) A monotonic, differentiable link function g
g (μ ) = η .
The inverse of the link function, g−1 (⋅), should form a mapping from R to [0, 1], which is exactly the range of μ . To achieve this, a variety of link
functions can be used, such as the logit link

eη ⎛ μ ⎞
μ= , η = log ⎜ ⎟,
1 + eη ⎝1 − μ ⎠

or the probit link

μ = Φ (η), η = Φ−1 (μ).
The parameter estimation can be implemented by two alternatives, namely, the method of least squares or the method of maximum likelihood
(Huang and Oosterlee, 2011).

References
Acharya, V.V., 2009. A theory of systemic risk and design of prudential bank regulation.
J. Financ. Stab. 5 (3), 224–255.
Acharya, V.V., Yorulmazer, T., Santos, J.A., 2010. Systemic risk and deposit insurance
premiums. Econ. Policy Rev. 16 (1), 89–99.
Adrian, T., Brunnermeier, M.K., 2016. CoVaR. Am. Econ. Rev. 106 (7), 1705–1741.
Basel Committee on Banking Supervision, 2011. Global Systemically Important Banks:
Assessment Methodology and the Additional Loss Absorbency Requirement. Bank
for International Settlements, Basel, Switzerland.
Basel Committee on Banking Supervision, 2013. Global Systemically Important Banks-
updated Assessment Methodology and the Higher Loss Absorbency Requirement.
Bank for International Settlements, Basel, Switzerland.
Bianconi, M., Hua, X., Tan, C.M., 2015. Determinants of systemic risk and information
dissemination. Int. Rev. Econ. Financ. 38, 352–368.
Brämer, P., Gischer, H., 2013. An assessment methodology for domestic systemically
important banks in Australia. Aust. Econ. Rev. 46 (2), 140–159.
Capponi, A., Chen, P.C., 2015. Systemic risk mitigation in financial networks. J. Econ.
Dyn. Control 58, 152–166.
Censor, Y., Zenios, S.A., 1997. Parallel Optimization. Oxford University Press, Oxford,
England.
Chan-Lau, J.A., 2008. Default risk codependence in the global financial system: was the
Bear Stearns bailout justified? In: Gregoriou, G.N. (Ed.), The Banking Crisis
Handbook 2008. CRC Press, Florida, 437–454.
Chen, Y., Shi, Y., Wei, X., Zhang, L., 2014. Domestic systemically important banks: a
quantitative analysis for the Chinese banking system. Math. Probl. Eng. 2014, 1–19.
Degryse, H., Nguyen, G., Interbank exposure: an empirical examination of contagion risk
in the Belgian banking system. International Journal of Central Banking 3 (2), pp.
123–171.
Drehmann, M., Tarashev, N.A., 2011. Systemic importance: some simple indicators. BIS
Q. Rev., 25–37.
Drehmann, M., Tarashev, N., 2013. Measuring the systemic importance of
interconnected banks. J. Financ. Intermed. 22 (4), 586–607.
Eisenberg, L., Noe, T.H., 2001. Systemic risk in financial systems. Manag. Sci. 47 (2),
236–249.
Elsinger, H., Lehar, A., Summer, M., 2006. Risk assessment for banking systems. Manag.
Sci. 52 (9), 1301–1314.
Everitt, B.S., Landau, S., Leese, M., Stahl, D., 2001. Cluster Analysis. Wiley, Hoboken,
USA.
Ferrari, S., 2010. Measuring the systemic importance of financial institutions using
market information. Financ. Stab. Rev. 8 (1), 127–141.
Fiala, T., Havranek, T., 2017. The source of contagion risk in a banking sector with
foreign ownership. Econ. Model. 60, 108–121.
Financial Stability Board, 2011. Policy measure to address systemically important
financial institutions. 〈http://www.fsb.org/2011/11/r_111104bb/〉.

11
Y. Yao et al.
Financial Stability Board, 2015. 2015 update of global systemically important banks (G-
SIBs). 〈http://www.fsb.org/2015/11/2015-update-of-list-of-global-systemically-
important-banks-g-sibs/〉.
Furfine, C.H., 2003. Interbank exposures: quantifying the risk of contagion. J. Money,
Credit Bank. 35 (1), 111–128.
Georg, C.P., 2013. The effect of the interbank network structure on contagion and
common shocks. J. Bank. Financ. 37 (7), 2216–2228.
Glasserman, P., Young, H.P., 2015. How likely is contagion in financial networks? J.
Bank. Financ. 50, 383–399.
Gordon, A., D., 1987. A review of hierarchical classification. J. R. Stat. Soc. 150 (2),
119–137.
Gravelle, T., Li, F., 2013. Measuring systemic importance of financial institutions: an
extreme value theory approach. J. Bank. Financ. 37 (7), 2196–2209.
Grilli, R., Tedeschi, G., Gallegati, M., 2014. Bank interlinkages and macroeconomic
stability. Int. Rev. Econ. Financ. 34, 72–88.
Huang, X., Oosterlee, C.W., 2011. Generalized beta regression models for random loss
given default. J. Credit Risk 7 (4), 1–27.
International Monetary Fund (IMF), Bank for International Settlements (BIS), Financial
Stability Board (FSB), 2009. Guidance to assess the systemic importance of financial
institutions, markets and instruments: initial considerations, Report to G20 Finance
Ministers and Governors.
Kanno, M., 2015. Assessing systemic risk using interbank exposures in the global
banking system. J. Financ. Stab. 20, 105–130.
Kaufman, G.G., 2010. Banking and currency crisis and systemic risk: a taxonomy and
review. Financ. Mark. Inst. Instrum. 9 (2), 69–131.
Kuzubaş, T.U., Ömercikoğlu, I., Saltoğlu, B., 2014. Network centrality measures and
systemic risk: an application to the Turkish financial crisis. Phys. A: Stat. Mech. Appl.
405, 203–215.
Ladley, D., 2013. Contagion and risk-sharing on the inter-bank market. J. Econ. Dyn.
Control 37 (7), 1384–1400.
Li, J., Liang, C., Zhu, X., Sun, X., Wu, D., 2013. Contagion in Chinese banking industry: a
transfer entropy-based analysis. Entropy 15 (12), 5549–5564.
Liu, X., 2016. Measuring systemic risk with regime switching in tails. Econ. Model..
http://dx.doi.org/10.1016/j.econmod.2016.09.015.
Lu, J., Hu, X., 2014. Novel three-bank model for measuring the systemic importance of
commercial banks. Econ. Model. 43, 238–246.
Martinez-Jaramillo, S., Alexandrova-Kabadjova, B., Bravo-Benitez, B., Solórzano-
Margain, J.P., 2014. An empirical study of the Mexican banking system's network
and its implications for systemic risk. J. Econ. Dyn. Control 40, 242–265.
Mistrulli, P.E., 2011. Assessing financial contagion in the interbank market: maximum
entropy versus observed interbank lending patterns. J. Bank. Financ. 35 (5),
12
Economic Modelling xxx (xxxx) xxx–xxx
1114–1127.
Molyneux, P., Schaeck, K., Zhou, T.M., 2014. ‘Too systemically important to fail’ in
banking—evidence from bank mergers and acquisitions. J. Int. Money Financ. 49,
258–282.
O’Hara, M., Shaw, W., 1990. Deposit insurance and wealth effects: the value of being “too
big to fail”. J. Financ. 45 (5), 1587–1600.
Pourkhanali, A., Kim, J., Tafakori, L., Fard, F.A., 2016. Measuring systemic risk using
vine-copula. Econ. Model. 53, 63–74.
Schuermann, T., 2004. What do we know about loss given default? Working Paper No.
04-01, Wharton Financial Institutions Center, Philadelphia, Pennsylvania, US.
Souza, S.R., Tabak, B.M., Silva, T.C., Guerra, S.M., 2015. Insolvency and contagion in the
Brazilian interbank market. Phys. A: Stat. Mech. Appl. 431, 140–151.
de Souza, S.R.S., Silva, T.C., Tabak, B.M., Guerra, S.M., 2016. Evaluating systemic risk
using bank default probabilities in financial networks. J. Econ. Dyn. Control 66,
54–75.
Tabak, B.M., Takami, M., Rocha, J.M., Cajueiro, D.O., Souza, S.R., 2014. Directed
clustering coefficient as a measure of systemic risk in complex banking networks.
Phys. A: Stat. Mech. Appl. 394, 211–216.
Tarashev, N., Tsatsaronis, K., Borio, C., 2016. Risk attribution using the Shapley Value:
methodology and policy applications. Rev. Financ. 20 (3), 1189–1213.
Tarashev, N.A., Borio, C.E., Tsatsaronis, K., 2009. The systemic importance of financial
institutions. BIS Q. Rev. Sept., 75–87.
Thomson, J.B., 2010. On systemically important financial institutions and progressive
systemic mitigation. DePaul Bus. Commer. Law J. 8 (2), 135–150.
Toivanen, M., 2013. Interbank exposures and risk of contagion in crises: evidence from
Finland in the 1990s and the 2000s. J. Appl. Financ. Bank. 3 (6), 45–66.
Upper, C., 2011. Simulation methods to assess the danger of contagion in interbank
markets. J. Financ. Stab. 7 (3), 111–125.
Upper, C., Worms, A., 2004. Estimating bilateral exposures in the German interbank
market: is there a danger of contagion? Eur. Econ. Rev. 48 (4), 827–849.
van Lelyveld, I., Liedorp, F., 2006. Interbank contagion in the Dutch banking sector: a
sensitivity analysis. Int. J. Cent. Bank. 2 (2), 99–133.
Wang, B., Chen, M., Chen, M., Yang, X., 2010. Distribution estimation of downturn LGD
and loan portfolio stress-test. Proceedings of 2010 International Conference on
Management Science and Engineering, Hubei, China.
Wells, S.J., 2004. Financial interlinkages in the United Kingdom’s interbank market and
the risk of contagion. Working Paper 230. Bank of England, London, United
Kingdom.
Zhou, C., 2010. Are banks too big to fail? Measuring systemic importance of financial
institutions. Int. J. Cent. Bank. 6 (4), 205–250.