macro stress testing on credit risk of...
TRANSCRIPT
Electronic copy available at: http://ssrn.com/abstract=1909327
1
MACRO STRESS TESTING ON CREDIT RISK OF COMMERCIAL BANKS
IN CHINA BASED ON VECTOR AUTOREGRESSION MODELS
Rongjie Tian*
Beijing Institute of Technology
Email: [email protected]
Jiawen Yang
The George Washington University
Email: [email protected]
Abstract
This paper develops a framework for stress-testing the credit risk of Chinese commercial banks to
macroeconomic shocks. Using data over the period 1985-2008, this study establishes a vector
auto-regression (VAR) model to describe the links between default rate and macroeconomic factors, and
then designs three stress scenarios to implement the stress testing by Monte Carlo simulation. As a result,
a credit loss distribution is generated. Our results indicate that the shocks in real property and CPI bring
long term and worst impact on credit risk to commercial banks in China.
JEL: G21,G32,E17
Keywords: Credit risk; Macro stress testing; Commercial bank; VAR model
* This research was conducted while Rongjie Tian stayed at the George Washington University as a visiting scholar
from September 2009 to September 2010.
Electronic copy available at: http://ssrn.com/abstract=1909327
2
MACRO STRESS TESTING ON CREDIT RISK OF COMMERCIAL BANKS
IN CHINA BASED ON VECTOR AUTOREGRESSION MODELS
1. INTRODUCTION
The global financial crisis that started in the United States in 2007 highlights the importance of
macroeconomic (macro) stress testing in the financial sector. Macro stress testing refers to a range of
techniques used to assess the vulnerability of a financial system to “exceptional but plausible”
macroeconomic shocks (see Virolainen, 2006). Most researchers design macro stress testing by modeling
the link between macroeconomic variables and credit risk measures. Sensitivity analysis, scenario
analysis, and extreme values theory are usually employed to implement the stress testing. In recent years,
a lot of effort has been put into stress testing on credit risk of banks all over the world. The New Basel
Capital Accord, Basel Ⅱ, imposes the rigorous stress testing requirements: banks which implement the
Internal Ratings-Based Approach (IRB) must conduct stress tests. The China Banking Regulatory
Commission (CBRC) also requires all Chinese commercial banks to implement stress testing to manage
credit risk. During the current financial crisis, stress testing on credit risk has become a reviving focus of
concern.
Within the framework of credit risk modeling and macro stress testing, we seek to address the
following issues in this paper: What are the most important macroeconomic variables that affect credit
risk for Chinese commercial banks? What is the specific relation between credit risk of banks and the
macroeconomic variables? How does banks’ credit risk react to macroeconomic shocks? We adopt the
default rate as our indicator of credit risk for Chinese commercial banks, and construct a Vector
auto-regression (VAR) model to generate a comprehensive indicator and then use an extended version of
Wilson (1997) model by imposing feedback effects between default rates and macroeconomic variables.
Scenario analysis is also employed as part of our stress testing. We design three “exceptional but
Electronic copy available at: http://ssrn.com/abstract=1909327
3
plausible” scenarios and use Monte Carlo simulation to get the credit loss distribution. Finally, we
compare the credit loss distribution under these stressed scenarios based on the concept of Value-at-Risk
(VaR).
The data employed in this study spans a time period from 1985 to 2008, which covers multiple
episodes of severe macroeconomic shocks such as the Asian financial crisis in 1997, the Russian crisis in
1998, and the current global financial crisis that started in 2007. The macroeconomic variables we have
examined include GDP growth, the Chinese currency exchange rate, nominal interest rate, real property
index, CPI, and unemployment rate. We find that GDP growth, exchange rate, nominal interest rate, real
property index, CPI and unemployment rate have prolonged impacts on the default rate. In order to
describe the feedback effects between variables, we employ vector auto-regression framework to assess
the impact of macroeconomic variables on firms’ probabilities of default.
The rest of the paper is organized as follows. Section 2 provides a brief literature review. Section 3
describes main features of Chinese commercial banks. Section 4 describes our methodology and data. We
lay out three specific methodological frameworks for our macroeconomic credit risk model: VAR model
with raw macroeconomic variables, VAR model with principal components analysis (PCA) and structural
VAR model. Section 5 presents our empirical results for the macroeconomic credit risk model with
different methodological framework. We also carry out our stress tests with impulse, scenario, and
variance decomposition. Section 6 concludes.
2. LITERATURE REVIEW
As Dovern et al. (2010) point out, as a field of academic research, macroeconomic stress testing is
relatively new. There are two approaches to estimating the linkage between credit risk and
macroeconomic factors. The first one is linked to the work of Wilson (1997a and 1997b), who establishes
a direct model based on sensitivity of many macro economic variables for default probability in each
4
industry. Studies by Boss (2002), Virolainen (2004), Choi, Fong and Wong (2006) have followed and
advanced the work of Wilson. In comparison, the second approach is based on the work of Merton (1974),
in which asset price changes are integrated into default probability evaluation (see Drehmann and
Manning (2004), Pesaran et al. (2006)). Compared with the Wilson (1997) methodology, the Merton
approach has some disadvantages, including high data and computational requirements. According to
Sorge (2006), in the Merton approach, equity prices may be noisy indicators of credit risk. Therefore, we
adopt the first approach in this paper.
Researchers have identified macroeconomic variables that affect credit risks according to specific
macroeconomic situations of the target countries involved. Boss (2002) finds that industrial production,
inflation rate, stock price index, nominal short-term interest rates, and oil prices are the determinant
factors of default probability in Australia. Virolainen (2004) reveals that GDP growth, interest rates,
corporate indebtedness, inflation, industrial production, real wages, the stock price index and the oil
prices are the related factors for the probability of default for banks in Finland. In investigating the
determining factors for credit risk in Hong Kong, Wong, Choi, and Fong (2008) recognize the importance
of GDP growth rates in both Hong Kong and mainland China, interest rates, and real estate prices. In
order to solve the problem of a certain degree of arbitrariness in choosing macroeconomic variables, Boss
(2009) applies a principal component analysis (PCA) to a set of Austrian macroeconomic variables.
Dovern et al. (2010) model the interaction between the banking sector and the macro-economy for
German banks and their VAR analysis indicates that the level of stress in the banking sector is strongly
affected by monetary policy shocks.
For the dependent variable, loan loss provisions (LLPs) and non-performing loans (NPLs) have been
used as a proxy of default rate. However, details of the default rate used for macro stress testing vary from
country to country, depending on the availability of data. For example, a central bank with limited access
to detailed data for individual banks tends to focus on aggregate data, such as LLPs in the banking sector
5
and the default rate of the corporate sector as proxies for the quality of bank loan portfolios.
Current studies of macro stress testing have mainly focused on advanced economies. Kalirai and
Schleicher (2002) use single equations and a large array of macroeconomic variables to test the soundness
of the Austrian banking sector. Hoggarth et al. (2005b) employ a vector autoregressive model for macro
stress tests of the banking sector in the UK. Dey et al. (2007) modify the Wilson (1997) and Virolainen
(2004) approach to investigate the delayed effects of macroeconomic variables on default risks for banks
in Canada. Sanvi Ai et al. (2009) simulate default probabilities for the French manufacturing sector for
several macroeconomic scenarios. Peter Spencer and Zhuoshi Liu (2010) develop a multi-country
macro-finance KVAR model to study international economic and financial linkage among the US, the UK
and other OECD countries.
However, studies of macro stress testing on credit risk of Chinese commercial banks are still at an early
stage due to special features of the Chinese bank system and data deficiencies. Xiong (2006) and Ren
(2007) employ a multiple Logit regression model on macroeconomic factors and finds that GDP, inflation
rate and interest rate are significant factors that affect the stability of the Chinese banking system. Li
(2008) may be the first one to test the relationship between probability of default (PD) and
macroeconomic factors in China based on the frameworks of Wilson (1997) and Virolainen (2004).
However, their results just provide a point estimation result based on stress scenarios. Simulation methods
have not been seen in stress testing for Chinese banks, which is accomplished in this paper.
3. MAIN FEATURES OF THE BANKING SYSTEM IN CHINA
China’s banking system has some unique features. First, four largest commercial banks, often referred
to as “the big four,” dominate China’s banking industry. Between 1985 and 2008, China’s banking sector
went through significant reforms. Our study reflects the performance and defaults of these big four banks
in China.
6
Second, the management and disposition of non-performing loans (NPLs) has been a main focus in
China’s banking reform. As this study chooses the default rate or NPL rate as an indicator of credit risk,
special attention has to be paid to the reforms involving NPLs in our analysis. From 1985 till the
mid-1990s, the big four, designated as specialized state owned commercial banks, took the responsibility
of financing most of the capital needs for investment and constructions by the state owned enterprises
(SOEs). They even had to lend to inefficient sectors, such as stagnant SOEs and, as a result, they
accumulated large volumes of NPLs (see Kumiko (2007)). In the mid-1990s, the Chinese government
decided to turn these specialized banks into truly commercial banks. Since 1999, China took measures to
recapitalize the state-owned banks and peel off NPLs from their balance sheets. As a result, the level of
NPLs has significantly declined ever since.
Third, as the big four had the responsibility to lend to SOEs for a long time, the NPLs in China have
been tightly related to macroeconomic policies in China. According to a survey by the central bank of
China, central and local government intervention and mandatory credit support for state owned
enterprises contributed 60% of the NPLs, while the banks’ own problems contribute for only 20% (Zhou,
2004). Therefore, there may be a more tight relationship between NPLs and macroeconomic factors in
China.
4. METHODOLOGY AND DATA
4.1 The macro VAR framework
We try different methodologies in our study to see which fits our data the best: (1) VAR model with
original or raw macroeconomic variables (Model A family), (2) VAR model with principle component
analysis (Model B family), and (3) structural VAR (SVAR) model with original macroeconomic variables.
The advantage of using original or raw data to establish VAR model is that the model could reflect all the
information of data and we could explain the result obtained easily. However, in order to avoid
7
multicollinearity among these variables, the number of macroeconomic variables may have to be limited.
In our case, four variables are selected. In order to save more macroeconomic information in our model,
we also consider the PCA method to integrate six variables into four principle components, which may
improve the performance of the VAR model. Thirdly, we estimate an SVAR model for the relationship
between NPLs and macroeconomic variables as such a model may reflect the contemporaneous
information of macroeconomic variables.
4.1.1 The macro credit framework – The VAR model. The macroeconomic credit risk model is based on
the Credit Portfolio View model proposed by Wilson (1997a and 1997b) and developed by McKinsey
(1998). This approach is well suited to macro stress tests because it relates the default rate in a given
economic sector to macroeconomic factors. Hence, when the model is estimated, the default rate can be
simulated through the effects of macroeconomic shocks applied to the system. In turn, these default rates
can be used to simulate the loss distribution for a given credit portfolio.
Our macroeconomic credit risk model is very close to the Wilson model. However, there is one major
difference -- we use a Vector Auto-regression model (VAR) with all variables. The VAR approach, made
popular by Sims (1980), has become an important tool in empirical macroeconomics. The popularity of
this approach arose both out of the inability of economists throughout the 1970s to agree on the true
underlying structure of the economy and from the Lucas critique, that changes in policy systematically
alter the structure of econometric models.
The macroeconomic credit risk model we employ (namely the Model A family) is described as follows:
1
( 1,2,..., )p
t t i
i t t
it t i
y ya Z W t N
X X
(1)
where yt denotes the default rate of banks, Xt is the (n1) vector of macroeconomic variables at time t
8
(Xt=(x1,x2,…,xn)’), Zt is the (m1) vector of exogenous variables, a is the ((n+1) 1) vector of constant, Φi
is an ((n+1)(n+1)) autoregressive matrix with lag i, for i=1,2,…,p, B is the (m1) coefficient vector of
exogenous variables, and Wt is a vector of serially uncorrelated innovations, vectors of length n. Wt is a
multivariate normal random vector with a covariance matrix Q, where Q is an identity matrix, unless
otherwise specified.
4.1.2 The principle component analysis (PCA). Principle components analysis (PCA) was invented in
1901 by Karl Pearson, developed by Hotelling (1933) and the best current time reference is Jolliffe (2002).
The aim of PCA is to reduce data dimensionality by performing a covariance analysis between variables.
As such, it is suitable for data sets in multiple dimensions. Instead of estimating the probabilities of
default by the changes of individual macroeconomic variables, we use a PCA and take the resulting
factors as input for the regression analysis. A PCA is an orthogonal linear transformation that places the
projection of the data with the greatest variance on the first coordinate. The other coordinates are chosen
subsequently, so that they explain the maximum remaining variance subject to the condition of
orthogonality. In this paper, the first four components are taken into account and they explain 95% of the
variance of the seven variables. The relationship between components and variables could be explained as
follows (Jolliffe, 2002).
1 11 1 12 2 1n nPCA u x u x u x
2 21 1 22 2 2n nPCA u x u x u x
(2)
1 1 2 2p p p pn nPCA u x u x u x
2 2 2
1 2 1 1,i i inu u u i p (3)
1 1 2 2 0 for all i j i j in jnu u u u u u i j (4)
9
where PCAi (i=1,…,p) is a set of principle components, xi (i=1,…,n) is a set of macroeconomic variables
and uij (i,j=1,…,n) is a set of unknown coefficients which need to be estimated. The first principal
component, PCA1, accounts for the maximum variance in the data, the second principal component, PCA2,
accounts for the maximum variance that has not been accounted for by the first principal component, and
so on.
In this paper, we also establish credit risk model with principle components (model B family namely),
which is computed from the original macroeconomic variables through principle component analysis. The
macroeconomic credit risk model with principle components we employ is described as follows:
1
( 1,2,..., )p
t t i
i t t
it t i
y ya Z W t N
PCA PCA
(5)
where yt denotes the default rate of banks, PCAt is the (n1) vector of macroeconomic variables at time t
(Xt=(x1,x2,…,xn)’), Zt is the (m1) vector of exogenous variables, a is the ((n+1) 1) vector of constant, Φi
is an ((n+1)(n+1)) autoregressive matrix with lag i, for i=1,2,…,p, B is the (m1) coefficient vector of
exogenous variables, and Wt is a vector of serially uncorrelated innovations, vectors of length n. Wt is a
multivariate normal random vector with a covariance matrix Q, where Q is an identity matrix, unless
otherwise specified.
4.1.3 The macro SVAR model. Structural VARs (SVAR) are an extension of traditional VAR analysis.
The structural VAR approach aims to provide the VAR framework with structural content through the
imposition of restrictions on the covariance structure of various shocks. Two features of the structural
form make it the preferred candidate to represent the underlying relations. First of all, error terms are not
correlated. The structural, economic shocks which drive the dynamics of the economic variables are
assumed to be independent, which implies zero correlation between error terms as a desired property. This
is helpful for separating out the effects of economically unrelated influences in the VAR. Second,
10
variables may have a contemporaneous impact on other variables. This is a desirable feature especially
when low frequency data is used since contemporaneous data can be included in the SVAR model.
In our paper, we consider the situation of n multi variables in structural vector auto-regression, as in
Sims (1980). A structural vector auto-regression model with lag p, SVAR(p), is presented as follows
(neglecting a possible intercept and/or exogenous variables for notational simplicity):
0 1 1 2 2t t t p t p ty y y y , 1,2, ,t T (6)
where the vector yt includes the variables of interest and εt is a vector of error terms with
variance-covariance matrix Σ.
12 1
21 2
0
21 22
1
1
1
n
n
,
( ) ( ) ( )
11 12 1
( ) ( ) ( )
21 22 2
( ) ( ) ( )
1 2
i i i
n
i i i
n
i
i i i
n n nn
, 1,2, ,i p ,
1
2
t
t
t
nt
A moving average representation of Equation (6) could be presented as follows.
( ) ,t tL Y u
( ') ,t t kE u u I
(7)
2
0 1 2( ) ... p
pL L L L ,
where Ψ(L) is the parameter matrix of lag polynomial L, Ψ0≠Ik. If Ψ0 is a lower triangular matrix, the
SVAR model is called a recursive SVAR model.
Consider a moving average representation of unrestricted VAR model,
11
( ) , 1,2,...,t tA L y t T (8)
Let A,B be invertible KK matrices, then multiply A
on the left to both sides of Equation (8),
( ) , 1,2,...,t tAA L y A t T (9)
If A,B satisfy the conditions as follows:
, ( ) , ( , ')t t t k t t kA Bu E u O E u u I , (10)
Equation (10) is called a SVAR model of AB type.
This SVAR model of AB type not only builds the contemporary relations among the endogenous
variables, but also reflects the response of the model to independent stochastic disturbance ut shock. If
matrix A is a identify matrix, it means that there is no contemporary relationship among the endogenous
variables and the response of every variable to orthogonal disturbance is simulated by matrix B. If matrix
B is a identify matrix, the contemporary relationship among the endogenous variables exists and is
decided by matrix A.
4.2 Data
The data for our study are retrieved from China Statistical Yearbook published by China Statistics Press
and Almanac of China’s Finance and Banking published by China Financial Publishing House. We use
the piecewise cubic Hermit interpolating polynomial method to match any missing data. Since the model
is based on the whole economic system in China, the average value of every variable is applied instead of
the value in each sector. When there are multiple changes in the interest rate in one year, the weighted
average is used for the relevant year. A line graph of NPLs is shown in Figure 1 and descriptive statistics
12
of all data are presented in Table 1.
[FIGURE 1 GOES ABOUT HERE]
[TABLE 1 GOES ABOUT HERE]
As shown in Figure 1, the trend of nonperforming loans in China’s commercial banks resembles a bell,
with 1999 and 2004 as major turning points. China's main commercial banks NPLs started to decline in
1999 and declined dramatically since 2004. In 1999 and 2004, the Chinese government injected funds
into the four state owned commercial banks to offset their large amount of bad assets, paving the way for
their initial public offering. Since the big four accounted for more than 60% of the total assets in China’s
banking sector, major changes in the asset quality of the big four exert a significant impact on changes of
the overall index.
According to the pairwise correlations between macroeconomic variables, we choose six
macroeconomic variables: Nominal GDP growth (GDPG), unemployment (UEP), nominal interest rate
(NR), exchange rate (ER), Consumer Price Index (CPI), and real property index (RPI). Nominal GDP
growth and unemployment rate allow us to investigate the effect of business cycle on default rates. We
use the 1-year nominal bank loan interest rate for the interest rate variable. This rate is linked to the
majority of loans taken by corporations in China. As is known, it includes both inflation and the real rate
of interest. Interest rate has a direct impact on the cost of loans and hence the credit risk of banks.
Exchange rate is measured by the ratio of RMB, the Chinese currency, to the U.S. dollar, which is
considered as an indicator of China’s international economic environment. An increase in the exchange
rate (a depreciation of the RMB) is viewed as a weakening of the Chinese economy relative to the rest of
the world. CPI is used to measure inflation.
The real property index (RPI) measures the construction units of real property. We consider this
variable for several reasons. First, housing mortgages have grown rapidly in China. Real estate loans in
13
China rose to RMB5.24 trillion in November 2008 with an annual growth rate of 10.3%. The proportion
of real estate loans in total bank loans has grown steadily as well. Second, when the real estate market
booms, even borrowers with financial hardships can pay mortgages or pay back their loans by selling
their properties. However, when the real estate market is down and house prices decline, the risk of
mortgage default increases. Third, the experience of the subprime mortgage crisis in the United States
tells us that banking stability collapses along with the housing market. Therefore, we should consider the
impact of real estate market on banks’ credit risk.
We use NPL ratio of main commercial banks in China is a proxy for default rate for two reasons. First,
the NPL data is available for the period 1985 to 2008. Second, the definition of NPL is similar with the
default rate that other researchers have used in their studies.
5. EMPIRICAL RESULTS AND STRESS TEST
5.1 Empirical results
5.1.1 Results from VAR with original macroeconomic variables. We start estimation of the VAR model
using a limited set of variables (Model A family). The pairwise correlations presented in Table 1 are taken
into account in our variable selection. This selection not only helps to identify which variables are more
correlated with default frequencies, but also to avoid possible multicollinearity problems. As a result, four
variables are selected: exchange rate (ER), nominal interest rate (NR), GDP growth (GDPG), and real
property index growth (RPI). Table 2 presents the results we have obtained using the VAR framework
with these four macroeconomic variables.
[TABLE 2 GOES ABOUT HERE]
In the first model (Model A1) presented in Table 2, the set of endogenous variables comprises the four
macroeconomic variables we have selected. As shown in the column of NR in Table 2, note that the
14
coefficient of NPL(-1) is negative and significantly different from 0.This result shows that an increase in
the default rate in the t-1 period is followed by a sharp cut in nominal interest rate in period t.
Consequently, the lagged values of the default rate affect the future path of some macroeconomic factors.
This is an ex-post corroboration of the use of a multivariate framework to model macroeconomic factors
and the default rate. In addition, the negative correlation between the nominal interest rate and the lagged
default rate could stem from the assumption that a higher default rate reflects a bad economic
environment. So, even though output is not a formal target of the central bank of China, the business
cycle seems to influence the conduct of the monetary policy. For instance, a cut in the nominal interest
rate could be viewed as the response of the monetary authorities to the deterioration of the
macroeconomic environment. Besides, the coefficient of NR(-2) in Table 2 describing the evolution of
NPLs is positive and significant, so that it confirms that an increase in the nominal interest rate in (t-2)
period leads to an increase in the default rate. This result is also in line with the expectation that the
default rate increases along with interest rates owing to higher borrowing costs. In this way, there is a
feedback effect between the default rate and the nominal interest rate. This is a second argument in favor
of a multivariate framework. Exchange rate in (t-1) period displays a negative coefficient to NPLs,
suggesting that RMB currency appreciation leads to a rise in NPLs. However, Exchange rate in (t-2)
period shows a positive effect on NPLs. We will discuss this problem later.
We find no significant relations between NPLs, GDP growth, and RPI growth in Model A1. In Model
A2, we modify GDP growth and RPI growth as an exogenous variable to see if the explanatory power of
the model increases. We also add two dummy variables, a macro dummy and a loan class dummy, to
Model A2. The loan class dummy is intended to control for the change from four to five categories of loan
quality classification in China, which is 1 for the year when the loan grating reform occurred and 0 for the
other years. The macro dummy is to control for the effects of capital injection by the Chinese government
in the banking industry, which is 1 when there was capital injection and 0 otherwise.
15
The AIC and SC values for Model A2 are smaller as compared with those for Model A1, indicating that
the fitting ability of Model A2 is better than Model A1. The VAR stability test is passed as well. In Model
A2, the coefficient of macro dummy is negative whereas the coefficient for the loan class dummy is not
significant. The result suggests that the effect of loan grating reform had no effect on NPLs but the capital
injection did. This finding confirms with a common notion that a change in loan classification has no real
effect on loan performance while capital injection changes the capital structure of the banking sector and
loan performance.
16
5.1.2 Results from VAR with principle components. We include six macroeconomic
variables in our principle component analysis: GDP growth, exchange rate, nominal interest
rate, real property index, CPI, and unemployment rate. The results are presented in Table 3.
The first four principle components can describe 95% of data information, so we employ
these four principle components as endogenous variables in VAR Model B1.Some coefficients
of the variables in Model B1are significant, suggesting that VAR Model B1 can describe the
character of the data. However, as Table 3 shows, there is at least one root outside the unit
circle, so the VAR model does not pass the stability test. We adjust PC1(-1) and PC4(-1) as
exogenous variables and add a macro dummy to get Model B2. Model B2 is a stable VAR
system although the fitting ability does not improve over Model B1, as shown by values of
AIC and SC. In Model B3, we make PC1 and PC4 as exogenous variables to test the
influence of variables in the current period. The coefficient of PC1 and PC4 become
significant and the fitting ability of Model B3 is much better than that of Model B2. Moreover,
Model B3 satisfies the VAR stability condition. Our result suggests that the PCA method does
help retain information and improve the VAR model with original variables. Model B3 shows
that, with PC1 and PC4 as exogenous variables, different linear combination of
macroeconomic variables produces different impact on the default rate. As for the fitting
ability, the results from PCA are better than those from VAR with original variables.
Nonetheless, the PCA models are less intuitive from an economic perspective because every
component is a linear combination of the original macro variables.
[TABLE 3 GOES ABOUT HERE]
5.1.3 Results from SVAR. To test if there is a contemporaneous relationship between NPLs
and the macroeconomic variables, we estimate a structure VAR (SVAR) model in AB type.
We use scoring as the estimation method and short-run pattern matrix as the restriction type.
Table 4 shows the type of A and B matrices we need to estimate. The estimation results are
presented in Table 5. The estimated coefficients of Matrix B are significantly different from
17
zero, indicating that Matrix B is not an identity matrix. However, the estimated coefficients of
Matrix A are not significant, suggesting that Matrix A is an identity matrix. The results of
Matrix A and Matrix B imply that there is no contemporaneous relationship among the
variables. Therefore, SVAR is not a proper model for our data.
[TABLE 4 GOES ABOUT HERE]
[TABLE 5 GOES ABOUT HERE]
Based on the above analysis, Models A2 and B3 are good fits for our data. Figure 2 and
Figure 3 depict the fitting charts of the two models respectively. Comparing the fitting charts
of NPLs, Model A2 is more accurate than model B3. Moreover, as mentioned before, Model
B3 is based on principle component analysis which is difficult to interpret economically.
Therefore, in the stress testing that follows, we choose Model A2 as our basic model.
5.2 Stress test on China’s banking system
5.2.1 Impulse response analysis. We use generalized impulse response analysis for
unrestricted vector autoregressive (VAR) and co-integrated VAR models proposed by Pesaran
and Shin (1998). This approach does not require orthogonalization of shocks and is invariant
to the ordering of the variables in the VAR. Our estimated impulse responses of NPLs to
macroeconomic variables are depicted in Figure 4.
[FIGURE 4 GOES ABOUT HERE]
When the nominal interest rate (NR) has a positive shock of one standard error, NPLs
decrease by 0.2 and start to increase and stay on a positive value after four years. It is
plausible that the rise in NR may discourage consumption and encourage savings in the short
run, leading to an increase loan repayment rate and a decline in NPLs. However, a rise in NR
increases the cost of loans in the long run, resulting in an increase in NPLs after several
periods. NPLs are negatively affected by exchange rate (ER) shocks, although the effect
18
fluctuates slightly at first. Our interpretation is that a rise in the exchange rate (home currency
depreciation) is a sign of bad economy. Turning to the NPLs shocks, they have persistent
effects over ten years on NR and ER, with NR trending down while ER trending upwards.
Finally, we find that NPLs have obvious external time lag effects. As Figure 4 shows, the
impulse of NPLs to ER produces longer time lag effects than NR. That is, a change in the
exchange rate impacts the default rate for a longer time than the nominal interest rate.
5.2.2 Variance decomposition. While an impulse response function describes the influence
of a shock in one endogenous variable on other endogenous variables in VAR models,
variance decomposition analysis is performed to identify the contribution of each shock to the
changes in the endogenous variables, which are usually measured by their variances. Hence,
variance decomposition presents the relative importance of every stochastic disturbance
which affects the variables in VAR models.
Figure 5 shows the result of our variance decomposition analysis. Changes in NPLs due to
an NR shock gradually rises up to 40% after 10 years, while the maximum percentage change
in the variance of NPLs due to an ER shock is 28% in the second year. It means the influence
of NR on NPLs is a long term effect and becomes more important than ER over time. It is
economically plausible that NPLs are more closely related to NR than ER.
[FIGURE 5 GOES ABOUT HERE]
5.2.3 Scenario analysis. Four scenarios of macroeconomic variables are selected for our
stress test on Chinese commercial banks as follows. Here, 2007:1 and 2007:2 represent the
first half and the second of the year of 2007 respectively while 2008:1 and 2008:2 are denoted
in the same way.
(1) The benchmark scenario, in which there is no shock;
(2) A fall in China’s nominal GDP growth by 5%, 7%, and 3% respectively in each of the
19
three consecutive half years starting from 2007:1;
(3) A rise of nominal interest rate by 400 basis points in 2007:1, followed by a rise of 500 bps
in 2007:2, and another rise of 300 basis points in 2008:1; and exchange rate increases by 5%,
8%, and 10% respectively;
(4) Reductions in real property index by 4.2%, 12.3%, and 20% respectively in each of the
three consecutive half years starting from 2007:1.
We have designed the scenarios according to several principles. First, there is a plausible
probability that these changes in the macroeconomic variables will occur. Second, the
scenarios are supposed to be extreme situations. That is, they appear rarely. Third, if one of
the scenarios occurs, it may bring high credit risk to Chinese commercial banks, which will
lead to enormous credit losses to them. All scenarios we have designed are more severe than
what actually happened in prior financial crises. In addition, every scenario has a group of
macroeconomic variables involved. That is, the scenario is more comprehensive than those
that consider shocks of a single macro factor. Last but not least, the macroeconomic variables
in every scenario are a sequence of shocks with a worsening trend. The situation is plausible
in the real world because economic stimulus policies may not be effective immediately to
stem the crisis.
We assume that these shocks occur only in 2007:1 and we analyze their consequences on
default rates and credit losses for 2007:2, 2008:1and 2008:2. In particular, we compare the
simulated default rates and loss distributions in 2008:2 with the ones obtained from the
benchmark or basic scenario (in which no shocks occur in 2007:1). Table 6 reports the
response of default rates for each macroeconomic scenario.
[TABLE 6 GOES ABOUT HERE]
Here we introduce the concept of Value at Risk (VaR) to measure the losses of a given
20
portfolio over time horizon H. Value at Risk (VaR) calculates the maximum losses expected
(or worst case scenario) on an given portfolio, over a given time period and given a specified
degree or confidence. We can describe it as follows (Philippe Jorion, 2006):
Pr( ) 1Loss VaR c (11)
where Loss means the losses of a given portfolio over time horizon H, c is the given
confidence level. From calculating VaR, the bank can know what is the most they can loss on
the NPLs.
Monte Carlo simulation is a good method to calculate VaR, which can get the loss
distribution of a given portfolio overtime horizon H. In this paper, we conduct Monte Carlo
simulation as follows.
According to Equation (1), we first produce a large number of random variables from the
multivariate normal distribution with mean being zero and variance-covariance matrix being
the estimated Q
, which represents a realization of the vector of disturbances Wt. Given the
current and past values of the X, macroeconomic variables, the default rate and the realization
Wt, the associated one-period-ahead values yt+1 and xt+1 can be calculated based on Model A2
we have described earlier.
Similarly, the two-period-ahead values can be calculated with another independently
random variables and one-period-ahead values previously obtained. Repeating the same
procedure yields a future path of the joint sector specific default rates, given the time horizon.
By simulating a large number of such paths, a frequency distribution of the horizon-end
default rates of benchmark scenario can be constructed.
In constructing the distribution of possible credit losses for a stressed scenario, we
introduce a hypothesized adverse macroeconomic development into the simulation procedure
above and generate another set of future path of the joint sector specific default rates, given
21
the time horizon. The difference from the benchmark scenario is that we change the value of
macroeconomic variables in the stressed year. A shock in one variable also leads to
disturbances in other variable through the transmission effect built in the model. In order to
estimate the loss distribution, the simulation is replicated over 2000 times. Table 7 and Figure
6 show the results of such simulations.
[TABLE 7 GOES ABOUT HERE]
[FIGURE 6 GOES ABOUT HERE]
In the benchmark scenario (no shocks), default rates are the real values in every half year.
Because we assume the macro shock begins at 2007:1, there is no response of default rate for
each shock in this period.
As shown in Table 6, compared to the benchmark scenario, the default rate in every other
scenario has strong responses to the specified shocks. From 1998, China’s GDP growth and
unemployment rate were increasing till 2007, so we design this scenario to test the respond of
default rate in China for the sustained increase of GDP growth and unemployment rate. From
Table 6, we can see that the GDP growth and unemployment rate shocks of Scenario 2 have
the strongest impact on the default rate: 7.93% in 2007:2 and 7.02% in 2008:1. But the
influence becomes much less, 3.93%, in 2008:2. The result means that the rise on GDP
growth and unemployment would lead to a quick respond to the default rate in the next period
and the effect would become weak in the next period.
Scenario 4, real property index and CPI shocks, has the longest influence on the default rate,
with the influence being the strongest in 2008:2. In recent years, price increases in
commodities and houses in China have become a major concern. Any price decrease, be it
caused by government policy or otherwise, may cause default rates to rise in the banking
sector. As shown in the Table 6, a decline in commodity and house prices would produce the
longest impact on the default rate.
22
Scenario 3, shocks in interest and exchange rates, has the same trend as scenario 4, but it is
less severe than scenario 4. Increasing interest rate and exchange rate will impose more
pressure on repaying debt owing to the cost of borrowing for both individuals and export
enterprises in China.
As shown in Table 7 and Figure 6, even in the worst situation, commercial banks in China
could still make profits. As shown in Table 7, however, at the confidence level of 90%, banks’
credit losses with shocks from different macroeconomic variables range from 3.64% to 4.29%
of the portfolios. At the 99% confidence level, the losses range from 4.05% to 4.78%. The
estimated maximum losses are very similar to the 5% loss of U.S. banks in November 2008,
when the U.S. market experienced severe financial crisis.
6. SUMMARY AND CONCLUSION
This paper describes a framework for macro stress-testing on credit risk in commercial
banks in China. The framework enables us to measure the vulnerability of commercial banks
against various macroeconomic shocks. The results show that the framework successfully
simulates the related responses of credit risk between severe financial crisis and subsequent
economic recovery.
A distinguishing feature of our study is that the sample period employed to estimate the
model includes several severe financial crises. Thus, we have avoided the shortcoming of
performing stress tests with a model based on benign historical data. Another distinctive
feature of the study is that the relationships between default rate and macroeconomic
variables are jointly estimated in a multivariate framework. Several models – VAR models
with original macroeconomic variables, VAR models with principal components, and
structure VAR model – are tested before we select the best one. We find that after principle
component analysis, VAR model ability could be improved significantly and structural VAR
model does not fit our data. Finally, we simulate the distribution of credit losses by the Monte
23
Carlo method and calculate VaR to report credit risk for commercial banks in China. The
macroeconomic credit risk model with explicit links between default rates and macro factors
is well suited for macro stress testing purposes. We use the model to analyze the impact of
stress scenarios on the credit risk of aggregated Chinese commercial banks. It should be noted
that using NPLs as a proxy of default rates is still a relatively crude approximation. For
further studies or extensions of the model, default rates for specific industry sectors may be
used. Company level rating information and credit rating transition probability matrix may
also be considered.
24
References
Boss, M. (2002). “A Macroeconomic Credit Risk Model for Stress Testing the Austrian Credit
Portfolio.” Financial Stability Report 4: The Oesterreichische National bank (OeNB), 64-82.
Boss, M., Fenz, G., Pann, J., Puhr, C., Schneider, M. Ubl, E. (2009). “Modeling Credit Risk
through the Austrian Business Cycle: an Update of the OeNB.” Financial Stability Report 17:
The Oesterreichische Nationalbank (OeNB), 92-108.
Dey, S., Misina, M., Tessier, D. (2007). “ Stress Testing the Corporate Loan Portfolio of
Canadian Banking Sector. ” Bank of Canada Financial System Review, June, 59-62.
Dovern, Jonas, Carsten-Patrick Meier, Johannes Vilsmeier. (2010). “How Resilient is the
German Banking System to Macroeconomic Shocks?” Journal of banking and finance 34(8),
1839-1848.
Drehmann,M. , Manning, M. (2004). “Systematic Factors Influencing UK Equity Returns. ”
Bank of England, Mimeo.
Hoggarth, G., Sorensen, S., Zicchino, L. (2005).“Stress Tests of UK Banks Using a VAR
approach.” Working paper 282, Bank of England.
Hotelling, Harold. (1933). “Analysis of a Complex of Statistical Variables into Principal
Components.” Journal of Educational Psychology 24(6 & 7), 417-441 & 498-520.
Li, J., Liu, L. (2008). “An Assessment on the Credit Risk for China Commercial Bank
System Based on Macro Stress Testing.” Modern Economic Science 30(6), 66-74.
Jolliffe, I. T. (2002). “Principal Component Analysis.” Second ed. New York: Springer
Series in Statistics.
Kalirai, H., Scheicher, M. (2002). “Macroeconomic Stress Testing: Preliminary Evidence for
Austria.”Financial Stability Report: Austrian National Bank 3, 77-98.
Merton, R. (1974). “On the Pricing of Corporate Debt: the Risk Structure of Interest Rates.”
Journal of Finance 29(7), 449-470.
Kumiko, O. (2007). “Banking System Reform in China: The Challenges of Moving Toward a
25
Market-Oriented Economy.” Rand National Security Research Division.
Pesaran, H., Shin, Yongcheol. (1998). “Generalized Impulse Response Analysis in Linear
Multivariate Models.” Economics Letters, Elsevier 58(1), 17-29.
Pesaran, M.H., Schnermann,T., Treutler, B. J., Weiner, S.M. (2006). “Macroecnomic
dynamics and credit risk: A global perspective.” Journal of Money, Credit and Banking 38(5),
1211-1262.
Peter, S., Liu, Z. (2010). “An Open-Economy Macro-finance Model of International
Interdependence: The OECD, US and the UK.” Journal of Banking and Finance 4(3),
667-680.
Pearson, K. (1901). “On Lines and Planes of Closest Fit to Systems of Points in Space.”
Philosophical Magazine Series 6 2(11), 559–572.
Philippe, Jorion. (2006). “Value at Risk: The New Benchmark for Managing Financial Risk.”
3rd edition. McGraw-Hill, ISBN 978-0071464956.
Sanvi, A., Caroline, J., Ludovic, K., Jeremy, M., Mireille, B. (2009). “Macro Stress Testing
with a Macroeconomic Credit Risk Model: Application to the French Manufacturing Sector.”
www.banque-france.fr.
Virolainen, K. (2004). “Macro Stress Testing with a Macroeconomic Credit Risk Model for
Finland.” Discussion Papers No. 18, Bank of Finland.
Sorge, M., Virolainen, K. (2006). “A Comparative Analysis of Macro Stress Testing
Methodologies with Application to Finland.” Journal of Financial Stability 17(2), 113-151.
Stock, J., Watson, M. (2001). “Vector Auto-regressions.” Journal of Economic Perspectives
15(4), 101-115.
Wilson, T. C. (1997a). “Portfolio credit risk (I).” Journal of Risk 9(10), 111-170.
Wilson, T. C. (1997b). “Portfolio credit risk (II).” Journal of Risk 10(10), 56-61.
Wong, J., Choi, K. F., Fong, T. (2006). “A Framework for Macro Stress Testing the Credit
Risk of Banks in Hong Kong.” Hong Kong Monetary Authority Quarterly Bulletin 10, 1-38.
Ren, Y., Sun, X., Cheng, G., Xian, E. (2007). “Applied Study on Stress Testing Approaches of
26
Credit Risk.” Statistics and Decision 14, 101-103.
Xiong, B. (2006). “An Assessment on the Stability of Chinese Bank’s System: Indicator
Analysis and Stress Tests.” Unpublished PhD Dissertation, University of Southwest, Chengdu,
China.
Zhou, Xiaochuan (2004a). “Some Issues Concerning the Reform of the State Owned
Commercial Banks.” Speech at the IIF Spring Membership Conference, Shanghai, 16 April
2004. Bank of International Settlement’s Website:
http://www.bis.org/list/cbspeeches/from_01072004/page_3.htm.
27
Table 1. Descriptive statistics of variables
NPL GDPG UEP NR ER CPI RPI
Mean 16.82 9.95 3.08 5.94 6.71 107.40 115.3519
Std.dev. 9.55 2.67 0.82 3,25 2.00 2.95 20.70
JB-test 0.35 0.61 0.42 0.39 0.18 0.66 0.34
NPL(-1) GDPG(-1) UEP(-1) NR(-1) ER(-1) CPI(-1) RPI(-1)
ADF-test -2.98* -4.19* -3.36* -2.47* -3.99* -4.00* -5.58*
Correlation NPL GDPG UEP NR ER CPI RPI
NPL 1 -0.29 0.06 -0.13 0.53 -0.33 0.22
GDPG -0.29 1 -0.08 0.12 0.01 0.3 -0.1
UEP 0.06 -0.08 1 -0.7 0.7 -0.46 0.74
NR -0.13 0.12 -0.7 1 -0.43 0.8 -0.75
ER 0.53 0.01 0.7 -0.43 1 -0.34 0.64
CPI -0.33 0.3 -0.46 0.8 -0.34 1 -0.62
This table shows descriptive statistics on the data series used in the empirical analysis. The
means, standard deviations (std.dev.) and correlations are computed over the period
1985-2008. Figures in row 4 refer to p-values corresponding to Jarque-Bera-tests on
normality of the variables. The figures in row 6 refer to test statistics from
Augmented-Dickey-Fuller-tests; an “*” indicates rejection of the hypothesis of a unit root at a
5% confidence level. Note that unit root exists in the level of every variable but there is no
unit root in the first difference.
28
This table shows the result obtained by using VAR with original macroeconomic variables.
Two models we constructed are listed in the table. In model A1, we take exchange rate(ER),
nominal rate (NR), GDP growth (GDPG) and real property index growth (RPI) as endogenous
variables. In model A2, we modify GDP growth and RPI growth as exogenous variables and
add two dummy variables. The macro dummy is to control funds injection by Chinese
government into the four state owned commercial banks in some years. The classes dummy is
to control loan grading change in 1998. In each model, each column represents a sub-equation
in Equation (1).
Table 2. VAR models using original variables
Model A1 Model A2
NPL ER NR GDPG RPI NPL ER NR
NPL(-1) 0.79*** -0.12 0.98*** -0.16**
NPL(-2) 0.14**
ER(-1) -3.00* 0.49*** 0.29***
ER(-2) 3.25* 0.48*** 0.98* 0.44***
NR(-1) -3.11** -2.84* 0.62***
NR(-2) 1.34*** 0.41*** 3.17**
GDPG(-1) 0.52** 0.55*** 0.52**
GDPG(-2) 0.14** 2.03** - - -
RPI(-1) -0.03** 0.45*** -0.04**
RPI(-2) -0.14** - - -
dummy
Classes
Macro -6.80***
C 6.51*** 25.18*** 3.99***
R-squared 0.95 0.96 0.95 0.71 0.93 0.97 0.96 0.94
Adj. R-squared 0.89 0.93 0.92 0.45 0.88 0.95 0.92 0.88
Determinant resid covariance (dof adj.) 150.71 1.33
Determinant resid covariance 4.71 0.16
Log likelihood -173.12 -73.91
Akaike information criterion 20.74 9.72
Schwarz criterion 23.46 11.35
VAR stability Yes Yes
Note: Standard errors are in parentheses.*, ** and *** indicate significance at the 10%, 5% and 1% levels
respectively.
29
Table 3. VAR models using PCA Model B1 Model B2 Model B3
NPL PC1 PC2 PC3 PC4 NPL PC2 PC3 NPL PC2 PC3
NPL(-1) 0.69*** 0.05*** 1.14*** -0.12*** 0.08** NPL(-1) 1.18*** -0.09*** 0.04***
NPL(-2) -0.34*** 0.06* NPL(-2) 0.04**
PC1(-1) 0.65*** -1.04** -0.26*** PC2(-1) -1.30** 0.87** 0.28*
PC1(-2) - - - PC2(-2) 3.03** -0.35***
PC2(-1) -3.76** 0.89** -1.39*** 0.54*** 0.59* PC3(-1) -2.89*** -0.55** 0.38**
PC2(-2) 2.96*** PC3(-2) 2.42**
PC3(-1) 0.57* -0.43** -3.26* PC4 -1.44*** 0.96***
PC3(-2) 3.19*** 0.48*** 2.49** 0.36*** PC1 -0.95*** 0.15*** -0.28**
PC4(-1) 2.46** 1.46*** -1.45** 1.56***
PC4(-2) - - -
Dummies Dummies
Macro -6.66*** Macro -7.40***
C 11.29*** -1.90*** 1.63** -1.84*** C 1.72** -1.35***
R-squared 0.96 0.99 0.77 0.87 0.89 0.96 0.76 0.86 0.95 0.90 0.87
Adj. R-squared 0.94 0.98 0.56 0.75 0.79 0.93 0.59 0.75 0.92 0.83 0.78
Determinant resid covariance (dof adj.) 1.72 E-04 0.43 0.12
Determinant resid covariance 5.38E-06 0.07 0.02
Log likelihood -22.61 -64.51 -50.75
Akaike information criterion 7.05 8.59 7.34
Schwarz criterion 9.78 10.08 8.82
VAR stability No Yes Yes
Note: Standard errors are in parentheses.*,** and*** indicate significance at the 10%,5% and 1% levels, respectively.
This table presents results obtained by using Principle Component Analysis (PCA). In each model, each column represents a sub-equation in Equation (5).
We add some variables in the set of macroeconomic variables in order to produce 4 or more principle component to analyze. We choose first four principle
components because they can describe 95% of data information. In Model B1, the first four principle components are endogenous variables without dummies.
In Model B2, we adjusted PC1(-1) and PC4(-1) as exogenous variables and add macro dummy to control for the influence of bank fund injection by Chinese
government in some years. In Model B3, we adjust PC1 and PC4 as exogenous variables to test the influence of variables in the current period.
30
Table 4. Structural VAR estimates assumption
This table shows the types of A and B matrices we need to estimate when we build the
structural VAR model in AB type (see Equation (10).
A = 1 0 0
C(1) 1 0
C(2) C(3) 1
B = C(4) 0 0
0 C(5) 0
0 0 C(6)
31
Table 5. Structural VAR estimates
Coefficient Std. Error z-Statistic Prob.
C(1) 0.04 0.09 0.41 0.67
C(2) -0.02 0.06 -0.28 0.77
C(3) -0.08 0.13 -0.59 0.55
C(4) 2.08*** 0.31 6.63 0.00
C(5) 0.93*** 0.14 6.63 0.00
C(6) 0.59*** 0.09 6.63 0.00
Log likelihood -96.79
This table shows estimated results from the structural VAR model. Standard errors, z-statistic
and probability (Prob.) are in Columns 2, 3, and 4. ***, **, * indicate 1%, 5%, 10% levels of
significance respectively.
32
Table 6. Expected default rate for each macroeconomic scenario (%)
Period Benchmark Scenario2 Scenario3 Scenario4
2007:1 3.10 3.10 3.10 3.10
2007:2 4.15 7.93 6.57 7.02
2008:1 4.35 7.02 5.33 5.96
2008:2 3.62 3.93 7.01 7.87
This table presents the response of default rates in each macroeconomic scenario. Numbers in
the table represent the percentage changes of default rate in the four periods.
33
Table 7. Loss distribution in half year horizon
Default rate%
Confidence% Benchmark Scenario2 Scenario3 Scenario4
80 3.16 3.52 4.13 3.69
90 3.64 3.81 4.29 3.70
95 3.79 3.94 4.38 3.79
99 4.13 4.22 4.78 4.05
99.9 4.39 4.57 4.93 4.34
Credit loss (billion RMB)
Confidence% Benchmark Scenario2 Scenario3 Scenario4
80 1079.75 1127.40 1263.04 1081.36
90 1127.57 1180.98 1317.90 1131.56
95 1176.39 1221.92 1359.65 1174.87
99 1254.01 1301.31 1446.56 1255.32
99.9 1330.21 1386.40 1528.67 1315.31
This table shows Value at risk (VaR) at different confidence levels in each scenario. We
design four scenarios which are explained in short as follows:
(1) Benchmark: no shocks;
(2) Scenario2: GDP growth and unemployment shocks;
(3) Scenario3: Interest rate and exchange rate shocks;
(4) Scenario4: real property index and CPI shocks.
34
Figure 1. Line graphs of variables
This figure presents the data that is used to estimate the VAR models. NPL figures are given
in percentage. NPLs are defined as the ratio of the sum of total non-performing loans to the
total amount of outstanding loans. GDP growth refers to the annual growth rate of real gross
domestic product for China. Interest rate is the 1-year nominal interest rate. Unemployment
rate refers to the ratio of the number of unemployed people to the total number of workers in
any given year. Exchange rate is expressed as the value of the U.S. dollar in terms of RMB
units.
35
Figure 2. Fitting graph of forecasting data by Model A2
This table shows the fitting ability of Model A2. The solid line represents the actual default
rate and the dotted line represents our forecast through Model A2.
36
Figure 3. Fitting graph of forecasting data by Model B3
This table shows the fitting ability of Model B3. The solid line represents the actual default
rate and the dotted line represents our forecast through Model B3.
37
Figure 4. Impulse responses to generalized one S.D. innovations
The figure shows the impulse responses to generalized one S.C. innovations based on VAR
Model A2 with three endogenous variables (GDP growth, nominal rate and exchange rate
respectively) and two exogenous variable (GDP growth and RPI). The blue lines indicate the
median of the impulses.
-2
-1
0
1
2
3
1 2 3 4 5 6 7 8 9 10
Response of NPL to NR
-2
-1
0
1
2
3
1 2 3 4 5 6 7 8 9 10
Response of NPL to ER
-0.50
-0.25
0.00
0.25
0.50
0.75
1.00
1 2 3 4 5 6 7 8 9 10
Response of NR to NPL
-.2
.0
.2
.4
.6
.8
1 2 3 4 5 6 7 8 9 10
Response of ER to NPL
Response to Generalized One S.D. Innovations
38
Figure 5. Variance decomposition
This figure shows the different contribution percentage of NPL due to shocks in NR and ER
shocks on the first row, and the contribution of NR and ER due to NPL shock in the second
row.
0
20
40
60
80
100
1 2 3 4 5 6 7 8 9 10
Percent NPL variance due to NR
0
20
40
60
80
100
1 2 3 4 5 6 7 8 9 10
Percent NPL variance due to ER
0
20
40
60
80
100
1 2 3 4 5 6 7 8 9 10
Percent NR variance due to NPL
0
20
40
60
80
100
1 2 3 4 5 6 7 8 9 10
Percent ER variance due to NPL
Variance Decomposition
39
Figure 6. Simulated loss distribution under benchmark and stressed scenarios
This figure reports the histogram of credit loss for each macroeconomic scenario. We design
four scenarios as follows:
(1) Benchmark: no shocks;
(2) Scenario2: GDP growth and unemployment shocks;
(3) Scenario3: Interest rate and exchange rate shocks;
(4) Scenario4: real property index and CPI shocks.
(a) Benchmark scenario (b) Scenario 2
2.2 2.4 2.6 2.8 3 3.2 3.4 3.6 3.80
10
20
30
40
50
60
70
80
Frenquency
credit loss(%)
2.2 2.4 2.6 2.8 3 3.2 3.4 3.6 3.80
10
20
30
40
50
60
70
80
90
(c) Scenario 3 (d) Scenario 4
2.2 2.4 2.6 2.8 3 3.2 3.4 3.6 3.8 4 4.20
10
20
30
40
50
60
70
80
90
100
credit loss%
Frequency
2.2 2.4 2.6 2.8 3 3.2 3.4 3.6 3.8 40
10
20
30
40
50
60
70
80
90
credit loss%
Frequency