A New Method for Early Warning System of Commercial Banks:
Hybrid of Trait Recognition and SVM
Yinhua Li
School of Management, Graduate University of Chinese Academy of Sciences
Beijing 100190, China
liyinhua07@mails.gucas.ac.cn, (O) +86-10-8268-0697, (F) +86-10-8268-0698
Jianping Li
Institute of Policy & Management, Chinese Academy of Sciences
Beijing 100080, China
ljp@casipm.ac.cn, (cel) +86-1366-138-4584
Yong Shi
Research Center on Fictitious Economy and Data Science, Chinese Academy of Sciences
Beijing 100190, China
and
College of Information Science and Technology, University of Nebraska at Omaha
Omaha, NE 68118, USA
yshi@gucas.ac.cn, (O) +86-10-8268-0697, (F) +86-10-8268-0698
1
A New Method for Early Warning System of Commercial Banks:
Hybrid of Trait Recognition and SVM
Yinhua Li1,2, Jianping Li3 and Yong Shi1,4*
1
Research Center on Fictitious Economy and Data Science, Chinese Academy of Sciences
Beijing 100190, China
2
School of Management, Graduate University of Chinese Academy of Sciences
Beijing 100190, China
3
Institute of Policy & Management, Chinese Academy of Sciences
Beijing 100080, China
4
College of Information Science and Technology, University of Nebraska at Omaha
Omaha, NE 68118, USA
Abstract
Commercial bank system is a core sector, which is crucial for the stability of financial system. As a phenomenon,
bank failure occurred frequently in most countries in the world (e.x. America, Russia and Turkey). In last twenty
years, the average number of failed bank almost reaches 175 per year in America. Based on existed models, the
paper aims to introduce a new method for prediction of bank instability based on the historical financial data of
316 American commercial banks. The method purposed in this paper is to combine information encoding
technique, which is used in trait recognition with SVM model. The conclusion suggested that our method
outperformed logistic model and trait recognition model in prediction accuracy in both training sample and
testing sample.
Key words
Early Warning; Commercial Bank; Failure; Hybrid; Trait Recognition; SVM
1. Introduction
Generally, it will bring chaos to economy order if commercial bank system collapses.
Therefore it becomes vital to forecast bank failure for both scholars and supervisors. To forecast
bank failure precisely, it is necessary to estimate early warning system of bank failure.
Based on the existing researches (Karels 1987, Kolari 1996), they demonstrate that historical
data from accounting sheets have the prediction power for future operating condition of banks. It
is far reaching that establishing an effective early warning model for commercial banks. Firstly, it
is propitious to allocate resources for central banks. If problem banks can be detected in time, it
can help reducing the cost of regulation and supervision. Secondly, it can make full use of
historical data. The attributes variables mainly come from the accounting sheets of the banks.
Based on the historical data, bank failure can be forecasted. Thirdly, it can provide evaluation
standard to supervisory board in order to judge if banks are given the proper rating.
As the development of computer technology, intelligent model are widely applied in the
area of bank failure prediction. The key problem will be dealt with in this paper are examining the
*
Corresponding author
2
effect of the trait recognition method in different data sets. According to the test result, a new
method is proposed to predict the commercial bank failure. The reminder of the paper is
organized as follows. The next section reviews the literatures on how to establish the early
warning system for bank failure prediction. The third section, we analysis and summarize the
deficiencies of the trait recognition method through describing and testing the general steps of
the traditional method. The forth section, based on the third section, we proposed the new
method hybrid of trait recognition and support vector machine method. And the robustness of
our method will be tested in different data sample. The last section, we get the conclusions.
2. Literatures review on early warning system EWS in bank failure
prediction
During the last twenty years, more than 10 countries have experienced banking and
currency crisis , with contractions in GDP of between 5% and 12%(Gutierrez et al ,2010).
Intelligent models are broadly used for prediction of bank failure. A knowledge tree is given to
describe main methods (Fig 1).
The most commonly methods include artificial neural network (Gelick and Karatepe,2007),
decision tree (Marais,1984), support vector machine (Vapnik,1995), trait recognition (Kolarl et
al.,1996; Lanine and Vander,2005), rough set technique (Pawlak,1982) and K-nearest neighbor.
The trait recognition technique is a non-parametric method. Kolarl et al(1996) first introduced
this technique into the research of bank failure early warning. He chose 1950 American
commercial bank as sample during 1984 and 1985. The ration between failed and nonfailed bank
is one to six. The main conclusions of Kolarl et al(1996) are summarized in table 1. The curacy in
both original split sample and holdout split sample is more than 95 percent. The relative work
(Gleb and Vennet 2006) investigate the 582 failed bank date from 1988 to 2004 in Russia. The
accuracy of their modified trait recognition model is 85.1 percent in holdout split sample.
The back
propagation neural
network (BPNN)
Artificial neural network
(ANN)
Gelik and
Karatepe (2007)
MLP
Intelligence
modeling
techniques
CL
SOM
Support vector machine
(SVM)
Vapnik (1995)
The principle
component neural
network (PCNN)
LVQ
Rough set technique (RST)
Pawlak (1982)



Rate of nonperforming loans
relative to total loan
Capital relative to assets
Profit relative to assets
Equity relative to assets
Idea: maximize the margin
between the two data sets
Trait recognition (TR)
Decision tree (DT)
Marais (1984)



Kolari et al.(1996)
Logit and TR for large US bank


Lanine and Vander Vennet
(2005) logit and TR for Russia

A nonparametric approach
Do not impose any
distribution assumption
Information about
interrelation of variables
Classification
Prediction
K-nearest neighbor
(K-NN)
Fig1 Knowledge tree of main methods for bank failure prediction
3
Table1 The main conclusions of existing work of trait recognition
Model
Sample
James,
Caputo
and
Wagner,
1996
1984
1985
Failed bank:126
Failed
bank:123
Nonfailed bank:
878
Nonfailed
bank:862
Gleb and
Vennet,
2006
Failed bank:582
from
1988-2004
Nonfailed
bank:2910
Data source
accuracy
FDIC(federal
deposit
insurance of
corporation)
1984
1985
Original split
sample:96.6%
Original split
sample:98.6%
Holdout split
sample:96%
Holdout split
sample:95.6%
CBR(central
bank of Russia)
Original
sample:91.6%
Hold
out
sample:85.1%
3. Background of Trait Recognition (TR)
The trait recognition technique is a nonparametric classification model to distinguish the
failed bank from nonfailed bank. The advantages of the TR technique including that it impose no
priori assumptions on the variables. TR emphasizes the interactions among the independent
variables. In TR procedure the following steps to system design are common (James Kolari,
Mighele Caputo and Drew Wagner, 1996):
Step1. Quantifiably measure the character or traits of the bank and encoding the information.
The first step of TR procedure is to construct a binary string according to financial ratios of each
bank. For example, consider financial ratio A and B reflected of bank financial condition. Based on
sample data, the distribution of these ratios can be plotted. Then for each ratio, two breakpoints
are established to segment the distribution of observation into three segments: 1.High only
non-failed banks (H, coded “00”); 2.Mixture of failed and non-failed banks (M, coded “01”); 3.
Only failed banks (L, coded “11”). Generally, the observation M is described by the binary
code M1M 2 ...M l , l is the length of the string. However, for some financial ratios, e.g. ROE,
either higher or lower value of ROE represents the failed banks. Therefore, the distribution of the
banks can’t be segmented into three parts by two breakpoints. So we adapt a method to solve
the problem (Gleb and Vennet 2006). When the cutoff points are chosen, we allow 10% failed
banks existing among the non-failed banks in the high section. In a similar way, 10% non-failed
banks are allowed to exist among the failed banks in the low section.
Step2. Establish the matrix of traits
To extract the distinctive feature that represent the common pattern of different groups of
observation, the string of binary code should be transferred into a matrix of trait. Each trait is
comprised of an array with six integers: T=p, q, r, P, Q, R. The letter p, q, r represents the positions
in the binary string from left to right. P, Q, R give the values of the binary code at the position p, q
and r. Take the binary string 0001 , for example, the matrix of traits for this bank is as follow:
4
Table 2 the matrix of traits for a bank
p
q
r
PQR
p
q
r
PQR
1
2
3
4
1
1
1
1
2
3
4
2
3
4
1
2
3
4
2
3
4
000
000
000
111
000
000
011
2
2
3
1
1
1
2
3
4
4
2
2
3
3
3
4
4
3
4
4
4
000
011
001
000
001
001
001
Step3. Feature selection
The traits matrix will become huge when the number of the financial variables increases. And no
doubt many traits are not able to be used as the features that distinguish between the failed
bank and non-failed bank. So it is necessary to select the feature from the large number of traits.
A feature is the trait present relatively frequently in the non-failed (failed) banks but relatively
infrequently in the failed (non-failed) banks. For example, the traits can be treated as a good
feature found in 80 percent of the non-failed banks and only 10 percent of the failed bank, and
vice versa in “bad” feature.
Table 3 The judgment standard of feature
Good feature
Bad feature
Failed
Non-failed
10%
80%
80%
10%
Step4. Voting
All sample banks are placed in a voting matrix with axes representing the number of good and
bad votes. A cutoff line then is chosen by visually inspecting the distribution of failed and
non-failed banks in the cells of the voting matrix. If a bank has more good features than bad
features, it will be classified as non-failed bank, rather than vise versa. If the number between the
good and bad features is not balance, the total times of features can be replaced by average
appearance times.
Although the trait recognition technique used for predicting bank failure has been improved
by Lanine (2006), it also exists some deficiencies. E.x., the robustness of the trait recognition
model need further testing in different period data sample. Besides, the process of selecting good
or bad feature is arbitrary. It will lead to missing the important information during the selection
process.
4. Hybrid of trait recognition and SVM
To overcome the deficiencies in existing trait recognition model, we combine information
encoding technique, which used in trait recognition with SVM model reasonably. The specific
process is described as followed: Select proper training data sample. Then the step1 and step2 of
trait recognition are continued to be used. After encoding as traits, substitute them into liner
5
SVM model, see equation (1) .
min

l
1 l l
yi y j ( xi  x j ) i j    j

2 i 1 j 1
j
l
s.t.
 y
i 1
i
i
0
0   i  C , i  1,..., l
(1)
According to the result of ten-fold cross validation, a high accuracy model is selected. In
order to demonstrate the robustness of the model, we substitute another test data set into the
model.
The new proposed method has three advantages. Firstly, it will not miss any useful
information. Secondly, it will give a clear result whether a bank belongs to failed group or
non-failed. And thirdly, it will consider correlation among two or three attributes through
encoding process.
5. Data description
The data is collected from Federal Deposit Insurance Corporation (FDIC), including both
failed and non-failed commercial bank. The number of the failed bank is 158. And we select the
same number of non-failed bank corresponding to the failed bank. Meanwhile, we collect
another 98 banks are used as test sample. Half of the test samples are failed after June 2010. And
for each bank, eight financial ratio are chosen including Profit margin, asset, Capital growth,
Loan funding, Provision rate, Short term gap, Capital ratio. These financial ratios cover six aspects
of financial condition, including profit, size, growth, loan risk, interest rate risk and capital.
Table 4 financial ratio for each bank
Category
Financial ratio
Remarks
Profit
Size
Growth
Profit margin
asset
Capital growth
Loan risk
Loan exposure
Loan funding
Provision rate
Interest rate risk
Short-term gap
Capital
Capital ratio
Net interest income/total assets
Total assets
(Total equity(t)-Total equity(t-1)/
Total equity
net loans and leases/Total assets
Net loans and leases/Total deposits
Provision for possible loan
losses/Total assets
(Assets adjustable 1 day to 1 year liabilities adjustable1 day to 1 year)/
Total assets
Total equity/Total assets
For each bank in training sample, there are 816 traits reflecting the information among any
two or three financial ratios. Follow the principle of feature selection, there 281 features, in
which there are 156 bad features and 125 good features among all the 113424 traits. And the
preliminary result is summarized in table 5.
6
Table 5 Preliminary result
Method
Accuracy
rate
(percent)
Logistic
Modified
TR
Hybrid of TR
and SVM
Training
sample
87.34
90
95.8
Testing
sample
85.8
71.4
88.2
6. Conclusions
Trait recognition is an important method to predict the commercial bank failure. However,
the robustness of the method performs not very well comparing to another two methods. The
accuracy of the hybrid of TR and SVM outperform the TR neither in training sample and testing
sample. It can make full use of information represented by financial data. The new method
proposed in the paper still has some deficiencies. The result of the method depends on the
training sample heavily. The selection of cutoff point is arbitrary.
Acknowledgement
This research has been partially supported by a grant from National Natural Science Foundation
of China (#70621001, #70531040, #70501030, #10601064, #70472074),National Natural Science
Foundation of Beijing #9073020, 973 Project #2004CB720103, Ministry of Science and Technology,
China and BHP Billiton Co., Australia.
Reference
Boyacioglu, M.A., Yakup K. and Omer K.B. Predicting bank financial failures using neural
networks, support vector machines and multivariate statistical methods: A comparative
analysis in the sample of savings deposit insurance fund (SDIF) transferred banks in Turkey.
Expert Systems with Applications, 2009, 36(2), 3355-3366
Celik AE, Kararepe Y. Evaluating and forecasting banking crises through neural network
models: an application for Turkish banking sector. Export systems with Applications 2007:33;
809-15
Honohan, Patrick. Banking System Failures in Developing and Transition Countries: Diagnosis
and Prediction. BIS WP NO39, 1997
Kaminsky, Graciela and Garmen Reinhart. The Twin Crises: The Cause of Banking and
Balance-of-Payments Problems, America Economic Review 89(3), Jun 1999, p. 473-500
Karels GV, Prakash AJ. Multivariate normality and forecasting of business bankruptcy. Journal
of Business Finance and Accounting 1987; 14(4)
Kolari J., Michele C. and Drew W. Trait recognition: an alternative approach to early warning
systems in commercial banking. Journal of Business Finance and Accounting 1996; 23(9)
Kumar, P.R. and V.Ravi. Bankruptcy prediction in banks and firms via statistical and intelligent
7
techniques- a review. European Journal of Operational Research, 2007, 180,1-28
Marais ML, Patel J. Wolfson M. The experimental design of classification models: an
application of recursive partitioning and bootstrapping to commercial bank loan classification.
Journal of accounting research 1984; 22; 87-113
Pawlak Z. Rough sets. International Journal of Computer and Information Science 1982; (11);
341-56
Rumelhart, D.E.,Hinton,D.E.,&Williams,R.J. Learning representation by backpropagating
errors. Nature 1986; 323(9)
Vapnik VN. The nature of statistical learning theory. Berlin: Springer; 1995
Zhang, G.P. Neural networks in business forecasting. Hershey, London: Idea Group
Publishing.2004
8