Recognizing and Studying the Nonperforming Loans of Chinese City
Commercial Banks
Guan Yong-sheng，Zhang Yu，Yu Guang
School of Management, Harbin Institute of Technology, P.R.China, 150001
Abstract: In recent years, a huge amount of nonperforming loans of commercial banks has become one of the biggest obstacles to the reform and development of China's commercial banks, so a method to find out the control of bank nonperforming loans is a core issue in the financial sector and continue to explore research. Based on the study of the existing nonperforming loans of commercial banks, using the data mining method, by comparing a large number of nonperforming loans and normal loan records, a significant feature extraction of the nonperforming loans of commercial banks and carries on the analysis, the early detection of nonperforming loans for the timely warning signal, capture, for commercial bank steady management has important practical significance.
Keywords: commercial bank; nonperforming loan; feature selection; feature analysis
nonperforming loan ratio of commercial banks has been on a downward trend since 2003, But the decrease is different in different years interval, from 2003 to 2006, The decline of NPLs is the biggest, from the original 17.8% down to the international warning line of 10%. Since 2008, nonperforming loan ratio decreased slowly, especially in the past two years The NPL ratio hovered around 1 percent, NPL ratio rising cases also appear in some quarters, this may be due to the nonperforming loans rate of Commercial banks are Close to the limit at the present stage of economic development, NPL ratio may decline further only through the development of economy healthy and sustainable.
Issuse of bad loans is a difficult problem faced by many countries, according to Japan's experience, nonperforming loans has significant harm. The biggest hazard of high rate of nonperforming loans is affecting the bank's ability to support the economy. In recent years, the loans have been controlled extremely careful by Chinese banks, because Bank lending capacity has been impacted by too many bad loans. If by issuing base money to solve the problem of bad loans, the inflation will be raised. If lightly, a high incidence of nonperforming loans will lead to social and moral hazard, if increased efforts to deal with bad loans and business chain bankruptcies will be caused increasing financial risks and social crisis. The presence of huge nonperforming loans will inevitably endanger the safe operation of the banking sector, reducing the bank's ability to resist risks, bank failures can be lead to and even financial crisis. So concerned commercial banks nonperforming loans, find the characteristic attributes of nonperforming loans of commercial banks, monitor and dispose of nonperforming loans of commercial banks, it has become the most important issue we have to attach great importance around the world.
Based on the research of the nonperforming of existing commercial banks Data mining methods has been used, By comparing a large number of nonperforming loans and regular loans recorded, significant features extracted from NPLs of commercial banks were analyzed, in order to determine the type of
1 Introduction
In recent years, a huge amount of nonperforming loans of commercial banks has become one of the biggest problems to restrict the reform and development of China's commercial banks. In the reform process, China's banking sector has accumulated a huge amount of nonperforming loans. 1999-2000, 1.3939 trillion yuan of bad loans for four state-owned commercial banks was stripped by four asset management companies, NPL(nonperforming loan) ratio dropped by nearly 10 percentage points, However, according to five-category classification standard, nonperforming loans of state-owned commercial banks still as high as 34.18%.By the end of 2001,nonperforming loans of Chinese state-owned commercial banks is 25.37% of the loan balance. According to the latest survey, at the end of 2012, the balance of nonperforming loans of the top ten listed banks was 492.9 billion yuan, 64.7 billion yuan has been increased, and even nonperforming loan’s ratio was 0.95%, down 0.01 percentage points. After so many years of reform banking and related sectors, the progress on the control of nonperforming loans made by commercial banks has been seen. The performance of
loans of China's commercial bank in advance, to provide a reference for prevent the occurrence of nonperforming loans[1-2]. It has important practical significance for capturing warning signals early, detecting nonperforming loans early, and the sound operation of commercial banks.
Nonperforming loans are loans abnormal or questionable loans, refers to the borrower could not repay the loan principal and interest to commercial banks as original loan agreement, or there are signs that it is impossible for the borrower to repay the loans of business Bank as the original loan agreement[3].
Yu shows that nonperforming loans are not formed overnight. Nonperforming loans in the conversion will produce many early warning signals. From the borrower's financial statements, the borrowers of the business with banks and borrowers of personnel and management of nonperforming [4]loans can be found early warning signal transformation.
Analysis of a prefecture level city in 1998 October - 2009 February a bank a total of more than 5 individual loan data, through the use of discrete variable regression and clustering, summarizes the characteristics of customer risk prone to bad loans, as banks for personal loans risk management reference. The study found that if borrowers have the following characteristics: loans for personal consumption or commercial, professional individual industrial and commercial households or private owners, the monthly income of less than 1500 yuan, more than 50 years of age, the monthly repayment amount of 5000 - 25000, the loan [5]period in 2-10 years, there is a great risk of bad loans.
Chi finds that bad loans is negative or malignant product of our country commercial bank credit risk, the Commercial Bank of China also has a lot of bad loans. The formation of nonperforming loans in addition to the impact of policy and institutional factors, but also affected by the economic development, the bank's own behavior and other factors. Using multiple regression model to analyze the influence, the nonperforming loan ratio of commercial banks by the GDP growth rate and negatively related, affected by the growth rate of the money supply is negatively correlated with the affected bank's asset liability ratio is positively related with the subject, loans accounted for the ratio of total liabilities and the influence of cases with a positive correlation, affected by the [6]relative size of the bank and correlated with it.
Based on the study of literature, by using fuzzy regression model, and presents a method to reduce the balance of nonperforming loans of banks. First, this paper analyzes the influence factors of nonperforming loans, selected loans this year, the cumulative loan, the loan should be the number of this year, investment in fixed assets of the 4 main factors affecting the balance of nonperforming loans minimization as the objective function, and find out the correlation function between the 4 factors as fuzzy constraints, is established between the balance of NPLs and 4 factors regression equation. Finally, according to the balance of NPLs to minimize the linear programming model, calculated the minimum value of bank nonperforming loan and the value of the 4 factors are given. The results show [7]that this method has better applicability.
An empirical study in the analysis of bad bank personal loans by decision tree technology in the analysis of personal loans has been formed to the risk of personal customer analysis, find out the customer characteristics for the formation of risk; the information obtained is beneficial to the risk management and control, to provide a [8]scientific basis for decision making of bank executives.
Zeng Yan pointed out some development trend of nonperforming loans of commercial banks in China since 2002 from studying the concept of nonperforming loans and its classification, by analysis of Chinese nonperforming loan situation, and the use of simple regression method in relations between NPLs rate and China macroeconomic development index. it is found that some macroeconomic indicators and the nonperforming loan ratio does exist a certain relationship, it makes a general analysis of financial condition and finally the use of discriminant analysis in the multivariate statistical analysis of the theory of enterprise bank lending, and proposed to use the method to improve the situation of bad loans should pay attention to the problem[9-11].
By using the relevant data in recent years of the Shenzhen development bank NPLs, the interest income, capital adequacy ratio, total assets, total amount, total deposits, inventory ratio for influence factors of bad loans, and set up multiple regression model to control the amount of bad loans, the application of Matlab residuals the analysis, combined with the concept of correlation degree of gray pre side, the influence of related factors of bad loans is concluded and analyzed[12].
Study on the causes of nonperforming loans of commercial banks has been very mature, scholars from different aspects summarized the causes of nonperforming loans of commercial banks generated mainly include: central government or local government intervention and their own business problems in two aspects. On this basis, the scholars also put forward to solve the problem of nonperforming loans countermeasures and suggestions[13].
The prediction and identification of nonperforming loans are seldom used by using the regression method, the complex data mining methods for prediction and recognition of nonperforming loans. This makes for some linear models or other regression model to explain the relationship between the amount of nonperforming loans and explanatory variables. But there is probably not a only simple linear regression or logistic regression
relationship between these variables and nonperforming loans[13]. The influence of various factors of nonperforming loans, in the end what variables can better describe the nonperforming loans’ characteristics, and the existing literature did not give the exact explanation.
2 Methodology
2.1 Feature selection method
Feature selection is to select the most effective features from a set of features and to reduce the dimension of feature space, feature selection is one of the key problems in pattern recognition, so feature selection result directly affects the accuracy and generalization capability of classifier[14].
Feature selection method based on whether the algorithm of independent learning, can be divided [15]into filter (Filter) and package type (Wrapper) two.
The feature extraction method used in this paper is the Relief algorithm of multivariable filter feature selection in. The Relief algorithm is proposed by Kira and Rendell in 1992, a multivariable filter feature selection algorithm is known, it is also an algorithm for computing feature weights based on sample learning[16].
Relief algorithm chooses classes by calculating the distance between samples. Due to the characteristics of the distance computation involved will affect the relative distance of the sample, thus affecting the neighbor selection. it will eventually play a role of feature weight evaluation. Relief to achieve the interaction between features in the process of computing the nearest neighbor, so it is a consideration of characteristics of interdependent effects of multi variable filter feature selection algorithm[17-19].
Relief judges the distinguish ability of characteristics between similar neighbor samples and heterogeneous samples. If the difference between the sample characteristics in the same is small, while in the heterogeneous sample of difference between the variables is great, [20-22]then the feature has strong ability to distinguish classes.
A sample set S={S1, S2,... , Sm}, each sample containing P feature, Si={Si1, Si2,... , Sip}, all the characteristics of value for the scalar or numerical type. Si labeled Ci in C, C={Cl, C2} for the label set. Two samples of Si and Sj (1≤i≠j≤m) in t (1≤t≤p) on the definition of difference.
If the characteristic of T is scalar type features as formula (1),
diff(t,s0sitsjti,sj) (1) 1sitsjtIf t is numeric special features as formula (2)
diff(t,si,sj)sitsjtmax (2)
tmint
Firstly, fthe sample set in a randomly selected sample of Si, the choice of a closest to the Si samples from the two types of samples. With the same kind of Si samples by Hit, Si and heterogeneous samples with Miss, using Hit and Miss according to the formula (3) to update weight (Wt) feature of t.
wtwtdiff(t,si,Hit)/rdiff(t,si,Miss)/r (3) To calculate the feature weights in the iteration
process, Si and heterogeneous samples in the feature diff(T, Si, Miss) /r of T minus diff(T, Si, /r, Hit) of Si with the same kind of sample in the feature on the T to distinguish categories the characteristics of stronger ability should be reflected in the differences between heterogeneous the larger and smaller differences between similar characteristics[23-26]. If T has the ability to distinguish, the weights of T should be positive. In order to avoid a random sampling, the iterative process should take r(r>1) times.
2.2 Feature analysis
Problem, when the data of serious imbalance, is always very difficult to satisfactory results, but due to the loan data of nonperforming loans and the proportion of components is few, so the characteristics of nonperforming loans [27]need to be compared to the data analysis division.
So we put forward the hypothesis, H1: normal loans and nonperforming loans have differences in the distribution of the sample data, and the difference is significant.
H2: because the number of loans is much higher than the number of nonperforming loans, resulting in the overall distribution of loan data from normal loan data distribution difference, namely normal loan data distribution Fp (S) and total loans data distribution of F (S) had little difference. It is showed by formula (4) and formula (5).
F(x)Fp(S)+Fn(s) (4) F(x)Fp(S) (5)
It can test whether the real obey the distribution hypothesis by mean and variance analysis method.
Further，through cluster analysis of normal loans and nonperforming loans, the characteristics of nonperforming loans would be well analyzed. Clustering method is [28-29]EM (Expectation Maximization) clustering algorithm.
What’s more, a analysis about content of the custemer Chinese names was undertook, in order to find the effective characteristic information[30].
3 Results
3.1 Data and data processing
This study uses the NPLs problem in China in 2015 as its empirical context. The data comes from a commercial bank database in Heilongjiang from January 2000 to March 2015. There are 96 features and 17112 instances in the data set. As aforementioned, both bank-specific variables and macroeconomic variables are
determinants to influence the situation of bank loan. According to the bank manager’s mark on each instance, the data is labeled with five categories. We follow the bank rule and label different types of loan as {0, 1, 2, 3, 4} to indicate pass, special Mention, substandard, doubtful, virtual loss and Loss respectively.
The data collected by the information contained in the nature of the loan classification, classification, loan period, the benchmark interest rate, currency rates, loan amount, loan date, expiration date, expiration date, the extension of the certificate, certificate of total loan amount, loan balance, cancellation guarantee, the cumulative interest income, interest, paid table of interest, internal interest receivable, table sheet, sheet paid interest paid interest, whether insolvent, operation date, loan certificate number, loan contract number, client code, code, item number, loan certificate number, the original loan category, special loan categories, borrowing, lending account, repayment, repayment period, guarantee form, sign the guarantee contract number, the four category loan classification, five tier classification of loans, in the form of interest five class five grade classification table, Waiqian information, the limitation that the date and information operator etc..
The data preprocessing is to extract attribute has an important effect on the target data from a large number of attributes to reduce the dimensionality of the original data, or deal with some bad data, so as to improve the quality of data and improve the speed of data mining, including data collection and sorting, cleaning, transformation etc.. Data cleaning is mainly in the data, eliminating errors and inconsistencies, and solve the object recognition process.
In the analysis of the obtained data, it found that some field value is empty, in the empty value, for the influence of some field is not the task of data mining, then delete. For the field and need to have a null value, divided into two kinds, one is directly delete records with null values, the benefits of doing so is to get the mining results more practical, but also reduces the number of records, can narrow the scope of mining, avoid blind search, improve the efficiency and quality of data mining, deficiency is to reduce the amount of data will be lost a large amount of information, the data mining results lack of persuasion. The two approach is to take up a null value, the advantage of this method is that can maintain a record number of more information, which is large enough, deficiency lies in the authenticity of the data at a discount, so as to reduce convincing results of data mining.
If the customer name and purpose of the loan is expressed in words, consider to be deleted in the empirical study. But later through the experiment included the customer name in the region and industry information plays a significant role for the research. The loan use this property in most sample value is empty, and fill in the content does not have credibility, thus excluding loan purposes, customer Chinese name was retentioned.
3.2 Feature selection result
According to the task of data mining, has been classified as nonperforming loans on bank loans can be analyzed, adverse factors for the formation of. The target variable is \"whether the nonperforming loan customers\". The predictive variables for the number, office number: IOU, account number, currency code, agency customer number, contract number etc.. In this paper, using the weka3.6 tool Select attributes of nonperforming loans is analyzed, the selection algorithm for the Relief algorithm, the determination of nonperforming loan customer analysis, find out the most effective features to the formation of nonperforming loans, risk prevention.
Parameters: numNeighbours=10, sampleSize=-1, seed=1, sigma=2. The result shows in Tab.1.
Tab.1 Feature selection result
Feature Weight Feature Weight Feature Weight Si1 0 Si9 0.05 Si17 0.008 Si2 0.013 Si10 0.024 Si18 0.009 Si3 0.007 Si11 0.055 Si19 0.001 Si4 0.009 Si12 0.047 Si20 0 Si5 0.034 Si13 0 Si21 0 Si6 0.031 Si14 0 Si22 0.003 Si7 0.016 Si15 0.008 Si23 0.001 Si8
0.023
Si16
0.037
According to the order from small to large, and weights of each attribute are as follows:
The characteristics of feature 11< feature 9< feature 12< feature 16< feature 5< feature 6< feature 10< feature 8< feature 7< feature 2< feature 18< feature 4< feature 15< feature 17< feature 3< feature 22< feature 19< feature 23.
We selected the weight threshold is 0.01, while the results of feature selection, Si11, Si9, Si12, Si16, Si5, Si6, Si10, Si8, Si7 and Si2, namely: overdue interest, interest bearing cycle, floating rate, execution rate, date, maturity, floating, the last day of interest, maturity and the type of loan, excluding other features.
This paper uses ten attributes of C4.5 decision tree to classify the feature selection. The recognition results are shown in Tab.4.
Tab. 2 C4.5 decision tree to classify result TP Rate FP Rate Recall ROC Area Class 0.987 0.257 0.987 0.927 0 0.768 0.006 0.768 0.925 1 0.626 0.016 0.626 0.903 2 0.67 0.002 0.67 0.893 4 0.943 0.224
0.943
0.924
Through the classification test show that the combination of the above variables can help loan classification effectively, but some of the data is directly
related to nonperforming loans.
3.2 Feature analysis result
In the feature analysis we selected data attributes for easy observation and practical significance: branch code, Chinese customer name, amount, loan date, due date, the last day of last calculating interest, open date and last modified date
The nonperforming loan of Class 4,2,1 distribute as Fig 1, Fig 2 and Fig 3. The well perform loan data feature is described as Fig 4.
Fig 1-a The nonperforming loan of Class 4 distribute of brh
Fig 1-b The nonperforming loan of Class 4 distribute of amt
Fig 1-c The NPL of Class 4 distribute of lnd_d
Fig 2-a The nonperforming loan of Class 2 distribute of brh
Fig 2-b The nonperforming loan of Class 2 distribute of amt
Fig 2-c The NPL of Class 2 distribute of lst_c
Fig 3-a The nonperforming loan of Class 1 distribute of brh
Fig 3-b The nonperforming loan of Class 1 distribute of amt
Fig 3-c The NPL of Class 1 distribute of lnd_d
Fig 4-a The well perform loan data distribute of brh
Fig 4-b The well perform loan data distribute of amt
Fig 4-c The well perform loan data distribute of lnd_d
Fig 4-d The well perform loan data distribute of last_c
The nonperforming loan of Class 4,2,1 of data feature data is described as follows in Tab.3, Tab.4 and Tab.5.
Tab.3 The nonperforming loan of Class 4 feature N=496 MIN MAX MEAN deviation brh 101 3206 1772.16 865.73 amt 90000 124000000 9079100. 12040154 lnd_d 20000201 20140806 20100819 37353.91 due_d 20010110 20150109 20112571 36137.16 lst_c 20050921 20150321 20129968 25842.76 opn_d 20000201 20140806 20100819 37353.91 lst_m 20050921 20150409 20130754 25445.77
Tab.4 The nonperforming loan of Class 2 feature N=1261 MIN MAX MEAN Deviation brh 101 3206 1858.88 910.779 amt 40700 549426724 9557203. 21811574 lnd_d 20030103 20141104 20112421 24225.54 due_d 20050102 20150410 20123046 22708.97 last_c 20041221 20150410 20123723 22862.29 opn_d 20030103 20141104 20112421 24225.54 lst_m 20050108 20150410 20125515 21587.55
Tab. 5 The nonperforming loan of Class 1 feature N=480 MIN MAX MEAN Deviation brh 101 3333 2031.89 1223.328 amt 450000 120000000 7291683. 11354688 lnd_d 20031202 20150121 20105676 26954.37 due_d 20050711 20160118 20118108 23385.56 lst_c 20050321 20150321 20114785 23944.07 opn_d 20031202 20150121 20105676 26954.37 lst_m 20050321 20150331 20118993 20120.12 The well perform loan data feature is described as follows in Tab. 6.
Tab.6 The well perform loan data feature
N=14875 MAX MIN MEAN deviation brh 8843 101 2359.99
1964.141 amt 300000000
25000
9013867. 17851641
lnd_d 20150410 20020611 20119880 21350.54 due_d 20241107 20050123 20130100 22221.14 last_c 20150410 20040921 20126865 21375.58 opn_d 20150410 20020611 20119880 21350.54 lst_m
20150412 20050108 20127928
20022.06
Feature of all loans is described as Tab.7.
Tab.7 Feature of all loans
N=17112 MIN MAX MEAN
Deviation
Brh 101 8843
2296.82 1872.353
amt 25000 549426724 9007489 17887389
lnd_d 20000201 20150410 20118379 22764.18 due_d 20010110 20241107 20128736 23130.80 lst_c 20040921 20150410 20126384 21815.31 opn_d 20000201 20150410 20118379 22764.18 lst_m
20050108
20150412 20127581 20387.09
Through the observation, we find that hypothesis 1 and hypothesis 2 are verified.
After the clustering the different kind of loans data, it was found that it was more easy to find congealing point for NPLs data than normal loans. The numbers of congealing points for different kinds of loans are shown in Tab.8.
Tab.8 Number of congealing points Normal NPL 1 NPL 2 NPL 4 1 7 5 7 The frequency of industry word in the customer name shows in different kinds of loans. these words are Agricultural, wood, Mine, oil, paper，food, coal, milk and Jewelry. The frequency is shown in Tab.9.
Tab.9 The frequency of industry words Normal NPL 1
NPL 2 NPL 4
Agri 1516 17 155 17 Wood 803 14 126 15 Mine 177 11 47 26 Oil 593 10 53 16 Paper 379 1 56 8 Food 945 8 42 21 Coal 722 13 147 54 Milk 25 0 50 0 Jewe
401
10
48
9
4 Conclusion
The present situation of the existence of a large number of nonperforming loans from the commercial bank at the present stage, based on the existing nonperforming loans of commercial banks, the relief algorithm used in data mining methods for feature selection on data sets, in the processing of a commercial bank loans to the public based on the data, through the comparison of a lot of bad loans and normal loan records, a significant feature extraction of nonperforming loans of commercial banks is analyzed, the experimental results show that the overdue interest, interest bearing cycle, floating rate, execution rate, date, maturity, floating interest on the way, the last ten characteristics, maturity and type of loan has good recognition ability to judge the nonperforming loans of commercial banks in china.
This paper from the actual data confirmed the theoretical hypothesis, hypothesis 1: abnormal distribution of different nonperforming loan data often loans, the overall distribution hypothesis 2 loan data and normal loan data distribution difference.
Through the analysis, It was found that the distribution amount of nonperforming loan data was related to time distribution, region distribution, and the theses issues appear on the density of central tendency. A loan subject tended to fail all of its loans, if it has a NPL. According to the manual data analysis, the research found that a large number of nonperforming loans in each subject on the same day got more than one loan, because the single paragraph approval system limit the bank's nonperforming loans. Before the occurrence of NPLs the issuer seemed to know their high risk.
Through the analysis of the loan object name in the information industry, the coal and mine industry nonperforming loan ratio is higher than most of others, the milk industry got the highest NPL ratio, by reaching more than 66%. Other industries such as agriculture, forestry, oil, jewelry and so on, had approximate distribution as overall data.
Further research will be to use the density analysis method with loan data detection, and we hope to be able to predict the NPLs timely and effectively by machine learning methods, we will utilize customer informations on nonperforming loans for further mining by technology of Natural Language Processing.
References
[1]LIU Yan. The causes and related factors of non-performing loans of Chinese commercial banks[J]. System Engineering, 2014, 32(5):66-75(in Chinese). [2]Winston T.H. Koh,Roberto S. Mariano,Andrey Pavlov,Sock Yong Phang,Augustine H.H. Tan,Susan M. Wachter. Bank lending and real estate in Asia: market optimism and asset bubbles[J]. Journal of Asian Economics, 2004, 5(6):1103-1118.
[3]Yang Changsheng. The definition of the non-performing loans of commercial banks and
recognition[J]. Finance and Accounting Monthly,2011,1(2):41-42(in Chinese).
[4]Yu Ming ye ,Yang Yun Peng, Wang Gui Liang. Analysis of the risk characteristics of the bank personal non-performing loans[J]. Oriental Enterprise Culture,2010,1(1):22-27(in Chinese).
[5]Chen Mu zi, Huang Yi qiu, Chen Min , Yang Xiao guang. Extreme zero recovery rate analysis based on family of logistic models[J]. Chinese Journal of Management Science, 2011,19(2):16-22(in Chinese). [6]Chi Qing yun, Chen Guang xing. An empirical
study on the decision tree technique in the analysis of non-performing loans in the individual[J]. Journal of TongHua Teachers College,2010,31(4):51-53.
[7] He M, Liu N. Discrimination for non-performing loans recovery: a method of support vector machines based on wavelet transform[C]. Information Science and Engineering (ISISE), 2010 International Symposium on. IEEE: 88-92.
[8]Campbell A. Bank insolvency and the problem of nonperforming loans[J]. Journal of Banking Regulation, 2007, 9(1): 25-45.
[9] Guya K. Non-performing loans[J]. Research and Economic Analysis, 2011,37(1): 7-8
[10] Kauko K. Esiternal deficits and non-performing loans in the recent financial crisis[J], Economics Letters, 2012,115(2): 196-199.
[11]Lin Jing. Feature extraction of machine sound using wavelet and its application in fault diagnosis[J]．NDT&E International．2001,34(2):25-30.
[12]Maggi B, Guida M. Modelling non-performing
loans probability in the commercial banking system: efficiency and effectiveness related to credit risk in Italy[J], Empirical Economics, 2011, 41(2): 269-291. [13]Parnes D. Modeling operational risk for good and bad bank loans[J]. Journal of Operational Risk Volume, 2012,7(4):13.
[14]Vapnik V, Golowich S E, Smola A. Support vector method for function approximation, regression estimation, and signal processing[J]. Advances in neural information processing systems, 1997,11(1): 281-287.
[15]Simon H Kwan. Operating performance of banks among Asian economies: An international and time series comparison[J]. Journal of Banking and Finance . 2003,27 (3):201-208.
[16] Dan Rosen,David Saunders. Risk factor contributions in portfolio credit risk models[J]. Journal of Banking and Finance . 2009,34 (2):336-349.
[17]Ling Hu. Dependence patterns across financial markets: a mixed copula approach[J]. Applied Financial Economics . 2006 ,16(10):717-729.
[18]Zenglian Z. Research of Default Risk of
Commercial Bank's Personal Loan Based on Rough Sets and Neural Network[C]. Intelligent Systems and Applications (ISA), 2011 3rd International Workshop
on. IEEE,2011:1-4.
[19]Zhang Z, Cao M. Research of credit risk of commercial bank's personal loan based on CHAID decision tree”[C], Artificial Intelligence, Management Science and Electronic Commerce (AIMSEC), 2011 2nd International Conference on. IEEE, 2011:7160-7164.
[20] Duffie D, Singleton K J. Credit risk: pricing, measurement, and management[M]. Princeton University Press, 2012.
[21] Gamble, W. Going bust: Overcoming a dysfunctional credit system[J]. Harvard International Review, 2002,25(2), 54-57.
[22]Li J, Ng C K. The Normalization of Deviant Organizational Practices: The Non-performing Loans Problem in China[J]. Journal of business ethics, 2013, 114(4): 643-653. [23]Isabelle Guyon,Jason Weston,Stephen Barnhill,Vladimir Vapnik. Gene Selection for Cancer Classification using Support Vector Machines[J]. Machine Learning, 2002,46 (3):389-395. [24]Wang Shitong,Korris F.L. Chung,Deng Zhaohong,Hu Dewen,Wu Xisheng. Robust maximum entropy clustering algorithm with its labeling for outliers[J]. Soft Computing . 2006,10 (7): 555-563.
[25]Marko Robnik-Šikonja,Igor Kononenko. Theoretical and Empirical Analysis of ReliefF and RReliefF[J]. Machine Learning , 2003,53 (2):23-69.
[26]Feng Liu,Ali Goodarzi,Haifeng Wang,Joanna Stasiak,Jianbo Sun,Yu Zhou,Jiucheng Xu,Lin Sun,Yunpeng Gao,Tianhe Xu. An ensemble feature selection technique for cancer recognition[J]. Bio-Medical Materials and Engineering, 2014,24 (1):25-26.
[27]Brian L. Mark,Yariv Ephraim. An EM algorithm for continuous-time bivariate Markov chains[J]. Computational Statistics and Data Analysis, 2013,57 (1): 524- 541.
[28]Yang Ji dong. The theory of EM algorithm and its application.School of Finance and Statistics[M],East China Normal University,Shanghai ,China,2013(in Chinese).
[29]Klaus Schaeck,Martin Cihák. Banking Competition and Capital Ratios[J]. European Financial Management . 2010 ,18(5): 836-866.
[30]Nazlia Omar,Nomariani A.Razik. Determining the Basic Elements of Object Oriented Programming Using Natural Language Processing[J]. Journal of Information Technology, 2008,3(10):201-207.
因篇幅问题不能全部显示，请点此查看更多更全内容