Bank Credit Risk Management Early Warning and Decision-making based on BP Neural Networks 1 Zhi-Yuan Yu, 2 Shu-Fang Zhao 1 Department of Economics and Management, Taiyuan Institute of Technology, Taiyuan Shanxi, China, economicer@sina.com 2 Institute of Computer Science and Technology, Taiyuan University of Science and Technology, Taiyuan, Shanxi, China Abstract Credit management, which is the basic of the credit application, is the most perfect embodiment in the bank credit application and asset supervision. The ultimate purpose of credit management is to ensure that credit fund is of safety, profit and fluidity. At present, it is extremely important of commercial banks to set up an early bank system. The author who makes great effort on the credit and its reason, besides bringing in the western commercial banks experience of bank credit management, makes researches on credit in a microcosmic view. The author sets up early indicators for commercial bank credit, and carries out the for the credit in advance with the help of artificial neural networks. The experiment has proved that this method is objective and effective. So it can provide theoretical basis, which is more scientific and credible for detection and early about commercial bank credit. 1. Introduction Keywords: BP Neural Networks; Credit Risk; Early Warning; Decision-Making Since any business are faced with various s, the commercial bank operating financial assets as a special enterprise, even operating its special management target, a wide range of social contact and a strong far-reaching influence becomes the focus of the of separation. Credit of commercial bank credit s faced by commercial banks, is the most basic, the most ancient and most dangerous [1]. Generalized credit refers to the loan principal security and income uncertainty and volatility, all kinds of uncertain factors lead to banking financial institutions in the business activities of loss or gain extra income a possibility. Special credit refers to the bank loans can not be recovered, resulting in the possibility of loss of credit funds. In actual operation, the bank 's first concern is to narrow the credit. Because only realized the value of assets, could they talk about the income. According to the present our country banking industry 's specific situation, bank credit assets makes most of the bank assets [2].This paper mainly studies the narrow credit. Credit activities mainly involves banks, borrowers and monetary fund. The maintain of relationship between banks and borrowers needs borrowing, which is with monetary fund as an intermediary. As a result of this lending relationship takes place in the economic environment, they are limited by the environmental factors and impacts. At present,the bank credit comes mainly from the following several aspects: (1) The market. Currency market fund price changes led to the bank the potential loss that market s, including interest rate, exchange rate and inflation. For the market interest rate as the fund demand fluctuations, so when the loan interest rate is higher than the credit contract interest rate, it will cause the credit assets relative loss. Exchange rate in general international credit activity occurs in. The is the main reason for China s interest rate and exchange rate management is still not fully open, belong to a kind of macroscopical management y (2) The operational is caused by not a sound internal control system, fault management, low quality of personnel and the personnel such as moral hazard. The operation of the employees by moral of fraud, will cause huge loss to the bank. (3) Credit refers to the that the debtor fails to repay the default due to various reasons causing the possibility of loss and its size. It is the most common, but also caused the loss of one of the biggest s in a commercial bank. When the bank carries on the credit, it will judge the borrower's credit level. But the borrower's credit level is not fixed, and it will have a credit, if out Advances in information Sciences and Service Sciences(AISS) Volume5, Number9, May 2013 doi:10.4156/aiss.vol5.issue9.51 428
of control of the bank. Credit is often caused by the financial crisis and it makes the bank faces huge loss. Therefore converting the bank credit measurement to measure the financial situation of enterprises is the effective way to reduce credit. The establishment of commercial bank credit early index, which used artificial neural network to forecast the credit of the credit management system is the reason that forms this context[3]. Enhancing credit management to respond to the new financial situation and credit quantitative research is the necessary task of commercial banks. Our country commercial bank loans to enterprises financial condition monitoring is not enough, usually after treatment rather than prevention, qualitative analysis quantitative analysis more, even when the credit assets and losses, the bank can not aware. When the bank credit assets are aware, has become bad debts and bad debts, resulting in the loss of the credit assets of commercial banks. Therefore, credit management system based on the BP neural network for commercial bank credit management[4], can any time analyze loan financial situation and management performance of enterprises to conduct a comprehensive, systematic analysis, through a simple statistical processing, resolution of the financial situation of enterprises, on the commercial bank credit precaution is necessary. From the bank credit management theory, credit has the diffusion and hidden features, if not controlled in time, will have a great impact on Commercial Bank management. And through the use of credit management system based on BP neural network,the of bank credit is to perfecting the management, it complements the commercial bank credit quantification in theory. From the information economics theory, credit is the fundamental cause of the asymmetry of information. For example, before the credit contracts signed, the borrower has more information in about their own financial status, use of the loan and the of investment projects ; and after it, they have more complete information in about the practical use of funds and the completion of the project[5]. As funding,the bank not directly involved in the actual operation of investment projects, they only through indirect channels to understand the statement investment project benefits and s. The asymmetric information in the financial intermediary and the nature of the financial intermediary intrinsic vulner has a special significance. Moral and adverse selection in this asymmetry is generated under the surroundings. In the bank credit market, facing the degree of different credit enterprises, banks often fail to identify enterprise project investment. And the use of credit management system based on BP neural network can provide the basis for 2. Construction BP neural networks principle and model 2.1. BP neural networks principle How to analysis the credit and is more important. From the point of the practice of bank credit management, the analysis of financial statement is the credit management important link, and the credit management system based on BP neural network can clearly reflect the customer's credit and financial situation[7]. Through quantitative analysis, it can objectively reflect the financial information. The use of credit management system based on the BP neural network can solve the credit management departments the poor information sharing, credit decision-making non-standard wait for a problem more standardized. The Application of credit management system based on BP neural network can monitor the credit continuously and effectively,making more scientific credit decision mechanism. BP neural networks is a nonlinear self-adaptive dynamic system, which simulates human s neural system structure. It is composed of a lot of collateral neural elements which have the of learning, memorizing, computing and intellectual handling. Generally include one input layer, connotative layers and one output layer. Each node between two close layers joins each other in single direction. In order to forecast the Credit BP neural network and to avoid the shortcoming of traditional methods, BP Algorithm that is improved is adopted to settle the problem in this thesis. Optimization methods are also introduced. It has already proved that using one input layer, many connotative layers and one output layer, it can realize the mapping from arbitrary M dimension to N dimension. So, in the neural networks algorithm of the bank's credit evaluation we generally choose three layers. Structure of three -layers neural 429
networks is shown in figure 1. I1 I2 Output I3 I4 Figure 1. Structure of three -layers neural networks The connotative layer node and the output layer node's transfer function uses the Sigmoid function: 1 ƒ(x)=. 1 e x The application of BP neural networks is the most extensive at present; BP neural network has more self-adaptive capacity. The application including training and testing two stepping course. The purpose of network training is to find a set of weights, and makes it minimum. In three BP neural networks, what is quite big to the network performance influence is the weight correction method, uses the following method revision weight: E Djk Djk Djk Djk D jk E Dij Dij Dij Dij D In the equation, (0< <1)is a positive constant, called studying rate,which reflects the adjustment speed of the weight.if is too small, the efficiency of study is relatively low. Conversely, if is too large, it may cause oscillation. For this reason, we introduce the momentum. Doing this can strain the high-frequency deviation of error-curved surface in weight space, and then make the interval of effective weight strengthen. Under the normal circumstances, while the momentum can reduce shaking, it makes the restraining speed of algorithm faster. 2.2. Construction of BP neural networks model The customer degree of comparison evaluation is the credit management foundational work. The evaluation content take credit capacity as a core, overall evaluation profit, business, factors and so on management, after the synthesis evaluation, obtains the customer degrees of comparison, establishes including the enterprise operational, management, the financial and the credit record early signal system, sets each target the marginal value, when some target tends a critical point, sounds can warn promptly. (1) Designing the input layer. The BP neural networks can only deal with the numerical data[6], so the quantitative index need be standardized and the qualitative index is quantized and standardized, generally in the limitative scope [0,1]. The nodes in the input layer correspondent to 14 indexes of 4 types in small and medium-sized enterprise credit evaluation. (2) Robert Hecht-Nielson proves that one BP networks with implicit layer can approach continuous function in the close block in 1989; therefore this paper contains an implicit choice of the three-layer BP neural network. The nodes number of the connotative layer is confirmed as 3 in experimentation. The number is the best when the total error of the system is minimum. While the number is confirmed as 3, with passing the test, the total error of the system is minimum. (3) Designing the output layer. The output node is 1.The police degree of early system has four kind of situations: normal, low-, medium and high-. It is the status output. Output is ij 430
normal (1000), low- (0100), Medium Risk Alert (0010), high- (0001). 3. Decision system of credit The early is a key link for credit control[8]. Early of the commercial bank credit is a process that the mode categorize: from the mapping relation along promise alert index alert feeling index and alert degree. Economic early is the course of a approximation of function; From the noise along Promise alert index alert feeling index and alert degree and Handling in the way with calling the police accurately, Economic early-- is the optimum course. So, a BP neural network that applies to early of the commercial bank credit is suitable. BP neural networks is composed of input layer connotative layer output layer. Input layer corresponding to alert index of promising, connotative layer corresponding to alert feeling index, output layer corresponding to alert degree. 3.1. Alert indexes The target early method is the commonly used early method. Alert index can be deemed to finance rate of input layer. Finance rate, an important aspect in quantities research on the corporation credit, is the main portion about the evaluation of the corporation credit. Moreover, credit is the key. Corporation credit evaluation system is mainly considered into four aspects that are the to refund, to profit, to operate and manage, developmental and potential, through the credit rating the credit transfers guarded against in anticipation before from supervised afterward. Especially, a bank will pay more attention to it about the valuation of the corporation credit. A company finance condition system that adapts to loaning has been utilized. In order to cope with the imprecise analysis and judgment and to make the operation easy. The system include four finance portions the to refund, to profit, to operate and manage. The following table 1 make a illustration about them in detail: Table 1. Alert index Criterion Index name Calculate formulae The to refund The to profit Management and administration asset-li ratio liquidity ratio rapidly ratio cash ratio debt ratio of cash in business activities current cash ratio rate of main business profit and financial expenses main business profit ratio net assets profit ratio assets profit ratio Turnover rate of the account receivable turnover rate of stock total debt/total asset*100% current assets/ current li*100% rapidly asset/ current li*100% cash assets/ current li*100% current cash in bushiness activities/total debt*100% total of cash net flux of management, investment and raise funds main business profit/ financial expenses*100% main business profit/ main business income*100% net profit/average net assets*100% total profit/total assets*100% main business income/average remaining of account receivable*100% cost of main business/average stock*100% Developmental and potential increased income ratio of main business rate of increase of the profit main business income in this issue/ main business income in last issue*100% (net profit of this issue-net profit of last issue)/net profit of last issue*100% 431
3.2. Early BP neural networks apply on loan early, which can make good deal with slowly changing information, and have great study and fault-tolerant. At first, this text carry on discussion with BP neural networks to early mechanism, showing that early system can say with BP model of 3 layers. Correspond to early rules of the index with BP model of 3 layers, And then to the finance early signal, the extraction credit capacity, profit, the management and operation, development and potentiality 4 aspects alert index correspond to input layer to BP,the alert index correspond to implies layer, alert degree is exported layer correspondingly. Alert degree is divided into the normal condition, the low, the middle, the high,as the output of exports layer. Node of connotative-layers act as n=sqrt(m+k)+a (a is a constant between 0 and 10) according to experience,means train and Test with the topological structure of 4 3 4. The financial crisis grade regard as similar separate question,for the good moral character of ANN,which has correct rate of classification analyzed more than common discrimination. 4. Experiment design According to the data offered by a commercial bank and feature sample with credit, the to refund, to profit, to operate and manage, developmental and potential can make the input of the networks input layer as four alert index. The alert can be divided into four scales, respectively are normal state, low, mid, high--, and all of them can be used as the input of the network input layer. The data that were inputted have been changed into experiment data. The experiment data are in the following table (table 2): Table 2. Experiment data Alert Enterprise The The Management and Developmental and degree number to refund to profit administration potential 1 0.51 0.32 0.56 0.41 Normal state Low Mid High-- 2 0.49 0.36 0.41 0.52 3 0.52 0.40 0.53 0.47 4 0.50 0.39 0.48 0.49 1 0.40 0.30 0.41 0.39 2 0.37 0.28 0.39 0.37 3 0.35 0.26 0.36 0.38 4 0.39 0.31 0.41 0.35 1 0.32 0.25 0.35 0.22 2 0.30 0.20 0.29 0.28 3 0.29 0.19 0.31 0.33 4 0.31 0.22 0.32 0.29 1 0.19 0.19 0.30 0.19 2 0.15 0.15 0.21 0.21 3 0.09 0.14 0.19 0.10 4 0.10 0.17 0.18 0.21 With the help of the toolbox of the Matlab software, the objective error is in the range of the demand. It is illustrated in detail in the following figure 2. 432
Figure 2. Simulation result In order to enable the network model has better to identify the difference, each additional sample will be retrained again. The three-layer BP network optimization model is illustrated in the Figure 1. The error which are used to make credit forecast is less than what is expected by using Matlab learning procedures. As is illustrated in the Figure, emulation trainings suggest: (1) If training samples are the same, the quantity of the connotative layer unit selected can generate a direct influence on network performance. The more connotative layer units are, the more accurate the system's prediction is, and the higher connotative layer units are. (2) If the number of connotative layer units in the same circumstances, the larger the number of training samples are, the better prediction performance BP neural network shows. With the parallel illation for the experiment design, the result can be obtained, using the Matlab neural network toolbox and the bank credit early procedures based on BP neural network, a better credit of early s is available. Furthermore, it can offer the result of an early ; early findings are showed in table 3. 5. Conclusions From the angle of theory combines with practice, this article analyze the profession and causation of loans in bank. And it get the mostly conclusion: (1) Back-propagation neural networks were applied in loans evaluation. It conquered the difficulty of making certain weight, and the character of non-linear was incarnated. In addition, the restrain the speed BP neural networks became faster. (2) Fault-tolerant. Because the network knowledge information adopts the distributional memory, the individual unit s damage cannot cause the output mistake. The fault-tolerant is stronger in the course of prediction and identifiably, and more credit. (3) Because of the limitations of condition and time, I can t consult and collect the sample data of large city and four commercial banks. It must have some limitation. 433
Table 3. Test results Enterprise number The to refund The to profit Management and administration Developmental and potential Deduce the result Alert degree 1 0.516 0.372 0.501 0.492 (1 0 0 0) 2 0.381 0.287 0.396 0.385 (0 1 0 0) 3 0.273 0.214 0.342 0.287 (0 0 1 0) 4 0.133 0.117 0.229 0.176 (0 0 0 1) Normal condition The low The middle The high 6. Acknowledgement The author would like to thank the anonymous reviewers for their valuable suggestion and comment on this work. Supported by The Shanxi province social science Joint Fund (SSKLZDKT2011055): On travel industry of Shanxi exploiting capital market research and the Shanxi Province philosophy and social science fund(2012280):research on the regional capital markets to establish and resources transformation in Shanxi. 7. References [1] Piramuthu S., Financial Credit Risk Evaluation With Neural and Neurofuzzy Systems, European Journal of Operational Research, Vol. 112, No. 2, pp. 310 321, 1999. [2] LI Rong-zhou, PANG SU-lin, XU Jian-min, Neural network credit- evaluation model based on back-propagation algorithm, Machine Learning and Cybernetics, Vol. 4, No. 3, pp.1702-1706, 2002. [3] Peltonen T, An application of panel estimation methods and artificial neural networks, Italia: European University Institute, 2002. [4] Chai Binghua, Liao Ningfang, Artificial neural networks performing the forward operation of color appearance model, Journal of Beijing Institute of Technology, pp. 54-57, 2003. [5] Daozhu Xu, Haiyan Liu, An Improved Algorithm for Creation of Delaunay Triangulation,Geomatic and spatial Information Technology, Vol. 30, No. 1, pp. 38-41, 2007. [6] Cairong Wu, Huaxing Huang, "Evaluation and Research on Sports Psychology based on BP Neural Network Model", AISS, Vol. 4, No. 10, pp. 355 ~ 363, 2012. [7] Shi Hui-bin, Li Hong, Liu Lu, Wang Li, "A credit assessment system for the small and medium enterprises in China", JCIT, Vol. 7, No. 2, pp. 277 ~ 284, 2012. [8] Baosen Wang, Xiaojun Ma, Yunfeng Cui, "Study on Risk Management of Commercial Bank Card of China", JCIT, Vol. 6, No. 9, pp. 186 ~ 191, 2011. 434