PoS(ISCC 2017)020. Credit Risk Assessment of Receivable Accounts in Industry Chain based on SVM. Speaker. Huan Sun 1

Transcription

1 Credit Risk Assessment of Receivabe Accounts in Industry Chain based on SVM 1 Schoo of computer and information, Hohhot Vocationa Coege Inner Mongoia, , China E-mai: sunhhvc@163.com Industria chain finance is a financia innovation that can effectivey integrate the financia needs of the upstream and downstream enterprises. The rapid deveopment of industry chain finance has put forward higher requirements for corresponding risk management. Among them, how to evauate the credit risk of industria chain receivabe accounts is a key ink of industria chain financia risk management. For the sma sampe size in a singe industry chain, a support vector machine (SVM) suitabe for sma sampe earning is introduced to assess the credit risk. The credit risk evauation index system for the industria chain accounts receivabe is constructed by principa component anaysis of the financia data. V-SVR (Support Vector Regression) mode is empoyed to evauate principa component of the receivabe accounts in teecom industry chain. By comparing the evauation resuts with historica data, it is found that the mode can we predict the credit risk of commercia banks, and has higher prediction accuracy than Logistic mode. ISCC December 2016 Guangzhou, China 1 Speaker Copyright owned by the author(s) under the terms of the Creative Commons Attribution-NonCommercia-NoDerivatives 4.0 Internationa License (CC BY-NC-ND 4.0).

2 1.Introduction Industria chain Finance is an extension and expansion of suppy chain finance. Suppy chain finance, simpy put, is to connect the core enterprises with upstream and downstream enterprises to provide a financing mode for the fexibe use of financia products and services. The industria chain finance is not ony imited to the horizonta upstream and downstream of the various inks, design standardization, personaized financia services products, but aso the vertica reorganization of mergers and acquisitions and other financia activities. For the enterprises in the industria chain, it is hepfu to improve the overa operation efficiency of the industria chain and promote the transformation and upgrading of the industry because of obtaining more abundant financia support. The estabishment and vigorous deveopment of industria chain finance has aroused extensive attention of many schoars. For exampe, some researchers have focused on an important branch of the industry chain Finance, suppy chain finance, from the suppy chain finance strategy [1], financia optimization [2] and suppy chain financia management [3], and so on. Of course, the deveopment of the industry chain Finance aso put forward the higher request to the corresponding risk management technoogy, among them, the credit risk assessment of the industria chain receivabes is an important probem that is needed to be soved before the financia institution. At present, the most infuentia credit risk measurement mode in the word has KMV Company s KMV mode, J. P Morgan's metrics mode and the McKinsey Portfoio view mode. However, the number of the sampes of the industria chain accounts receivabe is reativey sma, and the reated credit status of the effective data oss, resuting in these modes cannot be used in China's commercia banks credit risk assessment of the fied of effective appication. In view of the above characteristics of the credit risk assessment of the industria chain receivabes, this paper wi try to evauate the sma sampe earning agorithm using support vector machine. Since the support vector machine agorithm is suitabe for the cassification of sma sampe probems and has high precision, it has been favored by researchers at home and abroad, and some schoars have graduay appied it to the fied of credit risk assessment of commercia banks since its reated theories were founded. For exampe, for the Turkish commercia banking system, Meek et a. respectivey use neura network technoogy, support vector machine and mutivariate statistica methods to predict the possibiity of bank insovency, the different financia indicators to the data set to improve the predictive performance. The resuts show that the SVM mode can be regarded as the most successfu mode to predict the financia bankruptcy of commercia banks [4]. Support vector regression machine has been appied in some fieds of financia research and has high prediction precision. Harris compared support vector machine (SVM) based credit-scoring modes buit using Broad and Narrow defaut definitions, and the resuts of empirica reveaed that credit risk evauation at the Barbados based institution coud be improved if quantitative credit risk modes are used as opposed to the current judgmenta approach.[5]. After neary ten years of deveopment, support vector regression machine has been derived from many types, incuding ε-svr, C-SVR, v-svr and so on. Among them, v-svr is a kind of support vector machine with high accuracy, and it has been appied graduay in financia fied. In the prediction of gas price, by using the ε-svr, v-svr mode and artificia neura network mode, RBF network mode and asymmetric GARCH 2

3 mode, Hu and Trafais found that the prediction accuracy of v-svr mode with sigmoid kerne function is highest [6]. Therefore, aiming at the measurement of the credit risk of receivabes accounts in industria chain, it is necessary to forecast the indexes with high accuracy, and the sampe quantity is few, this paper attempts to use the v-svr mode to estimate the credit risk of the industria chain receivabes, and compare the resut with the forecast resut of Logistic mode to verify the predictive effect of the mode. 2.Support Vector Machine Mode Support vector machine (Support Vector Machine, SVM) [7] is a theory deveoped on the basis of statistica earning theory, which speciaizes in the study of machine earning in sma sampes, and is an important deveopment and suppement of traditiona statistics. 2.1Support Vector Cassifier Agorithm On the basis of the cassification method, Vapnik [8] proposed ε-support Vector Regression. He added and subtracted ε from y vaue of each training point in training set T, and obtained two sets of positive and negative points respectivey. Thus the hard ε-band hyperpane is equivaent to the hyperpane whose training set is correcty partitioned competey. The regression probem can be transformed into a cassification probem. On the basis of ε-svr, Schökopf et a. [9] proposed the ν-svr mode, and improved the support vector regression machine. The specific agorithm is as foows: Step 1: Given training set T ={( x 1, y 1 ),,( x, y )} ( R n R), where x i R n, y i R, i=1,2,, ; Step 2: Seect the appropriate kerne function K ( x, x' ), parameter v (0,1] as we as C>0; Step 3: Construct and sove the convex quadratic programming of optimization probems: 1 min α 2 ( ) i=1 s.t. i=1 j=1 (α s i )(α s j α j ) K ( x i, x j ) (α s j α j ) y j, j =1 (α s i )=0,0 α ( ) C /,i=1,2,,, (α s i +α i ) Cv (2.1) get soution: α ( ) =(α 1, α 1 s,, α, α s ) ; Step 4: Cacuate b : seect the two components α j and α k of α s in the open interva. Accordingy, b= 1 2 [ y + y [ i k (α s i )K (x i, x j )+(α s i ) K ( x i, x k )]]. (2.2) i=1 Step 5: Construct decision function g( x)= (α s i )K (x i, x)+b. (2.3) i =1 The kerne function K ( x, x ') transformed the noninearity into a inear cassification in high dimensiona space, so the kerne function pays an important roe in the noninear cassification. Considering the randomness of training sampes, the training sampes are non- i=1 3

4 ineary separate, thus using an inner kerne function agorithm. A Gauss radia base function (RBF) is seected in this paper, namey: K ( x, x ')=exp( x x' 2 /σ 2 ). (2.4) Compared with the traditiona ε-svr, ν-svr uses a quantitativey meaningfu parameters ν to repace the parameter ε, where ν is the upper bound of the percentage of the error training points and the ower bound of the number percentage of support vector machines. In addition, under certain conditions, when the number of training points + in training set T, v is cose to the ratio of the number of support vectors to the number of training points as the probabiity of 1. This design aows the mode to be optimized and to achieve higher computationa accuracy. Therefore, this paper mainy focuses on the training in ν-svr mode to test its effect in the credit risk forecast of the industry chain accounts receivabes. 2.2 Input and Output Indexes Leve 1 indicators Leve 2 indicators Leve 3 indicators Enterprise management status Basic situation of Enterprise scae Basic quaity of enterprises enterprises Enterprise status Financia discosure quaity ROE Profitabiity Ratio of income as a percentage of saes Operating capacity Turnover capacity Short term debt service Quick ratio Financia situation Interest protection mutipes of company Long term debt service Asset-iabiity ratio Long-term asset fitness ratio Deveopment potentia Saes revenue growth rate Net profit growth rate Tota assets growth rate Accounts receivabe Accounts receivabe turnover ratio Assets under Stock Inventory turnover ratio financing Prepayment turnover ratio Prepayment turnover ratio Basic situation of upstream Industria chain management enterprises Suppy chain Operation of ogistics operation condition Process risk enterprises Compiance Industria chain compiance risk Macro environment Macroeconomic situation Lega poicy environment Industry status Industry deveopment stage Industry deveopment Industry competition intensity prospect Industry environment Tabe 1: Input Index The indicators incude quaitative indicators and quantitative indicators, in which the quantification of quaitative indicators is divided into 0-5 vaues according to the five-eve quantification standard and the specific vaues are determined according to the reevant infuencing factors. 4

5 According to the principe of systematicness and comprehensiveness, fexibiity and practicabiity, conciseness and scientificaness and so on, the paper anayzes the infuencing factors of bank credit risk, and then draws on the reevant indexes of iterature at home and abroad. This paper constructs the evauation system of credit risk index of accounts receivabe in industria chain, seecting 14 financia indicators which are cosey reated to the credit risk as mode input indicators. The contents of the input indicators and their contents are shown in Tabe 1. Since the data units in this paper are not uniform, the data is normaized first: in the sampe data, the maximum vaue is max {x i }, the minimum vaue is min {x i }. The origina data is normaized as new data: x i =(x i min {x i })/(max {x i min {x i }). (2.5) After the data is normaized, the new data can be used as the input data of the v-svr mode, and the risk eve of the sma and medium-sized enterprises is the output of the mode. In order to avoid the high correation and the overapping of information between the financia indicators incuding bank credit risk information, and taking into account the efficiency of the support vector regression machine in training, this paper uses the principa component anaysis. Based on the bad debt provision of accounts receivabe as the output of the measure of Y, y {0,1 }. Y=0 means the company s bad debts are sma. To a certain extent, it refects ess bad debt reserves in accounts receivabe, which means the smaer risk eve. Here are the credibe enterprises. The same way, Y = 1 indicating that the company's bad debts, to a certain extent, refects more accounts receivabe need for bad debts, that is, high eve of risk. Here are "incredibe" enterprises. 3.Support Vector Machine Mode 3.1Data Preprocessing and Principa Component Anaysis Component Initia Eigenvauesa Extraction Sums of Squared Loadings Tota % of Variance Cumuative % Tota % of Variance Tabe2: Eigenvaue and variance contribution rate In order to evauate the credit risk of the receivabes in the industria chain more accuratey, this paper takes the sma and medium-sized enterprises with more frequent accounts receivabe in China's eectronic industry as the research object. Sampes are from isted sma and medium-sized enterprises whose market vaue is under 20 biion in 2016 eectronics industry The data of 75 eigibe sma and medium-sized enterprises is from the company discosed annua statements, JUYUAN database and TONGHUASHUN software. 5

6 The cumuative contribution rate of the extracting main component is up to 80%. According to it, so many origina dimensions of the indicators are reduced to a sma number of comprehensive indicators. The resuts of principa component anaysis of financia indicators using SPSS 17.0 software are as foows: Considering the principe that the eigenvaues are more than 1 and the cumuative contribution rate is more than 80%, the first eight factors are seected as the principa components, and the accumuated information reaches 83.6%. 3.2Logistic Regression Anaysis and Resuts First, the ogistic regression mode is used to evauate the credit risk of accounts receivabe so as to compare with the cassification effect of v-svr mode. From the above, this paper uses the principa component anaysis method to dea with the 16 indexes reated to the credit risk of the industria chain financia receivabes, and seects the first eight factors as the main component. Furthermore, Logistic regression is performed on the eight principa components. The regression method uses forward stepwise regression (Wad method). Finay the first main factor enters the equation, the others are removed. B S.E. WALD df Sig Exp(B) FAC1_ Step1a Constant a. The variabe entered on step 1:[%1:, 1:] Tabe 3: Variabes in equations The resuts of the ogistic regression avaiabe from Tabe 3 are as foows: where P= ea 1+e A, A= x x x x x x x x x x x x x x x x x x 18. Observed vaues Predictive vaue y 0 1 Step 1 y Overa percentage 77.3 Correct percentage a. The spit vaue is 500 Tabe 4: Logistic regression resuts cassification According to the above formua, the cassification resuts cacuated by SPSS software are shown in Tabe 4. The accuracy rate of the ogistic regression mode is 77.3%, and the accuracy rate of identifying "incredibe" enterprises of the sma and medium-sized in the eectronic 6

7 industry is 96.6%, and the accuracy rate of identifying credibe" enterprises is 11.8%; that is, the first type of error rate of 3.4% and the second type of error rate of 88.2%. Among them, the first type of error refers to the judgment of "incredibe" sma and medium financing enterprises (1) as "credibe " enterprises (0). The second type of error refers to the judgment of "credibe" sma and medium financing enterprises (0) as "incredibe " enterprises (1). 3.3 Anaysis of v-svr Mode Predicted Resuts In this paper, Matab2016a software is used for experimenta anaysis, and the v-svr mode is trained in an SVM toobox. Considering training accuracy and time, the parameters C and v in the main modes of the v-svr mode and the parameter σ in the kerne function are optimized. Appropriate precision is seected, and traversa searches are performed respectivey at the steps of 0.1, 0.01, and 0.01, so as to achieve the best training resuts. A sampe set is constructed according to the foregoing anaysis; x R 8 is an input index, that is, an eight-dimensiona vector; and y R is an output index of the sampe, that is, provision for bad debts of accounting receivabe. Eight principa components Factot1- Factot8, which are cacuated by using the principa component anaysis are used as input variabes of the v-svr mode. 0 is used to indicate that sma and medium-sized financing enterprises in the eectronics fied beong to the Trusted credit risk customer type, 1 is used to indicate that the sma and medium-sized financing enterprises beong to the Untrusted credit risk customer type, and {0,1} is the output variabe of the mode. The v-svr mode derives from the machine earning method in the statistica earning theories, and the method needs to divide the sampes into two categories, that is, the training sampe and the test sampe. First, the training sampe is input and output, and the mode automaticay earn and get the aw. Then, the aw is tested during test sampe input and output, so as to determine whether the resuts of mode training can be we promoted, that is, generaization capabiity of mode. In order to avoid the occurrence of contingency resuts, this paper draws the two categories of sampes ten times by random seection, and obtains 10 groups of different training sampe sets and test sampe sets. The sampe set distribution and the test resuts of the ten tests are as foows respectivey: sampes Sampe sets Training Scae of the sampe sets Number of the Trusted sampes Test sampes A sampes Tabe 5: Distribution of the first group of sampe sets Number of the Untrusted sampes The distribution of the other nine sampe sets is the same. Sampe number Entire Training sampes Test sampes sampes TypeⅠ TypeⅡ TypeⅠ TypeⅡ TypeⅠ 7

8 1 3.5% 8.3% 0% 0% 2.8% 2 0% 0% 4.7% 20% 2.8% 3 0% 0% 0.9% 11.1% 2.8% 4 0% 0% 2.3% 14.3% 2.8% 5 1.4% 11.1% 0% 0% 2.8% 6 0% 0% 4.6% 25% 2.8% 7 1.8% 7.1% 0% 0% 2.8% 8 0% 0% 3.2% 14.3% 2.8% 9 2.5% 7.7% 0% 0% 2.8% 10 0% 0% 4.1% 20% 2.8% Average 2.3% 3.42% 3.3% 10.47% 2.8% Tabe 6: Test resuts of the ten sampes It can be seen from the tabe above that in the entire sampes, the error rate of Type Ⅰ is 2.8%, the error rate of Type Ⅱ is 5.9%. This means the accuracy rate of distinguishing the Untrusted sma and medium-sized financing enterprises is 100% and the accuracy rate of distinguishing the Trusted sma and medium-sized financing enterprises is 94.1%, which shows that the mode is remarkaby effective. The empirica resuts indicate a high accuracy rate of the credit risk assessment that is performed, by using the PCA-SVM mode, on the financia accounting receivabe in industria chain of sma and medium-sized financing enterprises. 3.4Comparison Between v-svr and Logistic Regression Resuts The resuts of the credit risk assessment performed on the accounting receivabe by using the two modes are compared as foows: Entire sampes Entire sampes Industria chain assessment Accuracy rate of overa method The first type of error rate cassification Logistic regression 77.3% 3.4% v-svr mode 94.1% 2.8% Tabe 7: Comparison between the two modes As can be seen from the tabe above, under the index system of credit risk assessment of financia accounting receivabe in the industria chain, the accuracy rate of the overa cassification of the entire sampes by using Logistic regression is as high as 77.3%, with the first type of error rate of 3.4% and the second type of error rate of 88.2%, and the overa cassification of the entire sampes by using the SVM mode has an accuracy rate of 94.1%, with the first type of error rate of 2.8% and the second type of error rate of 29%. Therefore, the SVM mode is superior to the Logistic mode, both from the overa prediction accuracy rate and from the error rates of the two types. For financia institutions such as trust, the first type of cassification errors is more important. It can be seen from this paper that the first type of error rate in a v-svm evauation mode is ower than the first type of error rate in a Logistic regression mode. Therefore, the v- SVM evauation mode more meet the requirements of trust and other financia institutions to forecast the credit risk of sma and medium-sized financing enterprises. 4.Concusion Support Vector Machine (SVM) is a genera earning agorithm based on sma sampe earning theory. It has a rigorous theoretica basis and can sove practica probems such as sma sampes, non-inearity, high dimension and oca minimum points. After years of deveopment, 8

9 the support vector machine has got various improved forms, and v-svr is a kind of support vector regression machine with high prediction accuracy. In view of the ow number of isted companies in the teecom industry, this paper introduces the v-svr to assess the credit risk of the accounting receivabe in the industria chain. The assessment index system of credit risk of the accounting receivabe is estabished by using the principa component anaysis, and the credit risk assessment mode of the credit risk of the accounting receivabe in the industria chain is constructed based on v-svr. Compared with the Logistic mode, it is found that the ν- SVR mode has higher accuracy in training and prediction. Therefore, the credit risk assessment mode of the accounting receivabe in the industria chain based on ν-svr provides a more usefu auxiiary too for the credit risk management of commercia banks, and can be promoted to more areas for risk assessment. References [1] M Christopher, L Ryas. Suppy chain strategy: its impact on sharehoder vaue[j]. The Internationa Journa of Logistics Management, 1999, 10(1): [2] H C Pfoh, M Gomm. Suppy chain finance: optimizing financia fows in suppy chains[j]. Logistics Research, 2009, 1(3-4): [3] M L Gomm. Suppy chain finance: appying finance theory to suppy chain management to enhance finance in suppy chains[j]. Internationa Journa of Logistics: Research and Appications, 2010, 13(2): [4] A B Meek, K Yakup, K B Omer. Predicting bank financia faiures using neura networks, support vector machines and mutivariate statistica methods[j]. Expert Systems with Appications, 2009, 36(2): [5] T Harris. Quantitative credit risk assessment using support vector machines: Broad versus Narrow defaut definitions[j]. Expert Systems with Appications, 2013, 40(11): [6]Y Hu, T B Trafais. New kerne methods for asset pricing: appication to natura gas price prediction[j]. Internationa Journa of Financia Markets and Derivatives, 2011, 2(1-2): [7] C Cortes, V Vapnik. Support-vector networks[j]. Machine Learning, 1995, 20(3): [8]V N Vapnik. The Nature of Statistica Learning Theory. Springer-Verag, New York, [9] B Schökopf, R C Wiiamson, P L Bartett. New support vector agorithms[j]. Neura Computation, 2000, 12(5):