Font Size: a A A

Personal Credit Evaluation In Consumer Lending Based On Support Vector Machine

Posted on:2006-04-23Degree:DoctorType:Dissertation
Country:ChinaCandidate:C H ShenFull Text:PDF
GTID:1116360152981117Subject:Management Science and Engineering
Abstract/Summary:PDF Full Text Request
In this paper we do some reseachers on Support Vector Machine(SVM) by the optimization theory.We point out some basic drawbacks in SVM's apllication into personal credir evaluation and make some improvements to it; and We build a personal credit system on SVM to research evaluating consumer credit deeply in three levels based on the analysis of the exsiting methods of personal credir evaluation; SVM's application is generalized. Meanwhile, in order to apply theory to reality, methods are tested through data experiment in practice and get good results. The main works in this paper are followed:1. According to the fact that the standard SVM doesn't deal with different cost data, a model of SVM based on different costs is presented ,the research on this model from optimization point of view is at the beginning in the world .for the first time.2. The fomula are proposed for any cases to compute the thresholds of decision functions in this new model.3. The methods of selecting the parameters in the new model are also given according to knowing the two class costs or not.4. Updating the output of the new modelto give every sample not only a class but also a probability belonging to a class.5. Some researches are done on data preparation for consumer credit evaluation. By using of information gain in information subject and ROC curve in signal detection theory, a method of selecting index for personal credit evaluation is constructed based on information gain and AUC for the first time. In combination single imputation of missing data with multiple imputation, a new missing data imputation—KNNMI is proposed. The methods are proved good through experiment, therefore bear very important reference value for commercial banks.6. The first level—Decision of 'Default or not'. Three different models are created for credit evaluation according to different pernal credit data. (1).A model based on SVM for credit evaluation is proposed. The assumption behind that is: the error prediction of 'bad' risks is equal to the risks of mis-classifying 'good'; (2). A new method-- NN-C-SVC-KNN is presented to improve the accuracy of SVM classifier according to the one class samples serious intermixed in another class;(3).Based on the fact that the importance of the two types of errors is obviously different in the credit evaluation problem, a new model based on different costs and SVM for credit evaluation is presented for the first time and make the results better.7. The second level—Caculating of 'Probability of Default(PD)'. On the basis of the first level a model of calculating PD based on different cost and SVM is proposed for the first time to to caculate PD. The research on credit evaluation is deeper.8. The third level-'Measure of Credit Risk'. On the basis of the first level and the second level , a concept of 'Measure of Credit Risk' is proposed based on the nature of credit risk (the probability ofloan becoming bad debt). A method of predicting 'Measure of Credit Risk' is designed based on Support Vector Regression (SVR) for the first time to predict the probability of loan becoming bad debt, so as to research personal credit evaluation overall.The above reseaches are not only theoretically significant in enriching the content of SVM , but also realistically applicable in pushing the building of our country's credit system forward, increasing the level of credit risk management of consumer credit lending of our country's financial institutions, reducing the not good loan ratio of consumer credit lending of our country's financial institutions, promoting the further development of consumer lending market of our contry.
Keywords/Search Tags:Support Vector Machine, Personal Credit Evaluation, Default or not, Probability of Default, Measure of Credit Risk
PDF Full Text Request
Related items