Credit Evaluation Model Research Based On Mars-SVM

The importance of credit card for individual, enterprise and even national is self-evident. At present, the main challenges of the credit card is credit risk, in a general lack of credit management mechanism of Chinese this question seems particularly outstanding. How to effectively control risk, maximization in income, become a credit card issued institutions are facing a major problem. Therefore, from the theoretical level to practice level, the credit evaluation theory has its broad developing space. The credit evaluation purpose lies in using the existing customer attributes including social attributes and nature will credit card applicant is divided into two classes:for to better perform the obligation of reimbursement applicants is divided into "good customer", agree to issue credit CARDS. For possible default or refused to repayment of the applicant is divided into "bad customer", and refused to credit card application.Early credit evaluation depends on the empirical qualitative analysis, lack of efficiency and vulnerable to operating personnel’s subjective influence. For this reason, many experts and scholars try to design appropriate credit evaluation model for quantitative processing credit risk problem. Discriminant analysis and logistic regression is the most commonly used (parameters) statistical analysis method. With the development of computer technology, to data driven as the core thought of machine learning theory has become more and more popular, the decision tree and neural network, in the credit evaluation problems have achieved considerable success.Design a suitable credit evaluation model is the main research content, therefore, this article first introduces the significance of existence in the credit evaluation model. Secondly in literature review is given of the detailed establish credit evaluation model of each step as well as the current research status. And then the common credit model of carding points out the advantages and disadvantages of its existence.At last, it puts forward the MARS-SVM model, make full use of MARS global processing variables, and to the advantages of variable importance in order to make up for a SVM can not carry out the defects of feature selection, so as to get a higher forecasting ability of the hybrid model:MARS is the modern regression analysis method, the data distribution demand is not high, through the forward stepwise introducing variable, and gradually backward delete not important variable way to establish the regression model. So MARS to the importance of the variable sequence with global optimality. SVM using grid point search method and using cross validation way sure punishment parameters and kernel function parameters. So although its algorithm characteristics can avoid the "dimension disaster", but too much predictor variable will affect the work efficiency, MARS is the reasonable added. The core of the SVM part is the choice of kernel function, Rbf kernel function since it has universality, operability advantages of the most popular. But at the same time due to the Rbf may result in feature space sample information loss so in Rbf kernel function is put forward based on KOBF kernel function. This paper will use at the same time as KOBF and Rbf kernel function of SVM with contrast classification effect.In order to validate model MARS-SVM prediction ability, this paper made the contrast experiment. Using logistic regression, classification decision tree, and neural network for the same sample data set to do the classification processing. The results showed that MARS-SVM model has good prediction ability...