Credit Scoring Model Based On Support Vector Machine And Particle Swarm Optimization

Credit risk has become one of the key risks faced by commercial banks, especially in credit cards, mortgages, auto loans and other consumer credit business, credit risk management is become one of the most important problem which need to be solved. The appropriate credit scoring model is one of the effective approaches to reduce the credit risk.Currently in the area of credit scoring, many scholars have made exploration and research using statistical learning, machine learning, data mining and artificial intelligence methods to build the credit scoring models. The accuracy of the credit scoring model related to bank some important decisions, so this thesis conducted an in-depth study on improving the accuracy of credit scoring. SVM has been rapidly developed since its publication and we applied it to the areas of credit scoring to solve classification problems. Nevertheless, when SVM is selected to train the model, two problems are faced. One is optimizing the parameters of kernel functions and the other is selecting feature subsets. This paper solves these both problems by particle swarm algorithm (PSO) which is a kind of swarm intelligence algorithms with a lot of advantages on complex optimization problems. In order to further improve the accuracy of the model classification, the study constructed new multiple kernel functions instead of SVM used RBF kernel, then adding chaos mechanism to enhance particles search capacity on the basis of retaining the original advantage of the PSO algorithm. The main contents of this thesis are listed as follows:Researching the parameter optimization and feature selection problems in SVM and focusing on the analysis of key parameters c,γ of RBF kernel function.Combined the single kernel function SVM and basic PSO algorithm to establish the hybrid credit scoring model. The classification results of the model are compared with the currently used model.Currently, most of kernel functions of SVM are single kernel functions. Based on proposed kernels, the study combines local and global kernel functions to build a new multiple kernel function. So the distant and near data sample points all can affect the value of the new kernel function.Based on the basis of PSO, we introduced chaos mechanism into it and proposed Chaos Particle Swarm Optimization algorithm (CPSO) to improve the optimization capabilities of PSO. CPSO overcomes the disadvantages of early elementary PSO in local optimization problems.The study proposed the hybrid credit scoring model MKSVM-CPSO based on multiple kernel functions of SVM and CPSO. Moreover, an empirical analysis of real credit datasets was made to test the model. And experiment results were compared with the classification results of existing models. Finally, statistical methods are used to illustrate the validation of the model.