Personal Credit Risk Rating Based On Cost-sensitive Bayesian Classification

With the credit consumption concept gradually entered people’s daily life, the demand of credit risk rating of individual is more and more urgent, while personal credit risk rating of domestic commercial bank is still in the initial stage. At present, because of the commercial secrets of bank and the protection of customer privacy information or some other reasons, most of the studies were based on the foreign public test credit data, such as Germany, Australia and Britain. But the credit consumption behavior and people’s life style in our country are different from people abroad. Though classification models have good predictions on foreign country data, it does not mean that those models have the same performance in our country. This paper is to build models based on the real credit data of domestic commercial bank, and come to conclusions which can be applied to domestic commercial banks.In this paper, the experimental data comes from the sampling of the credit card customer’s credit data of Guangzhou bank. In the progress of data preprocessing, this paper will use the optimal segmentation method to discrete data, which combined the theory of decision tree. The analysis of the sample data will prove that, even if the traditional equidistant segmentation method perform well in the logistic regression model, but the optimal segmentation method has better predictability in the Bayesian classification.A misjudgment must exist in classification problems, and different fault will cause different costs of loss. In the problem of personal credit risk rating, a "normal" customer is mistaken for a "default" customer and a "default" customer is mistaken for a "normal" customer will come to different costs of loss. Although the former will lose some benefits, but the risk of loss the latter brought is immeasurable. This paper first introduce the concept of cost sensitive into personal credit risk rating problem, and construct the cost function and risk function. This paper will determine the optimal parameters of the cost function through lots of experiments, and establish the cost sensitive Naive Bayesian Model and the cost sensitive TAN Model. The experimental results show that two models can greatly reduce the probability of a "default" customer mistaken as a "normal" customer on condition that the correct rate was declined slightly.