| This paper focuses on the commercial bank credit scoring models built to handle the loan application of SME and mainly considers the variable selection of the model. The variable selection of model is very important not only from the viewpoint of practical modeling, but also from statistical. There are some commonly actual used indicators screening methods in the model of credit scor-ing:Logistic stepwise regression method, information entropy method, hypoth-esis testing method, contingency table analysis, This paper briefly describes the principle, points out that these methods are almost for low-dimensional problems and have bad effect on the high-dimensional problems, so we intro-duce two methods that have good performance in high-dimensional statistical processing and their theoretical results have been rather perfect:Lasso method and Elastic Net method. Lasso method is proposed in1996by Tibshirani, Elas-tic Net method is an improved method of Lasso, proposed in2005by Hui Zou and Trevor Hastie. We will introduce the principles and algorithms of these two methods, and then use a simulation example to describe the place that the Elastic Net is superior to Lasso, and then focus on these two variable selection methods used in Logistic regression model which is commonly used model in credit scoring models. Then we mainly focus on the empirical analysis, use the actual data of some commercial bank to compare the two methods and the traditional stepwise regression method.The comparison principles will not only consider the quantitative validation containing the ability to distinguish risk and the estimated accuracy, but also analysis the result from the economic point. The final conclusion is that in the bank credit scoring models, under some condition using the Lasso method for variable selection is better than using the Elastic Net method and stepwise regression method, we can foresee the Lasso method will be very useful in the actual credit scoring. |
PDF Full Text Request