Research On Cox Credit Risk Measurement Based On SCAD Variable Selection

This paper focuses on the credit risk measurement model which established by commercial banks in dealing with the loan application of the real estate listed companies,mainly considers the variables selection and parameter estimation in the model.Based on the domestic and foreign research results,this paper uses the open financial data of real estate listed companies to establish the Cox credit risk measurement model.It contains a lot of redundant information Because of the high dimension and strong correlation of the selected indexes,So it is very important to select the Covariate variables which really affect the response variables to establish the Cox model.In this paper,the stepwise regression method of statistical learning theory and the coefficient compression methods in dealing with high-dimensional problems,such as Lasso penalty and SCAD penalty,are applied to the Cox model.In order to improve the interpretability and accuracy of the results,the coefficients of the response variables are less than 0,and only the coefficients with larger coefficients are preserved.This paper introduces the principle,the corresponding algorithm and the choice of adjustment parameters of three kinds of variable selection methods in Cox model.Under certain conditions,it is proved that the SCAD estimator has the property of Oracle.The simulation results show that the Cox model based on SCAD method can select the real model,and the effect is better than the traditional stepwise regression method and the Lasso method.In the application of the model,we selected the real estate listed companies in China's Shanghai and Shenzhen A shares 2005-2015 as samples.Among them,ST 33,non ST company 88.Taking into account the availability of data,we select 28 financial indicators as the initial variables.According to the relevant data of the sample T-3,establish the traditional regression based on Cox model and Lasso method based on Cox model and SCAD Cox model to measure the punishment based on credit risk,through the comparison of three kinds of model prediction accuracy,prediction accuracy of SCAD penalized likelihood method is the highest,reaching 84.2%,a a certain reference value.Finally,the qualitative test of the model,verify the three methods of risk distinguish ability by using the ROC curve,the results show that the SCAD method based on Cox model is better than the stepwise regression method and the Lasso method of the Cox model.