The application and assessment of consumer credit scoring models in measuring consumer loan issuing risk of commercial banks in China.
详细信息
文摘
With the impressive growth of economy,various consumer credit businesses have been expanding in China rapidly during the past two decades. However,due to the low level of risk management methods and relatively laggard techniques being currently used by Chinas commercial banks,and because there was no mature credit risk control system based on credit scoring models in China,the further development of consumer credit business is seriously hindered. In order to solve this problem,consumer credit scoring,which is a predictive modeling method widely used to predict the probability of customer default,should be used when the commercial banks making decisions in granting credit to individuals. During the last 5 years,China has devoted to building the worlds largest personal credit database and established its own personal credit scoring system. But this system still falls behind those system in developed countries such as the U.S.A and needs to be improved. Several scholars in China have tried applying traditional consumer credit scoring techniques to the data in China,among which they found logistic regression was the most robust and neural network had the highest prediction accuracy. In recent years,three new and more advanced credit scoring methods: Support Vector Machine SVM),Random Forest RF) and Partial Least Squares Regression PLSR) came into use in the industry in the U.S.A and have been proved to have higher prediction accuracy and stability. In this dissertation,all the three latest modeling techniques have been introduced and applied to a dataset obtained from one commercial bank in China. This dataset with Chinese characteristic variables and large proportion of missing values has been carefully analyzed and handled. For the purpose of comparison,three typical and traditional models: logistic regression,classification tree and multi-layer-perceptron neural network models have also been built with the same dataset. Each of the six models has been assessed and tested respectively. A comparison based on various criteria: classification error rate,Kolmogorov Smirnov KS) test statistic and Area Under Curve AUC) value has been made between six models. The result of comparison has shown that the Random Forest model outperformed other five models no matter which criterion was applied. Thus Random Forest is a method recommended to be implemented by Chinas commercial banks in the future.