An Empirical Research On Personal Credit Evaluation Based On Fusion Model

With the rapid and steady development of China's economy,people's living standards are constantly improved and their consumption concepts are constantly upgraded,which has greatly promoted the development of China's consumer finance industry.However,in the past two years,the national regulatory authorities have gradually strengthened the supervision of the consumer finances,which has led many customers to flow back to banks.Thus,it provides new opportunities for commercial banks to develop retail business-credit cards.But faced with the fierce market competition and complicated domestic and foreign economic environment,how to evaluate personal credit more effectively and quickly has became one of the key issues that banks have focused on in recent years.At present,more and more scholars prefer to use fusion models to improve the predictive accuracy of personal credit evaluation,which has obtained some achievements.However,most of the studies directly specify the basis models in the fusion model.But the basis models should have some differences.For this problem,this paper builds a personal credit evaluation fusion model based on a credit default data set.The building process includes:data preprocessing,balancing positive and negative samples,variable screening,building single model,screening the basis model,building a fusion model and comparing the performance of different models.In the part of building model,firstly,this paper considers the predictive accuracy of models(AUC value and KS value)and the difference between models(Q statistic).Hereafter,two models—BP Neural Network and Decision Tree are selected as the basis models in the fusion model from five single models which are Logistic Regression,Decision tree,BP Neural Network,K-Nearest Neighbor and Support Vector Machine.Then,the two basis models are trained by 5-fold cross-validation and the predicted probability values of each sample are added to the original data set as new feature variables.So far the first layer learning of the fusion model is completed.In order to ensure the stability of the fusion model,this paper trains the Logistic Regression model based on the new combined data set,which completes the second layer learning of the fusion model.The empirical analysis shows that the predictive performance and stability of the fusion model built in this paper are superior to those single models and integrated models(XGBoost,Random Forest).Therefore,the process of building a fusion model has some technical reference for the research of personal credit evaluation.