Research On Post-loan Risk Early Warning Of Commercial Banks’ Credit Cards

According to the analysis of various public data since 2015-2021,at present,China’s consumer finance industry has shifted from the incremental era to the stock era.Banks and consumer financial institutions need to use their own services to "increase and save",and also need risk control to avoid losses and find high-quality customers.In this ecological pattern,compared with the consumer finance platform with Internet empowerment,banks are faced with the disadvantage of lack of user portraits and behavioral data of third-party platforms,and need to face higher challenges to achieve high efficiency of risk control and other businesses.In this context,this study is mainly about the post-loan risk early warning research of credit card carried out according to the business needs during the internship of a provincial branch of a commercial bank,and the follow-up in-depth research in the bank test environment due to personal interest,and finally online in the formal environment,and provide model help for the follow-up post-loan strategy personnel.This paper mainly starts from the actual business during the internship period,takes the customers whose credit cards are overdue for one month as the target customer group,selects January to April 2022 as the performance period,selects 25 initial variables(including derivative variables)from seven internal data tables as the characteristic variables,and predicts the final repayment intention of customers who are overdue for one month and the influencing factors through the LSR-LRboost model designed in this paper.LSR-LRboost model mainly includes feature engineering(LASSO feature variable selection,balanced data set optimization generation based on SMOTE and ROSE algorithm,WOE feature variable coding,Ⅳ value feature variable selection and correlation analysis)and model prediction(scorecard model,Xgboost model,LightGBM model and Logistic-Xgboost combined model comparative analysis).The research results of this paper are as follows:In terms of LSR-LRboost model prediction,Xgboost model has the highest AUC value,accuracy and accuracy of 0.948,0.87 and 0.87 respectively,which means that the overall resolution of the model is the strongest.The combination model of Logistic and Xgboost has the highest recall rate of 0.88,indicating that it has better predictability for positive samples.The overall indicators of LightGBM model are lower than Xgboost model,but the memory and running speed are faster than Xgboost model.All indicators of the scorecard model are the worst of the four models,but they are explanatory for the characteristic variables.The banking staff can choose the appropriate model combination according to their actual needs.In terms of variable interpretation,the five variables of family status,total level of personal credit line,number of customer transactions,annual income and age are statistically significant,while the number of bulls and working years are not statistically significant.The seven variables have a positive correlation with the overall credit card repayment intention,which means that the higher the index value,the higher the overall repayment probability and the lower the probability of overdue default risk.The strategists of commercial banks can carry out post-loan risk early warning business according to corresponding indicators.