Research On Credit Card User Default Prediction Model Based On ANP-lightGBM Algorithm

At present,the per capita holding amount and transaction amount of credit cards in China continue to rise.Credit Card business has become an important profit item for all major banks,but the risk of credit card default has also increased rapidly with the expansion of the credit card market,therefore,banks and various lending institutions have invested a lot of resources to constantly find and optimize the default risk prediction algorithm.Due to the complexity and imbalance of credit card data,it is difficult for banks to establish accurate default risk prediction models,but with the advent of big data and the rapid development of science and technology,the advanced machine learning algorithm provides a new method to prevent credit card default risk.Therefore,using machine learning algorithm to build a high-stability,high-precision default prediction model has a strong practical significance.Based on the existing research and the machine learning algorithm,this paper establishes a more comprehensive credit card user default risk prediction model from three aspects: Feature Engineering,credit scoring table and optimization prediction model.In feature engineering,numerical variables and category variables are screened in different ways according to data types and structural features,in which regression,random forest and Boruta are used to output the importance values of each variable,finally,14 variables were selected to enter the model,and the accuracy of variable selection was improved by using the method of calculating IV value.In the establishment of the credit scoring table,the personal credit index system is constructed by using the reserved indexes,and the indexes of the criterion layer and the network layer are calibrated according to the analytic network process(ANP),the credit scoring table is established by the comprehensive weight of indicators,and the credit score is calculated and classified to evaluate the quality of users intuitively.In the aspect of optimizing prediction model,the logical class variable,the ordered class variable and the disordered class variable are processed by different coding methods,and the model parameters are adjusted by 50% cross-validation,the user credit score result is added into Light GBM model as a new variable for training,and the Light GBM model based on the ANP result is established through the fusion of the two models,and the performance of the model is tested,finally,the accuracy,applicability,effectiveness and stability of the combined model and the base classifier model are compared and analyzed.The results show that the fusion of the two models not only retains the ANP method to systematize complex problems,but also retains the high efficiency and high precision of Light GBM,This improved default prediction model performs better than the single prediction model in risk user identification and prediction,and is more suitable for application in banking or related financial fields,at the same time,the model also provides a new research idea for the bank credit card user default prediction problem.