Customer Credit Default Prediction Research Based On Deep Learning

Credit economy is an important source of funds for individuals and enterprises and plays an important role in stimulating economic growth.With the improvement of national credit economy policies,the number of institutions engaged in credit business in China is increasing,the types of credit products are increasingly rich,and the credit balance is growing rapidly.Behind the rapid growth of credit economy is the increasing demand for intelligent risk control.Credit risk control generally includes pre-loan review,in-loan risk management,post-loan collection and other links,among which the most core is the pre-loan review stage.A good pre-loan risk prediction model can minimize future risks.Therefore,this paper takes the pre-loan approval of customer credit as the main research object,and tries to establish an efficient and accurate customer credit default prediction model.This paper firstly reviews the research results of domestic and foreign scholars in the field of customer credit default prediction,and finds that the current mainstream approach is mainly to use machine learning for modeling,with few cases of deep learning.Therefore,this paper attempts to use deep learning model TabNet to predict customer credit default risk,so as to enrich the application achievements of deep learning technology in this field.In this paper,open credit default data set is used to model,KS index and AUC index are used to test the effect of the model,and Bayesian optimization is used to search the model's hyperparameters.After the data set is preprocessed and feature constructed,the data is put into TabNet model,and the model performs well,with KS index of 0.408 and AUC index of 0.771,which indicates that the model has certain practical value.In addition,two ablation experiments were designed in this paper,and the experimental results show that TabNet model is better than Random Forest and Light GBM in selected data sets.Although the performance of TabNet model is slightly inferior to XGBoost,the model training efficiency is higher than that of XGBoost.Feature construction(feature construction based on business understanding,statistical feature construction)plays a significant role in improving model effect.