Based On Improved Stacking's Commercial Bank Customer Credit Default Risk Prediction Model

In recent years,with the rapid development of the credit market,commercial banks have paid more attention on credit risk management.Many scholars and banks have tried to use Random Forest,XGB and Neural Network to build credit default prediction models.They have made some achievements on the accuracy of prediction,but there is still some space to improve the result of the model.Based on customer credit data of a commercial bank,this paper uses ensemble learning method to build a credit default prediction model to obtain a more accurate customer credit default probability.Based on the study of the Stacking model fusion structure,this article proposes three improvements to the second layer of the Stacking structure.First,the feature selection of the second layer is improved.On the basic of the output of the first layer,the original data with all features or features selected by feature importance are added as another part of samples for second layer to obtain a higher feature complexity of second layer;Second,the depth of second layer is improved,this improvement will do- training,the sample of the new round is combined by the original data of second layer and the prediction results in the previous round of the model;The third improvement is based on the second improvement,the sample data of new round is concatenated by the sample data of previous round with the prediction results of the previous round.Based on the single algorithm model,the original Stacking structure and the improved Stacking structure,a total of four sets of comparative experiments were performed in this paper.Through a comparative analysis of the experimental results,the improved Stacking structure has significantly improved the prediction results of credit defaults,and the second improvement method is more effective.