Research On Personal Credit Risk Assessment Based On Stacking Fusion Model

In recent years,the credit scale of commercial banks has been growing rapidly,and the issuance of credit cards has been increasing every year,but the problem of credit risk has also become prominent.In this situation,in the face of a large number of credit applications,how to quickly and accurately complete the credit risk assessment of applicants has not only become the focus of commercial banks,but also attracted the attention of more and more scholars in the academic circle.As for credit risk assessment,the application of machine learning algorithms has been gradually recognized by the academia and industry.However,domestic researches are mostly based on a single model to carry out evaluation,and there are few researches based on the perspective of multi-model fusion,which obviously greatly affects the improvement of evaluation efficiency.Based on the above background,this paper from the perspective of model fusion,uses a variety of machine learning algorithms to build a more efficient credit risk assessment model.On this basis,the credit data set of domestic A commercial bank is used as the sample data to empirically evaluate the credit risk of applicants.More specifically,data preprocessing were carried out on the user label table in the sample data,and the RFM model was used to derive the characteristics of the transaction behavior table.Then,based on the feature selection method combining Light GBM algorithm and Pearson correlation coefficient,the important features were selected.Then,Borderline?SMOTE algorithm was used to process the unbalanced samples,the Logistic regression,KNN,Random forest and Light GBM 4 group single models were constructed.Finally,KNN,Random forest and Light GBM models with AUC value greater than 0.85 were used as primary learners,and Logistic regression model with good stability was used as secondary learner.The final credit risk assessment model was constructed through two-layer stacking integration framework,and the corresponding empirical evaluation was made.This paper draws the following conclusions: In terms of credit risk assessment indicators,the characteristics of users' trading behavior make an important contribution to credit risk assessment.In the ranking of feature importance of Light GBM,there are 14 trading derivative features in the top 20.In the construction of credit risk assessment model,by comparing the prediction results of four groups of single models,it shows that the overall performance of Light GBM model is better than the other three groups of models.At the same time,the overall performance of Stacking fusion model constructed in this paper is further improved compared with Light GBM model,and the fusion model has good scalability,so the effectiveness of the fusion model established in this paper is verified.In the application of the Stacking fusion model,this paper uses the prediction probability obtained from the Stacking fusion model to draw the promotion chart and establish the credit score card.The results show that the fusion model has strong customer grouping differentiation ability and prediction ability.At the same time,the fusion model is applied to the unbalanced new samples in line with the real business scene,which verifies that the fusion model has a certain practical value.