Effect Analysis of Resampling Techniques on the Performance of Customer Credit Scoring Models

Abstract

In customer credit scoring, the class distributions of credit scoring datasets are usually imbalanced, which severely affects the performance of credit scoring models. To solve the problem, we introduced 4 evaluation criteria, used 9 classification methods, analyzed the performance of 5 commonly used resampling techniques by extensive experiments on 3 real credit scoring datasets, and performed nonparametric tests on the results. We can conclude that: resampling improves the performance of credit scoring, and the degrees of improvement depend on the evaluation criteria selected; generally, SMOTE performs best; the best resampling technique varies with different models and SMOTE works well with most models. Therefore, the application of resampling techniques should be based on evaluation criteria and models selected.