An Evaluation of Interpretable algorithms for Credit Scoring Model

In credit risk management, credit scoring models are used to evaluate the borrower's repayment ability. Machine learning, as an important branch of artificial intelligence, is introduced to develop credit scoring model to improve predictive accuracy. Nevertheless, for the reason of regulatory requirements, banks need to deeply understand the principles of the credit scoring model in order to make compliance decisions. In this article, we compared the performance of multiple algorithms to develop credit scoring model, including Logistic Regression, Random Forest and XGBoost. Although XGBoost showed the best performance in accuracy, it's a black-box model for implementation. We utilized LIME and SHAP algorithms to analyze the explain the XGBoost model. To identify which features have a significant impact on the model, the feature importance is defined as the number of times for each feature serves as a branch node in the sub decision tree. Then compared the feature importance with the coefficients of logistic regression to verify the rationality. Moreover, LIME and SHAP interpretation algorithms were used to explain why a certain customer be judge as a default or non default one. By applied the combination of XGBoost, LIME and SHAP approach, a more accurate prediction can be obtained with an acceptable interpretability.