Machine Learning Models for Credit Scoring: Performance Assessment

This paper evaluates the risk scoring performance of three main machine learning models: XGBoost, Random Forest and Support Vector Machines. Various evaluation metrics are used in this process: feature importance, predictive score distributions, and Receiver Operating Characteristic curves. When compared to other models, the results show that XGBoost is both more accurate and has a higher capability than others in distinguishing risk levels with different levels of prediction reliability. In addition, the interaction between loan amounts and scores obtained from prediction experiments was considered and more information were provided. The results demonstrate the benefits of boosting algorithms such as scoring tools for applications based on scoring and suggest which is the optimal models to use when modeling financial risk assessment.