Explainable Machine Learning Applications for Predicting Corporate Default Risk in Vietnam

Abstract

The explainability of machine learning models is becoming increasingly important, especially in the field of risk management. In this study, we utilized the SHAP and LIME methods to explain the default prediction results from the XGBoost machine learning model on a dataset of listed companies in Vietnam from 2018 to 2023. The results indicated that SHAP identified significant factors such as the interest coverage ratio, the retained earnings to total assets ratio, the cash flow to interest ratio, the cash ratio, the debt to total assets ratio, and the size of the company as having a strong impact on the prediction outcomes and also described the nonlinear relationship of these impacts through SHAP values. Additionally, LIME was used to explain the factors affecting a specific default case and showed the alignment of the prediction results with the actual situation of the company.