A comprehensive explainable approach for imbalanced financial distress prediction

Ruhao Chen, Tong-Yu Lu, Jiyuan Min and Wenfu Xu

Abstract

Financial distress prediction (FDP) models are important risk prevention and control tools for monitoring the financial status of companies. To simultaneously address the issue of imbalanced data and the explanation of “black box” models in FDP, we propose an explainable approach for imbalanced FDP through focal-aware extreme gradient boosting (FL-XGBoost). In this approach we address the difficulty of identifying minority-class samples in imbalanced data sets by introducing the focal loss function into the XGBoost framework. Validation tests of different FDP models show the strong performance of our proposed model. Moreover, we use three explanatory methods to generate explanations for different FDP model users. The experimental results based on different kinds of imbalanced data sets show that our proposed method, FL-XGBoost, has strong classification performance. Further, the explanatory methods can increase the transparency and credibility of our proposed black box model.