Analyzing False Positives in Bankruptcy Prediction with Explainable AI

Abstract

With the rise of powerful machine learning solutions, it has become easy to create highly accurate solutions for financial services, yet they fail to comply with financial regulations as they lack transparency and explainability. Bankruptcy prediction is one of the major issues in finance and in the bid to create a highly efficient model which minimizes false negatives where we correctly classify companies that are going to be bankrupt, we see a tradeoff with an increase in false positive cases where companies that are not going to be bankrupt are also flagged. In this paper, we have used a post hoc model explainability technique called SHAP to explain the ML-based bankruptcy prediction model on Taiwan’s bankruptcy dataset and Polish company dataset by generating local as well as global explanations. We have also used the SHAP model to understand how different features contributed to false cases and compare feature attribution with overall model feature relevance to generate an in-depth study of false positive cases.