A decision-support framework for bankruptcy prediction using explainable ensemble learning

Bankruptcy prediction is a challenging problem in the field of risk management of financial assets due to the rarity of bankruptcy events, the class imbalance that results from such rarity, and the regulatory requirements regarding the interpretability of AI-based decision systems. Given the gradual development of bankruptcy, it is necessary to use AI-based models that can capture non-linear relationships among financial metrics and detect early signs of issues in a company's financial health. These requirements suggest using AI models beyond linear models and financial metrics alone. In this article, a Stacking Ensemble model is developed with both linear and non-linear models in order to investigate the ability of the models to predict bankruptcy with an emphasis on analyzing prediction trade-offs under severe class imbalance, as well as utilizing methods from the field of Explainable Artificial Intelligence (XAI) to investigate the models within the ensemble framework to determine the reasons for the ensemble's performance on the evaluation metrics. Results indicate that the model has good discriminatory power, but is conservative in its decisions to recognize financial distress within companies. However, the requirement for model interpretability is still met, and the model's performance across different evaluation thresholds is considered in the article.