Explainable gradient boosting for corporate crisis forecasting in Italian businesses

Abstract

Scholars have taken a keen interest in predicting corporate crises in the past decades. However, most studies focused on classical parametric models that, by their nature, can consider few predictors and interactions and must respect numerous assumptions. Over the past few years, the economy has faced a severe structural crisis that has resulted in significantly lower income, cash, and capital levels than in the past. This crisis has led to insolvency and bankruptcy in many cases. Hence, there is a renewed interest in research for new models for forecasting business crises using novel advanced machine learning techniques. This study aims to develop a model that achieves state-of-the-art accuracy while being fully interpretable, overcoming the limitations of previous research. The model demonstrates excellent predictive performance on par with black-box approaches while maintaining complete transparency by leveraging Explainable Boosting Machines, an intrinsically interpretable tree-based ensemble method, and hyperparameter optimization. The approach automatically considers all possible interactions and uncovers relevant aspects not considered in past studies. This line of research provides compelling results that can bring new insights to the literature on corporate crisis prediction. The interpretable nature of the model is a key advancement, enabling practical application and a deeper understanding of the factors driving corporate financial distress.