Unlocking the Black Box: Machine Learning Explainability for Financial Crises Prediction

Abstract

The increasing adoption of machine learning (ML) techniques in financial risk prediction has accentuated the trade-off between predictive performance and model interpretability, often referred to as the black-box problem. This study addresses this challenge by developing an interpretable Early Warning System (EWS) for predicting systemic financial crises, integrating the Ran-dom Forest algorithm with Shapley Additive Explanations (SHAP). Using the Jordà-Schularick-Taylor Macrohistory Database, the proposed model identifies key predictors – such as global yield curve slopes, credit growth, and debt service ratios – and achieves high accuracy and robustness. The in-corporation of Explainable AI (XAI) techniques bridges the gap between model performance and transparency, facilitating clear communication of predictions to policymakers and financial regulators. These findings under-score the potential of XAI to enhance machine learning adoption in critical ap-plications by mitigating interpretability concerns. Nevertheless, the model's reliance on historical data and its limited applicability to emerging econo-mies define key areas for future research, including exploring alternative data sources and further refining explainability frameworks.