Predicting financial crises: an evaluation of machine learning algorithms and model explainability for early warning systems

This paper addresses the critical challenge of detecting financial crises in their early stages given their profound economic and societal consequences. It investigates the efficacy of machine learning models by comparing the standard econometric model of Logistic Regression with k-nearest Neighbours, Random Forest, Extremely Randomised Trees, Support Vector Machine, and artificial Neural Network models in a cross-validation experiment. The study utilises financial crisis observations between 1870 and 2020 from the MacroHistory database and up to 14 early warning indicators from different theoretical backgrounds. The results demonstrate that advanced machine learning models, particularly Random Forest and Extremely Randomised Trees, outperform Logistic Regression, as measured by the area under the receiver operating characteristic curve (AUROC). This finding holds across different crisis data sources, early warning indicator sets, and model specifications. Furthermore, this study emphasises the importance of model explainability for both researchers and policymakers by employing Accumulated Local Effect (ALE) plots to unravel the complex predictive processes of advanced machine learning models and provide deeper insight into the underlying crisis dynamics. Nonlinear connections to financial crises were identified for several early warning indicators, which are difficult to capture by linear models. Several U-shaped relationships were observed, highlighting the disruptive role of strong economic changes for financial stability. The results emphasise the potential of advanced machine learning methods to limit the societal impact of financial crises by enabling earlier intervention by policymakers and public institutions.