New forecasting methods for an old problem: Predicting 147 years of systemic financial crises

Abstract This paper develops new forecasting methods for an old and ongoing problem by employing 13 machine learning algorithms to study 147 years of systemic financial crises across 17 countries. Findings suggest that fixed capital formation is the most important variable. GDP per capita and consumer inflation have increased in prominence whereas debt‐to‐GDP, stock market, and consumption were dominant at the turn of the 20th century. A lag structure and rolling window both improve on optimized contemporaneous and individual country formats. Through a lag structure, banking sector predictors on average describe 28% of the variation in crisis prevalence, the real sector 64%, and the external sector 8%. Nearly half of all algorithms reach peak performance through a lag structure. As measured through AUC, and Brier scores, top‐performing machine learning methods consistently produce high accuracy rates, with both random forests and gradient boosting in front with 77% correct forecasts, and consistently outperform traditional regression algorithms. Learning from other countries improves predictive strength, and non‐linear models generally deliver higher accuracy rates than linear models. Algorithms retaining all variables perform better than those minimizing the influence of variables.