ABSTRACT To examine if one machine‐learning model can consistently elucidate financial vulnerabilities, both over time and across levels of development, this paper applies 13 machine‐learning algorithms to evaluate comparative forecasting performance across several banking crises. The study concurrently contributes to the literature on early warning signals, in general, and modelling frameworks, in particular. Four decades of banking crises are appraised through a vector of 12 leading indicators encompassing real, banking and external sectors, and accounting for a representative sample of 19 emerging markets and developed economies. Through quarterly time series supporting expedient policy responses, findings suggest that bank deposits, capital output ratio, exchange rates, gross domestic product, consumption expenditure, and interest rates represent the most prominent leading indicators. Evaluated by applying root mean squared error, mean absolute error, and receiver operating characteristics with area under curve estimates, random forests exhibit highest predictive strength in minimising the prediction error across a panel format, individual country format and out‐of‐time format, and outperform all models in recursively predicting banking crises out‐of‐sample and out‐of‐time. On aggregate, one algorithm consistently fits most crises, which is advantageous for a global policymaker, however results from the individual country dimension highlight a more nuanced approach to algorithmic selection.
Emile du Plessis (Fri,) studied this question.