Towards Better Banking Crisis Prediction: Could an Automatic Variable Selection Process Improve the Performance?

This study proposes using the Least Absolute Shrinkage and Selection Operator (LASSO) method with cross‐validation to automate the variable selection process of the conventional multivariate logit early warning system (EWS), the purpose being to improve the prediction of systemic banking crises. Using a dataset covering 23 OECD countries with quarterly data from 1970Q1 to 2018Q3, model performance is evaluated in a recursive out‐of‐sample forecasting exercise, taking policy‐makers' preference of missed crises and false alarms into account. The results suggest that the automatic variable selection process can enhance the predictive performance of the EWS. It also highlights the importance of extracting information from variable interactions and lags that may not be easily identified and accessed by typical subjective variable pre‐selection. This simple approach is easy to interpret and is transparent, which are important aspects for effective policy communication. Five variables, namely credit growth, domestic and global credit gaps, real house price growth and the real effective exchange rate, are identified as the most important key indicators of systemic banking crises.