The paper develops an early warning model to identify liquidity crises for Italian banks by comparing the predictive performance of three machine learning methods (logistic LASSO, random forest and Extreme Gradient Boosting). Using an innovative dataset, built using data from the Bank of Italy's Emergency Liquidity Assistance (ELA) operations, carried out to fund banks facing temporary liquidity shortages, our early warning models' signals are calibrated using an optimization method based on preferences between type I error (missing a crisis) and type II error (false alarm).
The three estimation methods used show an excellent predictive performance, which improves when they are combined by computing a simple or weighted average. The combined models achieve a low percentage of missed crises (false negatives), while at the same time limiting the number of false alarms (false positives).