A Liquidity Risk Early Warning Indicator for Italian Banks: A Machine Learning Approach

The paper develops an early warning system to identify banks that could face liquidity crises. To obtain a robust system for measuring banks’ liquidity vulnerabilities, we compare the predictive performance of three models – logistic LASSO, random forest and Extreme Gradient Boosting – and of their combination. Using a comprehensive dataset of liquidity crisis events between December 2014 and January 2020, our early warning models’ signals are calibrated according to the policymaker's preferences between type I and II errors. Unlike most of the literature, which focuses on default risk and typically proposes a forecast horizon ranging from 4 to 6 quarters, we analyse liquidity risk and we consider a 3-month forecast horizon. The key finding is that combining different estimation procedures improves model performance and yields accurate out-of-sample predictions. The results show that the combined models achieve an extremely low percentage of false negatives, lower than the values usually reported in the literature, while at the same time limiting the number of false positives.