Abstract:
This paper develops an early warning system for predicting distress for large European banks. Using a novel definition of distress derived from banks' headroom above regulatory requirements, we investigate the performance of three machine learning techniques against the traditional logistic model. We find that the random forest model shows superior performance both out-of-sample and out-oftime. Unlike previous studies, we also employ a series of sampling techniques showing that they significantly improve the ability to identify distress events irrespective of the model used. Moreover, we show that ensemble techniques can help improve performance relative to the single best performing model. Finally, using the latest machine learning interpretability tools, we show that the variables closely tied to bank profitability and solvency are important drivers for predicting bank distress. Overall, our paper has important practical implications for bank supervisors and macroprudential authorities who can utilise our findings to identify bank weaknesses ahead of time and adopt pre-emptive measures to safeguard financial stability.