Predicting bank distress in the UK with machine learning

Using novel data and machine learning techniques, we develop an early warning system for bank distress. The main input variables come from confidential regulatory returns, and our measure of distress is derived from supervisory assessments of bank riskiness from 2006 through to 2012. We contribute to a nascent academic literature utilising new methodologies to anticipate negative firm outcomes, comparing and contrasting classic linear regression techniques with modern machine learning approaches that are able to capture complex non-linearities and interactions. We find the random forest algorithm significantly and substantively outperforms other models when utilising the AUC and Brier Score as performance metrics. We go on to vary the relative cost of false negatives (missing actual cases of distress) and false positives (wrongly predicting distress) for discrete decision thresholds, finding that the random forest again outperforms the other models. We also contribute to the literature examining drivers of bank distress, using state of the art machine learning interpretability techniques, and demonstrate the benefits of ensembling techniques in gaining additional performance benefits. Overall, this paper makes important contributions, not least of which is practical: bank supervisors can utilise our findings to anticipate firm weaknesses and take appropriate mitigating action ahead of time.