Bayesian Models for Early Warning of Bank Failures

The focus of this research is to demonstrate how probabilistic models may be used to provide early warnings for bank failures. While prior research in the auditing literature has recognized the applicability of a Bayesian belief revision framework for many audit tasks, empirical evidence has suggested that auditors' cognitive decision processes often violate probability axioms. We believe that some of the well-documented cognitive limitations of a human auditor can be compensated by an automated system. In particular, we demonstrate that a formal belief revision scheme can be incorporated into an automated system to provide reliable probability estimates for early warning of bank failures. The automated system examines financial ratios as predictors of a bank's performance and assesses the posterior probability of a banks financial health (alternatively, financial distress). We examine two different probabilistic models, one that is simpler and makes more assumptions, while the other that is somewhat more complex but makes fewer assumptions. We find that both models are able to make accurate predictions with the help of historical data to estimate the required probabilities. In particular, the more complex model is found to be very well calibrated in its probability estimates. We posit that such a model can serve as a useful decision aid to an auditor's judgment process.