Forecasting High-Risk Composite CAMELS Ratings

Abstract

We investigate whether statistical learning models can contribute to supervisors' off-site monitoring of banks' overall condition. We use five statistical learning and two forecast combination models to forecast high-risk composite CAMELS ratings over time (1984-2015), where a high-risk composite CAMELS rating is defined as a CAMELS rating of 3, 4, or 5. Our results indicate that the standard logit model, which is already widely used to forecast CAMELS ratings, comes close enough to be an adequate model for predicting high-risk ratings. We also find that the overall accuracy of the individual forecasts could be modestly improved upon by using forecast combination methods.

Keywords: Bank supervision and regulation, early warning models, CAMELS ratings, machine learning

JEL Classification: G21, G28, C53