Abstract
We compare the out-of-sample accuracy of three methodologies—the time-varying hazard model of Shumway (2001), the static probit model used by Cole and Gunther (1998), and a static logistic regression model similar to Cole and White (2012)—in forecasting U.S. bank failures. When we limit all models to financial data available at the time of prediction, we find that the logistic and probit models outperforms the hazard model, indicating that the superior performance of hazard models documented in previous empirical research is attributable to use of more timely financial data rather than to incorporation of time-varying covariates. We also find that the logistic model outperforms the probit model. Finally, we also find that a parsimonious specification fit to data over 1985-1993 performs well in forecasting bank failures during 2009-2012—evidence that the characteristics of “distressed banks” have experienced little change over the past two decades despite substantial changes in structure and regulation of the U.S. banking industry. Our findings support supervision focusing on banks' traditional CAMELS risk ratios. We also add to the literature finding that simpler models outperform more complex models in out-of-sample forecastin