Bank regulators have worked to develop statistical models predicting bank failures, but such models cannot be estimated during periods of few failures. We address this problem using an alternative approach, forecasting the leverage ratio as a continuous variable that avoids the small sample problem. The leverage ratio is a natural choice in this setting both because of its historically consistent ability to predict failures and because of regulators' primary focus on bank capitalization. Our model selection draws on both the earlier literature and more recent stress-testing studies. Out-of-sample performance shows promise as a supplement to the standard approach.