This paper introduces classification tree ensembles (CTEs) to the banking crisis forecasting literature. I show that CTEs substantially improve out-of-sample forecasting performance over best-practice early-warning systems. CTEs enable policymakers to correctly forecast 80% of crises with a 20% probability of incorrectly forecasting a crisis. These findings are based on a long-run sample (1870-2011), and two broad post-1970 samples which together cover almost all known systemic banking crises. I show that the marked improvement in forecasting performance results from the combination of many classification trees into an ensemble, and the use of many predictors.