Journal of Applied Econometrics, March 2017, Volume 32, Issue 2.
To improve the detection of the economic ”danger zones” from which severe banking crises emanate, this paper introduces classification tree ensembles to the banking crisis forecasting literature. I show that their out-of-sample performance in forecasting binary banking crisis indicators surpasses current best-practice early warning systems based on logit models by a substantial margin. I obtain this result on the basis of one long-run- (1870-2011), as well as two broad post-1970 macroeconomic panel datasets. I particularly show that two marked improvements in forecasting performance result from the combination of many classification trees into an ensemble, and the use of many predictors.