This paper employs a recently developed statistical algorithm in order to build an early warning model for banking crises in emerging markets. The procedure creates many ‘artificial’ samples by iteratively perturbing the original data set and estimates many models from these samples. The final model is constructed by aggregation, so that, by construction, it is flexible enough to accommodate new data for out‐of‐sample prediction. Out of a large number (540) of candidate explanatory variables, ranging from macroeconomic variables to balance sheet indicators, our procedure selects a handful of indicators (and their combinations) that is sufficient to generate accurate out‐of‐sample predictions of banking crises. Using data covering emerging markets from 1980 to 2010, the model identifies two banking crisis' ‘danger‐zones’, e.g. economic configurations that are conducive to crises. The first occurs when high interest rates on bank deposits, possibly reflecting liquidity risks and solvency fears, interact with credit‐booms and capital flights; the second occurs when an investment boom is financed by a large rise in banks' net foreign exposure. We compare our model to models derived by standard econometric techniques, and find that our approach delivers much better out‐of‐sample predictions. Copyright © 2016 John Wiley & Sons, Ltd.