Predicting systemic financial crises with recurrent neural networks

•We consider LSTM and GRU neural nets that take advantage of the recent advantages in deep learning.•The gated recurrent neural net models consistently outperform more basic neural nets and the benchmark logistic model.•The LSTM neural net produces less than half the amounts of false alarms in comparison to the logit model.•The success of the new models is based on their ability to handle time series data.•We characterize the drivers of the neural net predictions using a Shapley value decomposition.