Early Warning System for Currency Crises using Long Short-Term Memory and Gated Recurrent Unit Neural Networks

Currency crises, recurrent events in economic history for developing, emerging and developed countries, generate disastrous economic consequences. This paper proposes an early warning system for currency crises using sophisticated recurrent neural networks like Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU). These models were initially used in language processing where they performed well. Such models are increasingly used in forecasting nancial asset prices, including exchange rates, but they have not yet been applied to the prediction of currency crises. As for all recurrent neural networks, they allow to take into account non-linear interactions between variables and the inuence of past data in a dynamic form. For a set of 68 countries including developed, emerging and developing economies over the period 1995-2020, LSTM and GRU outperformed our benchmark models. LSTM and GRU correctly sent continous signals within a two-year warning window to alert 91% of the crises. For LSTM, false signals represent only 14% of the emitted signals compared to 23% for the logistic regression, making them ecient early warning systems for policymakers.