Predicting currency crises in emerging markets: A deep learning approach with LSTM networks and attention mechanism

Exchange rate crises in emerging markets can unfold with extraordinary speed; however, the macroeconomic vulnerabilities feeding into them typically build over years rather than overnight. This study asks whether machine learning — specifically, a Long Short-Term Memory network augmented with Bahdanau attention — can expose the temporal structure of that build-up in a way that conventional early warning models cannot. Using a panel of 30 emerging economies from 1998 to 2023 (780 country-year observations and nine macroeconomic predictors), Exchange Market Pressure indices are constructed to generate crisis labels cross-validated against Laeven and Valencia (2018). Head-to-head testing against Logistic Regression (AUC = 0.985) and Random Forest (AUC = 0.968) showed that the LSTM-attention model (AUC = 0.886) did not win on raw discriminative metrics, a result reported transparently rather than apologetically. The model's value lies elsewhere: attention weights decompose each crisis probability estimate across the three-year look-back window, revealing that conditions one year before a crisis carry roughly 40 percent of the total explanatory weight, with meaningful contributions from two and three years prior as well. Real effective exchange rate misalignment and interest rate differentials emerge as the most informative predictors. These findings support the view that temporal interpretability — understanding not just whether a crisis is probable but how vulnerability accumulates — constitutes a distinct and operationally useful addition to existing surveillance toolkits.