An eXplainable early warning system to predict systemic financial crises

This paper develops an interpretable Early Warning System (EWS) to detect early signals of systemic financial crises one year in advance, combining deep learning techniques with an eXplainable Artificial Intelligence (XAI) framework. Using data from 18 European countries over 2001–2020, we compare multiple Machine Learning models and identify the Long Short-Term Memory (LSTM) neural network as the most effective architecture for capturing cross-country and temporal dependencies. The model is designed to capture temporal dependencies in macrofinancial indicators, while interpretability is enhanced through Shapley value analysis, which allows us to identify the variables contributing most to the predicted crisis signals. The results suggest that the inclusion of global market variables improves the model’s forecasting performance and contributes to reducing false alarms. In particular, gold prices, sovereign bond spreads, and treasury rates emerge as the most influential early-warning indicators. These findings provide useful insights for policymakers and financial intermediaries by linking predicted crisis signals to economically interpretable drivers of systemic vulnerability, thereby supporting macroprudential monitoring and preventive policy assessment.