A machine learning framework for short-horizon monitoring of systemic financial stress in Europe

Early detection of systemic financial stress is essential for financial-stability monitoring in environments characterized by rapid market adjustment, nonlinear transmission, and rare stress episodes. This study develops a high-frequency machine-learning framework for short-horizon monitoring of systemic financial stress in Europe using daily financial, banking, interest-rate, exchange-rate, commodity, volatility, and sentiment-related indicators. Systemic financial stress is measured with the European Central Bank’s Composite Indicator of Systemic Stress (CISS), which is transformed into binary stress targets using fixed, rolling-sigma, and rolling-percentile threshold rules. The empirical design evaluates alternative model classes, dataset specifications, feature representations, forecast horizons, and stress definitions within a strictly chronological out-of-sample validation framework that avoids look-ahead bias. Model performance is assessed using rare-event metrics, with primary emphasis on precision–recall AUC, recall, and F1 score. The results show that predictive performance is strongest under the fixed CISS threshold and at short horizons, particularly the 5-business-day forecast window. The preferred specification is a transparent logit model using CISS- and VIX-based predictors with level and change features. Sentiment and uncertainty-related indicators also provide meaningful short-horizon information, while broader standalone market blocks perform more moderately. Nonlinear classifiers are competitive in several settings but do not materially dominate the linear benchmark. Robustness checks show that rolling thresholds preserve the qualitative short-horizon pattern but reduce performance and increase specification instability, especially at longer horizons. Overall, the framework is best interpreted as a disciplined decision-analytics tool for high-frequency systemic-stress monitoring rather than as a threshold-invariant long-horizon stress-prediction model.