Forecasting Financial Crises with Public Macro-Demographic Indicators: A Comparison of Logistic, Tree-Based and LSTM Models

Abstract

This paper examines how far financial crises can be anticipated using only publicly available macroeconomic, macro-financial and macro-demographic indicators. We construct a monthly US panel from OECD and Federal Reserve sources and transform standard aggregates into a rich feature set capturing volatility, momentum, higher-order moments, drawdowns and structural breaks. Crisis risk is modeled in a hazard-style early-warning framework: for each month in expansion, we define binary labels for crisis onsets within 6-, 12- and 18-month horizons, combined with lead-time weights that reward earlier, more operationally useful signals. Using this common information set, we compare three families of models: regularised logistic regressions, gradient-boosted decision trees (LightGBM and XGBoost) and a bidirectional LSTM with attention fed by fixed-length feature sequences. Models are evaluated with expanding-window cross-validation and strictly out-of-sample holdout tests. Across horizons, penalised logistic regressions deliver the most accurate and stable forecasts, achieving holdout ROC-AUCs up to 0.99 and F1 scores up to 0.86 at the 18-month horizon, while tree-based methods are competitive only at longer horizons and the Bi-LSTM substantially overfits, adding little incremental predictive power. These results suggest that, in small and highly imbalanced crisis datasets built from open macro-demographic indicators, well-regularised linear models can match or surpass more complex machine-learning and deep-learning approaches, and offer greater transparency for macroprudential policy use.