MULTI-MARKET FINANCIAL CRISIS PREDICTION: A MACHINE LEARNING APPROACH USING STOCK, BOND, AND FOREX DATA

Abstract

Early detection of financial crises remains a persistent challenge because traditional single-market indicators, focused separately on equities, bonds, or foreign exchange, cannot fully capture cross-market contagion and volatility spillovers that drive systemic breakdowns. This study develops a unified machine learning framework for multi-market crisis prediction in the United States by integrating data from the S&P 500 index, U.S. Treasury bond yields and prices, and EUR/USD exchange rates spanning 2010–2025. The pipeline constructs engineered features that encompass daily and rolling log-returns, multi-horizon volatilities, bond yield-curve slopes, contagion fractions, and pairwise rolling correlations, alongside principal component decompositions to capture latent systemic factors. Using rolling time-series cross-validation, we benchmark Logistic Regression and XGBoost models under realistic temporal splits and class imbalance adjustments. Results show that multi-market fusion markedly improves predictive power over single-market baselines, with XGBoost achieving superior ROC-AUC and precision–recall scores while maintaining calibration stability. SHAP-based explainability identifies volatility clustering, yield-slope inversions, and stock-bond correlation spikes as dominant early-warning features. Economic backtests further demonstrate that a crisis-aware portfolio allocation strategy informed by model probabilities reduces maximum drawdown and enhances risk-adjusted returns relative to buy-and-hold benchmarks. Collectively, these findings establish that machine-learning-driven, cross-market integration provides a transparent and operationally feasible foundation for systemic-risk monitoring and financial-stability forecasting within U.S. markets.