Building a composite early warning index for financial market crises using machine learning and macroeconomic-political uncertainty indicators

Abstract

The accurate and timely prediction of financial market crises remains a persistent challenge for economists, policymakers, and investors. Traditional early warning systems (EWS) often rely on low-frequency macroeconomic indicators and static econometric models, limiting their effectiveness in dynamic market environments. This study proposes to fill this gap by developing a novel framework for crisis prediction through constructing a Composite Early Warning Index (CEWI) that integrates daily data from financial markets, macroeconomic fundamentals, and political uncertainty indicators. Principal Component Analysis (PCA) was employed to synthesize these diverse variables into a single latent factor, capturing the underlying systemic risk. Machine learning algorithms, including Logistic Regression, Random Forest, and XGBoost classifiers, were trained on historical data spanning from 2000 to 2025 to predict crisis periods, defined by sharp equity market declines and official recession declarations. The XGBoost model achieved superior performance with an ROC-AUC of 0.953. Feature importance analysis utilizing SHAP values identified market volatility (VIX), gold prices, and oil prices as the most influential predictors. The results demonstrate that combining high-frequency financial and political indicators with advanced machine learning techniques significantly enhances crisis prediction accuracy. The proposed CEWI-based framework offers a powerful tool for early risk detection and has important implications for financial regulation, investment strategy, and economic policy design.