Early Warning and Prediction of Systemic Financial Risk Using Machine Learning Methods

Abstract

Machine learning techniques offer significant advantages over traditional economic modeling methods, particularly in capturing complex nonlinear relationships, making them well suited for systemic financial risk analysis. This study develops a machine learning-based framework for monitoring and early warning of systemic financial risk, supported by both theoretical analysis and empirical evidence. Early warning indicators are constructed across eight dimensions, including macroeconomic fundamentals, monetary conditions, fiscal status, financial markets, price dynamics, foreign exchange markets, leverage levels, and the banking system. Five representative machine learning models, along with ensemble learning approaches, are employed to forecast systemic financial risk. Empirical results indicate that machine learning models consistently outperform traditional linear models in both in-sample and out-of-sample settings. While the Lasso model achieves superior short-term forecasting performance, the SVM model demonstrates stronger predictive capability over longer horizons. Ensemble models effectively balance predictive accuracy and robustness. Furthermore, partial dependence analysis enhances model interpretability by revealing nonlinear effects and key risk drivers. Exchange rates, money supply, market interest rates, and industrial product prices emerge as critical determinants of systemic financial risk. Targeted monitoring of these variables can support timely risk identification and early intervention.