Tail Risk Early Warning System for Capital Markets Based on Machine Learning Algorithms.

Area of Science:

• Financial Risk Management
• Econometrics
• Machine Learning Applications

Background:

• Effective tail risk measurement and early warning are critical for identifying and controlling major capital market risks.
• Traditional models often fall short in accurately assessing and predicting extreme market events.
• China's capital market requires robust tools for managing systemic financial vulnerabilities.

Purpose of the Study:

• To develop and validate a novel tail risk measurement index for China's capital market using the autoregressive conditional Fréchet (AcF) model.
• To construct and optimize a machine learning-based early warning system for tail risks.
• To assess the predictive capabilities of the developed model compared to traditional methods.

Main Methods:

• Application of the autoregressive conditional Fréchet (AcF) model to construct a tail risk measurement index.
• Utilization of machine learning algorithms (Random Forest with oversampling and double sampling) for optimizing the early warning system.
• Comparative analysis of the AcF model and machine learning approach against traditional risk measurement and crisis identification models (e.g., Logit).

Main Results:

• The AcF model significantly enhances tail risk measurement efficiency compared to traditional approaches.
• Tail risk synergies between equity and bond markets are more pronounced than yield synergies, with the index acting as a leading indicator.
• Machine learning-optimized models achieved high out-of-sample crisis warning accuracies: 81.94% for the stock market and 90.20% for the bond market, outperforming the Logit model.

Conclusions:

• The AcF model provides a superior method for tail risk measurement in capital markets.
• The developed tail risk early warning model, optimized with machine learning, accurately identifies significant risks and crises.
• Specific machine learning algorithms (oversampling-random forest for stocks, double sampling-random forest for bonds) are optimal for effective early warning systems.