美联储-线性与非线性计量经济学模型与机器学习模型的对比_已实现波动率预测（英）-2025_68页_982kb

Summary

This study compares traditional econometric models and machine learning (ML) techniques in predicting realized volatility (RV) for the S&P 500 index using high-frequency data. The research evaluates models such as ARFIMA, HAR, regime-switching variants (THAR, STHAR, MSHAR) against ML methods including XGBoost, deep neural networks, and recurrent neural networks (RNNs).

Key Findings:

1. Econometric Models Outperform ML: Simple regime-switching models, particularly THAR and STHAR, consistently outperform all ML techniques across statistical accuracy, risk forecasts, and economic utility.
2. ML Limitations: While ML models capture some nonlinear patterns, they provide no consistent advantage over parsimonious econometric alternatives, especially during market turbulence and with limited predictors.
3. Relevant Prediction Context: Results hold even when additional predictors are included, demonstrating the robustness of econometric models.
4. Implications: Regime-switching econometric models are recommended for their simplicity, interpretability, and effectiveness in volatility forecasting, particularly for risk management and portfolio decisions.

Recommendation: Focus forecasting and risk modeling on regime-switching diagonal regression models due to their superior balance of accuracy, simplicity, and practical utility in financial applications.