美联储-期权定价的局部估计_利用市场状态信息改进预测（英）-2025_53页_1mb

Key Findings:

This paper introduces a local estimation framework for option pricing models. Instead of altering the underlying model structure (e.g., Heston-Nandi GARCH or Heston SV), the method reweights historical observations using kernel-based functions (with bandwidths selected via a validation procedure), emphasizing the relevance to current market conditions.

Main Results:

• Forecasting Accuracy: The local estimation significantly outperformed traditional non-local methods in forecasting near-term option implied volatilities (IV) and prices for both GARCH and SV models, especially during low-volatility periods (e.g., 2017) and across the entire option surface (cross-sectional errors).
• Model Confidence Set: Results from Diebold-Mariano tests and Model Confidence Sets consistently showed that local estimators (using state variables like VIX for GARCH and time/RV5/VIX for SV) yielded the lowest out-of-sample errors and were included in the 95% MCS.
• Adaptation: Local estimation allows for better adaptation to changing market regimes by assigning more weight to observations closer to the current state (e.g., VIX level, time point). This reduced the impact of model misspecification and improved the alignment between model-implied risk-neutral distributions and nonparametric benchmarks. Parameter estimates also exhibited greater time-variation under local estimation.
• Economic Implications: Expected option returns calculated using the local estimation method aligned more closely with realized returns compared to the non-local benchmark.
• Robustness: The performance advantage of local estimation over non-local estimation was robust to different performance metrics, longer-term horizons (3-month), and whether future underlying return information was assumed to be known (VIX and time remain top state variables).
• Compare GARCH/SV:
  • GARCH: Benefited greatly from incorporating VIX or time information.
  • SV: Showed improvement with time, RV5, and VIX state variables.

Conclusion:

The framework demonstrates that simply modifying the estimation method (weighting relevant past data) to reflect current market states can yield substantial improvements in out-of-sample forecasting accuracy and robustness, even without complicating the model structure. It is particularly valuable for applications requiring accurate near-term forecasts, such as risk management and margin calculations. The importance of time and volatility-related state variables (VIX, RV5, Time) is empirically supported.