Background: Accurate forecasting of realized volatility (RV) remains a cornerstone of risk management and option pricing. Hybrid volatility models, such as the HAR-RV-CARMA framework, have proven effective at fusing discrete-time persistence with continuous-time dynamics, utilizing the standard Kalman Filter for weight allocation. The model operates on a reactive feedback loop that relies on past residuals, often leading to a "tracking lag" during rapid market transitions, in which the model fails to shift weights quickly enough among its constituent components.

Methods: This study introduces a novel architectural enhancement to the HAR-RV-CARMA model by adding an Entropy-Conditioned State-Space Layer. We utilize Sample Entropy as an external complexity indicator to track the information regularity of the volatility series in real time. By linking entropy fluctuations to the noise covariance system in the Kalman Filter, we create a proactive regime-switching mechanism. This enables the model to identify breakdowns in structural persistence and speed up weight re-adjustments before prediction errors accumulate in full.

Results: Using daily realized volatility for major U.S. equity ETFs (SPY, QQQ, and IWM), the models are estimated over a training period from 2015 to 2022 and a test period covering 2023 to 2024. The proposed entropy-conditioned hybrid model is assessed against the base HAR-RV-CARMA model. Forecast performance is evaluated using MAE, MSE, QLIKE, and directional accuracy. The results show that the entropy-enhanced model consistently outperforms the base hybrid model across all evaluation metrics.

Conclusion: Overall, the proposed framework offers a flexible and econometrically sound approach to volatility forecasting that balances model simplicity with regime-sensitive learning. The Entropy-Conditioned HAR-RV-CARMA model effectively reduces the lag present in standard dynamic weighting, providing a more robust forecasting tool for highly volatile markets.