Deep reinforcement learning with particle filtering for latent volatility inference

This study proposes a novel hybrid reinforcement learning framework for volatility estimation, integrating policy gradient algorithms with particle filtering techniques to achieve dynamic and data-driven inference of latent volatility. Our approach reformulates volatility forecasting as a sequential decision-making problem, enabling an agent to learn adaptive, stochastic policies from evolving market patterns without reliance on rigid parametric specifications. Crucially, the integration of particle filtering addresses the challenge of unobservable volatility by providing robust, unbiased state calibration in non-linear environments. Extensive out-of-sample experiments demonstrate the model’s robust generalization and stronger predictive performance compared to traditional econometric baselines on both synthetic and real-world data. Furthermore, the framework’s effectiveness is validated using data from major global financial markets, where the inferred latent signals successfully uncover cross-market volatility transmission and lead-lag structures. By combining adaptive policy learning with probabilistic state estimation, this work enhances the practical potential for advanced financial risk monitoring and decision support.