Résumé
Estimating market risk in volatile environments remains a major challenge, as traditional GARCH-type models often struggle to capture nonlinear dynamics. This paper proposes a hybrid Value-at-Risk (VaR) framework that integrates GARCH volatility forecasts with a Double Deep Q-Network (DDQN) reinforcement learning classifier. By reframing VaR estimation as a classification problem, the model adaptively adjusts risk thresholds based on predicted low- and high-risk return regimes. Using more than 16 years of Euro Stoxx 50 data, the framework achieves 79.4% test accuracy and substantially reduces both the frequency and the temporal clustering of VaR violations. Backtesting confirms compliance with the Kupiec and Christoffersen tests, while Extreme Value Theory supports its ability to model tail risk. The resulting approach offers a statistically robust, capital-efficient, and regulatory-aligned solution for proactive financial risk management.