GARCH‑DDQN hybrid hits 79% accuracy

Value-at-Risk is a daily estimate of how much money a portfolio could lose in a bad market move, and banks use it to set capital buffers. (springer.com) A GARCH model is the old workhorse for that job: it watches how volatility clusters, like storms arriving in streaks rather than at random. The paper says those models can miss nonlinear market shifts during crises. (arxiv.org) (springer.com) The new study adds a Double Deep Q-Network, or DDQN, a reinforcement-learning system that learns by updating choices after wins and losses. The authors use it to classify markets into low-risk and high-risk regimes before adjusting the VaR threshold. (arxiv.org) (springer.com) That hybrid model was developed by Fredy Pokou of Inria and the University of Lille, Jules Sadefo Kamdem of Montpellier University, and François Benhmad of Montpellier University. Their paper first appeared on arXiv on April 23, 2025, was revised on August 18, 2025, and was published by *Computational Economics* on January 6, 2026. (arxiv.org) (springer.com) The test bed was daily Euro Stoxx 50 data spanning more than 16 years, including crisis periods and other high-volatility stretches. The paper reports 79.4% test accuracy for the classification framework. (arxiv.org) (springer.com) The authors say the model also cut the number of VaR breaches, the days when actual losses exceed the forecast limit, and reduced the tendency for those breaches to cluster together. In risk management, clustered breaches are a red flag because they suggest the model is failing repeatedly when markets are stressed. (springer.com) They report the backtests met the Kupiec and Christoffersen checks, two standard tests used to see whether VaR exceptions happen at the right rate and without serial dependence. The paper also says Extreme Value Theory supported the model’s handling of tail risk, the rare, very large losses that sit outside ordinary market noise. (springer.com) (arxiv.org) The paper frames the trade-off in practical terms: overestimating VaR can force firms to hold too much capital, while underestimating it leaves them exposed to losses. The authors argue the hybrid approach can improve both regulatory compliance and capital efficiency. (arxiv.org) (springer.com) What the study does not show in the abstract is a head-to-head table against every competing modern risk model, so the 79.4% figure is best read as a result inside this paper’s Euro Stoxx 50 setup, not a universal ranking. The core claim is narrower: a classical volatility model and a reinforcement-learning classifier worked better together than either old assumptions or raw machine learning alone. (arxiv.org) (springer.com)