Hybrid VaR: GARCH + DDQN

Banks use Value-at-Risk as a daily loss limit, and a new paper says a hybrid model that mixes GARCH with Double Deep Q-Networks estimated that limit more accurately on Euro Stoxx 50 data. (arxiv.org) Value-at-Risk, or VaR, is a probability cutoff: at a 99% confidence level, it estimates a one-day loss that should be exceeded about once in 100 trading days. Basel’s backtesting framework checks those forecasts by comparing model warnings with actual trading losses. (bis.org) GARCH is the older part of the system. It tracks how market volatility clusters over time — quiet days followed by quiet days, turbulent days followed by turbulent days — a pattern the paper says standard risk models often capture only imperfectly in crises. (arxiv.org) Double Deep Q-Network, or DDQN, is the newer part. It is a reinforcement-learning method introduced in 2016 to reduce the tendency of standard Deep Q-Networks to overestimate action values, and the paper uses it as a classifier for low- and high-risk market regimes. (ojs.aaai.org) The paper, “Bridging Econometrics and AI: VaR Estimation via Reinforcement Learning and GARCH Models,” is by Fredy Pokou, Jules Sadefo Kamdem and François Benhmad. The arXiv version lists the manuscript date as August 19, 2025, and a Springer version was published in 2026 in *Computational Economics*. (arxiv.org) (link.springer.com) The authors tested the model on more than 16 years of daily Euro Stoxx 50 data. STOXX describes that benchmark as a 50-stock blue-chip index covering eight euro-area countries and widely used in ETFs, futures and options. (arxiv.org) (stoxx.com) Their headline result is 79.4% test accuracy. The paper also says the hybrid approach reduced both the number of VaR breaches and the tendency for those breaches to cluster during volatile periods. (arxiv.org) The regulatory claim rests on two standard backtests: Kupiec, which checks whether the number of exceptions is about right, and Christoffersen, which checks whether exceptions arrive independently rather than in streaks. The paper says the model passed both. (arxiv.org) That matters for bank capital because a VaR model that misses too often can trigger supervisory penalties under Basel’s backtesting regime. The Basel Committee’s framework ties model performance to capital consequences by counting exceptions against model forecasts. (bis.org) The paper also says Extreme Value Theory supported the model’s handling of tail risk, the rare losses in the far edge of the distribution that standard averages tend to miss. The authors frame the method as a way to make risk limits adapt faster when market conditions shift. (arxiv.org) The result is still a research paper, not a new banking standard. But it lands in a market where firms already use the Euro Stoxx 50 as a liquid benchmark, and where risk teams keep looking for models that cut exceptions without inflating capital. (stoxx.com) (arxiv.org)