Deep reinforcement learning for an empirical approach to Value-at-Risk

Risk measurement is central to modern financial risk management. Market developments have highlighted to the financial data's complexity, with the result that most assumptions underpinning econometric models become obsolete. This means that their results are no longer valid, nor can they be correctly interpreted. For this reason, in this paper, a robust empirical approach to Value-at- Risk has been proposed based on VaR-GARCH model (resp. VaR-GJR-GARCH) and enriched with information from a directional forecast. The standard directional prediction problem was transformed into an imbalanced classification problem solved using Double Deep Q-Network (DDQN) classifier, a deep reinforcement learning algorithm. The model's performance in this paper is assessed using daily Eurostoxx 50 price data, which covers a number of major crises and shocks, enabling us to test its robustness in addition to statistical tests. Double Deep Q-Network (DDQN) makes better forecasts of risk levels of returns, enabling value-at-risk to be reduced when risk levels are low, or increased when they are high. Results obtained prove that this approach generates the most accurate VaR estimates.