Adaptive Learning for Financial Markets Mixing Model-Based and Model-Free Rl for Volatility Targeting

Model-Free Reinforcement Learning has achieved meaningful results in stable environments but, to this day, it remains problematic regime changing like financial markets. In contrast, model-based RL is able capture some fundamental and dynamical concepts of the environment but suffer from cognitive bias. work, we propose combine best two techniques by selecting various approaches thanks Deep Learning. Using not only past performance volatility, include additional contextual information such as macro risk appetite signals account for implicit changes. We also adapt traditional methods real-life situations considering data training sets. Hence, cannot use future our set implied K-fold cross validation. Building on statistical methods, walk-forward analysis, which defined successive testing based expanding periods, assert robustness resulting agent. Finally, present concept difference's significance a two-tailed T-test, highlight ways models differ more ones. Our experimental show that approach outperforms baseline portfolio Markowitz model almost all evaluation metrics commonly used mathematics, namely net performance, Sharpe Sortino ratios, maximum drawdown, drawdown over volatility.