Uncertainty-Aware Reinforcement Learning for Portfolio Optimization

We explored the use of Reinforcement Learning (RL) combined with risk assessment for optimizing investment portfolios. The dynamic nature of trading, compounded by market frictions, the responses of other market participants, and uncertainties, poses challenges to portfolio optimization. The financial market's intricacies make it difficult to model accurately, compounded by regulatory requirements and internal risk policies mandating risk-averse decisions to avoid catastrophic outcomes. To address this, we proposed risk estimation for investor's risk tolerance threshold. Moreover, modern Deep Learning models are adept at approximating complex relationship between abundant data, however, the main drawback we face now a day is generalization of the relationship to the unseen data. Therefore, the epistemic uncertainty can pose risk to the decision making system. This uncertainty is further addressed using a Variational Autoencoder (VAE) to estimate, and Cost Network to backpropogate riskiness through the model to learn actions with safe results. The actions with stable result or lower reward will be avoided due to reward optimization of RL. Consequently, we successfully managed to reduce the risk and uncertainties in the agent testing process. Our risk-constrained RL algorithm demonstrated zero violation of the constraint in the testing phase. This suggests that adopting a risk-averse RL approach could be beneficial for portfolio optimization, particularly for risk-averse investors.