A risk-aware approach to stock portfolio allocation based on Deep Q-Networks

Abstract

Reinforcement Learning techniques have shown a great potential in the active allocation of stock portfolios. However, state-of-the-art solutions show limited stability and fairly high sensitivity to volatile market conditions. To tackle these issues, this paper presents a new risk-aware approach based on Deep Q-learning Networks. It leverages Quantile Regression DQNs to mitigate the underlying market risks and an action branching architecture to effectively handle high-dimensional stock spaces. Furthermore, it also introduces noise perturbations to the network’s weights aimed at self-tuning the degree of exploration for each input dimension. Based on the empirical simulations, which were carried out on the Dow Jones-30 stocks over a three-year period, the proposed system performs better than state-of-the-art RL solutions in terms of cumulative return, stability, and sharpe ratio.