High-dimensional multi-period portfolio allocation using deep reinforcement learning

This paper proposes a novel investment strategy based on deep reinforcement learning (DRL) for long-term portfolio allocation in the presence of transaction costs and risk aversion. We design an advanced portfolio policy framework to model the price dynamic patterns using convolutional neural networks (CNN), capture group-wise asset dependence using WaveNet, and solve the optimal asset allocation problem using DRL. These methods are embedded within a multi-period Bellman equation framework. An additional appealing feature of our investment strategy is its ability to optimize dynamically over a large set of potentially correlated risky assets. The performance of this portfolio is tested empirically over different holding periods, risk aversion levels, transaction cost rates, and financial indices. The results demonstrate the effectiveness and superiority of the proposed long-term portfolio allocation strategy compared to several competitors based on machine learning methods and traditional optimization techniques.