Deep Reinforcement Learning Approach to Portfolio Optimization

Abstract: This paper evaluates whether a deep reinforcement learning (DRL) approach can be implemented, on the Swedish stock market, to optimize a portfolio. The objective is to create and train two DRL algorithms that can construct portfolios that will be benchmarked against the market portfolio, tracking OMXS30, and the two conventional methods, the naive portfolio, and minimum variance portfolio. We evaluate all the portfolios on a five-year period, from the start of 2016 to the end of 2020, in terms of returns and risk-adjusted returns. The two DRL algorithms implemented are Advantage Actor-Critic (A2C) and Deep Deterministic Policy Gradient (DDPG), they are also compared against each other. The results of this study show that the A2C constructed portfolio significantly outperform the market and all of the other benchmark portfolios, in terms of returns and risk-adjusted returns. The A2C portfolio also outperforms the DDPG constructed portfolio. Even though the DDPG constructed portfolio performs less than the A2C constructed portfolio, it still significantly outperforms all of the other benchmarks on the whole testing period. Thus, concluding that a DRL approach can be implemented, on the Swedish stock market, to optimize a portfolio. Moreover, the study shows that the two DRL agents can pick up on market trends and profit from them. However, applying the methods in a real-world environment does come with some data-processing caveats. Even though the models may come with caveats linked to them, the results of this study underline the usefulness of machine learning methods in portfolio management.