Achieving Mean–Variance Efficiency by Continuous-Time Reinforcement Learning

We conduct an extensive empirical analysis to evaluate the performance of the recently developed reinforcement learning algorithms by Jia and Zhou [11] in asset allocation tasks. We propose an efficient implementation of the algorithms in a dynamic mean-variance portfolio selection setting. We compare it with the conventional plug-in estimator and two state-of-the-art deep reinforcement learning algorithms, deep deterministic policy gradient (DDPG) and proximal policy optimization (PPO), with both simulated and real market data. On both data sets, our algorithm significantly outperforms the others. In particular, using the US stocks data from Jan 2000 to Dec 2019, we demonstrate the effectiveness of our algorithm in reaching the target return and maximizing the Sharpe ratio for various periods under consideration, including the period of the financial crisis in 2007-2008. By contrast, the plug-in estimator performs poorly on real data sets, and PPO performs better than DDPG but still has lower Sharpe ratio than the market. Our algorithm also outperforms two well-diversified portfolios: the market and equally weighted portfolios.