Reinforcement learning and portfolio allocation : challenging traditional allocation methods?

We test the out-of-sample trading performance of model-free reinforcement learning (RL) agents and compare them with the performance of equally-weighted portfolios and traditional mean-variance (MV) optimization benchmarks. By dividing European and U.S. indices constituents into factor datasets, the RL-generated portfolios face different scenarios defined by these factor environments. The RL approach is empirically evaluated based on a selection of measures and probabilistic assessments. Training these models only on price data and features constructed from these prices, the performance of the RL approach yields better risk-adjusted returns as well as probabilistic Sharpe ratios compared to MV specifications. However, this performance varies across factor environments. RL models partially uncover the nonlinear structure of the stochastic discount factor. It is further demonstrated that RL models are successful at reducing left-tail risks in out-of-sample settings. These results indicate that these models are indeed useful in portfolio management applications.