Portfolio Optimization using Reinforcement Learning

Reinforcement Learning has contributed to the automation of various tasks, including electronic algorithmic trading. The main benefit of Reinforcement Learning in financial applications is the relaxation of assumptions on market models and hand-picked features, allowing a data-driven, automated process. The performance of Reinforcement Learning agents on the portfolio management problem is measured. A finite universe of financial instruments such as stocks is selected and the trained agent is constructing an internal representation (model) of the market, allowing it to determine how to optimally allocate funds of a finite budget to those assets. The agent is trained on the actual market prices. The performance metrics are then compared with those of standard portfolio management algorithms on a dataset that has not been used before. A successful agent can be used as a consulting software for portfolio managers or it can be used for low-frequency algorithmic trading. Moreover, it can allow us to identify the missing pieces of the existing models and suggest directions to improve them.