Trading financial indices with reinforcement learning agents

Intelligent agents are often used in professional portfolio management. The use of intelligent agents in personal retirement portfolio management is not investigated in the past. In this research, we consider a two-asset personal retirement portfolio and propose several reinforcement learning agents for trading portfolio assets. In particular, we design an on-policy SARSA (λ) and an off-policy Q(λ) discrete state and discrete action agents that maximize either portfolio returns or differential Sharpe ratios. Additionally, we design a temporal-difference learning, TD(λ), agent that uses a linear valuation function in discrete state and continuous action settings. Using two different two-asset portfolios, the first asset being the S&P 500 Index and the second asset being either a broad bond market index or a 10-year U.S. Treasury note (T-note), we test the performance of different agents on different holdout (test) samples. The results of our experiments indicate that the high-learning frequency (i.e., adaptive learning) TD(λ) agent consistently beats both the single asset stock and bond cumulative returns by a significant margin.