Low- and high-beta stock rotation trading by reinforcement learning policies

The financial markets are becoming more complicated and unstable, which makes it even more important to find portfolio management solutions that aremore flexible and effective. Mean-Variance Optimization, Risk Parity, and Equal-Weighted Portfolio are examples of conventional models that do not always take into account how markets change over time. This is why more advanced methods like Reinforcement Learning (RL) are needed. This study looks at how three RL algorithms, SARSA, Q-Learning, and Greedy-GQ, can manage a portfolio with two stocks, one that is very volatile and one that is not. This study makes a unique contribution as it compares these algorithms and looks at how adding extra information, like technical indicators, the Relative Strength Index, and Moving Averages, can help the models make better decisions. This research evaluates how well these methods perform compared to a benchmark, the equally weighted portfolio, using data from the New York Stock Exchange (NYSE) from October 2021 to October 2022. During the financial crisis of 2007–2008, a robustness check showed that these methods still worked well even when the market was very unstable. A second robustness test with higher transaction costs showed once more that the reinforcement learning methods can work in more realistic crisis trading environments. The results showed that, on average, Greedy-GQ did better than SARSA and Q-Learning. But the performance was different for each configuration, and some choices of hyperparameters had a big effect on the results. When more market information was added to the state vector, all three algorithms worked better.