Comparing RL approaches for applications to financial trading systems

In this paper we present and implement different Reinforcement Learning (RL) algorithms in financial trading systems. RL-based approaches aim to find an optimal policy, that is an optimal mapping between the variables describing an environment state and the actions available to an agent, by interacting with the environment itself in order to maximize a cumulative return. In particular, we compare the results obtained considering different on-policy (SARSA) and off-policy (Q-Learning, Greedy-GQ) RL algorithms applied to daily trading in the Italian stock market. We both consider computational issues and investigate practical solutions applications, in an effort to improve previous results while keeping a simple and understandable structure of the used models.