A Q-learning approach for investment decisions

Abstract

This work deals with the application of the Q-learning technique in order to make investment decisions. This implies to give investment recommendations about the convenience of investment on a particular asset. The reinforcement learning system, and particularly Q-learning, allows continuous learning based on decisions proposed by the system itself. This technique has several advantages, like the capability of decision-making independently of the learning stage, capacity of adaptation to the application domain, and a goal-oriented logic. These characteristics are very useful on financial problems. Results of experiments made to evaluate the learning capacity of the method in the mentioned application domain are presented. Decision-making capacity on this domain is also evaluated. As a result, a system based on Q-learning that learns from its own decisions in an investment context is obtained. The system presents some limitations when the space of states is big due to the lack of generalization of the Q-learning variant used.