Reinforcement Learning for Finance

Chapter 3. Financial Q-Learning

Today’s algorithmic trading programs are relatively simple and make only limited use of AI. This is sure to change.

Murray Shanahan (2015)

The previous chapter shows that a deep Q-learning (DQL) agent can learn to play the game of CartPole quite well. What about financial applications? As this chapter shows, the agent can also learn to play a financial game that is about predicting the future movement in a financial market. To this end, this chapter implements a Finance environment that mimics the behavior of the CartPole environment and trains the DQL agent from the previous chapter based on the requirements of the Finance environment.

This chapter is brief, but it illustrates an important point: with the appropriate environment, DQL can be applied to financial problems basically in the same way as it is applied to games and in other domains. “Finance Environment” develops step-by-step the Finance class that mimics the behavior of the CartPole class. “DQL Agent” slightly adjusts the DQLAgent class from “CartPole as an Example”. The adjustments are made to reflect the new context. The DQL agent can learn to predict future market movements with a significant margin over the baseline accuracy of 50%. “Where the Analogy Fails” finally discusses the major issues of the modeling approach and the Finance class when compared, for example, to a gaming environment such as the CartPole game.

Finance Environment

The goal in this section is to implement ...