Systematic Trading Strategies with Reinforcement Learning

Abstract

Intertemporal choice is fundamental in many economic and financial decisionmaking problems. In this thesis we first present reinforcement learning theory and then show how it can be used as a tool to model and approach two of these problems fully automated. The first is the classical utility maximization problem framed as an investor performing portfolio optimization trading an equity index. Here we show that a risk-averse agent tends towards a classic buy and hold strategy the longer it trains. The second is a method to develop automated market-making trading strategies. We do this by simulating a stock market in an agent-based model, which lets us mitigate some common assumptions, such as no market impact and the absence of transaction cost, as well as model the dynamics of the order book. Here we see that the agent learns and improves its performance through time, but slowly. And due to the lag of computing power, we have not been able to run the experiment for as long as desired