Abstract

Reinforcement learning techniques have raised attention from financial industry, especially by employing reinforcement learning in portfolio managements. In this project, we explored three state-of-art reinforcement learning algorithms, including policy gradient (PG), deep deterministic policy gradient (DDPG) and proximal policy optimization (PPO). The goal is to train an intelligent agent that can continuously trade in stock market by allocating on different assets. In the course of this project, we performed hyper-parameter tuning and feature selection for each algorithm. Furthermore, we compared the return performance of different algorithms using both U.S. and China stock market data. As shown in the project report, we formally define the portfolio management problem as a Markov decision process, describe our detailed methodology, and demonstrate our experimental comparison and results.