The purpose of this repo is to determine the daily action and make the best profit for the future trading in the stock market.
The training_data.csv and testing_data.csv are provided. Aside from the given open-high-low-close, I add additional one-hot encoded features to provide the information of holding or shorting stock and two-days-ahead prediction for more precise decision.
According to the instructions, the aim is to maximize the resulting accumulated profit via deciding the Buy, NoAction, and Sell actions based on the current open-high-low-close prices. Thus I turned to reinforcement learning (RL), which learns a policy through the resulting rewards. The deep q network (DQN) is a RL technique with Bellman equation as the update function and is suitable for continuous state space. Gated recurrent unit (GRU) is used in the DQN for modeling the q-values.
The accumulated profit over each episode are drawn below: