This repository contains the code used to apply Deep Reinforcement Learning (DRL) to the problem of financial trading. The main contribution is the use of C51 algorithm (traditionally used for reinforcement learning on the Atari games) in a problem that has an inherently recurrent component: more specifically, the architecture of the DRQN (Deep Recurrent Q-Network) has been modified to support the C51 algorithm by adding another linear layer before the LSTM module.
The data was taken from Lixu Wang - DRQN Stock Trading: starting from his work, I organized the repository in a different way and I added all that concerns the C51 algorithm. I refer to his repository for the original code, which was the implementation of the paper "Financial Trading as a Game: A Deep Reinforcement Learning Approach", by Huang C.Y.
The algorithm was tested on 10000 episodes and, in some cases, there was an increase of 38% in the model performances with respect to the author of the code, as shown below.
