README
Overview
This project implements a Stock Trading Bot, trained using Deep Reinforcement Learning, specifically Deep Q-learning. Implementation is kept simple and as close as possible to the algorithm discussed in the paper, for learning purposes.
Introduction
Generally, Reinforcement Learning is a family of machine learning techniques that allow us to create intelligent agents that learn from the environment by interacting with it, as they learn an optimal policy by trial and error. This is especially useful in many real world tasks where supervised learning might not be the best approach due to various reasons like nature of task itself, lack of appropriate labelled data, etc.
The important idea here is that this technique can be applied to any real world task that can be described loosely as a Markovian process.
Approach
This work uses a Model-free Reinforcement Learning technique called Deep Q-Learning (neural variant of Q-Learning). At any given time (episode), an agent abserves it's current state (n-day window stock price representation), selects and performs an action (buy/sell/hold), observes a subsequent state, receives some reward signal (difference in portfolio position) and lastly adjusts it's parameters based on the gradient of the loss computed.
There have been several improvements to the Q-learning algorithm over the years, and a few have been implemented in this project:
- Vanilla DQN
- DQN with fixed target distribution
- Double DQN
- Prioritized Experience Replay
- Dueling Network Architectures
Results
Trained on GOOG 2010-17 stock data, tested on 2019 with a profit of $1141.45 (validated on 2018 with profit of $863.41):
You can obtain similar visualizations of your model evaluations using the notebook provided.