This project aims to develop a reinforcement learning (RL) agent capable of learning stock market trading strategies, specifically focusing on "reading the tape." It's targeted at individuals interested in applying RL to financial markets, offering a practical exploration of RL concepts in a real-world, albeit simulated, trading environment.

How It Works

The project utilizes Deep Q-Networks (DQN) and Policy Gradients (PG) for trading. The agent learns by making decisions (hold, buy, short) based on historical stock data. Training is episodic, with rewards calculated based on the final portfolio value at the end of each episode, simplifying reward calculation compared to step-by-step rewards. The architecture starts with simple feed-forward networks, with potential for convolutional networks, using ReLU activations in hidden layers and sigmoid output for action probabilities.

Quick Start & Requirements

• Install/run: Navigate to tensor-reinforcement directory and run python dqn_model.py or python pg_model.py.
• Prerequisites: Requires data from provided Google Drive links (stock market data, Nifty data, Nifty futures) to be copied into the tensor-reinforcement directory. A saved_networks directory must be created.
• Links:
  • DQN/PG code: https://github.com/deependersingla/deep_trader/tree/master/tensor-reinforcement
  • Thought journal: https://github.com/deependersingla/deep_trader/blob/master/deep_thoughts.md
  • Data: https://drive.google.com/file/d/0B6ZrYxEMNGR-MEd5Ti0tTEJjMTQ/view, https://drive.google.com/file/d/0B6ZrYxEMNGR-Q0YwWWVpVnJ3YmM/view?usp=sharing

Highlighted Details

• Explores both DQN and Policy Gradient RL algorithms.
• Uses episodic training with terminal rewards for simplified learning.
• Discusses potential use of feed-forward and convolutional neural networks.
• References foundational RL literature and AlphaGo's impact on the project's direction.

Maintenance & Community

The project's author states they are no longer able to answer questions due to running a company on RL trading. No active community channels or recent development activity are indicated.

Licensing & Compatibility

The README does not explicitly state a license. The project's code is available on GitHub, implying a permissive license, but this should be verified.

Limitations & Caveats

The project is presented as a personal development journal with the author no longer available for support. The data sources are from 2016, and the code may not be compatible with current TensorFlow versions or market conditions.