BTGym is a Python library designed for reinforcement learning (RL) research in algorithmic trading. It provides a scalable, event-driven backtesting framework that integrates with the OpenAI Gym API, enabling researchers and developers to build and test RL agents in simulated trading environments. The library aims to facilitate experimentation with complex, non-stationary stochastic environments common in financial markets.

How It Works

BTGym builds upon the backtrader library, leveraging its event-driven architecture for backtesting. It wraps trading environments as OpenAI Gym-compatible environments, allowing seamless integration with standard RL algorithms and libraries. The framework supports both discrete and continuous action spaces, modeling trading scenarios with riskless and risky assets, transaction costs, and optional exogenous data feeds. Data sampling strategies include random, sequential, and sliding time-window approaches to manage non-stationarity and improve data efficiency.

Quick Start & Requirements

• Install via pip: pip install -e . after cloning the repository.
• Requires Matplotlib version 2.0.2.
• LSOF utility must be installed on the OS.
• See Documentation and Examples.

Highlighted Details

• Supports discrete actions (hold, buy, sell, close) and continuous actions for portfolio optimization.
• Offers multiple data sampling methods (random, sequential, sliding window) for training.
• Integrates with RL algorithms like A3C, UNREAL, and PPO.
• Features multi-asset trading capabilities and multi-modal observation spaces.

Maintenance & Community

• Active development with frequent updates noted in the README.
• Community support available via a Slack channel.

Licensing & Compatibility

• The README does not explicitly state a license. Users should verify licensing for commercial or closed-source use.

Limitations & Caveats

• The code is research/development grade, potentially unstable, and may have bugs.
• Not optimized for Windows.
• Multi-asset continuous action spaces are in beta and may have convergence issues.
• Resource-intensive, especially for multi-asset setups.