GMGN Trading Bot

Automated cryptocurrency trading bot that learns optimal buy/sell strategies using reinforcement learning. The bot connects to GMGN.ai to monitor new token launches and makes trading decisions based on a PPO-trained agent.

Project Overview

This project consists of three main components:

1. logger_c/ - Real-time C-based token logger that connects to GMGN.ai WebSocket API
2. ai_model/ - Reinforcement learning trading agent with curriculum learning
3. trading_algorithm/ (Future) - Production trading system integration

Quick Start

Training the RL Agent

The AI model uses reinforcement learning (PPO) with curriculum learning to discover profitable trading patterns.

cd ai_model
python src/data/preprocess.py  # Prepare training data
# Then use train_rl_agent.ipynb in Google Colab for GPU training

See ai_model/README.md for detailed instructions.

Running the Token Logger

The C-based logger connects to GMGN.ai and streams new token data:

cd logger_c
make
./build/gmgn_logger

See logger_c/README.md for configuration options.

AI Model Approach

Why Reinforcement Learning?

Initial experiments with classification (predicting BUY/SELL/HOLD labels) failed due to:

• Unrealistic label generation (64% of data labeled as SELL)
• Features couldn't predict 20-second future price movements
• Model accuracy was worse than random (14-34% vs 33% baseline)

Solution: Reinforcement learning lets the agent discover profitable patterns through trial and error, rather than fitting to hand-crafted labels.

Key Features

• Curriculum Learning: Gradually increases trading fees from 0% to 100% during training
• Hindsight Rewards: Penalizes agent for missing profitable opportunities
• Asymmetric Rewards: Bigger bonus for winning trades than penalty for losses
• PPO Algorithm: Proximal Policy Optimization with high entropy (0.05) for exploration

Training Results

The agent learns to:

• Make active trades (prevents "do nothing" policy collapse)
• Identify profitable entry/exit points
• Manage risk with proper position sizing

Training takes ~1M steps (~2-3 hours on T4 GPU).

Project Structure

gmgn_trading/
├── logger_c/              # C-based WebSocket token logger
│   ├── src/               # Source files
│   ├── include/           # Headers
│   └── Makefile           # Build configuration
├── ai_model/              # RL trading agent
│   ├── data/              # Training data
│   ├── notebooks/         # Jupyter training notebooks
│   ├── src/
│   │   ├── config/        # Configuration
│   │   ├── data/          # Data loading and preprocessing
│   │   ├── rl/            # RL environment and training
│   │   └── utils/         # Helper functions
│   └── models/            # Saved models
├── trading_algorithm/     # (Future) Production trading
└── logger/                # Legacy JS logger

Dependencies

AI Model

• Python 3.10+
• PyTorch 2.0+
• Stable-Baselines3
• Gymnasium
• NumPy, Pandas

Logger

• C compiler (gcc/clang)
• libwebsockets 4.x
• JSON-C

Development Guidelines

This project follows strict C coding standards documented in CLAUDE.md:

• Snake_case naming for functions/variables
• Comprehensive documentation for all functions
• Memory safety and leak prevention
• Professional code quality

Known Issues & Lessons Learned

GMGN WebSocket API

Critical discoveries documented in CLAUDE.md:

• Correct subscription format requires action, f, and id fields
• Connection URL: wss://ws.gmgn.ai/quotation
• Use new_pool_info channel for new token launches

Training Issues Solved

1. Policy Collapse: Agent learned to do nothing → Fixed with curriculum learning
2. Poor Labels: Classification failed → Switched to RL
3. Low Exploration: Added higher entropy coefficient (0.05)

Future Work

• Production trading algorithm integration
• Real-time inference pipeline
• Risk management system
• Multi-DEX support beyond Pump.fun
• Advanced reward shaping with Sharpe ratio optimization

License

MIT License - See LICENSE file for details

Contributing

Contributions welcome! Please follow the coding standards in CLAUDE.md.

Disclaimer

This is experimental software for educational purposes. Cryptocurrency trading carries significant risk. Use at your own risk.