⚡ Tyche

An AI-powered Reinforcement Learning framework for high-frequency trading research and market simulation.

Tyche is a research-focused quantitative trading system that trains a PPO (Proximal Policy Optimization) reinforcement learning agent on high-frequency cryptocurrency market data. The project includes a complete training pipeline, backtesting engine, real-time monitoring dashboard, persistent model storage, and market stress testing for evaluating trading strategies under different conditions.

---

Features

• 📈 Reinforcement Learning based trading agent (PPO)
• ⚡ High-frequency trading environment using 1-second OHLCV data
• 📊 Interactive real-time dashboard built with Streamlit
• 🧠 Automatic model checkpointing and persistent training
• 💾 MongoDB integration for trade and episode storage
• 📉 Historical backtesting engine
• 🌪️ Adversarial market simulation (flash crashes, volatility spikes, etc.)
• 📦 Modular architecture for experimentation
• 🚀 GPU acceleration with PyTorch (CUDA supported)

---

Architecture

                 Historical Market Data
                          │
                          ▼
                Data Download & Loader
                          │
                          ▼
             Custom Trading Environment
                          │
                          ▼
               PPO Reinforcement Agent
                          │
                          ▼
              Training / Learning Loop
                          │
         ┌────────────────┴────────────────┐
         │                                 │
         ▼                                 ▼
 Model Checkpoints                  Trade History
 Training Logs                      MongoDB Storage
         │                                 │
         └────────────────┬────────────────┘
                          ▼
                Flask REST API Server
                          │
                          ▼
               Streamlit Live Dashboard

---

Project Structure

Tyche/
│
├── dashboard_app.py          # Live monitoring dashboard
├── server.py                 # Flask API
├── main.py                   # Training entry point
├── backtest.py               # Performance evaluation
├── download_data.py          # Dataset downloader
├── data_engine.py            # Market data processing
├── generate_weights.py       # Initial model generation
├── bootstrap_weights.py      # Weight initialization
├── restore.py                # Restore checkpoints
├── inject.py                 # Utility functions
├── requirements.txt
│
├── memory/
│   ├── agent_weights.pt
│   ├── best_weights.pt
│   ├── training_log.json
│   ├── trainer_status.json
│   └── recent_trades.json
│
└── tyche/
    ├── trainer.py
    ├── environment.py
    ├── agent_gpu.py
    ├── adversarial.py
    ├── mongo_store.py
    ├── data_loader.py
    └── config.py

---

Technologies Used

• Python
• PyTorch
• Stable-Baselines3
• Streamlit
• Flask
• MongoDB
• Pandas
• NumPy
• Matplotlib
• yFinance
• Binance Historical Data

---

Trading Environment

The environment simulates high-frequency cryptocurrency trading using:

• 1-second OHLCV candles
• Portfolio management
• Transaction fees
• Position sizing
• Multiple tradable assets
• Reward shaping
• Observation vectors
• Action space optimized for RL

---

Dashboard

The Streamlit dashboard provides live visualization of:

• Portfolio value
• Current profit/loss
• Agent status
• Trade history
• Asset prices
• Episode statistics
• Database connectivity
• Recent trading activity

---

Reinforcement Learning

Tyche uses Proximal Policy Optimization (PPO) to train an autonomous trading agent.

Training pipeline includes:

• Environment interaction
• Reward optimization
• Policy updates
• Automatic checkpoint saving
• Continuous learning
• GPU acceleration (CUDA)

---

Backtesting

The backtesting module allows trained models to be evaluated on historical market data.

Performance metrics include:

• Equity curve
• Cumulative profit
• Episode reward
• Portfolio growth
• Historical trade replay

---

Data Pipeline

Market data can be collected from:

• Binance historical datasets
• Yahoo Finance
• Custom CSV datasets

Data is processed into feature-rich observations before being fed into the reinforcement learning environment.

---

Installation

Clone the repository

git clone https://github.com/yourusername/Tyche.git
cd Tyche

Install dependencies

pip install -r requirements.txt

---

Running the Project

Start training

python main.py

Launch the dashboard

streamlit run dashboard_app.py

Run the Flask API

python server.py

Run a backtest

python backtest.py

---

Persistent Training

Tyche automatically saves:

• Latest model weights
• Best-performing weights
• Recent trades
• Episode logs
• Training statistics

This allows interrupted training sessions to resume without losing progress.

---

Research Goals

This project is intended for experimentation in:

• Reinforcement Learning
• Quantitative Finance
• High-Frequency Trading
• Portfolio Optimization
• AI Decision Making
• Market Simulation
• Financial Machine Learning

---

Disclaimer

This project is intended for educational and research purposes only.

It is not financial advice and should not be used for live trading without extensive testing, validation, and risk management.

---

Future Improvements

• Live paper trading
• Multi-agent reinforcement learning
• Transformer-based market prediction
• Hyperparameter optimization
• Risk-adjusted reward functions
• WebSocket market streaming
• Multi-exchange support
• Docker deployment
• Distributed training

---

Author

Ayush Rai

Computer Science Student • AI & Machine Learning • Quantitative Finance • Reinforcement Learning

---

If you found this project interesting, consider giving it a ⭐.