Chart to Code

A self-hosted trading assistant powered by a trained vision-language model (VLM) for trend analysis and algorithmic trading decisions.

• PyPI package: https://pypi.org/project/chart-to-code/
• Repository: https://github.com/Ruby-xantho/chart-to-code
• Free software: MIT License
• Documentation: https://github.com/Ruby-xantho/chart-to-code#readme

About

Chart-to-Code is a vision-language model trained specifically for financial chart analysis and trend trading. The system analyzes live market data, generates technical analysis charts, and provides trading signals through a trained VLM that understands price action, momentum indicators, and market trends.

Key Features

• Custom-Trained VLM: Fine-tuned vision-language model specialized for financial chart analysis
• Real-Time Market Analysis: Live data fetching from cryptocurrency exchanges via CCXT
• Multi-Panel Chart Generation:
  • Price charts with moving averages (EMA13, EMA21, SMMA14)
  • Awesome Oscillator momentum analysis
  • Stochastic RSI indicators
• Automated Trading Signals: AI-generated buy/sell/hold recommendations
• Rule-Based Validation: Hybrid approach combining ML predictions with algorithmic rules
• Interactive Web Interface: Streamlit-based dashboard for real-time analysis
• Exchange Integration Ready: Framework prepared for future live trading connections

Technical Architecture

The system combines several components:

• VLM Core: Custom-trained Qwen2.5-VL model for chart interpretation
• Data Pipeline: Real-time market data via CCXT (Binance, others)
• Chart Generation: Multi-panel technical analysis visualization
• Signal Processing: Rule engine for trade signal validation
• Web Interface: Streamlit app for user interaction

Installation

Basic Installation

git clone https://github.com/Ruby-xantho/chart-to-code.git
cd chart-to-code
pip install -e .

If you have access to a NVIDIA GPU like an A40, you can install it this way to self-host the model.

pip install -e ".[gpu]"

Requirements

• Python: 3.10+
• For local hosting the model and running it with VLM inference: NVIDIA GPU with ≥32 GB VRAM (A40, A100, etc.)
• Memory: ≥4 GB RAM for basic usage, ≥32 GB for model training and self hosting the model

Usage

Run the Trading Assistant

Make sure you're in the chart-to-code directory and run:

streamlit run app/trading_assistant.py

The web interface allows you to:

1. Select cryptocurrency pairs for analysis
2. View real-time technical analysis charts
3. Get AI-powered trading signals
4. Monitor multiple assets simultaneously

Supported Markets

Currently supports cryptocurrency markets via CCXT:

• Bitcoin (BTC/USDT)
• Ethereum (ETH/USDT)
• Major altcoins (SOL, ADA, LINK, etc.)

Model Training

The VLM was trained on a custom dataset of:

• 1000+ labeled chart examples
• Multiple timeframes (1h, 4h, 1d)
• Five signal categories: Sell Signal, Possible Buy Entry, Bullish, Bearish, Inconclusive
• Rule-based ground truth labels for supervised learning

Project Structure

chart-to-code/
├── src/chart_to_code/     # Core package
├── app/                   # Streamlit web application
├── model_training/        # VLM training pipeline
├── exchange_bots/         # Trading bot implementations
├── validate_model/        # Model validation tools
└── tests/                 # Test suite

Future Development

• Live Trading Integration: Connect to exchange APIs for automated execution
• Additional Markets: Expand beyond crypto to forex, stocks, commodities
• Advanced Strategies: Multi-timeframe analysis and complex trading logic
• Portfolio Management: Risk management and position sizing
• Backtesting Engine: Historical strategy performance analysis

Contributing

Contributions are welcome! Please read CONTRIBUTING.md for development setup and guidelines.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Disclaimer

This software is for educational and research purposes. Trading involves significant risk of loss. Always do your own research and never invest more than you can afford to lose.

This package was created with Cookiecutter and the audreyfeldroy/cookiecutter-pypackage project template.