Based on repository activity, growth velocity and community engagement.
50
Growth16/30
Activity7/25
Popularity15/25
Trust13/20
2,800
Stars
High
Sentiment
Votes
2,800
README.md
<div align="center">
<h2>QuantHarness: Price-Driven Multi-Agent LLMs for High-Frequency Trading</h2>
</div>
<div align="center">
<div style="position: relative; text-align: center; margin: 20px 0;">
<div style="position: absolute; top: -10px; right: 20%; font-size: 1.2em;"></div>
<p>
<a href="https://machineily.github.io/">Fei Xiong</a><sup>1,2 ★</sup>
<a href="https://wyattz23.github.io">Xiang Zhang</a><sup>3 ★</sup>
<a href="https://scholar.google.com/citations?user=hFhhrmgAAAAJ&hl=en">Aosong Feng</a><sup>4</sup>
<a href="https://intersun.github.io/">Siqi Sun</a><sup>5</sup>
<a href="https://chenyuyou.me/">Chenyu You</a><sup>1</sup>
</p>
<p>
<sup>1</sup> Stony Brook University
<sup>2</sup> Carnegie Mellon University
<sup>3</sup> University of British Columbia <br>
<sup>4</sup> Yale University
<sup>5</sup> Fudan University
★ Equal Contribution <br>
</p>
</div>
<div align="center" style="margin: 20px 0;">
<a href="README.md">English</a> | <a href="README_CN.md">中文</a>
</div>
<br>
<p align="center">
<a href="https://arxiv.org/abs/2509.09995">
<img src="https://img.shields.io/badge/💡%20ArXiv-2509.09995-B31B1B?style=flat-square" alt="Paper">
</a>
<a href="https://Y-Research-SBU.github.io/QuantHarness">
<img src="https://img.shields.io/badge/Project-Website-blue?style=flat-square&logo=googlechrome" alt="Project Website">
</a>
</p>
</div>
A sophisticated multi-agent trading analysis system that combines technical indicators, pattern recognition, and trend analysis using LangChain and LangGraph. The system provides both a web interface and programmatic access for comprehensive market analysis.
<!-- - **Multi-Agent Analysis**: Four specialized agents working together: -->
Indicator Agent
• Computes five technical indicators—including RSI to assess momentum extremes, MACD to quantify convergence–divergence dynamics, and the Stochastic Oscillator to measure closing prices against recent trading ranges—on each incoming K‑line, converting raw OHLC data into precise, signal-ready metrics.
Pattern Agent
• Upon a pattern query, the Pattern Agent first uses the agent draws the recent price chart, spots its main highs, lows, and general up‑or‑down moves, compares that shape to a set of familiar patterns, and returns a short, plain‑language description of the best match.
Trend Agent
• Leverages tool-generated annotated K‑line charts overlaid with fitted trend channels—upper and lower boundary lines tracing recent highs and lows—to quantify market direction, channel slope, and consolidation zones, then delivers a concise, professional summary of the prevailing trend.
Decision Agent
• Synthesizes outputs from the Indicator, Pattern, Trend, and Risk agents—including momentum metrics, detected chart formations, channel analysis, and risk–reward assessments—to formulate actionable trade directives, clearly specifying LONG or SHORT positions, recommended entry and exit points, stop‑loss thresholds, and concise rationale grounded in each agent’s findings.
# For OpenAI
export OPENAI_API_KEY="your_openai_api_key_here"
# For Anthropic (Claude)
export ANTHROPIC_API_KEY="your_anthropic_api_key_here"
# For Qwen (DashScope, based in Singapore — delays may occur)
export DASHSCOPE_API_KEY="your_dashscope_api_key_here"
# For MiniMax (204K context, OpenAI-compatible API)
export MINIMAX_API_KEY="your_minimax_api_key_here"
🚀 Usage
Start the Web Interface
python web_interface.py
The web application will be available at http://127.0.0.1:5000
Web Interface Features
Asset Selection: Choose from available stocks, crypto, commodities, and indices
Timeframe Selection: Analyze data from 1-minute to daily intervals
Date Range: Select custom date ranges for analysis
Real-time Analysis: Get comprehensive technical analysis with visualizations
API Key Management: Update your OpenAI API key through the interface
📺 Demo
🔧 Implementation Details
Important Note: Our model requires an LLM that can take images as input, as our agents generate and analyze visual charts for pattern recognition and trend analysis.
Python Usage
To use QuantHarness inside your code, you can import the trading_graph module and initialize a TradingGraph() object. The .invoke() function will return a comprehensive analysis. You can run web_interface.py, here's also a quick example:
from trading_graph import TradingGraph
# Initialize the trading graph
trading_graph = TradingGraph()
# Create initial state with your data
initial_state = {
"kline_data": your_dataframe_dict,
"analysis_results": None,
"messages": [],
"time_frame": "4hour",
"stock_name": "BTC"
}
# Run the analysis
final_state = trading_graph.graph.invoke(initial_state)
# Access results
print(final_state.get("final_trade_decision"))
print(final_state.get("indicator_report"))
print(final_state.get("pattern_report"))
print(final_state.get("trend_report"))
You can also adjust the default configuration to set your own choice of LLMs or analysis parameters in web_interface.py.
if provider == "anthropic":
# Set default Claude models if not already set to Anthropic models
if not analyzer.config["agent_llm_model"].startswith("claude"):
analyzer.config["agent_llm_model"] = "claude-haiku-4-5-20251001"
if not analyzer.config["graph_llm_model"].startswith("claude"):
analyzer.config["graph_llm_model"] = "claude-haiku-4-5-20251001"
elif provider == "qwen":
# Set default Qwen models if not already set to Qwen models
if not analyzer.config["agent_llm_model"].startswith("qwen"):
analyzer.config["agent_llm_model"] = "qwen3-max"
if not analyzer.config["graph_llm_model"].startswith("qwen"):
analyzer.config["graph_llm_model"] = "qwen3-vl-plus"
elif provider == "minimax":
# Set default MiniMax models (204K context window)
if not analyzer.config["agent_llm_model"].startswith("MiniMax"):
analyzer.config["agent_llm_model"] = "MiniMax-M3"
if not analyzer.config["graph_llm_model"].startswith("MiniMax"):
analyzer.config["graph_llm_model"] = "MiniMax-M3"
else:
# Set default OpenAI models if not already set to OpenAI models
if analyzer.config["agent_llm_model"].startswith(("claude", "qwen", "MiniMax")):
analyzer.config["agent_llm_model"] = "gpt-4o-mini"
if analyzer.config["graph_llm_model"].startswith(("claude", "qwen", "MiniMax")):
analyzer.config["graph_llm_model"] = "gpt-4o"
For live data, we recommend using the web interface as it provides access to real-time market data through yfinance. The system automatically fetches the most recent 30 candlesticks for optimal LLM analysis accuracy.
Configuration Options
The system supports the following configuration parameters:
agent_llm_model: Model for individual agents (default: "gpt-4o-mini")
graph_llm_model: Model for graph logic and decision making (default: "gpt-4o")
agent_llm_temperature: Temperature for agent responses (default: 0.1)
graph_llm_temperature: Temperature for graph logic (default: 0.1)
Note: The system uses default token limits for comprehensive analysis. No artificial token restrictions are applied.
You can view the full list of configurations in default_config.py.
🤝 Contributing
Fork the repository
Create a feature branch
Make your changes
Add tests if applicable
Submit a pull request
📄 License
This project is licensed under the MIT License - see the LICENSE file for details.
🔖 Citation
@article{xiong2025quantharness,
title={QuantHarness: Price-Driven Multi-Agent LLMs for High-Frequency Trading},
author={Fei Xiong and Xiang Zhang and Aosong Feng and Siqi Sun and Chenyu You},
journal={arXiv preprint arXiv:2509.09995},
year={2025}
}
🙏 Acknowledgements
This repository was built with the help of the following libraries and frameworks:
This software is for educational and research purposes only. It is not intended to provide financial advice. Always do your own research and consider consulting with a financial advisor before making investment decisions.
🐛 Troubleshooting
Common Issues
TA-Lib Installation: If you encounter TA-Lib installation issues, refer to the official repository for platform-specific instructions.
LLM API Key: Ensure your API key is properly set in the environment or through the web interface.
Data Fetching: The system uses Yahoo Finance for data. Some symbols might not be available or have limited historical data.
Memory Issues: For large datasets, consider reducing the analysis window or using a smaller timeframe.
Support
If you encounter any issues, please:
Try refresh and re-enter LLM API key
Check the troubleshooting section above
Review the error messages in the console
Ensure all dependencies are properly installed
Verify your API key is valid and has sufficient credits
📧 Contact
For questions, feedback, or collaboration opportunities, please contact:
