Summary: DeepFund tackles the critical question of whether Large Language Models (LLMs) can achieve professional proficiency in fund investment. It offers a comprehensive, unified research environment designed to rigorously evaluate LLM trading capabilities across a spectrum of financial markets. The system empowers LLMs to ingest diverse external information, orchestrate sophisticated multi-agent systems, and formulate actionable trading decisions within a simulated arena. This platform serves as an invaluable tool for researchers and engineers aiming to benchmark and advance AI applications in quantitative finance and algorithmic trading.

How It Works: The core architecture of DeepFund centers on an LLM acting as a sophisticated orchestrator. This LLM ingests a variety of external financial data streams and market signals, then leverages this information to drive a dynamic multi-agent system. These agents collaborate to analyze market conditions, identify opportunities, and generate trading strategies. The entire process culminates in a simulated trading arena, providing a standardized and reproducible environment for evaluating LLM performance, decision-making logic, and overall trading acumen in complex, simulated financial scenarios. The project is structured for progressive development, aiming towards production readiness with a current self-rated TRL of 4.

Quick Start & Requirements: To initiate, clone the repository (git clone https://github.com/HKUSTDial/DeepFund.git) and navigate into the project directory. Environment setup can be managed via Conda (conda env create -f environment.yml) or uv (uv sync), followed by virtual environment activation. Crucially, users must configure API keys for LLM providers (e.g., OpenAI, DeepSeek) within a .env file. DeepFund supports both Supabase (default, requiring SUPABASE_URL and SUPABASE_KEY) and local SQLite databases (activated with the --local-db flag and set up via src/database/sqlite_setup.py). The system is launched from the src directory using python main.py --config xxx.yaml --trading-date YYYY-MM-DD [--local-db].

Highlighted Details:

• Accepted as a poster presentation at NeurIPS 2025.
• Received an Oral Paper Award at the IJCAI 2025 FinLLM Workshop for its vision paper.
• Recognized with the Best Open-source Award at the AI Agent 2025 competition.
• Features an interactive DeepFund & Arena Leaderboard for demonstration purposes.
• Designed for TRL 1-9 progression, currently self-assessed at TRL 4.

Maintenance & Community: The project actively welcomes collaborations and general inquiries via email, with specific contacts provided for general questions (cli942@connect.hkust-gz.edu.cn) and research partnerships (yuyuluo@hkust-gz.edu.cn). While explicit community channels like Discord or Slack are not detailed, the project's stated goal of progressing through TRL levels suggests an ongoing development and engagement strategy.

Licensing & Compatibility: The specific open-source license governing the DeepFund project, along with any associated compatibility notes for commercial use or integration into closed-source systems, is not explicitly stated in the provided README documentation.

Limitations & Caveats: DeepFund is designated strictly for research purposes and does not facilitate or execute real-world financial trading. LLM performance is evaluated within a simulated trading environment, and its outcomes are inherently subject to the limitations and assumptions of such simulations, rather than direct real-market validation. The project's current self-rated TRL of 4 indicates it is in an experimental phase, not yet ready for production deployment.