DeepFund

DeepFund evaluates an LLM on the whole active fund-management loop rather than on an isolated prediction: analyst signals, planning, and portfolio decisions, run against real market data for a chosen trading date that falls after the model’s training cutoff. Runs are configured in YAML, with a planner_mode switch deciding the shape of the workflow — when true, a Planner agent orchestrates which analysts run; when false, every workflow analyst runs in parallel. Output lands in four tables: Config, Portfolio, Decision and Signal. Market data arrives through pluggable providers behind a router in src/apis — akshare, Alpha Vantage, financialdataset, JoinQuant and yfinance.

Running it takes real credentials. The environment template expects OpenAI, Anthropic, DeepSeek, Qwen and Zhipu keys alongside an Alpha Vantage key, and Supabase is the default database, so a fully local run means passing --local-db against a local SQLite file. The benchmark is simulated throughout: the README states plainly that the project is for research purposes only and does not trade.

The paper is an accepted NeurIPS 2025 Datasets and Benchmarks poster, confirmed on the conference’s own listing (arXiv:2505.11065); the authors evaluate nine flagship LLMs and report that even strong models such as DeepSeek-V3 and Claude-3.7-Sonnet take net trading losses inside the live environment. Two caveats accompany the repository. It is dormant — the last commit is March 2026 — and the live arena leaderboard the README advertises no longer resolves, so that demonstration cannot currently be used or checked. Separately, the README’s claims of an IJCAI 2025 FinLLM Workshop oral award for the companion vision paper and of a best-open-source award are author-stated and were not independently confirmed. The code itself remains MIT-licensed and free, alongside a 2026 announcement of collaboration with a fintech startup.