AI Hedge Fund

What it is

AI Hedge Fund is a proof of concept by Virat Singh (virattt) that explores using AI to make trading decisions. You install it as the aihf command, and it opens an interactive terminal app. A fund there is a set of strategies. Each strategy blends the views of analysts, which are LLM investor agents such as Warren Buffett and Charlie Munger, or quant models such as post-earnings drift. The README says the project is for educational purposes only and that the system does not actually make any trades.

1. Point-in-time market and fundamentals data
2. Analysts give each stock a conviction and a thesis
3. The portfolio blends their views into target weights
4. Hard risk limits clamp the book
5. The broker fills the orders at the next close
6. Paper funds append the session to a hash-chained ledger

Who it suits

Developers, students and quant hobbyists who want to see how an LLM-plus-quant trading system is built, run it on a few tickers and read or extend its code. The roadmap invites new analysts, strategies and data sources. Skip it if you want a bot that trades for you. The README says it is not for real trading, and live brokers are only planned.

Where to use it

It runs in a terminal on Python 3.11 or newer. The maker’s README and docs give no steps for Claude Code, Codex, Cursor or other coding agents, and no skill, plugin or MCP server.

• Terminal app: aihf opens an interactive app for paper trading and backtesting
• Command line: the engine README shows aihf backtest and aihf paper commands, run through Poetry from a checkout, and the roadmap calls the CLI a thin machine client with JSON on stdout
• Model providers: Anthropic, OpenAI, DeepSeek, Google, xAI, Kimi and TypeSafe (Jev)
• Market data: Financial Datasets, with an API key, or on the main branch x402 pay-per-call from a Base wallet (last 365 days of data, non-professional use)

How to set it up

Get it saves a zip of Minda’s setup notes in your downloads folder, with the install and key steps from the maker’s README and docs. You install the aihf package with pipx, uv or pip, then run it.

1. Run the install line from the README.
2. Run aihf to open the app.
3. Enter the keys when it asks. It saves them to ~/.hedge-fund/.env, and keys exported in your shell always win over that file.

pipx install aihf
aihf

The README also gives uv tool install aihf, and pip install aihf into an environment of your choice. For the x402 option the README on the main branch gives this line. The PyPI release, aihf 2.6.4, lists no x402 extra and its README does not mention x402, so until a release adds it, the x402 option is only on the repository’s main branch.

pip install 'aihf[x402]'

How to use it

Start with aihf and pick a mode. In the paper trading list, highlight a fund to see what its next run will do, then press enter to run it through the next completed session. An approval step shows the exact decision before anything trades.

Fund definitions live in ~/.hedge-fund/mandates/ and carry no tickers. You pick the universe per backtest, or when you deploy a paper fund. You can build a new fund by choosing its strategies, capital, cadence and tickers.

Example

You want to see how a value-and-drift fund behaves. The repository’s sample mandate has two strategies. Deep-value is a long-only pod led by the Graham agent, with Buffett and Munger beside it. Earnings-drift is a systematic pod that runs the post-earnings drift model, long and short. The fund starts with 100000 in capital, rebalances weekly, is measured against SPY and lets no single name pass 25% of equity.

You point it at tickers and a window. The engine README gives this backtest line, written for a checkout. With an x402 wallet this 2024 window fails: x402 serves only the last 365 days, and older data needs a Financial Datasets API key.

poetry run aihf backtest ~/.hedge-fund/mandates/example.yaml \
    --universe AAPL,MSFT,NVDA --start 2024-01-02 --end 2024-06-28

It prints the full result JSON, a record for every session, and saves a copy to ~/.hedge-fund/research/. The docs do not say whether an install puts example.yaml in ~/.hedge-fund/mandates/. The repository’s sample mandate is hedge_fund/fund/example.yaml. To paper trade the same fund, create it and tick it once per completed session.

poetry run aihf paper create alpha --mandate ~/.hedge-fund/mandates/example.yaml --universe AAPL,MSFT
poetry run aihf paper tick alpha

A tick that finds a new completed session adds one record to the fund’s ledger.

For newcomers and experts

If you are new. In this app a mandate is a file that names a fund’s strategies, risk limits, capital and how often it trades. The money is pretend. Paper trading runs a fund on real market days with fake money, and a backtest replays it over past days. Start with aihf and the two modes, and read the disclaimer first. Nothing here places a real order.

For experts. Every session is one verb, advance, called once per completed session in both backtest and paper mode. A decision made at session T’s close is executed at T+1’s close. The engine README says stepping a paper fund through the same sessions produces byte-identical records.

Language models never touch the trade. Agents form views and write theses, deterministic code sizes and places orders, and risk limits are hard gates. The engine README makes point-in-time data a principle: on a simulated date only data filed by then is visible, filtered on filing date and not report period. The roadmap still marks point-in-time data correctness as in progress.

The ledger is append-only and hash-chained. Replay verifies the chain and refuses an altered record, and the paper broker’s book is reconciled against the ledger before anything trades. A mismatch raises BookMismatch rather than being repaired, and any failure inside advance halts the fund.

To add an analyst, implement AlphaModel.predict(ticker, date, data_client) and return a Signal with a value from -1 to +1. A new strategy needs no code: a YAML file in strategies/ bundles existing models with a blend policy.

Good to know

Market data comes from Financial Datasets and model calls go to the provider you choose. The README gives no cost for a run.

• Education only: the README says it is not for real trading or investment, gives no investment advice or guarantees, and accepts no liability for financial losses. Past performance does not indicate future results
• Keys: they are saved to ~/.hedge-fund/.env. For the x402 option the key is a wallet’s private key, and the key template says to use a small, separate wallet
• Agents are approximations: VISION.md says the investor agents are stylized approximations of public philosophies, not the actual individuals and not endorsements
• Still being built: the engine README carries a work-in-progress note, and the roadmap marks live brokers and a validation gate as planned and the built-in scheduler as in progress
• Files: the app keeps its keys, fund definitions, results, ledgers and a cache of API responses under ~/.hedge-fund/
• License: MIT, Copyright (c) 2024 Virat Singh