ai-hedge-fund: can AI think like Buffett or Graham?

If you hear "this AI picks stocks like Warren Buffett", be skeptical. Still, it's useful to know what the ai-hedge-fund project actually does, because the idea inside is quite simple and worth learning from.

How do we know? We opened the code. Each investor has a file, and inside is a text instruction for the AI: "You are so-and-so, work through these questions". There's no other secret. We didn't run it, so we won't talk about results.

The authors state one thing openly: these aren't the people themselves, just simplified versions of their public philosophy. They don't endorse the project either.

What the five look at

Buffett. Judges a company as an owner, not a trader. Is the business understandable, is it protected from competitors, does management use money well, is debt low, is the price sensible? And finally: could you hold it for ten years?

Graham. Margin of safety before anything else. Is the price low relative to demonstrated earnings and book value? Is the balance sheet strong, are earnings stable? He's suspicious of paying up for future growth.

Lynch. First sorts the company into a category: fast grower (earnings up more than 20%), stalwart (10–12%), slow grower or turnaround. Then compares the P/E with earnings growth.

Munger. Thinks backwards: what would make this investment fail? If the numbers don't paint a clear picture, it goes in the "too hard" pile and he stays neutral. That's our favourite item: if you don't understand it, leave it alone.

Druckenmiller. Looks at direction rather than level: is growth speeding up or slowing down? He only acts when the upside is much bigger than the downside. This persona also allows shorting. Many scholars consider shorting impermissible, and we mention it only to say what's in the project.

How far ahead they look

Each persona also has a stated horizon for judging its call. Four use one trading year (252 days), Druckenmiller half of that:

The part that isn't a persona

There's also one systematic model: post-earnings drift, the way prices keep moving after a report. It isn't an AI "opinion", it's a rule-based model. The README says both kinds of strategy can run in one fund, like separate desks at a real firm.

Limitations

• No real trading. The README says so directly: the system doesn't make actual trades, and the project is for education and research. It has paper money and backtest modes.
• Data may cost money. You need a Financial Datasets key for prices and reports, plus a model key for the AI.
• It's being rewritten. The new core's docs mark it as a work in progress, so details may change.
• A name is no guarantee. An instruction with Buffett's name on it doesn't give you Buffett's experience. It's just a well-written list of questions.

The most useful part, in our view

You don't even have to install anything. Take the personas' questions and apply them to your own stock. Ask like Lynch, "which category is this company in?", or like Munger, "what would make this fail?"

If you want to see what real investors actually did, their SEC filings are on our superinvestors page.

For information only, not investment advice. This is a third-party project and we're not affiliated with it.