About benchmarks.finance

benchmarks.finance tracks how AI systems do on finance tasks. It lists the finance benchmarks that exist, quotes their published scores with the source and date, and hosts Moneda Bench, six benchmarks written by Moneda.

What you can do here

  • Leaderboards: every benchmark's results, with the metric and source for each score.
  • Best by use case: which models lead on research, advice, trading and the other use cases.
  • Directory: the finance benchmarks we track, with what each one tests and who made it.
  • Compare: two models side by side on every benchmark where both have a score.

Scores are never added up across benchmarks. Each number is shown with its own benchmark, metric and date.

Why we built this

Finance has no shortage of benchmarks. Nearly all of them test a model on its own: give it a filing or an exam question, grade the answer. Nobody meets AI in finance that way. You use a product, with tools, on live data, in a conversation about your own money.

So we read every benchmark we could find (the directory is that list) and looked for what they leave out. Five things:

  • Live figures. Ask for a price now and a product can fetch it, say it cannot, or make one up. The hallucination benchmarks hand the model the document first, so none of them scores that choice.
  • Market structure. Breadth, sector rotation, fund flows, trading hours across markets, the events calendar, and whether a move gets a cause that was never invented. The market benchmarks that exist ask for single figures.
  • Sources. Do the pages a product cites say what it claims? The method exists in research, but it was run once, on analyst tasks, with no public board. Nobody runs it across consumer products on everyday prompts, every quarter.
  • The advice line. A product should not tell a particular person what to buy or sell, and it should not refuse to explain finance either. The safety benchmarks test the first half only, so over-refusal is invisible to them.
  • Clean knowledge. Every CFA-style question set is mock-exam material that may already sit in the training data, and none of them checks.

Moneda Bench is six benchmarks: one for each of these, and one for portfolio decisions. They run on products as well as models, keep a private split next to the public one, and test for memorised answers. How Moneda Bench works has the method.

Who makes it

The site is created and maintained by Moneda. Moneda also builds Numa 1, which is entered in Moneda Bench.

Scores for other benchmarks come from their authors' papers or official boards, or were run by Moneda with the benchmark's own grader and rules. Each one says which.

Corrections and submissions

If a score, date or credit is wrong, tell us. To have a system tested, submit it. Anything else goes through the contact form.

Agents and scripts can do all of this without the form: one JSON endpoint, POST /api/submissions, takes corrections, model, benchmark and product submissions, and messages. GET the same URL for the schema of each kind, or start from /llms.txt. Every submission is read by a person.

What this site is not

Nothing here is investment advice. Moneda is not a regulated financial institution and does not manage portfolios.