Metaculus

What it is

A question site with a scoring system attached, run by Metaculus, Inc. You read a question, you enter a probability, and at resolution you are scored against a proper scoring rule. There is no balance, no position and nothing to sell before close — the only thing you accumulate is a record.

The scoring is the product, and Metaculus documents it in more detail than most vendors document their pricing. Everything is built on the log score, which is proper: the way to maximise it on average is to state the probability you actually believe. Because a raw log score is always negative on a binary question and reads badly, two rescalings sit on top of it. The Baseline score compares a forecast to a fixed chance benchmark and is positive when you beat chance — on resolved binary questions the median observed Baseline score is about +17. The Peer score compares a forecast to every other forecast on the same question; the sum of all participants' Peer scores on a question is always zero, which makes it the harsher and more informative of the two. A spot version of either ignores how long the forecast stood and reads only the value at one moment, usually when the Community Prediction is revealed.

Scores are time-averaged over the life of a question, so a forecast left standing counts for every day it stood. That is the mechanical reason the site rewards updating and not simply arriving late with the answer.

The Community Prediction is a time-weighted median of forecasters, and it scores better than most of the people feeding it — its own Peer score is positive by construction.

Availability

Open worldwide, free, no identity check. The terms of use state that the service is offered to individuals aged 13 and over, that Metaculus operates it in the United States, and that it makes no representation that the service is appropriate for use elsewhere — a reader outside the US is told they are responsible for their own local law. There is no money involved on the forecasting side, which is why none of the jurisdiction machinery that governs an event exchange applies here.

Tournament prizes are a separate matter: they are paid out, and payment mechanics and eligibility are handled per tournament rather than by a blanket rule on the site.

Pricing

Free to forecast on, and free to read. Metaculus earns from organisations rather than from forecasters — tournaments commissioned by a partner, private instances of the platform deployed inside an organisation, engagements with its Pro Forecasters, and workshops. None of that is something a reader of this card buys.

The one number a forecaster sees is a tournament prize pool. These are public and vary widely; on 19 September 2026 the tournaments index showed pools from none at all up to $50,000 for the bot tournament, with several in the $3,000 to $15,000 range. Your share is proportional to the square of your sum of Peer scores, each one weighted by its question's weight in the tournament. A negative sum pays nothing, and in tournaments ending after 1 June 2025 a share that works out below $50 is not paid at all — that money goes to the forecasters above you. The coverage term that older descriptions still quote belongs to the scheme this one superseded in March 2024.

Markets & resolution

Questions are written by staff and by the community, grouped into tournaments and question series. Subject matter skews to what research funders care about — infectious disease burden, AI capability and policy, geopolitics, energy and climate, macro indicators, space — rather than to whatever has an order book somewhere.

Resolution is by Metaculus against the criteria published on the question, and the interesting case is the one no venue has an equivalent for: a question whose criterion turns out not to be evaluable is annulled rather than forced to an answer, and annulled questions do not score. That is cheap to do when nobody holds a position, and it is one of the real differences between this and a venue, where a resolution has to land somewhere because money is sitting on it.

Some questions are collaborations with named institutions, and Metaculus says so on the tournament page. The Respiratory Outlook 2025/26 is described on its own page as a collaboration with the US Centers for Disease Control and Prevention, with the note that community forecasts are shared with the CDC and do not necessarily reflect CDC views. The Horizon Scanning Tournament is described as part of a study group run with Renaissance Philanthropy and the UK's Advanced Research and Invention Agency. Those are the two partner claims that check out against Metaculus's own pages; a great deal else gets asserted about Metaculus's institutional users in places that are not Metaculus.

Integrations

The API is the thing to check before planning anything, because it changed. Every request now requires an account token in an Authorization header — an unauthenticated call is rejected with a message saying so. With an ordinary account you get your own data, open question text, the text and resolution of closed questions you predicted on, staff comments, and the current Community Prediction on roughly 50 questions. On everything else, aggregation data is simply absent from the response. A separate tier for bot developers widens that to around 250 open and 250 resolved questions, by application. Commercial use requires a written agreement, and the terms of use prohibit using the data to train or evaluate machine-learning models without written permission.

The platform code is open source under BSD-2-Clause and actively committed to — the repository was last pushed to on 18 September 2026 — and the API documentation notes that the access restrictions above do not apply to a custom instance you run yourself. Metaculus also sells hosted private instances as a service.

There is no first-party Python client. The metaculus package on PyPI is not Metaculus's: it is version 0.0.2, last uploaded in February 2019, published by an unrelated author and pointing at a GitLab repository. Installing it will not error, and it is not the vendor's.

Limitations

The API restriction is the limitation that matters, and it is recent enough that most third-party descriptions of Metaculus are wrong about it. If your plan was to pull the Community Prediction across thousands of questions for analysis, that is now an application rather than a GET.

There is nothing to arbitrage against, by design. A number on Metaculus cannot be traded anywhere, so the usual argument that money keeps a probability honest does not apply — what keeps it honest is a proper score and a public record, which works on people who care about the record and not at all on people who do not.

The accuracy record is a leaderboard about forecasters, not a claim about the aggregate being right in any particular case. And the empirical score figures Metaculus publishes for Baseline and Peer are marked as of November 2023, so the benchmarks you compare your own score against are three years old.

Alternatives

Good Judgment Open is the same job with a different scoring rule — Brier rather than log — and a commercial firm on the other end of it recruiting from the leaderboard. Manifold is the opposite trade: many more questions, far faster, resolved by whoever wrote them, and scored by profit in a play currency rather than by a proper scoring rule.