What to use instead of PredictionBook

The product

PredictionBook was a notebook for predictions with the probability written down next to them. You typed a sentence about the future, gave it a number between 0 and 100 percent and a date by which it would be known, and on that date somebody marked it right or wrong. Your profile then drew the only chart the site cared about: of everything you called 60 percent, how much of it actually happened. Anyone could add their own number to a public prediction, so a question often carried a column of credences from half a dozen strangers and a comment thread arguing about its wording. Nothing was traded, nothing was priced, and there was no currency of any kind.

It ran for about seventeen years — the footer of the last working capture reads "PredictionBook © 2008-2025" — and was built by TrikeApps, the development arm of the Investling Group, which later merged into Bellroy. The code has been public since 6 March 2013, when the repository was created under that company's GitHub account; its README names a BSD licence, and there is no licence file in the repository. Both of those were read on 20 September 2026.

How it ended, with dates. On 22 August 2023 the operator opened issue 262 in the project's own repository and wrote that the level of active use no longer justified the resources required for maintenance. Effective 10 January 2024 the site would become read-only, and "one year after becoming read-only, PredictionBook.com is scheduled for permanent shutdown." Read-only arrived on schedule: the repository's final commit and its archiving both carry the date 10 January 2024, and from then on every page served its whole history under a banner with no way to add to it.

The permanent shutdown came about a month late and with no second announcement. The Internet Archive's last successful capture of a prediction page is 10 February 2025 at 07:46 UTC, and it still shows the read-only banner, the resolution date and the individual estimates. The archive's first capture of the retirement page is 12 February 2025 at 22:51 UTC. This page records 12 February 2025 as the end, because that is the first moment in the public record at which the site served nothing else — the true date is somewhere in the 63 hours between the two, and nobody published it.

What to use instead

Nothing replaces it whole, because it was two products sharing a database: a private scoring notebook and a public question board. Pick by the one you used.

If you were keeping your own record, go to Fatebook. It is what the retirement notice itself recommends, and the act is the same one — write the question, set the deadline, give it a number, get a Brier score and a calibration chart when it resolves. It is also made by Sage rather than by the people who shut PredictionBook down, which is worth knowing rather than reassuring. The working difference is the default: PredictionBook was public unless you made a prediction private, and Fatebook is private unless you share it.

If the thing you actually want is that this does not happen to you again, run Confido yourself. It is AGPL, self-hosted, and gives you rooms, calibration curves and a normalised Brier score on a server you control. PredictionBook was open source too, which did not save it: the repository is archived, and the operator's stated reason for shutting the site was that nobody had the time or the skillset to keep a Rails application safe as its dependencies aged. Self-hosting moves that job to you; it does not abolish it.

If you were there to read other people's numbers, Metaculus is the serious version of that — a public question list you forecast alongside everyone else, scored on a proper rule, with a track record on your profile that a stranger can check, and staff who resolve each question against the source named in its own criteria. What you do not get is the midnight question about your own life: you forecast the questions Metaculus decided to ask.

If what you liked was that anyone could ask anything, that is Manifold. Anyone opens a question, the creator resolves it, and an audience turns up within the hour. The shape of the answer is different — a play-money price rather than a list of who said what — and the incentive changes with it, because on Manifold there is a position to hold and on PredictionBook there was only a statement to be scored on.

If the point was the calibration training rather than the questions, Quantified Intuitions is the fastest route: it serves questions that have already resolved, so the feedback loop that used to take a year takes ninety seconds. Use it beside one of the others, not instead.

What you give up

A column of credences instead of a price. This is the loss nobody advertises. On PredictionBook eight people could put eight numbers on one sentence and every number stayed visible with a name and a timestamp beside it; disagreement was legible, and you could see who was usually right about what. Manifold aggregates that into one price and the individuals disappear behind it, Metaculus publishes a community distribution on questions it chose, and Fatebook is mostly you alone. The crowd on an arbitrary question you wrote yourself, stated as numbers rather than as a market, is the thing that has no direct replacement here.

Seventeen years of record, unless you personally exported it. There is no dump. Between 24 August and 23 September 2023, users in the announcement thread asked twice for one — first a static copy, then a database dump with private data stripped; the operator answered that keeping the application running to serve it was not possible, pointed at the Internet Archive, and said a pull request implementing the stripped dump would be welcome. None landed before the site went dark. Private predictions were never public and are in no archive at all.

Every link anyone ever made to it. Prediction numbers ran past 212,000, and the Internet Archive's index holds roughly 14,700 distinct numbered prediction pages — under a tenth of them, and counted from the index rather than from opening each one. Nothing redirects and nothing is paraphrased: a decade of citations in forum threads, blog posts and bibliographies now resolves to the same two-line notice, served with a 403.

Migration notes

The dedicated importer is dead, and the page offering it is still up. Fatebook still lists "Import from PredictionBook" and still asks for a PredictionBook API token. Its implementation, which we read on 20 September 2026, pages through the my_predictions endpoint on the subject's own domain using that token — and that endpoint returns the retirement page with HTTP 403, as every other path there does. The import will therefore find nothing, and the failure looks like an empty account rather than an error. Do not spend an evening on it.

If you do have an old CSV, use the spreadsheet import instead. Fatebook's spreadsheet route asks for a question title, your forecast as a number between 0 and 100 percent, a question created date, and a resolution that is YES, NO, AMBIGUOUS or blank, with optional columns for the resolved and resolve-by dates, tags and notes. A PredictionBook export has all of that under different column names, so this is a rename job in a spreadsheet, not a data-recovery job.

The Internet Archive is the only remaining copy, and it is searchable by index rather than by guesswork. The CDX API will list what exists before you start opening pages:

curl "http://web.archive.org/cdx/search/cdx?url=predictionbook.com/predictions*\
&output=text&fl=original,timestamp&collapse=urlkey"

Grep that for the prediction numbers you remember, then fetch https://web.archive.org/web/TIMESTAMP/URL for each hit. Captured pages preserve the question text, the deadline, the resolution and the individual estimates. They do not preserve your login, your private predictions or your calibration graph, and the archive's own content servers were intermittently unavailable on the day we checked, so expect to retry.

Running your own copy is possible and probably not what you want. The code is archived but still on GitHub, and the founders said plainly in September 2023 that anyone was welcome to stand up their own instance on their own domain. What you would be standing up is a Rails and Postgres application whose last commit is dated 10 January 2024, with the dependency rot that the operators named as their reason for stopping. If the appeal is self-hosting rather than this particular codebase, Confido is the maintained answer.