How accurate are these predictions?

The three forecasters, scored separately

Three different things predict on this site and they are never mixed into one score. The model is arithmetic over public facts. The AI forecaster is a language model whose reasoning is published beside its number. Members are people. Each is counted on its own, and the AI forecaster is never counted as a member.

• The member line stays a participation count until 30 different people have forecast. Below that a percentage would describe a handful of individuals, and a handful of individuals can be worked backwards out of a percentage that moves.
• Only questions that at least 3 members forecast on are inside the member figure at all. Each time the number moves it has to move by three or more people at once, or the movement itself would identify one of them.
• Forecasts recorded after the answer was already public are kept and shown to whoever made them, and left out of every figure on this page. That includes our own AI forecaster: a run produced after the news broke is marked on the question below and is not in the line above. Otherwise "our members called 80% of them" would include people who read the news first.

Every question, settled

The audit trail under the numbers above. Each row carries what was published while the question was open, what actually happened, and where we read it.

No question has been settled yet. When one is, it appears here permanently — including the ones we get wrong, and the ones we withdraw.

How the scoring works

1. Brier, because it cannot be farmed

   Take the difference between the probability given and what happened, and square it. Say 90% and it happens: 0.01. Say 90% and it does not: 0.81. Say 50% and it is 0.25 either way, which is the price of not committing. It is a proper scoring rule — exaggerating a forecast makes your score worse, not better — and that property is the whole reason a published number computed this way is worth quoting.

2. Points are only ever added

   A settled question awards round(100 × (1 − Brier)) points, floored at zero. There is no balance, nothing is at risk, and no question takes a point away from anyone. A member’s total is the sum of what they have earned, computed fresh each time it is shown rather than stored — a stored total is a second version of the truth waiting to disagree with the first.

3. A withdrawn question scores nothing, for anyone

   If both declared sources are gone by the deadline, the question is withdrawn with the reason printed. Nothing is scored and nothing changes, because nothing was ever at risk. Withdrawn questions stay listed below rather than disappearing.

4. Late forecasts do not count towards anything published

   When a question is settled we record the moment the answer became publicly knowable, taken from the source article. Any forecast recorded after that instant is marked and excluded from every figure on this page — a member’s, and our own AI forecaster’s, by the same test. It is still kept and still shown to whoever made it, and a late AI run stays in the table below labelled as excluded — what lateness disqualifies is a claim made to the public, not somebody’s own record.

5. A correction is added, never substituted

   If a resolution turns out to be wrong it is corrected: the revision number goes up, every score is computed again, and a new published figure is appended beside the old one. The page then carries both dates. Nothing in the history is edited, because the point of a public record is that it cannot be quietly improved.

Free game · points have no cash value

Episode Club is not affiliated with, endorsed by or sponsored by Netflix, Apple TV+, Rotten Tomatoes or any other company named on this page. Those names appear only to say whose show it is and where a figure came from.

Forecast on one yourself

Episode Club members can add a forecast to any open question. It is free, your answer is never shown to anyone else, and accuracy is scored here when the question settles.