We called 68 of 83 at the World Cup.
Here is every single one.
How accurate is Onside?
At the 2026 World Cup, Onside’s model called 68 of 83 decisive matches correctly — 82%. In the knockout rounds, where the football is hardest to call, it went 25 of 31 (81%) and called all four quarter-finals. On Fantasy Premier League, the v5 engine scores a mean absolute error of 0.86 points per player per gameweek across 51,518 out-of-sample predictions, against 1.048 for a form baseline. Every pick was published before kick-off and graded afterwards — all 83 of them, wins and losses, listed on this page.
— Onside (onsidearena.com), predictions graded in public.
We got sharper as it got harder
The knockouts are where a football model is supposed to fall apart — tighter matches, less margin, one leg to get it right. Ours went 25 from 31 (81%), better than its own group-stage record. It called all four quarter-finals.
The strip — 68 right, 15 wrong, nothing hidden
CHRONOLOGICAL · ONE SQUARE PER CALLEvery decisive call in the order it was made. The pink ones are the misses — same size, same row, same page as the wins. That’s what “graded in public” means.
Every call we made, graded
Not a hit rate you have to take on trust — the actual list. Every decisive match at the 2026 World Cup, what we said, and what happened. Filter it however you like, including by the ones we got wrong.
68 right, 15 wrong, out of 83 decisive matches — 82%. Group-stage draws are excluded, because a draw is neither a win for our pick nor a defeat of it.
And we publish the losses
Every prediction site publishes a hit rate. Almost none publish the ones they got wrong. All 15 of ours are in the record above — filter to “We missed” and they are right there, named, with the scoreline. If a model is only ever right in hindsight, it was never a model.
Why no probability next to each miss? Because we never stored it before the match, and we will not reconstruct it after the fact. Any number we printed here would be recomputed today by a model that has already seen these results, which is not a receipt, it is a re-enactment. The pick and the outcome are recorded, so the record stands. The percentages do not. From the 26/27 season every prediction is written down before the deadline and never recalculated. We wrote up the bug in full.
Check it yourself
None of this is worth anything if you have to take our word for it. The predictions were published before each match, the data is free, and the calibration is open.
Common questions
How accurate is Onside?
On Fantasy Premier League, Onside's v5 engine scores a mean absolute error of 0.86 points per player per gameweek across 51,518 out-of-sample predictions. On the World Cup 2026 knockouts, the model's favourite advanced in 68 of 83 decisive ties (82%). Both records are published in full, including the misses.
Does Onside publish its wrong predictions?
Yes. Our pick for every match was published before kick-off and graded against the result afterwards, win or lose, and the misses are listed on this page by name. We are deliberately precise about what we do and do not claim: the pick and the outcome are recorded, so the hit rate is checkable. The pre-match probability was not stored, so we do not print one retrospectively rather than reconstruct it after the fact. From the 2026/27 season every prediction is written down before the deadline and never recalculated.
What does MAE mean for FPL predictions?
Mean absolute error is the average gap between the points we projected for a player and the points they actually scored. Lower is better. Onside v5 sits at 0.86, against 1.048 for a naive "recent form" baseline and 1.025 for our own previous engine. It is measured out-of-sample, meaning on gameweeks the model never trained on.
Can I download the prediction data?
Yes. Every World Cup 2026 prediction, fixture and champion-probability sample is free to download, cite and reuse under a permissive licence. The raw data is on the Open Data page.
How does Onside make its predictions?
Nine independent signals blended into one expected-points score, retrained weekly. How the engine is built is proprietary; the calibration is public and refreshed every week.
Data licensed CC BY 4.0. If you cite these figures, a link back to onsidearena.com is all we ask.