How accurate is the SeerPlay AI?

Premier League 2026-27 · live scorecard

FIFA World Cup 2026 · final record

Recorded live across the tournament on this page; the World Cup edition is archived at wc.seerplay.com.

Why the archive can differ.The archived match pages were rebuilt after the final, so their per-match hit and miss markers reflect the AI's current draw-margin rule. Count them and you get three more correct calls than the figure above, all of them group-stage draws. The number we publish is the one locked live, match by match, as the tournament ran, because that is what was actually called before each kick-off. We would rather show the lower, older number than the flattering one.

Calibration · were the probabilities honest?

Accuracy only asks whether the call was right. Calibration asks the harder question: when the AI said 60%, did it happen about 60% of the time? Below is every probability published for the World Cup, sorted into bands and scored against what actually happened. 312 forecasts from 104 matches, three per match.

Read a row as: of the 20 forecasts the AI put in the 80-90% band, it said 86% on average and 90% of them happened. Across every band the average gap between what it said and what happened is 4.6 percentage points. The Brier score is the standard measure of forecast quality, where lower is better and 0.222 is what you would score by ignoring the match and always guessing the base rate. Log loss, the other standard measure, is 0.459 against a base-rate reference of 0.637.

Honest limits: this is one tournament, and the thin bands carry few forecasts, so read those rows loosely. The numbers come from the probabilities published on the match pages at wc.seerplay.com, scored against the full-time result of each match. The Premier League season gets the same treatment here as results land.

The methodology, in plain language

An Elo-style team-strength rating feeds a Poisson goals model that turns the gap between two sides into win/draw/win probabilities. The model is draw-aware (tuned on historical seasons where plain Poisson under-prices draws) and re-fits as results land. The AI backs the favourite unless the draw sits within a set margin of it, in which case it calls the draw, so the published call is not always the highest single probability.

The full step-by-step explanation is on how the AI works.

Three rules keep the record honest: every pick locks before kick-off, no pick is ever edited after the fact, and every settled match is scored — including the misses you can see in the tables above.

Quick answers