its co-favorite throughout; a 50.7% / 49.3% call in the final itself
Point-in-time forecasting · 2026 season, run live
104 matches. Forecast before kickoff. Scored after.
A Bayesian scoreline model that published a probability for every match of the 2026 World Cup before it was played — and publishes the scorecard, including the parts that were only average.
FIG. 01 Headline calls
its top two in title probabilities from the group stage onward
inside its projected quarterfinal eight, published before the Round of 32 finished
13.9% — the single most likely scoreline; Spain won in extra time
FIG. 02 The record
The record
Modal-outcome hits are against the model's own 1X2 pick. Point-in-time: each match priced by a model that had seen only earlier results.
| Stage | Matches | Mean RPS | Modal hits |
|---|---|---|---|
| Group | 72 | 0.157 | 46 / 72 |
| Round of 32 | 16 | 0.116 | 12 / 16 |
| Round of 16 | 8 | 0.188 | 6 / 8 |
| Quarter-finals | 4 | 0.134 | 4 / 4 |
| Semi-finals | 2 | 0.196 | 2 / 2 |
| Third place | 1 | 0.333 | 0 / 1 |
| Final | 1 | 0.271 | 1 / 1 |
| All knockout | 32 | 0.153 | 25 / 32 |
Knockout ties by which team advanced: 26 of 32 (81%). Quarter-finals onward: 7 of 7.
FIG. 03 Calibration
Calibration
Favorites: predicted 69.7%, realized 76.8%
across 56 favorite-band matches — calibrated, with the error on the humble side
Draw rate: predicted 23.4%, realized 23.1%
Goal-margin tails within ~1.5 points at every threshold
≥2 goals: 42.8% predicted / 43.3% real · ≥3: 21.5 / 21.2 · ≥4: 10.2 / 8.7
DISCLOSURE — not a figure
What it didn't do
Over the full 104-match replay, the model and a plain Elo baseline are statistically tied — 0.1561 against 0.1557. It was ahead through the group stage as played live (0.157 against 0.163), and the two converged across the knockouts. Both numbers are published because a forecaster that only reports its good window is not a forecaster.
It has no demonstrated edge against sharp betting markets. An early audit called that a tie; a later preregistered test on 217 fixtures put the market ahead of the model by about 0.010 in mean RPS, and the honest reading is that the market forecast better. The full account is here. Nothing on this site is betting advice, and nothing on it is sold as one.
FIG. 04 For publishers
For publishers
An embeddable forecast surface for your football coverage — probabilities, not picks.
- 01 Every number ships with its uncertainty no bare percentages anywhere; a missing input renders as a stated gap, never as a zero
- 02 Point-in-time by construction a forecast is a read at a timestamp; a result that wasn't knowable then cannot change it
- 03 Your styling CSS-variable theming, your domain, no cookies and no tracking of your readers
- 04 A public archive behind every claim each forecast is timestamped and republished; performance claims link to the record
Two tiers — the tournament ladder and schedule, or the full per-fixture detail. Availability by tournament window.
Enquire about a tournament window