The pre-tournament model made a call on every match with zero knowledge of any 2026 result. Here is how those blind calls held up against what actually happened — accuracy, sharpness, and how well its confidence matched reality. 104 matches so far; this updates as the tournament plays out.
Rank every team by the model's pre-tournament title odds, then see how many of its top N actually reached the last N. The call sharpens the deeper the tournament runs.
🎯 Its top two — Argentina & Spain — are the two finalists, and its top four were the exact four semi-finalists. All before a ball was kicked.
Every dot bundles all the calls at one confidence level. If the model is honest, when it says 60% the thing happens ~60% of the time — so the dots hug the dashed line. Dot size = how many calls sit in that bucket.
| Model said | avg | happened | calls |
|---|---|---|---|
| 0–10% | 7% | 6% | 17 |
| 10–20% | 15% | 10% | 59 |
| 20–30% | 26% | 19% | 112 |
| 30–40% | 34% | 28% | 32 |
| 40–50% | 44% | 60% | 25 |
| 50–60% | 55% | 75% | 28 |
| 60–70% | 65% | 81% | 21 |
| 70–80% | 75% | 85% | 13 |
| 80–90% | 85% | 60% | 5 |
Winners called by stage — group: 61% (72) · knockout: 84% (32).
Poisson (Dixon-Coles) + Elo · blind pre-tournament model · worldcup2026ml.pt