World Cup 2026 — AI Model Accuracy & Trust
An open scoreboard for the prediction model: how often it calls the result, how close the scorelines land, and what it is learning from finished matches.
Model accuracy so far
Does the prediction need updating? Only 4 matches are in so far — too small a sample to judge. Accuracy and the update signal will firm up as the group stage plays out.
How to read these numbers
Exact football scores are very hard to predict — the single most likely scoreline usually carries only around a 10–12% chance, so a low exact-score rate is normal for any model. Result direction (home win / draw / away win) is the more meaningful accuracy measure.
The Brier score grades how well the published win/draw/win probabilities matched reality: 0 is perfect and about 0.667 is an uninformed three-way guess. Lower is better. We track it openly so the model is measured against what actually happened rather than judged on memorable calls.
Confidence on a match page reflects how decisive the rating gap is, not a promise. A high-confidence match has a wide gap; a low-confidence match is close to even and more sensitive to a single event. These are model estimates for research and entertainment, not certainty.
Latest model learning notes
- 🇨🇦 Canada 1–1 🇧🇦 Bosnia and Herzegovinamodel call missed · score error 1 goal
- 🇺🇸 United States 4–1 🇵🇾 Paraguaymodel call missed · score error 3 goals
- 🇲🇽 Mexico 2–0 🇿🇦 South Africamodel call correct · score error 0 goals
- 🇰🇷 South Korea 2–1 🇨🇿 Czechiamodel call correct · score error 1 goal