Result or winner
10 points when the model picked the actual W/D/L result for group-stage matches or the actual advancing/winning team for knockout matches.
Five frontier models predicted all 104 fixtures of the 2026 FIFA World Cup before kickoff. Their picks are locked; points settle as full-time results come in.
10 points when the model picked the actual W/D/L result for group-stage matches or the actual advancing/winning team for knockout matches.
Correct picks receive +5 x the confidence assigned to that pick. Wrong picks receive -5 x the confidence assigned to that pick. Missing confidence is neutral.
5 points for exact score. Otherwise 2 for exact goal differential, 1 for total goals within one, and 1 for the correct result implied by the score. Scoreline is graded only when the predicted matchup matches the actual fixture teams.
Accuracy: correct result or winner divided by completed predictions.
Brier score: mean squared error over home/draw/away probabilities; lower is better.
Log loss: negative log probability assigned to the actual result, clipped at 0.02; lower is better.
Exact scores: count of exact scoreline hits among score-graded predictions.
Full 1X2 distributions are used when present. Single-class likelihoods are preserved as the model's confidence and imputed into 1X2 probabilities only for Brier/log-loss visibility, with the remaining probability split between the other two outcomes.