How we're doing.
Running calibration of the v1 ensemble model against actual results. Brier score, log-loss, winner accuracy, calibration plot. We update after every match. We publish misses honestly.
0
Matches scored
—
Winner accuracy
—
Brier score
—
Log-loss
No matches yet. The tournament starts on 2026-06-11.
This page goes live the moment the first group-stage match finishes. Until then, our predictions sit on the main predictions page without comment — locked in, dated, and unable to be edited after kick-off.
What we publish here once results start landing:
- Winner accuracy. Of the matches where we picked a clear favourite, what fraction did the favourite actually win? A blunt headline number.
- Brier score. Sum of squared probability errors. Random guessing gives ~0.67; a good calibrated model on World Cup matches typically scores 0.18-0.24.
- Log-loss. Penalises confident wrong predictions much more than uncertain wrong predictions. The right metric if you want to punish overconfidence.
- Calibration plot. If we say something has a 70% chance of happening, does it happen 70% of the time? Bucket every prediction by its probability and check.
- Worst misses. The three or four matches where we were most wrong, with a one-line honest explanation of why.
We will not delete or rewrite predictions after the fact. Every prediction this site publishes is timestamped and immutable from the moment kick-off occurs.
How this updates
As results come in, the numbers update — no edits to predictions.
When a match finishes, the score lands in our results record and this page recomputes the running totals. The original prediction is never re-run, never softened, never re-calibrated to look better in hindsight.