How We Predict

Updated live from full-time results. Brier and log-loss are standard probabilistic scoring rules — a perfect forecaster scores 0 on both.

⚠ Small sample (N=3). Accuracy, Brier, and log-loss are noisy below ~20 graded matches — treat as directional only.

Every team starts at 1500. Pre-tournament we seed the 32 teams from their FIFA / world-football Elo history; new entrants stay at baseline. Elo updates after each full-time result using the classic formula.

The TFI is a 0–100 score for one team arriving at one match. It penalises four things:

• Flight distance — 8 points per 1,000 km, capped at 50.
• Time-zone shift — 5 points per hour, capped at 25.
• Altitude jump — 2 points per 100 m gained, capped at 15.
• Short rest — 15 points for every day below 4 days between matches, capped at 30.

We compute the haversine distance between consecutive venues, the tz delta via IANA zones, the altitude delta, and the rest period between fixtures. The first match of a team's tournament uses a baseline TFI of 25 (one long-haul flight in).

We apply a TFI penalty of 1.2 Elo points per TFI point to each side, then convert the rating gap to a no-draw win probability using the standard logistic (base 10, scale 400). Draw probability is bounded between 18%and 30% based on the absolute rating gap.

Expected goals are anchored at 1.35 per side, then nudged ±0.4 by the rating gap and fatigue. xG is informational — match outcome is what the table above scores.

• · Squad availability, injuries, suspensions.
• · Weather and pitch conditions.
• · In-tournament form / momentum beyond Elo updates.
• · Referee or VAR effects.

These are deliberate omissions — they add complexity without clear signal at this scale. We publish misses in the table above so you can see exactly when the model is wrong.