The World Cup archive model: sample, filters, confidence, limits.

The baseline for each archived World Cup statistic is the combined record of the 2018 and 2022 World Cups (128 matches). We chose tournament play specifically because it is the closest reference to 2026: neutral venues, knockout pressure, condensed schedule, mixed-confederation matchups. Club-league data was deliberately excluded.

On top of the historical baseline we apply five adjustments, all of them visible on the statistic hubs and per-match pages:

• Stage adjustment. Group, Round of 32, last-16, quarter-final, semi-final and final each carry their own historical multiplier (intensity, fouls, shots on target rise into knockouts; goals tighten).
• Possession profile. Possession-dominant sides generate more throw-ins. We nudge throw-in projections up when both teams skew possession-heavy.
• Deep-defender bias. A team that defends in a low block historically draws more fouls in their own third. We raise foul projections for matches with a clear deep-defender side.
• Strength mismatch. When the FIFA-rank gap is wide we apply a capped nudge: large enough to matter, small enough that one rating point does not flip a projection.
• Climate downgrade. For matches flagged as high-heat or high-altitude, tempo-sensitive statistics (corners, shots on target, throw-ins) get a confidence downgrade rather than a speculative point estimate.

Every projection is tagged High, Medium or Low. Those labels are retained on the archived World Cup projection pages. The rating reflects three things: historical reliability of the statistic, agreement between the 2018 and 2022 sub-samples, and how many filters fired against the projection.

In the archived calibration, throw-ins were the most consistent statistic: the 2018 and 2022 sub-samples agree within roughly two percentage points. Goals statistics are noisier and more often land in Medium.

Offsides was tested as a statistic and dropped: the historical signal did not survive cross-tournament robustness testing and the per-match variance was too wide to support a calibrated number. Publishing it would have meant publishing a guess. We would rather ship five honest statistics than six with one we cannot defend.

• Two World Cups is a small sample. We treat it as the best available reference, not a law of nature.
• The model was calibrated before the 48-team format was played. Group dynamics, third-place qualification and the new Round of 32 were uncertainties in that calibration; the archive preserves the original projections.
• Heat and altitude effects are modelled from prior tournament data, not measured for this event. That is why we downgrade confidence rather than push a precise number.
• Squad changes inside the tournament (injuries, suspensions) are not yet modelled at player level. Where it materially matters we annotate it as a caveat on the match page.

During World Cup 2026 we published projected-versus-actual reviews. They are kept in News and on the accuracy page. The reliability rankings in World Cup statistics and the statistic hubs are preserved archive records, no longer updated.

The archived World Cup model is not evidence of performance in club competitions. The current free club predictions use an independent Poisson prototype based on stored season scoring, with minimum samples and disclosed limitations.

Read the club prediction methodology for its inputs, calculations and untested performance. It does not inherit the World Cup accuracy record.

The closed World Cup ledger remains at the archived accuracy page: 528 published projection lines across 88 matches. The full tournament schedule contains 104 fixtures; the remaining fixtures were not projected.