How it works

Every morning the model takes everything that has actually happened, treats those results as fixed, and simulates the rest of the tournament 40,000 times. Counting how often each team wins, reaches a given round, or tops its group turns into the probabilities you see across the app.

Each team starts from a pre-tournament World Football Elo rating (roughly 1500–2100). A separate pedigree index, built from World Cup results since 1986 and weighted toward recent editions, nudges historically strong sides. After every real result, both teams' Elo ratings update — winners gain, losers lose, by more when the margin is larger. That drift powers the Rating Drift chart on Trends.

The adjusted rating gap between two teams is converted into expected goals for each side, then scorelines are drawn from a Dixon–Coles bivariate Poisson distribution — a Poisson model with a correction that captures the real-world clustering of low scores (0–0, 1–0, 1–1). Knockout ties that finish level go to simulated extra time and then penalties.

Before the goal calculation, each match is adjusted for the things that actually move results:

• Host advantage for the USA, Mexico and Canada on home soil.
• Altitude at Mexico City (~2,240 m) and Guadalajara (~1,566 m).
• Heat for hot-climate venues with midday kickoffs.
• Rest and travel from each team's previous fixture and the distance between venues.

The simulator uses a seeded pseudo-random generator, with the seed derived from the date. Given the same results, the same day always produces the same probabilities — which is what makes day-over-day trend charts and movers meaningful rather than noise. The model runs once daily at 5:00 AM ET, writes an immutable snapshot, and every page reads from snapshots.

Groups rank by points, then goal difference, goals for, head-to-head record, and finally a deterministic drawing of lots. The 12 group winners, 12 runners-up and the best 8 third-placed teams advance to the Round of 32, using the official 2026 bracket mapping.

• openfootball — canonical schedule, groups and results (public domain, no key).
• football-data.org — live results during the tournament (free tier).
• Wikipedia REST — country snippets. TheSportsDB — squads/badges.
• Manual overrides can be dropped in to correct any feed; the model self-heals by replaying from corrected actuals.

• It is a model, not a prophecy — a 25% favorite still loses three times in four.
• Elo and pedigree don't see injuries, suspensions, tactics or in-form individuals.
• Squads, country facts and team history are static for this release.
• Penalty shootouts are modeled as near-coin-flips with a small edge to the stronger side.