2026 World Cup
Two views of the same model. PREREG PROJECTED is the bracket as frozen 1 day after kickoff — it never changes from here, so the post-tournament scorecard stays honest. LIVE folds in every completed match and updates every 10 minutes during match windows. Home, Draw, and Away (HDA) win probabilities come from our Elo ratings, which retune on the same cadence.
Projections
—Live: updates every three hours during the tournament. Pre-Tournament: frozen 1 day after kickoff. Differences between the two are examples of where the model is learning.
Today's Slate LIVE
refreshed every 3hToday's match docket pulled from ESPN — scheduled times, in-progress scores with minute, and same-day finals. H · D · A percentages on the right are our LIVE model odds (form-Elo updated each cron run, NOT bookmaker lines).
Recent Results LIVE
—Advanced, Bubble, and Eliminated LIVE
refreshed every 3hAll 48 teams sorted by group, color-coded by whether they have advanced, are on the bubble, or have been eliminated from the knockout stage. Click any team for a stat summary.
Group-Stage Matches
72 games · win/draw/loss probabilitiesEvery group-stage match with our pre-reg per-outcome probability (P(A) / P(draw) / P(B)). Bars are sized by win probability; the gold flag marks upset alerts. As matches finish, each gets stamped with the winner and our pick correctness — green ✓ if we had it, red ✗ if we whiffed. Draws count as wrong unless the pre-reg gap between sides was <1.5pp (true coin-flip credit, gold ✓).
Methodology + Honesty Notes
pre-reg locked 2026-06-07What's in the bracket model (v1):
- Per-team Elo at tournament start (FIFA-style with K=32, blended with club-form proxy).
- Monte Carlo bracket sim — group stage + knockout simulated together; probabilities are share-of-sims outcomes.
- Host-nation boost applied to USA / Mexico / Canada.
- Confidence flag per team (high / medium / low) based on data coverage.
v3 calibrated: we built our own club Elo from 28,967 matches across 5 European leagues × 16 seasons. For each historical WC, we counted how many squad members had a club ranked in our top-20 at tournament start (pct_in_top20). Fitted a 1-feature linear model to outcome ordinal (champion = 4, finalist = 3.5, ..., group-stage = 1.5). Robustness retest: 4 of 5 N-values beat v2 on champion hits.
Bold v3 contrarian shifts vs v1 Elo:
- ★ UPGRADES: Ghana 47→17 (Δ+30), Ivory Coast 34→12 (Δ+22), Sweden 38→18 (Δ+20), Belgium 21→7 (Δ+14), Senegal 22→10, US 26→13, Cape Verde 44→30
- ★ DOWNGRADES: Colombia 7→35 (Δ-28), Iran 23→39, Japan 15→27, Australia 24→36, Argentina 2→6, Brazil 5→9
Why we surface both: v1 has the cleaner per-round bracket probabilities (Monte Carlo). v3 has stronger historical fit for the squad-club-form signal but doesn't natively produce a full bracket. They disagree most where v3's squad-quality lens differs from v1's Elo summary — those disagreements are surfaced rather than averaged away.
Cardinal rules: raw data only as model inputs (no third-party predicted-winner models consumed as features); pre-reg before compute; post-tournament scorecard locked at pre-reg time, evaluated honestly when the tournament closes.