Every forecast on this site comes from 10,000 independent simulations of the entire 48-team tournament. Here's exactly how the model works, what it gets right, and where its limitations are.
We use a hybrid model. The spine is international football Elo — the same methodology used by FiveThirtyEight, the World Football Elo Ratings, and (since 2018) FIFA's official ranking system. On top of that we layer proprietary Betminer signals: form delta, margin distribution, and defensive solidity. Every match outcome is sampled from a probability distribution, not picked deterministically. We simulate the full tournament 10,000 times and aggregate the results.
Elo ratings convert team strength into a single number. When two teams meet, the rating difference produces a probability for each outcome — home win, draw, away win. We rate every international football team by replaying the last eight years of competitive international matches: World Cups, qualifiers, continental tournaments (Euros, Copa America, AFCON, Asian Cup), Nations League, and friendlies.
Different match types contribute different weights to the rating. World Cup knockout matches use a K-factor of 70, World Cup group matches 60, continental tournaments 40, qualifiers 30-35, friendlies 20. This means a friendly upset shifts a team's rating less than a World Cup upset — exactly as it should.
Home advantage is worth 65 Elo points in our model, applied at probability-generation time only. The stored rating represents pure team strength regardless of venue. For neutral-venue tournament matches — which is most of the World Cup for most teams — no home advantage is applied.
Pure Elo is defensible but generic. Three Betminer-specific signals adjust the base probabilities:
For each team, we compare actual recent results against what Elo would have expected. A team beating opposition it "should" beat just keeps its rating — but a team consistently over-performing against Elo expectations is genuinely in form. We weight the most recent five matches at 2x the weight of the previous five, then convert the over/under-performance into a probability nudge of up to ±10 percentage points.
When the simulator picks a result, the scoreline isn't drawn purely from a Poisson distribution. Instead we use the team's empirical margin distribution — how often they win by 1, by 2, by 3, by 4+. A team that grinds out 1-0 wins gets a different scoreline distribution from a team that wins 3-0. This preserves the tactical character of each side rather than treating them as interchangeable goal-generators.
Clean sheet rate and failed-to-score rate feed into the goal-count sampling for the losing side. A defensively solid winner means a tighter scoreline; a leaky winner means a higher-scoring one.
Host advantage applies for USA, Mexico, and Canada in their group matches and any knockout matches at home venues. The total signal-layer adjustment is capped at ±15 percentage points so no single factor can override the Elo spine entirely.
For every match, the model produces home/draw/away probabilities — say 48% / 27% / 25%. We then draw a random number to pick the actual outcome, weighted by those probabilities. Run the same match 100 times and you'll get roughly 48 home wins, 27 draws, 25 away wins. This is the critical difference between our model and a deterministic predictor: we represent uncertainty honestly rather than always picking the favourite.
Goal counts come from a Poisson distribution calibrated to international football's average of roughly 2.7 goals per match, adjusted for the strength gap between the teams.
Group stage matches can end in a draw. Knockout matches cannot — the simulator runs extra time if needed (each side gets a fresh draw from a low-lambda Poisson distribution for the additional 30 minutes), and if still tied, penalty shootouts. Penalties are essentially a coin flip in our model with a tiny weighting toward the higher-Elo side, because shootouts are notoriously low-signal.
The 2026 World Cup is the first 48-team tournament ever held — 12 groups of 4, top 2 from each group plus the 8 best third-placed teams advance. The bracket runs as a 32-team knockout from there. This format has never been used before, so we can't back-test the format itself — only the underlying match-level predictions.
Our simulator implements FIFA's published tiebreaker rules for ranking the third-placed teams: points, goal difference, goals scored, then a deterministic team_id tiebreak (where the official rules use drawing of lots).
Before publishing 2026 predictions we ran the model against the 2022 World Cup using only data available before 19 November 2022 (the day before that tournament started). The pass criteria — defined in advance — were that Argentina (the eventual winner) should land between 6% and 18% probability, and the top 3 teams collectively should account for 30-45% of the probability mass. The pre-tournament bookmaker market had Argentina at roughly 11-14% implied probability.
Back-test results will be added here once the 2022 simulation has been completed and reviewed.
We don't model injuries, suspensions, last-minute lineup changes, weather, referee assignments, individual player form, or in-tournament momentum effects. The model assumes squads are at full strength and the tournament runs as scheduled. A starting XI injury crisis on the eve of the tournament would invalidate parts of the forecast.
We also don't model what bookmakers do — there's no market price input. The model is purely a structural simulation based on team strength and recent form. If our numbers and the bookmakers' implied probabilities disagree wildly, that's information either way; we don't try to anchor to one or the other.
Every simulation in our run uses a deterministic random seed stored alongside its result. In principle the entire 10,000-simulation batch can be reproduced exactly by re-running with the same seeds. The model code, schema, and methodology are documented internally and available for review on request.