Everyone has a hot take on who wins the World Cup. We wanted something colder than a take — a number. Here is the exact system we use to turn 104 fixtures into win probabilities, goal lines, and title odds.
The honest truth about football prediction: nobody knows what happens. A deflection, a red card, a goalkeeper having the night of his life. What a model can do is price the uncertainty — tell you a match is closer to a coin flip than a coronation, and by how much. That is the whole game. Three steps.
Step 1 — rate every team
Each of the 48 teams gets two numbers: an attack rating (how many goals they tend to score against an average opponent) and a defence rating (how many they concede). We build those from three ingredients, weighted by how much signal each carries.
50%
Recent results
last ~24 months, decayed
30%
Elo / SPI base
long-run strength
20%
Market priors
bookmaker-implied
Recent form gets the most weight but decays — a 4–0 from two years ago barely moves the needle today. Elo keeps a long memory so a one-off bad week does not tank a genuinely elite side. And we anchor it all to the betting market, which aggregates more information than any single model, then fade it slightly because markets overreact to narratives.
Step 2 — turn ratings into expected goals
For any matchup we combine one team’s attack with the other’s defence to get an expected goals figure (xG) for each side, then add a home or host bump. Goals in football follow a Poisson distribution closely enough to be useful — feed in the two xG numbers and you get the probability of every scoreline, which sums up into win, draw, and loss.
Brazil project to 1.9 expected goals, Mexico to 0.9. Run that through Poisson and Brazil win 58% of the time — but Mexico still nick it or draw 42% of the time. Favourites are favourites, not locks.
xG here is a forecast of goals from team strength, not the post-match shot-quality xG you see on broadcasts. Same idea, measured before kickoff.
Step 3 — simulate the whole tournament 40,000 times
A single match is one probability. A tournament is thousands of them tangled together — group tables, tiebreakers, a knockout bracket where the draw matters. You cannot eyeball that. So we play the entire World Cup out 40,000 times. Each simulation rolls every match using its scoreline probabilities, fills in the group standings, advances teams through the bracket, and records who lifts the trophy.
Count how often each team survives each round across all 40,000 runs and you get clean probabilities: advance from the group, reach the quarters, win it all. Run it again and the numbers barely move — that stability is how you know it is signal, not noise.
Why probabilities, not predictions
Bonus — reading the betting odds
Want a quick gut-check without a model? Flip the bookmaker’s odds into a probability. Decimal odds of 2.00 imply 50% (1 ÷ 2.00). Odds of 4.00 imply 25%. The catch: add up every outcome’s implied probability for a match and it comes to more than 100% — that overround is the bookmaker’s margin, the vig. Strip it out proportionally and you have the market’s real read.
- Decimal 1.50 → ~67% — heavy favourite.
- Decimal 2.50 → ~40% — live underdog or a tight match.
- Decimal 6.00 → ~17% — a genuine longshot.
That is the toolkit. In the next pieces we point it at the actual 2026 bracket — the title race and the host nation’s road through it.