The whole World Cup, all 104 matches and the knockouts, simulated 50,000 times, five different ways, from cold form to the betting market.
Each bar is a team's chance of winning the tournament under the selected model. The dots mark Opta's supercomputer and the betting market, as a sanity check.
How far each team gets: how often they reached each round across the 50,000 simulations. The odds pile onto the favorites as the field thins.
Each group's live standings next to where the model predicted the teams to finish, so you can see who is running ahead of the forecast and who is behind. Expand a group for the match-by-match predictions and results.
The bracket view arrives once the group stage finishes and the round of 32 is set.
Every fixture by kickoff time, with the model's predicted score. Real results slot in as games are played. Knockouts join once the bracket is set.
How the Netherlands' tournament might play out under the selected model: a likely first-goal minute, half-time and full-time score for each game. Knockout rounds fill in once the bracket is set.
Each card shows a single scenario, so the minute, the half-time and the full-time line up, all under the selected model. Under Pure Market the bookmakers fancy the Netherlands in all three group games, even against Japan, where the data-only models call it a coin flip.
Most to least:
None of it has ever been tested on a real World Cup.
How France's tournament might play out under the selected model: a likely first-goal minute, half-time and full-time score for each game. Knockout rounds fill in once the bracket is set.
Each card shows a single scenario, so the minute, the half-time and the full-time line up, all under the selected model. France are clear favorites in Group I: Norway (with Haaland) and Senegal are the real tests, Iraq the softest.
Most to least:
None of it has ever been tested on a real World Cup.
Five ways to call a result, all on the exact same tournament: 104 games, the eight best third-placed teams, the official FIFA bracket (all 495 round-of-32 line-ups), and the knockout through the final with extra time and penalties, played out times. Only the match estimate changes between them. Simplest first.
Pure Elo. Ranks teams on results alone, like a chess rating, then turns the gap into expected goals. Hard to fool with one freak scoreline, but blind to how a team plays, and it backs the favorites hardest. (Ratings from eloratings.net.)
Pure Goals. Learns each team's attack and defense from 15,431 internationals since 2010 (recent games weigh more), so it predicts scorelines, not just a winner. It found a home-field bump of about 1.31x on goals, straight from the data. The catch: it rewards running up the score on weak teams, so it overrates the likes of Japan and Algeria. (Dixon-Coles is a standard football scoring model.)
Hybrid. The two above, averaged, so each patches the other's blind spot. Tested on 1,230 unseen internationals from 2025 and 2026, it grades at least as well as either alone, with the steadiest confidence of the three:
The honest read. Three things make a forecast good: how often it calls the result, how surprised it was by what happened (log-loss, lower is better), and whether its confidence matches reality (calibration). The Hybrid wins or ties on all three, and is the best calibrated: when it says 60 percent, about 60 percent happens. It clearly beats the goals model but edges plain Elo only by a whisker, well inside the noise. So the honest claim is modest: blending is at least as good as the best single model and steadier, mostly by calming the goals model down. More complicated is not the same as more accurate.
Hybrid + Market. The Hybrid nudged halfway to the betting market: take the bookmakers' title odds, back out the team ratings they imply, blend 50/50 with the model, re-run the tournament. Roughly how many professional models work. The effect: Argentina down, France up.
Does following the bookies pay off? The World Cup market could not be tested directly, since nobody kept the historical odds, so the idea was checked on 5,327 club matches where closing odds exist. The market beat a solid Elo model on every measure, and a 50/50 blend improved on the model but never beat the market alone. So leaning on the market makes a model better, not better than the market. The caveat: that was club football. The World Cup market is probably just as sharp, but that is not proven here.
Pure Market. Just the bookmakers, no model. Their ratings are calibrated so the simulated title odds match the published ones, then the bracket plays out, so you see the market's read on every group and knockout tie, not only the winner. Given the test above, the one to trust. The catch: it is the market through this engine's bracket, pinned to published title odds, not a live feed of every match price.
None of these models was trained on Opta's numbers, so this is a fair outside check. The selected model's title odds sit right next to Opta's supercomputer and the betting market:
Why models that rate single games almost identically still disagree wildly on the trophy. A tiny edge in one match compounds across seven knockout rounds and a 48-team field, so a small rating gap balloons into a big gap in who lifts the trophy. Spain's number swings from about 10 to 29 percent across the five, far wider than the match odds ever differ. Read the five less as rival predictions, more as a stress test: how much does the answer depend on the method?
This engine models team scores, not players, so it has no top-scorer odds. For reference, roughly where the betting market has it, with the margin stripped out:
It backs the favorites too hard. A fixed-strength sim cannot see a tournament turn (an injury, a player catching fire, a red card), so it piles probability onto the big teams. Pure Elo most, the market least. (The goals model spreads things out more, but that is goal-padding, not wisdom.) The head-to-heads stay sensible: Spain against Argentina is a coin flip.
What the test can and cannot prove. Those 1,230 test games are mostly qualifiers and friendlies, often one-sided, where any decent model looks clever. Tight ties between top teams, the games a World Cup turns on, barely feature. And the test grades the match calls, not the title odds, which nobody can check until the trophy is handed out. Encouraging, not gospel.
Injuries are not in it. The model knows results up to 6 June 2026 and nothing about who is fit for kickoff. Brazil, for one, is rated at full strength.
It draws a bit less than real life. International football, full of lopsided qualifiers, draws less than club leagues, and the model learned that. Its group draw rate is around 23 to 24 percent, just under the World Cup norm of 26 to 28. The honest fitted number, not tuned up.
The single "most likely score" is fragile. In an even game 1-1, 1-0 and 0-0 sit within a few percent of each other, so the headline score flips on a whim. Trust the win/draw/loss split, not the exact number.
Small print. The blend weight and the Elo-to-goals conversion were tuned on an Elo rebuilt from the same match data, which lines up 0.97 with the official ratings used live; any flat gap cancels out in a given matchup.
This is a toy, not an oracle. A transparent baseline from public data and standard methods. Great for the shape of the tournament. Not for staking your mortgage.