The whole World Cup, all 104 matches across the group stage and knockouts, simulated 50,000 times, five different ways, from past results to the bookmakers' odds. Switch models below and explore each tab.
Each bar is a team's chance of winning the tournament under the selected model. On the Day 0 view the dashed ticks mark Opta's supercomputer and the bookmakers as outside comparisons (both are pre-tournament figures, so they line up with Day 0, not the live odds). On the live view a tick marks where each team started, so you can see how far it has moved, and the bars show only the teams still in it; switch to Day 0 for the full pre-tournament field.
How far each team gets: how often they reach each round across the 50,000 simulations. The odds pile onto the favorites as the field thins. The live view narrows to the teams still in it and the rounds still to play; switch to Day 0 for the full pre-tournament table.
Each group's final standings next to where the model predicted the teams to finish, so you can see who beat the forecast and who fell short. Expand a group for the match-by-match predictions and results.
The model's pick and advance odds for every knockout match, with the real result slotting in color-coded, round by round. Predicted scores are the most likely final result from simulating the tie (90 minutes, plus extra time when level, then penalties if still level), so they are usually decisive; each prediction also shows the chance it goes to extra time and the smaller chance it goes to penalties.
Every fixture by kickoff time, with the model's predicted score. Real results slot in as games are played.
The bookmakers' pre-tournament pick for the Golden Boot, next to who is actually scoring. The model rates teams, not players, so the expected column comes from the bookmakers and does not move during the tournament.
One tournament, played out times. Each run plays the 72 group games, ranks the eight best third-placed teams, fills the official FIFA round of 32, and plays the bracket through to the final, with extra time and penalties when a tie needs them. The engine underneath stays the same. What changes between the five is how a single match is rated, from the ones that use only past results to the ones that use the betting market.
The first three use results only. Pure Elo treats a team like a chess rating, going on who has been beating whom. It is steady and hard to fool with one odd scoreline, it has no feel for how a team plays, and of the five it leans on the favorites most. Pure Goals learns each team's attack and defense from internationals since 2010, where the home side scores about 1.31x as often, so it works in scorelines rather than just winners. It also rewards a team for scoring heavily against weaker opponents, which lifts teams like Japan and Algeria. Hybrid averages the two, and each covers for the other's blind spot. On 1,230 internationals it had not seen, it scores at least as well as either one on the probability measures, with the steadiest confidence of the three:
The three columns each grade a forecast a different way (the same three appear in both tables). Acc (accuracy, higher is better) is the plainest: how often the model's most likely outcome is the one that happened. Log-loss (lower is better) grades the probabilities themselves, not just the pick, and punishes a confident call that turns out wrong far more than a hedged one. RPS (ranked probability score, lower is better) does the same but gives partial credit for being close, since a home win sits nearer a draw than an away win. Hybrid leads on log-loss, ties Pure Elo on RPS, and gives up a point of accuracy to Pure Goals. On the probability scores it clearly beats the goals model and is just barely ahead of plain Elo, close enough that either could be best, so blending lands at least as well as the best single method and a little steadier. A more complicated model did not mean a more accurate one.
The last two bring in the bookmakers. Hybrid + Market reads the ratings the betting odds imply and meets Hybrid halfway, which pulls Argentina down a little and lifts France. Pure Market drops the model and runs the bookmakers' implied ratings through the same bracket, so their read covers every tie and round, not just the title. It is the bookmakers as they stood before kickoff, played forward and then held fixed.
Those two are absent from the table above for a reason: they are built from the betting market, and there is no history of international match odds to hold out and score them on. The nearest test is club football, where every game is priced. Across 5,327 held-out club matches with real closing odds:
The market beat a competent Elo model on every measure, and blending the two improved on the model without catching the market, so the page opens on Pure Market. The one caveat: this was club football; the World Cup market is likely just as sharp, though that is not proven here.
None of these models was trained on Opta's numbers, so they make a fair outside check. Here are the selected model's Day 0 (pre-tournament) title odds beside Opta's supercomputer and the bookmakers, all three at the same point before kickoff:
The Bookmakers reference and Pure Market are not quite the same line, though they sit on top of each other on the title bar. The Bookmakers figure is the published title odds, champion only. Pure Market takes the ratings those odds imply and plays the tournament out 50,000 times, so it matches them on the title and also has a number for every round and match, which the raw odds do not.
It backs the favorites too hard. A fixed-strength simulation cannot see a tournament turn the way a real one does, an injury, a sudden run of great form, a red card at the wrong moment, so probability piles onto the big teams (Pure Elo most, Pure Market least). The head-to-heads still look right: Spain against Argentina is close to a coin flip.
The live odds only see the score. Once a game is played the odds re-simulate with that score locked in, but the underlying ratings stay at their pre-tournament strength, frozen on June 6. Winning by a little and winning by a lot move them the same amount, and injuries, current form, and a re-pricing market do not show up at all.
The validation only goes so far. The 1,230 test games are mostly one-sided qualifiers and friendlies that almost any model gets right, with very few of the tight knockout matches a World Cup turns on. And the test grades individual match calls; the title odds themselves cannot be checked until the trophy is lifted.
The single most likely score is shaky. In an even game 1-1, 1-0 and 0-0 sit within a few percent of each other, so the headline score can flip on very little. The win, draw and loss split is firmer than the exact number. The model also draws a little less than real life, around 23 to 24 percent in the groups against a 26 to 28 percent norm.
There is no top-scorer model. The engine rates teams, not players. The Top Scorers tab puts the bookmakers' pre-tournament pick next to whoever is actually scoring.
It is only a baseline. Built from public data and standard methods, it is good for the shape of a tournament and a feel for the odds. Do not stake anything on it you would miss.
Five match models over a shared -tournament Monte Carlo with the official FIFA Annex C round of 32 (495 line-ups). Dixon-Coles fit by maximum likelihood on internationals (2010 to 2026), the hybrid weight chosen by out-of-sample log-loss, bookmaker margins stripped out and the Pure Market ratings calibrated to them. Opta and the bookmakers (Polymarket, Kalshi, FanDuel) are shown for comparison only.