Predicting a World Cup is harder than predicting a league. Here's how we simulate the entire tournament 50,000 times every single day — and what the numbers can and can't tell you.
Predicting a football match is hard. Predicting an entire World Cup is a fundamentally different problem. Here's how we model 48 teams, 12 groups, and 104 matches running from June 11 to July 19 — and why we rerun the whole tournament 50,000 times every single day.
If you haven't read how our Premier League model works yet, start there. The core stack is the same: Dixon-Coles Poisson for match outcomes, Elo ratings for team strength. But tournament prediction adds a layer that domestic leagues don't have. That's what this article is about.
Leagues are forgiving. Over 38 Premier League matches, a team's true strength dominates. A lucky bounce here, a red card there, none of it matters enough to overturn the underlying truth. Manchester City don't win the league because they caught a break; they win because they're the best team across 9 months.
Tournaments are ruthless. 48 teams play 3 group matches, and then it's single-elimination. One bad half against a well-drilled underdog and you're on a plane home. A Cristiano Ronaldo penalty miss, a set-piece on 89 minutes, a bad VAR call — these moments outweigh 300 minutes of tactical superiority.
This means tournament predictions must explicitly account for variance. A team that's 10% better than their opponent will win a league table over 38 matches. In a single knockout match, they might lose 45% of the time. Across a full tournament path, those compounding single-match risks matter enormously.
That's why we use Monte Carlo simulation, not closed-form math.
Imagine you could play the entire World Cup 2026, from kickoff to final whistle, on a computer. Every match, every group, every knockout round. At the end, one team lifts the trophy.
Now imagine you could do that 50,000 times.
Each simulation plays out slightly differently because match outcomes are probabilistic, not deterministic. Spain beats England in one simulation, loses in the next. Morocco makes the quarter-finals in 20% of simulations, crashes out in the group stage in the other 80%.
By the end of 50,000 simulations, we have 50,000 "champions." Count them up and you get your probability distribution:
That's the championship race you see on the World Cup 2026 page. Every number traces back to a specific count of how often that team won.
We re-run the full 50,000 simulations every day. If a key player picks up an injury, if a team's form improves dramatically, if the model learns from a recent international match — the probabilities shift. Usually by small amounts. Occasionally by a lot.
Every simulation starts with the group stage. 48 teams are drawn into 12 groups of 4. Each team plays the other three in their group. Top two advance automatically; eight best third-placed teams also advance to give us 32 for the knockout.
For each group-stage match, we run the same Dixon-Coles Poisson model we use for the Premier League. The inputs:
We simulate the match to produce a scoreline. We tally points (3 for a win, 1 for a draw), goal difference, goals scored. At the end of three matches, we rank each group.
For the eight best third-placed teams, we compare across all 12 groups using the actual FIFA tiebreakers: points, goal difference, goals scored, and so on. This matters because who finishes third in Group A plays a completely different R32 opponent than the third-placed team in Group J.
Club football Elo and international football Elo are not the same beast. A few reasons:
Match density. Premier League teams play every 3-4 days for 9 months. National teams play 10-12 competitive matches a year, spread across qualifiers, friendlies, and tournaments. Elo adjustments that work for clubs are too aggressive for international teams — you end up with wild swings after a single loss to a minnow.
We use a slower K-factor for international matches. One upset doesn't overturn years of accumulated strength.
Friendlies vs competitive matches. A friendly against Luxembourg tells us less than a qualifier against Italy. We weight match importance — competitive tournaments and qualifiers count more than friendlies.
Confederation adjustments. This is the one that actually matters for tournament prediction, and it's controversial enough to deserve its own section.
Here's an observation backed by decades of World Cup data: teams from CONMEBOL (South America) tend to overperform their Elo rating at World Cups, especially in the knockout rounds. Teams from weaker confederations tend to underperform.
Why? Nobody knows for sure. A few plausible theories:
Whatever the cause, the pattern is real. A naive model that treats Elo as the only input would systematically underrate Brazil, Argentina, and Uruguay at World Cups. We include a confederation adjustment — a small, empirically-derived boost to South American teams in tournament settings. It's not large (a few percentage points on win probability), but over the course of 7 knockout rounds, it compounds.
We also apply smaller adjustments based on historical tournament performance by confederation. CONCACAF teams get a small home boost in 2026 because the tournament is in North America. African teams have outperformed their Elo in recent World Cups, so they get a nudge. Asian teams have underperformed against their Elo, so they get a small negative adjustment.
Are these adjustments biased? Arguably. They're based on historical patterns that may not persist. We tune them to fit past tournaments and we'll update them based on 2026's results. This is a judgment call where reasonable modellers could disagree. You should know we're making it.
Once the group stage simulation completes, we have 32 teams, seeded into the Round of 32 by FIFA's bracket rules. From here, it's single-elimination: R32 → R16 → QF → SF → F.
Each knockout match uses the same Dixon-Coles Poisson model. Scores are level at 90 minutes? We simulate 30 minutes of extra time, tuned to historical extra-time goal rates (which are lower than regular-time rates — teams tire, defences compact). Still level? Penalties. We model penalty shootouts as a weighted coin flip, slightly favoring the team with the better penalty record and slightly penalizing teams whose key penalty takers are fatigued or carrying knocks.
The compounding effect is dramatic. Consider a team that's 55% favorite in each of their 5 knockout matches. Sounds like they should cruise through, right?
Their probability of winning ALL FIVE is: 0.55 × 0.55 × 0.55 × 0.55 × 0.55 = 5.0%
Only a 5% chance to run the bracket, despite being a favorite every time. This is why the championship race percentages you see are so compressed. Even the tournament favorite usually sits around 15-18% — because the odds of winning 5 in a row are punishingly low, even when you're the better team every time.
Two teams with identical Elo ratings can have very different championship probabilities, because the bracket draw means they face different opponents.
A team in the "weaker" half of the bracket might face a 1600-Elo opponent in the QF while the other half's QFs feature a 1950-Elo giant. Our simulations capture this automatically — by playing out every match, we correctly weight the expected path each team faces.
When you look at the "Path to Glory" section on a team's share page, you're seeing stage-by-stage probabilities that account for this. France's 65% chance of reaching the Round of 16 reflects not just their strength, but the specific opponents they're likely to face.
On top of the domestic-league blind spots (injuries, referees, tactical changes), international tournaments add new unknowns:
Squad cohesion. National teams have 4-6 weeks together before a tournament. A group of talented individuals doesn't always mean a cohesive team. Our model can't see whether the manager has fallen out with the star player.
Travel and climate. Matches across the US, Canada, and Mexico span dramatic climate shifts. A team playing in Mexico City one day and Vancouver a few days later faces a physical challenge no model I've ever seen accounts for.
Penalty shootouts. We model these as weighted coin flips with small adjustments. But penalty shootouts are notorious for defying prediction. Germany have an 83% historical shootout win rate; England 14%. Is that causal (technique, temperament, preparation) or just a small sample? Probably both, and we honor the pattern without overweighting it.
Refereeing interpretations. VAR has added consistency but introduced new variance. A handball rule change between 2022 and 2026 could swing tight matches in ways we can't predict.
Two ways to engage:
Champion Pick — lock one team you think will win. Your probability is frozen at pick time. If you're right, leaderboard glory and 32 bonus points in the knockout pick'em.
Knockout Pick'em — predict winners round by round as the bracket unfolds. Score doubles each round. Maximum 80 points from picks, 112 with the champion bonus.
The AI plays alongside you, using its own simulations to make picks. If you beat the AI on the leaderboard, that's a real achievement — it's not setting the bar low. Over 50K simulations, the model is making informed probabilistic choices; beating it means you made better single-choice calls.
If you want to see how accurate our predictions have been historically, the calibration page is our honest scoreboard. Perfect calibration is rare. Overconfidence is the more common failure mode. We publish ours.
This is a prediction tool and a free game. It's not financial advice. It's not a betting tip service. It's a statistical model, playing out in public, run by humans who will definitely be wrong sometimes.
If you're betting on these matches: bet small, bet informed, bet what you can afford to lose. If betting isn't fun anymore, BeGambleAware, GamCare, and GamStop are free, confidential, and 24/7.
Enjoy the tournament. 48 teams. 104 matches. One trophy. 50,000 simulations every 24 hours, refined continuously, published publicly.
See you on June 11.
More reading: How the Premier League model works · Methodology page · Calibration chart