Run a tournament 50,000 times and you don't get a prediction — you get a distribution. Here's the difference, why it matters, and what to do with it.
We say it on the homepage. We say it on the methodology page. It's in the metadata of every blog post.
50,000 daily Monte Carlo simulations.
It's a real number — we've verified it in our own code, and you can see the audit. But the number itself doesn't matter as much as what it represents. This post is about the actual idea behind running a tournament 50,000 times instead of trying to predict it once.
If you've read the World Cup model post, this is the deeper version of one of its sections. If you haven't, this should still make sense on its own.
When someone asks "who's going to win the World Cup?", the wrong answer is a single team name.
The right answer is a probability distribution: "France 12.4%, Spain 13.2%, England 11.6%, Argentina 13.2%, Brazil 8.6%, ..." and so on across all 48 teams.
This sounds like a hedge. It isn't. It's the actual answer to the question, given the structure of tournaments.
Tournaments don't have one outcome. They have many possible outcomes, weighted by probability. France winning is one outcome. England winning is another. A surprise CONMEBOL team running the bracket is a third. Each of these is more or less likely. Saying "France will win" treats one possibility as if it were the answer; saying "France 12%" describes the actual landscape.
In principle, you could calculate tournament probabilities directly. For each possible bracket path, multiply the per-match probabilities. Sum across paths. Read off the answer.
This works for simple tournaments. For a World Cup with 48 teams, 12 groups, 32 knockout slots, multiple group-stage tiebreaker rules, and possible extra time and penalties, the math gets impractical fast. You'd need to enumerate millions of distinct bracket paths and probability-weight each one. It's possible but tedious, error-prone, and inflexible — every rule change forces a redo.
Simulation sidesteps this. Instead of calculating, you generate.
One run of our simulator does this:
That's one simulation. It produces one tournament outcome. France beats Brazil in the final 2-1 after extra time, say.
A single simulation tells you almost nothing. It's just one possible tournament. Out of all the worlds where 2026 plays out, you've witnessed one of them.
Do that 50,000 times.
Now you have 50,000 distinct tournament outcomes — 50,000 champions, 50,000 group stage exits, 50,000 quarter-final pairings.
Count what happens.
These aren't predictions. They're frequencies in the simulated multiverse, given the team strengths and rules we fed in. If our team strengths are roughly right and our match-outcome model is roughly right, these frequencies approximate the real-world probabilities of each outcome.
This is the core of Monte Carlo simulation. You don't predict; you simulate, and let the frequencies become your forecast.
The number is a trade-off.
Too few simulations, too much noise. With 100 simulations, your "France 15%" might actually be France 11% or 19% — you've got too few data points to be confident in the third significant digit. With 1,000 simulations, you're closer but still meaningfully noisy.
Too many simulations, diminishing returns. Once you're at 50,000+, the marginal gain from each additional simulation is tiny. The headline probabilities for the major teams stabilise to within 0.1-0.3 percentage points. Going to 500,000 would tighten this further but gain you almost nothing in practical predictive value.
50,000 is in the sweet spot for a daily-refresh model. It's enough to produce confident probability estimates without taking so long to compute that we can't refresh it daily.
We re-run the full set every 24 hours. As Elo ratings shift, as recent international matches feed into the model, as the bracket draw is finalised, the simulator picks up the new state and produces a new distribution.
For each team, every simulation produces a path through the tournament. Across 50,000 simulations, we get the full distribution of where each team ends up.
This is what the Path to Glory cards on team pages show:
Each number traces back to the count of simulations where that team reached that stage. The probabilities decline along the path because each round adds another opportunity to be eliminated.
A tournament favourite typically sits at 12-18% to win the World Cup. People sometimes ask why it's not higher.
The answer is the compounding effect of single-elimination matches. A team that's a 60% favourite in each of their 5 knockout rounds wins all five 7.8% of the time:
0.60 × 0.60 × 0.60 × 0.60 × 0.60 = 7.8%
To get to 25%+ probability of winning the tournament, you'd have to be roughly a 75% favourite in every single knockout match. No team is that dominant against the field of teams that survive to the knockout stage.
This is also why specific match win probabilities seem more confident than tournament championship probabilities. France beating Senegal in a single match? Maybe 65%. France winning the whole thing? 12-15%. Single-match confidence and tournament confidence are different scales of the same model.
Some things that simulations don't capture, even with 50,000 runs:
Things that aren't in the input data. If our team strength estimates are wrong because of an injury we don't know about, the simulator confidently runs the wrong tournament 50,000 times. Garbage in, garbage out.
Tail events that the model doesn't allow. Our Dixon-Coles model can produce 0-0 through 9-9 scorelines, but the probability of a 7-0 thrashing is set by the parameters. If the model under-weights extreme blowouts, no number of simulations recovers that.
The specific tournament that's about to happen. The simulations describe the probability landscape; the actual 2026 World Cup is one specific draw from it. Even if our 12% on France is exactly right, France will either win or not win — the 12% doesn't manifest as a slow-motion frame-by-frame in reality.
This last point matters. After the tournament, France either won or didn't. Saying "the model said 12%, France won, the model was wrong" misunderstands what the model says. The model said 12%. That's a high enough probability that France winning is consistent with the prediction. It's not a guarantee, just like 88% wasn't a guarantee they'd lose.
Three honest applications:
Compare against bookmaker odds. Bookmakers also produce a distribution (after stripping their margin). Where their distribution and our distribution disagree meaningfully, that's a signal — either we're seeing something the market hasn't priced, or the market knows something we don't. Both are possible.
Track how it shifts over time. Daily probability updates show which teams are gaining or losing strength as new data comes in. Big movers are interesting. Small movers across 6 weeks accumulate into something meaningful.
Use it to think probabilistically. Most football conversations are deterministic ("they'll win" or "they'll lose"). Simulation outputs train you to think in distributions. After enough exposure, you stop saying "France will win" and start saying "France will probably win, but the bracket variance is brutal."
That's the actual value of the model. Not a tip. A way of seeing.
50,000 simulations isn't a magic number. It's enough to produce stable estimates without absurd compute. Done daily, it becomes the engine for everything else.
OddsIQ provides AI analysis, not financial or betting advice. Past performance does not guarantee future results. Gamble responsibly: BeGambleAware, GamCare, GamStop.