First published on 8 June 2026 to explain the model before the tournament. Rewritten in October 2026 with how it performed.
During the 2026 World Cup, Mundialoop published a win probability for every team and a forecast for every match. None of it was a pundit's hunch: it all came from one statistical model, run the same way every time. Here is how it worked, in plain language, and how it held up against what actually happened.
Step 1 — Elo ratings measured team strength
Every national team carried an Elo rating, a single number summarising its strength, borrowed from chess. When two teams met, the gap between their ratings translated into an expected result: a 100-point edge meant a clear favourite, while teams a few points apart were close to a coin flip. Ratings rose after good results against strong opponents and fell after poor ones, so they tracked recent form rather than reputation. Because the tournament was in North America, the United States, Canada and Mexico received a modest host advantage.
Step 2 — A Poisson model turned strength into goals
Knowing one team is stronger is not enough: football is decided by goals, and goals are scarce and streaky. Each side's strength was converted into an expected scoring rate, and the Poisson distribution — the standard tool for counting independent events in a fixed period — gave the probability of every plausible scoreline. That is how the model could say a match was most likely to end 1–1 while still giving real weight to 2–0 or 0–0.
Step 3 — Thousands of simulated tournaments
To put a number on each team's title chances, the model played the whole tournament thousands of times: all 104 matches, through the real 12-group draw, the best-third-placed calculation under the real tie-break rules, the Round of 32 and every round to the final. A team's title probability was simply the share of simulations it won. After every real result, ratings moved and the bracket filled in, so the odds always reflected the latest state of the competition.
How it did: the 104 matches
We grade forecasts with the ranked probability score (RPS), the usual measure for win–draw–loss predictions: it punishes a confident wrong call more than a cautious one, and lower is better. Scored on the 90-minute result of all 104 World Cup matches, using only probabilities published before kick-off, the tournament model scored 0.1552. A plain Elo model on the same matches would have scored 0.1670, and an Elo rating built from every national team in the world, 0.1622.
In the check we ran after 75 matches (the whole group stage and the first three Round of 32 ties), the outcome the model rated most likely — home win, draw or away win — happened in 47 of them, and it called the exact score in 9.
The clearest lesson was about confederations. A traditional Elo underrates African teams and overrates CONCACAF teams when they face other regions: over 2019–2026, CAF sides beat their expected goal difference by 0.45 goals a game in inter-confederation matches, and CONCACAF sides fell 0.41 short. A World Cup is full of those match-ups. Adding a per-confederation adjustment to Elo closed the whole gap to the tournament model.
What changed after the World Cup
Since September 2026, our national-team predictions come from a newer model that blends an Elo rating — adjusted for venue, altitude, travel and confederation — with a goals-based strength model refitted every week. We chose it on a much wider test: 6,964 internationals between January 2019 and August 2026, each predicted without seeing the result. It scored 0.1626 there, against 0.1671 for the best Elo. On the 104 World Cup matches it ties the tournament model (0.1549 against 0.1552, a gap well within noise); its advantage shows up in the rest of the calendar, at neutral venues, in inter-confederation games, the Nations League and smaller cups.
What the numbers mean, and what they don't
A 12% title probability does not say a team will fall short; it says that if the tournament could be replayed many times from that point, the team would win roughly one in eight. Upsets are not model failures but the low-probability events the model expects to happen sometimes — Brazil going out to Norway in the round of 16, or third-placed Paraguay eliminating Germany. The model also knows nothing about a late injury, a red card or the weather on the day. Treat its percentages as a calibrated baseline to argue with, not a crystal ball.
Every prediction on Mundialoop is a statistical estimate published for editorial and entertainment purposes. It is not betting advice, and we are not affiliated with FIFA or any sportsbook.
The full comparison, with calibration tables, is on our Model accuracy page. For upcoming national-team matches, the Predictions page shows the current model's probabilities.