World Cup 2026 Predictions

The prediction system has three layers.

The first layer is a Dixon-Coles bivariate Poisson model, fit on international match results from 2005 through June 2026 — around 48,000 matches across 259 national teams. Each team gets an attack rating and a defense rating; each match has a home advantage adjustment (γ = 0.232, set to zero for neutral venues). Expected goals for the home team are a function of the home team's attack, the away team's defense, and the home advantage. Same for the away team, mirrored. Given expected goals for both sides, the probability of any specific scoreline is the product of two Poisson distributions — with one twist. Independent Poissons over-predict outcomes like 2-2 and under-predict 0-0, 1-0, 0-1, 1-1. Dixon-Coles introduces a small correlation parameter (ρ = -0.058) that nudges those four low-score cells to match observed frequencies. The full match distribution over scorelines follows from there.

The second layer is a time-weighted Elo rating that assigns each team a strength score, updated after every match, decayed by a two-year half-life. This weights recent form more heavily than historical form and provides a second view on team strength that's more responsive to short-term changes than the Dixon-Coles fit.

The third layer is a Monte Carlo bracket simulation. For each of 10,000 simulated tournaments, the model plays out every group-stage match using the Dixon-Coles score distribution, resolves standings by FIFA tiebreakers, seeds the knockout bracket per FIFA's rules, and simulates each knockout match. Championship probabilities are the fraction of simulations each team wins.

Data quality matters. The training set covers major tournaments (World Cup, continental championships), qualifiers, and friendlies. Matches are weighted by recency but not by competition importance in the Dixon-Coles likelihood — a gap discussed below. Ratings are refit whenever new results come in; the version powering this page uses data through June 10, 2026.

Full code and data are available at https://github.com/amin8448/worldcup2026. The methodology is a variant of the approach in Dixon & Coles (1997), extended for tournament simulation.

Every model has known limitations. Naming them is more useful than hiding them.

Team strength is not the 2026 squad. The model rates "Argentina the national team" as a continuous entity whose strength evolves through results over time. It doesn't know that Messi is 38, that specific players are injured, that certain managers have made tactical changes, or who's in June form. For a tournament that hinges on 23-player squads and last-minute fitness, this is a real structural blind spot. The two-year half-life on ratings mitigates the historical drag, but the model still can't see the roster.

Championship percentages are not backtestable. The single-match probabilities are well-calibrated — reliability diagrams across probability bins line up with observed frequencies. But "Argentina 14.1% to win the tournament" is the product of seven survival probabilities compounded together. Small systematic biases in single-match probabilities multiply across rounds, and there is no way to validate the compounded product empirically. There's only ever one tournament per cycle. Match-level calibration is real; championship-level calibration is a best estimate.

Knockout simplifications favor underdogs. The knockout simulation makes three simplifications that all point the same direction: no extra time (drawn knockouts go straight to penalties), penalties resolved as a 50-50 coin flip, and neutral venue enforced for all knockout matches. Real knockouts have 30 minutes of extra time where the stronger team is more likely to score, penalties have a small skill component, and hosts often play at home into the round of 16 and quarterfinals. The combined effect is that the model is too kind to underdogs in knockouts and too harsh on hosts. USA, Mexico, and Canada's championship odds are biased downward for this reason.

Data-quality weighting is inconsistent. The Elo layer weights matches by competition importance — World Cup games count 3x more than friendlies. The Dixon-Coles fit that actually drives the published predictions has no such weighting, only time-decay. A low-stakes June friendly where the manager rests starters counts as much as a World Cup qualifier for a team's attack and defense ratings. Adding competition weighting to the DC likelihood is the single most fixable real problem in the current version.

Rating uncertainty is not propagated. Monte Carlo captures one source of uncertainty — which team wins given fixed ratings — but nothing about whether the ratings themselves are right. If a team's rating is systematically off, it's off in all 10,000 simulations. Teams with thin match histories (Jordan, Uzbekistan, Cape Verde) have noisier point estimates, but the simulation treats them as certain. A bootstrap over the Dixon-Coles fit would give honest error bars; they'd be wider than the current numbers suggest.

The practical takeaway: treat close probabilities as close. Argentina at 14% and Spain at 12% are effectively co-favorites, not a genuine ranking. The confidence you should place in single-match predictions is higher than the confidence you should place in the tournament winner.