A Dixon-Coles model fitted to the expected-goals record of all 102 matches of the 2026 World Cup. Attack and defense estimated from xG rather than goals, opponent-adjusted, extra time normalized to 90 minutes, low-score dependence fitted to actual results.
EVERY PREDICTION POSTED BEFORE KICKOFF · BACKTEST: FITTED ON THE GROUP STAGE ALONE, THE MODEL CALLED 23 OF 28 KNOCKOUT WINNERS (82%) · LOG LOSS 0.543 V 0.693 COIN FLIP
Graded · Semifinal one · Arlington. Model said Spain 56, France 44. The market favored France. Final: Spain 2, France 0, on 1.63 expected goals to 0.30 from ten shots each.
Graded · Semifinal two · Atlanta. Model said Argentina 56, England 44. The market favored England. Final: Argentina 2, England 1, on 1.84 expected goals to 0.53. Both calls landed, both against the market.
What this number is graded on. The backtest scores advance probability, and advance is provably invariant to the low-score dependence parameter ρ: it equals ½ + (win − loss) / 2, and the correction shifts win and loss by the identical quantity. So 82% and 0.543 validate the base rate, the attack and defense ratings, and the shrinkage. They do not validate ρ.
Graded directly against the knockout rounds, ρ is untestable at this sample size. Twenty-eight matches cannot separate ρ = 0 from ρ = -0.27. The knockouts went level after 90 in 25% of matches; the model expects 33%. The lift figure above is therefore the number this page stands behind. The draw probability is commentary, not an edge.
| TEAM | ATTACK | DEFENSE | NET |
|---|---|---|---|
| Spain | +0.57 | +1.17 | +1.74 |
| Argentina | +0.44 | +0.68 | +1.12 |
| France | +0.63 | +0.47 | +1.10 |
| Brazil | +0.64 | +0.28 | +0.91 |
| Colombia | +0.22 | +0.56 | +0.78 |
| Canada | +0.38 | +0.39 | +0.78 |
| Senegal | +0.60 | +0.14 | +0.73 |
| Mexico | +0.20 | +0.43 | +0.63 |
| Uruguay | +0.31 | +0.28 | +0.58 |
| England | +0.38 | +0.19 | +0.57 |
| Norway | +0.49 | +0.07 | +0.56 |
| Switzerland | +0.30 | +0.24 | +0.54 |
Attack is the log-rate a team adds to its own expected goals against an average opponent. Defense is the log-rate it subtracts from its opponent's. Spain's defensive rating of +1.17 is the largest single parameter in the tournament, and Spain's attack is not the best in the field. The rating is carried by suppression. Argentina's semifinal moved it past France into second on net.
Advance probability splits drawn matches evenly through extra time and penalties, a simplification that slightly understates the stronger side. Ratings are relative to this 48-team field and are not portable to club football. Published xG depends on a provider's model; the record used here is internally consistent, which is what matters for relative ratings, but it is not the only defensible set of numbers.
Estimates come from a six- and seven-match sample with ridge shrinkage. They are honest, not certain. A lean is not a promise.