Round of 32
South Korea
48%
Canada
52%
Brazil
63%
Japan
37%
Germany
70%
Australia
30%
Netherlands
56%
Morocco
44%
Ivory Coast
41%
Norway
59%
France
87%
Egypt
13%
Mexico
52%
Uruguay
48%
England
70%
Austria
30%
Iran
52%
Senegal
48%
United States
32%
Bosnia-Herzegovina
68%
Spain
64%
Algeria
36%
Portugal
66%
Croatia
34%
Switzerland
68%
Sweden
32%
Paraguay
41%
Belgium
59%
Argentina
71%
Saudi Arabia
29%
Colombia
87%
Ghana
13%
Round of 16
Canada
39%
Germany
61%
Brazil
69%
Norway
31%
Netherlands
42%
France
58%
Mexico
42%
England
58%
Spain
55%
Portugal
45%
Iran
78%
Bosnia-Herzegovina
22%
Belgium
43%
Colombia
57%
Switzerland
41%
Argentina
59%
Quarterfinals
Germany
43%
Brazil
57%
Spain
72%
Iran
28%
France
61%
England
39%
Colombia
45%
Argentina
55%
Semifinals
Brazil
45%
Spain
55%
France
45%
Argentina
55%
Final
Spain
48%
Argentina
52%
Prediction Theory
Elo ratings
Every team starts at a base rating and is updated after each historical match using a standard Elo formula: the favorite gains less for an expected win, the underdog gains more for an upset. A home-field bonus is applied for non-neutral matches, and a margin-of-victory multiplier means a 4–0 win shifts ratings more than a 1–0 win. Ratings are computed walk-forward through ~49,000 historical international matches, so each prediction only ever uses information that would have been available before that match was played.
Recent form
Alongside Elo, each team's points-per-game over its last 10 matches (3 for a win, 1 for a draw) is tracked as a separate signal — capturing short-term momentum that a slower-moving Elo rating alone can miss.
The model
A gradient-boosted classifier is trained on Elo ratings, Elo difference, form, form difference, and a neutral-venue flag for every historical match, predicting the probability of a home win, draw, or away win. For knockout matches — which can't end in a draw — the draw probability is split evenly between the two teams before picking a winner.
How reliable is this? (backtested accuracy)
This same model (retrained from scratch on data available before each tournament, so no future leakage) was run against the actual 2014, 2018, and 2022 World Cups:
| Tournament | 3-class accuracy | Knockout winner accuracy |
|---|---|---|
| 2014 Brazil | 65.6% | 75.0% |
| 2018 Russia | 53.1% | 56.2% |
| 2022 Qatar | 46.9% | 62.5% |
| Combined (192 matches) | 55.2% | — |
Roughly 55% exact-outcome accuracy and ~55–65% knockout-winner accuracy across three real tournaments — clearly better than chance (33% for a random guess among home/draw/away) and better than always picking the favorite, but far from certain. Confidence is reasonably well calibrated in the middle of the range (matches the model called at 60–70% confidence were right 69% of the time) but gets noisier at the extremes. Football stays genuinely upset-prone: in 2022 alone, Saudi Arabia beat Argentina, Japan beat both Germany and Spain, and Morocco reached the semifinal.
Why this bracket is a projection
The group stage is still being played. Any group match that hasn't kicked off yet is resolved using the model's single most-likely outcome, and standings are updated accordingly to determine each group's winner, runner-up, and whether its third-place team is among the 8 best third-place finishers overall. The official FIFA rule for which third-place finisher lands in which Round-of-32 slot is a published allocation table this app doesn't have access to — instead, it finds *a* valid assignment respecting each slot's eligible groups. Every later round then follows automatically from "if the model's pick wins every prior match." Treat this as "the most likely single path through the bracket," not a probability-weighted forecast of who's most likely to win the tournament overall.