In brief
Spain was our number-one favorite and Argentina our number two. They met in the final. But a strong pre-tournament ranking is not enough to validate a model. We therefore reviewed all 104 matches, 500 picks, confidence badges, likely scores and several advanced metrics to measure what genuinely worked — and what still needs improvement.
The review in 30 seconds
The data covers the entire tournament, including the final. Across 104 matches and 500 picks, 352 were correct: 70.4% accuracy across all markets and 64.4% on the 1X2 result alone. The six key figures are summarized in the hero image.
These figures provide an initial view of performance. They do not yet tell us whether the model mainly understood the tournament's overall balance of power, match outcomes or the precise scenario of each game.
Before the tournament, Spain and Argentina were already on top
You can read the predictions we published before the tournament. They have not been changed since and serve as the reference point for this review. Our pre-tournament article did not name a certain winner. It presented a probabilistic ranking based on tournament simulations.
Spain ranked first with a 14.6% chance of winning the title. Argentina followed at 11.7%, just ahead of France at 11.2%.
| Initial rank | Team | Title probability | Actual result |
|---|---|---|---|
| 1 | Spain | 14.6% | Champions |
| 2 | Argentina | 11.7% | Runners-up |
| 3 | France | 11.2% | Fourth |
| 4 | Brazil | 6.7% | Round of 16 |
| 5 | England | 5.7% | Third |
| 6 | Colombia | 5.4% | Round of 16 |
| 7 | Portugal | 5.2% | Round of 16 |
| 8 | Netherlands | 3.9% | Round of 32 |
| 9 | Germany | 3.6% | Round of 32 |
| 10 | Mexico | 3.2% | Round of 16 |
The top of the ranking held up remarkably well: the first two favorites reached the final, and three of the top five finished in the final four.
The full ranking was not perfect, however. Brazil, Colombia and Portugal did not progress beyond the round of 16, while Germany and the Netherlands went out in the round of 32.
The pre-tournament simulations identified both finalists and much of the final four remarkably well, without forecasting every individual run with the same precision.
352 correct picks out of 500
Across the five markets analyzed, Foresportia generated 500 picks during the tournament. 352 were correct, representing 70.4%.
This overall rate should not be treated as a homogeneous metric: a double-chance pick covers two of the three possible outcomes, while a 1X2 pick must identify a single result.
The real question is therefore where the model was genuinely strong and where its performance was much more ordinary.
Not all markets performed equally
Double chance: the most solid market
Double chance produced 92 correct picks out of 104, or 88.5%. The rate even reaches approximately 97% for 1X picks.
Draw No Bet: a good read on the balance of power
Of 84 decided picks, 67 were correct, or 79.8%. Home DNB picks reached approximately 87%.
BTTS and 1X2: correct, but less dominant
Both Teams to Score reached 65.4%. The 1X2 result produced 67 correct predictions out of 104, or 64.4%.
Over/Under 2.5 goals: the weakest market
The Over/Under 2.5 market reached only 55.8%. The model therefore identified relative team strength better than the expected goal total in individual matches.
The more conservative markets clearly performed best. Double chance and Draw No Bet made good use of the model's reading of relative strength, while goal scenarios remained more erratic.
Did the confidence levels actually help?
The expected hierarchy appears across all 500 picks: Stable first, then Correct, then Risk.
The picture is more nuanced for the 1X2 market alone because the very small Stable sample was heavily affected by two major surprises: Spain vs Cape Verde and Ecuador vs Curaçao.
Overall, the badges ranked reliability sensibly, but high confidence remains compatible with an upset.
Spain vs Cape Verde: when a Stable prediction fails
On June 15, Spain faced Cape Verde. A Spanish win was assigned 80.41%, with a Stable badge.
Result: 0–0.
An 80% probability does not mean victory is certain. It still leaves roughly one chance in five for the other outcomes. The real question is whether the draw was simply a rare event or whether the model underestimated a low-tempo match against a highly defensive outsider.
The model's main trap: the draw
Of the 37 recorded 1X2 errors, 20 were draws. More than half of the incorrect picks therefore came from matches in which the model identified a favorite that failed to win.
The model was not always wrong about which team was theoretically stronger. It was often wrong about that team's ability to turn superiority into a win.
The main 1X2 problem was not a systematic reversal of the balance of power, but a repeated underestimation of draws involving favorites.
A much more difficult group stage
| Stage | Correct outcomes | Matches | Accuracy |
|---|---|---|---|
| Group stage | 42 | 72 | 58,3 % |
| Round of 32 | 13 | 16 | 81,3 % |
| Round of 16 | 6 | 8 | 75,0 % |
| Quarter-finals | 4 | 4 | 100% |
| Semi-finals | 2 | 2 | 100% |
| Third-place play-off | 0 | 1 | 0% |
| Final | 0 | 1 | 0% |
The small samples in the later rounds prevent strong conclusions, but the contrast with the group stage is clear.
The group stage was the hardest. The model performed better once the tournament narrowed to better-known teams and more direct matchups.
A final anticipated at tournament level, but poorly read at match level
Before the competition, Spain was the number-one favorite and Argentina number two. They did indeed reach the final.
On the match itself, only one of the five picks was correct: double chance 12. The 1X2, Over 2.5, BTTS and DNB picks all failed.
Foresportia read the tournament hierarchy very well, but misread the final's match scenario.
The higher the probability, the more often the picks succeeded
The markets included in these ranges do not all have the same difficulty. This is therefore not a perfectly homogeneous scientific calibration curve.
The signal is nevertheless clear: picks assigned the highest probabilities were generally the most reliable.
The final score appeared in the top five in more than half of matches
For 101 matches, the score observed after 90 minutes appeared among the five most likely scenarios in 53 cases, or 52.5%.
This does not mean the model predicted the exact score in half of all matches. Offering five scores naturally covers a wider share of possible outcomes than selecting one exact score.
Clean sheets, wins by 2+ goals and win to nil
Clean sheets: an excellent overall estimate
The average predicted probability was 26.97%, compared with 27.23% actually observed.
Wins by two or more goals: another well-estimated frequency
The model predicted 20.65%, compared with 20.79% observed.
Win to nil: a point of caution
The average predicted probability was 20.59%, compared with 18.32% actually observed. The strongest convictions on this market performed particularly poorly.
The model often estimated the overall frequency of an event better than the specific matches in which that event would occur.
The overall goal volume was right, but not always in the right match
The model predicted approximately 2.95 goals per match. The actual average after 90 minutes was approximately 2.88 goals.
Yet the 2.5-goal market reached only 55.8%. A model can therefore correctly predict the average number of goals across 104 matches while distributing those goals poorly between individual games.
Foresportia estimated the tournament's overall attacking level better than the exact scenario of each match.
What these results do not prove
Accuracy alone is not enough. We would need the odds available at the exact time of each prediction, predefined staking rules and rigorous treatment of voids.
This article measures Foresportia's internal consistency and performance, without a direct comparison against a market consensus frozen at the same time.
The World Cup represents 104 matches in a very specific context. These results must be compared with other tournaments and longer periods.
Spain vs Cape Verde is a reminder that probability is never certainty.
What Foresportia can improve after this tournament
- Better identify draws involving favorites.
- Review high win-to-nil probabilities.
- Improve match-by-match allocation of expected goal volume.
- Keep coverage markets as a core strength.
- Continue publishing mistakes and limitations.
So, did our AI predict the 2026 World Cup?
At tournament-favorite level: yes, remarkably well
Spain and Argentina were the top two teams in the pre-tournament ranking. They reached the final, which Spain won. Three of the top five favorites finished in the final four.
At match level: a solid but imperfect record
Foresportia correctly predicted 67 of 104 1X2 outcomes, or 64.4%. Performance was considerably stronger in double chance and Draw No Bet.
At detailed-scenario level: mixed results
The final score appeared in the top five for more than half of the usable matches. Aggregate clean-sheet and two-goal-margin frequencies were remarkably close to reality.
But the 2.5-goal market remained weak, some favorites were backed too strongly, and four of the five picks on the final were wrong.
The final assessment is therefore neither that of an infallible AI nor that of a simple lucky guess. It confirms the model's strengths without hiding where it still needs to improve.