Four matches played. Four results in. The World Cup model has its first receipts.
The scorecard
| Match | Model (H/D/A) | Result | Outcome probability |
|---|---|---|---|
| Mexico vs South Africa | 59.5 / 27.2 / 13.4 | Mexico 2-0 | 59.5% (top pick) |
| South Korea vs Czechia | 40.1 / 26.5 / 33.4 | South Korea 2-1 | 40.1% (top pick) |
| Canada vs Bosnia | 57.9 / 24.5 / 17.6 | 1-1 draw | 24.5% (second pick) |
| USA vs Paraguay | 32.2 / 29.4 / 38.3 | USA 2-1 | 32.2% (third pick) |
Two favourites won. One draw. One upset. That is roughly what you'd expect from a well-calibrated model: the most likely outcome happens most of the time, but not every time.
USA 2-1 Paraguay: the model's least likely outcome
The model rated the USA at 32.2% to win, making them underdogs in their own opening match. Paraguay were slight favourites at 38.3%, with the draw at 29.4%.
Gio Reyna settled things early: a header from a corner in the 3rd minute. Paraguay equalized six minutes later through Alex Arce, also a header. But Folarin Balogun restored the lead in the 70th minute from a Reyna assist, and the USA held on.
The model's Elo layer was even more bearish on the USA: just 23.4%. The xG and squad components pulled that up to 32.2%. Neither layer had the USA as favourites.
Prediction markets were closer to the model here. Most consensus estimates had Paraguay as slight favourites too, though the gap was smaller than the model's. Both got the direction wrong. The USA won despite being nobody's top pick.
This is not a model failure. A 32.2% probability means the USA win roughly one in three times. It happened. The question is whether the model's 32.2% was more accurate than the consensus 35-40%, and that answer takes more than one match.
Canada 1-1 Bosnia: the draw that prediction markets underpriced
This is the result the draws thesis predicted. The model had the draw at 24.5%. Prediction markets had it closer to 18%. The gap was structural, not match-specific: markets compress draws when there is a clear favourite, because traders think in binaries and draws have no narrative.
Bosnia's Lukic headed in from a corner in the 20th minute. Canada's Larin equalized in the 78th, turning and finishing from close range. The late equalizer turned what looked like a minor upset into the draw.
From the video analysis: Bosnia's set-piece delivery was consistently dangerous. Three of their five best chances came from dead balls. Canada controlled possession but struggled to create in open play until the final 20 minutes when Bosnia's defensive shape started to fray.
The model's draw probability, higher than the consensus, was closer to what happened. One match does not validate a thesis, but the pattern is the one we described before kickoff.
Mexico 2-0 South Africa: the expected result, with drama
The model's top pick (59.5%) landed. Quinones scored early (8'), Jimenez added a header (66'), and the scoreline reflected the gap the model saw.
But the match had more edge than 2-0 suggests. South Africa had two red cards (Sithole in the 49th minute for denying a goal-scoring opportunity, Zwane in the 83rd) and Mexico picked up one of their own (Montes, 91st). Five yellow cards total.
From the video analysis: South Africa goalkeeper Williams made three saves in the first half, two of them from close range. Mexico's quality told eventually, but the first 45 minutes were more competitive than the scoreline reveals.
The draws thesis called this match's draw probability at 27.2%, and prediction markets had it around 20%. Mexico won, so neither the model nor the market's draw estimate was tested. But the match's competitive first half suggests the model's reading of a closer contest was directionally right.
South Korea 2-1 Czechia: the comeback
The model's top pick (40.1% South Korea) landed, though via a comeback. Krejci headed Czechia in front from a set piece. South Korea's Inbeom equalized, and Oh Hyeongyu's second-half strike completed the turnaround.
The per-model breakdown is interesting. The Dixon-Coles component actually favoured Czechia (38.3% vs 32.2% for South Korea). The Elo and historical-performance layers tipped the ensemble toward South Korea. When the components disagree, the ensemble usually lands somewhere reasonable, and this time the Elo-influenced answer was the right one.
Early calibration
Four matches is too few to calculate a meaningful Brier score. But the directional pattern is worth noting:
The model's top-rated outcome happened twice (Mexico, South Korea). The second-rated outcome happened once (Canada draw). The third-rated outcome happened once (USA win).
That distribution, 2/1/1 from four matches, is roughly consistent with a model that assigns ~50% to its top pick, ~25% to its second, and ~25% to its third. Not proof of calibration. But not a red flag either.
The prediction market gap, early evidence
Before the tournament, we published a draws thesis: prediction markets systematically underrate draw probabilities, especially in matches with a clear favourite.
From the opening weekend:
- Mexico vs South Africa: model 27.2% draw, consensus ~20%. Result: not a draw, but the competitive first half supported the model's narrower gap.
- Canada vs Bosnia: model 24.5% draw, consensus ~18%. Result: 1-1 draw. The compressed outcome landed.
- USA vs Paraguay: model 29.4% draw, consensus ~27%. Gap was smaller here because there was no clear favourite. Result: not a draw.
- South Korea vs Czechia: model 26.5% draw, consensus ~24%. Another small gap for a balanced match. Result: not a draw.
One draw from four matches (25%) is right around the model's average draw probability across these fixtures (26.9%). The consensus average was lower (~22%). One match proves nothing, but the model's draw calibration is not embarrassing itself on day one.
What to watch next
The model's biggest remaining contrarian positions from the five calls post are still untested. Ecuador above Germany, Iran's 81% advancement probability, and Raphinha as the top anytime scorer all need more matches before they face real scrutiny.
For matchday 2, watch the draw probabilities in matches with heavy favourites. Germany vs Curacao (model draw: 13.1%) and Qatar vs Switzerland (model draw: 18.4%) both have large favourite-underdog gaps. The prediction market draw compression should be most visible there.
All probabilities are frozen pre-match model outputs from the ensemble behind every fixture page. Video analysis observations are from post-match highlight review. The model publishes probabilities, not recommendations. Methodology: /docs/methodology/. Full Terms of Use.