We Asked Maia to Predict the World Cup. Here's How It Did

Back in June, we published Maia's pre-tournament predictions for the 2026 World Cup. The final has now been played. Time to see how the model held up.

Before the tournament, Maia gave Spain a 6.87% chance of winning the World Cup. Highest of all 48 teams, more than double the next side. On June 15 I wrote that up and put it out there.

Spain won the World Cup. They beat Argentina 1-0 in the final.

The model called the champion from the very first run, cold, before a ball was kicked. And once it had a bit of real tournament data to work with, it went further than I expected: the exact final, the exact score, and Spain's correct opponent in every single knockout round on the way to the trophy.

Here's how it did, round by round, and what I changed along the way.

The Champion the Model Came Back To

I built the model in a few versions over the weeks around kickoff, refining it as I went. Base version first, pre-tournament, with nothing but historical form to go on. Then, once Round 1 matches had finished, including Cape Verde's draw, I reran it on that early data. Then another version once Round 2 was done.

Spain topped the very first run. But the model didn't hold that line without wobbling. After Round 1, on that smaller slice of fresh data, it swung away from Spain entirely and had France beating England 2-0 in the final. It took the Round 2 rerun, and a fix to how the model was scoring games, to pull it back to Spain. That version had Spain beating Argentina 1-0, the exact scoreline that played out.

So the honest version of the story isn't that the model never doubted itself. It's that the version built on more complete data, and a corrected model, got it right, while the version built on a partial slice and an uncorrected flaw didn't. That's the pattern worth trusting, not blind consistency, but a model that gets sharper as it gets more to work with and as its mistakes get fixed.

A Really Good First Prediction

The version I published in June had Spain beating England 1-0 in the final.

As an England supporter, losing a final was never going to be the outcome I was hoping for. But reaching one? I'd have taken that gladly, and on the numbers alone it was a genuinely strong prediction. England had a brilliant tournament. They beat Norway in the quarters and pushed Argentina all the way in the semi before going out 2-1. A Spain-England final was well within reach, and in fact both of the model's early versions, June's and the Round 1 rerun, had England reaching it.

What changed was the data. After Round 2, I fed the actual results back in, fixed the scoreline issue, and reran everything. That version, with real matches to learn from and the fix in place, swapped the finalist: Spain 1-0 Argentina. Same champion, same margin, and this time the opponent that reality delivered.

That's the model working as intended. It made a smart call on limited information, wobbled when it got a little more, then sharpened once it had enough. All three predictions taught me something. The third one just had the most to go on.

The Knockout Run

Once Round 2 was in and the final version was locked, the knockout predictions were sharp in a way that still surprises me.

From the quarter-finals onward, it got every winner right. All of them. Spain past Belgium, France past Morocco, England past Norway, Argentina past Switzerland. Then Spain over France, Argentina over England. Then the final.

It also mapped Spain's entire route to the trophy. Austria, Portugal, Belgium, France, Argentina, round by round, the correct opponent every time. I've built prediction models for two tournaments now and I've never seen a bracket hold together like that.

Tracking every match from the round of 32 onward, the model called winners 79% of the time, climbing to 100% across the knockouts from the quarters in. Across all 102 matches it landed 71%.

The One Thing I Had to Fix

The model's real weakness was scorelines, not a blanket bias against underdogs. It did back the occasional upset. It had Morocco beating the Netherlands, and Morocco did. But when it favoured the stronger side, which was most of the time, it sometimes predicted a beating rather than a win. After Round 2, it had England beating Panama 9-0, a scoreline tournaments almost never produce.

That's what pushed me to fix it. A team that's outmatched on paper usually raises its game when it matters, digs in, and keeps the score respectable even in defeat. Early on, the model didn't know that.

So I made a few adjustments. The big one was changing how team strength gets measured. Instead of leaning on goal average, I weighted teams that score more goals and teams that concede fewer more heavily, which gave a truer picture of how good a side actually is at both ends. That, along with some smaller tweaks, pulled those runaway scorelines back toward realism and let the model treat underdogs the way tournaments do. It's the same fix that helped bring the Spain prediction back after Round 2.

It still couldn't perfectly predict a genuine shock, a Paraguay knocking out Germany, say, or Norway beating Brazil. Nothing built on historical form ever will, because those results are the moments when form goes out the window. But after the fix, the model stopped punishing underdogs on principle and started giving them a fair shout, which is exactly what you want.

Euro 2028: Spain, and Then England

I fed Spain's post-World Cup rating into the same machinery and ran it forward to Euro 2028, hosted across the UK and Ireland. The number is almost rude.

Spain: 80.75% to win. The most dominant pre-tournament favorite the model has ever produced. Winning the World Cup pushed their rating so high that, on paper, the rest of Europe is playing for second.

England are that second, at 11.37%, ahead of France at 6.57%. As a fan, I'll take being the model's clear runner-up. Two years out, with a home tournament and the numbers pointing our way for once, I'm allowing myself some optimism.

Ask me again after the group stage, though. If this World Cup taught me anything, it's that the model gets better, and sometimes changes its mind entirely, the moment it has real results to learn from.

The Real Story Isn't Spain

The prediction was never really the point. The point is that I built, ran, and rebuilt a full tournament model, three versions of it, over a few weeks, just by talking to Maia. When Round 1 finished and the picture shifted toward France, that wasn't a lost afternoon of debugging. It was a rerun. When Round 2 finished and I wanted to see how the real results, and a model fix, would reshape the bracket again, that wasn't a fortnight of reworking pipelines either. It was another conversation and another rerun.

That's the shift. In 2021, when I built something like this for the Euros, the bottleneck was my hours. This time the bottleneck was how quickly I could think of the next question. Spain winning is a nice headline. The fact that I could rebuild the whole model three times in a matter of weeks, feed it new data the moment I had it, and watch it change its mind and then get sharper, that's the story.

Rebuilding the model three times in a few weeks took an afternoon each time, because the work of a data team became the work of a conversation. That's what Maia does with your data, not just football brackets.