At least, that's what the model said.
A few weeks ago, someone shared a World Cup prediction model. It looked good: past international results, player stats, team performance metrics, thousands of Monte Carlo simulations running the whole tournament over and over. There was some clever analysis here.
The model's results were clear. Its top four favourites to lift the trophy in 2026 were Japan, Argentina, Ecuador and Iran.
With the quarter-finals now underway, Japan, Ecuador and Iran are well into their post-tournament holidays, and they will have as much of an impact on the final stages as I will. That's three of the model's champions all out before the likes of Egypt, Canada and Paraguay. Iran didn't make it out of the group stage. So what happened?
To be clear, the person who built this model wasn't claiming Japan would win, or that they were the second coming of Pep Guardiola. The question was simple: should you trust the model, or your gut? They were upfront about the flaw, too - heavily weighting recent form makes teams from easier qualifying routes look like the Invincibles reborn. It's a great question, and a hard one to get right in day-to-day life.
Allez les Bleus
This may shock you given my name is Pierre, but I am French. And I'm tempted to say there's an easy fix for this model. It probably needs a touch more weighting on elite attacking talent. Maybe an adjustment for recent World Cup pedigree. You could stress the importance of a coach with World Cup-winning experience... a goalscorer with European and World Cup golden boots to his name, a Ballon d'Or winger with two recent Champions League wins, or a playmaker who led the Champions League in assists.
Then a small correction for... you see where this is going. I could tune it, scientifically of course, until it predicted France lifting the trophy and Mbappé scoring a hat-trick in every match. And in a way, is this not making the model more accurate than it currently is?
Of course, all this would do is build a mirror of my own assumptions.
Two doors, both wrong
Here's the trap, and it has two doors.
Door one: you trust the model completely. It's data, it's objective, it ran ten thousand simulations. Who are you to argue? So you predict that Iran are favourites, and that the eternal Enner Valencia will fire Ecuador to glory.
Door two: you trust your gut completely. Every time the model surprises you, you overrule it - more data, different weights, "that can't be right" - until it politely agrees with everything you already believed. Nice one. You've cloned yourself and called it a model.
Most people think avoiding door one is the whole job. Sniff-test the model, apply a bit of football knowledge, ask whether the forecast is even plausible. And that's half of it. Anyone who's watched a World Cup knows Iran aren't winning it (we all know France are).
But door two is the trickier one, because the entire reason we build models is that our gut is biased. How many years has football been coming home? Surely Harry Maguire and Jesse Lingard would prove too strong for Messi and co. Our gut has a lot of knowledge baked in, but it is not neutral. Mine wears a Zidane '98 shirt. Yours might be in a Kane one, still convinced that this time it really is coming home. Either of us could be proven right, but neither should be allowed to meddle subjectively with the weightings.
So investigate your features
How do you find the right balance? Not by picking a side. The answer isn't a third door. It's waiting until you've asked the most important question: why is the model saying this?
Because a surprising answer looks identical whether the model is broken or whether it's spotted something you're too biased to see. "Japan will win the World Cup" felt wrong. But before the start of the 2025/26 season, so did "Mo Salah will score seven Premier League goals this season". One of those was nonsense. The other is what happened. You can't always tell them apart by whether you like them.
In this case, the person who built the model had already found it. The simulations rewarded teams who'd rolled through weaker qualifying opposition, without properly adjusting for who they'd actually beaten. Beat Indonesia 6-0, look like world-beaters. That's not Japan being the best team in the world, that’s the model measuring the wrong thing.
Investigate the Salah forecast, though, and you might have found the opposite: a physical decline in your thirties is hardly unheard of, and usually comes with forewarning. Same surprise, opposite answer. The only way to know which one you're facing is to look further.
Most football fans can tell which is which without thinking, because they've watched hundreds (or even thousands) of games. And that’s the point: the instinct is really domain knowledge. You need the same thing in business.
The point
There's a famous line from the statistician George Box: all models are wrong, but some are useful. The wrongness was never the problem. The model that had Japan winning was wrong, and it's possible that so is the one in my head where France win 6-0 every week. My 5-0 prediction against Paraguay (which ended 1-0, thanks to a late penalty) didn't quite come off. The danger isn't that a model is wrong. It's forgetting to ask which way it's wrong, and whether the mistake belongs to the model or to you.
This is the trick to interpreting performance data: knowing when to trust it, and when to raise an eyebrow and dig deeper. It's also the difference between a model that informs a decision and one that just flatters your opinion.
The good news for Japan is that there's always 2030. The good news for France is that by then I'll have new reasons to justify why they'll win, and I'll believe every one of them.
Allez les Bleus. Allez Mbappé. Come on England?
n.b. This was written before France's semi-final. Will I wish this article could be taken down when they inevitably lose their next match? Maybe.
