> Applied prospectively to the in-progress 2026 World Cup from the Round of 32, the model identifies Argentina (28.0%) and Spain (21.1%) as the leading championship candidates.
Seems weird to wait to run the "prospective" simulation until the World Cup is already in progress. Although it seems that the model also needs to use "the actual bracket and group-stage performance". So it's not prospective?
Yes. I don't like phrasing this as being prospective for the World Cup as a whole. It's for the knockout stage. (Which the abstract says! But the title doesn't.)
That almost makes it worse—like they're vaguely aware that training too heavily on too small a data set makes badly trained models, but are unaware that it has a name and is an actual identified problem.
now back your claim with money and bet accordingly on betting sites to see if you uncovered some actual alpha here
> Applied prospectively to the in-progress 2026 World Cup from the Round of 32, the model identifies Argentina (28.0%) and Spain (21.1%) as the leading championship candidates.
Seems weird to wait to run the "prospective" simulation until the World Cup is already in progress. Although it seems that the model also needs to use "the actual bracket and group-stage performance". So it's not prospective?
It predicts likely winners based on the round of 32 performance (plus prior data). That's still "prospective" with respect to the finals
Yes. I don't like phrasing this as being prospective for the World Cup as a whole. It's for the knockout stage. (Which the abstract says! But the title doesn't.)
Probably the human pseudonym of Paul the Octopus (https://en.wikipedia.org/wiki/Paul_the_Octopus).
It's worth noting that there has only been 24 world cups
> the model identifies Argentina (28.0%) and Spain (21.1%) as the leading championship candidates
A good AI would calculate refereeing decisions and put Argentina at 100% unless England can pull off a miracle against FIFA today.
!remindme tomorrow
How does this paper not even mention the word "overfitting"?
The abstract does say
> limitations, principally the small number of tournaments available for validation and the risk of in-sample weight selection
But I agree this model is no more valuable than Paul the Octopus.
That almost makes it worse—like they're vaguely aware that training too heavily on too small a data set makes badly trained models, but are unaware that it has a name and is an actual identified problem.
"They've done studies, you know. 60% of the time, it works every time." - Brian Fantana, Anchorman