[peter-row]

I think he's saying it's a souped-up poisson model, not an agent based model. It's not trying to predict individual behavior.

A poisson model is probably valid, but it should perhaps be refined a bit. Maybe it should account for a systematic drift if the whole team plays badly (or a vital player, like the goal keeper, has a bad day). But it's probably really hard to calibrate that, and it's probably not too relevant unless you want to model major upsets.

[gbhn]

Wouldn't that refinement basically be identifying the tails as quite a bit fatter than you'd thought before? But in that case, the tails are probably substantially fatter for everyone, and so the numbers 538 is quoting may not be far from the "real" ones.

My reading of this 538 World Cup analysis is that the error bars are taken as being very large, and the relative numbers are mostly interesting as a means of comparing the impacts of various players and other factors in the model. Obviously there can still be major flaws in this method, chief among them being that the method tends to atomize the team's contributions. But that's not a horrible first pass, and still can yield some interesting insights.

[peter-row]

Yes and yes. It's still probably the biggest upset ever, but not as unlikely as their model suggests.