My guess is that you're misinterpreting Silver's work and assuming he was "predicting" that Brazil "would" win. He wasn't, and didn't claim to. He said Brazil was more likely to win, and gave numbers to that effect (that in this case happened to be wildly wrong).
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.
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.