Maybe Silver was never good at the polling game to begin with. He got 2008 right, and then just coasted after the upset.

Correlation of a large amount of data isn’t like some quantifiable thing. The larger the data set, the more knobs get turned, and the outcome can be almost whatever you want it to be for however you process that. The reason NS never wanted to open source his stuff is because it was largely just tweaked here and there, and didn’t offer much in NEW insights. It was nothing, and largely inaccurate, because the polling data is inaccurate, or not a representative of the actual populace going to vote.