If you read on the methodology of some of these 'election models', you'll understand there's a lot of narrative chasing that goes on (or even just "herding towards the least controversial number").

For example, from Nate Silver's blog:

> The Silver Bulletin polling averages are a little fancy. They adjust for whether polls are conducted among registered or likely voters and house effects. They weight more reliable polls more heavily. And they use national polls to make inferences about state polls and vice versa. It requires a few extra CPU cycles — but the reward is a more stable average that doesn’t get psyched out by outliers.

All this weighting and massaging and inferencing results in results that are basically wrong.

Come Election Night he basically threw the whole thing in the trash too!