The whole thing is dumb, especially for the Senate races compared to Presidential, and even then a lot of the same problems exist. Only a few states actually matter, and trends tell you like 95% of the information.
You don't even really need a sophisticated model, RCP's average is just a straight average of polls, including shitty ones, that get dropped after so many months or if a new poll comes out from the same pollster and it really doesn't do too much worse than these models.
RCP says 47-45 GOP with 8 toss ups, 50-49-1 if forced to pick, so 50-50 if Orman is with Dems.
Wang says 51-49 Dem.
Silver's most likely is 52-48 GOP. But he says there's a 60% chance that Republicans will have between 53 and 49 seats. An 80% chance between 54 and 48.
So we have basically two or three seats that we're talking about where every single model says "it could go either way" and another three or four that are leans but could change.
The most anyone is going to be wrong is like probably two seats. We're talking about an extremely narrow set of possibilities at stake.
It's like wow! Amazing! So you would be a bit less wrong!
If your model says one team is going to win 95 out of a 100 times and they lose, that's not an indictment of your model anymore than it saying 53 times and they lose. And it doesn't mean there's anything to necessarily correct for. Especially if you
can't with the available data.
Reminds me of this much funnier kerfluffle in some ways:
http://www.vox.com/2014/9/2/6088485/how-political-science-conquered-washington
http://www.salon.com/2014/09/14/all...ournalism_and_the_phony_washington_consensus/
http://www.vox.com/2014/9/17/6291859/area-pundit-angry-at-political-science-for-proving-him-wrong