A few thoughts on FiveThirtyEight:
1) Ignoring Polls-Plus: the link on the front-page goes to polls-only, so I am going to cite it as the main forecast. The polls-plus model did worse than the polls-only model in the primary, so I guess they have given up on it. And, it is not OK to pick and choose from multiple models each day. I get the difference between the three models (including now-cast), but there needs to be topline model they believe in. I assume that to be the one they link to. Thus, I will ignore polls-plus model.
2) Polls-Only is Really Conservative: today FiveThirtyEight has Pennsylvania as 56% for Clinton. She is up by 4 pp on Pollster and 4.6 pp nationally. Not to mention the Democrats have won PA every year since 1992. I am not sure how they justify this probability. They must be assuming some really insane variation.
3) Polls-Only is Not Believable: I believe their entire process is under-identified. They simply have too many variables for too few outcomes. They include ridiculous polls and have some crazy scheme for weighting polling companies. I used to use FiveThirtyEight as solid polls-only forecasts, but at this point, I think they are just doing to much to be believable.
I will continue to use Pollster’s trend as a much better approximation of the current state of the election (for polling-only, obviously we are huge fans of including prediction market data!) in any given state. And, use my very simple aggregation of the Pollster average into a probability. I am confident that the Pollster trend combined with a simple standard deviation derived empirically will create a more accurate polls-only forecast than all of their work. The method is here and the output is here.