Swing State Overview (270 to win, 269 to tie)
(Swing States are defined as states closer than 5.0%. Before you ask, if a state isn't listed, it isn't that close right now.)
I noticed, with interest, Nate Silver's takedown of Real Clear Politics earlier today. I have to say, Nate makes a an extremely compelling case that RCP is internally inconsistent in which polls they include in their averages, making it more difficult to take them seriously as a polling resource.
Then again, I stopped using RCP's polling averages a long time ago. They always seemed to be using arbitrary dates for their polling averages, and the information provided at Pollster.com left them behind a while ago. RCP is useful in the way they display the most recent state polls, and I will continue to use them for that purpose.
This also seems like as good a time as any to bring up my disagreements with other forecasting methodologies, including Nate's. While I am fully aware that most, if not all, other forecasters have a far better grasp of statistics than I do, in short I still believe that virtually all other methodologies are simply too complicated for their own good. My specific problems can be found in the extended entry.
Applying different weights to different polling firms based on past accuracy seems like a big mistake to me. This is because the degree to which polls were wrong in the past is not a very good indicator of how they will be wrong in the future. Polls with margins of error--which is all polls--will be consistently wrong in different ways. Apply different wieghts to different polls based on the past performance of the given polling organizations is, essentially, an attempt to predict how wrong a poll will be in the future. That strikes me as a permanent unknowable, and so I won't go down that path.
Adjusting for the "house effects" of different polling firms also feels like a big mistake. A given polling organization will rarely skew the exact same amount in favor of one party or another twice. However, attempts to adjust for House effects are assuming that a poll will skew in favor of a candidate or party exactly the same amount every time. The degree to which a poll will skew in favor of one candidate or the other strikes me as unknowable, so once again that is something I won't do.
Adding in non-polling information is something that should simply never, ever be done. Many forecasting sites, such as CNN, throw in meaningless "data" like the opinions of political "experts," past election performance, and the amount of money candidates are spending on ad buys. This information is simply irrelevant and unscientific, so I will not include it.
This also applies to demographic regressions. While I understand and sympathize with the idea behind using such numbers, they are still a deductive attempt to predict the composition of the electorate without using actual polling data. And so I won't use it.
Finally, unless no other data is available, I also don't believe in assigning any weight to polls older than a week, especially during the height of the campaign season when more voters are paying attention. Polls only function as a snapshot of public opinion during the time periods when they were taken, making older polls useless in determining the current state of the campaign.
In the end, all of this means that I just average recent polls (unless no recent data is available). It is extremely simplistic, and I bet the simplicity of it probably seems suspicious to many people. I really don't care, as I think everything I have listed above is unscientific gambling with election forecasting. Based on past election results, I am also convinced that my simplistic methodology is the most accurate. Just looking at polls from the last week of the campaign, and not adding any special sauce, provides a very accurate view of how close elections will turn out. And, when elections aren't close, who needs election forecasters anyway?
In about five weeks, I guess we will see if I am right. Still, even if another forecaster edges me out by a few tenths of a percent, if they did so using any of the methodologies I listed above, then my feeling is that they will simply have made a lucky guess / gamble, rather than revealed a deeper truth on how to predict elections.
STATE BLOGS
