Take the simple mean of virtually all polls where the majority of interviews were conducted over the last 15 days.
When available, use "likely voter" poll models.
Do not include Zogby Interactive polls due to their horrendous past performance.
Do not include Strategic Vision polls, as it is starting to seem likely those are not real polls.
Include campaign-funded and partisan polls. Further, if there is more than one poll from a single organization, include all of them.
In the interests of a smoother average, polling outfits that release new generic ballot polls every week, such as Rasmussen, The Economist, and Daily Kos, will always have the same number of polls in the average (two).
The basic idea is to cram as many polls with sound methodologies into the averages as possible, and weight them evenly to include more overall data in the sample. Because voter preferences don't really change that much in high-profile elections, I thought this method might produce a more accurate result through logic of regression to the mean. It seems to work pretty well, as my research has shown so far.
This is different from my 2006 and 2008 methodology in that it includes polls from 15 days out from an election, instead of only 8. Further, campaign funded polls, and multiple polls from a single polling firm, are now included. All of these changes were made to include more polls in the averages, since my previous methodology was about 10-20% less accurate than Pollster.com and fivethrityeight.com. Since they had already raised the bar so high, and since they will probably improve their methodologies for 2010 even more, it was time for Open Left to step it up.
