Also known as, predict 95% of election results in a half hour!
Before Nate Silver became a blogging superstar at 538, he was a prominent baseball blogger, primarily known for his work on PECOTA, the most accurate baseball projection system around. PECOTA takes into account scores of factors in order to project a players’ performance next year, from his handedness, to the size of his signing bonus, to the ballparks he played in last year. This stands in marked contrast to Tom Tango’s Marcel the Monkey, which takes a players’ previous stats, adjusts them for age, and regresses them to the mean. That’s it. Now, PECOTA has been more accurate than pretty much any projection system over the past few years. But Marcel the Monkey has been pretty close. The point here is, you can get almost all of the benefit of a baseball projection system by combining only a few simple factors. Is the same thing true for political projection systems?
I attempted to answer this on the Monday before the election by creating the Dumbest Projections Ever. I took 2004’s election results, in which John Kerry lost by 2.5 points. I looked at RCP and Pollster’s national polling averages, and figured that Barack Obama was ahead by roughly 7 points. I took John Kerry’s showing in every state and adjusted it upwards by 9.5 points. Thus, the DPEs were created. The point here isn’t to be the best system, it’s to provide a minimal bar over which even the monkeys in Dick Morris’s basement should be able to leap over. The full results are available here.
So, how does this system do? Pretty well, actually. The dividing line for where McCain is predicted to beat Obama is right in between Arkansas and Virginia. In other words, the system got 47 states + DC right, missing only MO, IN, and NC, which I think is pretty good. This accuracy boils down to the accuracy of the underlying components; poll averages performed very well in 2008 (again), and the vast majority of voting attitudes can be explained by prior voting + nationwide shifts. The 2004 results correlated with DPE and 2008 at an r^2 of .945. This compares to 538’s correlation of .983, which is substantially better. Still, 48/51 outcomes and a .945 correlation coefficient is nothing to sneeze at.
The other advantage of the DPEs are that they are essentially the individual states’ results normalized for national trends. Thus, we can see where Obama over and underperformed Kerry, given the national electoral climate. We can also get an idea of which projection techniques were particularly inspired, and which may be overfitting. More on this to come.