So… this only took two years to get around to…
My pWP stat is just an easy way to understand and aggregate multiple polls into a single, simple statement: Candidate X is (blank) percent likely to win his or her race. I’ve written a bunch of stuff on pWP, just in case you need to catch up.
A majority of the time, I was only paying attention to the presidential race in 2012. I did not keep up with the various other races. In the presidential race, with its many polls and good data, I went 50-51 in predicting where electoral votes would go (I did not pay attention to Maine or Nebraska’s split-vote system, as there was no need, there were no swing congressional districts in either state). The only state that pWP got wrong was Florida. Which makes sense, Florida was the closest election, with .88% (i.e. less than 1%) difference. However, looking back, had I eliminated an obvious outlier poll, or if I had looked at the median pWP instead of just the mean, I could have gotten closer to the right answer. Bad polls essentially remove the basis of pWP, so finding and eliminating them is a key challenge.
In addition to the presidential race, I made predictions on election night in seven other tight races (6 Senate, 1 Congressional) and I went six for seven. The race I got wrong was the North Dakota Senate race between Berg and Heitkamp. Once again, this was a close race, within 3000 votes. It was the closest Senate race, and relatively poorly polled. There were no polls done in November, and Heitkamp had closed a large gap in the last month of the race. About the only way I could have avoided being wrong in my prediction would have been to not make a prediction at all. Trendlines, using a rolling average, would have projected a 50-50 race. Basically, if there’s little to no current polling, and previous polling shows a tight race, skip the prediction.
The end result of using pWP over the last several election cycles has shown the stat has been a better predictor of the outcome of the race than it should, based on its probabilistic premise. Which means pollsters are doing, at least lately, a better job than even they suggest when giving out their margins of error. I don’t know if this is accidental or purposeful. It could be the way I’m aggregating the polls (when available). I don’t know. But good news is good news.
The bad news is those same polls that were really accurate over the last few election cycles show Dayton and Franken cruising to easy victory. The few partisan polls available in the 7th and 8th Minnesota congressional districts show easy DFL victories as well.
Update: Spoke too early, a non-partisan poll has Mills up on Nolan by almost four standard deviations. That’s over 90% pWP, but I would put it closer to 75% because the strong support for the Green candidate will fall, and the 11% undecided number is too high.