Thursday, October 9, 2014

The Art and Science of Forecasting an Election

Since Nate Silver began his prediction site FiveThirtyEight over six years ago, there have been multiple political/news websites that have included an election forecasting model, including this very site.  For example, in 2012, we correctly predicted the winner in 49 of the 50 states with an average margin of error of only 1.8%.  

This year is the first that we've had a running Senate forecast, which began in mid September.  As I have explained before, our model takes into account polling, state fundamentals, and other news events in order to calculate a prediction for a candidates chances in winning.  

Other websites use similar formulas, but by simple modifications and tweaks, the different models can give drastically different results.  For example, the Washington Post Election Lab has GOP chances of winning the Senate at 94%, while we currently have the GOP's odds at about 55%.  Why the huge disparity?

This is due to how much weight forecasters give to either polls or state fundamentals. Additionally, a site like WaPo likely assumes that a candidate with a two point deficit in the polls will have an approximate 20% chance of winning, while our model will assume an approximate 40% chance for the same polling deficit.  

Therefore, there are fewer scenarios in which seats can be in play.  While most forecasters have sites like Colorado as a true tossup, the Washington Post rates it as "likely Republican".  

Below, we have a chart showing how the most competitive Senate races are forecasted by Battleground270 as well as 538, the NYT, WaPo, and Huffington Post.  

As you can see, Battleground270 correlates most closely to the results from 538, with an average difference of only 4.5 points.  The largest differences between BG270's and 538's predictions are in Michigan (+7) and North Carolina (-14).  There's not really a great explanation to those larger gaps, but they will obviously change (and possibly converge) as the election nears.

What we do know is that everyone cannot be correct.  On election day, these four sites as well as Battleground270 will have published a final Senate forecast with predicted vote shares as well as likelihood of each party to take control of the Senate.  On November 5th, we'll know who had the better formula.

No comments:

Post a Comment