Out or sample prediction errors

Because the model in my AER-paper uses in-sample errors, it may underestimate the uncertainty in the upcoming election results. To see how important this is, I estimated the out-of-sample errors. The average absolute prediction error in the state level vote share increases from 3 percent in-sample to 3.7 percent out-of-sample. Most of this is due to larger out-of-sample national swings.

This has large impacts on the estimated probability of winning. For example, in the estimates made October 1, using variance estimates based on the out-of sample errors, Obama’s chance of winning drops by around 10 percent compared to when using in-sample predictions. This is because large national prediction errors are more likely when measured out of sample. Obama's average vote-share and number of electoral votes remain at the same levels.

While using the out-of-sample errors does make a difference for the probability of winning, it does little to the optimal campaign strategies. Graph 3 to the right shows the equilibrium visit shares for based on variances using the distribution of the out-of-sample errors.

For the exact description of how I proceeded, see here.