My sanity check alarms started going off here: > For the record, the Primary Model, with slight modifications, has correctly predicted the winner of the popula

My sanity check alarms started going off here: > For the record, the Primary Model, with slight modifications, has correctly predicted the winner of the popular vote in all five presidential elections since it was introduced in 1996 ... Also note that for all elections from 1912 to 2012 the Primary Model picks the winner, albeit retroactively, every time except in 1960. It can't be true that this model correctly predicts both the election winner and the popular vote winner, since Bush won one and lost the other in 2000. I am extremely skeptical of this model, partly because I find the idea that voters pay attention to the primaries, but not to anything that happens later, very implausible. But much more importantly, it has a huge overfitting problem. There are tons of degrees of freedom, not just in the model weights but in the choice of which primaries to consider. The authors picked NH in previous elections, and this year fairly arbitrarily decided to include SC as well. It's no surprise that you can make good retroactive predictions when you have many effective model parameters to play with.