And indeed, when it comes to my specialties like building election models, that objection still holds. Election forecasting is emphatically not a “big data” problem; just the opposite—there’s only one election every four years, so the data is exceptionally sparse. In cases like these, you need to bake a lot of structure (or if you prefer, assumptions) into a model—for instance, that the order of states from red to blue stays about the same from election to election, so Pennsylvania will probably be bluer than Wyoming but redder than Vermont.
