Here, we address the issue of forecasting from statistical models, and how they might be improved. Our real-world example is the forecasting of US presidential elections. First, we ask whether a model should be changed. To illustrate problems and opportunities, we examine the forecasting history of different models, in particular our own, which has tried to foresee presidential selection since 1984. We apply what we learn to the question of whether our Jobs model, which offered an accurate ex ante point estimate for 2004, should be changed for 2008. We conclude there is room for judicious, theory-driven adjustment, but also raise a caution about inadvertent curve-fitting. Some evidence is offered that simple core models, based on strong theory, may perform almost as well as more stretched models.