Forecasting non-incumbent presidential elections: Lessons learned from the 2000 election

As the 2008 election approaches, we offer a reexamination of the 2000 election -- its place in history, political science, and presidential forecasting models. This is especially relevant since 2008, like 2000, will be an election without a president seeking reelection. How should forecasting models deal with such elections? Looking carefully at 2000 we evaluate the utility of "weighting" candidates in non-incumbent elections. Using Bayesian Model Averaging, we find that weighting helps to better predict 2000, but also produces a poorer model fit over a wider set of elections. For other non-incumbent elections, weighting only improves predictions for the 1960 election. Moreover, we find that the 2000 election is anything but ordinary; attempts by forecasters to change the specification of models to better fit the 2000 election are ultimately harmful to the forecasting exercise. We conclude that presidential forecasts are best when they ignore whether or not an incumbent is running.