The forecasting model I wrote about in the October 2004 issue of PS (Holbrook 2004) is best described as a referendum model. It incorporates presidential approval (Gallup polls averaged from June through August), a measure of aggregate satisfaction with personal finances weighted by the tone of economic news¹

(averaged from June through August), and a dummy variable coded “1” for years in which the incumbent party had held the White House for at least two consecutive terms and “0” for all other years. The first two variables are intended to capture the political and economic performance of the incumbent administration, while the latter (borrowed from Abramowitz 1988) is based on the idea that it may be easier to convince voters that it is “time for a change” if the incumbent party has held the White House for at least two consecutive terms.

The parameter estimates used to generate the forecast for the 2004 presidential election are found in the first column of coefficients in Table 1. Based on the values of the independent variables going into the election, the predicted outcome for the 2004 election was 56.1%²

of the two-party vote for President George W. Bush. Since Bush ended up with 51.2%, the forecast error for 2004 was 4.9 percentage points. This represents a substantial error, more than twice the size of the average absolute out-of-sample forecasting error.³

Incorporating the 2004 data into the forecasting model changes the parameter estimates (second column of coefficients in Table 1) somewhat, though not dramatically. All coefficients are still positive and statistically significant and the slope for presidential approval increases somewhat while that for the tenure variable is reduced somewhat. The overall accuracy of the model is somewhat weakened, as revealed by increases in the standard error and the average absolute out-of-sample forecasting error, as well as a decline in the adjusted R². Figure 1 displays the accuracy of the out-of sample forecasts generated by this model. Here we see that the general pattern is that the actual vote follows the predicted vote very closely. While the error in the 2004 forecast is substantial, most other outcomes are appreciably closer to the predicted values.

One way of viewing the results of the 2004 election, especially in the context of this forecasting model, is that Senator John Kerry did better than expected. In fact, given that most academic models predicted a wider margin of victory for Bush than he received, it is not unreasonable to conclude that Kerry “beat the spread.” While this may say something about systematic error in the forecasting models, it could also say a lot about the relative efficacy of the two campaigns. Given the context of the election, Bush was favored to win in more convincing fashion than he did. So perhaps instead of wringing their hands in this period of post-election introspection, Democrats should feel good about doing as well as they did. Little comfort to them, I'm sure, but still a point that should be considered.