The Jobs Model of presidential election forecasting predicted well in 2004. The model, based on data available in August 2004, generated an error of only 1.3 percentage points when forecasting the incumbent share of the two-party popular vote (Lewis-Beck and Tien 2004). In contrast, the median forecast from seven teams of statistical modelers was off 2.6 percentage points (Campbell 2004, 734). We believe that the Jobs Model was more accurate because it broadened measurement of economic performance, a conceptual variable lying at the core of most of these efforts. Take, as a representative example, the Growth Model in Table 1, Column 1. Its forecast for George W. Bush was 54.0% (almost exactly at the median for the above-mentioned group of forecasters). This model was earlier reported by us, but rejected on grounds of specification error (Lewis-Beck and Tien 2004, 754). We argued that the changing nature of the American economy required attention to a hitherto neglected variable—job creation. When this variable, new jobs over the presidential term, is added to the Growth Model, the fit statistics improve dramatically (see Table 1, Column 2,).

The Jobs Model correctly forecast an extremely close race (49.9% for Bush to 50.1% for Senator John Kerry).¹

This forecast was preliminary, based on an official but early (August) estimate of GNP growth for the second quarter (yielding an estimate of 1.32% growth, non-annualized, across the first two quarters of 2004). With the more accurate GNP growth number (1.47% growth for these two quarters), available in September, the final model forecast was 50.2% for Bush v. 49.8% for Kerry. In other words, the prediction was a Bush victory, the error only one percentage point (i.e., actual 51.2%— predicted 50.2%).

Our final point estimate compares quite favorably to other estimates, regardless of the forecasting approach. With respect to the forecasts from the above seven teams, all of whom employ an aggregate statistical model, it bests all except the first Wlezien and Erikson forecast of 51.7% (Campbell 2004, 734). Besides these macro-models, there are micro-models (based on vote intention or vote expectation). With respect to the polls, which examine vote intention, the Final Gallup Pre-Election Survey results were allocated 49% Bush, 49% Kerry, 1% Ralph Nader (Gallup Poll, November 1, 2004). In other words, they placed Bush and Kerry in a dead heat for the two-party vote, at 50-50 each, giving Gallup a final error of 1.2 percentage points.

Another micro-approach comes from the Iowa Electronic Market (IEM), which reflects candidate forecasts daily through stock trader expectations. This year, the IEM gave Bush 50.5% on November 1, the day before the election, for an error of only 0.7 of a percentage point. A final micro-model, also based on expectations of voter behavior, is “citizen forecasting” (Lewis-Beck and Skalaban 1989). It has been demonstrated that survey respondent guesses about who will win the presidential election can be highly predictive of the outcome (Lewis-Beck and Tien 1999). According to data from the Pew Research Center's Final Pre-Election Survey (October 27–30), 75% of those with an opinion declared that Bush would win. That translates into a prediction of 52.9% for the incumbent Bush (Lewis-Beck and Tien 1999, 181). Hence, this citizen forecast also does well, with an error of only 1.7 percentage points.

Reviewing these rival approaches, certain conclusions are possible. First, in general, forecasters, regardless of method—equations, polls, markets, citizens—did well. That is, they almost all saw it would be a close race, with the win going to Bush. Second, and more specifically, our statistical model bested all other macro-models except that of Wlezien and Erikson, and all micro-models except that of the IEM. The IEM gets the edge, with an error of 0.7 of a percentage point, as opposed to our error of one percentage point. However, that slight gain in precision from the IEM comes at the cost of considerable lead time. IEM's 50.5% forecast was available November 1, the day before the election, whereas our 50.2% forecast was available September 29. In forecasting, the advantage a model has in lead time should not be disregarded, for it is here that the real value of forecasting lies (Lewis-Beck and Rice 1992, 96).

What of the future? The Jobs Model, updated with the 2004 observations, appears re-estimated in Table 1, Column 3. Including these current data strengthens the statistics (over Column 2): the standard error of estimate (SEE) is lower, the t-ratio for the Jobs variable is higher, and the Durbin-Watson statistic shows a reduced auto-correlation threat. We are encouraged that this specification is essentially sound, at least as far as it goes. What does it suggest about Republican chances in the 2008 contest? When the incumbent party is running the president for reelection, it gains, on average, over two percentage points in the popular vote. The Republicans will lose this incumbent party advantage, since George W. Bush cannot compete again. Thus, to bolster party prospects, President Bush could deliver a better job performance at the end of his second administration, compared to the end of his first (measured by presidential popularity at 47). The implication is that perceptions of the war on terror and in Iraq must improve. To further compensate, President Bush could boost the economy, in particular creating many more new jobs this time around. These are tall orders. The Republicans appear to have an uphill battle for the presidency in 2008.