Forecasting Elections with Real-Time Economic Data

This post is jointly written with Anton Strezhnev, a very bright Georgetown undergraduate.

One of the challenges in forecasting elections is that economic data are often inaccurate when first released. Some of the adjustments are substantial.  Just to illustrate this point, the image below (source) shows the change from original issue to current estimate in a composite index of economic performance: the Chicago Fed National Activity Index (CFNAI).

The magnitude of some of these adjustments could potentially affect forecasts in what the models predict to be a close election. Moreover,  there is serial correlation in the direction of the errors. So, if you are rooting for Obama you may think that the more recent positive adjustments mean that Obama has a slightly better chance than the models predict. If you are a forecaster, the serial correlation may allow you to better predict adjusted values.

Economists have long recognized that the use of real-time versus ex post adjusted values can make important differences when analyzing data. Similarly, we could use these differences to investigate whether using the data available at the time of the election (information) or actual economic performance shapes election outcomes. For this purpose, we re-estimated Douglas Hibbs’ bread and peace model using the “vintage” per-capita disposable income data that was available around the time of each election. Unfortunately, we could only go back as far as 1964 with our data.


The model with vintage data returns a slightly better fit (R2 of .884 vs. .867) but we should not make too much of this with so little data (indeed, we shouldn’t make too much of any model with this little data but this is a blog post and broad sweeping statements will follow).  As shown below, the model with real-time data gets 1988 and 1980 wrong but it gets 2000 and 2008 right.

Plotting the errors suggests another interesting pattern. While the real-time model tends to perform more poorly in pre-1992 elections, the pattern has reversed in recent years. Remarkably, for the previous four elections, real-time rather than revised economic data tends to be a better predictor. Since we are in the business of making large generalizations based on little data, we attribute this to increased information and media consumption.

What does this model say about President Obama’s re-election prospects? The coefficients are slightly different, thus estimating the model with vintage data yields slightly different predictions. While it is still clear that the weak economy will remain a large burden, the model based on the real-time data generates a slightly more optimistic outlook for Obama. The table below shows that, compared to Douglas Hibbs’ original predictions, the re-estimated “vintage” model returns slightly higher expected vote shares – roughly a 1 to 1.5 percentage point bump across all of the growth scenarios. That’s substantial in a close election and may provide a glimmer of hope for Democrats in dark days.