Predicting the 2012 US Presidential Election - The Day After

In our paper prepared for APSA 2012 (Here) we used Ensemble Bayesian Model Averaging to combine 9 prominent political science forecasting model into a single prediction for the 2012 presidential election. The EBMA prediction was based on the track record of these models in the past and their 2012 forecast. 

Our model predicted 50.48% of the two party popular vote for President Obama. And while that is just about right on for the popular vote percentage, it is under by 7/10 of a percent from the current figures, which give President Obama 51.2% of the two-party popular vote (Results still changing, November 8, 11:00 am EST).

The EBMA model places a lot of weight on the prediction generated by Professor Alan Abramowitz, and we’d do a lot worse without his model in the ensemble. We are likely to have under predicted President Obama's vote share because most of the models did under predict, especially those favoring a Romney victory. Both of these factors drew our aggregate prediction below what actually occurred. But not by too much.

It is now very clear that the ensemble aggregation approach, applied to polls by Drew Linzer and Simon Jackman, among others, did very well in predicting the outcome in terms of state results and electoral votes--even if one looks at their results from last summer, ignoring election eve updates. 

While the EBMA prediction is not necessarily the "best" prediction for any single observation (election), we contend that it outperforms single predictive models over many observations. Stay tuned and in four years, we’ll try again.

Ensemble Predictions of the 2012 US Presidential Election

We use ensemble methods to combine ten various forecasts of the US election, the most recent being almost two months prior to the November 2012 election. Based on this combination, and the component models, we estimate a 0.60 probability that the popular vote for the incumbent will be greater than 50%.