With the simple procedure of averaging forecasts within and across four component methods, Polly provided highly accurate forecasts for the past five U.S. Presidential Elections from 1992 to 2008. For example, compared to a randomly selected poll that was conducted around the same time that Polly released her forecast, the PollyVote reduced the forecast error on average by 58%. In addition, the PollyVote was more accurate than each of the other component methods alone.

Throughout the last two elections in 2004 and 2008, Polly regularly posted her forecasts on this website more than half a year before Election Day. On each single day, Polly was correct with her prediction of who will win the election. With her final forecast, released on Election Eve, Polly missed the candidates’ actual two-party vote-shares by 0.3 percentage points in 2004 and 0.7 percentage points in 2008 – an average error of only 0.5 percentage points.

How does Polly know?

Polly takes advantage of one of the best procedures to improve accuracy: combining forecasts. A meta-analysis of 30 studies found that combining reduced the error of the typical forecast on average by 12% (Armstrong 2001). For the past five elections, the PollyVote has shown that the gains from combining can be even higher, if the forecasting situation meets the conditions under which combining is most beneficial (Graefe et al. 2010):

1. a number of evidence-based forecasts are available,
2. the forecasts draw upon different methods and different data, and
3. there is uncertainty about which individual forecast is most accurate.

References

Armstrong, J.S. (2001). Combining forecasts, In: J. S. Armstrong (Ed.), Principles of Forecasting: A Handbook for Researchers and Practitioners, Norwell: Kluwer, pp.417-439.

Graefe, A., Armstrong, J. S., Jones, R. & Cuzán, A. G. (2010). Combining forecasts for predicting U.S. Presidential Election Outcomes, Presented at Bucharest Dialogues on Expert Knowledge, Prediction, Forecasting: A Social Sciences Perspective, November 2010.