Polly the parrot is back for forecasting the 2012 U.S. Presidential Election

Andreas Graefe writes:

After her great success in 2004 and 2008, Polly the parrot is back for forecasting the 2012 U.S. Presidential Election. Polly’s daily updated forecast, widely known as the PollyVote, is based on evidence-based principles in forecasting.

Combining forecasts

Polly’s favorite rule for improving forecast accuracy is the principle of combining. According to this rule, one can achieve large reductions in forecast errors by combining forecasts from different methods that draw upon different information. In following this simple advice, Polly combines forecasts within and across currently four groups of methods: (1) polls, (2) the Iowa Electronic Markets prediction market, (3) econometric models and (4) index models. Expert judgment, the fifth component, will be added in the fall of 2011.

For the past two U.S. Presidential Elections, this combining procedure has led to highly accurate forecasts of the final election outcome. The Election Eve PollyVote forecast missed the final vote share on average by 0.5 percentage points. In addition, Polly never strayed from the candidate that eventually won the election (Graefe et al., 2011).

Index models as decision aids

Since the 2008 election, further advances have been made to improve forecast accuracy, in particular the development of forecast models based on the index method.

Bio-index: The bio-index model predicts the outcome of U.S. Presidential Elections based on information about candidates’ biographies. This model correctly predicted the winner in 27 of the last 29 U.S. presidential elections from 1896 to 2008. Thus far, Polly coded 14 potential Republican nominees. According to the bio-index model, none of these candidates would beat Obama in 2012, except for one: Rick Perry. The bio-index can be used to advise candidates on whether they should run for office or parties on who they should nominate (Armstrong & Graefe, 2011).
Issue-index: The issue index model predicts the outcome of U.S. Presidential Elections based on information about how voters expect the candidates to handle the issues facing the country. This model correctly predicted the winner in 9 of the last 10 elections from 1972 to 2008. The issue-index can be used to advise candidates on which issues to emphasize in a campaign (Graefe & Armstrong, 2011).

Big-issue model: The big-issue model predicts the outcome of U.S. Presidential Elections based on information about how voters expect the candidates to handle the most important issue facing the country. This model provided accurate forecasts for the outcome of the last 10 elections from 1972 to 2008. The issue-index can provide fast advice on which issue to emphasize in a campaign (Graefe & Armstrong, 2010).

Forecasting the 2012 election

Although still more than a year to go until Election Day, Polly has already found a number of models that provide early forecasts of the election outcome and will continuously update her forecast on a daily basis. In addition, for the first time, Polly also provides forecasts of who will win the Republican nomination.

Check out the latest forecast at www.pollyvote.com.

References

Armstrong, J. S. & A. Graefe (2011). Predicting elections from biographical information about candidates, Journal of Business Research, 64, 699-706.

Graefe, A. & Armstrong, J. S. (2010). Predicting elections from the most important issue: A test of the take-the-best heuristic, Journal of Behavioral Decision Making, doi: 10.1002/bdm.710.

Graefe, A. & Armstrong, J. S. (2010). Forecasting elections from voters’ perceptions of candidates’ ability to handle issues, Presented at the Bucharest Dialogues on Expert Knowledge, Prediction, Forecasting: A Social Science Perspective, Bucharest 2010.

Graefe, A., Armstrong, J. S., Jones, R. J. Jr., Cuzan, A. G. (2011). Combining forecasts for predicting U.S. Presidential Election Outcomes, Presented at Bucharest Dialogues on Expert Knowledge, Prediction, Forecasting: A Social Science Perspective, November 2010.

I’d like to add a couple comments of my own.  First, I haven’t read the reference on the bio-index model, but based on the work of Rosenstone and others, I’m skeptical that candidate details matter much.

Second, as I’ve written again and again and again (the first time was here), I hate hate hate hate hate evaluations based on how many elections were predicted correctly (above it says something about correctly predicting the winner in 27 of the last 29 and 9 of the last 10 elections). 1960, 1968, and 2000 were ties, and 1976 was pretty close to a tie also. It is meaningless to give a model credit for correctly predicting a coin flip.

Finally, when predicting national elections, you’ll want to separately predict the national trend and the relative positions of the states (see this paper with Kari Lock).