This was the first national election for which projected results were based on polling data streams that ran in something close to real time. Mid-day on Election Day I took a look at what these last-minute projections were predicting, and it was fun today to go back and check out these final projections to see how close these sites came to calling the actual results of the presidential race.
The websites I consulted were 270towin, fivethirtyeight, and realclearpolitics. All three sites are powered by Monte Carlo software. Although the actual final results will not be known until the question of Missouri–predicted by all three sites to be the most “unpredictable” state–is resolved, it is still illuminating to check out the sites’ performances. As I compare the familiar red/blue maps for all three sites, it seems that fivethirtyeight’s risk analysis came closest to today’s provisional tally of electoral votes. The actual numbers in the current tally are Obama 364, McCain 163; fivethirtyeight’s prediction was 348.6 and 189.4.
I think the power of this projection is impressive. Any prediction is a decision made under uncertainty, and in this case, which involved millions of unpredictable humans, there were millions of uncertainties. Statistical analysis of the probabilities involved in these millions of chancy scenarios is a daunting task. But evidently it didn’t daunt the pollsters.
The near-accuracy of the election poll sites makes me wonder why the risk assessments done by the financial institutions packaging and selling securitized debt couldn’t have been as close to the mark as the risk analyses done by the polling organizations.