Google and the election

This paper argues that Google searches prior to an election can be used to predict turnout in different parts of the United States. Change in October search volume for ”vote/voting” over a four year period explains 20-40 percent of state-level change in turnout rates. The predictive power is little affected by changes in registration rates or early votes over the same period. This information might prove useful in predicting candidate performance beyond what is contained in polls. Because high turnout is usually good for Democratic candidates, unusually high (low) voting search volume might predict better (worse) Democratic performance than state polls suggest. Further, the information might be used to predict what demographics are likely to turn out above or below their usual levels. In 2008, for example, the Google data would have successfully predicted that African-Americans would turn out at greatly elevated levels. Finally, comparing search volumes in Democratic compared to Republican cities and towns within a state may give information about a state’s expected party-specific turnout. I make out-of-sample state and demographics turnout predictions for the 2012 election.

That is from Seth Stephens-Davidowitz's paper.