Forecasting elections based on the Twittersphere is a tweet in progress.
Social-analytics firm Attensity predicted former Massachusetts governor Mitt Romney would take Massachusetts, Virginia, Idaho, Ohio, Oklahoma, Tennessee and North Dakota in Republican contests on Super Tuesday, according to a Twitter analysis Attensity conducted for USA TODAY.
Attensity also projected that among the other candidates vying with Romney for the Republican presidential nomination, Newt Gingrich would win Georgia, Rick Santorum would take Vermont and Ron Paul would triumph in Alaska.
"We got it half right," says Rebecca MacDonald, Attensity's vice president of marketing. "The fact that people are talking about candidates on Twitter doesn't necessarily correlate to those people going out and voting."
In fact, voters came out to hand victories to Santorum — not Romney — in Tennessee, Oklahoma and North Dakota. Romney won Alaska and Vermont. Gingrich took Georgia, as expected by Attensity's forecast.
The results underscore that the science of applying predictive analytics to Twitter is still in its infancy, analysts say.
Twitter executives declined to comment.
Part of the problem lies in a lack of location-based data about Twitter users' tweets. Such information is "scarce" on Twitter, says Michael Wu, principal scientist of analytics for Lithium, a social-analytics firm. That's because Twitter users would have to turn on the "location" feature in their mobile devices.
A vast pool of location-based tweets would enable analytics experts to better connect tweets to where they come from across the nation. In the case of Super Tuesday, that would mean more localized information on tweets about candidates.
"We had a pretty good sampling (more than 800,000 tweets), but the geo-location data was a small percentage of that," says MacDonald.
Another challenge is sorting out multiple tweets by the same person or robo-tweets that are automatically generated by some users on the service, like spam messages.
Plus, such predictive analytics is a science that requires much human interaction with the data.
"Research needs to be done to correlate the sentiment data to the act of voting," says Wu. For now, social media is better used by candidates to mobilize their communities, rather than to predict election outcomes, he says.