Skip links

Can Twitter Predict Elections?

In the aftermath of the 2016 election, one couldn’t be blamed for looking beyond polls for indicators about who might be favored in difficult-to-predict elections. That’s not necessarily to supplant polling, which we still see as the best way to measure an electorate and its intentions, but rather to augment it. Additional sources may be particularly helpful when the polls are contradictory or when the likely electorate is hard to model, as may be the case in the Alabama special election between former state Supreme Court Chief Justice Roy Moore (R) and former U.S. Attorney Doug Jones (D).

It’s in that spirit that we wanted to introduce Crystal Ball readers to an interesting new entrant into the field of political prognostication: Cognovi Labs.

Cognovi made an against the grain, early prediction in the Alabama Senate race way back on Oct. 22, when most expected that the controversial Roy Moore was still a big favorite to win (and weeks before the sexual assault allegations that rocked his campaign):

We’ll get into why Cognovi made that prediction, and how the firm sees the race now, but first a little background:

The Ohio-based company uses artificial intelligence and behavior modeling to try to predict the future by measuring the emotional decomposition and intensity on both sides of an election through Twitter (they also try to predict corporate performances, marketing campaigns, stock price performance, and other non-political outcomes). The technology behind Cognovi Labs was developed at Wright State University by Dr. Amit Sheth, executive director of Ohio Center of Excellence in Knowledge Enable Computing (Wright State is located in Dayton, OH).

The firm’s technology can analyze information from Twitter, Facebook, Instagram, Reddit, and other sources, but prefers to use Twitter for their political projections because it can be a more accurate gauge of people’s actual emotions than, say, Facebook. A difference between Facebook posts and Twitter posts is that Facebook users often try to show themselves in the best possible light (be it posting pictures of a wedding or a vacation) while Twitter users, who often have anonymous accounts, traffic in a broader range of emotions, including sadness and rage (trust us, as regular Twitter users, we are well aware of the vitriol that emerges on that platform, although Facebook comment streams can get plenty nasty, too). So even though Twitter has a significantly smaller base of users than Facebook does, Cognovi believes the emotional undertone and passion they pick up from Twitter can give them a sense of where an election is headed because the Twitter conversation is more uninhibited and thus closer to reality than the conversation on, say, Facebook.

What Cognovi does is more complicated than just measuring, for instance, the share of Twitter chatter about one side or the other. Rather, Cognovi analyzes the actual content of individual tweets and picks up on the emotional intensity and behavioral patterns in both support of and opposition to candidates. This makes sense intuitively. Elections can be all about the motivation driving turnout, and if one can pick up on a greater surge in emotional intensity from one side, that could be an indication of an advantage.

Cognovi tracks several primary emotions and then constructs behavioral signals to capture people’s intent and motivation. One of the most debilitating emotions for a candidate’s emotional brand is sadness, as it often leads to disillusionment. That also makes sense: Anger can be a great motivator for voters, depending on where it is directed, whereas sadness can suggest apathy.

Cognovi can discern the location of the tweets, important when trying to project a state-level result. The firm also tries to ignore bots, of which there are many on Twitter. Theoretically, one way to game a prediction product like Cognovi would seem to be a well-orchestrated bot army, but Cognovi says it can filter out bots in multiple ways. In addition, it is able to correct for the difference in voter distribution against the age distribution of social media users, separate out the voting patterns of men vs. women, and isolate behavioral signals by a person’s political leaning.

While Cognovi is relatively new to the election forecasting game, some of their early results have been promising. They did well in the U.S. presidential election last year and picked up the United Kingdom’s Brexit result before the results were reported.

We met with Cognovi for the first time about a week before the recent Virginia gubernatorial election, when Cognovi’s COO Jim Levites and CEO Beni Gradwohl told us that they saw Ralph Northam (D) as a bigger favorite over Ed Gillespie (R) than the then-tightening polls suggested. However, as Election Day neared, Cognovi also hedged in subsequent reports prior to the election, when its metrics, like the polls, also picked up a tightening race — so they would have been closer to the mark had they stopped updating after their report to us six days before the election. Still, Cognovi identified the high Democratic enthusiasm in Virginia that turned what seemed like a close race into a nearly double-digit victory for Northam.

That brings us to Alabama, where Cognovi favors Jones. In October, when they made their early call in Alabama, they noticed that Jones had a significantly greater share of the emotionally charged Twitter conversation in Alabama, indicating to Cognovi that he was favored in the election.

More recently, Cognovi observed that the horrible sexual assault allegations against Roy Moore may be driving a greater level of emotional intensity among his supporters, and thus may have actually improved Moore’s chances: “While the sexual harassment claims have tarnished Moore’s reputation nationally, Cognovi AI indicates that in Alabama they have energized his base,” Cognovi noted in a recent short update (which you can read in full here for more information about the Alabama race and Cognovi in general). That seems counterintuitive, although when one considers how much antipathy many Republicans have for the press, it makes some sense. Cognovi also finds that while Moore has a “toxic emotional brand,” Jones “has not been able to further fortify his position and improve his brand” in the wake of the allegations against Moore.

Anyway, our intent here is not necessarily to endorse Cognovi’s methodology or to assert that they have cracked the election prediction code: Only time will tell how effectively their methods work over the long haul, and they face a test in Alabama, where they stuck their necks out for Jones fairly early in the general election period. Our own rating remains a Toss-up in Alabama. But we were sufficiently intrigued to tell you about them, and we’ll be tracking how Cognovi does on Tuesday and beyond as we (and others) look for whatever shreds of information we can find in advance of the elections we cover.