Kentucky Derby

As students filed into Anthony Fowler’s Science of Elections and Campaigns class the day after the November 2015 elections, everyone knew who the winner was. Not the winner of the Kentucky gubernatorial election, but rather the winner of the election prediction competition Fowler had organized. 

The week before, everyone in class had laughed off the only student who had predicted that the Republican gubernatorial candidate would win the election in Kentucky. All the polls gave Democratic candidate Jack Conway a clear edge in the contest against Republican challenger Matt Bevin and Independent Drew Curtis. But now, after Bevin’s surprising victory, building a prediction model that could get close to the actual outcome of an election seemed even more of a challenge than before.

“I wanted to demystify the process of forecasting,” says Fowler, an assistant professor at Chicago Harris whose research focuses on causal questions about political representation. “I wanted students to look under the hood and realize that they already have the skills to be effective forecasters.” 

For the contest Fowler asked his students to develop a prediction model for the elections in Kentucky and Mississippi. The winner, second-year student Andrew Angeles, came closer to the actual result than even the most seasoned election pundits did.

Angeles’s model used a combination of linear regression and polling averages. Using data from 10 states close in ideology to those being examined, he looked at the impact of states’ per capita GDP, the number of Democratic governors in the country, and incumbency status on election returns from 1990 to 2012. Overall, he found GDP and incumbency to have a significant effect, and was able to predict that the Democrats would win 44 percent of the vote in Kentucky and 34 percent of the vote in Mississippi. 

When asked if he was initially concerned about being the only student to predict a Republican victory in Kentucky, Angeles responded, “I was slightly worried when I saw the rest of the classes’ predictions. However, there are upsets in elections. In these two cases my model worked in predicting the gubernatorial election, but it may not have done so well in other elections.”

Another student, Elliot Balch, also developed an impressive model using GDP and incumbency. “The predictions are based on economic growth rates, polling information, and whether the incumbent is on the ballot,” Balch said. “Besides generating election predictions, my model reveals that being from the incumbent party, even in the absence of any economic growth, accounts for a substantial share of votes for a candidate. In fact, the effect of being from the incumbent party is several times stronger than that of personally being the incumbent.”

Though almost all of the students in the class failed to predict the unlikely outcome in Kentucky, the competition was a valuable exercise. According to Fowler, “Harris students learn about a lot of analytical tools in the abstract, but putting these tools to practice may lead to a deeper level of understanding. They are forced to think about prediction in ways that they wouldn't if they just discussed these concepts in the classroom.” 

—Mikia Manley