Elections Are More Predictable than They Look

As politics in America become more polarized and each voting cycle seems more fraught than the last, a common narrative in the media is that each election is unique and increasingly unpredictable. But new research shows that congressional elections follow strong statistical patterns, even when individual races produce surprises.

"There are a lot of generalities and systematic things happening in elections," says Jonathan N. Katz, Caltech's Kay Sugahara Professor of Social Sciences and Statistics. "It's not to say that there aren't interesting stories, but, overall, elections are very predictable, and I don't think the general public thinks about them that way. Patterns help estimate the odds, but close races—and control of Congress—can remain uncertain."

To help illustrate the effects of different data points on predictions and give people a practical tool with which they can build their own Election Day scenarios, Katz teamed up with former graduate student Danny Ebanks (PhD '24), now a research associate at Harvard's Institute for Quantitative Social Science (IQSS), and political scientist Gary King, the Weatherhead University Professor at Harvard University and director of IQSS, to build a forecasting website called ElectIt. The interactive tool draws on decades of election data that the team assembled—with the help of AI—covering every US House election since 1946 and Senate election since 1954 as well as the current races.

Katz has been modeling election dynamics for more than three decades. The forecaster is one product of a larger project with Ebanks and King: building the first generative statistical model that can simulate whole elections and that accurately describes congressional elections. King, a longtime collaborator of Katz's (he was Katz's postdoctoral advisor at Harvard), had co-developed what became the canonical model for legislative elections in the 1990s with statistician Andrew Gelman. Seven years ago, Katz, Ebanks, and King tried to port that model to a modern programming language and noticed a major flaw.

"The model actually didn't fit very well; results that it treated as less than one-in-10,000 events in individual races showed up in almost every other congressional election cycle," Katz explains. "For example, it badly underestimated how often popular incumbents lose their seats. We tried lots of different models to make one that worked better and finally came up with a framework that fits the data. The forecaster is one application of it."

Unlike many election-prediction models seen in popular media, ElectIt's core model uses no polls at all. Instead, it draws on decades of statistical data—such as past results; which candidates are running and whether one is an incumbent; and the midterm penalty, which is the tendency of the president's party to lose support in midterm elections. That information is used to build three types of forecasts: the "post-nomination forecast," which acts as a baseline, using the candidates and past results but no polls; the "model + generic ballot," which accounts for changes in the national political environment by adding national surveys that ask which party's congressional candidate people plan to vote for; and the "model + all polls," which adds available state- and district-level polls.

"It's a fun tool with lots of bells and whistles that let you adjust and play around with things, like you can call different elections, and it will update the entire map," says Katz, who hopes the public might use it during watch parties on November 3. "On election night, when the Associated Press starts calling races, you can put them into our app, and it will tell you how the probabilities update."

Iterating on the model required rewriting it to run on graphics processing units (GPUs)—a switch made with the help of AI coding tools and substantial GPU computing time. AI tools also helped build the website and construct new measures from historical data, such as statewide election results broken down by congressional district.

"No model is perfect. But one of the key things that we were able to do with AI programming is backfill data on things like gubernatorial elections and presidential elections," Katz says. "Most of this was possible without tools like Claude and Codex, but painfully slow. Without them, we couldn't have iterated on the model the way we did."

The researchers tested their new model by fitting it to earlier elections and checking its forecasts against later results. After the election, Katz says the results will provide another test of the model.

"We can compare the results with our predictions to learn where the model works and where it needs improvement," he explains. "An unexpected result can prompt further investigation, but it is not by itself evidence that anything went wrong with the election." Surprises, he adds, can have ordinary explanations, such as a scandal or an unusually strong challenger.

Katz notes that the team's approach emphasizes transparency: Election data, forecasts, and assessments of historical accuracy are available on the ElectIt forecasting website. A paper that describes their methods, "If a Statistical Model Predicts That Common Events Should Occur Only Once in 10,000 Elections, Maybe it's the Wrong Model," is under review at the journal Political Analysis.

"I'm not in this for forecasting elections," Katz says. "What Gary and Danny and I are interested in is understanding American politics, understanding elections. But in order to make plausible claims about things, we think you have to have a generally accurate model, and ours does forecast really well."

As of October 21, the ElectIt forecasting site was averaging 500,000 Google impressions a day, according to Katz. Build your own forecast at: https://gking.harvard.edu/electit.