The idea for 24cast.org began with a few friends trying to predict the outcome of the 2022 midterms on a whiteboard. Now, the group has dropped the whiteboard for something a bit more sophisticated: machine learning.
The model, created entirely by Brown students and trained on election data from 2002 to 2022, runs 100,000 simulations daily to forecast the most likely outcomes of the 2024 presidential, congressional and gubernatorial elections.
The website presents an interactive map that breaks down predictions by state. Users can see the two most likely outcomes — as estimated by the software’s simulations — as well as how the predictions have changed over time.
According to Asher Labovich ’26, the founder of 24cast.org, the model uses over a hundred features, ranging from previous elections results to campaign finances and voting accessibility in each state, to make its predictions.
The methodology used by 24cast.org is fundamentally different from other predictive models, Labovich said.
While all of them use similar data — such as polls, past elections and expert ratings — traditional election forecasters determine the weight of each factor manually. Instead, the Brown team “uses machine learning from the get-go because we don’t want to make assumptions,” Labovich explained.
“Rather than us determining what’s important, we let machine learning determine it based on the data that exists,” Ariel Shifrin ’27, the team’s head of operations, said.
Labovich created 24cast.org to replace his past model forecasting the 2022 senate midterm elections, which was published on the Brown Political Review website.
24cast.org is also affiliated with BPR, but uses a different approach from Labovich’s original election forecaster. “It’s an entirely new code … We believe it to be more accurate, more rigorous,” Shifrin said.
Methodology
The model uses mostly open source data, which can be found on the team’s Github. “We’re trying to provide something that’s very transparent,” Shifrin told The Herald.
24cast.org’s methodology page describes computational methods used to achieve daily predictions and lists any changes made to the code. The website also displays the most predictive factors for every outcome that users can interpret.
“We explain how we get to our predictions and provide something that no other model does, called Shapley values, that allow people to (see) exactly what factors contributed to each of our predictions,” Shifrin said.
Labovich told The Herald that a growing distrust of forecasting models resulted in the team’s dedication to transparency. Shifrin added that the team hopes their clarity creates a level of trust for users.
“For the last few elections, we’ve seen a lot (of people) lose confidence in these election models because they feel that the people creating them have very clear political biases, or feel they’re skewing the data,” Labovich said.
Paul Testa, assistant professor of political science, noted that there are tradeoffs to “trusting an algorithm.” It “needs a lot of data to produce the results you would expect,” he said. But “elections are fundamentally a rare event.”
Testa also noted that though the machine learning algorithm is data driven, there will still be human input in the model. “The amount of work that went into getting the data set up, how to transform it; those are human decisions,” he said. “But that’s true of all forecasting.”
“What they’ve accomplished is really impressive,” Testa added, saying the potential of the model for future forecasting is “really promising and interesting.”
Election day
On Nov. 5, 24cast.org will unveil the technology at an event, which will include student and faculty panels and field new election data from 7 p.m. to 2 a.m. via a decision desk, according to Labovich.
Shifrin said that the team is currently working on increasing campus and community engagement with 24cast.org. “We’re working on developing a series of events to center the technology that we’ve built, and also bring people together at the confluence of politics and technology,” he explained.
There are also plans to make the 24cast.org website a more interactive experience on election night. Alex Wick ’25, the web operations lead, said that it will include a “needle-like feature” that displays which candidate is leading as data comes to the decision desk.
There will also be a way to view how the prediction model stacked up compared to the actual results. “I’m pretty excited about creating a cool, interactive, engaging way to see” how the model performed, Wick said.
