When we released our last presidential update, the consensus was that Donald Trump was favored to win. Since President Biden’s departure from the race, however, the electoral picture has changed considerably. Kamala Harris’ entrance into the race ushered in a boom in Democratic enthusiasm and revitalized the party’s fortunes in the presidential race.
Until now, we have created our presidential forecast using a data-informed, qualitative approach. To make our predictions more statistically rigorous and internally consistent, we decided to develop a quantitative model for the 2024 presidential election to accompany our pre-existing House and Senate models. Between now and the election, we’ll be updating each of our three quantitative forecasts weekly and writing a summary column of the big picture in each chamber.
You can find our presidential forecast here, where it’ll be updated once per week, along with a table of outputs and state-level probabilities.
What Does the Model Say?
Our model gives Kamala Harris a 62% chance of winning the election. In other words, Harris is a slight favorite, but the race is still quite close. Even a relatively small change in the polls could pull the election back toward being a pure tossup; alternatively, if Harris’ polls continue to improve in line with the post-debate trajectory, she may find herself in an even better position.
Harris has a slight advantage in the electoral college thanks to her leads in the most likely tipping point states. Pennsylvania is the tipping point state in 34% of our model’s simulations, making it the most likely one by a substantial margin. Harris currently has a 58% chance of winning this commonwealth, in large part to her 1.6-point lead in our polling averages.
Michigan and Wisconsin are the next most likely tipping point states. Collectively, they are about as likely to decide the election as Pennsylvania. Given that Wisconsin was the tipping point state in 2020 for a 269–269 tie, it wouldn’t be too surprising to see it play a similar role in November.
More surprisingly, however, North Carolina is the third most likely tipping point state according to our model. While Democrats have not won the Tar Heel State at the presidential level since 2008, Harris has benefited in recent weeks from a series of favorable polls here, allowing her to take a narrow lead in our state averages.
Given the state’s slight Republican lean and the fact that polls in demographically-similar states like Georgia point to a slight swing right in the region from 2020, the model is slightly more skeptical of her chances in North Carolina. Nonetheless, it forecasts the state to be the closest in the country, with the model projecting a 50–50 split, though with her narrow polling lead, we will nominally give the vice president the nod here.
Overall, the model sees 2024 as something of a “2020 redux,” with the Harris/Trump matchup looking very similar to the Biden/Trump faceoff four years ago. In some states, this is indicative of a return to form after an unusual 2022 that saw candidate quality playing an outsized role in an extremely polarized political climate.
In Florida, for example, Republicans routed Democrats by double digits in every 2022 race, marking a substantial improvement over Trump’s 3-point victory in 2020. This year, however, the model sees Trump as a significant, but not overwhelming, favorite to win the Sunshine State, with a projected margin of Trump +4. Similarly, while Democrats performed extremely well in Michigan’s 2022 elections, Harris is projected to match Biden’s 2020 3-point margin.
As it stands, five of the seven “significant” swing states are all tossups, with Michigan and Wisconsin being the exception. Our polling averages have the election quite close in all seven swing states, and our model reinforces this finding. In other words, while Kamala Harris is a slight favorite to win the election, her edge is razor-thin, and Donald Trump retains a good shot at returning to the White House.
How Does the Model Work?
The model has three major components and is fit on data from 50 years of presidential elections. The first is a state’s polling average, where we calculate an average of the polls taken in each state. You can read about how we do that here, but the upshot is that we weight on recency, sample size, population, and pollster quality, and we adjust for house effects for partisan pollsters.
This makes up the bulk of each state’s forecast. This is unsurprising, of course — the single best tool to understand how a state will vote is to look at high-quality state-level surveys, as this provides the most direct (if imperfect) measurement of voter intent. In states like Michigan and Wisconsin, where we have an abundance of polling, this makes up roughly 80% of the forecast on its own.
The next component of the forecast is state correlations. States tend to move together, so we calculate how similar states are to each other based on demographic, region, and partisan similarities. We then adjust each state’s polls based on how similar states are polling.
One example of this phenomenon in effect is in North Carolina. Trump won this state by 1.4 points, and although polls show Harris up by 0.3, the model projects a tie, because polls in demographically-similar states like Georgia and Virginia point to a slight swing right in this region from 2020. As a result, the model is slightly skeptical of such a pronounced leftward swing isolated to the state, and it adjusts her numbers downward as a result.
The last component of the forecast is national polling. In states and clusters where we have a paucity of survey data (like Oregon and the Pacific coast in general), this is an especially invaluable component, because it allows us to at least gain a rough picture of how the state is likely to shift if the nation is moving to the right as a whole. This national environment is blended with a economics-based fundamentals estimate (the weight of which declines as the election nears).
We then simulate the election 1,000 times, making sure to simulate error across the nation, at a cluster level, and at a state level alike. This allows us to simulate a wide variety of outcomes and helps us calibrate our error bands better.
Data Acknowledgement
Our demographics data comes from a mixture of the US Census (for demographics) and Urban Stats (for density). We use presidential data from 1972 onwards in fitting and testing the model. All polls are sourced from FiveThirtyEight.
I’m a computer scientist who has an interest in machine learning, politics, and electoral data. I’m a cofounder and partner at Split Ticket and make many kinds of election models. I graduated from UC Berkeley and work as a software & AI engineer. You can contact me at lakshya@splitticket.org
I am an analyst specializing in elections and demography, as well as a student studying political science, sociology, and data science at Vanderbilt University. I use election data to make maps and graphics. In my spare time, you can usually find me somewhere on the Chesapeake Bay. You can find me at @maxtmcc on Twitter.
I make election maps! If you’re reading a Split Ticket article, then odds are you’ve seen one of them. I’m an engineering student at UCLA and electoral politics are a great way for me to exercise creativity away from schoolwork. I also run and love the outdoors!
You can contact me @politicsmaps on Twitter.
My name is Harrison Lavelle and I am a co-founder and partner at Split Ticket. I write about a variety of electoral topics and handle our Datawrapper visuals.
Contact me at @HWLavelleMaps or harrison@splitticket.org
I’m a political analyst here at Split Ticket, where I handle the coverage of our Senate races. I graduated from Yale in 2021 with a degree in Statistics and Data Science. I’m interested in finance, education, and electoral data – and make plenty of models and maps in my free time.