How our 2028 nomination forecast works: the betting markets and the primary polls

Summary

Who will win the 2028 Democratic and Republican nominations? This far out, two sources speak to it and neither is enough alone. The national primary polls measure support among the names pollsters choose to ask about. Fifteen months before Iowa or New Hampshire, that mostly means how well known a name is. The betting markets, Kalshi and Polymarket, price something the polls cannot: whether a candidate will run at all.

Our published chance blends the two. The markets count for three quarters while the first contest is more than 300 days away. Their share falls in a straight line to one half on the day of the first contest. The other part comes from our own model, which runs the polls through the 2028 calendar state by state, 5,000 times, with the bounces, dropouts and delegate rules of each party.

On 8 November 2026 the markets made Jon Ossoff the Democratic favorite, at about 25%. He is on 6% to 8% in the polls. Kamala Harris led the polls, on about 30%, and stood at about 6% in the markets. A forecast built on the polls alone would have put the markets' favorite fifth or sixth. That is why we blend.

Why polls alone are not enough this early

Our model was tested on the open nomination races from 2004 to 2020, the ones with no sitting president of the party running. Before any votes, it gave the eventual nominee 35% to 39% on average. Simply reading a candidate's poll share as their chance gave 30% to 36%. The leader in the national polls before voting went on to win 3 of those 7 races. After eight contest dates the model gave the eventual nominee 78%.

So the polls do carry information, and the calendar adds to it. But the backtest has a gap that matters in 2028. In every race it covers, by the time the model was measured the field was mostly known: the candidates had declared or were about to. In November 2026 nobody major has declared. A poll can include a name who will never run, and leave out one who will. A candidate who says they are not running may still run (of about 35 who ruled out a run in recent cycles, about 1 in 4 went back on it), and a candidate who looks set to run may not.

The polls also lag. A name that has just come to national attention, such as a senator after a high-profile re-election, will be underpolled for months: many pollsters do not list them, and those who do find them little known.

Markets have the opposite weaknesses. They are thin, they can be moved by a few large traders, and their prices include a margin. But they price the question we are asking, who will be the nominee, and they react within hours to news about who is running.

The blend

For each candidate:

• Markets. The middle of the bid and the ask on Kalshi and on Polymarket, averaged where both list the candidate, then scaled so that a party's candidates add up to 100%. Placeholder and inactive markets are left out. The prices are read every 30 minutes.
• Poll model. The candidate's chance in our model with everyone polled in the field (described below).
• Weight. The markets' share is 75% while the first contest is 300 or more days away. It falls in a straight line to 50% on the day of the first contest.

Published chance = market weight × market chance + (1 − market weight) × model chance.

A candidate the markets list but the polls do not ask about takes the market part alone. A candidate in the polls whom the markets do not list takes the model part alone. If the markets cannot be read, the published chance is the model's.

The weights are a judgement, not a fitted value. There is no backtest of betting markets for nomination races before 2016 that we can run on equal terms with our model, and two or three cycles are too few to fit a weight. The reasoning is this: while the field is unknown, the main uncertainty is who runs, which the markets price and the polls do not; once votes are close, the field is known and the polls measure a real choice between real candidates, so the two count equally. We will review the weights when candidates declare.

Each row on the page shows its parts, for example "markets 24% · poll model 3% · 6.1% in the polls". A reader can see when the markets and the polls disagree, and by how much.

The poll model

Starting support

We take the national primary polls for each party (Elections.Core/World/UsPrimary2028.cs). Each pollster counts once, with its latest poll. The window is the 120 days before the newest poll while the first contest is more than 90 days away, and 60 days after that. In the backtest, 120 days did best at 90 days or more before voting, and 60 days at 30 days. A median or trimmed mean did no better than the plain mean, so there is no outlier damping.

A poll that did not list a candidate counts as half a point for them. Anyone averaging under 1% is left out. Each candidate's share among the field is their starting national support.

Uncertainty before voting

National support this far out is very uncertain. In each simulated race every candidate's starting support is moved by a random amount on the log scale. The spread was measured by horizon in the backtest: 0.3 up to 180 days before the first contest, rising to 1.2 at a year. Beyond a year, past the tested range, we let it keep widening, to 1.5 at 540 days. That is wide: at 1.5, a candidate on 10% in the polls is as likely to start a simulated race on 2% as on 45%.

State by state

Each simulated race then runs through the 2028 calendar:

1. Each state's vote. A state's shares are the national support moved by the state's own random noise: a standard deviation of 0.67 (log scale) in primaries and 0.94 in caucuses, with 0.25 more on the first three dates, where small electorates and retail campaigning make results harder to call. A candidate gets a home-state lift of 0.67.
2. Delegates. Democrats: proportional among candidates over 15%. Republicans: each state's 2024 rule (winner-take-all; winner-take-all above a trigger, otherwise proportional; or proportional above a threshold).
3. Momentum. After each date, a candidate who won a state gains national support, most when they were small (0.75 − 2.4 × their share, at least 0.05). Candidates who won nothing fall by 0.15. Everyone drifts a little (0.08).
4. Dropouts. A trailing candidate may drop out: 35% after each date under 5% nationally, 15% under 15%, and 30% once 40% of delegates are allocated for anyone holding less than a third of the leader's delegates. The leader never drops out.

The nominee is whoever holds a majority of delegates at the end. If nobody does, the race is counted as contested and the candidate with the most delegates is taken as the likeliest nominee. The page shows how often that happens.

All these values were measured on the open races 2004 to 2020, leaving each race out in turn.

The calendar

The Democrats' first six states are on the dates the party has set. The Republicans' first four are on their 2024 dates moved to 2028. Every other state uses its 2024 date, delegates and (for Republicans) rules until the parties publish their 2028 delegate tables.

Field scenarios

Because nobody has declared, the model also runs other fields: without each of the two poll leaders, and fields we set where a candidate has tied their decision to another's (for example, Gavin Newsom said he would not run if Harris did). These appear as "Who runs decides it" on the pages. The published chance uses the model with everyone polled; the markets weigh up who runs.

What we tested and left out

• FEC fundraising. Money is a known early signal in the academic literature. In our backtest, adding it made the early odds worse, so it is left out.
• Endorsements. Too few are made this early to test on 2028.
• Favorability. Not tested; it is asked inconsistently and mostly measures recognition at this stage.

Limits

• The backtest covers seven open races. Every figure from it carries wide error.
• The markets' weight is set by judgement, not fitted.
• Markets can be thin and moved by a few traders. A sudden move in one market is softened by averaging Kalshi and Polymarket and by the poll model's share, but not removed.
• The calendar is 2024's outside the early states until the parties set 2028's.
• A candidate who enters late with no polling history is only in the forecast through the markets until pollsters start to ask about them.

Where to see it

• The 2028 primaries: both parties, every candidate's chance and its parts.
• The Democratic primary and the Republican primary: the model's detail, the early states, field scenarios and the polls.
• The 2028 primary calendar.

Glossary

• Open race. A nomination race with no sitting president of that party running.
• Market chance. The price of a "yes" share in a candidate winning the nomination, read as a probability, after scaling so a party's candidates add up to 100%.
• Log scale. Support is moved by multiplying rather than adding, so a candidate on 2% and one on 40% can both double or halve.
• Bounce. The national support a candidate gains from winning a state.
• Contested convention. No candidate holds a majority of delegates after the last contest.

How to cite

Polling Forecast (2026). How our 2028 nomination forecast works: the betting markets and the primary polls. Retrieved 9 November 2026 from https://www.pollingforecast.com/research/us-nomination-forecast. Live figures change as new polls arrive; the date above is when the text was last revised.