First, an important distinction
Odds and vote share are not the same thing
You will see three topline numbers on most race pages. They sound similar, but they are answering three different questions.
- Win chanceHow often a candidate wins across all of the Election Days simulated by the model.
- Projected voteThe model’s average estimate of how much of the vote each candidate or party receives.
- Projected marginThe expected difference between the two main sides of the race.
If a candidate has a 70% chance of winning, we are not saying they will receive 70% of the vote. We are saying they win about seven out of every ten times the model runs the election.
The short version
Every race starts with two estimates
At its core, the forecast is a blend of polling and fundamentals. Polling tells us what voters are saying right now. Fundamentals give us a second estimate based on the state’s political history, the national environment and, once the matchup is actually settled, the candidates themselves.
How much we trust each side of that blend depends on the race. Five recent polls should tell us more than one survey from three months ago. A race with no usable polling can still be forecast, but the model will be much more dependent on fundamentals—and appropriately less certain.
- Start with the polls.Average the qualifying surveys and put older results in the context of the national mood at the time.
- Build the fundamentals.Use state history, partisanship, national conditions and any meaningful candidate information.
- Combine the two.Give stronger and more recent polling more influence.
- Run the election thousands of times.Account for polling misses, campaign movement and surprises that can affect several states together.
The polling side
What are voters telling us right now?
We start with the qualifying public polls in each race. Democratic, Republican, other and undecided voters are kept separate. That is important because “undecided” is not another word for “third party,” and it certainly does not mean we can quietly give all of those voters to whichever candidate is already ahead.
An old poll is not useless—but it needs context
Take a state like Alaska, where it is entirely possible to go six weeks or longer without a new survey. If the national political environment changed during that gap, the old state poll may no longer describe where the race stands today. The model looks at national polling from when that state poll was conducted, compares it with the current national average, and applies the change to the state.
And if the latest state poll is several months old, the comparison can go several months back. We are not just grabbing the second-most-recent generic-ballot reading and calling it close enough.
Show the polling adjustment
Each party moves by its own change in national support. The national margin is never copied directly into a candidate’s state vote share.
The fundamentals side
What would we expect if the polls disappeared?
Polls are useful, but one strange survey should not be able to rewrite everything we know about a state. Fundamentals are the model’s reality check—and in races with no polling at all, they become the forecast’s starting point.
Every state has its own political history
Senate and governor races do not always behave the same way, even in the same state. So for a Senate forecast, recent Senate results receive most of the historical weight. For a governor forecast, recent governor results do. The state’s broader partisan lean and results from the other statewide office fill in the rest.
One technical point matters here: this produces a margin, not two separate bonuses. If the state-history calculation is D+10, the model adds five points to the Democratic share and subtracts five from the Republican share. The overall margin moves by ten—not twenty.
Show the state-history weights and formula
Republican adjustment = −state-history margin ÷ 2
The national environment matters, but it has limits
The national starting point comes from the generic congressional ballot. We also use the HDR Analytics Consumer Sentiment Index—which combines the University of Michigan, Conference Board and Gallup confidence series—as a modest signal about the environment facing the president’s party.
To be clear, a terrible consumer-confidence number cannot single-handedly turn Wisconsin into a 20-point Democratic or Republican state. The adjustment is intentionally softened: it can move things near the middle, but even an extreme reading can never create more than a four-point shift in the national two-party margin.
Show the consumer-confidence formula
That shift is split evenly between the two major parties. Other and undecided voters do not move, which keeps the full vote at 100%.
Fundraising is a small signal, not a requirement
Once the general-election matchup is settled, fundraising can tell us something about candidate strength. It is still only a modest part of the fundamentals. If the matchup is unsettled—or reliable fundraising data simply is not available—the model treats it as neutral instead of inventing an advantage.
Putting it together
Not every polling average deserves the same weight
A polling average built from several recent surveys should matter more than one built from a single old poll. That is the basic idea behind the polling weight: it rises with more polls and falls as the latest poll gets older.
There is also a nationwide ceiling that changes with time. Even an unusually well-polled race cannot receive Election-Week confidence months before Election Day, when there is still plenty of time for the campaign to move.
“Other” and “undecided” are not interchangeable
This sounds obvious, but it caused a real problem in the earlier version of the model. Many generic-ballot polls ask only whether someone prefers a Democrat or Republican, then place everyone else in an unassigned group. Those voters are undecided; they are not automatically supporting a third-party candidate.
If a national survey does not measure other candidates directly, we reserve a conservative 2% starting share for minor-party and write-in votes. Everyone else stays undecided. If a race poll actually names an Independent or third-party candidate, that race-specific evidence takes priority.
Democratic, Republican, other and undecided shares are blended separately. At the end, the ordinary D-versus-R model allocates the remaining undecided vote 40% Democratic, 40% Republican and 20% other. In a race where an Independent replaces one of the major-party sides, a neutral proportional rule is used instead so the Independent is not given an artificial boost just for occupying that slot.
Show the Other-vote rule
Race-specific Other polling takes precedence over this national prior. Named Independent candidates are modeled from their own available evidence; the residual Other bucket covers remaining minor-party and write-in support.
Show the polling-weight formulas
The model uses the lower of those two numbers. If there are no qualifying polls, polling receives 0% and fundamentals receive 100%.
Turning votes into odds
A projection is the middle of the forecast, not the end
Once we have a projected margin, we still need to ask how wrong that estimate could be. The model creates thousands of plausible Election Days around it. If the Democrat wins 6,300 of 10,000 simulations, their published win chance is 63%. It is that straightforward.
The forecast is wider early for a reason
A lot can happen between now and Election Day: campaigns can improve or collapse, the national environment can move, and the polling itself can change. The model leaves more room for those possibilities early in the cycle, then gradually narrows the range as voting gets closer.
Show the time-uncertainty formula
This is only one part of the total uncertainty. Polling quality and shared political surprises are handled separately.
A bad night for one party probably will not stop at one state
Races are connected. If polls underestimate Republicans nationally, they are likely to miss in the same direction in more than one state. A regional surprise can do the same thing on a smaller scale. Each simulated election therefore includes national, office, regional and state movement, along with noise unique to each race.
The stacked dots on a race page show those simulated margins. Democratic margins sit to the left of even, Republican margins sit to the right, and the middle 80% highlights the central range. Outcomes beyond that range are still possible; they are just less common.
From shares to raw votes
Turnout changes the vote total—not who is favored
This distinction is easy to miss. Polls and fundamentals determine the projected vote shares and win chance. The turnout model takes those shares and estimates how many actual votes they represent. A higher turnout estimate does not, by itself, help either candidate.
The baseline uses statewide ballots counted, citizen voting-age population and, when a state reports a comparable number, active registration from the 2018 and 2022 midterms. We carry population and registration forward at their recent pace, but cap four-year growth between −8% and +12%. That keeps one unusual registration change from taking over the estimate.
For most states, 80% of the turnout estimate comes from recent ballots per active registrant and 20% comes from ballots per voting-age citizen. North Dakota does not have a comparable active-registration series, so it relies on the complete population-and-ballot history instead of being left out.
States do not move in a vacuum
After the state estimates are built, each region is moved 25% toward its recent relationship with national turnout. Think of this as a guardrail. It keeps one state from drifting far away from the broader region based on a noisy input, while preserving both the national total and each state’s share within its region.
Show the turnout calculation
The saved turnout model is turnout-v4.1-automatic. Until a complete current-cycle primary signal exists, national turnout intensity stays at the neutral two-cycle baseline rather than being guessed from partial results.
From individual races to Senate control
The chamber is simulated as one election
We do not calculate Senate control by simply adding up a list of isolated race probabilities. In every simulation, the model combines the winners of all contested seats with the seats each party already holds. Because races share national and regional movement, an upset in one state can tell us something about what probably happened elsewhere.
The Election Mosaic shows 100 representative outcomes. The curved chamber and seat distribution answer a slightly different question: how many seats each party is likely to hold, and how much uncertainty remains around that average.
Build your path to Senate control lets you test your own scenario. Choose winners in uncertain races and the page filters the model’s saved simulations to the Election Days that best fit those choices. It is a way to explore the forecast, not a new forecast of its own. Hit reset and you are back to the published odds.
What counts as a projected flip?
A seat is labeled a projected flip when a party other than the current holder has a greater than 50% chance of winning. The overall chance that a seat changes hands can be even higher if multiple challengers have a path. For an open seat, we use the party that held it entering the election. An Independent winner stays in the Other column unless that candidate has an explicit chamber-alignment setting.
What changed
How is this different from the 2024 model?
The basic idea is still the same: calculate a polling estimate, calculate a fundamentals estimate and blend them. I did not throw out the 2024 approach. The goal was to keep what worked while making the process more consistent, more transparent and much easier to update.
A 2026 win chance is not perfectly interchangeable with one from the old system. The new simulation models uncertainty across races together, so race odds and chamber-control odds can change even when the central vote estimate looks similar.
Race calculation
Polls and fundamentals were calculated separately, then combined with one polling weight.
The same sequence remains, but every public forecast saves its inputs, intermediate steps and final estimate so changes can be traced.
Polling dates
Race polls and the matching national comparison required more manual upkeep.
Race polling is imported and matched to the national environment from the period it represents—even when the latest state poll is several months old.
Consumer confidence
The three confidence readings fed the national environment without a firm limit on an extreme result.
The same three-source signal remains, but its effect is softened and capped at four points of national two-party margin.
Candidate phase
Candidate details and fundraising were maintained race by race.
The forecast follows nominees, withdrawals and Independent candidates explicitly. Fundraising is neutral before a matchup is settled, and an Independent winner is not counted toward either major party’s Senate total.
Uncertainty
Central race estimates fed a separately maintained simulation process.
Race odds, seat totals and Senate control all come from the same reproducible simulations, with shared national and regional surprises represented together.
Turnout and vote totals
Projected turnout could be supplied race by race when a raw vote estimate was needed.
A versioned statewide turnout model now creates the estimate automatically from recent turnout, population and registration history. It changes projected vote counts, not vote shares or win chances.
Forecast history
Publishing updates involved more manual handoffs.
Published updates retain their date, assumptions and results, making historical charts and past-date maps possible.
Sources and transparency
Where does all of this data come from?
The short answer is: several places. Public race averages come from HDR Analytics polling. The consumer-confidence adjustment—and all three series behind it—is published on the HDR Analytics CSI page. State history comes from certified statewide election results, while the partisan baselines are saved as versioned model inputs.
For turnout, we use ballots counted and active-registration data from the U.S. Election Assistance Commission’s Election Administration and Voting Survey. Citizen voting-age population comes from the Census Bureau’s five-year ACS CVAP files. The state-level dataset is saved alongside the model so the turnout estimate and its backtests can be reproduced.
Every public page is built from the data file included with that forecast update. That file contains the race projections, probabilities, ranges, turnout estimates and the supporting values used by the charts. In other words, the numbers on the page are not typed into the design afterward.
Download the current forecast data →Browse forecast updates →
The honest answer
Yes, the forecast can be wrong
Polls can miss together. A primary can produce a nominee nobody expected. A major event can move the race faster than new surveys can measure it. Turnout can also look very different from the last two midterms.
That is why we publish odds and ranges instead of pretending the number after the decimal point is destiny. A 10% outcome should happen about one time in ten. It is unlikely, but it is absolutely not impossible. And the earlier the forecast, the more caution it deserves.
One final caveat: timing does not prove causation. If a race moves after a new poll or an economic report, the saved calculation can tell you exactly what changed inside the model. It cannot prove that one event changed the minds of actual voters.
A special case
What changes in a ranked-choice race?
Most races are simple: whoever receives the most votes wins. Ranked-choice elections are different because a candidate can lead the first round and still lose after lower-performing candidates are eliminated.
The RCV model is currently limited to Alaska Senate and Alaska Governor general elections, along with Maine Senate general elections that have at least three qualified candidates. Maine governor races, primaries and every other office still use the regular model.
We begin with the same polling-and-fundamentals blend used everywhere else. That gives us the first-choice projected vote. It does not give us the eventual winner. If nobody reaches 50%, the lowest candidate is eliminated and their ballots can transfer to someone else—or become exhausted if no remaining candidate is ranked.
What if there is no later-round poll?
The forecast still runs. It uses official transfer patterns from past Alaska and Maine elections, adds an explicit exhausted-ballot share and widens the uncertainty. This is not as strong as having a poll of the exact final matchup, so the page says when later-round polling is unavailable.
If a polling page does include a second- or third-round question, that result informs what voters might do with the candidates who remain. We do not count it as a brand-new poll. It is another question from the same survey, and counting it twice would give that poll too much weight.
In each simulation, the lowest candidate is eliminated and their votes move according to a sampled transfer pattern. The process continues until someone has a majority of the ballots still active. Independents are not automatically split evenly—or at all—between Democrats and Republicans.
Show the RCV technical details
The public page reports candidate win probabilities, first-choice shares, common final matchups, round probabilities, expected exhausted ballots and the final decisive-round margin range.
For the legal framework, see the Maine ranked-choice statute and the Maine election-results archive. Alaska’s official tabulations are used when available.
To be clear, the historical sample is still small. Nominee quality, ballot exhaustion and voters’ second choices can all matter a lot. That is a reason to show more uncertainty—not a reason to pretend the election is decided by the first round.