The fundamentals: partisanship, the national vote and incumbency
Abstract. Every race has a prior from its fundamentals: the state's presidential lean, the national House vote and whether an incumbent is standing again. Fitted on 420 Senate and 322 governor races since 2000, Senate results follow the lean closely (0.84 per point) and governors much less (0.39), while incumbency counts for more in governors' races. The fits leave about 15 points unexplained, so polls soon outweigh them, but in eleven unpolled races the prior is the whole forecast. The paper shows the priors at work in thinly polled races and their record in the backtests.
Summary
Every race in the forecast has a prior: what we would expect from how the state or district votes for president, the national House vote and whether an incumbent is standing again, before reading any poll of the race. This paper describes those priors, how they are fitted, how wide they are, and what they do in races with few or no polls.
For Senate and governor races the prior is a regression fitted on every contested race since 2000: 420 Senate and 322 governor races. Senate results follow the state's presidential lean closely (0.84 points per point of lean) and the national House vote strongly (0.73); governors follow both much less (0.39 and 0.54) and depend more on incumbency (15.1 points for a governor standing again, against 9.6 for a senator). What these fits leave unexplained is large, about 15 points of margin, so a single good poll outweighs the prior. In the six Senate races and five governor's races that have no poll of the nominees, the prior is the whole forecast, and in races resting on one polling firm it carries a fifth to a third.
House districts have their own fit, on 732 contested districts in 2022 and 2024, which is much tighter (5.9 points) because district results follow the presidential vote almost exactly; From the generic ballot to 435 districts describes it.
The lean
A place's partisanship is its presidential lean: how far it voted from the nation at the last two presidential elections, with the most recent counting most. For a race in year \(y\), with \(p_1\) the last presidential year before \(y\) and \(p_2 = p_1 - 4\),
where \(m\) is the state's presidential margin (Democrat minus Republican) and \(M\) the national popular-vote margin. For 2026, \(p_1\) = 2024 (national R+1.5) and \(p_2\) = 2020 (D+4.5). Kansas has a lean of R+15.8, Iowa R+12.0, Michigan R+0.4, Maine D+7.5 and Idaho R+35.1. If only one of the two elections is available, it is used alone.
The weights of three quarters and a quarter are a choice, not a fitted value: they keep the most recent election in charge while damping one-off shifts.
The statewide fit
For every contested Senate and governor race from 2000 to 2025 (both main parties above 5%, finishing first and second), we fit
where \(m_r\) is the result margin, \(L_r\) the state's lean before the election by equation (1), \(E_r\) the national House vote margin of that year (odd-year governors take the year before), and \(I_r\) is +1 if a Democratic incumbent stood again, −1 for a Republican, and 0 for an open seat. Fitting by least squares gives:
| Office | Races | \(a\) (intercept) | \(b\) (lean) | \(e\) (national vote) | \(i\) (incumbent) | Residual |
|---|---|---|---|---|---|---|
| Senate | 420 | +2.92 | 0.84 | 0.73 | 9.60 | 15.27 |
| Governor | 322 | −1.49 | 0.39 | 0.54 | 15.07 | 14.39 |
Table: The statewide fundamentals, fitted on contested races 2000–2025. Points of margin; the residual is the root mean square of what the fit leaves unexplained.
Three features matter.
Senators vote with their state. A Senate race moves 0.84 points for each point of presidential lean; the Senate has become a federal office voted on national lines. A governor's race moves only 0.39: states regularly elect governors of the party their presidential vote opposes (Vermont, Maryland, Massachusetts, Kansas and Kentucky have done so in this period).
Incumbents matter more for governors. Because \(I\) is coded +1 and −1, the difference between a Democratic and a Republican incumbent in the same state is twice the coefficient: about 19 points for a Senate race and 30 for a governor's race. Part of this is selection: incumbents who expect to lose often retire, so those who stand again are disproportionately safe.
The national vote passes through. For each point the national House vote moves, a Senate race moves 0.73 points and a governor's race 0.54. In the fit \(E\) is the actual national House vote of the year; in the forecast it is our projection of it, with its error, which is why the prior's width includes the national error (equation 3).
Incumbency
An incumbent is standing again if the winner of the same seat's previous race is among this race's candidates for the same party: for a Senate seat the race six years before, for a governorship the last election in the state within four years. Candidates are matched by surname and party. Where a result file marks the incumbent directly, that mark is used. For 2026 the status comes from the candidate lists: a race's incumbent is the sitting holder if listed as running again.
Some cases the rule cannot see. An appointed senator standing for the first time (as in Florida and Ohio's special elections this year) is counted as an incumbent if listed as one; a former governor returning after a gap is not.
Independents
Where an independent is the main challenger to the Republican (Nebraska, Idaho, South Dakota and Montana this year), the independent takes the Democrat's place in equation (2), and the residual is multiplied by 1.5, because the fit has too few races of this kind to measure how far they stray from it.
The width of the prior
The prior's standard deviation combines the fit's residual with the uncertainty of the national vote:
where \(s\) is the residual from the table, \(\eta\) is 1.5 for an independent challenger and 1 otherwise, \(e\) the national-vote coefficient and \(\sigma_E\) = 5.17 points the standard deviation of our national House vote (The generic ballot: the national vote behind every race). Today \(\sigma_P\) is 15.7 points for a Senate race, 23.2 for a Senate race with an independent challenger, and 14.7 for a governor's race.
That is wide. A prior of R+15 with a standard deviation of 15.7 gives the Democrat a chance of about one in six before any poll. The fit's residual is large because it is fitted on every contested race, including many one-sided ones in which candidate quality, scandal and money produce margins far from the state's lean; it is an honest measure of how far a race can stray from its fundamentals, and it means the prior gives way quickly to polls.
The blend, and what the prior does
The prior and the poll average are combined by their precisions (State polls: the pollster-adjusted average and the blend, equation 7). With a prior this wide, a race's polls usually carry 80% to 90% of the forecast, and the prior's main work is in three kinds of race.
Races with no polls
Six Senate races and five governor's races have no poll of the nominees in their average. Their forecast is the prior alone:
| Race | Lean | Incumbent | Prior | Chance |
|---|---|---|---|---|
| Colorado Senate | D+11.6 | D (Hickenlooper) | D+27.1 | 96% D |
| Delaware Senate | D+15.8 | D (Coons) | D+30.6 | 97% D |
| New Jersey Senate | D+8.4 | D (Booker) | D+24.4 | 94% D |
| Oregon Senate | D+14.7 | D (Merkley) | D+29.7 | 97% D |
| West Virginia Senate | R+41.1 | R (Capito) | R+36.4 | 1% D |
| Wyoming Senate | R+45.2 | open | R+30.2 | 3% D |
| Colorado governor | D+11.6 | open | D+6.5 | 68% D |
| Hawaii governor | D+24.7 | D (Green) | D+26.7 | 96% D |
| Oklahoma governor | R+34.0 | open | R+11.0 | 22% D |
| South Dakota governor | R+28.4 | R (Rhoden) | R+24.0 | 5% D |
| Wyoming governor | R+45.2 | open | R+15.3 | 15% D |
Table: Races resting wholly on their fundamentals, 15 October 2026. Margins, points; chances from the simulation.
The governor's fit's weak dependence on lean shows here. An open governorship in Wyoming, R+45 on the presidential vote, has a prior of only R+15 and a Democratic chance of 15%; in Oklahoma 22%; in Colorado, D+12 on the presidential vote, the Democrat's chance is 68%. Those chances are what the fit says about open governor's races across 25 years, in which states did elect governors far from their presidential lean. They are probably generous to the underdog in the most one-sided states, where the fit has few races, and these races are unpolled because nobody expects them to be close; What we don't know yet: limitations of the US forecast records this.
Races resting on one firm
When a race's polls come from one or two firms, the average's error is wide (State polls: the pollster-adjusted average and the blend) and the prior carries more:
| Race | Effective firms | Poll average | Prior | Polls' share | Forecast |
|---|---|---|---|---|---|
| Illinois Senate | 1.0 | D+11.7 | D+18.2 | 81% | D+13.0 |
| Kentucky Senate | 1.0 | R+13.8 | R+16.9 | 77% | R+14.5 |
| Oklahoma Senate | 1.0 | R+28.5 | R+20.8 | 74% | R+26.5 |
| Tennessee Senate | 1.0 | R+23.0 | R+25.5 | 77% | R+23.6 |
| Arkansas governor | 1.0 | R+5.0 | R+24.5 | 73% | R+10.3 |
| Idaho Senate (independent) | 2.7 | R+5.3 | R+31.3 | 65% | R+14.4 |
Table: Races with thin polling, 15 October 2026. Points of margin.
Arkansas's governor's race shows the blend at work: its one poll, by Hendrix College, puts Sarah Huckabee Sanders 5 points ahead after adjustment; the fundamentals say 24.5. The forecast is R+10.3, with the poll still carrying most of the weight because the prior is so wide.
Races where incumbency and lean disagree
| Race | Lean | Incumbent | Prior | Polls | Forecast |
|---|---|---|---|---|---|
| Vermont governor | D+32.5 | R (Scott) | R+0.5 | R+5.5 | R+4.6 |
| New Hampshire governor | D+3.9 | R (Ayotte) | R+11.5 | R+10.3 | R+10.4 |
| Arizona governor | R+4.1 | D (Hobbs) | D+15.6 | D+9.5 | D+10.2 |
| Georgia Senate | R+1.6 | D (Ossoff) | D+16.0 | D+11.1 | D+11.6 |
| Pennsylvania governor | R+1.0 | D (Shapiro) | D+16.8 | D+23.4 | D+22.6 |
Table: Incumbents against their state's lean. Points of margin.
In each of these the fundamentals, with the incumbency term, point the same way as the polls, and often about as far: Phil Scott in Vermont is the clearest case, a Republican governor of the most Democratic state in the table, whom the fit puts level and the polls 5.5 points ahead.
How the fundamentals have done
The backtests of 2018, 2022 and 2024 forecast each statewide race three ways, with every parameter fitted without the year tested: the fundamentals alone, the final polls alone (partisan polls adjusted) and the blend (Correlated error and the backtests of 2018, 2022 and 2024).
| Year | Office | Races | Fundamentals alone right | Polls alone right (races polled) | Blend right |
|---|---|---|---|---|---|
| 2018 | Senate | 32 | 27 | 26 of 29 | 30 |
| 2018 | Governor | 36 | 30 | 30 of 34 | 33 |
| 2022 | Senate | 33 | 30 | 23 of 23 | 32 |
| 2022 | Governor | 36 | 33 | 27 of 29 | 34 |
| 2024 | Senate | 32 | 29 | 23 of 24 | 30 |
| 2024 | Governor | 11 | 10 | 8 of 8 | 11 |
Table: Winners called right in the backtests. The fundamentals use the final generic ballot of each year as the national vote.
The fundamentals alone called 159 of 180 races right (88%), the blend 170 (94%). Where a race was polled the polls alone did about as well as the blend, and the fundamentals' value is in the unpolled and thinly polled races, which the polls alone cannot call at all.
The backtest uses a simpler blend than today's forecast: the poll average's variance is the mean squared final-poll miss of the other years, shrinking with the number of polls, and the prior's is its residual plus nine. Today's forecast adds the firm count, the pollster adjustments and the national trend (State polls: the pollster-adjusted average and the blend). The backtest therefore tests the structure, not every detail.
The fundamentals beyond 2026
The same fits drive the early projections of the 2028 Senate class III and the governors of 2027 and 2028, with the national vote taken from the 2028 presidential baseline and incumbency left out where it cannot be known (The 2028 presidential forecast: a party contest with candidates on top).
Limits
A wide residual. About 15 points of margin for statewide races. The prior is honest about it, but in an unpolled race a 15% chance for a Wyoming Democrat is a statement about the spread of past results, not about Wyoming.
Hindsight in the fit. The fit uses the actual national House vote of each year. In the forecast, our projection of it carries its own error, added in equation (3).
Incumbency by name. The rule matches surnames and party, so a relative with the same surname, or a married name, can be mistaken; the 2026 statuses come from candidate lists and are not affected.
One fit across 25 years. The Senate's dependence on presidential lean has grown since 2000; one coefficient averages the change.
Figures as of 15 October 2026. The pages they come from are rebuilt with each new poll.
How to cite
Polling Forecast (2026). The fundamentals: partisanship, the national vote and incumbency. Retrieved 17 October 2026 from https://www.pollingforecast.com/research/us-fundamentals. Live figures change as new polls arrive; the date above is when the text was last revised.