What we don't know yet: limitations of the US forecast

Summary

This paper collects the known weaknesses of the United States forecast in one place, says how much each matters, and what evidence would change our approach. The British counterpart is What we don't know yet: limitations and open questions; the method itself is set out in Forecasting the 2026 midterms: an overview and the papers it links to.

The largest weaknesses are in the evidence, not the arithmetic. The corrections rest on twelve cycles of state polls and thirteen final generic ballots, and the statewide correction passes its test by 0.06 points. The polling record shows larger misses in the most Republican states, which the forecast does not correct for. The simulation's regional errors are sized on presidential swings, not on polling errors, and for well-polled statewide races they add more spread than the race's own error allows. The House model is fitted on two elections, applied to 173 districts on new lines, and does not yet handle an independent who is the main challenger in a House district. The backtests count winners, not probabilities. And several inputs that readers watch closely (special elections, approval, early votes) are kept out because no test can yet show that they help.

A short record with a weak lean

Everything the forecast knows about polling error comes from 550 Senate and governor races with final polls since 2000, grouped into twelve cycles with five or more races, and from final generic-ballot polls at thirteen House elections.

The statewide correction is marginal. Taking half the other cycles' mean miss off every race brought the average miss from 4.60 to 4.55 points, closer in 6 of 12 cycles (Measuring American polling error, 2000–2024). It passes the rule, but by an amount smaller than the difference between any two reasonable versions of the test: a forward-only test, in which each cycle learns only from earlier cycles, gains 0.02 points and is closer in 4 of 10. Two cycles like 2006 to 2012, when the polls understated the Democrats, would reverse it. A check on the 2025 governor's races of New Jersey and Virginia, where the final polls understated the Democrats by 7.6 points on average, shows the correction can widen a miss.

The generic-ballot record is thin. The correction passed at 11 of 13 elections, but five of them rest on one final poll and the record before 2010 on four pollsters. It is measured on the polls' decided-vote shares against official shares that include minor parties and give a party nothing in districts it did not contest, so part of what it learns is a difference of basis rather than the polls' lean (The generic ballot: the national vote behind every race).

What would change our mind. Both tests are rerun every time the forecast is built, and both switch off on their own if a new election reverses them. The 2026 results will be the thirteenth statewide cycle and the fourteenth generic-ballot election.

The red-state pattern

Since 2018 the final polls of states at least 15 points more Republican than the nation have overstated the Democrats by 4.3 points on average, against 0.7 in swing states and −1.0 in the most Democratic states (Measuring American polling error, 2000–2024). The Democratic path to a Senate majority this year runs through Iowa, Kansas, Ohio, Texas, Alaska and Nebraska, all in or near the reddest band.

The forecast applies one correction to every race and lets the simulation draw regional errors in random directions. It does not lean the miss towards red states, because no regional correction has passed a leave-one-out test. A site article ran the stress test: applying the post-2018 band misses to every race moved the Democratic chance of Senate control from 59% to 42% in that run, and using the whole 2000–2024 record instead gave 54%.

What would change our mind. A regional correction that beats the single correction on the same leave-one-cycle-out test. The band pattern was flat from 2000 to 2008, and 2024's misses were not largest in the reddest states, so the test is not obviously passed.

The simulation

Regional sizes from presidential swings. The division (2.6 points), state (3.9) and demographic (2.5) errors are measured on how far states' presidential swings strayed from the national swing from 2004 to 2024 (Correlated error and the backtests of 2018, 2022 and 2024). A swing is real change; a polling error is a failure to measure it. The two are related but not the same, and we have not fitted the regional parts on polling errors.

Well-polled statewide races are widened. For a statewide race polled by many firms, the national and regional parts together exceed the race's blended error, and its own error is held at the floor of 1.5 points. Michigan's Senate forecast has a standard deviation of 4.8 points but a simulated spread of about 6.2, so its chance in the simulation (76%) is lower than its blended error alone would give (about 81%). This makes the chances of well-polled races more cautious than the race pages' error figures suggest. Whether that is right depends on whether the regional sizes are right.

Normal tails. Every draw is normal. A national miss twice the size of 2020's (6.9 points in the final statewide polls) is very rare in the simulation.

No test of the probabilities. The backtests of 2018, 2022 and 2024 count winners called and average misses; three elections cannot show that races given a 70% chance are won about 70% of the time.

What would change our mind. Regional error sizes refitted on the statewide polling misses themselves, with the floor reconsidered; and a calibration table once enough cycles have been forecast with this model.

The House

Two elections. The district fit (intercept 1.75, slope 1.00, incumbency 2.29, residual 5.9) comes from 2022 and 2024 (From the generic ballot to 435 districts). It uses the 2020 presidential lean; the forecast applies it to a blend of 2024 and 2020.

New lines. In 173 districts in nine states the 2024 presidential result is an estimate carried by 2020 shares, and in Louisiana it rests on scaling to state totals.

Independent challengers. The House model takes the Democratic nominee as the challenger in every district. In Alaska's at-large district this year an independent, Bill Hill, polls between 31% and 48% against Nick Begich, while the Democratic nominee polls under 12%; the model reads the district's polls against the Democrat and puts Begich more than 20 points ahead. The statewide model already handles independent challengers (Nebraska, Idaho, South Dakota, Montana); the House model does not yet.

District polls. Three quarters of the 2026 district polls are partisan, mostly from the Democratic side, and many decisive districts have no independent poll. Halving partisan polls and moving them back by 2.52 points is the same rule as statewide, measured almost entirely on statewide races.

Specials left out. The eleven House special elections since January 2025 swung 17.8 points to the Democrats on average against the 2024 presidential margin. No special elections of earlier cycles are held, so their value as a predictor cannot be tested.

What would change our mind. A third election (2026) for the fit; the independent-challenger rule extended to House districts; a record of special elections from past cycles, which would let the specials be tested as an input.

The fundamentals

A wide residual. The statewide fits leave about 15 points of margin unexplained (The fundamentals: partisanship, the national vote and incumbency). In unpolled races the forecast is the prior alone, and in the most one-sided states it is probably generous to the underdog: an open governorship in Wyoming, R+45 on the presidential vote, has a Democratic chance of 15%.

One fit across 25 years. Senate results follow presidential lean more closely now than in 2000; one coefficient averages the change.

What would change our mind. A fit that lets the lean coefficient change over time, or that weights recent cycles more, tested on left-out cycles.

Pollsters and polls

House effects are relative. They measure a firm against its peers, so a lean every firm shares is invisible to them (Rating and adjusting American pollsters).

One partisan lean for both sides. Republican-side final polls leaned 2.84 points, Democratic-side 1.70; the model uses 2.52 for both.

Set, not fitted. The weights for registered voters (0.7) and all adults (0.5), and the extra discount for registered voters where an independent is the main challenger, are judgements.

Weighted generic ballot. The generic-ballot average is weighted and adjusted like race polls. Its tests favour weighting on national presidential polls (2.37 against 2.65 points) and not on generic-ballot polls (3.87 against 3.83) or on 2026's new polls (2.59 against 2.53).

Secondary sources. Most race polls and past final polls start from Wikipedia's tables, and only some can be upgraded from the pollster's own release.

What would change our mind. A longer record of partisan polls by side; a measured gap between likely and registered voters in past midterms.

The trend move

Every race poll is moved by the full change in the national trend line since its fieldwork (State polls: the pollster-adjusted average and the blend). That assumes a state moves as far as the nation. On the races of 2008 to 2024 the move made the election-eve averages more accurate overall (4.16 to 3.83 points) but slightly less accurate in 2012 and 2024, years in which states moved differently from the national polls. Today it moves the polls in use by 1.1 points towards the Democrats on average, because the national line has risen since August; if the line has overshot, every older poll inherits the overshoot.

What would change our mind. A share below one, or a share that varies with how far a state's swings have followed the nation's, beating the full move on the past races.

The readings

The "miss like" readings move every race by one number, a past cycle's mean statewide miss, though in each past cycle races missed by different amounts. The reading of "the polls as they stand" undoes our corrections approximately, by moving every race by half the generic-ballot correction and half the statewide one, rather than by rebuilding each race without them. Both are shown as readings, not forecasts, and neither is a probability.

Turnout

Nothing in the forecast models turnout directly. Polls of likely voters are preferred, registered-voter polls are discounted, and the national vote's uncertainty includes past misses that were partly turnout misses, but an unusual midterm electorate, larger or smaller than the polls' likely-voter screens assume, enters only through the error terms.

Candidates and timing

Nominee lists. The 2026 nominees come from Wikipedia's race lists, checked against the official list only in Colorado. A wrong nominee, a missed withdrawal or a wrong incumbent status changes a race's forecast directly.

Publication dates. Most poll rows hold the end of fieldwork, not the publication date, so the forecast history assumes a fixed delay.

The backtests

The backtests of 2018, 2022 and 2024 use a simpler blend than today's forecast: no house effects, no trend move, no firm count, and the uncorrected final generic ballot as the national vote, where today's forecast corrects it. The partisan lean they use is fitted on every year, the year tested included. They test the structure of the method, not each of its current parts. The 2018 House test, which has no presidential result by district, moves each district from its previous House margin and missed by 25 seats.

Election day and night

Drift is scaled linearly in the days left, from measurements at two distances.

The live model has not run on a real night. Its backtest replays the 2024 Senate races in a simulated count order. When early votes lean to one side, as they did from 2020 to 2024, the live model is less accurate than the pre-election forecast until about three quarters of the vote is in (From a month out to election night).

Seventeen states have readable official feeds; the rest of the country is read through the swing seen in them.

What is shown but not used

Early-vote party splits (Why early-vote party splits are not in the forecast), special-election swings, the president's approval and campaign finance are shown on the site and kept out of the forecast. In each case the reason is the same: the forecast already has polls, and no test on our data shows that the input makes it better.

The 2028 forecast

The 2028 model fails two of its own gates, both shown on its method page: it does not beat a dead-heat forecast at 18 months out, and its voter-group layer made the backtests slightly worse (The 2028 presidential forecast: a party contest with candidates on top). Its fundamentals rest on seven elections. Some 2028 pages and the midterm method page still describe the down-ballot projections' national vote as today's generic ballot pulled back towards 2024, the method of the projection withdrawn in September 2026, while the figures shown come from the party baseline.

Figures as of 15 October 2026. The pages they come from are rebuilt with each new poll.

How to cite

Polling Forecast (2026). What we don't know yet: limitations of the US forecast. Retrieved 17 October 2026 from https://www.pollingforecast.com/research/us-limitations. Live figures change as new polls arrive; the date above is when the text was last revised.