Research
How the forecast works, in full
The 13 papers behind our United States forecast (26,732 words): how far the polls have missed, how we correct them, the seat model and the tests.
These papers set out every part of our method: what we measure, how we correct the polls, how votes become seats, how uncertain the result is, and what we have tested and rejected. Figures marked Live are computed from the database each time the page is built, so they always match the forecast.
Also: France · Germany · Spain · Italy · Netherlands · Poland · Canada · Australia · Great Britain
13 papers · about 27k words · Short summary of the method · How we adjust today's polls · Data sources
United States
- Forecasting the 2026 midterms: an overview The United States section forecasts the midterms of 3 November 2026: all 435 House seats, 35 Senate seats and 36 governorships. It applies the rule used for every country: correct the polls only where predicting each past election from the others shows the correction helps. The final generic ballot has overstated the Democrats at 11 of 13 House elections, so our national House vote is D+6.6 where the polls say D+8.7; state polls get a narrow half-correction of 0.7 points. This paper sets out the pipeline, today's projection, where the uncertainty lies and how the other twelve papers fit together.
- Measuring American polling error, 2000–2024 Our record holds 4,688 final polls of 806 Senate, governor and presidential races by state since 2000. The average of a race's final polls misses the result by about 4.6 points on the margin, and much of the miss is shared: in 2020 the final Senate and governor polls overstated the Democrats by 6.9 points on average. The miss has leaned towards the Democrats in 8 of 12 cycles, most of all in the most Republican states since 2018. We define the error, give a worked example, split shared from single-poll error and show how each number enters the model.
- The generic ballot: the national vote behind every race The generic ballot sets the national environment that every House district, Senate race and governor's race moves with. We average it like race polls: house effects and partisan lean taken off, weighted by recency, sample and record. Today nine polls by seven pollsters give D+8.7. Because the final generic ballot overstated the Democrats at 11 of the 13 House elections since 2000, and halving that miss passed the leave-one-out test at 11 of them, our national House vote is D+6.6, with a standard deviation of 5.2 points. The paper also covers the trend line and how the national vote reaches each race.
- Rating and adjusting American pollsters We do two separate things with American pollsters. For readers, every firm with a final poll on record is rated out of 100 on accuracy, its miss on the margin and openness; 502 are graded. For the model, three measured adjustments: a house effect for each firm against its peers, a record weight from its typical final-poll miss, and the treatment of partisan and sponsored polls, which leaned 2.52 points to their side across 556 final polls and are moved back by that amount and counted at half weight. The paper also explains the split of Hart Research and Public Opinion Strategies into three codes.
- State polls: the pollster-adjusted average and the blend How the polls of one Senate race, governor's race or House district become one number with an error. Each poll's margin on the decided vote loses its pollster's house effect, the 2.52-point partisan lean where it applies and the 0.73-point statewide correction, and is moved with the national trend since its fieldwork, which made past election-eve averages more accurate (4.16 to 3.83 points). Polls are weighted by recency, sample, record and the people sampled, and the average's error is counted in polling firms, not polls. The average is then blended with the fundamentals by their measured errors.
- The fundamentals: partisanship, the national vote and incumbency 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.
- From the generic ballot to 435 districts How the national House vote becomes a forecast in each district. Each district starts from its presidential lean on the 2026 lines, moves point for point with the national vote and gains 2.3 points for an incumbent standing again, coefficients fitted on 732 contested districts in 2022 and 2024. Nineteen seats are settled before polling day, and district polls, mostly partisan, are blended in at reduced weight. The paper covers the nine redrawn states, special elections (shown, not used), the tipping point and the backtests, in which 2024 was replayed with 384 of 396 contested districts called right.
- Correlated error and the backtests of 2018, 2022 and 2024 Polls miss together. Each of 5,000 simulated elections draws a national error shared by every race, errors shared by a census division, a state and a demographic profile, sized from how far states' presidential swings strayed from the nation's since 2000, and each race's own error. This paper sets out the draws, how Senate control is counted with independents, today's results, the readings that replay past cycles' misses on every race, and how the method did when replayed on 2018, 2022 and 2024 with every parameter fitted without the year tested: 170 of 180 statewide races called right.
- From a month out to election night How far American polls move before polling day, and what the site does when the count begins. Statewide averages taken 30 to 45 days out missed by about 6.9 points on the margin, against 6.1 two weeks out; the 3.2 points of drift are added to every race's error in proportion to the days left. On election night the site reads 17 states' official results feeds and a live model turns each count into a chance. Replayed on the 2024 Senate races, it was well calibrated with a neutral count order but less accurate than the forecast early in the count when early votes leaned to one side.
- Why early-vote party splits are not in the forecast Early-vote figures show who has voted, by registered party, not how they voted. In the 12 state races of 2022 and 2024 for which we hold final early-vote splits, reading the split as the result missed by 14.7 points on average; correcting each state's 2024 figure by its 2022 gap still missed by 7.5, against 2.7 for the final polls. A wider 75-race test found no state-level signal that survives a shuffle test. The paper sets out the data held, the mechanisms that separate ballots from votes, the 2026 figures so far and what early votes are good for.
- The 2028 presidential forecast: a party contest with candidates on top The 2028 forecast is a contest between a generic Democrat and a generic Republican, with named candidates added as small adjustments that shrink with time. The national level blends the governing party's record, which since 2000 has followed the president's June approval, with the polls, each weighted by how well it did at the same distance from past elections. Replayed on 2000 to 2024, it called 331 of 357 state results two years out and missed the national margin by 3.0 points, against 5.9 for repeating the last result. It fails two of its seven gates, both shown.
- The US data: sources, grades and checks Where every number in the US section comes from: results from the Federal Election Commission, the Clerk of the House, the MIT Election Data and Science Lab and J. Cha's governors file; presidential results by district built from precinct returns, estimated in the nine redrawn states; census data; candidates; 8,654 polls by 782 firms, each graded A, B or C by source; early-vote and election-night feeds from the states' own offices; and the inputs of the 2028 model. The paper also lists the checks the data pass, the corrections made and what is never used.
- What we don't know yet: limitations of the US forecast The known weaknesses of the US forecast in one place. The corrections rest on a short record, and the statewide one passes its test by 0.06 points. The polls' larger misses in the most Republican states are not corrected for. The simulation's regional errors are sized on presidential swings and widen well-polled statewide races. The House model rests on two elections, applies to 173 districts on new lines and does not yet handle independent challengers. The backtests count winners, not probabilities. For each weakness we say how much it matters and what evidence would change our approach.