The final polls of the 2026 midterms understated the Democratic margin in every one of twelve Senate races examined here, by 4.3 points on average. That error sits inside the range recorded since 1998. What has grown is the volume of probabilities built on such numbers, which experiments show many readers take for predictions.
On the eve of the midterm elections of 3 November 2026, the FiftyPlusOne polling average for the Senate race in Texas put the Democrat, James Talarico, 3.7 points ahead of the Republican, Ken Paxton.1 He won by 12.1 points, according to the Associated Press count reported the following afternoon.2 Across 27 Senate and governor races, Pew Research Center found that the polls understated the Democratic margin by about 4 points.3
The thesis is this: American election polls are no less accurate than they were. Their 2026 error, about four points in one direction, falls inside the range recorded since 1998. What has grown is the supply of certainty: forecast models and prediction markets turn uncertain numbers into probabilities, which many readers, in experiments and surveys, take for predictions of the outcome or of its size.
Three tests can each pass or fail on available data.
- Error. The mean absolute error of the final averages in twelve 2026 Senate races lies within the range of the same measure for Senate races from 1998 to 2022, computed by Titanica at 3.6 to 6.5 points. The test fails if 2026 lies above that range.
- Misreading. Published experiments or surveys show readers misreading probabilistic forecasts in measurable ways. If none exist, the claim narrows to the polls alone.
- Markets. On eight 2026 Senate races, Kalshi prices on election evening are scored against a probability derived from the polls alone. If the markets score clearly better, the thesis must say so.
The 2026 miss
On the Kalshi exchange, a contract paying one dollar if Talarico won traded at 81 cents on election evening.4 The polling average named the winner and missed the margin by 8.4 points; the market price named the same winner and made no statement about the margin. The day after the vote, Newsweek published a list of the biggest polling misses of the cycle, and an analysis of prediction markets credited them with the stronger record at identifying winners.25
Exhibit 01 sets out the twelve Senate races in which the FiftyPlusOne site published a final average for a contest between a Democrat and a Republican and in which a winner had been called by 6 November.16 Alaska, where the count and the ranked-choice tabulation were incomplete, and Nebraska and Montana, where the leading challengers were independents, are excluded. The results are unofficial counts compiled by Ballotpedia from the New York Times and Decision Desk HQ, except in Texas, where the Associated Press count reported by Newsweek is used.72
| State | Final average | Result | Understatement of Democratic margin |
|---|---|---|---|
| Florida (special) | R+6.1 | R+2.9 | 3.2 |
| Georgia | D+9.6 | D+18.0 | 8.4 |
| Iowa | D+0.7 | D+4.1 | 3.4 |
| Kansas | D+4.5 | D+5.0 | 0.5 |
| Maine | D+4.7 | D+7.7 | 3.0 |
| Michigan | D+5.4 | D+7.0 | 1.6 |
| Minnesota | D+6.2 | D+10.1 | 3.9 |
| New Hampshire | D+8.8 | D+14.2 | 5.4 |
| North Carolina | D+11.6 | D+19.6 | 8.0 |
| Ohio (special) | D+2.9 | D+4.8 | 1.9 |
| South Carolina | Tie | D+4.4 | 4.4 |
| Texas | D+3.7 | D+12.1 | 8.4 |
| Mean (median) | 4.3 (3.7) |
The direction is uniform. In all twelve races the Democrat ran ahead of the final average, so the mean absolute error and the mean signed error coincide at 4.3 points. The size is not. In Kansas the average was within half a point of the result, and in Michigan and Ohio within two points. In Georgia, North Carolina and Texas it was off by eight points or more.
Pew’s wider set, which includes governors’ races and uses averages from RealClearPolitics, Silver Bulletin and FiftyPlusOne, gives the same picture: a mean understatement of about 4 points, a range from under half a point to about 9, and one exception, the governor’s race in Pennsylvania, where the polls overstated the Democratic margin by a little under 3 points.3 An analysis of 215 individual polls published in the final three weeks by the software firm Pushbuild found that 168 understated the Democrat and that the average poll was 4.1 points too Republican on the margin.8 For the House, Pew reported that the RealClearPolitics generic-ballot average showed a Democratic lead of nearly 9 points, against a lead in the counted House vote of about 12 points as of 5 November, which the New York Times expected to reach 13 once late-counting West Coast states finished.3
Measured by winners, the averages in Exhibit 01 did well. The candidate leading the final average won in all eleven races with a leader, and the twelfth, South Carolina, was tied. The miss lay in the margins, and the margins all moved the same way.
Pew described its own figures as preliminary, noted that votes were still being counted, and said it was too early to identify the source of the error.3 It listed possibilities for later analysis: Democratic supporters may have been more willing to take pre-election surveys; enthusiasm gaps were wide; and weighting by party or by past vote, along with likely-voter models, may not have captured the differences between the electorates of 2024 and 2026.3
A quarter-century of error
Whether 4.3 points is large depends on the record. Exhibit 02 sets the 2026 figure against the same measure for every even-numbered year from 1998 to 2022, computed by Titanica from the poll-level database that FiveThirtyEight (538) published on GitHub with its pollster ratings.9 For each Senate race between a Democrat and a Republican, the polls whose median field date fell within 21 days of the election are averaged without weights, and the average is compared with the actual margin recorded in the database. The database covers 382 such races and 2,557 polls.
| Year | Races | Mean absolute error | Mean signed error (+ = polls overstated the Democrat) | Races where polls understated the Democrat | Generic-ballot signed error |
|---|---|---|---|---|---|
| 1998 | 25 | 4.9 | −2.8 | 68% | +2.8 |
| 2000 | 30 | 4.5 | −2.5 | 60% | +2.3 |
| 2002 | 27 | 5.2 | +1.2 | 44% | +5.3 |
| 2004 | 26 | 4.6 | +2.4 | 35% | +4.6 |
| 2006 | 27 | 3.6 | −2.6 | 63% | +6.3 |
| 2008 | 34 | 4.5 | −0.5 | 53% | −1.9 |
| 2010 | 34 | 4.6 | −1.2 | 50% | −0.3 |
| 2012 | 28 | 3.6 | −3.3 | 93% | −0.8 |
| 2014 | 34 | 4.3 | +3.1 | 26% | +3.2 |
| 2016 | 33 | 5.9 | +5.2 | 9% | +3.6 |
| 2018 | 27 | 3.6 | +0.4 | 48% | −1.2 |
| 2020 | 29 | 6.5 | +6.4 | 10% | +4.9 |
| 2022 | 28 | 4.9 | +1.0 | 50% | +2.3 |
| 2026 | 12 | 4.3 | −4.3 | 100% | about −3 to −4 (provisional) |
The mean absolute error has moved between 3.6 and 6.5 points without a trend. The lowest value, 3.6 points, was recorded in 2006, 2012 and 2018; the highest, 5.9 and 6.5 points, in 2016 and 2020. The 2026 value of 4.3 points is below the average of the thirteen earlier cycles, 4.7 points. On size, the error test passes. The generic-ballot error of three to four points is also inside the range of the earlier cycles, which runs from 0.3 points in 2010 to 6.3 points in 2006.
The signed column carries the distinctive feature of 2026. The Senate polls also understated Democrats on average in six earlier cycles, by more than 2 points in 1998, 2000, 2006 and 2012, and in 2012 they did so in 93 per cent of races. The 2026 understatement of 4.3 points is larger than any earlier one in that direction. In the other direction it is smaller than the overstatements of 2016 and 2020, 5.2 and 6.4 points, when nine races in ten moved the same way. The share column shows the pattern that matters most for readers: in some years the errors cancel across races, and in others they move together.
The professional association of survey researchers reaches comparable conclusions with different methods. The task force of the American Association for Public Opinion Research on 2020 found that national presidential polls missed the margin by 4.5 points on average, the largest error in 40 years, that state presidential polls missed by 5.1 points, the same as in 2016, and that Senate and governor polls had an average absolute error of 6.7 points, the largest in at least 20 years.10 Its 2024 task force reported an average absolute error on the two-party margin of 3.3 points across offices, down from 5.3 points in 2020 and 5.2 in 2016. It found that polls had overstated Democratic margins by 2.7 points in 2024, by 4.6 in 2020 and by 3.1 in 2016, and that they overestimated Republicans by 0.6 points in 2022 and Democrats by 0.1 points in 2018.11
The longest comparative series points the same way. Will Jennings and Christopher Wlezien assembled 30,916 national polls from 351 general elections in 45 countries between 1942 and 2017 and found that “recent performance of polls has not been outside the ordinary”. Their mean absolute error on party and candidate vote shares was about 2 points over the whole period, and 2.0 points since 2000.12 Their measure is the error on each party’s share; an error on the margin between two parties is roughly twice as large, which is the scale used in Exhibits 01 and 02.
Exhibit 02 has limits. The historical rows average individual polls; the 2026 row uses FiftyPlusOne averages whose weighting is not described on the page consulted, and the 2026 results are unofficial. Twelve races are fewer than in any earlier cycle. A different selection of 2026 races would give a somewhat different figure, as the 27-race estimate of about 4 points from Pew shows. Neither choice moves 2026 outside the earlier range.
What a probability says
A polling average is a measurement with a known record of error. A forecast converts it into a probability. A simple illustrative conversion makes the step visible: if the error on a Senate average follows the spread seen in the 538 data from 1998 to 2022, whose root mean square is 6.3 points, then a lead of 3.7 points, the Texas figure, corresponds to a 72 per cent chance of winning for the candidate ahead.9 The same arithmetic gives the trailing candidate a 28 per cent chance, and puts the chance of a winning margin of 12 points or more, the Texas result, at about one in eleven. On these assumptions the Texas outcome was an unusual draw from the expected range of outcomes, though inside it.
The difficulty lies on the reader’s side of the number. Sean Westwood, Solomon Messing and Yphtach Lelkes tested it directly in experiments reported in the Journal of Politics in 2020.13 In the first, 4,151 members of Pew’s American Trends Panel saw a hypothetical Senate race described by a vote-share projection, a win probability, or both. Win probabilities made respondents about 7 per cent more certain, on the study’s scale, about who would win. Respondents shown an 87 per cent chance of victory estimated the leading candidate’s vote share at 64.6 per cent, against 56.5 per cent for those shown a 55 per cent vote share, and about 38 per cent of respondents gave the same number for vote share and for the likelihood of victory.13 Many read a probability of winning as a share of the vote, and so as a margin.
The second experiment, with 1,171 respondents from an online panel, turned this into a cost. Participants were given $15 and could pay $1 in each of five rounds to vote for their team, with $2 gained for a win and $2 lost for a defeat. As the displayed win probability moved away from 50:50, they voted less; the authors estimate that a win probability of 70 per cent, 20 points over even odds, lowered voting by 3.4 per cent, and a 40-point advantage by 6.9 per cent. The vote-share display had no detectable effect.13 The game is a laboratory model and the authors treat it as such; it measures how a displayed probability changes a costly choice, and it is not a measure of turnout in any election.
Surveys of how people judge a forecast after the event show a related pattern. In 2020 YouGov asked Britons to assess a hypothetical tennis model. When a player given a 49 per cent chance won, 42 per cent of respondents said the model had been right and 21 per cent that it had been wrong. When a player given a 29 per cent chance won, 29 per cent said the model had been right and 34 per cent that it had been wrong. When a player given a 1 per cent chance won, 49 per cent said the model was wrong.14 The 29 per cent case is the mirror of a 71 per cent favourite losing, and in that case more respondents called the model wrong than right. The British sample limits the inference for American readers, but the survey’s authors drew the general conclusion that perceptions of probability “are varied and complex”.14
The confusion predates elections. In a study of pedestrians in Amsterdam, Athens, Berlin, Milan and New York, Gerd Gigerenzer and colleagues found that only in New York did a majority give the meteorological reading of “a 30 per cent chance of rain”, that in three out of ten cases like tomorrow there will be at least a trace of rain; in the four European cities that reading was judged the least appropriate.15 Verbal probabilities are no cure. In a survey of 1,700 respondents reported by Andrew and Michael Mauboussin, most respondents placed “a real possibility” somewhere between about 20 and 80 per cent.16
| Study | Design | Finding |
|---|---|---|
| Westwood, Messing & Lelkes (2020) | Survey experiment, 4,151 members of Pew’s American Trends Panel | Shown an 87% win probability, respondents estimated a 64.6% vote share; shown a 55% vote share, 56.5% |
| Westwood, Messing & Lelkes (2020) | Same experiment | About 38% gave the same number for vote share and for likelihood of victory |
| Westwood, Messing & Lelkes (2020) | Incentivised voting game, 1,171 online-panel respondents | A displayed win probability of 70% rather than 50% lowered voting by 3.4%; vote share had no detectable effect |
| YouGov (2020) | Survey of Britons, hypothetical tennis model | After a 29% outsider won: 29% said the model was right, 34% wrong |
| YouGov (2020) | Same survey | After a 1% outsider won: 49% said the model was wrong, 18% right |
| Gigerenzer et al. (2005) | Pedestrians in five cities | Only in New York did a majority give the standard reading of “30% chance of rain” |
| Mauboussin & Mauboussin (2018) | Survey, 1,700 respondents, 23 words and phrases | “Real possibility” placed between about 20% and 80% by most respondents |
The misreading test passes. The evidence is experimental and partly non-American, and none of it measures how readers interpreted the 2026 forecasts. What 2026 supplies is a sample of how accuracy was reported. An analysis published by LiveNOW from FOX concluded on 4 November that the markets “had the stronger record at identifying winners”.5 Kalshi said on the morning after the vote that its favourites had won 479 races, about 98 per cent.1718 Both measures count winners, and a count of winners gives the same credit to a 95 per cent call in a safe seat and a 55 per cent call in a close one.
The markets’ record in 2026
The third test needs a comparison on equal terms. Exhibit 04 takes the eight Senate races for which DeFi Rate, an industry publication, recorded a Kalshi price between 6:49 and 7:03 p.m. Eastern time on election day and which also appear in Exhibit 01. Alaska, where no winner had been declared, Nebraska, where no Democrat was on the ballot, and Louisiana and Mississippi, which lack a FiftyPlusOne average, are excluded.4 For Florida, where only the Republican price was shown, the Democratic price is taken as its complement. Each market price is set against the illustrative poll-based probability described above. The comparison uses the Brier score, the mean squared difference between the stated probability and the outcome, scored 1 if the Democrat won and 0 otherwise; lower is better, and a forecaster who always said 50 per cent would score 0.25.
| State | Final average | Poll-based chance, illustrative | Kalshi price | Winner | Brier, polls | Brier, market |
|---|---|---|---|---|---|---|
| Florida | R+6.1 | 17% | 14% | R | 0.028 | 0.020 |
| Iowa | D+0.7 | 54% | 63% | D | 0.208 | 0.137 |
| Kansas | D+4.5 | 76% | 74% | D | 0.057 | 0.068 |
| Maine | D+4.7 | 77% | 75% | D | 0.053 | 0.063 |
| Michigan | D+5.4 | 80% | 86% | D | 0.039 | 0.020 |
| Ohio | D+2.9 | 68% | 81% | D | 0.105 | 0.036 |
| South Carolina | Tie | 50% | 39% | D | 0.250 | 0.372 |
| Texas | D+3.7 | 72% | 81% | D | 0.078 | 0.036 |
| Mean | 0.102 | 0.094 |
The market scored 0.094 and the poll-based probability 0.102, a difference of 8 per cent. That is a small advantage on a small set, and Titanica does not treat it as clearly better; the test passes, narrowly. The direction of the result belongs in the thesis all the same: on this set the market did slightly better. Its gains came mainly from Texas, Ohio, Iowa and Michigan, where traders priced the Democrat above the polls and the Democrat won by more than the polls showed. Its losses came from South Carolina, where the Republican traded at 62 cents against a tied average and the Democrat won by 4.4 points, and to a lesser degree from Kansas and Maine. Both methods named the winner in seven races; in South Carolina the market chose the loser and the polls declined to choose.
Two features of the comparison favour the market. The prices were recorded on election evening, around the first poll closings in the East, so they could reflect information that no final poll contained, from early-vote reports to the news of the day. And eight races is a small sample: had Iowa, the closest race in the set, gone the other way, the poll-based probability would have scored 0.113 and the market 0.126, reversing the ranking. A broader review by DeFi Rate, scoring prices observed at midnight entering election day, found that Kalshi favourites won 489 of 507 scored events, 96.4 per cent, and that favourites on Polymarket US won 408 of 417, 97.8 per cent.18 Its list of misses includes Greg Abbott in the Texas governor’s race, priced at 60 per cent, and the South Carolina Senate race, priced at 67 per cent for the Republican on Kalshi.18 Kalshi’s own research paper reports that its prices “matched or exceeded polling averages at every horizon where they existed”, with a Brier score of 0.016 at poll close.19 That paper scores the races called by 6:26 a.m. Eastern time on 4 November, a different set observed at different times, and its figures are not comparable with those in Exhibit 04.
The case for the markets
The strongest counter-argument runs as follows. The polls missed in the same direction across most races in 2016, in 2020 and again in 2026, and readers who discount them are responding to that record. Markets combine the polls with early-vote data, money and local knowledge, and in 2026 they priced Democrats above the polls in four of the eight races in Exhibit 04 and were rewarded. A reader who followed the markets would have been less surprised by Texas.
Part of this holds. The errors of 2016, 2020 and 2026 were correlated across races, and a reader who treats each race as an independent measurement will underestimate the chance that all of them miss together. The market did slightly better in 2026, and Kalshi’s research reports a similar pattern across horizons. The rest does not hold on the evidence assembled here. The size of the 2026 error lies inside the range of the past quarter-century. Its direction has changed from cycle to cycle: the Senate polls understated Democrats by more than 2 points on average in 1998, 2000, 2006, 2012 and 2026 and overstated them by more than 2 points in 2004, 2014, 2016 and 2020, so a reader could not have known in advance which way to correct. The market’s advantage on the defined set was 8 per cent of the Brier score, with one race, South Carolina, going the other way. And the market states its view as a probability, so it carries the reading problem of Exhibit 03 with it: a contract at 81 cents is a 19 per cent statement that the favourite loses.
The certainty business
The growth of election markets in the United States followed a court decision. On 12 September 2024 Judge Jia Cobb of the federal district court in Washington ruled that the Commodity Futures Trading Commission had exceeded its statutory authority when it blocked Kalshi from listing contracts on control of Congress.20 The commission appealed, then moved to drop the appeal in May 2025.21 Kalshi’s chief executive, Tarek Mansour, said: “Today is historic.”21
The volumes that followed are reported by the companies themselves or by analysts working from public market data, and they measure different things. Exhibit 05 sets them out with their nature.
| Date | Item | Figure | Nature |
|---|---|---|---|
| 12 Sep 2024 | District court rules the CFTC exceeded its authority in blocking Kalshi’s congressional contracts | – | Court decision (press report) |
| Nov 2024 | Polymarket, 2024 presidential-winner market | $3.64bn | Platform data (press report) |
| 2024 cycle | Kalshi, election contracts | $1.3bn | Company statement |
| May 2025 | CFTC moves to drop its appeal | – | Court filing (press report) |
| 22 Jul 2026 | Kalshi launches a midterm portal combining polls and market prices | – | Company announcement (press report) |
| to 29 Oct 2026 | Texas Senate race, wagers | nearly $67m | Analyst estimate (ACDC) |
| to 2 Nov 2026 | Kalshi, all midterm contracts | $1.33bn | Company figure |
| to 2 Nov 2026 | Polymarket, congressional markets | about $145m | Analyst estimate (ACDC) |
| by 9 Nov 2026 | Kalshi, all midterm contracts | more than $2.5bn | Company statement |
A single Polymarket contract on the 2024 presidential winner traded $3.64 billion.22 Kalshi says its election contracts traded $1.3 billion in the 2024 cycle and more than $2.5 billion on the 2026 midterms, nearly double.17 Reuters reported $1.33 billion on Kalshi’s midterm markets by 2 November, about $145 million on Polymarket’s congressional markets according to the Anti-Corruption Data Collective, a nonprofit, and nearly $67 million on the Texas Senate race alone by 29 October.23 In July Kalshi launched a portal combining polls, fundraising data and its own prices, which it said would make it the primary reference for US election odds.24
Probabilities were already prominent in election coverage before the markets grew. According to a figure in the paper by Westwood, Messing and Lelkes, cable news mentioned election probabilities about 16 times a day in 2016; the authors also counted more than 9,800 Google News articles using the language of probabilistic forecasting between 1 August and 7 November that year, and 281,661 tweets linking to major probabilistic forecasts against 28,416 linking to the main collections of polls.13 The markets add a price to that form, and the companies that run them publish their records as counts of winners.
The record in this Dispatch does not establish that the growth in probabilities caused any change in what readers believe. It establishes the coincidence of three things: a polling error of ordinary size, election markets that by Kalshi’s account traded more than in any earlier cycle, and experimental evidence that such probabilities are often read as margins or as promises.
What the numbers can carry
The Texas average of 3.7 points was a measurement, and its error of 8.4 points was large by the standard of a single race and ordinary by the standard of a cycle in which all twelve Senate errors ran the same way. The polls did what the record of 1998 to 2022 says polls do. Under the illustrative conversion they gave the Democrat a 72 per cent chance; the market gave 81 per cent. Both numbers were statements about uncertainty, and the outcome confirmed neither and refuted neither. In South Carolina the market’s 39 per cent for the eventual winner was the same kind of statement, and on that evening it was the one that looked wrong.
The House vote showed the same thing in different units. The generic ballot missed by three to four points, inside the range of earlier years, and in the same direction as the Senate polls. Pew’s caution, that the sources of the error cannot yet be identified, applies to the explanations now in circulation as much as to the error itself.3 The 2026 figures will change as states certify their counts, and the comparison in Exhibit 02 should be recomputed then. On the evidence available on 9 November, the polls erred by the amount they usually err; the probabilities built on them were read, by many, as something more.
- FiftyPlusOne, “Polling averages” (Senate and governor averages for 2026; the page is undated and does not describe its weighting) (analyst estimate). https://www.fiftyplusone.news/averages (accessed 9 November 2026) ↩︎
- Sam Stevenson, “Biggest polling misses of the 2026 midterm elections”, Newsweek, 4 November 2026 (press report; RealClearPolling final averages compared with Associated Press counts as of 2 p.m. ET on 4 November; Texas Senate: Talarico 55.1%, Paxton 43.0%). https://www.newsweek.com/biggest-polling-misses-2026-midterm-elections-12666486 (accessed 9 November 2026) ↩︎
- Scott Keeter, “A First Look at Polling Performance in 2026”, Pew Research Center, 6 November 2026 (analyst estimate; 27 Senate and governor races, averages from RealClearPolitics, Silver Bulletin and FiftyPlusOne, unofficial counts as of 4 November). https://www.pewresearch.org/short-reads/2026/11/06/a-first-look-at-polling-performance-in-2026/ (accessed 9 November 2026) ↩︎
- DeFi Rate, “2026 Midterm Election Odds: Kalshi vs Polymarket”, Kalshi Senate prices observed 3 November 2026 between 6:49 and 7:03 p.m. ET (press report of exchange prices; live page, content as consulted). https://defirate.com/odds/2026-midterms/ (accessed 9 November 2026) ↩︎
- Chris Williams, “Prediction markets vs. polls: Who got the election right?”, LiveNOW from FOX, 4 November 2026 (press report). https://www.livenowfox.com/news/prediction-markets-vs-polls-election-2026 (accessed 9 November 2026) ↩︎
- Ballotpedia, “United States Senate elections, 2026”, status as of 4 p.m. ET, 6 November 2026: seven seats changed party, two races uncalled (reference compilation). https://ballotpedia.org/United_States_Senate_elections,_2026 (accessed 9 November 2026) ↩︎
- Ballotpedia race pages for the 2026 Senate elections (reference compilation of unofficial counts reported by the New York Times and Decision Desk HQ; totals may change until certification): Florida https://ballotpedia.org/United_States_Senate_special_election_in_Florida,_2026 ; Georgia https://ballotpedia.org/U.S._Senate,_Georgia_election,_2026 ; Iowa https://ballotpedia.org/U.S._Senate,_Iowa_election,_2026 ; Kansas https://ballotpedia.org/U.S._Senate,_Kansas_election,_2026 ; Maine (first round of ranked-choice count) https://ballotpedia.org/U.S._Senate,_Maine_election,_2026 ; Michigan https://ballotpedia.org/U.S._Senate,_Michigan_election,_2026 ; Minnesota https://ballotpedia.org/U.S._Senate,_Minnesota_election,_2026 ; New Hampshire https://ballotpedia.org/U.S._Senate,_New_Hampshire_election,_2026 ; North Carolina https://ballotpedia.org/U.S._Senate,_North_Carolina_election,_2026 ; Ohio https://ballotpedia.org/United_States_Senate_special_election_in_Ohio,_2026 ; South Carolina https://ballotpedia.org/U.S._Senate,_South_Carolina_election,_2026 (accessed 9 November 2026) ↩︎
- Pushbuild, “How accurate were the 2026 midterm polls?”, 7 November 2026 (analyst estimate; polls ending in the final three weeks in 31 Senate and governor races, against unofficial MIT Election Data and Science Lab results). https://pushbuild.com/apis/analysis/2026-polls-vs-results/ (accessed 9 November 2026) ↩︎
- FiveThirtyEight (538), pollster-ratings data, raw_polls.csv, GitHub repository fivethirtyeight/data (dataset; general-election Senate polls 1998–2022 used). https://github.com/fivethirtyeight/data/tree/master/pollster-ratings (accessed 9 November 2026) ↩︎
- American Association for Public Opinion Research, Task Force on 2020 Pre-Election Polling (chair Josh Clinton), “An Evaluation of the 2020 General Election Polls” (professional-body report; undated PDF). https://aapor.org/wp-content/uploads/2022/11/AAPOR-Task-Force-on-2020-Pre-Election-Polling_Report-FNL.pdf (accessed 9 November 2026) ↩︎
- American Association for Public Opinion Research, Task Force on 2024 Pre-Election Polling, Executive Summary (professional-body report; undated PDF). https://aapor.org/wp-content/uploads/2025/10/Task-Force-on-2024-Pre-Election-Polling-Report-Executive-Summary.pdf (accessed 9 November 2026) ↩︎
- Will Jennings and Christopher Wlezien, “Election polling errors across time and space”, Nature Human Behaviour 2 (4), 2018, doi:10.1038/s41562-018-0315-6 (peer-reviewed; author manuscript consulted). https://eprints.soton.ac.uk/421260/1/JenningsWlezien_PollingErrors.pdf (accessed 9 November 2026) ↩︎
- Sean Jeremy Westwood, Solomon Messing and Yphtach Lelkes, “Projecting Confidence: How the Probabilistic Horse Race Confuses and Demobilizes the Public”, The Journal of Politics 82 (4), October 2020, pp. 1531–1544 (peer-reviewed). https://www.journals.uchicago.edu/doi/10.1086/708682 ; text consulted: https://www.asc.upenn.edu/sites/default/files/2021-03/Projecting%20Confidence-%20How%20the%20Probabilistic%20Horse%20Race%20Confuses%20and%20Demobilizes%20the%20Public.pdf (accessed 9 November 2026) ↩︎
- Matthew Smith, “How likely is likely?”, YouGov, 1 October 2020 (survey report). https://yougov.com/en-gb/articles/32255-how-likely-likely (accessed 9 November 2026) ↩︎
- Gerd Gigerenzer, Ralph Hertwig, Eva van den Broek, Barbara Fasolo and Konstantinos V. Katsikopoulos, “‘A 30% chance of rain tomorrow’: how does the public understand probabilistic weather forecasts?”, Risk Analysis 25 (3), 2005, pp. 623–629 (peer-reviewed). https://researchonline.lse.ac.uk/id/eprint/16975 (accessed 9 November 2026) ↩︎
- Andrew Mauboussin and Michael J. Mauboussin, “If You Say Something Is ‘Likely’, How Likely Do People Think It Is?”, Harvard Business Review, July 2018 (commentary reporting a survey of 1,700 respondents). https://hbr.org/2018/07/if-you-say-something-is-likely-how-likely-do-people-think-it-is ; text consulted in a republication: https://theyouthtimes.com/news-details/2549/2018-10-14 (accessed 9 November 2026) ↩︎
- Enna Lee, “Kalshi Says 2026 Midterm Contract Trading Tops $2.5 Billion”, TokenPost, 9 November 2026 (press report of a company statement). https://www.tokenpost.com/news/investing/52265 (accessed 9 November 2026) ↩︎
- Cheryle Shepstone, “Election Market Review Finds Kalshi Favorites Won 96.4% of Scored Races”, DeFi Rate, 5 November 2026, updated 6 November 2026 (press report and analysis; prices at or before midnight ET entering 3 November). https://defirate.com/news/election-review-kalshi-favorites-won/ (accessed 9 November 2026) ↩︎
- Kalshi Research, “Midterms calibration and accuracy”, November 2026 (company research). https://kalshi.com/research/publications/midterms-calibration-and-accuracy (accessed 9 November 2026) ↩︎
- Sarah Wynn, “Judge rules CFTC exceeded authority in blocking Kalshi’s election contracts”, The Block, 12 September 2024 (press report of a court decision, U.S. District Court for the District of Columbia). https://www.theblock.co/post/316176/judge-rules-cftc-exceeded-authority-in-blocking-kalshis-election-contracts (accessed 9 November 2026) ↩︎
- Cheyenne Ligon, “CFTC Drops Appeal in Kalshi Election Betting Case”, CoinDesk, May 2025 (press report of a court filing). https://hotfix.coindesk.com/policy/2025/05/06/cftc-drops-appeal-in-kalshi-election-betting-case (accessed 9 November 2026) ↩︎
- Ryan S. Gladwin, “Polymarket Resolves U.S. Presidential Election Market After $3.6 Billion Volume”, Decrypt, 6 November 2024 (press report of platform data; $3.64 billion). https://decrypt.co/290374/polymarket-donald-trump-us-presidential-election (accessed 9 November 2026) ↩︎
- Douglas Gillison and Michelle Price, “Prediction markets face midterms test as election bets soar”, Reuters, 3 November 2026, as published by the Honolulu Star-Advertiser (press report; Kalshi figure and Anti-Corruption Data Collective estimates). https://www.staradvertiser.com/2026/11/03/breaking-news/prediction-markets-face-midterms-test-as-election-bets-soar/ (accessed 9 November 2026) ↩︎
- Lydia Beyoud, “Kalshi Moves to Go Big on US Election Bets Ahead of Midterms”, Bloomberg Law, 22 July 2026 (press report). https://news.bloomberglaw.com/securities-law/kalshi-moves-to-go-big-on-us-election-bets-ahead-of-midterms (accessed 9 November 2026) ↩︎
- Context
- The midterm elections of 3 November 2026, in which the final polling averages understated the Democratic margin in every Senate race examined, and the first assessments of polls and prediction markets published in the following week by Pew Research Center, Newsweek, Kalshi and DeFi Rate.
- Causes
- The size of polling error, measured on Senate races from 1998 to 2022 with FiveThirtyEight data; the shared direction of error within a cycle, which Pew provisionally links to differences in survey response, enthusiasm and weighting; the conversion of polling margins into probabilities by forecast models and markets; the growth of US election markets after the 2024 court ruling for Kalshi.
- Actors
- FiftyPlusOne, RealClearPolitics, Silver Bulletin and FiveThirtyEight; Pew Research Center and the American Association for Public Opinion Research; Kalshi, Polymarket and the Commodity Futures Trading Commission; Judge Jia Cobb; Sean Westwood, Solomon Messing and Yphtach Lelkes; Will Jennings and Christopher Wlezien; Gerd Gigerenzer and colleagues; YouGov; the Anti-Corruption Data Collective.
- Perspectives
- Separates the size of polling error from its direction; compares 2026 with the historical record on one defined measure and labels where the comparison is not strict; scores market prices and an illustrative poll-based probability on the same races with the Brier score; treats experimental evidence on misreading as evidence about readers in general, not about 2026; makes no partisan reading of the results.
- Relationships
- Opens the Mathesis Domain. Links to Demos (Thirty-Two Percent) and Dry Powder (AI's Money Loop Concentrates Risk on Few Balance Sheets).
