Skip to content
Outcome
Sign in
Explore notebooks

Kalshi’s 2026 Senate Forecasts: Accuracy, Revision, and Reversals

A 2026 midterm prediction-market audit using Outcome trade data: what 97% winner accuracy misses about forecast revision and election-night reversals.

8views 0

Plain text

This retrospective examines Kalshi’s forecasts for the 2026 U.S. Senate elections: which outcomes the markets favored, how their probabilities changed during the campaign, and whether election-night confidence held up. Later sections compare vote-margin estimates, chamber-control forecasts, and other forecasting systems.

Stop South Carolina’s election-night chart at the right moment and Kalshi looks like it saw the upset coming. Move forward one point and it favors the losing side.

The points are five-minute trade averages. The Democratic contract averaged 67% in the 7:40 p.m. Eastern bucket and 48% in the next. By 8:23 p.m., a trade had reached 90 cents. The Democrat eventually won. There is a successful prediction in that chart, provided you choose when to look.

The main analysis follows 35 Kalshi contracts, each asking whether a Democrat would win a particular Senate race. We construct the scorecard from Outcome’s latest executed YES price strictly before November 3 at 4:17 p.m. Eastern, matched to AP’s calls available by November 5. A price above 50 cents favors a Democratic win; below 50 cents favors a non-Democratic win, which can include an independent. These are last observed trades, not synchronized bid/ask quotes.

Thirty-three of those races had been called at the results cutoff. Alaska and Mississippi had not, so they remain in the price histories but receive no outcome score. Among the 33 called races, the favored outcome won 32: 97% directional accuracy. South Carolina was the only miss because the scorecard uses the afternoon snapshot, when Kalshi still favored the Republican, before the election-night reversals above.

The same Outcome trade data lets us reconstruct how those forecasts developed. South Carolina supplied nearly a third of the sample’s reduction in probability error over the final month while remaining the underdog at both monthly checkpoints. Maine repeatedly retreated after reaching high confidence. The winner tally is the starting point; the rest of the analysis follows the price changes that it leaves out.

Reported results are frozen to November 5, 2026, and settlement records were checked November 6. The 33 scored contracts resolved; Alaska and Mississippi remained unresolved at those cutoffs. Times in the text are Eastern; interactive charts use the viewer’s local time. AP call chronology; outstanding races.

Where Kalshi’s Senate Forecast Error Accumulated

Winner accuracy reduces each probability to a binary prediction at 50%. To retain information about confidence, we also use the Brier score: the mean squared difference between a forecast probability and the observed outcome, coded as 1 for a Democratic win and 0 otherwise. Lower scores indicate less error. The Outcome last-trade snapshot scored 0.036392.

We first divide the 33 contracts by their Democratic probability. Twenty-four were below 10% or above 90%; all 24 favored sides won. The remaining nine were between 10% and 90%, inclusive. Eight of those nine favored sides won, but this group generated 98.3% of the total squared error.

Democratic probability

Contracts

Favored side correct

Share of squared error

Below 10% or above 90%

24

24 / 24

1.7%

10%–90%, inclusive

9

8 / 9

98.3%

All scored contracts

33

32 / 33

100%

This grouping describes the price of a Democratic win. It does not independently measure electoral competitiveness. Nebraska illustrates the distinction: a strong independent candidate can make a race competitive while the Democratic contract remains near zero.

With all 24 tail favorites winning, their squared errors were necessarily small. The individual-race decomposition identifies where the remaining error sits. South Carolina contributed 39.6% of total error and Iowa another 20.0%. Together, a directional miss and a directional success accounted for 59.6% of the probability score.

Two races account for 59.6% of Senate forecast error

Chart configuration unavailable. Cached query result data is shown below.

Races ranked by squared error

Cumulative error (%)

0 races

0

1 · SC

39.64394557596549

2 · IA

59.63661798258031

3 · KS

69.83696104717971

4 · ME

77.3310906456609

5 · OH

83.85931020700451

6 · TX

89.48823421653039

7 · MI

94.28447715955836

8 · LA

96.69092544173732

9 · FL

98.32298033207323

10 · MN

99.07447499458759

11 · NH

99.35458890535747

12 · AR

99.5232068213233

13 · KY

99.6252102519693

14 · GA

99.70523090245976

15 · NC

99.76593335220744

16 · AL

99.81389578163773

17 · TN

99.85419754525621

18 · OK

99.87551418055857

19 · NJ

99.89183472946193

Showing the first 20 of 34 rows.

Cached: 2026-11-08T16:56:53.02Z.

The 33 resolved contracts are ordered from largest to smallest squared error; each point adds one race to the cumulative total. South Carolina alone contributes 39.6%, and adding Iowa brings the share to 59.6%. Inputs are the latest Outcome trades strictly before November 3, 4:17 p.m. Eastern. Alaska and Mississippi are unscored.

For Iowa, the squared error was (1 − 0.51)² = 0.2401. For South Carolina, it was (1 − 0.31)² = 0.4761. The winner tally assigns the first forecast full credit and the second none, while the probability score preserves the difference in confidence.

This does not imply that a 31% forecast is disproved by a win. Events assigned 31% probability should occur approximately 31 times in 100 across a calibrated set of forecasts. One election can identify where realized error occurred; it cannot establish persistent miscalibration or partisan bias. The 97% result describes directional performance within this particular sample.

When the Favorites Changed

Kansas’s Adam Hamilton moved from 17% in early September to 40% on October 28 and 63% at noon on November 2. The Democratic outcome had become the favorite before Election Day, following a 46-percentage-point revision. Kansas history; November 2 odds.

Texas moved earlier. James Talarico was priced at 54% on September 15, 67% on October 7, and 75% on November 2. South Carolina remained below 50% in the observations used here: the exchange publication put it around 25% on October 28, and its latest Outcome trade before our election-day cutoff was 31%. The three Democratic victories therefore represent different forecasting histories: an early favorite, a late crossover, and an outcome that remained a long shot at the final pre-results observation. Early Texas price; later Texas observations; South Carolina odds.

The Texas governor’s race provides a similar example outside the Senate sample. Hinojosa rose from roughly 8% in mid-July to around 40% on November 2 and then won. The price registered a large increase in the probability of a Democratic victory while continuing to favor the Republican. Governor observations.

Daily volume-weighted trade prices show the complete 35-contract sample below. Twenty lines end at verified YES payouts of 100% and 13 at NO payouts of 0%. Alaska and Mississippi receive no forced endpoint. UTC-day bucket starts are displayed in local time; payout points are separate from trade averages. The legend can isolate a race.

All 35 Senate races · Daily Democratic YES price (%) through settlement

All 35 matched Democratic Senate contracts · daily trade VWAP, Sep 1–Nov 4 in UTC-day buckets, displayed in local time. Final 0% / 100% points are verified payouts; Alaska and Mississippi remain unresolved.

Cached source data for the selected chart columns, before chart aggregation, bucketing, normalization, top-N, zoom, and plot ordering. Values retain their source precision.

Time bucket

Alabama

Alaska

Arkansas

Colorado

Delaware

Florida

Georgia

Idaho

Illinois

Iowa

Kansas

Kentucky

Louisiana

Maine

Massachusetts

Michigan

Minnesota

Mississippi

Montana

2026-09-01T00:00:00.000000Z

4.3999999999999995

64.20495362342616

7.664093132459401

—

—

8.410959493884762

92.94061225147423

1.4

—

39.952752707581226

18.946660184866346

6.24647355163728

8.245965608465609

69.27534619696677

—

63.71080835490424

91.62542425907544

6.615131609042919

1.5

2026-09-02T00:00:00.000000Z

4.381292676082686

63.54111940182105

7.466996081640361

—

—

9.028190851074314

93.36630093730638

—

—

39.64178087681334

18.18931402450392

5.868307498952661

5.225461904761905

68.43109461193426

98.10000000000001

63.804061472304895

92.24965999009817

6.542381489841987

1.5441783783783785

2026-09-03T00:00:00.000000Z

—

63.93717437000348

6.817172675521823

—

96.39999999999999

9.093082645249412

93.51397253815782

1.4

—

37.1496571758122

17.33466804369563

4.7219904458598725

6.515892420537898

68.59344084311226

—

63.95332159948798

93.09882729501497

6.6358625322631974

—

2026-09-04T00:00:00.000000Z

1.5

63.465886158184276

6.324014750950177

96.41377859965914

—

8.726951772596614

92.30233032551969

—

—

36.719658949867785

17.636481365995824

4.348860194484297

7.611242360322354

68.19725298162402

98.1

63.43759695188585

92.30554806293799

6.2314364453899325

1.676626217944886

2026-09-05T00:00:00.000000Z

—

63.62237777345693

6.469585253456222

96.99999999999999

—

8.94492186910244

92.43524343284152

—

98

35.90065344999446

17.034597286531582

5.8999999999999995

6.062019350037212

68.86418768097104

98.2

62.72613838370026

92.09715607014465

6.043199999999999

1.3622

2026-09-06T00:00:00.000000Z

4.356988052236733

64.68342093993967

6.868086890711611

96.5

—

8.90792014069004

92.72218313171413

—

98

35.714464214572864

17.000000000000007

5.10015993602559

7.886662337953116

68.31034594265981

—

63.17949664259841

91.80310277957336

6.394018915180742

1.4

2026-09-07T00:00:00.000000Z

—

64.75513698630137

6.899300894854586

96.43071871127634

—

8.90801600927234

93.00844254906042

—

97.1

36.06640260535032

17.645825298985695

6.000372737946586

4.565748621498175

68.75592138910547

98.19999999999999

63.98401320901323

92.5549873050417

6.256534055918216

1.2444444444444447

2026-09-08T00:00:00.000000Z

1.5

64.80581178352439

6.947846910785041

96.39999999999999

96.4

9.01185500253936

93.34663341645886

—

—

36.50042825726301

20.227752145582578

5.454107869663559

6.61190669156771

67.62632269248492

97.30285714285714

63.419123676600954

90.84660278819379

6.3986522002830375

1.6983317823827144

2026-09-09T00:00:00.000000Z

1.7413751611516972

64.99433374256301

6.87222972972973

96.89999999999999

98.9

8.900001236797191

93.25076215087263

—

97.1

37.02319033911394

20.640532091172336

4.615312464397992

4.1000000000000005

66.4806521318453

96.95954660847104

62.90341698065797

90.20293987012491

6.447362842578104

1.3222222222222222

2026-09-10T00:00:00.000000Z

—

64.96921656148992

7.153089495826676

96.9

—

8.977260709403167

92.46993540416048

—

97.06546363131592

35.9997691348957

21.091967220794004

4.7

7.6306304917527354

67.23346633801347

96.94708328335531

63.97325826919793

90.08360921578198

6.587656944176164

—

2026-09-11T00:00:00.000000Z

—

65.02690099292579

6.158678912045214

96.95319333658908

—

8.968508036238214

93.08582939041757

—

—

38.42383962786653

20.9502671252642

5.726768975200728

7.53708516727819

66.95126230012124

96.83297920341278

62.77184241587601

90.3765989940252

6.627117070985085

—

2026-09-12T00:00:00.000000Z

—

64.43290514322713

6.70307917888563

—

—

8.782548005604086

92.99362830623213

—

—

39.27084117468858

20.349053123968453

5.348543290931473

7.041161291140685

66.3980915845973

—

62.58821804334879

90.2396442405276

7.253209943210174

0.9795226925797685

2026-09-13T00:00:00.000000Z

—

64.86519385894263

7.217325086301557

—

96.82941780821916

8.871266517489108

94.05593449327314

1.2000000000000002

—

38.4138597389301

21.086655689060095

6.010682110682111

—

66.74782325605146

—

61.222603906011784

90.00468332541101

6.240594362745098

1

2026-09-14T00:00:00.000000Z

4.2862068965517235

65.68059906889587

7.260122636897

95.8

96.61795387773172

9.58180171600165

93.34578886882286

—

97.67139194139196

39.26397013371509

21.146635638320763

6.399014179608373

7.299999999999999

66.04743278018749

97.1

60.98215104281498

89.56912376113635

7.758474155069583

0.9943980004254412

2026-09-15T00:00:00.000000Z

—

64.59246295032342

6.432490240363762

97

97.35199999999999

9.408647687288084

93.18390715475856

—

97.3

38.77091181746467

22.84134797801489

5.921165592127586

7.255918367346941

65.46495675696087

—

61.66651912522817

89.76618685336214

7.135709718670076

1.6982513661202188

2026-09-16T00:00:00.000000Z

4.379350264740196

65.99694845712054

6.871624936164545

96.99999999999999

98.39938347718866

9.44175287456986

93.90467071803454

—

98.9

38.90657235269279

23.351361178979214

5.7221945560522895

—

65.98475543592066

97.2402376910017

62.98665920046007

90.54013311182213

7.4577788982730375

0.6994854202401372

2026-09-17T00:00:00.000000Z

4

68.6377069125547

6.9796385105612035

97.0056363180188

98.3

9.225832569429864

94.41368437443232

0.8859374999999999

99

39.806219427474645

22.979886871387375

6.272800684528583

—

67.78203687841508

—

64.2838854777719

90.81847608106081

8.2981591471301

0.5715678194440044

2026-09-18T00:00:00.000000Z

—

68.99020520103825

6.79240020519446

—

98.264

10.981786151489622

94.14130966952266

0.6

97.20077900334843

40.00119833642095

24.875596082580426

6.077029278743157

6.753579935547148

67.79455130007686

97.55201592832256

64.04736814869024

90.24713177145117

7.502689905862616

1.4427737352239063

2026-09-19T00:00:00.000000Z

—

69.81226389016862

6.841311196937785

—

98.5

12.365193022917577

93.85727978617557

0.2

98.00846994535517

40.052443511075694

26.955138983313482

5.6761171032357485

6.938419243986255

67.59102691924221

—

64.9916025223725

90.12986295052117

8.195571427957859

—

2026-09-20T00:00:00.000000Z

3.9

70.932106785171

6.878221062914389

—

—

11.74504434308327

93.54239621866009

—

97.3

40.992688531643616

26.96312533882867

6.080498564526294

—

67.82175905056187

98.00709706546274

65.64775945841501

90.29997364768717

8.286618776641639

1.4531122104650513

Showing the first 20 of 98 rows. Showing the first 20 of 36 columns.

Cached: 2026-11-08T16:56:53.72Z.

We examine Iowa, Kansas, Maine, Ohio, South Carolina, and Texas more closely below. These six contracts were selected for their repricing and all settled YES. Comparing them with the complete sample above separates the timing of successful revisions from the behavior of losing and unresolved contracts.

Case studies · Six Senate races through settlement

Six selected Senate races · daily trade VWAP in UTC-day buckets, September 1–November 4. Final 100% points are verified YES settlement payouts, not trades.

Cached source data for the selected chart columns, before chart aggregation, bucketing, normalization, top-N, zoom, and plot ordering. Values retain their source precision.

Time (UTC)

Race

Democrat (%)

2026-09-01 00:00:00

Iowa

39.952752707581226

2026-09-01 00:00:00

Kansas

18.946660184866346

2026-09-01 00:00:00

Maine

69.27534619696677

2026-09-01 00:00:00

Ohio

53.79310256747808

2026-09-01 00:00:00

South Carolina

13.217095026822255

2026-09-01 00:00:00

Texas

50.26089825409275

2026-09-02 00:00:00

Iowa

39.64178087681334

2026-09-02 00:00:00

Kansas

18.18931402450392

2026-09-02 00:00:00

Maine

68.43109461193426

2026-09-02 00:00:00

Ohio

53.00665518342362

2026-09-02 00:00:00

South Carolina

13.892467380110746

2026-09-02 00:00:00

Texas

49.13311573765942

2026-09-03 00:00:00

Iowa

37.1496571758122

2026-09-03 00:00:00

Kansas

17.33466804369563

2026-09-03 00:00:00

Maine

68.59344084311226

2026-09-03 00:00:00

Ohio

52.99874583705813

2026-09-03 00:00:00

South Carolina

13.630682132806593

2026-09-03 00:00:00

Texas

48.02959701907447

2026-09-04 00:00:00

Iowa

36.719658949867785

2026-09-04 00:00:00

Kansas

17.636481365995824

Showing the first 20 of 396 rows.

Cached: 2026-11-08T16:56:53.202Z.

The dated publication snapshots provide a separate view of the campaign. Each mark is a reported observation, not a daily trade average. TX, KS, and SC denote Senate races; TXG denotes governor. Same-day a/b labels distinguish observations without asserting their intraday order. Exchange observations.

How the favorites changed: dated observations

Discrete published snapshots. TX/KS/SC = Senate; TXG = governor. Same-day a/b are separate source observations.

Cached source data for the selected chart columns, before chart aggregation, bucketing, normalization, top-N, zoom, and plot ordering. Values retain their source precision.

observation

Democrat (%)

TX·9/15

54

TX·10/7

67

TX·10/28

66

TX·10/31

71

TX·11/2a

75

TX·11/2b

76

KS·9/4

17

KS·10/14

37

KS·10/28

40

KS·10/30

51

KS·11/1

53

KS·11/2

63

TXG·10/28

36

TXG·11/2a

38

TXG·11/2b

40

TXG·11/3

34

SC·10/28

25

SC·11/2

33

Cached: 2026-11-08T16:56:53.099Z.

Learning Without Changing the Favorite

Which revisions produced the improvement? We hold the same 33 resolved contracts fixed and reconstruct 721 hourly forecasts from the latest trade strictly before each cutoff. The final point is the same Outcome snapshot used in the scorecard above. Over the final month, directional accuracy rose from 30 of 33 to 32 of 33, while mean Brier error fell 41.45%, from 0.06216 to 0.03639.

We decompose this change by subtracting each contract’s final squared error from its initial squared error. Kansas’s probability rose from 34% on October 4 to 65% on November 3, accounting for 36.8% of the net improvement. South Carolina rose from 14% to 31%, accounting for 31.0%. The two contracts supplied 67.8% of the improvement across the entire cohort.

South Carolina improved the score while remaining the underdog at both monthly checkpoints. Moving from 14% to 31% reduced the squared error attached to the eventual Democratic outcome from 0.7396 to 0.4761, even though the Republican remained favored at both endpoints. We use learning here to mean this retrospective reduction in realized error. The decomposition measures which revisions improved the score; identifying what information drove them requires separate evidence.

The hourly series also shows that improvement was uneven. Its lowest sampled Brier score, 0.03385, occurred at 3:17 a.m. Eastern on election day. Error rose to 0.03845 by 1:17 p.m. before declining to 0.03639 at the 4:17 p.m. cutoff. The final observation was 1.79% worse than the same time on the previous day. Most of the month’s progress had already occurred before the final afternoon.

How the Senate forecast learned · hourly

721 hourly snapshots · same 33 Senate contracts. Source window: October 4–November 3, 2026, with UTC cutoffs. Last-trade Brier score; lower is better.

Cached source data for the selected chart columns, before chart aggregation, bucketing, normalization, top-N, zoom, and plot ordering. Values retain their source precision.

horizon

Brier

2026-10-04 21:17

0.06215527272727272

2026-10-04 22:17

0.06257212121212122

2026-10-04 23:17

0.06261260606060605

2026-10-05 00:17

0.06323330303030303

2026-10-05 01:17

0.06235175757575757

2026-10-05 02:17

0.062466878787878787

2026-10-05 03:17

0.06306687878787878

2026-10-05 04:17

0.0638541212121212

2026-10-05 05:17

0.06337839393939393

2026-10-05 06:17

0.06330542424242423

2026-10-05 07:17

0.06279575757575756

2026-10-05 08:17

0.06262239393939394

2026-10-05 09:17

0.06365269696969696

2026-10-05 10:17

0.06288603030303029

2026-10-05 11:17

0.06325681818181815

2026-10-05 12:17

0.06342506060606058

2026-10-05 13:17

0.06349896969696968

2026-10-05 14:17

0.06341766666666665

2026-10-05 15:17

0.0628289696969697

2026-10-05 16:17

0.062337606060606066

Showing the first 20 of 721 rows.

Cached: 2026-11-08T16:56:53.193Z.

721 hourly observations of the same 33 resolved contracts. Each uses the latest executed price strictly before its cutoff; the series ends before election results. Its final Brier score is 0.036392, the same snapshot used in the primary scorecard. Alaska and Mississippi are excluded throughout.

Election-Night Reversals

A first threshold crossing measures when a price reaches a specified level. It does not measure how long that confidence persists. South Carolina’s first observed trade at or above 50 cents after 6 p.m. occurred at 7:34:21 p.m. Eastern. Its first trade at or above 90 cents followed at 8:23:04 p.m., less than 49 minutes later.

Within that interval, the five-minute volume-weighted price fell from 67.27% in the 7:40 p.m. bucket to 47.52% in the next. The decline was 19.76 percentage points using unrounded values. Those two buckets contained 1,704 trades and 246,570 contracts. The reversal was large enough to change the favored side between consecutive five-minute averages.

Louisiana exhibited a similar reversal on a contract that eventually settled NO. Its Democratic price reached a five-minute average of 34.3% at 8:50 p.m., then fell to 9.7% in the 9:15 p.m. bucket. The 24.6-point decline spanned six buckets containing 932 trades. Louisiana never became a Democratic favorite under this measure, but the losing contract experienced a substantial temporary rally.

The five-minute histories below cover all 35 contracts. Each trade point is the volume-weighted executed price within a five-minute bucket, plotted at the bucket’s start. Verified settlement payouts follow the trade observations. Lines connect observed samples; missing buckets contain no inferred price. Louisiana, Mississippi, and Alaska provide cases beyond the six selected winning contracts.

All 35 Senate races · Five-minute Democratic YES price (%) through settlement

All 35 matched Democratic Senate contracts · five-minute trade VWAP. Local times. Final 0% / 100% points are verified payouts, not trades; Alaska and Mississippi remain unresolved.

Cached source data for the selected chart columns, before chart aggregation, bucketing, normalization, top-N, zoom, and plot ordering. Values retain their source precision.

Time bucket

Alabama

Alaska

Arkansas

Colorado

Delaware

Florida

Georgia

Idaho

Illinois

Iowa

Kansas

Kentucky

Louisiana

Maine

Massachusetts

Michigan

Minnesota

Mississippi

Montana

2026-11-03T14:00:00.000000Z

—

69.33770414201184

—

—

—

13

96.7

—

—

52.052003715114886

62.88697838244786

—

—

72.34531612639954

—

79.95963460918311

90.80000000000003

14.154644351464437

—

2026-11-03T14:05:00.000000Z

—

—

4.7

—

—

13

96.37676192539185

0.7000000000000001

—

52.14416251046634

63.000000000000014

3.9982908904460777

14

72.12625049792292

—

79.3696526988914

90.75250023498448

14.065640474145884

0.9806404150695011

2026-11-03T14:10:00.000000Z

—

69.93051297772327

4.8712883093160135

—

—

12.002589763277417

96.69999999999997

—

98.80666666666664

52.68300978464837

62.769146293162436

4

—

72.60587623504941

—

79.65116803463516

90.8

14.994446090047393

—

2026-11-03T14:15:00.000000Z

2.864610488954164

70

—

—

—

12.627419605751493

96.69999999999999

—

—

51.85282110610233

62.200855528367995

—

14.000000000000004

72.34370699693024

—

79.73802563525976

90.71066525248955

14.358134605193426

—

2026-11-03T14:20:00.000000Z

—

69.07227583433854

4.839735132685432

98.8

99

13

96.6

—

98.8

52.15165544110922

63

—

14.000000000000004

72.25618518518519

—

79.04656854825684

90.7945569344655

14.656941709921844

0.9965033575714228

2026-11-03T14:25:00.000000Z

—

69.8999785515121

4.157272312406284

—

—

12.437034281976633

96.23428472145436

—

—

51.67234699238246

62.98609008144343

4

13.000000000000002

72.26004660479946

99.19589860713482

79.47179517703478

90.79697081640191

14.880295483798758

—

2026-11-03T14:30:00.000000Z

—

69.14272034727836

4.508659921745406

—

—

12.70124646644786

96.6

0.7000000000000001

—

51.46756762563652

62.970418732382264

4

13.000000000000002

72.2803252074092

—

79.91265273397391

90.80000000000001

14.003073886619898

0.8

2026-11-03T14:35:00.000000Z

—

69.96178599049827

3.700503313701464

—

—

12.890710679189965

96.6

—

98.8

51.99603677550444

62.83853795253653

3.9296369265409505

12.758739255014325

72.99865696474032

99.2

79.86402586957317

90.80000000000001

14.782740940297211

1

2026-11-03T14:40:00.000000Z

—

70

—

—

—

13

96.60000000000001

—

—

51.992967822242576

62.880167909884335

3.9834881320949433

—

72.67879901750287

99.2

79.42149551330388

90.70401144318281

13.999999999999998

—

2026-11-03T14:45:00.000000Z

—

69.85907553551297

—

98.79999999999998

—

12.057172598159436

96.50132211281745

—

98.79999999999998

51.75440679666854

62.99999999999999

4.20173227739561

12.999999999999998

71.7256793638186

—

79.14167889527177

90.76003224432682

14.000000000000004

—

2026-11-03T14:50:00.000000Z

2.5999999999999996

70.00000000000001

3.9745222929936306

—

—

11.843032219583831

96.5401363510246

—

—

51.882043397631975

62.56105025773196

3.999999999999999

13.000000000000002

72.7598329946221

99.2

78.99379590081506

90.79999999999998

14.000000000000002

—

2026-11-03T14:55:00.000000Z

—

70

3

—

—

11.595525872348524

96.50051664465215

—

98.30000000000001

51.39003056170303

62.90763492948087

4

13.000000000000002

72.81878497923812

98.97711732084208

78.89908178408459

90.70056829452137

14.000000000000002

0.8991538274404145

2026-11-03T15:00:00.000000Z

—

69.43843137254902

3.8007468744926127

—

—

11.992130944237479

96.49999999999999

0.7000000000000001

—

51.98810601944093

62.26983345009153

3.9965852825678674

13

71.98063455400879

—

78.73025566886655

90.71820933997247

13.774294387829952

0.8999999999999999

2026-11-03T15:05:00.000000Z

—

69.41999999999999

3.8999999999999995

—

—

11.068537941674245

96.50000000000001

—

—

51.87598562152133

62.635524738022006

4

13.000000000000002

71.49189411040425

—

78.21776130523284

90.7905191323422

13.997546534070466

—

2026-11-03T15:10:00.000000Z

—

70

3.9566294919454776

—

—

10.874673050707846

96.49999999999997

—

—

51.602751715983594

62.97215190334153

4.000000000000001

12.999999999999998

72

—

78.47062050528213

90.54180401069091

14

—

2026-11-03T15:15:00.000000Z

2.5999999999999996

69.24143202327531

—

—

—

10.942448476610018

—

0.7

—

51.955048472654106

63.00000000000001

3.927700329892519

13

72

—

77.22988782547931

90.45621826118587

14.000000000000002

—

2026-11-03T15:20:00.000000Z

2.5999999999999996

70.00000000000001

3.9

—

—

10.980160522631822

96.2061064185441

—

98.79999999999998

51.94673302358133

62.82476862788694

—

12.024449877750616

71.70723591366225

—

78.3905593596946

90.397987052647

14

0.8

2026-11-03T15:25:00.000000Z

—

70

3.44149115726132

—

—

10.93829522034777

96.60000000000001

0.7

—

51.11405576176919

62.89724743729731

4

12.974161542039173

71.62493145539905

—

78.94855388322914

90.19999999999999

13.950526766199072

—

2026-11-03T15:30:00.000000Z

—

69.2410270422065

3.9166398349458413

98.8

—

10.066287932350797

96.60000000000001

—

—

50.71087672781862

62.43826007278184

4.005833015630957

12.772813328284741

71.32735588242326

99

78.96834134007791

90.35333083130043

13.711855305230781

0.8975785775329977

2026-11-03T15:35:00.000000Z

—

69.9959260164589

4.177502818943299

98.79999999999998

—

9.585025204683806

96.58125826912475

—

—

50.0737705084326

62.13239481588548

—

12.870744679636351

71.44653887598841

99

78.73096059680952

90.39825348865641

14.000000000000004

0.8488532110091743

Showing the first 20 of 279 rows. Showing the first 20 of 36 columns.

Cached: 2026-11-08T16:56:53.733Z.

Unresolved Contracts

Mississippi was the third-most-traded Senate contract in our election-night window, behind Texas and Michigan. It does not appear in the scorecard because it remained unresolved. It recorded 8.93 million contracts, or 12.5% of the sampled total, from November 3 at 9 a.m. Eastern to November 4 at 5:30 a.m.

Mississippi’s five-minute average rose from roughly 14% in the opening morning bucket to 82% in the final 5:25 a.m. bucket. The opening bucket contains only four trades, so we also checked the endpoints using 30-minute averages. They retain the rounded 14%-to-82% change. The contract remained active and unresolved in the November 6 audit and therefore contributes no outcome score.

Alaska came close to certainty and backed away. Its five-minute average reached 98.1% at 1:30 a.m. before declining to 86.1% by the end of the sampled window. It also remained unresolved at the audit cutoff. Even a price near 100% left room for a substantial reversal.

Volume here counts executed contracts. It does not measure dollars, unique traders, or available liquidity. Resolved contracts stop contributing at settlement, making the Mississippi ranking a comparison of window totals rather than trading intensity over equal durations.

The selected close-up contains the six Senate case studies and both chamber-control contracts. All eight settled YES. Their 100% endpoints are verified payouts; the preceding observations are five-minute trade averages.

Case studies · Election night: six races and chamber control

Six selected Senate races plus both chamber-control contracts · five-minute trade VWAP. Source window: November 3–4 UTC. Final 100% points are verified YES settlement payouts, not trades.

Cached source data for the selected chart columns, before chart aggregation, bucketing, normalization, top-N, zoom, and plot ordering. Values retain their source precision.

Time (UTC)

Race

Democrat (%)

2026-11-03 14:00:00

House control

93.89615110562907

2026-11-03 14:00:00

Senate control

76.64999706348742

2026-11-03 14:00:00

Iowa

52.052003715114886

2026-11-03 14:00:00

Kansas

62.88697838244786

2026-11-03 14:00:00

Maine

72.34531612639954

2026-11-03 14:00:00

Ohio

71.99999999999999

2026-11-03 14:00:00

South Carolina

31.262844046283853

2026-11-03 14:00:00

Texas

73.07093143508361

2026-11-03 14:05:00

House control

93.88976581518348

2026-11-03 14:05:00

Senate control

76.63434846191343

2026-11-03 14:05:00

Iowa

52.14416251046634

2026-11-03 14:05:00

Kansas

63.000000000000014

2026-11-03 14:05:00

Maine

72.12625049792292

2026-11-03 14:05:00

Ohio

71.69290042642326

2026-11-03 14:05:00

South Carolina

30.81308143571231

2026-11-03 14:05:00

Texas

72.11501546200273

2026-11-03 14:10:00

House control

93.8077496397591

2026-11-03 14:10:00

Senate control

76.97613013954185

2026-11-03 14:10:00

Iowa

52.68300978464837

2026-11-03 14:10:00

Kansas

62.769146293162436

Showing the first 20 of 1416 rows.

Cached: 2026-11-08T16:56:53.489Z.

Iowa’s five-minute average fell from 87.48% at 8:45 p.m. to 72.78% at 9:05 p.m. Maine raises a further question: after a price reaches 90%, how often does it fall back?

How Long Did High Confidence Take to Persist?

We compare the first observed five-minute average at or above 90% with the start of the final observed run at or above 90%. For Maine, those times were 8:45 p.m. and 9:55 p.m.: a 70-minute gap. Six intervening averages fell below 90%, across three retreats, with a low of 87.77%. All 15 five-minute buckets spanning the two endpoints contain trades.

Maine’s delay remains 70 minutes at an 89% threshold and becomes 100 minutes at 91%. Ohio’s corresponding 90% delay is 25 minutes, but its lowest later average was only 89.75%, so the result is sensitive to the exact threshold. Iowa, Kansas, South Carolina, and Texas had no later sub-90% average after their first qualifying bucket.

The final run is identified retrospectively from observed bucket averages. It is not a rule a trader could have recognized in real time, and averages above 90% do not imply every trade or quote stayed above 90%. In Maine, the first crossing still left more than an hour of retreats ahead.

Markets and Newsroom Calls

Kalshi’s recap reports House control reaching 99% at 8:10 p.m. and Senate control passing 99% at 9 p.m. AP called Senate control at 11:36 p.m. and House control at 1:24 a.m. the next morning. The reported lead times were 2 hours 36 minutes for Senate control and 5 hours 14 minutes for House control. Exchange account; AP chronology.

Those gaps compare a market threshold with a newsroom decision. Our persistence check illustrates one reason the standards differ; it does not adjust the reported lead times, which use different contracts and a 99% threshold. A first crossing establishes when a price reached a level. It leaves open how stable that reading was and how much of the relevant information was already public.

When markets reached 99%—and AP called

Minutes after 18:00 ET on November 3. Reported market thresholds and AP calls use different standards.

Cached source data for the selected chart columns, before chart aggregation, bucketing, normalization, top-N, zoom, and plot ordering. Values retain their source precision.

chamber

Market 99%

AP call

House

130

444

Senate

180

336

Cached: 2026-11-08T16:56:53.325Z.

Minutes after 6 p.m. Eastern on November 3. Exchange-reported 99% thresholds and AP calls use different standards. Outcome’s first observed 99-cent trade times differ slightly from the recap; the appendix reconciles the measurements.

What Was the Market Forecasting?

Predicting who wins Texas is a different task from predicting the margin. Kalshi got the Senate winner right. Its separate margin market put the Democrat much closer to the Republican than the provisional count did.

Winner Contracts and Separate Margin Markets

Kalshi News reported a +2.70-point Democratic margin estimate from its margin market on November 3. The provisional DDHQ count recorded the following evening was +12.09, a 9.39-point gap. The margin estimate comes from a different market; we do not derive it from the binary winner price. Reported margin-market estimate; provisional result.

The Texas governor’s October 28 margin-market estimate was −2.70 points against a provisional +7.12 result. Kansas Senate’s estimate on that date was −1.30 against +5.00. All three publisher-reported estimates fell below the provisional Democratic margins. Their dates differ and counts were unfinished, so they are selected comparisons rather than a matched error ranking. We retain the publisher’s reported estimates without assigning an undocumented mean or median interpretation. Governor estimate and count; Kansas estimate and count.

Relative to DDHQ’s 2024 presidential margins, the provisional Senate results shifted 25.78 points toward Democrats in Texas and 21.12 in Kansas. These between-election changes establish the scale of the results. Differences in candidates, offices, and turnout prevent interpreting them as individual voter switching.

Separate margin-market estimates vs provisional results

D−R percentage points: dated market forecasts versus provisional results. Forecast dates differ.

Cached source data for the selected chart columns, before chart aggregation, bucketing, normalization, top-N, zoom, and plot ordering. Values retain their source precision.

race_and_forecast_date

Market forecast

Provisional result

TXSen

2.7

12.09

TXGov

-2.7

7.12

KSSen

-1.3

5

Cached: 2026-11-08T16:56:53.392Z.

Separate, publisher-reported margin-market estimates versus provisional vote margins, in Democratic-minus-Republican percentage points. These estimates are not transformations of the binary winner probabilities. Forecast dates differ; exact timestamps are in the appendix.

The Probability of a Wave

Add the individual race probabilities and you get the expected number of wins. That calculation works whether the races move together or independently. Across the 33 scored contracts, Democratic probabilities summed to 17.582 wins; Democrats won 20. The probability of winning at least 20 depends on which outcomes tend to occur together.

A set of individual probabilities does not determine a unique probability of reaching a seat threshold. We use the same Outcome snapshot to illustrate how a shared influence across races can change the distribution while leaving every Democratic-win probability fixed.

The illustrative proxy adds 34 fixed Democratic-caucus holdover seats to 35 simulated Democratic-win contract outcomes. Each contract uses its unadjusted Outcome last-trade probability, including Alaska and Mississippi; no Republican-price normalization is applied. A new independent winner could add a Democratic-caucus seat without settling a Democratic-win contract YES. Because that contribution is omitted, the proxy is not a calibrated forecast of caucus membership or Senate control.

With independent races, the proxy reaches 51 or more seats with 88.7% probability and 54 or more with 24.6%. At an illustrative latent correlation of 0.4, those probabilities become 73.8% and 38.2%. The expected proxy seat count remains 52.41. Greater dependence moves more probability toward both a disappointing result and a larger victory; a lower chance of clearing 51 can coexist with a higher chance of reaching 54.

Latent correlation describes the correlation between the model’s underlying hypothetical race influences, not the correlation of the binary wins themselves. The value 0.4 is an example, not an estimate from this election or a fit to a traded control price. The model specification is in the appendix.

Direct wave contracts specify joint thresholds rather than an expected total. A November 3 publication priced at least 235 House seats together with at least 51 Senate seats at 68%, and at least 240 House seats together with at least 54 Senate seats at 32%. The second event is nested within the first. Neither probability follows from adding individual race odds alone: both depend on the joint distribution across chambers. These contracts use a February 1, 2027 reference date and were not settled results at the November 5 cutoff. Definitions; published prices.

Illustrative seat-count proxy: 51+ vs 54+ seats

Chart configuration unavailable. Cached query result data is shown below.

Assumed latent correlation

51+ proxy seats (%)

54+ proxy seats (%)

ρ = 0

88.70956648745512

24.60894709809771

ρ = 0.02

86.98915034800723

26.24800057331902

ρ = 0.04

85.48363018721328

27.626307509456772

ρ = 0.06

84.15905215491792

28.807345461081717

ρ = 0.08

82.98642864276168

29.835176655291384

ρ = 0.1

81.94186391643889

30.74142698066841

ρ = 0.12

81.00587447273887

31.549449719234147

ρ = 0.14

80.16258461219138

32.276914912315576

ρ = 0.16

79.39900531084629

32.93747874168378

ρ = 0.18

78.70444230464662

33.54189412189052

ρ = 0.2

78.07002675072664

34.09877016475591

ρ = 0.22

77.48834795106644

34.615104307423486

ρ = 0.24

76.9531668545

35.09666331179165

ρ = 0.26

76.45919215502141

35.54826141520169

ρ = 0.28

76.00190453008567

35.973967015294974

ρ = 0.3

75.57741787669397

36.37725876659149

ρ = 0.32

75.18236907129834

36.76114527294111

ρ = 0.34

74.81382983809203

37.128258200441685

ρ = 0.36

74.46923586826847

37.48092573938271

ρ = 0.38

74.14632950312081

37.821231386015576

Showing the first 20 of 41 rows.

Cached: 2026-11-08T16:56:53.468Z.

Illustrative seat-count proxy using all 35 Outcome Democratic-win probabilities, with expected proxy seats fixed at 52.41. The 51+ and 54+ thresholds concern this proxy, not a calibrated caucus-control forecast. New independent winners are omitted. Latent correlation varies from 0 to 0.8 and is never estimated from this election.

Why the Forecaster Ranking Changes

The leaderboard changes when the question changes. FiftyPlusOne’s provisional assessment gave Kalshi and Polymarket identical House-race Brier scores of 0.026, compared with 0.030 for Silver Bulletin and 0.034 for FiftyPlusOne. On the two chamber-control outcomes, the models had assigned greater confidence to Democrats and scored better. Published assessment.

The reversal follows from evaluating individual House races and chamber outcomes separately. A system can score better across district-level forecasts while placing less probability on the eventual national result. This published comparison uses market forecasts from November 2 at 11:59 p.m. and some projected outcomes as of November 4 at 7:53 p.m.; its race universes vary across metrics and differ from our Senate cohort.

Before declaring one forecaster better, check that the comparison uses the same races, the same observation time, the same outcome definition, and the same scoring rule.

The leaderboard depends on the question

Rank 1 is best; stronger color means a worse rank. Markets tie on House-race Brier. Provisional published comparison.

Cached source data for the selected chart columns, before chart aggregation, bucketing, normalization, top-N, zoom, and plot ordering. Values retain their source precision.

metric

50+1

Silver

Polymarket

Kalshi

House ctrl

1

2

3

4

Senate ctrl

2

1

3

4

House Brier

4

3

1

1

Cached: 2026-11-08T16:56:53.427Z.

Rank 1 is best; stronger color indicates a worse rank. Markets tie on House-race Brier in this provisional publisher comparison. Raw values are retained in the query results.

Discussion

Twenty-four contracts were already outside the 10–90% probability band and all favored the correct outcome. The unresolved uncertainty was concentrated in nine contracts, which generated 98.3% of realized squared error. Iowa’s nominally correct call contributed one-fifth of that error because its price was still close to a coin flip.

Revision was concentrated as well. Kansas and South Carolina supplied 67.8% of the net monthly reduction in Brier error. Election night introduced another distinction: a price could cross a high-confidence threshold and then retreat. In Maine, 70 minutes separated the first 90% five-minute average from the beginning of the final observed run above 90%. Mississippi’s substantial trading remained outside outcome scoring altogether.

Change the races, the timestamp, or the outcome being predicted and the verdict can change. The markets’ district-level House scores and their chamber-control scores even reverse the ranking against the models. A claim about forecast accuracy needs to say which of these questions it answers.

One election cannot establish long-run calibration or persistent superiority over other forecasts, and trade averages cannot establish returns after spreads and fees. South Carolina remained the only losing favorite call in our frozen Senate sample while supplying 31% of the month’s improvement in realized probability error. Counting it only as a miss discards a substantial part of what changed before the result.

Methodological Appendix

The primary Senate scorecard, campaign and election-night trade histories, and dependence illustration use Outcome trade records. Dated exchange publications, provisional election counts, and the publisher’s forecaster comparison remain separately attributed. Reported election results and provisional margins are frozen to November 5, 2026; contractual settlement timestamps were checked against Kalshi’s market API on November 6. Reported calls, exchange resolutions, and certified results are different things. Alaska and Mississippi remain unresolved in this audit, and House totals remain unfinished.

The scorecard and its universe

The core scorecard takes each contract’s latest executed YES price from Outcome strictly before November 3, 4:17 p.m. Eastern (21:17 UTC), matched to the frozen AP outcome labels. It scores Democratic versus non-Democratic outcomes in the 33 called Senate races. Identical trade timestamps use the retained deterministic trade-ID tie break. The median last-trade age is 118 seconds; the oldest observation is Rhode Island’s at 27 minutes 9 seconds. These are last-trade probability proxies, not synchronized order-book quotes. Brier score averages (probability − outcome)²; log loss penalizes confident mistakes more sharply. Both are lower-is-better scores. Highly favored binary events can dominate aggregate winner accuracy, and unresolved exclusions can change the result.

Frozen 33-race scorecard

Value

Democratic wins

20

Sum of Democratic probabilities

17.582

Wins above the probability sum

+2.418

Democratic/non-Democratic favorite accuracy

32 / 33 — 97.0%

Mean Brier score

0.036392

Mean log loss

0.132597 nats

The 35 matched histories include 20 settled YES contracts, 13 settled NO contracts, and Alaska and Mississippi unresolved. The learning curve uses the same fixed 33-contract resolved cohort at every cutoff. The six case-study races are Iowa, Kansas, Maine, Ohio, South Carolina, and Texas; they were selected for repricing and all settled YES. The election-night case-study view adds the two YES-settled chamber-control contracts. This is a Kalshi reconstruction; the Polymarket and model comparison is a separately attributed publication.

How the trade charts were built

Campaign points are contract-volume-weighted executed YES prices in UTC calendar-day buckets from September 1 through November 4. Election-night points use five-minute buckets in the November 3, 14:00 UTC to November 4, 10:30 UTC source window. Points are placed at bucket starts; axes and tooltips convert those instants to the viewer’s local time. Lines connect observed samples and do not establish continuous quotes. Missing trade buckets are not filled with invented prices.

The full campaign view contains 1,895 observed daily buckets plus 33 settlement endpoints; the full election-night view contains 4,352 observed five-minute buckets plus 33 endpoints. The six-race campaign close-up has 390 trade buckets plus six payouts. The election-night close-up has 1,408 trade buckets plus eight payouts. Each resolved contract’s trading observations stop before its verified settlement instant, and the terminal 0% or 100% point is a payout with zero trade count and contract volume. Payouts never enter the VWAP calculation, and unresolved races receive no forced endpoint.

Settlement times are retained explicitly in the saved SQL because canonical settled_time metadata was unpopulated during the audit. An endpoint is displayed only while current Outcome metadata agrees with the verified settled result. Contracts were matched to states by titles and rules: Kentucky is SENATELA-26-D, while Louisiana is KXSENATELA-26NOV-D. A ticker-substring match would silently mislabel Kentucky.

The hourly learning curve

The 721 hourly cutoffs run from October 4 at 21:17 UTC through November 3 at 21:17 UTC, always at minute :17. Each point uses the latest executed YES price strictly before the cutoff for all 33 contracts. These forecast snapshots carry the last observed trade forward until another arrives; the campaign and election-night VWAP series do not fill missing buckets. Five contracts with heavily favored outcomes needed more than a 24-hour lookback at the first cutoff. The final hourly snapshot is also the primary scorecard. This is a last-trade proxy, rather than a synchronized order-book snapshot.

Snapshot

Brier

Correct D/non-D calls

Median trade age

Oldest trade

30 days before

0.06216

30/33

103 minutes

2.97 days

7 days before

0.05441

30/33

30 minutes

43.37 hours

1 day before

0.03575

32/33

15 minutes

3.83 hours

Election day, 16:17 ET

0.03639

32/33

2 minutes

27 minutes

The fixed 21:17 UTC checkpoints correspond to 5:17 p.m. EDT in October and 4:17 p.m. EST on November 2–3. The largest one-hour improvement was a 0.00406 drop between November 2, 00:17 and 01:17 UTC (November 1, 7:17–8:17 p.m. EST). The chart describes observed changes without assigning a news cause. Its final last-trade score is 0.036392, identical to the primary scorecard before rounding.

The Trade Snapshot and Its Freshness

At the common November 3, 4:17 p.m. Eastern cutoff, each of the six selected contracts had traded within the preceding 31 seconds. The table shows the last executed YES price, its age at the cutoff, and the number of trades in the preceding hour. The full 33-contract freshness range is reported above:

Race

Last executed YES price

Trade age at cutoff

Trades in preceding hour

Texas

74%

0.127 seconds

1,813

Kansas

65%

2.600 seconds

310

South Carolina

31%

7.346 seconds

291

Iowa

51%

1.195 seconds

345

Maine

70%

2.542 seconds

235

Ohio

72%

30.972 seconds

211

One-hour volume-weighted prices were Texas 74.55%, Kansas 63.96%, South Carolina 30.17%, Iowa 50.34%, Maine 69.74%, and Ohio 72.14%. Those are averages over the preceding hour, rather than forecasts frozen at its final instant. Separately, daily trade averages put Kansas at 18.95% on September 1 and 59.29% on November 2, and Texas at 50.26% and 72.26%. Kansas’s November 2 average rose 7.86 points from the preceding UTC day, its largest increase in that September 1–November 2 sample. These daily trade measures differ from the dated published observations in the campaign section.

For election-night control, Outcome’s first observed trade at or above 99 cents was 8:18:04 p.m. Eastern for the House and 9:01:52 p.m. for the Senate. Kalshi’s recap reports 99% at 8:10 and 9 p.m. respectively. Executed prints and reported market odds can differ through quote measurement or rounding; the House gap does not establish an error in either source. A first threshold trade does not imply sustained confidence or continuous executable depth.

The margin comparisons

All margins below are Democratic minus Republican, in percentage points. Forecast dates are unmatched and vote counts are provisional. The source links appear beside the cases in the main text; exact timestamps are retained here:

Race

Reported margin-market D−R estimate

Provisional D−R margin

Difference

Texas Senate

+2.70 pp, Nov. 3

+12.09 pp, Nov. 4 10:11 p.m. ET

+9.39 pp

Texas governor

−2.70 pp, Oct. 28

+7.12 pp, Nov. 4 3:47 p.m. ET

+9.82 pp

Kansas Senate

−1.30 pp, Oct. 28

+5.00 pp, Nov. 5 1:20 a.m. ET

+6.30 pp

Decomposition and Robustness Checks

The 24-versus-nine grouping uses each contract’s frozen Outcome last-trade probability, with 10% and 90% included in the middle band. Both groups retain the same 33 called contracts and outcome labels as the primary scorecard. The group shares divide each group’s summed squared error by the total; this is a probability band, not an independent classification of electoral competitiveness.

Monthly improvement contributions use the existing 33-contract last-trade snapshots at October 4 and November 3, both 21:17 UTC. For each contract, subtract final squared error from initial squared error, then divide by the sum of those differences across the cohort. Kansas contributes 36.8273%; South Carolina contributes 30.9933%. The denominator is net improvement, including contracts whose scores worsened. Excluding the five contracts with initially older-than-24-hour prices leaves Kansas’s share at 36.9032%.

The Louisiana, Mississippi, and Alaska statistics use observed five-minute rows from the saved full-universe election-night query. The common window is November 3, 14:00 UTC through November 4, 10:30 UTC exclusive, with resolved contracts truncated at actual settlement. Payout rows are excluded from volume, peaks, and changes. The Mississippi rank sums executed contract counts, not dollars; contracts remain eligible for different durations because they settle at different times.

Mississippi’s opening five-minute bucket contains only four trades. Its independent 30-minute opening and final averages are 14.4236% and 82.0333%, supporting the rounded endpoints used in the text. Louisiana’s five-minute peak contains 218 trades and more than 26,000 contracts; its 30-minute peak is 28.7962%. Alaska’s five-minute peak contains 92 trades; its 30-minute peak is 96.5104%. Reaggregating the saved five-minute data matches the independently saved 30-minute results to floating-point precision. These checks establish that the described moves survive broader buckets; they do not establish a news cause or executable return.

First Crossings and Final Observed Runs

The confidence-persistence check uses the saved five-minute trade VWAP rows, excluding payout endpoints. A first crossing is the first observed bucket in the source window at or above the chosen threshold. The final observed run begins with the first qualifying bucket after the last observed below-threshold bucket. It is selected using the full observed path, so it is a retrospective description rather than a real-time decision rule. Safe races can start the window already above the threshold, and unresolved races can end without a final qualifying run.

Selected race

First ≥90% bucket

Final observed ≥90% run starts

Gap

Iowa

9:35 p.m.

9:35 p.m.

0 min

Kansas

8:30 p.m.

8:30 p.m.

0 min

Maine

8:45 p.m.

9:55 p.m.

70 min

Ohio

7:55 p.m.

8:20 p.m.

25 min

South Carolina

8:25 p.m.

8:25 p.m.

0 min

Texas

8:05 p.m.

8:05 p.m.

0 min

All times in this table are November 3, 2026, Eastern bucket starts. All six selected contracts have a populated five-minute grid over their observed spans. Maine’s 15 buckets from the first crossing through the final-run start contain 3,537 trades and approximately 1.262 million contracts; six averages are below 90%, in three episodes, with a minimum of 87.77%. Its final run has 38 populated buckets through the November 4, 1 a.m. bucket, which intersects market close. Observed averages do not establish continuous quotes, and no confidence path is inferred from trading close to settlement.

Threshold sensitivity matters. Maine’s first-to-final-run gap is 70 minutes at 89%, 70 at 90%, and 100 at 91%. Ohio’s 25-minute gap at 90% disappears at 89% and 91%; its lowest subsequent 90%-test average is 89.75%. These race-level five-minute measurements must not be subtracted from the separate chamber-control 99% lead times reported by Kalshi.

The Illustrative Seat-Count Proxy

For each of the 35 races, pᵢ is its unadjusted Outcome executed YES price strictly before November 3, 21:17 UTC. No cross-party normalization is applied. Let Xᵢ = √ρ Z + √(1 − ρ) εᵢ, where Z and each εᵢ are independent standard-normal variables. The simulated Democratic outcome is 1 when Xᵢ ≤ Φ⁻¹(pᵢ). This keeps each marginal win probability fixed while imposing correlation ρ between the latent X variables. The correlation between binary wins generally differs. Proxy seats equal 34 fixed holdovers plus the simulated Democratic wins. An independent winner may caucus with Democrats without settling its Democratic-win contract YES; omitting that contribution prevents treating the proxy as a calibrated caucus-membership or Senate-control model. Alaska and Mississippi enter through their observed prices without being assigned election outcomes.

The saved sensitivity output gives P(proxy seats ≥51) = 88.7096% and P(proxy seats ≥54) = 24.6089% at ρ = 0; at ρ = 0.4 they are 73.8431% and 38.1511%. Expected proxy seats equal 52.412 in both cases. The common factor is integrated with Simpson quadrature over [−8, 8] using 800 intervals, with exact Poisson-binomial convolution at ρ = 0. All 41 curve points were checked at both 400 and 800 intervals; the maximum change in a tail probability was below 10⁻¹², and marginal, total-mass, and expected-seat checks passed. No correlation was estimated, preferred, or fitted to a traded chamber price.

Open the saved-query controls on each chart to inspect the inputs, ticker filters, cutoffs, and calculations. Public-source charts retain their archival observations; Outcome history queries retain the observed trades and settlement treatment. Historical bid/ask quotes are not exposed in the public schema, and none of these calculations establishes achievable trading returns after spreads, fees, and timing.

Limitations and Sensitivity

The primary calculations retain the raw Democratic YES probabilities. They score the contract’s event, including an independent winner as a non-Democratic outcome, without converting it into a two-party probability. State matching uses the audited market titles and rules; the Kentucky/Louisiana ticker distinction is documented in the appendix. A contract’s price does not by itself specify the caucus affiliation of every possible winner.

Alaska and Mississippi remain unscored. Assigning each of their four hypothetical outcome combinations to the fixed Outcome prices gives a possible eventual 35-race Brier range of 0.037618 to 0.069047. These are bounds across unresolved outcomes, not a confidence interval or assigned winners. The 33-race result is therefore provisional. The same small cohort, exposed to shared national uncertainty, cannot establish long-run calibration or persistent partisan mispricing. Uncalled-race status.

Future comparisons should hold races, timestamps, outcome definitions, and scoring rules fixed. The unresolved questions are how reliably markets adapt across elections, how accurately they price margins and seat tails, how they compare on sustained election-night confidence, and whether any apparent opportunity was executable. Those are extensions of this audit, rather than conclusions supplied by a single night.

Updated
Data freshness

Data saved between and .

Results update when the author refreshes.

Explore more research