AFOS Weekly · USA-2026 midterms Political Risk Weekly

AFOS Analytics

Real-time political pricing
Prediction market × polls × press
no smoothed averages

Harvard Dataverse · DOI 10.7910/DVN/XRUT8U
United States·Issue №9·Week of 2026-09-24 to 2026-10-01·published thursday 07:00 BRT

What the prediction market did

The eight midterm contracts hold USD 23.01M, up USD 0.91M on the week. Four of the five published prices closed exactly where they opened, and the House pair summed 100.00% at both ends. The one move was Senate Republicans, and it survives normalization at 0.63pp rather than the full point.

House · Democrats

92.50%

→ 0.00pp on the week

On USD 7.26M, the deepest of the eight books. It closed at 92.50% on all eight days, and across 45 readings its high and its low were the same number.

House · Republicans

7.50%

→ 0.00pp on the week

On USD 5.49M, which took USD 0.30M of new money, more than the Democratic side. It did not sit still to get there: it reached 8.50% on September 28 and spent two days there before coming back.

Senate · Democrats

62.50%

→ 0.00pp on the week

On USD 3.20M. It closed at 62.50% on all eight days, but opened the week at 66.50% in the small hours of September 24, the highest this series holds since April 14, and was back by midday.

Senate · Republicans

36.50%

↓ 1.00pp on the week

On USD 2.32M, the thinnest control book. The only price that moved. Its week ran from 34.50% on September 24 up to 38.50% on September 28 and down to the close: a 4.00pp spread around a 1.00pp result.

The fifth published price, on the election happening on schedule, closed at 98.10%, unchanged on the week, on USD 0.50M. It moved more than the two control pairs did: down to 97.80% on September 29 and up to 98.30% on September 26. Weekly figures here are close to close.

What the polls did

The disagreement comes before the average, because it is the finding. Three national rounds finished fieldwork this week, and they are the two ends of the range plus one in the middle. Thirty-six rounds from twenty-nine institutes sit in the thirty-day window, averaging D+8.11.

Widest

Angus Reid Global · D+14

Narrowest

Harvard/Harris · D+2

Spread between institutes

12 points

Both ends were fielded this week, three days apart. Angus Reid Global surveyed 1,041 registered voters through September 25 and has Democrats at 49 and Republicans at 35. Harvard/Harris surveyed through September 28: its likely-voter screen reads 51 to 49, and its full sample of 2,200 registered voters is an exact 50 to 50 tie. The median round is D+8 and the middle half runs D+6 to D+10.

What the press reported

What the fixed-list outlets covered, September 25 to October 1

The outlets spent the week auditing the polls themselves. On September 29 The Hill asked how bad midterm polls are for Republicans, and on September 30 the Washington Post ran an analysis headlined whether you can trust the polls, alongside an opinion piece arguing the polls may already know how many House seats Democrats will win. On October 1 NBC News published an explainer on how it collects its own polling data. Four pieces in three days, the same week three pollsters landed twelve points apart.

Senate control carried the campaign coverage. The Washington Post published the nine seats most likely to flip on September 28 and the races that could surprise on September 29. On the same day the New York Times reported that in nine key Senate races, five Republicans would not commit to campaigning with the president.

The economy and the Iran war ran underneath it. Fox News reported Republicans fearing they had done nothing to lower prices on September 29, and The Guardian covered a 15 billion dollar Iowa steel plant announcement framed as a bid to lift flagging midterm hopes on September 28. On September 30 The Hill carried the president's daughter-in-law saying the Iran war could cost him the midterms.

Our measurement: On direction the rounds agree: all thirty-six put Democrats ahead. What a single round cannot show is the width, and this one runs D+2 to D+14. Whether any one poll is right is not something we can measure. How far apart they are is, and this week it was twelve points.

The crossing of the week

Three pollsters, one week, twelve points apart

Angus Reid closed fieldwork on September 25 at D+14. Harvard/Harris closed on September 28 at D+2, with its full sample tied at 50 to 50. Both are national, both ask about Congress, and both are in our average. We do not pick between them, and the gap is not noise to average away: it is the measurement.

The market closed the week where it opened, and it was not sitting still

All five prices ended within 1.00pp of their September 24 closes and three ended identical. But House Republicans travelled to 8.50% and back, and the Senate pair opened at 66.50% and 34.50% before settling. A flat week and a quiet week are different things, and this was the first.

The coverage asked the question the spread was already answering

Three outlets published pieces on whether polls can be trusted between September 29 and October 1. In the same window the distance between the highest and lowest round in our average was twelve points. We do not read one from the other, and we never subtract a market probability from a polling margin: the market prices who controls a chamber and the polls measure a lead in votes.

How to read this number

Why we print the spread before the average

An average of thirty-six rounds is one number, and it cannot tell you whether the pollsters agreed. This week they did not: the same question, asked by three houses within three days, produced D+14, D+12 and D+2. The average is D+8.11 and it is true, but a reader who sees only it learns nothing about how far apart the pollsters were. That is why the range comes first here: when the spread is wide, the average summarises a disagreement rather than describing a settled fact.

Sources

Prediction market, October 1 closes

Polling, primary sources

Press cited in sections 4 and 5

Method

Market prices in this issue are October 1 closes, the last point recorded that day in our daily database backup, compared with the September 24 closes. The House series opens on July 29, 2026 and the Senate on April 14, 2026, so every superlative here means since those dates. Series highs and lows are measured against the full recorded series in that backup, not against the public history route, which caps its window at 90 days and would hide the start of the Senate series. A certified double reading was also taken at 22:08 UTC on October 1 and is reported in the market footnote. The polling average is the simple mean of every national round fielded in the last thirty days, one round per institute per wave, choosing likely voters over registered voters over adults when a house publishes more than one screen. A wave that a house publishes with a different instrument, such as a ballot naming the actual candidates in the respondent's district rather than the generic question, is excluded and listed with its proof. Press coverage comes from a fixed list of 23 outlets, at most two items per outlet, with no political labelling of any outlet.

AFOS
AFOS Analytics
Global Political Risk Intelligence · GLOBAL BY DESIGN
www.afos-analytics.com · contact@afos-analytics.com · github.com/AFOS-Analytics

Get it by email

AFOS Weekly follows the US midterms every Thursday, written for voters: where the prediction market, the polls and the press disagree, and what each one measures.

No spam. Unsubscribe anytime.

After signing up you choose whether to receive it in English, Portuguese or Spanish.

AFOS Weekly · Issue No. 9 · Three pollsters asked the same question this week and came back twelve points apart | AFOS Analytics