United States panel

Midterm elections · November 3, 20263 days to go

This panel places side by side what the prediction market prices and what the polls measure about the race for the American Congress.

The two are neither added nor subtracted, because they measure different quantities: the prediction market gives the probability of a party controlling the chamber, and the poll gives the lead in vote points. What is crossed here is direction and movement, never level.

AFOS reports the crossing. It does not forecast, does not recommend a position, and does not say who will win. Market from Polymarket. Polls attributed to the pollster that published them, each with a link to the primary source.

Prediction market: midterms

Polling: House vote

The caveat comes before the number

The market prices the PROBABILITY of a party controlling the House. Polling measures a VOTE MARGIN in points. These are different quantities, and this panel does not subtract one from the other: the result would have no meaning. In 2012 Democrats won more votes and fewer seats, because the gap can be entirely geography. What is cross-read here is DIRECTION and MOVEMENT, never level.

AFOS average · House vote intention

last 30 days, since 09/30 · 33 polls from 23 institutes
Dem
51.09%
Rep
43.45%

Democratic margin: D+7.64

These two numbers are the national vote intention for the HOUSE. Voters are asked whom they would back for Congress in their own district, and the answer is read as the national House vote. It is not the Senate, and it is not the two added together.

There is no equivalent for the Senate. Each seat is contested state by state and polled that way, so a national Senate average would be an invention. That is why the prediction market here shows House and Senate, while the polls show only the House.

How this average is calculated

média aritmética simples, sem ponderação e sem exclusão de instituto. Quando o mesmo instituto publica mais de um recorte da mesma rodada, entra um só, pela hierarquia provável votante > eleitor registrado > adultos. Quando o índice lista a mesma casa sob dois nomes, o de quem executa e o de quem publica, ou duas grafias, ela conta uma vez, pelo nome de quem executa, por tabela declarada casa a casa. Onda que o instituto publicou com OUTRO instrumento, como cédula com nomes de candidatos, fica de fora e é listada em excluidasPorInstrumento com a prova.

A Wikipédia é usada como ÍNDICE de quais pesquisas existem e quem publicou. Os números são atribuídos ao instituto e cada linha traz o link da fonte primária. O AFOS não redistribui tabela de terceiro.

Decision Desk HQ · FiftyPlusOne · RealClearPolitics · Silver Bulletin · VoteHub: são modelos de terceiros; o AFOS calcula a própria média simples e não republica modelo alheio.

Most recent polls

InstituteFieldworkSampleMoEDemRepMarginSource
Tufts University/YouGov10/23–10/27499 LV±4.65240D+12open
Tufts University/YouGov10/23–10/271,000 A±3.64736D+11open
Harvard/Harris Poll/HarrisX10/24–10/26— LV—5248D+4open
The Economist/YouGov10/23–10/261,123 LV—5241D+11open
The Economist/YouGov10/23–10/261,263 LV—5041D+9open
The Economist/YouGov10/23–10/261,430 RV—4640D+6open
Reuters/Ipsos10/22–10/261,164 LV±2.95344D+9open
Reuters/Ipsos10/22–10/261,280 RV±2.85041D+9open
NPR/PBS/Marist University10/19–10/26— LV±3.85244D+8open
NPR/PBS/Marist University10/19–10/261,289 RV±3.75242D+10open
Cygnal (R)10/23–10/241,500 LV±2.55243D+9open
CBS/YouGov10/21–10/231,749 LV—5345D+8open

LV = likely voter · RV = registered voter · A = adults

Provenance and limitations

  • Não existe registro público obrigatório de pesquisa nos EUA, ao contrário do TSE no Brasil.
  • A ponta da Wikipédia pode atrasar alguns dias em relação ao que os institutos já publicaram.
  • Pesquisa americana é proprietária: exibida com atribuição e link, nunca redistribuída.

Index read on 10/30 · 535 polls on file · en.wikipedia.org/wiki/2026_United_States_elections (CC BY-SA 4.0 (Wikipédia))

World Bank indicators about the country, not about the election. They sit alongside the signal to give scale to the terrain, never as a predictor: nothing here forecasts an electoral result.

Structural context

Governance (0–100 scale)

Political stability64
Voice & accountability72
Rule of law74
Government effectiveness78
Regulatory quality76
Control of corruption70

Economy

Population340.1M
GDP$28.75T
GDP per capita$84,534
Inflation2.9%

Education

Public education spending (% GDP)5.4%
Expected years of schooling16.3 years

Source: World Bank — Worldwide Governance Indicators + World Development Indicators (2024). Annual structural indicators that contextualize the country; they do not predict the electoral outcome. WGI ↗ · WDI ↗

Declared limitations

What follows is not a footer. It is the list of what this panel cannot claim, written before anyone asks.

On the market

  • The volume is Polymarket’s, not the entire American prediction market. It is one venue among others, and the number shown is the probability it prices. Other venues may sit at a different price at the same moment.
  • Kalshi is out, and that is a declared choice, not an oversight. Relying on a single venue is a real limitation of this panel, and we would rather say so than add up venues with different rules.
  • Thin band markets move on little money. Where volume is low, a single participant shifts the price, and the reading is worth less than the number suggests. That is why volume appears next to every market.
  • A market’s bands only become numbers on screen if they total between 95% and 105%. Outside that range they cannot be read as probability, and the card shows shape only, that is, the relative size of the bands, with no percentage. The measured total stays in plain sight, along with how much it moved over the last 24 hours.

On the polls

  • There is no mandatory public poll registry in the United States. Nothing matches Brazil’s TSE protocol, so there is no way to know what is in the field and has not been released yet.
  • Wikipedia is used as an index of which polls exist, and the number is always attributed to the pollster, with a link to the primary source. AFOS does not redistribute anyone’s table. The edge of that index can lag a few days behind what pollsters have already published.
  • The average is our own and it is a simple mean, unweighted and excluding no pollster. Weighting would require a public quality yardstick, and none has existed since FiveThirtyEight closed in March 2025. Inventing ours would be a value judgement dressed up as method.
  • American aggregators are not part of this average. Those are third-party models, and republishing someone else’s model as if it were our own reading is the opposite of what AFOS does.

On the crossing

  • The market gives the probability of a party controlling the chamber. The poll gives the lead in vote points. These are different quantities, and this panel does not subtract one from the other: the result would carry no meaning.
  • In 2012 Democrats won more votes and fewer seats. The gap between votes and seats can be entirely geography, and no arithmetic done here would separate the two.
  • Until the market that measures the same quantity as the polls matures, what is crossed here is direction and movement, never level.
  • AFOS reports the crossing. It does not forecast, does not recommend a position, and does not say who will win.

On this reading

522 rows were read from the index and 535 published. None were discarded, and 2 lack a link to the primary source. The house average covers 33 polls from 23 pollsters over the last 30 days. Index read on 10/30/2026.