Brazil Election Polls vs Polymarket Odds: What the Numbers Mean

A candidate can lead a close poll while trading at a much higher winner-market price. The poll’s percentage is an estimate of voting intention. The market’s percentage is a price interpreted as a chance of satisfying that contract’s winning condition. A difference between them does not, on its own, show a pricing mistake.

This matters in Brazil’s runoff. The Polymarket election page, checked on 19 October 2026, displayed a 77¢ Yes offer for Flávio Bolsonaro and a 24¢ Yes offer for Lula. The dated poll below measures something else, and was conducted several days earlier. This is a comparison of definitions and observation times, not a calibrated forecast.

Datafolha’s 16 October release reported 47% for Flávio Bolsonaro and 45% for Lula among all responses; 6% chose blank or null and 2% were undecided. On its valid-vote presentation, the two candidates were at 51% and 49%.

Fieldwork ran on 13–14 October: 2,520 in-person interviews in 153 municipalities. Datafolha reported a margin of error of two percentage points and a 95% confidence level, with TSE registration BR-00466/2026. The institute described a technical tie. This is one identified survey, rather than an average of all available polls.

Download table figures:CSVJSON with sources

The arithmetic illustrates the denominator change. Using the published rounded all-response figures, 47 ÷ (47 + 45) is approximately 51.1%; the other candidate’s share is approximately 48.9%. Those calculations are illustrative checks on rounded inputs, rather than unrounded estimates from the pollster’s underlying data.

Removing other responses changes the scale of the comparison. It does not discover new votes, resolve how undecided respondents will behave, or turn a sample into the eventual electorate.

Winning a two-candidate contest depends on finishing ahead, not on reaching the market price as a vote share. Imagine a set of possible elections in which a candidate usually finishes just above half the valid votes. That candidate’s average vote share could remain close to half while the fraction of scenarios they win is much higher.

Turning a poll into that fraction requires a forecast model. The model must make assumptions about polling error, turnout, changes after fieldwork and how errors relate across surveys. The published sample margin is not a complete measure of those uncertainties. A two-point polling lead alone does not supply a unique probability of victory.

A market also does not explain its own reasoning. Its price may reflect information, positioning, liquidity constraints or trading incentives. Observing a move is not enough to attribute it to a debate, endorsement or policy announcement.

The two Yes offers in the observed page display totalled 101¢. That is not a 101% forecast. They were offers in separately quoted contracts, with spreads and rounding. A best ask is the lowest currently displayed selling price; a best bid is the highest displayed buying price. They need not be the same.

Do not silently normalise executable offers into a two-candidate probability chart. First decide whether the comparison uses last trades, midpoints or offers, make the observation times comparable, and state the transformation. Normalising a pair of numbers also does not make either price executable for an arbitrary order size.

The Brazil election room is where to inspect current available readings. The 19 October prices above remain a dated page observation; this article does not relabel them as live.

The question is whether the two sources disagree after their definitions and timing have been made comparable. Start with the following checks before drawing that conclusion.

• Match the outcome: eventual president, first-round leader, margin bracket or another condition.
• Match the time: use fieldwork dates for a survey and an observation time for a quote.
• Name the denominator: all responses and valid-vote shares have different bases.
• Name the price: last trade, midpoint, best bid and best ask are different readings.
• Keep uncertainty visible: sampling error, turnout assumptions, spread, depth and missing observations each limit the comparison.

For the candidates, dates and source map, open the Brazil election guide. On election night, use the official-results guide; the electoral count and contract settlement follow different processes.

For execution mechanics beyond this election, our order-size study explains why the quote at the top of a book can differ from an average fill. That is a separate question from which candidate wins.