AFOS Analytics
Real-time political pricing
Prediction markets × polls × news
no smoothed averages
BrazilThe week of August 17 to 21 was the first since July in which the Brazilian presidential market moved in a sustained direction. The distance between Lula and Flávio Bolsonaro closed from 33.10pp to 29.30pp in the presidential contract, a compression of 3.80pp across five sessions, and the move came more from the runner-up rising (+1.80pp, from 31.40% to 33.20%) than from the leader falling (-2.00pp, from 64.50% to 62.50%). Both figures are daily closes for Aug 17 and Aug 21, and the choice of endpoint matters: measured from Monday's intraday high against Friday's low, the same move would read as 7.00pp, almost double. This issue uses daily closes on every contract, and says so because the criterion changes the headline.
What makes the week relevant to anyone pricing risk is not the compression itself but what the measuring instrument did while it happened. On Aug 21, the last session of the window, two national polls were released the same day and they invert on the runoff: Datafolha (n=2,058, margin of 2pp) gives Lula 47% against 43%, and Veritá (n=3,840, margin of 2pp) gives Flávio Bolsonaro 47.3% against 42%. That is 9.30pp between two instruments measuring the same thing, with overlapping field periods, released hours apart. In the first round the spread across the three national polls of the week reaches 6.80pp, from 0.2pp at Veritá to 7.0pp at Datafolha. The price moved 3.80pp in one direction while the measurement lost the ability to say what the level is.
Two methodological caveats close the picture. The first is capture: the house guard rejected the second-place contract on one of Sunday's runs, with the two readings disagreeing by 0.65pp, and cleared it on the next run; the value published here is from the approved run. The second is structure: the five largest accumulated volumes in the book add up to USD 55,733,279, or 42.3% of the total, and carry 4.05% of combined probability. Accumulated volume records the money that passed through since April, not today's conviction, and in a book with that distribution the liquidity read has to say which of the two it is measuring.
1.Executive Summary
The theme of the week is the distance between the firmness of the price and the instability of the measurement. The presidential market compressed 3.80pp monotonically, with no reversal in any of the five closes, and the third-place contract followed, with Renan Santos going from 52.50% to 55.50%. Over the same period, three registered national polls produced readings ranging from a statistical tie to a seven-point lead in the first round, and two of them swap the runoff winner. For a desk reader the implication is direct: in this window the relevant confidence interval is not the sample margin declared by each poll, it is the distance between polls. Neither uncertainty is captured by an average.
Against the previous issue the contrast is one of regime. Issue №13 stamped the distance at 35.60pp on Aug 16 and described a price that was practically still; this issue stamps 29.65pp on Aug 23. That is 5.95pp between stamps, over a seven-day window that does not coincide with the session window used in the grid, and the two calculations are reported separately on purpose. The spread across institutes fell from 13.7pp to 6.80pp over the same interval, but with seven national polls in the previous window against three in this one: part of the drop is the composition of the poll sample, not convergence among them.
2.Why AFOS does not smooth
This week's case is the cleanest the panel has recorded for the central AFOS argument, and it does not depend on any editorial choice of ours: the two national polls of Aug 21 went to field in overlapping windows, were registered with the electoral court, and reached opposite conclusions about who wins the runoff. Veritá's field ran Aug 16 to 20 and Datafolha's Aug 18 to 20, with three days of overlap, so the difference is not explained by days between collections.
<strong>Why it matters:</strong> the average is not conservative here, it is assertive. It produces a statistical tie that none of the three measurements supports, and it erases the only actionable fact of the week, which is the disagreement between instruments about the same quantity on the same day. Anyone pricing off the average is pricing a scenario nobody measured.
The house reliability scale, published per poll on the panel, gives Datafolha 5 and Veritá 2, and Veritá has the larger sample of the two. Larger sample and lower reliability point in opposite directions, and the panel does not resolve that conflict by weighting: it displays both readings with the scale beside them, so the reader makes the choice while declaring the criterion. Weighting here would embed in the number a decision that belongs to the reader.
The same framework applied to Brazil, measuring the distance between prediction markets and polls and checking the signal against the actual election result, has already been tested against eleven elections across four continents, among them India, France, the United Kingdom, Mexico and the United States (2024), South Korea, Chile, Germany and Canada (2025) and Peru and Colombia (2026). India is the only case in which market and polls erred together in the same direction. The method reports convergence and divergence with no retroactive adjustment. The dataset is open under Apache 2.0 and carries a permanent DOI at the Harvard Dataverse.
3.Weighted scenarios for the week
The three scenarios below refer to the window of August 24 to 30, in which five national polls are registered for release, adding up to 14,800 interviews. The probabilities are the writer's approximations about price behaviour, not model output, and are not a recommendation.
The five national polls of the week arrive distributed around the Datafolha and BTG/Nexus range, with the first round between five and seven points, and Veritá stays isolated as the extreme point of the window. In that case the price compression loses traction near 29pp to 30pp, because the 29.65pp distance already embeds a reading more favourable to the runner-up than the median of the registered polls. The third-place contract tends to stay stable, with Renan Santos at 56.00% against 41.00% for Ronaldo Caiado.
The national polls of the week confirm the narrowing measured by Veritá, and the first round starts showing a contest inside the margin at more than one house. In that case the 29.65pp distance stops being a floor and becomes a ceiling, and the compression continues. The point to watch is that this scenario requires the institute with the lowest reliability on the house scale to be right against two with higher reliability, which has happened in other series and is not dismissible.
The JOTA Jornalismo poll, registered for Aug 26 with n=6,000 and a field running from July 27 to August 24, has the largest sample and the longest field window on the calendar. A field of almost a month measures a period in which the price moved 3.80pp, so it may publish a number that corresponds to no specific moment of the series. If it comes in outside the range of the others, the effect on the price tends to be larger than that of an ordinary poll, because of the sample, and less informative than that of an ordinary poll, because of the field.
4.Indicator Grid
| Contract | Current | Δ week | Vol USD acc. | Implied reading |
|---|---|---|---|---|
| Lula, presidential | 62.50% | -2.00pp | USD 8,895,288 | Closed all five sessions without a gain. Sat at 63.50% from Aug 18 to 20 and gave up 1.00pp at Friday's close. |
| Flávio Bolsonaro, presidential | 32.85% | +1.80pp | USD 8,795,662 | The entire weekly gain came from a single session: 31.40% at the Aug 17 close to 33.20% at the Aug 18 close, and the rest of the week oscillated inside that range. |
| Renan Santos, presidential | 3.75% | +0.10pp | USD 10,698,439 | Second largest individual volume in the book, above both front-runners, and it pays less than four points. |
| Renan Santos, third place in the first round | 56.00% | +3.00pp | USD 192,789 | Largest percentage move in the grid. Rose from 52.50% to 55.50% on the week while his presidential price stood still. |
| Ronaldo Caiado, third place in the first round | 41.00% | 0.00pp | USD 58,397 | Closed the week at the level it opened, despite appearing at 5% in two of the three national polls of the window. |
| Flávio Bolsonaro, second place in the first round | 86.50% | -1.50pp | USD 363,372 | Gave up ground on the week while rising in the presidential contract. Thin book: the capture guard rejected this contract on one of Sunday's runs and cleared it on the next. |
| Supreme Court justice impeached before 2027 | 3.35% | 0.00pp | USD 83,737 | No variation across the five sessions, including the week in which the Supreme Court ruled on amendments directed by party presidents. |
| PL with the largest Senate caucus | 78.00% | +2.00pp | USD 259,384 | Rose from 77.00% to 79.00% on the week and gave part of it back over the weekend. MDB is second, at 16.55%. |
| 2026 CPI between 5.00% and 5.49% | 51.05% | +3.20pp | USD 14,547 | The central band went from 47.10% to 50.30% on the week, taking weight from the 5.50% to 5.99% band, which gave up 1.20pp. |
5.Liquidity and market structure
The five largest volumes in the book add up to USD 55,733,279, or 42.3% of everything traded, and carry 4.05% of combined probability in the confirmed reading of Aug 23. The two front-runners, who together carry 95.35% of probability, account for USD 17,690,950, or 13.4% of the book. The correct read is not that the market is wrong: accumulated volume measures how much money passed through the contract since it opened in April, not today's conviction. Tarcísio de Freitas opened a re-election campaign in São Paulo on Aug 16 and is not a presidential candidate; his volume is the record of a thesis that was traded and unwound, not of a live position. The exception on the list is Renan Santos, second largest volume in the book and the only one of the five with probability above one point: he pays 3.75% in the presidential contract and 56.00% in the third-place contract, which makes his volume consistent with a current position and not only with historical turnover.
Accumulated volume is informational, not directional. Book depth is not cited in this publication, by editorial decision of May 21, 2026: in a continuously arbitraged market, low liquidity does not imply a wrong price, and exposing the number without that context produces a skewed reading.
6.Calendar of price-relevant prints
| Date | Sample | Why it matters | |
|---|---|---|---|
| Aug 24 | Nexus | n=2,000 | First national poll after the contradictory pair of Aug 21. Protocol BR-09028/2026. |
| Aug 25 | Gerp | n=2,400 | House with a series on the panel since July, which allows internal comparison. Protocol BR-03547/2026. |
| Aug 26 | JOTA Jornalismo | n=6,000 | Largest sample on the calendar and a field from Jul 27 to Aug 24, almost a month. Protocol BR-07806/2026. |
| Aug 26 | Indexa | n=2,000 | Second release on the same day, which repeats the configuration of Aug 21. Protocol BR-06366/2026. |
| Aug 27 | PoderData | n=2,400 | Closes the window of five national polls in four days. Protocol BR-04974/2026. |
Release dates as registered with Brazil's electoral court, filtered to national scope and a sample of at least 1,000. Another 58 registrations from the same period were left out by scope, 57 of them state-level and 1 with no scope identified. Registration precedes release: institutes may delay or cancel, and confirming actual publication requires two primary sources before citing a number.
7.Watch list, week triggers
- If two of the five national polls of the week repeat a first round inside the margin the Veritá reading stops being an isolated point and becomes the second observation of a narrowing. It is the trigger that would most quickly reprice the 29.65pp distance.
- The JOTA poll on Aug 26, with n=6,000 and a 29-day field has the largest sample and the longest window on the calendar. A long field over a period in which the price moved 3.80pp measures an average of the move, not the current level.
- Two releases on the same day on Aug 26, JOTA and Indexa repeat the configuration of Aug 21. If they also diverge, the spread between institutes stops being an event and becomes a feature of the window.
- The third-place contract, which rose 3.00pp on the week moves more than the presidential one and on volume thirty-five times smaller. A large move in a thin book requires confirmation before it becomes a reading.
- The Aug 18 jump in the runner-up contract concentrated the entire weekly gain in one session, on a day when the news flow was dominated by state-level polls from Paraná. The panel records the coincidence and does not assert causation: the AFOS scope is national, and state polls do not enter the scale.
8.Methodology
AFOS does not weight polls or build averages of institutes. It reports the prediction market price, each national poll registered with Brazil's electoral court along with its sample, margin and house reliability scale, and press coverage, leaving visible where the three disagree. Every published price goes through two independent readings eight minutes apart and is released only when they agree within 0.20pp; the guard certifies book by book, and a rejected book is not published.
The weekly variations in this issue use each day's close rather than the first and last point of the window, because the second choice would have returned a compression of 7.00pp instead of 3.80pp for the same move. The series comes from the daily database backup, not from the history route, which for the Brazilian presidential market blends contracts with similar prefixes. Code and method are open under Apache 2.0 on GitHub, and the datasets carry permanent DOIs at the Harvard Dataverse.
9.Additional reading · macro coverage
AFOS is the primary source for this analysis. The references below are complementary macro context and do not support any of the measurements above.
- Folha de S.Paulo · Coverage of the Aug 21 Datafolha poll, the house's first national survey after the campaign began and the first to include Pablo Marçal in the prompted scenario.(paywall), www1.folha.uol.com.br/poder/
- Valor Econômico · Tracking of the electoral calendar and the effect of polls on domestic assets.(paywall), valor.globo.com/politica/
- Poder360 · Aggregation of polls registered with the electoral court, with protocol and field dates, useful for confirming actual release., www.poder360.com.br/
- Reuters Brazil · English-language coverage of the electoral cycle, a reference for readers outside the country., www.reuters.com/world/americas/
- Central Bank of Brazil, Focus · Weekly market expectations report, an institutional counterpoint to the inflation contract cited in the grid., www.bcb.gov.br/publicacoes/focus
Get it by email
AFOS Tradeoff is the weekly read on Brazilian political risk, for people who decide on data rather than headlines.
No spam. Unsubscribe anytime.
After signing up you choose whether to receive it in English, Portuguese or Spanish.