AFOS Analytics
Real-time political pricing
Prediction markets × polls × news
no smoothed averages
BrazilThe week of 7 to 11 September did not compress the presidential spread, it inverted it. The gap between Lula and Flávio Bolsonaro in the presidential contract went from 18.60pp in favour of the former on Monday to 6.70pp in favour of the latter on Friday, a shift of 25.30pp across five sessions. Measured over the 576 paired captures in the series opened on 14 April, that is the largest weekly shift ever recorded in this book, against 18.00pp in the week of 11 to 15 May, which was the only other one to invert. The lead changed hands on 10 September at 13:00 UTC and it was the first change in 120 days, since 13 May.
The mechanics are not one-sided, and that is what separates this week from the two before it. The leader gave up 13.00pp, from 58.50% to 45.50%, and the runner-up rose 12.30pp, from 39.90% to 52.20%. Both sides moved almost half and half, which had not happened since the week of 24 to 28 August. And the first-round runner-up contract moved in mirror, which is the internal confirmation that the book repriced who wins and not noise: Flávio Bolsonaro lost 13.00pp of probability of finishing second, from 88.50% to 75.50%, and Lula gained 15.00pp, from 8.40% to 23.40%.
The eight national polls released in the window explain the repricing and, at the same time, refuse the simple reading. In the runoff pair, five of the eight put Flávio Bolsonaro ahead, two measure a tie and only one, the Datafolha poll of 11 September, gives the pair to the first-round leader. In the first round the picture is the reverse: six of the eight keep Lula ahead, with a dispersion running from Flávio +3.0 points, in the Gerp poll of 9 September, to Lula +7.0 points, in the Quaest/Genial poll of 7 September. That is 10 points of range between houses on the same question in the same week. The two rounds do not contradict each other: they measure different questions, and the presidential contract prices the second one.
1.Executive Summary
The theme of the week is an inversion, and the useful part of it is not that it happened, it is the coherence between the instruments that accompanied it. The winner contract turned, the runner-up contract turned in mirror, and five of the eight national polls in the window were already placing the runner-up ahead in the runoff pair before the lead changed in the price. Three independent measurements pointing the same way in the same week is rare in this series, which spent all of August with price and polling moving in opposite directions. What remains open is the first round, where six of the eight houses keep the previous leader ahead, with 10 points of dispersion between the highest and the lowest reading.
2.Why AFOS does not smooth
The eight national polls released between 7 and 11 September carried the same first-round question, overlapping fieldwork and samples of 1,500 to 5,000 interviews. A weighted average of them produces a number that looks like consensus and erases the central fact of the week: two houses of comparable reliability measured different winners.
Why it matters: the 10-point range is larger than the sum of the margins declared by any pair of houses in the window, which means the difference between them is not explainable by sampling error. The average of 2.6 points corresponds to what none of the eight measured, and the presidential contract, which prices the runoff and not the first round, turned in the direction of the five houses the average dilutes.
Issue №16 assigned roughly 60% to the base case, in which the presidential spread would settle between 13pp and 17pp without breaking any recent boundary, and roughly 30% to the case running against the pricing, which said that the houses measuring the runner-up ahead or tied in the runoff were not the exception and that the reading would repeat. What happened was the 30% case, and it happened more strongly than it was written: it was not two houses, it was five of eight, and the spread did not merely break below 13pp, it inverted its sign. The base case was wrong at both ends of the range it declared. The 10% tail, which expected the Supreme Court impeachment contract to return above 15.10% and hold there, also did not materialise: it closed the week at 7.30%. Recording this matters more than recording the hits, because a 30% case that materialises and a 60% case that fails in the same week say the distribution was displaced, not that the reading was right by another route.
3.Weighted scenarios for the week
The week of 14 to 18 September is the most crowded release calendar ever recorded in this series. The Superior Electoral Court registry carries twelve national polls scheduled for the window, five of them concentrated on 17 September and the Veritá poll of 40,500 interviews on 18 September. On 21 August two national polls released on the same day went as far as inverting the runoff pair between them, and now there are five on the same day.
The twelve national polls confirm the pattern five of the eight already measured, with the runoff pair inside the declared margin and the first-round runner-up ahead or tied in most of them. In that case the presidential spread settles between 2pp and 10pp in favour of the side that turned, without breaking the 14.30pp the week produced as its high, and the book starts treating the inversion as a level rather than an excursion. The first round keeps giving the previous leader the advantage, because the transfer of votes from eliminated candidates is what separates the two scenarios, and a market that prices the winner has no obligation to agree with the first round.
The concentration of five releases on 17 September gives back what the previous week took. The Datafolha poll of 11 September already moved that way, measuring 46 to 44 for the first-round leader after seven readings that did not give him the pair, and it is the most recent reading in the window. If two or three of the twelve repeat that pattern, the winner contract has to choose between the inversion it just priced and the highest-reliability house of the week, and the spread crosses back through zero. The series has already shown that this boundary is crossable: it was crossed twice within four days, on 12 September in the morning and again before midday.
The Supreme Court session scheduled for 15 September produces an act that puts the impeachment contract back above the 8% band where it spent the week. The book holds USD 90 thousand accumulated and took in USD 3 thousand in the window, so it is thin enough for isolated orders to produce double-digit shifts, like the 11.50pp of the week of 31 August and the 7.80pp given back since then. What makes this tail different from the previous ones is that the session is scheduled and public, which removes the calendar uncertainty and leaves only the uncertainty of content.
4.Indicator Grid
| Contract | Current | Δ week | Vol USD acc. | Implied reading |
|---|---|---|---|---|
| Flávio Bolsonaro, presidential | 52.20% | +12.30pp | USD 10.40M | Took the lead of the book on 10 September at 13:00 UTC, the first change in 120 days. Intra-week high of 55.10% on Thursday. Took in USD 562 thousand of new money in the window. |
| Lula, presidential | 45.50% | -13.00pp | USD 10.74M | His largest weekly fall in the series. Came off 58.50% and recorded a low of 43.50% on Thursday. Took in USD 811 thousand, the largest single inflow in the book on the week. |
| Presidential spread, Flávio minus Lula | +6.70pp | +25.30pp | USD 148.54M | Largest weekly shift of the 576 paired captures since 14 April. Path: -18.60, -7.40, -5.90, +11.60, +6.70. The high of 14.30pp was recorded at 06:30 UTC on Friday. |
| Flávio Bolsonaro, first-round second place | 75.50% | -13.00pp | USD 731 thousand | Mirror of the winner contract, and the symmetry is the internal confirmation: whoever rises in the probability of winning falls in the probability of finishing second. |
| Lula, first-round second place | 23.40% | +15.00pp | USD 602 thousand | Almost tripled across five sessions, from 8.40% to 23.40%. Took in USD 127 thousand, the largest inflow in the runner-up book on the week. |
| Renan Santos, first-round third place | 38.50% | +11.00pp | USD 297 thousand | High of 44.00% on Wednesday. A large move in a book that took in USD 17 thousand on the week, and a book that size moves double digits on a small order. |
| Ronaldo Caiado, first-round third place | 4.50% | -6.00pp | USD 118 thousand | Second straight week of decline in that contract, after the 21.50pp lost in the previous window. Came off 10.50% and closed below 5%. |
| Supreme Court impeachment before 2027 | 7.30% | -1.00pp | USD 90 thousand | The weekly delta hides the path: 8.30, 4.40, 4.90, 4.60, 7.30. Against the close of 15.10% on 4 September, that is 7.80pp given back over two weeks. |
| PL, Senate majority | 83.50% | +1.00pp | USD 261 thousand | The quietest contract on the panel in one of the most mobile weeks of the series. It took in no new money in the window, and the closing range ran from 82.50% to 83.50%. |
5.Liquidity and market structure
The book took in USD 3.76M of new money on the week, and the second largest single inflow did not go to either of the two names that moved. Romeu Zema took in USD 638 thousand and closed all five sessions at 0.20%, with a recorded range of 0.20% to 0.20%, meaning no variation at all. At the other extreme, Tarcisio de Freitas holds the largest accumulated volume in the book, USD 14.07M, and took in exactly zero in the window, with the price frozen at 0.10%. The two cases are opposite in nature and the same reading applies to both: accumulated volume is trading history and it is not current conviction. Anyone reading the volume ranking as a probability ranking is reading the book backwards, and the top two by volume pay 0.10% and 1.80%.
Volume is accumulated since each contract opened. The week's values are measured at the recorded closes of 7 and 11 September; the prices in the table are the closes of 11 September. AFOS does not quote order-book depth in editorial text, because low liquidity in a continuously arbitraged market does not mean the price is wrong.
6.Calendar of price-relevant prints
| Date | Sample | Why it matters | |
|---|---|---|---|
| 14/Sep | Quaest/Genial and NEXUS | n=2,004 and n=2,000 | Tier 1 on Quaest. It measured 41 to 41 in the runoff on 7 September, the tie that opened the sequence, and it is the first of the two with a comparable reading inside its own series. |
| 15/Sep | MDA and Indexa | n=2,002 and n=2,000 | Two houses with no reading in the previous window. They arrive without a recent series of their own, which makes them a test of level rather than of trend. |
| 16/Sep | Datatrends and American Analytics | n=2,000 each | Same case as the previous day. Their value lies in widening or narrowing the 10-point range the week of 7 to 11 produced. |
| 17/Sep | AtlasIntel, PoderData, Gerp, Datafolha and 100 Cidades | from n=2,000 to n=5,000 | FIVE national polls on the same day, three of them Tier 1. It is the largest concentration of releases ever recorded in this series, and the three with their own series measured mutually incompatible readings in the previous week. |
| 18/Sep | Veritá | n=40,500 | A sample twenty times the median of the window. Fieldwork from 13 to 18 September, meaning it straddles the Supreme Court session of 15 September, which makes it the only one of the week with fieldwork on both sides of that event. |
Registration with the TSE is not a confirmed release: institutes can delay or cancel. The table lists the twelve national polls in the window with a sample of 1,000 or more; 56 state-level registrations from the same period are left out by scope. Every registration number is searchable in the public TSE registry.
7.Watch list, week triggers
- The spread crossing back through zero The inversion is four sessions old and the boundary was already crossed twice on 12 September, in the morning and again before midday. A spread oscillating around zero is not the same as a spread that has settled on one side, and the distinction only shows in the count of captures, not in the close.
- The second Tier 1 house to repeat Datafolha The Datafolha poll of 11 September was the only one of the eight to give the runoff pair to the first-round leader, by 46 to 44. If AtlasIntel or Quaest repeat that pattern on 17 September, the price has to choose between the inversion it just priced and the two highest-reliability houses in the window.
- The 10-point dispersion in the first round The range between Gerp, which measures the runner-up 3.0 points ahead, and Quaest, which measures the leader 7.0 points ahead, is larger than the sum of the margins declared by any pair of houses. Twelve releases in five days will either narrow that range or confirm it, and both answers are informative.
- The third-place book after two shifts Renan Santos rose 11.00pp this week after rising 5.00pp in the previous one, and Ronaldo Caiado lost 6.00pp after losing 21.50pp. The order of the chasing pack changed twice in two weeks in a book of USD 760 thousand, and a move of that amplitude in a book that size does not require large flow.
- The impeachment contract on the eve of the 15 September session It spent the week between 4.40% and 8.30% after closing 4 September at 15.10%. The session is scheduled and will be public, which removes the calendar uncertainty. A book of USD 90 thousand that has already produced 11.50pp in a week has room to repeat in either direction.
8.Methodology
AFOS does not produce a weighted average and does not arbitrate between sources. It publishes the three signals side by side, contract price, registered poll and press flow, and reports the distance between them when it exists. A market price only goes out after two independent readings eight minutes apart confirm within 0.20pp, contract by contract, and the reference series are measured in the daily database backup rather than in the public interface, which caps the window at ninety days. Every claim of an extreme declares the window in which it was checked.
The polls cited are those registered with the Superior Electoral Court, with registration number, sample, margin and fieldwork dates declared, and national scope is a condition of entry into the comparison table. The code and the method are open under the Apache 2.0 licence, and the datasets are published with a permanent identifier. Every factual claim in this issue links to the primary source.
9.Additional reading · macro coverage
Complementary macro context. AFOS is the primary source on divergence between market and polling; the links below cover the economic and institutional backdrop of the window.
- Valor Econômico · Daily coverage of the 2026 election and of the fiscal impact of the electoral calendar.(paywall), valor.globo.com/politica/eleicoes-2026/
- Reuters Brazil · English-language flow on politics and the currency, useful for non-resident investors., www.reuters.com/world/americas/
- Bloomberg Línea Brasil · Market analysis focused on Brazilian assets and the political risk premium.(paywall), www.bloomberglinea.com.br/
- Poder360 · Tracking of registered polls and of the Supreme Court institutional agenda., www.poder360.com.br/poder-eleicoes-2026/
- Public TSE registry · Official search of registered electoral polls, the primary source of the registration numbers cited., divulgacandcontas.tse.jus.br/divulga/
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.