AFOS Analytics
Real-time political pricing
Prediction markets × polls × news
no smoothed averages
BrazilThe week of 14 to 18 September did not undo the previous week's inversion, it firmed it up, and it started out doing the opposite. At 12:30 UTC on Monday the gap between Flávio Bolsonaro and Lula in the presidential contract stood at 1.60pp, the narrowest distance since the lead changed hands, and the inversion priced in the previous week had almost vanished in a single session. Over the next four sessions it moved the other way and closed Friday at 14.00pp, a shift of 7.60pp across five sessions, the sixth largest of the 23 weeks with four or more sessions in the series opened on 14 April.
This week's mechanics are not last week's. The price leader rose 4.60pp, from 51.90% to 56.50%, and his opponent gave up 3.00pp, from 45.50% to 42.50%: both sides moved, but the one on top moved more, whereas the previous week the two had moved almost half and half. The first-round runner-up contract confirmed the mirror move for the second week running, and that is the internal confirmation that the book repriced who wins: Flávio Bolsonaro lost 5.50pp of probability of finishing second and Lula gained exactly the same 5.50pp.
The ten national polls released in the window say two different things at the same time, and that is the week's information. In the first round nine of the ten keep Lula ahead, and the only exception is Gerp. In the runoff pair the count inverts: six of the ten put Flávio Bolsonaro ahead, and the interval between the extremes is 14.3 points, from CNT/MDA, which measures Lula by 7.3, to Gerp, which measures Flávio Bolsonaro by 7.0. Five of them were released on the same day, 17 September, and on their own they span 9.0 points of that interval. The presidential contract prices the runoff, and it moved with the six.
1.Executive Summary
The week's theme is the separation between the two rounds, and it became measurable because the window carried ten national polls instead of the two or three of an ordinary week. The first round is almost unanimous in one direction, with nine of the ten putting the previous leader ahead. The second is a majority in the opposite direction, with six of the ten giving the pair to the first-round runner-up. Both statements are compatible, because they measure different questions, and the transfer of votes from eliminated candidates is what separates one from the other. The presidential contract pays whoever wins the election, that is, the runoff. What the week adds is not the direction, already priced in, it is the size of the disagreement between pollsters on the same question on the same days.
2.Why AFOS does not smooth
The ten national polls released between 14 and 18 September asked the same runoff pair, with overlapping field dates and samples of 2,000 to 5,018 interviews. A weighted average of them returns half a point of advantage to one side, a number that reads as a tie and that erases the window's central fact: the pollsters do not disagree about magnitude, they disagree about who is ahead.
Why it matters: the pollsters at the extremes declare margins of 2.2 and 2.0 points, whose sum is 4.2, and the interval between them is 14.3 points, that is 3.4 times larger. The difference is not explainable by sampling error, and the average of 0.6 point describes a tie that six of the ten pollsters did not measure.
Issue №17 assigned about 55% to the base case, in which the spread would settle between 2pp and 10pp in favour of whoever had flipped, without breaking the 14.30pp of the previous week. The direction was right and the band was narrow: four of the five closes landed inside it, between 6.40pp and 9.80pp, and Friday's came in at 14.00pp, above the declared ceiling and still below 14.30pp. About 30% went to the contrarian case, which expected the spread to cross back through zero after the five prints of 17 September. It did not happen, and it missed narrowly on the first day: at 12:30 UTC on Monday the spread stood at 1.60pp, and from there on it only widened. About 15% went to the tail, which expected the Supreme Court session of 15 September to put the impeachment contract back above 8%. Here the distinction matters: the contract touched 9.00% at 09:30 UTC that day, above the band, and closed the day at 5.00% and the week at 4.60%. Measured at the close, the tail did not happen; measured in the recorded readings, it happened for part of one session and did not hold.
Erratum to issue №17. That calendar listed twelve national polls scheduled in the window and ten were published. The two that were missing deserve a record for the reason. American Analytics, scheduled for 16 September, had its registration pulled from the electoral court's public archive on the very day it was due to release. Veritá, scheduled for 18 September, went in with the line n=40,500 and the reading that it was a sample twenty times larger than the window's median. That sample is not from a national poll: the registration of 40,500 interviews is the consolidation of twenty-seven state registrations from the same house, one per federal unit, and the sum of the twenty-seven comes to exactly that number. The two genuine national polls from Veritá that sit in the panel's comparative table, from 21 August and 6 September, carry 3,840 and 3,804 interviews. The number that should have raised the doubt was in our own table, ten times smaller.
3.Weighted scenarios for the week
The week of 21 to 25 September repeats the shape of the previous one with an even heavier day of concentration. Three national polls release on Monday the 21st, among them Quaest and NEXUS, which opened the last window, and the registry carries six national polls scheduled for 24 September. The three houses that opened the previous window get a comparable reading within their own series again.
The national polls of 21 and 24 September keep the previous window's split, with a simple majority giving the runoff pair to the current price leader and a consistent minority giving it to the opponent, and dispersion between pollsters stays above 10 points. In that case the spread settles between 12pp and 20pp, which is the band it entered the week in, and the book stops repricing on every isolated print. The first round keeps giving the previous leader the lead at most pollsters, because the transfer of votes from eliminated candidates is what separates the two questions.
The six national polls of 24 September repeat as a block what Datafolha, CNT/MDA, BTG/Nexus and Indexa measured in the previous window, and the runoff pair goes back to the other side in the majority. Datafolha came in three rounds running with the same 46 to 44 and is the window's pollster with the highest reliability declared in the panel; CNT/MDA measured 7.3 points on its side, with a margin of 2.2. If two of the three largest samples on the 24th confirm the pattern, the winner contract has to choose between the inversion it consolidated over two weeks and the block that does not follow it, and the spread has already crossed from 1.60pp to 14.00pp within the same week.
The impeachment contract goes back above 8% and this time closes the day there. The distinction matters because on 15 September it already touched 9.00% and gave it all back in the same session. The book has USD 92 thousand accumulated and took in USD 894 over the whole week, so it is thin enough for isolated orders to produce double-digit shifts, and the 5.40pp range between 3.60% and 9.00% happened on that volume. What separates this tail from the previous one is that it requires persistence, not a touch.
4.Indicator Grid
| Contract | Current | Δ week | Vol USD acc. | Implied reading |
|---|---|---|---|---|
| Presidential spread, Flávio minus Lula | +14.00pp | +7.60pp | USD 151.46M | Sixth largest shift of the 23 weeks with four or more sessions. Path: +6.40, +6.40, +9.80, +9.70, +14.00, with an intra-week low of 1.60pp at 12:30 UTC on Monday. |
| Flávio Bolsonaro, presidential | 56.50% | +4.60pp | USD 11.02M | Intra-week low of 50.10% on Monday and high of 56.50% at Friday's own close. Took in USD 339 thousand over the window. |
| Lula, presidential | 42.50% | -3.00pp | USD 11.36M | Intra-week high of 48.50% on Monday and low of 42.50% at 03:30 UTC on Friday. Took in USD 347 thousand, the book's largest single inflow. |
| Flávio Bolsonaro, 2nd place in the 1st round | 70.00% | -5.50pp | USD 0.95M | Second week running mirroring the winner contract. High of 79.00% at 10:00 UTC on Thursday, before giving up 9.00pp over two sessions. |
| Lula, 2nd place in the 1st round | 29.30% | +5.50pp | USD 0.81M | He gained exactly what his opponent lost, and the 5.50pp symmetry is the internal confirmation that the repricing was about position. |
| Renan Santos, 3rd place in the 1st round | 49.00% | +11.50pp | USD 0.35M | Jump from 39.00% to 49.50% in a single session, between Wednesday and Thursday. High of 53.50% at 04:00 UTC on Friday, range of 18.50pp. |
| Ronaldo Caiado, 3rd place in the 1st round | 12.00% | -3.50pp | USD 0.14M | Third week running of decline. The whole book holds USD 852 thousand accumulated, so a double-digit move here does not require large flow. |
| Supreme Court impeachment before 2027 | 4.60% | -0.70pp | USD 0.09M | Touched 9.00% at 09:30 UTC on 15 September, the day of the Supreme Court session, and gave it all back the same day. Range of 5.40pp on USD 894 of new money. |
5.Liquidity and market structure
Three of the book's five largest accumulated volumes belong to contracts that closed the week paying 1.00% or less, and the largest of them all, Tarcisio de Freitas with USD 14.08M, took in USD 4 thousand over the window, or 0.3% of the new money. The flow tells another story: the book took in USD 1.63M, and after the two names that moved come Romeu Zema, with USD 278 thousand, and Jair Bolsonaro, with USD 137 thousand, both closing at 0.20% without leaving that level across five sessions. Together they took a quarter of the week's new money without producing any price change. The reading holds for both extremes: accumulated volume is trading history since the contract opened and is not current conviction, and anyone who reads the volume ranking as a probability ranking is reading the book backwards.
Volume is accumulated since each contract opened. The week's new money is the difference between the volumes recorded at the closes of 14 and 18 September, and the prices in the table are those of 18 September. AFOS does not cite 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 | |
|---|---|---|---|
| 21/Sep | Quaest, NEXUS and American Analytics | n=2,004, n=2,000 and n=2,000 | The first two opened the previous window and return with field dates of 17 to 20 and 18 to 20 September, which gives a comparable reading within their own series. The third is the house whose earlier registration left the electoral court's archive on 16 September. |
| 23/Sep | AtlasIntel | n=5,000 | Largest sample of the week, field dates of 17 to 22. In the previous window it measured the narrowest runoff distance among the ten, 0.4 point, with the smallest declared margin, 1.0. |
| 24/Sep | Datafolha | n=2,002 | Field dates of 22 to 24, the shortest and the most recent of the day. It measured three rounds running with the same 46 to 44 and on its own sustains the minority side among the Tier 1 houses. |
| 24/Sep | PoderData and PALVER | n=3,000 and n=5,000 | PoderData measured 2 points of distance in two rounds running. PALVER comes in with no reading in the previous window, so it is a test of level and not of trend. |
| 24/Sep | 100 Cidades and ALFA | n=2,000 and n=2,700 | The first measured 4.4 points in favour of the price leader in the previous window, and it was the reading that inverted the Datafolha of the same day. |
| 24/Sep | Real Time Big Data | n=2,000 | It enters with the label under review: the national scope comes from the sampling-plan field, and this house's plan declares national in 18 of 18 registrations that the methodology field resolved as state-level. |
Registration with the electoral court is not confirmed release, and the previous window showed both ways a promise fails to become a publication: of twelve national polls scheduled, ten came out, one had its registration pulled on the day it was due to release and another carried as its headline figure the sum of twenty-seven state registrations. The table lists national polls with samples of 1,000 or more and excludes one registration already pulled from the court's archive with release scheduled for 23 September.
7.Watch list, week triggers
- The spread at 17.55pp as the week opens The certified reading of 21 September at 03:06 UTC puts Flávio Bolsonaro at 59.05% and Lula at 41.50%, or 17.55pp. The value exceeds the high of 17.50pp recorded at 01:30 the same day, which was already the largest of the 641 paired captures since 14 April.
- The three houses measuring against themselves Quaest, NEXUS and AtlasIntel have a round this week and one in the previous week, with roughly a week between field dates. It is the first window in which dispersion can be measured house against itself, and that calculation isolates movement in the electorate.
- The house that left the registry and returns today American Analytics had its 16 September registration pulled from the electoral court's archive on the very day of the scheduled release, and the same house has a registration set for today, with the same signature of sample, cost and methodology. The registry does not state a reason for a withdrawal, and an identical signature does not prove re-registration.
- Third place after two double-digit weeks Renan Santos rose 11.50pp after 11.00pp the previous week, and reached 53.50% in the early hours of Friday. The book holds USD 852 thousand accumulated and took in USD 88 thousand. Two weeks like that in a book of that size either establish a level or precede a giveback.
- Impeachment and the difference between touching and closing The contract touched 9.00% on the morning of 15 September and closed the day at 5.00%. As long as the book takes in less than a thousand dollars a week, excursions of that size will keep showing up in the recorded readings and vanishing at the closes.
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 news flow, and reports the distance between them when it exists. A market price goes live only after two independent readings eight minutes apart that agree within 0.20pp, contract by contract, and the reference series are measured in the daily backup of the database and not in the public interface, which caps the window at ninety days. Every claim of an extreme declares the window in which it was checked, and every value later than the close of the covered week carries the hour of the reading.
The polls cited are those registered with the Superior Electoral Court, with protocol, sample, margin and field dates declared, and national scope is a condition of entry into the comparative table. When the scope label depends on a text field that the house repeats without distinguishing, the issue says so in the line instead of publishing the label as fact. The code and the method are open under the Apache 2.0 licence. Every factual claim in this issue links to its primary source.
9.Additional reading · macro coverage
Complementary macro context. AFOS is a primary source on divergence between market and polls; the links below cover the economic and institutional backdrop of the window.
- Valor Econômico · Daily coverage of the 2026 election and the fiscal impact of the electoral calendar.(paywall), valor.globo.com/politica/eleicoes-2026/
- Reuters Brazil · English-language flow on politics and currency, useful for the non-resident investor., www.reuters.com/world/americas/
- Poder360 · Registry of electoral polls with protocol and history by pollster., www.poder360.com.br/pesquisas/
- Gazeta do Povo · Coverage of poll releases with verification of the electoral court registration., www.gazetadopovo.com.br/eleicoes/
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.