AFOS Analytics

📌 Jul 28 was the inverse of the day before, and the inversion is the finding itself. On Monday there was national polling and the favourite's price stood still. On Tuesday there was no national polling and the price rose. Lula went from 62.50% to 63.50% (vol USD 7.69M), up 1.00pp, and touched the ceiling of the AFOS series again, which is 63.50%, marked on Jul 26 at 22:30 UTC. It MATCHES the ceiling, it does not pass it, and the difference between those two sentences is exactly the error the panel made on Jul 26 and corrected on Jul 27. The gap over Flávio went to +39.65pp, still below the widest gap in the series, +39.80pp, from those same early hours. Flávio held at 23.85% (vol USD 7.69M) for a second session. Since the price moved with no national polling to anchor it, the panel records the move and does NOT pick a cause: the same session carried the report that Trump will extend the national emergency over Brazil by one year, the Brazilian government's request for consultations at the WTO, three state polls and the echo of a public statement about re-election probability made on Jul 24. Picking one of those would invent precision the data does not have. THE DAY'S POLLING was state-level, which is why it does not enter the panel's national set, but its content matters: the race is geographically split. Genial/Quaest in Pernambuco (n=900, fieldwork Jul 22 to 26, 3pp margin, BR-03810/2026 and PE-09649/2026) has Lula at 55% against 21% for Flávio in the first round. Genial/Quaest in Minas Gerais (n=1,482, fieldwork Jul 22 to 26, 3pp margin, BR-09333/2026 and MG-03490/2026) has a statistical tie in both rounds, with 30% to 29% in the first and 38% to 35% against Flávio in the runoff, and with Romeu Zema numerically AHEAD of Lula by 39% to 37%, inside the margin. Real Time Big Data in Rio de Janeiro (n=2,000, fieldwork Jul 23 to 27, BR-06074/2026) has Flávio at 46% against 42% for Lula in the runoff. Adding up what Poder360 compiled from the published state polls, Lula leads in 12 states and Flávio in 7. THE RARE CROSS-READING of this week is not between market and poll, it is between market and model, and it deserves a separate note, with the caveat that the remark is from Jul 24 and not from today. Felipe Nunes, director of Quaest, said at Expert XP that Friday that Lula's probability of re-election went from around 38% to close to 60%, in a model tying government approval to re-election probability, with the turn starting in May (InfoMoney, Jul 25; Brasil 247, Jul 27). Polymarket is at 63.50%. It is the first time in this panel that the two sources estimate the SAME quantity: probability of victory. Normally the panel crosses implied probability with voting intention, which are not the same thing and therefore cannot be subtracted from one another. Here they can be compared, and the panel shows them side by side without subtracting one from the other, because close to 60% is not an exact value. That validates neither: they are models with different assumptions that landed close, and closeness is not proof of being right, just as divergence would not be proof of error. The third axis is the third-place realignment, now in its second day. Caiado went from a range of 19.00% to 19.50% up to 26.50% in third place in the first round, which adds 11.00pp over two sessions counting the 15.50% of Jul 26, while Renan Santos fell from 62.00% to 60.50%. And here the panel needs to separate two things that look like one: in the WINNER contract, Caiado FELL 0.05pp, to 1.90%. In other words, the market is repricing who finishes third and not who wins the election. Renan, on the other side, fell in ALL THREE books where he appears, 8.75% in the winner contract, 12.15% in second place and 60.50% in third, which is different from yesterday's mixed signals and describes money leaving on every front. The factual counterpoint on Caiado came from Poder360 and is geographic: he leads in Goiás, the only state where the third way looks competitive, and Goiás is worth 3.2% of the national electorate. That does not contradict the national statistical tie BTG/Nexus measured on Jul 27, and the panel shows both with the same prominence. The fourth axis is institutional and carried more news than the day before, with the price moving against intuition. Moraes ordered the PGR to weigh in on Flávio Bolsonaro's cancelled testimony in the inquiry into alleged slander against Lula, after the senator failed to appear at the Federal Police and sent a written defence arguing that the president again linked him to traitors of the homeland and mentioned hanging. Judges' and lawyers' associations responded formally to Javier Milei's attack on Moraes. And the Banco Master congressional inquiry route remains dammed: O Globo reported that Kassio Nunes Marques has for four months been sitting on the writ that decides whether the committee is installed, and Edson Fachin rejected, as filed out of time and without ruling on the merits, the request by four senators to remove him from the case. In the Master case, Valor Econômico published an exclusive report that André Mendonça authorised the Federal Police to trace abroad the assets of Daniel Vorcaro and others involved, information that enters attributed to the outlet and without an independent second source. Even with that accumulation, the market on the impeachment of an STF justice FELL 0.60pp to 2.80%, in a book of USD 83 thousand. Recording the direction is mandatory; building a narrative on it is not, because that book moves on very little money. The fifth axis is inflation, and it comes with two caveats before the number. The modal band flipped: 4.50% to 4.99% went to 39.25%, against 26.30% yesterday, and moved ahead of the 5.00% to 5.49% band, which sat at 38.15%, with 5.50% to 5.99% at 12.20%. In one sentence, the centre of the distribution stepped down one notch and came up against the ceiling of the target instead of sitting half a point above it. The caveats: the whole book adds up to USD 79 thousand, and the inflation market is NOT covered by the double-reading lock, which watches only the five electoral books. That number did not go through the same verification as the others and enters with the warning. A capture note, this time a positive one: the lock PASSED today, without a single divergence across the five books it covers, unlike yesterday's two blocked rounds. Winner and placement prices all enter as firm, and there is no range in this update. Finally, a record of an absence, which is data as much as a presence: AtlasIntel had publication declared to the TSE for today, with n=5,000 and fieldwork Jul 22 to 27 (BR-08602/2026), and would be the largest national sample in the set. It was NOT released by the close of this capture, and no story about it appeared among the 927 items collected today. A TSE filing is not a guarantee of publication, and the panel does not cite numbers from a poll that did not come out. The week still brings PoderData on Jul 30 (n=2,400) and Vox Brasil on Jul 31 (n=2,100) to say whether the 63.50% level has polling behind it.