AFOS Analytics
Real-time political pricing
Prediction markets × polls × news
no smoothed averages
United StatesThis is the first issue with a weekly window in the United States series. Issue №1, on 31/Jul, was an opening piece: our market collection had only just been switched on, and reporting 'the week' with three days would have been selling coverage we did not have. From here on the window is the previous Monday to Friday, on the same cadence as the Brazil edition.
The 27-31/Jul window has one finding and it is the absence of movement. In the contract that pays if Democrats take the House, the price marked 85.50% on every single day, eight readings across six days of series. This is not a small oscillation: it is the same number, repeated. In the Senate contract, Republicans have been at 55.50% since 22/Jul, thirteen days without moving.
The only price that moved was the Democratic side of the Senate, from 43.50% to 44.50% on 02/Aug, already outside the window. It rose without the Republican side giving ground, and the effect is arithmetic: the two sides now add to exactly 100.00%, where they added to 99.00%. There was no change in the reading of who wins, there was a closing of the gap between the two ends of the contract itself.
On the other side, those who measure vote intention do not agree. Our house average for the House is D+5.69, across 26 polls from 23 pollsters in a 30-day window. But the distance between the highest and the lowest reading is 9 points, from D+11 at Big Data Poll and Emerson College to D+2 at Reuters/Ipsos, RMG Research and Harvard/Harris. Two first-rank houses, CNN/SSRS and Quinnipiac, were in the field on exactly the same days, 23 to 27/Jul, and came back with D+8 and D+7.
That is the crossing of the week: the price is standing still and those who measure the electorate are split by 9 points. The two things measure different quantities and neither corrects the other, but the reading of where the uncertainty sits is direct. It sits in the declared ballot, not in the price. The set of eight contracts this panel follows holds USD 16.74M.
1.Executive Summary
The week of 27 to 31/Jul had no price movement in the two contracts that decide control of Congress, and that is what this issue reports. A standing price is usually read as an inattentive market, and here the volume says otherwise: there is USD 4.93M on the Democratic side of the House alone. A deep contract that does not move is not a contract without an audience, it is a contract where the audience saw nothing that would change the answer.
The divergence of the week is in the ballot. Twenty-three pollsters measured the same race in the same window and came back 9 points apart from each other. And there is a third reading, the most interesting one in this issue: two different contracts about the Senate give different answers to the same question, 4.80 points apart. The market numbers are from 03/Aug at 17:11 UTC and cleared the capture guard, which takes two readings eight minutes apart and only releases them if they agree within 0.20pp.
2.Why this brief does not subtract the market from the poll
This brief publishes a polling average and never subtracts it from the market price. It looks like a technical detail and it is not: it is what separates a defensible reading from an invented number.
It is worth declaring what weakens this issue, and there are two things. The House series is six days long, starting 29/Jul, and six days do not distinguish a stable contract from one that has not yet had a reason to move: any superlative here is 'of the series', with the start date declared, never 'of the cycle'. The Senate series is longer, starting 06/May, and it is the only one in this piece that supports a historical comparison.
The second is about collection, and it is larger. The polling base was rebuilt today, before this issue went out. Our reading of the source page was destroying 43 of the table's 352 rows, and the effect was not losing a row: it was one pollster's name swallowing the entire row of the previous one. Because of that the Morning Consult tracker never came to exist in the base and stayed out of the average. No percentage was wrong, and that is precisely why no gate complained: the defect was in the label, not in the number. Once fixed, the reading went from 307 to 351 rows and the average fell from D+6.00 to D+5.69, because the three recovered Morning Consult rounds mark D+3, D+3 and D+4, all below the previous average. The consequence for anyone following the series: the D+5.57 published in Issue №1 came out of the defective reading and does not compare with the number in this issue. The polling series starts here.
3.Weighted scenarios for the week
The scenarios below are not our narrative: they come from a contract that already prices every band of Republican Senate seats. Where in Brazil the hypothesis would have to be built, here it is enough to read what the market pays. The bands add to 99.20%, inside the 95% to 105% coherence gate this panel requires before publishing a distribution, and the contract holds USD 2.74M.
📌 One rule of the American system changes the arithmetic: in a Senate tied 50 to 50, the tie-breaking vote belongs to the vice-president. That is why the 50-seat band counts on the Republican side.
Adding the bands of 51 seats or more (37.20%) to the tie at 50 (13.50%), which the vice-president's tie-break resolves, the seat-count contract gives Republicans 50.70%. The two heaviest individual bands on that side are tied at 13.50%, exactly 50 and exactly 51, followed by 52 at 12.25%. This is a narrow majority, not a comfortable one: everything above 52 seats adds up to very little.
The bands where Republicans end up with 49 seats or fewer add to 48.50%, and the heaviest of them is 47 or fewer, at 23.50%, nearly a quarter of all the probability concentrated in a single outcome. In other words, when this market imagines Democrats winning, it does not imagine them winning narrowly. 🔴 And this is where the two Senate contracts contradict each other, which is set out in the grid below.
The bands of 53 seats or more add to 11.45%, with 54 at 2.25% and 57 or more at 1.65%. This is the tail that practically no national poll would describe today, with the average at D+5.69, and even so there is money behind it: the 54-seat band alone holds USD 763 thousand, the largest individual volume in the entire seat-count contract. High volume in a low-probability band is an old position left open, not conviction right now.
4.Indicator Grid
| Contract | Current | Δ week | Vol USD acc. | Implied reading |
|---|---|---|---|---|
| House · Democrats control | 85.50% | 0.00pp | USD 4.93M | The same number across the eight readings of the six days of series, which starts on 29/Jul. It is the deepest contract in the set and the one that moved least |
| House · Republicans control | 13.50% | 0.00pp | USD 4.04M | Also standing still across the six days. The two sides add to 99.00%, not 100%, which is normal in a book with an open spread |
| Senate · Republicans control | 55.50% | 0.00pp since 22/Jul | USD 1.52M | Thirteen days on the same number. The series is long, starting 06/May, with a high of 58.50% on 27/Jun and a low of 49.50% on 06/May |
| Senate · Democrats control | 44.50% | ↑1.00pp on 02/Aug | USD 2.15M | The only price that moved, and already outside this issue's window. It rose without the Republican side giving ground, so the two now add to 100.00% |
| Senate by seats · Republicans with 51 or more | 37.20% | no daily series | USD 2.36M | Adding the tie at 50, which the vice-president's tie-break resolves, it reaches 50.70%. The control contract says 55.50% for the same question |
| Senate by seats · Republicans with 47 or fewer | 23.50% | no daily series | USD 168 thousand | The most likely individual band in the entire seat-count contract, and it is one of a wide Democratic win |
| Governorships · most likely Republican band | 22 or 23, at 31.00% | no daily series | USD 680 thousand | The second distribution that cleared the gate, adding to 101.65%. Holding 22 governorships or more adds to 84.15%, and holding fewer than 22 adds to 17.50%. It is the only race on this panel where the market prices continuity rather than a flip |
| Polls · average for the House | D+5.69 | no comparison with Issue №1 | not applicable | 47.77% x 42.08%, a simple average of 26 polls from 23 pollsters in a 30-day window since 04/Jul. ⚠️ There is deliberately no Δ against Issue №1: the base was rebuilt today and that issue's D+5.57 came out of the defective reading. Comparing the two would be inventing a movement that never happened |
| Polls · distance between the highest and the lowest pollster | 9.0 points | no series | not applicable | From D+11 at Big Data Poll (field 27-29/Jul) and Emerson College to D+2 at Reuters/Ipsos, RMG Research and Harvard/Harris. It is the number the average hides |
| Election held as scheduled | 96.25% | ↑0.20pp since Issue №1 | USD 269 thousand | The quietest contract in the set. The move is exactly the size of our own capture guard tolerance, meaning it is indistinguishable from noise |
5.Liquidity and market structure
6.Electoral calendar through November
| Date | Sample | Why it matters | |
|---|---|---|---|
| 04/Aug | Michigan state primary | a primary, not a general election | The race that dominated the week's coverage. It decides who Democrats put in the Senate race in a state that appears in both sides' arithmetic |
| Through August | Other state primaries | each state on its own calendar | They do not enter this panel by estimate: every state has its own calendar and its own rules, and they are not in our base. When they enter, they enter with a source |
| Sep and Oct | National trackers | no mandatory registration | Unlike Brazil, there is no advance registration of polls in the United States. There is no print date to anticipate, only the cadence of whoever publishes |
| 03/Nov | Election | the whole House, a third of the Senate | The date on which every contract on this panel settles |
There is no equivalent to Brazil's electoral court registration in the United States, so this calendar is the election's, not one of poll publication. The polls used by this brief come from an index that says which ones exist, and each number remains attributed to the pollster that produced it, with a link to the primary source. In this round, none of the 39 readings in the window was left without a primary source.
7.Watch list, week triggers
- Two contracts about the Senate disagree by 4.80 points. The control one gives Republicans 55.50%. The seat-count one, adding 51 or more (37.20%) to the tie at 50 (13.50%), gives 50.70%. It is the same question with two answers, and the difference is not thin market against deep market: the seat-count one carries USD 2.74M against USD 1.52M in the control one. **The contract with more money behind it is precisely the one giving Republicans the lower chance.**
- The House series is six days long and marked the same number on every one of them. 85.50%, with no variation at all since 29/Jul, across eight readings. Watch whether that is stability or absence of a reason, because six days do not separate the two. The first reading that moves this price is worth more than any analysis of it standing still.
- The distance between pollsters is larger than all the price movement of the period. Nine points between the highest and the lowest reading, against 1.00pp of everything the four main prices have moved since 22/Jul. Until that distance closes, the D+5.69 average is a fragile summary, and that is why this brief publishes the dispersion before it.
- The polling base was rebuilt hours before this issue, and the average changed. Our reading of the source table was destroying 43 of its 352 rows, and the Morning Consult tracker never entered the average because of it. Once fixed, the reading went from 307 to 351 rows and the average fell from D+6.00 to **D+5.69**. No percentage was wrong, which explains why no gate caught it: the defect was in the label. What to watch now is the opposite, that **Morning Consult enters the window three times while a pollster with a single round enters once**, meaning whoever publishes more weighs more in an average that declares itself unweighted. It is a question of method, it is open, and it is not settled in the middle of an issue.
- Michigan decides on Tuesday 04/Aug. It is the primary that concentrated the coverage of the outlets on the AFOS fixed list in this window: 7 of the 10 stories collected on 03/Aug are about it. The panel does not price state primaries, and their effect shows up indirectly, in whichever candidate is left for the November race, which is what the contracts in this brief measure.
8.Methodology
Market: Polymarket, read through the platform's own proxy, captured every 30 minutes and with a guaranteed daily point even when the price does not move, so the series records time and not only change. Every price cited cleared the capture guard: two independent readings eight minutes apart, released only if they agree within 0.20pp. Polls: simple arithmetic average over a 30-day window, unweighted and excluding no pollster, built from an index that says which polls exist, while each number remains attributed to whoever produced it. When the same pollster publishes more than one subsample of the same round, only one enters, but different rounds from the same pollster each enter, which is why the window holds 26 polls from 23 pollsters. ⚠️ Collection defect fixed in this round, and declared because it changed a published number: the reading of the source table flattened multi-line citations so as not to cut a cell in half, and in that flattening it treated the abbreviated reference, which has no closing tag of its own, as though it were an open citation. The effect was to sweep away the line breaks up to the next citation and destroy 43 of the table's 352 row separators. The rows did not disappear: they were absorbed into the name of the following pollster. Fixed today, the reading went from 307 to 351 rows, the 51 corrupted labels went to zero and the average fell from D+6.00 to D+5.69. No percentage was wrong before or after, and that is why the value gates caught nothing: the defect was in the label. The form gate now refuses any pollster name containing a pipe, a percentage or a cell attribute. Distributions: a distribution only reaches this page when its bands add to between 95% and 105%. In this round the Senate seat count (99.20%) and the governorships (101.65%) clear it; the House seat count (110.90%), turnout (91.40%) and popular vote margin (148.25%) stay out. Book depth is not cited: low liquidity on Polymarket does not mean the price is wrong, because the market is arbitraged continuously, and putting that number on display produces a skewed reading. What this brief does not do: it does not forecast a result, it does not say which measurement is right and it issues no investment recommendation.
9.Additional reading · macro coverage
Stories from the week in outlets on the AFOS fixed list. This brief is a primary source for market and polling; what follows is context reading. Only links that resolve to the outlet's own site are included: of the 10 items collected on 03/Aug, 4 had a canonical address and are these; the other 6 existed only through an aggregator and were left out, with the exclusion declared here.
- ABC News · El-Sayed and Stevens make their final Democratic Senate primary push in Michigan churches and union halls (03/Aug), abcnews.com/Politics/wireStory/el-sayed-stevens-make-final-democratic-senate-primary-135316495
- The Washington Post · The decisive Senate race could come down to Michigan's Black voters (03/Aug), www.washingtonpost.com/politics/2026/08/03/black-voters-divided-progressives-are-pivotal-michigan-senate-primary/
- Fox News · The first showdown in the race for Lindsey Graham's Senate seat puts the Republican hopefuls to the test (03/Aug), www.foxnews.com/politics/first-showdown-race-lindsey-grahams-senate-seat-puts-gop-hopefuls-test
- Politico · Michigan's Republican primary for governor is getting weird (03/Aug), www.politico.com/news/2026/08/03/michigans-gop-governor-primary-john-james-01020176
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.