Methodology
Anteo reports what real-money markets believe. That is only worth anything if you can check how we got there — so here is the whole method, including the parts we haven't finished.
Where the numbers come from
Every probability on this site traces back to a market where people have money at risk. We read prices from prediction markets, where a contract's price is the market's probability: a contract trading at 31¢ means the market puts the outcome at 31%. Nothing is a house estimate, and nothing is a poll.
When more than one venue prices the same event, we keep all of them and show where they disagree. The disagreement is not noise to be averaged away — it is often the most informative thing on the page.
Why sportsbook odds get corrected
Sportsbook odds are not probabilities. A book quoting −110 on both sides of a two-way market is implying 52.4% on each — a total of 104.8%. That extra 4.8% is the book's margin, called the vig, and it is revenue, not belief.
Before a book can be compared to a prediction market we strip that margin, so its outcomes sum to exactly 100%.
HOW THE MARGIN IS REMOVED →
remove the 4.76% margin → 50% / 50%
We use a power method rather than simple proportional scaling, because books load relatively more margin onto longshots — proportional scaling leaves that bias in and systematically understates favourites.
One event, one clock, one delta
Numbers drift when they are typed in more than one place. So no displayed figure on this site is written by hand: percentages, volumes, deltas, and chart geometry are all derived from the same underlying record, and a reconciliation pass re-derives them and checks them against what is shown.
THE THREE RULES →
- A binary market's outcomes sum to exactly 100%.
- A price chart ends at the price printed beside it.
- A quoted move equals the change across the series actually drawn — it is recomputed from the line, never asserted separately.
How The Brief is composed
The Brief is fifteen stories, and it is selective on purpose: a general-interest front page, not the top of every section. The ranking is still the activity behind each market weighted by how far the price actually moved — dollars where a venue reports dollars, contracts where it reports contracts, each measured against the busiest market in its own unit and never converted into the other — but an eligibility standard decides what may appear at all, and because a composed page presented as a raw leaderboard would be a quiet lie, every rule is stated here and on the page itself. Sections carry each category in depth and Search reaches the wider catalogue; nothing here removes a market from either.
THE RULES →
- General-interest eligibility. The test: would a broadly informed reader who does not follow this category expect to know about the event? Geopolitics, economics and politics generally qualify; sports only for events non-fans know are happening (a Super Bowl, a World Cup, an Olympics — never a fixture, a line, or an ordinary championship outright); crypto and technology only for regulatory, structural or historically significant developments, never daily price action or leaderboards; esports and markets the taxonomy cannot place do not qualify. The rules are fixed patterns anyone can read — no model, no score, no per-market judgment call — and each market gets a stated reason either way.
- A live question with a reading. A market whose event has already happened is paperwork, not news, and leaves the front page a day after the venue's own end time whatever its price says. A market whose book is too wide to read a probability from has the question and not the answer, and does not lead with it. Both stay in their sections, in Search and at their own address.
- One story per event. Where the venue itself files several markets under one underlying event — every strike of one Fed meeting, every candidate in one race — the event takes one front-page position. A venue-defined race is shown as the race: the event's own title with its leading options priced, grouped on the venue's own event structure. A genuine standalone question stays standalone.
- One story across venues, two markets. When two venues price the same question, they stay two markets — each with its own price, its own settlement terms, its own history and its own page; we never average them or call one the truth. The front page headlines the question once, with the other venue as a row under it carrying that venue's name and own figure, and states any difference as a difference. Two markets are grouped this way only when the question is identical word for word, the venues' own clocks agree, the shape and section match, and a runner matches a runner — never on similar wording alone. When any of that is missing, two stories stand.
- No section sweep. At most three stories from one section in the top ten, and at most five on the whole page. Both caps defer, never remove; the rules know no section's name and reserve nothing for anyone. When the eligible pool cannot fill fifteen slots from other sections, the caps yield rather than leaving the front page short — the page fills to fifteen whenever the pool supports it and never pads itself with ineligible stories to reach a number.
What these rules never do: score a market's importance by opinion, promote a story the ranking did not, or hide a market. The eligibility standard is deliberately an editorial position — stated, fixed, and the same for every reader — and every open market remains at its own address, in its section, and on the everything view, whatever the front page shows.
What we will not draw
A line on a probability chart is a factual claim about what a market did. So when we have not observed a market long enough to know, we say so rather than drawing something plausible. A market we have seen once shows a flat line reading NO HISTORY YET — not an illustrative squiggle.
Charts are also plotted on an even time grid, so a market that sat still for an hour and then jumped is drawn that way, rather than as a smooth climb. Today's move means since today's session open, not a rolling twenty-four hours that quietly borrows from yesterday evening.
How we score ourselves
We report calibration, not a win rate. Anyone can look accurate by only calling near-certainties. Calibration asks a harder question: when the market said 70%, how often did it actually happen?
Across 44 resolved markets since 2026-08-03: when the market said 46%, it happened 48% of the time.
The Scoreboard carries the curve behind that figure — every price band, and how many calls sit in each.
Misses are published in the same place and the same format as the hits. The most recent logged miss: US CPI prints below 3.0% year-on-year? — priced at 16% — resolved Yes.
What we don't know yet
Price history is recorded from the moment a market enters our feed, so a market we picked up an hour ago has an hour of history — not a month. Longer ranges fill in as we observe them.
Cross-venue matching groups markets by the wording of the question, and it is deliberately conservative: it would rather show the same event twice than merge two questions that only look alike. When we cannot be sure, the venues are simply not compared. If you spot a wrong grouping, that is a bug worth reporting.
WHEN TWO VENUES COUNT AS THE SAME QUESTION →
Venues word the same bet differently — an exchange asks “Detroit beat Cleveland?” and answers yes or no, while a book quotes the two team names. We line those up only when the mapping cannot be wrong: a clean head-to-head, both sides named, the question leading with one of them, and nothing left over once the names are removed. A margin bet that mentions both teams looks identical to a moneyline by every other test, so it is refused. Where a quote has been restated, the Signal says which venue said what.
Resolution is the venue's, not ours. We report what a market settled at; we do not adjudicate outcomes ourselves.
We measure conviction, not attention. Conviction is money at risk and we read it from prices; we have no source for social attention at all, so nothing here shows the two side by side.
Once an observation or a grade is recorded it is never edited or deleted. A correction appends a new row that supersedes the old one, so the earlier claim stays visible.
WHY THE RECORD CANNOT BE REWRITTEN →
Not by us, and not by anything holding a connection to the database, which refuses the attempt outright rather than trusting the application to behave. A record that can be quietly rewritten is not a record, and a calibration figure computed from one would be worth as little as a typed one.