39 Days, 100 Matches, One Ledger

What only the window could see

At 20:45:04 UTC on 15 July, Kalshi had England at 68.5 percent to reach a World Cup final. Sixty-two seconds later they were 25 percent and the draw was 58. Six minutes after that, at 20:52:35 UTC, Argentina went from 14.5 percent to 81.5 percent in one 63-second step of the same tape.

There is a box score for that match. There are highlight reels, ratings, and a hundred thousand words of reaction. None of them contain those five numbers, because none of them were watching the price. We were. For 39 days we recorded what every major prediction market and 9 sportsbooks believed about every World Cup match, at 60-second resolution across the whole board and at one-second resolution on 95 of the 100 matches, and then we graded every one of those beliefs against what actually happened.

This is that record, published in full, with the underlying dataset attached so you can recompute every figure below rather than take our word for it. It is not a tournament recap. It is the answer to a different question: what did the market believe, when did it change its mind, and what did it cost you to agree with it. We publish the misses in the same table as the hits, because a record that only contains the hits is not a record.

The record, in full

Every settled match was scored against its pregame price on a public three-way ledger. Home, draw, away, graded on regulation time, which is how these markets resolve. Here is the whole thing.

A three-way Brier score of 0.478 is the single most useful number here, and it needs a reference point to mean anything. A forecaster who knows nothing and splits every match evenly across three outcomes scores 0.667. Lower is better. So the market carried real information, a little under a third of the way from ignorance toward certainty, or 28% of that distance, across 99 matches. That is a good score. It is not a clairvoyant one, and anybody selling you the second thing is selling you something.

The more interesting reading is the calibration. When the market said a favourite was a 65 percent shot, was it right 65 percent of the time? Mostly yes, with one loud exception.

4 of the 7 bands land within about two and a half points of the truth, which is a genuinely well-behaved market. The loud exception is the 60 to 70% band: across 20 matches the market averaged 65.0% and those favourites won 80.0% of the time. 20 matches is a small sample and we will not pretend otherwise, but it is the one place in the whole tournament where the board was systematically cheap on its own opinion. The top band is thinner still, so read that row as an anecdote rather than a finding.

And then the draw. Regulation draws happened 26 times in 100 matches. The market's average draw price across those same matches was 22.9%. That gap of about three points is small, persistent, and it is the seam this entire tournament was decided through. In 100 matches the market named the draw as its single most likely outcome exactly 2 times, at Paraguay against Australia and Algeria against Austria, and all 2 of those matches ended in draws. Two for two. It almost never made the call, and when it did, it was right.

Four nights the window saw it

Every line below is a literal row from our capture, quoted with its UTC timestamp. Nothing here is reconstructed after the fact from a box score. One caveat we should state before you read them rather than after: these rows are the Kalshi leg of a cross-venue capture, not a cross-venue consensus. On 15 July in particular our Polymarket tape carried 540 rows and no usable prices at all, so nothing on this page claims that two venues moved together on that night. One venue, quoted honestly, beats a consensus we cannot evidence.

1. The 67-point snap. England vs Argentina, 15 July, kickoff 19:00 UTC.

This was the tightest board of the tournament. The startup record of our own capture for that match logged a kickoff consensus of England 35.6, draw 33.3, Argentina 31.1, all three outcomes inside five points of each other. Then, in fourteen minutes, the market changed its mind twice.

Sixty-three seconds separate the third and fourth rows. In that minute the market moved from believing the most likely result was a draw to believing Argentina were an 81.5 percent chance. That is what a winning goal looks like when you are pointing a camera at the price instead of the pitch.

2. The suffocation. France vs Spain, 14 July, kickoff 19:00 UTC.

France went in as the nominal favourite at 37.0 percent, which is to say barely a favourite at all. The market took twenty-five minutes to abandon that position completely, and then spent ninety more never once revisiting it.

This is the shape of a favourite genuinely beaten rather than merely held, and it is visually unmistakable on the tape: one violent repricing early, then a slope with no recovery in it. Compare that against the third night.

3. The one that came back. France vs England, 18 July, kickoff 21:00 UTC.

The third-place match is the game nobody watches, and it produced the most violent price action of the tournament. France opened as a 56.3 percent favourite. England then went to near certainty, and gave almost all of it back.

A market that is 97.5 percent certain and then falls to 40.5 percent and then returns to 99.5 percent inside seventy minutes is not a market being stupid. It is a market doing exactly its job, repricing on information as fast as the information arrives. The value of recording it is that the 97.5 print and the 40.5 print both existed, and anyone who only saw the final score will tell you the result was never in doubt.

4. And the final, which the market called correctly and still got wrong.

Spain against Argentina on 19 July produced the rarest read of the whole tournament: total agreement. Polymarket, Kalshi and 9 sportsbooks landed on Spain 41.9, draw 31.3, Argentina 26.8, a spread of 0.33 percentage points across every venue we track, on roughly 4.6 million dollars of Polymarket liquidity. 11 independent pools of money, no outcome above 42 percent, and nothing to arbitrage.

Then the market spent two hours quietly walking to the one conclusion it had refused to reach all summer.

The pregame favourite was Spain, so the ledger records the final as a miss. Live, the market was excellent: it found the draw, committed to it at 88 percent, and was right. Spain lifted the trophy after regulation. Both of those statements are true, and the reason we publish both is that a scoreboard which only grades the pregame price would tell you the market failed on the biggest night of the tournament, and that is not what the tape says.

Which sets up the number that closes the record. Across the last four matches of the 2026 World Cup, the pregame favourite did not win once.

0 for 4, in the four matches where the most money in the world was paying attention. Read the full settled scoreboard if you want all 100 lines, or take the raw file and check our arithmetic.

Where we were wrong

This is the section that makes the rest of the article worth reading. Four corrections against ourselves, and one honest limitation.

Correction 1: we overstated our own pattern.

In The Window Pays Off we wrote that the two semifinals were “the first regulation favourite defeats of the entire tournament.” The full settled record says that is false. There were 8 earlier ones, starting with Australia beating a 56.5% Turkiye on 14 Jun. What was true, and what we should have written, is the ratio: through the quarterfinals the tournament produced 25 regulation draws against 8 outright favourite defeats. When the chalk cracked, it stalled about 3.1 times more often than it broke. That is a real and unusual pattern. The sentence we published around it was not.

Correction 2: we had a fee schedule wrong for a week, and a counterparty caught it.

Our entire product rests on showing the all-in cost of a position, so a stale fee table is not a cosmetic error, it is the product being wrong. On 10 July we discovered that Polymarket had raised its sports taker fee from 3 percent to 5 percent and that our pricing layer had been carrying the old schedule. For roughly a week, every Polymarket sports position on Tater looked cheaper than it was. We did not find it in a monitoring alert. An exchange counterparty asked us, in passing, whether our lens accounted for the increase. Fee schedules drift silently, our validation had been a snapshot rather than a monitor, and we are building the drift check that should have existed first.

Correction 3: our own live brief invented a divergence that did not exist.

During the final, at 21:18 UTC, our automated brief generator published this: “Kalshi lurched suddenly toward the draw, pushing consensus to 99.5 percent, but Polymarket Argentina still remains at 84.3 percent, an 84pp gap the tape does not explain.” There was no gap. Our Polymarket mapping for that market had the draw token sitting in the Argentina slot. The leg quoted at 84.3 was bid 84.0 and offered 84.5 on the draw, and it settled at 99.75 four minutes later, which is precisely what happened on the pitch. The prices were correct. The label was wrong, and the label is what the reader sees. It is fixed, and we are stating it here rather than deleting the brief.

Correction 4: the first draft of this article got its own headline number wrong.

The draft of this piece carried a favourites record of 64-34 and a footnote claiming that two group-stage fixtures had no resolvable pregame price. Neither was true. The record is 65-35 across all 100 matches, every one of which carries a priced favourite, and the footnote had been invented to reconcile an arithmetic slip. The cause was mundane and worth naming: the figures had been transcribed by hand from an earlier snapshot, and then the ledger moved and the article did not. This version transcribes nothing. The record, the averages, the Brier score, the calibration bands, the draw counts, the coverage counts and both tables above are computed at build time from the published dataset, so the page cannot disagree with the file even if we want it to. We spent the tournament arguing that a claim without a checkable number behind it is worth nothing. It would be a poor look to exempt ourselves.

The limitation: coverage was uneven.

Of the 100 settled matches, our stored pregame snapshot carries a full three-way Kalshi price on 99, a Polymarket price on 31, and sportsbook prices on 37. All three venues appear together on 30. Our cross-venue claim is strongest on the knockouts and the marquee ties, where every venue had a market, and it is thinnest in the group stage, where a Curacao fixture simply was not listed everywhere. The live tape is broader than the pregame snapshot, but the honest headline is that the 100-match Brier score is mostly a Kalshi score. We would rather you know that than find it yourself.

Two methodological notes in the same spirit. Every venue tape flattens to 0.500/0.500/0.500 once a market settles or suspends. Those rows are settlement artifacts, not prices, and no extreme quoted anywhere in this article or any of our coverage is taken from them. And the Brier score above is computed from the per-venue pregame snapshots in the published dataset, which is what you can reproduce yourself. Our scoreboard generator carries its own summary figure of 0.477, because it scores a slightly different close-window blend. The difference is one thousandth. We are telling you rather than quietly picking the prettier one.

And where the market was wrong.

The market's five heaviest losses all had the same shape: a big favourite failing to win in regulation, and every one of the five ended in a draw rather than an upset. Nobody beat these teams. They were held.

The heaviest of them: Argentina were priced at 85.3% and the match was drawn. The market gave that outcome a 10.8% chance. A 10.8% event is not a scandal, it is a Tuesday, and it happened often enough across 39 days to move the whole tournament. This is the practical meaning of a 0.478 Brier score: informative, and nowhere near safe.

What it cost to agree

Everything above is about belief. The other half of our record, and the half nobody else publishes, is price. A market's stated probability and the amount you hand over to take that side are two different numbers, and the gap between them changes by venue.

Take the final. Spain was offered at 42.5 cents on both major prediction venues before kickoff. Identical stated belief. Not identical cost.

Half a cent sounds trivial until you look at the shape of the fee. Because it is charged on P times one minus P, it is heaviest in the middle and lightest at the edges in absolute terms, but as a share of what you actually stand to win it runs the other way. On Kalshi a 50-cent contract costs 1.75 cents in fees against 50 cents of potential profit, which is 3.5% of the upside. A 90-cent favourite costs 0.63 cents against 10 cents of upside, which is 6.3%. The fee is close to twice as expensive, in profit terms, on the safe side of the board. Backing favourites feels conservative and is quietly the more heavily taxed position.

Sportsbook prices work differently again: what you see is already all-in, with the margin baked into the odds rather than charged on top. That is why comparing a Kalshi cent price to a sportsbook decimal without adjusting for either is meaningless, and why we built the two-layer lens in the first place. If you want the full mechanics, they are in How to Read Tater.

The lens was never about football

Nothing in the previous six sections is a football technique. Not one line of it. Recording every venue at once, blending them into a single honest read, catching the second the market changed its mind, grading the pregame price against the settled result, and separating what a market believes from what it charges you to agree. Those are properties of markets, not of sport.

The World Cup was our proving ground, and for a while we mistook it for our product. It was a good place to prove things because it was hard: 100 fixtures in 39 days, three-way markets rather than binaries, venues that listed different subsets of the same tournament, and a hard finish line to be judged against. We finished it with a public ledger, a downloadable dataset, and four corrections we wrote against ourselves.

The same ledger is now pointed at the rest of the calendar, and the next marquee arc is the November midterms. That is not an arbitrary pick. It is the arc where this lens has the least competition anywhere on the board. US political event markets sit outside the sportsbook affiliate channel, so the whole incentive structure that produces odds content for sport does not reach them. We have not found a forecast-model publisher that puts the all-in cost of taking a side next to its probability. And political markets on the two regulated venues carry a different fee schedule again, 4 percent rather than the 5 percent that sport pays, so the same belief prices differently in a way essentially nobody is showing you.

Two exchanges disagreeing by a cent or so per dollar on control of the House, with the fee model made explicit, the dataset attached, and the whole thing graded in public afterward including the misses, is a story that does not currently exist. It is the same story you just read, with the football taken out. Everything in this article stays online permanently, because a record you can check is the only reason to believe the next one.

The live board is at taterit.com/pulse, the same one that watched two semifinals and a final decide themselves. The midterms ledger is being built in the open at Tater Research: Midterms, and the World Cup archive stays exactly where it is, at /research/world-cup.