The pressure signal did not work
We built three strategies on the idea that watching a match tells you something the price does not. Measured properly, the filter selects nothing — and the best rows are the ones it throws away.
Measured on: db/037 and db/040, 494 fixtures on the full-match arm, 338 on the first half
The premise was reasonable. A live match produces shots, corners, possession and expected goals within seconds. Polymarket's price is set by people, some of whom are not watching. If a team is visibly battering the other and the score is still 0-0, maybe the over is cheap.
We built three strategies on that premise and ran them on paper for weeks. This is what the data says.
The gate selects nothing
The filter is a pressure index on a 0–100 scale. The test is simple: sort every observation by how much pressure we read, and see whether the ones we called "high pressure" did better than the ones we ignored.
Realised frequency minus our fair value, by pressure quintile, at minutes 75–88:
| quintile | 1 (quietest) | 2 | 3 | 4 | 5 (loudest) |
|---|---|---|---|---|---|
| real − fair | +4.82pp | +10.82pp | +13.84pp | +9.61pp | +7.29pp |
It peaks in the middle. A signal that worked would rise from left to right. This one goes up and then comes back down, which is what noise plus a table bias looks like.
Restricting to clean books makes it worse, not better: entries above the threshold gave +2.35pp against +5.28pp for everything below it. The gate is throwing away the better rows.
The first-half arm agrees, tested separately
Different market, different table, out-of-sample five-fold cross-validation, paired difference in log-loss, bootstrap interval — against a control that knows only the minute and the pre-match total.
- pressure as a mean: +0.00380, CI [+0.00278, +0.00498] — worse than the control, and confidently so
- pressure at its maximum: +0.00265, CI [−0.00218, +0.00750] — indistinguishable from zero
Terciles are flat within ±9pp. Whatever the live reading contains, the minute and the pre-match total already had it.
How much is actually in the box score
We measured the ceiling directly. Over and above what is free — the minute, the score, the pre-match total — the entire apparatus of shots, corners, possession and xG adds +0.0008 pseudo-R².
For scale: the Polymarket price itself carries roughly sixteen times that.
There is a version of this result that sounds like a bug in our feed, so it is worth saying that we chased that too. A separate test asked whether the movement of the ask carried anything beyond its level. It came back −0.00042, CI [−0.00104, +0.00019] — a zero.
What was left standing
Not the signal. What survived the review was something we were not looking for: book quality, which turned out to separate cleanly where pressure did not. That is its own piece.
Why the thresholds are still in the code
A fair question, given the above. The answer is that removing them is a decision about what these strategies are, not a bug fix — three arms defined by a signal, with the signal removed, are a different experiment wearing the same name. So they stay, and they stay recorded, and every row carries the control arm beside the treatment so the comparison remains answerable from the data.
None of the three is near its verdict gate, which is 200 settled entries with an interval clear of zero after fees. At the current rate that is months away. We would rather publish this now than publish a number later.
Read next
What an $85,000 Polymarket wallet actually did
One trader made $85,423 in 105 days on a book that never exceeded $29,000. Nearly all of it came from one leg, and identifying that leg needs no model at all.
The spread is the tell, not the depth
A book quoting an ask of 0.99 behind $30,117 of depth is not a deep book. It is not a book at all — and every depth filter we had passed it.
Get these by email
What we measured, what it said, and what it did not. Including the things that did not work — those are most of them.
Your address is stored and nothing else is done with it. No issue has been sent yet — there is no mailer wired up, and this page will not pretend otherwise. When the first one goes out it will say it is the first.