The spread is the tell, not the depth

We ran a live in-play strategy to exactly 100 paper entries and then stopped to look at it. The yield was −2.07% gross, about −4.6% after Polymarket's taker fee, with a confidence interval of [−22.4%, +18.2%].

That interval is the first thing worth saying out loud. It is ±20 points wide against an effect we would be happy to find at 2–4 points. A hundred bets cannot answer the question the strategy was built to ask. So the settlement arm was set aside, and everything below comes from the calibration arm instead — one row per fixture, minute and score, which has two orders of magnitude more power because it does not have to wait for a bet to resolve to learn something.

What the calibration arm found

We compared the realised frequency of the event against the price you could actually have paid — real − ask, clustered by fixture so one busy match cannot carry the result. Split by how wide the book was at that moment:

spread at the quote	real − ask
0–3pp	+3.64pp
3–6pp	+0.92pp
6–10pp	−4.26pp
20pp+	−38.38pp

6,449 observations across 494 fixtures.

The bottom row is not a market. It is 430 observations on 221 fixtures, quoting an ask near 0.90 on an outcome that resolved at 0.529. It is a single sell order parked a long way from any bid, with nothing on the other side to argue with it.

The quote that changed what we filter on

CA Mineiro against EC Vitória, 29 August, fifteen minutes in. The first-half over 0.5 book quoted:

• bid 0.55
• ask 0.99
• on $30,117 of depth

Two minutes later it traded at 0.56.

Every depth filter we had passed that quote, because by depth it was one of the healthiest books on the board. A liquidity floor asks how much is there. It never asks whether the two sides agree on anything, and on a thin in-play market that second question is the whole of it.

Depth still matters — but only on one of the three

To be fair to depth: on the full-match over arm it did come out monotone, and the floor was raised from $50 to $1,000 on the strength of it — −4.72pp under $200, −0.49pp to $1,000, +3.63pp to $5,000, +4.61pp above. On the two first-half arms the same cut comes out non-monotone and the floor was left where it was.

So the honest summary is narrower than "depth does not matter". It is: the spread separates on every market we measured, and depth separates on one of them. A rule that only had the second one was going to keep buying quotes like the Mineiro book.

What this buys, and what it does not

With the spread gate on, the ask is fair to slightly cheap at every minute from 70 to 89 — between +2.1pp and +4.7pp. Without it, the same market looks expensive, by around 5 to 6 points late on.

That apparent expense was never a property of the market. It was an artefact of including books that were not books.

⚠️ And it is worth being precise about what that means, because it is tempting to read it as an edge. Every one of those positives still has a confidence interval crossing zero. Deep, tight books also belong to the larger competitions, and that confound is not controlled for here. The gate removes a measured bleed. It does not find an edge. Those are different claims and only the first one is supported.