A BTC five-minute contract resolves at a fixed second. Sixty seconds before that second, two in five order books already have an empty side. Ten seconds out it is nearly nine in ten, and on the side that is about to win the book is not thin at all — the top level holds a median of 42,254 shares at 0.99. There is simply nothing to buy at it.

We published the mechanism on an 80-contract sample in August. This post re-measures it at scale: 298 resolved BTC five-minute contracts drawn from the 8,349 that settled between 2026-08-26 and 2026-09-23, and 17,880 one-second order book snapshots across their final forty hours' worth of final minutes.

The gradient is real

For each sampled contract we read the one-second book on the Yes outcome over the sixty seconds before settlement and counted snapshots where only one side of the book was resting (any bid on the Yes token with no ask, or an ask with no bid):

seconds to close   snapshots   one-sided
1-10s              2,980       86.9%
11-20s             2,980       80.1%
21-30s             2,980       72.5%
31-40s             2,980       62.4%
41-50s             2,980       50.9%
51-60s             2,980       41.3%
final 60s         17,880       65.7%
final second         298       90.3%

Each bucket holds 2,980 snapshots — 298 markets, ten seconds each. The shape matches the earlier 80-contract study (94.0 / 79.1 / 73.8 / 64.6 / 53.4 / 39.7 percent across the same buckets); ours runs a few points lower in the last ten seconds and stays one-sided 86.9% of the time there, 90.3% in the very last second. No market went fully silent — both sides were never simultaneously empty in 17,880 snapshots. The book does not vanish; it becomes one-directional.

One contract, the last forty-eight seconds

This is the Yes book of the contract resolving at 23:30:00 UTC on 23 September (question: "Bitcoin Up or Down, 7:25PM-7:30PM ET"), a market where the Up side won. Sparse rows until the bell:

time      best bid           best ask
23:29:10  0.950    800.0      0.960     39.6
23:29:11  0.970    230.0      0.980    580.3
23:29:12  0.990  4,615.6       --
23:29:25  0.990  25,057.4      --
23:29:39  0.990  44,194.8      --
23:29:53  0.990  55,067.3      --
23:29:59  0.990  67,571.5      --

Through 23:29:11 it is an ordinary close: 0.97 / 0.98, moving, with a few hundred shares per side. At 23:29:12 the ask disappears, and from that moment nobody is willing to sell the winner. The bids do not dry up — they grow, at a fixed 0.990, from 4,616 to 67,571 shares. For the final forty-eight seconds you could sell into that wall, but buying was impossible at any price. The direction is the whole story.

How deep the winner quotes

Because the two outcomes mirror each other, we can measure the winning side on both resolutions from the Yes book alone (a Yes best ask of 0.01 is a No best bid of 0.99 with the same size). Across all 298 contracts, at the last second before settlement:

final second, 298 markets
winner side >= 0.95                277 / 298   93.0%
  of which >= 0.99                              98.6%
  median top-of-book size                       42,254
  90th percentile                               65,081

last 15 seconds, 4,030 near-certain snapshots
  median top-of-book size                       28,689
  90th percentile                               52,387
  other side empty                               83.0%

The certainty is almost complete by the bell — 93% of markets had the winning side at 0.95 or better in the final second, and 98.6% of those at 0.99 or 0.995. The size is the new number: a median of 42,254 shares, p90 of 65,081, resting at a near-certain price. And 83% of the near-certain snapshots in the last fifteen seconds had nothing on the other side. The market is telling you the outcome, loudly, and it is also telling you that this is information you cannot trade on — there is no counterparty.

The moments that still look two-sided

34.3% of the final-minute snapshots do show both sides. They are not a clean market either. Of those 6,137 two-sided snapshots, 14.1% are strictly crossed — the best bid above the best ask — plus 2.1% exactly zero spread; 4.8% of all final-minute snapshots, and 14.1% of the final ten seconds' worth, are crossed. The moment a resting ask exists near settlement it is often a quote that has not been reconciled against a higher bid. A backtest that fills at the midpoint of a two-sided snapshot will report executions through crossed or zero-spread books at exactly the time of day when it is most eager to trade.

Is it stale?

The archive materialises 1-second buckets with last-value carry, so a repeated book can mean "unchanged" or "unobserved." To rule out that the one-sidedness is an artifact, we counted how many final-ten-second snapshots actually changed the top of book versus the prior second: 62.5% did. The near-closed book is still updating; the empty side is a living state of the market, not a gap in the capture.

What this changes about a backtest

The practical guard from the earlier study still holds: before treating a fill as possible, check that the relevant side of the book is populated and holds the size you need. Two additions from this measurement:

  • The error is one-directional. Selling the eventual winner deep into settlement is easy — the wall is there. Buying it is impossible. Any strategy that enters near the winning side will fail to fill in reality while filling instantly in a price-based simulation; exits into the winner, conversely, are the one trade that genuinely could have happened.
  • The size at 0.99 is not liquidity in your direction. A 42,000-share top level looks liquid in an aggregate depth metric. It was, for sellers. A depth figure that does not record which side holds the size will mislead every order sized against the wrong direction of the close.

Polymarket's own free price API is the right tool for prices, and we say so in our API pricing guidance. This is a book statement: a 1-second price series says the market closed at the winner's price; the book says the other side had left the building.

What we are not claiming

We did not attach a cause to the ask's disappearance, and we did not measure order flow or queue position — the archive records state, not events. The sample is 298 of the 8,349 resolved BTC five-minute contracts in the window (every 28th, sorted by close time); we have not measured hour-long or strike contracts here, though we would expect the shape to differ. The winning side is inferred from the complementary outcome's mirror at 0.99 or 0.01, a property that holds structurally.

Checking it

Every figure above comes from the order book history endpoint. The recipe: list resolved events on series btc-up-or-down-5m (id 10684), take every 28th market, and request 1-second books over the sixty seconds before its settlement timestamp (the window close, which is five minutes after the slug time; end_ts is exclusive). The 80-contract dataset behind the original study remains publicly downloadable, so the two measurements can be compared rather than taken on trust.

The last minute of a five-minute market is not a low-liquidity minute. It is a one-liquidity minute: a deep, certain bid on the eventual winner and no way to get to the other side. Price data sees a settled market; the book sees a door that only opens one way.