Polymarket microstructure data: depth, crosses, and the last seconds

Market microstructure — who rested depth where, how the book behaves near resolution — needs stored books, not live deltas. Replaying a public delta archive against its own snapshots diverged at 67.8% of 381,093 checkpoints, with 6.4% crossed and extra levels outnumbering missing ones 4.2 to 1. A stored snapshot archive does not inherit that.

PolyOrderbooks publishes the raw capture — 897,192 snapshots across 805 resolved markets, CC BY 4.0 on Zenodo — so the measured rates below are reproducible by anyone.

Across 298 resolved BTC 5-minute contracts (17,880 one-second snapshots), 90% of final-second order books had a single side, and 87% were already one-sided in the last ten seconds; the winner's side rested a median 42,254 shares at 0.99 (settlement depth post). A separate 80-contract study found 94% of final-ten-second snapshots had an empty bid or ask side (final seconds post).

In the final minute of a 5-minute market, 76.2% of snapshots have an empty bid or ask side — nobody offers the losing outcome and nobody bids the winning one. Other measured repricing behaviors: a strike-ladder market repriced roughly one second before its strike with a book running ~0.99 wide (post).

• Guard the head of the ladder — index bid_prices[0] without guarding and one-sided rows fail; treat empty-side rows as data (order book data).
• Resample spread to 60s — 250ms captures include transient quote noise; 60-second series represent executable levels (spread analysis).
• Model slippage from depth — walk the ladder at fill time and compute VWAP (slippage estimation).
• Study repricing — measured posts on final seconds, the strike ladder, and same-second Binance repricing (blog).

Microstructure is the layer of measurement that stored books make possible: who rested depth where, how the ladder bent near resolution, and what the executable price did around reference prints. The working material is the 250ms capture — full L2 on both sides — and the measured rates come straight from those frames.

The headline numbers are reproducible from the public set: 5-minute books are one-sided 16.9% of the time and 3.24% of snapshots are crossed; across 298 resolved BTC contracts, 90% of final-second books had a single side and the winner's side rested a median 42,254 shares at 0.99. A separate 80-contract study found 94% of final-ten-second snapshots had an empty side.

That same three-part structure generalizes to other event shapes: a strike-ladder market repriced roughly one second before its strike with a book running about 0.99 wide, which is the same anatomy as a settlement — the executable price migrating to the bound ahead of the event.

The work loop is: measure a rate on the stored books, name its window, publish the raw frames as the citation, and let the reader rerun. The 897,192-snapshot Zenodo release exists so that loop stays closed end to end.

The one-sided share is the venue's fingerprint: 16.9% of 5-minute snapshots carry a single side, so any model that assumes two-sided execution is wrong 16.9% of the time by construction.

Settlement microstructure is the strongest signal family: final-minute books emptying (76.2%), final-second single-sidedness (90%), and median 42,254-share rests at the bound all describe the same punctuation mark.

When studying repricing, keep the reference and the book on the same clock; the strike-ladder one-second finding is only visible because both were recorded on the same timestamp grid.

Publish methodology with every number: a rate without its window and dataset definition is a claim, not a measurement, and the open archive exists to turn it into one.

• Order book data
• Order book archive
• Liquidity data
• Resolved markets
• Final seconds study