Polymarket Order Book Replay Data

The published replay-accuracy study compares event-archive reconstruction against direct capture for an hour sampled 2026-05-02 15:00 containing 1,498 captured rows. On-chain reconstruction reproduced only 36.6% of the recorded quote stream in that hour — the rest of the book never touches the chain.

Root cause: the full-depth Polymarket book is off-chain. Executions and settlement reach Polygon; cancellations, re-pricings, and resting levels do not. Any reconstruction from events is therefore a subset of what happened.

• Continuous capture at a dense grid — the 250ms interval this archive keeps; an event-driven parquet log is the alternative shape.
• Full ladders, not top-of-book: sports ladders measured at 35–36 levels per side must replay level by level.
• Explicit frames and defect flags so crossed or one-sided states are preserved, not smoothed.
• Boundary alignment with reference prices (Binance/Chainlink) for settlement-adjacent replay.

With recorded rows, replay becomes mechanical: given a timestamp, restore bids and asks at that frame, then walk the book for fills at your limit price, queue-safe, with spread cost.

The alternative — mid-based replay over reconstructed series — silently grants fills the book never offered. That is the difference between a starting point and a result.

• Select a one-hour window (the recorded 2026-05-02 15:00 hour is the canonical one) and pull the book frames at 250ms for the period; expect the 1,498-capture scale of material.
• Reconstruct the on-chain order book state from the frames for a keyword window and compare against the chain-derived reconstruction; the documented reproduction rate of 36.6% sets the honest expectation for how much microstructure a reconstruction can capture.
• List the frames that could not be reconstructed and describe why (missing side, crossed quotes, boundary frames); the explainable gap is the difference between a replay claim and a replay guess.
• Save the replay as a small report: window, captures, reconstruction rate, and the top five unexplained frames — the audit trail every replay should carry.

Replay is the reproduction primitive of this entire catalog: an exact as-of reconstruction of the book at any past second, built from monotonic sequences and full ladder frames rather than from sampled mid totals.

The 36.6% reproduction rate is a feature if it is stated correctly: it says the reconstructable core of a one-hour window is a genuine subset, and honest research begins by naming the subset it is studying.

Replaying the archive is how every claim in this catalog gets tested — a mid report, a fill benchmark, a settlement flag all resolve to the same as-of question.

The number to quote honestly is the reconstruction rate of your own window, not the archive's; measure yours, then extrapolate.

Start with the one-hour canonical window; it is small enough to replay end-to-end and rich enough to teach the limits of reconstruction before you scale.

Replay data is the verification layer of the catalog: every mid, spread, and slippage claim resolves to an as-of reconstruction, which is what the honest-workflow pages assume when they tell you to replays.

Each measured case in this page is stored at 250ms with the same field names as every other dataset, so the switch from one page's material to another is a query-parameter change, not a migration.

That shared format is the reason a family can be studied across settlement, repricing, replay, final-seconds, and cross-market frames with one pull and one schema.

A replay that fails to reconstruct a frame is data about the reconstruction, not about the market; record the failure and the reason, because unexplained frames are exactly where the honest audit stops.

• Research blog
• Order book data
• Backtesting data
• Crypto data