Polymarket trades data: the tape tells you what filled, not what was offered

Polymarket's Data API serves executed trades with full history, free and keyless, alongside Gamma metadata and CLOB prices. For what happened after the fact, trades are the honest record — they tell you what filled, not what was available to fill against.

What fills do not record is the liquidity on the other side of the tape: every other price level and size that stood when the fill happened. That part of the market is forward-only — the ladder only exists if it was captured while it stood — which is why the official book endpoint has no history and why a stored archive records it instead.

On the published archive, 5-minute crypto books are one-sided 16.9% of the time and 3.24% of snapshots are crossed (best bid above best ask), with crossed episodes lasting a median 2 seconds and 98.2% clearing within 60 seconds. A tape-only backtest that fills at the last trade or the mid would score those fills as if liquidity existed on both sides.

A fill model that walks the book behaves differently and measurably: 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. Backtest AI and the backtest API replay on the stored ladder rather than the tape.

• Polymarket Data API — trades — executed trades, full history, free and keyless. Starthere when a tape is all you need (what the API serves).
• CLOB prices-history — per-token price series at a 1-minute floor, free (price history).
• PolyOrderbooks books — what stood at each level: full L2 ladders every 250ms, queryable 1 day down to 250ms, free Starter with 3 days of history.
• Open sample — 897,192 snapshots across 805 resolved markets on Zenodo, CC BY 4.0, to see the depth side of the tape before building.

Executed trades are the public layer of the venue: what filled, at what price, size, and time, served free through the official Data API, with full history. For anything that fills, the tape is authoritative and free; the gap is depth — the resting ladder behind those fills, which has no official historical endpoint.

The research consequence is subtle but sharp: fills recorded at the last trade price or the mid flatter a backtest, because on this venue the executable price is the bound-side floor and the spread hugs the tick grid. A strategy measured against last-trade prints assumes liquidity the book frequently did not have.

The reconcile is the habit that pays: take one minute of tape, match each trade to the stored book for its second, and confirm the printed size existed on the resting side. Where the tape and the ladder agree, the record is whole; where they disagree, the difference is slippage or reconstruction.

The two sources are complementary rather than competing: the official Data API is the assembly line of what happened, and the stored archive is the warehouse of what was available before it happened. Both belong in the same study, joined on the UTC second.

Read volume against depth, not alone: a spike printed against a thin ladder is a different event from the same spike against a deep one, and the stored book is the only way to tell them apart.

Never price a backtest off the mid or the last trade alone — with 16.9% one-sided 5-minute books and 3.24% crossed snapshots, an executable price is a floor-side fact, not an average.

Watch resolved-market trades during the final minute, where the tape runs out: 76.2% of final-minute snapshots have an empty side, so late prints are settling, not transacting.

Keep the tape and the archive on one clock; a trade whose second cannot be joined to a book belongs to neither source.

• Order book archive
• Order book data
• Backtest API
• What the API serves
• Price history