astrnvk/polymarket-orderbook-data-preview

Polymarket Microstructure — Free Preview

A free sample of the Polymarket CLOB dataset, so you can inspect the exact schema, file layout and manifest format before buying. Same Parquet + sha256-manifest packaging as the full dataset.

Two parts, and they show different things:

• daily/ — four full days, 26–29 June 2026, all four underlyings (BTC, ETH, SOL, XRP). Three tables, top-of-book only, deliberately our roughest coverage.
• sample/2026-07-21/ — 30 minutes of BTC, 13:00–13:30 UTC, all six tables including full L2 depth. This is what the paid data actually looks like today.

The full history is sold separately at astrnvk/polymarket-orderbook-data.

→ Browse, price and buy at the storefront: polymarket-microstructure — $5/day for 1–6 days, $4/day for 7–29, $3/day beyond a month, all four assets included. Incomplete days cost proportionally less, and the storefront shows the full arithmetic before you pay.

The 30-minute depth sample (sample/2026-07-21/)

June cannot show you order-book depth, because there was none — book_depth starts 20 July 2026. So this slice exists to show the rest of the schema.

BTC, 21 July 2026, 13:00:00–13:30:00 UTC. Six tables, 4.4 MB, ten markets (seven 5-minute, three 15-minute):

File	Rows	What it shows
book_depth	2,771,148	Full L2, both sides, up to 118 levels per side
book_meta	28,141	tick_size, min_order_size, neg_risk, book_hash
clob_trades_ws	13,493	WebSocket trades with tx_hash, fee_rate_bps
book_snaps	9,582	Best bid/ask with a real spread
clob_trades	7,201	REST trades with taker side
spot_klines	1,800	1-second Binance candles

The window is deliberate. On 21 July at 12:51 UTC the collector switched to change-only writing with a heartbeat, and to millisecond exchange timestamps. This slice sits entirely after that switch, so it shows the current format rather than the legacy one. Take a morning slice of the same day and you would be looking at the old era and drawing the wrong conclusions.

n_obs is NULL here, and that is correct. The column counts how many raw rows were collapsed into one during deduplication. Days before 21 July are exported from the deduplicated table, where each row summarises roughly 80–150 identical raw snapshots. From 21 July the collector writes change-only itself, so nothing is collapsed and every row already is one observation. NULL means "no collapsing happened" — not "value missing".

Why the June days

These are the earliest days of collection, and deliberately the roughest — they show you the format and our worst coverage, so there are no surprises after purchase:

Date	Coverage	Note
2026-06-26	~61%	Collector went up mid-day
2026-06-27	90–95%	Collection still being tuned
2026-06-28	90–95%	Collection still being tuned
2026-06-29	~34%	Down 00:42–16:00 UTC for a planned migration

The paid dataset (from 30 June onward) does not carry these particular gaps: measured uptime there runs from 95.8% to 100%, and every day is priced from its own figure.

What's in each file

Table	Time column	Contents
book_snaps	ts	Best bid / ask stream — not a full ladder (level/size are placeholders)
clob_trades	ts	Trades with taker side
spot_klines	ts	Spot candles for the underlying

Three tables in June, not six — and that is the point of showing these days. book_depth starts 20 July 2026; book_meta and clob_trades_ws start 21 July. June shows you the schema and the packaging, but it cannot show you order-book depth, because there is none to show. That is what sample/2026-07-21/ above is for.

The June days also predate the 21 July migration, so timestamps are receipt time quantized to whole seconds, ts_ms and spread are NULL, and level/size are placeholders. Paid data from 21 July onward uses millisecond exchange time, carries full depth, and populates spread.

One honest limitation on timestamps in both eras: there is no independent ingest clock. In the June era there is only the receipt second; from 21 July ts is derived from ts_ms (ts == floor(ts_ms/1000) exactly), so the two columns are one timestamp at two resolutions, not source-versus-arrival. If you need to measure your own feed latency against ours, this dataset cannot do it.

Every Parquet file ships with a <file>.manifest.json (sha256, row count, interval bounds).

Terms

Collected from the public Polymarket API. Not affiliated with Polymarket. This preview is provided for evaluation under CC BY-NC 4.0. Nothing here is trading advice.