Free Polymarket order book datasets

• Crypto up/down order books, three contract lengths→897,192 snapshots805 marketsParquet, 43 MB8 coinsThe research dataset. Full L2 ladders at 1-second resolution with the winning outcome on every row, across 5-minute, 15-minute and 4-hour contracts. Carries a DOI, so it can be cited.Download directlyCC BY 4.0
• BTC 5-minute single-market sample→One full lifecycle1 marketCSV + JSONBTC 5-minuteOne market from open to settlement, every second. The quickest way to see the shape of the data before committing to anything larger.Download directlyCC BY 4.0
• Crypto 5-minute L2, Hugging Face mirror→158,745 snapshots352 marketsParquet5-minute contractsFive-minute contracts only. Loads directly with datasets.load_dataset, which is convenient inside an existing ML pipeline.Download directlyCC BY 4.0
• Crypto 5-minute order books, Kaggle mirror→158,745 snapshots352 marketsParquet per coin5-minute contractsThe same five-minute capture, split one file per coin. Useful if you already work in Kaggle notebooks. The account requirement is Kaggle's, not ours.Needs a free Kaggle accountCC BY 4.0
• Provider comparison→CSVProviders, not marketsCoverage, resolution and pricing across the providers in this space, as a flat file rather than a marketing table.Download directlyCC BY 4.0

Every file is CC BY 4.0: commercial use, modification and redistribution are all fine, and attribution is the only requirement. The research dataset carries a DOI, so it is valid in a data availability statement — what that means for a paper.

• Books go one-sided near settlement

  In the final minute of a 5-minute contract, 76% of snapshots have an empty bid or ask side, because nobody offers the losing outcome. Real behaviour, but it breaks code that assumes two sides.

• Some snapshots are crossed

  Best bid at or above best ask, which is not tradeable. Around 3% on 5-minute markets. Flagged rather than removed, so the rate stays measurable rather than hidden.

• Contract length changes the data

  4-hour books are one-sided 0.5% of the time against 17% for 5-minute ones. Conclusions from one length do not transfer to another.

These are stated here rather than left for you to discover. The crossed flag in particular is published rather than filtered out, so the rate is something you can measure instead of something we assert.

import pandas as pd
df = pd.read_parquet("updown_5m.parquet")

(df.best_bid >= df.best_ask).mean()            # crossed rate
(df.bid_prices.str.len() == 0).mean()          # empty bid side
df[df.seconds_to_close <= 60].crossed.mean()   # and near settlement

These are windows of a larger archive — over 781 million order book snapshots across 815,000 markets. If you need a different period, different markets, or a longer history than the published files, the API queries the archive directly at the same 1-second resolution.

• The free tier queries at 1-second resolution, the same as paid plans, over a shorter history window.
• Every way to get Polymarket data, including from Polymarket itself — their API is free and better than ours for prices.
• What an L2 snapshot actually contains, with a worked slippage example.