DineshKumar8399/polymarket-orderbook-dataset

Polymarket Order Book Dataset

Order-book snapshots from a prediction market, collected continuously between 2026-07-10 and 2026-08-19: 281,268,633 quote observations across 172,036 markets, plus settlement outcomes and a separate high-frequency feed that records actual traded prices.

It is published so other people can build and train on it without first spending a month running collectors. Everything here is an independent observational recording of publicly displayed market data.

Read Known issues before you train anything on this. The collection has a four-day outage, one column that dies partway through, and a label set with time censoring. All three are documented, none are hidden, and each one will quietly wreck a model if you miss it.

---

Get the data

The parquet files are not in this git repo — they are rebuilt daily, and committing them would grow the history by roughly 12 MB a day forever. They live in two places instead, both refreshed every night:

Hugging Face (browsable, has a dataset viewer, resumable):

pip install huggingface_hub
hf download DineshKumar8399/polymarket-orderbook-dataset --repo-type dataset --local-dir polymarket-data

import duckdb
duckdb.sql("SELECT * FROM 'polymarket-data/quotes/**/*.parquet' LIMIT 5").show()

GitHub Releases (a single dated tarball, ~149 MB):

gh release download data-2026-08-20 --repo DineshKumar8399/polymarket-orderbook-dataset
tar --zstd -xf polymarket-orderbook-*.tar.zst

Each release is a frozen snapshot, so data-2026-08-20 is reproducible: cite the tag and anyone can reconstruct the exact data you trained on. Latest build: 2026-08-20.

---

Contents

File	Rows	What it is
quotes/dt=YYYY-MM-DD/*.parquet	281,268,633	Book quotes for every tracked market, partitioned by date
markets.parquet	172,036	One row per market: question text, category, coverage
labels.parquet	146,022	Binary settlement outcomes, with a source column
watch_quotes.parquet	1,691,703	High-frequency feed — the only table with traded prices
data_quality.parquet	7	The known issues below, as queryable rows

Total: about 149 MB of ZSTD-compressed Parquet.

---

Quick start

import duckdb

con = duckdb.connect()

# the whole quote history — the partition layout means you can slice by date
# without reading the rest
con.sql("""
    SELECT * FROM read_parquet('quotes/**/*.parquet', hive_partitioning=1)
    WHERE dt = '2026-08-01' AND slug = 'some-market-slug'
""").show()

# join quotes to outcomes, using ONLY the authoritative labels
con.sql("""
    SELECT q.slug, q.ts, q.bid, q.ask, q.mid, l.y
    FROM read_parquet('quotes/**/*.parquet', hive_partitioning=1) q
    JOIN read_parquet('labels.parquet') l USING (slug)
    WHERE l.source = 'api'
""").show()

With pandas or polars:

import pandas as pd, polars as pl

markets = pd.read_parquet("markets.parquet")
one_day = pl.read_parquet("quotes/dt=2026-08-01/*.parquet")

---

Schemas

quotes/

One row per market per poll of the order book.

Column	Type	Notes
ts	timestamp	When the snapshot was taken
slug	string	Market identifier, joins to markets and labels
category	string	sports, politics, climate, culture, …
bid	double	Best bid. NULL when no bid was resting
ask	double	Best ask. NULL when no ask was resting
mid	double	Midpoint; NULL unless the book was two-sided
spread	double	ask - bid as reported upstream
volume24hr	double	Mostly unusable — see issue 2
segment	int	1 = before the outage, 2 = after. See issue 1
dt	date	Partition key

dt is stored in the directory name, not inside the parquet files, so it only materialises as a column when the reader is told to parse the partitions — hive_partitioning=1 in DuckDB, automatic in pandas.read_parquet / pyarrow.dataset when you point them at the quotes/ directory rather than at individual files. The Hugging Face viewer does not parse it, so dt is absent there; slice on ts instead when browsing.

Prices are probabilities in [0, 1]: a market at 0.35 implies a 35% chance. A YES share pays 1.00 if the event happens and 0.00 otherwise, so the price is also the cost per unit of payoff.

markets.parquet

Column	Type	Notes
slug	string	Primary key
question	string	Human-readable question
category	string
n_snaps	bigint	Quote rows present for this market
first_ts / last_ts	timestamp	Coverage window

question is stored here rather than on every quote row — repeating it 281,268,633 times is most of why the raw CSV was 32 GB.

labels.parquet

Column	Type	Notes
slug	string	Joins to quotes / markets
y	int	1 = resolved YES, 0 = resolved NO
source	string	api = authoritative. convergence = inferred, biased

Filter on source. See issues 5 and 6 — this is the single easiest way to get a wrong answer out of this dataset.

watch_quotes.parquet

Column	Type	Notes
ts	timestamp
slug	string
bid / ask / mid	double
last_traded	double	Last traded price — the only print data here

---

How it was collected

Two independent collectors ran continuously on a dedicated machine:

Broad sweep — polled the full market list roughly every two minutes and recorded the top of book for every market it could see. This produced quotes/. It is wide (every market) but shallow: it records what was quoted, never what traded.

Watchlist — polled about twelve actively-traded markets every twenty seconds, rotating the selection every thirty minutes, and recorded the last traded price alongside the book. This produced watch_quotes.parquet. It is narrow but deep, and it is the only place in this dataset where you can ask whether a trade actually happened.

That split matters more than it sounds. Displayed quotes are not the same thing as executable prices, and nothing in quotes/ can tell you whether a given quote could have been filled.

---

Known issues

Also shipped as data_quality.parquet so you can assert on them in a pipeline.

1. A four-day collector outage

No rows exist between 2026-07-17 17:26:09 and 2026-07-22 10:32:48. The machine lost its storage enclosure and stopped writing.

The dataset is therefore two disjoint series, not one 35-day window. Any per-market price path that spans the gap is broken, and any price change computed across it is meaningless — you would be measuring a 4-day-17-hour jump as if it were a normal interval.

The segment column marks which side each row falls on. Restrict to a single segment, or handle the discontinuity explicitly.

2. volume24hr dies partway through

94.4% NULL overall, and 0.0% populated after 2026-07-22 — the upstream API stopped returning the field. It is 26–40% populated before the outage.

Any feature built on volume or liquidity silently becomes all-NULL for the larger part of the dataset. Use spread, or per-market snapshot frequency (markets.n_snaps) as a rough activity proxy, and label them as proxies.

3. No traded prices in quotes/

The broad sweep records book quotes only. Its last column was 100% NULL, so it has been dropped rather than shipped as an empty column named last.

If your question is "did this actually transact" — fill realism, execution modelling, print-versus-quote — it is only answerable on watch_quotes, which covers 5,022 markets rather than 172,036. Note last_traded is itself 85.6% populated, not 100%.

4. Crossed books

A small number of rows have ask < bid, which is not physically meaningful and reflects the two sides being read a moment apart. Filter with ask >= bid if your method is sensitive to it.

5. Labels are time-censored

Most source = 'api' labels come from a one-off backfill run on 2026-07-23/24. So "has a label" correlates strongly with "settled before Jul 24" — 110,173 markets (64%) carry an authoritative label, and they are not a random 64%.

This bites hardest on walk-forward validation: naively splitting train/test on a late date can leave you with an empty test set and a script that reports success anyway. Check your split sizes.

6. Convergence labels are biased — prefer source = 'api'

Rows with source = 'convergence' were inferred by watching the price settle toward 0 or 1. That method systematically mislabels markets whose books died before converging, and it selects for markets that converged at all.

On this data that bias was large enough to manufacture an edge that did not exist — a backtest showed a substantial per-share profit that vanished entirely when the same cell was recomputed on authoritative labels. They are included because throwing away data is worse than labelling it, but treat them as a weak-supervision signal, never as ground truth, and never blend the two sources without checking how much the choice moves your result.

7. watch_quotes is not a random sample

The watchlist deliberately tracked the most active markets, rotating every thirty minutes. Anything you measure there describes liquid, high-attention markets — typically live in-play sports — and will not generalise to the long tail in quotes/.

---

Coverage

Markets by category:

   category  slugs
     sports 168274
   politics   1443
    climate    977
    culture    936
      macro    154
 technology     93
    finance     82
     crypto     51
    science     14
geopolitics     12

Sports dominates by design: it is the bulk of what the venue lists and the bulk of what trades.

---

License

CC BY 4.0 — use it, remix it, build commercial things on it; just give credit. Full legal code in LICENSE; attribution and disclaimer in NOTICE.

Polymarket Order Book Dataset (2026), Dinesh Gopalakrishnan.
Licensed under CC BY 4.0.
https://github.com/DineshKumar8399/polymarket-orderbook-dataset

Disclaimer

Research and educational use. This is an independent observational recording of publicly displayed data and is not affiliated with, endorsed by, or supplied under agreement with any exchange or venue. Nothing here is financial advice. Past market behaviour does not predict future market behaviour, and a backtest on this data is not a trading strategy.