Polymarket dataset for academic research: citable L2 order books

The record is on Zenodo, registered with DataCite, and hosted independently of this site. Download needs no account.

Row counts are the wrong comparison. Depth recorded once a minute cannot answer what an order would have cost, how long size rested, or how liquidity moved through the final seconds before settlement — a minute is longer than the phenomenon. Volume follows from resolution and coverage; it is not independent evidence of usefulness.

We cover crypto only, and say so. A narrower category captured at 1-second resolution with reconciliation running is more use for microstructure work than every category at minute granularity.

Several providers hold historical Polymarket order book data. Very little of it can go in a paper, because appearing in a data availability statement requires more than access:

The last row matters more than it looks. Publishing a defect flag rather than filtering the affected rows means the internal consistency rate is measurable by a reader instead of asserted by us — which is the difference between a dataset a reviewer accepts and one they question.

Each row is one outcome token at one second: the full bid and ask ladders as typed list columns, the derived top of book, seconds to settlement, and the winning outcome. Parquet with zstd, 43 MB in total, and it loads in pandas, Polars or DuckDB without a parsing step.

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

(df.best_bid >= df.best_ask).mean()          # internal consistency rate
df[df.seconds_to_close <= 60].crossed.mean() # and near settlement

• Execution cost and slippage

  Full ladders mean an order can be filled level by level against real resting size. Midpoint series cannot answer what a trade would have cost.

• Liquidity around resolution

  Three contract lengths over the same window let you separate the effect of time-to-settlement from the effect of contract design.

• Price discovery and calibration

  Every market is resolved with the winning outcome attached, so forecast accuracy can be scored directly without joining an external source.

• Data quality methodology

  The crossed flag supports work on order book data integrity itself — a rate that can be computed on any archive and used as a diagnostic.

Depth has to be captured while a market trades; it cannot be reconstructed afterwards from an archived event stream. The reason is structural rather than a question of implementation care: Polymarket's feed is unsequenced, so a dropped message leaves no gap to detect, and a replay of a closed market has no full book to re-read against.

We tested this by replaying a public archive against its own periodic snapshots — a design that needs no external reference, since the source grades its own homework. The measurement is written up here, and a formal note with the analysis code is in preparation.