Polymarket Stylized Facts Dataset

This dataset provides nine stylized-fact measurements (SF1–SF9) computed on a stratified sample of 13,314 Polymarket binary-event markets that resolved during 2026-04-21 to 2026-04-27 UTC. It is released as the empirical foundation for Paper 1 of the four-paper Event-Linked Perpetuals research programme and as a shared baseline for prediction-market microstructure research.

Corpus

Attribute	Value
Total markets	13,314
Source archive	PMXT v2, 168 files
Date range (resolved_at)	2026-04-21 to 2026-04-27 UTC
Subsample rule	Stratified-by-day, seed 20260505
Snapshot cutoff	2026-04-27T23:59:59Z
Format	Parquet (primary), JSON (aggregates)

Note: The companion paper (Paper 1) reports 13,298 markets; the released dataset contains 13,314 due to a more complete UMA oracle cache available at build time (2026-05-07). Stylized-fact headline values are unaffected. See CHANGELOG.md.

Counts by event class

Event class	Count	Share of total	Share of three classes
sports	6,800	51.1%	77.9%
other	4,584	34.4%	(excluded)
crypto	1,518	11.4%	17.4%
politics	412	3.1%	4.7%
Total	13,314	100%	—

Sports account for 77.9% of the three-class total (excluding other), triggering the pre-registered ≤70% threshold consequence rule on the analysis sample.

Stylized facts headline values

Stylized fact	Headline value	Pre-registered floor	Passed
SF1 boundary depth asymmetry ρ (base)	1.72	≥ 1.5	✓
SF1 boundary depth asymmetry ρ (resume)	1.65	≥ 1.5	✓
SF2 terminal jump magnitude |Δ|	0.50	≥ 0.10	✓
SF3 news vs control basis	0.0132 / 0.0367	descriptive	—
SF4 mid half-spread	0.27	descriptive	—
SF8 crypto surge factor	24.62×	descriptive	—
SF8 politics surge factor	0.68×	descriptive	—
SF9 12h-3h → 3h-1h depth ratio	4.91	descriptive	—

Files

File	Format	Description
data/markets-stylized-facts-v1.parquet	Parquet	One row per market; per-market SF1, SF2, SF4 columns
data/aggregates.json	JSON	Pooled and per-class aggregate values for all nine SFs
data/sf7-class-hour-v1.parquet	Parquet	SF7 hourly activity — 96 rows (4 classes × 24 hours)
data/sf9-bucket-aggregate-v1.parquet	Parquet	SF9 depth by time-to-resolution bucket — 5 rows

Schema: markets-stylized-facts-v1.parquet

One row per market. SF1, SF2, and SF4 carry per-market columns; SF3, SF5–SF9 are aggregate-only (in aggregates.json).

Field	Type	Description
market_id	string	Polymarket condition ID, lowercase 0x-prefixed hex
question	string	Market question text
event_class	string	One of sports, politics, crypto, other
tags	list[string]	Polymarket tag names
created_at	string	ISO 8601 UTC
closed_at	string | null	ISO 8601 UTC; null if not reported
resolved_at	string	ISO 8601 UTC (UMA OO settlement)
resolution_outcome	int8	0 (NO) or 1 (YES)
volume_total_usdc	float64 | null	Cumulative trading volume
is_negrisk_member	bool	Whether part of a Polymarket negRisk group
negrisk_group_id	string | null	Group identifier if applicable
sf_pass	string	CC-004 pass: resume or none
sf1_rho	float64 | null	Boundary depth asymmetry ratio; null if no boundary observations (88% of markets)
sf2_terminal_jump_magnitude	float64 | null	|Δ index| over [restime − 1h, restime]; null for 23% illiquidity cohort
sf4_half_spread_boundary_low	float64 | null	Median half-spread when index < 0.10; null in v1 (not per-market in CC-004)
sf4_half_spread_low	float64 | null	Median half-spread, index in [0.10, 0.30)
sf4_half_spread_mid	float64 | null	Median half-spread, index in [0.30, 0.70]
sf4_half_spread_high	float64 | null	Median half-spread, index in (0.70, 0.90]
sf4_half_spread_boundary_high	float64 | null	Median half-spread when index > 0.90; null in v1 (not per-market in CC-004)

Quick start

import pandas as pd

df = pd.read_parquet("data/markets-stylized-facts-v1.parquet")

# Median terminal jump magnitude by event class (SF2)
jump_by_class = (
    df[df["sf2_terminal_jump_magnitude"].notna()]
    .groupby("event_class")["sf2_terminal_jump_magnitude"]
    .median()
    .sort_values(ascending=False)
)
print(jump_by_class)
# crypto    0.9995
# sports    0.5000
# politics  0.5000
# other     0.5000

# Median boundary depth asymmetry (SF1) by class
rho_by_class = (
    df[df["sf1_rho"].notna()]
    .groupby("event_class")["sf1_rho"]
    .median()
)
print(rho_by_class)

import duckdb

con = duckdb.connect()

# SF9: depth growth approaching resolution
con.execute("""
    SELECT bucket, bucket_lower_h, bucket_upper_h,
           pooled_median_depth_within_200bps_usdc,
           pooled_n_market_observations
    FROM read_parquet('data/sf9-bucket-aggregate-v1.parquet')
    ORDER BY bucket_lower_h DESC
""").df()

@dataset{pmxt2026,
  author    = {Nechepurenko, Maksym},
  title     = {{Polymarket Stylized Facts Dataset}, v1},
  year      = {2026},
  publisher = {GitHub},
  url       = {https://github.com/ForesightFlow/datasets/tree/main/pmxt-stylized-facts-v1},
  note      = {Tag: pmxt-stylized-facts-v1. Snapshot cutoff: 2026-04-27.}
}