Datasets:

vgregoire
/

polymarket-users

language:
  - en
license: cc-by-4.0
pretty_name: Polymarket Users
tags:
  - prediction-markets
  - polymarket
  - finance
  - cryptocurrency
  - user-behavior
configs:
  - config_name: markets
    data_files:
      - split: train
        path: markets.parquet
  - config_name: events
    data_files:
      - split: train
        path: events.parquet
  - config_name: predictions
    data_files:
      - split: train
        path: predictions.parquet
  - config_name: user_features
    data_files:
      - split: train
        path: user_features.parquet
  - config_name: user_pnl_summary
    data_files:
      - split: train
        path: user_pnl_summary.parquet
  - config_name: pnl_daily
    data_files:
      - split: train
        path: pnl_daily/**/*.parquet
  - config_name: pnl_daily_resolved
    data_files:
      - split: train
        path: pnl_daily_resolved/**/*.parquet
  - config_name: pnl_daily_no_fee
    data_files:
      - split: train
        path: pnl_daily_no_fee/**/*.parquet
  - config_name: pnl_daily_resolved_no_fee
    data_files:
      - split: train
        path: pnl_daily_resolved_no_fee/**/*.parquet
  - config_name: pnl_change_daily
    data_files:
      - split: train
        path: pnl_change_daily/**/*.parquet
  - config_name: pnl_change_monthly
    data_files:
      - split: train
        path: pnl_change_monthly.parquet
  - config_name: pnl_category_daily
    data_files:
      - split: train
        path: pnl_category_daily/**/*.parquet
  - config_name: pnl_category_daily_resolved
    data_files:
      - split: train
        path: pnl_category_daily_resolved/**/*.parquet
  - config_name: pnl_category_daily_no_fee
    data_files:
      - split: train
        path: pnl_category_daily_no_fee/**/*.parquet
  - config_name: pnl_category_daily_resolved_no_fee
    data_files:
      - split: train
        path: pnl_category_daily_resolved_no_fee/**/*.parquet
  - config_name: trades
    data_files:
      - split: train
        path: trades/**/*.parquet
  - config_name: ohlcv_1d
    data_files:
      - split: train
        path: ohlcv_1d.parquet
  - config_name: ohlcv_1h
    data_files:
      - split: train
        path: ohlcv_1h/**/*.parquet
  - config_name: ohlcv_5m
    data_files:
      - split: train
        path: ohlcv_5m/**/*.parquet

Polymarket Users

Trading activity, profits, and behavioral features for every user on Polymarket, the largest on-chain prediction market. The dataset spans from Polymarket's launch on 2022-11-11 to 2026-03-29 and covers all reconciled end-user trades, daily mark-to-market PnL, and a wide set of user-level behavioral features. Built from on-chain CTF Exchange events on Polygon.

This is the research dataset behind:

Akey, P., Grégoire, V., Harvie, N., & Martineau, C. (2026). Who Wins and Who Loses In Prediction Markets? Evidence from Polymarket. Working Paper. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6443103

Documentation: https://www.vincentgregoire.com/polymarket-users-data

Disclaimer. This is an independent academic research dataset. The authors are not affiliated with, endorsed by, or sponsored by Polymarket. "Polymarket" is a trademark of its respective owner; it is referenced here only to identify the source platform of the underlying public on-chain data.

Quick start

With `polars` (recommended — fast, lazy, scans Hive partitions natively)

import polars as pl

# Markets metadata (one row per market)
markets = pl.read_parquet("markets.parquet")

# Wide one-row-per-user terminal PnL across five variants
users = pl.read_parquet("user_pnl_summary.parquet")
print(users.sort("pnl_total", descending=True).head(10))

# Sparse daily PnL, Hive-partitioned by year/month/day
pnl = pl.scan_parquet("pnl_daily/**/*.parquet", hive_partitioning=False)

With the `datasets` library

Install it first:

pip install datasets
# or with uv:
uv add datasets

from datasets import load_dataset

markets = load_dataset("vgregoire/polymarket-users", "markets")
events = load_dataset("vgregoire/polymarket-users", "events")
user_pnl = load_dataset("vgregoire/polymarket-users", "user_pnl_summary")

Dataset structure

The dataset is a collection of related tables. Each row of the table below maps to a config_name in the loader interface above.

Metadata

Subset	Layout	Rows per	Description
`markets`	`markets.parquet`	market	Per-market metadata including category (classifier-derived), parent event, resolution outcome, fee flags, and lifecycle timestamps.
`events`	`events.parquet`	event	Per-event metadata: tag-based category, number of constituent questions, total trading volume.
`predictions`	`predictions.parquet`	conditional token	Per-token lookup mapping `prediction_id` → parent `market_id`, outcome label and index, number of outcomes, the complementary token id, and the resolution flag. Useful for joining `trades` back to market metadata at the token granularity.

Per-user features and terminal PnL

Subset	Layout	Rows per	Description
`user_features`	`user_features.parquet`	user	Full-sample behavioral feature vector: trade counts, volume, maker/taker share, holding durations, category concentration, distribution metrics, etc. (87 columns)
`user_pnl_summary`	`user_pnl_summary.parquet`	user	Wide terminal PnL across five variants — base, resolved-only, no-fee, spread-adjusted, spread-adjusted resolved-only — each with a total and a per-category breakdown over seven categories (Sports, Crypto, Finance, Politics, Tech, Culture, Weather). The base variant also carries `portfolio_value` (mark-to-market value of open positions) and `usdc_balance` (cash account) so the realized/unrealized split is recoverable.

Daily PnL panels

Subset	Layout	Notes
`pnl_daily`	`pnl_daily/year=YYYY/month=MM/day=DD/data.parquet`	Sparse delta encoding: one row per `(user, day)` where PnL changed. To reconstruct a dense daily series, forward-fill from the last observation.
`pnl_daily_resolved`	`pnl_daily_resolved/year=YYYY/month=MM/day=DD/data.parquet`	Same as `pnl_daily`, restricted to markets that resolved on or before the sample end.
`pnl_daily_no_fee`	`pnl_daily_no_fee/year=YYYY/month=MM/day=DD/data.parquet`	Same restricted to markets with no taker fees (predates the Q4 2024 fee introduction).
`pnl_daily_resolved_no_fee`	`pnl_daily_resolved_no_fee/year=YYYY/month=MM/day=DD/data.parquet`	Intersection of the two filters above.
`pnl_category_daily`	`pnl_category_daily/year=YYYY/month=MM/day=DD/data.parquet`	As above but split by market category. Markets without a category label are excluded.
`pnl_category_daily_resolved`	`pnl_category_daily_resolved/year=YYYY/month=MM/day=DD/data.parquet`	Same as `pnl_category_daily`, restricted to markets that resolved on or before the sample end. Needed to reproduce the resolved-only variants of paper-profits exhibits (concentration, spread decomposition, probit) without inflating the per-(user, category) denominator with users who had no resolved positions in that category.
`pnl_category_daily_no_fee`	`pnl_category_daily_no_fee/year=YYYY/month=MM/day=DD/data.parquet`	Same restricted to markets with no taker fees (predates the Q4 2024 fee introduction). Companion of the no-fee variant of paper-profits exhibits.
`pnl_category_daily_resolved_no_fee`	`pnl_category_daily_resolved_no_fee/year=YYYY/month=MM/day=DD/data.parquet`	Intersection of the two filters above.

PnL change panels

These store the daily/monthly delta in user PnL (pnl_change), not the level. Useful for return-style analyses where you'd otherwise have to first-difference the level series yourself.

Subset	Layout	Rows per	Notes
`pnl_change_daily`	`pnl_change_daily/year=YYYY/month=MM/day=DD/data.parquet`	`(user, day)`	Per-user daily PnL change.
`pnl_change_monthly`	`pnl_change_monthly.parquet`	`(user, month)`	Per-user monthly aggregation of the same.

For a single point-in-time terminal PnL per user (no forward-filling needed), use user_pnl_summary above — it carries the same values plus all five variants and the portfolio_value / usdc_balance decomposition.

Trade-level data

Subset	Layout	Notes
`trades`	`trades/year=YYYY/month=MM/day=DD/data.parquet`	All reconciled end-user trades, one parquet per day. End-user maker/taker addresses are recovered from the CTF Exchange `OrderFilled` event stream. Schema: `trade_id`, `timestamp`, `market_id`, `event_id`, `prediction_id`, `outcome`, `winner`, `category`, `category_original`, `price`, `quantity`, `maker_address`, `taker_address`, `taker_bought`.
`ohlcv_1d`	`ohlcv_1d.parquet`	Per-token daily OHLCV bars (open, high, low, close, volume, trade count) plus daily open interest.
`ohlcv_1h`	`ohlcv_1h/year=YYYY/month=MM/day=DD/data.parquet`	Per-token hourly OHLCV bars, day-partitioned. No open interest column (no matching position cache).
`ohlcv_5m`	`ohlcv_5m/year=YYYY/month=MM/day=DD/data.parquet`	Per-token 5-minute OHLCV bars, day-partitioned. No open interest column.

Forward-filling sparse PnL

The pnl_daily and pnl_category_daily tables are delta-encoded — they only contain rows where PnL changed on that day. To reconstruct a dense daily panel:

import polars as pl
from datetime import date

sparse = pl.scan_parquet("pnl_daily/**/*.parquet", hive_partitioning=False)

# Date grid: every day of the sample period
grid = pl.LazyFrame({
    "snapshot_time": pl.date_range(date(2022, 11, 11), date(2026, 3, 29), "1d", eager=True)
})

# Cross-join users × dates, then asof-join the sparse data
users = sparse.select("user_address").unique()
dense = (
    users.join(grid, how="cross")
    .sort("user_address", "snapshot_time")
    .join_asof(
        sparse.sort("user_address", "snapshot_time"),
        on="snapshot_time",
        by="user_address",
    )
    .collect()
)

Sample, scope, and provenance

Source: Public on-chain data from Polygon. Reconciled from OrderFilled events emitted by the CTF Exchange contract. End-user identification uses Polymarket's proxy/safe wallet pattern.
Sample period: 2022-11-11 — 2026-03-29 (UTC).
Universe: All Polymarket markets, including binary and multi-outcome. Wash trading is detected (via counterparty HHI) but not filtered out of the base PnL — see the paper for the methodology, and use the resolved variant of user_pnl_summary if you want to restrict to fully-settled markets.
PnL methodology: Mark-to-market portfolio_value + usdc_balance from the reconstructed cash account. The spread_adj variants in user_pnl_summary net out a fixed half-spread on each fill (paper default 0.005, i.e., half the 1¢ tick).

Time convention

All timestamps are in UTC, and all daily / monthly bucketing is anchored at midnight UTC. Two labelling conventions are used across panels:

Event panels — natural timestamps. trades and ohlcv_* rows are timestamped at the moment the event occurred (block timestamp) or at the start of the bar (OHLCV). A row with timestamp = 2025-06-15 12:00 UTC is activity on the calendar day 2025-06-15, and the Hive partition path matches (day=15).

Snapshot panels — +1 day right-boundary. pnl_daily, the three pnl_daily_* filtered variants, pnl_category_daily and its three variants, pnl_change_daily, and pnl_change_monthly all label rows with the right boundary of the period they summarize. A label X 00:00 UTC means "state or change up to (but not including) that boundary" — i.e., the close of day X − 1.

Column	A label `X 00:00 UTC` means…
`pnl_daily.snapshot_time` (and category / variant siblings)	State at the close of day X − 1
`pnl_change_daily.day`	Change accumulated during day X − 1
`pnl_change_monthly.month`	Sum of `pnl_change_daily` rows whose `day` falls in calendar month `X` (those daily values are themselves +1-shifted)

The Hive partition path always matches the column value. The convention is chosen for compatibility with polars.join_asof against a daily price grid keyed at midnight UTC. Worked example: a user whose first trade is at 2025-06-15 18:18 UTC first appears in pnl_daily at snapshot_time = 2025-06-16 00:00 UTC (Hive partition day=16) and in pnl_change_daily at day = 2025-06-16 00:00 UTC.

Citation

@unpublished{akey2026prediction,
  title  = {Who Wins and Who Loses In Prediction Markets? Evidence from Polymarket},
  author = {Akey, Pat and Gr{\'e}goire, Vincent and Harvie, Nicolas and Martineau, Charles},
  note   = {Working Paper},
  year   = {2026},
  url    = {https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6443103}
}

License

The processed data in this release — reconciled end-user trades, computed PnL panels and summaries, classifier-derived market categories, behavioral user features, and OHLCV aggregates — is released under CC-BY 4.0. Use, modify, and redistribute freely, including commercially; please cite the paper.

Scope. CC-BY 4.0 covers the authors' contribution: cleaning, reconciliation, classification, computation, and curation. It does not cover fields that originate from the Polymarket API (e.g., market question text, descriptions, slugs, raw platform tags, lifecycle timestamps). Those fields are included for convenience and reproducibility but remain subject to Polymarket's terms of use. Users who plan to redistribute those fields should consult Polymarket's terms directly.

No warranty. As CC-BY 4.0 §5 states, this dataset is provided "AS IS" without warranty of any kind, express or implied.

Changelog

v1.3 (2026-07-06) — Corrected the two directional tail-price features frac_longshot and frac_sureshot in user_features. In v1.0–v1.2 both columns were shipped as byte-identical copies of frac_extreme_price — the two-tail share of a user's fills at an extreme price (price < 0.10 or price > 0.90) — because a placeholder in the feature-engineering code aliased both to the two-tail measure instead of splitting the tails. They are now computed as documented: frac_longshot is the share of a user's fills at price < 0.10 (the long-shot / low-priced side) and frac_sureshot the share at price > 0.90 (the sure-shot / high-priced side), so that frac_longshot + frac_sureshot = frac_extreme_price for every user. frac_extreme_price itself was always correct and is unchanged, as is every other column in user_features and every other table. None of the accompanying paper's exhibits use the two corrected columns (they use only frac_extreme_price), so no published results change. With thanks to Marcos Cardozo (Universidad Católica del Uruguay) for identifying and carefully documenting the issue.
v1.2 (2026-06-14) — Removed 16 maker/taker terminal-PnL columns (pnl_maker_total, pnl_taker_total, and their per-category siblings) from user_pnl_summary. These were shipped undocumented in v1.1 and the per-category role columns did not recombine to the base per-category totals for users who acted as both maker and taker (≈17.5% of users), so they have been withdrawn. The five documented variants (base, resolved, no-fee, spread-adjusted, spread-adjusted resolved-only), their per-category breakdowns, and the portfolio_value / usdc_balance decomposition are unchanged. No other table is affected. A user's maker-vs-taker activity is still fully recoverable from the trades table (each row carries maker_address and taker_address); see the documentation recipes.
v1.1 (2026-06-01) — Corrected PnL panels and per-category breakdowns. Every PnL-related table (pnl_daily, the three pnl_daily_* variants, pnl_category_daily and its three variants, pnl_change_daily, pnl_change_monthly, and user_pnl_summary) was rebuilt from corrected position snapshots. Sample period and the other panels (markets, events, predictions, trades, user_features, ohlcv_*) are unchanged from v1.0.

What changed and why:
1. Phantom-settlement bug fixed. A zero-position row was being dropped before settlement logic ran, which caused round-trip trades on already-resolved markets (buy + sell back within the same day) to produce a ghost USDC settlement for a position the user no longer held. The fix moves the zero-row filter to after settlement decisions. Effect on pnl_daily (base): aggregate user-population terminal PnL moved from $114,798 to exactly $0 (zero-sum restored); the median user's terminal PnL moved by $0.0007; top-5 winners' PnL is bit-identical to 6 decimals.
2. Analysis variants now filter at the market level. The pnl_daily_resolved, pnl_daily_no_fee, pnl_daily_resolved_no_fee variants (and their pnl_category_daily_* counterparts) previously applied the variant predicate only at PnL aggregation time, which produced asymmetric USDC leakage of $1–5 B per variant. The variants are now built from positions that exclude trades on markets outside the variant's subset, so each variant is structurally zero-sum within its subset (modulo platform-collected fees and non-user counterparties). Effect: each variant's aggregate terminal PnL moved by +$1 B to +$4.7 B, from large negative leakage totals down to ≈ $0.
3. Sample composition. A small number of users (≤ 0.05 % of the 2.48 M user base) drop in or out of the panel depending on which variant subset they touch. For the variants, user counts drop by 54 k–127 k as users who only traded outside the variant's market subset now correctly disappear from the panel.
The headline statistics that appear in the accompanying paper round to the same values as in v1.0 — top 1 % capture 76.5 % of profits, top 0.1 % capture 51.2 %, ~70 % of users lose money — but the underlying panels are now internally consistent and pass a whole-cache maker_pnl + taker_pnl = base_pnl invariant across all 2.26 billion daily snapshots.
v1.0 (2026-05-19) — Initial release. Sample 2022-11-11 → 2026-03-29.