--- language: - en license: apache-2.0 size_categories: - 1M` suffix | | `condition_id` | string | CTF condition id (market key) | | `market_slug` | string | Polymarket market slug | | `token_id` | string | ERC-1155 outcome token id (uint256, decimal string) | | `side` | string | `BUY` or `SELL` | | `outcome` | string | Outcome leg the token represents (`Yes` / `No`) | | `price` | double | Execution price, 0–1 | | `size` | double | Shares traded | | `bet_value` | double | USD notional (`price × size`) | | `proxy_wallet` | string | Lowercase 0x-address whose position changed | ### `markets.parquet` | Column | Type | Description | | :--- | :--- | :--- | | `condition_id` | string | CTF condition id | | `slug` | string | Polymarket market slug | | `title` | string | Market question | | `category` | string | Polymarket category | | `end_date_ms` | int64 | Scheduled market end, epoch ms | | `resolved_outcome` | string | Winning outcome (`Yes` / `No`) | | `resolved_at_ms` | int64 | On-chain resolution time, epoch ms | | `final_price` | double | Final YES price | | `fill_count` | int64 | Fills in this file | | `volume_usd` | double | Total `bet_value` | | `first_fill_ms` | int64 | First fill in this file | | `last_fill_ms` | int64 | Last fill in this file | ## Quickstart ```bash pip install duckdb ``` ```python import duckdb con = duckdb.connect() FILLS = "hf://datasets/Crowdintel/polymarket-historical-data/fills_2024_us_presidential.parquet" # Hourly volume and VWAP around election night con.sql(f""" SELECT date_trunc('hour', to_timestamp(timestamp_ms / 1000)) AS hour, outcome, count(*) AS fills, round(sum(bet_value), 0) AS volume_usd, round(sum(price * size) / sum(size), 4) AS vwap FROM read_parquet('{FILLS}') WHERE timestamp_ms >= epoch_ms(TIMESTAMP '2024-11-05 00:00:00') GROUP BY 1, 2 ORDER BY 1 DESC """).show() ``` `hf://` reads only the row groups the query touches, so this is fast without downloading the file. To work offline, download the parquet with `huggingface_hub.hf_hub_download("Crowdintel/polymarket-historical-data", "fills_2024_us_presidential.parquet", repo_type="dataset")`. Runnable end-to-end example: [quickstart.py](https://github.com/crowdintel-xyz/polymarket-historical-data/blob/main/quickstart.py). ## Provenance - Source: Polymarket CTF `OrderFilled` events on Polygon, indexed by CrowdIntel from the exchange contracts and joined to Gamma market metadata and on-chain resolution. - `tx_hash` is unique across the file (one row per fill). The file is sorted by `timestamp_ms`. - First and last fills, row counts, and volume in the table above were computed from the published file, not from an estimate. - Resolution rows come from `market_outcomes`, populated from the on-chain condition resolution, not from an off-chain leaderboard. Independent checks a reviewer can run: ```python # No duplicate fills con.sql(f"SELECT count(*) - count(DISTINCT tx_hash) AS dupes FROM read_parquet('{FILLS}')").show() # Every fill sits inside the market's trading window con.sql(f""" SELECT count(*) FROM read_parquet('{FILLS}') WHERE timestamp_ms < epoch_ms(TIMESTAMP '2024-01-04 20:36:24') OR timestamp_ms > epoch_ms(TIMESTAMP '2024-11-06 12:07:20') """).show() ``` ## Known limitations - **Two markets, not the whole exchange.** This is a closed-event extract chosen for end-to-end backtesting, not a sample of all Polymarket activity. - **No maker/taker split.** Polymarket fills a maker order against a taker; the public tape records the wallet whose position changed (`proxy_wallet`). It does not attribute the resting side. - **No order book history.** L2 snapshots are not part of this release. Do not infer depth from fills. - **`market_slug` repeats per row** so the file is self-contained. Join `markets.parquet` for metadata rather than parsing the slug. - **One fill is excluded.** A 2024-10-21 fill in the Trump market has no token id in the source, so its outcome leg cannot be attributed. 7,655,165 of 7,655,166 source rows ship. - **`outcome` is derived from the CLOB token map, not the legacy column.** The stored `trades.outcome` was inverted for older backfills; this file joins `token_registry`. Join on `token_id` and you get the same answer. - **Prices are floats.** They come from the CLOB's fixed-point prices; values are exact to the tick, not the original integer representation. ## License Apache-2.0. Attribution to CrowdIntel is required for redistribution; a citation is appreciated: ```bibtex @dataset{crowdintel_polymarket_2024, title = {Polymarket Fills — 2024 US Presidential Election}, author = {CrowdIntel}, year = {2026}, url = {https://huggingface.co/datasets/Crowdintel/polymarket-historical-data} } ``` ## Need more than one election? CrowdIntel maintains the full tape (1.9B+ fills, 2.2M+ resolved markets) plus the reconciled wallet graph: PnL after NegRisk unwrapping, shrinkage-adjusted win rates, and funding-cluster ids. API, ClickHouse pipes, and weekly delta drops: [crowdintel.xyz/data](https://crowdintel.xyz/data). ## Changelog - **2026-10-19 (2)** — Corrected `outcome`: the legacy column had Yes/No inverted for these markets (verified against `clob.polymarket.com/markets/{condition_id}`). Rebuilt from the token map; dropped one fill with no token id. - **2026-10-19** — Replaced the synthetic sample with the real closed-event tape (7.65M fills) in Parquet, added `markets.parquet`, corrected every claim on this card, removed the synthetic order-book subset.