Hugging Face's logo

Datasets:
kinzikdza
/
polymarket-updown-microstructure

Modalities:
Tabular
Text
Formats:
parquet
Languages:
English
Size:
1M - 10M
Tags:
finance
prediction-markets
microstructure
polymarket
order-book
adverse-selection
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
polymarket-updown-microstructure / README.md
kinzikdza's picture
kinzikdza
fix reproduce: pip install honest-backtest[parquet] + parquet dir (no sqlite); card is parquet-only
05d7404 verified
preview code
|
Raw
History Blame Contribute Delete
5.99 kB
metadata 
license: cc-by-4.0
language:
  - en
tags:
  - finance
  - prediction-markets
  - microstructure
  - polymarket
  - order-book
  - adverse-selection
pretty_name: Polymarket Crypto Up/Down Microstructure (5m/15m)
size_categories:
  - 1M<n<10M
configs:
  - config_name: slots
    data_files: parquet/slots.parquet
  - config_name: book_snapshots
    data_files: parquet/book_snapshots.parquet
  - config_name: pm_trades
    data_files: parquet/pm_trades.parquet

Format. Three tables are published as parquet (under parquet/) for the Hub viewer and pandas/polars/datasets users — pick a table from the config dropdown above. The honest-backtest loader reads this parquet/ directory directly via its parquet adapter (adapters.parquet_pm.load_corpus) for one-command reproduction of the paper's results — see Reproduce the headline result below.

Dataset card — Polymarket crypto up/down microstructure (5m/15m)

Six weeks of real order-book microstructure, trade tape, and oracle-settled resolutions for Polymarket's BTC/ETH/SOL/XRP "Up or Down" binary markets (5-minute and 15-minute slots). Captured live, May 27 → Jun 24 2026.

Built following Datasheets for Datasets (Gebru et al.).

Motivation

Why does retail lose the near-close crypto up/down game even when their signal "predicts"? The honest answer needs real microstructure: the book you saw, the trades that actually printed, and the ground-truth outcome — at sub-second resolution near the close. This dataset is that substrate, released so the honest-backtest results are reproducible and so others can study adverse selection on a real venue.

Composition

Three parquet tables; the schema is exactly what the honest_backtest parquet adapter reads:

table one row per key fields
slots market (a single up/down slot) condition_id, coin, duration, open_ts, close_ts, strike, spot_at_open/close, yes_token_id, no_token_id, resolved_side, fee_rate
book_snapshots top-of-book sample (~30s steady, ~1Hz near close) ts_ms, secs_to_close, yes/no best bid/ask + sizes, top-5 ladder strings, spot, ds_spot
pm_trades a real trade print ts_ms, token_id, price, size, taker_buy (1=buy/0=sell)

condition_id/token_id/slug are Polymarket-public market identifiers; they map to real, already-closed markets. There is no account, wallet, order, or trade-of-ours information of any kind (see Anonymization).

Collection process

A single always-on recorder subscribed to Polymarket's CLOB market WebSocket (book + trades) and an independent spot/oracle feed. book_snapshots sampled top-of-book on a cadence that tightens to ~1Hz inside the final ~5s of each slot. pm_trades is the raw public trade tape.

Ground-truth resolution (slots.resolved_side) comes ONLY from Polymarket's gamma outcomePrices (the authoritative settlement oracle), backfilled by a 15-minute cron — never inferred from spot. A YES token pays $1 iff resolved_side='Yes'; a NO token pays $1 iff 'No'.

Anonymization

The export whitelists three market-observation tables and a fixed column list. Everything tied to the operator was excluded by construction: our orders, fills, positions, round-trips, redemptions, on-chain tx hashes, strategy labels, and maker/shadow decision logs. The exporter asserts: the output contains exactly {slots, book_snapshots, pm_trades}, no column name matching order/tx/wallet/strategy/pnl, and no wallet/tx-hash-shaped string in text fields.

Known caveats (read before modeling)

  • Crossed / phantom book (~98% of raw deltas; visible in snapshots too). Displayed ladders are frequently crossed (yes_ask + no_ask ≠ 1; sums < 1 seen routinely). The inside market must be anchored on pm_trades (real executions), not taken at face value from the book. This is the central data- quality fact — a naive "buy at displayed ask" backtest is fiction.
  • Stale levels rarely removed. Old ladder levels persist; depth grows over a slot's life. Filter to recent/large sizes or anchor on the trade tape.
  • Window starts May 27 deliberately. An earlier capture era (May 11–23) had ~39% of tight-market resolved_side corrupted by a spot-based inference bug, fixed by moving to gamma-only resolution. That era is excluded here.
  • spot is a Binance-frame consensus, ~10bp off the data_streams oracle Polymarket settles on. For cross-probability features, use spot moves vs spot_at_open, not absolute spot-vs-strike. Grading uses resolved_side regardless, so outcomes are unaffected.
  • Order-side data is intentionally absent (only the public trade tape is here); the order-rejection / fill-selection asymmetry is studied in the paper via the operator's private ledger, not in this public release.

Recommended uses

Microstructure research; adverse-selection and execution-quality studies; realistic backtesting with honest-backtest; calibration of fair-value models for short-horizon binary markets. Not a source of a profitable trading signal — the accompanying paper shows realizable edge is ~0 on this venue by construction.

Reproduce the headline result 

pip install "honest-backtest[parquet]"      # numpy core + pandas/pyarrow for parquet

# fetch the parquet tables from this dataset
huggingface-cli download kinzikdza/polymarket-updown-microstructure \
    --repo-type dataset --local-dir pm_data

# run the calibration anchor against the parquet/ directory
python -m honest_backtest.examples.no_overpriced pm_data/parquet
# expect headline edge_real ~ 0 / negative (live anchor was -0.004)

License

CC BY 4.0. Polymarket market identifiers are public; this release adds no proprietary or personal data.