alphalab-research/polymarket-2024

Polymarket 2024 — NegRisk Wash-Flagged Trade Dataset (v3.1)

Maintainer: Alpha Lab Research · research@alphalab.dev Methodology stamp: v3_theta_tuned_per_market License: CC-BY-NC-4.0 (research / non-commercial). Commercial use requires a separate license — contact the maintainer.

Scope (read first)

This is the Phase 1 honest-scope release:

• Coverage window: 2024-01-05 → 2024-12-07 (≈11.1 months, 16,959,221 trades, $3,545,492,028 total notional, 199 unique Polymarket V1 NegRisk markets).
• Source: Polymarket V1 CTFExchange via the Goldsky subgraph backfill, filtered to the WS-N family-reconstruction universe (the 38 highest-priority NegRisk markets).
• Schema and family bridge: 78-digit decimal market_id (the ERC-1155 CTF token id) is bridged to a 32-byte NegRisk family_id via the CTFPositionDeriver (Alpha Lab WS-BF.2). Rows whose token id resolves through the family scan get family_id_unknown=False; rows that don't resolve still ship with family_id set to the token id itself plus family_id_unknown=True.

Extended version coming Q3 2026. The 2025-01 → 2026-05 window (covering pre-V2 history and the post-V2 cutover) lands as polymarket-2024-2026 once the AMS3 universal archive is unblocked. This release does NOT include those windows.

Methodology summary

Per-row wash flag = output of the V3WashDetectorWithThetaTuning (sirolly_alg2 strategy, base θ₀=0.70). The detector composes round-trip count, size parity, side parity, and volume share into a [0, 1] wash_score, then thresholds against a per-market θ that compresses logarithmically with market size:

theta_market = 0.70 / log10(max(market_volume_usd / 1000, 10))

A $10k market keeps θ=0.70; a $1M market gets θ≈0.23; a $1B market gets θ≈0.12. The compression matches the logarithmic decay of the maximum achievable volume-share feature on whale-dominated markets — pair-level wash signal that a flat 0.70 cutoff misses on $100M+ venues.

This per-market θ tuning is the methodological novelty of the companion arXiv preprint (Alpha Lab Research, 2026). The pair-level scorer is a strict subset of the wallet-level Sirolly et al. (SSRN 5714122) detector; on this window the overall pair-level wash rate is 0.22%.

Schema

Column	Type	Source	Description
trade_id	string	Goldsky id	${txhash}_${orderhash} shape
market_id	string	Goldsky	78-digit decimal CTF token id
trade_at_utc	timestamp	derived	UTC trade timestamp
proxy_wallet	string	Goldsky maker	hex address, 0x-prefixed
counterparty_wallet	string	Goldsky taker	hex address
side	string	derived	BUY_TOKEN / SELL_TOKEN
size	double	derived	USD notional from the USDC leg
transaction_hash	string	Goldsky	dedupe key part 1
order_hash	string	Goldsky	dedupe key part 2
family_id	string	WS-BF.2 bridge	NegRisk family ID (zero-byte form) or fallback token id
family_id_unknown	bool	derived	True if no NegRisk question resolved
question_id	string	WS-BF.2 bridge	32-byte NegRisk questionId, NA if unresolved
question_index	int / NA	WS-BF.2 bridge	per-family question slot, NA if unresolved
wash_score	double	V3 detector	[0, 1] continuous score
wash_flag	bool	V3 detector	wash_score ≥ theta_used
theta_used	double	V3 detector	per-market θ applied to this row
wash_reasoning	string	V3 detector	human-readable explanation
wash_method_version	string	V3 detector	v3_theta_tuned_per_market
archiver_run_id	string	provenance	static-export build ID (synthetic, one per build)
fetched_at_utc	timestamp	provenance	source parquet mtime (synthetic, one per build)

The dedupe key is (transaction_hash, order_hash). Provenance columns are synthetic for this static export — the Goldsky-V1 backfill predates our streaming archiver, so archiver_run_id is one constant per build and fetched_at_utc is the source parquet's file mtime. The columns are included so the schema matches the upcoming streaming-era dataset and downstream code does not need to branch.

Partitions

year=2024/month=01/data.parquet
year=2024/month=02/data.parquet
...
year=2024/month=10/data.parquet

Hive-partitioned. datasets.load_dataset("alphalab-research/polymarket-2024") picks up year/month automatically (they are not duplicated in the file body).

Limitations (you should read this before citing)

1. SMKS December 2024 60% peak is OUT OF WINDOW. We do not confirm or refute it. The follow-up extended dataset is the right artifact to cite for the peak.
2. Per-category breakdown is OUT OF SCOPE. Goldsky carries no category field; reproducing SMKS's Sports / Election / Politics / Crypto split needs the Gamma /markets REST join, deferred to v4.
3. Pair-level vs. wallet-level. Our wash_flag is the deterministic pair detector — the loud "ping-pong" subset of SMKS's wallet-level iterative scorer. Numbers here will read lower than headline SMKS wallet-level numbers on the same window; that's expected by construction.
4. Wash detector is a heuristic, not a ground-truth labeller. Flagged wallets are not necessarily operated by a single beneficial owner; the detector identifies the network pattern, not the operator.

Citation

If you use this dataset, please cite the companion preprint:

@misc{alphalab2026wash,
  author = {Alpha Lab Research},
  title  = {Replicating Network-Based Wash Trading Detection on Polymarket V1
            NegRisk (2024-01 to 2024-10), with a per-market θ tuning
            extension},
  year   = {2026},
  note   = {arXiv preprint; HuggingFace dataset:
            alphalab-research/polymarket-2024}
}

Provenance + integrity

This dataset was built by alpha_lab/saas/hf_dataset_v3_publish.py at 2026-05-26T17:45:35.063794+00:00 from 199 Goldsky parquet files. The build is deterministic on the source parquets — re-running with the same inputs produces byte-identical partitioned parquet output.

No personally-identifying information is included. Wallet addresses are on-chain pseudonyms and are public on the Polygon blockchain.

Daemons untouched. The build does not interact with the Alpha Lab real-time archiver (Goldsky 91128) or the arb monitor (2192); it reads static parquet from local disk only.