Polymarket-v1 Database

Battle Count: 71

Highly relevant for quantitative trading in prediction markets and binary options. Key contributions: (1) Demonstrates that standard trade classification algorithms (tick rule, BVC) achieve near-random accuracy (~50%) in prediction markets, invalidating traditional TCA and informed trading detection methods; (2) Shows systematic price-level gradient in classification errors driven by positive direction autocorrelation and concentrated market-making; (3) Documents that classification errors propagate into VPIN, OFI, and spread estimates, distorting liquidity and adverse selection measurements; (4) Establishes that ground-truth microstructure quality (True VPIN, Gibbs spread) predicts forecasting accuracy (Brier scores) in ways classified proxies cannot recover; (5) Provides empirical evidence on fee reform impacts on market quality via staggered DiD. Directly applicable to algorithmic execution, market making, and risk management in prediction market environments.

Moderate complexity. The dataset itself is straightforward (parquet files with clear schema). However, the analytical pipeline involves: (1) Ground-truth direction normalization across binary complementarity (YES/NO token equivalence); (2) Relayer filtering (excluding ~53% of nominal trades); (3) Standard Binary market identification (neg_risk=false, both legs trading, ≥30 trades); (4) VAR/VECM estimation with Cholesky identification requiring sufficient price observations; (5) Staggered DiD with TWFE and event-study specifications; (6) Gibbs sampling for effective spread estimation; (7) Multi-graph cycle detection for wash trading. The econometric methods are standard but the data engineering (1.2B records, 42 partitions, metadata joins) requires significant computational infrastructure.

Dataset is publicly available on HuggingFace (https://huggingface.co/datasets/TimeSeventeen/Polymarket-v1). Complete on-chain trade tape with 100% ground-truth direction coverage. Frozen metadata snapshot ensures reproducibility. 42 parquet partitions with clear documentation of filtering criteria (relayer exclusion, Standard Binary definition). All econometric specifications (DiD, VAR, VECM, OLS) are fully described with equations.