Polymarket Data Quality Explained

The quality report is public: recorded books alone reproduced 36.6% of on-chain settlement outcomes in the replay family; replay divergence from the reference chain runs 67.8% at 381,093 checkpoints (6.4% crossed) in the earlier corpus; and the venue's own texture numbers (one-sided 16.9%, crossed 3.24%) are measured, not assumed.

The standard this site holds is symmetrical: every number it quotes is one it could be asked to reproduce from stored frames.

Quality lives in the artifact, not the prose: the Zenodo dataset ships with capture cadence, defect flags, and methodology; the DataFrame/dataset receives a DOI; and each factual page carries a verified-as-of date.

A citable dataset is the audit trail made legible — someone can stand on the reconstruction script and rerun it, which is the definition of quality that survives review.

Quality is measured against what a record can, honestly, contain: the archive states how much of a market's truth is in its books versus its reference feeds, publishes the defect flags, and refuses to filter the row that embarrasses a summary.

The quality story on this archive is that defects are flagged, not filtered: crossed books appear at 3.24% of 5-minute snapshots and one-sided books at 16.9%, and both stay on the rows so a study can measure its own exposure. A dataset that silently drops the crossed frames would flatter every mid-based metric it publishes.

That is the honesty test for any provider: does the documentation name the defect and keep the rows, or hide them? The archive on this site publishes both the flags and the measured rates.

• Fidelity
• Citation guide
• Open data