Polymarket study reveals fill-side behavior is unimodal, not multi-archetype

A new empirical study by Maksym Nechepurenko examines non-retail trading on Polymarket, analyzing 13,356,931 OrderFilled events from 77,204 addresses across 43,116 markets during April 21–27, 2026. The paper, the fourth in a four-part program on prediction market microstructure, reveals a structural limitation: Polymarket's off-chain CLOB architecture prevents address-level quote lifecycle attribution because OrderPlaced and OrderCancelled events are not publicly archived. This G-QUOTE-LIFE failure forces researchers to restrict analysis to a fill-side vector of six features, ruling out traditional microstructure metrics like quote intensity or posted spreads.

Applying DBSCAN clustering across 15 sensitivity configurations on the fill-side vector yields a single dense cluster with zero noise—meaning fill-side behavior is unimodal, contradicting the pre-registered hypothesis of four to five separable archetypes. However, a clustering-independent feature-tier stratification successfully separates retail from non-retail: whale-tier, high-frequency-operator, and power-trader tiers together hold 81.4% of total notional value while comprising only 12.6% of active addresses. The author withdraws claims about market-making detection and reduces spoof manipulation detection to market-level book diagnostics. A companion dataset (PMXT Bundle 3) is deposited on Zenodo.