A long-shot screen on historical Polymarket fills: base rates and a correction

Can a screen built on public fills find insider trading on Polymarket? This study tests one against 714,691,643 fills on settled markets between 1 January 2025 and 28 April 2026, from the Polymarket-v1 dataset by Boka Qin and Rui Yang. A bet is one wallet-market-side episode whose rolling one-hour stake reached $2,500 on a side priced at 35 cents or less. Across the platform that bet type returned −3.67%; in politics it returned +16.87%. The tight screen keeps bets by wallets under two weeks old that placed almost nothing else and acted within 48 hours. Its corrected version, published 8 October 2026, holds 1,316 bets, and against matched bets their gap is −5.40 cents per share, with a market-bootstrap 95% interval from −10.53 to −0.07 cents. The correction replaced raw token notional with directional acquisition cost in the concentration gate. The earlier sensitivity tables stay on the page as historical results and are not to be combined with the corrected figures. The two charged cases of 2026 are described alongside the base rates a screen has to beat: wallets with perfect records against chance, what a success-rate filter selects, and how many new wallets had traded before (134 of 250 first seen in the record's first month). The record holds the page at https://vultax.com/research/polymarket-longshot-insider-screen-base-rate as a research note (PDF) and page (HTML, PDF) (27 October 2026), the base-rate CSV, the wallet-age audit folder (CSV, results, methods, reproduction zip), the proof-scope record and a CITATION.cff. Wallets appear as hashes.