Whether Is Not When: Separating Occurrence from Timing in Polymarket Prices (arXiv (Cornell University))

Summary

• Between 2022 and September 2026, over 3.1 million accounts traded contracts with a face value of about USD 110 billion in 1.24 billion trades across 2.37 million Polymarket markets.p.1
• One day before public Web reports, prediction market prices correctly predicted 56% of 193 geopolitical events but all 51 scheduled central-bank decisions.p.5
• Network-based filtering identifies that wash trades accounted for 13.8% of traded shares across the analyzed sample markets.p.4
• Timing profits are highly concentrated, with the top 1% of accounts with positive timing profits capturing 80% of all timing profits.p.7
• Reading near-term warnings from an entire deadline ladder yields an average precision of 0.30 and a Brier score of 0.033, outperforming single contract readings.p.7

Key insights

1. One day before public Web reports, prices correctly predicted only 56% of geopolitical events compared to 100% of central-bank rate decisions.p.1
2. Information regarding when an event happens accounts for only about one eighth (13%) of total price information across deadline ladders.p.6
3. Net buying by new or large accounts does not help predict whether an event will happen within days beyond current market prices.p.7
4. Network filtering shows that 13.8% of traded shares in the studied prediction markets were wash trades.p.4
5. One day before an event, median contract prices ranged from 0.10 for deadlines within one day to 0.78 for deadlines 90 days or more after the event.p.6
6. Evaluating whole deadline ladders achieves a 44.5% warning precision for events within seven days compared to 24.1% for single contracts.p.7
7. A small group of 337 accounts trading in at least ten ladders demonstrated statistically significant timing skill, posting liquidity for 62% of their volume.p.7