3.14% of Polymarket Accounts Drive Market Accuracy

A working paper posted to SSRN on April 20 and revised April 25, 2026, analyzed the complete transaction history of Polymarket. The authors-Roberto Gomez-Cram, Yunhan Guo and Howard Kung of London Business School, and Theis Ingerslev Jensen of Yale-examined 98,906 events across 210,322 markets, $13.76 billion in trading volume and 1.72 million accounts. The study used a sign-randomization statistical test to separate trading skill from chance and to classify accounts by performance.

The analysis classified 3.14% of accounts as skilled. Those accounts generated persistent profits and order flow that predicted both next-period price moves and final market outcomes at statistically significant levels. Skilled accounts traded across an average of 79 markets each. The paper reports that a one-percentage-point increase in skilled net buying corresponded to an eight-basis-point rise in the probability of the correct final outcome. Accounts that had positive balances by chance did not show meaningful predictive power in the tests.

The authors tested persistence by splitting events into training and test sets. Forty-four percent of traders labeled skilled in the training period retained that status in the test set. By comparison, a parallel test for skilled mutual funds showed a 10% retention rate. The study also found persistence among poor performers: 51% of unskilled losers in training remained unskilled in the test set.

In event-timing tests, skilled traders tended to trade first around scheduled information releases. The paper reports that only the skilled group shifted order imbalance in the direction of surprises for Federal Open Market Committee announcements and corporate earnings within a narrow window around releases; other groups did not show a consistent response.

Polymarket activity grew sharply over the study period. Monthly trading volume rose from $3.3 million in December 2023 to $1.98 billion in December 2025, and active accounts increased from roughly 1,600 to more than 519,000. The study found that unskilled and unlucky losers made up 67% of accounts and absorbed aggregate losses. Market makers and skilled takers together represented fewer than 3.5% of accounts but captured more than 30% of total gains.

The paper identified 1,950 accounts that met timing and conviction criteria consistent with trading on non-public information. Those accounts averaged about $15,000 in profits each and produced large price effects when active. The authors document one episode in which three accounts took positions on a contract tied to Venezuelan President Nicolas Maduro hours before a secret U.S. military operation on Jan. 3, 2026, collectively earning more than $630,000. On April 23, 2026, the Commodity Futures Trading Commission filed a complaint alleging an active-duty U.S. Army service member used one of those accounts.

The authors wrote that the insider events they identified were concentrated and noted uncertainty about whether identified skilled traders will continue participating if platforms grow and fees rise.