Is that Polymarket trader skilled or lucky?

What the 2026 studies found

Four independent looks at Polymarket’s population landed on the same shape this year. A London Business School and Yale team studied the universe of Polymarket transactions, 1.72 million accounts and $13.76 billion in volume, in “Prediction Market Accuracy: Crowd Wisdom or Informed Minority?” and concluded that the platform’s celebrated accuracy comes from roughly 3% of accounts. That skilled minority kept its classification out of sample 44% of the time, against 10% for skilled mutual funds, so the label reflects persistent ability rather than a good quarter. CoinDesk’s read of the same data reported that under 1% of wallets take half the profits.

The population everyone else belongs to looks like coin flips at posted prices. Pew Research examined 11,989 active accounts over the six weeks from May 7 to June 19, 2026. The median trader placed 46 trades, and the average trader finished within 2 dollars of break-even. And when PANews dissected 27,000 transactions from the top ten profit-ranked whales, reported win rates of 50 to 83% fell to 50 to 58% once open losing positions were counted, a range the authors described as similar to flipping a coin. Headline profits, famous wallets, and high win rates all survived until someone graded them against prices. Then most of them stopped being evidence.

Why profit and win rate cannot answer the question

Profit measures bankroll and variance as much as judgment. One oversized position that resolved well dominates a P&L ranking, and the whale study above shows how open positions let a public record flatter itself. Win rate fails for the opposite reason. A wallet buying 90 cent favorites wins nine times in ten by construction, so a 90% win rate can carry zero information beyond what the market already priced. The folk thresholds that circulate, 60 or 65% win rate as a mark of skill, sit comfortably inside what chance plus favorable prices produces. The win-rate evidence from the graded population documents this with numbers. Among wallets with 1,000 or more graded events, the 90th percentile win rate is 99.9%, and the median record in that group sits within about one win per hundred of what its entry prices predicted (as of 2026-09-27).

The test that answers it

Every entry price states a probability. A position bought at 30 cents claims the market understates a 30% outcome; if it resolves favorably, that is a win the market priced at three-in-ten. Summing those stated probabilities across a wallet’s resolved positions gives the number of wins its own prices predicted. The record’s actual wins sit either inside or outside the interval that chance produces around that prediction, and because each position carries its own probability, the interval comes from the Poisson binomial distribution rather than a single coin-flip rate. One adjustment matters most. Each set of mutually exclusive outcomes is one event, so fifty candidate markets in one race count once at most, however many fills they took, and a wallet that bet on more than one of the candidates has that race left out of the test. Separate yes-or-no markets each count, even on one subject, except a repeating series such as one price threshold asked for several dates, which we count as one event.

The result reads as one of three verdicts. Distinguishable means the excess wins clear the 95% interval. Consistent with luck means they sit inside it, which profitable wallets manage regularly. Insufficient evidence means too few independent resolved events for the interval to say anything at all, the verdict every short hot streak earns regardless of how it looks. The full method, including the independence rules and the out-of-time validation that tested whether high grades persist, is on the methodology page and the research page.

The three ways eyeballing a record goes wrong

• Favorite collecting. Long win streaks built at 90 cents and up perform as priced. The record is real, the skill claim is not supported, and only the entry prices reveal the difference.
• One correlated cluster. A wallet that swept dozens of markets on a single outcome holds far less evidence than its market count suggests. Screenshots of such sweeps circulate weekly as proof of genius. The checker counts a multi-outcome question once at most, and leaves it out when the wallet bet on more than one of its outcomes. Separate yes-or-no markets settled by the same outcome each still count, so the number of separate subjects a record spans is what to look for.
• Short samples. Twenty resolved events is the floor at which a verdict interval becomes meaningful. Most records that go viral are far below it.