Skilled or lucky? Run the significance test
In 2026, researchers studying Polymarket concluded that a small minority of traders drives prediction-market accuracy while the median account loses a little. The method behind that conclusion is testable on any record: compare the wins against what the entry prices themselves predicted. Type a record below, yours or one you are curious about, and the calculator runs that test. No wallet address, no signup, nothing stored.
Past accuracy does not predict future outcomes. Nothing here is investment advice.
Three records, three verdicts
- 30 wins from 50 events at 55 cents average. A 60% win rate, and the verdict is consistent with luck: chance at those prices produces this often. The interval runs from -8.8 to +18.8 excess wins per 100.
- 24 wins from 25 events at 90 cents average. A 96% win rate, still consistent with luck: the prices already predicted 22.5 wins, so beating them by 1.5 proves little.
- 140 wins from 214 events at 55 cents average. Distinguishable: about 10 excess wins per 100 events with an interval of +3.8 to +17.1, clear of zero.
Where this test comes from
The price-relative principle is the backbone of the 2026 academic work on prediction-market skill: the SSRN study “Who Wins and Who Loses in Prediction Markets” and the coverage of it by CoinDesk and Yale Insights popularized the finding that accuracy concentrates in a few percent of traders. OVERROUND applies the same principle at production scale. It grades wallets’ resolved positions against the prices they paid, with the method described on the methodology page and its out-of-time validation published in full on the research page.
How does the test work?
The entry price is treated as the forecast. Under the luck hypothesis, an event entered at 55 cents is won 55% of the time. Your win count is compared with what your own prices predict, using the Poisson-binomial distribution's normal approximation. If the gap clears the 95% interval that chance produces, the record is distinguishable from luck; otherwise it is consistent with it.
Why does it ask for independent events instead of bets?
The legs of one question are one piece of evidence, however many there are. A hundred positions on one race's candidates count once. Counting legs as events is the single most common way records overstate themselves, and the reason a 12-leg run on one question proves nothing.
Is a 60% win rate good?
It depends entirely on the prices. Winning 30 of 50 events at an average entry of 55 cents is consistent with luck: the interval spans -8.8 to +18.8 excess wins per 100. Winning 140 of 214 at the same price is distinguishable. Same win rate arithmetic, different evidence. Our win-rate guide covers the full population data.
Why does the calculator refuse small samples?
Below 20 independent events, or with a book of near-certain prices, the normal approximation is not trustworthy and the honest answer is insufficient evidence. The public wallet checker applies the same refusal.
For an on-chain record, skip the typing
The free wallet checker pulls any Polymarket wallet’s graded record and runs this test with event-family collapsing applied automatically.
Check a wallet →And the test applied to our own output, every alerted position graded in public: the receipts ledger.
The Weekly Skill Report. Each Monday, one free email counts the positions the alert feed first flagged last week and how many have since won or lost. It names the busiest categories and the ledger’s running total, and links to every graded row.
Free. Unsubscribe anytime. Privacy.