Who Wins and Who Loses in Prediction Markets?

Read from the public on-chain record, the honest answer has three parts. Most resolved Polymarket records cannot be separated from chance at retail sample sizes. Many of the largest P&Ls are concentration stories, with one event accounting for most of the net result. And a P&L screenshot is not a skill receipt; the receipt is the realized entry edge with its 95% interval against zero. Each of those reads is reproducible on any public wallet with the free Polymarket wallet analyzer. No signup, no wallet signature, public data only.

Looking for the PDF by this title?

A widely searched academic paper by that title examines Polymarket trader outcomes. We do not host the PDF and we are not affiliated with its authors; search the exact title on Google Scholar for the authors' copy. What follows is Convexly's own complementary read of the same question, computed from public on-chain data with a method you can run yourself on any wallet.

Read one: most records cannot be separated from chance

The test we apply to a record is simple to state: compute the realized entry edge (how much more often the wallet's entries resolved true than the price it paid implied) across its resolved positions, then ask whether the bootstrap 95% interval around that edge clears zero. On 2026-06-09 we ran our skill read on our own published top-50 cohort; 35 of 50 had enough resolved positions to test; exactly one interval cleared zero on the positive side, roughly what chance predicts across 35 tests at a 2.5% threshold (about 0.9 false positives expected), and it did not survive a multiple-comparisons correction.

The cohort behind that scan is our own published top-50 board: wallets drawn from the frozen 8,656-wallet reference cohort, each with at least 20 resolved positions, refreshed daily and ranked by Edge Score V3b. Edge Score V3b is a descriptive in-sample composite; its only out-of-sample forward test held a +0.11 rank correlation (95% CI 0.05 to 0.18) against future PnL and did not clear the pre-set +0.30 bar. That forward miss is exactly why the scan reads realized edge intervals rather than the score itself, and why we publish the result even though it is unflattering to our own board.

Read two: concentration drives many large P&Ls

Many of the biggest winning records on Polymarket are one-event stories: a single resolved market accounts for most of the net result. A record like that may be a brilliant call or a coin flip that landed; the resolved data cannot distinguish the two. Convexly's free check flags any record where a single event is at least 60% of the net result as too concentrated to read, and that flag overrides any flattering point estimate. When you see a leaderboard P&L, the first question worth asking is not “how big” but “how spread”.

Read three: the screenshot is not the receipt

A P&L screenshot is a point estimate with the uncertainty cropped out. The honest read of a record is the realized entry edge with its bootstrap 95% interval against zero: if the interval clears zero on the positive side, the record is consistent with an edge; if it includes zero, the record cannot be told apart from chance, whatever the dollar figure says. On retail sample sizes the interval is usually wide, which is why “cannot tell” is the most common honest verdict and why we built the check to say it plainly.

Diagnostics on public on-chain records, not investment advice; a past read is not a forecast.