What is a cohort audit?

The answer first

A cohort audit scores a defined set of Polymarket wallets together rather than one at a time. Each wallet's realized entry edge is read with a bias-corrected and accelerated (BCa) bootstrap 95 percent interval and its resolved-position count, never a bare point estimate. Then two things happen that a single-wallet check cannot do: a false-discovery-rate correction is applied across the whole set, and a size-matched negative control is run. The report states which wallets, if any, are distinguishable from chance at the frozen bar. This is the enterprise engagement deliverable.

Why a cohort needs a correction

The reason you cannot just run the single-wallet test 40 times is multiplicity. If each wallet has a 5 percent chance of clearing an uncorrected bar by luck, then screening 40 wallets with no real skill still throws up two positives on average. Rank a leaderboard by the raw point estimate and the top of the board is exactly where that luck concentrates. A cohort audit answers the multiplicity problem head on: the bar each wallet must clear to be called distinguishable from chance rises with the number of wallets tested. That is the false-discovery-rate correction, applied here with the Benjamini-Hochberg procedure across every wallet in the set.

The negative control

A correction tells you the bar is high enough in theory. A negative control checks it in practice. Convexly runs a size-matched placebo cohort, records with no reason to carry skill, through the identical pipeline. If the placebo cohort produces about as many positive reads as the real one, the method is finding noise and the real result is discounted accordingly. The negative control is also what makes a null cohort audit, in which nobody clears the bar, a genuine and reportable outcome rather than a failure to find something.

Worked reference: our own top-50 cohort

We ran this on ourselves. In the frozen 2026-06-09 scan of our own published top-50 cohort (full table at /research/top50-skill-scan), 35 wallets had enough resolved positions to test. Each was scored with its realized entry edge, its BCa 95 percent interval, and its resolved-position count. After the Benjamini-Hochberg correction across all 35 tests, zero wallets cleared the corrected bar. One wallet showed +5.6 probability points across 137 resolved positions with an interval of [+1.2, +9.5] that clears zero on its own, but as one uncorrected test among 35 it is consistent with chance at the cohort level, so it is not in the corrected cleared set. That is the cohort audit doing its job: a board that looks full of winners can hold zero wallets that survive the multiplicity correction.

What a cohort audit does NOT do

It does not rank dollar PnL, which on a fat-tailed market is dominated by a few large positions and by luck at these sample sizes. It does not forecast any wallet's next trade: the read is retrospective and in-sample, and the per-wallet temporal holdout on this statistic did not clear the filed threshold, so a cleared wallet is not a prediction and not a signal to copy any trader. And a cohort audit does not manufacture a winner to satisfy a request; a null result, where nobody clears the bar, is a valid deliverable.

Read a single wallet first

A cohort audit runs as an enterprise engagement on a wallet set you supply, returning the cohort with per-wallet intervals, resolved-position counts, the FDR-corrected four-state verdicts, and the negative control. To try the underlying per-wallet statistic yourself, paste any Polymarket address into the free analyzer. Same interval, same resolved-position count, one wallet at a time.