Before you copy a Polymarket trader, check skill vs luck

Entry edge is not money

This is the one idea to take from the page. The honest read of a wallet's past record is its realized entry edge: how much more often its entries resolved true than the price it paid implied, reported with a bootstrap 95% interval against zero. That number describes the entries. It does not describe the sizing, the timing, or the exits.

A wallet can show a positive realized entry edge with its confidence interval and still be net-negative on those same entries. The reason is that a copier inherits the entries but applies their own sizing, and sizing is a separate decision the edge number never captures. How much was staked on each entry, and in what order, is where most of the realized outcome actually comes from on a venue with a tail this heavy.

How heavy is the tail? Hill alpha on per-wallet PnL in our V1 cohort is 1.28 (95% bootstrap CI 1.20, 1.36), well below the equity-market value near 3.5 and borderline on whether even the population mean of per-wallet PnL is well-defined. When the tail is that heavy, sizing dominates the result, and a copier who matches entries but not sizing has copied the part of the record that travels least. So even a measurable positive edge is not a transferable profit; it is a property of one wallet's entries, computed on its own past.

Why the leaderboard is not a copy-trade shortlist

Polymarket ranks wallets by realized PnL. That is the natural thing to rank by and it is the wrong filter for the question “will copying this wallet work,” for three reasons that compound.

First, a single concentrated bet can produce a full career of PnL for one wallet. The top of the board is often one resolved event, not a repeatable process. If the trade has already resolved, the edge is already gone; if it has not, the reader is copying one event.

Second, the board only shows survivors. Wallets that took concentrated risks and lost are not on it. The leaderboard selects for the winning tail of a variance process, and a copier who follows only today's winning tail has no reason to assume that tail repeats tomorrow.

Third, across the full 8,656-wallet cohort, calibration and realized PnL correlate at just Spearman +0.148. A high PnL is weak evidence of a calibrated process underneath. The leaderboard is a record of what happened, not a forecast of what a copier would get.

How to check a wallet before you copy

Paste any public 0x address into the free wallet check (a Polymarket profile URL contains the address). The check reads resolved positions from public data and returns the numbers that actually bear on the question. Read three of them in order:

1. The realized entry edge with its 95% interval. If the interval crosses zero, the record cannot be told apart from variance, which is the most common honest outcome at retail sample sizes. “Cannot tell” is a real answer, not a missing one.
2. The concentration share. How much of the record sits in a single event. A record dominated by one resolved event is one bet, not a process, and copying it forward copies nothing the wallet can repeat.
3. The resolved-position count. Below a few dozen resolved positions, almost no edge interval will separate from zero, so the read should be read as “not enough evidence,” not as a green light.

Whatever the interval supports, remember the centerpiece above: the edge describes the entries, not the sizing, and the sizing is where the PnL came from.

What our own scan found

We ran this read on our own published top-50 cohort on 2026-06-09. Of the 35 wallets with 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. A meaningful share of the published cohort was net-negative. In the same scan, a wallet with a very large public PnL that people cite as proof of skill showed an edge of +0.4 probability points, 95% interval [-16.0, +15.6], across 38 resolved positions, an interval far too wide to separate from chance. The full per-wallet table is at the top-50 skill scan.

Does Convexly recommend wallets to copy?

No. We publish the methodology, the coefficients, the cohort snapshot, and the ranking itself, all free and open, and we run a free per-wallet check. We do not publish a recommended-wallet list, and we do not recommend wallets to follow or mirror. Convexly places no orders and is non-execution. The reason is narrow: naming a wallet to copy would require a forward- performance claim, and we have filed forward-skill pre- registrations precisely because that claim is not ours to make yet; both tests are pending and a null is a valid result. The correct primitives are the per-wallet check, the PnL-vs-Edge comparison, and the open V1 methodology. Build your own read; the data is there.