Why is a profitable record not proof of skill?

Because on a few dozen bets, chance alone produces plenty of profitable records. If 1,000 traders bet at fair prices with no skill at all, roughly 500 finish ahead, and a handful finish far ahead. Ranking by profit finds the luckiest records, not the most skilled ones. Reading a record as skill takes a measure that nets out the prices paid, an uncertainty interval, and a correction for how many records were screened.

Did Convexly test this on its own leaderboard?

Yes. In the frozen 2026-06-09 scan of Convexly's own published top-50 board, 35 wallets had enough resolved positions to test. After the multiple-comparisons correction across all 35, zero cleared the corrected bar. That is no FDR-corrected evidence of skill on our own descriptive board, published as-is at /research/top50-skill-scan.

Does any Polymarket wallet show measurable skill?

Platform-wide, yes: of 3,871 Polymarket wallets with at least 30 resolved discretionary positions, 178 clear a false-discovery-corrected test at q = 0.10 for positive realized edge over their entry prices, with expected false discoveries among those 178 of about 17.8. That result is in-sample: it reads the past record and is not a forecast that those wallets keep winning.

Why profit is not proof

The answer first

A profitable record is not proof of skill. On a few dozen resolved bets, chance alone produces plenty of winners, and the biggest winner in any crowd is usually the luckiest, not the sharpest. Before you trust a record, you need three things: a measure that nets out the prices paid, an uncertainty interval around it, and a correction for how many records you screened to find it. The rest of this curriculum builds those three things one lesson at a time.

The intuition, with one worked example

Imagine 1,000 wallets that bet with no skill at all, every entry at a fair price. After 40 bets each, about half are in profit, purely by chance. A few are far ahead. Sort them by profit and the top of the list looks like genius. It is a coin-flipping contest read as a talent show.

We ran this logic against ourselves. Our own published top-50 wallet board is ranked by a descriptive composite, and its rows are full of profitable records. In the frozen 2026-06-09 scan, 35 of those wallets had enough resolved positions to test. After the correction for testing 35 records at once, zero cleared the corrected bar. That is no FDR-corrected evidence of skill on our own board, and we published the scan as-is at /research/top50-skill-scan. Platform-wide the same test does find something: of 3,871 wallets with at least 30 resolved discretionary positions, 178 clear it, about one in twenty-two, in sample. Profit is everywhere. Readable skill is rare.

The actual method

Why profit specifically is so hard to read on prediction markets:

Fat tails. Realized profit is dominated by a few large positions. One win can carry an entire record, so the total says little about the typical decision.
Small samples. Most wallets have well under 100 resolved positions. At that size, chance swamps skill in the raw totals.
Selection effects. You only hear about the winners. Screening thousands of records and showcasing the best one manufactures heroes out of noise.

The fix is not a better profit number. It is a different measurement: the realized edge over entry prices (lesson 2), an interval on every estimate (lesson 3), a four-state verdict that demotes unreadable records (lesson 4), and a false-discovery correction whenever many records are tested at once (lesson 9).

Where you see this on the site

The free wallet analyzer is this lesson in tool form: paste a wallet and it answers with an edge, an interval, and a verdict instead of a profit number. The paper Profit Is Not Proof runs the full argument on our own leaderboard, and /learn/luck-share covers the companion term.

What this does NOT mean

It does not mean profitable wallets are unskilled. Zero cleared on our board is not evidence of no skill; several of those records carry wide intervals that include large positive values, so the honest read is that the records cannot be separated from chance either way. It also does not mean skill is absent from the platform: 178 of 3,871 tested wallets clear the corrected test, in sample. What it means is narrower and more useful. A profit figure, on its own, tells you almost nothing. Ask for the receipt behind it.

Next lesson: entry price and realized edge

Convexly publishes new methodology research roughly every 6-8 weeks plus the /learn series on a rolling cadence. Get the next paper in your inbox when it ships: