Which Polymarket Leaderboard Should You Trust?

None of them, if the question is “who should I follow”. A leaderboard ranks outcomes; skill is a separate question, and the two disagree in the data. On Convexly's own published top-50 Edge Score cohort (refreshed daily; snapshot 2026-07-02), 11 of 50 wallets are net-negative on realized PnL. And when we ran a formal skill test on the cohort as it stood on 2026-06-09, none of the 35 readable wallets survived a multiple-comparisons correction at q=0.10. The honest use of any board is as a starting list for wallets to check, not as a verdict.

PnL boards and skill reads are different instruments

A PnL leaderboard answers “who made the most money”. Dollar PnL is dominated by position size, single-event concentration, and timing, none of which require a repeatable forecasting process. A wallet can top a PnL board on one large resolved event and never show a readable edge anywhere else in its record. A skill read answers a narrower question: did this wallet's entries, taken together, beat the prices it paid by more than chance would produce? Those are different measurements, and ranking by one tells you little about the other.

Convexly publishes the disagreement rather than hiding it. The PnL vs Edge Score comparison puts the top 20 wallets by realized PnL next to the top 20 by Edge Score so you can see how little the two lists overlap. The wallet rankings board ranks by Edge Score V3b, keeps the realized-PnL column visible, and gates the rendered list to positive-PnL wallets while the underlying cohort artifact stays published in full, losers included.

What our own board looks like under its own test

The sharpest evidence against leaderboard-as-verdict is our scan of our own cohort. On 2026-06-09 we computed the realized entry edge, with bootstrap 95% intervals, for every wallet in the published top-50 Edge Score cohort. 35 of 50 had enough resolved positions to read. 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 that lone hit did not survive the Benjamini-Hochberg correction at q=0.10. Its net PnL was negative. The per-wallet table, method, and concentration reads are public at the top-50 skill scan.

Read that plainly: on a board we built, using a test we chose, zero wallets cleared the corrected skill screen. A leaderboard operator who tells you its top rows are proof of repeatable skill is telling you something its own data almost certainly cannot support.

Where Edge Score fits, stated honestly

Edge Score V3b is a descriptive composite of behavior (calibration posture, conviction, discipline) fitted on a frozen 8,656-wallet cohort. In its one out-of-sample forward test it held a +0.11 rank correlation (95% CI 0.05 to 0.18) with forward PnL, below the +0.30 bar set in advance. So the board presents it as a profile of the resolved record, not as a forecast, and the skill question is delegated to the per-wallet realized-edge read with its interval.

A five-point checklist for any board

• Ranking basis. PnL, volume, or a composite? If a composite, is the method published and frozen, or adjustable after the fact?
• Inclusion floor. Convexly's cohort requires at least 20 resolved positions per wallet. A board without a floor ranks noise.
• Losers visible. Does the realized-PnL column stay visible even when it cuts against the ranking? Ours shows 11 of 50 net-negative as of 2026-07-02.
• Refresh date. A board without a generation timestamp may be ranking a market that no longer exists.
• Uncertainty. Any skill claim needs an interval or a corrected test in the same breath. A rank number alone is not one.

The practical workflow: use any leaderboard, including ours, as a list of candidates, then run each candidate wallet through the free check to see whether its record separates from chance. Most will read “cannot tell”, and that is the honest answer.