Polymarket all-time profit board (top 50), position by position

The subject is the top 50 of the board, not any one trader.

Subject

Polymarket ranks wallets by all-time profit. A prediction-market record has no capital series, so the wallet check's t does not apply. This audit asks of every resolved position: was the price paid too low for how things turned out?

Method

• Under no skill, a share bought at price p pays 1 with probability p. So edge = outcome − p has mean 0 and variance p(1 − p).
• Positions on the same event (an election, a season) share an outcome, so they are one bet, not many. The headline statistic sums edge per event and uses the spread across events: z = Σ edge / √(Σ over events of (event's summed edge)²).
• Positions are equal-weighted. avgPrice is the average buy price, so a position partly sold before resolution is scored on what was bought.
• The board reaches about 10,050 wallets and reports no total, so the bar is computed for N = 10,050: z ≥ 4.59, a floor (the true N is larger and its bar higher). Update 2026-10-16: the board is far larger than that floor. Filtering it by username returns a wallet ranked 450,853, so N is at least 450,853 and the bar at least z ≥ 5.33. No finding changes: the highest z among the complete records is 2.40, below either bar.
• Data: data-api.polymarket.com (v1/leaderboard, closed-positions, positions), public, no key, fetched 2026-10-15. Script: tools/audit_polymarket.py.

Finding

Of the 38 top-50 wallets whose full history could be read, none clears the bar; 8 reach z ≥ 2. Twelve of the most active wallets have more resolved positions than the script reads (3,000 per list) and are left out rather than scored on part of their record.

• Half of the 38 bet on 8 or fewer distinct events. Much of the top of the board is a few large, related bets: the #1 wallet's 22 positions are all on the 2024 US election (z = 2.05).
• With so few events, z is itself imprecise; "cannot be told apart from luck" is the safe reading, not "no skill".

The trap this audit fell into first

Polymarket lists sold or redeemed positions under closed-positions. A losing share is worth nothing, so most are never redeemed: they stay under positions, marked redeemable with a final price of 0. Scored on closed positions alone, 11 of the top 50 cleared the bar, with z as high as 18.3. Adding the resolved, unredeemed positions:

Any wallet tracker that reads only the closed list will show the same mirage: a record of winners, because the losers were never cashed in.

What it would take

At even odds, z = edge × √n / 0.5, so a real 5-point edge per bet needs about 2,100 independent events to reach z = 4.59. The complete record with the most events here (0x5966…E507: 168 events, 7.4 points of edge per position) reaches 2.2. The top of the profit board mostly has a handful of events.

Recheck: quarterly. A per-event calibration test (Brier against the price paid) is the natural next step; Convexly runs that kind of analysis for Polymarket.

Update, 2026-10-16

• Board size. The leaderboard's username filter (/v1/leaderboard?userName=…) returned a wallet ranked 450,853, so the board ranks at least that many wallets, not the 10,050 that paging by offset reaches. The bar for the best of 450,853 is z ≥ 5.33. Every count above holds under either bar.
• Check any wallet. The same test now runs in the browser at /polymarket/, on any wallet or username. For the wallets above it reproduces this audit's figures (for example 1.08 for DrPufferfish, 18.31 on closed positions only).