Polymarket copy trading: what the graded data shows

Does copying a wallet five minutes late work? We tested it

In tests pre-registered and published on our research page, every scored fill on resolved Polymarket markets was re-priced after a five-minute delay. Each wallet’s record in each category was split into earlier and later events, and the wallet-categories whose earlier record passed a 95% test were compared with the rest on later events, against the price curve and before fees.

The first test priced a follower at the first trade on the same outcome at least five minutes after the wallet’s fill, whichever side started it. In the deciding run of 2026-09-25, the 13,422 wallet-categories that passed averaged +0.12 cents a share on later events (95% interval +0.04 to +0.21), and the other 76,960 averaged −0.83. A first run on 2026-09-24, on data that filed most tennis and esports markets under Sports, had the passing group at +0.08 (−0.01 to +0.17), short of the pre-registered rule. It stays published beside the rerun, which cleared it.

The second test, run on 2026-09-28, priced each fill at the taker price, the first purchase at the asking price at least five minutes later (for a sale, the first sale at the bid). That trade cost 0.33 cents a share more than the first test’s on average, counting any price move in between. At the taker price, the 13,290 wallet-categories whose earlier records passed at the first test’s price came out at −0.21 cents a share on later events (95% interval −0.30 to −0.12), below the price curve, before fees. Counting the fee recorded on those trades lowers that to −0.42 or −0.60, depending on how purchases are charged. The rest came out at −1.26.

In both tests the earlier record sorted later results, and the wallet-categories that passed did better than the rest by about a cent a share. At the taker price, even the group that passed did not beat the price curve. The figures are measured against the price curve and before fees, so they do not show whether copying made money. The rest of this page covers how to read a record. The check below grades a record at the prices the wallet itself paid, before any follower’s delay.

The selection problem is the whole problem

Every copy-trading pitch begins with a leaderboard, and most leaderboards rank by profit. Our statistics page documents what that metric hides. Of the 10 largest all-time P&L wallets on Polymarket, 4 clear OVERROUND’s public evidence floor. Among those, the median skill-adjusted rank is outside the top 1,000 (re-verified 2026-09-26). The Polymarket leaderboard guide walks that official profit board in detail. Ranking by profit and ranking by demonstrated accuracy answer different questions.

Win rates mislead in a specific, measurable way

Across the 385,301 graded wallets with 20 to 99 events in which they held a single market on one side through resolution, the median win rate is 72.0% while the median price-implied win rate is 71.3%, so for a typical record the win rate mostly restates the prices paid. At 95% confidence, 1.8% of such records are ahead of their prices and 3.6% are behind them (full distribution here, as of 2026-09-27). The folk criteria that circulate in copy-trading guides, a 60% win rate or a green month, sit comfortably inside what chance plus favorable prices produces. The 2026 SSRN study “Who Wins and Who Loses in Prediction Markets” found from account-level data that a small minority of participants drives accuracy.

What a record has to survive before it means anything

• Price-relative grading. Wins count against the odds actually paid; favorite-buying stops looking like skill.
• Event independence. Each set of mutually exclusive outcomes counts once at most. Positions on several candidates in one race score as one result on the leaderboard, and the free checker leaves that race out.
• Evidence depth. The share of records that are distinguishable from their own prices and ahead of them climbs from 1.8% at 20 to 99 events to 42.9% at 1,000 or more (95% confidence, as of 2026-09-27); shallow records are unratable, however green.
• A graded history of the recommender. Whoever surfaces wallets to you should publish what their surfacing returned, losses included. That is the standard this site holds itself to.

The parts no leaderboard shows

Two structural gaps sit between a copied wallet’s record and a copier’s result. Entry lag: the price a copier gets is the price after the wallet’s own move, so part of any edge is spent before the mirror fills. And style risk: some profitable records are market-making or spread-capture patterns that produce constant small fills a copier cannot replicate at their size. Both are reasons the graded, price-relative record of the wallet matters more than its dollar total, and why this site’s alerts carry the entry price and the wallet’s graded context rather than an instruction.