Skill or luck? What $67B of Polymarket trades says about copying winners

Most of the money made on Polymarket is won by a tiny minority — and they win in a way a copier structurally can't reproduce. A 2026 working paper that analysed roughly $67 billion of Polymarket trading found that among users who finished in profit, the top 1% captured about 76.5% of all the gains, and that the winners systematically made markets — resting limit orders that resolved in their favour — while the losers took markets, hitting those orders at market price. A copy-bot is, by construction, a market-taker that arrives after the signal. That puts it on the losing side of the single clearest divide in the data. Copying can still pay — but only for the narrow kind of edge that survives being copied, and in our own scored set that's about 1.3% of active wallets (July 2026).

What a $67-billion study of Polymarket actually found

The paper reports three things that matter to anyone thinking about copying a wallet. First, profits are hyper-concentrated: of the users who finished ahead, the top 1% took roughly 76.5% of the total, leaving the other ~23.5% to be split among everyone else in the black. Second, winners and losers separate cleanly by how they trade — the consistent winners provide liquidity with limit orders that resolve favourably to the realised outcome, while the consistent losers take liquidity with market orders. Third, month-to-month performance is only "modestly persistent," and the authors caution that even that may reflect selection — survivors of variance — rather than durable skill.

One more finding is worth stating plainly, because it cuts against the easy narrative: the authors conclude that insider trading is unlikely to explain the performance of the largest winners. The biggest accounts aren't mostly tipped-off insiders — they're patient liquidity providers. (Insider cases on Polymarket are real but rare, and they're their own kind of uncopyable edge.) A liquidity-provision edge is a harder thing to copy, not an easier one.

Why this is bad news specifically for copiers

A copier is a taker, and the study says takers lose. When a copy-bot mirrors a wallet, it watches for a fill that already happened and then chases it — paying the spread, paying taker fees, and arriving milliseconds to seconds late into a book that the original trade just moved. The wallet you're copying earned its edge by posting a resting bid at a good price; you inherit none of that, because you buy at market after the price has adjusted. The very behaviour that made the leader a winner — providing liquidity — is the behaviour your bot can't replicate.

This is the same gap we've measured from the cost side: a wallet's headline return overstates what a copier keeps by 20–40% once fees, slippage, and latency are paid. The study explains why the gap is structural rather than incidental — and why it is widest for exactly the limit-order-driven wallets that look best on a leaderboard. Sharp traders also actively hide their entries from copy-bots, which widens it further.

Skill or luck: does last month's winner stay good?

Mostly you can't tell from the record alone — which is the whole problem. "Modestly persistent," with a selection caveat, means a wallet that crushed last month has only a weak, noisy tendency to repeat it, and part of that tendency is just the survivors of a coin-flip looking skilled in the rear-view mirror. A high win rate doesn't rescue you: on a prediction market the break-even win rate is simply the price you paid, so a gaudy hit-rate can sit on top of a negative edge. The only defence is to judge consistency the way it is actually earned — across many resolved, independent markets, not one lucky event-week — which is why consistency is a scored factor, not a vibe.

What our own scoring shows

Run the actively-traded population through the tests a copier should actually apply and the survivors are scarce. In the June 2026 snapshot charted below, of 1,297 active wallets under coverage, 995 had enough recent history to judge, 606 were farming-clean, 214 cleared a realistic post-fee edge above 1%, and only 28 — about 2.2% — passed every test at once (the July re-score tightened that to 1.3%). That bottom number is the practical version of the study's point: the set of wallets whose edge is real, clean, and the transferable kind is tiny, and none of them stand out on a raw-profit board.

How to copy anyway: screen for the transferable kind of edge

Copying isn't hopeless; it's selective. The wallets worth copying are the ones whose edge doesn't depend on being first to a resting order — repeatable, post-fee, and located in a category your execution can actually reach. Read a candidate the way you'd read it as a quant: is the return risk-adjusted or one oversized bet; is the edge still there after fees; is it spread across enough independent markets to be skill rather than a streak? Then run the full vetting checklist before any capital moves. The per-wallet Copy Verdict packages those checks — post-fee edge, risk-adjusted return, drawdown, consistency, and the farming veto — into one read, and the published base rates show how rare a clean pass is.

The honest takeaway: "find a winner and copy them" fails because most winning is concentration and liquidity provision, and you can copy neither. The job is to find the few whose edge survives being copied — and to verify it before you fund it.

CopyGrade is analysis-only — it never executes trades or holds funds, and a Copy Score is a documented research opinion, not a statement of fact about any trader or financial advice.