Most "top traders" on Polymarket are farming win rate. I backtested it.

If you've looked at a "top traders" leaderboard on Polymarket and thought about copy-trading the top wallets, I have some bad news: a lot of those flawless 95%+ win-rate records are farmed, not earned. I pulled the on-chain data and backtested it, and the raw leaderboard turns out to be a pretty bad signal to copy. Here's the breakdown, with reproducible notebooks at the end.

The farming trick

It's simple. You only ever buy outcomes that are already basically decided, priced around 95-99c.

Think of a market like "Will X happen?" sitting at 99c the day before it resolves. You put up 99c to make 1c. You win almost every time. Do this across enough near-decided markets and you end up with a 95%+ win rate and a clean, green, very impressive-looking record.

But the edge is basically zero, and more importantly for a copy-trader: it's unfollowable. You can't get filled at 99c at any real size, and you're risking 99 to make 1. The win rate is real in the data and useless in practice.

Raw P&L and win-rate leaderboards reward exactly this behavior. So the wallets that look best are often the ones you least want to copy.

Measuring it properly

A few things matter when you compute this from raw on-chain data, or the numbers come out wrong:

• NegRisk consolidation. Polymarket's multi-outcome markets settle through a NegRisk adapter. If you don't consolidate the legs, you double-count and P&L inflates. (The public data API's cashPnl has this problem - I don't use it as cost-basis truth.)
• FIFO cost basis, per wallet per market. Matching sells against buys in order, not averaging blindly.
• Win rate per resolved market, not per trade. A wallet that makes 40 trades in one market that resolves in its favor didn't win 40 times.

With that in place, I split the leaderboard into two cohorts to copy:

1. Raw top 8 by cumulative P&L (farmers included).
2. Filtered top 8 - same board with farmers and hedgers removed.

Then I backtested copying each one as an equal-weight equity curve across resolved markets.

The result

Honest read: for most of the window the two curves basically track each other. The filtered cohort only pulls ahead near the end, and not by a huge margin (~33M vs ~27M in this run). I'm not going to oversell "filtering wins on P&L" - that gap is small and probably sample-dependent. Both curves are lumpy too; a few big resolutions carry most of the move, which is kind of the point - these are near-decided markets resolving.

What I'd actually stand behind: the raw leaderboard is built on those unfollowable ~99c fills, so copying it is a bad idea regardless of the cumulative number. The win rate on those boards is mostly noise. If you're building a copy-trading strategy, "who has the highest win rate" is close to the worst possible wallet-selection signal.

Small sample (8 per cohort), so take the exact numbers with a grain of salt. The interesting part isn't the dollar figure, it's the mechanism.

Reproduce it

The notebooks are open source (MIT) so you can re-run this on a different set of wallets or a different time window and check whether the filtered edge holds out of sample - genuinely curious if it does:

github.com/orcalayer/polymarket-backtest-cookbook

Three notebooks: follow one wallet's resolved bets as an equity curve, follow a smart-money cohort by category, and the raw-vs-filtered comparison above.

The data comes from OrcaLayer, which is a data layer I build over Polymarket (every trade read straight from the Polygon chain, then classified - farmer filtering, NegRisk-correct win rates, FIFO P&L). The leaderboard endpoint is keyless so the notebooks run free; per-wallet backtests use an API key.

If you go deeper into prediction-market data, I'd love feedback on the methodology - especially the NegRisk and FIFO handling, which is where most of the subtlety lives.