How to backtest copying a Polymarket wallet (and what a backtest can't prove)

Backtesting a Polymarket copy strategy means replaying a wallet's real, historical fills as if you had been copying them — with your capital, and with the costs a copier actually pays layered on: taker fees on both legs, slippage, and the latency between the leader's fill and yours. Done honestly, it answers the one question a leaderboard never can: not "did this trader make money?" but "would copying this trader have made me money, after everything?" Done dishonestly — zero-cost, zero-latency, on a wallet you picked because it just won — it manufactures confidence, which is worse than having none. This guide covers how to run the honest version and the five traps that make backtests lie.

The tool: replaying real fills

CopyGrade's Copy Simulator does the mechanical half for you: pick any wallet in the index and it replays that wallet's actual historical fills against the capital you specify, discounted for Polymarket fees, slippage, and copy-bot latency — each exposed as a knob you control rather than an assumption you inherit. A compare mode runs two wallets side by side under identical settings, which is the honest way to break a tie between finalists (the finalist framework itself is in choosing between two traders). It's free with an account, and every run is URL-encoded so a result can be reloaded or shared rather than remembered.

Whatever tool you use, the inputs that matter are the same: real fills (not a synthetic equity curve), fees on both legs (the 2026 schedules are real money now), and latency you set pessimistically — a copy-bot follows at a delay, and on fast markets that delay is where a paper edge dies.

Reading the result

Two numbers matter more than the headline return:

• Net-of-costs return vs the wallet's own. The gap between what the wallet made and what the replay says you'd have made is the copy haircut, made concrete. A wallet whose edge survives the haircut is rare — the gap is the entire thesis of headline vs realistic returns.
• The drawdown path, not the endpoint. A replay that ends +20% but spent six weeks at −35% is a strategy you would almost certainly have abandoned at the bottom. Size from the drawdown, and treat the replay's worst stretch as a preview of the night you'll want to quit.

The five traps

1. Hindsight selection. You are backtesting this wallet because it already won — the selection itself is look-ahead bias. A backtest can never rescue a wallet from this; only forward evidence can. Pick candidates by vetting criteria, then backtest, never the reverse.
2. Short samples. Replaying a hot month proves the month was hot. Demand months of fills across enough resolved markets that variance has had a chance to hurt — and read the record's observation age before believing any of it.
3. Optimistic frictions. Zero latency and zero slippage are not "neutral defaults"; they're the most flattering possible assumption. Set latency to what your actual bot achieves at worst, not best — if you don't know, that's a finding too.
4. Open positions don't exist yet. A replay scores what resolved. A wallet heavy in long-dated futures has most of its eventual P&L still invisible — the replay is silent about exactly the exposure you'd be copying. (This is why pre-season sports records mislead.)
5. The past isn't a forecast. The deepest limit: a perfect replay proves the edge existed, not that it persists. We publish our own version of this humility — CopyGrade's forward validation of whether high scores predict copier returns has so far returned a null result. A backtest is a filter that removes bad candidates; it is not a machine that produces good ones.

Where the backtest fits

The loop, in order: shortlist from the graded boards or the live shortlist → vet → backtest the survivors in the simulator → size from the drawdown → write the exit rules → only then, a bot. The backtest also earns its keep after entry: rerunning a wallet you're copying and comparing the replay against your live results tells you whether underperformance is the wallet's edge fading or your own execution — two problems with two different fixes.

The simulator replays public trade history under stated assumptions; its output is an estimate, not a promise, and past performance — replayed or otherwise — doesn't guarantee anything forward. CopyGrade is analysis-only and never executes trades or holds funds. Not financial advice.