Copy Score performance
Does any of this actually predict the future?
A score that explains the past is worthless. The only test that counts is whether it works on data that did not exist when the score was calculated. We scored 1,796 traders using only information available at a fixed cut-off, sealed those scores, and watched what those traders did over the months that followed. A second test asks a harder question: does that survive the trip to our copies?
The sealed test
1,796 traders, scored before the clock started
Each wallet had at least 50 settled positions and $2,000 of cost basis in both windows. Trailing window: 17 March 2026 to 15 June 2026. Forward window: 15 June 2026 onward. The score used in the test was rebuilt from only the trailing window.
Traders tested
1,796
At least 50 settled positions and $2,000 of cost basis before and after the cut-off.
Skill persists
+0.26
Rank correlation between a trader's past and future returns. Real, and clear of zero.
The score predicts
+0.21
Rank correlation between a point-in-time Copy Score and what the trader did next.
Top tenth returned
+14.5%
Median return per settled position over the following quarter, against −0.10% for the bottom tenth.
What +0.26 means, plainly. It is a score from −1 to +1. At +1, the best trader last quarter is always the best next quarter. At 0, the past tells you nothing. At +0.26 there is a real, dependable tendency for better traders to keep being better, with plenty of individual exceptions. In markets, a durable +0.26 on out-of-sample data is a signal worth building on. Anyone claiming much higher is usually measuring the past twice.
Past results, then what they did next
Sort traders into groups by past results, then watch
Group 1 is the weakest past performers, group 10 the strongest. Measured on settled positions after the cut-off. We publish the four groups we measured, not a filled-in ladder.
Traders grouped by past results, what they returned next
Both signals
The score adds something beyond who did well before
Past returns alone are a decent guide. The question is whether Copy Score improves on them. We split traders by both at once. Median return per settled position over the following quarter.
| Trader group | How many | What they did next |
|---|---|---|
| Strong results + strong Copy Score | 695 | +3.71% |
| Strong results, weak Copy Score | 203 | +1.30% |
| Weak results, strong Copy Score | 203 | +0.13% |
| Weak results, weak Copy Score | 695 | +0.35% |
Among traders who already performed well, the ones Copy Score also rates highly went on to do nearly three times better than those it rates poorly. The score is not a restatement of the leaderboard. It separates traders the leaderboard treats as identical.
Equally telling is the third row. A strong Copy Score on a trader with weak results is worth almost nothing. Both signals have to agree. That is a constraint we hold ourselves to, not a caveat buried in a footnote.
The fee curve
The same trade costs ten times as much at 5¢ as it does at 90¢
Prediction-market fees are not a flat percentage. They are largest where the odds are closest to even and smallest near certainty. Slippage on a cheap share is a huge fraction of the money you put in. Expressed as a share of that money, the cost of copying looks like this.
18.2% of your money, at 5¢. 1.7% at 90¢. A trader who specialises in cheap long-shots is dramatically more expensive to follow than one who trades favorites, even if their headline returns are identical. Copy Score prices each trader against the curve they actually trade on.
How we validate
The things we do to stop fooling ourselves
Backtesting is easy to do badly. Nearly every impressive result in trading is an artefact of testing something on the data that produced it.
Point-in-time scoring
A trader's score is rebuilt from only what was known at the time. Using today's score to pick traders for a historical test would be choosing winners with hindsight, and it produces spectacular results that mean nothing.
Measured friction, not assumed
The price slippage in the simulator comes from more than a hundred thousand of our own real fills. A resting order only fills when the market comes to it, so favourable prices are rarer than unfavourable ones.
Benchmarked against doing nothing clever
Every result is compared with picking at random, and with simply trusting the market price. A strategy that cannot beat the market's own forecast is not a strategy, however good its P&L chart looks.
Overfitting is scored, not hoped away
The more variants you test, the more likely the best one is luck. We compute the probability that our selection is fooling us, and reject results that fail the bar.
The right unit of uncertainty
Ten trades on one market share one outcome. They are one observation, not ten. Treating them as independent makes confidence intervals roughly five times too narrow. Ours cluster on the market.
We try to break our own findings
Promising results get a randomisation test: reshuffle the traders thousands of times and ask how often chance produces something as good. If it does so often, the finding is noise and we discard it.
This is not decoration. We use it.
An earlier version of this analysis found what looked like a large, clean advantage for high-scoring traders, holding up across two independent variations. It failed the randomisation test: four in ten random reshuffles produced a result just as strong, and a single unusual trader was carrying the whole thing. We retracted it. The findings on this page are the ones that survived that process.
Retraction
Why the old page came down
The previous version of this URL published win rates and P&L by Copy Score range from a study of the trader's own resolved trades. It claimed that high-scoring trades won more often. That study was a correlation on the leader's tape. It was not a test of our copies.
When we scored our Auto Copy fills with that old grade, the trades it liked most lost more than an unfiltered book, not less. Publishing those tables would be repeating a claim our own fills refuted. We took them down. If you cited a win rate by score range, or a letter grade, from this page or from /copy-score, that figure described the retired score. Do not use it.
What our copies showed
Skill persists. The trip to your account is the hard part.
The sealed test above is about the trader. This section is about our Auto Copy fills, with each trader's Copy Score built only from history before the fill.
Did the old per-trade grade pick winners on our book?
No. Filtering to the trades it called better than the market price made copier results worse.
Does the new Copy Score work as a skip?
Yes, narrowly. Skipping traders who are hard to copy beat copying everyone. Treating the top band as a ranked "best traders" list did not turn the book profitable. Even the clearer band lost money dollar-weighted on this tape. That is why the product is a warning, not a picker.
Does category matter?
Yes, for the same person. Copying them in a market type they already had a record in beat copying them in a market type they did not. That supports the category mark. It does not support a public "top NFL by Copy Score" board.
We are not going to publish the fill-level scoreboard. The point of this page is the decision those tests produced, not a recipe for rebuilding the measure. Full explainer: How Copy Score works.
Limits
What we will not claim
- That a high Copy Score means our copies will profit.
- That Copy Score is a ranking of the best traders on Polymarket.
- That the old 0 to 100 letter-grade study still applies.
- Win rate by Copy Score band on the leader's tape, dressed up as copier performance.
Discover can pre-filter to people who are not hard to copy, then still rank by P&L. Profiles show the number and the qualifier. Auto Copy uses Copy Score to filter the traders, not which click to skip. None of those surfaces sort the internet by Copy Score and call it a leaderboard. The states that read as a “no” are Caution and Thin history.
FAQ
Common questions
Does Copy Score actually predict the future?
On 1,796 traders, scored only from information available before a fixed cut-off, the score's rank correlation with what they did next was +0.21. Among traders who already had strong results, those with a strong Copy Score went on to +3.71% per settled position, against +1.30% for those with a weak score. That is a statement about the trader, not a promise our copies will match it.
Does Copy Score work on our copies?
As a warning, on our Auto Copy fills: skipping traders who are hard to copy beat copying everyone. As a "best traders" ranking, no. The top band was not a profitable picker on that tape. We ship the warning, not the ranking.
Why did you take down the backtest tables?
They measured the retired per-trade score against the trader's own outcomes. On our copies, that score failed. Leaving the tables up would keep a claim we no longer stand behind.
Will you publish a new win-rate table?
Not a table that invites people to rank by Copy Score. The tables on this page are out-of-sample trader returns and a 2-by-2 split, not a recipe for sorting Discover.
Explore traders
Use Copy Score as a veto, then copy people in the categories they already have a record in. Require both a strong record and a strong score.
Risk
Copy trading does not guarantee profits
A trader who clears copy costs on their own tape can still lose in the next month, and our fills can differ from theirs. Copy Score is a decision-support indicator, not financial advice.
Even after skipping hard-to-copy traders, our Auto Copy book on the test window was not a profitable guarantee. Copy Score reduces a class of bad follows. It does not make copy trading safe.
Figures on this page come from measurements on settled positions and real executed fills. Out-of-sample tests cover 1,796 traders across a fixed cut-off. Past performance is not a guarantee of future results. Copy trading carries risk, including the loss of the amount you commit.
Polycopy does not tell you to place a trade.