We re-scored 11,721 Polymarket wallets. 155 are worth copying. The leaderboard has none.
This is the fourth edition of our recurring report on the state of Polymarket copy trading, and the first whose trend line we can print. We re-scored the full universe under CopyGrade coverage — 11,721 actively-traded wallets, up from 2,017 in July. 35.0% carry a farming-risk flag. The median wallet keeps −1.4% after modelled trading costs. 155 wallets — 1.3% — pass every test a copier should apply. And of the 191 wallets on Polymarket's public leaderboards, none does.
The blended flag rate fell from 47.8% to 35.0%. Read that as a cleaner market and you will be wrong, and most of this edition exists to show why: the 2,019 wallets we already covered in July are flagged at exactly 47.8% again. Every point of the drop is who we added.
A note on the sample: 11,721 wallets that trade actively enough to be scored, snapshotted on 30 September 2026 — every score computed within the preceding seven days, 99% within the preceding 24 hours — each graded from its public Polymarket trade history by the same CopyGrade Score model as every prior edition. The weights and thresholds are unchanged since the 30-day validation we published as a null result; one gate was added on 2 September and is disclosed below.
The headline numbers, and the first honest trend line
| July 2026 (2,017 wallets) | October 2026 (11,721 wallets) | The July wallets, re-scored (2,019) | |
|---|---|---|---|
| Median CopyGrade Score | 42 | 48 | 42 |
| Farming-flagged | 47.8% | 35.0% | 47.8% |
| — high-confidence (severe) | 37.3% | 24.2% | 36.8% |
| Negative fee-adjusted edge | 74.6% | 77.0% | 76.1% |
| Median fee-adjusted edge | −4.8% | −1.4% | −5.6% |
| Worth copying (all tests at once) | 23 (1.1%) | 155 (1.3%) | 13 (0.6%) |
The right-hand column is the wallets that were under coverage at the July snapshot, re-scored today. (2,019 by discovery date against the 2,017 the July report scored: two had been discovered but not yet graded when that snapshot was taken.) The July edition promised this column. Editions one to three were computed on a trade tape that stored duplicate copies of fills, which inflated the repetition the farming detectors read, so their flag rates were not comparable with anything computed after the 25 July correction. Both the July figures shown here and today's are computed end to end on the corrected tape. This is the first cross-edition comparison we consider honest.
What it says is that on the same wallets, nothing moved. Flag rate 47.8% to 47.8%. Severe 37.3% to 36.8%. Median score 42 to 42. The median wallet's edge slipped from −4.8% to −5.6%, and the copyable count fell from 23 to 13 — more on both below. A quarter is a long time in a prediction market, and the incumbents' farming profile did not shift by a point.
Where the 12.8-point drop came from: who we added
Coverage grew from 2,017 to 11,721 wallets in one quarter. That is not the market growing; it is our discovery widening. Since July the sync has swept the participants of the most active markets, where before it mostly followed the leaderboards, and that changes who is in the denominator.
| Cohort | Wallets | Farming-flagged | — severe | Median fee-adj. edge | Median score | Worth copying |
|---|---|---|---|---|---|---|
| Already under coverage in July | 2,019 | 47.8% | 36.8% | −5.6% | 42 | 13 |
| Newly covered since | 9,702 | 32.3% | 21.6% | −1.0% | 48 | 142 |
The new cohort is fifteen points cleaner and its median wallet loses far less. Both are composition. By the route a wallet arrived:
| Where the new wallets came from | Wallets | Farming-flagged | Median fee-adj. edge | Median trades, last 90 days |
|---|---|---|---|---|
| Participants in the most active markets | 8,824 | 29.0% | −0.7% | 14 |
| Polymarket's profit and volume leaderboards | 808 | 66.3% | −20.9% | 242 |
| Graded on request by a visitor | 69 | 65.2% | −7.2% | 376 |
One further wallet arrived through a board-entry alert and is left out of the table. The 808 new leaderboard names are dirtier than any cohort we measured in July. The 8,824 market participants are ordinary traders: a median of 14 trades in the last 90 days, a median record 47 days long. They are cleaner — and thinner, which turns out to matter more.
Clean, or just quiet?
A farming detector reads patterns in a trade history. A wallet with 14 trades gives it very little to read. So we split every wallet by whether it clears the activity gate, 20 or more trades in the last 90 days:
| Active (20+ trades, last 90 days) | Thin (under 20) | |
|---|---|---|
| Already under coverage in July | 978 wallets · 67.1% flagged | 1,041 wallets · 29.7% |
| Newly covered since | 4,658 wallets · 52.7% flagged | 5,044 wallets · 13.5% |
| Everyone | 5,636 wallets · 55.2% flagged | 6,085 wallets · 16.3% |
Half the new cohort is thin, and the thin half is flagged at 13.5%. Among wallets active enough to judge at all, the flag rate is 55.2% — and 52.7% even inside the new, "cleaner" cohort. Some of the headline drop is innocence; a good part of it is silence. The active-set figure we publish is 35.0% and we say in the same breath which denominator it is over. But the number a copier should carry is not 35. It is 55: among wallets with enough recent history to be worth copying at all, more than half carry a farming-risk flag.
We ruled out our own inputs first, because last edition's headline turned out to be our own bug. The farming detectors have not been touched since 25 July — no commit to that module — and every wallet in every table above was re-scored on the same code within the week. One scoring change shipped in the interval: a 30-day floor on the length of the record, which caps a wallet with fewer than 30 days between its first and last fill at the caution band. It changes scores, not flags, and the incumbents' records are a median 115 days long, so the right-hand column's 47.8% owes it nothing.
The vetting funnel: 11,721 in, 155 out
Each row is how many wallets pass that single test; the last row is how many pass all four at once.
| Test | Wallets passing | Share |
|---|---|---|
| Scored | 11,721 | 100% |
| Enough recent history to judge (20+ trades, last 90 days) | 5,636 | 48% |
| Farming-clean | 7,619 | 65% |
| Fee-adjusted edge above 1% | 979 | 8.4% |
| CopyGrade Score ≥ 75 | 159 | 1.4% |
| All four simultaneously — worth copying | 155 | 1.3% |
Two things changed in the shape of this table. The farming-clean row now sits above the history row — the thin wallets again: a record too short to judge is also a record too short to flag. And in July the last two rows were the same 23 wallets, which we said then was a property of that population rather than a shortcut in the method. This quarter they part: four wallets score 75 or better while failing one of the other gates. Same four tests, applied separately, as always.
278 wallets (2.4%) are clean, active and genuinely profitable after costs before the score bar is applied; the bar takes it to 155. That is a longer shortlist than July's 23, from a universe nearly six times larger. The rate barely moved: 1.1% to 1.3%.
The headline-edge trap, restated: 1,749 wallets currently show a positive headline edge, and 689 of them — two in five — carry a farming-risk flag. In July it was three in five. The difference is composition once more: among the July incumbents the overlap is still 58% (190 of 327 positive-headline records); among the new cohort it is 35%.
What happened to July's 23
The 23 wallets that passed every test in July are all still under coverage, so we can say what a shortlist looks like one quarter on.
| July's 23 copyable wallets, today | Wallets |
|---|---|
| Still pass all four tests | 6 |
| Fell to the 60–74 band | 17 |
| Fell below 60 | 0 |
| — of the 17: now under 20 trades in 90 days | 8 |
| — of the 17: still active and clean, edge no longer above 1% | 5 |
| — of the 17: still clear every gate except the score bar | 3 |
| — of the 17: picked up a farming-risk flag | 1 |
Six of 23 survived the quarter. Not one blew up: none fell below 60, and one acquired a flag. The rest went quiet — eight now trade too little to judge — or their edge thinned to nothing after costs. The median of the 23 today is a score of 64 and a fee-adjusted edge of +2%. That is the copier's real hazard in numbers: a good target does not usually turn bad, it turns stale, and a bot keeps mirroring it either way.
The leaderboard cut: zero
The wallets on Polymarket's public profit and volume boards — 191 under coverage at this snapshot — remain the worst slice we measure, and this quarter none of them clears the bar:
| Full set (11,721) | Current leaderboard (191) | July's leaderboard (189) | |
|---|---|---|---|
| Median CopyGrade Score | 48 | 37 | 37 |
| Farming-flagged | 35.0% | 61.3% | 59.8% |
| — severe | 24.2% | 44.5% | 46.0% |
| Median fee-adjusted edge | −1.4% | −18.9% | −9.9% |
| Worth copying | 155 (1.3%) | 0 | 1 (0.5%) |
The roster churns as the windows roll, so the July column is a level to set beside today's, not a baseline for a trend. Read as levels, the board is where it was: a median score of 37, six in ten flagged, a median edge after costs that has now reached nearly −19%, and the one wallet that survived in July has not been replaced. The boards rank raw profit, raw profit rewards exactly the behaviours the model penalises, and this quarter not one board name passes every check. If your shortlist is the leaderboard, your base rate is zero in 191 — see the graded leaderboard for the rank-versus-reality gap on today's board.
What this means if you copy
The base rates are the strategy, and this quarter they come with a denominator lesson attached.
- Vet before you automate. 155 of 11,721 pass; the pre-automation checklist exists because the other 98.7% fail at least one test.
- Read edge net of costs. 77% of active wallets are negative before you add slippage and a bot fee. The headline return is a ceiling you will never reach.
- Treat a thin clean record as unproven, not clean. Thin wallets are flagged at 16%; active ones at 55%. Most of the gap is that the detectors had nothing to read. Demand the history first, then the behavioural signatures.
- Re-check what you already copy. Seventeen of July's 23 fell out of the copyable band inside a quarter, almost all by going quiet or losing their edge rather than by doing anything visible. Alerts exist so that reaches you before your bot keeps mirroring a stale target.
- Do not shop the leaderboard. Zero of 191.
How we got these numbers
Point-in-time snapshot, 30 September 2026: 11,721 actively-traded wallets under CopyGrade coverage, each re-scored within the preceding seven days by the current CopyGrade Score model from public Polymarket trade history. "Already under coverage in July" means discovered before 26 July 2026, the day after the July recompute; the July column repeats the figures that edition published after its recompute. The leaderboard cohort is the 191 distinct wallets on Polymarket's profit and volume boards at capture time. Edge figures are net of Polymarket's published trading costs; we quote medians throughout, which the long negative tail does not distort. Farming flags are algorithmic risk assessments — our documented opinion from disclosed public data, not verified conduct or a statement of fact about any trader — and each wallet's page carries a dispute path. The model and the pipeline behind it change in public, with dated notes; the one scoring change in the interval is disclosed above and does not touch the flag rates. Prior editions: July 2026, June 8, June 2. Not financial advice.
CopyGrade is analysis-only — it never executes trades, holds funds, or custodies keys, and a CopyGrade Score is a documented research opinion, not financial advice.