Polycopy watches every trade on Polymarket, ranks wallets by whether they are profitable to follow two minutes late, and copies a narrow few of them. Across the first half of 2026 it returned +12.7% per month on average on a $10,000 book — after fees, after modelled slippage, and after refusing 91.5% of its own signals.
Read this first Every performance figure on this page is a backtested simulation on a $10,000 book. Nothing here is live-traded or paper-traded PnL. Real order-book data appears in one section only (§07), and is marked as observed. No live capital has ever been deployed.
Three numbers carry the case. The first is the headline: June 2026 was an unseen month, run with the rules frozen beforehand and the harshest assumptions yet.
1% fee on every fill, reduced $3,500 synthetic book depth, a 120-second execution delay, and full risk caps. Rules were frozen before the month was run.
The largest simulated month, and the only one where Book A contributed materially ($1,201.02 of it). Also the widest sample.
The same June, re-run with book depth cut to $2,500. Eight of fourteen fills vanish — and it still clears. This is the downside case, not the headline.
Each month is an independent simulation starting from a fresh $10,000 book. Returns are not compounded and no equity curve is drawn here, because chaining independent runs into one would overstate the result.
| Month (2026) | Net PnL | Return | Fills | Book A | Book B |
|---|---|---|---|---|---|
| January | +$929.99 | +9.30% | 20 | — | +$929.99 |
| February | +$45.54 | +0.46% | 29 | −$50.23 | +$95.77 |
| March | +$989.96 | +9.90% | 5 | — | +$989.96 |
| April | +$459.01 | +4.59% | 36 | +$27.64 | +$431.37 |
| May | +$3,365.40 | +33.65% | 66 | +$1,201.02 | +$2,164.38 |
| June — frontier | +$1,849.86 | +18.50% | 14 | +$224.93 | +$1,624.93 |
| H1 2026 | +$7,639.76 | +12.7% avg | 170 | +$1,403.36 | +$6,236.40 |
Book C appears in no row: it has never produced a signal. The H1 total is the sum of six separate $10,000 runs, not the result of compounding one account.
Before the current rule set, the same engine returned +6.84% ($683.73, 23 fills) over the whole of 2025, and +13.72% ($1,372.41) across H2 2025 alone. 2024 was approximately flat — a sparse-data era with near-zero activity. These ran on a thinner configuration and a much smaller stored tape; they are context, not evidence.
Nine stages, from a firehose of trades down to a handful of positions. The narrowing is the whole point: almost everything that enters the top of this pipeline is thrown away.
A wallet is not ranked by what it earned. It is ranked by what a copier would have earned entering two minutes late, at prices the order book would actually have given them.
Ingest every Polymarket trade in near-real-time, plus order books, minute-level prices and resolutions.
Flag interesting wallets: size anomalies (3–5× their own median), repeated pre-news timing, surprise-winner forensics, fresh-wallet longshots, leaderboards.
An LLM sorts each market into niches and flags the ones where somebody could plausibly know the outcome early — 822 so far.
Replay every historical entry as a follower copying 120 seconds late with $100, walking real order books where stored. Keep only wallets profitable to copy.
Reject wash trading, reward farming and one-lucky-win concentration; demote wallets whose skill decays. Raw wallet PnL is never the ranking metric.
Three independent books — operator consensus, single-wallet conviction, insider conviction — each with its own budget and kill switch.
A web-enabled LLM may only veto or shrink Books B and C. The code clamps it: it can never create a trade or enlarge one.
Fractional Kelly sizing under stacked hard caps — $500 per trade, $3k per market, $6k per operator, $5k per niche, $9k total — plus daily and weekly circuit breakers.
Walk the book up to 250 bps past signal price; full fill or skip. Bail if the edge decayed. Positions settle at resolution; every signal, skip and fill is logged.
Each book has its own budget, its own kill switch and its own bar for entry. Today one of them is the engine, one is still proving itself, and one has never fired a shot.
Fires when several independent qualified operators converge on the same outcome inside a 15-minute window, with copycat wallets de-duplicated so one idea cannot masquerade as a crowd. Rare by construction. Its best month remains May ($1,201.02 on 6 fills).
A qualified wallet enters fresh at anomalous size, in sports or elections niches, between 0.50 and 0.95. This is the working core: 88% of June’s PnL and $2,164.38 across 60 fills in May. If Polycopy has an edge, it currently lives here.
Buys “knows-the-outcome” wallets in insider-decidable markets and holds to resolution. Coded, hard-gated and intentionally extreme — it has never met its own bar, not once in three years of tape. It is included here because a strategy that never triggers is still part of the honest picture.
Book B 87.8% · Book A 12.2% · Book C 0.0%
Fourteen fills consolidated into nine closed positions. All nine won. That is a genuinely good month and a genuinely small sample — both statements are true at once.
| Market | Side | Book | Result | Net PnL |
|---|---|---|---|---|
| Spain to win — World Cup 2026 match | No | B | Won | +$388 |
| Knicks — NBA Finals | Yes | B | Won | +$380 |
| IR Iran to win — World Cup 2026 match | No | B | Won | +$372 |
| Spurs to win the title | No | A | Won | +$225 |
| All nine positions | Mostly buying “No” | 9 won / 0 lost | +$1,849.86 | |
The pattern is worth naming: the profitable side was usually “No” on a favoured team in a single World Cup match. That is a structurally cheap position that pays often — and one that would look very different in a month where the favourites all held. Nine winners out of nine is not a hit rate you should extrapolate from.
June raised 164 signals and took 14 of them — a 8.5% acceptance rate. Each square below is one signal. The bot saying “no” 150 times is not the risk system failing to find trades; it is the risk system working.
| Outcome | Count | Share |
|---|---|---|
| Filled | 14 | 8.5% |
| Thin book | 48 | 29.3% |
| Insufficient edge | 39 | 23.8% |
| Insufficient depth | 32 | 19.5% |
| Edge decayed | 12 | 7.3% |
| Signal invalidated | 12 | 7.3% |
| Missing price history | 7 | 4.3% |
| Total signals | 164 | 100% |
Note that three of the top four refusal reasons are liquidity, not disagreement with the signal. Polycopy passes on trades it likes whenever the book cannot fill $500 inside 250 bps — which is precisely the discipline that makes §07 matter.
Everything above is simulated. This section is not. On 22–23 July 2026 the bot began recording real Polymarket CLOB order books, so the central assumption of the backtest — that a $500 order could actually have been filled — could finally be tested against observation.
This is the closest thing to proof the project currently has. It does not validate the returns — it validates the fill assumption underneath them. The backtests priced fills on modelled depth because real books were not being recorded before 22 July 2026; these observations say that the modelled behaviour was, on the majority of books, achievable. That is a bridge, not a verdict.
Everything that would make a careful reader discount the numbers above, stated without softening. If any single item here disqualifies the project for you, that is a reasonable conclusion to reach.
Every performance figure is a simulation on a $10,000 book. Not live trading, not paper trading. No capital has been at risk. A backtest is a hypothesis about the past, and it was written by the same person who chose the rules.
Real order-book collection only began on 22 July 2026. Every backtested fill was priced against synthetic depth layered on top of real prices. The observed statistics in §07 are a validation bridge, not a substitute — they were gathered after the fact, on different markets, in a different week.
Between 5 and 66 fills per month; June, the headline month, had 14 fills and 9 closed positions. At that sample size a 9-for-9 month and a 4-for-9 month are not far apart in probability. Months are independent $10,000 runs and do not compound.
Book B produced 88% of June’s PnL and essentially all of the history. Book A is still validating. Book C has never produced a single signal in three years of tape. This is a one-strategy engine wearing a three-strategy coat.
The paper runner is live and collecting, but it has not yet produced a single signal. Until it does, there is no out-of-sample evidence of any kind — only frozen-rule backtests, which are weaker.
Capacity is tuned for roughly $10,000. Copy-trading edge on a venue this thin does not scale linearly — §07 shows the median executable book at $633. Larger size would move prices, miss fills, and compete with the very wallets being copied.
Not another backtest. The rules have now been frozen and the paper runner is live against the real tape — the next meaningful evidence is prospective paper PnL, accumulated forward, on signals nobody has seen yet.
That is a slower and much less flattering way to make the case, which is exactly why it is the right one. A sixth consecutive profitable simulated month adds very little; a single month of honest forward paper trading would add more than all of §02.
Until then, the correct reading of this report is: a plausible, carefully-risked copy-trading system with encouraging simulations and no live evidence.