Methodology
How the Copy Engine ranks — and what a simulation of copying it produced
Every copy-trading product ranks wallets by profit and hopes profit means skill. It doesn’t. This page sets out exactly what our engine measures instead, every gate it applies, and a historical simulation of copying its daily picks scored against the two obvious alternatives — copying the P&L leaderboard, and copying random active wallets. Today’s board.
SIMULATED BACKTEST
Simulated $50 copies, 2026-07-01 to 2026-08-11 — backtested historical results, not realized user returns. The simulation assumes a fill at the leader’s entry VWAP plus 8¢ — the cost we measured on the tape for a copier arriving 1–5 seconds behind the leader — and a hold to resolution; real execution differs. Nothing here is a forecast, a guarantee, or investment advice. Full method and caveats.
01 / Score
What the engine ranks on
For every position a wallet resolved in the trailing window we take its excess: the outcome (1 or 0) minus the price it paid. A wallet that buys at 30¢ and wins scores +0.70; buying at 95¢ and winning scores +0.05. Paying the right price is the skill, not picking the winner.
Those excesses are then demeaned by category and day, so a day when every favourite won lifts nobody, and summed with Bayesian shrinkage toward zero — the sum is divided by (number of positions + τ). A wallet with three lucky resolutions cannot outrank one with two hundred measured ones. The shrinkage constant is load-bearing: too little chases noise, none at all just ranks by volume, and both lose money on the same window.
Alongside the skill score, a second sleeve ranks on realized return inside the validated entry-price band, which is where the copyable, high-probability flow lives.
02 / Gates
The gates
A high score is not enough. A wallet only reaches the board if it clears every gate below. Each one exists because removing it made the simulated result worse, or because the wallets it admits cannot actually be followed.
Skill window
14 days
Shrinkage
τ = 25
Rebuild
06:45 UTC daily
Board size
Top 25
- Sniper gate> 30 min
Median minutes from first buy to resolution must exceed the threshold — you cannot copy a fill placed minutes before the market settles.
- Liveness≤ 48 h
Hours since the last taker fill. A stopped wallet cannot lose its streak — a stale leader is survivorship bait, not a signal.
- Decay switchoff (a 10pp drop at Welch t ≤ −2.5 trips it)
Trailing 7-day mean per-position return collapsing versus the prior window — the eviction signal a trailing average hides.
- Taker-visible> $0
The copyable (taker-side) slice of the in-band book must itself be profitable — maker-only edge is uncopyable by construction.
- Entry-price band0.65 – 0.9
The validated entry-price window the in-band sleeve measures its return over.
- Position cost basis$10 – $100
Cost-basis range counted by the in-band screen.
- Minimum resolved positions (skill sleeve)≥ 10
Resolved positions required in the skill window before a wallet is eligible at all.
- Minimum positions (in-band sleeve)≥ 30
In-band positions required before the in-band sleeve considers a wallet.
Showing the engine’s documented default parameters — sign in to read the live values.
03 / Cadence
Daily refresh
Measured edge half-life is about two days, so a weekly board would spend most of its life stale. The engine rebuilds every morning and each build is recorded as its own walk-forward observation — the score a wallet had on a given day cannot be quietly revised later.
- 02:15 UTCNightly metrics chain refreshes resolved-position and fill inputs.
- 06:45 UTCCopy Engine rebuilds the day’s scores and gates. The build refuses to publish a day computed on stale position data rather than freezing bad numbers into the record.
- 07:15 UTCThe daily evaluation arms read the fresh engine output, so every day is recorded as a new walk-forward observation.
04 / Backtest
Simulated track record
Each arm below picked 25 wallets a day and was scored under an identical copy protocol over 35 scored pick days (2026-07-01 to 2026-08-11; 7 days inside the window were not scored, on a stride fixed in advance). 174,579 simulated copies in total.
SIMULATED BACKTEST · simulated $50 copies, 2026-07-01 to 2026-08-11. Not realized user returns.
| Arm | $ / copy | Copies | Win % | Day-grain t | Days + |
|---|---|---|---|---|---|
Polyrank Copy Engine 14-day shrunken category-excess skill (τ=25), gated | +$8.38 | 10,970 | 64.3 | +6.59 | 28/35 |
P&L leaderboard control Top 25 by trailing-30d realized P&L — the naive copy-trading default | −$11.20 | 68,141 | 56.4 | −59.72 | 0/35 |
Random active wallets control Deterministic-hash random pick from active wallets | −$10.77 | 3,838 | 55.8 | −9.33 | 1/35 |
“Day-grain t” treats each pick day as one observation rather than each copy — the honest unit, because copies made on the same day are not independent. “Days +” counts the days on which that arm’s mean was positive, out of 35. Day-bootstrap 95% interval on the engine arm (2000 iterations): +$24.70 to +$40.30 per day-mean.
The ranking is what moves the number
Same days, same copy rules, only the wallet selection differs — which isolates the ranking as the cause rather than the market regime.
SIMULATED BACKTEST · simulated $50 copies, 2026-07-01 to 2026-08-11. Not realized user returns.
Engine vs P&L leaderboard
+$37.50
per simulated $50 copy · paired t = +9.40 · 35 matched days, identical copy rules
Engine vs random active wallets
+$41.50
per simulated $50 copy · paired t = +8.64 · Sniper-filtered variant, matched days only
Both halves of the window, scored separately
SIMULATED BACKTEST · these rows are the sniper-filtered variant (entries more than 30 minutes before resolution — the ones a copier could realistically have taken), so they are not the same cut as the unrestricted headline table above. Simulated, 2026-07-01 to 2026-08-11.
| Period | Days | Engine $ / copy | t | P&L board $ / copy | t |
|---|---|---|---|---|---|
| July 2026 | 26 | +$35.15 | +6.99 | −$5.50 | −12.25 |
| August 2026 | 9 | +$24.03 | +4.82 | −$5.56 | −11.72 |
05 / Protocol
How the simulation was scored
Identical for every arm. Publishing the protocol is the point: a backtest whose rules you cannot inspect is a marketing number.
- Clip size
- $50 per copied position, fixed
- Entry
- The picked wallet’s taker-buy positions first filled the next day, at that day’s entry VWAP
- Slippage assumption
- +8¢ added to the entry price on every copy — the cost measured on the tape for a copier arriving 1–5s after the leader
- Fees
- 7% charged on winnings — net = payout/(vwap+slip) − 1 − 0.07·(1 − (vwap+slip))
- Exit
- Hold to resolution — no discretionary exits, no stop-losses
- Exclusions
- INVALID and still-unresolved markets are dropped (0.4–2% on mature days)
- Unit of significance
- The day — each pick day is one observation (day-grain t-statistics)
06 / Robustness
What we tried to break it with
SIMULATED BACKTEST · every figure below is from the same simulation, 2026-07-01 to 2026-08-11.
- Not one wallet: the top contributor accounts for 18.8% of simulated engine profit. Removing it leaves +$24.13 day-grain (t = +8.27), still positive on 35 of 35 days.
- Removing the top three wallets leaves +$22.05 per simulated copy.
- 60 of the 85 wallets the engine picked were net-positive over the window.
- The sniper filter is not doing the work for the engine (unrestricted ≈ sniper-filtered). It is load-bearing for the random control, whose "wins" are resolution snipes a copier could not have taken.
- Reconstructed picks match the live experiment: on the two overlapping days, backtest and live day-means agree within noise.
07 / Caveats
Limits of this evidence
- These are simulated results. No user capital was deployed and no user held these positions. Past simulated performance does not indicate future results.
- Execution is modelled, not promised. Every taker-visible entry is filled at the leader’s same-day VWAP plus 8¢ at a $50 clip — 8¢ is what we measured on the tape for a copier arriving 1–5 seconds behind the leader (median +7.0¢), and the engine’s simulated edge stays positive up to ~15¢. Real fill probability and book depth still vary by market.
- The defensible core is the ranking separation — engine versus controls on identical days under identical rules — not the absolute dollar figure.
- Capacity beyond a $50 clip is untested here.
- Most of the window is a same-formula reconstruction of what the shipped engine would have picked, computed with leak-safe as-of bounds. Genuinely live logged picks exist for only the last four days of the window.
- Seven days inside the window were not scored — a deterministic agent stride, fixed in advance, not a choice made after seeing results.
- Nothing here is investment advice, a signal to trade, or a forecast. Polyrank is read-only analytics.