Levandor

Copy-Trade Cross-Category Sweep (2026-06-29) — Handoff to cohort

3 min read

Copy-trade cross-category sweep (2026-06-29) — handoff to cohort

Cross-category validation of “follow a public Polymarket leaderboard sharp” as a standalone copy-edge. Run read-only by the crypto lane (crypto) as a handoff to the cohort lane (cohort_algo). One uniform method applied to weather, politics, and sports.

Following public leaderboard "sharps" is NOT a robust standalone copy-edge

Validated across weather, politics, and sports with a uniform method: the follow-edge is statistically insignificant and non-persistent everywhere we could measure it. This independently reproduces the prior “luck-dominated / no demonstrated edge” finding in copy-trade-algorithm. The lone exception is crypto-updown, which cohort already validated (+18–30%, t≈9) and owns. Do not build a category copy-follower.

Method (reusable, validated)

For each candidate wallet’s BUY entries on resolved markets:

  • Follow-edge (post-entry drift): won − price_at(entry + follow_lag), measured off the CLOB prices-history series so it strips the wallet’s own market impact. won is derived from the token’s resolved final price (>0.95 win / <0.05 loss).
  • Significance: bootstrap 90% CI on the follow-edge.
  • Persistence vs luck: early/late split — does the edge hold out-of-sample within the wallet’s own history.
  • Capacity: post-entry fillable $ at ≤ entry+0.02 within the follow window, via /trades?market=<conditionId>.

API gotcha

On the trades endpoint, ?asset=<token> does NOT filter — only ?market=<conditionId> does. Discovery uses data-api.polymarket.com/v1/leaderboard?category=<cat>&orderBy=PNL|VOL (9 fixed categories).

Results by category

Weather (daily temp, 11 buckets)

A copyable class of ~9 slow-drift wallets does exist, but on the firmed sample (n≈120/wallet) the follow-edge shrank to +0.004…+0.047/share with all CIs including 0 (insignificant). Capacity is the killer — median fillable only $4–26 (thin longshot books). Verdict: NO.

Reference wallets: Pityok1, 1f66796b…, b40e8967…, 686880…, cbbc5e03…. See copytrade-top-weather-traders and weather-top50-decypher-2026-06-19.

Politics

Inconclusive — data-starved. Top-politics wallets’ large PnL was earned on past (2024) events beyond the 3000-trade window; current positions sit on still-open future markets (e.g. one wallet: 95 politics entries / 0 resolved). Only 4 of 20 had ≥12 resolved samples; none significant, 3/4 fail persistence. Slow resolution makes politics un-backtestable from public data.

Sports (games, fast resolution, deepest books)

Capacity is ample (median fillable 16,717) — but follow-edge is insignificant and non-persistent: of 5 scored wallets all CIs span 0, 4/5 fail persistence, the two largest samples (n=45, 33) are negative, and edge decays to ~0 by +1d. Verdict: NO.

Category triage (PnL × depth)

Politics 92M top-10 PnL with deep books; weather is the runt (107M) — cohort’s lane.

Conclusion + recommendation

Don’t build a category copy-follower. The edge that looks present in small samples is luck-dominated noise — CIs include 0, persistence fails, and point estimates shrink with n (one weather wallet’s follow-edge went +0.131 → +0.043 → +0.017 as the sample grew). The only durable copy-edge found anywhere is cohort’s crypto-updown.

The surviving usable pattern elsewhere is “detect, don’t mirror”: sharp-wallet appearance as a confirmation input to a producer’s own signal (as the weather producers already do), never a standalone follower.

For cohort (lane owner)

Method + scripts are reusable if you want to re-run with better tooling:

  • proper per-category market enumeration (instead of keyword filters);
  • settlement-truth via Gamma (instead of the final-price proxy).

All candidate addresses + per-wallet metrics are saved in the analysis scratchpad. The crypto lane (crypto) ran this read-only; the build, if any, is cohort’s (cohort_algo). See cohort-mirror-feature for the existing per-trader mirror signal source.

Graph View