how polyscope measures

Polymarket ranks traders publicly by profit. Profit is not the same as being predictive. A trader can be profitable from a handful of oversized wins while being systematically wrong on most of their diverse positions — including the ones where they disagree with the market.

PolyScope asks a narrower, more useful question: when top-ranked traders take positions that diverge from the market price, does their direction agree with how the market eventually resolves?

1. Every 5 minutes, scan 500 active Polymarket markets meeting quality thresholds (max open interest or 24h volume ≥$50K, and ≥$10K 24h volume).
2. For each market, fetch current positions held by the top-100 ranked traders from Polymarket's leaderboard.
3. Compute a weighted consensus: each trader's implied YES probability is weighted by inverse rank × alpha ratio × log-scaled size × category-skill multiplier.
4. If the gap between consensus and market price exceeds 10% and the composite signal score passes threshold, emit a signal. For the top ~80 candidates we refine using trade-weighted consensus with a 24h exponential half-life decay.
5. Persist per-signal, per-trader attribution: who held which direction, at what size, with what weight. This is the foundation of per-trader accuracy scoring.

The first ~98K resolved signals were where the original strategy was finalized: we flipped the aggregate SM direction universally (after finding raw consensus was anti-predictive at 8.3%). The flipped strategy hit 96% at the headline level — but a follow-up per-skew backtest showed that was almost entirely a composition effect from heavily lopsided markets. (Capture has since grown past 139K resolved signals; the live numbers above always reflect the current dataset.)

The per-skew breakdown forced a correction. On tight 40–60% markets, fading SM lost 0 of 17 unique markets. On moderate, 3 of 21. SM was actually predictive when it dissented on those bands — going with SM would have won. On very-lopsided markets (≥90% or ≤10%), fading gives a near-100% hit rate but ~0% ROI — the favorite almost always wins anyway.

The live signal view now labels that finding: very-lopsided markets as composition-risk, and other bands as top-trader-side observations. On the 1,556-market backtest at finalization, this flipped net ROI from −2.6% to +3.9% at the same 97% headline hit rate. §4 below shows live filter performance against the current dataset.

The per-band breakdown after correction:

Headline hit rate is a misleading summary when the sample is dominated by one skew band. We report it because it is what people ask for; per-band numbers and ROI are what actually matter.

Aggregating top-100 traders into a single consensus hides per-trader skill. We score each contributor individually on their resolved divergent positions. A trader qualifies as predictive when all three hold: at least 30 resolved observations, point accuracy above 50% (genuinely above coin flip), and a Wilson-95%-CI lower bound ≥ 40% (not purely noise).

Computing live filter performance…

An earlier version used only the CI-lower gate and misreported a much rosier ROI. The issue: high-volume anti-predictive traders (47.5% accuracy on 1,000+ observations) have a Wilson-CI lower bound that still crosses 40%, so they cleared the gate and appeared on nearly every signal. Requiring the point estimate itself to be above 50% restores the filter's meaning.

Caveats: the eligible-contributor list is currently short — none yet cleared the 40% Wilson-lower threshold at time of writing. The filter's value grows as per-trader observations accumulate. Signals are tagged predictive-backed when they qualify.

Three findings hold up:

1. Polymarket's P&L leaderboard is not a predictive-skill leaderboard. Top-ranked-by-profit traders cluster near 45–55% individual accuracy on divergent signals. Most are not reliably predictive.
2. A handful of individuals clear a meaningful edge. none yet currently have Wilson-CI lower bounds above 40% on 30+ resolved observations. This is the population the predictive filter surfaces.
3. Aggregation changes the picture per-band. Weighted consensus of top-100 traders is predictive enough to overcome individual noise on tight and moderate markets; it is composition-bound on very-lopsided markets.

The per-trader leaderboard is on /traders — predictive ↔ anti-predictive, with sample sizes and confidence intervals shown so you can weigh the evidence yourself.

The builder identity endpoint reports whether this deployment exposes PolyScope's public Polymarket builder code. When configured, orders routed through PolyScope carry a bytes32 identifier and attribute volume to us on-chain.

Full transparency page with live order log: /builder.

• Per-trader accuracy capture began April 12, 2026. Early per-address samples are small — treat anything below n=25 as preliminary.
• Resolved-outcome coverage depends on Polymarket's market-close cadence. Fast-resolution markets (sports, daily prices) dominate the resolved sample; long-horizon markets are underweighted.
• Signals are generated every 5 minutes. By construction they cannot capture intraday micro-structure or react faster than that.
• Category labels come from Polymarket's tags and are inconsistent across similar markets.
• PolyScope offers an optional non-custodial order-routing UI. Your wallet signs every order locally and submits directly to Polymarket; PolyScope never holds keys or funds. Signals are research, not trading advice. Past accuracy does not imply future performance. See /terms.