PnL vs Edge Score

Across the full 8,656-wallet Polymarket cohort in our V1 methodology paper, the Spearman rank correlation between a wallet's Brier score (calibration) and its realized PnL is +0.148 (in-sample, descriptive). The top 1 percent of wallets by absolute PnL captured 36.2 percent of all signed profit on the platform. In a market that concentrated, a PnL leaderboard is heavily shaped by a small number of large outcomes.

Convexly's Edge Score is a composite of three behavioral pillars fitted against signed log PnL: posture, conviction, and discipline. The original V3b fit's in-sample out-of-fold Spearman with signed log PnL was +0.514, but that figure shares positions with the PnL it is scored against and its dominant input is PnL-derived, so it overstates predictive skill; the constants were also refit on 2026-07-13, so that fit statistic does not describe the current column (dated method-change note at /methodology). In the composite's one forward test, a per-wallet temporal holdout with the threshold set before the analysis ran, the frozen score held an out-of-sample Spearman of +0.11 (95% CI [0.05, 0.18]) with forward PnL. That did not clear the +0.30 bar. Treat the right-hand column as a descriptive behavioral profile, not a validated skill ranking.

The divergence you see above therefore shows that the two orderings disagree, not which one is right. Wallets in the left column but not the right have a PnL rank ahead of their composite rank; wallets in the right column but not the left have a composite rank ahead of their PnL rank. The honest read of any single record is whether its realized-edge 95% interval clears zero, which neither column displays. A wallet-by-wallet realized-edge scan of the published top-50 cohort is being prepared as a research note; see the top-50 skill scan in the research index.

The V1-M extension paper published 2026-04-22 adds a 15,106-user Manifold cohort and a paired-window sweepcash sensitivity analysis. The original public bundle reported median concentration 8.9 percentage points lower in the sweepcash window on 1,647 paired users, with concentration delta defined on n=333. A 2026-05-04 recovered-cohort rerun excluding political and election markets reports -9.2pp, 95% CI [-12.8, -3.6], Wilcoxon p=0.0021 on n=515 of 1,208 paired users with defined concentration delta. The concentration shift survives the sensitivity check, but causal attribution to sweepcash alone remains limited by the catalog confound.