Methodology · AsPredicted #288610 · 2026-05-01

Edge Score V2-Perps

Cross-substrate skill-ranking methodology for crypto perpetual futures, equity perps, compute futures, AI benchmark markets, private-company valuation futures, and prediction markets. Four-pillar composite, locked form, coefficients pending Hyperliquid 90-day cohort fit.

The four pillars

The composite score for trader w at evaluation time is a linear combination of four standardized pillar inputs:

score(w) = c1·z(posture) + c2·z(conviction) + c3·z(discipline) + c4·z(fcr)

• Posture (CRPS): Continuous Ranked Probability Score reduction relative to a baseline forecast. CRPS is a strictly proper scoring rule that reduces to Brier on binary outcomes (Gneiting and Raftery 2007 Eq. 21). Same pillar covers continuous PnL on perps, score-at-deadline AI benchmarks, and binary-resolution prediction markets.
• Conviction: Share of realized signed PnL attributable to the trader's single largest event. Ports unchanged from V1.
• Discipline: log(1 + n_positions) with a negative coefficient (fewer-but-larger bets score higher). Optional holding-time-variance term gated on a pre-registered BIC ablation.
• Funding-Capture Ratio: Time-weighted alignment of trader exposure with adverse funding-rate sign during their hold. Range[-1, +1]; positive means harvesting funding asymmetry, negative means paying funding consistently. Drops to NaN on non-perp substrates; the composite falls back to the three-pillar form there.

Substrates covered

Section 6 of the methodology spec extends the form across seven venue kinds. CRPS implementation choice (Gaussian closed form vs ensemble Hersbach decomposition vs binary identity) and FCR pillar applicability are documented per substrate.

Validation gates (pre-registered)

Seven gates locked ex-ante at AsPredicted #288610. Each gate either passes, fails, or triggers a methodology amendment per the priorities published in the pre-registration.

1. OOS Spearman vs forward 30-day signed PnL exceeds +0.30 lower-bound on a held-out fold (paired bootstrap N=10,000 over wallets).
2. Each pillar coefficient (c1..c4) has a 95% bootstrap CI excluding zero with HC1-robust standard errors.
3. FCR pillar adds incremental Spearman of at least +0.03 absolute over the three-pillar fallback on perp substrates; otherwise the FCR pillar is dropped.
4. Multi-venue cohort: per-venue OOS Spearman positive and within ±0.10 of pooled estimate.
5. Skill-weighted aggregation of trader forecasts on continuous-outcome substrate fails its pre-registered TOST equivalence test (V2.8.2 negative-result transfer).
6. Negative-control permutation: random wallet-label shuffle within fold (N=10,000); 95% CI excludes the observed value.
7. Multiple-testing correction: BH-FDR across H1 + H2 (4 pillars) + H3 + H4 (per venue) + H5; FDR-adjusted thresholds reported alongside per-test p-values.

What ships when

• Now: Form-locked methodology + spec + AsPredicted filing + scoring modules (crps.py, funding_capture.py, edge_score_v2_perps.py). Coefficients raise CoefficientsPendingError until the cohort fit lands.
• Pending Hyperliquid 90-day cohort fit: Wallet panel ingestion (public-API only); cohort filter (n_positions ≥ 20); OLS fit with HC1-robust standard errors; freeze-coefficient write-back.
• Pending validation run: All seven gates evaluated; per-pillar significance + per-venue invariance reported; methodology paper published with public PDF + data bundle + audit-chain anchor.

References

• Gneiting, T. and Raftery, A. (2007). “Strictly Proper Scoring Rules, Prediction, and Estimation.” JASA 102(477), 359-378. Eq. 21 (Gaussian closed-form CRPS).
• Hersbach, H. (2000). “Decomposition of the Continuous Ranked Probability Score for Ensemble Prediction Systems.” Wea. Forecasting 15(5), 559-570.
• Litterman, R. and Scheinkman, J. (1991). “Common Factors Affecting Bond Returns.” Journal of Fixed Income 1(1), 54-61. Level / slope / curvature decomposition for term structure.
• Lakens, D. (2017). “Equivalence Tests: A Practical Primer for t Tests, Correlations, and Meta-Analyses.” Social Psychological and Personality Science 8(4). TOST protocol used in H5.
• López de Prado, M. (2018). Advances in Financial Machine Learning. Wiley. Ch. 7 (purged k-fold) for fold geometry.
• Convexly Edge Score V1 + V1-M papers at /research/edge-score-methodology-v1 and /research/edge-score-methodology-v1m.

This work used AI tools (Claude, GPT-4) as research aids during methodology design and pre-publication review. All claims on this page are reproducible from the public spec at docs/research/v2-perps/methodology-v2-perps-spec.md and the freeze-commit code modules under services/api/app/engine/. No claim on this page is taken as true on the basis of an AI tool's output; every quantitative result is recomputable from the freeze commit with the documented random seed.