Audit infrastructure

Negative-result registry

Every pre-registered hypothesis Convexly has tested and rejected, with the AsPredicted ID, the effect-size confidence interval, the specific claim disproved, and the open questions that remain. Indexed in machine-readable form at /research/negative-results.json and verifiable via the audit chain.

Last updated 2026-04-30.3 entries

Entry #001

AsPredicted #287983

Skill-weighted aggregation as a per-market price prior

Rejected

Twenty-four pre-registered skill-weighted aggregator variants (linear, exponential, rank-based, top-k, conviction-weighted, posture-weighted, calibration-weighted, multiplicative, et al.) were tested as forecasts. All twenty-four were worse than the market price. The wash-trader-filtered variant Brier delta CI [+0.16028, +0.19287] sits inside the pre-registered TOST equivalence range [+0.154, +0.204]; movement after wash filtering was +0.00243 Brier (1.4% relative).

Specific claim disproved
Aggregating Polymarket trader forecasts by any skill-weighted shape produces a per-market probability prior that is more accurate than the market price.
Effect-size CI (Brier delta)
+0.179 [+0.16028, +0.19287]
Direction
Aggregator was WORSE than market price by 17.9% Brier on average.
Pre-registered TOST equivalence range
[+0.154, +0.204]
Filed / ran
Filed 2026-04-25, ran 2026-04-26, wash-filter robustness 2026-04-29

What remains open

  • Per-wallet ranking (V1, OOF Spearman +0.514, NOT closed by V2.8.2)
  • Within-wallet z-score for insider-pattern detection MVP (NOT closed)
  • Wash-trading detection (Sirolly methodology, NOT closed)
  • Structural / constraint-projection inference (CME V0.1, uses MARKET PRICES not skill-weighted aggregation as forecast input, NOT closed)

Implications for product

Bright line: when evaluating any new commercial proposal, check first 'does it use skill-weighted aggregation as a per-market price prior?' If yes, reject unless there's a substantive reason to believe the prior shape was wrong vs the 24 already-rejected shapes.

Entry #002

AsPredicted #287368

V1 V3b composite survives strict per-wallet temporal holdout

Rejected

V1.5 experiment E2 (pre-registered): refit V3b OLS on a 2024-01-01 to 2025-09-30 training window with a 14-day purge, score on a 2025-10-15 to 2026-04-15 held-out window. Frozen V1 coefficients applied to training-window pillars produced Spearman ρ = +0.111 with 95% CI [+0.046, +0.175] and p ≈ 0.001. Positive and the CI excludes zero, but well below the pre-registered ρ ≥ +0.30 pass threshold. The fail is reported per the pre-reg's no-post-hoc-re-analysis rule.

Specific claim disproved
The V1 V3b composite, fit on full-cohort data and applied to training-window pillars, predicts forward held-out PnL at Spearman ρ ≥ +0.30 on the V1-M cohort.
Effect-size CI (Brier delta)
Spearman ρ = +0.111, 95% CI [+0.046, +0.175], p ≈ 0.001
Direction
Positive but below the +0.30 pre-registered threshold.
Pre-registered TOST equivalence range
Pre-registered ρ ≥ +0.30 pass criterion; observed +0.111 fails the threshold.
Filed / ran
Filed 2026-04-25, ran 2026-04-27

What remains open

  • V3b as a CROSS-SECTIONAL ranker (the per-period ranking + S4 partial Spearman +0.494 controlling for log capital remain open)
  • V3b as a per-wallet TEMPORAL forecast: closed by E2; do not market it as such

Implications for product

V3b is positioned as a cross-sectional ranker, not a per-wallet temporal predictor and not a forecast-aggregation weight. External framing should never claim 'V3b predicts your forward PnL across windows' without citing this E2 rejection.

Entry #003

AsPredicted #287368

V1 V3b per-quarter IC stability across the pre-registered window

Rejected

V1.5 experiment E7 (pre-registered): bucket V1-M positions by quarter, recompute V3b with quarter-local standardization, compute Spearman of frozen V3b vs signed log PnL in each quarter. Pre-registered window 2024Q1 to 2025Q2 (six quarters); 5 of 6 quarters present (2024Q1 had 29 wallets, below the ≥5-position cohort filter). Median per-quarter Spearman = +0.038, range across quarters [-0.164, +0.155]; 3 of 5 quarters positive. Pre-registered pass criterion required median ρ ≥ +0.30 AND ≥5/6 positive quarters. Both legs failed.

Specific claim disproved
The frozen V1 V3b composite produces a stable per-quarter Spearman with realized log PnL across the 2024Q1 to 2025Q2 window.
Effect-size CI (Brier delta)
Median per-quarter Spearman = +0.038, range [-0.164, +0.155] across 5 quarters
Direction
Per-quarter IC is unstable; one quarter (2025Q1) is meaningfully negative with CI [-0.221, -0.102].
Pre-registered TOST equivalence range
Pre-registered: median ρ ≥ +0.30 AND ≥5/6 positive quarters. Observed: median +0.038, 3 of 5 positive. Both legs fail.
Filed / ran
Filed 2026-04-25, ran 2026-04-27

What remains open

  • Exploratory extension to 2025Q3 to 2026Q2 (NOT in pre-reg) shows 2026Q1 ρ = +0.281, CI [+0.249, +0.314]; whether IC stability returns in more recent windows is open and pre-registered for replication as #11 in the 12-month plan
  • Cross-quarter autocorrelation of per-quarter ρ values is +0.366 lag-1, so when the V3b signal is on it tends to persist; the failed E7 test was on the level of the IC, not its persistence

Implications for product

V3b is not marketed as a stable per-quarter IC predictor. Forward-validation rolling-Spearman publication (#11 in the 12-month plan, ships 2026-10-01) is the next opportunity to revisit the per-quarter stability claim with a fresh pre-reg.

Registry policy

  • Addition rule: An entry is added the moment a pre-registered hypothesis is rejected by the test it specified. Entries are never removed; corrections to the analysis are added as additional fields, not by overwriting prior content.
  • Verification: Each entry's effect-size CI and AsPredicted ID is independently verifiable from the linked paper data bundle. The audit chain at https://www.convexly.app/research/verify covers the input + output hashes for each registered run.
  • Credibility-loaded-term audit: This file passes the credibility-loaded-term audit per CLAUDE.md: every effect-size estimate has a CI in the same line; 'rejected' refers to a formal pre-registered test, not informal opinion; 'robust' and 'near zero' do not appear without their CI / TOST counterpart; only pre-registered tests are listed (V1.5 S6 was an exploratory follow-up, not a pre-registered hypothesis, and is therefore not a registry entry).