Bayesian priors results: H5/H4/H6/H10/H11 REJECTED, H9 direction-correct
Path C (xG calibration, 136K shots): - H9 player finishing: Brier -0.00040 direction correct but BELOW -0.002 threshold - H10 opponent context: +0.00002 (hurts slightly) REJECTED
Path C (xG calibration, 136K shots):
- H9 player finishing: Brier -0.00040 direction correct but BELOW -0.002 threshold
- H10 opponent context: +0.00002 (hurts slightly) REJECTED
- H11 team tier: +0.00010 (hurts slightly) REJECTED
Path A (EPL, 3 seasons 2022-25):
- H5 Elo warm-start: GW1-5 delta +0.0026 REJECTED (target -0.003, wrong direction)
- H4 Standings: GW1-5 delta +0.0052 REJECTED (target -0.003, wrong direction)
- Both priors help GW6+ slightly but hurt GW1-5 for data-rich leagues
Path B:
- H6 Marcel early prior: Post Phase-0 fix, marginal entry-adj ROI +0.0%
(base with signal = base without = +6.9%), bootstrap p=0.48 REJECTED
Key finding: The hypothesis that priors help early-season for leagues with 6+ years of solver history is wrong. The cold-start solver converges fast enough that the prior is noise, not signal. Priors may still help leagues with less historical data (needs broader multi-league backtest) but the EPL-level evidence is clear.
Per the decision matrix in docs/specs/bayesian-priors-experiment-spec.md: no path fully passes. Remaining untested: H1 (squad-value warm-start), H2 (regularization target), H3 (squad-value early-season recovery), H7 (promoted team lambda adjust), H8 (ensemble, gated on 2+ passing).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>