PropStein Track Record & Evidence

Full-board close calibration (Week 07)

Headline metric: every graded board decision (N = 37):
• 2 of 37 beat the close (5.4%, 95% Wilson CI: [1.5%, 17.7%]; Clopper–Pearson exact: [0.7%, 18.2%])
• 34 of 37 tied the close (91.9%, line unchanged)
• 1 of 37 closed worse (2.7%, -1.0 yd drift)
Average line delta: +0.05 yds (95% CI: [-0.23, +0.33] yds). Average CLV EV: 0.82%.

Headline honesty: Across the full board only 2 of 37 decisions beat the close that week. That is what an honest calibration record looks like early. We publish full-board calibration first so there is zero cherry-picking.

Graded ledger (preliminary sample, n = 8)

7-1-0 over 8 graded records — win rate 87.5% (95% Wilson CI: [52.9%, 97.8%]), beat-close rate 100.0%, average close beating value 4.29% of implied probability, net P&L 549.64 on 790.0 staked (ROI 69.57%). Ledger root hash 178e29974bf8748c.

Read this honestly: 8 records is a tiny sample. None of these rates are meaningful evidence of edge yet. We publish them below the full board so the methodology is auditable from day one, and so there is no incentive to quietly drop bad weeks.

Published model evaluation (holdout)

Decision cutoff enforced: yes; 862 snapshots evaluated, 0 future snapshots excluded to prevent leakage.

Brier score (model)

0.2813

Brier score (naive baseline)

0.25

Brier skill

-12.51%

Log loss (model)

0.758

Log loss (baseline)

0.6931

The model currently underperforms the naive baseline on this holdout. We publish that because it is true. Recalibration is an active workstream; roles and thresholds carry provisional labels everywhere in the product until the skill score turns positive.

Market-Model Edge Fit & Guarantee Status (Phase 2 & 3)

Model has no edge over market closing lines at this time. All projections should be treated as informational.

Chronological regression fitting p = w · pmodel + (1 - w) · pmarket on active trigger decisions yields:
• ŵ = 0.6742 (Standard Error: 0.3957, 95% CI: [-0.1013, +1.4497], N = 37)
Because the 95% confidence interval spans zero, there is no statistically significant evidence of closing line alpha over the market consensus.

Public actions in effect:
• Beat-Close Guarantee Suspended: Beating or tying a close on a small sample guarantees nothing about win rate or expected value. The marketing guarantee is publicly withdrawn.
• Paid Signups Frozen & Quant Tier Waitlisted: Commercial paid upgrades remain suspended until N ≥ 300 graded board decisions achieve positive CLV bounded strictly above zero with 95% confidence.
• Uncensored Research Access: Projections and live pricing calculations are made available without paywalls in observer research mode.

What we publish as the sample grows

Calibration by predicted-probability bucket; win rate with sample size; ROI with a confidence interval; maximum drawdown; splits by sport, market and model version; and losing weeks next to winning weeks, every week. The commitment is the record, not the highlight reel.