# Week9 model research — challengers remain separate

Run: 20261025T173456Z. Next slate: October27–31,2026 (America/Chicago),56 games. Designated live baseline remains EPA-OLS; displayed incumbent ensemble uses its original training pack.

## Common retrospective cohort

All model families train on 2016–2025 (5,070 rows) and evaluate the same229 games from2026 Weeks5–8. These are retrospective research results, not backfilled pregame selections. Saved pregame records are graded separately in the Week8 report. Market MAE on this cohort is11.516 points.

| Family | Winner accuracy | Brier | Log loss | Native margin MAE |
| --- | ---: | ---: | ---: | ---: |
| 01 Ridge margin | 68.12% | 0.2115 | 0.6102 | 11.992 |
| 02 RandomForest | 69.00% | 0.2090 | 0.6058 | 12.028 |
| 03 XGBoost | 65.50% | 0.2403 | 0.7256 | Unavailable: probability model |
| 05 LogisticRegression | 65.07% | 0.2111 | 0.6084 | Unavailable: probability model |
| 07 Stacked ensemble | 65.50% | 0.2111 | 0.6094 | Unavailable: probability model |

Only Ridge and the direct-margin random forest predict native margins. Classifier inverse-normal pseudo-margins use training outcome-margin dispersion; their MAE/ATS values do not establish a native score or cover model. The shuffled stacking folds are the maintained research implementation, not a chronological promotion gate. Multiple comparisons and one partial season prevent choosing a new baseline from these results.

## Prospective Week9 comparison

The maintained pipeline produced the incumbent pack-only and expanded-through-Week8 ensemble forecasts for all56 unstarted Week9 games. The incumbent exactly reproduces the live display probabilities (maximum absolute difference0). Expanded training has5,299 rows, changes mean home-win probability by0.0237 in absolute terms (maximum0.0964), and flips no predicted favorite. Expanded forecasts are research only; no promotion.

The independently supplied official Week9 feature CSV was unavailable; this run uses the repository’s own feature builder and labels that basis explicitly. Huber has no reviewed runnable implementation, and the FastAI tutorial has no maintained bounded Week9 runner. Held SRS, neutral, failed chronological-stack and other stopped lanes remain held.

## Artifact identity

- 2026 frame SHA256: `fa6ed63b3481e6bc4990f3ee221fac7dd590e153bf6a0296f8e5b9a210eb86d8` (229 rows; through Week8).
- Incumbent training SHA256: `0837fbbedb3a684b201340cc993067c34179ef0f7ea0e6cc49da8424378a4dd9`.
- Expanded training SHA256: `43b40f44d9c422a10c8cdad4861daaf49409ce9caf5490dca4c26bac8bfdd282`.
- Feature captures: results `20261025T172319018190Z`; adjusted EPA `20261025T172420386352Z`. No Week9 outcomes exist in training.
- Full normalized prediction tables and identical expanded training input are retained in the private research run. Licensed raw captures and fitted scratch models are excluded from Git.

## Next experiment — proposal only

Freeze a fixed50/50 blend of primary and saved market-implied margin prospectively, compare against both parents on identical future games/lines, and evaluate margin error and probability calibration independently of ATS thresholds. Do not tune weights on Week8 or replace the baseline before the existing promotion gates.