South Ark Daily Picks / Model Performance & Prediction History
Model Performance & Prediction History
Measured chronological backtests and an immutable forward prediction ledger, kept separate from betting results.
Published Pick Record
Every published pick →CFB and NFL selections count against their original spread or total. Pending picks become wins, losses or pushes only after confirmed final results. NBA and NHL remain in private model training.
| Sport | W–L–P | Pending | Win % |
|---|---|---|---|
| College Football | 2–0–0 | 16 | 100.0% |
| NFL | 0–0–0 | 0 | Unavailable |
Units and ROI require recorded odds and stakes; missing prices are not invented.
Research backtests
| Sport | History | Latest final | Test season / games | Score MAE | Brier | Log loss |
|---|---|---|---|---|---|---|
| NFL | 919 | 2026-10-05 | 2026 / 64 | 7.7 | 0.226 | 0.642 |
| NBA | 0 | 2026-06-13 | Unavailable | Unavailable | Unavailable | Unavailable |
| NHL | 0 | 2026-10-07 | Unavailable | Unavailable | Unavailable | Unavailable |
| MLB | 0 | 2025-09-28 | Unavailable | Unavailable | Unavailable | Unavailable |
Score MAE is mean absolute error per team; lower is better. Brier and log loss evaluate home-win probability on decisive games. Each test prediction uses only earlier calendar dates; coefficients refit chronologically. Season thresholds are fixed in advance and not optimized on reported results.
Historical ingestion times are unavailable: calendar-date exclusion is a proxy, not evidence of point-in-time injury or roster coverage. The preserved NBA 2013–2015 and MLB 2023–2025 archives are historical studies. Once connected, the NBA model trains separately on API-SPORTS results; archive rows are retained. None establishes profitable betting performance.
Probability calibration
NHL
| Probability bin | Sample | Mean estimate | Observed home wins |
|---|---|---|---|
| 0.0%–20.0% | 0 | Unavailable | Unavailable |
| 20.0%–40.0% | 0 | Unavailable | Unavailable |
| 40.0%–60.0% | 0 | Unavailable | Unavailable |
| 60.0%–80.0% | 0 | Unavailable | Unavailable |
| 80.0%–100.0% | 0 | Unavailable | Unavailable |
Forward prediction ledger
0 distinct graded professional games; 0 immutable forecast revisions. Revising a pregame estimate creates a new record rather than changing the original.
| Published | Sport / game | Original score | Team win probability | Result | Model |
|---|---|---|---|---|---|
| No frozen predictions in this filter yet. Backtests are kept separate. | |||||
The published CFB and NFL record tracks locked spread and total selections, including pending picks, wins, losses and pushes. Research winner forecasts above remain separate. Units and ROI require prices and stakes.
Preserved CFB original-line pick record and validation →