MLB · 2026
Model Results
Last successful check: 2026-09-21T23:27:24.883957Z. Enable JavaScript to view current quote availability.
Every market has an independent outcome check. These results measure probability accuracy, not betting profits.
What was tested
3,178 completed games supply the history. Model weights use games through 2026-07-22; model selection and probability calibration use the next period through 2026-08-21. The untouched regular-season test runs 2026-08-22 to 2026-09-20.
Rolling player and team features update before each test date using earlier dates only. Live features now include completed games through 2026-09-20. Same-day results never enter a forecast.
Regular-season holdout
| Market | Games | Model Brier ↓ | Reference Brier ↓ | Relative gain | Calibration gap ↓ | Status |
|---|---|---|---|---|---|---|
| Pitcher strikeouts | 405 | 0.1545 | 0.1709 | +9.6% | 1.6 pp | Passed initial checks |
| Pitcher outs | 405 | 0.1849 | 0.2006 | +7.8% | 5.3 pp | Passed initial checks |
| Batter hits | 405 | 0.1500 | 0.1521 | +1.4% | 0.9 pp | Passed initial checks |
| Total bases | 405 | 0.1861 | 0.1886 | +1.3% | 1.4 pp | Passed initial checks |
| Home runs | 405 | 0.1019 | 0.1031 | +1.2% | 1.4 pp | Passed initial checks |
| RBIs | 405 | 0.1511 | 0.1522 | +0.7% | 1.5 pp | Passed initial checks |
| Game totals | 405 | 0.2325 | 0.2313 | -0.5% | 5.5 pp | Research only |
| Moneylines | 405 | 0.2483 | 0.2500 | +0.7% | 5.7 pp | Passed initial checks |
| Run lines | 405 | 0.2297 | 0.2301 | +0.2% | 4.9 pp | Passed initial checks |
2025 postseason audit
| Market | Games | Model Brier ↓ | Reference Brier ↓ | Relative gain | Calibration gap ↓ | Status |
|---|---|---|---|---|---|---|
| Pitcher strikeouts | 46 | 0.1771 | 0.2101 | +15.7% | 5.2 pp | Passed initial checks |
| Pitcher outs | 46 | 0.2210 | 0.2308 | +4.2% | 11.8 pp | Passed initial checks |
| Batter hits | 47 | 0.1483 | 0.1504 | +1.4% | 1.4 pp | Passed initial checks |
| Total bases | 47 | 0.1834 | 0.1862 | +1.5% | 1.8 pp | Passed initial checks |
| Home runs | 47 | 0.1052 | 0.1071 | +1.8% | 1.8 pp | Passed initial checks |
| RBIs | 47 | 0.1374 | 0.1402 | +1.9% | 2.3 pp | Passed initial checks |
| Game totals | 47 | 0.2160 | 0.2551 | +15.3% | 4.6 pp | Passed initial checks |
| Moneylines | 47 | 0.2462 | 0.2500 | +1.5% | 4.7 pp | Passed initial checks |
| Run lines | 47 | 0.2322 | 0.2346 | +1.0% | 3.2 pp | Passed initial checks |
How to read these numbers
Brier score is the average squared difference between a predicted probability and the result (0 or 1). Lower is better. The reference uses the training period’s overall outcome distribution, without player or matchup information. Relative gain compares the two scores; a negative gain means the model was worse. A small positive gain is not proof of a reliable advantage.
The calibration gap is the weighted difference between predicted and observed outcomes in ten probability buckets. Tests use fixed line grids, with several thresholds per forecast. Games are the number of distinct games; thresholds from the same game are correlated and are not independent trials. We have not established statistical significance.
Initial eligibility requires a lower Brier score than the reference, at least 150 regular-season games and a calibration gap no greater than 6 percentage points. Postseason checks require 30 games and at most 12 points. Those broader postseason limits reflect the small sample; they do not make playoff forecasts equally reliable.
The postseason audit used a separate earlier model: training through 2025-09-07, calibration through 2025-09-28. It never trained on the later regular-season holdout or on those playoff results. This is a limited historical stress test, not a validation of the current weights on future playoffs.
What this does not establish
We do not have timestamped historical book prices for this audit. There is no claimed historical return, win-rate-at-offered-odds, or proof that these forecasts beat the sportsbook market. Current picks are experimental estimates. A forward record begins with the archived forecast snapshots.
- Fixed threshold grids, not archived historical bookmaker lines; no historical ROI claim.
- Historical evaluation conditions on actual starters and original batting order; archived pregame announcements are unavailable.
- Postseason audit is small and uses an earlier model fitted before that postseason; upcoming playoff accuracy is unproven.
- No explicit handedness, weather, injury, umpire or tactical pinch-hit adjustments.
- Park estimates are shrunk venue scoring averages and can include home-team effects.
- Joint team-run approximation conditions out ties; extra-inning dynamics are not simulated pitch by pitch.
Model version: mlb-v1.0-ffe831160a077d5d-2026-09-20. Download the full validation report.