Situational win probability
Score, clock, possession and field state · historical league model
Estimate the eventual winner from a regulation pre-snap scrimmage state. Historical league model, fitted on 2023–2024 and checked on separate 2025 games. Teams are generic: strength, timeouts, weather and home-field advantage are not modeled.
Clock is time remaining in the entire regulation game (30:00 = 1,800 seconds). Field position is always measured toward the possessing team’s opponent goal. Use the score before the snap. No kickoffs, tries, live-ball or between-play states.
Estimated win probability
Home 53.8%
Away 46.2%
1,351 earlier plays across 235 games in this clock/score/down band; 207 games with a winning offense and 167 with a losing offense (a game can contribute both). This is marginal support, not an exact-state sample or individual confidence interval.
Late-game states are estimated only with earlier support; the last 14 seconds are unavailable. Clutch checks below measure historical performance, not clock-management skill. Omitted timeouts and team strength can matter greatly. No per-state calibrated confidence interval is available. These probabilities do not value alternative plays and are separate from R04’s EP scenarios and the production pregame forecasts.
Later-game evaluation
Fixed fit through 2024-12-14T20:00:00+00:00; evaluation starts 2025-08-23T16:00:00+00:00. No 2026 outcomes. Final scores label the winner only; they are not predictors. Games are disjoint across fit and test.
| 2025 states | Plays / games | Brier / score-clock baseline | Game-balanced Brier (90% game bootstrap) |
|---|---|---|---|
| All bounded states | 106,045 / 762 | 0.152 / 0.155 | 0.148 (0.141–0.156) |
| Available states | 98,183 / 762 | 0.164 / 0.167 | 0.155 (0.148–0.163) |
| Available: last 5 minutes | 4,531 / 425 | 0.143 / 0.154 | 0.102 (0.089–0.113) |
| Available: last 2 minutes | 1,987 / 284 | 0.161 / 0.177 | 0.129 (0.115–0.144) |
Lower Brier is better. Baseline: independently fitted score and clock-scaled score only. Fits weight games equally; play-level scores and game-balanced scores differ. Bootstrap resamples held-out games; its interval is for aggregate Brier, not a prediction. 7,862 bounded 2025 states fail the earlier-support gate.
Available-state and clutch calibration
Aggregate improvement does not establish local calibration. Inspect the clutch bands: late-game estimates can be materially miscalibrated, and repeated plays from the same game are not independent evidence.
Available states
| Offense probability band | Plays | Predicted | Won |
|---|---|---|---|
| 0–10% | 13137 | 4.0% | 3.6% |
| 10–20% | 8192 | 14.8% | 17.2% |
| 20–30% | 8699 | 25.3% | 27.6% |
| 30–40% | 7959 | 35.0% | 36.7% |
| 40–50% | 9051 | 45.3% | 44.9% |
| 50–60% | 13769 | 54.7% | 54.4% |
| 60–70% | 9165 | 64.7% | 63.9% |
| 70–80% | 8010 | 75.2% | 73.4% |
| 80–90% | 8470 | 85.0% | 83.1% |
| 90–100% | 11731 | 95.5% | 96.3% |
Available: last 5 minutes
| Offense probability band | Plays | Predicted | Won |
|---|---|---|---|
| 0–10% | 655 | 5.9% | 4.4% |
| 10–20% | 595 | 14.3% | 24.4% |
| 20–30% | 464 | 25.1% | 28.9% |
| 30–40% | 381 | 34.6% | 37.5% |
| 40–50% | 276 | 44.5% | 45.7% |
| 50–60% | 269 | 55.2% | 58.4% |
| 60–70% | 284 | 65.0% | 66.5% |
| 70–80% | 315 | 75.3% | 79.0% |
| 80–90% | 537 | 85.5% | 88.3% |
| 90–100% | 755 | 94.8% | 97.7% |
Available: last 2 minutes
| Offense probability band | Plays | Predicted | Won |
|---|---|---|---|
| 0–10% | 173 | 7.5% | 9.8% |
| 10–20% | 285 | 14.4% | 29.8% |
| 20–30% | 272 | 25.3% | 27.6% |
| 30–40% | 217 | 34.3% | 38.7% |
| 40–50% | 132 | 44.4% | 39.4% |
| 50–60% | 145 | 55.2% | 57.9% |
| 60–70% | 155 | 64.7% | 69.0% |
| 70–80% | 126 | 75.3% | 84.1% |
| 80–90% | 193 | 85.2% | 94.3% |
| 90–100% | 289 | 94.5% | 99.3% |
Model, support and sources
Eventual game winner, including overtime, from a regulation pre-snap state. Regular-season FBS versus FBS league states. No team strength, home advantage, timeouts, weather or personnel inputs. Not CFBD's provider WP model or a production forecast.
Logistic regression with score, clock-scaled score, time, field position, distance, down and late-state interactions. No provider WP is copied or distilled. This is not the CFBD retired or revamped WP family. No public provider calculator API is assumed.
Availability requires at least 100 training plays, 30 games and 5 games for each outcome in the clock/score/down band. This gate does not establish joint support for every field-position/distance combination.
208,847 fit states / 1503 games; 51 hashed supplied inputs.
league-regulation-logistic-v1 · 30584f78f1590adf106e286d9088f861fe25d28fd5bddf4934ee8c7d5ffa3208
Download coefficients, source hashes, support, exclusions and evaluation