Go, field goal and punt · explicit next-score expected-points scenarios
Next scoring event before halftime; zero if a proven-complete half ends without another score. Positive values favor the offense. Values compare explicitly modeled scenarios; they do not establish an optimal action or win probability.
Regular season, both teams FBS, regulation, score within eight, more than five minutes left in half. Historical league model; not a win-probability or causal policy model. The score/time inputs enforce this scope; they do not adjust a league field-position curve. Unparsed names remain league-only. No current roster or PAAR-to-make-probability inference.
Go
+1.11 EP
Scenario envelope: +0.62 to +1.60 · n=176
Conversion 50.6%. Success assumes exactly the line to gain, then first down. Failure assumes no net gain/return: opponent 65 yards to goal.
Field goal
+0.84 EP
Scenario envelope: +0.47 to +1.17 · n=179
52-yard attempt; make 53.0%. League distance-band estimate. Unreturned dead-ball miss: opponent 65 yards to goal. Blocks/returns are not projected.
Punt
-0.33 EP
Scenario envelope: -0.73 to +0.12 · n=102
Empirical net-gain distribution in this 20-yard field band, including observed return-TD risk. Predicted opponent start averages 87.4 yards to goal. Punt touchbacks use own 20. Unmatched possession transitions are excluded.
Envelopes combine 90% component probability bands, game-bootstrap possession bands and punt sampling variation. They are sensitivity ranges, not calibrated joint confidence intervals. Action selection, gain-on-conversion assumptions, weather, team strength and model misspecification remain outside these ranges.
Later-game component checks
Fixed 2023–2024 estimation; 2025 regular-season evaluation, with disjoint games. Training ends 2024-12-14T20:00:00+00:00; testing starts 2025-08-23T16:00:00+00:00. No 2026 outcomes or production forecast changes.
Component
2025 n
Check
First-down next-score EP
14530
MAE 4.325 points; constant baseline 4.544. Training n=33402.
Fourth-down conversion
1151
Brier 0.227; log loss 0.646.
League kicking
1086
Brier 0.149; log loss 0.462.
Shrunk kicker
1086
Brier 0.150; 164 matched earlier player/band records; remaining kicks use league fallback.
Punt transition
2233
Next-score MAE 5.022 points; landing MAE 8.0 yards. 2558 unmatched/unsupported punt states excluded across all three seasons.
Conversion and kicking calibration
Conversion
Probability band
n
Predicted
Observed
0–20%
0
Unavailable
Unavailable
20–40%
94
38.0%
35.1%
40–60%
504
50.9%
50.8%
60–80%
553
71.4%
71.8%
80–100%
0
Unavailable
Unavailable
League kicking
Probability band
n
Predicted
Observed
0–20%
0
Unavailable
Unavailable
20–40%
0
Unavailable
Unavailable
40–60%
240
55.0%
60.4%
60–80%
175
64.7%
69.1%
80–100%
671
87.4%
88.5%
Shrunk kicker
Probability band
n
Predicted
Observed
0–20%
0
Unavailable
Unavailable
20–40%
0
Unavailable
Unavailable
40–60%
240
55.0%
60.4%
60–80%
182
65.3%
70.3%
80–100%
664
87.5%
88.4%
Model assumptions and sources
Conversion uses distance/field bands and Beta(1,1) smoothing. FG uses five-yard attempt bands and the same smoothing; named kickers need at least five parsed earlier attempts overall and use 20 league-prior attempts per band. A name/team pair parsed from play text is not a verified player ID. Historical PAAR is not used. Possession EP is monotone by yards to goal on first downs in proven-complete halves; 80 game-cluster bootstrap resamples describe fit variation. Punts use a 20-yard-field-band empirical net-gain distribution and observed terminal-return risk. Alternative-action outcomes are unobserved: these checks validate components on selected historical actions, not the counterfactual ranking.
Failed-FG spot assumes an untouched kick beyond the neutral zone that becomes dead; live returns, blocks and enforcement can change the spot. NCAA 2025 rules, 8-4-2b. A made FG ends the next-score horizon at +3; TD scenarios assume seven including the try.