Win probability

Game Prediction Model

20seasons of data2006–2025
5,133games trained on4,166 more held back for testing
64.7%called correctlyon games the model never saw
±2.1points of calibration driftwhen it says 70%, it happens about 70% of the time

23.2%decided by 3 points or fewer

47.5%decided by 7 points or fewer

8points, the median margin

A quarter of NFL games come down to a field goal, which is the ceiling every forecaster runs into. Predicting every game decided by more than a touchdown perfectly, and coin-flipping the rest, would still only reach 76.3%. Across 5,133 games, 2006–2025.

Where the model was wrong

Every miss from 2023 to 2025, out of sample. The model predicted 854 games over those three seasons and got 292 of them wrong, at 65.8% accuracy. What follows is the 292, sorted, explained, and in one section excused.

The finding that reframes the rest

Of the 292 misses, the betting market picked the same losing side in 253 of them (87%). Only 39 games , or 4.6% of the slate, are failures unique to this model. The rest are games the sharpest-priced market in sport also called wrong. Judging the model by its raw miss count measures the difficulty of predicting football, not the quality of the model.

Games predicted: 854
Got wrong: 292
Median losing margin: 5 pts
Model-specific: 39

Where the misses concentrate

By confidence band

Band	Games	Wrong	Accuracy
Coin flip, no call	145	62	57.2%
Low confidence	164	61	62.8%
Medium confidence	163	59	63.8%
High confidence	382	110	71.2%

The bands behave as they should. The model is not confidently wrong; it is uncertain and wrong, which is the honest failure mode, and the one a reader can act on, by ignoring the coin flips.

By how badly it lost

3 points or fewer107 · 36.6%
4–782 · 28.1%
8–1442 · 14.4%
15 or more61 · 20.9%

107 of 292 misses were lost by a field goal or less. Those are not misjudgments. A one-score game is near a coin flip whatever was known beforehand. The 61 lost by fifteen or more are the ones worth interrogating.

Its best and worst seasons

Every season the model was tested on, 2010 to 2025, covering 4,166 games, each predicted by a model trained only on the seasons before it. The bar is the model; the tick is the market on the same games. Bars start at 58%, not zero, because every season lands in a twelve-point range and a full axis would hide the differences.

201060.5%-5.8
201167.5%+0.4
201262.4%-4.1
201364.0%-6.2
201468.6%+0.4
201561.3%-2.1
201662.7%-2.0
201769.9%+0.4
201864.6%-0.8
201968.6%+3.9
202066.0%-1.1
202158.1%-4.2
202264.2%-2.1
202366.0%-1.4
202467.4%-3.2
202564.1%-1.8

Best season · 2017

69.9%

Market, same games: 69.5%
Model minus market: +0.4 pts
Misses: 78 of 259
Shared with the market: 66 (85%)
Model-specific: 12
Lost by 3 or fewer: 18
Calibration error: 0.066

The market was near its own best that year too, and the season was unusually clean: 20% of games finished within a field goal, the lowest share in the record. Only 12 of its 78 misses were games the market got right. This is what the model looks like when football cooperates.

Worst season · 2021

58.1%

Market, same games: 62.3%
Model minus market: -4.2 pts
Misses: 119 of 284
Shared with the market: 103 (87%)
Model-specific: 16
Lost by 3 or fewer: 39
Calibration error: 0.117

The market fell to 62.3% the same year, its worst season in the record, so most of this was the season, not the model. But not all: the model was also at its most confident (66.3% average) and its worst calibrated (error 0.117, against 0.066 in 2017). Confident and wrong is the expensive combination, and it is the one year the record shows it.

Why the good years are good

Across all 16 seasons, the model’s accuracy tracks the market’s at r = 0.63. Its worst season, 2021, was also the market’s worst season. That is the honest reading of the chart above: a good year is mostly a season that was predictable, and a bad one is a season that was not. The same model, fit the same way, produced 2017 and 2021. What changed was the football.

Measured against the market instead of against itself, the picture changes. 2019 is the model’s real high point, +3.9 points clear of the market on 258 games, the only season in the record it beat the closing line by more than a point. 2013 is the real low point: 64.0% looks respectable until you see the market at 70.2% on the same slate.

What the model could not see

Four candidate explanations, tested against the 292 misses. Three of them fail, which is what makes the fourth worth believing.

Turnovers

The explanation

When the model was right, its pick won the turnover battle by +0.82 per game. When it was wrong, that team lost it by -0.97, a swing of 1.79 turnovers per game between a correct prediction and an incorrect one, and 180 of 292 misses featured the picked team losing that battle. This is the largest gap between what the model knows and what decides games, and it is largely irreducible: the model carries takeaway and giveaway rates as rolling inputs, but a rate is not an event. Nothing pre-game predicts which Sunday a quarterback throws three picks.

Weather

Not the culprit

The intuition that bad weather breeds upsets is not supported. Ugly conditions appear to suppress variance: fewer possessions and a narrower playbook favor the better team.

Condition	Games	Accuracy
All games	854	65.8%
Outdoors	568	65.8%
Wind 15 mph or more	52	75.0%
Freezing (32°F or below)	55	67.3%

Known injuries

Handled

In the 96 games with a quarterback listed out, the model went 74.0%, better than its overall rate. A known absence is priced correctly because it is on the injury report and the model reads it. The injuries that hurt are the ones at 1:20pm: an in-game exit has no pre-game feature.

Divisional games

Real, but modest

Divisional games run 63.1% across 293 of them, against 67.2% on the 561 non-divisional. The familiarity cliché holds, but it is worth 4.1 points of accuracy, not the mythology attached to it.

The twenty worst

Ranked by confidence multiplied by losing margin, so these are the games where the model was both sure and badly beaten. “TO” is the turnover margin for the team it picked, and it is negative in most of them, which is the section above stated one game at a time.

Game	Picked	Conf	Final	TO	Leading performer
2023_05_NO_NE	NE	60%	NO 34–0 NE	-3	Alvin Kamara (NO): 80 rush yds, 17 rec yds
2024_04_WAS_ARI	ARI	69%	WAS 42–14 ARI	0	Jayden Daniels (WAS): 233 pass yds, 1 TD, 47 rush yds
2025_11_LAC_JAX	LAC	65%	LAC 6–35 JAX	0	Travis Etienne (JAX): 73 rush yds
2024_02_NO_DAL	DAL	76%	NO 44–19 DAL	-1	Alvin Kamara (NO): 115 rush yds, 65 rec yds
2025_03_ATL_CAR	ATL	61%	ATL 0–30 CAR	-3	Bijan Robinson (ATL): 72 rush yds, 39 rec yds
2025_08_MIA_ATL	ATL	73%	MIA 34–10 ATL	-1	Tua Tagovailoa (MIA): 205 pass yds, 4 TD
2023_19_CLE_HOU	CLE	54%	CLE 14–45 HOU	-2	C.J. Stroud (HOU): 274 pass yds, 3 TD, 1 rush yds
2025_14_WAS_MIN	WAS	53%	WAS 0–31 MIN	-3	J.J. McCarthy (MIN): 163 pass yds, 3 TD, 19 rush yds
2023_19_PHI_TB	PHI	66%	PHI 9–32 TB	0	Baker Mayfield (TB): 337 pass yds, 3 TD, 16 rush yds
2023_14_HOU_NYJ	HOU	63%	HOU 6–30 NYJ	1	Breece Hall (NYJ): 40 rush yds, 86 rec yds
2024_15_TB_LAC	LAC	63%	TB 40–17 LAC	0	Mike Evans (TB): 159 rec yds
2023_16_CIN_PIT	CIN	62%	CIN 11–34 PIT	-3	George Pickens (PIT): 195 rec yds
2023_04_CIN_TEN	CIN	59%	CIN 3–27 TEN	0	Derrick Henry (TEN): 2 pass yds, 1 TD, 122 rush yds, 11 rec yds
2023_05_CHI_WAS	WAS	70%	CHI 40–20 WAS	-2	DJ Moore (CHI): 230 rec yds
2023_17_CAR_JAX	CAR	53%	CAR 0–26 JAX	-1	Travis Etienne (JAX): 102 rush yds, 16 rec yds
2024_03_DEN_TB	TB	72%	DEN 26–7 TB	-1	Bo Nix (DEN): 216 pass yds, 0 TD, 47 rush yds
2023_03_HOU_JAX	JAX	68%	HOU 37–17 JAX	-2	Tank Dell (HOU): 145 rec yds
2024_22_KC_PHI	KC	75%	KC 22–40 PHI	-2	Xavier Worthy (KC): 157 rec yds
2024_06_JAX_CHI	JAX	71%	JAX 16–35 CHI	-1	Caleb Williams (CHI): 226 pass yds, 4 TD, 56 rush yds
2024_07_SEA_ATL	ATL	66%	SEA 34–14 ATL	-3	Kenneth Walker III (SEA): 69 rush yds, 24 rec yds

The honest limit

253 games the market also lost (87%): shared blind spots, not model defects.
107 decided by three points or fewer, outcomes no pre-game information could separate.
180 where the picked team lost the turnover battle, the largest single cause and close to irreducible.
39 genuine model-specific failures, and its confidence in those was modest.

Strip out the shared market misses, the one-score games and the turnover luck, and what remains is a small residue. That residue, not the 292 headline number, is the model’s real error, and it is not large enough to close the gap with a market that prices injury news and sharp money within minutes. The productive direction is not more features: wider sets scored worse. It is better probabilities on the games it already calls.

Put plainly: the model is not failing to understand football. It is running into the fact that a quarter of NFL games are decided by a field goal, and a fifth turn on a bounce nobody can forecast.

model-misses-2023-2025.html (prediction pipeline), reconciled against artifacts/oos_predictions.csv, model_dataset.csv and report_per_fold.csv.

Calibrated Logistic Regression (core + market features) · walk-forward by season (train on all prior seasons, test on the next)