Results & validation

How good are the forecasts?

Every number here was produced by the evaluation code on real data. There are two kinds of results, reported separately. Retrospective backtests rebuild forecasts for past seasons from team data available before each game, but with two approximations: the quarterback who actually started is used as a stand-in for the expected starter, and historical market lines are untimed (approximately closing lines). This makes them somewhat optimistic. Prospective results are forecasts that were generated before kickoff and archived with timestamps in the 2026 season; only these are genuine live tests.

Plain-English verdict. The market point spread and total are very hard to beat. The combined model (market + football) is statistically indistinguishable from the market alone: marginally better in 2022–24, marginally worse on the untouched 2025 test. The football-only model, which never sees the market, is clearly better than naive baselines but about 0.3 points of margin error worse than the market. Probabilities are well calibrated and the 80%/95% ranges cover close to 80%/95% of outcomes.

2025 locked test · margin error (MAE)9.70combined · market 9.67 · football 9.97 (n=285)

2025 locked test · win-probability log loss0.6327market-only 0.6286 · football 0.6444 (lower is better)

2025 locked test · interval coverage80%of margins inside the 80% range · 95% range: 97%

Live (prospective) results

Each game is scored once per horizon, against the latest valid version generated before the cutoff. "Publicly verifiable" means GitHub's own record shows the forecast file was public before kickoff; forecasts generated before kickoff on this project's machine but published later are reported separately and corrected on their game pages.

All forecasts generated before kickoff

Horizon	Games	Margin MAE	Market MAE	Football MAE	Total MAE	Log loss	80% coverage
early	42	8.64	8.61	8.28	9.97	0.7885	88%
final_pregame	59	8.23	8.20	8.38	10.21	0.7680	85%

Publicly verifiable before kickoff

Horizon	Games	Margin MAE	Market MAE	Football MAE	Total MAE	Log loss	80% coverage
early	42	8.64	8.61	8.28	9.97	0.7885	88%
final_pregame	58	7.93	7.91	8.04	10.29	0.7630	86%

Live head-to-head, 2026

The three forecasts on the same finished games, using only forecasts that were publicly verifiable before kickoff. Margin error uses the last forecast before kickoff (the market forecast is the line's own margin); the record vs the market line uses the locked version of each pick. Lower margin error is better. So far the lowest season-to-date margin error is the market line (7.91 points over 58 games): too few games to separate the forecasts with confidence.

Week	Games	Average margin error (points)			Vs market line
Week	Games	Model	Stats-only	Market	Model pick	Stats-only
Week 3	15	7.08	7.85	7.10	8–6–1	5–9–1
Week 4	16	7.26	6.58	7.22	7–8–1	12–3–1
Week 5	15	10.42	10.55	10.40	7–7–1	8–6–1
Week 6	12	6.76	7.08	6.71	3–8–1	6–5–1
Season to date	58	7.93	8.04	7.91	25–29–4	31–23–4

Cumulative average margin error

ModelStats-onlyMarket

Record vs market line: wins–losses–level results for the model pick and for the side of the locked line where the stats-only margin fell (a tracking record, not the official model pick). The market line has no record against itself.

Retrospective benchmark, not live results. The historical tables below are the original walk-forward benchmark (preserved unchanged). They rest on two approximations that live forecasts do not have: (1) the actual starting QB stands in for the expected starter at the final horizon, and (2) historical market lines are a single untimed line per game, roughly the closing line. Both make the historical numbers somewhat optimistic compared with genuinely prospective forecasting, which is reported separately above.

Point-forecast accuracy (final pregame horizon)

Seasons	Model	Games	Margin MAE	Margin RMSE	Total MAE	Total RMSE
2019–21 (tuning)	Combined (selected)	821	10.23	13.13	10.64	13.27
2019–21 (tuning)	Market only	821	10.22	13.12	10.63	13.26
2019–21 (tuning)	Football only	821	10.41	13.34	10.79	13.50
2019–21 (tuning)	Naive home average	821	11.59	14.83	11.18	13.98
2022–24 (development)	Combined (selected)	854	9.48	12.47	10.11	12.90
2022–24 (development)	Market only	854	9.49	12.48	10.11	12.95
2022–24 (development)	Football only	854	9.75	12.72	10.27	13.13
2022–24 (development)	Naive home average	854	10.63	13.80	10.80	13.66
2025 (locked test)	Combined (selected)	285	9.70	12.29	10.53	13.32
2025 (locked test)	Market only	285	9.67	12.24	10.42	13.24
2025 (locked test)	Football only	285	9.97	12.70	10.54	13.28
2025 (locked test)	Naive home average	285	11.00	14.11	11.00	13.83

Walk-forward: each season is predicted by models trained only on earlier seasons. Historical market lines are a single untimed line per game (approximately closing), so market comparisons apply to the final-pregame horizon. The early (72-hour) horizon can only be evaluated for the football-only model historically; the table in the full report covers it.

Cumulative margin error, 2019–2025

Running mean absolute margin error across all retrospective walk-forward forecasts (lower is better), shown once at least 100 games have accumulated.

Combined (selected)Market onlyFootball onlyNaive home average

Paired comparison with the market-only benchmark

Difference in mean absolute error on the same games (model minus market; negative = better than the market), with 95% intervals from a season-week block bootstrap. Intervals that include zero mean no demonstrated difference.

Seasons	Model	Target	Games	Difference	95% interval
2022–24 (development)	Football only	margin	854	+0.258	[0.119, 0.397]
2022–24 (development)	Football only	total	854	+0.162	[0.021, 0.315]
2022–24 (development)	Combined (direct)	margin	854	+0.068	[-0.004, 0.138]
2022–24 (development)	Combined (direct)	total	854	+0.095	[-0.020, 0.202]
2022–24 (development)	Combined + injuries	margin	854	-0.005	[-0.024, 0.014]
2022–24 (development)	Combined + injuries	total	854	+0.004	[-0.081, 0.093]
2022–24 (development)	Combined (boosting)	margin	854	+0.095	[-0.031, 0.215]
2022–24 (development)	Combined (boosting)	total	854	-0.047	[-0.174, 0.091]
2022–24 (development)	Combined (selected)	margin	854	-0.008	[-0.035, 0.019]
2022–24 (development)	Combined (selected)	total	854	-0.004	[-0.092, 0.083]
2025 (locked test)	Football only	margin	285	+0.301	[-0.029, 0.593]
2025 (locked test)	Football only	total	285	+0.122	[-0.178, 0.442]
2025 (locked test)	Combined (direct)	margin	285	+0.066	[-0.081, 0.200]
2025 (locked test)	Combined (direct)	total	285	-0.029	[-0.215, 0.151]
2025 (locked test)	Combined + injuries	margin	285	+0.014	[-0.073, 0.103]
2025 (locked test)	Combined + injuries	total	285	+0.087	[-0.102, 0.271]
2025 (locked test)	Combined (boosting)	margin	285	-0.014	[-0.192, 0.149]
2025 (locked test)	Combined (boosting)	total	285	+0.290	[0.007, 0.573]
2025 (locked test)	Combined (selected)	margin	285	+0.026	[-0.113, 0.163]
2025 (locked test)	Combined (selected)	total	285	+0.106	[-0.082, 0.292]

Probability calibration

Combined model, final horizon. Dots on the dashed line mean the probabilities meant what they said; dot size grows with the number of games.

2025 locked test

2022–24 development

Win-probability scores

Seasons	Model	n	Log loss	Brier
2019–21 (tuning)	Combined (selected)	554	0.6291	0.4232
2022–24 (development)	Combined (selected)	854	0.6225	0.4219
2019–21 (tuning)	Market only	554	0.6286	0.4232
2022–24 (development)	Market only	854	0.6234	0.4228
2019–21 (tuning)	Football only	554	0.6424	0.4327
2022–24 (development)	Football only	854	0.6466	0.4434
2025 (locked test)	Combined (selected)	285	0.6327	0.4307
2025 (locked test)	Market only	285	0.6286	0.4269
2025 (locked test)	Football only	285	0.6444	0.4387

Interval coverage (combined model)

Seasons	Target	Nominal	Actual	Mean width (pts)
2022–24 (development)	margin	80%	79.7%	31.8
2022–24 (development)	margin	95%	95.3%	52.1
2022–24 (development)	total	80%	82.0%	33.5
2022–24 (development)	total	95%	94.7%	51.1
2025 (locked test)	margin	80%	80.0%	32.5
2025 (locked test)	margin	95%	97.2%	52.0
2025 (locked test)	total	80%	78.2%	32.7
2025 (locked test)	total	95%	94.0%	51.4

Candidate feature groups (evaluated separately before promotion)

Promotion rule, fixed before results were seen: the combined model must improve the group's target loss in the development seasons with a 95% interval entirely below zero, and also improve it in the tuning seasons. The 2025 locked season is not used. Negative differences mean the group helped.

non-QB injuries (final weekly report; expected lost snap share by unit): not promoted (rule not met)

Inputs: final weekly injury report (strictly available before the final horizon). Early horizon: not evaluable: mid-week report versions are not retained historically (now being collected).

Model	Seasons	Games	Loss difference	95% interval	Total RMSE without → with
Combined	tuning	821	0.035	[-0.221, 0.307]	13.269 → 13.269
Combined	development	854	0.271	[-0.473, 0.936]	12.896 → 12.902
Football only	tuning	821	-0.560	[-1.885, 0.774]	13.505 → 13.483
Football only	development	854	-0.495	[-1.833, 0.916]	13.128 → 13.139

weather (exposure-weighted wind, gusts, cold, precipitation): not promoted (rule not met)

Inputs: RETROSPECTIVE: Open-Meteo Historical Forecast API (stitched short-lead forecasts), 2019+.

Model	Seasons	Games	Loss difference	95% interval	Total RMSE without → with
Combined	tuning	285	-2.843	[-5.970, 0.252]	13.358 → 13.251
Combined	development	854	-0.465	[-1.467, 0.557]	12.940 → 12.922
Football only	tuning	285	-3.167	[-6.726, 0.094]	13.540 → 13.423
Football only	development	854	-2.017	[-4.571, 0.287]	13.106 → 13.029

Coaching changes are deferred (no dated source). Weather forecasts and every injury-report version are now collected before each game, so both groups can later be re-evaluated on strictly as-of data.

Quarterback start-probability audit

What the historical rates measure: whether the depth-chart QB1 started. That is not the same as playing or being active. Rates carry 90% intervals, and each test season is predicted from earlier seasons only.

Final status	Final practice	Cases	Started	Played, not started	Declared inactive	P(start), 90% interval
Doubtful	DNP	20	0	0	12	0.02 (0.00–0.09)
Doubtful	Limited	9	0	0	7	0.05 (0.00–0.19)
Doubtful	Full	2	0	0	2	0.17 (0.00–0.57)
Questionable	Limited	131	68	5	39	0.52 (0.45–0.59)
Questionable	Full	41	36	1	3	0.87 (0.78–0.94)
Questionable	DNP	21	8	0	11	0.39 (0.23–0.56)

Estimator	Cases	Log loss	Brier	Mean predicted	Observed
status beta11	Questionable + Doubtful (162)	0.6246	0.2245	0.565	0.475
status jeffreys	Questionable + Doubtful (162)	0.6216	0.2244	0.563	0.475
status x practice	Questionable + Doubtful (162)	0.6019	0.2143	0.535	0.475

The practice-aware estimator (highlighted) is used. Across all QB1 games it is better with a 95% interval excluding zero; on the 162 Questionable/Doubtful cases the gain is not statistically established. All estimators over-predict starts for listed QBs in recent seasons (the rate has drifted down); this is reported, not adjusted.

Archive: 2025 retrospective forecasts

Retrospective walk-forward backtest (not published before games). Combined model shown; these were not published before the games.

Show 285 games

Wk	Game	Forecast	P(home)	Final	Margin error
1	ARI @ NO	25.9–18.9	25%	20–13	0.0
1	BAL @ BUF	26.1–25.6	47%	40–41	1.4
1	CAR @ JAX	20.4–23.5	60%	10–26	12.9
1	CIN @ CLE	27.1–20.0	25%	17–16	6.1
1	DAL @ PHI	20.6–30.7	81%	20–24	6.1
1	DET @ GB	23.7–25.2	54%	13–27	12.5
1	HOU @ LA	20.7–25.4	65%	9–14	0.3
1	KC @ LAC	25.2–22.4	39%	21–27	8.7
1	LV @ NE	20.6–24.7	64%	20–13	11.2
1	MIA @ IND	22.9–25.0	56%	8–33	22.9
1	MIN @ CHI	23.8–22.9	45%	27–24	2.1
1	NYG @ WAS	19.9–26.7	72%	6–21	8.3
1	PIT @ NYJ	19.8–18.1	42%	34–32	0.4
1	SF @ SEA	22.9–20.0	38%	17–13	1.1
1	TB @ ATL	24.4–22.3	41%	23–20	1.0
1	TEN @ DEN	17.7–27.5	80%	12–20	1.8
2	ATL @ MIN	20.6–23.3	58%	22–6	18.7
2	BUF @ NYJ	27.4–19.6	23%	30–10	12.2
2	CAR @ ARI	20.0–26.3	71%	22–27	1.4
2	CHI @ DET	21.3–27.8	71%	21–52	24.4
2	CLE @ BAL	17.5–31.5	88%	17–41	9.9
2	DEN @ IND	23.5–21.6	42%	28–29	2.9
2	JAX @ CIN	22.5–26.7	64%	27–31	0.2
2	LAC @ LV	25.4–21.4	34%	20–9	7.0
2	LA @ TEN	22.9–17.9	31%	33–19	9.0
2	NE @ MIA	21.0–23.7	58%	33–27	8.7
2	NYG @ DAL	20.1–24.9	66%	37–40	1.8
2	PHI @ KC	23.5–22.6	45%	20–17	2.1
2	SEA @ PIT	19.2–23.2	63%	31–17	18.0
2	SF @ NO	21.8–19.5	40%	26–21	2.7
2	TB @ HOU	20.0–22.1	56%	20–19	3.1
2	WAS @ GB	23.8–27.5	62%	18–27	5.4
3	ARI @ SF	21.2–22.7	54%	15–16	0.6
3	ATL @ CAR	24.7–19.2	29%	0–30	35.5
3	CIN @ MIN	19.9–22.7	59%	10–48	35.2
3	DAL @ CHI	26.3–23.6	39%	14–31	19.6
3	DEN @ LAC	22.2–24.2	56%	20–23	1.0
3	DET @ BAL	25.1–28.8	62%	38–30	11.7
3	GB @ CLE	24.3–15.1	19%	10–13	12.2
3	HOU @ JAX	20.8–23.6	59%	10–17	4.2
3	IND @ TEN	24.4–18.4	28%	41–20	15.1
3	KC @ NYG	25.3–19.1	27%	22–9	6.8
3	LA @ PHI	21.7–26.1	65%	26–33	2.5
3	LV @ WAS	20.0–23.4	61%	24–41	13.6
3	MIA @ BUF	20.3–32.5	85%	21–31	2.2
3	NO @ SEA	16.6–24.8	76%	13–44	22.8
3	NYJ @ TB	18.5–25.6	73%	27–29	5.2
3	PIT @ NE	22.5–21.5	45%	21–14	6.0
4	BAL @ KC	25.1–22.7	40%	20–37	19.4
4	CAR @ NE	17.3–23.4	70%	13–42	22.9
4	CHI @ LV	24.4–25.0	51%	25–24	1.7
4	CIN @ DEN	19.1–27.4	76%	3–28	16.7
4	CLE @ DET	17.6–29.0	83%	10–34	12.6
4	GB @ DAL	26.2–20.0	28%	40–40	6.1
4	IND @ LA	23.6–27.6	63%	20–27	3.0
4	JAX @ SF	20.1–24.6	65%	26–21	9.6
4	LAC @ NYG	23.7–18.4	30%	18–21	8.3
4	MIN @ PIT	22.5–21.6	45%	21–24	3.9
4	NO @ BUF	17.0–33.1	91%	19–31	4.1
4	NYJ @ MIA	21.3–24.3	59%	21–27	3.0
4	PHI @ TB	23.8–20.8	37%	31–25	2.9
4	SEA @ ARI	22.9–21.7	44%	23–20	1.8
4	TEN @ HOU	16.5–24.9	76%	0–26	17.7
4	WAS @ ATL	20.5–22.5	56%	27–34	4.9
5	DAL @ NYJ	23.3–24.5	53%	37–22	16.3
5	DEN @ PHI	21.2–25.4	64%	21–17	8.2
5	DET @ CIN	29.3–17.7	14%	37–24	1.4
5	HOU @ BAL	21.1–20.6	46%	44–10	33.4
5	KC @ JAX	23.6–21.4	40%	28–31	5.2
5	LV @ IND	19.8–27.7	75%	6–40	26.0
5	MIA @ CAR	23.1–20.7	40%	24–27	5.5
5	MIN @ CLE	17.9–15.3	39%	21–17	1.4
5	NE @ BUF	22.0–30.2	76%	23–20	11.2
5	NYG @ NO	19.6–21.9	57%	14–26	9.8
5	SF @ LA	18.4–27.6	78%	26–23	12.2
5	TB @ SEA	20.3–24.3	63%	38–35	7.0
5	TEN @ ARI	17.5–26.0	77%	22–21	9.5
5	WAS @ LAC	23.6–24.9	53%	27–10	18.3
6	ARI @ IND	19.6–29.1	79%	27–31	5.4
6	BUF @ ATL	27.1–21.9	30%	14–24	15.2
6	CHI @ WAS	22.5–28.7	70%	25–24	7.2
6	CIN @ GB	16.1–30.9	89%	18–27	5.8
6	CLE @ PIT	17.4–23.8	71%	9–23	7.5
6	DAL @ CAR	25.2–20.9	33%	27–30	7.3
6	DEN @ NYJ	23.8–17.9	28%	13–11	3.9
6	DET @ KC	24.0–26.1	56%	17–30	10.9
6	LAC @ MIA	24.9–22.1	38%	29–27	0.8
6	LA @ BAL	24.7–18.1	26%	17–3	7.4
6	NE @ NO	25.3–22.7	39%	25–19	3.4
6	PHI @ NYG	23.6–16.6	25%	17–34	24.1
6	SEA @ JAX	23.5–22.7	46%	20–12	7.3
6	SF @ TB	21.0–24.6	61%	19–30	7.4
6	TEN @ LV	19.7–23.1	61%	10–20	6.6
7	ATL @ SF	21.6–24.0	57%	10–20	7.7
7	CAR @ NYJ	18.7–20.3	54%	13–6	8.6
7	GB @ ARI	25.1–18.3	25%	27–23	2.8
7	HOU @ SEA	18.8–23.5	65%	19–27	3.3
7	IND @ LAC	23.2–24.8	54%	38–24	15.6
7	LA @ JAX	22.8–20.5	40%	35–7	25.7
7	LV @ KC	15.0–29.9	89%	0–31	16.1
7	MIA @ CLE	15.8–16.8	52%	6–31	24.1
7	NE @ TEN	23.4–17.1	27%	31–13	11.7
7	NO @ CHI	20.0–24.5	65%	14–26	7.5
7	NYG @ DEN	15.8–24.0	76%	32–33	7.2
7	PHI @ MIN	24.5–21.0	36%	28–22	2.5
7	PIT @ CIN	25.4–19.2	27%	31–33	8.2
7	TB @ DET	24.5–30.2	69%	9–24	9.3
7	WAS @ DAL	27.4–27.6	49%	22–44	21.8
8	BUF @ CAR	26.5–18.4	22%	40–9	22.8
8	CHI @ BAL	22.0–24.1	56%	16–30	11.9
8	CLE @ NE	16.7–25.6	78%	13–32	10.1
8	DAL @ DEN	24.3–26.9	58%	24–44	17.3
8	GB @ PIT	24.9–22.2	39%	35–25	7.3
8	MIA @ ATL	18.5–24.9	71%	34–10	30.3
8	MIN @ LAC	22.5–25.1	58%	10–37	24.5
8	NYG @ PHI	18.8–25.9	73%	20–38	10.9
8	NYJ @ CIN	18.7–24.5	69%	39–38	6.8
8	SF @ HOU	19.9–22.9	59%	15–26	8.0
8	TB @ NO	25.4–21.4	34%	23–3	16.0
8	TEN @ IND	16.7–33.7	92%	14–38	7.1
8	WAS @ KC	19.2–30.0	82%	7–28	10.2
9	ARI @ DAL	25.2–28.1	59%	27–17	13.0
9	ATL @ NE	19.7–26.8	73%	23–24	6.0
9	BAL @ MIA	30.6–22.4	22%	28–6	13.8
9	CAR @ GB	15.8–29.5	87%	16–13	16.7
9	CHI @ CIN	27.7–24.4	37%	47–42	1.7
9	DEN @ HOU	19.2–21.6	58%	18–15	5.5
9	IND @ PIT	28.0–23.8	34%	20–27	11.2
9	JAX @ LV	23.1–20.3	38%	30–29	1.8
9	KC @ BUF	28.8–26.9	42%	21–28	8.8
9	LAC @ TEN	26.9–16.9	18%	27–20	3.0
9	MIN @ DET	20.8–30.7	80%	27–24	12.9
9	NO @ LA	17.3–31.9	89%	10–34	9.4
9	SEA @ WAS	23.7–21.5	40%	38–14	21.8
9	SF @ NYG	25.4–21.7	35%	34–24	6.3
10	ARI @ SEA	19.5–26.5	73%	22–44	14.9
10	ATL @ IND	20.2–28.2	75%	25–31	2.0
10	BAL @ MIN	27.2–21.5	29%	27–19	2.4
10	BUF @ MIA	30.4–20.6	18%	13–30	26.8
10	CLE @ NYJ	19.9–19.0	45%	20–27	7.9
10	DET @ WAS	28.1–19.1	20%	44–22	13.0
10	JAX @ HOU	18.2–17.6	46%	29–36	7.6
10	LA @ SF	27.4–22.5	31%	42–26	11.1
10	LV @ DEN	16.9–26.8	80%	7–10	6.9
10	NE @ TB	23.7–26.6	59%	28–23	7.9
10	NO @ CAR	16.7–22.6	69%	17–7	15.9
10	NYG @ CHI	20.3–25.2	66%	20–24	0.9
10	PHI @ GB	21.7–24.1	57%	10–7	5.4
10	PIT @ LAC	21.6–23.5	55%	10–25	13.1
11	BAL @ CLE	22.5–14.2	21%	23–16	1.4
11	CAR @ ATL	19.1–23.7	65%	30–27	7.7
11	CHI @ MIN	23.9–25.2	53%	19–17	3.3
11	CIN @ PIT	21.9–26.6	65%	12–34	17.3
11	DAL @ LV	26.7–21.5	30%	33–16	11.8
11	DET @ PHI	22.9–24.9	56%	9–16	5.0
11	GB @ NYG	24.4–16.0	21%	27–20	1.5
11	HOU @ TEN	21.6–16.8	32%	16–13	1.8
11	KC @ DEN	23.6–20.7	38%	19–22	6.0
11	LAC @ JAX	22.8–20.8	41%	6–35	31.0
11	NYJ @ NE	16.4–28.2	84%	14–27	1.1
11	SEA @ LA	23.6–27.7	63%	19–21	2.1
11	SF @ ARI	26.1–22.3	35%	41–22	15.3
11	TB @ BUF	20.6–27.1	71%	32–44	5.5
11	WAS @ MIA	22.5–24.3	55%	13–16	1.2
12	ATL @ NO	20.3–22.2	55%	24–10	15.8
12	BUF @ HOU	25.5–19.0	27%	19–23	10.5
12	CAR @ SF	20.2–28.8	77%	9–20	2.3
12	CLE @ LV	17.4–21.6	64%	24–10	18.2
12	IND @ KC	23.0–26.2	60%	20–23	0.2
12	JAX @ ARI	24.3–22.4	42%	27–24	1.1
12	MIN @ GB	17.9–26.1	76%	6–23	8.8
12	NE @ CIN	28.4–21.6	26%	26–20	0.8
12	NYG @ DET	19.1–32.5	87%	27–34	6.4
12	NYJ @ BAL	16.0–30.7	89%	10–23	1.7
12	PHI @ DAL	25.2–22.6	39%	21–24	5.6
12	PIT @ CHI	21.6–25.7	63%	28–31	1.1
12	SEA @ TEN	26.2–13.7	13%	30–24	6.5
12	TB @ LA	21.4–30.1	77%	7–34	18.4
13	ARI @ TB	20.3–24.9	65%	17–20	1.6
13	ATL @ NYJ	20.6–18.0	39%	24–27	5.6
13	BUF @ PIT	24.0–20.8	37%	26–7	15.8
13	CHI @ PHI	20.4–26.5	70%	24–15	15.1
13	CIN @ BAL	24.0–30.4	71%	32–14	24.4
13	DEN @ WAS	23.9–17.9	28%	27–26	5.0
13	GB @ DET	24.3–26.2	56%	31–24	8.9
13	HOU @ IND	20.8–26.6	69%	20–16	9.7
13	JAX @ TEN	23.3–17.3	28%	25–3	16.0
13	KC @ DAL	28.9–25.2	35%	28–31	6.7
13	LA @ CAR	26.9–17.6	19%	28–31	12.3
13	LV @ LAC	16.4–26.6	81%	14–31	6.8
13	MIN @ SEA	15.6–29.2	87%	0–26	12.4
13	NO @ MIA	19.5–25.2	69%	17–21	1.7
13	NYG @ NE	19.7–27.3	74%	15–33	10.4
13	SF @ CLE	20.2–13.9	27%	26–8	11.7
14	CHI @ GB	19.9–27.3	74%	21–28	0.4
14	CIN @ BUF	25.4–30.2	66%	34–39	0.2
14	DAL @ DET	27.2–29.5	57%	30–44	11.7
14	DEN @ LV	24.4–16.2	22%	24–17	1.2
14	HOU @ KC	18.1–24.1	69%	20–10	16.0
14	IND @ JAX	23.6–20.2	36%	19–36	20.4
14	LA @ ARI	29.5–19.8	18%	45–17	18.2
14	MIA @ NYJ	22.6–19.7	38%	34–10	21.1
14	NO @ TB	17.2–25.6	76%	24–20	12.4
14	PHI @ LAC	21.9–20.7	44%	19–22	4.3
14	PIT @ BAL	19.3–25.5	70%	27–22	11.3
14	SEA @ ATL	25.9–20.0	28%	37–9	22.1
14	TEN @ CLE	14.8–20.2	68%	31–29	7.5
14	WAS @ MIN	22.9–20.3	39%	0–31	33.6
15	ARI @ HOU	17.1–27.6	82%	20–40	9.4
15	ATL @ TB	18.6–24.6	69%	29–28	6.9
15	BAL @ CIN	26.9–24.7	40%	24–0	21.8
15	BUF @ NE	27.3–24.0	37%	35–31	0.7
15	CAR @ NO	21.4–20.0	43%	17–20	4.4
15	CLE @ CHI	15.4–25.5	81%	3–31	17.9
15	DET @ LA	24.8–31.1	70%	34–41	0.8
15	GB @ DEN	21.6–20.5	44%	26–34	9.1
15	IND @ SEA	14.2–28.1	88%	16–18	11.9
15	LAC @ KC	17.7–24.5	72%	16–13	9.8
15	LV @ PHI	13.1–27.7	89%	0–31	16.4
15	MIA @ PIT	20.3–23.7	61%	15–28	9.6
15	MIN @ DAL	21.3–27.1	69%	34–26	13.8
15	NYJ @ JAX	13.7–28.0	88%	20–48	13.7
15	TEN @ SF	16.3–30.5	88%	24–37	1.1
15	WAS @ NYG	23.3–23.5	49%	29–21	8.3
16	ATL @ ARI	24.9–22.1	38%	26–19	4.2
16	BUF @ CLE	26.6–15.2	15%	23–20	8.5
16	CIN @ MIA	26.0–21.6	33%	45–21	19.7
16	GB @ CHI	21.9–24.0	56%	16–22	4.0
16	JAX @ DEN	20.5–25.3	66%	34–20	18.8
16	KC @ TEN	19.5–17.7	42%	9–26	18.8
16	LAC @ DAL	24.3–26.9	58%	34–17	19.6
16	LA @ SEA	21.2–22.2	52%	37–38	0.0
16	LV @ HOU	12.7–29.6	92%	21–23	14.9
16	MIN @ NYG	22.8–19.9	38%	16–13	0.1
16	NE @ BAL	22.7–27.5	66%	28–24	8.8
16	NYJ @ NO	16.9–23.8	72%	6–29	16.1
16	PHI @ WAS	24.6–18.5	28%	29–18	5.0
16	PIT @ DET	23.5–30.6	73%	29–24	12.1
16	SF @ IND	25.2–21.2	34%	48–27	17.0
16	TB @ CAR	23.1–20.9	40%	20–23	5.2
17	ARI @ CIN	23.3–29.8	71%	14–37	16.5
17	BAL @ GB	17.3–21.5	64%	41–24	21.3
17	CHI @ SF	24.1–28.6	65%	38–42	0.5
17	DAL @ WAS	29.0–18.8	17%	30–23	3.3
17	DEN @ KC	23.6–10.5	12%	20–13	6.1
17	DET @ MIN	26.8–18.4	21%	10–23	21.4
17	HOU @ LAC	19.6–22.8	60%	20–16	7.2
17	JAX @ IND	25.4–23.4	41%	23–17	4.0
17	LA @ ATL	28.1–20.3	23%	24–27	10.8
17	NE @ NYJ	27.3–15.4	14%	42–10	20.1
17	NO @ TEN	20.9–20.4	47%	34–26	7.6
17	NYG @ LV	22.9–18.7	34%	34–10	19.8
17	PHI @ BUF	21.5–25.7	64%	13–12	5.2
17	PIT @ CLE	20.1–15.1	31%	6–13	12.0
17	SEA @ CAR	23.9–18.9	31%	27–10	12.0
17	TB @ MIA	24.4–19.4	31%	17–20	8.1
18	ARI @ LA	18.4–33.3	89%	20–37	2.1
18	BAL @ PIT	23.1–19.6	36%	24–26	5.5
18	CAR @ TB	20.3–23.3	59%	14–16	1.0
18	CLE @ CIN	19.1–30.4	83%	20–18	13.3
18	DAL @ NYG	27.2–22.8	33%	17–34	21.5
18	DET @ CHI	23.8–28.1	64%	19–16	7.3
18	GB @ MIN	13.8–26.3	86%	3–16	0.5
18	IND @ HOU	14.2–24.1	80%	30–38	1.9
18	KC @ LV	20.2–16.0	33%	12–14	6.2
18	LAC @ DEN	11.1–27.6	91%	3–19	0.5
18	MIA @ NE	16.1–29.8	87%	10–38	14.3
18	NO @ ATL	20.2–25.7	68%	17–19	3.6
18	NYJ @ BUF	13.4–25.6	85%	8–35	14.8
18	SEA @ SF	24.5–23.9	46%	13–3	9.4
18	TEN @ JAX	17.3–30.2	86%	7–41	21.1
18	WAS @ PHI	17.8–19.8	56%	24–17	9.1
19	BUF @ JAX	25.1–27.7	58%	27–24	5.6
19	GB @ CHI	23.6–23.0	46%	27–31	4.7
19	HOU @ PIT	19.8–19.9	49%	30–6	24.1
19	LAC @ NE	21.2–27.8	72%	3–16	6.4
19	LA @ CAR	26.4–17.8	21%	34–31	5.6
19	SF @ PHI	18.8–24.9	70%	23–19	10.1
20	BUF @ DEN	21.9–23.9	56%	30–33	1.1
20	HOU @ NE	18.8–24.4	69%	16–28	6.4
20	LA @ CHI	25.9–23.4	39%	20–17	0.5
20	SF @ SEA	18.6–26.0	74%	6–41	27.5
21	LA @ SEA	21.1–25.7	65%	27–31	0.6
21	NE @ DEN	22.2–20.6	43%	10–7	1.5
22	SEA @ NE	24.4–21.6	38%	29–13	13.1

The 2025 locked test was run once, on 2026-09-24, after the production model was frozen. Later changes (quarterback availability, validation and publication handling in releases) affect live forecasts only; the evaluated model itself is unchanged. Any future evaluation of a revised model on 2025 will be labelled a re-evaluation, not an untouched test.

Evaluation runs: walk-forward bt_20260924T234801Z · locked test locked_20260924T234835Z (code 39e057cbc51b). Full reports: reports/backtest/LATEST.md and reports/locked_test/LATEST.md in the repository.