NFL Lab

Experimental models running beside NFL-v1.1. Each kind of evidence is kept separate, because they are not worth the same.

EXPERIMENTAL · test models only — not official forecasts, not betting advice Details

Nothing on this page is a production forecast, and nothing here is a betting recommendation — no side is being advised. NFL-v1.1 remains the only official model; Current, History and Performance are unaffected by anything shown here. Promotion requires a human decision.

Finalists were chosen from 2704 configurations across 3 recency weightings (8112 walk-forward evaluations) on seasons [2018, 2019, 2020, 2021, 2022, 2023]. Seasons [2024, 2025] were held back and read once, after the finalists were fixed. Agreement between development and holdout ranking across the top 30 candidates: +0.30 (1.00 would mean the search generalised perfectly, 0.00 that it did not generalise at all).

Challenger A — Market-AnchoredPROMISING

Starts from the market line and adds independent football information on top: the starting quarterback's career efficiency and snap-weighted personnel availability. Notably, the fit chose to drop NFL-v1.1's own disagreement with the line entirely — on 2018-2023 that disagreement carried no information the line had not already priced (beta +0.04, t +0.65).

Status: Beats NFL-v1.1 on the untouched holdout for both probability accuracy and margin error. Live sample so far: 311 decided.

HOLDOUT

UNTOUCHED HISTORICAL HOLDOUT n=543
probability accuracy 0.207 — better than NFL-v1.1 by 0.038 · average margin miss 9.7 pts · 68.3% of picks correctSeasons deliberately set aside and read once, after the finalists were already chosen. The strongest historical evidence available here.

WALK-FORWARD

HISTORICAL WALK-FORWARD n=1067
probability accuracy 0.212 — better than NFL-v1.1 by 0.025 · average margin miss 9.9 pts · 66.9% of picks correctRefit on seasons before each test season, then scored on it. Genuine out-of-sample — but this candidate was chosen from many, so some of its edge may be luck in the search.

SHADOW

2026 LIVE SHADOW n=311
probability accuracy 0.242 · 176 of 311 correct (56.6%)Forecasts written before kickoff, on games nobody had seen. No selection risk at all — and the smallest sample.

USES

qb_epa_delta, avail_deltaExperiment tournament_A_unconstrained

Challenger A — Market-Anchored — feature evidence classes and fitted parameters

PIT_SAFE: reconstructible from data timestamped before kickoff. OPTIMISTIC: the value is clean but the fact it applied to this game is recorded after it. UPPER_BOUND: historically an observation where a live forecast would only have a forecast.

qb_epa_delta: OPTIMISTIC
avail_delta: PIT_SAFE

Parameter	Value
n_train	2095
fit_seasons	[2018, 2019, 2020, 2021, 2022, 2023, 2024, 2025]
mode	anchored
margin	{'terms': ['intercept', 'qb_epa_delta', 'avail_delta'], 'coefs': [-0.13314808261893504, 1.8907446585546934, 0.3605516418295857]}
total	{'terms': ['intercept', 'wind', 'cold', 'avail_total', 'qb_epa_delta_abs'], 'coefs': [0.5765538519859024, -0.15367635584114925, -0.041595801727398164, 0.5936126731121204, -2.603946894792646]}

Challenger B — Football+PROMISING

NFL-v1.1's own projection plus three things it ignores: starting quarterback quality, whether the starter changed, and snap-weighted availability lost to injury. No market information at all, so it is a straight test of whether better football information improves the football model.

Status: Beats NFL-v1.1 on the untouched holdout for both probability accuracy and margin error. Live sample so far: 311 decided.

HOLDOUT

UNTOUCHED HISTORICAL HOLDOUT n=543
probability accuracy 0.235 — better than NFL-v1.1 by 0.011 · average margin miss 10.9 pts · 60.4% of picks correctSeasons deliberately set aside and read once, after the finalists were already chosen. The strongest historical evidence available here.

WALK-FORWARD

HISTORICAL WALK-FORWARD n=1067
probability accuracy 0.227 — better than NFL-v1.1 by 0.010 · average margin miss 10.4 pts · 63.1% of picks correctRefit on seasons before each test season, then scored on it. Genuine out-of-sample — but this candidate was chosen from many, so some of its edge may be luck in the search.

SHADOW

2026 LIVE SHADOW n=311
probability accuracy 0.246 · 180 of 311 correct (57.9%)Forecasts written before kickoff, on games nobody had seen. No selection risk at all — and the smallest sample.

USES

qb_epa_delta, qb_change_delta, avail_deltaExperiment tournament_B_football_plus

Challenger B — Football+ — feature evidence classes and fitted parameters

qb_epa_delta: OPTIMISTIC
qb_change_delta: OPTIMISTIC
avail_delta: PIT_SAFE

Parameter	Value
n_train	2095
fit_seasons	[2018, 2019, 2020, 2021, 2022, 2023, 2024, 2025]
mode	residual
football_margin	{'features': ['qb_epa_delta', 'qb_change_delta', 'avail_delta'], 'coefs': [12.731895832600213, -3.048535943750645, 1.1500630439545154], 'ridge': 0.1, 'n': 2095}
football_total	{'features': ['wind', 'cold', 'avail_total', 'qb_epa_delta_abs'], 'coefs': [-0.17486154379031396, -0.17515295371546152, 0.5995953150108153, -1.6817245582812814], 'ridge': 0.1, 'n': 2095}
market_margin	{}
market_total	{}
sd_scale	1.2600000000000005

ProductionNFL-v1.1 — frozen, unchanged by anything on this page

ChallengersChallenger A — Market-Anchored, Challenger B — Football+ — frozen parameters, shadow only

Shadow record10716 rows from 2026-09-16 15:21 UTC weeks 2–9

Decided311 predictions per model (moneyline 104, spread 102, total 105)

Enough to judge?Past the 30-prediction floor this page uses before it reads anything into a live difference

2026 live shadow scoreboard

Same games, same frozen pregame data, one official pregame forecast each. Overall first, then by market — a model can lead overall purely because moneyline probabilities are the easiest of the three.

Model	n	Correct	Probability accuracy	Log loss	Claimed / actual
NFL-v1.1 (production)	311	173 of 311 (55.6%)	0.2513	0.6970	52.4% / 56.6%
Challenger A — Market-Anchored	311	176 of 311 (56.6%)	0.2420	0.6773	56.7% / 56.6%
Market baseline	311	172 of 311 (55.3%)	0.2431	0.6796	55.2% / 56.6%
NFL-v1.1 (production)	311	173 of 311 (55.6%)	0.2513	0.6970	59.0% / 57.9%
Challenger B — Football+	311	180 of 311 (57.9%)	0.2461	0.6859	59.1% / 57.9%
Market baseline	311	170 of 311 (54.7%)	0.2431	0.6796	54.3% / 57.9%

By market — moneyline, spread and total scored separately

Market	Model	n	Correct	Probability accuracy	Claimed / actual
Moneyline	NFL-v1.1 (production)	104	62 of 104 (59.6%)	0.2394	57.9% / 62.5%
Moneyline	Challenger A — Market-Anchored	104	65 of 104 (62.5%)	0.2270	64.1% / 62.5%
Moneyline	Market baseline	104	65 of 104 (62.5%)	0.2291	65.0% / 62.5%
Spread	NFL-v1.1 (production)	102	52 of 102 (51.0%)	0.2571	49.2% / 55.9%
Spread	Challenger A — Market-Anchored	102	57 of 102 (55.9%)	0.2477	52.1% / 55.9%
Spread	Market baseline	102	50 of 102 (49.0%)	0.2506	50.3% / 55.9%
Total	NFL-v1.1 (production)	105	59 of 105 (56.2%)	0.2575	50.0% / 51.4%
Total	Challenger A — Market-Anchored	105	54 of 105 (51.4%)	0.2512	53.8% / 51.4%
Total	Market baseline	105	57 of 105 (54.3%)	0.2496	50.1% / 51.4%
Moneyline	NFL-v1.1 (production)	104	62 of 104 (59.6%)	0.2394	60.7% / 64.4%
Moneyline	Challenger B — Football+	104	67 of 104 (64.4%)	0.2311	61.6% / 64.4%
Moneyline	Market baseline	104	65 of 104 (62.5%)	0.2291	62.5% / 64.4%
Spread	NFL-v1.1 (production)	102	52 of 102 (51.0%)	0.2571	56.3% / 52.9%
Spread	Challenger B — Football+	102	54 of 102 (52.9%)	0.2530	56.8% / 52.9%
Spread	Market baseline	102	49 of 102 (48.0%)	0.2506	50.2% / 52.9%
Total	NFL-v1.1 (production)	105	59 of 105 (56.2%)	0.2575	59.9% / 56.2%
Total	Challenger B — Football+	105	59 of 105 (56.2%)	0.2543	58.7% / 56.2%
Total	Market baseline	105	56 of 105 (53.3%)	0.2496	50.1% / 56.2%

Each slice is a third of an already small sample. THIN marks any cell below 30 decided predictions, where a difference of a few games moves every number on the row.

This week — how each model sees the same games

Unresolved forecasts, side by side. This is not a recommendation and no side is being advised.

Game	Market	Prediction	NFL-v1.1	Challenger A — Market-Anchored	Challenger B — Football+	Market
ARI @ SEA	Moneyline	ARI ML	32.9%	25.4%	24.3%	21.4%
ARI @ SEA	Spread	ARI +8.5	60.6%	51.9%	44.8%	50.2%
ARI @ SEA	Total	UNDER 39.5	39.8%	48.6%	42.2%	49.6%
BUF @ MIN	Moneyline	BUF ML	60.9%	60.2%	63.4%	60.9%
BUF @ MIN	Spread	BUF -3	52.3%	55.2%	59.4%	51.1%
BUF @ MIN	Total	UNDER 44.5	58.1%	50.5%	56.8%	50.0%
CIN @ ATL	Moneyline	CIN ML	50.9%	59.7%	52.8%	60.2%
CIN @ ATL	Spread	CIN -3	41.7%	54.7%	48.8%	50.4%
CIN @ ATL	Total	UNDER 52.5	63.4%	50.5%	61.1%	50.2%
CLE @ NO	Moneyline	CLE ML	38.5%	40.9%	37.8%	40.3%
CLE @ NO	Spread	CLE +3	47.0%	53.3%	47.7%	49.0%
CLE @ NO	Total	UNDER 42.5	74.5%	48.1%	69.6%	50.0%
DAL @ IND	Moneyline	DAL ML	40.8%	67.1%	46.5%	67.4%
DAL @ IND	Spread	DAL -5.5	25.7%	50.6%	33.6%	49.2%
DAL @ IND	Total	UNDER 51.5	40.6%	51.1%	43.2%	50.0%
DEN @ CAR	Moneyline	DEN ML	57.3%	58.3%	57.9%	57.4%
DEN @ CAR	Spread	DEN -2.5	52.3%	53.2%	53.9%	51.9%
DEN @ CAR	Total	UNDER 44.0	51.3%	51.2%	52.6%	49.8%
DET @ MIA	Moneyline	DET ML	56.1%	71.2%	59.3%	74.8%
DET @ MIA	Spread	DET -7	32.8%	51.8%	43.2%	50.0%
DET @ MIA	Total	UNDER 50.5	44.4%	49.6%	46.3%	50.2%
GB @ NE	Moneyline	GB ML	37.6%	41.6%	40.8%	40.0%
GB @ NE	Spread	GB +3	46.0%	54.0%	50.8%	49.2%
GB @ NE	Total	UNDER 44.5	36.7%	49.4%	39.4%	50.0%
HOU @ LAC	Moneyline	HOU ML	51.0%	66.6%	51.8%	67.4%
HOU @ LAC	Spread	HOU -5.5	34.5%	50.0%	38.5%	49.6%
HOU @ LAC	Total	UNDER 40.5	69.2%	47.9%	64.8%	49.8%
JAX @ BAL	Moneyline	JAX ML	53.8%	78.8%	68.0%	80.2%
JAX @ BAL	Spread	JAX -9.5	26.1%	52.4%	45.2%	50.2%
JAX @ BAL	Total	UNDER 39.5	34.0%	43.6%	34.2%	48.9%
LA @ WAS	Moneyline	LA ML	61.4%	64.1%	61.2%	66.6%
LA @ WAS	Spread	LA -4.5	46.4%	49.2%	49.3%	51.1%
LA @ WAS	Total	UNDER 48.5	41.7%	49.7%	43.8%	51.1%
LV @ SF	Moneyline	LV ML	14.9%	23.3%	17.4%	20.0%
LV @ SF	Spread	LV +9.5	38.1%	50.4%	36.3%	49.8%
LV @ SF	Total	UNDER 51.5	66.9%	51.4%	64.2%	48.9%
NYG @ PHI	Moneyline	NYG ML	45.2%	58.4%	48.5%	60.1%
NYG @ PHI	Spread	NYG -3	35.9%	53.4%	44.6%	51.5%
NYG @ PHI	Total	UNDER 38.5	28.8%	48.8%	33.1%	49.6%
NYJ @ KC	Moneyline	NYJ ML	29.3%	22.2%	29.4%	18.8%
NYJ @ KC	Spread	NYJ +10	60.2%	53.2%	55.2%	50.4%
NYJ @ KC	Total	UNDER 47.5	56.6%	50.0%	55.7%	50.4%
TB @ CHI	Moneyline	TB ML	46.1%	33.1%	48.2%	29.6%
TB @ CHI	Spread	TB +6	64.9%	53.1%	64.1%	49.8%
TB @ CHI	Total	UNDER 47.5	51.7%	49.1%	51.5%	50.4%

What the evidence labels mean

HISTORICAL TRAINING — The model was fitted on these games. It proves nothing about the future and is shown only for completeness.
HISTORICAL WALK-FORWARD — Refit on seasons before each test season, then scored on it. Genuine out-of-sample — but this candidate was chosen from many, so some of its edge may be luck in the search.
UNTOUCHED HISTORICAL HOLDOUT — Seasons deliberately set aside and read once, after the finalists were already chosen. The strongest historical evidence available here.
2026 RETROSPECTIVE — Applied after the fact to 2026 games. Useful for sanity, but the outcome was already known when the calculation was run.
2026 LIVE SHADOW — Forecasts written before kickoff, on games nobody had seen. No selection risk at all — and the smallest sample.

These are never combined into a single figure. A number that pooled a training score with a live one would be the most misleading thing this system could publish.

No model on this page can alter a production forecast, enter the official History record, or affect grading. There is no automatic promotion anywhere in this system: promotion means a human editing the model version and coefficients in a commit.
Generated 2026-11-05 23:48 UTC.