NFL Prediction System

Honestly-backtested moneyline / spread / total predictions, benchmarked against the closing line.

This is a research/analytics project, not betting advice. Realistic ceilings (see README): moneyline ~66-69% straight-up, ATS ~52-54%, totals near breakeven vs the market. The backtest below shows this build tracks the market rather than beating it -- picks reflect the model's calibrated view, not a demonstrated edge. Predictions are generated pre-kickoff and immutable once written.

This week: Week 1, 2026 season

Snapshot generated at 2026-09-14T02:35:29.483768+00:00 (site rebuilt 2026-09-14 02:36 UTC).

How to read a card: each row is the pick for that market -- the side/total the ensemble (model + de-vigged market) currently favors, with its probability and a confidence bar (green ≥70%, blue 58-70%, gray <58%, essentially a coin flip). "Edge vs market" is how far the ensemble's probability sits from the market's own implied probability -- a large edge is interesting, not proof; the backtest above shows this system does not currently demonstrate a beat-the-market edge on average, so treat a big edge as "the model disagrees with the market," not "the model is right."

DAL @ NYG 2026-09-13

ML DAL 60% edge -0.4pp vs market

SPREAD NYG +3.0 50% edge +0.7pp vs market

TOTAL OVER (47.5) 52% edge +1.3pp vs market

ML: model 53% / market 40% (home=NYG) · Spread: model margin +1.0 (home=NYG) vs market line -3.0 · Total: model 50.8 vs market 47.5

DEN @ KC 2026-09-14

ML KC 53% edge -2.4pp vs market

SPREAD DEN +2.5 53% edge -2.9pp vs market

TOTAL UNDER (43.5) 52% edge -1.3pp vs market

ML: model 44% / market 55% (home=KC) · Spread: model margin -1.3 (home=KC) vs market line +2.5 · Total: model 42.4 vs market 43.5

This week's suggested parlay

Combining picks multiplies risk, not just reward: 2 legs at these probabilities give only a 31.9% chance every leg hits -- each leg's own probability may look reasonable, the combined number is what actually matters for a parlay slip, and it drops fast as legs are added. This is the model's most-confident combination this week, not a recommendation to place it.

#	Game	Market	Pick	Model prob.	Market prob.	Edge
1	DAL @ NYG	ML	DAL	60.3%	60.0%	+0.4pp
2	DEN @ KC	SPREAD	DEN +2.5	52.9%	50.0%	+2.9pp

Combined (2 legs)	Probability	Fair odds (implied, not a real book price)
Model	31.9%	+213
Market-implied	30.0%	+234

One leg per game only (highest-probability market for that game), drawn only across different games -- same-game legs (e.g. a team's moneyline and its own spread) aren't independent enough for the combined-probability math to mean anything, so they're never combined here.

Moneyline backtest (walk-forward, leak-free)

Model	N	Accuracy	Log loss	Brier	ECE
elo_only (raw)	4162	0.6425	0.6321	0.2211	0.0293
elo_only (isotonic-calibrated)	4162	0.6430	0.6633	0.2216	0.0188
logistic_regression (n=4162)	4162	0.6470	0.6262	0.2183	0.0208
xgboost (nested-CV tuned, n=4162)	4162	0.6497	0.6523	0.2196	0.0210
market_implied (n=4161)	4161	0.6655	0.6094	0.2110	0.0165
home_always (baseline)	4162	0.5543	0.6877	0.2473	0.0168
favorite_always (baseline, n=4161)	4161	0.6657
ensemble (logistic+market, n=4162)	4162	0.6648	0.6093	0.2109	0.0146

Spread & total backtest (walk-forward, leak-free)

Market	N	Accuracy	Model MAE	Market MAE	ECE
Spread (ATS), model only	4069	0.5072	10.33 pts	10.08 pts	0.0599
Spread (ATS), ensemble	4069	0.5139	—	—	0.0042
Total (O/U), model only	4134	0.5160	10.61 pts	10.45 pts	0.0458
Total (O/U), ensemble	4134	0.5104	—	—	0.0205

Baselines -- ATS: home-always 0.491, favorite-always 0.489. O/U: always-over 0.496, always-under 0.504. "Ensemble" blends the model with de-vigged market juice (spread/total odds) on the log-odds scale, weights learned walk-forward, same discipline as the moneyline ensemble.

Calibration & home-field advantage

On CLV (closing-line value)

Historical nflverse data provides closing lines only, not opening lines, so true historical CLV (did we get a better price than where the market closed) cannot be reconstructed retroactively. What the walk-forward backtest above measures instead is whether the model's calibrated probabilities beat the market's closing-line-implied probabilities on log loss / Brier / accuracy -- and, as reported above, they generally do not by a meaningful margin. Starting this week, every prediction snapshot is timestamped pre-kickoff; as the season progresses we will append realized closing lines to results_log.csv and publish a genuine prospective CLV chart here.