NFL Predictor

A neural-network ensemble that predicts NFL win probabilities using only information available before kickoff. It retrains itself weekly, freezes every prediction at kickoff, and publishes picks, a live track record, power rankings and playoff odds to the site. Vegas lines are a benchmark, never an input.

Results

Walk-forward backtest: each season is predicted by models trained only on earlier seasons. All tuning used 2018–21; the 2022–26 games were scored once, after the lineup was locked.

Held-out 2022–26 (1,218 games)	Accuracy	Log-loss
Ensemble (this project)	66.2%	0.624
MLP neural net	66.8%	0.625
Vegas moneyline	67.4%	0.610
Elo rating	63.9%	0.635
Always pick home team	55.0%	0.688

Confidence is calibrated: held-out picks made at 75%+ confidence won about 81% of the time. Every 2026 prediction since launch is also logged, frozen at kickoff, in docs/data/live_log.csv.

How it works

flowchart LR
  A[nflverse<br/>play-by-play, rosters,<br/>snaps, injuries] --> B[Features ~250/game<br/>opponent-adjusted EPA,<br/>ridge ratings, Elo, QB,<br/>pregame roster strength]
  B --> C[Models<br/>MLP + Siamese nets,<br/>logistic, ridge, GBM]
  C --> D[Ensemble<br/>log-odds average]
  D --> E[Freeze at kickoff<br/>live log]
  E --> F[Site<br/>picks, explanations,<br/>rankings, playoff odds]

• Features:
  • Short- and long-window form with opponent adjustment, ridge power ratings and Elo.
  • A shrunk QB EPA rating that carries across teams and seasons, plus QB-change flags.
  • Roster strength from player snap share and draft capital over the pregame active roster, with missing starters counted against the team.
  • Rest, travel and body clock.
• Models:
  • An MLP with a linear skip path, and a Siamese matchup network with a shared team encoder. Both are multi-task (win logit + point margin), heavily L2-regularized and 8-seed ensembles.
  • Logistic regression, ridge-margin and gradient boosting complete the ensemble.
  • Every game is trained from both sides, so predictions are symmetric.
• Explanations: the linear member splits each prediction exactly into grouped feature contributions ("Grounds" on the site).
• Ratings and playoff odds:
  • A Kalman filter gives each team a rating with uncertainty.
  • 10,000 Monte Carlo season simulations give playoff, division and #1-seed odds.

Keeping it honest

• Pregame features only. Every feature is computed as of kickoff. tests/test_leakage.py checks this on a synthetic league: features for a week must be identical whether or not later results exist. Those tests caught (and fixed) a subtle leak of whole-dataset league averages into early-season priors; see REPORT.md.
• Same roster view in testing and live. A player is available if he's on the 53-man roster and not listed Out or Doubtful on Friday. Game-day inactives were never used in training.
• Frozen predictions. Live predictions train only on earlier weeks and freeze at kickoff (tests/test_freeze.py).
• Measured experiments. Many ideas didn't make the cut, and LOOP_LOG.md records each with its numbers: weather, coaching, kicking, Kalman features, QB age, calibration, stacking, a GRU sequence model and per-player production ratings.

Run it

pip install torch --index-url https://download.pytorch.org/whl/cpu
pip install -r requirements.txt
python src/download_data.py         # ~450 MB from nflverse
python src/run_weekly.py            # rebuild features, retrain, predict the next week
python src/build_site.py            # rebuild docs/ (site, live log, rankings, playoff odds)
pytest                              # leakage, symmetry and freeze tests

Backtests: python src/evaluate.py --phase dev|test, then python src/export_results.py.

Code map (src/)

File	Role
download_data.py	nflverse schedules, play-by-play, weekly rosters, snap counts, depth charts, injuries, player IDs
team_games.py	Per-team per-game stats from play-by-play, plus QB game logs
players.py	Player ratings and pregame roster strength by unit, including missing regulars
build_features.py	The game-level feature table (form, ratings, Elo, QB, stakes, rest/travel, Vegas benchmark)
kalman.py	State-space team ratings with uncertainty
dataset.py, models.py, config.py	Symmetric matchup matrices, the five models, the final lineup
backtest.py, evaluate.py, export_results.py	Walk-forward evaluation and site result exports
experiments.py, exp_features.py, exp_calibration.py	Experiment grids and A/B harness
predict.py, explain.py, run_weekly.py	Weekly predictions, frozen at kickoff, with explanations
live_log.py, playoffs.py, build_site.py, pick_sheet.py	Live record, playoff simulation, site and pick-sheet builders

More: REPORT.md (findings), MODEL_CARD.md, LOOP_LOG.md (experiment log).