Weekly NFL predictions, built from five seasons of play-by-play.

A logistic-regression / gradient-boosting / random-forest ensemble trained on rolling team and QB form, injuries, and situational context — bagged over block-bootstrap resamples so every prediction carries an uncertainty interval, not just a point estimate.

Current slate

This week's games

Each matchup with the model's predicted win probability and the de-vigged market probability, for both teams.

Loading…

Browse

Other weeks

Pipeline

How a pick is made

Features, strictly lagged

Every in-game statistic enters only as a rolling average of earlier games. Four independent guards assert this, and they have caught real bugs.

Bagged prediction with an interval

500 block-bootstrap resamples over season-week blocks give each game a probability and a 95% interval — roughly 0.16 wide.

De-vig the market

Shin's method removes the book's 3.7% margin. Comparing against a raw implied probability would invent edge on every bet.

Require the interval to clear it

One-sided: the whole interval must sit above the fair price. A point estimate beating the market is never enough.

Walk-forward cross-validation, time-based split, 2025 held out and touched once. No random k-fold anywhere.