My NFL pick combos

1. Game model probability

Each leg's raw probability is locked before kickoff. For the NFL, team strength ratings (offensive/defensive EPA, home advantage, rest days, quarterback status) feed a logistic function that outputs win probability.

2. Track-record shrinkage

p′ = 50 + k · (p − 50)

k is fitted by least squares on graded locked model probabilities: k = Σ(p−50)(hit rate−50) / Σ(p−50)². k=1 means displayed values match reality; k<1 means overconfidence, so values shrink toward 50%. k never exceeds 1 (never boosts confidence), and small samples are pulled toward 1 with a 100-pick prior. If a league has no samples, it falls back to the whole league, then the football family.

3. Combo probability

P(combo) = p′₁ × p′₂ × … × p′ₙ

Combos pick one leg per game, and the joint probability is the product of the shrunk leg probabilities. Correlation adjustment stays off until 30 combos are graded.

Data sources

• Schedule, live scores, final results: ESPN public API
• NFL results and EPA: nflverse (api.nfldata.org)
• College football supporting metrics: CFBD
• Grading: model probabilities locked before kickoff are compared against final scores (hit/miss/push)