NFL Timing Study:
Bayesian Analysis (Patriots)

The Question

Popular belief: the Pats are "better at night." Reality check: is the timing effect statistically credible once we account for confounders?

Data Pipeline

• Sources: Multi-year game logs (opponent, location, score), kickoff time, rest days.
• Scraping: Requests/BeautifulSoup for static pages; Selenium fallback for JS-rendered pages.
• Cleaning: Standardized team names, parsed kickoff as night/day, created features: home, night_game, rest_days, optional elo_diff.
• Storage: CSV + SQLite snapshot per run with idempotent writes and de-duplication.

Model

• Outcome: binary win (1/0).
• Link: logistic regression in a Bayesian framework.
• Predictors: night_game, home, elo_diff (or opponent strength proxy), rest_days.
• Priors: Normal(0, 1) on coefficients; weakly informative intercept.
• Inference: PyMC NUTS; 4 chains, 2k draws; checked r_hat, ESS, and energy diagnostics.
• PPC: Posterior predictive checks to validate calibration.

Results

• Night Game Coefficient: median near zero with wide credible interval overlapping 0 (i.e., no strong evidence of a timing effect once controls are included).
• Home Field: positive and credible, as expected.
• Opponent Strength: meaningful driver of outcome; rest days small/modest effect.

Interpretation: any "prime-time magic" is likely narrative, not signal—at least for the studied seasons.

Takeaways

• Bayesian framing makes the uncertainty explicit; we report credible intervals, not just p-values.
• Scraper + schema-first exports make reruns robust to site changes.
• Template generalizes to other teams and hypotheses with minimal changes.