Overview
A full data → model → evaluation → deploy project that predicts win probabilities for NFL games. The narrative spine is to rebuild and extend FiveThirtyEight's retired NFL Elo, then approach it with a gradient-boosted model on EPA-based features.
The honest framing matters: the goal is calibrated probabilities that approach market efficiency, not "beating Vegas." Claiming a real edge over the closing line is a credibility red flag — calibration and transparency are the whole story. Every prediction is graded against the market's own implied probability, in public.
Features
- Walk-forward backtest by season — no shuffling, because temporal leakage is the silent killer of sports models.
- Honest benchmarks — Brier score, log loss, and calibration curves, measured against the devigged moneyline (market-implied probability).
- Self-grading live tracker — a public scoreboard that predicts each slate, auto-resolves results, and shows the running Brier/calibration vs. the market.
- Weekly auto-refresh — a GitHub Action re-runs the full pipeline every Tuesday and redeploys only when a prediction actually changed (a semantic JSON diff, not a timestamp bump).
- Free data only — nflverse play-by-play (EPA) and schedules; no scraping, no paid feeds.
Built with a Python pipeline (data pull, features, models, backtest) that publishes a single JSON to a static Tailwind + Chart.js tracker — no server, no database. The site includes companion pages on the architecture, the anatomy of a single prediction, and the model's honest limitations. Off-season today; it flips from a season-replay stand-in to live predictions at the 2026 kickoff.