College football forecasting model

A model that predicts college football point spreads and tests them against sportsbook lines, with an honest track record.

Site: https://mattrudy13.github.io/cfb-picks/

At a glance

• What it does: rates every FBS team from game results and play-by-play efficiency (Elo, an opponent-adjusted margin, EPA and success rate), turns the ratings into a predicted spread for each game, and compares it with the betting market. Picks are saved before kickoff and graded automatically.
• How it was tested: each version's pass/fail bar was written down before its backtest ran (SPEC.md), and each backtest season was predicted only from earlier seasons.
  • v1: 52.1% against the opening line on 266 picks (the bar is 52.4%). Missed.
  • v2: 52.8% on 863 picks, but profitable in only 2 of 4 seasons (the bar is 3). Missed.
  • v3: 55.3% on 172 picks over 2022–2025, profitable in 3 of 4 seasons. Passed, narrowly. Against the closing line it was 49.1%, so its live picks start as provisional.
• Live record (as of Oct 21, 2026): v3 is 6–10 on its first 16 picks (week 7), with 15 week 8 picks still to play. Picks become official only if, over 100+ live picks, lines move toward them by the close by more than luck would explain.
• Built with: JavaScript (Node 22), GitHub Actions on a schedule, GitHub Pages, and the CollegeFootballData and The Odds API data services. No server and no dependencies.

Details

Status: model v3 passed its pre-registered backtest bar narrowly (55.3% at the opener on 172 picks, 2022–2025, but 49.1% at the close; see SPEC.md and the Backtest page), so its spread picks are provisional. They become official only if, over 100+ live picks, lines move toward them by the close by more than luck would explain (95% confidence). Earlier models (v1, v2) missed the bar; v2's week 7 picks are kept as a separate record. Underdog moneylines were tracked through October 11, 2026, then stopped.

• Best Bets (index.html): this week's picks and the record (win rate, ROI, line move).
• Board (board.html): every FBS game with the model's line and win chance next to the market's.
• Ratings (ratings.html): power ratings for every FBS team, plus the Elo, scoring, EPA and success-rate ratings behind them; SP+ and FPI for comparison.
• Team (team.html?team=<school>): a team's schedule, model and market lines, results and rating over time.
• Backtest (backtest.html): how both models did on 2022–2025 and the go-live checklist.

How it works

No server and no build step: GitHub Actions write JSON into data/, and Pages serves the HTML plus that JSON.

update-data.yml    Sat noon–Sun 3 AM every 3 h, Sun/Mon/Wed 6 AM ET   build-data → grade-picks
snapshot-odds.yml  Sun 9 PM, Tue + Fri 10 AM, Sat 9 AM ET             build-data → fetch-odds → record-picks
check-assets.yml   on push                                           stamp-assets --check

File	What it is
model.js	Ratings (Elo, opponent-adjusted margin, EPA and success rate), fitted preseason priors, blend. Shared by the scripts
common.js, style.css	Browser helpers and styles for every page
scripts/backtest.mjs, scripts/backtest-v2.mjs, scripts/backtest-v3.mjs	The v1, v2 and v3 backtests. v2 and v3 also write their model's settings (data/backtest-v*.json → live); build-data runs v3's if it passed its bar, else v2's
scripts/build-data.mjs	Fetches this season from CFBD and writes data/<season>.json; also caches closing lines and weekly SP+/FPI snapshots in data/raw/
scripts/fetch-odds.mjs	The Odds API snapshot (US-licensed books only) → data/odds-<season>.json
scripts/record-picks.mjs, scripts/grade-picks.mjs	Maintain data/picks-<season>.json; grading also measures each pick's line move to CFBD's closing line
scripts/fetch-history.mjs	One-time CFBD history into data/raw/ (skips cached files)
scripts/build-ledger.mjs	Research: every FBS game's model and book lines at each snapshot (from git history), CFBD open/close and result → data/ledger-<season>.json (gitignored)

Secrets: CFBD_API_KEY (CollegeFootballData, free tier shared with h2h) and ODDS_API_KEY (this project's own key; te-tracker has a separate one). Set with gh secret set <NAME> -R mattrudy13/cfb-picks.

Running locally

node scripts/build-data.mjs --offline   # rebuild from data/raw/ (no API calls)
python3 -m http.server 8000             # then open http://localhost:8000
node scripts/backtest-v3.mjs            # rerun the live model's backtest (changes the live model's settings)

After editing common.js or style.css, run node scripts/stamp-assets.mjs before committing. Test offline odds with node scripts/fetch-odds.mjs --fixture <file> in a scratch copy of the repo, never in data/.