Matchday

A personal analytics hub for college football and men's college basketball.

Live at matchdayterminal.com · by @MatchdayTRM

What this is

This is one person's analytical playground, published in the open. It exists so the method and the record can be checked by anyone — not as a service, not as a tipster, and not as advice. Everything on it is my own opinion and my own work.

It covers two sports and nothing else: college football and men's college basketball. Coverage narrowed to those deliberately. They share a data source, a season calendar that runs almost year-round between them, and a field of several hundred programmes that is genuinely hard to price.

Where the numbers come from

Matchday does not compute ratings. The power rating, the model probabilities and the rest of the modelling come from a separate, private engine. This repository reads its published output and renders it; it never recalculates a number.

• The power rating is an edition with a date on it, shown as published rather than recomputed, with strength of schedule beside every rating.
• My Top 25 is my own ballot, ranked on each team's résumé. It is an opinion, kept apart from the power rating.
• Match forecasts are live shadow forecasts: they keep moving until kickoff, carry no pregame lock receipt, are not official picks, and are not graded.

After new engine output lands, rebuild the site's snapshot:

python build_cfb_snapshot.py

Rules this project holds itself to

• Projections are labelled. Brackets and power rating editions say "projected" until real selection and championship results exist. A power rating published before a season starts describes the previous completed season, and says so in its own note.
• Nothing is rewritten. A settled score is never revised; a graded pick is never rescored. archive/games/ is the raw record and refuses revisions.
• Only locked picks are graded. The graded record comes from the engine's handoff (trdlab_picks.json → scorecard); a pick shown after its lock can never be graded as though it had been an official call.
• Model-market disagreement is reported, not endorsed. On graded college samples a wider gap predicted worse results — the sign is inverted — so nothing here ranks, filters or stakes on it.

The board

Upcoming games · Results · Scorecard · Bracket · Conferences (power rating and My Top 25) · Community.

The scorecard is deliberately two numbers, picks won and picks lost. Brier, log loss, calibration and closing-line value are still computed by the research modules; they are not on that page because a reader asking "is it any good" wants a record.

Running it

python fetch_data.py      # refresh provider data
python -m http.server 8743   # serve locally
python -m unittest discover -p "test_*.py"

Provider keys live in config_keys.py, which is gitignored. See docs/SETUP.md.

Notes

AGENTS.md documents the repository's working rules. docs/PROVIDER_COMPLIANCE.md records data licensing decisions. legal.html carries the privacy policy, terms and data-source credits.

Backup plan notes

Preserve the engine, fitted weights, calibration and manual ballots; recollect public fixtures and history after a damaged or deleted database. See BACKUP-PLAN-NOTES.md for the recovery checklist and the engine-only backup command. Provider credentials are restored separately.