An Elo-based model with home-field advantage and recent-form weighting. We publish what we'd say before kickoff, then archive what happened — Brier score in public.
Matches where the model disagrees with the market closing line by 3 points or more. Red rail = high-conviction edge.
Rolling accuracy + Brier score from archived predictions. All results checked the day after.
Each team carries an Elo rating updated after every finished match. The match prediction is a logistic map of the Elo difference plus home-field advantage, with a Gaussian draw zone. No ML, no black box — every coefficient is in plain Python, audit it at /opt/sports-ingest/elo_model.py.