South Ark Daily Picks / Methodology
Methodology
Reproducible projections with chronological testing and visible missing data.
SE v3: regularized scoring regression
The score model uses up to eight prior games per team. Features include points scored and allowed, passing and rushing efficiency, yards per play, third-down and red-zone rates, turnovers, takeaways, plays per game, defensive efficiency, recent margins, rest, home/away/neutral site and day/night classification. Opponent profiles enter separately. Metrics shrink toward disclosed baseline priors with two pseudo-games; missingness indicators distinguish absent observations.
Continuous features are standardized on the training set. Ridge penalty 20 is fixed in advance, not tuned on the reported holdout. Predictions are nonnegative. Weekly rolling-origin refits evaluate games from Week 4 onward. Same-kickoff results never enter one another’s feature history. An eight-hour availability proxy is a limitation until verified retrieval times exist.
Probability and uncertainty
With at least 30 prior held-out margin residuals, the model estimates home win probability and an empirical 80% margin interval. Earlier predictions show unavailable rather than invented confidence. Calibration and Brier score use only prior residuals for each test game. These are estimates from a short sample, not a claim of proven calibration.
Weather, injuries and market data
Verified NWS forecasts and official injury reports are attached as context. They receive no numeric adjustment until historical observations support tested coefficients. Travel is omitted without reliable team and venue coordinates. Market gaps are arithmetic differences only; no implied probability, EV or ROI is invented.
Pick publication policy
New automated selections are disabled pending CFBD reconciliation and validation of a selection threshold. Original v1/v2 picks remain frozen and labeled with their original model; no new model overwrites them. Games do not enter the score model without two matched final-score rows. Missing fields stay null.
See measured validation and every pick →