The model. Each team gets a power rating equal to its scoring differential per game, shrunk toward the league average early in a season. A game's projected margin is (home rating − away rating) + home advantage, converted to a win probability via a normal distribution calibrated to each sport's typical scoring variance. The pick is simply whichever side that probability favors; the confidence tier (Toss-up / Lean / Moderate / Strong) reflects how lopsided the probability is.
Horse racing uses a different, market-based model: morning-line odds are converted to implied win probabilities and the betting overround is removed by normalizing across the full field. The pick is the horse with the highest resulting probability; confidence reflects how far above an even share of the field (1/N horses) that probability sits. This model doesn't yet incorporate speed figures, trainer/jockey stats, or track bias — see "what's next" below.
Tracking & the accuracy tracker. Every pick this page makes is logged with its date, probability, and confidence tier before the contest happens. Once results are known, each pick is graded correct/incorrect and folded into the Accuracy Tracker — overall hit rate, per-league accuracy, calibration by confidence tier, and home/away and favorite/underdog splits. Nothing is regraded or removed after the fact.
How the model "learns." On a regular cadence, each sport's home-advantage and variance constants are nudged (in small, bounded steps) toward whatever values would have minimized error on that sport's own graded picks so far — a lightweight self-calibration loop, not a black-box neural net. Changes require a minimum sample size per sport before they're allowed to move, so early results won't cause wild swings.
Known limitations: off-season leagues show last season's ratings with no games to project; NFL preseason uses prior-season ratings regressed toward average; NCAA Football is scoped to the AP preseason Top 25, and unranked opponents are treated as replacement-level (likely overstating ranked teams' edge against a strong-but-unranked opponent); horse racing is currently limited to major graded stakes races where morning-line odds are publicly available. This is a transparent, continuously-graded model, not a guaranteed or professional handicapping service.