For each projected starting hitter, the model estimates the probability that the player scores zero runs in the game — i.e. the player goes Under 0.5 runs scored. The card shows that probability as P(zero).
The five feature groups
The probability blends five weighted groups of inputs:
On-base — the hitter's ability to reach base (OBP and recent form).
Lineup context — batting slot and the quality of the hitters behind them who would drive them in.
Pitcher suppression — how well the opposing starter prevents baserunners and runs.
Environment — park factors and game conditions that depress scoring.
Team offense — the overall run-scoring strength of the hitter's lineup.
Calibration
Raw model scores are passed through an isotonic calibration step so the published P(zero) lines up with the true historical rate — a 70% prediction really does cash about 70% of the time.
Edge & the two filters
Edge = model P(zero) − the sportsbook's implied probability, in percentage points. A pick is issued only when the edge clears the configured minimum of 3.0 ptsand the Under price is short enough — under odds at or below -160 — to reflect a genuinely favored outcome.
On this page you can narrow the slate yourself with two independent filters: a minimum-edge slider and an under-odds range (min/max). A side filter and sort round out the controls.
The sufficiency gate
Recommendations stay hidden until the model has proven itself on a live sample: at least 150 graded picks with a live Brier score at or below 0.18. Until then the slate still shows probabilities and lines, but no UNDER call-to-action — and a yellow "Building sample" banner reports progress.
Locking & grading
Predictions update through the day as lineups and lines move, then lock shortly before first pitch. Once the game is final, each pick is graded against whether the player actually scored a run. Wins return profit at the locked odds; losses are −1 unit; voids (scratched / cancelled) are zero.