How the Model Works

SharpSDA uses a machine learning model to estimate each team's win probability for every MLB game. Here is what goes into it — without revealing the specific features, weights, or exact methodology.

Input categories

The model looks at several categories of game-day information:

• Starting pitcher performance — recent and season-to-date metrics measuring how effectively each starter retires batters, generates weak contact, and limits runs.
• Team batting — offensive production relative to league average, adjusted for the matchup (left-handed vs. right-handed pitcher splits).
• Bullpen — recent relief pitching quality for both teams.
• Park factors— each stadium's historical effect on run scoring and outcomes.
• Weather — temperature and wind conditions at first pitch.
• Rest and schedule — days of rest, travel, and back-to-back game flags for each team.
• Market signals — live line movement and sharp-money indicators captured at predict time, used as supplemental context (not as primary features in the base model).

Calibration

The model's raw output is passed through a calibration layer that aligns predicted probabilities with real historical frequencies. This is what allows the probabilities to be used directly for bet sizing via the Kelly Criterion.

Quality gates

Every model version must pass accuracy and calibration benchmarks on an out-of-sample holdout before any picks can be published from it. A human reviews and approves each new version before it goes live.

What the model does not do

• It does not incorporate injury reports in real time (coming in a future update).
• It does not use public betting percentages in the base prediction (only as a supplemental market signal).
• It does not guarantee any outcome — baseball has irreducible randomness.

This page describes categories of inputs only. Specific features, exact formulas, training windows, and model weights are not disclosed. This page should be reviewed by a human editor before publish.