Projection Methodology

What the model predicts

For every probable starting pitcher on the day's MLB schedule, PitchProphet projects four quantities: strikeouts (K), outs recorded (from which innings pitched follows), walks (BB) and earned runs (ER). Each is a point estimate — the model's expected value for that start — not a floor, a ceiling, or a prediction of a specific outcome.

How a projection is built

Four layers compose in order. Each is a multiplicative adjustment on the one before, and every weight is clamped so a single layer cannot dominate.

• Layer 1 — Baseline. The pitcher's own rate stats (K/9, BB/9, ERA) and expected innings set a per-start baseline. Live 2026 rates from the MLB Stats API are used once the sample clears 10 innings and 2 starts; below that the calibrated preseason baseline is used instead, because a two-start sample moves a rate stat far more than it informs it.
• Layer 2 — Weather and air density. Temperature, wind speed and wind direction relative to the park's orientation adjust ball flight. The effect is scaled by the pitcher's stuff quality: a pitcher who misses bats is less exposed to batted-ball environment than one who pitches to contact. Domes zero this layer out.
• Layer 3 — Opponent splits. The opposing lineup's strikeout rate, walk rate and offensive index against the starter's handedness. This is why a verified throwing hand matters: the entire layer is selected by it, so a guessed hand silently applies the wrong split.
• Layer 4 — Park factors. A per-venue run and strikeout environment tunes the final K and ER lines.

How the weights are set

The layer weights are not hand-tuned. After each game finalises, the real pitching line is pulled from the box score, every projection is graded, and the residuals feed a bounded ordinary-least-squares fit that nudges the weights toward what the data supports. Two guardrails apply: each weight is clamped to a fixed band around its calibrated baseline, and any weight that drifts to the edge of that band is flagged publicly on /results rather than silently accepted.

The model version increments when the weights change. This document describes version v10.

Coverage: what "verified" means

A projection is verified only when three things hold: the starter resolves to a real MLB player identity, the inputs are that pitcher's own rates rather than league averages, and the throwing hand comes from data rather than a default. Anything short of that is a baseline fallback: still shown, always labelled, and never eligible for Prophet Locks.

We show fallback starters rather than hiding them because omitting a start from the slate is its own kind of dishonesty — but a fallback projection carries no pitcher-specific information beyond the opponent and the park, and should not be read as though it does.

How Prophet Locks are selected

Locks rank verified starters by a composite edge score: the projection's departure from its own baseline on K (weighted 0.45) and ER (0.30), plus bounded adjustments for arsenal quality and opponent weakness. The top ten are captured before first pitch and frozen. A slate is never re-locked after a game has started — rebuilding a settled slate with post-hoc weights and grading it against known results would manufacture a record with hindsight.

Known limits

• The model has no bullpen, no manager tendency and no pitch-count policy input, so it systematically misses starts cut short for reasons unrelated to performance.
• Weather weights are calibrated against physical ball-flight models, not against historical weather-outcome data, and are the least evidenced layer.
• Stuff+ and arsenal values are refreshed periodically rather than live, so they lag in-season changes in a pitcher's repertoire.
• The directional hit rate is not a pure skill score; its baseline shifts with each metric's distribution. Mean absolute error and bias are the honest measures. See /accuracy.

Owner: PitchProphet editorial  |  Last reviewed: 2026-08-16  |  Model version: v10  |  Report a data issue: /contact