Accuracy & Selection Universe

What gets graded

This is the part most accuracy claims leave out, so it is stated precisely.

• The universe is every projection the engine logged, not a selection. Projections are snapshotted automatically for every confirmed starter on the slate, before first pitch. There is no step at which a projection can be withdrawn because it looks likely to miss.
• Grading is automatic and from the box score. When a game goes final, the pitcher's real line is pulled from the MLB Stats API and compared to the stored projection. No manual entry is involved in the published record.
• A start is voided only if the pitcher did not start it. Scratches and rotation changes are marked void and excluded — they are not misses, because the projection was for a start that never happened. Voids are visible in the ledger rather than deleted.
• Nothing is excluded for being wrong. Blowups are in the record. So are the fallback-coverage starts, which is why the verified/fallback split is published above rather than quietly netted out.

The current record

Of 286 graded starts, 274 ran on verified pitcher inputs and 12 on league-average fallbacks. The full per-metric breakdown, including bias and the weight-calibration health check, is at /results.

How to read the numbers

Mean absolute error (MAE) is the average miss, in the metric's own units — the honest measure of how close the model gets. Bias is the average signed error; a positive bias means the model systematically projects too high, which is a fixable calibration problem rather than noise.

The directional hit rate — how often the actual result beat the projection in the favourable direction — is published because it is the number people ask for, but it is the weakest of the three. Its baseline moves with each metric's distribution shape, so 52% on strikeouts and 52% on walks do not mean the same thing. Exact ties are treated as pushes and excluded, so projections landing on round numbers cannot pad the rate. Judge the model by MAE and bias.

What this record is not

It is not a betting record. The model is graded against its own projection, never against a sportsbook line, so a "win" here means the model's estimate was directionally right — not that a wager would have profited. No closing-line value, vig or market comparison is implied anywhere on this site.

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