CORE Actual — research

What the tournaments actually established

These estimators were selected by sealed-holdout tournaments. Those comparisons now carry paired bootstrap confidence intervals, and the intervals change the reading — the phrase tournament-validated has been removed from the banner above because it overstated two of the three results.

Four parallel partial-pooling claims are published across this project, so the family column applies a Bonferroni correction for a 5% family-wise error rate.

B-CORE — batting runs above average per 600 PA (min 400 PA)

P-CORE starters — runs prevented above average per 750 BF (min 400 BF)

P-CORE relievers — runs prevented above average per 250 BF (min 150 BF)

“Fitted PA” and “Fitted BF” are the opportunities this metric was actually fitted on, which is fewer than the season totals shown on Season. Aaron Judge has 679 plate appearances there and 655 here. The difference is the censored classes the run-expectancy model excludes — bottom of the ninth, extra innings under the automatic-runner rule, and truncated final half-innings — which the counting stats include and this metric never sees. Two different questions, two different denominators; neither is wrong, and until now the page called them both “PA”.

Pitcher values still contain team-defense and park effects — disclosed, unmodeled in v1. Starter and reliever tables are different workloads and are never ranked together.

±SE is the standard error of the published value — the shrunken estimate. It previously carried the standard error of the unshrunken mean, which is 1.5× to 3.7× wider and belongs to a number this page does not show; for relievers that stated interval was 2.5× the entire spread between relievers, so every row covered the whole table. The league mean shrunk toward is treated as known, so these are conditional intervals. An interval covering zero means this estimator cannot separate the player from league average on this sample — not that the player is average.