Reference

Glossary

Every technical term used on this site, in one place. Dotted-underlined words across the site open these same definitions in a small card; this page holds the full versions. New to the subject entirely? The reader’s guide builds the concepts up in order.

Bucketing Circularity Ex ante Ex post Leave-team-out Marcel projection Opponent-adjusted Pre-registration Reliability Shrinkage Split-half correlation Team slope (δ)Transfer coefficient wOBA ZiPS

Bucketing: Every version of the split statistic starts by dividing pitchers into quality groups: terciles, quartiles, ERA thresholds. The choice is arbitrary, and rankings move by about five places (out of thirty) when the definition changes. How the groups are defined also decides whether the statistic is circular: grouping by a pitcher's season results lets a team's own hitting move pitchers between its buckets.See also: Circularity · Ex post
Circularity: TV-style splits classify pitchers by their season results, which include what the observing team did to them. A team that hammers a mid-tier pitcher pushes his ERA up, reclassifying him from "good" to "bad", which simultaneously inflates the team's numbers against bad pitching and drains its numbers against good pitching. In 2026 this artifact alone accounts for roughly half of the Brewers' famous 93-point split. The fix is to classify pitchers by information the team cannot move: projections made before the season, or statistics computed with that team's plate appearances left out.See also: Ex ante · Leave-team-out · Bucketing
Ex ante: An ex-ante quality label is fixed before any of the plate appearances being studied happen, so a team's results cannot leak into its own comparison groups. This project uses two: a transparent Marcel-style projection built from the previous three seasons, and FanGraphs' pre-season ZiPS. Under either, the 2026 Brewers rank near the top of baseball against good pitching.See also: Ex post · Marcel projection · ZiPS
Ex post: Ex-post quality labels (season ERA, season wOBA-against) are the broadcast convention, and they are the circular ones: the team being evaluated helped write them. Not every ex-post statistic is equally contaminated. xFIP, built from strikeouts, walks, and fly balls, moves far less under one team's damage than ERA does, and the Brewers' split largely evaporates under it.See also: Ex ante · Circularity
Leave-team-out: The surgical fix for circularity: when classifying pitchers for the Brewers' split, compute every pitcher's quality from his plate appearances against the other 29 clubs. The pitcher pool and the season are identical; only the contamination is removed. Applying just this correction cuts the Brewers' 2026 pooled split roughly in half.See also: Circularity · Ex post
Marcel projection: Tom Tango named it after the monkey from Friends: even a monkey-simple projection captures most of what is knowable. This project's ex-ante pitcher quality is a Marcel-style projection of wOBA allowed: three seasons of history, 5/4/3 weights, and a strong pull toward the league mean. It is transparent and reproducible; its one weakness is rookies, who get projected as league average, which is why the independent pre-season ZiPS bucketing is also reported.See also: Ex ante · ZiPS
Opponent-adjusted: October offense collapses league-wide, but two-thirds of the collapse is simply that postseason lineups face aces and elite relievers. Opponent-adjusted performance subtracts that expectation: it asks how a team hit relative to what anyone would be expected to hit against those exact pitchers. This is the fair yardstick for postseason offense, and the outcome variable for the pre-registered 2026 test.See also: Team slope (δ) · Pre-registration
Pre-registration: One analysis here produced a suggestive result (the transfer coefficient) that could easily be noise. Rather than promote or bury it, the study froze every team's 2026 slope estimate before the playoffs, committed the exact regression to be run when the postseason ends, and stated what each hypothesis predicts. The result will be reported either way. The frozen values live in the repository at a recorded commit.See also: Transfer coefficient
Reliability: A statistic that reflects a stable trait should correlate with itself: the same team's value in odd-numbered dates vs even-numbered dates (split-half), or this season vs next season. Team good-vs-bad-pitching slopes fail both tests almost exactly at zero (split-half r = −0.003, year-over-year r = +0.047). By comparison, a hitter's overall wOBA has year-to-year reliability near 0.50. Zero reliability means a ranking built on the statistic reshuffles randomly every time you re-measure it.See also: Split-half correlation · Shrinkage · Team slope (δ)
Shrinkage: Empirical-Bayes shrinkage answers "how much of this extreme number should I believe?" with a weight computed from the data: the ratio of true spread to true-plus-noise spread. For team pitching-quality slopes that weight averages 0.12: a team that appears 100 points better than average against good pitching should be believed to the tune of about 12 points. The caterpillar chart on this site shows what survives.See also: Reliability · Team slope (δ)
Split-half correlation: The cleanest reliability test there is: no roster changes, no aging, no park moves. The same team, the same season, odd dates against even dates. If "struggles against good pitching" described anything real about a roster, the two halves would agree at least somewhat. The correlation across 780 team-seasons is −0.003.See also: Reliability
Team slope (δ): Everyone hits worse against better pitchers: the league-wide slope. The claim behind "they can't hit good pitching" is that some teams decline faster than that: a team-specific slope deviation, written δ. This project estimates δ for all 780 team-seasons since 2000 and finds its reliable component is approximately zero: the estimates are almost entirely sampling noise.See also: Reliability · Shrinkage
Transfer coefficient: The pre-registered postseason test regresses opponent-adjusted October performance on each team's regular-season slope times the quality jump it faced. Across 26 postseasons the estimate is 0.92 with a standard error of 0.38, nominally favoring transfer. But the design can barely detect even full transfer, the measured reliability implies the coefficient should be near 0.12, and the effect is absent from series outcomes. It is reported as an unresolved anomaly with a committed 2026 out-of-sample test.See also: Pre-registration · Opponent-adjusted
wOBA: wOBA credits a walk about 0.7, a single about 0.9, and a home run about 2.0 (weights derived each season from how many runs each event actually produced) and scales the result to look like on-base percentage (league average ≈ .320). This site's weights are rebuilt per season from play-by-play run expectancies and match the published FanGraphs constants within a few thousandths. Differences of 10 points of wOBA are meaningful over a season; differences of 90 points over a third of a season's plate appearances are mostly noise.See also: Bucketing
ZiPS: ZiPS projects every player before the season from multi-year performance, aging curves, and minor-league data. Because the pre-season snapshot exists before any 2026 games, bucketing pitchers by it is immune to circularity by construction, and unlike this project's in-house Marcel it has real projections for rookies. Under pre-season ZiPS buckets the 2026 Brewers rank eighth against good pitching and eighth against bad: no split at all.See also: Ex ante · Marcel projection