Keystone Stuff

Count-neutral expected run value per pitch, pitcher perspective. Batted balls are replaced by the average outcome for their exit-velocity × launch-angle bin, which removes defense and sequencing luck.

• Velocity, spin rate, spin axis, extension, release height and side (arm-side normalized)
• Induced vertical and horizontal break; arm angle
• Vertical approach angle, adjusted for plate location so the model cannot see location
• Velocity, movement, and axis differentials vs. the pitcher’s primary fastball

• Gradient-boosted trees (scikit-learn HistGradientBoosting) on 2024–26 MLB Statcast
• Leave-one-season-out cross-fitting: every graded pitch comes from a model that never saw that season
• A production model trained on all seasons grades hypothetical pitch designs and the Stuff landscape, and grades every season of a player’s own history when Keystone compares his seasons (cross-fit folds differ by season, which would show drift that isn’t his). A change is the production model’s; a level is anchored to the published season grade, so a pitcher-season has one Stuff number everywhere on the site
• The production model was fit on every pitcher it grades, so a Stuff drop is also graded by two pitcher-held-out models (each fit on half the pitchers, grading the other half). Pitch Lab flags a drop of 5+ points only when both agree, and puts runs at stake on the smaller of the two changes
• Aggregated to pitch-type and pitcher level, then scaled within each season: 100 = that season’s league average (scaling all seasons together left 2026 grades about 2 points high)
• Forward test: a model trained on 2024 only grades 2025; that grade tracks 2026 results about as well as 2025 results do (see the Pitch Lab and Track Record). The cross-fit grades’ own next-season correlation is optimistic, because the 2025 fold’s model was trained on 2026
• Unseen pitchers: trained on half the pitchers and tested on the other half, the forward correlation is lower than for pitchers the model has seen, so on a thin sample Stuff is a useful early read, not a better one than results
• Hypothetical shapes (pitch designs, the landscape) get the approach angle their release, velocity and rise imply (a least-squares fit on real pitches), never a blank the model didn’t see in training
• Landscape reality check: a map’s best shapes are compared with the real pitches at that shape and velocity (±2 mph) from similar slots, graded by the same model pitch by pitch. A mark above their 90th percentile, or with fewer than five of them, is labeled on the page as the model extrapolating or untested
• Dead weight: a weak pitch with below-average expected results moves toward his best pitch only against a batter side where that pitch wins by 2+ standard errors (his own results against the side, combined with Stuff and the league’s platoon split for the pitch), valued with those side values, never below 35% fastballs
• Pitch design: a new pitch is built from his own fastball plus comparable pitchers’ typical differences, graded against the pitch it would compete with (never his primary fastball; first his weakest pitch of the same kind that breaks the same way) under the production model and every cross-fit model, and queued only when it wins by 5+ under all of them and grades no higher than the comps’ 90th percentile, both on their published grades and with the comps graded the same way; otherwise it is labeled fragile. Only for arsenals of four or fewer main pitches, never a pitch he has thrown 100+ times, with 15+ comparable pitchers

• Location and command are excluded by design. A pitcher can have great stuff and poor results because of where he throws.
• Sequencing, tunneling, and deception aren’t modeled.
• No biomechanics, so it can’t say whether a design is physically safe for a given arm.
• Pitch-type labels come from MLB’s classifier. Misclassified pitches blur the grades.
• Minor-league grades exist only where Hawk-Eye tracks (Triple-A, parts of the FSL).
• A built pitch design tends to grade high: the typical shape on his fastball out-graded every comp’s real pitch for about two in five of the ideas that passed the every-model test, so those are labeled fragile, and the best of several candidate pitch types is chosen, which flatters the winner a little more.
• A grade isn’t exact: refit the model on different seasons and a pitcher’s grade, and his change since last season, move a few points. The Pitch Lab publishes the measured spread and marks a Stuff alert inside it as borderline.
• For a pitcher whose slot or shape has few neighbours, the published grade leans on his own other seasons: the Pitch Lab marks a leader that a model which never saw him grades 15+ points lower.

• Prospective: run-value change for pitchers given a model-informed plan vs. matched controls
• Retrospective: share of dead-weight flags whose pitch the pitcher later reduced, and the results that followed