Stuff or command? Splitting pitching into what the ball does and where it goes
The finding Location+, a grade for where each pitch crossed the plate in the count it was thrown, is a steady trait: it held from one half of a season to the other at r = 0.69 and from 2025 to 2026 at r = 0.70, but it did almost nothing to predict next season’s run value, and adding it to Stuff+ didn’t improve the forecast (324 MLB pitchers).
So what
Forecast with Stuff+. Adding location (Pitching+) predicted next season no better (0.32 vs. 0.30), so the bigger model adds nothing to a projection.
- Read Location+ as a description of a season, not a forecast. Within 2026, among the 468 MLB pitchers with 500+ pitches, it tracked results only weakly (r = 0.09, against 0.29 for Stuff+ and 0.38 for both together).
- Don’t project a Triple-A pitcher on his command grade. Triple-A Location+ said nothing about the same pitchers’ MLB run value the next season (r = −0.05, n = 148). Triple-A Stuff+ carried more (r = 0.15), though that sample is small too.
Three grades, four tests
Stuff+ grades a pitch on what the ball does: velocity, movement, spin, and release. It ignores where the pitch goes. Location+ is the mirror image: it sees only where the pitch crossed the plate, the count, and whether pitcher and batter are on the same side. Pitching+ sees everything both models see. All three are trained on the same target (run value, with batted-ball luck removed) and put on the same scale, where 100 is an average qualified MLB pitcher from 2025 and 10 points is one standard deviation. Actual results (run value per 100 pitches) are the baseline to beat.
Pick a test. Each one asks a different question of the same pitchers.
Location is real, but it doesn’t carry forward
The stability numbers say Location+ measures something about the pitcher, not just luck. Split-half reliability is 0.69, well below Stuff+ (0.99, which is nearly fixed by a pitcher’s arm) but far above actual results (0.20). Year to year, Location+ repeats at 0.70, against 0.19 for results.
That steadiness doesn’t turn into a forecast. Location+ on its own predicted next season’s run value at r = 0.01, with an interval (−0.08 to 0.10) that runs across zero: no detectable signal. Pitching+, which adds location and count to the same stuff inputs, predicted it at r = 0.32, against 0.30 for Stuff+ alone. The difference (+0.02, interval −0.05 to +0.09) is indistinguishable from zero: adding location did not improve the forecast. Pitching+ still beat last season’s results (0.19), with an interval for the gap of +0.02 to +0.22 that stays above zero.
Why would a steady trait predict so little? The data here can’t settle it, but a few explanations fit. Among those same 468 pitchers, Stuff+ and Location+ were slightly negatively related (r = −0.24): pitchers with less stuff tend to live on the edges, so a high location grade often comes attached to stuff that hitters can square up when a pitch misses. Location+ also grades where the ball ended up, not where the catcher set up, so it can’t tell a well-executed pitch from a lucky miss. And the value of a location depends on sequencing and on how hitters adjust, neither of which the model sees.
The top five by Location+ in 2026 among MLB pitchers with 500+ pitches: Cody Bradford (TEX, 124); Miles Mikolas (WSH, 122); Jacob deGrom (TEX, 121); Tyler Mahle (SF, 121); José Urquidy (CWS, 121). By Pitching+: Adrian Morejon (SD, 133); Jhoan Duran (PHI, 131); Louis Varland (TOR, 130); Mason Miller (SD, 128); Alex Vesia (LAD, 125). Every pitcher’s three grades, by date and by pitch, are in the pitch-grade leaderboard and on each player page.
Triple-A
The models grade Triple-A pitchers on the MLB scale. The average qualified Triple-A pitcher sits at 88 on Location+ and 85 on Pitching+, below big leaguers on both. Location+ didn’t carry up a level: for 148 pitchers who moved from Triple-A in 2025 to MLB in 2026, Triple-A Location+ correlated with MLB run value at r = −0.05, against 0.15 for Stuff+ and −0.02 for Triple-A results. The sample is small, so only the broad pattern is worth reading: stuff carried a little, location and results didn’t.
Method and caveats
Data. Pitch-level Statcast from Baseball Savant: MLB 2024–2026 and Triple-A 2025–2026, the same pitches as Stuff+. Batter heights come from the MLB Stats API.
Location+ features. Horizontal location, mirrored by batter side so positive always means inside; height, scaled to the batter’s rulebook strike zone (0 at the bottom, 1 at the top); balls; strikes; and whether pitcher and batter throw and hit from the same side. One model per pitch group (fastballs, breaking balls, offspeed), so a slider at the knees is judged against other breaking balls.
The strike zone changed in the data. In 2026, Statcast’s zone top and bottom are fixed for each batter at 53.5% and 27.0% of his listed height, the rulebook zone used by the automated ball-strike system. In earlier seasons they were set from the batter’s stance on each pitch, which gives a taller zone. Using those fields would grade the same pitch differently across seasons, so every season here uses the height-based zone instead.
Pitching+ features. Every Stuff+ input plus every Location+ input, in one model per pitch group.
Target, model, and scale. Same as Stuff+: run value from the pitcher’s side with batted-ball luck replaced by expected outcomes, LightGBM with wide leaves (min_data_in_leaf = 500), trained on MLB 2024–2025, and scaled against qualified 2025 MLB pitchers, frozen for the season. On that scale 2026’s qualified MLB pitchers average 98 on Location+. Pitch-level grades are scaled within pitch group.
Validation. The year-over-year test refits all three models on 2024 only, grades 2025, and predicts 2026 run value, so neither season touches training. Intervals are 90% bootstrap intervals over pitchers (2,000 resamples), with each gap computed within the same resample. Differences: Pitching+ minus Stuff+ +0.02 (−0.05 to +0.09); Location+ minus Stuff+ −0.29 (−0.43 to −0.14); Pitching+ minus results +0.13 (+0.02 to +0.22).
Checks. Pitch-level R² on the 2026 holdout is 0.034 for Location+ and 0.036 for Pitching+, against 0.0006 for Stuff+. That gap doesn’t mean location is the better grade: where a pitch lands and the count decide whether it’s a ball or a strike, so location explains more of any single pitch, but it averages out over a season in a way stuff doesn’t.
What this can’t tell you. Location+ is not command. It sees where the pitch ended up, not where it was aimed, so a pitcher who misses his spot into a good location gets credit. It also doesn’t account for catcher framing, the umpire, sequencing, or the hitter’s plan. Correlations here are between seasons for the same pitchers; they don’t say that improving location wouldn’t help a pitcher, only that last season’s location grade didn’t predict this season’s results.
Code:baseball/pitch_quality/