Which farm systems make players better?
The finding Across 30 organizations, how much an org’s Triple-A players improved from the first half of the season to the second in 2025 said little about the same org in 2026. None of the six measures showed an org effect that repeated from 2025 to 2026 by more than chance: every interval for the year-to-year correlation includes zero. Fastball velocity came closest (r = 0.42, interval −0.03 to 0.72).
So what
Don’t credit or blame a farm system for one season of Stuff+ gains. An org’s 2025 figure said little about its 2026 figure (r = 0.24, interval −0.01 to 0.49).
Velocity is the closest thing to an org trait, and it’s still unconfirmed. Org velocity gains repeated at r = 0.42, but the interval (−0.03 to 0.72) reaches zero. Worth a look at what the top orgs do, not a ranking to act on.
- Judge development player by player, with the sample size in view. A typical org had 12 qualifying pitchers and 6 qualifying hitters in 2026, too few to separate a system’s effect from who happened to be on the roster. The drill-down below and the player pages show each player’s own change.
Does an org’s improvement repeat?
Each dot is one organization: its players’ average change in 2025 (across) and in 2026 (up), after removing the league-wide change each season. If farm systems had a lasting effect, the dots would run from lower left to upper right.
The org scores, with their uncertainty
Every org’s score is shrunk toward the league average in proportion to how noisy it is, and comes with a 90% interval and the number of players behind it. In 2026 Stuff+, the top org was MIA (+1.2, interval +0.5 to +1.9, n = 13) and the bottom was NYM (−1.5, interval −2.7 to −0.4, n = 16). Those two intervals don’t overlap, so within 2026 the gap between them is more than sampling noise. In 2025, MIA had ranked 18th of 30 on the unshrunk average and NYM 20th. Pick an org to see its players.
Inside one organization
How each player in the org changed between halves. For 2026 pitchers, the date slider steps through the season week by week, with each pitcher’s last-30-days grade against the org average and the average of every pitcher in the sample.
Why one season isn’t enough
An org’s score averages over a small group of players (12 pitchers for the typical org in 2026), and each player’s half-season change is noisy on its own: a pitch-mix change, a minor injury, or a few hundred pitches of sampling error all move it. Within a single season, the share of the spread between org scores that reflects real differences between orgs, rather than sampling noise, ranged from 0% to 51% across the six measures and two seasons. The year-to-year test asks the harder question: whether the orgs that did well once did well again. For Stuff+, the answer was r = 0.24 (−0.01 to 0.49); for Pitching+, 0.02 (−0.19 to 0.24); for chase rate, 0.02 (−0.33 to 0.33); for zone contact, −0.11 (−0.33 to 0.11); and for EV90, −0.12 (−0.47 to 0.29).
That doesn’t prove farm systems don’t matter. Two seasons and 30 orgs is a small test, an org can change its coaches and methods between years, and development that shows up as a player’s whole-season level rather than a first-to-second-half change isn’t measured here. What it does show is that a one-season ranking of farm systems by in-season improvement isn’t evidence of much on its own.
Method and caveats
Data. Pitch-level Statcast for Triple-A, 2025 and 2026, plus MLB for the same seasons (to find promotions and big leaguers on assignment). Each season is split at its median Triple-A pitch date (June 24 in 2025, June 21 in 2026), so the halves hold about the same number of pitches. Each player is credited to the MLB organization his Triple-A club was affiliated with that season (MLB Stats API).
Who’s included. Pitchers with 100+ Triple-A pitches in each half (361 in 2025, 367 in 2026) and hitters with 100+ PA in each half (194 and 191). Excluded across both seasons: players who played for more than one Triple-A club in the season, which means they changed orgs midseason (238 pitchers, 118 hitters), and players with meaningful MLB time that season before their first Triple-A game, mostly rehab assignments and optioned big leaguers (100+ MLB pitches or 50+ MLB PA; 72 pitchers, 17 hitters).
Measures. Pitchers: Stuff+, Pitching+, and primary-fastball velocity (the more-thrown of his four-seamer and sinker). Hitters: chase rate, zone contact, and EV90 (90th-percentile exit velocity), with the strike zone set from each batter’s height. Change = second half minus first half, oriented so positive is always better (for chase, a drop counts as a gain), in Stuff+ points, mph, or percentage points. Each change has the league-wide average change for that season subtracted: the average included pitcher’s Stuff+ moved −0.5 in 2025 and −1.1 in 2026, and fastball velocity +0.12 and −0.05 mph.
Org score. The average adjusted change of the org’s players, shrunk toward the league average by empirical Bayes: the spread between orgs that’s larger than sampling noise alone would produce is estimated from all 30 orgs, and each org’s average is pulled toward the mean by how much of its deviation is likely noise (more for small or noisy orgs, none when the between-org spread is no bigger than noise). The 90% interval comes from 2,000 bootstrap resamples of the org’s players, with the shrinkage factor held fixed.
Consistency test. The correlation between each org’s 2025 and 2026 average adjusted change (unshrunk, so shrinkage can’t manufacture or hide agreement), with a 90% interval from 2,000 bootstrap resamples of the orgs.
Survivorship. Players who earn a promotion leave the second-half Triple-A sample, and an org that promotes its improvers would look worse for it. Pitchers who qualified in the first half and threw 100+ MLB pitches in the second instead: 75 in 2025 and 66 in 2026; hitters: 34 and 39. Each org’s count is in the table. Adding the promoted pitchers back, using their MLB second half (the grades are on the same scale at both levels), gives year-to-year correlations of 0.16 for Stuff+, 0.06 for Pitching+, and 0.42 for velocity. Promoted hitters aren’t added back, because MLB swing and contact rates aren’t comparable to Triple-A ones without a translation.
What this can’t tell you. Whether an org’s methods work for any one player; development that changes a player’s level before he reaches Triple-A or across a whole season rather than between halves; anything about the lower minors. Correlation across orgs isn’t causation in either direction: an org’s roster construction, park, and promotion habits all feed its score.
Code:baseball/farm_systems/