Pitch Sequencing & Tunnels
Trajectories rebuilt from Statcast kinematics, and a placebo test that separates the pitch from the pair: beyond the pitch, the order adds only a sliver.
What it predicts
Whiff on swings and chase on pitches outside the zone at the second pitch of a pair, and run value per pitch; results (xRV) beyond Stuff for predictability.
Inputs
- Every consecutive pitch pair in a plate appearance, 2025–26 (Statcast velocity and acceleration at 50 ft, plate location)
- Keystone Stuff (expected run value) for the predictability test
- Today’s 40-man and the Winter Plan’s 2027 roles (decisions feed) for the Pittsburgh list
Method
- Solve each pitch’s time to the plate, back-integrate from the plate to its position 0.167 s earlier (the decision point)
- Pair model: logistic regression of whiff on swings with pitch-type, previous-type, count, velocity and velocity-gap controls, the second pitch’s own movement and where it crossed the plate, the platoon matchup and the season, plus tunnel distance and late separation; average marginal changes with pitcher-clustered standard errors, re-fit on each season alone
- Placebo: the real previous pitch replaced by a random pitch of the same type by the same pitcher that season from another plate appearance (seeded). It “predicts” whiffs, chases and run value as well as the real one, so the tunnel effect belongs to the pitch, not the order
- Arsenal tunnel: each pitch’s average distance at the decision point (and at the plate) from 12 random pitches of the previous pitch’s type. With it in the model the real pair adds only a sliver: in-sample, nothing to whiffs from the tunnel, a few tenths of a point of chases and about 0.1 runs per 100 pitches; forward (fit 2025, scored 2026), a detectable but tiny gain in whiff log-loss, about a tenth of the arsenal tunnel’s, and none on chases
- The verdict is computed from every outcome and the forward test (2 standard errors to count, and a forward gain at least half the arsenal tunnel’s to say the order matters), not from whiffs alone
- No pair-level “lean on” calls; pair results are shown with standard errors, colored only at 2.5 standard errors (100+ pairs), same-pitch repeats marked
- Predictability: smoothed conditional frequencies of the next pitch type given count, previous pitch and batter hand, fit on alternate pairs and scored on the rest within each season, averaged over seasons; stray types (under 3%), pitchouts and intentional balls dropped
- Predictability tested against results minus Stuff, both seasons stacked, with a pitcher-clustered bootstrap interval; tendencies (the likeliest next pitch by context) scored out of sample
How well it works
Computed by the pipeline from the data on this site
- Tunnel effect on whiffs (561,742 swings on the second pitch of a pair, with controls for 8 other factors)
- 3 in tighter at the decision point +0.41 points of whiff rate (standard error 0.07) · 3 in more separation at the plate +0.17 (standard error 0.04), on a base of 22.7% Each season alone (tighter / more separation): 2025: +0.43 / +0.19 · 2026: +0.35 / +0.12 A pooled 2025–26 fit with a season term, and each season alone, before the placebo control (the pitcher’s arsenal-level tunnel). The Track Record row below quotes the same estimate with its 95% interval (about two standard errors each way), and the pair-specific part that survives the control.
- Do predictable pitchers get hit harder? (predictability vs. results beyond Stuff in the same season; 2025 and 2026 stacked, 541 pitcher-seasons)
- r = −.01 (95% −.09 to .06, a bootstrap over pitchers) · by season 2025: r −.06 (275), 2026: r .05 (266) The interval spans zero: being predictable doesn’t show up in results. Both are same-season tests, not forward ones; the Track Record row quotes the same stacked test.
- Track Record: Tighter pitch-to-pitch tunnels produce more whiffs
- pair-specific (placebo-adjusted), 3 in tighter: −0.07 pts whiff (95% interval −0.21 to +0.06); before the placebo control: +0.4 pts (95% interval ±0.13); 2025: +0.4, 2026: +0.3; arsenal-level tunnel +1.8 pts Verdict: fail.
- Track Record: Pitch order adds to results beyond the pitch itself
- In-sample, pair terms that clear: whiffs (pair separation −0.12 pts), whiffs on different pitch types (pair separation −0.13 pts), chases (pair tunnel +0.45 pts), chases (pair separation +0.19 pts), run value (pair tunnel +0.10 runs/100). Forward: whiffs +11% of the arsenal term’s gain (clears the noise); chases 0% of the arsenal term’s gain (inside the noise) Verdict: partial.
- Track Record: Predictable pitchers get hit harder
- r = −.01 (95% interval −.09 to .06, bootstrap over pitchers); by season 2025: −.06 (275), 2026: .05 (266) Verdict: fail.
Data: Statcast 2025–26 pitch pairs (561,742 swings on the second pitch of a pair); this model doesn’t use 2024.
What it can't see
- The arsenal tunnel is an association with controls, not a tested design change: it mixes pitch design with location the controls don’t fully capture.
- No batter effects; location intent is private (the catcher’s target).
- Results beyond Stuff barely repeat (r ≈ 0.2), which caps any predictability test.
- Decision point is a fixed 0.167 s; real hitters vary.
How we'd measure impact
- Coaching time moved off pitch order
- Whiff and chase rates of pitches redesigned to track the arsenal, 2027 (a future Pitch Lab test: the tunnel isn’t in the Pitch Lab or Stuff yet)
- Predictability re-tested every season