Does pitch order matter, and does being predictable cost us?
Is it the pair? The placebo test
Effect of 3 in tighter at the decision point (separation: 3 in further apart at the plate), ± one standard error (clustered by pitcher). Same controls in every model: pitch types, count, velocity and velocity gap, the second pitch’s movement and location, platoon, season
| Outcome | Real pair | Placebo pair | Real pair, beyond the arsenal | Separation, beyond the arsenal | Arsenal tunnel | Rows |
|---|---|---|---|---|---|---|
| Whiff on swings (pts) | +0.41 ± 0.07 | +0.56 ± 0.06 | −0.07 ± 0.07 | −0.12 ± 0.04 | +1.84 ± 0.14 | 561,742 |
| Whiff, different pitch types (pts) | +0.49 ± 0.07 | +0.49 ± 0.07 | +0.05 ± 0.07 | −0.13 ± 0.04 | +1.43 ± 0.14 | 367,070 |
| Chase, different pitch types: swings at pitches outside the zone (pts) | +5.58 ± 0.09 | +6.02 ± 0.09 | +0.45 ± 0.09 | +0.19 ± 0.06 | +12.22 ± 0.18 | 372,587 |
| Run value per 100 pitches, different pitch types (pitcher’s view) | +0.41 ± 0.03 | +0.45 ± 0.03 | +0.10 ± 0.04 | +0.02 ± 0.02 | +1.14 ± 0.05 | 694,719 |
The placebo swaps the real previous pitch for a random pitch of the same type, thrown by the same pitcher that season in another plate appearance. It “works” as well as the real one, so the tunnel effect isn’t about the order. The arsenal tunnel is each pitch’s average distance, at the decision point, from 12 random pitches of the previous pitch’s type by the same pitcher; with it (and its plate-separation twin) in the model, the real pair’s tunnel is what the sequence adds. Forward, fit on 2025 and scored on 2026 (share of the log-loss saved, ± one pitcher-clustered standard error): adding the real pair to a model that has the arsenal tunnel saves +0.017% ± 0.006% on whiffs and −0.007% ± 0.005% on chases; adding the arsenal tunnel to the pair saves +0.153% ± 0.015% and +2.966% ± 0.095%. In-sample, beyond the arsenal, these clear two standard errors: +0.45 ± 0.09 points of chases (4% of the arsenal’s); −0.13 ± 0.04 points of whiffs on different pitch types from late separation (the opposite sign to the arsenal’s); +0.19 ± 0.06 points of chases from late separation (4% of the arsenal’s); −0.12 ± 0.04 points of whiffs from late separation (the opposite sign to the arsenal’s); +0.10 ± 0.04 runs per 100 pitches (8% of the arsenal’s). Giving every pitcher his own baseline (pitcher fixed effects, a linear probability model) barely moves the real pair beyond the arsenal: −0.02 ± 0.06 points of whiffs against −0.03 ± 0.06 without them, and +0.64 ± 0.10 of chases against +0.57 ± 0.10, so differences between pitchers aren’t what that column measures. 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. A whiff is worth 0.13 runs to the pitcher over another swing.
Whiffs by tunnel distance
Swings on the second pitch of a pair with different pitch types, all MLB 2025–2026: raw rates (bars) and the pair’s tunnel alone through the model with the arsenal in it (marks)
Raw rates rise with tunnel distance (bars) because wide tunnels finish far apart and land in different places: the raw link is confounded. Through the model, with where the pitch went and how it tracks his other pitches held as they were, the pair’s own tunnel moves whiffs from 22.4% to 22.2% across the whole range (marks). Same-pitch repeats, 35% of swings, are left out: there the “tunnel” is only a location change (no measurable change in whiffs per 3 in closer to the last one: −0.4 ± 0.3 points).
Pittsburgh’s pitchers: how predictable?
Today’s 40-man pitchers with their 2027 status (decisions feed). Log-loss gained guessing the next pitch from count, previous pitch and batter hand, scored within each season and averaged; league median 0.059
| Pitcher | Pairs | Types | Predict. | Pctile | 2025 | 2026 |
|---|---|---|---|---|---|---|
| 2027: bullpen | 1,510 | 4 | 0.108 | 87th | 0.135 | 0.068 |
| 2027: bullpen | 860 | 5 | 0.095 | 80th | — | 0.095 |
| 2027: bullpen | 1,479 | 4 | 0.086 | 74th | 0.069 | 0.092 |
| 2027: bullpen | 1,526 | 4 | 0.086 | 74th | 0.044 | 0.117 |
| 2027: rotation · injured | 3,403 | 6 | 0.066 | 58th | 0.067 | 0.065 |
| 2027: rotation | 4,346 | 6 | 0.066 | 58th | 0.068 | 0.064 |
| free agent | 1,618 | 4 | 0.059 | 50th | 0.030 | 0.083 |
| 2027: bullpen | 2,693 | 6 | 0.048 | 38th | 0.059 | 0.038 |
| 2027: rotation | 2,464 | 4 | 0.040 | 30th | 0.061 | 0.031 |
| 2027: bullpen | 1,441 | 3 | 0.028 | 19th | 0.031 | 0.025 |
| 2027: bullpen | 1,132 | 4 | 0.022 | 15th | 0.033 | 0.009 |
| free agent | 1,125 | 2 | 0.015 | 10th | 0.022 | 0.009 |
| 2027: call-up | 1,892 | 5 | 0.006 | 6th | 0.014 | 0.000 |
| 2027: rotation | 2,349 | 6 | 0.004 | 5th | 0.036 | −0.001 |
| 2027: rotation | 1,218 | 4 | −0.010 | 1st | — | −0.010 |
Too few pitch pairs to score yet (under 600): Wilber Dotel 525 (2027 bullpen), Noah Murdock 346 (2027 call-up), Hunter Barco 298 (2027 minors), Khristian Curtis 262 (2027 call-up), Antwone Kelly 212 (2027 call-up), Thomas Harrington 150 (2027 minors). Deeper arsenals show more structure on this measure: the league median is 0.016 with 2 pitch types and 0.085 with 7 pitch types. Scores still rise a little with sample size (r = 0.17 with the number of pitch pairs, log scale). It isn’t sampling noise: the two halves of each season agree so closely that every score here is at least 97% signal (half signal at about 21 pairs), so shrinking by sample size leaves r at 0.17. Scored on a fixed 300 pairs a season, the link reverses (r = −0.20): part of the rise is the smoothing, which holds small samples near usage, and under it pitchers with more pairs are a little less predictable. Compare scores between pitchers loosely. Pitch types under 3% of a pitcher’s pitches, pitchouts and intentional balls are left out. Since predictability shows no cost league-wide (the test rules out correlations beyond about 0.09 either way), it is context for game-planning, not a development goal. Results beyond Stuff repeat only at r = 0.18 from season to season, which limits how large a link any test could find.
Paul Skenes: how his pitches track each other
Each pitch at 5%+ of his mix (the Types column above counts pitches at 3%+): its arsenal tunnel, how far it is at the decision point from his other pitches (mean over his different-type pairs, 2025–2026), against league pitchers with the same pitch
| Pitch | Use | From his other pitches | League median | Tracks his other pitches |
|---|---|---|---|---|
| Four-seam | 41% | 11.4 in | 11.5 in | closer than 60% of league four-seamers (499) |
| Sweeper | 15% | 10.9 in | 11.7 in | closer than 91% of league sweepers (202) |
| Changeup | 14% | 12.3 in | 11.9 in | closer than 32% of league changeups (295) |
| Splitter | 11% | 10.4 in | 12.0 in | closer than 97% of league splitters (78) |
| Sinker | 10% | 10.2 in | 10.9 in | closer than 90% of league sinkers (311) |
| Slider | 8% | 11.6 in | 11.5 in | closer than 42% of league sliders (317) |
A pitch that stays close to his other pitches until the hitter commits goes with more whiffs and chases across the league, beyond its movement and location. That makes it a candidate pitch-design input, not yet in the Pitch Lab or in Stuff, and not a measured effect of changing a pitch. There are no “lean on this pair” calls: beyond the arsenal, the order adds only a sliver (in-sample, at most +0.45 ± 0.09 points of chases; forward, a detectable but small gain on whiffs, 11% of what the arsenal tunnel saves), and no pairs of his differ from his own average by 2.5 standard errors.
His pitch pairs: the 14 most frequent of 17 thrown 80+ times (descriptive)
| Sequence | Times | Tunnel | Separation | Whiff | vs. his avg | Runs / 100 | vs. his avg |
|---|---|---|---|---|---|---|---|
| 4S → 4Ssame pitch again | 831 | 8.0 in | 13.5 in | 31% | +4.5 ± 2.0 | +1.9 | +0.2 ± 0.7 |
| ST → 4S | 276 | 10.7 in | 17.2 in | 21% | −5.2 ± 3.5 | +0.5 | −1.2 ± 1.6 |
| 4S → CH | 237 | 13.1 in | 23.4 in | 33% | −4.2 ± 4.4 | −1.7 | −2.2 ± 1.6 |
| 4S → ST | 226 | 10.5 in | 17.5 in | 26% | −2.3 ± 4.2 | +1.4 | −0.7 ± 1.2 |
| CH → 4S | 223 | 13.2 in | 23.4 in | 32% | +5.5 ± 3.6 | +2.7 | +0.9 ± 1.5 |
| FS → 4S | 199 | 9.7 in | 18.5 in | 17% | −8.9 ± 4.1 | +1.7 | 0.0 ± 1.3 |
| ST → STsame pitch again | 174 | 10.2 in | 15.8 in | 38% | +9.8 ± 4.8 | +1.5 | −0.5 ± 1.4 |
| 4S → FS | 151 | 10.8 in | 21.0 in | 14% | −4.4 ± 4.3 | +1.0 | +0.7 ± 1.8 |
| 4S → SL | 138 | 12.5 in | 22.6 in | 32% | +2.5 ± 5.2 | +2.2 | +1.2 ± 1.7 |
| SI → 4S | 124 | 10.2 in | 18.5 in | 20% | −6.3 ± 5.5 | +3.4 | +1.7 ± 1.5 |
| 4S → SI | 123 | 9.6 in | 17.0 in | 12% | −0.5 ± 4.1 | +0.5 | −0.9 ± 2.1 |
| FS → FSsame pitch again | 117 | 7.6 in | 12.6 in | 21% | +2.1 ± 4.6 | −3.2 | −3.5 ± 2.8 |
| SL → 4S | 115 | 11.2 in | 19.9 in | 22% | −4.6 ± 5.1 | +2.9 | +1.2 ± 1.6 |
| CH → CHsame pitch again | 114 | 9.7 in | 15.6 in | 29% | −8.3 ± 6.0 | +1.7 | +1.2 ± 1.8 |
± is one standard error; a value is colored only at 2.5 standard errors (with 100+ pairs, 40+ swings for whiffs). Runs per 100 are from the pitcher’s side (positive = good for him). Muted rows repeat the same pitch, where the tunnel is just a location change.
Advance scouting: any pitcher’s tendencies
Where a pitcher’s next pitch is most often one type, beyond his usage (count, previous pitch, batter side; 25+ pitch pairs, 2025–2026), and how often the likeliest pitch came on pairs the frequencies weren’t fit on
Kevin Gausman (CHC): predictability 0.042, 31st percentile, 3 pitch types, 4,469 pairs.
- 0-2, after a Splitter, vs. LHB: Splitter 77% of the time (his usage 43%; 90 pairs). Out of sample, the likeliest pitch came 77% of the time.
- 0-2, after a Slider, vs. RHB: Slider 39% of the time (his usage 7%; 33 pairs). Out of sample, the likeliest pitch came 30% of the time.
- 3-1, after a Four-seam, vs. RHB: Four-seam 83% of the time (his usage 51%; 36 pairs). Out of sample, the likeliest pitch came 83% of the time.
These are tendencies, not tells: being predictable hasn’t cost pitchers runs league-wide, so treat them as a prior for the at-bat, not a plan.