Baseball Research · Pitch-Level Study

The Shape of the Modern Pitch

Ball-flight geometry — not velocity, and emphatically not spin rate — is the pitch property that carries information about outcomes. It is also the only one measured well enough to judge a pitcher on a small sample.

Abstract

I break down every tracked pitch outcome from the 2015–2025 regular seasons into the unique contributions of velocity, spin rate, and pitch shape. Specifically: induced vertical break, horizontal break, and vertical approach angle. Shape carries 4.5× more out-of-sample explanatory weight than velocity for run value and 3.4× more for whiffs; residual spin rate, once velocity is partialed out, provides no distinguishable contribution to outcome even when granted first claim on shared variance. The result holds for all seven pitch types tested, even when adjusted for pitch location, the most relevant variable.

The evaluation of consequence is focused on reliability rather than effect size. Shape metrics carry split-half reliability of r ≈ 0.99 within a season; run value per pitch has r = 0.20. Because shape is measured almost without error, it forecasts a pitcher's next season better than his own results do below roughly 80 pitches, and continues to add information on top of results at every sample size. I state the boundary of the claim explicitly: shape does not replace outcome data on full seasons, and command — not any pitch property — remains the single largest determinant of pitch outcomes.

1Introduction

For fifteen years, a pitcher's value has been synonymous with two statistics that are easy to put on a scoreboard: how hard the ball was thrown, and how fast it was spinning. Velocity is intuitive. Spin rate arrived with the introduction of Statcast in 2015, quickly rising to analytical prominence attached to another new concept made accessible by the same technological breakthrough — the "rising" fastball — and became shorthand for pitch quality. A high-spin arm was a good arm.

While velocity and spin are undoubtedly important properties of any pitch, neither alone instructs a hitter on how to respond. A hitter cannot perceive in real time revolutions per minute, or the difference between a 93 and 95 mph fastball. What he perceives is the trajectory: where the ball appears to be going, how it deviates from the expected path, and at what angle it arrives. Velocity and spin matter only insofar as they produce that trajectory.

This study asks the question in the form that can actually be answered. Rather than asking whether spin "matters," which is unanswerable because spin correlates with everything else a pitcher does, it asks: after a model already knows everything else about a pitch, how much does each property add? That question has a clean empirical answer, and the answer is that pitch shape dominates. It also has a practical corollary that is more useful than the finding itself, and which occupies the second half of this paper: because shape is a near-deterministic property of how a pitcher throws, it is knowable from a handful of pitches, at a point when every outcome statistic is still noise.

I have attempted to write this so it can be readily fact-checked and discredited, if need be. Section 6 collects the results that constrain the thesis, including one that contradicts its strongest form. The code and the exact commands that produce every number are in Section 9.

2Definitions

This paper largely revolves around findings related to four pitch properties. Each is defined technically, for precision, and then again in plain language, for clarity.

2.1Induced vertical break (IVB)

The vertical deviation of the pitch, in inches, from the path a spinless ball thrown on the same initial vector would follow — that is, the movement caused by the ball's spin, with gravity's effect removed. Statcast reports it as pfx_z in feet; this paper converts to inches.

In plain English

Backspin counteracts the inherent downward effect of gravity. IVB measures how much of gravity's pull the spin cancelled out. A four-seam fastball with 18 inches of IVB doesn't actually rise — it just falls about a foot and a half less than the hitter's eye expects, which is why it's described as "riding" or "carrying."

2.2Vertical approach angle (VAA)

The angle, in degrees below horizontal, at which the ball crosses the front edge of the plate. I compute it from Statcast's release-point velocity and acceleration vectors by solving the constant-acceleration trajectory for the moment the ball reaches y = 17/12 feet, then taking the arctangent of the vertical over the horizontal velocity component at that instant.

In plain English

How steep the angle of descent is when the ball arrives at the plate. A hitter's swing travels slightly upward through the zone. A pitch that arrives flat — descending at, say, 4° instead of 6° — meets that swing plane at an awkward angle and gets missed or hit under. This is the real mechanism behind "the high fastball plays."

Crucially, VAA is not as simple as merely throwing high in the zone. Two pitchers can put the ball in identical spots but arrive at different angles, because of how tall they are, how far down the mound they release, and how much the ball carried on the way. Those factors contextualize the pitch, and they are the part that belongs to the pitcher.

2.3Horizontal break (HB)

Spin-induced horizontal deviation in inches, mirrored by handedness so that positive values always mean arm-side movement — otherwise right and left-handed pitchers cancel each other out in any pooled model.

In plain English

How much the pitch moves sideways, measured so that "toward the pitcher's own arm side" is denoted as positive whether he's a lefty or a righty.

2.4Residual spin rate

Raw spin rate (rpm) with the component linearly predictable from release velocity removed, computed within each season and pitch type.

In plain English

High-velocity pitchers inherently spin the ball more, simply because a faster arm imparts more rotation. Meaning, raw spin rate is partly a function of velocity. Residual spin removes that shared component by asking: given how hard this pitcher throws, is he spinning it more or less than you'd expect? That residual is the only version of "spin" that can be tested independently, and it is the spin metric used throughout this paper.

Figure 1 Two pitches, same release point, same destination, different geometry
strike zone mound home plate release steep VAA ≈ −6.5° flat VAA ≈ −4.2°
Both trajectories leave the same hand at the same speed and cross the plate at the same height — an identical box score, an identical location chart. The pitch with more induced vertical break resists gravity, so it can be thrown on a flatter line and arrives at a shallower angle to the hitter's swing plane. Everything this paper measures is the consequence of that difference.

3Data

The sample is every pitch tracked in the regular seasons of 2015 through 2025, pulled from Baseball Savant: 7,483,321 pitches. Run value (delta_run_exp) is present on 99.5–99.8% of rows in every season.

Cleaning removes pitches missing tracking fields and physically impossible readings (release speed outside 60–108 mph, spin outside 500–3600 rpm), retaining roughly 92% of pitches. The primary analysis uses four-seam fastballs — the pitch where spin most dominates the discussion — with all seven modeled pitch types examined separately in Section 5.4.

The 2026 season is deliberately excluded. Given that it was in progress at the time of writing, a partial season would skew the per-season trend analysis in §5.7 with incomplete data.

Spin measurement changed from Trackman to Hawk-Eye before the 2020 season. Values on either side of that boundary are not perfectly comparable, and as a result, per-season models are fit separately and never pooled across it.

The outcome variable is run value from the pitcher's perspective: the change in run expectancy attributable to the pitch, sign-flipped so positive always means good for the pitcher. A secondary outcome is whiff rate on swings. Both are used because they answer different questions — missing bats is not the same as preventing runs, and a pitch that largely misses bats but gets punished on contact is ultimately not a good pitch.

4Methods

4.1The attribution ladder

Feature blocks are added to a gradient-boosted model one at a time, each building on the last:

  1. Controls — count, plate location, distance from the middle of the zone, platoon matchup.
  2. Velocity — release speed and extension.
  3. Spin — residual spin rate.
  4. Shape — IVB, HB, VAA, horizontal approach angle, release height and width.

Because each block enters a model that already contains the previous ones, its improvement in held-out accuracy is its unique contribution. The ladder is then run in reverse, with spin entering before velocity, so neither variable's apparent importance is attributable to the order it was tested. Reporting both orderings is not optional: the gap between them is the collinearity.

4.2Guarding against leakage and self-deception

Four design choices do the real work, and each exists because its absence paints an incomplete picture:

  • Folds are grouped by pitcher. With random pitch-level splits, the same pitcher's fastballs appear in both training and test data, and the model memorizes the arm instead of learning the physics.
  • Every rung is scored on identical rows. Features have different missingness — release extension is null on a couple of percent of pitches — so a per-rung dropped-row policy would quietly hand later rungs an easier sample. All rungs use one common complete-case set.
  • Gains are paired across folds. Because the fold assignment is deterministic, each block's improvement can be differenced fold by fold, yielding a standard error and a t-statistic. This matters enormously: pitch-level run value has a standard deviation of 0.23 runs around a mean of zero, and at that noise level a block can "gain" +0.0004 R² by luck. Nothing in this paper is called real below t = 2.
  • Location enters first. Shape is only credited with what it explains after the model already knows where the pitch was thrown.

4.3Separating shape from location

The most serious objection to any shape-based finding is that vertical approach angle is mechanically tied to pitch height: any pitch at the top of the zone arrives flatter than the same pitch at the knees. A model conflating "shape" with VAA might rediscover that high fastballs play well.

To test this, VAA is regressed on plate height (quadratic, within season and pitch type) and the residual retained — how flat the pitch arrived relative to other pitches at the same height. Every result in Sections 5.4 through 5.6 uses this height-adjusted version.

5Results

5.1Attribution at the pitch level

On four-seam fastballs, with folds grouped by pitcher and gains paired across identical folds, shape is the only block that meaningfully influences the model.

Figure 2 Unique out-of-sample explanatory power added by each block
Velocity Residual spin Shape
Predicting run value
Velocity 0.00024 Spin 0.00004 Shape 0.00108 0 .0006 .0012 incremental CV R²
Predicting whiffs on swings
Velocity 0.0058 Spin 0.0017 Shape 0.0195 0 .010 .020 incremental CV R²
Four-seam fastballs, 2015–2025. Each block is added on top of all previous blocks, so the bar is that block's unique contribution. Note the two panels use different axis scales — whiff outcomes carry far more signal than run value at the pitch level. Exact values, standard errors, and t-statistics are tabulated below.
Table 1 — Attribution ladder, four-seam fastballs, both orderings
Block addedCV R²GainSEtVerdict
Run value · velocity entered first
Controls0.03667———baseline
+ Velocity0.03691+0.000240.00007+3.63real
+ Spin0.03694+0.000040.00003+1.15not significant
+ Shape0.03802+0.001080.00010+10.75real
Run value · spin entered first
+ Spin0.03664−0.000030.00006−0.46nothing
+ Velocity0.03694+0.000300.00008+3.63real
+ Shape0.03802+0.001080.00010+10.75real
Whiffs on swings · velocity entered first
Controls0.07718———baseline
+ Velocity0.08298+0.005800.00055+10.56real
+ Spin0.08467+0.001690.00013+13.08real
+ Shape0.10419+0.019520.00076+25.57real

5.2Spin is a proxy, not a cause

The decisive test is the reversed ladder. Residual spin was handed first claim on every scrap of variance it shares with velocity and shape — the most generous treatment the design can give it — and it returned −0.00003 ± 0.00006 (t = −0.46). Nothing. Velocity, entering afterward on the leftovers, still cleared its noise.

This is not a claim that spin is physically irrelevant; it is a claim about what spin rate tells you that you didn't already know. Spin's entire contribution is already expressed in the movement it produced. Once you can see IVB and VAA, the rpm figure is redundant.

Why this matters mechanically

Spin only moves a baseball when the spin axis is perpendicular to the direction of travel. Pure gyroscopic spin — a bullet-style rotation — does not move the ball at all. A pitch can register 2,800 rpm and break like a brick. This is why "high spin" is an unreliable indication of the value of a pitch, and why the movement it produces is the honest measurement.

5.3Shape is not location in disguise

The strongest available objection: VAA is confounded with pitch height, so perhaps the shape block is a laundered location effect. Replacing raw VAA with the height-adjusted residual described in Section 4.3 leaves the result essentially intact.

Table 2 — Shape block with raw versus height-adjusted approach angle
Shape block specificationGainSEtRetained
Raw VAA+0.000990.00011+9.38—
Height-adjusted VAA+0.000920.00009+10.3693%

Ninety-three percent of the shape effect survives the removal of everything attributable to where the pitch was thrown, and the t-statistic actually improves — stripping the location component removes noise as well as signal. Shape is a property of the pitcher, not of his aim.

5.4The result generalizes to every pitch type

Skeptics of fastball-only findings are right to be skeptical, since the meaning of spin is at odds on a four-seamer and a curveball. The ladder was therefore refit independently on all seven pitch types with adequate sample.

Table 3 — Incremental CV R² by pitch type (t-statistic in parentheses)
PitchRun valueWhiffs
VelocitySpinShapeVelocitySpinShape
Four-seam+.00025 (2.2)+.00020 (1.9)+.00085 (6.0)+.00540 (9.8)+.00129 (4.3)+.01923 (24.8)
Sinker+.00008 (0.5)−.00000 (0.0)+.00217 (9.9)+.00300 (3.4)+.00061 (3.4)+.01085 (6.2)
Cutter+.00034 (2.4)−.00014 (−2.3)+.00117 (6.7)+.00184 (5.9)−.00031 (−1.2)+.01145 (12.4)
Slider−.00030 (−3.3)+.00012 (1.3)+.00099 (23.9)+.00102 (4.4)−.00009 (−0.4)+.00563 (14.4)
Sweeper−.00031 (−1.4)−.00027 (−2.4)+.00168 (6.7)+.00206 (4.6)−.00026 (−1.0)+.00755 (6.8)
Curveball+.00010 (1.2)−.00021 (−1.6)+.00051 (2.1)+.00496 (7.5)+.00025 (0.6)+.00561 (10.5)
Changeup−.00025 (−2.7)+.00012 (0.9)+.00200 (9.1)+.00105 (6.5)+.00032 (1.8)+.00463 (7.4)

Shape is the largest block in all fourteen pitch-type-by-outcome combinations, and significant in every one. Residual spin fails to clear significance in eleven of fourteen and is negative in six. Velocity's contribution to run value is inconsistent across pitch types — genuinely negative on sliders and changeups, where the pitch's job is deception rather than speed — while remaining reliably positive for whiffs.

5.5Reliability: the finding that changes evaluation

Everything to this point concerns what explains the outcome of a pitch. Evaluation asks a different question — what tells me how good this pitcher is, not his individual pitches — and the answer is governed less by effect size than by measurement error.

Splitting each pitcher-season randomly in half and correlating the halves against each other separates signal from noise without needing a second season. The result is a difference not of degree but of kind.

Figure 3 Year-over-year correlation: what actually persists about a pitcher
Shape & delivery Outcome statistics
Release height .956 Extension .949 VAA (adjusted) .940 Velocity .925 Residual spin .915 Horizontal break .886 Induced vert. break .836 Whiff rate .628 Run value / pitch .216 0 0.25 0.50 0.75 1.00 correlation between consecutive seasons (1,889 pitcher-season pairs)
Every shape and delivery metric persists from one season to the next at r ≥ 0.84. Whiff rate manages 0.63; run value per pitch, the statistic closest to "how well did he actually pitch," manages 0.22. Within a single season the gap is starker still: split-half reliability is ≈ 0.99 for shape metrics versus 0.20 for run value.
Table 4 — Reliability, four-seam fastballs (3,348 pitcher-seasons; 1,889 consecutive pairs)
MetricKindSplit-half rSpearman–BrownYear-over-year r
Release heightshape1.0001.0000.956
Extensionshape0.9991.0000.949
Release speedshape0.9991.0000.925
VAA (height-adjusted)shape0.9990.9990.940
Horizontal breakshape0.9980.9990.886
Residual spinshape0.9980.9990.915
Induced vertical breakshape0.9960.9980.836
Whiff rateoutcome0.6580.7940.628
Run value per pitchoutcome0.2000.3330.216
What reliability means here

If you split a pitcher's season in half at random and his induced vertical break in the first half predicts how his second half shakes out almost exactly (r = 0.996), the metric is essentially free of measurement error. If his run value in the first half barely tells you anything about the second (r = 0.20), then most of what you are looking at when you look at his results is luck — sequencing, defense, where the ball happened to land.

This is the whole argument in one line: shape is a measurement, results are a sample.

5.6The information crossover

The practical payoff follows directly. The practical payoff follows directly. Because shape is measured nearly exactly from the very first pitch while outcome statistics need hundreds of pitches to begin accumulating into something resembling a statement on a pitcher's ability, there must exist a sample size below which shape is the better predictor of a pitcher's future.

To locate it, each pitcher-season was subsampled to exactly N four-seamers, metrics computed from only those N pitches, and next season's performance predicted from them. The target is always measured on a full sample, so only the quality of what we know now varies.

Figure 4 Predicting next season's whiff rate, by how much you have seen
Shape only His own past results
0.40 0.30 0.20 0.10 0 crossover ≈ 80 pitches shape past results 30 60 125 250 500 four-seamers observed (log scale) — each point built from exactly this many pitches
Shape's predictive power is flat: it is just as informative from 30 pitches as from 500, because it is measured without error the moment the pitch leaves the hand. Outcome statistics start near-useless and improve steadily as the sample accumulates, overtaking shape at roughly 80 pitches. Below that line, a pitcher's own results tell you less about his future than his ball flight does.
Table 5 — Out-of-sample R² predicting the following season
Pitches seenNext-season whiff rateNext-season run value
ShapeResultsBothShapeResultsBoth
300.1880.0940.2310.0360.0220.046
600.1930.1600.2660.0370.0380.057
1250.1920.2500.3190.0440.0620.078
2500.2060.3210.3660.0510.0870.099
5000.1880.3620.3780.0370.0950.087

Facts deserve equal billing. Below roughly 80 pitches, shape beats results — at 30 pitches, by a margin of two to one. Above that, results win, the margin growing as the sample increases. But in every row, combining the two provides more value than either alone: at a full 250-pitch sample, adding shape to results lifts next-season whiff prediction from 0.321 to 0.366, a 14% relative improvement that is free for the taking.

5.7League context: velocity has compressed

One further trend reframes the velocity conversation. Between 2015 and 2025, average four-seam velocity rose from 93.11 to 94.48 mph (t = +9.40). Over the same window, the standard deviation across the league fell from 2.82 to 2.53 mph (t = −7.22).

The value of one additional mph shows no significant trend (t = −0.83). What changed is the supply. Velocity is worth what it always was; there is simply far less variation left within the league, so it separates one pitcher from another less than it used to. That is the defensible sense in which velocity "matters less" — not that hitters adapted, but that the style of pitching has converged uniformly toward higher velocity.

6Threats to validity, and where the thesis had to be revised

This section exists because a paper that only reports confirmations of its hypothesis lacks credibility. Three findings below genuinely constrain the argument, and the first two forced me to narrow it.

6.1Command dwarfs everything, including shape

In permutation importance on run value, distance from the middle of the zone scores 0.050. The best physical trait, induced vertical break, scores 0.00097 — roughly fifty times smaller. Ball-strike count is the second largest factor. On any honest accounting, where the pitch is thrown matters vastly more than what the pitch does.

Revision forced by this result

The claim is not "shape is the most important thing about a pitcher." It is: among the intrinsic properties of a pitch — the things a pitcher brings with him to each game, rather than executes on a given night — shape dominates. Command sits outside that set and outranks all of them. Any framework that leads with shape and ignores command is incomplete, and this paper's framework says so explicitly in Section 7.

6.2Shape does not replace results on full seasons

My initial expectation was that shape's reliability advantage would make it the superior evaluation input across the board. It does not. At 250 pitches, past results predict next-season whiff rate at R² = 0.321 against shape's 0.206; at 500 pitches it is 0.362 against 0.188. Results win comfortably once there is a significant enough sample.

The reason is not that shape degrades — it is flat — but that outcome statistics contain everything: shape, command, sequencing, deception, the pitcher's ability to execute under pressure. Given enough observations to average out the luck, the broader perspective wins.

Revision forced by this result

Shape is not a replacement for outcome data. It is the correct input when the sample is small, and an additive input once the sample has grown larger. Any presentation of this research that claims shape supersedes results is overselling it, and will be dismantled by the first analyst who runs the full-season comparison.

6.3An earlier version of this experiment was wrong

The first crossover test filtered to pitcher-seasons with at least N pitches rather than subsampling to exactly N. Because most seasons passing a 30-pitch filter still contain hundreds of pitches, that design never tested small samples at all, and it produced the false conclusion that results beat shape everywhere. The corrected version — subsampling each pitcher-season to exactly N — produces Figure 4. I report the error because the corrected and uncorrected designs give opposite answers, and anyone reproducing this should know which one they are running.

6.4Remaining limitations

  • True spin efficiency is not measurable from public data. Statcast's spin_axis is inferred from observed movement, not measured from Hawk-Eye's 3D axis. Every public "spin efficiency" figure, including any derivable here, is merely an inference. This is the single largest gap between this analysis and what a club can do internally.
  • Everything here is observational. Nothing establishes that adding two inches of IVB to a specific pitcher's fastball would improve his results.
  • Shape features are collinear with each other and with velocity. The block's contribution is well identified; individual coefficients within it are not, and should not be quoted as effect sizes.
  • The height adjustment removes a main effect, not an interaction. Chamberlain (2021) shows that VAA's influence is concentrated at the vertical extremes of the zone and that VAA and pitch height interact non-linearly. The quadratic residualization in §4.3 strips out the average relationship between approach angle and height, but it does not model that interaction. A specification allowing the VAA effect to vary by zone region would likely find a larger shape contribution than reported here, so the §5.3 figure is better read as a conservative floor than as a point estimate.
  • Sequencing and tunneling are absent. A fastball plays off the pitch before it. Some of what is credited to shape is plausibly the arsenal-level interaction that shape enables.
  • Survivorship. Reliability figures come from pitchers who threw at least 250 four-seamers — pitchers good enough to keep being used.
  • The Trackman–Hawk-Eye transition before 2020 makes spin values non-comparable across that boundary; models are fit per season, never pooled across it.
  • Absolute effect sizes are small. Pitch-level run value is mostly irreducible noise; a shape gain of 0.00108 R² is real but modest. It is meaningful because it accumulates over thousands of pitches, not because any single pitch is well predicted.

6.5The adjusted thesis

Stated so it can be checked

Pitch shape is the dominant pitch-intrinsic driver of outcomes, ahead of velocity and far ahead of spin rate, on every pitch type and both major outcome measures. It is also the only pitch property measured reliably enough to support judgement on a small sample, which makes it the correct primary input below roughly 80 pitches and a persistent additive input above that. It does not outrank command, and it does not replace outcome data on full seasons.

7A framework for evaluating the modern pitcher

The findings above imply a specific, ordered procedure. The organizing principle is that the right evaluation input for a given circumstance depends on sample size, and the crossover falls between the 60- and 125-pitch samples, interpolating to roughly 80.

7.1Triage by sample size

Under ~80 pitches of the pitch type
Read the shape. Ignore the results.

Applies to: a prospect's first look, a reliever early in the year, a pitcher who has just added or redesigned a pitch, a rehab assignment, a trade-deadline target you have limited recent data on.

At this sample, his results carry roughly half the informational value his ball flight does. Whiff rate, ERA, and run value are all still inconclusive at this point. Shape, by contrast, is already fully measured — IVB and VAA are as knowable after 30 pitches as after 3,000.

What to do: rank the pitch on IVB, height-adjusted VAA, and horizontal break against the league distribution for that pitch type. Treat the outcome line as essentially uninformative.

Roughly 80 to 400 pitches
Blend, weighting results upward as they accumulate.

Results have overtaken shape but have not rendered it obsolete, and the combination beats either alone by a wide margin — at 250 pitches, shape plus results predicts next season 14% better than results alone.

What to do: use both. Where they disagree, the disagreement is the signal: a pitcher with elite shape and poor results in this window is usually a command or sequencing problem, which is a coachable problem. The reverse — ordinary shape, excellent results — is the profile most likely to regress.

Full season and beyond
Lead with results. Use shape to explain and to forecast change.

With 500+ pitches, a pitcher's in-game outcomes are the better single predictor, and it is not close. Shape's role shifts from estimating talent to explaining it and predicting change.

What to do: lead with the outcome measures, but keep shape in the model — it still adds value to the data. Use shape as the early-warning system: because it is measured exactly, a two-inch drop in IVB or a flattening release is detectable in a single outing, long before it shows up in results.

7.2What to measure, in priority order

  1. Command first, always. Location outranks every pitch property by roughly fifty to one. Any evaluation that starts with shape has already skipped the biggest term. Shape tells you the ceiling of the pitch; command tells you his ability to reach it.
  2. Induced vertical break. The single most important shape variable and the top physical feature in permutation importance. Read it relative to the pitch type — high IVB is an asset on a four-seamer and a liability on a sinker.
  3. Height-adjusted vertical approach angle. The largest standardized coefficient of any physical trait (0.0123 runs per standard deviation, an order of magnitude above velocity). Use the height-adjusted version — raw VAA's statistical value can be corrupted by location.
  4. Horizontal break, handedness-mirrored. Always convert to arm-side positive, or lefties and righties will cancel each other out in any pooled comparison.
  5. Release height and extension. Nearly perfectly stable year to year (r = 0.956 and 0.949), these are the geometric inputs that determine approach angle. Extension also buys effective velocity.
  6. Velocity. Real and consistently positive for whiffs, but a weaker differentiator than it was — the league's spread has compressed 10% since 2015. Worth about 5.4 runs per standard deviation over a starter's season of four-seamers.
  7. Spin rate — as a diagnostic only. It adds nothing to outcome prediction once movement is known. Its legitimate use is mechanical: a spin rate that is high relative to the movement it produces indicates inefficient, gyro-heavy spin, which is a coaching target, not an evaluation input.

7.3Errors this framework is designed to prevent

  • Ranking pitchers by raw spin rate. It contributed nothing in eleven of fourteen pitch-type-by-outcome tests, and was negative in six.
  • Comparing shape across pitch types. Eighteen inches of IVB is elite on a four-seamer, but a mistake exploited by batters on a sinker. Every comparison must be within pitch type.
  • Using VAA unadjusted for location. Roughly 7% of its apparent value is location, and reporting it unadjusted invites the accusation that you have rediscovered the high fastball.
  • Trusting a reliever's ERA. At 60 innings, run value per pitch has a year-over-year correlation of 0.22. It is close to noise. This is the single most common evaluation error the reliability table exposes.
  • Treating shape as destiny. Shape sets the ceiling. Command determines how much of it is realized, and command is the larger term.

7.4A worked reading

Consider a 26-year-old reliever called up in August who has thrown 84 four-seamers with a 4.91 ERA and a 19% whiff rate. His fastball shows 18.9 inches of IVB and a height-adjusted VAA a full degree flatter than league average, at 95.1 mph.

The framework's reading: the sample is below the crossover, so the ERA and whiff rate carry less information than the ball flight does — the results line should be heavily discounted. The shape profile is well above average on the two variables that matter most, on metrics already measured almost exactly despite the small sample. The correct inference is a pitcher whose stuff is real and whose results have not yet caught up, with the open question being command — which is where the remaining evaluation effort should be, because it is both the largest term and the one this data does not settle.

8Conclusion

The industry spent a decade organizing pitcher evaluation around two numbers that are one step removed from the property that actually causes them. Velocity and spin are inputs to a trajectory; the trajectory is what a hitter faces. Measuring the trajectory directly — induced vertical break, horizontal break, and the angle at which the ball arrives — explains four and a half times more variation in run value than velocity does, and residual spin adds nothing at all once movement is accounted for.

But the more useful finding is about measurement error rather than effect size. Shape is a property; results are a sample. A pitcher's induced vertical break is known almost exactly after thirty pitches, while his run value is barely known after a full season. That asymmetry, not the size of any coefficient, is what should change how pitchers are evaluated. This change would be felt precisely where evaluation is hardest and most valuable: the prospect, the call-up, the new pitch, the trade target with sixty pitches of recent data.

The honest boundary is worth restating. Command still outranks every pitch property by a wide margin, and a full season of a pitcher's own results still beats his shape profile. Shape is the best available answer to a specific and common question — what do I make of this arm before the results mean anything? — and a persistent, additive input after that.

9Reproduction

Every figure in this paper is produced by the analysis code, from data downloaded directly from Baseball Savant. No number here was entered by hand.

All code is available at github.com/scobobo/modern-pitcher-evaluation, released under the MIT licence. The Statcast cache is not committed — it is rebuilt by the fetch script below.

Environment

pip install pandas numpy scikit-learn matplotlib pyarrow
pip install --no-deps pybaseball pygithub pyjwt Deprecated wrapt pynacl
pip install requests beautifulsoup4 lxml tqdm attrs python-dateutil cffi

pybaseball pulls in cryptography through a Retrosheet module this project never uses; the two-step install above avoids that build.

Fetch the data (~7.5M pitches, one parquet file per season)

python src/fetch.py --seasons 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025

Reproduce each section

# Section 5.1 — attribution ladder, both orderings
python run_analysis.py
python run_analysis.py --target is_whiff

# Sections 5.3–5.5 — VAA robustness, pitch-type generality, reliability
python run_paper_analysis.py

# Section 5.6 — the sample-size crossover
python run_sample_size_test.py

Results are written to output/ as CSV. The verdict function prints whichever conclusion the numbers support, including "hypothesis not supported" — it was written before the results were known and was not adjusted afterward.

10References, data, and tools

10.1Data and software

  • Statcast pitch-tracking data. MLB Advanced Media, via Baseball Savant (baseballsavant.mlb.com). 2015–2025 regular seasons, retrieved August 2026. Tracking hardware: TrackMan through 2019, Hawk-Eye from 2020.
  • pybaseball. Open-source Python library for retrieving baseball data, used here for the Statcast pulls.
  • Analysis code for this paper. github.com/scobobo/modern-pitcher-evaluation, MIT licence.
  • scikit-learn. Pedregosa et al., "Scikit-learn: Machine Learning in Python," Journal of Machine Learning Research 12 (2011), 2825–2830. Provides the gradient-boosting, ridge, and cross-validation implementations.
  • Run expectancy framework. Tango, Lichtman & Dolphin, The Book: Playing the Percentages in Baseball (2006), the basis for the change-in-run-expectancy outcome used throughout.
  • Baseball aerodynamics. Alan M. Nathan's published work on the physics of the baseball in flight, which underlies the Magnus-effect reasoning in §2.1 and the gyro-spin caveat in §5.2.

10.2Prior work on approach angle

Vertical approach angle was not developed here. It was established as a public analytical concept largely through the following, which informed the framing in §2.2 and the height-adjustment procedure in §4.3:

  • Alex Chamberlain, "Where Vertical Approach Angle Seems to Matter Most," FanGraphs, 7 January 2021. Finds that VAA's effect on whiffs is concentrated at the vertical extremes of the strike zone rather than through its middle, and that VAA and pitch height interact non-linearly — a flat pitch thrown low does not behave like a flat pitch thrown high.
  • Alex Chamberlain, "A Visualized Primer on Vertical Approach Angle (VAA)," FanGraphs, 1 February 2022. Documents the strong correlation between VAA and pitch height that motivates the height-adjusted measure used throughout this paper, and maintains a public VAA Above Average leaderboard built on the same reasoning.

The height-adjusted VAA used here (§4.3) is a simpler construction than Chamberlain's leaderboard measure — a quadratic residual within season and pitch type — and is used because it can be recomputed from raw Statcast fields by anyone reproducing this analysis.

10.3Author contributions and use of AI tools

I directed, interpreted, revised, and published this study; parts of its implementation and drafting were produced with AI assistance. The division of labor is set out below so readers can weigh it for themselves rather than guess.

What I contributed

  • The research question and the original hypothesis — that spin rate had displaced velocity as the meaningful measure of a pitch. This study tested that hypothesis and rejected it.
  • The decision to redirect the study toward pitch shape once the evidence pointed there, and to define its scope: a pitch-level attribution analysis, a reliability analysis, and a practical evaluation framework, rather than a descriptive result alone.
  • The falsifiability standard the project ran under. I required from the outset that any finding capable of discrediting the thesis be surfaced and the theory adjusted to fit the evidence, rather than the reverse. Sections 6.1 and 6.2 exist because of that requirement, and both narrowed the claim I started with.
  • The decision to publish the disconfirming results — the command finding that outranks shape fifty to one, the full-season comparison that shape loses, and the experimental design error in §6.3. The argument needs none of them; it is more honest for all three.
  • Interpretation of the results, the structure of the evaluation framework in §7, the boundary conditions in §6.5, and editorial revision of the manuscript.

Editorial contribution

Jimmy Stanley reviewed the full manuscript and edited it throughout. The argument is clearer and more precise for it. Responsibility for the analysis, the claims, and any remaining errors is mine.

What AI assistance contributed

Anthropic's Claude, via Claude Code, was used under my direction to build the data-acquisition, cleaning, and feature-engineering pipeline; to implement the grouped cross-validation and paired-fold significance testing described in §4.2; to execute the robustness checks the thesis had to survive, including the location confound in §5.3 and the corrected sample-size experiment in §6.3; and to generate the figures, tables, and draft prose, which I revised.

What I hold myself accountable for

I can defend every methodological choice in §4: why cross-validation folds are grouped by pitcher, why every rung of the ladder is scored on identical rows, why incremental gains are paired across folds rather than compared as point estimates, and why the attribution ladder is run in both orderings. Each exists because its absence manufactures false positives — and §6.3 documents where an earlier version of this analysis fell into exactly that trap, reported here rather than quietly corrected.

Using capable tools well is not the same as outsourcing judgement. The tooling accelerated the implementation; the decisions about what to test, what would count as evidence, what to believe, and what to retract remained mine throughout.

The verification standard

Readers are invited to verify rather than to trust. Every quantitative claim in this paper is reproducible from the open-source code in §9, against public data, by anyone with a laptop. The analysis was deliberately constructed so that its central hypothesis could fail — the reporting function prints "hypothesis not supported" when the numbers say so, and was written before the results were known.

That is the standard by which this work should be judged, and it is a stronger claim than authorship: none of it requires taking my word for anything.