Pitch Shape Explorer

Pitch Shape Explorer

Who is out-throwing their shape?

Filter the pitcher-season population by year, age, pitch type, handedness, and volume. Every row is scored by the same validated model.

The models are fixed, not refit. Filtering changes which pitchers you see, never what either model learned.
What the columns mean
IVB
Induced vertical break, in inches. How much backspin resists gravity. A fastball with high IVB does not rise, it just falls less than the hitter expects, which is why it gets swung under.
VAA
Vertical approach angle, adjusted for pitch height. How flat the ball is travelling when it reaches the plate, compared with other pitches thrown to the same spot. Flatter is generally better on a four-seamer.
Whiff
Share of swings the hitter missed entirely, as a percentage.
Shape exp
The whiff rate this pitcher's ball flight alone predicts, ignoring what actually happened to him. Think of it as what the pitch should do.
Grade
The shape expectation on the 20–80 scouting scale, where 50 is average and every 10 points is one standard deviation. Graded against the same pitch type in the same season, because a curveball's break is negative and a four-seam's is not, and league velocity has moved nearly two miles an hour across the window. This is not a stuff model. Stuff+, PitchingBot and the Driveline model use far more information and are fitted against different targets; this grades ball flight and nothing else. It knows nothing about location, sequencing or how a pitch plays off the rest of an arsenal, and location outranks every shape metric here by roughly fifty to one.
vs FB
How far a pitch sits from the fastball the pitcher throws most: miles per hour slower, inches of movement apart, and feet of release point apart. These are what the model uses to judge secondary pitches, on the reasoning that a slider misses bats by how late it stops resembling the fastball rather than by its own shape.
Edge
What the shape argues for beyond simple regression to the mean. Positive means the stuff is better than the results; negative means the results are running ahead of the stuff.
Why filtering
does not refit
The shape expectation is fit out-of-fold and grouped by pitcher, so no pitcher's expectation comes from a model that has seen him. From 2020 it also uses spin efficiency — what fraction of a pitch's spin is tilted to actually move the ball, rather than pointing down the flight path and doing nothing. That is a different measurement from spin rate, which the underlying paper found adds nothing once velocity is accounted for. Efficiency does add something: it improves the expectation on sliders (t = +5.9), cutters (+3.8) and four-seamers (+2.6), and nothing measurable elsewhere. It comes from Hawk-Eye and does not exist before 2020, so earlier seasons use ball flight alone; each season is fit separately, so nothing pools across that boundary. The edge is trained on the four most recent complete seasons rather than every season back to 2015, because the game drifts: league fastball velocity is up nearly two miles an hour since 2015, and including seasons that old made recent predictions measurably worse. Both are fit once, here, and then applied to every row. If the model refit itself on whatever you filtered to, a narrow selection would quietly report a model trained on a handful of rows with no validation behind it.

Progression candidates

The ball is doing what a more effective pitch does and the outcomes have not followed. Historically about half of a gap like this closes the next season.

whiff rate the shape predicts whiff rate he actually got

Regression candidates

Results are running ahead of what the pitch supports. Some of that is real skill the model cannot see, such as command, but some of it comes back.

whiff rate the shape predicts whiff rate he actually got

Shape says / results say

Each point is one pitcher-season. The diagonal is parity. Above the line, results beat what the shape supports; below, the shape is better than the results. Distance from the line is the part that partly reverses.

Ranked

Click any row to open that pitcher's full arsenal — every pitch he throws, side by side, with a movement chart. Click a column heading to sort. Edge is in whiff-rate points: +0.02 means shape argues for two percentage points more than mean reversion alone predicts.