Love, Death, and Pitching Robots, Pt. 2: Coming to Grips with New Technology

Whether there is a luddite rebellion, a scouting counter-revolution, or some other attempt at rolling back the technological advances the game has seen in recent years, last week I detailed why it makes sense for pitchers to adjust to new technology right now. Sure, hurlers could wait for a tech nullification or, in its absence, a new kind of tech to level the playing field, but as things stand, the scales are tipping.

Throughout the pitch-tracking era, hurlers have stood to benefit more than hitters from analytics thanks to highly customizable pitching plans. In recent years, motion-capture systems have helped optimize pitchers’ mechanics in addition to their repertoires. But this year, a significant number of teams have unearthed another use of limb-tracking software: in-game pitch tipping. Essentially, machine learning identifies subtle differences in muscle activation in real-time, typically due to different grips across pitch types, while the pitcher is still holding the ball in his glove. After a quick relay system, that information reaches the field, and the batter can then look to the dugout or a base coach for some indication of what’s coming.

To counteract this, I suggested pitchers mix up their wrist action and finger pressure right before release; by the time the KinaTrax systems pick up what’s going on, it’ll be too late. Sadly, I don’t have access to metrics like breadth of wrist action and finger pressure relative to grip, so I decided to come up with my own way of identifying pitchers who are already doing what I would recommend. I theorized that spin axis could be a proxy for grip, since the point about which a thrown baseball spins is heavily reliant on how the pitcher holds it. The direction and magnitude of movement is also closely related to spin axis, but there isn’t a one-to-one correspondence, so I hypothesized that pitchers with wide variability in movement despite minimal spin axis variability could be adding some extra mustard after separation.

Unfortunately, the pitchers I identified — namely, Martín Pérez and Kevin Gausman — achieved their wide ranges of movements relative to their small ranges of spin axes through a multitude of different grips (Pérez) or a pair of very different grips (Gausman). This used to be a good technique for keeping a batter guessing, given hitters’ reliance on spin axis (or seam orientation) as a cue for pitch movement. The importance of spin axis as a cue is supported by a number of studies, articles, and hitting clinics, as well as coaches like the Yankees’ Dillon Lawson, who in a recent interview described how he stresses to his hitters “instead of seeing the ball longer, try to see it earlier” so that they can pick up on spin and other signals. But where Pérez and Gausman may have previously found success by taking the spin-axis cue away across different pitch types, their differences in grip across those same offerings have left them vulnerable to motion-capture systems.

So I headed back to the drawing board, in need of a new way to control for grip in addition to spin. Ultimately, I settled on looking at each pitcher’s movement variability relative to spin axis variability within single pitch types rather than across entire arsenals — the thinking behind this being that it’s exceedingly rare for an individual pitcher to have two entirely different grips that result in a pitch of the same class.

Due to issues that arise when working with circular statistics (i.e., those given in degrees), I separated the angular spin axis into two components: sine and cosine. In my last piece, I found that on the arsenal-level, the standard deviation (SD) of sine was highly correlated with the SD of horizontal movement, and the SD of cosine with the SD of vertical movement. The same should go for the pitch-level:

And sine:

Something’s not right here. Even considering I took out the gyro pitches (those with a mean spin efficiency of 25% or less) that aren’t well understood by Statcast’s two-dimensional spin axis measure, as well as pitches with a measurement error (more than 100% spin efficiency by Alan Nathan’s calculations), the correlations aren’t anywhere near what I was getting last week. The major difference being that I’m categorizing by pitch type this time.

I’ve long taken issue with MLB’s classification system, which relies on a combination of supervised and unsupervised machine learning algorithms. By “supervised,” I mean that some of their algorithmic training involves using manually-entered pitch types, which the model then bases its future output on. This goes hand-in-hand with MLB’s stated desire to have pitch types on record that match what each pitcher (or their coach) calls the offering, but it’s likely what’s messing with the spin axis/movement correlations above — for example, there is a lack of standardization in terms of what a slider means across different pitchers and coaches.

For the task at hand, it’s worth considering MLB’s system because it’s informed by the pitchers themselves, who are in turn informed as to what pitches they’re throwing by their own grips and wrist action. Further, in the absence of direct quotes about pitch types, MLB uses still images of pitch grips to inform their classifications. Still, I chose to compare their classifications to my own model to serve as a check both on MLB’s system and the idea that there’s still a crucial relationship between spin axis and movement on the pitch-level. This will in turn inform how I go about investigating which pitchers are best equipped to survive the new wave of pitch tipping technology.

There have been myriad attempts at unsupervised classification, but I drew most from Professor Brian Mills’ work. His demonstration of model-based clustering appealed to me the most because his model chooses its own number of groupings for each pitcher (though I capped the potential number of pitches at nine for computational purposes and because those are the max that MLB has anyone — Yu Darvish — throwing), its own minimization criterion (from 14 different options) for each pitcher based on the shape of their pitch distribution across the relevant variables (standardizing for density instead), and where in the pitch-cluster space to start optimizing from. The only other limitation I imposed was a Bayesian prior in an effort to normalize the distribution of the number of pitch types per pitcher.

For MLB’s model, once each pitcher has a large enough big league sample, they are classified based on a customized neural network trained only on their pitches, which can make the groupings valid at least for comparisons within each pitcher. This is a similar idea to allowing for a (customizable) variety of different minimization criteria and starting points in the optimization problem; the major philosophical difference between the two models is that mine is choosing its own number of pitch types for each pitcher. It does so without any input from the pitcher, solely based on the distributions of velocity, spin rate, spin axis in three dimensions (sine, cosine, and spin efficiency), and horizontal and vertical movement.

The downside to this strategy is that there is no direct grip-based input, but it makes up for that in not requiring each pitcher to have a high level of self-awareness; a slight tweak they make for a slower or tighter version of a breaking ball, for example, might cause the pitch to diverge into two types in my model while MLB still considers it under one umbrella. On the other hand, there may be some instances in which a pitcher thinks they throw two different breaking balls, but my model disagrees.

It turns out that the former was more common than the latter. Among the 248 pitchers who threw at least 500 pitches through Saturday, my model identified 831 distinct pitch types thrown at least 100 times, while MLB has identified 773, with both figures including gyro pitches, of which there were 186 and 163, respectively. The added flexibility in my model resulted in better correlations between cosine and vertical movement (r-squared of .49 vs. .24):

And sine and horizontal movement (.28 and .07):

As well as, for gyro pitches, spin efficiency and vertical movement (.50 and .14):

And spin efficiency and horizontal movement were virtually the same (.58 and .56):

Note that the scales are different on all of these graphs — it’s more aesthetically pleasing that way, but it’s also the reason some of the differences in correlations look smaller or larger than they actually are. Another move I made for aesthetically-pleasing purposes was removing the one major outlier — a Rich Hill (no one is surprised) cluster that my model identified. Removing it actually worsened the correlations for my model, but that cluster was really out there: its horizontal movement and vertical movement SDs were both nearly four inches greater than the second-highest. It also had the highest standard deviation of spin axis sine and cosine, as well as spin efficiency.

This is an example of my model struggling with the fact that Hill has thrown eight changeups, 36 sinkers, and three sliders this season. All three of those pitches have at least something in common with a subset of his cutters (of which he’s thrown 208): the sinker and changeup have similar velocities (low 80s) and similar amounts of rise (from six to eight inches), while the slider has a similar spin efficiency (32.9% vs. the cutter’s 17.2%) and spin rate (2421 vs. 2305 rpms). It so happened that optimization in this case meant lumping all of those offerings together; generally, curtailing the number of pitch types produced better solutions for my model, but not in this case.

Despite the occasional wayward solution granted through unsupervised learning, I still think that my model (or a similar one) should be used in concert with MLB’s pitch classification system because the latter simply doesn’t do the best job at grouping pitches according to the characteristics that underlie pitch movement and batter pitch recognition. Yet, to ultimately identify the pitchers who can weather advances in pitch tipping, we need to look at movement variability in the context of spin axis (i.e., which pitchers can take away the spin cue for hitters) and grip; the value in MLB’s system is that it serves as one of the only ways to factor grip into the equation. Stay tuned for more on the pitchers who put together the best of both worlds.