Written By: Luke Smailes

Over the last few months, we’ve added two tabs to our TrackMan SYNC tile in the 643 interface – Pitching Models and Hitting Models. The Pitching Models tab features a trio of pitch quality models including Stuff, Command, and Expected Run Value (xRV). The Hitting Models tab features a model for hitter swing decisions and an estimated bat speed calculation. In this blog, I’ll provide a high-level overview of what each of the models represent and their construction, while also breaking down how to use/interpret them for scouting, player development and recruiting efforts.

TrackMan SYNC is an interactive visualization suite that brings hitting, pitching, catching, defense and umpire TrackMan data to life while also being mapped to Synergy video where available.

Pitching

With each of the pitching models, we scaled them into “plus” stats to provide intuitive context. With plus stats, each point above or below 100 represents one percentage point better or worse than the league average.

Each model is represented in two ways to tell two different stories. The interface displays Stuff+, Command+ and xRV+, but also pStuff+, pCommand+ and pxRV+. To use Expected Run Value as an example, xRV+ compares that pitch’s performance against all other pitch types. One good use-case for this method of scaling the pitch quality metrics is pitch usage optimization among a given pitcher’s arsenal. If a pitcher’s Slider xRV+ is 125 and his Fastball xRV+ is 98, you can quickly ascertain that his Slider is expected to perform 27% better. His Fastball is 2% worse than league average when compared to all other pitches, while his Slider is 25% better, and thus his pitch usage of the two should reflect that.

However, if this pitcher or his coach wants to see how that Fastball with an xRV+ of 98 stacks up against just other Fastballs, they can look at the pitch’s pxRV+ to understand the pitch’s quality just in the context of others with of the same pitch type. If this pitcher’s Fastball pxRV+ is 110, we can conclude that it’s 10% better than the average fastball, but still 2% worse than the average pitch thrown in D1 baseball.

Now, let’s get into the actual models. Here’s a layout of the respective predictors.

Stuff+

Stuff+ quantifies the “nastiness” of each pitch using only it’s physical characteristics and ignoring pitch location. A big question with Stuff models is what the target variable should be. What exactly does good stuff produce, and how do we train an algorithm to predict that? We decided that we wanted to target whiffs and called strikes, with the latter being to a certain extent. We know that a large component of generating a called strike is the pitch’s location and the game context. To control for this, we weighed called strikes based on the average in-zone swing rates in each respective count to proxy how much of a called strike is due to the pitch’s “stuff” characteristics and how much is based mainly on the game’s situational context. The result is a dependent variable that we refer to as the pitch’s Weighted Called Strike + Whiff, or wCSW.

We also include a seam-shifted wake component (or the presence of non-magnus induced movement and potentially “late break”) by comparing the observed spin axis to the inferred spin axis (sometimes referred to as the movement axis) for each pitch, and this was a significant predictor of wCSW.

Breaking balls and offspeed pitches produce a higher wCSW than fastballs on average, so they will have higher Stuff+ values on average.

Among pitchers in 2023 that threw at least 500 in-game pitches in front of a TrackMan, the top arsenals included Iowa’s Brody Brecht (128), Tennessee’s Chase Burns (124), Oregon’s Grayson Grinsell (123), Florida State’s Doug Kirkland (123), and LSU’s Ty Floyd (122).

At the pitch type level (min. 100 pitches), the top pitches included:

Fastballs: Nicholls’ Jacob Mayers (128), Tennessee’s Jake Fitzgibbons (126), Duke’s James Tallon (126), Clemson’s Reed Garris (125), and Wake Forest’s Sean Sullivan (125).

Breaking Balls: Iowa’s Brody Brecht (167), Florida’s Hurston Waldrep (152), Tennessee’s Chase Burns (150), Coastal Carolina’s Teddy Sharkey (150), and Texas A&M’s Brandyn Garcia (149). (all Sliders)

Offspeed: Georgia’s Charlie Goldstein (143), Boston College’s Joey Ryan (138), South Carolina’s Noah Hall (138), Stanford’s Joey Dixon (137) and Indiana’s Craig Yoho (136). (all Changeups, except Ryan’s Splitter)

Command+

Command+ is essentially the inverse of Stuff+ while also not attempting to control for any factors of deception. After controlling for the relevant context (count state and the batter/pitcher handedness relationship), the model predicts the pitch’s run value leveraging only its location coordinates.

For non-batted balls, each pitch is assigned a run value that reflects its impact on the game state. For batted balls, instead of converting its actual wOBA value to a run value, we convert the batted ball’s xwOBA to an “expected run value” of sorts to more accurately capture and train on the quality of contact instead of the result, a detail that we bring to xRV as well.

xRV+

xRV serves to be the most comprehensive pitch quality model of the three, and its main addition is how it attempts to model pitch deception. The factors that influence Stuff+ and Command+ are largely a science. You generally want to throw hard while differentiating speeds, make it move a lot, and locate it well. But deception is a lot less straight forward in terms of putting your finger on exactly what it is and what it looks like in the wild. It’s more of an art.

We model deception in five distinct ways:

  1. Weighted Pitch Mix

This component answers the question “does the count have a large impact on your pitch type selection as a pitcher?” Pitchers with non-count dependent pitch type usage are more deceptive because hitters are less able to sit on specific pitches to do damage by simply being aware of the count.

These pitch sequencing Plinko trees are a good way to visualize this component. The first image is the 2023 pitch mix of Arkansas’ Hunter Hollan, who you can see had very non-count dependent pitch type usage. Other than 3-0 and 3-1 counts, his pitch selection stayed pretty constant.

Hunter Hollan – Arkansas

Contrast that image with this one below from Southern Miss’ Tanner Hall, who was much more comfortable throwing his nasty Changeup with count leverage. The dichotomy between his Fastball and Changeup usage can be seen just based on the result of the first pitch.

Tanner Hall – Southern Miss

  1. Predicted Movement Differential

Collegiate hitters, to this point in their careers, have seen a lot of pitches over their lifetimes. Over that span, they develop expectations as to how a pitch should move based on the pitcher’s release characteristics. This has less to do with knowing a scouting report, and more about innate expectations that you develop from hitting against thousands of different pitchers. Pitchers with a high or “over-the-top” arm slot generally work north to south with their pitches (i.e., more extreme vertical break values than horizontal break), while a ¾ slot pitcher will generally be the opposite.

Pitchers that can deviate from these release:movement relationship norms create an abnormal look for opposing hitters, and are thus deceptive in the regard of unexpected movement. A good example of this is Indiana’s Craig Yoho, who releases his Slider from a high arm slot yet generates over 18 inches of horizontal break on it. Hitters had a tough time with it in 2023.

This process uses a Nearest Neighbor algorithm to compare each pitcher to their peers in terms of that release-to-movement relationship, and those with a larger distance between their “nearest neighbors” are more deceptive.

  1. Vertical Approach Angle vs. Expected

This is a similar concept to the predicted movement differential, but this focuses on vertical approach angle. VAA is a very important attribute on its own in the xRV model, but those who create angles that are unexpected based on release point and velocity tend to fool hitters as well.

The hitter-pitcher matchup in general largely boils down to creating and matching angles. Pitchers with abnormal approach angles given release parameters are deceptive. Many think of VAA only in the regard of flat Fastballs, but pitchers that create steeper or flatter than expected angles for hitters with their breaking balls (e.g., Brody Brecht’s Slider) or their offspeed pitches (e.g., Hurston Waldrep’s Split-Change) also grade well here.

  1. Release Anomaly Score

We could almost call this the “Sullivan Score” after Wake Forest’s southpaw Sean Sullivan. Prior to being drafted in the second round of the 2024 MLB draft, Sullivan generated a 37% whiff rate on his Fastball (league average = 18%) despite average velocity, below average vertical break, and a lot of pitches thrown right down the middle. However, he created funky angles with a low and left-exaggerated release point combined with a 7’2” extension.

xRV, without the inclusion of the Anomaly Score, didn’t think that Sullivan’s fastball should have generated anywhere near a 37% whiff rate given the aforementioned pitch characteristics. Though, watching him pitch and breaking down his results, it’s clear that he was tough to hit.

So, the Anomaly Score is just measuring how different your release point and resulting approach angles are from your peers, regardless of performance. Part of what made Sullivan so successful through the 2023 season was how unique the look he gave to hitters was.

This feature addition, however, wasn’t just an add to help boost Sullivan alone, as arms like Princeton’s Jacob Faulkner, Georgia’s Dalton Rhadans, and TCU’s Benjamin Abeldt represent others with high and influential Anomaly Scores, an attribute that ended with significant feature importance to the model.

If you can just be different, you create a larger margin for error with your stuff and command.

  1. Pitch Tunneling

The last factor of deception is pitch tunneling, and our version compares the trajectory of each pitch to the average trajectory of the pitcher’s primary pitch. The actual comparison happens at two distinct points through the trajectory of the pitch – the decision point (approximately 167 milliseconds from home plate, i.e., when the hitter must commit to swinging or not) and the point at which the pitch crosses the plate. When tunneling pitches, the theory here is that you want the two pitches to be as close as possible at the hitter’s decision point to maximize indistinguishability before dispersing as they cross the plate.

The ratio becomes the tunnel score for the pitch that’s ultimately fed into the gradient boosting model that spits out an xRV for each pitch.


In the interface, we display the breakdown of each metric in the interactive table below.

We also present interactive time series plots to analyze trends. This could be beneficial when analyzing a grip or arm slot change, or what an injury might be affecting.

 


Hitting

Swing Decisions

On the hitting side, the main component in the models tab is swing decisions, measured by SD+. To model swing decisions, we are strictly focusing on the decision to swing or not. We do not care about the results. Hitters can homer on awful swing decisions and strikeout on elite ones.

To model the value of each decision, we take “stuff” metrics such as velocity, movement, spin, etc., as well as pitch location and factors to control who player handedness and game context. We use these features to predict an expected run value for both a swing and a take on every pitch before taking the difference between the two to get a value of the decision.

Finally, we scale each decision value based on the probability that the given pitch is swung at in D1 baseball. This is done by training a separate classification model trained on the same attributes. The reasoning here is to ensure that “easy takes” are not overvalued due to the massive difference in the expected run value of a take vs. a swing, and vice versa for pitches that are chased but would qualify as “nasty”.

Some of the best swing decision makers in 2023 were Loyola Marymount’s Connor Blough, Bucknell’s Jacob Corson, Michigan State’s Trent Farquhar, Charlotte’s Shane Taylor, and Oregon State’s Travis Bazzana.

To display results in the 643 interface, we partition by a few different things to contextualize the results as best as possible. We start with results grouped by pitch type in a table that lists the SD+ values for each unique pitch type (and also on all pitches) along with other pieces of context to better help the user tell the story of why the swing decision results are what they are.

Bat Speed

The last addition to the new models’ tabs is the Estimated Bat Speed model. This is done by utilizing some of baseball physicist Alan Nathan’s batted ball research over the last few decades and applying it to TrackMan data. We end up with a bat speed estimate at the barrel of the bat that’s solely from TrackMan data, without the requirement of a bat sensor attachment or a marker-less biomechanical measurement. We’re not declaring this estimation to be more accurate than systems that are more explicitly measuring bat speed, but this allows us to wind up with estimates over a large sample size of swings.

Without getting into too much of the intricate math behind the calculations, the main task was to calculate each’s hitter’s unique collision efficiency (EA) based on their batted balls to date. We then use a linear decay function to estimate EA for each batted ball based on the batted ball parameters. This is important because a batted ball with a 100 mph exit velocity, for example, does not mean the same thing for each hitter, meaning that it could be the result of the maximum EA for hitter A, but in the 75th percentile for hitter B, so they need to be treated differently with respect to how much of that exit velocity is due to bat speed.

Some of the top average bat speeds through the 2023 seasons were LSU’s Dylan Crews, Georgia’s Charlie Condon, Notre Dame’s Vincent Martinez, Miami’s Yohandy Morales, and Wake Forest’s Brock Wilken.


The goal with these models is to tell a more in-depth story than what’s possible without leveraging machine learning and advanced modeling. They were curated over the last year and went through various iterations before we settled on the algorithms presented here. As we’ve delved into the intricacies of the newly integrated predictive analytics within TrackMan SYNC, we recognize that no model is ever “correct”, and it’s certainly possible that we iterate on these in the future if new insights emerge.