From Stuff to Strategy: Improving MLB Pitch Profiles and Optimizing Usage
- Sean
- 39 minutes ago
- 21 min read
The 2025 Major League Baseball (MLB) season has been over for a few weeks now, but players’ quest for improvement never really ends. Even as many enjoy a bit of rest and recovery, they’re also beginning to plan how to elevate their performance in 2026. For pitchers, that often means exploring how to evolve the profiles of their current pitches, or even add entirely new ones, to better miss bats, induce softer contact, or improve command.
I’ve always been fascinated by how pitchers, coaches, trainers, and analysts decide what exactly should change in a pitch to make it more effective. That curiosity led me to build a data driven approach to identify which pitch profiles a pitcher should emulate to improve performance. That’s the project I’ll share with you here.
I pursued this for two reasons. First, I enjoy the strategic side of baseball and understanding how a pitcher’s intent, such as getting more groundballs with a specific pitch, can be achieved through targeted pitch profile adjustments. Second, I wanted to create a framework that is simple, repeatable, and adaptable for anyone interested in doing similar work.
In the sections that follow, I’ll cover three things:
A brief look at relevant public research on pitch shapes, arsenal development, and pitch mix usage.
A detailed breakdown of my methodology for identifying comparable pitch profiles, simulating their impact, and suggesting pitch mix adjustments.
Real examples using current White Sox pitchers and this past year’s Cy Young Award winners.
Let’s get into it!
Relevant Public Work
Diving into public pitching research can be quite overwhelming due to the quality and quantity of work that is available online. I am not going to attempt to perform a literature review of the area, but instead will lightly cover work related to the attributes of pitches and pitch design, how pitches “play” off each other within an arsenal and why that is important, and how pitchers seek to sequence and optimize their pitch usage.
I wouldn’t be able to take myself seriously if I didn’t start off this section talking about Stuff+. Stuff+ is a model that evaluates the physical “nastiness” of a pitch based solely on measurable traits like release point, velocity, vertical and horizontal movement, and spin rate, comparing each pitch to a pitcher’s primary fastball. Trained on run values, it captures deception and weak contact ability and uses nonlinear relationships to identify which pitch shapes most consistently generate strong outcomes for pitchers. Stuff+ has become a popular model for established and up-and-coming baseball analytics professionals to replicate and put their own spin on due to the availability of the necessary data via MLB as well as the fairly simple feature engineering and modeling processes that can be used to generate a desired outcome. It’s also worth noting that Stuff+ isn’t the only model that is used to evaluate pitchers. Location+, which is a count and pitch type adjusted judge of a pitcher’s ability to put pitches in the right place, and Pitching+, which uses the physical characteristics, location, and count of each pitch to try to judge the overall quality of the pitcher’s process, are two that are well established within the analytics community. While these models help us understand what makes a pitch effective, they do not necessarily tell us how to change a pitch to reach that level.
Bridging that gap between evaluation and actionable development is where pitch design research becomes essential. In a 2021 article for Driveline Baseball, Sam Bornstein demonstrated how modern tracking technologies such as Rapsodo, TrackMan, and Yakkertech allow analysts and coaches to precisely measure pitch characteristics and uncover key physical tradeoffs, particularly between velocity, movement (sweep and depth), and spin efficiency. Building on those measurements, Driveline introduced a model that predicts how much velocity a breaking ball should lose based on its spin components and a pitcher’s fastball profile. This provides an objective baseline for determining whether a pitch is underperforming, performing as expected, or exceeding expectations for its specific shape. Complementary work by Kevin Goldstein at FanGraphs on fastball shape and pitch movement further reinforces the importance of how deviation from a pitcher’s expected movement profile, driven by arm slot and spin, can meaningfully alter outcomes such as whiff rates, contact quality, and batted-ball profiles. Together, this body of research forms the foundation for the tradeoff aware individualized pitch design framework I apply in this project.
Recent public research makes it clear that evaluating individual pitches is only part of understanding what makes a pitcher effective. Models like Stuff+ or Baseball Prospectus’ StuffPro can describe the quality of a single pitch quite well, but they often miss the value created when those pitches work together. Stephen Sutton-Brown, of Baseball Prospectus, highlighted this gap by discussing how many pitchers consistently outperform the sum of their individual pitch grades because their repertoires create deception, uncertainty, and timing disruption that individual “stuff” models cannot measure. A pitcher can have solid pitch shapes, but the real advantage can come from how difficult it is for a hitter to recognize those shapes early or anticipate where the ball will finish even after multiple trips through the order.
Sutton-Brown frames this idea through two pathways. Pitchers with deeper arsenals hold their performance longer as hitters see them again because familiarity builds more slowly when a pitcher can offer multiple looks. They also perform better when their pitches share similar release cues and early trajectories but still separate cleanly in movement or velocity later in flight. That combination forces hitters into worse decisions and increases the likelihood of whiffs or weak contact. In short, a pitch has value on its own, but it has even more value when it enhances the deception or effectiveness of the rest of the repertoire.
A similar perspective shows up in a recent project by Alex Britton, Josh Hejka, Jack Lambert, and Marek Ramilo of Driveline Baseball, that was presented at Sabersminar 2024. They introduced Arsenal+, a new pitching evaluation model designed to overcome the limitations of traditional “stuff” models by assessing pitches within the context of a pitcher’s entire repertoire. The core idea is that the combined value of a pitch arsenal is greater than the sum of its isolated parts, reflecting how pitches interact to deceive a batter. The model’s methodology involves four steps: creating average 3D trajectories for each pitch, using Kernel Density Estimates (KDEs) to profile each pitch’s plausible distribution, calculating entropy to proxy batter expectation, and finally quantifying two key interaction effects: Mix+, which measures the arsenal’s breadth and diversity by calculating the distance between pitches, and Match+, which measures the arsenal’s depth and deception by rewarding pitches that “tunnel” or mirror each other for a longer duration. Ultimately, Arsenal+ offers applications beyond player valuation, including optimizing pitch usage to combat the “familiarity effect” and systematically testing pitch design recommendations.
The last topic I wanted to explore was how others have approached evaluating or optimizing pitch usage. There isn’t much public facing research that deals directly with pitch mix optimization, although there has been a fair amount of work on pitch sequencing. In 2021, Patrick Brennan used Shannon entropy to show that unpredictability increases when a pitcher has more pitch types and uses them more evenly. Ajay Patel and I took a similar approach with our Pitch Type Sequencing Similarity Ratio. Even with these efforts, there are still not many examples of people trying to actually determine an optimal mix of pitches.
There are a few recent exceptions. Robert Frey shared a series of tweets where he suggested optimized pitch usage plans for specific pitchers. He described his method as a nonlinear optimization model that selects pitch mixes within realistic bounds, incorporates handedness splits, and aims to maximize expected run value for the pitcher. In the Arsenal+ presentation I mentioned earlier, Ramilo also mentioned Driveline’s Paint Mixer framework. Paint Mixer generates pitch usage recommendations by combining Arsenal+ with other important key metrics. The goal is to help pitchers break familiar patterns that hitters learn over time and to take advantage of a “buyback effect”, where certain secondary pitches make a pitcher’s primary offerings more effective.
Methodology: What I Did
As a reminder, my goal was to build a data driven approach to identify which pitch profiles a pitcher should emulate to improve performance. For this iteration, I anchored my evaluation on Whiff Rate (Whiff%) and used Groundball Rate (GB%) as a guardrail. I know that many projects use run value as their north star metric, but I wanted to ensure explainability and I feel that saying that we expect a lift in expected Whiff% (xWhiff%) is easier to grasp than we expect an improvement in run value. The beauty of this framework is that you can swap in any metrics that you’d desire. Without further ado, below is a concise framing of my methodology and below that, I will go into more detail.
Identify a target pitch to improve. For a given pitcher and handedness split, I flag pitches with below average whiff rates and pick one to work on.
Find realistic “inspiration” pitches. I narrow down the universe of potential pitch profiles to pitches thrown by similar pitchers, based on release characteristics and historically realistic year-over-year pitch attribute changes.
Simulate counterfactual pitch profiles. I swap those candidate pitch profiles into the pitcher’s 2025 pitch data and use models to estimate whiff and groundball rates.
Evaluate and select improved profiles. I compare modeled actual vs counterfactual outcomes at both the pitch and overall level, keeping only profiles that meaningfully improve whiffs while maintaining or improving groundballs.
Optimize pitch usage. For viable profiles, I select one and simulate alternative pitch mixes using leaguewide usage patterns and estimate how these mixes would affect overall and pitch level whiff and groundball rates across the season.
Detailed Walk Through
Step One: Pick a pitcher and a pitch to improve
I start by selecting a pitcher of interest. For example, suppose I choose Shane Smith of the Chicago White Sox. The goal is to find an adjusted pitch profile (velocity, spin rate, spin axis, induced vertical break, and horizontal break) that increases his whiff rate while either improving groundball rate or at least avoiding a substantial decrease.
I first split his pitches by batter handedness (vs RHB and vs LHB). Within each split, I compare his pitch types to league averages for whiff rate and groundball rate. Any pitch that is clearly below league average becomes a candidate for improvement. From that short list, I pick one pitch to work on, such as his Slider.
Step Two: Find “inspiration” pitchers using release similarity
Once I have a target pitch, I need a realistic pool of pitchers that he could plausibly emulate. I build this pool by focusing on pitchers who deliver the ball in a similar way.
To do that, I construct a release vector for like-handed pitchers using three components: horizontal release position of the ball (x), vertical release position of the ball (z), and release extension. I standardized these release features across all right-handed pitchers, then computed the Euclidean distance between Smith’s standardized release vector and every other right-handed pitcher.
I keep only pitchers whose release vectors fall within 1.5 standard deviations of Smith’s in a standardized space. This choice was arbitrary, but after some tinkering it felt like a reasonable threshold. If I went too small, then it would severely limit the candidate pool. If I went too large, it would introduce candidates that did not make biomechanical or practical sense. These pitchers form the “inspiration” group. Any pitch profiles I consider must come from this group, which helps keep the suggested changes at least somewhat biomechanically realistic.
Step Three: Filter pitch profiles using realistic year-over-year change thresholds
Within that inspiration group, I now have a bunch of possible pitch profiles for the same pitch type (for example, Sliders from similar right-handed pitchers). Not all of them are realistic for Smith to reach in a single offseason.
To impose realism, I use historical year-over-year changes in pitch attributes as guardrails. For each pitch type, by handedness, I have typical ranges for velocity, spin rate, spin axis, induced vertical break, and horizontal break actually change, for a given pitcher, from one year to the next. For example, a 4-Seam Fastball might reasonably see about a +/- 4 percentage point change in velocity(mph) year over year.
I treat these historical changes as thresholds. For each candidate inspiration pitch, I compare its profile to Shane Smith’s current pitch and filter out any that would require a bigger jump than those year-over-year limits. What remains is a set of pitch profiles that look realistically attainable.
It is also important to note that clearing these statistical guardrails does not guarantee biomechanical feasibility. Even if a candidate profile falls within historically realistic year-over-year changes, a pitcher may still be unable to consistently achieve the required shape if they cannot comfortably adjust the underlying mechanics that drive pitch movement and velocity. Subtle factors like wrist position, finger placement, and finger pressure, all influence pitch behavior, and not every pitcher can manipulate those inputs in the same way. For that reason, I view these filtered profiles as plausible exploration targets, not as guaranteed outcomes.
Step Four: Build counterfactual seasons using suggested pitch profiles
For each surviving candidate pitch profile, I create a counterfactual version of the pitcher’s 2025 season.
Sticking with the Shane Smith Slider example, I replace Smith’s actual slider profile with the average features of the suggested Slider profile while keeping everything else the same. For every pitch thrown in 2025, I then regenerate the features that feed into my whiff and groundball models. These features include:
The pitch’s own attributes (velocity, spin rate, spin axis, IVB, and HB)
How the pitch compares to his primary pitch (generating deltas for velocity, spin rate, etc)
How the current pitch compares to the previous pitch in the plate appearance, as a simple way to capture sequencing and basic arsenal effects
Before continuing, I want to mention that I am not including location data, regarding where the pitch was thrown, in the models. This was intentional as I was hoping to home in on, as best as I could, the impact of a pitch’s profile and how it compared to the previous pitch and the pitcher’s primary pitch. Additionally, I did not want to assume that a counterfactual pitch profile would land in the same exact location as the actual pitch did. Location features likely would have aided in the models’ performance, but I was able to get an acceptable enough level of performance to move forward without it.
Using these features, I run two sets of model predictions:
Actual Scenario: Predictions based on Smith’s real 2025 pitch attributes
Counterfactual Scenario: Predictions where only the target pitch (for example, the Slider) has been replaced with the suggested profile.
This gives me modeled whiff and groundball probabilities for both the actual and counterfactual worlds on a pitch by pitch basis.
Step Five: Compare actual vs counterfactual outcomes
Next, I aggregate these predictions to evaluate the impact of the suggested pitch profile. At the season level, I roll up the results across all pitches and by the pitch of interest (Smith’s Slider) to see how the new profile affects whiff and groundball rates, holding pitch usage fixed at the 2025 rates.
The key outputs here are the modeled lifts in expected whiff rate and expected groundball rate. If a candidate pitch profile increases whiff without tanking groundballs, that is encouraging. If it improves one outcome but clearly hurts the other, that is still useful information but less attractive from a recommendation standpoint. One caveat to mention is that I think it is best to treat the evaluation from a directional perspective as opposed to an absolute one. For example, if the modeling exercise suggests that the lift would be two percentage points, then it’s best interpreted as “this should improve things by a few points” versus “this will result in a two point improvement”.
Step Six: Select a profile and optimize pitch usage
For any pitch profile that shows a positive improvement (for example, clear whiff gains with stable or slightly improved groundballs), I pick the one that balances these outcomes in the way I care about most for the pitcher.
Once a profile is chosen, I move from “what should the pitch profile look like” to “how often should it be thrown”. To do this, I:
Derive realistic usage patterns. Using leaguewide data, I look at pitch usage rates by pitch type and handedness.
Sample candidate pitch mixes. I sample pitch usage from these distributions, enforcing that each sampled mix sums to 100 percent and stays within plausible ranges for each pitch type. For this exercise, I did this 1,000 times.
Simulate seasons under each mix. For each candidate mix, I resample the pitches in Smith’s 2025 dataset according to the sampled usage and simulate a full season’s worth of pitches, using my whiff and groundball models to project outcomes. I repeat this process 5 times per mix to stabilize the estimates.
For each mix, I compute average whiff and groundball rates across simulations and compare them to his actual 2025 outcomes under his real pitch usage. If a particular usage mix, paired with the new pitch profile, shows clear gains in whiff with acceptable groundball results, that mix becomes a recommended usage strategy to test. If the simulated mixes do not meaningfully outperform the real 2025 usage, the recommendation is to keep usage similar to the baseline, even if the profile itself is viable.
This process is effectively a two-stage Monte Carlo simulation: first sampling candidate pitch usage mixes, and then simulating pitch-by-pitch seasons under each mix. While this is not the most computationally efficient approach, generating 1,000 candidate mixes and simulating each 5 times provided enough stability for directional insights given the scope of the project.
Applying the Framework on Chicago White Sox Pitchers
Shane Smith
Shane Smith was one of the more unexpected success stories for the Chicago White Sox this season. Originally acquired from the Milwaukee Brewers via the Rule 5 Draft, Smith went on to post a 4.10 FIP and 2.2 WAR in his rookie campaign while also earning an All-Star selection. Despite that success, there is always room for improvement. Examining his pitches relative to league average Whiff%, I identified his Slider against both handedness splits and his Changeup against left-handed batters as potential targets. Because the Slider was both used more frequently and underperformed against league average whiff rates versus both right and left-handed hitters, it became the focus of this exercise.
As shown in the graphics, several candidate Slider profiles emerged as viable options. While each profile listed in the table projects an expected lift in Whiff%, only the Slider profiles from Carlos Vargas, Andre Pallante, and Bubba Chandler also maintained favorable groundball projections. I ultimately selected Vargas’ profile because it offered the best balance of increased Whiff% and Groundball%. Although this profile does not appear to be a perfect fit for driving whiffs at first glance, as seen in the SHAP feature summaries, the combined increases in velocity, spin rate, and horizontal break drive the expected improvements in swing-and-miss rates.
The suggested pitch usage changes produced several interesting insights. Against left-handed hitters, most usage rates remained within similar ranges to Smith’s actual 2025 tendencies. However, the optimized mix reduced 4-Seam Fastball usage from nearly 50% to 26% while substantially increasing Slider and Sweeper usage. It is worth noting that Smith threw his Sweeper only three times in 2025, but one of the strengths of this framework is its ability to simulate how a pitch might perform if given meaningful usage. Drawing on earlier research around arsenal interactions, the Sweeper and Slider appear well positioned to work off one another from a tunneling perspective. Their similar velocities imply comparable hitter decision points, and because both feature glove-side horizontal break without drastic separation in induced vertical break, distinguishing between the two could become especially difficult for hitters. From a combined movement and velocity vector-space perspective, the new Slider also creates greater separation from Smith’s other pitches, with the exception of his Curveball. Overall, the optimized mix is noticeably more balanced than his 2025 usage, which alone increases the difficulty for hitters attempting to anticipate what is coming.
A similar pattern appears against right-handed hitters. The optimization framework again recommends reducing 4-Seam Fastball usage, this time from over 40% to 23%, while increasing Slider and Curveball usage. From a modeling perspective, this directly alters the velocity-delta feature relative to Smith’s primary pitch. Because the Slider is substantially slower than the 4-Seam Fastball, shifting primary designation reduces the typical velocity gap that hitters experience. The increased Slider and Curveball usage also introduces more glove-side horizontal movement into the mix and pushes sequencing toward pitch combinations that are farther apart in movement-velocity space. As with the left-handed split, the final recommended mix is well balanced, with usage rates spanning roughly 11% to 32%, making it far more difficult for hitters to confidently anticipate Smith’s next offering.
Davis Martin
Davis Martin also enjoyed a fairly successful 2025 season with the Chicago White Sox, which marked his first full year in the major leagues. He posted a 4.64 FIP and 1.3 WAR across 26 appearances and 25 starts. When evaluating Martin’s pitch-level Whiff% performance, however, several offerings graded below league average. His Sinker, Changeup, and Slider all underperformed against both right and left-handed hitters, while his 4-Seam Fastball and Cutter were below league average specifically against left-handed batters. I elected to focus on his Changeup because it was one of his most frequently used pitches against both splits and performed particularly poorly from a swing-and-miss standpoint (–14 percentage points versus RHBs and –11 versus LHBs relative to league average).
As shown above, the counterfactual results were more mixed than in the Shane Smith example. Reese Olson’s and Mason Englert’s Changeup profiles both projected gains in Whiff% for Martin, but those improvements came with notable losses in Groundball%. Conversely, the Changeup profiles from Huascar Brazoban and Victor Vodnik projected improving Groundball% at the expense of Whiff%. Among these options, Brazoban’s profile was the only one that suggested a meaningful net benefit across a combined view of Whiff% and Groundball%. That projected improvement is driven primarily by modest gains in both velocity and horizontal movement, so I ultimately recommended Brazoban’s Changeup profile for Martin.
The pitch usage optimization step also produced several interesting recommendations. Against left-handed hitters, the framework suggested increasing Changeup, Slider, Sinker, and Curveball usage, primarily at the expense of the 4-Seam Fastball and, to a lesser extent, the Cutter. Anchoring the arsenal around the Changeup as the primary pitch introduces more usage of a pitch that stretches arm-side horizontal coverage against left-handed hitters while also lowering the overall velocity profile relative to the 4-Seam Fastball and Sinker. Both effects increase the separation between pitch profiles in combined movement–velocity space. From a trajectory-mirroring standpoint, it is also plausible that the counterfactual Changeup’s slight increase in induced vertical break allows it to more closely resemble the Sinker early in flight. If that effect is limited, the Cutter-Slider pairing may instead provide stronger tunneling benefits. While the resulting usage against left-handed hitters remains somewhat top-heavy, with the Changeup projected at 43%, the non-primary pitches are distributed fairly evenly, with usage rates between 9% and 14%. In practice, that balance would make it difficult for hitters to confidently anticipate anything other than the Changeup itself.
Against right-handed hitters, the recommended usage adjustments appear similarly reasonable and result in a well-balanced overall mix. The model again suggests reducing Cutter and 4-Seam Fastball usage and redistributing that volume across the Sinker, Changeup, Slider, and Curveball. Using the Cutter as the primary pitch in this split creates strong interactions with the Sinker, Four-Seam Fastball, and Changeup, as its glove-side horizontal movement and induced vertical break increase the separation from those pitches in movement space. The Cutter is also slower than those three offerings, which further increases the distance between their velocity–movement profile vectors. From a trajectory-mirroring perspective, the Cutter most closely resembles the Slider, which could allow those two pitches to tunnel effectively before separating via the Slider’s greater depth and glove-side movement. As with the left-handed split, the Sinker–Changeup pairing may also benefit from early-flight similarity. Overall, the final usage mix is well balanced, with no single pitch dominating the distribution. Usage rates range from roughly 9% to 22%, a structure that would likely score favorably under Brennan’s entropy-based sequencing framework.
Sean Burke
Coming into 2025, Sean Burke was hoping to build on a strong September 2024 call-up in which he posted a 3.32 FIP across 4 appearances and 3 starts. Unfortunately, Burke struggled for much of his 2025 campaign, finishing with a 4.92 FIP and 0.7 WAR across 28 appearances and 22 starts. When evaluating Burke’s pitch level Whiff% performance, several candidates emerged as potential targets for improvement. His Slider and Curveball both underperformed league average against right and left-handed hitters, while his Sinker and Changeup were below league average against right-handed batters. I ultimately focused on his Slider because it was the most frequently used pitch among this group and showed substantial room for improvement (-4 percentage points versus RHBs and -16 percentage points versus LHBs, albeit on a smaller sample).
Selecting a single counterfactual Slider profile proved more challenging than in the previous cases, as nearly every candidate suggested some degree of Whiff% improvement. I ultimately landed on Shohei Ohtani’s Slider profile for Burke because it projected not only a modest +2 percentage point lift in overall Whiff%, but also a sizable +10 percentage point increase in overall Groundball%. Comparing Burke’s current Slider to Ohtani’s profile, the gains in velocity and spin rate appear to drive the Whiff% improvement, while the reduction in induced vertical break likely contributes to the increase in groundball rate.
As with the other pitchers discussed, the pitch usage optimization step suggested several meaningful shifts. Against left-handed hitters, the framework recommended large reductions in 4-Seam Fastball and Curveball usage, paired with substantial increases in Slider and Sinker usage. Interpreting these changes through the lens of arsenal interaction research, the new Slider does not appear to materially increase the overall breadth of Burke’s repertoire. Relative to the 4-Seam Fastball, movement separation increases while velocity separation decreases. Relative to the Curveball, the opposite occurs: velocity separation increases while movement separation narrows. In effect, these changes largely net out from a breadth perspective. However, the altered movement profile of the Slider may increase its potential for tunneling with the Curveball, making those two pitches more difficult to distinguish early in flight. Given that the suggested mix remains heavy on the Slider, 4-Seam Fastball, and Curveball, this interaction could still pose challenges for hitters. The resulting usage distribution is somewhat top-heavy, with the Slider projected at 30% usage and the 4-Seam Fastball at 26%, while the non-primary pitches are more evenly distributed between roughly 11% and 17%.
Against right-handed hitters, the largest recommended usage reduction is to the Slider, with the 4-Seam Fastball also seeing a moderate decrease. The Curveball and Sinker gain modestly, while the Changeup experiences a large relative increase from its 2025 baseline. Focusing on the 4-Seam Fastball, Curveball, and Changeup combination, this mini arsenal grades very well from a breadth perspective, as each pitch occupies a distinctly different region of movement-velocity space. From a trajectory mirroring standpoint, the Curveball likely separates cleanly from both the 4-Seam Fastball and Changeup early in flight, while the 4-Seam Fastball and Changeup themselves may benefit from stronger early deception. The Slider and Sinker appear to function as effective “bridge” pitches within this mix, helping to buy back additional effectiveness for the primary offerings through sequencing and interaction effects. Overall, the right-handed usage recommendation mirrors the left-handed split in structure, with the primary difference lying in which pitches occupy each usage tier rather than in the distribution itself.
I am not going to get into the same level of detail for Jonathan Cannon and Mike Vasil, but will provide their graphics below.
Jonathan Cannon
Jonathan Cannon endured a difficult 2025 season with the Chicago White Sox, posting a 5.13 FIP and 0.4 WAR across 22 appearances and 17 starts. His performance represented a clear step back, and his role with the club remains uncertain heading into 2026. From a pitch-level Whiff% perspective, his Sinker emerged as the optimal target for improvement. The pitch performed below league average against both right-handed batters (-3.5 percentage points) and left-handed batters (-1 percentage point) and was also his most frequently used offering among those that underperformed relative to league average. A summary of the suggested profile change and its projected impact is shown below.
Mike Vasil
Mike Vasil was another rookie contributor for the Chicago White Sox in 2025. Claimed off waivers from the Tampa Bay Rays following Spring Training and added as a Rule 5 player, Vasil posted a 4.32 FIP and 0.4 WAR across 101 innings pitched. From a pitch level Whiff% standpoint, his Sweeper emerged as the primary target for improvement. The Sweeper graded below league average against both right and left-handed hitters and was used at a meaningful rate, particularly against right-handed batters. A summary of the suggested profile change and its projected impact is shown below (note that no sampled usage mix outperformed his 2025 pitch mix, so no chart is provided).
Applying the Framework on the 2025 MLB Cy Young Award Winners
American League Winner: Tarik Skubal
If you are reading this, Tarik Skubal likely needs no introduction. He was exceptional during his 2025 Cy Young campaign, posting a 2.45 FIP and 6.6 WAR across 31 starts. I fully acknowledge that it may seem a bit silly to run one of the game’s best pitchers through this type of exercise, and I will also apply the same process to Paul Skenes later. But part of the appeal of the framework is seeing what it suggests even at the extremes of performance. For Skubal, I focused on his Slider, as its Whiff% graded below league average against both right and left-handed hitters. A summary of the suggested profile adjustment and its projected impact is shown below.
National League Winner: Paul Skenes
Paul Skenes’ first full season in the major leagues was nothing short of extraordinary. He posted a 1.97 ERA, a 2.36 FIP, and 6.5 WAR across 32 starts on his way to winning his first Cy Young Award. For Skenes, I explored several potential pitches to evaluate but ultimately focused on his Four-Seam Fastball, which graded below league average in Whiff% against right-handed hitters. I selected this pitch not because it was necessarily the weakest offering in his arsenal, but because proposed profile changes from mechanically similar pitchers were more feasible for the Four-Seam Fastball than for some of his other pitches, where the suggested adjustments would have required unrealistically large departures from his current profile. A summary of the suggested change and its projected impact is shown below (note that no sampled usage mix outperformed his 2025 pitch mix, so no chart is provided).
I will also note that one of the primary comparison profiles comes from Chase Dollander, who spent much of his recent pitching time in Colorado. Because altitude is known to influence pitch movement, that context is important to keep in mind when interpreting the result, and it adds an additional layer of uncertainty to the recommendation.
Closing Thoughts
I really enjoyed working through this project and feel like I came away with a much better understanding of pitch design, arsenal development, and usage optimization. A lot of that growth didn’t come from anything I “discovered” myself, but from spending time reading, watching, and learning from the people who have already done great work in this space.
Being fully transparent, I know there are plenty of places where someone more experienced than me could poke holes in parts of the approach. That said, one of the things I like most about this framework is how flexible it is. If my release-vector threshold feels too wide, it’s an easy tweak. If my year-over-year pitch attribute bounds feel too loose, those can be tightened as well. The structure itself holds up even as those dials get adjusted.
It’s also important to acknowledge that even if the model suggests a pitch profile is better on paper, that doesn’t mean it’s actually achievable for a given pitcher. A player may not be able to consistently change wrist position, finger placement, pressure, or any of the subtle mechanical inputs that influence shape and velocity. That’s why I think of this project less as a prescriptive tool and more as a way to kickstart the brainstorming process. It’s meant to help answer questions like, “What kinds of changes might be worth exploring?” rather than to guarantee outcomes.
One direction I went back and forth on was whether to explore adding completely new pitches to a pitcher’s arsenal. I ultimately decided not to boil the ocean and kept the focus on tweaking pitches a pitcher already throws. That said, I do think this same framework could be extended to evaluate new pitch additions and better understand what kind of whiff or groundball lift might be on the table.
This project also didn’t start where it ended. Early on, I tried to build an arsenal-level model using features like movement spread, convex hull area, velocity spread, and pitch mix entropy to predict whiffs and called strikes directly. That model was… not good. Interestingly, that lined up with what Sutton-Brown has noted as well - that while arsenal level effects clearly matter, it’s still very hard to incorporate them in a way that reliably improves predictions. That realization is what pushed me toward a pitch level approach, which ultimately gave me enough signal to move forward with this work.
Thanks again for checking this out and if you’d like to learn more, do not hesitate to reach out!