2026 SABR Analytics Conference Research Presentations
SABR is excited to announce the list of accepted research presentations, presented by Teamworks Intelligence, for the SABR Analytics Conference, which will be held on February 27-March 1, 2026, in Phoenix, Arizona. Register today to join us!
Click on a link below to scroll down for abstracts, presenter bios, and schedule:
The SABR Analytics Conference brings together the leading minds in baseball analytics to discuss and debate topics relevant to the quantitative analysis of the game of baseball, including aspects of innovation and technology.
All student biomechanics research submitted for consideration for the SABR Analytics Conference will be eligible for the Dr. Mike Marshall Baseball Biomechanics Research Award, which comes with a cash prize of at least $2,000 per year.
As the leader in integrated technology for elite sports, Teamworks brings the industry-leading analytics team behind Teamworks Intelligence to this year’s research initiative. SABR has long set the standard for advancing baseball scholarship, and together they are highlighting cutting-edge research that continues to push the boundaries of performance, strategy, and innovation in the game.
General Track presentations
Friday, February 27
2:20-2:50 p.m. MST
RP1: Introduction to Causal Analytics, Collision Geometry Deviation (CGD) and CGD+
Ken Cherryhomes and Stacy Silvernail
Modern baseball analytics rely on descriptive metrics such as xwOBA and BABIP to quantify offensive productivity. These models evaluate hitting events based on output variables like exit velocity (EV) and launch angle (LA). However, these metrics describe effects rather than causes. Consequently, descriptive models cannot quantify whether a hitter achieved the correct geometric solution. This research introduces Causal Analytics to evaluate the geometric inputs of a collision. The central metric is Collision Geometry Deviation (CGD). This metric explicitly excludes EV and LA because they do not speak to the directional intent relative to the pitch location. An aligned swing can still produce high velocity. A perfect geometric decision can result in an out.
The methodology utilizes a spatial swing map to define deterministic optimal collision depth and barrel orientation for twenty-five pitch zones. A middle pitch neutral barrel orientation serves as the reference anchor. CGD measures the weighted Euclidean distance between observed and expected geometry. By isolating depth and orientation, the metric evaluates the specific directional intent required for each zone. Preliminary analysis separates geometric precision from variance. It contextualizes Hard-Hit Rate, proving extreme velocity on incorrect vectors creates predictable failure, not misfortune. Furthermore, this resolves standard BABIP assumptions by proving that much of what is currently labeled luck is actually quantifiable geometric error/obedience. This causal clarity enables a more powerful regression for predictive analytics. By translating mechanical reality into computational geometry, Causal Analytics isn’t just adding new stats; it is fixing a structurally incomplete way of thinking.
Ken Cherryhomes is a hitting coach with 25 years of experience, an inventor, and the founder of X Factor Technology. He holds multiple granted patents in sports technology and has trained players across several MLB organizations. His work spans timing training systems, novel force-plate software, and the creation of Causal Analytics. He builds hitter evaluation tools that treat contact as measurable collision physics.
Stacy Silvernail is the CEO of X Factor Technology and a technologist focused on applying measurement and computation to baseball performance analysis. A former collegiate athlete with a B.S. in Computer Science, she began her career at Microsoft and later earned SABR Analytics Certification. Her work centers on building the data architecture and analytical tooling behind X Factor’s hitting evaluation systems.
2:50-3:20 p.m. MST
RP2: Solving Ball Flight Physics
Conner Pelletier, Josh Hejka, and Sam Ehrlich
It has become increasingly clear in baseball over the last decade that leveraging knowledge of ball flight physics allows pitchers to optimize their performance and gain an edge over hitters. Most recently, a better understanding of the effects of spin, seam orientation, and non-Magnus movement (Seam Shifted Wake) has had tangible impacts on the pitches designed, trained, and thrown most often by MLB pitchers. The rise of four-seam fastball usage in the late 2010s, followed by the rise in sweepers and sinkers with seam shift effects in the early 2020s, demonstrates this effect. While the impact of pitch physics on baseball has been growing rapidly, there are still many important questions that have been left unanswered.
Publicly available research on this subject has largely plateaued within the last few years, and the technology used for older, foundational research has become outdated. Specifically, research detailing the relationship between lift coefficient and spin factor (spin rate/pitch velocity) used methodologies that could be improved upon with modern pitch tracking devices. Additionally, this research primarily studied pitches with only transverse spin (Dr Alan Nathan, 2015, 2018, 2020). This relationship is believed to change when gyrospin is introduced (Dr Alan Nathan, 2021) but has never been measured experimentally.
Research on non-Magnus movement also exists but is equally limited. For a constant spin rate, pitch velocity, and 3D spin axis, it is known where seams must be present most often relative to the baseball’s trajectory in order to induce seam-shifted movement (Andrew Smith, 2020). It is likely that changing these variables would change the necessary seam location to induce shifted movement, and it is also probable that these variables interact among themselves to further alter the mapping of ideal seam locations needed to create non-Magnus movement (Andrew Smith, 2021).
We intend to leverage the Trajekt machine in our Arizona gym to help fill these gaps in understanding. This machine grants control over all relevant ball flight parameters (seam orientation, spin rate, pitch velocity, release point, and 3D spin axis), allowing us to design and “throw” pitches with various combinations of these variables. Our Rapsodo unit will directly measure the spin parameters of each pitch while our Trackman unit will directly measure ball trajectory and movement. We plan to use this data to answer two key questions:
- How does gyro degree impact lift and drag coefficients?
- How do spin metrics and seam orientation interact to impact ball trajectory in ways beyond the Magnus force?
By answering these questions, we hope to understand how both Magnus and non-Magnus forces combine to change ball flight behavior in the pitch physics parameter space. In doing so, we hope to streamline the process of pitch design during training and uncover previously-unexplored areas of the parameter space that could result in valuable new pitch shapes.
Conner Pelletier is a Forward Deployed Engineer at Driveline Baseball Arizona, partnering with coaches and technical teams to drive research initiatives, product development, and technology integration throughout the facility. He graduated from University of California, Davis with a B.S. in Statistics.
Josh Hejka is a member of the R&D team at Driveline Baseball and a pitcher in the Texas Rangers organization. His research at Driveline has included quantitative pitch and arsenal quality modeling, pitch usage and location optimization for Major League clients, and tools to enhance the pitch design process. Before signing with the Rangers, he also was a member of the New York Mets and Philadelphia Phillies organizations. He graduated from Johns Hopkins University in 2019 with a BS in Computer Science.
Sam Ehrlich is a Baseball Operations Analyst at Driveline Baseball, where he is responsible for MLB deliverables and building tools that enable trainers to enhance athlete development. His work focuses on developing analytics infrastructure and data-driven solutions that support Driveline’s training and enterprise offerings. He holds a bachelor’s degree in sport management and a master’s degree in data science and analytics from the University of Missouri-Columbia.
3:20-3:50 p.m. MST
RP3: Contact Depth Affects Most of the Other Hitting Metrics We Analyze
Kevin Giordano
Background: Outcomes, such as launch angle, spray angle, and exit velocity are now routinely used as predictors of hitting success and likewise, are used by pitchers to create attack plans for each hitter. However, contact depth has theory to substantially influence all of these variables, but is not released publicly, and has nearly no attention in the academic literature. Misaki et al. (2016) determined lower contact points led to less contact precision with a lower attack angle, but on a tee. Recently, Nakashima et al. (2025) determined bat paths affect the timing error range for a hitter, which implies contact depth is important, but does not directly address it. Therefore, our purpose was to explore the relationship of contact depth with exit velocity, max swing speed, swing speed at ball contact, vertical attack angle, and horizontal attack angle.
Methods: Kinematic, bat and ball tracking data were collected using fixed, in-stadium KinaTrax markerless motion capture cameras and TrackMan technology on 2779 batted balls put in the field of play, excluding bunts and check swings across 239 NCAA Division I hitters over the course of the 2024 and 2025 college baseball seasons. We defined contact depth as the number of inches in front of the hitter’s center of mass in the direction of home plate to the pitcher’s mound that the bat made contact with the ball. A simple quadratic linear regression was used to assess the influence of contact depth on exit velocity. Simple linear regressions were used to assess the influence of contact depth on max swing speed and swing speed at ball contact. Due to high expected collinearity of vertical and horizontal attack angle, partial least squares regression was used to assess the influence of contact depth on the combination of horizontal and vertical attack angles.
Results: Average contact depth was 20.1±7.2” in front of the hitter’s center of mass with a range of -4.3-47.9”. Exit velocity was weakly explained by contact depth (r2=0.04, p<0.001, however positive residuals were abnormally distributed around the mean. Contact depth predicted max bat speed (r2=0.020, p<0.001). Contact depth had no meaningful influence on bat speed at ball contact (r2=0.01, p<0.01). Partial least squares regression confirmed high covariance between vertical and horizontal attack angle with the first component explaining 68.7% of the variance between vertical and horizontal attack angles. The first component explained 80.0% of the variance in contact depth, meaning there was a very strong relationship between contact depth and attack angles.
Conclusion and Application to Baseball Analytics: Contact depth strongly influences attack angles, weakly predicts max bat speed and exit velocity, but does not affect bat speed at ball contact. When analyzing hitters’ spray or launch angle distributions, contact depth should be accounted for. If designing hitting training programs aimed at better contact, contact depth’s influence likely means time series analysis of bat and ball path is superior to discrete ball contact measures.
Kevin Giordano is faculty in the physical therapy department at Louisiana State University’s Health Science Center, where he teaches foundational kinesiology and exercise physiology. His research focuses on orthopedic shoulder biomechanics and baseball injury and performance. Within baseball, he leverages in-game data and industry partnerships to analyze large volumes of data with advanced quantitative methods.
Saturday, February 28
2:20-2:50 p.m. MST
RP4: Assessing the Accuracy of In-Stadium and Portable Multi-Camera Markerless Motion Capture for Baseball Pitching Kinematics and Kinetics
Arnel L. Aguinaldo
Markerless (ML) motion capture has emerged as a viable option to marker-based (MB) motion capture in estimating movement biomechanics, but limited data exists on the accuracy of ML systems during high-speed throwing. This study evaluated the accuracy and reliability of an in-stadium (Hawk-Eye) and a portable (Theia3D) ML motion-capture system in quantifying baseball pitching kinematics and kinetics relative to an MB reference. Eighteen collegiate pitchers were simultaneously recorded using all three systems. Mean per-joint position error (MPJPE), statistical parametric mapping (SPM), root mean square error (RMSE), Bland-Altman analysis, and concordance correlation coefficients (CCC) were used to assess agreement. Both ML systems demonstrated measurable discrepancies across variables, with MPJPE values of 56.6 ± 9.4 mm (Hawk-Eye) and 52.0 ± 12.3 mm (Theia3D). Stride length exhibited the strongest agreement with MB in both systems (CCC > 0.85), whereas shoulder rotational variables showed greater variability. Error magnitudes in joint positions and kinematic waveforms were comparable to those reported for other ML systems during dynamic movements. These results highlight the influence of system configuration, camera deployment, and pose- estimation models on biomechanical accuracy. Overall, both configurations showed potential for estimating pitching biomechanics, underscoring the trade-offs between criterion and ecological validity in markerless motion capture.
Arnel Aguinaldo is a Full Professor of Biomechanics, Athletic Training, and Sport Science and the Director of the Padres Biomechanics Laboratory at Point Loma Nazarene University. He has earned an undergraduate degree in Bioengineering and a Ph.D. in Health and Human Performance and is a board-certified athletic trainer. Dr. Aguinaldo currently serves as the President of the American Baseball Biomechanics Society and as a scientific advisor to Major League Baseball. His research aims to understand the biomechanical implications of injury risk and player performance in baseball pitching and hitting.
2:50-3:20 p.m. MST
RP5: Individual Muscle Capacity to Generate Elbow Varus Moment during a Fastball Pitch
Kristen M. Stewart
Introduction: The extreme loads exerted on the body during pitching pose a high risk for injury. Specifically, a high valgus load is placed on the elbow and requires the muscles surrounding the elbow to counter with a varus moment. The incidence of elbow injuries is increasing, particularly at the ulnar collateral ligament (UCL). Cadaver studies have furthered the understanding of the mechanism of a UCL injury, but these studies may not elucidate the interaction between active elbow muscle force and the UCL.
Musculoskeletal modeling and simulation can be used to assess how muscles contribute to the varus moment throughout the pitch. Previous modeling studies have shown how elbow muscles may contribute to the varus moment quasi-statically or through subject-specific simulations, but it is unclear what the individual muscle capacities are to generate a varus moment, which could be used to generalize targeted strengthening programs. Thus, the purpose of this study was to investigate individual muscle capacities to generate elbow varus moment during a fastball pitch.
Methods: Kinematic data from a high-school pitcher were previously collected at the American Sports Medicine Institute from 38 retroreflective markers during a fastball pitch off a mound. A musculoskeletal model was adapted from a previously developed upper-limb model. A varus- valgus degree-of-freedom (DOF) was added at the midpoint of the humerus and ulna contact space oriented orthogonal to the elbow flexion and pronation-supination axes, resulting in 14 total DOF. Muscles not spanning the elbow were removed, torque actuators were applied to all DOFs without muscle actuation, and a baseball with appropriate inertial properties was fixed to the hand. The model was scaled to match the subject-specific anthropometry and inverse kinematics determined the joint angles that track the experimental data. Trunk motion was prescribed and varus deviation was locked at zero degrees for subsequent analyses. To assess muscle capacity, each muscle’s excitation was assumed to be 100% throughout the pitch. Each muscle’s maximum isometric force and force-length-velocity relationship were used to calculate muscle force and multiplied by the moment arm to the varus axis at each state to quantify its capacity to produce a varus moment.
Results: The wrist flexors and pronator teres had the highest capacity to generate a varus moment, and the wrist flexors had the largest varus moment arms. Due to forearm pronation during the stride phase, the biceps transitioned from a varus to valgus moment. The brachialis and biceps provided valgus moments during the stride phase, but this occurred while the elbow was extending, which suggests these muscles may not realistically reach full activation during this phase.
Implications: These results provide insight into the capacities of the elbow muscles to provide a protective varus torque throughout the fastball pitch and which muscles may be optimal for targeted strengthening programs to help offload the UCL. Furthermore, this provides a framework to understand how muscle capacity changes across conditions (e.g., pitch type) to identify greater UCL injury risk.
Kristen M. Stewart is a PhD candidate in mechanical engineering with an emphasis in biomechanics at the University of Texas at Austin. She currently works in the Neuromuscular Biomechanics Lab studying how assistive devices influence joint mechanics using motion capture, musculoskeletal modeling, and simulation. She has also applied her modeling and simulation expertise to baseball, analyzing muscle function during pitching. She will complete her PhD in Summer 2026 and is passionate about pursuing a career in baseball biomechanics.
Sunday, March 1
9:00-9:30 a.m. MST
RP6: Lead Leg Ground Reaction Force Patterns and Their Association with Rotational Velocities and Bat Speed in Baseball Hitting: A Statistical Parametric Mapping Analysis
Wesley Gawel
Baseball hitting demands exceptional power transfer through the kinetic chain, with the lead leg serving as a critical blocking mechanism. While previous research established correlations between peak lead leg ground reaction forces (GRF) and bat speed, the temporal importance of when these forces occur remains unexplored.
Methods: We employed statistical parametric mapping (SPM) regression to analyze lead leg force-time curves in 96 baseball hitters (bat speed: 32.57±1.53 m/s). Force data across magnitude and three directional components (x, y, z) were compared against bat speed and torso angular velocity throughout the entire swing phase (101 timepoints, 0-100% of swing).
Results: SPM regression revealed no significant temporal clusters between lead leg forces and bat speed across any force direction (all p>0.05). However, torso angular velocity showed significant associations with medio-lateral forces (Force_x: 20.8-100% of swing, p<0.0001, peak t=3.464) and anterior-posterior forces (Force_y: 9.5-38.8% of swing, p=0.0091, peak t=3.088).
Conclusion: Lead leg force-time characteristics did not directly predict bat speed when analyzed continuously throughout the swing. The significant relationship between lead leg forces and torso angular velocity suggests the lead leg’s primary role may be facilitating trunk rotation rather than directly influencing bat speed, challenging traditional assumptions about force transmission in baseball hitting.
Wesley Gawel is a member of the sports science team within the R&D Department at Driveline Baseball, where he primarily handles Launchpad operations including conducting biomechanical analyses with motion capture systems. He received his PhD in Sport Physiology and Performance from East Tennessee State University.
9:30-10:00 a.m. MST
RP7: Torpedo-Style Bat Mass Redistribution: A Statcast-Calibrated Simulation of Exit Velocity and HR Proxy Effects
Siddharth Ramakrishnan
We evaluate whether “torpedo” bats—defined here as a localized mass bulge positioned at a hitter’s contact hotspot—produce meaningful gains in exit velocity (EV) or home-run probability proxy (HR-proxy) while respecting MLB-legal wood-bat constraints. Using 2025 Statcast swings for three MLB hitters (Rafael Devers, Willy Adames, Jung Hoo Lee), we infer each player’s contact hotspot, simulate bat inertia, swing speed, and ball–bat collision outcomes, and quantify EV distributions under mass-preserving and +2 oz mass-allowing designs. We report bootstrap confidence intervals and sensitivity sweeps over key physics parameters. Across all players, torpedo designs increase mean EV by about 0.1 mph when mass is preserved and by about 0.8 mph when +2 oz is allowed, with P90 EV gains around 0.1–0.2 mph and about 1.0 mph respectively. HR-proxy gains are near-zero (<0.01 percentage points) in all conditions. These results suggest that hotspot-centered mass redistribution alone is unlikely to materially affect HR outcomes under the modeled assumptions.
Siddharth Ramakrishnan is a venture capital investor at Scale focused on AI applications and infrastructure, where he partners with founders building across the machine learning stack. He evaluates technical differentiation in ML systems and supports companies on product and go-to-market strategy. A lifelong Bay Area sports fan, he has applied machine learning directly to sports, including working with the San Francisco Giants on player evaluation models and building deep learning video systems to help athletes analyze game footage. He also researched machine learning in sports at Columbia University, focusing on how ML methods can augment scouting and player development.
11:15-11:45 a.m. MST
RP8: Kinematic Consequences of Pull-Centric Swing Approaches
Jessica Fackler
Introduction and Literature Review: Recently, MLB has seen a large emphasis placed on power hitting and homeruns, causing a league-wide trend in increasing pull rate. Previous literature has examined this pull-centric hitting approach and its effects from varying angles. With previous work quantifying the experienced success and drawbacks of pulling a pitch in MLB (Blengio, 2014; Clemens, 2024), Driveline explains the physical concepts that support the intuition behind success in this approach (Driveline Baseball, 2022). Other research has looked more specifically at the changes in mechanics and kinematic changes associated with pulling pitches relative to hitting to different parts of the field (Lim et. al, 2015; Gelinas, 1988). Nonetheless, previous literature lacks evaluation of the effect pitch location has on both biomechanical changes and potential success. With the use of computer vision, this study aims to evaluate kinematic differences in a swing given the pitch location and spray angle, and understand any potential tradeoffs with on-field success.
Methodology: This study uses broadcast footage from the 2025 MLB season and YOLO-based computer vision models to extract 2D batter pose estimations, which are then uplifted to 3D using MotionBERT. Biomechanical features are then derived using inverse kinematics and vectorized math. With batter identity and handedness as random effects and pitch location as the fixed effect, functional mixed-effects models are built for each biomechanical feature. Kinematic differences are thus attributed to pitch location, given that all batted balls were hit in the same direction. To understand how these differences associate with on-field success, this study recreates xwOBAcon to incorporate spray angle and ballpark, in addition to other batted ball features. Using data from 2015–2023, an ordinal LightGBM was built and assigns an adjusted xwOBAcon value to each batted ball, allowing swing kinematic variation to be directly linked to performance.
Results and Contributions: The output of the computer vision pipeline produced temporally aligned 3D skeleton reconstructions of all swings. One figure shows two reconstructions, both of Isaac Paredes pulling an inside pitch and an outside pitch. The second figure shows a stacked view of the mean predicted curve of the angular velocity of the lead elbow and that of the transverse pelvic angular velocity. For inside pitches that are pulled, batters are rotating their hips and extending their elbow much more rapidly than if it were an outside pitch. Additionally, with outside pitches, we notice less explosive swinging behaviors when pulling these pitches, aligning with intuitive understandings of swing mechanics. These kinematic variations are then linked to corresponding changes in on-field success to assess the true value a batter achieves through such adjustments. By linking kinematic differences to on-field value, this study not only uncovers the true impact of pulling pitches, but also establishes a framework for the baseball analytics community to derive biomechanical data from single-angle video. With certain data lacking public availability, this study leverages the recent emergence of computer vision to collect this data, expanding resources for baseball data scientists to conduct new analyses.
Jessica Fackler is entering her second season with the Cincinnati Reds as an incoming Quantitative Analyst following graduation. Currently, she is a third-year senior at Syracuse University completing a dual-major in both Sport Analytics and Physics. She serves as president of the Baseball Statistics and Sabermetrics Club on campus.
11:45 a.m.-12:15 p.m. MST
RP9: Prediction with Transparency: Offensive Value in Baseball
Mark Kritzman and David Turkington
Overview: We describe a model-free prediction system called relevance-based prediction (RBP). Unlike linear regression analysis or machine learning models, which work by estimating model parameters and then applying those parameters to new tasks, RBP forms a prediction as a weighted average of observed outcomes in which the weights are based on a rigorously defined and theoretically justified statistic called relevance. We illustrate this novel methodology by predicting wRC+ for MLB players for the 2025 season.
Methodology: RBP has three foundational features: relevance, fit, and grid prediction. Relevance measures the importance of a player to a prediction. It is composed of similarity and informativeness, which are both measured as Mahalanobis distances. Similarity measures the multivariate distance of a prior occurrence from the current prediction circumstances. Informativeness measures the multivariate distance of a prior occurrence from average. By measuring informativeness as a difference from average, we recognize that unusual players contain more information than typical players. Fit measures the extent to which there are useful patterns in a dataset. It is measured as the average standardized alignment between the relevance weights of a given prediction task and the outcomes for every pair of observations that go into it. Fit gives advance guidance about the reliability of each prediction. R-squared only provides information about a model’s average quality. Fit is a mathematically exact prediction-specific decomposition of R-squared. Grid prediction creates a reliability-weighted composite prediction by blending the predictions that result from different combinations of players and predictive variables. The columns in the grid represent different combinations of predictive variables, and the rows represent different subsamples of players based on relevance thresholds. Each cell in the grid provides a prediction 2 and a measure of reliability. The prediction grid diversifies the composite prediction across many calibrations in a way that bends toward those that are more reliable.
Expected Contribution to MLB Teams: RBP can support MLB teams in several crucial ways:
Efficacy
- In forming predictions, RBP can address complexities that are beyond the reach of conventional prediction models such as linear regression analysis.
- RBP can extract as much information from complex datasets as machine learning models such as neural networks.
Transparency
- RBP gives visibility into the formation and reliability of each individual player prediction.
- RBP reveals the reliability of each player prediction before the prediction is made, thereby enabling teams to discard or treat more cautiously predictions known in advance to be untrustworthy.
- RBP identifies the previous players who are most relevant to the formation of a current player’s prediction, thus enabling teams to compare this information with knowledge obtained from other sources.
- RBP shows the contribution of each predictive variable to the value of each individual player prediction as opposed to each variable’s average contribution across all players.
- RBP shows the contribution of each predictive variable to the reliability of each individual player prediction as opposed to each variable’s average contribution across all players.
These insights are unobtainable from conventional models as well as machine learning models.
Mark Kritzman is a Founding Partner of Cambridge Sports Analytics. He is also a Founding Partner of Windham Capital Management and State Street Associates, and he teaches a graduate finance course at MIT Sloan. He has published more than 100 articles in peer-reviewed journals and is the author or co-author of eight books including Prediction Revisited. Mark holds a Bachelor of Science degree in Economics from St. John’s University and a Master of Business Administration degree with distinction from New York University.
David Turkington is a Founding Partner of Cambridge Sports Analytics. He is also Senior Managing Director and Head of State Street Associates, State Street Corporation’s Cambridge-based innovation hub. Dave is the author of more than 45 peer-reviewed scholarly articles and co-author of three books including Prediction Revisited. Dave graduated summa cum laude from Tufts University with a Bachelor of Arts degree in Mathematics and Quantitative Economics.
Student Track Presentations
Saturday, February 28
10:05-10:20 a.m. MST
SP1: Building a Better Ballpark: Optimizing the Placement of the Outfield Wall in an MLB Stadium
Peter Culley
Adjusting ballpark features to further home field advantage is a relatively new concept being explored by MLB teams today. The Baltimore Orioles pioneered this idea in the 2021-2022 MLB offseason when they started tinkering with their left field wall to limit opposing right-handed hitters’ slugging rates. According to team-wide changes in several underlying metrics, the new wall appeared to be largely responsible for the Orioles’ thirty-one-win increase from 2021 to 2022. After their initial success with the new wall, the Orioles again changed their left field wall prior to the 2025 season when they landed their own right-handed power bat in Tyler O’Neill (Wilytics, 2025). The effectiveness of this new wall is still to be determined.
The Orioles are not the only team to have made significant alterations to their park to help their team gain an advantage. Prior to the 2024 season, the Cleveland Guardians removed several shipping containers from their right field upper deck and instantly created a powerful wind tunnel, pushing fly balls hit that direction out of the park at a much-increased rate (Strack, 2024). Without major changes to their offense for most of the season, the Guardians improved from hitting 0.67 to 1.22 home runs per home game on average, largely thanks to the wind tunnel (StatMuse, 2024). More recently, in early January 2026, the Kansas City Royals announced plans to shift their outfield wall inward 9-10 feet to allow for more offensive production at a notoriously pitcher-friendly park while minimally impacting their pitching staff’s success (Rogers, 2026).
Since games are won not based on who hits more home runs but on who scores the most, this research was focused on optimizing a ballpark’s outfield wall dimensions to maximize the gap between the home team’s run expectancy and their opponent’s run expectancy. By constructing a stochastic, operational model to describe the objective and any wall constraints set forth by MLB or the team itself, we can determine the optimal distance from home plate to each wall, the angle at which they should intersect the foul lines, and their height. Implementing this approach suggests the Arizona Diamondbacks could have altered the walls at Chase Field so that they could have won an additional three home games in 2024, thus clinching a playoff spot.
While some of the specific techniques used here were tailored to fixing Chase Field, the general approach could be expanded to several other ballparks as well and provides a unique solution for teams trying to win without spending on top-of-the-market free agents. Future suggested work includes the development of a brand-new statistic called wWAR (Wall Wins Above Replacement) to help further the comparison between spending on players and spending on moving the wall.
Peter Culley is an independent baseball analyst with a focus on major league strategy and optimal resource allocation. He graduated in December 2025 from Arizona State University with a master’s degree in Industrial Engineering, providing him with a strong, mathematical background to attack modern baseball challenges with a unique approach. Applying deterministic and stochastic operations research techniques as well as statistical analysis and discrete event simulation, he has experience solving problems ranging from roster construction to field geometry. An alumnus of last year’s MLB In-Person Cohort at the SABR Analytics Conference, he aspires to become a Research and Development Analyst with an MLB team.
10:25-10:40 a.m. MST
Room 240, 2nd floor
SP2: Relationship Between Fastball xCTRL and Kinematic Determinants of Arm Slot in NCAA Division I Baseball Pitchers
Joseph H. Caplan
Background: Pitch control is paramount when evaluating fastball execution in baseball pitchers. However, challenges arise when assessing control without prior knowledge of a pitcher’s intended target location. Conventional measures, such as strike percentage, do not account for the pitcher’s intent. In contrast, a new expected control (xCTRL) metric addresses this limitation by measuring pitch location relative to a pitcher’s historical tendencies, providing a single value that reflects control in relation to expected placement. This approach involves establishing “zones” and calculating the probability-weighted Euclidean distance between the actual pitch location and the prospective zone, or the nearest zone when multiple are defined. In complex movements, such as pitching, numerous kinematic factors and their associated variabilities influence both control and positional consistency. For example, recent trends of pitchers lowering their arm slots warrant investigation alongside xCTRL to explore potential influences on accuracy further. With limited existing research on the impact of upper extremity kinematics on control, this study aimed to examine how xCTRL is affected by arm slot kinematics.
Methods: Kinematic data from 42 NCAA Division I pitchers (height: 1.87 ± 0.05m, mass: 93.2 ± 7.49kg), who have accumulated a minimum of 50 total fastballs throughout the full span of data collection, were collected at 300hz using an eight-camera markerless motion capture system. Data were processed and filtered using proprietary KinaTrax software. Frontal plane kinematics associated with arm slot were measured with respect to global position at ball release (BR) (torso angle, upper arm angle, forearm angle, and hand angle). Means and standard deviations of torso and upper extremity kinematics were subsequently calculated for each participant. Aside from pitch count inclusion criteria, xCTRL was calculated per Ludwig et al. using data retrieved from a TrackMan V3 Game Tracking unit (pitcher handedness, batter handedness, count, pitch location coordinates). A backward multiple linear regression (α = .05) was used to identify the kinematic predictors of xCTRL.
Results: The average fastball xCTRL from the cohort was 12.33 ± 1.04. After the backward elimination, one predictor, torso angle deviation, explained 11.0% of the variance in fastball xCTRL (F(1,40)=4.96, R2 =.110, p=.032). As torso angle standard deviation increased by 0.332°, xCTRL increased by 1 (β=0.332, t(40)=2.23, p=.032).
Conclusion: These results indicate that greater variability in torso angle at ball release is significantly associated with higher xCTRL values, suggesting reduced control. No significant effects were observed for the mean or standard deviations of other arm slot-related kinematic variables. These findings support the hypothesis that enhancing torso angle consistency during fastball delivery may improve control. Future research should also explore lower extremity kinematics or rotational movements, and their relationship to xCTRL.
Joseph H. Caplan is a first-year master’s student at Auburn University, where he is researching baseball and softball biomechanics in Dr. Gretchen Oliver’s Sports Medicine and Movement Lab. His research interests include pitching performance improvement and injury risk mitigation by use of both markerless and marker-based, electromagnetic motion capture systems (KinaTrax and Ascension, respectively). He is a graduate of the University of Miami, where he studied Exercise Physiology and Sport Administration, with a research focus on exercise interventions for aging and diseased populations, such as those with Parkinson’s.
10:25-10:40 a.m. MST
SP3: Modeling When a Third Base Coach Should Send a Runner Home
Shane Sullivan
Third base coaches are some of the most underrated and undervalued individuals on a baseball field. Previous work on the subject of baserunner decision-making has largely focused on baserunner/fielder timing and run expectancy, but not on the actual decision-making of third base coaches themselves and the value they can provide. Third base coaches make some of the most critical decisions in a game, potentially being the difference between a win and a loss on any given day. Yet, there is no WAR, or ranking, or anything of the sort for third base coaches.
This gap is what inspired my examination into third base coach decision-making. This research was conducted through a project submission to the 2025 SMT Data Challenge, where player tracking data from an unspecified MiLB level was provided. The first step was modeling when a third base coach should send a runner home or not. This was based on variables such as player distances from home, baserunner speed, outfielder momentum, and game state gathered from provided data. With these variables, a variety of model types and parameters were tested to find one that best fit the data. Then the performance of third base coaches as a whole, and individual team third base coaches were evaluated based on the decisions the model would have made with an expected runs lost statistic. Finally, a comparison was made between how each variable was weighed in the created model, and a model fit to the actual decisions made by third base coaches to find where they could make improvements.
Overall, it was found that third base coaches did typically make the correct decision, but they still lost their teams several expected runs throughout a season from their incorrect decisions. As a whole, they were not as aggressive as they should have been with sending runners home. They also tended to weigh the baserunner’s top speed too heavily, while not weighing the momentum of the outfielder and the game state enough. This evaluation was also done for three individual teams, with noticeable variation between them in decision quality and how much they weighed different variables.
Overall, I believe this research could be built upon and used in a variety of ways, such as quantifying the impact of a good third base coach. This could also be used by teams to aid in improving the decision making process of their own third base coaches.
Shane Sullivan is a junior at The College of New Jersey majoring in Mathematics with a specialization in Data Science and Statistics. He plans to pursue a graduate degree in data analytics, then a career in sports analytics post graduation. He has competed in the 2025 Garden State Undergraduate Math Competition, 2025 SMT Data Challenge, and 2026 Sumer x Shrine Bowl Analytics Competition. He also continues to play the game of baseball as player and stat keeper for his school’s club baseball team.
10:40-10:55 a.m. MST
Room 240, 2nd floor
SP4: Beyond the Aging Curve: A Bayesian Framework for Distinguishing and Quantifying Injury vs. Natural Decline in MLB Pitchers
Nickolas Bartle
Previous Work: Pitcher aging curves have been extensively studied in baseball analytics, with researchers establishing that velocity and overall effectiveness typically decline after age 30-32. However, existing approaches treat all performance declines as equivalent, failing to distinguish between natural aging patterns and injury-related deterioration. This limitation significantly reduces the actionable value of aging models for front offices evaluating contract decisions and player development staff managing workload. While changepoint detection methods exist in the broader statistical literature, their application to pitcher performance evaluation particularly when integrated with injury data remains largely unexplored in public baseball research.
Research Methodology: This study develops a hierarchical Bayesian framework using 2015-2024 Statcast data to model individual pitcher aging trajectories while quantifying the uncertainty inherent in performance projections. The methodology employs three integrated components: First, we fit individual aging curves for velocity, movement, and command metrics using mixed-effects models that allow each pitcher to deviate from population-level trends while borrowing statistical strength across similar player profiles. Second, we implement Bayesian changepoint detection algorithms to identify the precise timing of performance regime shifts with associated probability estimates, distinguishing gradual decline patterns from sudden drops. Third, we integrate historical injury list data to create features that help differentiate injury-related declines from natural aging patterns, examining velocity recovery trajectories and the temporal relationship between IL placements and detected changepoints.
The model outputs posterior probability distributions rather than point estimates, providing decision-makers with uncertainty quantification (e.g., “75% probability this pitcher’s decline began in June 2023” rather than merely “pitcher is declining”). We validate the framework through retrospective analysis of known decline cases, measuring both the model’s predictive accuracy and the calibration of its confidence intervals.
Preliminary Results: Analysis of 400+ pitcher-seasons reveals that changepoints in velocity occur with varying magnitudes and recovery patterns. Preliminary findings suggest that sudden velocity drops followed by partial recovery correlate strongly with documented injuries, while gradual declines (i.e velo per year) better match expected aging trajectories. The hierarchical structure successfully captures individual variation pitchers like Justin Verlander show measurably different aging patterns than population averages. We are currently finalizing the integration of comprehensive injury data to strengthen the causal distinction between decline mechanisms.
Expected Contribution: This research provides baseball operations departments with a probabilistic early-warning system that quantifies decline risk while acknowledging uncertainty. Unlike existing aging models that simply project future performance, this framework identifies when decline began, how confident we should be in that assessment, and why the decline likely occurred. The methodology offers immediate applications for contract negotiations, roster construction, and workload management decisions where distinguishing recoverable injury declines from permanent aging effects carries substantial financial implications.
Nickolas Bartle is a graduate analyst for Wake Forest Baseball with a primary focus on predictive modeling within pitching analytics. Previously he worked for the University of Alabama baseball team as a data analyst, Mercedes-Benz as a data engineer co-op, and USA baseball as an operations intern. He holds a Bachelor of Science in Management Information Systems from the University of Alabama, and will complete his Master of Science in Business Analytics at Wake Forest University in May 2026.
10:40-10:55 a.m. MST
SP5: “Stomping” Out Fielder Pitch Tipping
Tejas Rama and Sam Cowan
In this paper, we study whether defensive position players can unintentionally tip the upcoming pitch through their pre-pitch movements or positioning. Very little public research exists on this topic; prior to this study, the only documented discussion came from a 2022 interview with Arizona Diamondbacks personnel. In that interview, manager Torey Lovullo and infielder Geraldo Perdomo described noticing patterns in which middle infielders would subtly shift or “stomp” in anticipation of certain pitch types. Lovullo coined the term stomping to describe these pre-pitch cues, and Perdomo noted that hitters occasionally exploited them, with one saying, “I stayed on that pitch because I saw the second baseman move.” While these anecdotes suggest that fielder-induced pitch tipping may be real and actionable, no quantitative research had investigated the phenomenon before this work.
Research Methodology: The data used is anonymized 2-D player-tracking and 3-D ball tracking data from across 2 MiLB seasons provided by SMT. We first calculated velocity and horizontal break and classified pitches as fastball or offspeed with glove- or arm-side run. After accounting for pitcher- and batter-handedness, we then used logistic regression models and plots faceted by pitch type and team to conclude which positions and (anonymized) teams had pre-pitch movement patterns that correlated strongly with pitch type.
Results: Our findings suggest there is a real link between fielder positioning/motion and pitch type/ We conducted 112 statistical tests with a threshold of p < 0.05, and 13 of these tests were found to be significant. The chance of all of these being false positives is about 4.56% which is statistically significant by the 0.05 threshold, suggesting that at least some of the observed effects of fielder movements and positioning on pitch type are real. One example of such an effect is when facing left-handed hitters against left-handed pitchers, first basemen showed a significant pre-pitch movement tendency to cheat towards the first base line (margin = -0.09323, p = 0.0106, N = 421). This aligns with the defensive intuition of moving closer to the line on offspeed pitches, while playing further off on fastballs.
Expected Contribution to the Field of Baseball Analytics: Some of the effects found are large enough to be exploited by hitters – if a scouting team finds that a team’s third baseman tends to take a step backwards on offspeed pitches to righties, batters can be told to watch for this. On the other side, a team might be able to catch themselves tipping in this manner and correct themselves before their opponents catch on. This is an understudied area of research, and the effects at the college and professional levels could be exacerbated by the use of the PitchCom device, which can grant any three fielders knowledge of the pitch type (typically the 2B, SS, and CF, but other arrangements are possible).
Tejas Rama is a dual-degree student at Boston University pursuing a dual degree in Statistics and Finance. He interned with D.C. United of Major League Soccer, applying data to support recruitment and player valuation. He has conducted sports analytics research at Questrom and served as a data analyst for the BU women’s soccer team. He is the Co-Founder and Co-President of the Sports Analytics Group at Boston University.
Sam Cowan is a Data Science student at Boston University and a founding board member of the school’s Sports Analytics Group. He has conducted baseball research in the SMT Data Challenge, won the 2025 Ohio State Sports Analytics Hackathon, and volunteered for SABR’s Games Project and BioProject.
10:55-11:10 a.m. MST
Room 240, 2nd floor
SP6: A Kalman Filter Approach to Pitcher Mechanical Consistency in Major League Baseball
William Lee and Bach Nguyen
The 2024 MLB Pitcher Health Report highlights a troubling trend: pitchers pursuing higher velocity are throwing with increasingly inconsistent mechanics, contributing to rising arm injury rates. Monitoring how pitcher mechanics evolve within games and across seasons could identify which pitchers face injury risk and what distinguishes high performers from poor ones. Yet despite prior work establishing that mechanical consistency meaningfully contributes to pitcher effectiveness, no framework currently exists to quantify how pitcher mechanics drift and change over time.
We introduce Drift Variance and Drift Rate—a quantitative measure of how pitcher mechanics evolve pitch-by-pitch and game-by-game. We employ an Extended Kalman Filter to track pitcher mechanical states (release point, velocity, spin characteristics) recursively through each outing. The filter works by generating predictions of mechanical state for each pitch, then updating those predictions based on observed Statcast data. Drift Variance measures the inherent mechanical variability a pitcher exhibits, while Drift Rate quantifies how mechanical stability changes within a game as fatigue or adjustment occurs. A key benefit of this approach is that it filters measurement noise, allowing us to isolate true mechanical changes. We validate the framework by confirming that prediction errors behave as random noise without systematic patterns, indicating the model captures authentic mechanical evolution.
Our analysis reveals that mechanical drift alone shows weak correlation with short-term pitching performance. However, meaningful relationships emerge in specific contexts: pitchers with higher mechanical variability struggle to sustain velocity late in games, and greater pitch-to-pitch mechanical instability corresponds with degraded location accuracy. These findings suggest that while mechanical drift is not a dominant performance driver, it manifests in ways that matter—particularly in late-game execution. These drift metrics open pathways for investigating pitcher durability and injury risk, areas where mechanical instability may prove more predictive than traditional performance metrics.
William Lee is a sophomore at UC Berkeley pursuing a degree in Physics along with minors in Mathematics and Data Science. He is currently involved in cosmology research through the National Energy Research Scientific Computing Center, where he develops surrogate machine learning models for measuring dark matter halo masses around luminous galaxies. Outside of physics, he is an avid fan of baseball and basketball, and serves as a project manager and data journalist at Sports Analytics Group at Berkeley
Bach Nguyen is an undergraduate senior at the University of California-Berkeley, majoring in statistics. His research outside of class includes the intersection between medical and data science, and he has been a member of the Sports Analytics Club at Berkeley since 2023. He is pursuing either graduate studies or a position in the medical research field after graduation.
10:55-11:10 a.m. MST
Room 240, 2nd floor
SP7: When to Challenge a Pitch? A Reinforcement Learning Approach to ABS Challenge Strategy
Christopher Martinez
Starting in 2026, MLB will implement an Automated Ball-Strike (ABS) Challenge System that allows teams two pitch challenges per game. If the team making the challenge is successful, they retain the challenge. This raises a strategic question: when should a team use one of its limited challenges to maximize expected runs or win probability?
Previous work on ABS strategy has focused on describing the new rule and the inaccuracy of human umpires. In terms of deciding when to challenge, there are existing “greedy” models that look at the immediate change in win expectancy of an overturned call and the probability that the call was incorrect. ABS challenge strategy has not been treated as a full sequential decision problem. There is a need to explore how teams should optimize their limited challenges.
This project frames ABS challenge strategy as a Markov Decision Process (MDP) and solves it with reinforcement learning. The state space consists of 2,592 discrete states encoding the innings, outs, count bucket, base-runner configuration, score differential, challenges remaining, and pitch distance from the zone edge. The reward function is defined as the change in run expectancy (RE24) if the call is successfully overturned. Using pitch-by-pitch data from MLB Statcast of 93,206 pitches across 1,568 games from June-September 2024, I simulate the ABS challenge environment by modeling challenge outcomes based on pitch location relative to the strike zone boundary. I train a Q-learning agent with Dyna-Q model-based updates to learn an optimal policy. I compare the Q-learning policy against two baselines: a situational threshold heuristic that uses pitch proximity to the zone edge and game context (inning, outs, baserunners, score) to prioritize higher-leverage moments, and a greedy heuristic that challenges whenever the estimated immediate expected value is positive.
Across a held-out test set of games, the learned Q-learning policy produces 0.12 expected runs per game, compared to 0.04 for both the greedy policy and the situational threshold heuristic baseline. This projects to an improvement of approximately +1.2 wins per season over these baselines. The results demonstrate that accounting for the option value of reserving challenges for high-leverage situations outperforms greedy policies that challenge any positive-EV opportunity.
This work provides a baseline framework for teams to evaluate their own in-game pitch challenge protocols before Opening Day. This research demonstrates that sequential decision models can outperform immediate-value models when optimizing the challenge system. Future work can integrate pitcher-specific tendencies, umpire-specific accuracy models, and validation on real 2026 ABS data.
Christopher Martinez is a Master’s student in Computer Science at the Georgia Institute of Technology. He has built backend systems and data pipelines for baseball analytics, with interests in software engineering, large-scale data processing, and machine learning. His current work explores practical applications of AI and ML methods to baseball in-game decision making. He is pursuing engineering roles in baseball R&D.
Poster Presentations
Presented by Teamworks Intelligence
Saturday, February 28
4:35-5:15 p.m. MST
P1: Quantifying Pitcher’s Movement Profile and Evaluating Its Relationship with Arm Angle
Choi Min Seok
This study examines the relationship between arm angle and pitchers’ movement tendencies using Statcast data from 2020–2025. Movement characteristics are summarized through two scalar metrics constructed from league-standardized horizontal and vertical pitch movement computed separately across pitcher–batter handedness matchups. Movement Bias represents overall movement tendency by aggregating usage-weighted standardized components, and scalarized using a softsign transformation. Separation Bias is quantified by measuring maximum movement differences between two pitch types and transforming the horizontal–vertical contrast into a single interpretable value. Generalized additive modeling (GAM) is used to assess the measurable effects of arm angle on each movement metric. Modeling results indicate a substantially stronger structural relationship: lower arm angles are associated with horizontally separated repertoires, higher arm angles with vertically separated patterns. Arm angle alone explains 20~46% of the variance. These findings demonstrate that arm angle meaningfully contributes to pitch movement and repertoire structure. This arsenal-independent framework provides empirical evidence for how arm angle shapes pitch movement tendency.
Choi Min Seok is a third-year undergraduate student in the Data Science department at Hanyang University. He has a strong passion and interest in analyzing baseball from mathematical and probabilistic perspectives and aims to develop his own original analytical models. He is particularly interested in pitching metrics and hopes to work for a baseball team after graduation, building analytical models. He was also a columnist for YAGONGSO, the largest baseball analytics community in Korea.
P2: The Relationship Between Energy Flow, Ball Velocity, and Upper Extremity Joint Loading in Collegiate Baseball Pitchers
Brandon Muczynski
Introduction: Baseball pitching is one of the most demanding overhead motions in sport, exposing the shoulder and elbow to large repetitive loads. Among male pitchers, 46–57% experience throwing-related shoulder or elbow injuries, with pitchers accounting for 39.6% of shoulder and 56.9% of elbow injuries. To generate high ball velocity, energy must be efficiently transferred through the kinetic chain, from the lower extremities and trunk to the throwing arm. Previous work has shown that the trunk contributes the most mechanical work to both elbow valgus torque and ball velocity, yet how energy is generated, absorbed, and transmitted across the full body remains unclear. Prior research on energy flow during pitching has largely focused on isolated segments whereas, this study performs a comprehensive analysis of joint force power, joint torque power, and energy-transfer efficiency (ETE) across the lower extremity, trunk, and upper extremity in collegiate baseball pitchers. ETE quantifies how effectively mechanical energy produced at a proximal segment is transmitted to a distal segment, which may serve as an indicator of both performance and joint stress. It is hypothesized that higher ETE will be associated with greater ball velocity and lower upper extremity loading.
Methods: Forty collegiate pitchers threw in the University of Nebraska Omaha Pitching Lab. Kinetic data were collected using three in-ground force plates (Bertec, 320 Hz) and kinematics with a 16-camera motion capture system (Qualisys, 320 Hz) using a 41-marker full-body model. A 15-segment linked model was processed in Visual3D (HAS-Motion) and MATLAB to calculate joint force power, torque power, and work as the time integral of power. Energy- Transfer_Efficiency (ETE) was defined as the ratio of distal (upper extremity) to proximal (lower extremity + trunk) positive work. Data were normalized to the pitch cycle (Setup→Ball Release), and Spearman correlation (p<0.05) was used to evaluate relationships between ETE, ball velocity, and joint kinetics.
Results: Energy transfer efficiency (ETE) showed a weak, non-significant positive correlation with ball velocity (r = 0.25, p = 0.12). ETE was not correlated with maximal elbow varus torque (r = –0.01, p = 0.97) or maximal shoulder distraction force (r = 0.07, p = 0.73). Regional work ratios of the lower extremity, trunk, and upper extremity showed no significant relationships with ball velocity.
Significance: This study examined the relationships between energy transfer efficiency (ETE), performance, and upper extremity loading in collegiate baseball pitchers. Overall, our findings suggest that ETE and specific region energy contributions do not significantly predict pitching velocity or upper-extremity joint loading in collegiate pitchers. Despite previous literature examining efficient energy flow with throwing performance, total mechanical efficiency may not directly influence throwing speed or joint loading magnitude. Instead, timing and coordination of energy transfer among body segments may play a greater role in pitching performance and arm stress. Future work should include temporal aspects of energy flow and include various pitch types to better understand variability in efficient energy transfer and injury risk.
Brandon Muczynski is currently in the final semester of his master’s degree in Biomechanics with a concentration in Sports Biomechanics at the University of Nebraska at Omaha. He earned his bachelor’s degree in Kinesiology from Purdue University, where his initial research experience involved using EEG to examine the cognitive benefits of physical activity. His current research focuses on the relationship between athlete mechanics, performance outcomes, and injury prevention primarily in baseball pitching. Using motion capture technology, Brandon investigates the biomechanical factors that influence a pitcher’s mechanics. He is also expanding his research into overhand sports, using markerless motion capture to analyze volleyball hitting mechanics. After completing his master’s degree, Brandon plans to pursue a PhD in Biomechanics with the goal of working in professional baseball, ideally with a Major League Baseball organization.
P3: Kinematic Predictors of Pure Fastball Quality in Division I Pitchers
Adam Hoechstetter
Background: The effectiveness of a fastball depends on several variables, including velocity, spin rate, horizontal break, induced vertical movement (IVB), and extension. While velocity is the most studied determinant of fastball success, spin rate and movement display comparable influence on pitch effectiveness. Pure Fastball Quality (PFQ) was developed to summarize these physical pitch metrics into one composite statistic utilizing 2025 MLB pitch metrics and an expected weighted on-base average. Because PFQ can be calculated solely from pitch metrics, it allows for both on-field and training applications. The purpose of this study was to examine kinematic variables related to PFQ in collegiate pitchers.
Methods: Three hundred ten Division I pitchers (1.89±0.06m, 92.04±9.99kg) were included. In-game kinematic data were collected at 300Hz using an eight-camera markerless motion capture system and were processed and filtered using KinaTrax software. Kinematics of interest included stride length, stride width, and shoulder rotation at stride foot contact (FC); shoulder rotation and elbow flexion at maximum shoulder external rotation (MER); and arm slot and trunk flexion at ball release (BR). Peak kinematic values included horizontal shoulder abduction, hip-shoulder separation, rotational velocities (pelvis, trunk, shoulder), elbow extension velocity, forward center-of-mass (COM) velocity, and stride knee angular velocity. Ball velocity, spin rate, IVB, horizontal break, and extension were recorded using a TrackMan V3 Game Tracking unit. PFQ was calculated using these measurements, as a lower PFQ score indicates a higher-quality fastball. The average of each subject’s best five PFQ fastballs was included in the study. A backward multiple linear regression (α_OUT = 0.10) was used to identify the kinematic predictors of PFQ.
Results: The final regression model revealed seven significant variables, accounting for 44.3% of the variance (F(7,302) = 34.37, p < .001, R2 = .443). Significant predictors of PFQ include arm slot (β = .250, t(302) = 5.036, p < .001), stride length (β = -.401, t(302) = -8.847, p < .001), external shoulder rotation at FC (β = .130, t(302) = 2.950, p = .003), elbow flexion at MER (β = -.174, t(302) = -3.903, p < .001), trunk flexion at BR (β = .233, t(302) = 4.597, p < .001), trunk rotational velocity (β = -.107, t(302) = -2.384, p = .018), shoulder rotational velocity (β = -.155, t(302) = -3.510, p < .001).
Discussion/Conclusion: These findings indicate that higher arm slots, longer stride lengths, decreased external shoulder rotation at FC, increased elbow flexion at MER, increased trunk flexion at BR, greater trunk rotational velocity, and greater shoulder rotational velocity are associated with lower PFQ scores. Interestingly, previously established velocity predictors, such as lead knee angular velocity, forward COM velocity, and hip-shoulder separation, did not predict PFQ. In contrast, other factors, such as stride length and trunk and shoulder rotational velocities, showed significant relationships. Applying these findings may reduce a pitcher’s PFQ score, potentially enhancing fastball quality. While these results identified specific kinematics associated with PFQ, further investigation is recommended to explore the relationship between these kinematic variables and each component of PFQ in greater detail.
Adam Hoechstetter is a master’s student studying biomechanics at Auburn University. He is currently conducting research in the Sports Medicine & Movement Laboratory. His work utilizes KinaTrax and Theia markerless motion capture systems, along with Ascension electromagnetic marker-based motion capture, to examine the clinical application of biomechanics in baseball pitching to reduce injury risk and improve performance. He holds a bachelor’s degree in biology from Colby College in Waterville, Maine.
P4: Pure Fastball Quality (PFQ): An Interpretable PCA-Based Metric Predicting Batted-Ball Outcomes in MLB Pitchers During the 2025 Season
Kai-Jen Cheng
Background: Recent advances in baseball analytics have emphasized the need for interpretable approaches that describe the physical characteristics of a pitch. Principal component analysis (PCA) is well-suited for this purpose, as it groups correlated fastball characteristics into coherent mechanical dimensions. By reducing fastball data to a few orthogonal components, PCA reveals fundamental physical patterns that differentiate pitchers. These components can then be examined in relation to batting-outcome measures, such as Weighted On-Base Average (wOBA), to determine which aspects of fastball performance have the most significant impact on outcomes. The purpose of this study was to use fastball metrics from the 2025 MLB season to (1) derive a low-dimensional representation of intrinsic pitch characteristics using PCA and (2) identify which principal components significantly predict wOBA.
Methods: Fastball leaderboard data for 333 MLB pitchers in the 2025 season were obtained via Python-based web scraping from Baseball Savant, including full physical characteristics of fastballs. Physical characteristics of fastballs were used as inputs: ball velocity, spin rate, induced vertical break (IVB), horizontal movement, and release extension. All variables were standardized using z-scores before analysis. PCA was performed to identify the principal axes that describe the covariance structure of fastball traits, retaining components until the cumulative explained variance exceeded 95%. To identify performance-relevant components, wOBA was regressed onto the retained components using backward multiple linear regression with a removal criterion of αOUT = .10. Pure Fastball Quality (PFQ) was then constructed using the regression equation PFQ = β₀ + βᵀPC, representing the linear projection from PCA space onto expected wOBA.
Results: Five components (PC1-PC5) explained over 95% of the total variance in the physical characteristics of fastballs. PC1 reflected velocity, spin, and horizontal movement; PC2–PC3 captured vertical break and extension; PC4 represented spin and horizontal movement; PC5 combined velocity and horizontal break. Among these, PC1, PC4, and PC5 were negatively correlated with wOBA (PC1: r = −.145, PC4: r = −.302, PC5: r = −.316, all p < .01). These three components together explained 21.2% of the variance in wOBA scores (R2 =0.212). The predictive model incorporating these components was statistically significant (F (3, 329) = 29.58, p<.001). Regression coefficients indicated that lower PFQ scores were associated with lower wOBA values, with the final model expressed as: PFQ = .318 – 0.004*PC1 -0.011*PC4 – 0.013*PC5.
Conclusion: This study presents a clear, PCA-based framework for evaluating fastball quality through measurable pitch characteristics. By identifying the fundamental physical dimensions of fastball traits and linking them to batting performance metrics, PFQ provides a transparent, physically grounded measure of pitch effectiveness. Although developed in relation to wOBA, the PCA features represent intrinsic properties of fastballs independent of specific outcome variables. This approach captures key aspects of fastball performance, explaining approximately 21.2% of the variance in batting outcomes, and offers valuable insights for scouting and player development. Its simplicity, interpretability, and adaptability make it particularly useful in environments with limited data.
Kai-Jen Cheng is a scholar in biomechanics whose research examines how movement patterns influence performance and injury risk in throwing sports. He is currently pursuing his Ph.D. in Biomechanics at Auburn University under the mentorship of Gretchen D. Oliver. His work integrates motion analysis and applied sports science to identify efficient movement strategies that reduce mechanical stress on the body. He earned his master’s degrees from Texas Woman’s University and National Taiwan Sport University.
P5: The Role of Physiological Metrics in Baseball Pitching Performance
Brandon Doehne
Introduction: Physiological metrics can provide informative insights into an athlete’s state or response during a task. Previous cycling literature found that an individual’s average respiratory frequency was significantly correlated with their rating of perceived effort in self-paced cycling. Additionally, recent soccer research found that respiratory frequency demonstrated a faster response to supramaximal exercise during different work-rest ratios compared to heart rate. While literature has concluded that respiratory frequency may be superior to monitor at supramaximal intensities, utilizing heart rate in conjunction with respiratory frequency can provide more context for an athlete’s physiological state due to its stability and ability to quantify cumulative effects from a task. To the best of our knowledge, the use of both respiratory frequency and heart rate have not been monitored and utilized during an actual game in baseball. The purpose of this study is to investigate a pitcher’s respiratory frequency and heart rate during their in-game pitching appearance with their in-game performance.
Methods: Fifteen healthy college baseball pitchers (Height: 186.0 ± 7.4 cm, Mass: 88.38 ± 9.77 kg, BMI: 25.5 ± 2.2) were extracted from the UNO Pitching Lab database for preliminary analysis. Each pitcher wore a Hexoskin Pro Shirt (Hexoskin, Montreal, Canada) under their jersey while they pitched in a summer collegiate game or fall intrasquad game. During the pitcher’s appearance, manual annotations were created at each pitch release using the Hexoskin application. Pitching performance was recorded manually with pitch velocity (Stalker radar gun used during summer games and Trackman used for intrasquad games), strike or ball, pitch type, runner(s) on base, and outcome of the at-bat. Hexoskin’s internal software processed the original heart rate and respiratory frequency data which was output and down sampled to 1 Hz. Pitch windows were created and centered around the pitch release (t=0) and consisted of three seconds prior and two seconds following the pitch to encapsulate the physiological metrics during the pitch delivery. The average respiratory frequency and average heart rate were calculated for each pitch window throughout the game and served as general comparison among other pitchers to better understand the differences and intricacies among pitchers.
Results: Our analysis found that each pitcher had a unique respiratory frequency and heart rate profile while pitching. Among the pitchers, in-game success (number of pitcher-friendly outcomes) was not consistent pitcher-to-pitcher at similar physiological values. While two pitchers (Pitcher 3 and 4) had average heart rates near 165 bpm and similar average respiratory frequencies, one pitcher had five strikeouts within six batters, and the other had two earned runs with several walks.
Significance: While it was found that there is not one physiological state for all pitchers to be in to have better performance, the data suggests that there may be a physiological state specific to each pitcher that may enhance their performance. Future research will investigate the optimal physiology per pitcher to increase their performance.
Brandon Doehne is a Ph.D. student in biomechanics at the University of Nebraska at Omaha (UNO), expected to graduate in Spring 2028. He holds a bachelor’s degree in mechanical engineering from the University of New Mexico and a master’s degree in biomechanics from UNO. His work focuses on integrating biomechanics research with applied sports performance to deliver actionable insights for athletes and coaches. Brandon also has collegiate baseball experience as both a pitcher and hitter at the junior college and Division II levels, and as a pitcher at the Division I level.
P6: Amateur Baseball Profiler
Octavian Yuen
In this project, I want to classify pitchers into different prototypes based on their pitch usage tendencies in different counts, pitch movement profiles, and pitch results. I will be using a dataset with select Trackman data, a proprietary pitch-tracking system, from the 2025 season of the California Collegiate League, a summer league with 14 teams across California. Evaluating data from a summer league is especially unique because there are limited scouting resources, and the short seasons make efficient player evaluation particularly valuable.
I will use the Non-Negative Matrix Factorization (NMF) method to decompose an n by m matrix constructed from Trackman datasets. Each row will represent a player, and the columns will be organized by pitch movement profiles and pitch tendencies. This approach minimizes the Frobenius norm to achieve an optimal low-rank approximation of the data. The matrix will be factorized into a specified number of k components and will produce two interpretable matrices. The first one will have the size n by k and represent how strongly each pitcher corresponds to each of the k latent components. The second matrix will have the size k by m and will capture how each component loads onto the original features. To determine an appropriate value of k, I will analyze the rate of convergence of the Frobenius norm as k increases and approaches infinity. This will help identify the point where additional components yield diminishing improvements in reconstruction accuracy.
This methodology is inspired by Chapter 53 of Mathletics by Scott Nestler, Konstantinos Pelechrinis, and Wayne L. Winston, in which the authors applied a similar NMF-based approach to analyze NBA shooting patterns across different court areas.
Major League Baseball organizations and studies have explored clustering and classification methods to group pitchers by pitch type usage, velocity profiles, or movement characteristics. They used advanced tracking systems like Statcast to identify archetypes such as high-spin fastball pitchers, sinker/slider arms, and command-first profiles. However, such methods are rarely applied in amateur or developmental contexts, where data coverage is incomplete and variability in performance is high. This research extends existing frameworks into the summer league environment, which is an essential proving ground for collegiate players seeking professional opportunities.
Octavian Yuen is a second-year student at the University of California, Berkeley, studying Applied Mathematics and Economics. On campus, he serves on the board of the Sports Analytics Group, works as a data analyst for the softball team, and is helping develop a sports analytics course with Professor Graham of the Economics Department. This past summer, he was a Baseball Operations intern with the Alameda Merchants, a team in the California Collegiate League, and he currently helps with recruitment.
P7: Using Quasi-Experimental Methodologies to Analyze Pitcher Performance Post-Tommy John Surgery
Atul Venkatesh
A pitcher’s performance post-Tommy John surgery is a complete wildcard. How would they perform if they never needed surgery? With an inability to observe the counterfactual (if the pitcher had been healthy), we implement and compare two experimental methodologies: Difference-in-Differences (DiD) and Synthetic Control Method (SCM). The SCM constructs a weighted combination of healthy pitchers matched on pre-surgery statistics to create a synthetic version of the treated pitcher. By comparing the average performance of this synthetic doppelganger to the pitcher post-surgery, we can calculate how long it takes a pitcher coming back from injury to reach “healthy” levels and accurately assess the recovery process.
The analysis looks at 50 major league pitchers who underwent Tommy John Surgery from 2011 to 2022. These players must have pitched in at least 20 games or started 10 games each year. We then take the three most recent years before the player’s surgery, the pre-treatment period, and the first two years after surgery, the post-treatment period. For each player, a customized donor pool is created, consisting of pitchers who were fully healthy throughout the pre-treatment and post-treatment periods. Next, we apply the two techniques for the treated pitcher on two statistics: fielding independent pitching and strikeout rate. In other words, the synthetic pitcher should have similar strikeout rates and FIP to the treated pitcher before surgery. Lastly, the average treatment effect is calculated for each pitcher. For DiD, this is done by comparing the difference in pre-treatment and post-treatment years of the treated pitcher to the average difference in the donor pool in those same years. For SCM, this is done by comparing the synthetic pitcher’s performance to the pitcher’s actual performance post-surgery. The findings are two-fold. First is a case-by-case analysis on how well each pitcher recovers from Tommy John surgery relative to expectation. Second is the aggregated tendency of how pitchers tend to perform after undergoing the surgery. We found that, on average, the treated pitcher’s production dipped both one year and two years after surgery before climbing back to pre-surgery levels in year three.
This project has been conducted thanks to Dartmouth College’s Presidential Scholar program. I collaborated on this project with Professor Michael Herron of the Quantitative Social Sciences department.
Atul Venkatesh is a Junior at Dartmouth College majoring in Quantitative Social Science with minors in Statistics and Human-Centered Design. He is the President of Dartmouth Sports Analytics and enjoys applying technical concepts to research in sports. Atul is currently a Data Science intern for The 33rd Team and will be joining the Baltimore Orioles as an Analytics Fellow for the 2026 MLB season.
P8: De-confounding Location from Stuff+: A Location Invariant Pitch Quality Metric Using Gradient Reversal
Brody Chambers
The methods of analyzing pitch metrics has changed significantly in recent years with the advent of machine learning methods, particularly in “stuff+” models. Stuff+ models attempt to predict the run value of a pitch based solely on its physical characteristics (induced vertical and horizontal break, velocity, arm angle, etc), ignoring factors such as location or arsenal effects. This is done to isolate a pitcher’s stuff from things like poor command or sequencing, as well as good command or sequencing, in order to get a more accurate representation of how good their pitches are. There are a variety of models that have been publicized, such as Driveline’s proprietary model, FanGraphs’ Stuff and pitchingbot stuff; however, simply removing location data from the dataset used by a model will not remove all of the bias of location. Pitch shapes that typically must have better command/arsenal effects to perform well – for instance slower fastball shapes – could bias the model to evaluate pitches more favorably than they should objectively be. Thus, this paper proposes a new adversarial stuff+ architecture to combat this confounder. First, the model separately evaluates the value of the location of a given pitch. Then, the physical characteristic data is fed through a neural network based encoder which trains on both a run value head as well as an adversarial head that attempts to predict location run value. This adversarial head is put through a gradient reversal layer in order to remove the influence of location. This new model yields a more pure stuff metric, that can more robustly evaluate players. In the future, hopefully these findings can be further expanded to a model that removes sequencing and arsenal effects to even further isolate stuff.
Brody Chambers is a student at Oakmont High School in Roseville, California and pitches for his varsity baseball team. He has worked on a few independent projects with pitching analytics and hopes to one day work as an analyst. Outside of baseball, he enjoys math and chess.
P9: Do Plate Coverage Weaknesses Matter?
Jackson Scrimpshire, Willie Miracky, and William Haray
Understanding how a hitter’s performance varies across different regions of the strike zone is central to evaluating plate coverage and identifying potentially exploitable weaknesses. Previous work has examined the relationship between pitch location and offensive outcomes, including Pemstein’s (2015) analysis of how location influences batted-ball metrics and Roegle’s (2013) investigation of expected versus actual wOBA using pitch velocity, movement, and placement. Additional public analyses have mapped wOBA values across locations and explored how zone classification affects hitter performance. While these studies demonstrate the strong role of location in run value, none directly assess how intra-hitter variation in zone-specific performance relates to overall offensive productivity, nor the degree to which pitchers systematically target identified weaknesses. This gap leaves open important questions about how meaningful plate-coverage deficiencies truly are in practice.
The study uses pitch-level Statcast data from the 2024 MLB season to quantify the relationship between a hitter’s zone-specific performance and their overall offensive productivity. For each pitch, the Statcast provided expected wOBA components (woba_value and woba_denom) were aggregated by each of the 14-zone strike zone model to compute a hitter’s wOBA in each individual zone. League-wide and player-level summaries were then generated to evaluate how variation in zone performance reflects plate coverage ability.
To assess the importance of plate coverage, the study measured within-player variability in zone-specific wOBA as an indicator of coverage consistency. Higher variability indicates concentrated strengths and exploitable weaknesses, while lower variability reflects more uniform performance across the zone. To evaluate whether pitchers actively target these weaknesses, pitch frequency by zone was correlated with each zone’s relative “coldness” (defined as the difference between a hitter’s maximum zone wOBA and their wOBA in the given zone). Positive correlations indicate pitcher exploitation of hitter weaknesses.
This approach enables direct comparison between (i) location-specific offensive value, (ii) overall offensive performance, and (iii) the extent to which opposing pitches adjust their attack patterns to exploit coverage deficiencies.
Results show almost no relationship between a hitter’s overall wOBA and their degree of zone-to-zone variation (r = 0.05), suggesting that having distinct hot and cold zones does not meaningfully predict overall offensive productivity. However, pitcher behavior exhibited wide variability: correlations between pitch frequency and zone coldness ranged from –0.60 to 0.70, with a median of 0.11. The asymmetrical distribution indicates that some hitters are targeted heavily in their weakest zones, while for many others, pitchers do not (or cannot) systematically exploit these areas.
These findings challenge the common assumption that hitters with pronounced cold zones are inherently more vulnerable. By showing that zone-specific weaknesses have limited predictive value for overall production and are inconsistently exploited by opponents, this work offers new insight for player evaluation and development. Organizations may benefit from nuanced assessments of plate coverage rather than assuming that visible zone deficiencies automatically translate to reduced performance or tactical predictability.
Jackson Scrimpshire is a junior at Tulane University majoring in mathematics and computer science. He is the president of the Tulane Sports Analytics Club and works as a Data Scientist for Phoenix Rising FC. He plans to obtain his Master’s of Data Science at Tulane.
Willie Miracky is a junior at Tulane University majoring in mathematics and economics. He is a member of the Tulane Sports Analytics Club and aspires to work in soccer or baseball analytics. He plans to complete his Master of Science in Statistics at Tulane.
William Haray is a junior at Tulane University majoring in economics. He is a member of the Tulane Sports Analytics Club and works in the data department for Tulane’s baseball team.
For more information on the 2026 SABR Analytics Conference, visit SABR.org/analytics.