Poster Session (in alphabetical order by presenters)

1. PREDICTING NBA PLAYER SUCCESS BASED ON COLLEGE PERFORMANCE

Josh Avram, Saint Ann's School

Abstract: Can college basketball stats predict a player's NBA success if adjusted for competition level? By evaluating competition level and statistical translation, we aim to develop create an adjustment model to better predict NBA performance. Our methodology includes calculating opponent strength using historical data, adjusting college stats based on competition level, and comparing adjusted stats to first-year NBA performance. To accomplish this we developed a predictive model using regression/machine learning.

2. Schedule Path Ratios: A Graph-Based Approach to Quantify Schedule Balance

Luke Benz, Harvard T.H. Chan School of Public Health

Abstract: Schedules across college and professional sports display display substantial heterogeneity. In this work, we introduce a novel metric to quantify schedule balance in a manner that is agnostic to team strength. This metric, which we title schedule path ratios, compares graph-based distances in the observed schedule network to those in a network which represents an optimally balanced schedule. We illustrate this metric both globally and locally (within conferences and divisions) across major North American professional sports and NCAA football. Finally, using schedule path ratios, we conduct a simulation based analysis which demonstrates that a decline in schedule balance over the past 20 years has made it more difficult to identify the best teams, a salient point given the new college football playoff.

3. A Statistical Analysis of Tournament Structures

Zachary Allen Culp, Loyola University Chicago

Abstract: We run multiple different types of tournament structures to compare teams' true ranks to their simulation ranks. In order to numerically determine which structure is better in accurately outputting the correct ordering of teams, we create plots for the information (currently using mutual information) obtained by many simulations.

4. Plus-Minus Adjusted Impact with Generalized Estimation: Building a WNBA RAPM model

Amer El Zein, University of Connecticut

Abstract: Regularized Adjusted Plus Minus has become a cornerstone of modern basketball analytics, creating a measure of player impact by combining On-Off court differentials with regularized regression techniques. While RAPM and its successors such as RPM, EPM and LEBRON have shaped player evaluation in the NBA, no such model has been systematically applied and made public for the WNBA. This project presents the development of a custom RAPM model for the WNBA using publicly available play by play data spanning multiple seasons. To address the shorter season, RAPM is stabilized using box-score statistics as empirical priors, allowing the model to borrow from known performance indicators and stabilize estimates for players with sparse minutes. Furthermore, a time decay factor is applied, weighting recent seasons more heavily to reflect current player ability and trends. The result is a robust and transparent RAPM framework tailored to the WNBA, allowing reliable insights to player impact and laying the groundwork to advanced analytics in women's basketball.

5. Leveraging Alternative Data for Mixed Martial Arts Betting Markets

Eugene Han, Yale University

Abstract: The rise of sports betting has led to increased interest in developing data-driven strategies for identifying market inefficiencies. While predictive modeling for sports betting is well studied in team-based sports, individual combat sports such as mixed martial arts (MMA) present unique challenges due to limited data availability, outcome volatility, and the sport’s dynamic nature. Existing public research has relied primarily on data from a single source, UFC Stats, which provides detailed fight metrics but lacks broader contextual information. This study explores the use of alternative data sources—including fight history databases, betting odds aggregators, and community-driven metadata—to construct a more comprehensive dataset for modeling fights in the Ultimate Fighting Championship (UFC). By leveraging diverse data streams, this approach aims to uncover hidden signals that may provide an edge. Additionally, we experiment with ideas from conformal prediction and robust optimization to optimize bet sizing under uncertainty. To evaluate performance, our methods are tested on out-of-sample UFC fights from 2017 to 2024, with a focus on profitability and calibration. The results provide insight into the potential for systematic edges in MMA betting and a starting point for incorporating more diverse data sources into the modeling process.

6. Faster And Longer, The Swing That Makes You Stronger: Identifying Successful Swing Metrics In Major League Baseball Using K-means Clustering

Charlotte Imbert, Pomona College

Abstract: This study examines the relationship between bat speed, swing length, and baseball batting performance using k-means clustering. The goal is to determine whether distinct swing profiles are associated with more favorable offensive outcomes and how batters adjust their swings in different pitch contexts. Data from the 2024 MLB season were wrangled to address missing values and reduce multicollinearity. Composite swing metrics, including swing efficiency (bat speed/swing length) and swing combined (bat speed × swing length), were introduced. K-means clustering grouped swings based on bat speed and swing length, and batter performance was assessed using a batter score metric that incorporates both event value (e.g., single, home run) and changes in run expectancy. Statistical tests, including permutation tests and Kruskal-Wallis tests, were conducted to compare clusters. Three swing clusters were identified. The cluster with the highest bat speed and swing length exhibited the most favorable batting outcomes, while the cluster with the lowest values performed poorest. Swing adjustments were observed in high-pressure situations, with batters reducing bat speed and swing length when facing two-strike counts. The findings suggest that both high bat speed and longer swing length contribute to better offensive performance, with situational adjustments playing a role in approach at the plate.

7. Thirst for Victory: Exploring the Influence of Coca-Cola vs. Pepsi Sponsorship on College Basketball Teams' Performance in March Madness

Ryan Juricic, Belmont University

Abstract: This study examines the impact of corporate sponsorship by Coca-Cola and Pepsi on college basketball teams' performance in the 2024 NCAA March Madness tournament. Given the prominence of corporate sponsorship in collegiate athletics, understanding its influence on competitive outcomes is valuable for athletic programs and corporate partners.

Using data from the 2024 season, we analyzed whether a college's beverage sponsor correlated with its tournament performance. Sponsorship data was sourced from open web sources, and team data was obtained from Kaggle. Fishers exact test assessed the association between sponsorship and the round reached, while random forest and logistic regression analyses evaluated sponsorships predictive power.

Results indicate no significant relationship between sponsorship and tournament participation (p > 0.05). However, a statistically significant dependence was found between sponsorship and advancement (p = 0.01171). Notably, only Coca-Cola-sponsored schools reached the Elite Eight and beyond. Despite this, neither the random forest nor logistic regression identified sponsorship as an influential predictor of performance.

These findings suggest a potential, yet unexplained, relationship between corporate sponsorship and team success in the NCAA tournament. Future research should explore this association across multiple years and sports to determine its broader significance.

8. Investigating the Effect of Extenuating Life Circumstances on NBA Player Performance

Brady Kim, The Hotchkiss School

Abstract: How do extenuating life circumstances have an effect on player performance in the NBA? This project seeks to investigate the effect of such events on performance by using BPM (Box Plus/Minus) and PER (Player Efficiency Rating). For this project, a random sample of 100 current NBA players were researched using publicly available information on the internet. For each player, average BPM and PER was calculated for each player in two separate categories: seasons with extenuating life circumstances and normal seasons. Wilcoxon signed-rank tests and paired t-tests were conducted for the differences in BPM and PER between normal seasons and seasons with extenuating circumstances, which produced p-values of 0.000011, 0.000035 for BPM and 0.000037, 0.0012 for PER. Overall, when dealing with extenuating life circumstances 65% of players had a higher BPM, 30% had a lower BPM, and 5% experienced no change. For PER, 69% of players had a higher PER, 27% had a lower PER, and 4% experienced no change. These results suggest that extenuating life circumstances do have an effect on player performance, and that players can be categorized into three groups: players that thrive under stress and perform better, players that perform consistently under both normal and extenuating circumstances, and players that crumble under stress and underperform. For this sample, a majority of players actually performed better when dealing with extenuating life circumstances. This suggests that the mental aspect of how players perform under stress could be another helpful metric in player evaluations.

9. Profiling NASCAR Qualifying Performance with Functional Data Analysis

Joshua Lee, University of Connecticut

Abstract: NASCAR is the premier American motorsport, with millions of fans across the United States and beyond eagerly tuning in each week to see who comes out on top in a test of skill, endurance, and most of all, speed. As with any sport, winning is of paramount importance, and prior research has attempted to model important driver and team characteristics, as well as methods for optimizing driver performance towards this end. However, no prior work has investigated ways to improve driver qualifying performance. This problem is significant since qualifying well will naturally improve the likelihood of achieving a strong finishing position in the race that follows. To address this gap, we analyze qualifying lap data from the NASCAR AdventHealth400 through the use of functional principal components analysis, and agglomerative hierarchical clustering. This allows us to generate distinct groups of drivers from which typical behavior patterns can be extracted. In particular, we observe multiple patterns of braking, throttle application, and steering which are differentially associated with qualifying performance. By identifying the highest performing clusters, we successfully identify actionable insights which can be used to enhance the qualifying performance of other drivers and ultimately achieve stronger race results.

10. Optimizing Batting Orders : Monte Carlo Game Simulation Based on Batter Swing Clustering

Juwon Lee, Yonsei University

Abstract: Can analyzing a batters swing type provide deeper insights into their overall performance and potential? If so, how can teams leverage this insight for player evaluation and lineup optimization? Our analysis explores whether Swing Length and Bat Speed, newly introduced by Major League Baseball through Statcast, reflect intrinsic batter characteristics. By integrating these physical swing metrics with plate discipline indicators, we classified batters into distinct swing profiles and examined their correlation with key performance metrics such as OPS and wRC+.

To assess the strategic implications of swing analysis, we employed Monte Carlo simulations to estimate expected run production across various lineup configurations. Unlike traditional models that primarily focus on batted ball data, our approach incorporates both swing mechanics and decision-making tendencies. This framework enables teams to optimize their lineup by testing multiple configurations and selecting the one that maximizes run production.

In addition to lineup construction, we developed scenario-based simulations to guide roster decisions. One scenario evaluates the optimal placement of a newly acquired player within an existing lineup to maximize offensive outputproviding valuable insight for free-agent acquisitions. Another scenario identifies the ideal swing profile for specific lineup spots, helping teams address underperforming positions with the most suitable player types.

Our findings introduce a novel approach to batter evaluation and lineup optimization, integrating swing analysis with probabilistic modeling. By combining swing mechanics with strategic decision-making, this research offers practical insights for player development, scouting, and in-game management.

11. Assessing Bias by Athletic Conference in Ratings Percentage Index (RPI) Rankings for NCAA Division I Men's Soccer Teams

Conrad Lee, Yale University

Abstract: This study examines the hypothesis that the Rating Percentage Index (RPI) metric systematically overestimates teams from stronger conferences in NCAA Division I Men’s Soccer. Alternative RPI weightings are proposed and tested, first by using Bradley-Terry-Davidson simulations and then using real data from the 2024 NCAA season, to determine whether they reduce conference bias while improving predictive accuracy. Simulations show that the standard RPI formula is less accurate and tends to favor teams from stronger conferences when compared to alternative weights which emphasize performance over context. However, these findings do not hold in the 2024 season data, where context-heavy RPI formulas beat selected alternatives from an accuracy standpoint despite their preference for teams from stronger conferences. These results suggest that alternative RPI weights hold potential, but further multi-season analysis is warranted.

12. Quantifying Clutchness: A Data Driven Analysis of Performance Under Pressure in Baseball

Ethan Long, University of Connecticut

Abstract: The concept of “clutchness” in professional baseball has long been debated, yet remains difficult to quantify. This study aims to investigate whether high-pressure situations influence player performance and whether distinct “clutch” archetypes exist. Using regular season and postseason data from Baseball Reference, we analyze three key performance metrics: On-Base Plus Slugging (OPS), Total Bases (TB), and Championship Win Probability Added (cWPA)—to assess how players respond under varying levels of pressure. We employ hypothesis testing to determine statistical significance in performance differences, K-means clustering for classifying players into specific archetypes, and logistic regression models to forecast probability of success in high-leverage situations. By integrating statistical analysis with sabermetrics, this study seeks to provide a data-driven perspective on the existence of a “clutch gene” in baseball. Our findings could have implications for player evaluation, recruitment strategies, and in-game decision-making, offering a more objective framework for understanding performance under pressure.

13. The Effect Of Injuries On Bat Speed

Reese Mullen, Boston University

Abstract: This study explores the relationship between swing speed trends and injury-related absences among Major League Baseball (MLB) players. It was inspired by Rafael Devers’ injury- impacted 2024 season. Devers’ swing speed significantly declined following his midseason injury, motivating a broader investigation into whether similar trends could be identified across other MLB players. The analysis examined swing speed and length using rolling averages, highlighting potential injury and performance decline indicators.

14. Extracting & Analyzing Squash Match Patterns Using a Low-Cost Computer Vision Framework

Abhi Nagireddygari, Bowdoin College

Abstract: Despite advancements in sports analytics, there is limited research applying computer vision techniques to squash. This study presents a novel, low-cost AI system that uses a single-camera setup to detect players, the ball, and court boundaries using deep learning models. The system extracts spatial and temporal data from match footage, allowing for detailed performance analysis without the need for expensive multi-camera setups. Top-down heatmaps are generated to illustrate player positioning, movement trends, and time spent in different court areas. In addition, we can determine when the ball is hit, from where, and the location it lands with respect to the opposing player. These insights can help players and coaches refine strategies, optimize training, and improve decision-making. In addition, in our work, we analyze patterns of movement and shot selection by shared characteristics that may influence tactics or behavior such as gender, nationality, and skill level. By reducing hardware costs while maintaining accuracy, this method makes data-driven squash analytics more accessible. The project addresses a key research gap in squash-related AI applications, showcasing the potential of Computer Vision for match evaluation.

15. Analyzing Different Swing Types Using Bat Tracking Data

Danielle Napierski, Brett Cerenzio, and Dan Griffiths, Syracuse University

Abstract: For some time, there has been a disagreement among fans about the ideal swing type for batters. Some argue that batters are too selfish, citing low batting averages, rising strikeout rates, and increasing home run swings. Others argue that low batting averages and rising strikeout rates have to do more with the advancement of pitching and that hitting for power, rather than contact, typically yields better results in the long run. There are also those who argue that batters should change their swings in different situations. For example, they can swing for the fences when the bases are empty, but should look to shorten their swing and emphasize contact when there are runners in scoring position. For this project, we wanted to evaluate if different swing types are needed to maximize run production or if theres one swing type that provides a perfect balance between power and contact. We utilized k-means clustering to classify seven different swing types and then created two RBI matrices. One evaluates the expected number of RBIs on a given swing based on swing type, base situation, and number of outs, while the other finds the probability of at least 1 RBI scoring on a given swing based on swing type, base situation, and number of outs. We also created an XGBoost model that predicted the change in win probability based on swing type, base situation, number of outs, count, score, inning, and batting team score differential.

16. The Influence of Bat Speed and Swing Length on Fouling Off 2-Strike Pitches: A Bayesian Causal Analysis

Ebenezer O Olubayode, University of Oklahoma

Abstract: Introduction Fouling off 2-strike pitches is a strategic skill in baseball that extends at-bats, tires pitchers, and improves offensive outcomes. However, batter mechanics (bat speed, swing length) influencing this skill remain unclear. This gap prevents players, coaches, and analysts from developing true cause-and-effect strategies beyond simple correlations, limiting strategic adjustments in high-pressure situations. This study examines how these mechanics influence 2-strike fouling probability using Bayesian causal inference, providing actionable insights for player development and decision-making. Methods Pitch-level data from Baseball Savant was used to analyze bat speed, swing length, and pitch movement in 2-strike fouling behavior. Although limited to the 2024 MLB season, this methodology provides a foundation for future research. A Bayesian logistic regression model estimated fouling probability while accounting for uncertainty and confounders. Contrast means and High-Density Intervals (HDI) validated statistical significance, while Gelman-Rubin statistic & Effective Sample Size confirmed model robustness. Results Shorter swings significantly increased fouling probability, with batters using shorter swings fouling off 48% of 2-strike pitches, compared to 37% for longer swings. Higher bat speed had a smaller effect, with batters swinging faster fouling off 44% of such pitches, compared to 41% for slower swings. Vertical pitch movement amplified these effects, making short swings more effective against high pfx_z pitches. Conclusion This study integrates batter mechanics and Bayesian causal inference to provide a fuller picture of 2-strike fouling behavior. Findings offer practical insights for refining hitting strategies and pitching tactics, and supports the growing use of causal inference to enhances player evaluation and decision-making.

17. Data Science Meets Athletics: bridging academic learning and athletic excellence

Jessica Joy Pasteur, Goucher College

Abstract: Data analytics is playing a beneficial role in track and field. Wearable devices can easily generate personally actionable data ranging from biomechanical to weather conditions. Students in the Computer and Data Science Department at Goucher College are collaborating with the colleges Track and Field teams to develop personalized machine learning and analytics tools. We present initial steps in our collaboration with coaches to optimize training, help tailor athlete schedules, and ultimately predict race outcomes. This bridge between academic learning and athletic excellence is providing athletes with advanced analytical tools while giving students hands-on experience and generating real-world case studies.

18. A Bayesian Approach to Ulnar Collateral Ligament (UCL) Reconstruction

Brady Pinter, Belmont University

Abstract: Ulnar Collateral Ligament (UCL) reconstruction, commonly referred to as Tommy John Surgery, has seen a significant rise among Major League Baseball (MLB) pitchers in recent decades. While previous studies have explored correlations between pitching mechanics and UCL injuries using frequentist statistical approaches, the present study employs a Bayesian hierarchical model to better quantify the relationship between pitch characteristics and the likelihood of requiring UCL surgery. By incorporating prior knowledge and leveraging Bayesian ANOVA techniques, our approach offers a more structured and probabilistic assessment of the factors contributing to UCL injuries. Our findings aim to refine injury prediction models and provide a more comprehensive statistical framework for understanding the mechanics underlying UCL injuries in professional baseball.

19. Building Data-driven Tools for Yale Women's Soccer

Rebeka Róth, Yale University

Abstract: My project focuses on helping the Yale Women’s Soccer team by presenting their data in a way that is accessible to coaches. By developing an interactive user interface, the tool will enable coaches and staff to compare their players’ statistics within the team or with those of other Ivy League players, utilizing data visualizations like bar charts, radar maps, and scatter plots. This initiative addresses the need for data-driven insights for the team, which are currently limited to post-match reports and static spreadsheets.

20. F1 versus Indy: Analyzing a Unique Shared-Course Natural Experiment to Determine the Worlds Fastest Auto Racing Format

Simran Sanders, Jamesville Dewitt High School (Jamesville, NY)

Abstract: This study utilizes a unique shared-course environment to compare the average speeds of the worlds two fastest auto racing formats, Formula One (F1) and IndyCar. These two racing circuits are distinct in terms of their places of origin, where in the world they typically source drivers and hold races, and the locations of their respective fan bases. Though F1 and IndyCar cars share several common physical characteristics, their engines, chassis, tires, steering, and aerodynamics are distinct. While these two formats are commonly recognized as the two fastest forms of full-course auto-racing, the relative average speed capabilities of the two formats (e.g., over a neutral course) are unclear, as they have historically held official events on different sets of courses. In 2019, however, at the Circuit of the Americas in Austin, TX, F1 and IndyCar both held races over the same course, allowing for a natural experimental comparison in which year (state of technology), course, and even general weather conditions were common across the two races. By modeling qualifying times conditional on driver, race format, and other car characteristics, we statistically compare the relative average speeds of the two formats over the same course. In a set of fixed and mixed effects linear and polynomial regression models, we find that F1 was significantly and substantially faster than IndyCar. Average speed difference between the two formats was approximately 16.5 miles per hour over the course. These results inform not only race speed supremacy but returns to technological investment in the private sector.

Keywords: natural experiment, auto racing performance, F1, Indy Car, fixed effects, mixed effects

21. Investigating the Presence of Mechanical Deviations in Baseball Swings Over Time

Khushi Shah, University of North Carolina at Chapel Hill

Abstract: We participated in the competition and were selected to present our poster! "In professional baseball, every swing reflects a decision influenced by pitch characteristics, game context, and a batter's intent. However, deviations from expected swing mechanics may indicate a breakdown in consistency throughout a game. This study investigates the presence and progression of deviations from expected swing mechanics throughout a game, focusing on two key metrics: swing length residuals and the relationship between bat speed and launch speed. Using pitch-level data from Major League Baseball, we apply regression models and residual analysis to quantify these deviations. A logistic regression indicated a 26.4% increase in unintended swing likelihood per additional swing in an inning (p < 0.001). Swing length analysis revealed increasing mechanical variability, with mean squared error increasing significantly with swing count (Estimate: 0.00176, p = 0.00467). These trends suggest a pattern of mechanical inconsistency, potentially linked to fatigue or cognitive strain. This research provides a foundation for further exploration into batter performance trends and offers insights for coaching strategies aimed at injury prevention and fatigue management."

22. The Power of Proximity - Using Network Science to Analyze the Link Between Coaching Communities and Team Success in the National Football League

Atul Venkatesh, Siddharth Singh, and Charles Betts, Dartmouth College

Abstract: Coaching staff cohesion has been one of the most overlooked aspects of team success. A group of coaches that have experience working with each other should theoretically foster team success. But is this true in reality? To answer this question, we created a first-of-its-kind coaching network that maps connections between every head coach, coordinator, position coach, and quality control coach from 2010 to 2024. Our analysis revealed a statistically significant positive association between the closeness of coaching communities and key team performance metrics, including win percentage, playoff appearances, and division titles. These findings highlight the value of long-term coaching relationships, shedding light onto the potential benefits of front office patience and continuity. In addition to these findings, we employed modularity-maximizing clustering algorithms to identify distinct “coaching communities,” which represent the different coaching philosophies in the NFL today. The algorithm divided our coaching network into 19 different communities, 13 were found to be significant. Lastly, inspired by separation theory, we measured the closeness of every coach to Kyle Shanahan. Shanahan has been praised for his influence and success, with several of his disciples earning head coaching jobs. By analyzing the proximity of other coaches to Shanahan, we aimed to understand the extent of his influence on the broader coaching landscape. This could perhaps inform how connections to an influential figure impact team success and offer some power in predicting future head coaches. By examining the structure of coaching networks like never before, we uncover valuable insights that could influence decisions across the NFL.

23. Acing Predictions: Logistic Regression in the 2024 US Open Men's Tennis Championship

Krithin Visvesh, North Carolina State University

Abstract: Predicting match outcomes in tennis poses a significant challenge due to the sport's unpredictable nature and the influence of numerous factors on player performance. This study seeks to forecast the top 10 ranked athletes and their respective winning probabilities to win the 2024 US Open Men's Championship using Logistic Regression. The research analyzes data from US Opens from 2016 to 2023. The primary variables selected for the analysis are the winner's rank and the opponent's rank, applied in a logistic regression model using an 80/20 train-test split. The test accuracy was 68%. The probability of winning the US Open was also calculated for the top 10 ranked players, finding that the No.1 ranked players probability of winning the US Open was 3.1%. As No.1 seeds have won 28.9% of the mens US Open singles tournaments, this suggests that using the players and opponents rank is insufficient to determine the probability of an individual winning the US Open.

24. Regularized adjusted plus-minus: pseudo-Bayesian player evaluation in hockey and performance improvements through design matrix transformations

Jacqueline Wang, Yale University

Abstract: Regularized adjusted plus-minus (RAPM) models are foundational tool in hockey analytics for evaluating individual players in terms of their offensive or defensive contribution. We examine the performance of RAPM models with respect to matrix formats (dense, compressed sparse row, compressed sparse column, coordinate list) as well as collapsing portions of the design matrix into a value relative to the league average. We also investigate the use of box score statistics to predict RAPM coefficients as a step toward a fully Bayesian RAPM model and explore whether expected goals-based alternative metrics improve the results.

25. Residual Analysis of Pitch Sequences to Maximize Whiff Rate

Jack Whitney-Epstein, Brunswick School

Abstract: In baseball, one way to make pitchers more effective (without improving their mechanics) is to select the right pitches. I wanted to know which pitch sequences were the most unpredictable and difficult to bat against; specifically, I searched for those which generated the most whiffs. I implemented a generalized additive model to predict a pitch’s whiff probability given its pitch type - fastball, changeup, slider, or curve - and the hitter’s bat speed and swing length. Then, I looked at sequences of pitches to see which ones generated more swings-and-misses than expected. Some sequences emerged as more effective than others, demonstrating that pitch selection is not just a superficial task for the catcher but a strategic medium for the pitcher to gain an advantage. Results include the best overall sequence, the best starting pitch and follow-up, and specific combinations of pitches to throw or avoid.

26. Predicting The Next Pitch Type and Analyzing Batter-Pitcher Strategic Interactions

Claire Yang, New York University

Abstract: The strategic interaction between pitchers and batters has always been one of the highlights of MLB games. While it is well established that batters’ behavior are more passive, adjusting their actions in response to the pitcher’s movements, in this study, we aim to explore a more counterintuitive scenario: how pitchers, in turn, are influenced by batters and other on-field contexts. Using models such as Random Forest and BART (Bayesian Additive Regression Trees), we predicted pitch types and examined the causal effects of swing characteristics on pitching strategies. Our findings revealed that Random Forest achieved the highest accuracy (42.02%) in the predictive model. Additionally, we found facing the last bat speed outside the 50-75 mph range, pitchers appear highly inconsistent in their decision-making regarding fastball choice, indicating that extreme cases have a significant impact on their strategy.