CARNEGIE MELLON SPORTS ANALYTICS CONFERENCE

Albert Cohen is the interim Director of the Actuarial Science Program at Michigan State University, in addition to directing MSU’s Graduate Certificate in Sports Analytics. Albert received his Ph.D. from Carnegie Mellon University in 2007 under the supervision of David Kinderlehrer, and his research focuses on the interaction between probability and science, sports, and financial economics.

This focus began with his work on a stochastic approach to the coarsening of cellular networks, and has since branched out to develop new risk measures using stochastic optimal control. This approach has also led to models of online auction behavior, the structural modeling of default bond pricing under the incorporation of recovery processes, and (recently) the use of tools commonly found in risk management to analyze challenges in player contract valuation.

Mathematical Models of Player Contract Valuation

How does a team properly compensate a player for their on-field efforts? In this brief talk, we will outline foundational tools first proposed by Scully (with the onset of MLB free agency in the 1970’s) and expanded upon by Rockerbie and others. We will carry out two pricing examples, and no previous quantitative finance or actuarial tools are needed. In fact, a little bit of multi variable calculus and a desire to apply it to sports is all that is needed to follow this talk. We hope that audience members will want to follow up the presentation with further work in contract pricing within sports analytics.

Monnie McGee is an Associate Professor in biostatistics and bioinformatics in the Department of Statistics and Data Science and SMU in Dallas, Texas. She earned a B.A. in Mathematics and English from Austin College in Sherman, Texas, and M.A. and Ph.D. degrees in Statistics from Rice University. She specializes in analysis of high-throughput biological data and is an expert in background correction, normalization, and hypothesis testing for these techniques. She is also interested in sports analytics, particularly analysis of “individual-team” sports like track & field, swimming & diving, and gymnastics. For all the applications previously mentioned, she has applied and developed methods for compositional data analysis, particularly those methods using the Nested Dirichlet Distribution and related distributions. Dr. McGee has taught and developed many courses in statistics at the graduate and undergraduate levels, including proposing several full programs at all academic levels.

Sources of bias in diving competitions: prevalence and adjustments

Sports such as diving, gymnastics, and ice skating rely on expert judges to score athletic performance during competition. To assure that scores are fair and comparable, judges are trained to follow a set of standards for various components of a routine and align their scores to these standards. However, eliminating subjectively is impossible even for the most conscientious and experienced judges. Nationalistic bias, where a judge from a particular country favors participants from his or her country (while disfavoring others), is the most infamous type of bias, but it is not the only kind of bias. Other more subtle forms of bias include difficulty bias, where athletes will more difficult routines receive better scores, and round bias, where scores increase or decrease during the course of a competition. In order to determine the prevalence of various types of bias in diving competitions, we have assembled results from divemeets.com, a database of most diving competitions, from youth to college, in the US. In this talk, we present the results of our analysis and provide adjustments for such types of bias when necessary.

Devin is the Senior Director of Research and Development for Team Operations at MLSE, with leadership roles in both the Sport Performance Lab and SportsX programs. These groups are tasked with executing strategic research initiatives to develop competitive advantage for MLSE teams on the court, field, and ice. Prior to this cross-sport role, Devin led the analytics department at Toronto FC for 8 seasons, which included winning an MLS Championship in 2017. With Toronto FC, Devin led the integration of analytics and data science into opposition analysis, player recruitment, and performance forecasting. Before joining Toronto FC, Devin worked for Opta Sports, Major League Soccer, and as a soccer coach. Devin has a technical background in both Computer Science and Goalkeeping and is a regular contributor to the sports analytics academic community.

From Fixed to Fluid: A Model for Frame-by-Frame Player Role Classification

Traditional perspectives on soccer formations evoke images of rigid, symmetrical lines of players categorized into different nominal positions. Like a chess board. However, the contemporary demands of the sport necessitate that players dynamically adapt, often assuming multiple roles within the span of a single game. This phenomenon is accentuated by an emerging trend of teams deploying asymmetrical formations, with player rotations built directly into the game plan. A case in point: a left defender may occasionally required to position themselves as a left winger when the left right winger pinches inward to behave as an attacking central midfielder. In sophisticated systems, these sort of adjustments often trigger a cascading re-categorization across the field. This rapid positional interchanging, taking place within a dynamically evolving setting, poses challenges for cataloging and analysis. Gaining insights into such rotations is critical, especially for opposition analysis and in-game tactical adaptations. Currently, this is primarily achieved through traditional video analysis. Harnessing the potential of player tracking data, we have engineered a model that achieves frame-by-frame role classification for individual players. This approach discerns the frequency at which a player adheres to their assigned position but also deciphers the plethora of roles they transition into amidst the game's inherent chaos.

Sara is a senior at Pomona College majoring in Mathematics and Philosophy, specifically focusing on Statistics and Ethics. She is deeply interested in how we can combine statistics and ethics to help inspire good decision-making practices. After graduation, she plans to pursue a Ph.D. in Statistics and supplemental graduate coursework in Ethics. Outside of academics, she also enjoys running and is a member of Pomona-Pitzer's Varsity Cross Country and Track and Field teams.

Clustering Racehorse Movement Profiles to Discover Trends in Injured Horses

Between 2009 and 2021, over 7,200 horses died or were euthanized due to racing-related injuries (Fobar, 2023). Using horse profile data and horse tracking data from the NYRA, we identified horses who under-raced between 2019 and 2021 and clustered movement profiles for horses who raced in 2019 New York races to discover if certain profiles were more associated with injured horses. By fitting a negative binomial model on horse profile data and performing residual analysis, we discovered that at least 251 horses under-raced between 2019 and 2021. Additionally, clustering horse movement profiles revealed that a horse’s speed profile is most associated with its injury status; specifically, greater variation in speed is more associated with injury.

Jonathan Pipping is a senior at the University of Florida studying Statistics and Economics. In addition to his coursework, Jonathan serves as the VP of Projects for the university's Statistics Club and is actively working towards two undergraduate theses in his fields of interest. In his free time, he enjoys playing music, working on data challenges, and watching football with friends. After graduation, he plans to complete a Ph.D. in Statistics and Data Science before pursuing a career in sports, biomedical, or economic research.

Clustering Racehorse Movement Profiles to Discover Trends in Injured Horses

Between 2009 and 2021, over 7,200 horses died or were euthanized due to racing-related injuries (Fobar, 2023). Using horse profile data and horse tracking data from the NYRA, we identified horses who under-raced between 2019 and 2021 and clustered movement profiles for horses who raced in 2019 New York races to discover if certain profiles were more associated with injured horses. By fitting a negative binomial model on horse profile data and performing residual analysis, we discovered that at least 251 horses under-raced between 2019 and 2021. Additionally, clustering horse movement profiles revealed that a horse’s speed profile is most associated with its injury status; specifically, greater variation in speed is more associated with injury.

Kristopher Wilson is a senior at North Carolina State University (NC State) where he is majoring in Statistics with a minor in Industrial Engineering. His current research interests include social and behavioral science, sports analytics, and high-dimensional data, among others. At NC State, he is an officer of the Sports Analytics Club, and an active participant in the Grappling Club. After graduation, Kris plans to pursue a Ph.D. in Statistics. Outside of the classroom, he is an avid Atlanta sports fan and music enthusiast.

Clustering Racehorse Movement Profiles to Discover Trends in Injured Horses

Between 2009 and 2021, over 7,200 horses died or were euthanized due to racing-related injuries (Fobar, 2023). Using horse profile data and horse tracking data from the NYRA, we identified horses who under-raced between 2019 and 2021 and clustered movement profiles for horses who raced in 2019 New York races to discover if certain profiles were more associated with injured horses. By fitting a negative binomial model on horse profile data and performing residual analysis, we discovered that at least 251 horses under-raced between 2019 and 2021. Additionally, clustering horse movement profiles revealed that a horse’s speed profile is most associated with its injury status; specifically, greater variation in speed is more associated with injury.

I, Fungai Jani, am a senior at the College of Wooster, double majoring in data science and business economics. I enjoy working on data science, specifically sports visualizations and machine learning projects. I am deeply passionate about sports and all the insights that can be uncovered from data to help a franchise reach the next level. After graduation, I plan to pursue a career in data science. I am an avid basketball and soccer fan. I enjoy watching and playing any sport in my free time.

To Bet or Not To Bet?

In recent years, there has been an increase in bets placed on sporting events, especially in soccer. Soccer is the biggest sport in the world, attracting global attention, which puts pressure on betting companies. To build upon the already existing foundation of betting odds, we enriched our predictive models by adding two key components: player evaluation metrics and the teams' recent form based on their performances over the last five matches. Our primary predictive models initially rely on betting odds to forecast match results, constrained by the nature of the dependent variable, an ordered categorical variable representing match outcomes. This leads us to employ Generalized Additive Models (GAMs), Multinomial regressions, and Random Forest algorithms. Subsequently, we augment these base models with four independent variables: streak difference, expected goals (XGs) for the home team, XGs difference, and XGs difference based on defense and offense ratings of home/away teams. The incorporation of team ratings derived from player evaluation metrics forms a central objective of our project. To evaluate the accuracy of our predictive models, we employ the Brier score, akin to mean squared error for predicted probabilities. Our findings reveal that the GAMs model, particularly when incorporating XGs based on defense and offense ratings, consistently yields the lowest Brier scores, indicating superior performance. Conversely, models of the Random Forest type exhibit notably higher Brier scores. The results showcase significant improvements over standard betting probabilities, illuminating the potential of player and team evaluation metrics. Nevertheless, inherent limitations persist, stemming from the unpredictability of soccer matches, including factors such as last-minute injuries, weather, and unforeseen tactical adjustments. Moreover, the data's specificity to a single season underscores the need for caution when applying these findings to future seasons with potentially distinct trends and dynamics.

Maria Tsakalakos is a junior at Emory University, where she is currently enrolled in the Goizueta Business School. She is concentrating her studies in Information Systems and Operational Management, with a minor in Data Science. Maria is highly engaged in sports both on and off campus. She is an active member of the Emory Women's Club soccer team and is also formally associated with the Emory Oxford varsity soccer team. Data science is a newly discovered passion to her and after graduation, Maria aspires financial or data analyst in the sports industry.

To Bet or Not To Bet?

In recent years, there has been an increase in bets placed on sporting events, especially in soccer. Soccer is the biggest sport in the world, attracting global attention, which puts pressure on betting companies. To build upon the already existing foundation of betting odds, we enriched our predictive models by adding two key components: player evaluation metrics and the teams' recent form based on their performances over the last five matches. Our primary predictive models initially rely on betting odds to forecast match results, constrained by the nature of the dependent variable, an ordered categorical variable representing match outcomes. This leads us to employ Generalized Additive Models (GAMs), Multinomial regressions, and Random Forest algorithms. Subsequently, we augment these base models with four independent variables: streak difference, expected goals (XGs) for the home team, XGs difference, and XGs difference based on defense and offense ratings of home/away teams. The incorporation of team ratings derived from player evaluation metrics forms a central objective of our project. To evaluate the accuracy of our predictive models, we employ the Brier score, akin to mean squared error for predicted probabilities. Our findings reveal that the GAMs model, particularly when incorporating XGs based on defense and offense ratings, consistently yields the lowest Brier scores, indicating superior performance. Conversely, models of the Random Forest type exhibit notably higher Brier scores. The results showcase significant improvements over standard betting probabilities, illuminating the potential of player and team evaluation metrics. Nevertheless, inherent limitations persist, stemming from the unpredictability of soccer matches, including factors such as last-minute injuries, weather, and unforeseen tactical adjustments. Moreover, the data's specificity to a single season underscores the need for caution when applying these findings to future seasons with potentially distinct trends and dynamics.

Tseegi is a junior at Smith College double majoring in Mathematical Statistics and Computer Science. She is passionate about developing statistical inferences and machine learning to tackle real-life problems. Her interest in community development, sports analytics, and computer science have culminated in various types of projects and experiences. After graduation, she wishes to work in either sports analytics companies/teams or non-governmental organizations.

To Bet or Not To Bet?

In recent years, there has been an increase in bets placed on sporting events, especially in soccer. Soccer is the biggest sport in the world, attracting global attention, which puts pressure on betting companies. To build upon the already existing foundation of betting odds, we enriched our predictive models by adding two key components: player evaluation metrics and the teams' recent form based on their performances over the last five matches. Our primary predictive models initially rely on betting odds to forecast match results, constrained by the nature of the dependent variable, an ordered categorical variable representing match outcomes. This leads us to employ Generalized Additive Models (GAMs), Multinomial regressions, and Random Forest algorithms. Subsequently, we augment these base models with four independent variables: streak difference, expected goals (XGs) for the home team, XGs difference, and XGs difference based on defense and offense ratings of home/away teams. The incorporation of team ratings derived from player evaluation metrics forms a central objective of our project. To evaluate the accuracy of our predictive models, we employ the Brier score, akin to mean squared error for predicted probabilities. Our findings reveal that the GAMs model, particularly when incorporating XGs based on defense and offense ratings, consistently yields the lowest Brier scores, indicating superior performance. Conversely, models of the Random Forest type exhibit notably higher Brier scores. The results showcase significant improvements over standard betting probabilities, illuminating the potential of player and team evaluation metrics. Nevertheless, inherent limitations persist, stemming from the unpredictability of soccer matches, including factors such as last-minute injuries, weather, and unforeseen tactical adjustments. Moreover, the data's specificity to a single season underscores the need for caution when applying these findings to future seasons with potentially distinct trends and dynamics.

Mathew Chandy is a senior at the University of Connecticut, where he is double majoring in Statistics and Statistical Data Science, and minoring in Computer Science and Economics. His research interests include methods research, econometrics, and of course, sports analytics. He provides web support for the Connecticut Data Science Lab, which organizes the UConn Sports Analytics Symposium. Additionally, he is a co-founder of the UConn Joint Statistical Club. After graduating, he hopes to pursue a PhD in Statistics. He is very passionate about sports, and he especially enjoys watching and playing basketball.

Deal or No Deal? An NBA Recommender System for Team Composition and Salary Optimization

In the dynamic ecosystem of professional basketball, the ability to discern the true value of a player often lies at the intersection of performance on the court and the economic dynamics of the sport. By venturing beyond traditional evaluation metrics, we seek to shed light on the nuances of the game and offer insights that could redefine team strategies. Our investigation delves into whether player compensation aligns with their performance, revealing potential hidden gems. Recognizing that basketball is a team sport, we also aim to identify optimal player combinations for a hypothetical team. This challenge has been tackled by previous methods, but we seek to combine the two team goals of positive surplus value and complimentary playstyles into a single Shiny App. Through this project, we hope to offer a demonstration of a salary surplus model for an individual and an expected points model for a five-man lineup. We believe this demonstration can help teams make decisions about team-building and salary allocation to maximize wins.

Leo Cheng is a junior at Carnegie Mellon University, majoring in Economics and Statistics with a minor in Computer Science. He is passionate about the insight from data and his academic interest is to develop statistics, data science methodology, and computational simulations in data analytics. Beyond academia, he is a big fan of basketball and soccer. This enthusiasm in sports motivates his research in team composition & salary optimization at NBA level.

Deal or No Deal? An NBA Recommender System for Team Composition and Salary Optimization

In the dynamic ecosystem of professional basketball, the ability to discern the true value of a player often lies at the intersection of performance on the court and the economic dynamics of the sport. By venturing beyond traditional evaluation metrics, we seek to shed light on the nuances of the game and offer insights that could redefine team strategies. Our investigation delves into whether player compensation aligns with their performance, revealing potential hidden gems. Recognizing that basketball is a team sport, we also aim to identify optimal player combinations for a hypothetical team. This challenge has been tackled by previous methods, but we seek to combine the two team goals of positive surplus value and complimentary playstyles into a single Shiny App. Through this project, we hope to offer a demonstration of a salary surplus model for an individual and an expected points model for a five-man lineup. We believe this demonstration can help teams make decisions about team-building and salary allocation to maximize wins.

Lauren Okamoto is a junior at UC Berkeley studying Data Science and Cognitive Science. She is passionate about deriving meaning from data and using these insights to shape decision making. Her research interests include computational modeling of the mind, sports analytics, and the usage of generative AI in various applications. CMSAC along with her participation in Berkeley’s Sports Analytics Club has furthered her passion for sports analytics. Currently, she is interning for the United Soccer League and hopes to continue work in the sports industry following graduation.

Deal or No Deal? An NBA Recommender System for Team Composition and Salary Optimization

In the dynamic ecosystem of professional basketball, the ability to discern the true value of a player often lies at the intersection of performance on the court and the economic dynamics of the sport. By venturing beyond traditional evaluation metrics, we seek to shed light on the nuances of the game and offer insights that could redefine team strategies. Our investigation delves into whether player compensation aligns with their performance, revealing potential hidden gems. Recognizing that basketball is a team sport, we also aim to identify optimal player combinations for a hypothetical team. This challenge has been tackled by previous methods, but we seek to combine the two team goals of positive surplus value and complimentary playstyles into a single Shiny App. Through this project, we hope to offer a demonstration of a salary surplus model for an individual and an expected points model for a five-man lineup. We believe this demonstration can help teams make decisions about team-building and salary allocation to maximize wins.

Ethan Park is a senior at the University of Southern California studying Computer Science and Data Science. Ethan is highly passionate about the application of data science in sports and works as a Data Analyst for the USC Men's Baseball Team and the USC Sports Business Association, Analytics Division and various personal projects. After graduation, he plans on pursuing a career in sports analytics for a front office.

The Art of Sequencing: Utilizing Inter-Pitch Dynamics to Enhance Pitch Evaluation in Major League Baseball

Pitch evaluation metrics such as Stuff+, Location+, and Pitching+, have added dimension to the evaluation of pitch types and are indicative of a growing focus on individual pitch characteristics. While the discrete qualities of a pitch (such as movement, velocity, and spin rate) are undoubtedly useful in evaluating and explaining elite pitchers, the effects that sequencing, tunneling, and in-game pitching strategy have on outcomes cannot be understated. Therefore, integrating inter-pitch dynamics with discrete pitch characteristics might more accurately model a pitcher’s effectiveness. Our motivation behind this project is to evaluate current pitch metrics and consider how inter-pitch dynamics can enhance existing models.

I am from Pasadena, California and go to Elon University in North Carolina where I am majoring in Statistics. I have loved baseball my whole life and want to use numbers to help people better understand my favorite sport.

The Art of Sequencing: Utilizing Inter-Pitch Dynamics to Enhance Pitch Evaluation in Major League Baseball

Pitch evaluation metrics such as Stuff+, Location+, and Pitching+, have added dimension to the evaluation of pitch types and are indicative of a growing focus on individual pitch characteristics. While the discrete qualities of a pitch (such as movement, velocity, and spin rate) are undoubtedly useful in evaluating and explaining elite pitchers, the effects that sequencing, tunneling, and in-game pitching strategy have on outcomes cannot be understated. Therefore, integrating inter-pitch dynamics with discrete pitch characteristics might more accurately model a pitcher’s effectiveness. Our motivation behind this project is to evaluate current pitch metrics and consider how inter-pitch dynamics can enhance existing models.

Priyanka is a junior at Harvard University studying Applied Mathematics on the Computer Science track with a minor in Mind, Brain, and Behavior. She is an infielder on the Harvard Softball team and has always been curious about the intersection of athletics and analytics as an athlete. Priyanka is involved in Harvard Undergraduate Women in Business and works as a calculus course assistant. She is a diehard Yankees fan and enjoys watching college softball, baseball and football.

The Art of Sequencing: Utilizing Inter-Pitch Dynamics to Enhance Pitch Evaluation in Major League Baseball

Pitch evaluation metrics such as Stuff+, Location+, and Pitching+, have added dimension to the evaluation of pitch types and are indicative of a growing focus on individual pitch characteristics. While the discrete qualities of a pitch (such as movement, velocity, and spin rate) are undoubtedly useful in evaluating and explaining elite pitchers, the effects that sequencing, tunneling, and in-game pitching strategy have on outcomes cannot be understated. Therefore, integrating inter-pitch dynamics with discrete pitch characteristics might more accurately model a pitcher’s effectiveness. Our motivation behind this project is to evaluate current pitch metrics and consider how inter-pitch dynamics can enhance existing models.

Quinn Robnett is a senior at Syracuse University majoring in Sport Analytics and minoring in Information management and Technology. At Syracuse, he is the Director of Media for the Hockey Analytics Club and previously worked as an Analytics Intern for the DI Field Hockey team. After completing his undergraduate degree, he plans to pursue a Master’s in Applied Data Science at Syracuse. He is motivated by a deep love for hockey and data, and his long-term career goal is to work in the analytics department of an NHL team.

Examining the Decline of Save Percentage in the NHL

Goaltending is of vital importance in the National Hockey League (NHL). A goaltender's job is simple: to stop the other team from scoring. A recent trend that has been observed by many people throughout the hockey community has been that shots are being saved at declining rates over the most recent 7 seasons. It was hard to believe that the collective goaltender talent declined across the league, so we began to search for the true causes. Through the analysis on the league's play-by-play data, which is available dating back to the 2011-12 season, we began to find multiple reasons for the observed change. We were successfully able to attribute the change to a couple reasons. We determined that the best goaltenders on teams were playing less minutes in more recent seasons in a trend referred to as "load management". Secondly, we found that shots were being taken more often from closer and more straight on angles. And lastly, by creating expected goal logistic regression models, we found that situational descriptions of shots were influencing shooters' success rates differently across the years.

Luke Welsh is a senior at the University of Wisconsin - Madison, double majoring in Data Science and Economics. At Wisconsin, he is the founder of the sports analytics club, where he is able to spread his passions for the topic to his classmates. In his free time, he likes to keep busy hopping in to play any sport he can, as well as participating in his school's track, swing dancing, and log rolling clubs. After graduation, he is excited to begin his career working in data science, with aspirations to work for a sports team.

Examining the Decline of Save Percentage in the NHL

Goaltending is of vital importance in the National Hockey League (NHL). A goaltender's job is simple: to stop the other team from scoring. A recent trend that has been observed by many people throughout the hockey community has been that shots are being saved at declining rates over the most recent 7 seasons. It was hard to believe that the collective goaltender talent declined across the league, so we began to search for the true causes. Through the analysis on the league's play-by-play data, which is available dating back to the 2011-12 season, we began to find multiple reasons for the observed change. We were successfully able to attribute the change to a couple reasons. We determined that the best goaltenders on teams were playing less minutes in more recent seasons in a trend referred to as "load management". Secondly, we found that shots were being taken more often from closer and more straight on angles. And lastly, by creating expected goal logistic regression models, we found that situational descriptions of shots were influencing shooters' success rates differently across the years.