CARNEGIE MELLON SPORTS ANALYTICS CONFERENCE

Brooke Coneeny is a junior at Swarthmore College studying Mathematics (with an emphasis in Statistics) and Computer Science. She is a member of Swarthmore's varsity softball team and hopes to pursue a career in sports analytics, specifically softball or baseball.

Maximizing wOBA with Launch Angle and Exit Velocity

In the game of baseball, each batter’s goal is to be a high production player where they are getting on base and driving in runs. However, not all players have the same swing or physical power and therefore have different production capabilities. Players that can consistently hit with a high exit velocity can lift the ball and put it over the fence, but players that naturally have lower exit velocities would routinely pop out if they used the same launch angle as these higher exit velocity hitters. With this intuition, we created a model that recommends a swing plane (attack angle) based on a batter’s power profile. We began by taking the player’s balls that were hit into play and created a linear model that predicted how the balls would leave the bat should their attack angle change. We then took these new launch angles and used a generalized additive model to predict the quality - measured by wOBA - of this new hit. Unfortunately, this model did not take into account that when a player changes his attack angle, the pitches he receives and makes contact with will change. We attacked this problem by taking all balls a player swung at in a year and creating GAMs to predict which balls they would make contact with, and then predict which of these balls they would hit fair. These new pitches comprised the set of pitches we then ran through the original model and allowed us to advise an optimal attack angle to the specific player.

Grace Fain is a senior at The University of Oklahoma, where she is studying Management Information Systems and Management with a concentration in Sports Management. After graduation, she plans to pursue a career in data analytics with interests in sports and business intelligence.

Tired of Misattribution, Modeling Player Fatigue in the NBA

The prevailing belief propagated by NBA league observers is that the workload of the NBA season dramatically influences a player's performance. We offer an analysis of cross game player fatigue that calls into question the empirical validity of these claims.

Erin is a junior at Macalester College in Saint Paul, MN and is majoring in Statistics and Economics. She is working toward a career as an analyst and is especially passionate about analytics in the field of baseball and social good. She will be joining Macalester’s varsity track and cross country team this winter and also enjoys hiking and watching/playing sports.

Maximizing wOBA with Launch Angle and Exit Velocity

In the game of baseball, each batter’s goal is to be a high production player where they are getting on base and driving in runs. However, not all players have the same swing or physical power and therefore have different production capabilities. Players that can consistently hit with a high exit velocity can lift the ball and put it over the fence, but players that naturally have lower exit velocities would routinely pop out if they used the same launch angle as these higher exit velocity hitters. With this intuition, we created a model that recommends a swing plane (attack angle) based on a batter’s power profile. We began by taking the player’s balls that were hit into play and created a linear model that predicted how the balls would leave the bat should their attack angle change. We then took these new launch angles and used a generalized additive model to predict the quality - measured by wOBA - of this new hit. Unfortunately, this model did not take into account that when a player changes his attack angle, the pitches he receives and makes contact with will change. We attacked this problem by taking all balls a player swung at in a year and creating GAMs to predict which balls they would make contact with, and then predict which of these balls they would hit fair. These new pitches comprised the set of pitches we then ran through the original model and allowed us to advise an optimal attack angle to the specific player.

My name is Nicholas Ho and I'm a 3rd year at Arizona State University studying Computer Science and Mathematics. My current interests are in modeling biophysical dynamical systems using deep learning techniques. I am a researcher in the Structural Systems Biology lab at the Biodesign Institute.

An Interpretable Method of Learning Stochastic Game Dynamics

In soccer, modeling expected goals in a match is difficult without the use of player tracking data. Many models that attempt to make score predictions depend almost exclusively on the outcome of previous matches, and hence tend to do a poor job of capturing high score differentials (as in when a team wins by a substantial margin over another team). Relying on just the tracking data of the ball alone, we wanted to encapsulate the complex movement and forces acting on the ball into a much simpler object. This object is the potential function, which is simply an equation used to model underlying forces (e.g. gravitational potential functions). In a 2007 study, David R. Brilllinger wrote a paper on how to learn a potential function given a trajectory. The crux of this paper is that potential functions can be approximated using basis functions. In our case, these basis functions are a set of gravitational points, whose coefficients are based on the offensive and defensive movements of our teams, creating a potential function landscape unique to each team pairing. Through this “potential function landscape,” we are able to simulate games over time and create an averaged score prediction when two teams are pitted against each other. What we found is that predictions formed using this methodology capture high score differentials more reliably, and reduce both the MSE and residual variance compared to a Poisson Regression Model. We believe that this new method of modeling expected goals could also be used to determine player impacts in a game and provide a real-time game evaluation in the future.

Tej is currently a Junior at the University of Michigan majoring in Information Analysis and minoring in Applied Statistics. He enjoys working on sports analytics related projects, listening to true crime podcasts and spending time with family and friends.

Run vs. Pass Play Prediction: Incorporating NFL Tracking Data

In football, it is advantageous for the defense to be able to identify and gain insight into whether the offense is running a pass play or a run play. Models have been created in the past using situational factors of the play to predict play type. We sought to recreate these previous models but also incorporate NFL tracking data to improve these models. Incorporating tracking data from weeks 1-6 of the 2017 NFL season, we used various types of machine learning models to predict whether a play would be run or pass based on a multitude of situational factors and generated positional factors. Predictability of play type was compared across all 32 NFL teams. In an extension of our project, we created a model that updates play type predictions every tenth of a second from before the snap until 2.5 seconds after the snap.

I’m a third-year undergraduate at the University of Wyoming getting degrees in Computer Science and Statistics. I'm motivated by learning how data can offer clever insights into all kinds of challenging problems. Exploring this interest, I've worked on a range of projects including automated selection, modeling stock market volatility, and sports analytics. Beyond academia, I enjoy backpacking and trail running taking advantage of the nature in Wyoming and Colorado.

Tired of Misattribution, Modeling Player Fatigue in the NBA

The prevailing belief propagated by NBA league observers is that the workload of the NBA season dramatically influences a player's performance. We offer an analysis of cross game player fatigue that calls into question the empirical validity of these claims.

Sarah is a senior at Washington University in St. Louis majoring in Economics and Computer Science. Originally from Houston, Sarah grew up watching the Houston Astros which is where her love of baseball began. Her hobbies include equestrian showjumping and Krav Maga martial arts. After graduation she hopes to work as an analyst in baseball or business.

Maximizing wOBA with Launch Angle and Exit Velocity

In the game of baseball, each batter’s goal is to be a high production player where they are getting on base and driving in runs. However, not all players have the same swing or physical power and therefore have different production capabilities. Players that can consistently hit with a high exit velocity can lift the ball and put it over the fence, but players that naturally have lower exit velocities would routinely pop out if they used the same launch angle as these higher exit velocity hitters. With this intuition, we created a model that recommends a swing plane (attack angle) based on a batter’s power profile. We began by taking the player’s balls that were hit into play and created a linear model that predicted how the balls would leave the bat should their attack angle change. We then took these new launch angles and used a generalized additive model to predict the quality - measured by wOBA - of this new hit. Unfortunately, this model did not take into account that when a player changes his attack angle, the pitches he receives and makes contact with will change. We attacked this problem by taking all balls a player swung at in a year and creating GAMs to predict which balls they would make contact with, and then predict which of these balls they would hit fair. These new pitches comprised the set of pitches we then ran through the original model and allowed us to advise an optimal attack angle to the specific player.

My name is Sifan Tao and I'm a 3rd-year student at the University of Virginia studying statistics and mathematics. I'm currently interested in causal inference and statistical machine learning. During my free time, I love playing tennis and watching sports competitions.

An Interpretable Method of Learning Stochastic Game Dynamics

In soccer, modeling expected goals in a match is difficult without the use of player tracking data. Many models that attempt to make score predictions depend almost exclusively on the outcome of previous matches, and hence tend to do a poor job of capturing high score differentials (as in when a team wins by a substantial margin over another team). Relying on just the tracking data of the ball alone, we wanted to encapsulate the complex movement and forces acting on the ball into a much simpler object. This object is the potential function, which is simply an equation used to model underlying forces (e.g. gravitational potential functions). In a 2007 study, David R. Brilllinger wrote a paper on how to learn a potential function given a trajectory. The crux of this paper is that potential functions can be approximated using basis functions. In our case, these basis functions are a set of gravitational points, whose coefficients are based on the offensive and defensive movements of our teams, creating a potential function landscape unique to each team pairing. Through this “potential function landscape,” we are able to simulate games over time and create an averaged score prediction when two teams are pitted against each other. What we found is that predictions formed using this methodology capture high score differentials more reliably, and reduce both the MSE and residual variance compared to a Poisson Regression Model. We believe that this new method of modeling expected goals could also be used to determine player impacts in a game and provide a real-time game evaluation in the future.

Nicole Tucker is a junior at the University of Evansville where she is majoring in Mathematics and Data Science/Statistics with a minor in Communication. She is an avid football fan, and her main research interests lie in football analytics. Her passion lies in modern fan engagement, which has prompted her to find ways to use her statistics and communication background to connect fans with current sports analytics concepts. At the University of Evansville, Nicole is the president of the Statistics/Data Science Club and the Vice President of Academic Development for her chapter of Alpha Omicron Pi. Nicole currently works for Purple Aces Sports Properties (a division of Learfield) as an Associate Coordinator where she plays a role in University of Evansville game day operations and fulfillment of promotions.

Run vs. Pass Play Prediction: Incorporating NFL Tracking Data

In football, it is advantageous for the defense to be able to identify and gain insight into whether the offense is running a pass play or a run play. Models have been created in the past using situational factors of the play to predict play type. We sought to recreate these previous models but also incorporate NFL tracking data to improve these models. Incorporating tracking data from weeks 1-6 of the 2017 NFL season, we used various types of machine learning models to predict whether a play would be run or pass based on a multitude of situational factors and generated positional factors. Predictability of play type was compared across all 32 NFL teams. In an extension of our project, we created a model that updates play type predictions every tenth of a second from before the snap until 2.5 seconds after the snap.

My name is Adhvaith Vijay and I am a 4th-year student at UCLA studying Statistics. I currently work for the Digital Humanities department as an undergraduate researcher and aspire to pursue a career in machine learning and customer analytics. Outside of academics, I play Badminton and Table Tennis.

An Interpretable Method of Learning Stochastic Game Dynamics

In soccer, modeling expected goals in a match is difficult without the use of player tracking data. Many models that attempt to make score predictions depend almost exclusively on the outcome of previous matches, and hence tend to do a poor job of capturing high score differentials (as in when a team wins by a substantial margin over another team). Relying on just the tracking data of the ball alone, we wanted to encapsulate the complex movement and forces acting on the ball into a much simpler object. This object is the potential function, which is simply an equation used to model underlying forces (e.g. gravitational potential functions). In a 2007 study, David R. Brilllinger wrote a paper on how to learn a potential function given a trajectory. The crux of this paper is that potential functions can be approximated using basis functions. In our case, these basis functions are a set of gravitational points, whose coefficients are based on the offensive and defensive movements of our teams, creating a potential function landscape unique to each team pairing. Through this “potential function landscape,” we are able to simulate games over time and create an averaged score prediction when two teams are pitted against each other. What we found is that predictions formed using this methodology capture high score differentials more reliably, and reduce both the MSE and residual variance compared to a Poisson Regression Model. We believe that this new method of modeling expected goals could also be used to determine player impacts in a game and provide a real-time game evaluation in the future.

Hey I'm Matt Yep, I'm a fourth year at UC Berkeley double majoring in Economics and Data Science, and I think that Jordan Poole is going to win Most Improved Player. I like to fish in the SF bay, read books about behavioural economics, and hoop on the practice squad with the Cal Women's basketball team.

Tired of Misattribution, Modeling Player Fatigue in the NBA

The prevailing belief propagated by NBA league observers is that the workload of the NBA season dramatically influences a player's performance. We offer an analysis of cross game player fatigue that calls into question the empirical validity of these claims.