CARNEGIE MELLON SPORTS ANALYTICS CONFERENCE

Lotte Bransen is a Lead Data Scientist at SciSports, where she leads the Data Analytics team that develops analytical tools to derive actionable insights from soccer data. An avid soccer player herself, Lotte primarily works on developing machine learning models to measure the impact of soccer players’ in-game actions and decisions on the courses and outcomes of matches.

Player Chemistry: Striving for a Perfectly Balanced Soccer Team

Soccer scouts typically ignore the team balance and team chemistry when evaluating potential signings for their teams. Instead, they focus on the individual qualities of the players in isolation. To overcome this limitation of their recruitment process, this talk takes a first step towards objectively providing insight into the question: How well does a team of soccer players gel? In this talk I will introduce two chemistry metrics that measure the offensive and defensive chemistry for a pair of players, respectively. The offensive chemistry metric measures the pair's joint performance in terms of scoring goals, whereas the defensive chemistry metric measures their joint performance in preventing their opponents from scoring goals. Finally, I will explain how these chemistry metrics can be used to build a perfectly balanced soccer team.

Nathan Sandholtz has recently begun a postdoctoral fellowship at the University of Toronto, where he is working on inverse optimization applications in sport with Timothy Chan. In August, he completed his PhD in statistics from Simon Fraser University under the supervision of Derek Bingham and Luke Bornn. Nate is originally from Provo, Utah where he will be returning after the postdoc to join the statistics faculty at Brigham Young University.

Measuring Spatial Allocative Efficiency in Basketball

Every shot in basketball has an opportunity cost; one player’s shot eliminates all potential opportunities from their teammates for that play. For this reason, player-shot efficiency should ultimately be considered relative to the lineup. This aspect of efficiency—the optimal way to allocate shots within a lineup—is the focus of our paper. Allocative efficiency should be considered in a spatial context since the distribution of shot attempts within a lineup is highly dependent on court location. We propose a new metric for spatial allocative efficiency by comparing a player’s field goal percentage (FG%) to their field goal attempt (FGA) rate in context of both their four teammates on the court and the spatial distribution of their shots. Leveraging publicly available data provided by the National Basketball Association (NBA), we estimate player FG% at every location in the offensive half court using a Bayesian hierarchical model. Then, by ordering a lineup’s estimated FG%s and pairing these rankings with the lineup’s empirical FGA rate rankings, we detect areas where the lineup exhibits inefficient shot allocation. Lastly, we analyze the impact that sub-optimal shot allocation has on a team’s overall offensive potential, demonstrating that inefficient shot allocation correlates with reduced scoring.

Paul J. van Staden is a senior lecturer in the Department of Statistics at the University of Pretoria, South Africa. He earned his PhD in Mathematical Statistics from the University of Pretoria in 2014. His research focuses on statistical distribution theory, in particular the construction of generalized families of probability distributions in the quantile statistical universe. With respect to sports analytics, his research interests include team and player performance measures in cricket, modeling of score distributions in rugby and cricket, and rating and ranking systems. He has furthermore collaborated in various research projects ranging from tongue protrusion and bite mark analysis in forensic pathology to oral and dental diseases in cheetahs.

Points, rating & ranking systems in professional tennis

The Association of Tennis Professionals (ATP) for men’s tennis and the Women's Tennis Association (WTA) use rolling 52-week points-accumulation rating systems. These systems rank tennis players according to points awarded for their performances in tournaments without taking strength of opponents or score differences into account. Because of the coronavirus pandemic, the 2020 tennis calendar had to be revised with tennis tournaments cancelled or postponed. Consequently the ATP and the WTA froze their rankings on 16 March 2020 and have subsequently introduced revised rating systems, which are still based on accumulation of points. This paper develops a points-exchange rating system for tennis with the rating points of the competing players adjusted by equal, opposite amounts after each tennis match based on the margin of victory in that match. Since margin of victory cannot be directly calculated from traditional tennis scores, a points system for tennis matches is also presented in which tennis scores are converted into performance points. The proposed points and rating systems are illustrated with the 2019 Wimbledon Championships.

Chris Watkins is an analyst and contributing writer to sites such as HockeyGraphs, RawCharge and Blue Seat Blogs. In addition he is the occasional host of the Let's Do That Hockey podcast, which explores the intersection of analytics, on-ice strategy and race and society within the sport of hockey.

Racial Bias in Drafting and Development: The NHL’s Black Quarterback Problem

Coming soon!

Meghan Hall is a higher education data professional by day and an amateur hockey analyst by night. She contributes to Hockey-Graphs and also creates tutorials and resources, particularly for beginners, for R and Tableau because she loves helping other people learn these tools to make their work easier. She has presented her hockey analysis work, on league-wide trends in goalie pulling and features of the penalty kill, at the Seattle Hockey Analytics Conference, the Rochester Institute of Technology Sports Analytics Conference, the Ottawa Hockey Analytics Conference, and the Columbus Blue Jackets Hockey Analytics Conference. She also presented virtually at Hockey (Analytics) Night in Canada on how and why to learn a programming language.

Leveling Up With The Tidyverse (And Hockey Data)

In this workshop targeted to beginner and intermediate R users, we'll go through a sample hockey analysis question and discuss techniques from the tidyverse and beyond (including user-defined functions, RMarkdown, custom ggplot2 themes, and data manipulation with tidyr and dplyr) to learn how to level up your R programming and make your analysis more efficient and more reproducible.

Tom Mock works in Customer Success at RStudio, helping empower their High Technology and Sports customers to fully leverage open-source Data Science with R and Python via RStudio's Professional software. Tom holds a MS in Exercise Physiology from FAU, and a PhD in Neurobiology from UNTHSC. He founded #TidyTuesday to help newcomers and seasoned vets improve their Tidyverse skills. In his spare time, he gives back to the open source community by writing guided tutorials (mostly with nflfastR data) at themockup.blog.

Introduction to machine learning with the tidyverse and nflfastR data

This workshop provides a gentle introduction to supervised machine learning: concepts, methods, and R code. Participants will learn how to apply a few common methods on open-source NFL data via nflfastR. Along the way, I'll introduce several core tidymodels packages, which provide a grammar for modeling that makes it easy to do the right thing, and harder to accidentally do the wrong thing.

Alexandra Mandrycky is the Director of Hockey Strategy and Research for the Seattle Kraken. She and her group facilitate the use of data and technology throughout the Hockey Operations department. Alexandra previously spent four seasons as Hockey Operations Analyst with the Minnesota Wild. Alexandra graduated from Georgia Tech in 2013 with a B.S. in Industrial Engineering.

Alison Lukan is a data-driven storyteller covering the Columbus Blue Jackets and the NHL. She is a contributor at The Athletic and Hockey Graphs, has written at 1st Ohio Battery, FoxSports Ohio, and for the Blue Jackets' official team site.

Sara Civian is a Staff Writer for The Athletic, covering the Hurricanes. Before The Athletic she covered the Boston Bruins for WEEI 93.7-FM. She spent the five years prior at Penn State, where she served as managing editor for Onward State.

Along with being 1/3 of the Too Many Men Podcast, Shayna is a contributor for The Athletic and HockeyGraphs, an editor for Blueshirt Banter, and co-creator of Behind the Benches.

Doug Fearing is the Founder and President of Zelus Analytics, an Austin-based sports analytics startup creating the world’s best sports intelligence platform. Prior to founding Zelus, he spent four seasons building the R&D department for the L.A. Dodgers, reaching the World Series in 2017 and 2018. Doug has held tenure-track faculty positions at Harvard Business School and UT Austin’s McCombs School of Business. He received his Ph.D. in Operations Research from MIT and his B.S. in Computer Science at CMU.

Robert Nguyen is a PhD student in statistics at the University of New South Wales. He loves kicking the footy with his pet dog (Charlie), watching his AFL team win (Eagles) and encouraging reproducible sports analytics in the Australian sports space.

My name is Bria Cratty, and I am a senior at the University of Miami studying individualized general business with a minor in psychology. I am also in the Dean’s Scholar Program, an accelerated program that allows me to complete my undergraduate degree in three years and a Master of Science in Business Analytics my following fourth year. This summer I worked with Jack de la Parra on our final project for the Carnegie Mellon Sports Analytics Camp. After graduation, I plan to start a career in strategy analytics and/or consulting.

Redefining the Penalty Box: Does the Punishment Fit the Crime?

In soccer, the practice of awarding a penalty kick to a player fouled near the goal has been a long-standing method attempting to keep the game as fair as possible. Recently, watching games play out and stats accumulate has indicated that penalty kicks are converted to goals at a rate much higher than shots are scored during regular play. Previous studies have shown many inconsistencies in how and when fouls are called and the potentially monumental impact a penalty kick can have on the outcome of a match. This identified a need for the redefinition of the current penalty kick to create a scoring opportunity more comparable to the opportunity players would have had in the absence of the foul. Using StatsBomb open data at the event level for the 2018 Men’s World Cup, we used a generalized additive model approach to predict expected goals for different distances and angles to the goal. The goal was to change the location of the penalty kick, so the chances of scoring were closer to the chances during regular play. We proposed replacing the current penalty kick spot with a penalty kick arc. The format of the arc would position the kick-taker based on where the foul occurred within the box as well as extending the distance from which the kick is taken. We believe this new method of penalty kicks will allow for a fairer experience while keeping the traditional aspects of the game fairly constant.

My name is Raj Dasani and I am a 3rd year student at UC Berkeley studying Business and Data Science. I currently work for the women's basketball team at Cal as a data and film analyst and am aspiring to work in sports analytics whether for a professional basketball team or a sports-themed organization. Outside of academics and career, I play the trumpet as part of the University Marching Band!

Quantifying Passing: Using NBA Tracking Data to Create an Expected Assist Model

In the game of basketball, the success of a possession hinges not just on the quality of a shot but also on the quality of the preceding passes. Although much research has been done modeling the expected outcome of a shot, much less is devoted to the value added from a pass. This paper attempts to bridge that gap by offering an Expected Assists model. Using publicly available tracking data from the 2015-16 NBA season, we employ a combination of a rules-based approach and a Generalized Additive Model to identify when a pass occurs. Due to heavy computational limitations, our research concentrates on all Los Angeles Clippers games from December 2015. We then develop a shot model that projects the chance of a player making a shot given his location and the defense around him. We incorporate this into a passing model that takes into consideration the timing and location in order to create an expected points added metric for each pass. This metric allows us to identify passers who are creating good shooting opportunities for their teammates without penalizing them when their teammates are unable to convert. Isolating an individual player’s play-making from team success is a key component of equitable player evaluation.

Fiona is a junior at Kenyon College where she is studying Economics with a minor in Statistics. She is an avid Boston sports fan and is a member of the Kenyon Varsity Women’s Soccer team.

High Anticipation: Exploring Trends Between Public Perception and Player Value

Being a professional basketball player is one of the many ways one can be at the forefront of the public’s eye. The way they play, their actions off the court, and their social media presence are all factors that can affect how a player is perceived by the public. This research aims to determine the trends of a relationship between public perception and a player’s performance in a game. The research focuses on the 2018 NBA first round draft class. The data consists of a series of data scrapes from Reddit, Wikipedia, and Google to gauge public perception and NBA box score data for player performance data. Hierarchical clustering was the best type of analysis to accurately display our public perception data which consisted of sentiment and popularity metrics. Two clusters were created: players who were not very popular with low positive sentiments and players who are more popular with higher positive sentiment . An XGBoost classification model was trained with an 0.887 AUC score to predict which players would fall into each cluster based on basketball performance variables like number of 3-point goals and assists. The results showed that points scored by a player and minutes on the court were given the most importance by the model. These results indicate a relationship between a player’s performance and public perception but confirming this would require an expansion of the sample and complex modeling.

Meg is a senior at Kenyon College majoring in Psychology and Anthropology with a minor in Statistics and a concentration in Scientific Computing. She is an avid hockey fan (go Blue Jackets!), and hopes to pursue a career in statistics.

Draft Decisions in Uncertain Times: Valuing and Simulating NHL Draft Picks

The NHL draft lottery is complicated even at the best of times, but it is especially so this year, when the Pittsburgh Penguins must decide what to do with a conditional first-round pick traded to the Minnesota Wild. Using publically available data on player statistics and team performance, we built two models. The first model is a NHL draft slot value curve, which predicts a player’s contribution to his team based on his draft position, while the second model predicts a team’s performance in the upcoming season based on the previous season’s statistics. We used this second model to simulate league-wide season rankings and then first-round draft pick orders to generate a distribution of possible outcomes. Combining these two models, we created a one-number index to compare the value of this year’s pick with the probable value of next year’s pick. Based on our modeling and simulation, a solution to this dilemma is proposed to maximize the Penguins' chances of future success.

James is a senior sport analytics and neuroscience major at Syracuse, where he will be pursuing a master’s in applied data science next fall.

Quantifying Passing: Using NBA Tracking Data to Create an Expected Assist Model

In the game of basketball, the success of a possession hinges not just on the quality of a shot but also on the quality of the preceding passes. Although much research has been done modeling the expected outcome of a shot, much less is devoted to the value added from a pass. This paper attempts to bridge that gap by offering an Expected Assists model. Using publicly available tracking data from the 2015-16 NBA season, we employ a combination of a rules-based approach and a Generalized Additive Model to identify when a pass occurs. Due to heavy computational limitations, our research concentrates on all Los Angeles Clippers games from December 2015. We then develop a shot model that projects the chance of a player making a shot given his location and the defense around him. We incorporate this into a passing model that takes into consideration the timing and location in order to create an expected points added metric for each pass. This metric allows us to identify passers who are creating good shooting opportunities for their teammates without penalizing them when their teammates are unable to convert. Isolating an individual player’s play-making from team success is a key component of equitable player evaluation.

Alex Lagarde is a junior year student at Elon University, where he is majoring in statistics with a concentration in data analytics. After graduation, he plans to pursue a Master's Degree and a career in data analytics, with interests in sports and business intelligence.

Quantifying Passing: Using NBA Tracking Data to Create an Expected Assist Model

In the game of basketball, the success of a possession hinges not just on the quality of a shot but also on the quality of the preceding passes. Although much research has been done modeling the expected outcome of a shot, much less is devoted to the value added from a pass. This paper attempts to bridge that gap by offering an Expected Assists model. Using publicly available tracking data from the 2015-16 NBA season, we employ a combination of a rules-based approach and a Generalized Additive Model to identify when a pass occurs. Due to heavy computational limitations, our research concentrates on all Los Angeles Clippers games from December 2015. We then develop a shot model that projects the chance of a player making a shot given his location and the defense around him. We incorporate this into a passing model that takes into consideration the timing and location in order to create an expected points added metric for each pass. This metric allows us to identify passers who are creating good shooting opportunities for their teammates without penalizing them when their teammates are unable to convert. Isolating an individual player’s play-making from team success is a key component of equitable player evaluation.

I am a junior at the University of Scranton, where I am a double major in Applied Mathematics and Philosophy with a concentration in Data Science. My interests include applications of statistics to the humanities.

A Puck Above the Rest: Exploring the Effects of New Data on 2020 NHL Draft Decisions

Our work aimed to understand how the addition of new season data to existing player records would impact a player's place in the National Hockey League draft. Previous authors used generalized additive models in a similar context to measure time on ice as an indicator of player success, and many used the relative age of the player as one of their predictor variables. Our custom model incorporates both a logistic model and a generalized additive model to compare player's draft year statistics with their statistics from one year later. The indicator of a player’s performance was the percentage of a team’s points per game that he scored. The model outputs the product of the player's probability of being successfully drafted with the player's expected NHL contribution given that he is drafted. Future work may benefit from considering different subsets of players and assessing league strength to generalize the results.

My name is Jack de la Parra and I am a junior at Bucknell University studying Mathematical Economics with a concentration in Statistics. I worked alongside Bria Cratty this summer at the Carnegie Mellon Sports Analytics Summer Research Program. I have always had an interest in mathematics and statistics and look to begin a career in sports analytics, data or actuarial science.

Redefining the Penalty Box: Does the Punishment Fit the Crime?

In soccer, the practice of awarding a penalty kick to a player fouled near the goal has been a long-standing method attempting to keep the game as fair as possible. Recently, watching games play out and stats accumulate has indicated that penalty kicks are converted to goals at a rate much higher than shots are scored during regular play. Previous studies have shown many inconsistencies in how and when fouls are called and the potentially monumental impact a penalty kick can have on the outcome of a match. This identified a need for the redefinition of the current penalty kick to create a scoring opportunity more comparable to the opportunity players would have had in the absence of the foul. Using StatsBomb open data at the event level for the 2018 Men’s World Cup, we used a generalized additive model approach to predict expected goals for different distances and angles to the goal. The goal was to change the location of the penalty kick, so the chances of scoring were closer to the chances during regular play. We proposed replacing the current penalty kick spot with a penalty kick arc. The format of the arc would position the kick-taker based on where the foul occurred within the box as well as extending the distance from which the kick is taken. We believe this new method of penalty kicks will allow for a fairer experience while keeping the traditional aspects of the game fairly constant.

Caleb Peña is a Research and Development intern at NHL Seattle. He is currently finishing his senior year at California State University, Fullerton where he is majoring in mathematics with a concentration in probability and statistics.

Quantifying Passing: Using NBA Tracking Data to Create an Expected Assist Model

In the game of basketball, the success of a possession hinges not just on the quality of a shot but also on the quality of the preceding passes. Although much research has been done modeling the expected outcome of a shot, much less is devoted to the value added from a pass. This paper attempts to bridge that gap by offering an Expected Assists model. Using publicly available tracking data from the 2015-16 NBA season, we employ a combination of a rules-based approach and a Generalized Additive Model to identify when a pass occurs. Due to heavy computational limitations, our research concentrates on all Los Angeles Clippers games from December 2015. We then develop a shot model that projects the chance of a player making a shot given his location and the defense around him. We incorporate this into a passing model that takes into consideration the timing and location in order to create an expected points added metric for each pass. This metric allows us to identify passers who are creating good shooting opportunities for their teammates without penalizing them when their teammates are unable to convert. Isolating an individual player’s play-making from team success is a key component of equitable player evaluation.

I am a junior at Purdue University majoring in Applied Statistics and Computer Science. Over the last 4 years, I have been working on Machine learning and Data Science projects that have piqued my interest in a variety of applications of data.

High Anticipation: Exploring Trends Between Public Perception and Player Value

Being a professional basketball player is one of the many ways one can be at the forefront of the public’s eye. The way they play, their actions off the court, and their social media presence are all factors that can affect how a player is perceived by the public. This research aims to determine the trends of a relationship between public perception and a player’s performance in a game. The research focuses on the 2018 NBA first round draft class. The data consists of a series of data scrapes from Reddit, Wikipedia, and Google to gauge public perception and NBA box score data for player performance data. Hierarchical clustering was the best type of analysis to accurately display our public perception data which consisted of sentiment and popularity metrics. Two clusters were created: players who were not very popular with low positive sentiments and players who are more popular with higher positive sentiment . An XGBoost classification model was trained with an 0.887 AUC score to predict which players would fall into each cluster based on basketball performance variables like number of 3-point goals and assists. The results showed that points scored by a player and minutes on the court were given the most importance by the model. These results indicate a relationship between a player’s performance and public perception but confirming this would require an expansion of the sample and complex modeling.

Jill Reiner is a junior at Denison University majoring in Data Analytics with a concentration in Economics and a minor in Mathematics. With respect to sports analytics, she is interested in draft strategy and prospect analysis in all sports, especially hockey.

Draft Decisions in Uncertain Times: Valuing and Simulating NHL Draft Picks

The NHL draft lottery is complicated even at the best of times, but it is especially so this year, when the Pittsburgh Penguins must decide what to do with a conditional first-round pick traded to the Minnesota Wild. Using publically available data on player statistics and team performance, we built two models. The first model is a NHL draft slot value curve, which predicts a player’s contribution to his team based on his draft position, while the second model predicts a team’s performance in the upcoming season based on the previous season’s statistics. We used this second model to simulate league-wide season rankings and then first-round draft pick orders to generate a distribution of possible outcomes. Combining these two models, we created a one-number index to compare the value of this year’s pick with the probable value of next year’s pick. Based on our modeling and simulation, a solution to this dilemma is proposed to maximize the Penguins' chances of future success.

Thea Sukianto is a senior undergraduate student in Applied Mathematics at Boise State University. Her main research focus is spatio-temporal statistics with applications in agriculture and seismology. In sports analytics, she is interested in simulating game outcomes and constructing rating systems.

Evaluating Parametric Methods for Modeling European Soccer Team Goals

There is no doubt that soccer is one of the most watched and favored sports around the world. With this background, soccer betting has become a popular event for soccer fans and gamblers. To bet on the right team, we desire to know which team would score more goals in a specific match, i.e., we want to know which team has the best rating. In the realm of parametric models, zero-inflation and overdispersion have created challenges in using the traditional Poisson goal count-based model for ratings. As an alternative, previous research has proposed a negative binomial GLM to address overdispersion and zero-inflated Poisson to address zero-inflation. However, comparative studies for determining the best probability distribution to model such ratings are extremely rare. In this study, we compare the performance of four GLMs: Poisson, negative binomial, and their zero-inflated analogues, on predicting match outcome in various European leagues. Through maximum likelihood estimation and match simulation using model parameters, we found that the zero-inflated Poisson GLM yields the best prediction performance on Premier League matches during the 2015-16 season. Moreover, through the multi-category Brier score comparison, we found that this result could potentially be extended to other European leagues as well.

Lucy Ward is a fourth year undergraduate completing a degree in statistics at the University of Wyoming. She plans to pursue graduate studies and a career in statistics, with a focus on public health and medical research.

A Puck Above the Rest: Exploring the Effects of New Data on 2020 NHL Draft Decisions

Our work aimed to understand how the addition of new season data to existing player records would impact a player's place in the National Hockey League draft. Previous authors used generalized additive models in a similar context to measure time on ice as an indicator of player success, and many used the relative age of the player as one of their predictor variables. Our custom model incorporates both a logistic model and a generalized additive model to compare player's draft year statistics with their statistics from one year later. The indicator of a player’s performance was the percentage of a team’s points per game that he scored. The model outputs the product of the player's probability of being successfully drafted with the player's expected NHL contribution given that he is drafted. Future work may benefit from considering different subsets of players and assessing league strength to generalize the results.

Alana Willis is a graduating senior at Winston-Salem State University in North Carolina studying mathematics with a concentration in statistics. She is currently working as Teaching Assistant, treasurer for the Math and Stat Club, and served as a manager for the Women's Basketball Team. She has hopes of becoming a data analyst for the WNBA.

High Anticipation: Exploring Trends Between Public Perception and Player Value

Being a professional basketball player is one of the many ways one can be at the forefront of the public’s eye. The way they play, their actions off the court, and their social media presence are all factors that can affect how a player is perceived by the public. This research aims to determine the trends of a relationship between public perception and a player’s performance in a game. The research focuses on the 2018 NBA first round draft class. The data consists of a series of data scrapes from Reddit, Wikipedia, and Google to gauge public perception and NBA box score data for player performance data. Hierarchical clustering was the best type of analysis to accurately display our public perception data which consisted of sentiment and popularity metrics. Two clusters were created: players who were not very popular with low positive sentiments and players who are more popular with higher positive sentiment . An XGBoost classification model was trained with an 0.887 AUC score to predict which players would fall into each cluster based on basketball performance variables like number of 3-point goals and assists. The results showed that points scored by a player and minutes on the court were given the most importance by the model. These results indicate a relationship between a player’s performance and public perception but confirming this would require an expansion of the sample and complex modeling.

Senior student studying Mathematics and Statistics in University of Michigan, interested in bioinformatics and machine learning research.

Evaluating Parametric Methods for Modeling European Soccer Team Goals

There is no doubt that soccer is one of the most watched and favored sports around the world. With this background, soccer betting has become a popular event for soccer fans and gamblers. To bet on the right team, we desire to know which team would score more goals in a specific match, i.e., we want to know which team has the best rating. In the realm of parametric models, zero-inflation and overdispersion have created challenges in using the traditional Poisson goal count-based model for ratings. As an alternative, previous research has proposed a negative binomial GLM to address overdispersion and zero-inflated Poisson to address zero-inflation. However, comparative studies for determining the best probability distribution to model such ratings are extremely rare. In this study, we compare the performance of four GLMs: Poisson, negative binomial, and their zero-inflated analogues, on predicting match outcome in various European leagues. Through maximum likelihood estimation and match simulation using model parameters, we found that the zero-inflated Poisson GLM yields the best prediction performance on Premier League matches during the 2015-16 season. Moreover, through the multi-category Brier score comparison, we found that this result could potentially be extended to other European leagues as well.