CARNEGIE MELLON SPORTS ANALYTICS CONFERENCE

Adriana Gonzalez Sanchez is a senior at Slippery Rock University where she is majoring in Physics and Mathematics with a minor in Statistics. She is currently interested in causal inference and sports analytics, more specifically tennis analytics. At Slippery Rock University she is a member of the E-Board of the Statistics Club, and a member of the women’s tennis team. After graduation, she plans to pursue a PhD in Statistics and Data Analytics. Beyond academia, she enjoys swimming and watching soccer.

Does Icing the Kicker Work in the NFL?

In football, the term “icing the kicker” refers to a play immediately before a field goal attempt in which the defensive team calls a timeout in the hope that it will affect the kicker’s ability to make the field goal. This is important to look at because it is often a game deciding play. This project is designed to try to predict if icing the kicker (i.e., calling a timeout before a field goal) impacts a kicker’s ability to make the field goal. To approach this problem, a causal inference perspective was taken due to the nature of the data, which was non-random observational data. The data used for this project was NFL play-by-play data from the 1999 season to the 2021 season, the treatment variable was whether a kick was iced or not, and the set of potential controls was all non-iced kicks. There were three different models created during the process, one using all data without matching, another using matched data on win probability, and the last one using a propensity score for kick success. Then these three models were fit using logistic regression, more specifically, using a generalized additive model (GAM), and the results were compared. The results showed that there was no evidence that a kicker’s ability to make a field goal is affected by timeouts. However, in the future, factors such as roof or surface type, as well as a running clock will be considered to reach new conclusions.

I am a junior mechanical engineering student at the University of Oklahoma with an interest in sports analytics. I played competitive soccer through my freshman year of college, and have always loved sports, especially football, growing up in Texas.

Does Icing the Kicker Work in the NFL?

In football, the term “icing the kicker” refers to a play immediately before a field goal attempt in which the defensive team calls a timeout in the hope that it will affect the kicker’s ability to make the field goal. This is important to look at because it is often a game deciding play. This project is designed to try to predict if icing the kicker (i.e., calling a timeout before a field goal) impacts a kicker’s ability to make the field goal. To approach this problem, a causal inference perspective was taken due to the nature of the data, which was non-random observational data. The data used for this project was NFL play-by-play data from the 1999 season to the 2021 season, the treatment variable was whether a kick was iced or not, and the set of potential controls was all non-iced kicks. There were three different models created during the process, one using all data without matching, another using matched data on win probability, and the last one using a propensity score for kick success. Then these three models were fit using logistic regression, more specifically, using a generalized additive model (GAM), and the results were compared. The results showed that there was no evidence that a kicker’s ability to make a field goal is affected by timeouts. However, in the future, factors such as roof or surface type, as well as a running clock will be considered to reach new conclusions.

Vivek Shah is a senior in high school at the American School Bombay. He enjoys working on data science and machine learning projects. He is deeply passionate about sports and loves watching football as well as playing basketball, soccer, and golf.

Identifying Defensive Coverage Types in NFL Tracking Data

The proliferation of data in sports has led to the increased use of analytics in strategic decisions and the development of more specialized features. However, publicly available statistics for defensive players in the National Football League (NFL) are limited. Nevertheless, the NFL has added chips to players’ shoulder pads that monitor the position, speed, and direction of each player every tenth of a second. A subset of this data is released to the public and can be used to construct new measures to analyze player performance on a play-by-play level. The aim of this paper is to develop an unsupervised-learning based approach to predict player-level soft labels for man and zone coverage in the NFL. This is done through the use of a multivariate t mixture model on the tracking data. The model displays both a high Adjusted Rand Index (ARI) value and a high accuracy on a manually labeled dataset of man and zone coverage and can be used to explore how the coverage prediction for a defensive player changes over the course of a play. This paper also introduces a website that provides a simple interface that enables users to label if a defensive player is in man or zone coverage and thus can be used to create a large crowd-sourced dataset of defensive coverage labels. The dataset can be used to further improve the model and to explore anomalous defensive situations where the model’s predictions are radically different from human labels.

Eric Warren is a senior at North Carolina State University, where he is studying Statistics. After graduation, he plans to pursue a career in data analytics with interests in sports and finance. He is an avid football and hockey fan, who enjoys watching and playing any sport in his free time.

Paying Players What They Are Worth

The salary cap is something that all sports have. In the National Hockey League there is no luxury tax or roll over cap, which is present in other professional sports. For this reason, optimizing spending with player performance is a must, since all teams are on a set budget. This project looks at each player's statistical outputs during a season and determines what the salary they should be receiving compared to their actual salary.

Boyuan Zhang is a third-year undergraduate student at New York University, where he is joint majoring in Data Science and Computer Science, and double minoring in Mathematics and Business Studies. He is passionate about the insight from data and his academic interest is to develop, apply, and analyze machine learning techniques to solve real-world problems. Boyuan is currently a research assistant at the Visualization Imaging and Data Analysis Center (VIDA) at NYU Tandon School of Engineering. He is looking forward to pursuing graduate education in the field of data science and statistics.

A Regularized Adjusted Plus-Minus Model in Soccer with Box Score Prior

Evaluating the impact of individual players on their team’s performance is an important question to consider when analyzing team sports. Variants of the Adjusted Plus-Minus (APM) model have been viewed as all-in-one player’s performance evaluation matrices, and have been widely utilized in basketball and hockey. However, because of the low number of substitutions and scoring chances in soccer, the APM model has not been shown to be effective in identifying players' performance. This talk introduces a new kind of Regularized Adjusted Plus-Minus (RAPM) model, which incorporates priors generated from box score statistics into a regularized regression framework, performing point estimation on the player’s contribution to the expected goals per 90 minutes. In particular, using data from the 2021-2022 season of the English Premier League, we show that our RAPM model with box score prior has better predictability and interpretability than the APM model, RAPM model without priors, and RAPM model with FIFA ratings as prior. This model could be further utilized to evaluate the impact of player transfer, simulate teams' performance, and forecast players' market value.

Michele is a senior Computer Science and Statistical & Data Sciences major at Smith College in Northampton, Massachusetts. She is interested in hockey analytics and hopes to pursue a career in sports analytics post-graduation. She is captain of the Smith College club hockey team and enjoys playing hockey, lifting, and listening to music in her free time.

Does Not Playing Hockey Make You Worse At Hockey?

The COVID-19 pandemic disrupted all aspects of life when it began in 2020, and professional sports were no exception. In 2020-2021, many hockey leagues had shortened or canceled seasons due to the restrictions caused by the pandemic. Some players participated in alternative leagues or tournaments while others did not record any games for that season. Our goal in this research was to investigate the impact of pandemic restrictions on hockey player development in the OHL. We defined our treatment variable as whether a player participated in at least one game during the COVID-19 impacted season (2020-2021), and we assessed the impact on a player’s performance (points per game) the following year. We used a variety of approaches including multiple variations of regression (particularly ordinary least squares, gamma regression, and mixed-effects models) and causal analysis in the form of propensity score matching and Bayesian Additive Regression Trees. In this talk, we will share our findings and our plans for future study.

Jackie is a senior at Elon University majoring in Statistics and Music in the Liberal Arts. She is a member of the Baseball Analytics Club, part of Sigma Kappa Kappa Zeta, and an academic mentor. Her passions for analytics, music, and sports have culminated in many interesting projects that combine the fields. After graduation she hopes to pursue a career in baseball or hockey analytics. The headshot is attached.

Does Not Playing Hockey Make You Worse At Hockey?

The COVID-19 pandemic disrupted all aspects of life when it began in 2020, and professional sports were no exception. In 2020-2021, many hockey leagues had shortened or canceled seasons due to the restrictions caused by the pandemic. Some players participated in alternative leagues or tournaments while others did not record any games for that season. Our goal in this research was to investigate the impact of pandemic restrictions on hockey player development in the OHL. We defined our treatment variable as whether a player participated in at least one game during the COVID-19 impacted season (2020-2021), and we assessed the impact on a player’s performance (points per game) the following year. We used a variety of approaches including multiple variations of regression (particularly ordinary least squares, gamma regression, and mixed-effects models) and causal analysis in the form of propensity score matching and Bayesian Additive Regression Trees. In this talk, we will share our findings and our plans for future study.

Ben Wieland is a sophomore at the University of Virginia majoring in applied statistics with a concentration in data science. He is a member of the Sports Analytics and Statistics Laboratory and writes for Virginia sports blog Streaking The Lawn. His work focuses on making new data accessible and making current data understandable through effective data analysis and visualization. His goal is to work as a data analyst for a professional sports team and potentially pursue an M.S. in Data Science.

fRisbee: An Open-Source Package for College and Professional Ultimate Frisbee Data and Modeling

The popularity of Ultimate Frisbee at the collegiate level is skyrocketing: over 650 men’s and women’s teams competed in the 2021-22 season. fRisbee is an open-source R package designed to create a fast and accessible way to access collegiate men’s and women’s Ultimate Frisbee rankings, results, and historical data. The package provides ease of access to previously infeasible datasets, including access to aggregated game-by-game information on over 5,400 Ultimate Frisbee matches at the collegiate level during the 2021-22 season; it also includes web scraping functions wrapped around www.frisbee-rankings.com to access up-to-date results and data. The package also makes Ultimate Frisbee modeling simple: it comes equipped with two pre-trained and easily deployable models for win probability and margin of victory projections, and includes easy access to the aforementioned historical data repositories for users to train and fit their own models.

Akiva is a second-year undergraduate at the University of Virginia majoring in economics and statistics with a concentration in data science. At UVA, he is a member of the Data Science Club and plays on the Ultimate Frisbee Club team. After graduation, he hopes to pursue a career in data analytics with interests in sports and utilization of resources.

fRisbee: An Open-Source Package for College and Professional Ultimate Frisbee Data and Modeling

The popularity of Ultimate Frisbee at the collegiate level is skyrocketing: over 650 men’s and women’s teams competed in the 2021-22 season. fRisbee is an open-source R package designed to create a fast and accessible way to access collegiate men’s and women’s Ultimate Frisbee rankings, results, and historical data. The package provides ease of access to previously infeasible datasets, including access to aggregated game-by-game information on over 5,400 Ultimate Frisbee matches at the collegiate level during the 2021-22 season; it also includes web scraping functions wrapped around www.frisbee-rankings.com to access up-to-date results and data. The package also makes Ultimate Frisbee modeling simple: it comes equipped with two pre-trained and easily deployable models for win probability and margin of victory projections, and includes easy access to the aforementioned historical data repositories for users to train and fit their own models.

Robert Binion grew up in Atlanta, GA and now lives in Salt Lake City, UT with his wife and two children. He does data analysis and software quality engineering within a data lake team for a large restaurant corporation. Previously, he spent over ten years in non-profit and church work before coming to the world of data. His foray into the world of sports analytics started with charting play by play data for Georgia Tech football games and has expanded from there. He currently writes analytics-driven football content for Football Outsiders and From The Rumble Seat. He has a B.S. in Chemical Engineering from Georgia Tech and is currently pursuing a Master of Science in Analytics at Georgia Tech. Robert enjoys watching football, reading across all sorts of genres, and spending time in the mountains with his family.

The Racial Imbalance in College Football Coaching

College football has had a race problem since its inception. From issues of team integration to paying student athletes to demographics of the coaching community, race is never far afield from the preeminent issues that permeate the college football landscape. This paper presents a data-driven look at just how inequitable the current racial distribution is in FBS coaching and provides some insights into how this can change. This project uses regression, community detection, and classification models to show that there are persistent racial imbalances in the ranks of college football coaching. The imbalances are not due to performance but are perpetuated by reliance on connections within the coaching community and potentially exacerbated further by implicit or explicit racial bias. To mitigate these imbalances, significant shifts are needed within the paradigms that govern hires made at the highest level of the sport.

Mark Wood is from Atlanta, GA and is a design & construction Project Manager for Georgia Tech. He has previously worked on commercial and industrial projects throughout the southeast U.S. and in South Africa. He has a B.S. in Civil Engineering from Georgia Tech and is currently pursuing a Master of Science in Analytics at Georgia Tech. In his free time he enjoys spending time with his wife and two kids, watching all sports except baseball, and going to concerts.

The Racial Imbalance in College Football Coaching

College football has had a race problem since its inception. From issues of team integration to paying student athletes to demographics of the coaching community, race is never far afield from the preeminent issues that permeate the college football landscape. This paper presents a data-driven look at just how inequitable the current racial distribution is in FBS coaching and provides some insights into how this can change. This project uses regression, community detection, and classification models to show that there are persistent racial imbalances in the ranks of college football coaching. The imbalances are not due to performance but are perpetuated by reliance on connections within the coaching community and potentially exacerbated further by implicit or explicit racial bias. To mitigate these imbalances, significant shifts are needed within the paradigms that govern hires made at the highest level of the sport.