Tuesday, Aug 6: 8:30 AM - 10:20 AM
5093
Contributed Papers
Oregon Convention Center
Room: CC-E148
Main Sponsor
Section on Statistics in Sports
Presentations
In the last decade, the offensive and defensive philosophies employed by teams in the National Basketball Association (NBA) have changed substantially. As a result, most players can no longer be classified into only one of the five traditional positions (PG, SG, SF, PF, C) and instead spend a percentage of their playing time at multiple positions, making positional data compositional. Further, given the desirability for versatile players, an argument can be made that traditional positions themselves are archaic. Using data from the 2016-17, 2017-18, and 2018-19 seasons, I explore how Bayesian hierarchical models can be used to estimate team defensive strength in three ways. First, only considering players classified by their majority traditional position. Second, by using compositional traditional positional data. Third, using compositional data from modern positions (archetypes) defined by fuzzy k-means clustering. I find that the fuzzy k-means approach leads to a modest improvement in both the root mean squared error and median 95% posterior predictive interval width for the test data, and, more importantly, identifies 11 modern archetypes that, when combined, are correlated with
Keywords
compositional data
Bayesian
hierarchical models
basketball
NBA
If professional teams can accurately predict the order of their league's draft, they would have a competitive advantage when using or trading their draft picks. Many experts and enthusiasts publish forecasts of the order players are drafted into professional sports leagues, known as mock drafts. Using a novel dataset of mock drafts for the National Basketball Association (NBA), we explore mock drafts' ability to forecast the actual draft. We analyze authors' mock draft accuracy over time and ask how we can reasonably aggregate information from multiple authors. For both of these tasks, mock drafts are usually analyzed as ranked lists, and in this paper we propose ways to improve on these methods. We propose that rank-biased distance is the appropriate error metric for measuring accuracy of mock drafts as ranked lists. To best combine information from multiple mock drafts into a single consensus mock draft, we also propose a combination method based on the ideas of ranked-choice voting. We show that this method provides improved forecasts over the standard Borda count combination method used for most similar analyses in sports, and that either combination method provides a more accurate forecast across seasons than any single author.
Keywords
Expert elicitation
Rank-biased overlap
Borda count
Ranked-choice voting
Instant-runoff voting
Rank aggregation
Baseball is unique in the major US sports in that nearly every player who is drafted will spend significant time in the minor leagues (MiLB) before reaching the major league (MLB). Beginning in 2021, the MLB draft was cut in half from 40 rounds to 20, yet still most will spend years in MiLB and retire before making it to the big leagues. This research applies competing risks analysis to investigate how different draft day factors influence the time it takes draftees to either reach MLB or retire before doing so. The results suggest position, pick number, type (high school vs. college), and bonus as a proportion of slot are all important features. This approach can be used to quantify a draftee's likelihood of reaching MLB or retiring over time based on these features, which can be of immense use to the players, their agents, and even the teams drafting them.
Keywords
competing risks
Fine-Gray model
MLB draft
survival analysis
baseball
It is well-known that some ballparks in Major League Baseball are are more conducive to scoring than others. Estimation of "park factors" that quantify these differences has received considerable attention in industry and in the literature, but has not been without criticism. We make two contributions towards the improvement of estimating these effects. We compute, for each ballpark, runs and home runs achieved by all players (home and visiting) with plate appearances at the park, when visiting all other parks. This "elsewhere" measure of performance can be used to quantify offensive strength-of-schedule observed at each park. Secondly, we fit generalized linear models to test data to estimate probabilities of a variety of outcomes (e.g. home runs, doubles, foulouts) that are specific to batter-pitcher handedness combinations. These regression models use handedness-specific relative frequencies of events computed using training data as explanatory variables. The models are fit using test data and used to compute handedness-specific event probabilities adjusted to league averages of event probabilities which we define as park factors.
Keywords
Generalized linear models.
Regression. Analysis of covariance. Covariate-adjustment.
Baseball. Park Factors.
In this paper we introduce a novel inverse decision problem formulation which we call "Memory Learning". Given a data set of human decisions and their consequences (i.e. rewards), we consider the situation in which the decisions appear to be "sub-optimal" according to a statistical analysis of the data. Our proposed method seeks to explain these deviations by learning a reweighting of observations, or 'memories', such that the analytical model trained on the reweighted observations matches the observed human behavior. We interpret the reweighting of the observations as a representation of the memory bias inherent in the decision-maker's choices. To bridge the gap between theoretical models and real-world decisions we explore various strategies for learning optimal weightings, employing both analytical and simulation methods. Finally, we introduce a unique iterative resampling approach to apply our method to the well-studied fourth down decision in professional football. Remarkably, our research reveals that our Memory Learning approach outperforms traditional classification methods in predicting coach decisions.
Keywords
sport
inverse optimization
fourth down
bootstrap
resampling
Having looked at the full match statistics for the England-France 2022 FIFA World Cup Quarterfinal, one could come away thinking "England lost despite having played better than France": 16 to 8 shot attempts, 8 to 5 shots on target, 5 corners to France's 2, resulting in a 1-2 loss. What's disregarded is the score situation: in the 40 minutes when the match was tied (0-0, 1-1), France led in all of the mentioned statistical categories, while consciously ceding initiative to England in the 66 minutes when up a goal - a tactic we'll refer to as "prevent defense" (term borrowed from American football). We use match event sequencing data across five European club leagues over the past 15 years to study impacts of prevent defense on the aforementioned statistical categories and goal-scoring tendencies for teams when trailing, leading or tied. For that we leverage categorical and count response modeling approaches with predictors that could reasonably affect the likelihood of a given team implementing prevent defense tactic, which would include, besides scoring differential, such aspects as prematch booking odds (to gauge the relative levels of opponents), red cards, time in the match.
Keywords
Sports statistics
Categorical data
Multivariate regression