Welcome to Sabermetrics, Scouting, and the Science of Baseball, our ninth annual charity baseball research conference. This year, we are proud to support the Angioma Alliance, an organization by and for those affected by cavernous angiomas and their loved ones, health professionals, and researchers. We hope you have a fun time, learn some new things, think differently about a problem, and meet some great people.
Members of the Red Sox front office take questions about player development, analytics, and scouting.
The success of a pitcher depends on a number of variables including the quality and frequency of his different pitch types. Game theory can be used to find optimal pitch distributions but depends on having an accurate measure for the quality of each pitch. Linear weights pitch values have been used for this purpose but depend on observed outcomes and, as a result, have a low degree of repeatability. We develop a method that determines the optimal pitch mix for a pitcher against a given handedness of batter based on the physical characteristics of each pitch type. The method utilizes an objective function surface that is constructed using nonparametric regression. The underlying metric on pitch distributions is defined using a whitened earth mover's distance. We show that the new model accurately predicts the dependence of performance on changes in pitch distribution.
A pitcher may be in for a difficult outing when hitters are able to extend their at-bats past two strikes, be it past a full count or staying alive at 0-2. That being said, how and in what fashion do these situations impact the pitcher for the duration of his appearance in that particular game? Much like the ability of baserunners to affect a pitcher’s rhythm, hitters have a unique opportunity to increase a pitcher’s pitch count and throw off his pitch sequencing. In fact, based on PITCHf/X, 6.7% of all pitches in 2018 were fouled off in a count that already included two strikes. In regular season games, hitters saw more than 5 pitches in an at-bat approximately 25% of the time in 2017 and 2018. Once the strike count has gotten to 2, or the pitch count for an at-bat reaches more than a full count, how does a pitcher change his repertoire, and which pitch will he use to set up the last strike? Also, just how pivotal are such at-bats to the rest of the pitcher’s time in the game? These are questions I will answer and investigate through the use of, among others, Retrosheet, Statcast and PITCHf/X data. Additionally, using these data sets and other related research, I seek to find if pitchers regain composure and become more effective after encountering a long at-bat, or do they trend towards an issue-related outing, or even a meltdown?
Using data on free agent signings since 2006, I examine the evolution of MLB front offices’ player valuations over time. I compared the awarded contract lengths and amounts to each player’s underlying statistics in the years prior to the signing. For position players, this includes measurements of each player’s offensive and defensive value, as well as framing for catchers. For pitchers, this includes fastball velocities, pitch repertoires, injury history and past performance. I found that contract valuations have changed in recent years in several ways. These shifts correspond in some cases to known advances in our ability to measure player contributions to winning. For example, catcher framing was not significantly correlated with annual salaries in the period from 2006-2011, but became strongly correlated thereafter. There has been an overall decrease in the extent of correlation between defensive metrics and position player valuations, suggesting that teams may be coming to rely upon proprietary defensive indicators rather than the publically available statistics. For both hitters and pitchers, valuations have become less dependent on the recent observed performance of the players and more strongly correlated with intrinsic characteristics, like fastball velocity. Overall, free agent contracts are increasingly drifting away from the public consensus of player values, which may indicate that the gap between public and private valuations is widening.
Winning in a given baseball season requires prediction of future player performance based on historical player performance. Historical studies of player aging performance have concentrated on mean player aging to predict the future performance of a given player, many using the so-called delta method (Bradbury 2010, Zimmerman 2011-2014, Druschel 2015, etc.). Recent work in baseball research has tended towards probabilistic methods for prediction, including contract risk determination (Forman 2019), pitch selection (Melling 2017), and player psychological effects (the "hot hand", Arthur and Matthews 2017). In this work, I will propose a method for understanding player performance in a given year as a noisy sample of underlying true skill, which, from one year to the next is governed by an underlying process that is modeled using a probabalistic interpretation of the delta method. The resulting prediction with confidence bounds will give some insight into expected player performance and could provide a prior distribution for Bayesian team performance predictions.
The game of baseball has changed little over the past century, but how management builds a team has. Rewind 20 years, and teams were trying to sign big name hitters that had the highest average and most home runs, or pitchers with the most wins. In a sport with no hard salary cap, the big market teams would acquire these players while smaller market teams watched as their players left for more money. As a countermeasure the implementation of ‘Moneyball’ arose where players with high on base percentages were valued. As years passed, players became valued by many ‘invisible’ statistics; player metrics captured by the machine such as pitcher spin rate or exit velocity. Although these new measures serve a purpose, they are essentially meaningless if their use doesn’t improve a team’s wins. Player analytics and statistics are primarily used to increase win probabilities. They can be pointless and costly when a team fails in this objective. From our research we will propose a spending option for distributing player salaries that teams with less financial power can use to increase wins. Specifically, we will suggest how teams can best optimize spending on their pitching staff. We will illustrate how spending on specific less-costly pitchers can result in a team with either a smaller pitching staff or more money to spend on higher quality everyday players. Our research will focus on teams in the National League Central (NLC) using data from years 2000 to present.
Rule 5 Draft is a relatively low-risk, high-reward event to build up team rosters. The presentation aimed to determine strategic looks for Major League front offices to evaluate Rule 5 draftees, which only focuses on the Major League phase. The presentation covers draft history and player information from the past 13 years, begin from 2005 to 2018. With a total of 210 players drafted since 2005, including 151 pitchers (72%) and 59 position players (28%), Orioles and Phillies led the league in 16 players drafted in the past 13 years, while Dodgers have not drafted a single player since 2005. Further, team transaction strategies can reflect on the 40 Man roster, 21 players (10%) were originated from the Yankees, which shows the clubhouse has insufficient spots for young talents, will be discussed in the presentation. Time and age are crucial elements to baseball players. My research findings indicate that only 24.8% of players make the routine lineup/rotation 1.7 years after the draft. Players are averaged 24.3 years old when drafted, and who make the routine lineup/rotation averaged 24.0. Moreover, exclude players including R.A. Dickey and Darren O'Day who debuted prior the Rule 5 draft, the average declines .4 to 23.6. Which shows that teams should target younger players in the draft pool. Major leaguers usually have more than one tool that make them special. During the course of my research, I found that Rule 5 position players, who find themselves in the lineup, are generally more aggressive. The aggressiveness mirrors that of Plate Discipline statistics, and produces a relationship between offensive production and plate discipline. I ran a regression analysis for Rule 5 hitters whose Swing% is above MLB baseline and the relationship between wOBA and Swing% had an r^2 of .30 while the relationship between wOBA and Z-Swing% had an r^2 of .28. Josh Hamilton and Odubel Herrera are Great examples. Relievers with great stuff and an effective breaking ball are more likely to stick in the bullpen, with familiar names including Tommy Kahnle and Ryan Pressly. Kahnle ranks top 6% with a 11.7 wCH(Changeup runs above average) since 2014, while Pressly being the top 5% with a 17.1 wCB(Curveball runs above average) since 2013 among qualified relievers. Inning Eaters who possess an outstanding pitch are productive as well, including R.A. Dickey and Ivan Nova, while comparison of extended stats will be mention in the presentation. With over 75% of draftees fail to contribute on a routine basis to the big leagues, front offices should target certain tools when it comes to the right time to draft a player during the winter meetings. Future studies will be Rule 5 players’ value in trade and psychological points of view and eagerness to prove themselves along with Plate Discipline statistics.
J.T. Realmuto is doing things behind the plate that no other catcher is doing in the league. Muñoz will show you statistics, videos, and quotes from this player, and former MLB catchers discussing their approaches behind the plate.
A presentation on recently published data on THT and Fangraphs discussing ways to think about how we should model the strike zone.
This presentation will evaluate pitching data for evidence of injuries before they happen as well as quantify lost production due to injury. There are numerous measurements that can be taken to quantify a pitcher’s physical output including pitch speed, horizontal and vertical movement as well as spin. These measurements can be examined for sudden changes or irregular patterns in a pitcher’s physical ability. It is feasible that such patterns might occur before a pitcher is injured and provide a warning to coaches and trainers. Two statistical methods will be utilized to model these patterns, time series and time to event survival analysis. Time series models can be evaluated for structural breaks and sudden point changes. They can also be evaluated for changes in conditional variability. Using PITCHf/x data, these series can be constructed using pitches, innings, games or even seasons as time units and physical measurements or averages as point values. The point in time where the pitcher sustains an injury is identified and different lengths of time preceding that point can be tested for sharp departures from the general trend. The series can also be broken up into groups based on injury type in order to examine if certain injuries see different changes in physical output before occurring. Beyond the section preceding the injury, periods of time after the injury can be evaluated for structural breaks to determine how effective the recovery was. The time to event analysis will be used to create a survival curve for a pitcher that represents the average time before a pitcher is injured. Using this model, a measurement for how healthy a pitcher is can be developed. This project will examine a handful of parametric, semi and non parametric models. Using parametric models, pitch characteristics can be used to help explain the time that it takes for a pitcher to get injured and different models can be built for different pitcher groups such as right handers or left handers. The curves can then be used to estimate how much value is lost when a player is injured at different points in a season or career. We can create a synthetic lifetime using different groups of players and determine how long a player should last. Lost production can be estimated from this relative lifetime.
Ulnar collateral ligament reconstruction (UCLR) is a common procedure used in pitchers. In 2007, PITCHf/x (Sportsvision, Inc.) was introduced into Major League Baseball (MLB) stadiums in order to track pitching metrics, including velocity and movement. Studies have investigated the relationship between pitch metrics and UCLR, but none have investigated the effects of the injury on the movement of pitches following surgery. The purpose of this study was to use the PITCHf/x database to analyze the velocity and movement of the fourseam fastball, curve, and slider thrown before and after UCLR in MLB pitchers.
Musculoskeletal injuries in baseball players are a persistent and significant problem,1-4 with the greatest incidence attributed to the shoulder, elbow, and trunk. These injuries are common and injury rates are increasing. Recent estimates suggest Major League Baseball has lost 7 billion dollars in lost wages, which the majority are from upper extremity injuries. These statistics suggest current approaches to understanding and preventing upper extremity injuries within baseball are not effective. The majority of research has focused on the upper extremity, while foregoing the entire kinetic chain and psychological impact. Thus, it is imperative to study a combination of modifiable physical, psychological, and biomechanical factors to improve upper extremity injury prediction and enable targeted injury prevention programs. The purpose of this study was to determine the association between physical factors (range of motion, trunk and hip dissociation, and strength), psychological factors (grit and resiliency), 3D biomechanical pitching factors (torque, force, and kinematics) and shoulder and elbow injury in a cohort of high school pitchers.
Injuries can be devastating at any level, but, with millions of dollars on the line for professional athletes, they are in a critical league of their own. If the position and/or injury of a potential injured player is more crucial, then this is significant as it may affect the future of the series, season, and even the future of that athlete’s career. Therefore, we look at performance data in the past and various datasets from multiple years to see if various types of machine learning methods can statistically predict if a MLB player will be classified as being at risk of a future injury. Based on the performance of multiple training and testing models, we further extract which features will be the most impactful through a random forest variable importance plot. We look at variable features such as type of injuries, occurrence, time, days to recover, etc. to see which method will have the highest significant value or correlation to achieve the best accuracy percentages in prediction of injuries. Sports is inevitably unpredictable, however, this study uses numbers, statistics, and machine learning to see if we can make a difference by classifying a rank of at risk MLB players.
Athletes experience unique stressors that have been indicated to increase their vulnerability to mental health symptomology, including intensive travel, fatigue, and criticism from others. Although there has been a lack of systematic examinations of mental health in baseball, anecdotal evidence suggests baseball players are at least as vulnerable to mental health symptomology as athletes in other sports and non-athlete peers. Of course, mental health symptomology results in gross negative consequences, including relatively high medical costs, distractions, poor performance, injury and so on. Along this vein, the aim of the current proposal is to conduct a systematic literature review of all studies that have been conducted with athletes and identify areas of mental health that may be particularly relevant to baseball players. This comprehensive review of the literature will facilitate cost-efficient development and utilization of mental health assessments and interventions in baseball players; which is currently absent in the management of mental health in baseball.
The research will investigate the performance statistics of different MLB players. Players will be grouped and sorted based on whether they use the traditional helmet or the C-Flap helmet. An entire review of their performance statistics will be discussed including Statcast batted ball data alongside traditional performance outcome statistics. The paper is based on the premise that the C-Flap helmet sharpens cognitive focus and reduces distractions and that players with the new helmet will have heightened performance. The two groups will be compared on a variety of statistics including strikeout rate, walk rate, on-base percentage, slugging percentage and on-base plus slugging. Player’s swing tendencies will be investigated including pitch selection and swing tendencies by count. Statcast data including launch angles and exit velocities will be compared between the two groups. The ultimate goal of the research is to potentially determine if one helmet design offers a competitive advantage compared to the other.
This research aims to examine trends in games going to extra innings. Looking at the span of games beginning from just after the 2015 All-Star Break - the beginning of the home run surge - to the end of the 2018 season, I attempt to answer the question of where exactly to obtain the most value (i.e., minutes of baseball) for your attendance money via a system of analytical models that examine both park factors and longitudinal factors over time.
The comment appended to Rule 5.05(b)(1) allows for a base on balls “including an award of first base to a batter by an umpire following a signal from a manager.” Commissioner Manfred, as he has done with many changes, argues this is for pace of play, despite rare occurrences. However, this rule change, for all the time it saves, has tilted the balance of power at the plate – completely removing the batter from the equation – as Bryce Harper famously found out. Baseball action is defined by the batter-pitcher dynamic and everything possible should be done to preserve this. It robs fans of the possible well-off-the-plate hits, but these can be sacrificed if the rules make sense. With one twist, free passes could come with a charge: more passes, more bases. This presentation will look at the hypothetical consequences of the second intentional walk being worth two bases. Like fouls in basketball crossing a threshold, so too would intentional walks. In a time where there is concern about pitchers having too strong an upper hand, there is a tweak that could increase offense, either directly by encouraging managers to pitch to batters or indirectly by getting more players into scoring position or beyond. How often are there multiple intentional walks in a game? How different would these games have played out if this rule existed? Would there have been a difference in any games of consequence that the manager would not have called for an intentional walk? Would it add an element of managerial strategy in the wake of universal DH in the near future?
Weather has long been known to affect the flight of the baseball. Scientific effort has given us the ability to calculate flight trajectories in a laboratory setting. Previous analysis has been done with real-world baseball and weather data, but has been hampered by the lack of good weather data. We use improved weather data and the latest major league data sources to examine the effects and begin to build a model and a process for predicting outcomes if they had happened in a different stadium or at a different time.
In today's game, statistics has becoming more intertwined with sports more than ever. As Major League Baseball is trying to find ways to appeal to a more younger generation, the hope is that this application can be used to introduce some basic statistics to players at a younger age. This application's target use is for little league and high school coaches that will store basic statistics for hitters and pitchers of their respective teams. Then said players can login to see their statistics. This application aims to provide an approach to younger players that are more inclined to use their phones and hopes to get them interested in statistics. The presentation would showcase the functions of the application.
The purpose of this study is to evaluate the major physical disparities and boundaries between elite male and female baseball players and to explore if female baseball players have or will have the ability to be able to contest long-term as part of a Minor or Major League organization. Based on statistics from the 2018 Women’s Baseball World Cup (the pinnacle of women’s baseball), the top four teams’; Japan, Chinese Taipei, Canada & USA statistics were averaged and evaluated by Speed, Arm Strength and Hitting. These statistics were recorded by analysis from games played at the 2018 Women’s World cup. The data that was collated was then assessed against the 20-80 scouting scale from an undisclosed Major League Baseball (MLB) organization. Scouts use the grading scale to evaluate prospects. 20 is the lowest grade that a player can get, with 80 being the highest. Most prospects will hover around the 50 mark, which indicates the average grade of tools possessed to be a current MLB player. Although, in the case of this study, playing statistics were averaged by country not by individual to succinctly identify the playing level that women are at. The data that was captured evaluates the current performance of these female athletes at the world cup – which should be their peak performance. Therefore, ensuring that the data collected was of eminence. As the data is compared to 20-80 prospect scale, the study evaluates the major physical disparities of female baseball players and challenges if women will ever have the ability to contest at a men’s professional level. Factors such as; biomechanics, pitch speed, age, funding, and exposure to quality training & showcase events are all considered in a decision matrix. To conclude the study, a projection is predicted of how good female players may be once they are given similar opportunities and if they will have the ability to be able to develop as organizational players at a minor or independent league level.
Many NCAA baseball players, particularly those from power conferences like the SEC and PAC-12, have demonstrated the ability to rise rapidly through farm systems and impact Major League teams within a short time of being drafted. Despite its increasing sophistication and relevance, college baseball remains largely unexamined by the sabermetric community. In contrast to Major League Baseball, with large datasets readily exportable from a number online sources, college baseball’s datasets are relatively small and inconvenient to access. Observations for a limited number of variables are gathered and organized into tables, but many more lie locked in strings of Play-by-Play text on NCAA team websites: pitch counts, base-out states, batted ball locations, and more. Accessing, organizing, and making these datasets accessible can revolutionize the statistical understanding of college baseball. Using Python to scrape and process NCAA Play-by-Play data, and R to analyze and visualize it, we will identify newly available variables and explore the avenues of analysis they open. We will present findings on notable college draft picks from recent years, and discuss the implications the analysis may have for the ability of Major League organizations to optimize the way they scout college players in the future.
An update on MLB's Diversity Fellowship Program, which is a full-time, 18 month, supported opportunity to work in a participating club's front office. The application for the next class of fellows (http://www.mlb.com/fellowship) will go live on August 19, 2019.
Practically all MLB pitchers use an arsenal of multiple different pitches. Game theory suggests that if both pitchers and hitters are at Nash equilibrium then, on average, each type of pitch should have the same value. This should be true for any base-out-count state, though likely with a different optimal mix of frequencies. Assuming the hitters are behaving optimally, if one pitch has lower effectiveness than others, the pitcher should throw it less often making it less expected and therefore more effective; similarly, pitches with higher effectiveness should be thrown more often. Likewise, if a pitcher is behaving optimally, then the same logic applies to how batters expect a pitch. If either deviates from equilibrium, we would expect the other also to deviate to take advantage of the opponents’ suboptimal strategy. The first goal of this study was to determine if batters and pitchers operate at a Nash equilibrium. My research showed that the equilibrium is not being attained. After controlling for batter, pitcher, and base-out-count state, I found that relative to four-seam fastballs, sliders are worth more than half a run per 100 pitches. This difference is massive and statistically significant. Changeups, curveballs, and cutters have differences between 0.25 and 0.4 runs per 100 pitches that are also statistically significant. Two-seam fastballs, sinkers, and splitters had run values similar to the four-seam fastball. Knuckle-curves were better, but not significantly due to small sample size. Overall, these data show that either pitchers are not throwing enough off-speed pitches, or the hitters are not expecting enough of them. I then sought to determine what the optimal frequency mix for pitch types is. To do this, I looked at pitchers who pitched a sizeable sample in two consecutive years and regressed their values for off-speed pitches and fastballs in year 2 against those values in year 1 and the frequencies in both years. This analysis gave a rough estimate as to how much pitchers should change their fastball usage and how much improvement they could get gain from switching to the optimal frequency mix. On average, pitchers should throw 3.1% more fastballs than they currently do. That this number was positive was surprising, but largely came from the analysis showing that the value of off-speed pitches is more sensitive to changes in frequency. This means that hitters should expect more off-speed pitches in order to return to equilibrium. If pitchers followed the optimal mix, they would expect to see an improvement of roughly 0.05 runs per 100 pitches. While this may not seem like much, it corresponds to roughly one win per season for each team. Doing further analysis on each different pitch type may also improve this even further. Future research would analyze the optimal pitch mix for each pitch type, study how the deviation from Nash equilibrium changes based on situation, and control for changes in velocity and batter when determining the optimal pitch mix.
As analytics has become fundamentally integrated into much of professional baseball, the college game seems like the logical next frontier for a widespread analytics revolution. Over the last few months as a baseball research and development intern at The University of Texas, I have seen firsthand how colleges have started using ball tracking technology such as Trackman and Rhapsodo to gain insight into the same kinds of metrics that have taken Major League Baseball by storm. However, this transition has been slow, given the significant difference in resources and manpower that even the biggest college programs can devote to analytics. While Texas Baseball is still in the process of fully integrating analytics into the day-to-day operations of the team, the program has made great strides towards the cutting edge of amateur baseball. My presentation will focus on some of the work I have done in my first season with the Texas Baseball team, looking at the challenges I’ve faced and successes I have had in starting a baseball R&D department from the ground up. In entering a largely untapped area of baseball research, we've taken multiple approaches to evaluating college players, from tackling NCAA D1 run values with web scraping to analyzing and modeling events based off the data we’ve obtained from 40 teams’ Trackman systems, and looked for ways to present that information in a way the coaches can understand and use. Many of these projects, looking at things like pitch profiles, catcher framing, or hitter tendencies, started with existing research which used MLB data, but I quickly found that the college game introduces countless wrinkles that do not exist as prominently in professional baseball (such as large talent and competition level variance, small sample sizes, and, for many players, a lot of game-to-game inconsistency). I will talk about how I've adapted and adjusted some of the tools and metrics I've worked on for these aspects of the college game and discuss how students can be the driving force in bringing the data revolution to College Baseball.
While conventional pitching statistics rely on a conglomerate of play-by-play outcomes, objective pitch grades seek to solely analyze the pitcher’s skill set to more accurately evaluate his expected performance based on his past levels of fundamental skill sets. Integrating a foundation of traditional scouting principles with innovative pitch f/x data provides a tool to objectively evaluate each pitcher’s strengths and weaknesses, which cuts out noise from batted ball outcomes. Player value (specifically pitchers) is extremely volatile because many standard pitching statistics are based on outcomes that become riddled with noise and inconsistencies. These cloud our ability to predict the player’s true skill set, and sub sequentially, his future value. Algorithms that utilize data from player tracking technology will allow us to evaluate the player’s true skills and will prove to be more reliable tool to diagnose strengths and weaknesses. The proposed algorithm scrapes pitch f/x data from MLBAM’s gameday application, scrubs misclassified pitches to improve data quality, and benchmarks pitchers based on their pitch peripherals. Finally, to communicate with the greater baseball operation, each variable is distributed on a 20-80 scale, with 50 as the mean, and 10 as the standard deviation. Therefore, pitchers that have “80 Fastball Velocity” throw a fastball that is three standard deviations faster than the average MLB fastball.