Big Data Baseball Page 7
* * *
At the 1906 Plymouth country fair in England, eight hundred people took part in a contest to guess the weight of a particular ox. Statistician Frank Galton noted the average guess of 1,207 pounds was within 1 percent of the actual weight—1,198 pounds. The anecdote is contained in James Surowiecki’s book, The Wisdom of Crowds, which theorizes that decisions made by a diverse collection of individuals are likely to produce more accurate predictions than even those of experts. With the advent of the Web, never before could so much knowledge, so much data, and so much brainpower be thrown at a problem. The wisdom of crowds’ closely related kin are online open sourcing and crowdsourcing. With the Web, companies and institutions are able to outsource problem-solving, giving tasks once performed by employees to the public.
In 2006, Netflix put open sourcing to great use. The movie-streaming company’s primary goal was to connect people to the movies they liked and predict other movies users might like. The company asked the crowd to see if it could produce an algorithm better than its own, Cinematch, and offered a $1 million prize. In 2006, Netflix presented the interested parties with a data set of more than a 100 million ratings that 480,000 anonymous users gave to 18,000 movies. Netflix withheld over 3 million recent ratings from those same subscribers. Contestants were required to make predictions for the 3 million more recent ratings and better Netflix’s own algorithm’s predictive power by 10 percent to win the prize. By June 2007 more than 20,000 teams had registered for the competition from over 150 countries, according to Netflix. The algorithm had to incorporate millions of ratings, thousands of users, and the ever-evolving preferences of those users. In 2009, after three years of heavy collaboration, hundreds of e-mail exchanges, and long nights, BellKor’s Pragmatic Chaos became the first team to beat Netflix’s algorithm for predicting ratings by over 10 percent, at 10.6 percent.
Crowdsourcing and open sourcing began impacting an assortment of industries, including baseball. And the phenomenon benefited from baseball’s first great automated big-data collection tool, PITCHf/x.
* * *
Entering 2007, the Chicago-based company Sportvision was best known for enhancing telecasts, bringing the glowing hockey puck to NHL telecasts and the yellow first-down marker superimposed on the field to football. The camera-based, motion-tracking system PITCHf/x was developed to improve for ESPN its K-Zone product, which measured if pitches were in the strike zone.
PITCHf/x was functioning and collecting real-time pitch data in select parks in 2007, and every major league stadium in 2008. That season, three 60 Hz cameras were mounted inside every park. The cameras and object-recognition software capture images of a pitch’s flight from the time it leaves a pitcher’s hand until it crosses home plate. From the images the speed, trajectory, and three-dimensional location of the ball are calculated in real time. PITCHf/x has a margin of error of less than one mile per hour in speed and one inch of location. Moreover, PITCHf/x labeled every pitch type in real time. For the first time, the exact speed of a pitcher’s throws and the exact percentage of times he threw certain pitches could be tracked. Finally standard measurements existed for pitch speed, type, movement, and location, and this data was made widely available, such as the on Web sites FanGraphs.com and BrooksBaseball.net.
Wrote University of Illinois professor Alan M. Nathan in a 2012 paper on physics and baseball, “[PITCHf/x] records with unprecedented precision such quantities as the pitch speed and the location at home plate. But even more importantly, we have measures of quantities that we never had before.”
Database journalist Sean Lahman made a presentation titled “Baseball in the Age of Big Data” at a 2013 Society for American Baseball Research conference in Philadelphia. Lahman explained James’s Baseball Abstracts had spiked the sport’s total data points to more than two hundred thousand per season in the early 1980s. By 1990, Project Scoresheet’s play-by-play data tripled the game’s total data points to just under 1 million per season. Still, Lahman noted that big data doesn’t mean a lot of data. It means collecting every shred of data available and employing complex mathematical formulas to draw conclusions from it. Baseball did not have a true big data tool until PITCHf/x. “James clearly understood that to make advances in our understanding of the game, we needed to make a quantum leap in terms of the amount of information we had available,” Lahman said at the presentation. “As far as I’m concerned, that was his real genius.”
While Bill James and John Dewan understood the importance of creating more data points, they were manually gathering much of it. PITCHf/x was automatically generating nearly 20 million usable data points per year, nearly as much as all the data recorded in baseball in the twentieth century, a giant leap for the game.
The growth has only accelerated as Lahman estimates new player-tracking technology that was being tested in a handful of ballparks in 2014 could push data points to 2.4 billion per year.
With the creation of PITCHf/x, major league front offices had overnight more data than they knew what to do with and not enough manpower to think of creative ways to employ it. But outside the front offices the game had hundreds if not thousands of creative, statistically inclined hobbyists who, like teams, were curious to search for hidden value in the sport. And hobbyists stumbled upon a use of data that teams had not thought of.
Pitch framing is a catcher’s skill to influence borderline ball-strike calls. A batter and a home-plate umpire each have less than half a second to identify a 90 mph fastball as in or out of the strike zone. How a catcher receives the ball is a visual trick, a sleight-of-hand skill that can influence umpires to call borderline pitches as strikes. The ability to frame pitches had always been thought to carry value by managers, coaches, and players, but in the analytical community the skill had not been quantified, so its value was underappreciated. No one doubted that if the ability to turn a borderline ball into a strike existed, it would be immensely valuable. A hitter’s batting average changed dramatically based upon whether the count was in his favor or the pitcher’s. The difference between a two-ball and one-strike count versus a one-ball and two-strike count is nearly 200 points in batting average.
On April 5, 2008, an article appeared on the Web site BeyondtheBoxScore.com by Dan Turkenkopf. By day, Turkenkopf was a data architect for the software company Apprenda. By night, he was a baseball blogger. His was the first known effort to quantify the value of pitch framing based upon PITCHf/x data. He worked off PITCHf/x research from fellow hobbyists Jonathan Hale and John Walsh. They had published articles in 2007 that used PITCHf/x data to track how accurately major league umpires called balls and strikes. They were curious to identify individual umpires’ strike-zone biases. Turkenkopf was most interested in the gray areas, the borders of the strike zone, the areas where Hale and Walsh found, on average, umpires called pitches as strikes 50 percent of the time. This is where the value of pitch framing resided.
Wrote Turkenkopf, “Using Walsh’s strike zones, empirically defined as the areas where at least 50 percent of pitches are called strikes, I determined for each catcher how many balls should have been strikes and vice versa. This allowed me to calculate an average rate and then figure out how many strikes above or below average each catcher was.”
Reviewing his findings, originally Turkenkopf thought he had miscalculated. The results were staggering. The catchers who caught at least 120 innings in 2007 showed a tremendous difference in the value they added or subtracted in their ability to frame pitches. Gregg Zaun led all catchers by saving 0.85 runs per 150 pitches. Gerald Laird cost his team -1.25 runs per 150 pitches. While Zaun led in the rate stats, the catcher who successfully framed the most pitches in 2007 was Russell Martin, who finished fourth in rate value, saving .63 runs per 150 pitches.
“I’ll be the first to admit this is a much larger effect than I expected to see. In fact it’s so large that I have to think there’s something wrong in the analysis,” Turkenkopf wrote. “Over the course of 120 games (a reasonable est
imate for the number of games caught in a season by a starting catcher), the difference between Gregg Zaun and Gerald Laird is over 250 runs or 25 wins.”
Turkenkopf’s findings were astounding, and while imperfect, he was on to something. He had discovered something remarkable. He’d proved certain catchers were better than others at manipulating borderline strike calls and could swing the probable outcome of an at bat dramatically in favor of the pitcher. During the 1990s and early 2000s, catcher defense had been undervalued by analysts. Wrote Turkenkopf, “Maybe we’ve been wrong the whole time and catcher defense is really that important.” Other analysts soon followed with attempts to refine the valuation of pitch framing. BaseballProspectus.com analyst Mike Fast found that the best catchers in 2011 were saving their teams 15 to 30 runs per season through pitch framing, and the worst catchers had cost their teams roughly 15 runs per season. Ranking second on Fast’s leaderboard was Russell Martin, who had saved 70 runs through pitch framing over the five years from 2007 to 2011; ranking last was Pirates catcher, Ryan Doumit. Doumit had caught more than 2,800 innings over those five seasons and had cost the club thousands of favorable counts—and 65 runs—due to his faulty glove. Doumit’s receiving skills were so obviously poor he earned the nickname Ryan “No-Mitt” in Pittsburgh. The five-year study showed a 135 run difference between Martin and Doumit. The Pirates’ 2012 catcher, Rod Barajas, was also a below-average pitch framer according to the study.
Consider the following chart:
Max Marchi, then a writer for Baseball Prospectus, published more PITCHf/x research on pitch framing, including studies on how the skill improved or declined with age, and found that the skill aged well.
Realizing the value these hobbyists brought to the game, all three, Fast, Marchi, and Turkenkopf, were hired by major league teams to become data analysts or baseball systems developers. One could argue that an analyst such as Fast, who could be procured for far less than the minimum wage of a major league player, was in fact more valuable than many of the players on the rosters. An analyst could identify undervalued players and produce millions in dollars of value for his club.
Wrote NBC baseball analyst Craig Calcaterra in a January 2013 blog post: “For as much crap as the sabermetic and bloggy types take from the mainstream media about how they don’t truly know the game because they’re not out there at the park or interviewing players in clubhouses and stuff, ain’t it funny how the sabermetrics and bloggy types are continually hired by major league teams to work in baseball operations? And did you notice that teams never hire the guys who claim to know so much more about baseball and who continually slam advanced metrics and statistical analysis?”
Just as the game had been closed for so many decades, open to only those who had played, the sport was still exclusive early in the twenty-first century. Only those that played, or had an Ivy League degree, were typically found in front offices. But with the advent of big data, of PITCHf/x, and the accompanying hobbyist research, anyone who showcased talent and creativity could gain entrance to a front office. Data had moved the game closer to a true meritocracy. The race was on for teams to hire the best and brightest hobbyists and to create proprietary metrics. For example, in 2008 Dan Fox—an Iowa State undergrad—was the only data analyst or architect employed by the Pirates baseball operation side: by 2013, the Pirates had five full-time employees devoted to data analysis, collection, and architecture.
With the creation of PITCHf/x, Sportvision had incidentally created a job field in professional baseball that hadn’t existed before: data-science departments.
That these hobbyists had access to the PITCHf/x data was unintentional on the part of Sportvision, whose CEO, Hank Adams said, “People get the data by scraping the site.… We don’t send it out. It’s not downloadable per se, but amateurs went in there and figured out how to scrape the [MLB Advanced Media] site. Now many of them have made it publicly available. You’re not supposed to commercialize this, but most of these guys are amateurs and hobbyists. We could theoretically shut it down, but at the same time we recognize that these guys have done a lot of great work to highlight the value of the data.”
With PITCHf/x’s millions of data points entering the game, by 2013 an arms race was on for analysts who could make sense of them. Clubs were after analysts who could craft their own meanings from data using complex algorithms. The data entering the game in the 2010s would be on an entirely different level of sophistication than anything before.
“If [teams] can only use a standard set of data [like PITCHf/x], finding the answers and asking the right questions is going to be key. Teams are hiring very smart people to ask those questions and find those answers,” said Ryan Zander, general manager of baseball operations for Sportvision. “We can create the data but not how it’s interpreted and how it’s used and analyzed. Data has gone from being something where maybe you had one person looking at it, to what we are seeing now, where data is being used as an everyday process for a front office. I think a lot of decisions being made are using data we’re creating, whether it’s player development or coaching, training players, or understanding the value of players.”
In 2012, the Pirates added their second full-time data quantitative analyst. He would become exponentially more valuable than what he was paid.
As a sophomore at MIT studying chemical engineering, Mike Fitzgerald read Moneyball. Like so many other mathematically inclined young people, he was awakened to the idea that professional-sports front offices had jobs for people like him, people who had once been outside the game. Fitzgerald wasn’t sure exactly what path he wanted to pursue, or how to enter the field, but he had a passion for sports and a strong mathematical mind. When he was a child, Fitzgerald’s mother would tote him along to their Boston-area grocery store. She would tell him what every item cost and offer an approximation of the tax.
“At the end I would give what the total bill would cost to a penny,” Fitzgerald said. “It used to blow her mind.… I knew I could make a career out of [mathematics] after my sophomore year in college.”
Like Dan Fox, what set Fitzgerald apart from some of his classmates at MIT was an ability to communicate complex mathematical ideas in simple, relatable ways, an example of which was an assignment that impressed one of his professors. Fitzgerald and his classmates were tasked with writing a ten-page paper that someone with a minimal mathematical background could comprehend. Fitzgerald did not choose an obscure subject; rather, he tackled one of America’s most popular sporting and gambling events: the NCAA Division I men’s basketball tournament. Fitzgerald’s project was how to optimize selecting teams in a tournament. Being a basketball fanatic, he examined a strategy for selecting teams in a traditional bracket pool, where points are awarded for correctly predicting winners in each round. He then examined another optimization plan for a bracket where participants are awarded additional points for picking upsets. For instance, if a No. 12 seed beats a No. 5 seed, and the participant had selected the No. 12 seed (the underdog), the participant was awarded the seed differential of seven points. Fitzgerald chose a topic that made the mathematical concept he was studying—conditional probability—accessible to a wide audience.
Conditional probability, simply defined, measures the probability of an event given that another even has occurred. Conditional probability led Fitzgerald to the Pirates’ most important free agent signing in club history in the 2012–13 off season.
“The notion that if we are looking at a situation and we bring new information into the situation, how does that change the possible distributions of outcomes you could have?” Fitzgerald said. “So I think my mind works pretty quickly through those types of things, which is cool, because baseball is a game where you are constantly getting more and more input data, which changes our initial thoughts. There is so much data entering the game.”
Fitzgerald is an enthusiastic Boston Celtics fan. As a sophomore at MIT in 2008, Fitzgerald and several of his cousins traveled from Boston to wat
ch the Celtics play the Pistons in Games 3 and 4 of a play-off series in Detroit. They split a $100 hotel room in suburban Birmingham, Michigan. Across the street the Pistons were housed in luxury accommodations. In a park near the hotels, Fitzgerald and his cousins were tossing a football on an off day when to their delight Celtics forward Glen “Big Baby” Davis approached them. Soon a pickup game between Davis, Fitzgerald, his friends, and several local Michigan kids emerged. The spectacle drew the attention of ESPN television analysts Jeff Van Gundy and Marc Jackson, who were staying at the Celtics’ hotel and came down to the park. Van Gundy had been a coach with the Houston Rockets, and although he was fired after the 2007 season, he remained friendly with Rockets GM Daryl Morey, who had brought analytics to the NBA.
“Van Gundy asks me if I had ever thought about getting into analytics in basketball,” Fitzgerald recalled. “I had no idea [analytics] were breaking into basketball. I said, ‘That’s interesting.’”
That conversation helped Fitzgerald land an unpaid basketball internship with the Celtics that fall. He then took a paid internship with the Danish company TrackMan, a burgeoning rival for Sportvision. TrackMan had made its name in golf, using radar to track the trajectory of golf balls. In 2009, TrackMan was in the early stages of an effort to use radar to track pitches and batted balls in baseball. They had begun testing their technology in three major league parks.
“Basically the goal was to clean the data and come up with some very elementary information for the clubs,” Fitzgerald said. “The biggest thing that jumped out to me was the effective velocity. TrackMan readings are basically the same as PITCHf/x except they measure the [entire] flight of the ball instead of picking up the ball at [twenty] different points starting at fifty feet … and they get pitchers’ extension.”
Pitchers’ extension was one of the primary reasons teams were interested in the TrackMan product, along with tracking bat speed and exit speed of balls in play. PITCHf/x gives teams a pitcher’s vertical release point, but it does not produce a horizontal release point, meaning, it doesn’t read how closely a pitcher is releasing a ball to home plate. This is important because if a pitcher releases the ball closer to home plate, his effective velocity can become greater. For example, a 93 mph fastball traveling 53 feet instead of 55 feet is effectively a faster pitch.