Using Data Analytics to Improve Pitcher Performance in Major League Baseball

The game of baseball places exceptionally strong importance on statistics. Fans use statistics to better engage with the sport, the media to tell a story, and teams and players to gain an advantage in the never-ending arms race between batters and pitchers. New technology and data capture methods have led to the arrival of new statistics, technology and big data in pitching. Our analysis examined data gathered from TrackMan, a recently implemented technology that collects data from 30 Major League Baseball (MLB) teams and their more than 150 associated minor league affiliate teams, and further introduces novel metrics for the evaluation of pitcher performance.

First, we observed that fastballs are more complicated to analyze than simply measuring speed. The myths of the crafty “junk-ball” pitcher and the artless “fire-ball” pitcher stereotypes are unfounded. The most elite fastballs combine both speed and movement, not just one or the other. Further, we found that a player can improve in one of these dimensions without sacrificing the other, challenging conventional wisdom. What once was a “game of inches” is now a game of seemingly impossibly small measures of distance, position, angles and movement, imperceptible to the human eye, but not the latest technology. These measures will prove critical in the assessment of pitchers.

Second, we used K-means clustering to sort pitchers into cohorts based on statistical similarities. Year-over-year data allowed us to identify which cohorts possessed elite qualities and which were ripe for improvement. The insights from this finding direct teams on how to train current pitchers to maximize their effectiveness at a previously unattainable level of pitch performance. 

Data & Methodology

To perform our analysis, we used data collected from TrackMan, a 3D Doppler radar system that measures the location, trajectory and spin rate of pitched and hit baseballs. The technology captures 27 unique measurements of a pitch, including pitch release data, such as a pitcher’s extension and release point, as well as pitch movement data, such as velocity, vertical and horizontal movement, and the ball’s in-flight spin axis. Most importantly, as the TrackMan data is collected, the software can automatically define pitch type (e.g., four-seam fastball, two-seam fastball, curveball, slider). We focused the majority of our analysis on four-seam fastballs, as they are the most commonly thrown pitch and had the largest sample size.

Our dataset contained the details of every pitch thrown in a major or minor league baseball stadium since the beginning of the 2016 season through the 2019 season. In addition to the pitch attributes collected by TrackMan, our dataset included a calculation of pitch quality for each pitcher, which was determined by combinations of specific attributes. This allowed us to track a player’s improvement over time and at the level of the physics of the pitch.

Finding No. 1: Relationship between break and velocity. There are two major variables that determine the quality of a fastball – movement and velocity. Movement, otherwise known as break, is conventionally defined as how a pitch moves across the plate and up and down the strike zone. These two facets, known as horizontal break and induced vertical break, are measured as the distances between the actual trajectory of a pitch and the expected trajectory of a pitch if it was thrown straight from the pitcher’s hand into the catcher’s glove. Velocity is much more easily defined – it’s simply the speed at which the pitch is thrown.

With fastballs, there is a positive, moderately strong correlation between velocity and pitch quality (correlation = 0.664 for four-seam and 0.586 for two-seam). This finding suggests that pitchers who increase their velocity will see positive returns on their pitch quality, regardless of which fastball type is thrown. Regarding movement, the path to pitch quality improvement isn’t as clear cut.

First, for each type of fastball, there was an inverse relationship between how components of break impact pitch quality. For example, in four-seam fastballs, a higher amount of vertical break and lower amount of horizontal break led to a higher pitch quality. Pitch quality and vertical break were positively correlated at 0.155, while pitch quality and horizontal break were barely negatively correlated at -0.027. The inverse relationship was true for two-seam fastballs – pitch quality and vertical break were negatively correlated at -0.284, while pitch quality and horizontal break were barely positively correlated at 0.0246. But the most interesting finding about break and its impact on pitch quality was that unique break profiles (unique combinations of horizontal and vertical break) led to the highest pitch qualities, whereas generic break profiles (an average value of each break coefficient) led to lower pitch quality scores. In baseball terms, pitchers who excel at throwing fastballs with either extreme amounts of vertical or horizontal break will perform better than pitchers with average break.

In baseball, there’s a perception that pitchers seeking to improve the quality of their fastballs must either focus on increasing velocity or increasing movement, at the expense of the other trait. However, this perception does not reflect reality, as we found in the data. The data showed no such trade-off between vertical break and velocity and only a small trade-off between horizontal break and velocity. In fact, velocity and vertical break were slightly positively correlated at 0.066, while velocity and horizontal break were slightly negatively correlated at -0.103. This finding suggests that pitchers looking to improve the quality of their fastballs can do so by improving either break or velocity without worrying about erosion in the other pitch facet.

Finding No. 2: K-means clustering and machine learning. Acquiring proven pitchers is expensive because teams are required to either spend significant money to attract players to their team in free agency or give up valuable assets in exchange for elite pitchers in a trade. Therefore, developing talent in-house has traditionally been the most efficient use of a team’s resources, especially for cash-strapped teams. But to develop prospects, organizations need to correctly identify pitching prospects that are likely to see overall pitch quality improvement by altering facets such as break and velocity. By identifying prospects that stand to gain from modifying their pitches, teams can tailor individual player development to maximize the likelihood of reaching elite performance.

K-means clustering was used to algorithmically identify players in the data with similar pitching characteristics. These clusters were organized into deciles according to each pitcher’s aggregate break and velocity grades. Using these decile rankings, pitchers were identified for potential improvement. For example, pitchers with exceptional fastball velocity but a lower score on vertical break were deemed “vertical break improvement candidates” so long as those pitchers were within the average rate of improvement to the next decile. This methodology identified pitchers poised for velocity improvements.

The utilization of machine-learning techniques will allow for the systematic identification of players who are likely to be successful based on their pitch characteristics. As more data is accumulated and analyzed, we anticipate that machine-learning driven approaches will identify optimal player development plans, inclusive of pitching technique and the role of break and velocity, while also recommending pitches to add to a pitcher's arsenal. 

Conclusions

Using advanced computerized data collection and statistical techniques, we demonstrated that the perceived trade-off between fastball movement and velocity did not reflect reality. The implication this has on pitcher development is significant – pitchers can focus on improving either velocity or movement without sacrificing erosion in the other pitch component. Similarly, we found that unique break profiles on fastballs are more effective for pitchers than generic ones, meaning that an individualized approach to pitcher development would be beneficial in generating pitch quality improvements. The use of machine-learning techniques to identify opportunities for player development will direct baseball organizations on more targeted approaches to pitcher development.

Over time, as the MLB collects more TrackMan data and as the technology is spread throughout more levels of the sport, the predictive value of the data will improve. We anticipate more widespread adoption of the evaluation of pitches at such detailed levels and a new lens into the development of pitchers.

Acknowledgment

This work was made possible by contributions and industry expertise from Jesse Smith and the Seattle Mariners analytics department.