Exploring Outfielder Jump
Which aspect of an outfielder’s tracking performance is more likely to carry over into the following year, their burst to the ball or their reaction time? Does Ryan Braun’s quick reaction time or Max Kepler’s efficient routes give a better indication of overall fielding performance? These are some questions that I set out to answer in my study of outfielder jump data over the last couple of months. After outfielder jump data was added to Baseball Savant and Tom Tango featured graphs from that data on his Twitter, I was inspired to learn more. Using R, the statistical programming language, and Baseball Savant’s database, I set out to determine which part of an outfielder’s jump is the most “skill-like”, which best predicts an outfielder’s jump itself, and which aspect is the most relevant to actually making the play.
Before we get started, we need to understand outfielder jump terminology and how it works. On Baseball Savant, outfielder jump is measured by adding up feet saved against league average in 3 stages: reaction, the first 1.5 seconds after contact, burst, the next 1.5 seconds, and route efficiency over the first 3 seconds. It is only calculated on plays rated two stars or more, or plays with catch probability less than 90%. For more details, visit the website here.
When I loaded the data into R, I made a few changes to make studying it easier. First, I standardized (z-scored) every category I would end up using in the project, primarily to put everything on the same scale, but not just any scale, a scale that more people are familiar with. The fact that Ronald Acuna Jr. saved 0.4 feet vs average in the burst phase in 2018 may not resonate as much as knowing that his z score (amount of standard deviations away from the mean) was 0.622. Additionally, I put statistics that are reliant on volume like outs above average on a per play basis to prevent the amount of plays from adding uncertainty to the data. A link to all of my code is here.
After cleaning the data, I got to work and analyzed it. The first things I wanted to measure were the year to year correlations of each aspect of outfielder jump, as well as how the relationships would hold up over a long period of time.
Burst average r (correlation coefficient) = 0.701
Burst average R-squared ( how much of the variance can be explained by the model) = 0.493
Reaction average r = 0.853
Reaction average R-squared = 0.724
Route average r = 0.847
Route average R-squared = 0.715
With the highest correlations and the highest R-squared values, reaction seems to have the most carry over year to year, and functions the most like a defined skill. This notion makes sense, as reaction (and reaction time) is believed to be a skill, and tends to be thought of as one that is difficult to train and improve. Meanwhile, route also appears to act like a skill on a year to year basis, but its high correlation isn’t maintained when looking at a longer period of time. While it still could be a skill, the case for it isn’t as strong as reaction. However, burst seems to be the most unpredictable of the three. While the r and R-squared values are decent, they are low when compared to the other two phases of outfielder jump. One possible explanation is that as players age, their quickness and acceleration suffer.
After discovering those fascinating correlations, I turned my attention to finding out which phase of jump is best indicative of jump itself. Both axes are measured in feet saved vs. average (standardized).
Burst average r = 0.91
Burst average R-squared = 0.827
Reaction average r = 0.615
Reaction average R-squared = 0.374
Route average r = -0.164
Route average R-squared = 0.02
This time, burst has the highest correlation, and clearly predicts overall jump very well. Reaction has a positive overall correlation, but the points are more spread out and are further away from the regression line. Lastly, route actually has an inverse correlation with jump, meaning hypothetically, it would predict the opposite of jump better then jump itself.
Next, I set out to find which aspect of jump means the most with regards to actually making the play. I used OAA to represent an outfielder’s ability to make the play, but I also put it on a per play basis and standardized it.
Burst average r = 0.82
Burst average R-squared = 0.669
Reaction average r = 0.391
Reaction average R-squared = 0.148
Route average r = -0.101
Route average R-squared = 0.007
The relationships for aspects of jump with OAA wind up being similar to their relationships with total jump. Burst, with consistently high correlations, is a great indicator of how well an outfielder performed in a given year. Meanwhile, reaction can be rather inaccurate, but it is a far better predictor than route. Route should never be used to estimate overall outfielder production, considering its extremely low correlations, high standard errors and high p-values.
At this point in the study, it is worth considering whether any of the discoveries we have made matter. None of these findings would be useful or relevant if jump didn’t correlate well to making the play, or OAA.
Average r = 0.806
Average R-squared = 0.646
Thankfully for our purposes, jump is a great indicator of OAA. Clearly, Tom Tango did a great job building his model, as its consistency in predicting overall outfielder performance is remarkable.
From this investigation, three main conclusions can be drawn. First, performance during the reaction phase functions the most like a defined skill, with route relatively close behind and burst trailing both of the other two. Second, out of the three stats, burst predicts overall jump best with reaction close behind while route actually has a slightly negative correlation. And finally, all three stats project OAA similarly to the way they project overall jump, with burst being the best, followed by reaction and then route giving inverse indications.
These discoveries could be very useful for MLB players looking to improve their tracking abilities. First, even though burst doesn’t function like a trainable skill as much as reaction and route, it still has enough characteristics of one to be considered a skill. Additionally, per our findings, burst has an overwhelmingly positive correlation with OAA. With those notions in mind, outfielders should focus on training the muscles used in the burst phase as opposed to the instincts employed in the reaction and route phases in their offseason workouts. Simply by practicing the things that actually translate to success in games, outfielders can be much more productive.
Not only can players find the knowledge presented here useful, but front offices could use this information in their player evaluation and roster construction processes. One way in which analytics departments can use this data is in identifying players whose OAAs have been far below what their bursts would’ve predicted. These outfielders could be both undervalued and set to improve.
In spotting these players that are likely to regress up to their mean, teams could potentially exploit a market inefficiency. Players that fit that description could potentially be cheap targets that provide great return on investment. To determine how much a player was underperforming their OAA based on their burst, I created the discrepancy and average discrepancy metrics. Discrepancy is calculated simply by subtracting the standardized and rate based OAA from standardized burst, and the average is found by averaging a player’s discrepancy over several years. The graphic below features the ten players with the highest average discrepancy over the last two years.
Clearly, Ramon Laureano and David Dahl are in positions to have spectacular fielding performances in the 2020 campaign. One may notice that most of the players in the list above aren’t necessarily what we consider cheap targets, but most of them, like Mallex Smith and Kyle Schwarber, are known for their running or hitting, not their fielding. So, it is possible that their fielding isn’t recognized the same way that their other skills are.
Looking ahead to the future, there are a couple of extra things I would have loved to have done with this study but haven’t done yet that I could add to this project later. These include creating a model to predict OAA from the three aspects of jump, creating an algorithm to determine a player’s monetary value based on their burst, reaction and route, and implementing a machine learning model to predict gold glove award winners based on their jump statistics.
This article was originally made as a presentation/slideshow that can be viewed here
— Jake Federman
New York, NY
Email me at Jake_federman@horacemann.org or Jakefed1@hotmail.com
All data is updated as of September 14, 2019
All data is from Baseball Savant
Thank you to editors Aidan Resnick, Maxwell Resnick, Oliver Steinman and Richard Diamond, and other reviewers Kush Malhotra and Ryan Altman
Idea inspired by @Tangotiger on Twitter
Follow my Twitter @federmanjake
All rights reserved
