Proper after FanGraphs printed my piece on the Kirby Index, the metric’s namesake misplaced his contact. George Kirby’s trademark command — so dependable that I felt snug naming a statistic after him — fell off a cliff. Whereas the stroll price remained beneath management, the house run price spiked; he allowed seven dwelling runs in Could, all on pitches the place he missed his goal by a big margin.
Watching the namesake of my new metric flip mediocre instantly following publication was among the many many humbling experiences of publishing this story. However, I needed to revisit the piece. For one, it’s December. And writing the story led me down an interesting rabbit gap: Whereas I realized that the Kirby Index has its flaws, I additionally realized a ton about up to date efforts to quantify pitcher command.
However first, what’s the Kirby Index? I discovered that launch angles, in live performance with launch peak and width, virtually completely predicted the situation of a pitch. If these two variables advised you virtually the whole lot concerning the location of a pitch, then a measurement of their variation for particular person pitchers may theoretically present novel details about pitcher command.
This received a couple of folks mad on Twitter, together with baseball’s eminent physicist Alan Nathan and Greg Rybarczyk, the creator of the “Hit Tracker” and a former member of the Pink Sox entrance workplace. These two — notably Rybarczyk — took situation with my use of machine studying to make these predictions, arguing that my use of machine studying steered I didn’t perceive the precise mechanics of why a pitch goes the place it goes.
“You’re spot on, Alan,” wrote Rybarczyk. “The amazement that trajectory and launch parameters are strongly related to the place the ball finally ends up can solely come from individuals who see monitoring information as columns of digits fairly than measurements of actuality that replicate the underlying physics.”
Whereas the tone was a bit a lot, Rybarczyk had a degree. My “amazement” would have been tempered with a extra thorough understanding of how Statcast calculates the situation the place a pitch crosses dwelling plate. After publication, I realized that the nine-parameter match explains why pitch location may very well be so powerfully predicted by launch angles.
The placement of a pitch is derived from the preliminary velocity, preliminary launch level, and preliminary acceleration of the pitch in three dimensions. (These are the 9 parameters.) Launch angles are calculated utilizing preliminary velocity and preliminary launch level. As a result of the situation of the pitch and the discharge angle are each derived from the 9P match, it is smart that they’d be virtually completely correlated.
This led to an inexpensive critique: If launch angles are location data in a unique kind, why not simply apply the identical strategy of measuring variation on the pitch areas themselves? This can be a truthful query. However utilizing areas would have undermined the conclusion of that Kirby Index piece — that biomechanical information like launch angles may enhance the precision of command measurements.
Groups, with their entry to KinaTrax information, may create their very own model of the Kirby Index, not with implied launch angles derived from the nine-parameter match, however with the place of wrists and arms captured for the time being of launch. The Kirby Index piece wasn’t nearly creating a brand new method to measure command; I needed it to level towards one particular manner that the brand new information revolution in baseball would unfold.
However sufficient about that. It’s time for the leaderboards. I eliminated all pitchers with fewer than 500 fastballs. Listed here are the highest 20 within the Kirby Index for the 2024 season:
2024 Kirby Index Leaders
SOURCE: Baseball Savant
Minimal 500 fastballs thrown.
And listed here are the underside 20:
2024 Kirby Index Laggards
SOURCE: Baseball Savant
Minimal 500 fastballs thrown.
A number of takeaways for me: First, I’m so grateful Kirby received it collectively and completed within the prime three. Dying, taxes, and George Kirby throwing fastballs the place he needs. Second, the highest and backside of the leaderboards are satisfying. Cody Bradford throws 89 and lives off his elite command, and Joe Boyle — nicely, there’s a motive the A’s threw him in as a chunk within the Jeffrey Springs commerce regardless of his otherworldly stuff. Third, there are guys on the laggard listing — Seth Lugo and Miles Mikolas, specifically — who look misplaced.
Mikolas lingered across the backside of the leaderboards all yr, which I discovered curious. Mikolas, in spite of everything, averages simply 93 mph on his four-seam fastball; one would think about such a man would want to have elite command to stay a viable main league starter, and that league-worst command successfully can be a demise sentence. Complicated this additional, Mikolas averted walks higher than virtually anybody.
Why Mikolas ranked so poorly within the Kirby Index whereas strolling so few hitters may in all probability be the topic of its personal article, however for the needs of this story, it’s in all probability sufficient to say that the Kirby Index misses some issues.
An instance: Mikolas ranked second amongst all pitchers in arm angle variation on four-seam fastballs, suggesting that Mikolas is deliberately altering his arm angle from pitch to pitch, possible relying on whether or not the hitter is left-handed or right-handed. This is only one motive why somebody may rank low within the Kirby Index. One other, as I discussed within the authentic article, is {that a} pitcher like Lugo is likely to be aiming at so many various targets that it fools a metric just like the Kirby Index.
So: The Kirby Index was a enjoyable train, however there are some flaws. What are the alternate options to measuring pitcher command?
Location+
Location+ is the business customary. The FanGraphs Sabermetric library (an unimaginable useful resource, it have to be mentioned) does an amazing job of describing that metric, so I’d encourage you to click on this hyperlink for the complete description. The quick model: Run values are assigned to every location and every pitch sort based mostly on the depend. Every pitch is graded on the stuff-neutral areas.
Implied location worth
No one appears notably happy with Location+, together with the creators of Location+ themselves. As a result of every depend state and every pitch sort makes use of its personal run worth map to distribute run worth grades, it takes a brilliant very long time for the statistic to stabilize, upward of a whole bunch of pitches. It additionally isn’t notably sticky from yr to yr.
The latest model of Location+, which is able to debut someday within the close to future, will use an identical logic to PitchProfiler’s command mannequin. Basically, PitchProfiler calculates a Stuff+ and a Pitching+ for every pitcher, that are set on a run worth scale. By subtracting the Stuff+ run worth from the Pitching+ run worth, the mannequin backs into the worth a pitcher will get from their command alone.
Blobs
Whether or not it’s measuring the usual deviation of launch angle proxies or the precise areas of the pitches themselves, this methodology will be outlined because the “blob” methodology, assessing the cluster tightness of the chosen variable.
Max Bay, now a senior quantitative analyst with the Dodgers, superior the Kirby Index methodology by measuring launch angle “confidence ellipses,” permitting for a extra elegant unification of the vertical and horizontal launch angle elements.
Miss distance
The central concern with the Kirby Index and all of the blob strategies, as I said on the time, is the one goal assumption. Ideally, as a substitute of how intently all pitchers are clustered round a single level, every pitch can be evaluated based mostly on how shut it completed to the precise goal.
However targets are onerous to come back by. SportsVision began monitoring these targets within the mid-2010s, as Eno Sarris outlined in his piece on the state of command analysis in 2018. Lately, Driveline Baseball measures this working alongside Inside Edge. Inside Edge deploys human beings to manually tag the goal location for each single pitch. With these information in hand, Driveline can do a few issues. First, they created a Command+ mannequin, modifying the imply miss distances by accounting for the issue of the goal and the form of a pitch.
Utilizing supposed zone information, Driveline additionally exhibits pitchers the place precisely they need to purpose to account for his or her miss tendencies. I’m advised they are going to be producing this technique in a public put up quickly.
Catcher Targets (Pc Imaginative and prescient)
In an ideal world, computer systems would substitute human beings — wait, let me attempt that sentence once more. It’s costly and time-intensive to manually observe targets via video, and so for good motive, miss goal information belong to those that are prepared to pay the worth. Pc imaginative and prescient methods current the potential to supply the info cheaply and (subsequently) democratically.
Carlos Marcano and Dylan Drummey launched their BaseballCV challenge in September. (Drummey was employed by the Cubs shortly thereafter.) Joseph Dattoli, the director of participant growth on the College of Missouri, supplied a contribution to the challenge by demonstrating how laptop imaginative and prescient may very well be used to tag catcher targets. The one limitation, Joseph identified, is the computing energy required to comb via video of each single pitch.
There are some potential issues with any command measurement depending on goal monitoring. Targets aren’t all the time actual targets, extra like cues for the pitcher to throw towards that normal route. However Joseph will get round this concern by monitoring the catcher’s glove in addition to his middle of mass, which is much less prone to those kinds of dekes. Nonetheless, there’s a method to go earlier than this methodology scales right into a kind the place each day leaderboards are accessible.
The Powers methodology
Absent a raft of public details about precise pitcher targets, there as a substitute will be an effort to simulate them. In his 2023 presentation, “Pitch trajectory density estimation for predicting future outcomes,” Rice professor Scott Powers proposed a technique to account for the random variation in pitch trajectories, within the course of providing a framework for simulating one thing like a goal. (I’ll possible butcher his strategies if I attempt to summarize them, so I’d encourage you to look at the complete presentation when you’re .)
The Powers methodology was modified by Stephen Sutton-Brown at Baseball Prospectus, who used Blake Snell for instance of the best way these focusing on fashions will be utilized at scale to evaluate particular person pitchers. First, Sutton-Brown match a mannequin that created a world goal for every pitch sort, adjusting for the depend and handedness of every batter. Then, for every pitcher, this international goal was tweaked to account for that pitcher’s tendencies. Utilizing these simulated targets, he calculated their common miss distance, permitting for a separation of the run worth of a pitcher’s targets from the run worth of their command skill.
“Nothing”
On Twitter, I requested Lance Brozdowski what he noticed because the gold customary command metric. He answered “Nothing,” which sums up the issue nicely. This can be a difficult query, and all the present strategies have their flaws.
There are methods that the Kirby Index may very well be improved, however so far as I can inform, one of the best ways ahead for public command metrics is a few form of mixture of the ultimate two strategies, with energetic monitoring of the pc imaginative and prescient developments to see if constant targets will be established.
However one would think about the story is totally totally different on the group facet. By marrying the KinaTrax information with miss distance data, these strategies may probably be mixed to make some form of tremendous metric, one which I think about will get fairly near measuring the true command skill of main league pitchers. (In a video from Wednesday, Brozdowski reported on a number of the potential of those information for measuring and bettering command, in addition to their limitations.) The general public may not be fairly there, however so far as I can inform, we’re not that far off.