It’s a term in baseball that some people have trouble defining, and yet everyone knows it when they see it.
Everyone knows that this is what elite fastball command looks like. Just like everyone knows that this is what elite breaking ball command looks like.
It’s different from control; control is simply the ability to throw strikes. A decent indication of a pitcher’s control can be found in his BB%. It isn’t perfect but it gets the job done.
Command is, more or less, putting the ball where you want. All strikes are considered equal in the box score, but a strike right down the middle and a strike on the edge are not equal in-game. One pitch - the one down the middle, in this case - is likely easier to hit, and more often than not, that pitch right down the middle was not where the pitcher intended it to go. The ability to pick and choose your spots smartly and with consistency - that’s command. It’s often considered a subcategory of control itself.
That “pitch intention” - that is, where the pitcher intends for the ball to go - is a huge part of the reason that baseball and sabermetricians and front offices may never have a firm grasp on quantifying command. Unlike many other baseball stats today, there is an inherent human aspect that is impossible to eliminate.
In May of 2018, Eno Sarris of the Athletic tried his hand at quantifying command with a stat in which he dubbed Command+. It's a fantastic article, idea, and metric that I absolutely recommend that you read. Command+ attempts to quantify pitcher intent using the pitch type and where ideally that pitch should be according to historical uses, an idea developed by Kyle Cunningham-Roads, Greg Gifford and Adam Whitehouse at STATS.
Yet, while we know that command is a combination of intent and location together, and the methodology in determining intent in Command+ is sound, perhaps we’re better off ignoring something remarkably hard, maybe even borderline impossible to quantify such as intent. So that’s what I will be doing - focusing on pitch location and ignoring the “intent” altogether, to produce a Location Runs Prevented (LRP) for each pitcher. Let’s see what kind of results are generated.
How It Works
Each pitch crosses the plate at a certain point, and now with the help of StatCast, we can measure the X and Y of these locations better than ever. Baseball Savant has broken them down in what the minds at Baseball Savant at called “Attack Zones”, which are compartmentalized as shown below:
All 39 zones have their own number, but they can also be grouped into 4 major groups that are defined by the batter’s strike zone - the heart, shadow, chase, and waste zones.
Depending on the count, some attack zones are “better”, or more preferable than others. Most pitchers are likely going to aim for the shadow zone (aka the edge of the strike zone) a majority of the time, and maybe the Chase zone when ahead in the count. However, we know that pitchers miss their spots quite frequently and there are quantifiable differences between pitches in different zones.
LRP sets out to quantify these differences in every single pitch by analyzing 3 major components - the count, the attack zone, and the pitch type.
The count was one of the “core 2” variables I chose to analyze when I began this project along with attack zone, as it plays a crucial part in what we want to quantify.
Every pitch influences an at bat in either a positive or negative way; that goes without saying. That magnitude of that influence is heavily dependent on the current count, however. A strike is a strike in the box score, always, but a strike in a 0-0 count and a strike in a 0-2 count are not created equal in terms of net wOBA. Here are the wOBA’s on all counts in 2019:
So the difference between an 0-1 count and an 0-2 count is -0.058 wOBA for the pitcher while the difference between 2-2 count and a 3-2 count is +0.087 wOBA for the pitcher. A strike in a 2-strike count is recognizably huge (final wOBA drops to 0.000) just like a ball in a 3-ball count is huge (wOBA = 0.69). All of these net wOBA’s can be converted into run totals and accumulated, which is the basis on which LRP is designed.
The Attack Zones
In this project, the attack zones are the tool that is used to determine what the chances of those advancements are. For the purposes of example, I’m going to isolate a fastball in a 1-0 count.
We see from the pitch count wOBA expectancy chart above that a 1-0 count has a wOBA of 0.332. A strike moves the needle to 0.311 (- 0.021 wOBA) and a ball moves it to 0.362 (+ 0.030 wOBA). The chances of a strike or a ball or a ball hit into play are determined by the Attack Zone.
Here are the % chances of each of those outcomes (for fastballs) by each attack zone, including wOBA on contact:
Strike% is both called strikes and swinging strikes, which is why the Strike% for the Waste zone is not 0%. As you probably noticed, the Shadow zone was split into 2 zones, Shadow In and Shadow Out. The Shadow zone had a high number of total pitches relative to the rest of the zones, which confirms the earlier hypothesis that pitchers target that zone the point. Splitting the zone allowed for each to have a similar number of pitches to the other zones, and, though it may be risking too much granularity, there are notable differences between the two Shadow zones, so I feel confident this was the right choice.
So, for each pitch in our data, we have a count and an attack zone where the pitch was located. Example: for a 1-0 fastball in the Chase zone, our net runs would look something like:
(17.5% strike x -0.021 net wOBA from 1-0 to 1-1) +
(78.3% ball x 0.030 net wOBA from 1-0 to 2-0) +
(4.2% batted x 0.286 wOBAcon in the Chase zone) = + 0.032 wOBA for that pitch
On average, a fastball in the Chase zone on a 1-0 count will result in a +0.032 wOBA for the hitter, which is good for the hitter and bad for the pitcher, obviously. Every pitch gets analyzed this way, adjusted, converted into runs, and totalled by pitcher. The result is total Location Runs Prevented.
Originally, LRP was going to just consist of count and location. Then, as I began to bounce the idea off of various Six Man Rotation minds (just to make sure it wasn’t a stupid endevor), Jack Boulia said he had had ideas that overlapped with this idea in the past and that he thought pitch type was important in something like this. I figured I might try to include pitch type in future iterations, but upon sleeping on it, Jack was right.
Pitch type is too important to omit in an exercise such as this and couldn’t have been left out, even for a first iteration. This is the reason I specified “fastball” in the 1-0 count in the example above. Pitches in all zones are not created equal. Everything changes with pitch type. Depending on the pitch and the attack zone, rate of a strike rate, rate of a ball, and quality of contact change in one way or another, altering the net runs of each pitch.
So, in order to not reduce sample sizes too small, pitch types were broken down into 3 classifications - fastballs, breaking balls, and offspeed pitches. I may experiment with separating the “breaking” classification into a curveball family and a slider family in the future, but for now, they remain together.
Hopefully by now, the objective and methodology of LRP is clear; depending on count, attack zone in which the pitch is thrown, and pitch type, a “net wOBA” is derived, adjusted a little, and converted into net runs - which are very small on an individual basis - and then totalled for each individual pitcher.
LRP is the total Location Runs Prevented. LRP/M is on a per 1000 pitches basis to put everyone on the same scale (M is the Roman Numeral for 1000, so this says the same thing as LRP/1000). Here are the LRP/M leaders for both SP and RP:
You’ll notice that the RP have extreme high and extreme low LRP/M marks when compared to the tail ends of the SP leaderboards. This is because of the small number of pitches thrown by the relievers compared to the starters. Think of it as 1 month of hitting vs. 3 months; of course the batting leader after April is going to have a much higher mark than the leader in July. We know that the April marks do not show a “true” talent level.
Some testing has shown that the “reliability” point of LRP/M is somewhere around 2000 pitches, so from year to year, RP usually don’t hit that mark. Even with the small sample size for those players, LRP/M for all pitchers with 500+ pitches seems to be roughly normally distributed, which is a positive:
Might be slightly skewed right, but if all of these pitchers were up around 2000+ pitches, I would assume this would shift back to the left.
Being a location based metric, a large part of pitcher command lies with in it, and, since command is a subset of control, we should see some relationship between LRP/M and walk rate. We really don’t want too much, since LRP doesn’t paint the full picture of command, and command is not control, but if they are completely unrelated, we would have to address that.
Seems like a decent enough relationship. As expected, the spread of SP (who have thrown more pitches) is tighter than that of RP, but we would expect those to converge with a larger sample for the RP. So this looks fine.
Year Over Year
Like all stats/metrics, we want some year over year correlation. There should be some variance, as no one stat should have a perfect correlation, but it’s hopefully low enough to show that the metric is “working” and maybe even show that there is some “skill” behind how/when pitchers locate.
The 2018-to-2019 correlation of all pitchers past that rough, 2000-pitch reliability point (n=56) seems is pretty good. The correlation and fit (r^2 = 0.75) is strong while having some outliers, which is good, also! Players improve all the time and these numbers shouldn’t be totally static year to year.
Kyle Hendricks and Actual Command
After seeing Kyle Hendricks name at the back end of both the 2018 and 2019 LRP/M lists, I felt as though I needed to address him specifically and what he represents in this whole project.
When people think of Kyle Hendricks, the first thing that comes to mind (or maybe second after the changeup) is his “command”. He ranked #1 on Sarris’ Command+ metric in 2018, and he represents the biggest discrepancy between Command+ and LRP by a good amount. Other names on the top of the Command+ leaders from 2018 such as Tanaka, Nola, Mahle, Jordan Zimmerman, Alex Wood, deGrom, etc. graded out equally well in LRP, but for whatever reason, Hendricks was a huge disagreement.
It serves as a good reminder that LRP doesn’t paint the full picture of command - not totally, anyway. It hyper-analyzes the location aspect of command, but omits intention altogether. I (with some bias) personally think it does a great job of doing what it sets out to accomplish, but would be remiss if I didn’t mention the flaw of lack of intention. For someone like Kyle Hendricks, maybe location alone can’t possibly tell the whole story. Maybe Kyle Hendicks will deliver a pitch on an 0-2 with the full intention of throwing it middle-middle because he knows what a hitter will do better than we do.
All of this is very possible.
That said, the average net runs from his pitch locations tell a different story, as Hendricks has continuously thrown his pitches in less than ideal locations for a few years now. Maybe it’s by design, knowing full well he has good sequencing and can throw stuff with good movement and location matters less. Command+ sure finds that to be the case. Either way, LRP/M thinks he’s been doing himself a disservice with the location he’s chosen. Maybe he doesn’t have the command that we think? Or maybe his “misses”, however rare, are worse than the average pitcher.
All-in-all, I’m happy with both the project and the results. It has it's clear flaws, including but not limited to
- Omission of pitcher "intention"
- Certain pitches from certain pitchers being harder to hit than others
- Lack of batter handedness
However, Hendricks aside, a lot of the names at the top and bottom of the leaderboards make intuitive sense. I’ll be updating the leaders semi-frequently and adding some additional tabs as I go, as I have been with pCRA, mostly because I find the results interesting. Hopefully, this idea can stimulate some similar/additional ideas with regard to command and location from people that are much smarter than I, and hopefully LRP can be a helpful small step in that effort.