Pinned Tweet
I'll be using this account to publicly share the results of OpenCommand, a computer-vision based command tracker.
Ideas in mind:
- MLB command leaderboards
- Target maps
- Pitcher command trends
- Pitcher tactics changes
- Reply to this tweet if you have ideas!
Some news:
I decided to open source my command project. It's called OpenCommand.
This means anyone can see the code (and even the data)!
I hope this gives more people a chance to pursue and share cutting edge baseball research.
Anyways, here's how to measure command: 🧵
1/N
4
31
11,652
PSA: Open Command is at the GM meetings!
We're offering a wide range of services to teams, agencies, & players.
DMs & email open for inquiries
This was also kind of wrong because there IS a steep hockey stick when you look at RA9 instead of xERA.
The minimum command you need to reliably be in the big leagues is around a 12-inch mean miss!
As @pitchprofiler & others mentioned, maybe there's some "minimum command" to be a viable pitcher (throw enough strikes).
I managed to find *some* hockey stick going on but it looks like it's driven by K% dropping (not BB% going up disproportionately).
It's hard to show in MLB because these pitchers already proved they could throw strikes in the minors!
1
14
1,536
Yeah so command is correlated to RA9 - xERA but stuff isn't
This is the most absurd result I've ever seen
So turns out command is worth TWICE as much. Using xERA was a mistake and underestimated its importance.
On RA9, command is 2/3 as important as stuff.
Anyone have an idea on why such a big gap? I've always considered xERA to be RA9 but faster. But it's missing something?
Parker Messick had his 2nd worst command day of the year yesterday.
He dropped from 80th to 12th percentile command, walking 4 batters in 2.1 innings.
3 things are true at the same time:
- Command is worth 1/3 Stuff
- Command stabilizes almost as fast as Stuff
- Adding Command on top of Stuff doesn't dramatically improve ERA prediction
Why? 🧵
1/3
The problem is that command's influence is observable *conditional* on stuff. (Raw command is uncorrelated to ERA)
It's a huge survivorship bias (good stuff guys get away with bad command) - there's a negative correlation between stuff and command.
2/3
2
7
1,089
As @pitchprofiler & others mentioned, maybe there's some "minimum command" to be a viable pitcher (throw enough strikes).
I managed to find *some* hockey stick going on but it looks like it's driven by K% dropping (not BB% going up disproportionately).
It's hard to show in MLB because these pitchers already proved they could throw strikes in the minors!
For reference, going from worst stuff in the league to best is worth 3 ERA.
So, stuff is arguably 3x as important as command (at the MLB level)!
Going from worst command in the league to best is worth 1 ERA!
There's a strong correlation between xERA (unexplained by Stuff+) and command.
And it gets really interesting because some pitch types are way more dependent on command than others (see below)
Fastballs are almost twice as command dependent as Breaking balls. Anyone have an idea why?
2
6
1,708
Note: This wasn't obvious before I managed to remove location leakage from Stuff+
teapaw.org/tomdoyo/status/2097322…
The biggest remaining flaw with stuff models is that they're secretly location models.
Contrary to popular belief, there's a trivial way to adjust for this. 🧵
1
1,512
OpenCommand might actually be the most accurate command tracker in the world.
It estimates intended targets with the lowest error!
(TL;DR at the end)
🧵
TL;DR
Measurement errors increase avg estimated miss, so lower avg estimated miss is a good indicator for small measurement error.
OpenCommand's target estimation is about 4in (or less) away from true target each pitch.
This is among the best in baseball, despite lacking advanced scouting and situational adjustments.
1
1
157


















