THE BOOK--Playing The Percentages In Baseball

Yeah this stuff is awesome.  I followed up on that Wandy 40+ pitches in an inning look and found very similar results to Josh’s albeit looking at different aspects.  It seems that the pitchers who throw the hardest are affected the most by the long-pitch innings.  They rely on the velocity, not movement, to make their heater successful, and are more likely to use their fastball more often in the inning/game.  After the inning, they take a big hit in their velocity, though it sustains itself as the rest of the game progresses; the movement, however, steadily decreases.

So we have hard-throwing guys, 93.5+ on average, with so-so or below average movement, losing the velocity that makes the straighter pitch effective.

It’s also possible that one fastball type will bleed into another fastball type as the game progresses, based on the pitch data collected.

The result is that a two-seamer may fall out of the analysis of the two-seamers and instead make its way into the four-seamers.

So, as another parameters in terms of distinguishing pitches, you would need to know the pitch number.  The clustering technique therefore is not static for each pitcher, but dynamic based on whether it’s the 10th pitch or 90th pitch of the game.

In these kinds of analysis, you have to be very careful that you don’t discard or include an extreme piece of data that simply doesn’t belong for that bucket.

Most people can’t distinguish between four-seamers and two-seamers in the PITCHf/x data, Tango.  It’s only the people doing spin analysis, which I believe is Dan Brooks and I and occasionally Harry Pavlidis, really have a shot at telling the four-seamers from the two-seamers with any accuracy.  The spin axis should be fairly invariant under fatigue.

Mike,

I am curious if you think my pitch classification is really doing a poor job at identifying sinkers (two seamers)?  Yes it throws in the submariners in with the sinkers when they are four seamers but the action of the pitch clearly fits better with the sinker group than the fastball group.  Isn’t it more important to classify the pitch as the hitter sees it anyway?

I am curious if you think my pitch classification is really doing a poor job at identifying sinkers (two seamers)?  Yes it throws in the submariners in with the sinkers when they are four seamers but the action of the pitch clearly fits better with the sinker group than the fastball group.

Roy Oswalt throws a four-seamer and a two-seamer in roughly equal proportions.  You just show the four-seamer.

Miguel Batista mainly throws a two-seamer; you show only a four-seamer.

I’ve debated quite a bit about what Kevin Millwood actually throws, but my best guess is that it’s a two-seamer, with maybe an occasional four-seamer.  You show only a four-seamer.

Josh Beckett throws a four-seamer about twice as often as he throws a two-seamer; you show only a four-seamer.

Erik Bedard, I agree with you that he throws a four-seamer but not a two-seamer.

Joba Chamberlain added a two-seamer to his four-seamer this year; you show only a four-seamer.

Bartolo Colon throws a four-seamer and a two-seamer in roughly equal proportions.  I don’t see a 2008 card for him, but I agree with your 2007 card.

Kelvim Escobar threw a four-seamer and two-seamer in roughly equal proportions last year.  Your 2007 card shows only a four-seamer.

Eric Gagne, I agree with you that he throws a four-seamer but not a two-seamer.

Odalis Perez, I agree with you that he throws a four-seamer but not a two-seamer.

Jon Lester throws a four-seamer and a two-seamer in roughly equal proportions.  You show his four-seamer but call his two-seamer a splitter.

Yasuhiko Yabuta, I agree with you that he throws a four-seamer but not a two-seamer.

So anyway, that’s a sample of the first dozen pitchers I looked at.  Eight of them throw the two-seamer, of which you only identified Colon’s.

Isn’t it more important to classify the pitch as the hitter sees it anyway?

(1) That depends on what you want to do with the data, and (2) do we have any reliable information about how hitters see different pitches?

Josh, Mike & others - my pfx database is under construction, one major thing that is missing so far is the ability to classify pitches.

Are these algorithms proprietary, or do you share?

I’m fairly good at SQL, and have a grasp of the concepts, so I can do my own with enough time, but wanted to see if code was available.

I am planning an article on batter’s plate discipline, looking to see if some batters are better than others at recognizing the pitches that they have historically hit better, and then being able to put them in play.

In this I would agree with Josh’s last comment, in that I think what the batter percieves in location, speed and movement is more important than what the pitcher calls the pitch

1. see pitch
2. think “I can hit this well”
3. swing

There are two main objectives:
1. We want to know what the pitcher threw, and therefore we need to know:
a) pitch speed
b) spin axis (angle)
c) spin speed (rpm)
d) release vector (point and direction)

2. We want to know what the batters (thinks he) sees, so we need to know:
a) distance of ball to react (reaction distance)
b) incoming vector at reaction distance
c) incoming vector when crosses the plate
d) spin axis and spin speed when crosses the plate

I think if you have that, you have the information the pitcher needs to be able to throw a pitch, and the information a batter needs to be able to hit a pitch.

Am I missing anything obvious?

Brian, my automated algorithm is a work in progress.  Right now it gets about 80% of pitches correct.  I don’t find that terribly useful for most purposes, so I do most of my classification for real analysis and research by hand. 

Josh could maybe tell you if he’s measured the accuracy of his algorithm, but my feel is that it’s in the same ballpark as mine.

Regarding what the hitter sees, I’ll ask again, how do we know what the hitter sees?  If we have some reliable information about that, that would be a huge boon.  I guess what I’m saying is, how do you know that location, speed, and spin movement are what the hitter perceives?

I suspect that it is not.  I suspect, for one thing, that the hitter does not separate gravity from spin movement.  I also suspect that the hitter relies more on spin clues (e.g., the red dot of a slider) and trajectory clues (e.g., the hump of a curveball) than he does on the ability to perceive spin movement directly.  I have a feeling this is why pitches like Brian Bannister’s fastball are effective.

I also suspect this varies from hitter to hitter and depends on which pitches the pitcher throws.  But I don’t know any of this directly, and if anybody does, I’d like to know.  It would be a huge piece of the puzzle of figuring out what makes a pitch effective.

Mike-
“Joba Chamberlain added a two-seamer to his four-seamer this year; you show only a four-seamer.”

Okay, I had a question about this pitch, so I’m glad you brought it up.

Here is a quick look at his stuff from last start:
http://brooksbaseball.net/pfx/index.php?month=6&day=3&game=gid_2008_06_03_tormlb_nyamlb_1/&pitchSel=501955.xml&prevGame=gid_2008_06_03_tormlb_nyamlb_1/&prevDate=63
http://brooksbaseball.net/pfx/index.php?month=6&day=8&game=gid_2008_06_08_kcamlb_nyamlb_1/&pitchSel=501955.xml&prevGame=gid_2008_06_08_kcamlb_nyamlb_1/&prevDate=68

In his first start (against TOR) he throws 3 pitches that look a lot like 2seam fastballs. >90ish mph, 10-12 inch horizontal break, 240ish spin angle. It’s definitely not his normal 4seam fastball.

And then in his second start (against KCA), this pitch sort of disappears. Instead, he throws a changeup that breaks about the same amount but is a good 10mph slower than what appeared to be his 2seamer.

What happened to this pitch? Why did he go 80 pitches without throwing it? It’s like in one start he invents a 2seam fastball and then in start #2 he invents a changeup.

Tango-

Sorry for the double post.

2. We want to know what the batters (thinks he) sees, so we need to know:
a) distance of ball to react (reaction distance)
b) incoming vector at reaction distance
c) incoming vector when crosses the plate
d) spin axis and spin speed when crosses the plate

I think if you have that, you have the information the pitcher needs to be able to throw a pitch, and the information a batter needs to be able to hit a pitch.

Am I missing anything obvious?

Perception is way more complicated than that.

Hitting a baseball is a perceptual/motor task that literally occurs faster than 99.9% of people can correctly identify the color of a square. Really. I can bring people into the lab, show them a square, ask them to identify the color of it with one of two easily made button presses, and tell them they are rewarded with money if they beat their previously fastest time. And normal, able-bodied people are still slower to do this task than hitters react to hitting a baseball.

I think even if we know all of those things, we still have to know about perception. How does tracking the ball actually work? What perceptual illusions factor into the success of certain pitches? What heuristics do hitters use when they look for particular pitches? How do their timing mechanisms work?

I think there are lots of unanswered perceptual questions here beyond the physics of the thing.

Dan,

Those three two-seamers against Toronto were the ones I was thinking of, although I believe he has thrown the pitch in other games this year, too.  I don’t have my data in front of me.

Apparently he hasn’t used his changeup much this year, either.  In Josh’s data, it only shows up four times (the group 80-85 mph and with -10 to -15 pfx_x, labeled as sliders):
http://baseball.bornbybits.com/2008/Joba_Chamberlain.html

I looked through some of the images from his first start, and all I could see were his four-seam fastball:
http://www.daylife.com/photo/07h6egR7WHequ/New_York_Yankees
and what I believe is his slider:
http://www.daylife.com/photo/043Y6PtcnT17E/New_York_Yankees
Of course, rarely-thrown pitches are going to be harder to find that way.

I suspect, for one thing, that the hitter does not separate gravity from spin movement.  I also suspect that the hitter relies more on spin clues (e.g., the red dot of a slider) and trajectory clues (e.g., the hump of a curveball) than he does on the ability to perceive spin movement directly.

I didn’t say anything about spin movement.  I’m the last person who would champion the separation of gravity from the hitter’s perspective.

The spin speed (rpm) would be the spin clues as you are saying, along with the spin axis.

I think there are lots of unanswered perceptual questions here beyond the physics of the thing.

Maybe I’m not being clear enough.  If we had that information, we can then write a simulator to track that information from the batter’s perspective, and then we can see how the batter responds.  The information I provided in #2 was the “inputs” (which act in combination, natch) to which the batter will respond.

I’m not suggesting we write a simulator, but what I am suggesting is capturing enough input that that’s what you would need to write the sim.

I didn’t say anything about spin movement.  I’m the last person who would champion the separation of gravity from the hitter’s perspective.

I was responding to Brian/#6.  I didn’t see your post #7 until I posted #8.  I think most people who look at the PITCHf/x data don’t even consider gravity, so I wanted to be clear that I think gravity is an important effect to consider.

The spin speed (rpm) would be the spin clues as you are saying, along with the spin axis.

Yes, but it may (probably isn’t) as simple as saying that the hitter sees the spin axis and spin rate.  He can probably see obvious things about the spin rate to identify a very slowly rotating forkball or a knuckleball, for example, but I doubt he can the difference between a fastball spinning at 40 revolutions per second and a changeup spinning at 30 revolutions per second. 

Whether or not he can see the spin axis probably depends on the orientation of the ball in the pitcher’s hand and thus how the features (seams and white spaces) on the ball are rotating.  The simplest case of this is the red dot slider when the spin axis goes through or near a seam.  But I imagine there are other similar clues on some pitches but not on others.

Maybe I’m not being clear enough.  If we had that information, we can then write a simulator to track that information from the batter’s perspective, and then we can see how the batter responds.  The information I provided in #2 was the “inputs” (which act in combination, natch) to which the batter will respond.

I’m on board. =)

I’m just saying that visual perception is a pretty tricky process, and that understanding how a hitter reacts to and responds to a pitch probably requires lots of experimentation and little simulation. Even estimating the speed of a baseball or trying to judge it’s trajectory is an incredibly difficult perceptual problem for the hitter, so the “inputs” as you describe them might in fact be very different in the simulator from what’s used in practice.

I don’t know about you, but I’ve got about 40 years left in me…

I was not suggesting that a batter can be in the box and think “here comes a 87 mph slider”. I phrased it as “a pitch I can hit well”.

Many years ago a was a fairly poor amateur batter, but I still remember enough that a batter should be able to tell if a pitch is fast, medium or slow; in or out, high or low; straight or breaking.

Even at the highest professional level, some batters are going to be better at this than others. We have a thread here on mistake (straighter, slower) pitches being hit harder. Find the pitches that the league tends to hit best (letter high fastball over the middle?) and then see how much variance there is between batters on their swing rates and their contact rates. Some hitters (Bonds, Giles) are very disciplined, rarely swinging at pitches out of the zone, while haveing high contact rates for those pitches in the zone. Others chase just about everything

Hmm I guess you have made a very good point Mike.  I will have to go back to my pitch classifier and see what I can do.  I really thought it was doing a fair job of separating the two by movement but I guess not.  IT is due for an overhaul anyway as it takes far too long to run as it is.

Thanks.

Josh-

We had talked about this earlier, but pitch classification has come up in this post, so I thought I’d mention it. I looked up Phil Hughes (because I am apparently into self-mutilation and was reading a post at NYYfans.com)

http://baseball.bornbybits.com/2008/Phil_Hughes.html

It looks there like his changeup is being called a fastball. I have no idea if these were the same pitch classifications that you looked at, but you might want to check this out.

-Dan

Dan,

Yeah my pitch classification code is in need of an overhaul.  Because of the smaller statistics it right now is lumping clusters together when it shouldn’t.  This would somewhat fix itself shortly but I need to rework the sinker/fastball code anyway so I will probably alter everything.  Last season I did need to hand edit over 30 pitchers for things exactly like you say though.  I am not hand editing right now because every time I run things are changing underneath my feet!

Josh,

Is the main issue that you are trying to cluster every pitch, as opposed to setting those “too close to call” as simply “unknown”?

We should find some space on a forum where we can all talk about how best to classify pitches. It sounds right now like everyone has slightly different ideas - and while I’m not advocating groupthink, I think it’d be good to work together on this problem so that everyone doesn’t have a different model of how to do it or what they’re working with.

I started this thread for you…

http://www.insidethebook.com/ee/index.php/site/comments/technique_for_clustering_pitches_from_pitchf_x/