THE BOOK--Playing The Percentages In Baseball

Great, great stuff.  The next step is to add in expected value for control, pitch sequencing, and pitching to the game situation.  Those things are A LOT more difficult to handle!

One big problem with any “stuff” analysis:

The author says this:

“Two components determine how nasty a pitcher’s stuff truly is: velocity and movement. “

And of course it is really three things, although technically only the two he mentions are physical attributes of the pitch.  The third thing is “deception.” Deception includes all kinds of things like hiding the ball before it is released, arm angle and height off the ground, body movement which may or may not distract the batter, etc.

Those things can have significant effects on pitch value.  So, the question then becomes, if a pitcher is under or over-performing his “stuff” how much of that is due to command and sequencing, how much of that is random fluctuation, and how much is due to non-average deception.  That really throws a monkey wrench into the whole equation. If pitchers were robots or machines but they all threw different pitches with different velocities and had their own unique command and sequencing talent, we could nail their true talent simply by quantifying those things.

Because of “deception” we can never do that unless we can come up with an objective method for quantifying “deception” (which is possible to some extent).  Without that ability, we are left with having to assume at least some, and maybe most, of the difference between what you would expect a pitcher’s value to be based on his movement, velocity, command and sequencing, and what that value actually is, is because of his “deception.”

For example, let’s say that from the velocity and movement of all his pitches, plus his command and sequencing, we determine that Randy Johnson should have an overall value of -1.30 runs per 100 pitches, where minus is good.  But we find that he actually has -1.50 in some limited sample of data.  Is that just a random blip and we can assume that he is going to be -1.30 going forward, or is most of the difference between the -1.30 and -1.50 due to the fact that he is 6 foot 10 and slings the ball, thus having more deception than the average pitcher on at least some of his pitchers, even after controlling for velocity and movement?  Tough question to answer.  Which is why the usefulness of quantifying a pitcher’s “stuff,” using velocity and movement, but not including deception (which cannot really be objectively quantified - at least at this point) somehow, is always going to be of limited value.

Tango, thanks for the link.

MGL, I’m glad you liked the article.

I know you often stress deception as an important part of a pitch’s quality, but thinking strictly in terms of “stuff”, I’m not so sure that I consider deception. I suppose we define the term differently. Of course, I do agree that deception is an important part of a pitch’s quality.

Doesn’t matter how you define it.  The point, which I’m sure you agree, is that a pitch’s value, independent of command and sequencing, is a function of exactly 3 things:  Movement, velocity, and deception. Even though only 2 of those things are physical characteristics of the ball, I can’t think of any reason to separate the three things, other than we cannot at the present time objectively measure deception but we can measure velocity and movement.

Plus, I always want people to know that velocity and movement alone do NOT tell us how difficult a pitch is to hit.  That is probably the most important point in bringing up deception.  IOW, if we want to talk about who has the best fastball or curve ball or whatever, we HAVE TO include deception.

IOW, your numbers tell us how good each pitcher’s pitches are, given that they are being thrown by a league average pitcher with league average “deception.”

At the very least, you would have to use different values (in the regression equation) for lefty and righty pitchers, wouldn’t you?

MGL, it all really depends on what we’re trying to measure. I believe that what I was trying to measure using three variables--velocity, horizontal movement, and vertical movement--was the quality of force that each pitcher exerts on the ball in terms of velocity and spin, or something like that. I believe that that is what makes up “stuff.” That might be useful in and of itself, if, for example, we wanted to decide whether a pitcher should give up 2 MPH in velocity if it means getting 3 more inches of movement on his cutter, all else equal. And in this case, we don’t need to consider deception or release point or handedness. But of course, to measure the actual quality of a pitch, there are likely dozens of variables, many unquantifiable, that would be needed to be taken into account.

How does one define “deception” exactly?  Is it prehaps a high leg kick (el duque)?  Or can deception also be for example a changeup with exactly similar movement to the fastball?  IOW, I think deception can also have physical characteristics of the ball.

"deception” is automatically defined as the difference between the value of the average (across all pitchers in the league - perhaps by pitching hand) pitch from a pitcher, given its velocity and movement, and the actual value of that same pitch from that pitcher, independent of (or adjusted for) command and sequencing.

For example, let’s say that Dontrelle Willis’ fastball is 90 mph with 3 inches of downward movement and 2 inches of horizontal movement. And let’s say that the average LH pitcher with that exact same pitch (or we simply put those parameters into a regression equation) is -.50 runs per 100 pitches.

And let’s say that Willis’ actual value is -.75 runs per 100.  Negative is good for the pitcher. And let’s say that somehow these values are after controlling for command and sequencing. The difference of .25 runs, for the good, is Willis’ “deception” by definition.

Now, there are problems with that.  One, we never know a pitcher’s actual pitch values - only his sample ones.  And the smaller the sample size of pitches, the less reliable are those sample values, of course.

Two, even if we have a very large sample of pitches, the difference between a pitcher’s actual value and his “expected” value (based on velocity and movement) is clearly a combination of deception, command and sequencing. 

So how can we separate out command, sequencing, and deception?  I have no idea! Hence, one of the big problems with this train of analysis.  I suppose that we could try and quantify command and sequencing, factor them out, and whatever is left, is deception.  Or we can just lump them all together and call that “comseqeption.” The problem with that is we want to teach optimal sequencing, we want to track a pitcher’s command progression throughout his career, and we probably can’t change much a pitcher’s deception.

IOW, we are trying to figure out something significant about a pitcher’s true talent, independent of his results, especially if we don’t have many results to go by.  So we figure, “OK, we have a record of the physical characteristics of his pitches (pitch f/x).  Why not use that?  That pretty much tells us everything other than his command and his sequencing, which we can either get to later or we can teach (the sequencing at least).” But…

Those physical characteristics (velocity and movement) are only half (or 30& or 20% - I don’t know) of what gives his pitches value, independent of command and sequencing.  The other “half” is what I call “deception” (I could call it anything I want - don’t confuse a name with a definition - the name of something does not always define it), which is simply everything else other than velocity and movement that creates value in a pitch.

We could (almost) exactly quantify “deception” if we conducted this experiment.  We have a pitcher throw to some batters in a “game situation.” We tell the batter that the pitcher is going to throw fastballs 50% of the time and curve balls 50% of the time, randomly.  We have the pitcher do that 100,000 times (spread out over several sessions of course).  We do that for all the RH pitchers in the league and all the LH pitchers in the league.  We use pitch f/x to record the velocity and movement of each pitch.  Again, if we can somehow control for command and location, the difference between the result of each pitch for each pitcher and the average result for all same-handed pitchers given the same velocity and movement, is that pitcher’s “deception index” or whatever you want to call it.  It obviously is a function of his arm angle, release point, body movement, etc., etc.

I ran exactly this analysis, only I looked only at fastballs, I included location in the analysis, and I broke it down by component.  There some minor differences: I used a kernel density estimation algorithm instead of a local regression, but these are minor.

I got largely the same results.  I’m hoping to find a place to post them soon.  In addition to a list of the best individual pitches and pitchers, breaking down the analysis into components gives us a sense of the relative importance of velocity and movement.  Turns out velocity is hugely important in predicting outcome (45%), and X location is as well (28%) but not Z location.

A weakness of this analysis (and mine) is that it treats each pitch individually.  So, if Joel Zumaya and Jon Papelbon throw the same pitch out of the zone, they get the same credit. But the data clearly show that batters know Zumaya has no command, so they don’t swing at bad pitches.  When Papelbon throws a “bad” pitch, it often has a good result; when Zumaya pitches, it rarely does.  But this analysis can’t account for that.

MGL: Sequencing is concrete; we can deal with that.  Justin’s work is a great first step in that direction.  Control is also concrete; we can look at X/Z location in the zone.  My analysis accounted for that.  The only bugaboo is deception.

You’re defining “deception” as unexplained variance.  Much of that is random chance. You’re talking about deception as if its like clutch hitting--some undefinable quality that is difficult to demonstrate even exists.  How do you know that “deception” exists?

"How do you know that “deception” exists?”

Are you kidding me?  I don’t know how to even answer that.

I am treating deception as the sum total of the effect that a pitcher’s arm angle, position on the rubber, motion, length of arms, height of the ball of the ground, height and weight of the pitcher, etc., etc., has on the batter’s ability to do whatever he does or does not do that affects the value of the pitch.

Now, measuring that is another issue and obviously problematic.

If you think that those things don’t significantly effect the value of a pitch over and above the velocity, movement, command, and sequencing, then you literally have never watched a baseball game.

Have you ever seen Sid Fernandez throw an 88 mph fastball that appears to look like a 95 mph fastball because of his funky delivery?  Have you ever seen or heard someone speak about a pitcher who hides the ball well or has a funky delivery which makes it harder for the batter to “pick up” the ball?

So you think all of those things are myths?  Do you think that arm location, arm angle, position on the rubber, etc., etc., have no effect on the value of a pitch?

Or were you just asking?

Here are my top 24 broken into components. Clearly pitch F/X mistakenly classifies 2-seamers as 4-seamers.  Otherwise Joe Smith and Sean Green would be 8+ standard deviations from the mean in vertical movement. I’m rerunning the analysis w/ my own classification system, so these are preliminary results that will change… Its a proof of principle at least.

http://www.princeton.edu/~cdm/rankings.jpg

MGL/8: I like your definition, because I can’t really find in it any single factor that we can *really* argue is important. Are you sure that pitcher weight affects “deception”? Or position on the rubber?  How could we claim that?  I think you mean, “Its real, da**it, and its in the throw, somehow, but I can’t really say how.”

I believe some pitchers have funky deliveries that hurts the hitters ability to predict the timing of the pitch. Using that definition, we may have access to deception by using hit f/x data, seeing how often the batter is early or late. I think its likely that it only affects a small subset of pitchers with abnormal deliveries.

This looks like one of those things that is so ingrained in baseball “wisdom.” Much of what sabermetrics is about is questioning such wisdom.

"This looks like one of those things that is so ingrained in baseball “wisdom.” Much of what sabermetrics is about is questioning such wisdom.”

You’re preaching to the choir, baby…

I don’t think that is even close to being correct in this case, but I could be wrong.

One way to test this would be for someone to either watch video and put each pitcher into one or three categories: funky, average, or very straightforward. Or maybe 2 categories, funky and not funky.  Then see if we can find any differences in the pitch f/x data in the value of the various pitches, after controlling for as much else as we can, certainly velocity and movement.

Or what about we do the same thing for height (and BTW, by “weight” I simply meant the “bulk” of the pitcher on the mound).  Or how about by release point according to the pitch f/x data (although we would probably expect release point to affect movement).

I would be quite surprised if we didn’t find significant differences in pitch value based on all of these physical characteristics of the pitcher and his delivery.

But, this is a factual issue, so there is little use in debating it…

Interesting discussion. I am thinking about Keith Foulke. His delivery was a little unusual, but I think the main reason he was good is that batters couldn’t distinguish his fastball from his changeup. (That, and he had great control.) I would call that deception even if it isn’t part of MGL’s definition.

dcj/12: Funny thing you should mention Keith Foulke.  He has the best control of any pitcher, 2007-2009, who threw more than 100 fastballs. His fastball behaves a lot like a cutter; it has little/no acceleration away from righties.  The value of his movement was 1.8 SD above the mean, and his location was 1.5 SD above the mean. Could his “deception” just be great movement and location?

All of this with pretty graphs will see the light of day soon.

Sure, the difference in arm speed and motion between the various pitches is definitely part of “deception.” Deception - at least my definition - is anything that creates value in a pitch after controlling for speed, movement, command and sequencing.

I agree that we we might think of as “deception” is sometimes and could easily be movement.  Pitchers with a funky delivery or arm position or angle often produce funky spin and movement on the ball.  So we have to be careful about assuming that it is the delivery itself that is causing the value.

Then again, if I remember correctly, didn’t Foulke have kind of a short-arm funky delivery?  I would not be surprised if that delivery in and of itself had an effect on the value of his pitches.

If there was little truth to the idea that hiding the ball, position on the rubber, etc., affects how the batter sees the ball, and therefore the value of each pitch, then pitching coaches are teaching young pitchers some things that are a waste of time. Which is possible.