THE BOOK--Playing The Percentages In Baseball

There is some discussion at THT between Peter Jensen, MGL and me about whether or not a pitch attributes based metric would take out more skill than noise.  Peter and MGL think that the aspects of pitching not captured by Pitch f/x (sequencing, which can be looked at but is really hard!, deception, game theory/pitch selection, pitching to the batter’s weaknesses) are hugely important.  I agree with them on that front, but I’m just not sure if those factors have more of an impact than random variation from umpires and batters.  In in fact a pitch attributes based metric would take out more noise than skill, it is a net gain. 

The problem is there simply isn’t enough Pitch f/x data to investigate that question, at least I don’t think.

I don’t know that one comment consitiutes a “discussion” , but I agree with this, of course:

“In in fact a pitch attributes based metric would take out more noise than skill, it is a net gain.”

Whether there would be a net gain or not, I don’t know.  It would depend entirely on the exact methodology.

The “controversy” is VERY similar to using a DIPS or PZR-type system for hitters.  IOW, we might consider taking the parameters of a batter’s batted balls (location and speed essentially) and assign a value to them, and then use that value rather than the actual outcome of the batted ball to assess that batter for past “context-neutral” value or projection purposes.  This is similar to PZR for pitchers.  The problem, as with the pitch f/x discussion, is that it eliminates the positioning of the fielders from the equation, park effects, the ability of batters to possibly “hit ‘em where they ain’t,” and the differences in the speed/trajectory of the batted balls that are unique to the batters but are not picked up by the data - for example, it is likely that a “hard” line drive by Cecil Fielder is actually a lot harder than a “hard” line drive by Juan Pierre.

It was more with Peter and me, but then you joined in and I didn’t want to make you feel left out .

I don’t think it’s similar to DIPS for hitters.  Pitch f/x is VERY precise.  It gives, to a half an inch, the exact location of the ball, a very close estimate of velocity, it’s spin deflection, RPMS, etc.  A hitter based DIPS using BIS data or something wouldn’t have nearly the precision of Pitch f/x data.  Even if it used Hit f/x data, that would still include the performance of the pitcher (undoubtedly, some batters see tougher pitches over the course of a season than others).

With Pitch f/x, the only variables that are not considered are the ones you and Peter have already mentioned, mainly to do with approach and deception.  And of course there is the problem of establishing the bins (or regression function).

My experience with a lot of the Pitch f/x data (and more so BIS’s TLV data) is that there are more exceptions than there are “rules”, which makes bucketing data not so useful when trying to make a model that will end up correlating at least somewhat well with a pitcher’s actual results.

If Trever Hoffman can make his 85 mph fastball work and if you look generally at 85 mph fastballs and find that they are not such a good pitch, you’re just going to come up with fairly wrong results for Hoffman that claim he’s not a good pitcher.  That’s a bit more of a broader bucketing than what you’re talking about, but I think you’ll run into similar problems.

I think each pitcher, when it comes to his repertoire, pitch selection, etc… is in some sense his own special case and there can be such a variety in what actually works and doesn’t that creating an all-encompassing model is going to be quite difficult.

Imagine if instead of PitchFx, we were talking about “PunchFx”, which tracks the direction and speed of punches thrown in a boxing match.

We’d analyze jabs, hooks and uppercuts like we do fastballs, curves and changeups, and look at punches to the face and body blows like we do high and low pitches, etc.  With enough data, and some fancy math, we could attach an average value to each type of punch, an average value to punches thrown to certain spots, a differential value for left handed punches vs. right-handed punches, etc.  We’d know all there is to know about the value of different punches, and thus we could tell an aspiring boxer exactly which punch to throw in order to win, right?

Well, that wouldn’t really work, would it?  A left jab might on average yield 1 point of “knockout”, or whatever we decide to call the “currency” of punch value, but a jab thrown straight into the other guy’s gloves, or thrown when the other guy isn’t there yields nothing.  One punch at the right time, which by PunchFx yields on average a pedestrian 2.46 points of “knockout”, could put your opponent on the floor if it is thrown at the right time and the right place…

This sort of analysis will be able to shed light on whose punches are on average better (to support someone’s argument for entry into the boxing hall of fame), and it might shed some light on which of a boxer’s many punches is the best in his arsenal, and it could help you figure out if someone is losing his punching power, but it seems certain that it could never fully describe all of boxing, specifically the portion of boxing that is “art” rather than “science”: when to move, when to punch, what punch to throw in what sequence, etc. etc…

Similarly, I think PitchFx does a great job of characterizing what we’re watching, and helps us to understand lots of things, and it is immensely entertaining, but it’s never going to be able to tell the complete story (and I’m not saying anyone thinks that, btw). 

There will always be a residual portion of pitching (this most likely a very big portion of the whole) that can’t be understood by “merely” tracking the ball.

(and I say “merely tracking the ball” with tongue-in-cheek, offering the utmost respect for what a seminal accomplishment that is...)

Although the “evidence” is only anecdotal, I gotta agree with Dave above that if you use a methodology suggested in the article and you don’t account for other factors unique to the pitcher, especially pitch selection, you are going to come up with some horrendously wrong results, such as with Hoffman.  I assume that the only reason he has good results (I only assume they are good - maybe not) with his 85 mpg fastball is that he throws his change-up so often AND they are hard for the batter to distinguish.

I like Greg’s analogy.