Measuring pitch variability through PITCHf/x

For a while, I've been wondering what can be measured and analyzed using PITCHf/x data that hasn't already been measured and analyzed. A few things crossed my mind, but the most interesting thought was about the degree of variability of a pitcher's pitches.

It would be relatively easy to measure how much variability a pitcher has in velocity and movement if all things were equal. Of course, they aren't.

The two biggest problems for analyzing this type of variability are, as I see them, pitch type identification and park-to-park measurement error. Variability would mean little if half of a pitcher's "two-seam fastballs" are actually change-ups. Variability also runs into problems when parks like Kansas City -- whose radar gun readings are notoriously high -- are included in a data set with other ballparks.

Fortunately, if we only look at a single ballpark -- usually the pitcher's home ballpark because it has the greatest sample size -- park-to-park measurement error should be less of a factor. Without some form of park-to-park normalization, though, interpark comparisons shouldn't necessarily be taken at face value.

Additionally, 2010 saw a huge improvement in pitch type identification. While it still isn't 100% accurate, it is close enough on many pitchers to give me confidence while playing around with my ideas.

I haven't really dug into the numbers yet, but I will be looking to see if variability within a pitch type helps or hurts a pitcher. My gut feeling is that the number itself won't have much meaning.

To calculate the variability, I plan to capture the 95% window using two measurements of the standard deviation in both directions from the mean. By definition, this eliminates the outliers, but it will take some study to determine if that's really the measurement to use.

It may be beneficial to use a pythagorrean measure to find the variability for pitch movement; however, this would not appropriately model pitches that have greater variability vertically than horizontally (and vice versa, of course).

Look for a follow-up after I play around with this idea.

Thinking about run values

For some time, I've been looking for a way to appropriately integrate run values into the PITCHf/x database. I have read articles at Beyond the Boxscore, Inside The Book, and Cubs f/x, but I am no closer to getting what I want. Unfortunately, I lack the resources and time to find the answers myself.

Many tables have been published with run expectancies for the 12 ball/strike count states for various time periods. Tables have also been published for the 24 base/out states. Because the two tables contain different representations of the same data, there's no way to combine them. What I would like to see -- and I'm sure this makes me a sadist -- is a run expectancy table for the 288 ball/strike/base/out states.

Yes, that's one hell of a matrix to process, but there are two thougts that seem to be the beginning of arguments against the two relatively simple approaches:

  • The thought against only using the 12 ball/strike count states table: a first-pitch strike in a bases loaded, no out situation has to effect the run expectancy more than a first-pitch strike in a bases empty, two out situation, right?
  • The thought against only using the 24 base/out states table: an 0-2 single with a runner on first base has to effect the run expectancy more than an 3-0 single with a runner on first base, right?

Admittedly, I don't have the knowledge or skills necessary to issue either of those thoughts as facts, so I have posed them as questions. It seems logical, though, doesn't it?

I think an appropriate time period for the analysis to cover is 1998-present -- since the last expansion.

Does anyone know if anyone has tackled this subject, successfully or otherwise? Is this covered in a book that I have not yet read -- possibly even one that I have read?

Consider this an open call for help in this matter.