THE BOOK--Playing The Percentages In Baseball

Tango, isn’t also true that it takes some stats longer to stabilize because they occur less frequently.  A good example is triples per batted ball.  There is a very wide range of talent in triples, but it takes a long time to stabilize (relative to other, more frequent base hit types) because there aren’t a lot of triples hit.

Right, I was only talking about the primary reason.

In the “solving dips” article from a long time ago, I listed all the possibilities.  Number of opps was one, the rate being close to .500 was another.  I don’t remember the others.

So, if your rate is triples per PA, then, yes, that’ll take longer to stabilize.

I was thinking of that article.  Spent a lot of time on it--eventually made it about halfway through.

studes, before a batter can get a triple he first has to get to second base - so the proper denominator for triples is (doubles+triples), which for most players is 30 to 50 a season.

You’ve mentioned that before, Brian.  What would that do to the stabilization of the rate?

It’ll stabilize alot faster.  Not only because of the higher rate, but also because you remove a ton of noise in the denominator.

The idea is that you want a metric that does stabilize fast.

So is the idea that a skill that players are selected on will stabilize quickly? But it can’t be TOO important, like goalie save % or pitcher BABIP? In other words, if you graphed “importance” and “stability”, it wouldn’t be linear, but would be low then high then low again?

The idea is that you want a metric that does stabilize fast.

Okay, but why?  I understand that it’s more predictable, but does it also mean that you’re closer to something “significant?”

studes: yes.  By stabilizing faster, you remove the noise.  SB/(1B+BB) will stabilize faster than SB/H or SB/PA.  Because we know that a SB can’t happen on a HR or if the batter is already out.

anon: the more noise you remove, the faster it will stabilize if it’s a real skill, and the slower it will stabilize if it’s not a real skill.

For example, H/PA stabilizes faster than (H-HR)/BIP (i.e., BABIP).

But, you can get a little crazy on it.  You can make it stabilize with say only n=50, but if it takes you 10 years to get n=50, then it doesn’t really help you that you were able to isolate the skill with that metric.

So, they’ve already been selected based on BABIP, so they don’t necessarily selected based on K/PA.

It’s not intuitively apparent to me that this must be true, but I’m reminded of the data from the Hamilton College thread where Tangotiger demonstrated this persuasively.

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

---
[Can you tell us anything yet about the back-story as to why Hamilton College?]

I replied in the other thread…

Thanks, Tango.  So, back in the article, Derek found that IF/BIP (Infield fly per all contacted balls) stabilized faster than IF/FB (Infield fly per all fly balls).  I’m intrigued by this, because I would have thought that removing GB and LD would be removing “noise.”

Based on your opening post, I assume this is due to the fact that there is a wider range of IF/BIP than IF/OF, right?  Makes sense, given ground balls, etc. etc.  But, given this comment:

the more noise you remove, the faster it will stabilize if it’s a real skill, and the slower it will stabilize if it’s not a real skill.

...I guess means that IF/FB isn’t a “real skill.” Presumably because it follows FB/BIP so closely.

Make sense?

Well, maybe removing GB removes noise, but removing LD removes information.

It’s also possible that there is a relationship between IFFB and GB.

I’d like to see a matrix of all paired-combinations of IFFB, OFFB, LD, and GB.

studes/13 - This is just an intuitive comment, but I think it’s likely true that an OFFB is more similar to a LD than it is to an IFFB in terms of what has to happen to get the result. What I mean is that the difference between a line drive to the gap and a lazy fly ball to the gap, in terms of the batter’s swing, is not that significant compared to that of a swing that results in a pop up to the pitcher’s mound.

Is there a way to check this?

On one hand, if the batter makes contact, there are four possible outcomes(in Gameday), GB, LD, FB & PU, which overlap adjacent classifications when plotted vs vertical launch angle. You could then argue that all four could be measured as a function of balls contacted.

On the other hand, using a flow chart approach to get true talent rates of each type, work by Derk and Harry shows that GB stabilize the quickest (about 100 BC). Once you have determined the GB rate, what’s left is (LD+FB+PU). FB and PU both stabilize at about 300 BC, but FB is adjacent to (overlaps with) LD, the remaining classification, where PU is adjacent to FB but not LD. So then I’d regress PU based on PU/(LD+FB+PU). What’s left are LD and FB. Regress the FB rate on FB/(LD+FB), and whatever’s left is LD, which takes by far the longest to stabilize.

studes - I don’t have my tables handy, but IIRC I regress TR to about 100 or 150 XBH. At 30-50 XBH per season, that can be 3-5 seasons.