Are these stats defined anywhere?  For instance, is OF FB% a percentage of all balls hit that are outfield flies, or a percentage of fly balls that are outfield flies?  And what is SBO% and the other SB stats?

Last point: it would be nice to see references and comparisons with the many other studies of this that have been done in the past.

I’m not sure what your example means, but the R squared measures how much of the variation among all players in Year 2 can be attributed to the variation among those same players in Year 1.

I’m new to THT and it is fantastic, so bear with me if I can’t make as sophisticated inferences.

If LD% is so unstable, yet is has one of the strongest correlations with batting average/offensive succes (retrofitted), then is it the secret weapon in fantasy baseball drafting/projections?

In other words, if we see a player far off the mean LD% of 19%, could that be used as a primary indication as to how the player will perform the following season?

It’s almost as if it’s an anti-correlation in that it can be used to project performance in Year 2 if Year 1 is an outlier.

Thanks in advance for any clarifications.

Note: I’m not even a fantasy baseball player, but I figured it was an easy example of putting future projections in use.

Sorry for the confusion, Dave.  I added a quick glossary.  As to all the other studies, I’m sure there have been loads of them, so I knew I’d miss a whole bunch if I tried (if you have some links handy, though, I’d be happy to add them).  This isn’t anything new, just a quick reference for the readers who were looking for one.

The Real Neal,
Dave nailed it.  It’s a statistical tool that tells us how much of the variance for the player pool overall can be predicted by the one half of the data.  If you’d like a longer explanation, just let me know.

Thanks, Seth   (Also, please note that the contact rate I’m referring to is on a per-AB basis, not a per-pitch basis)

ThankYouMichaelLewis,
Glad to hear you’re enjoying THT.  I’m always willing to help people who want to learn, so feel free to ask away whenever you have a question.

As to this specific question, David Rasmussen pretty much nailed it.  LD% is a big driver of BABIP, but because it is so unstable, a LD% too far from league average is likely just good/bad luck itself.  While it tells us *something* about the hitter, if we were to try to predict his LD%, we’d need to include a heavy proportion of league average, so a guy like Bartlett’s projected LD% going forward might only be 20-21% or so.

We do need to note, though, that for pitchers, BABIP will generally regress to .300.  For hitters, everyone regresses to their own unique number (not necessarily .300!), so things become a little trickier to analyze.  This is a very important point to remember that many analysts still don’t understand.

Derek -
Great question that you ask here.
From what you’ve got here, I’m guessing you did an auto-regression with 1 lag estimated using OLS, no?
If so (and perhaps even if not):
R-squared isn’t exactly the metric that we want for measuring repeatability.  For example, you can have a high R-squared (meaning that the explanatory variables capture a lot of the explained data’s variance) and still have the coefficient on the lagged variable near zero (which means that next year’s statistic is likely to be near the league average even if this year’s wasn’t). In this case (high Rsq, low coefficient), the regression captures well that the stat is not repeatable.

Yeah, Jonathan, as I said, there are much better ways to do this sort of thing.  This is far, far from perfect or comprehensive or flawless.  All this is is a simple reference guide for those who haven’t seen anything like this yet.  There are definitely flaws, but I’m wasn’t looking to be super precise.  For comparative purposes, all I’m trying to do here is say “BA is unstable, contact rate is stable.  BABIP is unstable, HRs are somewhat stable.  LD% is unstable, GB% is stable.  etc, etc.”  The results this produces are roughly in line with what we get from a more complex study, which suits what I was going for.

It should be remembered that all of these correlations are artificially high due to the 350 AB cutoffs used - that substantially reduces the variance and therefor increases the correlation. This is why a weighted correlation is preferable.