Thanks, Happy New Year and a question

[MLBMovingAvg]

My question has to do with the determination between predictive and descriptive stats when applied to baseball. Not sure I'm asking this question clearly; I don't mean the definition of the two, but what garners that distinction? Can it change as more information becomes available and correlations change?

I know its a lot to unpack but Ive done some researching on my own and I couldn't find a comprehensive explanation.

Thanks a million for the membership Im trying to make the most of it
John

[Todd Zola]

It's all about sample size. The Holy Grail is knowing when a new skill can be considered real and not noise. The predictive/descriptive has spread in the lexicon since the introduction of Statcast data. The hope is since the data is more granular and in many cases, objectively determined, it is more predictive than what's out there.

Predictive/descriptive predates Statcast. The notion of stabilization points has been around for over a decade. (Notes: I write about this in the Rotowire Magazine, the piece will be posted in Platinum). Stabilization points are when half the outcome can be considered all skill. The assumption (and I was guilty of this) is the player now owns that skill so the original projection for that skill can be regressed towards the new level (higher or lower). This is still the common application - with regression commencing at the stabilization point.

The problem is, the person first coming up with stabilization points has since come out and said we're doing it all wrong (including the author himself). The explanation is the skill within each sample of X size is real, and it will be real in the next X. However, the skill in the first could be different than the skill in the second. Within each sample, the luck/skill ratio was 50:50, but the skill levels don't have to be the same. Most everyone either ignores this, is unaware or doesn't understand it. I'm not 100% sure I completely understand it.

The manner I previously applied this principle was as follows, using contact rate as an example. Contact "stabilizes" at 40 PA. Therefore, once a batter attained 40 PA, I averaged the projected skill with the new skill and carried that through in my "rest-of-season" projection.

The revised explanation says this isn't correct -- I may not be regressing to the player's new skill. I'm regressing to the skill exhibited in those specific 40 PA.

The $64,000 question is what is the point a sample can be considered real? To be honest, I haven't found anything definitive. Those capable of determining it are still doing it wrong. Maybe I just haven't found it, or don't understand the methodology well enough to make that statement, but logically, if it was known, wouldn't we all know it?

All that said, I think it's reasonable to assume the stats with faster stabilization points in fact become real sooner, we just don't know when. I still incorporate this into projections on a relative basis, with admitted guesswork when to apply the regression and to what extent.

FWIW -- stabilization points as defined above on Statcast data are indeed sooner than conventional data, so I think it's fair to say some of them are "more predictive", though we don't know the extent.

[MLBMovingAvg]

Thanks Todd that was excellent, thanks for your time.

If I could ask one more question, who is the author you referred to in regards of stabilization points?

Thanks again

[Todd Zola]

Do a search on pizza cutter russell carleton stabilization points