A Fanpost on the Clutch Stat

There was a side disucssion on how to apply Fangraphss' Clutch stat. First the defition from Fangraphs:

Clutch: A measurement of how much better or worse a player does in high leverage situations than he would have done in a context neutral environment.

How it’s calculated: WPA / pLI – WPA/LI

Why you should care: Unlike tradition clutch statistics (close & late), Clutch is a much more comprehensive statistic taking into account all situations that may or may not have been high leverage. Additionally, instead of comparing a player to the rest of the field, it compares a player to himself. A player who hits .300 in high leverage situations when he’s an overall .300 hitter is not considered Clutch.

A few important takeaways:

*Comparing the clutchness of players does not mean much. A player with wOBA of .255, maybe hit at a .290 level in "clutch spots".  At the same time a .390 wOBA player may hit at at .350 clip in "clutch spots". Player 1 will have a sizable clutch number, while player two will have a sizabel negative number. However, who wouldn't want player 2 and his .350 clip at the plate over Mr. Clutch at .290?

*There is not much year/year correlation with the Clutch stat which is why it is rarely used. Last  year's most clutch players were Ben Zobrist (,50) and BJ Upton (.62). The year Zobrist is -1.55 and Upton is at -.74. Last year's most unclutch player was Carl Crawford  at-1.66, this year he is at -.01. Jason Bartlett was at -1.3 last year and is at .31 this year.

Now remember clutch factors in all events (WPA) not just the high leverage, but those plate appearances are given the greatest importance. Look at Jason Bartlett, unclutch last year despite a high leverage wOBA of .347, but considered clutch this year with a .285 high leverage wOBA. You can't say "I want .31 clutch Bartlett at the plate over -1.3 Bartlett, because last year's Bartlett was a superior hitter. Its all about the personal benchmark, and for this reason its not useful to compare across players even when using it as a descriptive stat. There is little evidence to support it being a predictive stat.