Ran across a recent piece on the evolution of Jason Garrett's playing calling by Jonathan Bales of DC Times (an analytics based blog on the Dallas Cowboys and the NFL).
When I first began studying Jason Garrett’s play-calling in 2009, perhaps his most blatant tendency was allowing previous play-calls to dictate his current ones. That was particularly true on second down, when Garrett would often run if he had passed on first down and pass if he had run. The thinking, in all likelihood, was to “mix it up” in order to keep the defense guessing. In attempting to randomize his play-calling, however, Garrett was quite ironically becoming extremely predictable. You can consult my previous studies of Garrett’s second-down calls to see just how predictable he had become.
This harmonizes with an NBER paper by Kenneth Kovash and Steve Levitt on a wide set of pitch data from MLB and play calling from the NFL. Rather than well mixed plans, they find evidence of correlation from one decision to the next. However, Bales finds that from 2010 onward, Garrett's play calls across plays have become randomized better:
To Garrett’s credit (or perhaps more so to the credit of the analytics team the Cowboys brought in after that season), the coach’s second-down play-calling improved dramatically in 2010. It’s been above-average ever since.
Analytics becomes information not only about the other team but about one's own behavior that can improve behavior.