In addition to our own research, a major part of HSAC is the discussion of other projects in the sports world. Steven Levitt (of Freakonomics fame) and Kenneth Kovash recently released a working paper on game theory in the NFL and MLB.
National Bureau of Economic Research (summary)
National Bureau of Economic Research (working paper)
Last week, we discussed the baseball part of the paper. For other summaries of the paper, please see the following articles:
This week, we will consider the football part of the paper. After our discussion, we will post the HSAC analysis of the paper. Stay tuned…
The baseball portion of the analysis has very serious flaws, starting with the fact that they calculate OPS incorrectly. The result is that walks are severely undervalued (even more than in actual OPS), which in turn makes breaking pitches appear more effective than they truly are (because they lead to more walks).
More importantly, focusing on PAs that end on a given pitch also provides a very inaccurate read on outcomes, as sabermetricians have long understood. The authors do also look at subsequent outcomes, but that too can be misleading. The correct approach is to look at all outcomes on/after a pitch.
For good discussions of there problems, see the Sabermetric Research discussion: http://sabermetricresearch.blogspot.com/2009/09/game-theory-study-on-pitch-selection.html, and also discussion at The Book blog: http://www.insidethebook.com/ee/index.php/site/comments/game_theory_on_pitch_selection/
As it stands, the paper really tells us nothing about whether pitchers are optimizing their pitch selection.
You can also find a good discussion of the football analysis here:
Very good points, Guy. We brought up the OPS issue during our discussion. As you imply, even if they calculated OPS correctly, they would still be undervaluing walks. We wondered whether the entire result would disappear if they used linear weights.
I have sort of lagged on posting our notes, but I will get them up there this weekend. Thanks for checking out the site and thanks for the links.
If you look at the Book Blog discussion, you’ll see that Mike Fast put together some data using linear weights and “through counts”, i.e. all outcomes combined whether on or after the pitch. Most of the anomaly disappears, except on 2-strike counts. My guess is that once we controlled for platoon effect, the result would completely disappear (because pitchers throw more breaking balls on 2 strikes when they have platoon advantage), but I can’t say for sure. Hopefully, Mike will write up his results at some point.