Wednesday, October 14, 2009

The Empirics of Baseball Strategy

Standard game-theoretic utility maximisation assumes that in a competitive game where somebody is trying to predict your actions, the rational thing to do is to completely randomise your strategy: be entirely unpredictable. An implication of this is that there should be no correlation through-time (positive or negative) between your tactics. Results from behavioural experiments question the notion that people actually behave this way.

In the fine tradition of economists testing utility maximisation using data from professional sports, Kovash and Levitt have a new working paper based on "high stakes, real world settings that are data rich: choice of pitch type in Major League Baseball and whether to run or pass in the National Football League."
We observe more than three million pitches in baseball and 125,000 play choices for football. We find systematic deviations from minimax play in both data sets. Pitchers appear to throw too many fastballs; football teams pass less than they should. In both sports, there is negative serial correlation in play calling. Back of the envelope calculations suggest that correcting these decision making errors could be worth as many as two additional victories a year to a Major League Baseball franchise, and more than a half win per season for a professional football team.


Kevin Denny said...

Is it really rational to be entirely unpredictable?
Does this not depend on whether there are costs to switching? I am right-footed so it is easier for me to kick the ball to the left (& more accurate). So while I may want to left my opponent guessing, I don't think it follows that I should kick randomly to the left or right.
In chess for example grandmasters will try surprise an opponent with an opening that they haven't used before (Fischer playing Alekhine's against Spasski for example) but mostly they use one's that they do know.

Enda Hargaden said...

I think a conclusion along the lines of "These people know what they're doing and do not behave as von Neumann would predict" would be better than "We reckon von Neumann would win more games."

That's the main critique I have of Levitt's work:

Very inventive,
Clever econometrics.
Miss the feckin' point.