|
Abstract:
|
Sports is a great model system for understanding how to apply statistical methodology in practice; similarly, the field of Sports Statistics itself is a great model system for understanding how rigorous statistical methodology penetrates new fields. The recent growth of Sports Statistics has paralleled broader interest in Data Science, and in both cases, a core meme that has taken hold, often at the expense of other principles, is the notion that predictive accuracy can serve as a unifying yardstick for evaluating all quantitative analyses. In this talk, I highlight three cases where a myopic focus on predictive accuracy can paradoxically lead investigators astray: attribution problems, where we hope to attribute a change in expected outcome to particular decision; strategy development, where we hope to use data to design in-game strategies; and player evaluation, where we hope to use game outcomes to calculate player value. In each case, a more appropriate analysis can be framed in terms of an alternative prediction problem whose construction highlights a core statistical principle that should be advanced in combination with predictive accuracy.
|
