Douglas Fearing, Jason Acimovic, Stephen C. Graves
January 18, 2010
Abstract:
Existing performance metrics utilized by the PGA TOUR have biases towards specific styles of play, which make relative player comparisons challenging. Our goal is to evaluate golfers in a way that eliminates these biases and to better understand how the best players maintain their advantage. Through a working agreement with the PGA TOUR, we have obtained access to proprietary “ShotLink” data that pinpoints the location of every shot taken on the PGA TOUR. Using these data, we develop distance-based models for two components of putting performance: the probability of making the putt and the remaining distance to the pin conditioned on missing. The first is modeled through a logistic regression, while the second is modeled through a gamma regression. Both of these models fit the data well and provide interesting insights into the game. Additionally, by describing the act of putting using a simple Markov chain, we are able to combines these two models to characterize the putts-to-go for the field from any distance on the green for the PGA TOUR. The results of this Markov model match both the empirical expectation and variance of putts-to-go. We use our models to evaluate putting performance in terms of the strokes or putts gained per round relative to the field. Using this metric, we can determine what portion of a player’s overall performance is due to advantage (or loss) gained through putting, and conversely what portion of the player’s performance is derived off the green. We demonstrate with examples how our metric eliminates significant biases that exist in the PGA TOUR’s Putting Average statistic. Last, extending the concept of putts gained to evaluate player-specific performance, we show how our models can be used to quickly
Keywords: golf, putting, performance modeling, logistic regression, gamma regression, Markov models
[Source] MIT Sloan Research Paper No. 4768-10
댓글 없음:
댓글 쓰기