Validating Strokes-Gained Method for Measuring PGA Tour Player Success
Background
In 2016, the PGA Tour, the leading professional golf tour in the world, began using an upgraded version of the Strokes-Gained methodology pioneered by Mark broadie as the primary statistic used to describe player success on tour. The strokes-gained statistic is calculated by comparing the improvement in a playerβs position for each shot, compared to a baseline average improvement of the average PGA tour player from the same position. While a simplified version of the metric was used for several years, statistics are now being collected for 4 specific areas of the course: off the tee, approaching the green, around the green, and putting. By keeping a cumulative score in each area for each golfer, one can measure success in various parts of the game, or by summing across statistics, one can expect to measure the success of a golfer in a specific round vs the average golfer. The full details of the calculation can be found here: http://everyshotcounts.com/. More color about the metrics and the adoption of them by the PGA tour can be found here: http://www.pgatour.com/news/2016/05/31/strokes-gained-defined.html.
The resulting metrics provide a uniform way of explaining the success of a tour player vs the field despite the differences in the types of shots taken or the playerβs position on the course. Yet, for decades, the golf community has relied upon statistics separate from scoring to explain success in the sport. For instance, rather than talking about how many strokes have been gained off the tee, the golf community relied upon descriptors such as driving distance and driving accuracy to measure performance in this area of the game.. These statistics can offer more description of the actual shots taken, while the strokes-gained metric captures how much each shot improved the golfers relative position to the field without describing βwhyβ or βhowβ that shot was played.
Because of this, as strokes-gained has been adopted and used increasingly as the metric of choice by the Tour, there has been backlash from the fanbase about the transition. Many feel that the traditional statistics are just as sufficient as strokes-gained in describing a golferβs success, and provide qualitatively a better description of a Tour players game. So, do they have a point?
Using data from www.pgatour.com/stats, we can test whether strokes-gained can better explain success than traditional statistics. Using data from 2009 through 2015, we can try to build predictive models using traditional stats and these new stats and compare how well they perform at predicting a golferβs success in 2016. βSuccessβ on Tour can be measured by the number of Fedex Cup Points (FCPs) accumulated during the year. As we will be transforming the data to allow us to predict a normally distributed variable (the raw Fedex Cup Point distribution is not normally distributed), we will not be able to directly predict the number of FCPs accumulated, but will be able to predict the rank compared to the field and compare how well the models pcan predict a players standing.
Code used to analyze the data in R can be found below.
The Models
We built and optimized two multiple linear regression models in R, predicting Box-Cox transformed Fedex Cup Points per tournament. The best-fit model using the old statistics called for inclusion of many traditional statistics: driving distance, driving accuracy, greens in regulation, sand saves, proximity to hole around the green, putts made distance, and putts per round. However, when evaluating the model, there was evidence of multiple collinearity amongst several of the variables. By dropping the driving accuracy and putts made distance, we were able to produce a well-fit model with an adjusted R-Squared of 0.4977. Qualitatively, this model makes sense, as it uses data from the four major areas of the game, which correspond with the four areas of the game which are captured by the strokes-gained methodology: driving (off the tee), greens in regulation (approaching the green), proximity to hole around the green (around the green), and putts per round (putting).
In the strokes gained model, we attempted to predict the same Box-Cox transformed Fedex Cup Points data, using the four strokes-gained metrics. The best-fit model relies upon each metric with similar weighting and confidence for each, and upon inspection we see that there is no multiple collinearity. The model produces an R-Squared of 0.7548.
The evidence suggests that the strokes-gained model is a better predictor of success, as defined by relative Fedex Cup Points gained per round played, than the traditional model. However, the ultimate test is in predictability of results in the current year, 2016. Again using data from www.pgatour.com/stats, we are able to capture statistics for Tour players in 2016, and can predict their ranking in the average Fedex Cup Points gained per tournament.
The Conclusion
We tested each of the predictions by comparing the model fit to the actual standings in 2016. Applying the same Box-Cox transformation on the average Fedex Cup Points earned in 2016 provided a target for the model. When comparing the errors to the actual results vs the fit from the models, we found that the mean error squared for the strokes gained model is 0.38 with a standard deviation of 0.94, while the standard model returns a mean error squared of 0.75 and standard deviation of 0.98. It appears that the Strokes-Gained model is a better predictor of results in 2016 and we can conclude that the better fit model is in fact strokes gained. The PGA Tour has wisely decided to use this metric to describe a golferβs success.
The Code