This is the Beginning of the End
In my mini-series on course history so far, I’ve covered how to measure it, quantified its effect, and seen how well it does for specific courses. This is a pretty exhaustive list of all the things I’ve discovered about course history, but by no means do I consider the topic closed. Incorporating course history has so many imperfections baked into the process that it’s only natural to speculate about future areas to improve our understanding and application. I’ll close this series with an idle list of topics and questions for further study.
Understanding How to Weight Samples Within the Course History Window
Course history in its simplest form is taking an average level of relative performance at a course during an observation period. I don’t think that a simple average is the optimal way to weight past results: There should be some optimal weights to apply to each sample. But how do we know what those weights are?
An example: When Zach Johnson won the Masters, it was one of the coldest Masters ever played, which provided atypical conditions. Those conditions are probably not representative of future Masters conditions, so they should probably be weighed less in course history. But how much less? And what if it were a super windy Masters? Or what if the greens were slower than in years prior? How would those factors affect how much you weight results?
In order to crack this nut, I think that you need to have A) some prior expectations of what “typical” conditions are and B) some idea of what factors would make for atypical conditions. Once those conditions are quantified, you could back out some factor-to-weight curve for weighting prior course history results. The problem is in developing accurate expectations for typical conditions, since if you limit course history to a five-year window then small sample sizes prevent you from getting a robust baseline for typical conditions.
There’s a solution in there somewhere. It’s just a matter of working through the numbers.
Finding Patterns in Course History Across Multiple Courses
If a golfer dominates a given course — Sergio Garcia at Sawgrass, Jordan Spieth at Augusta, etc. — is he more likely to dominate similar courses? Answering that requires quantifying similarity between courses. It would likely involve identifying which course attributes are significant for similarity purposes (key stats for success, total distance, fairway acreage, grass type, etc.), clustering courses along those dimensions, and seeing how well players’ course histories correlate within each cluster. If course history from similar courses is in fact predictive, it could be used to supplement course-specific course history.
Understanding the Frequency and Magnitude of Course History Adjustments
I’ve shown that on average, course history correlates with future outcomes. What the “on average” part misses is how those adjustments are distributed.
Example: Let’s say that course history works one of two ways. For 10 percent of the time, you get an 80 percent boost to your performance — or 80 percent of the time you get a 10 percent boost to your performance. Over the long term, that adjustment will average out to an eight-percent boost, but that advantage will manifest itself in very different ways depending on how course history works.
Understanding this distribution is critical to determining if course history is more or less important for cash games or guaranteed prize pools. Cash games favor small predictable boosts, and GPPs favor the infrequent but huge boosts. (To some degree, that may be mitigated if course history influences ownership disproportionately, but that’s a separate subject.)
The End is What You Make of It
I don’t consider my research into course history done by any stretch, and I look forward to revisiting it down the road. Refining course history with some of these ideas probably allows for some third-level edge on how much to weigh course history, but in the meantime there are plenty of fundamentals left to cover for DFS golf analytics.
