In my first article on course history, I outlined some principles I’ve used to quantify and calculate course history. All of those principles are meaningless, however, if course history isn’t actually valuable in predicting out-of-sample outcomes. So in this article, we’ll take a look at how course history, as calculated from the principles in the previous article, holds up as a predictive metric.
For each tournament in my historical data set, I calculated how well I expected each golfer to do according to the principles from the first article. I used the residuals from those expectations to calculate a course history score. Then, for each tournament, I plotted each golfer’s course history score going into the tournament versus how they actually did relative to expectations (outlined previously in another article). Here’s what the average graph looks like:
Clearly, there’s something there. Keep in mind that course fit has already been accounted for in expectations (as best as the data allows), so this isn’t just a case of a year-over-year correlation of course fit. In fact, here’s how course history looks versus good, neutral, and bad course fits for each player:
Okay, so our mystery-ish metric is improving predictions, so it’s worth speculating a little bit about what’s going on with course history. What is it that we’re missing from general rules, but we can still pick up empirically to improve our predictions? There are a couple angles worth discussing:
Course history is course fit you didn’t pick up the first time
Definitely possible (and for that matter, anything else you didn’t pick up the first time). However, based on the above graph that shows course history’s effects on bad course fit players, I’m less inclined to believe that course history is just course fit that wasn’t properly picked up. I think there are too many other reasons why a player might fit a specific course beyond generic fit. In addition (and it’s admittedly anecdotal), I’ve seen other models and projections that don’t use course history, and the biggest differences between my data/Player Models and the other systems are on the course history-heavy players. If course history was just course fit that wasn’t picked up successfully, other systems that have that successful pickup should end up with the same conclusion as our Player Models, albeit with a more direct path. However, the player predictions are different enough that it reflects a fundamental disagreement on whether course history is predictive.
Confidence/fits the eye/the usual narrative stuff
Honestly, any of these narratives are as good an explanation as any of why course history matters. I’ve had plenty of impulses to dismiss these factors for the same reasons the analytics crowd usually does: They’re espoused by people that can’t quantify or prove it, the narratives are never thoroughly tested, etc. But in golf, I’m much more inclined to believe these factors are real. There’s no real way to quantify the degree, but golf is absolutely more of a mental game than most other sports. Mental stuff is easily the hardest variable to quantify, and at the same time can most definitely be swayed by intangible factors such as whether or not a golfer likes a course. While I’d love to be able to predict who will be mentally strong from week to week, I’ll settle for possibly picking up empirical results of mental stuff I missed the first time around.
Who cares, as long as it improves predictions
I’m actually very sympathetic to this one for a couple reasons. First, it’s kind of liberating knowing that you don’t necessarily have to know what you’re not picking up, as long as you can quantify it and fold it back in. Second, blending in empirical errors on a player-level basis is probably something every DFS Golf model should have. Models are basically sets of one-size-fits-all rules applied to all players. It’s not unreasonable that some individual players will have quirks or unique tendencies that a model doesn’t capture. If it happens consistently enough, you’re probably better off introducing correction factors for those players. That’s what course history is in a nutshell: A correction factor for when you’re empirically off.
