PGA Championship Boosting Methodology
First of all, what a fun DFGolf season it has been! This was my first season following fantasy golf, but I personally have found it to be a very enjoyable daily fantasy format, with lots of good data out there and a pretty practical application of some cool data science methods.
Now that we’ve got that out of the way… PGA Tour Championship. Yeesh. Our algorithms are built to project performance of players starting from the same start line, and probably work best under fairly stable pricing ranges. We have neither of those qualities in this PGA season finale. Up until this morning, we had projections that were run from our standard algorithm, which we should note does consider Vegas win odds, which are boosted for players starting with stroke leads. However, it’s reasonable to assume that the algorithm ingests these odds in a somewhat muted fashion, using them as indicators for how skilled a player is and projecting how they will drive total fantasy performance. We do have boosted Vegas win odds for Justin Thomas and other leaders, the algorithm probably reads this as: “this is a really good golfer who has this probability against a field that operates under the same variability that we see in all tournaments”. Variability where it is reasonable for the 20th most probable golfer in a field to outshoot the most probable golfer - they simply have to gain more strokes than the best golfer.
But with the new Tour Champ format, not only does the 20th most probably golfer have to gain strokes on the field leader Thomas, but they have to do it 5-to-10-fold (for most players). In reality, there are really only a small handful of golfer that can push JT off the top of the leaderboard, and it would take a catastrophic collapse and multiple improbable surges to push JT (and others) outside of the Top 5 or Top 10. It is because of this that relying simply on inflate Vegas win odds isn’t enough to handicap the lead that the top golfers have built in to the start (in my opinion).
So what have we done? Well, we started by looking at exactly how much of a players’ total fantasy point total we can expect from their finish bonus.
So as a players’ win probability increases, the expected finish fantasy point bonus we can expect from them becomes decreasingly addictive to their total score. Based on this relationship, we sought to project how much of each players’ unboosted projection was part of the finish bonus component. From that we said: “ok let’s leave the rest of the projection as is (because JT starting at -10 won’t help him gain hole or streak bonus fantasy points), and let’s boost the rest, the perceived “finish bonus component”.
We were next tasked with figuring out how much of each players implied Vegas win probability was added or subtracted due to their starting position. This is a bit tough to define precisely, but I felt a good approach was to look at their win probability from the BMW Championship, scale the win probability to the total win probability of the pool of 30 players that are playing in the Tour Championship this weekend, and calculate the percentage defat as (scaled BMW win prob - Tour Champ win prob)/(Tour Champ win prob). Then, through a log-log model of log(finish bonus) ~ log(win probability), using past tournaments for data, determined a coefficient for determining by what percent a players’ finish bonus increases with a x percent increase in win probability. The relationship between the scalar values of these two variables is plotted below.
From here, we can create a boosting or negative-boosting multiplier for every player based on (a) their win probability for this weekend - to determine how much of their base projection should be boosted and (b) the differential between their win probability this weekend and last weekend - to determine by how much that projection proportion should be boosted.
Like any non-scientific approach, it works under some assumptions. But hey, that’s statistics for you. I think given the information we have access to and how our algorithm functions, it is a reasonable boosting approach.
Looking ahead to 2020
When the next season kicks off, it is my full intention to continue running our weekly projections, updating our data with new tournament data, etc. However, with football kicking off in a couple weeks, and eventually the basketball tipping off and running simultaneously alongside NFL, we are going to be swamped this fall. Such that it will be very difficult to work on making significant changes to our PGA tools. I think we will take the approach of maintaining them, but not re-inventing them. However, I do have the desire to continue developing our PGA offering. To me that means developing new tools and improving the accuracy and predictability of our current tools. This is an undertaking we will hopefully get to in early 2020, when football season has ended, basketball season and content has stabilized and baseball season is still hibernating. I definitely have some ideas for ways to improve our projection algorithm and want to experiment on those, as well as spending time looking for and cleaning new data. So in summary, we won’t be taking a PGA hiatus per se, but we will be dialing back the new content creation for the rest of 2019. Hope you all had a fun and profitable season and we’re looking forward to picking back up where we left off next golf season.