Although I did not end up writing articles as much as anticipated in 2021, I did keep up as always with tracking IndyCar stats, including my Elo ratings. During the 2021 season I made some updates to how the ratings are calculated. This was partially an exercise to make sure my championship simulation was getting the best data possible, and also part of a regular process to make sure the model is tuned properly.
There were not any major improvements, but overall the Brier score, the metric I use most often to evaluate the Elo ratings, did improve for the seasons in my data after making the adjustments. 2021 ended up with the third best Brier score of the seasons in my dataset (2008-Present), which I am pretty happy with.
Going into 2021 I had a goal to hopefully have Elo correctly predict 70% of driver matchups. That is, for any two drivers in a race, I wanted to be able to predict the winner of that matchup 70% of the time. I ended up predicting 67.2% of matchups correctly, which is really not too bad. Considering I only made the one time update to the formula, I was more so hoping it would just have a better year than before without really doing much additional data work. To hit the 70% mark (which I only did in 2010), I will likely need a bit more behind the scenes work so I do not need to rely on just a good year for Elo to hit it.
I am not sure how much time I will devote to reworking Elo as opposed to working on some of the other projects I have in mind for The Single Seater, but any major updates or accomplishments with Elo will be relayed here. I’ll put a goal down of 70% again, but without much additional work, I expect to be in about the 65-67% range, which is pretty typical for my ratings.