Formula 1

Formula 1 is an interesting sport to test the framework with because it introduces a new paradigm where each team has to compete against multiple other drivers and constructors rather than most other sports, where a team only competes against one other team at one time. This means that the results are more sensitive to the performance of other teams, which is a good test of the framework.

The data is less clean than I had hoped, but better than most other sports. I did some manual cleaning where I spotted errors, but I think the results are good, despite likely having overlooked some errors. The graphs are interactive; I recommend isolating certain years or clicking on the legend to isolate certain drivers/constructors.

The scores themselves are somewhat meaningless absolutely, but are useful for relative comparison. For example, Lewis Hamilton was the best driver recorded since 1981 in his 2020 season. The scores are also useful for comparing drivers across different years. For example, Micheal Schumacher was dominant for a decade, whereas no other driver has been that dominant for multiple years in a row. We can also think of the scores as a combination of current performance and expected future performance. For example, Lando Norris has a very high score despite being quite young. We can also see this in George Russell’s anomalous 2022 season, when he moved to Mercedes and scored well, showing large upside potential, although his score reverts somewhat in 2023. Interestingly, if we run a Gaussian mixture model, we see 3 broad groupings for constructors—the top 1-2 teams, a middle group of ~4 teams, and the rest. Williams seems to be underrated in the community, being around the bottom of the middle group or top of the bottom group recently, rather than at the bottom of the bottom group.

I’m not terribly familiar with Formula 1 history and I imagine there are some anomalous results due to insufficient data cleaning, but I think the results are good enough to give you a general sense of how the framework works, its general use cases, and effectiveness. If you have any questions or suggestions, feel free to reach out over email.