Methodology
Long-time acquaintances of Jeff know he is a huge fan of the Elo rating system. Elo is a simple way to grade competitors based on a series of head-to-head results. Unlike most sports, Formula One is not a head-to-head sport; multiple drivers take to the track at the same time or in the same session. In order to make this work, each session or race is treated as if it was a round-robin one-on-one tournament. A driver who finishes second out of 12 cars is viewed as having gone 10-1, losing to the first-place finisher and defeating the rest.
All drivers are assigned Elo ratings going into each qualifying session and race, which represents their form at that particular moment. After each event, the driver's rating will changed based on the result. In general, finishing high helps you gain Elo points, while finishing low costs you Elo points. If a driver doesn't finish a race, Elo acts as though that driver never entered the race, but the driver's One-Lap Reliability is affected (this metric helps inform the model's expectation that a driver will finish a given race).
The average driver has an Elo score around 1000, and new drivers start with 1000 points (the best racers achieve ratings of 1300+). The “K-factors” in this version of Elo — which are multipliers that determine the sensitivity and fluctuation of a driver’s rating — are more extreme in the beginning of a driver’s career. Drivers start with a K-factor of around 19, then reduce as they gain experience. Drivers always gain Elo points after “defeating” another driver and lose ground after “losing” to them.
The overall system is zero-sum, in that the total number of points remains constant before and after a session or race but given that each session or race can include drivers with a range of K-factors, there can be asymmetric point gains and losses. We adjust for this by normalizing participants’ scores after each session. Without this normalization, it is possible in the short term both for Elo deflation to occur — a new driver does poorly and gives away more points than the opponents claim — and Elo inflation to occur — a new driver does well and gains more points than opponents lose. Given the rapid driver churn in F1CRL, these effects would quickly lead to skewed rating scales.
Special thanks to Justin Moore's posts on his IRL F1 prediction model and 538's post debating the best F1 driver of all-time (Justin Moore was also involved with this).