asdf

using data analysis to compare "pure" driving ability

hello all, as we're settling down from the danny ricc news, i found myself under the weight of a heavy workload morning and filled with angst about this long post-summer break. so of course, i decided to look at some F1 stats and thought up of this non-serious data analysis of F1 drivers for our collective amusement and/or outrage.

note, i cooked this up in literally a couple hours. i am not a data analyst, just an out-of-practice former software engineer with just enough knowledge to be dangerous. i'll try to be upfront about flaws in this but i wouldn't put too much weight on this either

the concept

i've always wondered how to compare a driver like hamilton with a driver like goatifi. they're both once-in-a-generation drivers but it's impossible to tease out their relative pure driving talent unless they're both using the same car. but hey! they both raced against russell in equal machinery, so we should be able to compare them by normalizing against russell.

the general idea is to come up with relative performance as a ratio b/w one driver to another on the same machinery. this ratio would be the trimmed mean of their laptimes in the most recent x races against one another under the same constructor.

a ratio of exactly 1 would mean that the two drivers are equivalent in driving skill. a value greater than 1 would mean driver1 is slower than driver2 (and vice-versa). the more recent the pairing, the better.

then, since many drivers end up driving for different teams and/or teammates, i can build a web (or graph) linking all those drivers together. finally, i can compare two drivers in that graph and find their relative performance by calculating the product of all those ratios as we traverse that graph from one driver to another.

the methodology

• first, data. i was amazed to learn that there is an amazing and updated store of f1 statistics via Ergast which collects some stats even from 1950. it's amazing — they have data available via an api, but i just opted to download their static tables (as of today) to load into python. i didn't do much in the way of data cleaning, and i did notice one crazy outlier by chance (and corrected it) but there could be others.

• some variables. i came up with 3 variables that i could adjust to make things easier (ie. smaller datatables for my poor ancient macbook pro).

  • MIN_YEAR: this is the lowest year i would use laptime data from. i settled on 2000 for now
  • TRIM_AMOUNT: to calculate the initial ratios, i take an average of laptimes for a given driver in a given race while attempting to exclude outliers. the trim amount is the % that want to exclude from the mean calculation. for the purposes of this analysis, i chose to set this to 0.1 meaning i would exclude the most extreme 10% of laptimes for a given race. the idea here is to remove laps where someone is pitting or on an outlap, or if there's some kind of event that causes them (and only them) to have a weird lap timing. this wasn't set scientifically, just plucked a number out of the air that felt right. 10% would be 6 laps in a 60 lap race.
  • MAX_COMP_RACES: i chose to only take a certain amount of the most recent races for any two drivers on the same machinery. i settled on 1.5x of a season's worth of races, or 36.

• data prep. i had to create many datatables for this analyis. one was for determining the unique driver pairs (UDP) under the same constructor. then i created a table calculating the trimmed mean laptimes for each of the drivers per race. some issues here are since i take the trimmed mean of each race and compare them b/w drivers driving for the same constructor in a race, it doesn't take into account things like different strategy, different tires, different starting grid positions, different upgrade packages, etc. the process assumes that everyone racing under the same constructor name in a race is on equal footing. then i calculate the ratios b/w the averaged laptimes b/w each UDP. finally, i take the (up to) 36 most recent races for each UDP and average those race ratios together to come up with a final ratio for each.

• the graph. i opted to use the networkx library as it pained me to try and conjure up dijkstra's algorithm from my memory of college courses like 20 years ago. using the UDPs and the time since they last raced together, i added drivers as vertexes and the time since their pairing as edge weights. the idea was to link two drivers together using dijkstra's to get the shortest path (as defined by time since pairing) b/w any two drivers. for example, to link hamilton to goatifi, dijkstra's would see that hamilton and russell raced together only a few days ago and russell and goatifi raced together almost 3 years ago and that's the shortest path between them. a more interesting case (and illustrative of some of the drawbacks) is linking hamilton to alonso. even tho they raced against each other in equal machinery more than a decade ago, the shortest path algorithm comes up with this route: <'hamilton', 'russell', 'latifi', 'de_vries', 'tsunoda', 'gasly', 'ocon', 'alonso'> since the sum of the times since each of those pairings comes out to a little less than 10 yrs. the difference in the final ratios is slight and at the very least in the same direction, so it seems to be generally consistent.

• the final link. here it was a matter of putting the pieces together and creating a function that can do all the math and spit out the relative performance metric. here's an example (WEIGHT is simply the number of years since their pairing, RATIO is the driver1_avg_laptime/driver2_avg_laptime, # is the number of races together that go into the computed ratio, MOST_RECENT is the last time they raced together

hamilton -> russell -> latifi:

DRIVER1              DRIVER2              CONSTRUCTOR  WEIGHT RATIO    #  MOST_RECENT
-------------------  -------------------  -----------  ------ -------- -- -----------
hamilton             russell              mercedes     0.0110 0.999316 36 2024-09-22
russell              latifi               williams     2.7899 0.986169 36 2021-12-12

overall ratio:  0.985494
OR hamilton IS 1.4506 % FASTER THAN latifi

that settles that once and for all!

here's one b/w liam lawson and danny ricc:

DRIVER1              DRIVER2              CONSTRUCTOR  WEIGHT RATIO    #  MOST_RECENT
-------------------  -------------------  -----------  ------ -------- -- -----------
ricciardo            tsunoda              rb           0.0110 0.996915 18 2024-09-22
tsunoda              lawson               alphatauri   0.9692 0.997232 5  2023-10-08

overall ratio:  0.994156
OR ricciardo IS 0.5844 % FASTER THAN lawson

and another one b/w max and lewis:

DRIVER1              DRIVER2              CONSTRUCTOR  WEIGHT RATIO    #  MOST_RECENT
-------------------  -------------------  -----------  ------ -------- -- -----------
hamilton             russell              mercedes     0.0110 0.999316 36 2024-09-22
russell              latifi               williams     2.7899 0.986169 36 2021-12-12
latifi               albon                williams     1.8508 1.005562 21 2022-11-20
albon                max_verstappen       red_bull     3.7864 1.007578 26 2020-12-13

overall ratio:  0.998485
OR hamilton IS 0.1515 % FASTER THAN max_verstappen

a fun one to project sainz v. albon:

DRIVER1              DRIVER2              CONSTRUCTOR  WEIGHT RATIO    #  MOST_RECENT
-------------------  -------------------  -----------  ------ -------- -- -----------
sainz                leclerc              ferrari      0.0110 0.999768 36 2024-09-22
leclerc              bearman              ferrari      0.5503 0.993133 1  2024-03-09
bearman              hulkenberg           haas         0.0301 1.001774 1  2024-09-15
hulkenberg           perez                racing_point 3.9589 1.008106 1  2020-10-11
perez                max_verstappen       red_bull     0.0110 1.020222 36 2024-09-22
max_verstappen       albon                red_bull     3.7864 0.992479 26 2020-12-13

overall ratio:  1.015308
OR sainz IS 1.5078 % SLOWER THAN albon

couple interesting notes here, there's a number of driver links with this combo where the driver only shared 1 race together (the ones with bearman and the hulk/checo pairing), which is a pretty weak link cumulatively. again, this is just for fun, so take it with a big wheelbarrow of salt. also, the checo/max perf disparity is one of the biggest of the current crop of UDPs, so any link that utilizes this pairing is going to be very skewed, fair or not.

the lowest hanging fruit here for improvement is probably to alter the edge weights in the graph to be some kind of combination of # of races together and time since racing together. there's also the question (brought up in the HAM/ALO linking) whether direct competition from many years ago is better than many links from recent times. happy to take suggestions here from those more knowledgeable than me.

final thoughts

this was a fun and interesting quick project to whip up. let me know if you want to see any other driver pairings (from drivers in the past 20 yrs). i'll do my best to keep my real work at bay to post any good pairings. it would take some time to put together a public google sheet to host ALL pairings and i might get to it at some point in the future, but this was just to get some feedback to see what i could improve. if anyone has some good questions/feedback on the methodology happy to look into those as well. thanks for reading!