Australian GP 2019

Australian GP 2019

The new season is starting and the teams have given us a first glimpse of their 2019 cars and performance. The most visible change to the cars this year will be the front wing. The new wings will be wider but, most importantly, simpler. The vortex-generating fins have been stripped off in an effort to minimize the turbulence created by the front wing and help the cars race closer to each other. Will this improve racing in 2019? I am not too optimistic but we’ll find out soon.

In the first week of the tests in Barcelona, the Ferraris seems to have the upper hand but without knowing the fuel loads and how much each team was sandbagging, we cannot be sure about their actual performance. Without the model viewing any of the test data, its predictions are only based on the past years performance of the drivers and cars. I’m expecting the model to get more accurate as the season goes forward since it will get access to any changes in competitiveness of the teams. That’s why you see some unexpected results in the table below. The model is confused, isn’t it?

Driver Qualifying Prediction Actual Qualifying Result Race Prediction Actual Race Result
Bottas 4 2 2 1
Hamilton 1 1 1 2
Verstappen 7 4 4 3
Vettel 5 3 3 4
Leclerc 8 5 5 5
Magnussen 19 7 8 6
Hulkenberg 14 11 14 7
Raikkonen 2 9 6 8
Stroll 20 16 18 9
Kvyat 12 15 15 10
Gasly 16 17 17 11
Norris 10 8 9 12
Perez 15 10 11 13
Albon 17 13 13 14
Giovinazzi 13 14 12 15
Russell 18 19 19 16
Kubica 6 20 20 17
Grosjean 9 6 7 R
Ricciardo 3 12 10 R
Sainz 11 18 16 R

3 thoughts on “Australian GP 2019

  1. Hi, loving your posts! I just have a bunch of questions 🙂 For driver features regarding previous season performance (e.g. drivers’ championship position last year), how do you deal with rookies that are new to F1? Also, considering that drivers may have moved to different teams for the current season, is it even worth having drivers’ past season performance as input features? Also, if you have features like driver’s no. of podiums this current season up to the race (which increases as the season progresses), is it possible that the model might think Hamilton is a bad driver after the first race of the season because he ‘only’ has 1 podium so far? Thanks so much

    1. Hi Alby. Great questions, let me answer them one by one.

      Features without proper values
      As you said, there are many cases where a feature does really apply. For instance, driver’s championship position last year or even driver’s last race position (in case he either DNF or even DNS). In all these cases, I simply replace the ‘null’ values with a fixed value (e.g. -1). An alternative would be to add new boolean features that have a value of 1 if the respective feature is missing, otherwise 0.

      Drivers moving teams
      I think it’s worth having driver’s past performances as features even if they changed teams. You could argue the same holds true even if a driver does not move, because the same constructor can improve from season to season. Intuitively, driver’s position from year to year do not change that much (unless something big happens), so I find it as a valid feature.

      Cumulative features
      Your thinking is right. But… there are some ways to mitigate this issue. One option would be to ‘normalize’ the no. of podiums to ‘percentage of podiums’ so that it kind of stabilizes after the first few races.
      Here is what I’m actually doing though. I keep the feature as-is (i.e. number of podiums). Then, during modeling, my model actually calculates for each possible pair of drivers the probability of driver_1 finishing above driver_2. The input features of this model is the subtraction of the drivers original features. So the final feature would show how many more podiums driver_1 has compared to driver_2. This still may mean that the features keeps increasing as the season progresses or may not. In any case, the model (a tree-based model) decides what is the optimal cut point. So maybe all values above 1 have the same impact on the final prediction. Makes sense?

Leave a Reply

Your email address will not be published. Required fields are marked *