About the FastF1 Generated Data + Tyres + Laptimes

[Samyaktg]

Hello! Recently I used the api to extract weather and race data for the entire 2023 season and conduct an analysis on it and noticed 3 things which I would be happy to get context about.

1. The FastF1 generated data: Is this data being produced via some ML model? if so can I test it for my project and if not then how is it being produced?
2. The laptimes are in day:hh:mm:second.millisecond format, is there some way to make it only seconds automatically?
3. Finally upon reading the documentation i realised the the tyre types from C1 to C5 are not being displayed, we only know if its soft,medium or hard. Is there some way to get the Cnumber here? if not I working on an ML model that can use the data to predict Cnumber and could it be considered as a future contribution?

Thanks a lot to anyone taking the time to answer this question

[Casper-Guo]

1, Sometimes whole laps are generated but this is only done in very rare cases such as a crash on the first lap (source). Otherwise there is a lot of interpolation in telemetry data but certainly none that can be called ML per se. As you'll see, this project can be classified as ML-unfriendly.

2, Perhaps something like dt.total_seconds in pandas?

3, This is a proposed feature (#332) and is part of the Ergast API succession plan. The attitude of the maintainer is pretty well summarized in #507 but FastF1 is only interested in providing accurate data. Unless you make a breakthrough in explainable ML and achieve something like 99.5+% accuracy, I don't think the maintainer will be open to having the C numbers be predicted by a black box model. Also, there are the following coverage issues.

• test compounds that occasionally pop up in practice sessions
• Tyre data goes back to 2018 so you have to contend with superhard and ultrasoft etc.
• Pirelli can bring updated compounds mid-season with no previous data
• The tyre naming scheme can very well change between seasons, again giving the model no training data to work with
• etc.

[Samyaktg]

Thanks a lot for the prompt response! you said that this project is ML unfriendly, may I know why is that the case?

[Casper-Guo]

1, Unexpected things happen often and ML is not suitable for handling that

2, FastF1 generally does very minimal augmentation to the data provided by the official API. Having to audit the outputs of ML models after races to ensure correctness is not a sustainable long-term solution. Plus, FastF1 cache queries by default so it is difficult to fix wrong data already delivered to users (this has required maintenance updates in the past IIRC).

[Samyaktg]

Understood! Thanks for the explanation!

[theOehrly]

@Casper-Guo gave a great explanation of the issue here.

One additional note here, FastF1 does do data corrections and augmentation when these changes can be done with mostly absolute certainty. One simple example is the calculation of sector times that are missing in the source data. Because the sum of the three sector times is exactly equal to the lap time, a missing sector time can be calculated from the lap time by subtracting the other two sector times. But even this simple correction has additional checks to ensure accuracy.
This isn't directly relevant to the original question, but it nicely shows what kind of data correction/augmentation is reasonable to do in FastF1.

[Samyaktg]

Understood! This is an amazing project thanks for making it! I'm looking forward to using it in 2024 season :)