aflaisler / f1_2017

This project aims to predict the fastest and average laps of a formula one driver for the season 2017.

2017 Formula one driver's performance estimator

This project aims to predict the fastest and average laps of a formula one driver for the season 2017.

Motivation

Formula one is by nature a data-driven sports. Using the latest technologies, engineers are streaming insane amount of data about the car and its surroundings in (almost) real time, from the temperature of the limited slip differential to the tire wear. Each races however, contain many analog properties which are hard to predict of which the weather is one of the most important. Teams comes out with different strategies to counter them. They have in 2017, a choice of 7 different tires types and 5 different fuel types. In parallel the regulations constantly evolve to make sure that team with lower budget can still compete. This year, amongst the many changes, the tires are 25% larger and the engines more powerful.

The idea behind this project is, are we able, to predict, with a reasonable error rate the performance of each driver, by adjusting for the weather, the driver, the circuits properties and the qualifications times.

The iteration modeling process

The model has been built following the CRISP-DM process (https://en.wikipedia.org/wiki/Cross_Industry_Standard_Process_for_Data_Mining).

Once the data - understanding/cleaning phase is done the iteration process constantly highlight how it essential to constantly review the business understanding after each iteration. By doing this I discovered properties that did not sound important at first (like the direction (clockwise or anti-clockwise) or the type of the circuit Street/Road/Race circuit).

Getting the data

The first and main source of the data comes from the Ergast motor racing results API (http://ergast.com) maintained by Chris Newell which summarize data from 1950 onwards.

External, data sources were added along the way using unstructured sources like @f1debrief who posted lap-times and tires compounds for 100 long runs on Twitter.

Understanding the signals

One of the challenge has been understanding which are the key components which are involved with the process of a driver making good time. Overall in this model they can be grouped, by:

• the team adjustments
• the drivers' performance
• the competition effect
• the circuits' properties
• the regulation Overall effect / year
• the pit stops strategies

Modeling against the residuals / GradientBoosting

The natural unpredictability of a driver not finishing a race is on its own a very hard problem to solve. Here we account for the probability of the combination driver/team of not finishing the race using a second model applied on the residuals.