Applying Machine Learning to Forecast Formula 1 Race Outcomes¶

L. García Tejada — M.Sc. thesis, Aalto University, August 2023

A pit stop is one of the few decisions in a Formula 1 race that a team fully controls, and getting the lap wrong can cost places that no amount of pace recovers. The question this thesis asks is whether the decision can be anticipated from race telemetry and standings alone — not to replace the strategist, but to see how much of the call is recoverable from data that is public while the race is still running.

The learning problem¶

Data points. One data point is a single driver-lap: one driver, on one lap, of one race, drawn from the 2019–2022 seasons. Laps run in wet conditions were removed, as were drivers who retired or were lapped more than three times, on the grounds that their pit-stop behaviour is driven by something other than strategy.

Features. Twenty-one variables per driver-lap, mixing categorical, numerical, binary and time-based information about the race, the driver and the lap itself. Categorical variables were one-hot encoded, numerical ones standardised; missing qualifying positions were filled with the starting position.

Labels. Two separate binary labels, each giving its own binary classification task. Has pit stop records whether the driver pitted on that lap. Good pit stop records whether the driver's position improved or held after pitting. Both are heavily imbalanced: only about 3% of laps contain a pit stop, and roughly half of those count as good. That imbalance shapes every modelling choice that follows.

Hypothesis space. Three families were compared rather than one chosen in advance: a random forest, a support vector machine with several kernels, and a feed-forward artificial neural network with ReLU hidden layers and a sigmoid output.

Loss function. The neural network was trained with binary cross-entropy, the logistic loss, and stopped early when validation loss failed to improve for five consecutive epochs. Because accuracy is close to meaningless at a 3% positive rate, model selection was driven by the F1 score throughout, with class weighting used to stop the majority class from dominating training. Hyperparameters were chosen by grid search under five-fold cross-validation, and the three families were compared on a held-out test set not used for tuning.

What it found¶

The random forest reached AUC values of 0.95 and 0.96 on the two labels, but the thesis is careful to point out that this is largely an artefact of the imbalance. On the F1 score, which unlike accuracy is not inflated by the majority class, recall was the weak point: the models miss a substantial share of actual pit stops, and performance on good pit stop was worse than on has pit stop.

The conclusion is appropriately modest, and is the interesting part: these models work as decision-support, flagging laps worth considering, rather than as standalone predictors. The residual uncertainty in a race is not a modelling deficiency to be engineered away.

Watch the defence¶

Full text¶

Read the thesis at Aaltodoc (open access)