Strategic and Environmental Variables as Predictors of Formula 1 Race Outcomes: A Machine Learning Approach

Abstract

With Formula 1 generating US$4 billion in annual revenue, race strategy is one of the most consequential decisions in motorsport, yet remains dependent on human judgment and proprietary simulation tools. This paper investigates whether machine learning models trained on historical race data can learn relationships between strategic and environmental variables and race outcomes. Data from the 2023, 2024, and 2025 Formula 1 seasons was collected through the OpenF1 API, covering 903 driver-race sequences across 23 circuits and 14 drivers. Features included tire compound, stint duration, tire age, track temperature, humidity, rainfall, and starting grid position. Four models were developed and compared: a Transformer encoder, xLSTM, Random Forest, and XGBoost classifier, each predicting finishing position across four classes. The Transformer achieved the highest mean accuracy at 64.8% (±1.6%), while sequential models consistently outperformed tabular models by approximately 5 to 8 percentage points. Mixed-effects logistic regression identified starting grid position as a statistically significant predictor across three of four classes. Predictability was strongest for podium (P1-3) and outside-the-points (P11+) finishes, while P4-5 and P6-10 outcomes showed substantially greater uncertainty.  These findings demonstrate that strategic and environmental variables contain predictive signals for race outcomes and that sequential models capture temporal patterns in race strategy.