This project started from my interest of F1 (Essere Ferrari) as well as wanting to learn more about ML. The goal of this model is to predict the finishing order of drivers using historical race data.
I trained a gradient-boosted decision tree model (LightGBM), selected for tis strong performance on tabular data, ability to model nonlinear feature interactoins, and built-in regularization. Input features inclued lap times, grid position, track conditions, and environmental variables. Because race outcome prediction is a ranking problem, the NDCG@1/3/10 metric was chosen as it emphasized correctness within meaninful outcome groups (winner, podium, points), rather than individual classification accuracy
Furthermore, the features were engineered with shifted windows in pandas to prevent data leakage, so that only information avilable prior to each race was used for prediction. The final model utilized 16 temporal features and achieved NDCG@1/3/10 scores of 0.98, 0.97, and 0.94 respectively.
Currently, I am trying to predict average race times using a seperate LSTM model architecture, that could also serve as another temporal feature to model my results. Additionally, I am performing robustness checks and validation experiments to investigate potential sources of my overly optimistic performance to ensure that the model generalizes beyond the training data.