Kaggle Tabular ML
F1 Pit Stop Prediction
A Kaggle Playground project for predicting next-lap pit stops. I organized it as a small experiment pipeline with race-aware cross-validation, saved out-of-fold predictions, comparable configurations, and a checked submission-writing step rather than relying on one notebook run.
Period
May 2026
Tools / Tech
PythonPandasLightGBMXGBoostCatBoostScikit-learnGroupKFold
Why I built it
Rows from the same race share context, so I focused on whether the validation split represented unseen races instead of choosing the highest score from a random split.
Links
What it includes
- Uses Race and Year together as the grouping key for the main cross-validation setup.
- Builds lap, tyre, race-progress, and categorical-frequency features for tabular models.
- Compares LightGBM, XGBoost, CatBoost, group statistics, and target encoding under the same validation logic.
- Stores experiment notes, out-of-fold outputs, selected configuration, and validated submission files.
What I worked on
- Chose grouped validation after finding that random stratified folds produced a score that looked too optimistic for the race structure.
- Kept a simpler regularized LightGBM setup when additional models and feature ideas did not improve the grouped result.
- Used AI assistance to help organize experiment code and documentation while I selected the validation design, compared outputs, and decided which result was credible enough to keep.
What I Learned
- Learned that validation design can matter more than adding another model to a structured competition dataset.
- Practiced keeping rejected experiments in the record instead of presenting only the best run.
- Improved at turning a notebook-style competition task into a rerunnable workflow.