Will they pit on the next lap?

Each row is a lap. The label is PitNextLap, yes or no. AUC is the score, a number from 0.5, a coin flip, to 1.0, perfect ranking. You want the cars that will pit near the top of your list.

I concatenated an extra F1 strategy table onto the official train file when it was attached, dropped the exact duplicate rows, and treated Year and PitStop as categories. Then n-grams: I glued pairs of category columns together so the trees could see driver-and-compound as one thing, not two lonely columns.

Three models, same 10-fold split: CatBoost, XGBoost, LightGBM. They vote. That is the ensemble. Two accepted csvs both came back 0.94962 public and 0.95000 private. Rank 1090 of 3022. Upper third, not a win.

I named the kernel P1 F1 PitStop Ensemble v1.0 because I wanted 0.95372, the number at the top when I started. I did not get there. 0.950 is what the file says.

What I would not claim. This is playground tabular, the kind of contest you use as a gym. The extra rows help only if they are the same sport. The n-grams are a cheap trick. The mix of three trees is the default that usually works, and this time it did, a little.