Hi, Thank you, I really appreciate that!

To be honest, I didn’t experiment with survival analysis during the competition. I approached the problem directly as three probabilistic classification tasks for the 7, 90, and 120-day adoption windows.

The competition framing made that approach quite natural since the targets were already defined as binary outcomes within fixed time horizons, and the evaluation metric (log loss / ROC-AUC) aligns very well with standard probabilistic classifiers.

Because of that, I focused most of my effort on feature engineering and iterative modeling with tree-based methods, and CatBoost worked particularly well given the tabular structure and categorical variables in the dataset.