cross-validation

Evaluation goes wrong when data splitting is treated like boilerplate. I stratify imbalanced targets, guard time order when necessary, and make sure preprocessing lives inside cross-validation. This is the difference between a model that looks good in