You are building a model for a consumer lending product to predict whether an applicant will default after origination. The raw data includes application details, basic bureau attributes, and recent account activity, but many fields are not immediately useful in their original form.
What is feature engineering in this context, and how would you decide which new features to create, transform, or exclude before training a classification model? Explain how your feature choices would affect model performance, interpretability, and risk of leakage.