Choosing an Attribution Window for Delayed Conversions

An ad platform predicts P(purchase | click). Analysis of historical data shows purchase delays after the click distributed roughly as: 50 % within 1 hour, 75 % within 1 day, 90 % within 3 days, 97 % within 7 days, 99.5 % within 30 days. The team retrains daily.

1. Propose an attribution window and explain the trade-off you are making.
2. Describe exactly which rows can be included in today's training set and today's evaluation set, and why.
3. The business complains the model under-predicts conversions for a new product category. What label-related cause would you check first?