Design a Discrete-Time Weekly Churn Model on Person-Period Data

A meal-kit company has 50,000 customers who started a subscription at some point in the last two years. For each customer you have the start date, the cancellation date if any, and weekly behavior (orders, skips, app opens, support contacts). The data cut is today; about 55 percent of customers have not cancelled. The business wants to know each week which customers are likely to cancel in the next 4 weeks, and the expected number of remaining weeks, for retention offers and LTV. The team has LightGBM and wants to use it.

1. Explain how you would build the training table: rows, label, features, and how censored customers are treated. Estimate the number of rows if the average customer has 20 weeks of observation.
2. How do you get "probability of cancelling within the next 4 weeks" and "expected remaining weeks" from the fitted classifier?
3. How do you split train and test so there is no leakage, and what happens if you split rows at random?
4. A teammate suggests a simpler alternative: one binary classifier, "cancelled within 8 weeks of start", on customers with at least 8 weeks of history. Compare it to your design.