Rolling 7-Day Spend Per Customer

This is the shape of nearly every fraud and churn feature in production, and it is where time-based features are most often written wrongly.

Given a DataFrame of transactions with columns customer_id, timestamp and amount, return a DataFrame with one row per input transaction and the columns:

• customer_id
• timestamp
• amount
• rolling_7d_avg: the mean transaction amount for that customer over the 7 days ending at that transaction, inclusive

The window is defined in time, not in number of rows: a customer with three transactions in one day and none for a month has a very different window from one who transacts daily.

Sort the output by customer_id then timestamp.

The trap: a transaction's own window must not include anything that happened after it. A feature that peeks into the future looks excellent offline and collapses in production.