Label Maturity and Selective Labels

Chargebacks on your platform arrive with this empirical distribution: 50 % within 14 days, 80 % within 45 days, 95 % within 90 days, 99 % within 150 days. Your team currently trains nightly on all transactions from the last 120 days, labelling anything without a chargeback so far as "legitimate."

1. What bias does this training setup introduce, and which transactions does it hurt most?
2. Propose a concrete maturity-window policy and explain the trade-off it makes.
3. Separately, the model declines a rising number of transactions from a segment over the last month. How would you find out whether that segment actually has more fraud, or whether the model is simply creating its own blind spot?