Collection Score: Thiết Kế Label, Treatment Và Recovery Strategy
Collection score không chỉ tìm người dễ trả. Nó là bài toán phân bổ capacity: gọi ai, dùng treatment nào, kỳ vọng thu được bao nhiêu và chi phí có đáng không.
On this page
Điểm cần nhớ
- Collection score là bài toán allocation: capacity, treatment, timing, recovery và cost.
- Label phải khớp objective: payment, cure, roll-forward, recovery amount hoặc uplift.
- Không nên trộn mọi DPD bucket vào một model nếu action và behavior khác nhau.
- Propensity trả lời ai dễ trả; uplift trả lời action có làm họ trả thêm không.
- Lift@topK, recovery captured và net recovery thường hữu ích hơn AUC trong vận hành.
Collection score is capacity allocation
A collection score is not only about predicting who will pay. It helps decide who gets SMS, who gets a call, who receives restructuring treatment, and who is not worth expensive effort right now.
Label follows the objective
Payment, cure, roll-forward and recovery are different objectives. They need different labels, windows and populations. A 1-30 DPD model should not be casually mixed with a 90+ DPD or write-off model.
Propensity is not uplift
Propensity predicts who is likely to pay. Uplift estimates whether an action makes payment more likely. Calling the highest propensity accounts can waste capacity if those accounts would have paid anyway.
Practical takeaway
For collection, evaluate lift@capacity, recovery captured, roll-forward reduction and net recovery after cost. AUC is useful, but it does not answer the operational question alone.
In This Series
Credit Risk Modeling & Decisioning
Bài 8 / 8