Nhãn Rủi Ro Tín Dụng: Trước Khi Train Model, Hãy Thống Nhất “Bad” Là Gì
Credit model không bắt đầu từ XGBoost hay feature engineering. Nó bắt đầu từ một thỏa thuận rất cụ thể: ai được xem là “bad”, quan sát trong bao lâu, và cohort đã đủ mature để dùng làm dữ liệu huấn luyện hay chưa.
On this page
Điểm cần nhớ
- Hãy xem nhãn rủi ro tín dụng như một thỏa thuận nghiệp vụ, không phải chỉ là một cột
bad_flag. - Một label tốt luôn cần bốn phần: điểm quan sát, sự kiện bad, outcome window và quy tắc maturity.
- Không dùng chung một khoảng thời gian quan sát cho mọi sản phẩm. BNPL, cash loan, thẻ tín dụng và behavior scorecard có nhịp rủi ro rất khác nhau.
- Kiểm tra cohort đã đủ mature, tức đã đủ thời gian để rủi ro xuất hiện, trước khi chia train/test.
- Gian lận, tất toán thỏa thuận, xóa nợ và tái cơ cấu cần rule riêng; nhét tất cả vào một flag “bad” sẽ gây khó khi theo dõi và review chính sách.
Sơ đồ dưới đây là cách nhìn tối thiểu về một label tín dụng. Điểm cần chú ý là label không xuất hiện ngay lúc giải ngân; bạn phải chờ đủ outcome window để rủi ro có cơ hội bộc lộ.
Loading diagram…
Hình 1. Outcome window nối thời điểm quan sát với thời điểm label đủ mature; nếu chốt label quá sớm, dữ liệu huấn luyện sẽ thiếu bad thật.
A label is a risk contract
A credit model does not start with XGBoost, a feature store, or hyperparameter tuning. It starts with a business contract: which risk is being modeled, where observation starts, how long outcomes are measured, and when a case is mature enough to be called Good or Bad.
If that contract is vague, a model can still have strong AUC and answer the wrong decisioning question. Risk may be reviewing 60+ DPD within 12 months, Product may care about early delinquency, while DS trains on 30+ DPD within 6 months.
This post covers the minimum set of decisions to settle before modeling: DPD, outcome windows, label maturity, Ever-90/FPD/MOB/roll-rate labels, and the traps that make a model look clean in a notebook but fail policy review.
A credit label is a contract, not a column
In a dataset, the label may look like a simple bad_flag. In production, that flag represents an agreement: where observation starts, how long outcomes are measured, which event counts as bad, and which cases should be excluded from credit-risk modeling.
DPD — Days Past Due
DPD measures how many days a borrower is overdue relative to their payment due date.
DPD = Current date − Most recent missed due date
Example: due date is March 1st; as of March 15th, no payment has been made → DPD = 14.
Common conventions:
| Label | DPD Definition | Notes |
|---|---|---|
| Bad-30 | DPD ≥ 30 at any point in window | More lenient; higher bad rate |
| Bad-60 | DPD ≥ 60 | Balanced between signal and sample |
| Bad-90 | DPD ≥ 90 | Strict; near write-off policy |
| Ever-90 | Ever reached DPD 90+ | Does not require consecutive |
No definition is correct in isolation. A label is only correct when it matches the portfolio risk appetite, product mechanics, and decision the model will support.
The “label zoo” in practice: choosing the right target
In real projects, “bad” is usually defined along three axes: (1) delinquency severity, (2) horizon/window, and (3) observation point (application vs after-booking).
Here are the most common label families you’ll see in the market:
1) Ever-delinquency within an outcome window (application PD-style)
- Ever-30/60/90 in 12M/24M: within 12/24 months from origination, the account reaches DPD ≥ 30/60/90 at least once.
- Used when the main objective is application scoring (approval decisions).
2) FPD / EPD (early-warning / early performance)
- FPD (First Payment Default): the borrower fails to make the first scheduled payment within a delinquency threshold (public sources often describe default in a 30+ DPD sense).
- EPD (Early Payment Default): delinquency/default occurring very early after origination, commonly described as within the first 3–6 months or 90–180 days.
3) MOB-based labels (months-on-book)
- MOB3 / MOB6 / MOB12: assign labels based on behavior within the first 3/6/12 months on book.
- Typical examples: “Ever-30 within MOB3” or “60+ by MOB6”.
- Useful when you need faster feedback loops or you want a model aligned to early performance.
4) Roll rate / next-cycle delinquency (revolving / collections)
- Roll rate measures the % of accounts that migrate from one delinquency bucket to a worse one in the next cycle (30→60, 60→90, etc.). This is common in credit cards and loss forecasting.
- Targets like “next-cycle 30+” or “roll 30→60” are often used for behavior scoring and collection strategies.
Outcome window: how long are you waiting for risk to show up?
The outcome window is the period from the observation point (usually loan origination) to when you assign the label.
Loading diagram…
Example with a 12-month window: a loan originated in Jan 2023 is observed through Jan 2024. If DPD ≥ 60 occurs during that period → label = bad.
This is a real modeling trade-off, not a cosmetic parameter:
- Short window (3–6 months): More data, faster training, but misses late defaults. Suitable for short-tenor products (BNPL, sub-6-month consumer loans).
- Long window (12–24 months): Richer signal, but you must wait for data to mature. Suitable for personal loans, mortgage.
Label maturity: do not split data before outcomes are ready
A cohort is considered mature when most cases have had sufficient time to exhibit bad behavior (if they are going to).
Signs a cohort is not yet mature:
- Bad rate is still rising steadily by observation month (not yet flattening).
- A high number of cases are still "open" or "pending outcome."
Practical check: Plot bad rate by vintage — one curve per origination cohort. If curves are still sloping upward at the right edge, the cohort is not mature.
Other bad events: do not hide everything inside one flag
| Event | Meaning | Typical Timing |
|---|---|---|
| Charge-off | Bank writes the debt off its books | Usually after DPD 90–180 |
| Write-off | Similar to charge-off; policy-dependent | Varies |
| Settlement | Borrower pays partial; bank closes account | After serious delinquency |
| Bankruptcy | Personal/corporate insolvency | May occur without DPD history |
| Fraud | Identity fraud, not a credit default | Must be excluded from label |
Pre-label checklist before modeling
Before starting model training, answer all 7 questions. If the answers are unclear, the problem is not the algorithm yet.
- Bad definition (DPD threshold) signed off by Risk?
- Outcome window chosen to match product tenor?
- Cohort is mature? (verify with vintage curve)
- Fraud cases removed?
- Charge-off / write-off cases: include or exclude? (per definition)
- Definition consistent between training set and monitoring (population stability)?
- Definition documented and shared with the full team before EDA begins?
Tham khảo / References
- Siddiqi, N. (2017). Intelligent Credit Scoring, 2nd ed. — Ch. 3: Bad Definition.
- Thomas, L. C. et al. (2017). Credit Scoring and Its Applications, 2nd ed. — Ch. 2.
- Anderson, R. (2007). The Credit Scoring Toolkit — Ch. 7: Data Preparation.
- Experian. What Lenders Need to Know About First Payment Default (FPD). https://experian.com/blogs/insights/first-payment-default
- CreditCards.com Glossary. Roll rate definition. https://www.creditcards.com/glossary/term-roll-rate/
- Oracle OFS Analytical Applications Docs. Delinquent Roll Rate Computation. https://docs.oracle.com/en/industries/financial-services/ofs-analytical-applications/loan-loss-forecasting/8.1.2.0.0/llfpug/delinquent-roll-rate-computation.html
- Bank of England (2024). Credit risk: definition of default (Supervisory Statement). https://www.bankofengland.co.uk/-/media/boe/files/prudential-regulation/supervisory-statement/2024/credit-risk-definition-of-default-supervisory-statement.pdf
In This Series
Credit Risk Modeling & Decisioning
Bài 2 / 8