Problem
A bank launching a new lending product has no defaults of its own, so it cannot train a scoring model. The usual options both lose money. It can issue loans with no model, only antifraud rules and a debt-burden limit, then wait 3–6 months for short targets, retrain, cut risk and repeat. Or it can borrow the model of the closest product, which restricts issuance and later hits applicants the model has never seen. A model on a 12-month target normally means waiting 12 months.
Example from the article: Ozon Bank (32+ million customers) had one credit product and wanted scoring for a new one from day one.
Solution in brief
Use credit bureau reports of customers you already have as a proxy. Find loans of the new product's type in each report, reconstruct what happened to them after opening from the payment string, and train on that history. Align the raw bureau fields with what production computes, so the model behaves the same live.
Questions this case answers
- How to build a credit scoring model for a new lending product before it has any outcomes of its own by using loan histories of the same type observed elsewhere
- How to choose between a short and a long default target for a new product when the long outcome period delays any model
- How to build a credit scoring model for a new lending product before launch, without waiting for defaults
- How to build a training set for a new product from credit bureau reports of customers of an existing product
Tools and methods
- Credit bureau reports as a proxy dataset — past openings of a similar product serve as training rows
- Payment string — month-by-month payment history of a loan, used to build the target
- Gradient boosting on bureau + internal features, plus vendor scores — final decision model
- Logistic regression — alternative final model, compared against boosting
- Raw-to-production feature alignment — own script instead of paid aggregation
Google or a one-time email code. No password.
