Problem
A telecom operator lets subscribers spend money they have not yet paid (a trust payment) and expects it back within a set number of days. The operator cannot collect debt: the only lever is to block services on the number. Before scoring, Altel decided amounts with a table of just two features, average monthly spend (ARPU) and subscriber lifetime. Wrong decisions mean direct monetary losses.
Example from the article: the old table gave one of four denominations (0 means refusal) by ARPU and lifetime only; the new model scores all subscribers on all available data.
Solution in brief
Train a model on all subscriber data, then convert its 0–1 score into risk grades (A, A+, B, … F) that map to denominations, with a ROC-based cutoff and hard filters on top. Features are selected by Information Value and checked for stability by PSI. Losses fell almost twofold, but the day-4 repayment target proved too harsh.
Questions this case answers
- How to turn a model score into approved amounts by mapping it to risk grades with a cutoff, when a wrong decision causes direct monetary losses
- How to train a binary risk model when the event rate is about 5 percent, choosing between oversampling and stratified sampling, and select features by information value and stability
- How to build a scoring model that decides which subscribers get a trust payment
- How to turn a 0–1 model score into approved amounts using risk grades
Tools and methods
- Information Value (IV) — feature selection: how well a feature separates defaulting and non-defaulting subscribers
- Population Stability Index (PSI) — stability of a feature between training data and live data; retrain or keep the model
- Risk grades A to F — turn a score into an approved denomination in an explainable way
- ROC curve threshold — minimum score below which nothing is approved
- Stratified sampling vs oversampling — handling a rare default class
Google or a one-time email code. No password.
