Scoring on Non-Bank Data

Evgeny Pateha25 May 2024

Problem

A non-bank company (marketplace, payments) wants to lend to its customers or use a risk score in its funnel. It has no credit histories, and a cautious first approval policy leaves too few bad borrowers for training. A model that has not seen bads in all their variety can give unpleasant surprises. Outside data is not unique and can disappear at any moment.

Example from the article: a call center offered installment cards by phone, and a significant share of customers who agreed were then rejected. Calling only customers with a high score raised the approval rate and cut acquisition cost several-fold.

Solution in brief

Build the score from the company's own behavioral data, with a proxy target when real defaults are missing (for sellers: a sharp drop in sales). Produce all features with one parameterized generator used in research and production, precompute them daily, and combine them with live application data at scoring time. Monitor input data and score distributions with PSI.

Questions this case answers
  • How to build a risk score when a company has no credit history and too few observed defaults, using a proxy target built from its own behavioral data
  • How to keep one feature pipeline shared by research and production scoring and monitor input drift with a population stability index
  • How to build a credit scoring model when a non-bank company has no credit history and too few bad borrowers
  • How to create a proxy target for scoring from the company's own behavioral data

Tools and methods

  • Proxy target — label sellers whose revenue fell sharply as potential defaulters, refine labels as real loans mature
  • Feature generator in Spark SQL — one script builds the full pool or one model's short list
  • Daily precomputation into Postgres — fast lookup for a sub-second microservice
  • Population Stability Index (PSI) — stability of the score, partner scores and every feature
  • Logistic regression — blending own score with external scores
Continue reading — it's free

Google or a one-time email code. No password.