Social Scoring and a Phone-Number-Based Approval Model

Georgy Smirnov4 Jun 2022

Problem

A bank collects many abandoned applications (drafts), but operators can contact only a random part of them. The bank wants to call first those who will pass approval, yet at that moment it knows only a phone number. Classical credit scoring does not fit: it needs the full application and queries to external services such as credit bureaus, which are not allowed before the client consents. Without a ranking, operator time is wasted on likely declines.

Example from the article: Alfa-Bank predicts approval of a credit product application from a phone number alone, and uses the prediction to choose which drafts to call.

Solution in brief

Build a client graph from bank-internal links (transfers, guarantees, referrals, shared salary project) and describe each phone number by three feature blocks: its own application history, aggregates over its graph neighbours, and a few other signals such as operator region and a monthly risk score. Train an approval model on these features and call the drafts with the highest predicted approval probability first.

Questions this case answers
  • How to rank a limited-capacity outreach list by probability of a positive outcome when the only identifier known about each person is a phone number and no external data requests are allowed
  • How to build features about an entity from the attributes of its neighbours in a relationship graph when little is known about the entity itself
  • How to prioritize abandoned loan applications for outreach when only a phone number is known?
  • How to predict whether an application will be approved before the customer submits it?

Tools and methods

  • Social graph of clients — vertices are clients, edges are typed connections, each with its own feature vector
  • Neighbour aggregation — weighted averages of neighbours' risk scores and transfer aggregates as client features
  • Approval model — predicts whether the application will be approved, using only phone-number-derived data
  • Draft prioritization — ranks abandoned applications by predicted approval probability for limited operator capacity
Continue reading — it's free

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