AI credit scoring: alternative data for thin-file Pakistanis and the bias problem

Three ways to see it

1. 1

   Credit scoring is the process of estimating the probability that a borrower repays. The traditional model uses bureau data: credit history, declared income, existing debt. In Pakistan that bureau is eCIB, run by SBP. The problem is structural: 70% of working-age Pakistanis have no eCIB record because they never had a formal loan or card. They are not bad risks; they are invisible risks. AI credit scoring tries to make them visible by using alternative data, which is any signal that correlates with repayment behaviour without sitting in the credit bureau.

2. 2

   Six families of alternative data are used in Pakistan today. One: telecom data, including airtime top-up patterns and bill payment regularity, available through telco partnerships. Two: mobile-money data from JazzCash, Easypaisa, and SadaPay, showing inflow consistency, savings behaviour, and merchant payments. Three: utility bill payments, available through 1Link aggregator queries. Four: smartphone metadata such as device model and app installation patterns, with consent. Five: e-commerce activity from Daraz and similar platforms, when shared by partner. Six: psychometric tests adapted to Urdu, where a fifteen-minute mobile quiz produces a behavioural score.

3. 3

   The model itself is usually a gradient-boosted tree (XGBoost or LightGBM) or a deep tabular network, trained on a labelled set of past borrowers with known repayment outcomes. The output is a probability of default and a tier (A, B, C, D), which the credit policy then maps to a loan size and price. Performance is measured by Gini coefficient, KS statistic, and crucially by the rate of default in each tier after deployment. A model that scores 0.4 Gini on paper but produces 25% default in tier C is worse than a simpler scorecard.

The why-tree

Try this with Claude

AI-edge prompt: 'I run credit underwriting at a Pakistani digital bank. I have data on 50,000 borrowers including telecom top-ups, mobile money inflows, eCIB record (where present), and 12-month repayment outcomes. Walk me through the scoring model I should build, the bias checks I must run before launch, the four most likely regulatory questions, and how I would defend the model in a press inquiry alleging gender discrimination.' Save the press defence; if it sounds defensive rather than principled, redesign the model.