Credit scoring fairness in a Pakistani context

Three ways to see it

1. 1

   Credit fairness has three layers. Direct discrimination: using a protected attribute (gender, religion) as a feature. Pakistani law and SBP norms forbid this. Indirect discrimination: using a feature that is a proxy for a protected attribute. Postcode often proxies for ethnicity or province; mobile phone tier often proxies for income. The model produces the same disparate outcome through a back door. Statistical bias: training data underrepresents some groups, so the model is simply less accurate for them. The third is the most common and the least discussed.

2. 2

   Way one to test in a Pakistani bank: cut decline rates by province and by gender. Compute decline rate for Punjab versus Sindh versus KPK versus Balochistan, controlling for income and product. Compute decline rate for male versus female applicants, controlling similarly. A 20 percent gap that cannot be explained by underlying credit quality is a finding. Document the gap, hypothesise the cause, propose a fix. Banks that do not do this are flying blind by SBP's likely 2027 standard.

3. 3

   Way two: thin-file applicants. A large share of Pakistani applicants have no formal credit history because they have never had a credit product. AI models trained on those with thick files will treat thin-file applicants as risk by default. This is a structural unfairness against women, rural applicants, and the young. The fix is alternative data with consent (utility bills, mobile wallet history, telco data) plus a separate thin-file model with calibrated thresholds.

The why-tree

Try this with Claude

AI-edge prompt to try: 'Acting as a credit fairness auditor for a Pakistani bank, propose ten test cuts and the statistical metric I should compute on each (decline rate, default rate, approval-to-application ratio). Order them by political sensitivity.' Use as a starter audit plan.