Fairness monitor
Gender and region are never scorecard inputs. They are held beside each application only so outcomes can be compared. The test that matters is parity at equal band.
Synthetic population of 1,500 applicants. We generated plausible differences by gender (more women in vegetables, salons and savings groups; more men in transport and hardware). These are assumptions, not survey data, so the monitor has proxy effects to find. Real monitoring starts with a lender’s applications.
- "Savings group" points correlate with gender (r = 0.36). Possible proxy; review the factor.
- "Guarantor / group" points correlate with gender (r = 0.26). Possible proxy; review the factor.
Overall
Adverse impact ratio 0.97 (lowest approval rate ÷ highest; alert below 0.8). An overall gap can reflect real differences in the inputs.
| Group | Applicants | Approved | Avg offer | Avg score |
|---|---|---|---|---|
| Men | 681 | 39.5% | US$563 | 598 |
| Women | 819 | 40.8% | US$632 | 603 |
Approval rate at equal band
Applicants in the same band should be approved at similar rates. Groups under 30 applicants are not compared.
| Band | Men | Women | Ratio |
|---|---|---|---|
| A | 93.9% · 49 | 81.8% · 55 | 0.87 |
| B | 80.0% · 140 | 74.1% · 197 | 0.93 |
| C | 53.4% · 208 | 54.4% · 263 | 0.98 |
| D | 0.0% · 195 | 0.0% · 219 | — |
| E | 0.0% · 89 | 0.0% · 85 | — |
Possible proxies
Correlation between each factor’s points and being in the “Women” group. |r| ≥ 0.25 is flagged for the credit committee to review: the factor may carry a group effect into the score.
- Savings group0.36
- Guarantor / group0.26
- Premises-0.19
- Income stability-0.02
- Loan size vs income0.02
- Officer site checks-0.01
- Cash buffer0.01
- Wallet activity-0.01
- Other digital loans-0.01
- Credit bureau0.01
- Betting-0.01
- Business tenure0.01
- History with you0.00
A flag is a prompt to investigate, not proof of unfairness. Savings-group and group-loan factors may genuinely lower risk; the question for the committee is whether they do so equally for everyone, which needs real repayment data.