Practical Fairness Constraints in Credit Scoring for Unbanked Populations: A Pareto-Optimal Analysis of Gender and Intersectional Disparities
This paper proposes a fairness-constrained credit scoring framework using mobile money data that significantly reduces gender-based approval gaps and regulatory violations while maintaining high accuracy, though intersectional analysis reveals that such constraints alone cannot fully resolve compound socioeconomic disparities affecting low-income women.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the modern financial world, a person's ability to borrow money often depends on a digital shadow they have left behind: a record of past loans, credit card payments, and bank transactions. For hundreds of millions of people, particularly in Sub-Saharan Africa, this shadow does not exist. They are "unbanked," meaning they have never used a formal bank account, so traditional systems have no way to judge whether they are trustworthy enough to lend to. This exclusion traps individuals and small businesses in a cycle where they cannot invest in their futures or protect themselves from economic shocks. To solve this, lenders are turning to alternative data, such as the history of mobile money transactions, which can reveal how a person manages money even without a bank account. However, as computers learn to read these new patterns, there is a growing fear that they might simply repeat old mistakes, unfairly denying loans to women or the poor because the data reflects historical inequalities rather than true ability to repay.
This is the challenge that researchers Muadh Abdullateef Olamilekan and Nguyen Anh Van Trang set out to address. They asked a fundamental question: Can we build a computer system that grants credit fairly to everyone, including those who have never had a bank account, without sacrificing the system's ability to predict who will actually pay back their loan? To find the answer, they did not use real mobile money records from Africa, as those are difficult to obtain for research. Instead, they used a large, public dataset of loan applications that mimics the complexity of alternative data. They trained a computer model to predict loan defaults and then tested three different methods to force the model to treat men and women equally. One method adjusted the data before training, another changed how the computer learned during training, and the third simply adjusted the final decision rules after the computer had finished learning.
The researchers discovered that the most effective approach was the one that adjusted the final decision rules. By carefully shifting the threshold for approval for different groups, they were able to nearly eliminate the gap in loan approval rates between men and women. In their tests, the gap shrank from nearly nine percentage points down to less than half a percentage point, a reduction of almost ninety-five percent. Crucially, this massive improvement in fairness came at a very small cost to the system's overall accuracy. The model's ability to correctly identify good borrowers dropped by only a tiny fraction, suggesting that lenders do not have to choose between being fair and being profitable. The system became significantly more inclusive, approving more loans overall, while still maintaining a high standard of risk management.
However, the study also revealed a deeper, more complex truth about fairness. While the computer model successfully treated men and women equally on average, it could not fix the inequalities caused by income. When the researchers looked at the intersection of gender and wealth, they found that the poorest women were still far less likely to get a loan than the richest men. The gap between these two specific groups remained large, even though the overall numbers for men and women looked equal. This finding suggests that simply telling a computer to be fair to one group, like women, is not enough to solve the broader problem of economic inequality. The model was fair in a narrow sense, but it could not overcome the structural disadvantages faced by the poorest applicants.
The researchers also noted that while the system made fair decisions, the confidence scores it generated were not always perfectly calibrated, meaning the exact probability numbers it gave might need adjustment before being used in a real-world bank. They emphasized that their work is a proof of concept, showing that fairness constraints can be applied effectively to credit scoring. They did not claim to have solved the entire problem of financial exclusion, but they provided a clear path forward. By using a method that balances the need for accuracy with the need for fairness, financial institutions can begin to extend credit to the unbanked without reinforcing the very biases that kept them out in the first place. The study concludes that with the right tools, it is possible to build a financial system that is both smart and just, though it will require constant vigilance to ensure that fairness extends to all layers of society, not just the surface level.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.