Algorithmic Bias in Credit Decision-Making: A reflexive thematic analysis of data practitioners in a South African Bank

Abstract

Algorithmic bias in credit decision-making is increasingly recognized as a socio-technical phenomenon where historical and structural inequalities become embedded in data-driven systems. In financial services, biased models can contribute to discriminatory lending and economic exclusion. This study explored how data practitioners at a South African bank perceive diversity and inclusivity in training data, population representation, and geographic/demographic bias, as well as the controls, considerations, and model training integrity practices used to prevent or mitigate bias. Exploratory qualitative interviews were conducted with 10 bank-employed data analysts/data scientists and analyzed inductively in Atlas—ti (v21) using thematic analysis and collaborative codebook development. Four themes emerged. Participants identified insufficient demographic diversity in training data as a threat to fairness and representativeness, noting testing of race, gender, and age but limited consideration of disability. They emphasized full-population representation and warned that narrow geographic sampling and proxies such as suburb “area rating” can encode socioeconomic and demographic bias. Controls included structural-bias testing, rules against modelling on narrow subpopulations, monitoring, and re-evaluation. Findings suggest that mitigating bias in banking requires integrated sociotechnical governance, inclusive representative data practices, transparency, human oversight, and diverse development teams, rather than reliance on fairness metrics alone.