Machine learning feature selection shapes which signals models use to predict creditworthiness, and those choices strongly influence whether a credit score reproduces or mitigates social bias. When features correlate with protected groups because of historical segregation, economic inequality, or differential access to services, automated selection can make discriminatory patterns central to decisions rather than incidental. Solon Barocas Microsoft Research and Andrew D. Selbst UC Berkeley argue that algorithmic choices embed social structures into technical systems, turning neutral-seeming predictors into mechanisms of disparate treatment. Cathy O'Neil, author of Weapons of Math Destruction, documents how opaque models that rely on correlated proxies can amplify inequality at scale.
Feature choice and proxy risk
Selecting features by statistical criteria alone—correlation, mutual information, or regularized coefficients from LASSO—tends to favor variables that most strongly predict past outcomes. That can be useful for accuracy but problematic for fairness: a zip code or utility-payment history may predict default because of historical redlining and income segregation, not because of borrowers’ creditworthiness. Automatically removing explicit demographic fields does not eliminate this problem, since many socioeconomic features act as proxies for protected characteristics. Dimensionality-reduction methods like principal component analysis increase efficiency but reduce interpretability, making it harder to identify and audit proxy influence.
Regulatory and social consequences
From a regulatory and societal perspective, biased feature selection yields real harms: denial of credit, higher cost of borrowing for marginalized groups, and reinforcement of territorial inequities in neighborhoods. The Consumer Financial Protection Bureau has examined how alternative data can both expand access and introduce new risks, stressing the need for oversight and validation. Beyond legal risk, lenders face reputational and operational costs when automated feature pipelines produce disparate impact against culturally or geographically defined groups.
Practical mitigation requires integrating fairness-aware selection and human expertise. Domain-informed constraints, counterfactual testing, and disparate impact audits help identify when a high-importance feature is serving as a proxy for a protected trait. Interpretable modeling and documentation allow communities and regulators to evaluate trade-offs between predictive power and equity. Ultimately, the technical choices about which features enter a credit model are not merely statistical; they are decisions about territory, history, and social justice—and they determine whether scoring systems perpetuate or reduce financial exclusion.