Breaking or Reinforcing the Cycle? Longitudinal Impacts of Bias-Correction Techniques on Feedback Loops and Sustained Financial Inclusion in Machine Learning Credit Scoring

Abstract

In machine learning–driven credit scoring, fairness interventions can produce unintended long-term effects because models reshape borrower behavior and data over time. Building on Juan C. Perdomo et al.’s performative prediction framework and Pagan et al.’s classification of feedback loops (sampling, feature, outcome, and model loops), this study examines how bias-correction techniques interact with dynamic lending environments. Using 1.2 million U.S. loan applications (2018–2024) plus synthetic emerging-market simulations, we modeled multi-cycle credit systems where decisions feed back into future borrower data. We evaluated adversarial debiasing, pre-processing reweighting, causal proxy mitigation, and threshold adjustments across metrics such as demographic parity, AUC-ROC, Brier scores, credit score progression, and inclusion indices. Results show stark trade-offs. Simple threshold adjustments increased initial approvals for Black and Hispanic applicants by 12–15%, but by later cycles, feedback effects widened disparities by 22% due to proxy discrimination and degraded alternative data. In contrast, dynamic resampling aligned with feedback-aware modeling sustained an 18% equity uplift with less than 3% rise in default rates, even under downturn simulations. Overall, static fairness fixes can backfire. Longitudinal, system-level design—combined with multi-horizon stress testing and richer data-sharing—is essential to achieve durable financial inclusion rather than short-term fairness gains.