Engineering Trust in Algorithmic Lending: Explainable, Fairness-Aware Credit Risk Modeling for Financial Inclusion

Published in: Proceedings of the 7th National HBCU Blockchain, FinTech, and AI Conference
Conference: 7th National HBCU Blockchain, FinTech, and AI Conference
Conference Location: Nashville, Tennessee
Date of Conference: November 8–10, 2026
Publisher: National Center for the Study of Blockchain and FinTech, Morgan State University

Published: 2026
Publication Type: Full Paper
Pages: 182-202
Paper ID: 10

Author(s)

Rezoan Sultan
Department of Electrical and Computer Engineering, Morgan State University

Marwan Hossain
Department of Electrical and Electronic Engineering, European University of Bangladesh

Abstract

Consumer credit decisions are increasingly influenced by machine learning, but its use creates a major conflict for financial inclusion because models that are geared for predictive accuracy may encode and magnify inequalities against the very groups that inclusion-oriented lending aims to assist. This study asks whether predictive performance alone is a sufficient basis for trust in algorithmic lending. Using the UCI Default of Credit Card Clients dataset (30,000 borrowers in Taiwan), we train and compare logistic regression, random forest and gradient-boosted decision tree (XGBoost) classifiers and subject the strongest model to SHapley Additive exPlanations (SHAP) interpretation, a demographic fairness audit across sex, education and age, a feature-ablation test and a post-processing mitigation. XGBoost attains a ROC-AUC of 0.775, on par with established benchmarks, confirming that raw accuracy is not the differentiating contribution. The audit reveals moderate but real disparities in false-positive rates, with the youngest and least-educated applicants most often wrongly flagged; we also show that naively including a very small subgroup substantially overstates apparent disparity. The assigned credit limit is strongly stratified by education and age, yet removing it does not reduce the disparity, illustrating that “fairness through unawareness” is insufficient because the signal is redundantly encoded in correlated features. Group-aware decision thresholds, a straightforward post-processing solution, lower the education false-positive gap from 0.082 to 0.005 at the expense of 0.4 percentage points of accuracy. We suggest a layered evaluation approach and contend that explainability, meticulous fairness auditing and mitigation are necessary conditions for reliable, inclusion-focused credit AI.

Keywords: credit scoring, explainable AI, SHAP, algorithmic fairness, financial inclusion, trustworthy machine learning.

Attachment: View Full PDF