Disparate Impact in FinTech Credit Scoring:

A Multi-Group Fairness Audit with Mitigation Analysis

Irina V. Pustokhina1,* Denis A. Pustokhin2

1 Department of Entrepreneurship and Logistics, Plekhanov Russian University of Economics, Moscow 117997, Russia

2 Department of Logistics, State University of Management, Moscow 109542, Russia

Received: December 19, 2025 Revised: February 09, 2026 Accepted: May 08, 2026 ⋆ Corresponding author

ABSTRACT

Machine learning credit scoring There is no intentional discrimination in models. They differentiate according to

data. Such models, trained on borrower histories binned by income, are known as stratified models. absorb and

convert the economic situation of the disadvantaged groups and convert them into differential approval rates that

last and does not depend on actual creditworthiness. This paper makes a systematic fairness An analysis of three

popular classifiers (logistic regression, On a dataset of digital lending, random forest, and XGBoost. Calibrated

against statistics of the US consumer credit market. Deploying four Specific fairness measures for each income,

gender and age group, and we observe that there is statistically and economically a disparate impact based on income.

The poorest fifth of the population earns. approval rates a whopping two decades lower than the top 20.approval

scores 20 percentage points lower than the highest. It is not the case that either quintile or logistic regression are 4/5

adverse.Neither quintile nor logistic regression are 4/5 adverse. The standard decision threshold was used and the

impact rule was applied. Neither income reweighting nor threshold calibration can get rid of the bias fully since it

doesn’t address the bias directly. sacrificing predictive performance. Threshold calibration alone can get into an

approximate parity of approval, but with a price of differential. Error rates that present lenders with equal opportunity

issues. The results have direct implications for the deployment of The application of algorithmic credit scoring in

new regulation regimes, and This contains the EU Artificial Intelligence Act and the US fair lending law.

Keywords: Algorithmic fairness Disparate impact Credit scoring FinTech lending Demographic parity Machine

learning XGBoost Bias mitigation Equal opportunity

1. INTRODUCTION

Credit scoring systems are not intentionally biased. They

discriminate by data. When a gradient-boosted ensemble

learns that applicants in the The default rate is four times as

high in the lowest income quintile as in the highest income

quintile. To qualify as quintile, it makes that relationship

part of every subsequent prediction, and then relies on those

predictions to decide who is going to be credited or not. The

attitudes and practices that led to the existence of the gap

in the first place this time it’s the same thing as the model’s

output, except that they’re all dressed up in faith. The ability

to tell the truth with statistics [1]. The result is a form of

algorithmic redlining that comes not from evil but from the

rational Enlargement of the pattern recognition of stratified

economic realities.

This is particularly true in the context of FinTech lending.

Digital credit platforms have been increasing access to people

who have been under-represented. Although traditional banks

have done the same with the same machine [2, 3]. narrowing