What does this research mean for the field?

Contrary to the dominant assumption, more complex and accurate machine learning models in credit scoring do not inherently amplify bias, but can instead produce significantly fairer outcomes on equalized odds than simpler models like logistic regression. Novelty: ClaimNovelty.CONTRADICTORY. Consensus alignment: ConsensusAlignment.CHALLENGES_CONSENSUS.

What question did this study set out to answer?

To evaluate whether machine learning models that achieve higher predictive accuracy in consumer credit scoring systematically amplify demographic bias across distinct empirical lending datasets.

September 5, 2026Journal of risk and financial managementOpen Access

The Fairness Illusion? A Cross-Dataset Audit of Accuracy and Demographic Bias in Credit Scoring Based on Machine Learning

View Full Paper

Ask AI

Bookmark

Authors

CEColin EllisUniversity of Saskatchewan

Discussion

Member takes

Implication

Comparative audit demonstrates that accurate machine learning models improve equalized odds in mortgage applicants, indicating that model complexity does not inherently amplify bias.

Key Points

To evaluate whether machine learning models that achieve higher predictive accuracy in consumer credit scoring systematically amplify demographic bias across distinct empirical lending datasets.
Evaluated four model families—logistic regression, random forest, XGBoost, and a multilayer perceptron—using 15-fold repeated cross-validation.
Analyzed two real-world datasets: the UCI Credit Card Default Dataset and the 2024 US Home Mortgage Disclosure Act (HMDA) database comprising over six million mortgage applications.
Assessed predictive accuracy alongside demographic parity difference, equalized odds difference, false positive rate difference, and calibration difference.
In the UCI dataset with modest baseline demographic disparities, model architecture choice showed negligible effects on fairness outcomes across evaluated metrics.
In the HMDA mortgage data with substantial baseline racial disparities, more accurate models achieved significantly fairer outcomes on equalized odds, while logistic regression performed worst across both accuracy and fairness metrics.
Demographic parity disparities persisted across all complex architectures, indicating residual feature-level bias embedded in the underlying data rather than model-driven distortion.

Cite This Study

Colin Ellis (2026) studied this question.

synapsesocial.com/papers/6a9bd4536b95aff0620ebe6a https://doi.org/10.3390/jrfm19090674

View Full Paper

Ask AI

Bookmark

Implication

Comparative audit demonstrates that accurate machine learning models improve equalized odds in mortgage applicants, indicating that model complexity does not inherently amplify bias.

The Fairness Illusion? A Cross-Dataset Audit of Accuracy and Demographic Bias in Credit Scoring Based on Machine Learning

Discussion

Member takes

Implication

Key Points

Cite This Study

The Fairness Illusion? A Cross-Dataset Audit of Accuracy and Demographic Bias in Credit Scoring Based on Machine Learning

Discussion

Member takes

Implication

Key Points

Cite This Study

Also Consider

Also Consider