When Simpler Wins: Interpretability, Performance, and Regulatory Fit in Machine Learning Credit Scoring
Journal for Social Science Archives · 29.03.2026 · 4 (1) · 2282-2295
DOI: 10.59075/jssa.v4i1.657 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal for Social Science Archives, online erschienen am 29.03.2026. Verfasst von Rafia Noreen und 2 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
The adoption of machine learning in credit scoring presents a persistent governance challenge: models that perform best on predictive metrics are often the least interpretable, and the models regulators most readily accept perform worst. This research empirically examines that trade-off using the German Credit dataset (UCI, n = 1,000), comparing Logistic Regression, Decision Tree, Random Forest, and XGBoost across Accuracy, F1-score, and AUC. Rather than treating explainability as a binary property, a composite Explainability–Performance Trade-off Index is constructed and applied across all four models. Results challenge the assumption that model complexity drives predictive dominance: Logistic Regression achieves the highest AUC (0.804) on this structured dataset, outperforming XGBoost (0.775) and Random Forest (0.779), while also scoring highest on the trade-off index (0.600 versus 0.165 for XGBoost). SHAP analysis of the XGBoost classifier reveals that its most influential predictors, checking account status, credit duration, and credit amount, are economically interpretable but recoverable only through post-hoc methods, underscoring why SHAP improves but does not resolve the transparency deficit. The study draws on these findings to argue that current regulatory frameworks, by specifying explainability in technology-neutral but practically constraining terms, impose compliance costs that fall disproportionately on borrowers. Technically specific, risk-proportionate explainability standards would enable deployment of higher-performing models in regulated environments without abandoning the consumer protection rationale underlying explainability requirements.
Zitieren
APA
Noreen, R., Hussain, A., & Ali, W. (2026). When Simpler Wins: Interpretability, Performance, and Regulatory Fit in Machine Learning Credit Scoring. Journal for Social Science Archives, 4(1), 2282-2295. https://doi.org/10.59075/jssa.v4i1.657
Vancouver
Noreen R, Hussain A, Ali W. When Simpler Wins: Interpretability, Performance, and Regulatory Fit in Machine Learning Credit Scoring. Journal for Social Science Archives. 2026;4(1):2282-2295. doi:10.59075/jssa.v4i1.657
BibTeX
@article{Noreen2026,
author = {Noreen, Rafia and Hussain, Asad and Ali, Wajid},
title = {When Simpler Wins: Interpretability, Performance, and Regulatory Fit in Machine Learning Credit Scoring},
journal = {Journal for Social Science Archives},
year = {2026},
volume = {4},
number = {1},
pages = {2282--2295},
doi = {10.59075/jssa.v4i1.657}
}
EUVIMED