On the Limits of Genetically-Optimized Homogeneous Ensembles for Credit Risk Classification
Journal of the Brazilian Computer Society · 25.06.2026 · 32 (1) · 1685-1698
DOI: 10.5753/jbcs.2026.6415 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of the Brazilian Computer Society, online erschienen am 25.06.2026. Verfasst von Ricardo Franceli Da Silva und 1 weiteren Autor:innen. Die Arbeit nennt 43 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Credit risk assessment is a challenging task with significant economic and financial impacts. It requires capturing complex nonlinear patterns and interactions between variables to accurately predict creditworthiness and minimize the risk of default. This study investigates the performance of a genetically-optimized homogeneous ensemble composed of five Multilayer Perceptron (MLP) models applied to a large-scale peer-to-peer lending dataset. Individual models achieved competitive precision scores (up to 80.57%); however, the optimized ensemble failed to surpass the best-performing individual model under economically viable conditions. Ensemble scenarios operating under a stricter classification threshold achieved higher precision gains but yielded negative Expected Profit per loan, rendering them impractical for real-world credit granting. This finding was confirmed by a robustness check, where the experiment was repeated after removing the top model. A pairwise error correlation analysis revealed consistently high correlations among base learners (0.762-0.918), with co-occurring error rates between 79.43% and 93.93%, providing empirical evidence that the base classifiers lack the predictive diversity necessary for synergistic ensemble gains. The results reveal a critical boundary condition for ensemble methods: when base classifiers share the same underlying learning algorithm, thereby lacking conceptual diversity, synergistic gains are unattainable; instead, the optimization process converges on weighting the strongest component. This study concludes that classifier diversity is a fundamental principle for an ensemble to deliver superior performance, regardless of the strength of its individual learners, challenging the assumption that optimized ensembles universally outperform their strongest individual components in machine learning.
Zitieren
APA
Silva, R. F. D., & Maciel, L. D. S. (2026). On the Limits of Genetically-Optimized Homogeneous Ensembles for Credit Risk Classification. Journal of the Brazilian Computer Society, 32(1), 1685-1698. https://doi.org/10.5753/jbcs.2026.6415
Vancouver
Silva RFD, Maciel LDS. On the Limits of Genetically-Optimized Homogeneous Ensembles for Credit Risk Classification. Journal of the Brazilian Computer Society. 2026;32(1):1685-1698. doi:10.5753/jbcs.2026.6415
BibTeX
@article{Silva2026,
author = {Silva, Ricardo Franceli Da and Maciel, Leandro dos Santos},
title = {On the Limits of Genetically-Optimized Homogeneous Ensembles for Credit Risk Classification},
journal = {Journal of the Brazilian Computer Society},
year = {2026},
volume = {32},
number = {1},
pages = {1685--1698},
doi = {10.5753/jbcs.2026.6415}
}
EUVIMED