Cost-Sensitive Ensemble Models for Microcredit Risk Prediction: A Comparative Evaluation
Abstract
Microcredit risk prediction is challenged by class imbalance and asymmetric misclassification costs. This study evaluates cost-sensitive ensemble models and examines model-selection stability across heterogeneous credit sources. Logistic Regression, Random Forest, XGBoost, and CatBoost were evaluated on one operational microcredit portfolio and three public credit datasets using development-only preprocessing, class-balanced fitting weights, and leakage controls. Model and threshold selection used validation data with a five-to-one false-negative cost ratio and minimum specificity of 0.70. XGBoost was selected for the operational, Credit Risk, and FIN ANA sources, while Logistic Regression was selected for German Credit. On the retrospectively held-out evaluation partitions, ROC-AUC values were 0.895, 0.940, 0.461, and 0.620, with recalls of 0.364, 0.839, 0.083, and 0.483, respectively. FIN ANA showed substantial degradation under broader distribution shift. The study jointly evaluates minority-class recall, asymmetric error costs, and validation-selected operating points, showing that cost-sensitive model selection is source-dependent and ROC-AUC alone is insufficient for practical risk detection. The operational analysis represents retrospective adverse-status screening rather than prospective default prediction because temporal feature snapshots and future outcome windows were unavailable.
Related
- KESurv: A Kernel Ensemble Method for Patient-Specific Survival Prediction
- A Multimetric Evaluation of Ensemble and Neural Models for Imbalanced Credit Scoring on the German Credit Dataset
- Measuring consistency via ensemble margin and local prediction variability: Auditing decision systems in the presence of predictive multiplicity
- A hybrid car-following behavior prediction framework for snowy weather based on stacking ensemble learning
- Which Metrics Save the Most Human Annotation? Prediction-Powered Evaluation and Meta-Evaluation
BibTeX
@article{reliya2026costsensitive,
title = {{Cost-Sensitive Ensemble Models for Microcredit Risk Prediction: A Comparative Evaluation}},
author = {Reliya Indah Okta Limabri Reliya and Ari Muzakir},
journal = {Journal of Information Systems and Informatics},
year = {2026},
month = oct,
doi = {10.63158/journalisi.v8i5.1902},
url = {https://doi.org/10.63158/journalisi.v8i5.1902},
note = {CC BY},
}
RIS
TY - JOUR
TI - Cost-Sensitive Ensemble Models for Microcredit Risk Prediction: A Comparative Evaluation
AU - Reliya Indah Okta Limabri Reliya
AU - Ari Muzakir
PY - 2026
DA - 2026/10/24
JO - Journal of Information Systems and Informatics
DO - 10.63158/journalisi.v8i5.1902
UR - https://doi.org/10.63158/journalisi.v8i5.1902
AB - Microcredit risk prediction is challenged by class imbalance and asymmetric misclassification costs. This study evaluates cost-sensitive ensemble models and examines model-selection stability across heterogeneous credit sources. Logistic Regression, Random Forest, XGBoost, and CatBoost were evaluated on one operational microcredit portfolio and three public credit datasets using development-only preprocessing, class-balanced fitting weights, and leakage controls. Model and threshold selection used validation data with a five-to-one false-negative cost ratio and minimum specificity of 0.70. XGBoost was selected for the operational, Credit Risk, and FIN ANA sources, while Logistic Regression was selected for German Credit. On the retrospectively held-out evaluation partitions, ROC-AUC values were 0.895, 0.940, 0.461, and 0.620, with recalls of 0.364, 0.839, 0.083, and 0.483, respectively. FIN ANA showed substantial degradation under broader distribution shift. The study jointly evaluates minority-class recall, asymmetric error costs, and validation-selected operating points, showing that cost-sensitive model selection is source-dependent and ROC-AUC alone is insufficient for practical risk detection. The operational analysis represents retrospective adverse-status screening rather than prospective default prediction because temporal feature snapshots and future outcome windows were unavailable.
SN - 2656-5935
SN - 2656-4882
ER -