Resampling And Cost-Sensitive Random Forest and Xgboost For Non-Performing Sharia Financing Prediction
Equivalent: Jurnal Ilmiah Sosial Teknik · 31.10.2026 · 8 (5) · 2071-2083
DOI: 10.59261/jequi.v8i5.467 ↗Kurz gefasst
Fachartikel in der Zeitschrift Equivalent: Jurnal Ilmiah Sosial Teknik, online erschienen am 31.10.2026. Verfasst von Febriany Syafitri und 1 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY-SA 4.0).
Abstract (Originalsprache)
Background
Class imbalance presents a significant challenge in predicting non-performing financing (NPF) in Islamic financial institutions, potentially reducing the effectiveness of minority-class detection.
Objective
This study aims to compare six resampling techniques combined with Random Forest and XGBoost under standard and cost-sensitive learning to identify effective configurations for NPF prediction.
Methods
A quantitative experimental approach was applied to 8,814 historical Sharia financing records from 2016 to 2025. Six resampling techniques—SMOTE, ADASYN, Borderline-SMOTE, SMOTE-ENN, Distributed Gaussian (D-GA), and Distributed Poisson (D-PO)—were evaluated using stratified cross-validation. Model performance was assessed using the area under the precision–recall curve (AUPRC), macro F1 score, NPF precision, and NPF recall, followed by evaluation on a reported holdout dataset of 1,763 records.
Results
D-GA combined with cost-sensitive XGBoost achieved the highest cross-validation AUPRC (0.5503). On the holdout dataset, Borderline-SMOTE combined with standard XGBoost achieved the highest AUPRC (0.5942) and NPF precision (100.00%), although NPF recall remained 27.75%. SMOTE-ENN combined with cost-sensitive XGBoost achieved the highest NPF recall (43.54%), whereas SMOTE combined with standard Random Forest achieved the highest macro F1 score (0.7089).
Conclusion
Different configurations demonstrate distinct trade-offs between NPF detection and prediction reliability. The findings support objective-specific model selection, although inconsistent class distributions, unavailable baseline results, and data from a single institution limit generalizability.
Zitieren
APA
Syafitri, F., & Mauritsius, T. (2026). Resampling And Cost-Sensitive Random Forest and Xgboost For Non-Performing Sharia Financing Prediction. Equivalent: Jurnal Ilmiah Sosial Teknik, 8(5), 2071-2083. https://doi.org/10.59261/jequi.v8i5.467
Vancouver
Syafitri F, Mauritsius T. Resampling And Cost-Sensitive Random Forest and Xgboost For Non-Performing Sharia Financing Prediction. Equivalent: Jurnal Ilmiah Sosial Teknik. 2026;8(5):2071-2083. doi:10.59261/jequi.v8i5.467
BibTeX
@article{Syafitri2026,
author = {Syafitri, Febriany and Mauritsius, Tuga},
title = {Resampling And Cost-Sensitive Random Forest and Xgboost For Non-Performing Sharia Financing Prediction},
journal = {Equivalent: Jurnal Ilmiah Sosial Teknik},
year = {2026},
volume = {8},
number = {5},
pages = {2071--2083},
doi = {10.59261/jequi.v8i5.467}
}
EUVIMED