EUVIMEDEUVIMED entdecken
Open Access · CC BY-ND 4.0

Application of Cost-Sensitive Focal-Loss, Imbalance-Aware Stacking Framework for Financial Fraud Detection

Huijuan Lin

Journal of Electrical Systems · 12.06.2026 · 22 (01) · 686-695

DOI: 10.52783/jes.9675 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Journal of Electrical Systems, online erschienen am 12.06.2026. Verfasst von Huijuan Lin. Der Volltext ist frei zugänglich (Open Access, CC BY-ND 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Financial statement fraud detection is a challenging computational problem because confirmed fraud cases account for well under one percent of firm-year observations and the discriminative signal resides in nonlinear interactions among financial indicators rather than in isolated ratios. This study proposes CSFL-Stack, a cost-sensitive focal-loss, imbalance-aware stacking ensemble for financial fraud detection. The framework combines in-fold SMOTE oversampling, a self-implemented focal-loss neural network, Random Forest, Gradient Boosting, and a logistic-regression meta-learner, augmented by Bayesian cost-sensitive threshold optimization and SHAP-based interpretability. Experiments are conducted on a large U.S. firm-year panel (146,045 observations, 1990–2014, 964 confirmed fraud cases) under a strict time-aware protocol: training on 1990–2009, validation on 2010–2011, and out-of-time testing on 2012–2014. Among the six models compared, Logistic Regression achieved the highest test AUC of 0.7495, while the proposed stacking framework attained 0.7047. Results show that under severe temporal distribution shift, architectural complexity alone does not guarantee superior generalization; probability calibration, cost-sensitive threshold design, and explanation quality are equally critical. SHAP analysis highlights equity issuance, soft assets, long-term debt issuance, and tax-related variables as the most influential fraud predictors, consistent with interaction-driven fraud signals. The study contributes a reproducible, leakage-controlled, and explainable computational pipeline that advances fraud-detection methodology beyond single-classifier benchmarks.

Zitieren

APA

Lin, H. (2026). Application of Cost-Sensitive Focal-Loss, Imbalance-Aware Stacking Framework for Financial Fraud Detection. Journal of Electrical Systems, 22(01), 686-695. https://doi.org/10.52783/jes.9675

Vancouver

Lin H. Application of Cost-Sensitive Focal-Loss, Imbalance-Aware Stacking Framework for Financial Fraud Detection. Journal of Electrical Systems. 2026;22(01):686-695. doi:10.52783/jes.9675

BibTeX
@article{Lin2026,
  author = {Lin, Huijuan},
  title = {Application of Cost-Sensitive Focal-Loss, Imbalance-Aware Stacking Framework for Financial Fraud Detection},
  journal = {Journal of Electrical Systems},
  year = {2026},
  volume = {22},
  number = {01},
  pages = {686--695},
  doi = {10.52783/jes.9675}
}
Kostenlos starten →