Enhanced cost-sensitive ensemble models with performance-weighted ensemble feature selection for financial statement fraud detection
Matin N. Ashtiani, Bijan Raahemi · Machine Learning with Applications · 2026
Detecting fraud in financial statements is challenging because fraudulent statements are rare, raw accounting variables and financial ratios are highly redundant, and practical deployment requires interpretable risk scores rather than black-box decisions. We study FSFD on a balanced AAER–Compustat case–control benchmark and propose a leakage-safe stacking framework that combines Random Forest (RF), XGBoost (XGB), and TabTransformer (TT) through a cost-weighted logistic-regression meta-learner and a performance-weighted ensemble feature selection module. The framework is designed to integrate granular raw accounting variables (RAV) with standardized financial ratios (FR) while controlling redundancy through nested cross-validation, out-of-fold stacking, fold-wise calibration, and within-fold feature selection. On the benchmark, the full stack achieves mean out-of-fold ROC–AUC of 0.858 and held-out ROC–AUC of 0.845, while a compact 60-feature variant ( ≈ 45 % fewer variables than the 109-feature model) attains comparable held-out discrimination. The results indicate that combining RAV and FR is more effective than using either representation alone, and that a curated subset of features can preserve performance while improving transparency and efficiency. Because the evaluation benchmark is case–control balanced, threshold-dependent precision and recall should be interpreted as benchmark-specific rather than as deployment estimates; practical use in lower-prevalence settings would require recalibration and threshold re-optimization. Overall, the framework is intended as a human-in-the-loop screening aid for audit and investigative review, not as a substitute for final professional judgment.