ERP-RiskBench: Leakage-Safe Ensemble Learning for Financial Risk

TL;DR

A rebuilt experimental framework for ERP financial risk detection using ensemble machine learning is presented, showing that procurement control features, especially three-way matching discrepancies, rank as the most informative predictors.

Abstract

Financial risk detection in Enterprise Resource Planning (ERP) systems is an important but underexplored application of machine learning. Published studies in this area tend to suffer from vague dataset descriptions, leakage-prone pipelines, and evaluation practices that inflate reported performance. This paper presents a rebuilt experimental framework for ERP financial risk detection using ensemble machine learning. The risk definition is hybrid, covering both procurement compliance anomalies and transactional fraud. A composite benchmark called ERP-RiskBench is assembled from public procurement event logs, labeled fraud data, and a new synthetic ERP dataset with rule-injected risk typologies and conditional tabular GAN augmentation. Nested cross-validation with time-aware and group-aware splitting enforces leakage prevention throughout the pipeline. The primary model is a stacking ensemble of gradient boosting methods, tested alongside linear baselines, deep tabular architectures, and an interpretable glassbox alternative. Performance is measured through Matthews Correlation Coefficient, area under the precision-recall curve, and cost-sensitive decision analysis using calibrated probabilities. Across multiple dataset configurations and a structured ablation study, the stacking ensemble achieves the best detection results. Leakage-safe protocols reduce previously inflated accuracy estimates by a notable margin. SHAP-based explanations and feature stability analysis show that procurement control features, especially three-way matching discrepancies, rank as the most informative predictors. The resulting framework provides a reproducible, operationally grounded blueprint for machine learning deployment in ERP audit and governance settings.

Figures (11)

• Figure 1: Synthetic ERP data generation and augmented test suite construction. CTGAN is fitted only on training folds. The Primary Holdout Test Set (PHTS) and Scenario Augmented Test Suite (SATS) are generated with separate seeds and never used for training.
• Figure 2: End-to-end leakage-safe experimental pipeline for ERP financial risk detection. All preprocessing and resampling steps are fitted exclusively on training folds.
• Figure 3: Nested cross-validation structure. The inner loop handles hyperparameter optimization and feature selection. The outer loop provides unbiased performance estimation. All resampling is confined to inner training partitions.
• Figure 4: Stacking ensemble architecture. Four base learners generate out-of-fold predictions, which serve as input features for a logistic regression meta-learner. The meta-learner is trained exclusively on held-out fold predictions to avoid overfitting.
• Figure 5: Precision-recall curves for selected models on BPI P2P under condition A5. AUPRC values shown in legend. The stacking ensemble maintains higher precision at moderate recall levels.