Machine Learning Models for Detecting Hidden Collusion Networks in U.S. Corporate Finance

Abstract

Hidden collusion networks in U.S. corporate finance present a significant challenge to market integrity and regulatory oversight. Such networks are difficult to detect due to the indirect and distributed nature of collusive behavior, which is often embedded within legitimate financial transactions and governance relationships. This study aims to develop and evaluate machine learning models capable of detecting latent collusion structures by combining firm-level financial indicators with relational and governance-based network information. The goal is to uncover clusters of firms that exhibit coordinated behavior while providing interpretable insights for regulatory decision-making. A multi-stage detection framework was implemented, incorporating classical machine learning classifiers, tree-based ensembles, graph neural networks (GNNs), and hybrid models combining feature-driven and structure-driven learning. Corporate entities were represented as nodes within financial and governance networks, with edges encoding ownership ties, shared executives, transactional dependencies, and temporal co-movements. Models were evaluated using precision, recall, F1-score, AUC-ROC, AUC-PR, and network consistency metrics, while robustness analyses examined performance under class imbalance and sparse labeling conditions. Graph-based models outperformed traditional baselines, achieving F1-scores up to 0.91 and AUC-ROC values up to 0.94. Hybrid ensembles that combined tree-based and graph-based predictions achieved the highest overall performance (F1-score = 0.93, AUC-ROC = 0.95, AUC-PR = 0.91). The models successfully identified densely connected collusive clusters and key hub firms, highlighting the importance of relational and temporal features over isolated firm-level indicators. Ablation studies confirmed that financial metrics alone were insufficient to detect coordinated behavior without network structure. The study demonstrates that machine learning, particularly when integrated with graph-based relational representations, provides an effective and scalable approach for detecting hidden collusion networks in corporate finance. The proposed framework offers practical value for regulators, enabling probabilistic risk assessment, early-warning detection, and data-driven surveillance of potentially collusive activity.