No Data Is Not No Risk: Visibility Aware Graph-Based Inference of Business Conduct Risk

Battle Count: 58

The paper addresses business conduct/ESG risk prediction which is increasingly relevant for quantitative strategies incorporating ESG factors, event-driven trading, and risk management. The graph-based approach to identifying hidden risk in firms without prior incident records could inform stock screening, portfolio risk models, and alpha signals related to ESG controversies. However, the paper focuses on investigative prioritization rather than direct trading signals, and the proprietary nature of the data limits direct applicability.

High complexity due to: (1) large-scale graph processing with 11.4M entities and 14.1M edges requiring sampled computation graphs, (2) custom HeteroGCNII architecture with 8 relation-specific propagation channels, (3) nnPU training objective with non-negativity constraints, (4) hub handling with bounded neighborhood sampling, (5) temporal train/test split with leakage prevention, (6) multiple baseline comparisons including label propagation and random forest. Requires expertise in GNNs, PU learning, and large-scale graph computation.

The study uses proprietary data from RepRisk AG (human-curated business conduct risk database and company relationships dataset). No public code or data repository is mentioned. The methodology is described in detail, but the 11.4M entity graph and 297K labeled entities cannot be replicated externally. Hyperparameters are fully specified (learning rate 1e-3, batch size 64, 3 layers, hidden dim 64, nnPU parameters pi_p=0.12, beta=0, gamma=1).