CausalGraphX: A Counterfactual Graph Neural Network Framework for Explainable Systemic Risk Assessment

Financial systems are highly interconnected, making them prone to systemic risk where individual institutional failures trigger widespread cascades. While Graph Neural Networks (GNNs) are effective at modeling these relational structures, they typically function as black boxes that capture correlations rather than causal mechanisms. This paper asks: how can we build a GNN-based model that accurately predicts systemic risk while providing transparent, causal explanations that regulators can use for stress testing and intervention?

The authors introduce CausalGraphX, a framework that combines two primary components. First, it uses a Graph Attention Network (GAT) architecture augmented with adversarial regularization. This training objective forces the model to learn representations that are predictive of defaults while remaining invariant to confounding variables, effectively filtering out spurious correlations. Second, it employs an optimization-based counterfactual generator. Once the model is trained, this module identifies the minimal changes to an institution's capital reserves or interbank exposures required to prevent a predicted default cascade, providing a clear, actionable path for policy intervention.

CausalGraphX significantly outperforms traditional risk models (like DebtRank) and standard deep learning baselines (like GCNs and standard GATs) in predicting default cascades across both Erdős–Rényi and Barabási–Albert network topologies. Beyond predictive accuracy, the framework generates counterfactual explanations that are highly sparse and plausible. In simulated stress tests, the model identified interventions that reduced systemic losses by over 78% with minimal capital injections, demonstrating that the model successfully captures the underlying causal drivers of shock propagation rather than just statistical associations.