GRAPH NEURAL NETWORKS FOR SYSTEMATIC RISK AND CLUSTER-BASED DIVERSIFICATION

Abstract

The paper presents a study of Graph Neural Networks as a superior alternative to Euclidean-based correlation models for understanding market interconnectedness. Traditional diversification strategies often assume that assets are independent or have linear correlations, failing to account for the complex, non-linear contagion paths in global markets. This study models the stock market as a Dynamic Heterogeneous Graph, where nodes representing assets are connected by edges that signify multiple relationship types, such as supply chain links, common institutional ownership, and lead-lag return correlations. The study proposes a Graph Attention Network architecture to learn Risk Embeddings that quantify an asset’s vulnerability to shocks propagating through the network.

The purpose of the work is to develop a topology-aware diversification framework that identifies hidden clusters of risk. By using Graph Neural Networks, the research seeks to move beyond sector-based diversification toward a data-driven graph-cluster approach that accounts for the structural complexity of the modern financial ecosystem and its inherent systemic dependencies.

The methodology encompasses constructing dynamic asset graphs using thresholding and k-Nearest Neighbors on Pearson and Spearman correlations, deploying Message Passing Neural Networks to aggregate risk signals from neighboring nodes, applying Hierarchical Risk Parity on the learned graph embeddings instead of raw returns, and correlating node centrality with asset beta and tail risk metrics.

The scientific novelty. The research introduces the Multi-Relational Risk Embedding, a novel vector representation of assets that captures higher-order dependencies. Unlike linear Principal Component Analysis, this graph-based embedding identifies how a shock to a specific node flows through the network to affect distant nodes, providing a predictive measure of systemic fragility based on the topology of the financial system.

The practical value lies in offering risk managers a stress-test visualization tool that identifies super-spreader assets within a portfolio. It enables the construction of anti-fragile portfolios that are structurally decoupled, reducing the likelihood of simultaneous collapses during liquidity crises and providing a more granular view of diversification.

Conclusions. Network topology is a critical dimension of market risk that is often ignored by classical models. The study finds that graph-informed clustering identifies risk concentrations that traditional methods miss, leading to a significant improvement in the Sortino ratio and a substantial reduction in maximum drawdown during periods of market contagion.