Mapping Systemic Cross-Curency Risk: Using Graph Neural Networks on Liquidity Flows

Traditional macroeconomic risk assessment tools and econometric methods struggle to model the highly interconnected and multi-dimensional nature of over-the-counter (OTC) foreign exchange and derivatives markets. While classic statistical and linear algebraic models excel at evaluating market indicators in isolation, they usually assume independence of separate currency pairs. As such, they fail to capture the complex structural dependencies and vulnerabilities that manifest in cross-currency derivatives. To bridge this gap, this paper proposes an unsupervised, topology-aware computational framework that maps global liquidity as a directed currency exchange multi-graph. In this setup, individual currencies serve as topological nodes, while the directed edges represent specific transaction properties which encompass, but are not limited to, instrument type, counterparty sectors or reporting origins. An unsupervised Graph Autoencoder is implemented over this model to map the global foreign exchange market into a low-dimensional latent embedding space. The proposed pipeline is empirically evaluated using the Bank for International Settlements (BIS) Triennial Central Bank Survey of foreign exchange and OTC derivatives markets. By isolating edges with a high reconstruction error, the model manages to flag systemic anomalies and hidden asset-concentration risks before they propagate into the market. Ultimately, this approach provides central banks and regulators with a novel, but proactive computational oversight tool.