BIMSA >
BIMSA Digital Economy Lab Seminar
Inferring the Global Currency Network Based on Deep Graph Learning
Inferring the Global Currency Network Based on Deep Graph Learning
Speaker
Time
Friday, March 6, 2026 3:00 PM - 4:00 PM
Venue
A3-2-303
Online
Zoom 435 529 7909
(BIMSA)
Abstract
Existing studies fail to trace out global currency network owing to the limited public data. This study proposes a novel deep learning model—Attentive Spatial-Temporal Hyperbolic Graph Convolutional Network (AST-HGCN)—to infer bilateral preferences for the USD and EUR using cross-border banking flows from the BIS. The AST-HGCN employs a sandwich structure of two temporal and one spatial module, integrating currency inertia and cross-border capital flows. The spatial module leverages Hyperbolic Graph Convolution to capture the core-periphery and small-world properties of global currency networks. Our model outperforms Euclidean-based methods, improving network structural fidelity by 13.9% and forecast accuracy by 9.5%. We empirically investigate the inferred USD and EUR networks from 2002–2022. Results show that the inferred preferences closely approximate available snapshots of USD- and EUR-denominated cross-border bank loans. The dynamic evolution of USD and EUR preferences diverges across income groups and varies with the global financial cycle. Moreover, both networks exhibit core-periphery and small-world structures that amplify liquidity shocks and systemic risk propagation. This suggests that our model innovatively captures the strategic complementarity between borrower and lender currency choices through core-node dominance and regional preference convergence.
Speaker Intro
Ji Qi is a postdoc at BIMSA & Tsinghua University. Her research focuses on the digital economy, graph neural network, and knowledge graph.