A temporal graph attention autoencoder framework quantifies geopolitical trade risk through the evolutionary embedding of international trade networks

The increasing complexity of global trade networks demands analytical tools capable of capturing higher-order relational structures and their co-evolution with geopolitical tensions. Building on the limitations of static graph embeddings such as DeepWalk and Node2Vec and on prior temporal models including EvolveGCN, DySAT and TGAT, we present a Temporal Graph Attention Autoencoder (TGAA) that integrates multi-head graph attention with Gated Recurrent Unit sequential modelling and reinforces the encoder with residual connections to mitigate over-smoothing on dense trade graphs. Using United Nations Comtrade data covering 178 countries from 2001 to 2023, the model is implemented in PyTorch 2.1 and torch_geometric 2.4 with all hyperparameters selected by grid search and cross-validation. Evolutionary features are distilled from the learned embeddings and synthesised into a three-tier Geopolitical Trade Risk Index (GTRI) whose indicator weights are derived through the entropy method rather than imposed a priori. Empirical results show that the TGAA model attains a link prediction AUC of 0.932, outperforms both static and contemporary dynamic graph baselines, and produces a GTRI that captures major geopolitical disruptions—including the 2014 Crimea crisis and the 2022 Russia–Ukraine conflict—while preserving robustness under sensitivity tests on β and λ. Augmented Toda–Yamamoto Granger tests, complemented by effective transfer entropy, confirm that the GTRI carries linear and nonlinear predictive content for the Caldara–Iacoviello GPR index across an expanded sample of twelve major economies. We interpret these findings as evidence that network-embedding-derived structural signals encode forward-looking information about geopolitical tensions that text-based indices alone cannot fully capture.