TradeGraph: Learning Global Semiconductor Market Dynamics from International Trade Networks using Hybrid Graph Neural Networks

Abstract

The global semiconductor industry has become a strategic pillar of the digital economy and mainly supports the artificial intelligence technology, cloud computing, advanced manufacturing, and modern electronic devices. Forecasting international semiconductor trade flows is therefore essential for governments, researchers, and policymakers. In this paper, we propose TradeGraph, a hybrid Graph Neural Network (GNN) architecture that combines Graph Attention Networks (GAT), Graph Convolutional Networks (GCN) and GraphSAGE through an adaptive gated fusion mechanism for bilateral trade flow prediction. This research used the BACI international trade database. First, we analyze the structural evolution of the global semiconductor trade network and identify its major trading communities. Then, we develop a hybrid GNN model that we compare with three baseline GNN architectures under a temporal train-validation-test split, to predict the semiconductor trade flow. Experimental results demonstrate that TradeGraph achieves the best predictive performance and exhibits lower variability across multiple independent runs. Our results reveal that the global semiconductor market is characterized by a limited number of major trade blocs and suggest an increasing importance of Asian economies in the global semiconductor trade network over the study period. These findings demonstrate the potential of hybrid Graph Neural Networks for analyzing and forecasting complex international trade networks.