Global trade network prediction based on spatiotemporal graph neural network and gradient boosting model

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

This paper addresses the challenges of nonlinearity, spatiotemporal dependence, and coupling with external factors in global trade network forecasting. A hybrid model (ST-GBDT) integrating spatiotemporal graph neural networks and gradient boosting decision trees is proposed. This model constructs a multi-channel spatiotemporal graph neural network module, employing graph attention mechanisms to capture non-Euclidean spatial dependencies, using gated temporal convolution to extract dynamic temporal features, and designing feature fusion gating to effectively integrate external driving factors. Based on this, the learned high-order spatiotemporal representation is used as enhanced features input to the gradient boosting decision tree for final prediction, forming an innovative architecture of “deep representation learning + gradient boosting optimization”. Empirical research based on bilateral trade data from 187 countries worldwide from 1995 to 2021 shows that the model significantly outperforms the baseline model in various prediction metrics on the test set. Specifically, the root means square error (RMSE) (1.521) is reduced by 10.1% compared to the best baseline model, and the coefficient of determination reaches 0.890. Ablation experiments further confirm that deep spatiotemporal representation features contribute significantly to the prediction performance, with a 15.6% increase in RMSE when these features are removed, and the model maintains good robustness under extreme events such as trade frictions and pandemics. This study not only provides new methodological tools for global trade forecasting but also offers a paradigm that can be referenced for modeling complex spatiotemporal systems.