Replication Data for: A Boltzmann Generator Framework for Modeling and Forecasting International Trade Networks by Spelta & Bosone (2025)

This paper introduces a Boltzmann Generator–inspired framework for modeling and forecasting the global trade network. Departing from traditional econometric gravity models, the proposed method employs a conditional deep generative architecture that maps macroeconomic and geographic covariates to full probabilistic distributions of bilateral trade flows. Its energy-based formulation bridges statistical physics and international economics, enabling realistic simulations that capture both dyadic dependencies and higher-order network effects. Using data on 206 countries from 2001 to 2020, the model achieves superior in-sample and out-of-sample predictive accuracy relative to standard econometric and machine learning benchmarks, with formal statistical tests confirming robust predictive dominance across time and country pairs. Beyond forecasting, the framework facilitates counterfactual policy analysis: simulations reveal distinct propagation mechanisms of GDP shocks through the trade network, highlighting China’s central role in transmitting global disturbances.

本文提出一种受玻尔兹曼生成器（Boltzmann Generator）启发的框架，用于全球贸易网络的建模与预测。与传统计量经济学引力模型不同，本文所提方法采用条件深度生成架构，将宏观经济与地理协变量映射为双边贸易流量的全概率分布。该方法的基于能量的建模范式搭建起统计物理学与国际经济学之间的桥梁，可实现高真实性的模拟，既能捕捉二元依赖关系，又能刻画高阶网络效应。本文使用2001年至2020年间206个国家的数据集开展实验，相较于标准计量经济学与机器学习基准模型，该模型在样本内与样本外预测精度上均表现更优；经严格统计检验证实，其在不同时序与国家对维度下均具备稳健的预测优势。除预测任务外，该框架还可支持反事实政策分析：模拟实验揭示了GDP冲击通过贸易网络传播的差异化机制，凸显了中国在传递全球扰动中的核心枢纽作用。