International Trade Demand Forecasting Model Based on Transformer-XGBoost-LightGBM

Wenhao Wang; Zhitao Yang; Tianwen Zhao

doi:10.47852/bonviewAIA52027024

Authors

Wenhao Wang Department of Trade and Logistics, Daegu Catholic University, South Korea https://orcid.org/0009-0000-7141-2545
Zhitao Yang School of Economics and Management, Qilu University of Technology, China https://orcid.org/0009-0003-0867-4616
Tianwen Zhao Department of Trade and Logistics, Daegu Catholic University, South Korea https://orcid.org/0000-0002-0528-6496

DOI:

https://doi.org/10.47852/bonviewAIA52027024

Keywords:

international trade forecasting, hybrid models, time series forecasting, explainable artificial intelligence, supply chain management

Abstract

The global economic situation is complex and constantly changing. Under such a macroenvironment, it is extremely difficult to accurately predict the demand for international trade. This study presents a hybrid prediction model relying on the Transformer-XGBoost-LightGBM framework, aiming to significantly optimize the accuracy and stability of the prediction. This model innovatively integrates the Transformer’s ability to identify temporal correlations and nonlinear patterns over long time spans, as well as the advantages of XGBoost and LightGBM in dealing with structured features. Moreover, the model achieves the integration effect of dynamic weights through SHAP values. From the experimental results, when this model is used for short-term prediction, its weighted mean absolute percentage error (WMAPE) is 9.2%, which is 23.5% higher than that of the LSTM model and 15.8% higher than that of the model using only the Transformer. When encountering extremely severe shock events like the COVID-19 pandemic, this model demonstrated strong stability, with WMAPE remaining below 19.8% all the time, which was 38.7% stronger than that of the traditional VAR model. This model demonstrates the key trade-driving factors (exchange rate and tariff) and the main trade links between countries (the weight of the China–US channel is 0.41) through the SHAP value and the attention mechanism, thereby endowing the model with a certain degree of interpretability. In the actual deployment of enterprise supply chains, this model has optimized the inventory turnover rate by 22%, reduced the out-of-stock rate by 18%, and brought about outstanding economic benefits. The net present value over three years reached 12.7 million US dollars, and the return on investment was 310%. This study provides a high-performance, interpretable, and stable advanced tool for international trade forecasting. It has significant theoretical and operational significance for global supply chain operation and risk avoidance.

Received: 1 August 2025 | Revised: 31 October 2025 | Accepted: 13 November 2025

Conflicts of Interest

The authors declare that they have no conflicts of interest to this work.

Data Availability Statement

The data that support the findings of this study are openly available in GitHub at https://github.com/Tianwen-Zhao/AIA25.10.30.

Author Contribution Statement

Wenhao Wang: Conceptualization, Formal analysis, Resources, Writing – original draft, Writing – review & editing, Supervision. Zhitao Yang: Conceptualization, Methodology, Validation, Formal analysis, Investigation, Resources, Data curation, Writing – original draft, Writing – review & editing, Visualization, Supervision. Tianwen Zhao: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Resources, Data curation, Writing – original draft, Writing – review & editing, Visualization, Supervision, Project administration, Funding acquisition.