Research on Deep Learning Models for Forecasting Cross-Border Trade Demand Driven by Multi-Source Time-Series Data

Abstract

Against the backdrop of rapid changes in the global foreign trade environment, cross-border commodity demand exhibits complexity and uncertainty influenced by multiple factors including seasonality, price indices, promotional rhythms, and exchange rate fluctuations. To enhance the accuracy of inventory and transportation capacity planning for foreign trade enterprises, this study constructs a deep learning forecasting model integrating multi-source data. This model is based on approximately 84,000 time-series data points spanning 3.5 years and covering 11 countries from an export enterprise. The model integrates features including historical orders, price indices, promotional schedules, international holidays, search trends, exchange rates, and shipping cycles, employing TCN combined with attention mechanisms for sequence modeling. Experimental results demonstrate that compared to benchmark models such as ARIMA, PROPHET, and LSTM, MAPE, SMAPE, and RMSE improvements range from 14% to 21%, with more robust performance in forecasting demand peaks. The study demonstrates that multi-source time series fusion effectively captures the dynamic characteristics of cross-border demand, providing reliable predictive support for digitalized foreign trade operations. Wine export operations serve as the validation scenario.