A Comparison of Deep Learning and Time Series Models in Forecasting Agricultural Exports: Evidence From Vietnam's Pepper Exports to the United States

This study analyzes and forecasts the export value of Vietnamese pepper to the United States by integrating traditional time series models, namely SARIMA-GARCH, with a deep learning approach based on the gated recurrent unit (GRU). The dataset consists of monthly time series observations spanning from January 2002 to December 2025, which are utilized to examine trends, seasonality, volatility, and nonlinear characteristics of export values. The SARIMA-GARCH framework is employed to model both the conditional mean and variance structures, whereas the GRU model is designed to capture nonlinear relationships and long-term dependencies inherent in the data. Empirical results from out-of-sample forecasts indicate that both models perform effectively; however, the GRU model demonstrates superior predictive accuracy in terms of MAE, RMSE, MAPE, and the coefficient of determination (R2). Furthermore, the Diebold–Mariano test and robustness analysis provide additional evidence supporting the stability and generalization capability of the GRU model. This study offers empirical insights into the effectiveness of combining time series and deep learning approaches in forecasting agricultural exports. It also provides valuable implications for policymakers and businesses in formulating sustainable development strategies for Vietnam's pepper industry.