Forecasting Foreign Trade in Türkiye’s Agri-Food Sector: A Comparative Analysis of SARIMA, LSTM, and GRU Models

Abstract

In a global economic conjuncture shaped by geopolitical tensions, climate shocks, and supply chain vulnerabilities, the accurate forecasting of foreign trade data underpins proactive policy design for national food security and economic stability. This study systematically compares the forecasting performance of Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) models, two modern deep learning approaches, and the traditional Seasonal ARIMA (SARIMA) model to forecast Türkiye's agriculture, food, and beverage foreign trade series. Monthly data for the period 2004-2020 were used to train the models, and their forecasting performance was measured using MAE, MSE, RMSE, and MAPE metrics over the test period 2021-2025. The results of the analysis indicate that the most appropriate model depends on the structure of the series. Whereas the SARIMA model outperforms in export series with regular seasonal patterns, the GRU model performs better in import series with complex nonlinear dynamics. These results underline the need to adopt a data-driven and problem-specific modeling strategy rather than relying on a universal \"best model\". In light of these findings, this study recommends the adoption of a hybrid \"model portfolio\" approach that combines SARIMA models for predictable seasonal trends and deep learning methods for managing volatile and shock-prone environments to enhance the resilience of national trade strategies.