The Impact of Training-Testing Proportion on Forecasting Accuracy: A Case of Agricultural Export in Indonesia

Abstract

Accurate forecasting of agricultural exports is crucial for supporting trade policy and ensuring economic stability in Indonesia. This study investigates the impact of training–testing proportions on the forecasting accuracy of six models: linear regression, decision tree, optimized decision tree, neural network, Auto Regressive Integrated Moving Average (ARIMA), and exponential smoothing. Using Indonesia’s agricultural export data, model performance was evaluated under two data-splitting schemes (80%:20% and 75%:25%) with error metrics including MAE, MSE, RMSE, and MAPE. The results consistently show that statistical time series models outperform regression-based and machine learning approaches. In particular, SES achieved the lowest forecasting errors across all evaluation criteria, with MAPE values as low as 0.93%, followed by ARIMA as the second-best performer. Machine learning models, on the other hand, produced relatively higher error values, suggesting their limited ability to capture temporal dependencies in the data. Importantly, the choice of training–testing proportion did not significantly alter the ranking of model performance, indicating that model selection plays a more critical role than data partitioning. Overall, this study highlights the robustness of exponential smoothing methods as reliable forecasting tools for Indonesia’s agricultural exports and provides evidence-based insights for policymakers in designing effective trade strategies.