BiGRU with Nadam Optimization for Modeling Indonesia's Non-Oil and Gas Trade Values

Indonesia's non-oil and gas trade sector has considerable potential, significantly contributing to the country's economic growth. However, fluctuations in exports and imports present complex challenges, particularly in time series forecasting. Complex data patterns and vanishing gradient issues in conventional models often hinder prediction accuracy. This study proposes a Bidirectional Gated Recurrent Unit (BiGRU) model optimized with the Nesterov-Accelerated Adaptive Moment Estimation (Nadam) algorithm to improve prediction accuracy and convergence efficiency. This method leverages BiGRU's ability to process sequential data bidirectionally and Nadam's advantages in stability and parameter update speed. The study utilizes monthly non-oil and gas trade data from January 1993 to October 2024, which were preprocessed using Min-Max Scaling and lag selection based on PACF. The BiGRU was trained on sequences of the previous 12 time steps. For export forecasting, the best model used 150 epochs and a batch size of 32, achieving an RMSE of 961.61 and a MAPE of 8.87%. For import forecasting, 150 epochs and a batch size of 8 yielded the best performance, with an RMSE of 1,238.26 and a MAPE of 9.02%. The results confirm that BiGRU with Nadam optimization effectively models Indonesia's non-oil and gas trade values, particularly for exports.