Forecasting Indonesia’s Oil and Gas Exports Using a VAR–NN Model

Abstract

Global economic instability and energy market volatility pose significant challenges to the accurate prediction of Indonesia's oil and gas export performance. This study develops a hybrid VAR–NN forecasting model that incorporates global oil prices and exchange rates as external variables, with the objective of improving predictive accuracy by capturing both linear and nonlinear dynamics in multivariate export time series data. A hybrid VAR–NN(11,19,8,4) model is proposed, in which the VAR component captures linear interdependencies among variables, while a feedforward Neural Network with one hidden layer of 8 neurons models the nonlinear residual patterns that the linear component fails to explain. The dataset consists of 296 monthly observations from January 2000 to August 2025, with 284 observations used for training and 12 for out-of-sample testing. Model accuracy is evaluated using MSE, RMSE, and MAPE. The hybrid model achieves MSE of 18,009.57, RMSE of 134.19, and MAPE of 8.61%, compared to 18,028.55, 134.27, and 8.61% for the standalone VAR model. While the hybrid model demonstrates marginally superior performance, the minimal difference across all metrics particularly the identical MAPE values indicates that linear dynamics largely dominate the dataset, and the contribution of the nonlinear component remains modest under the current model specification. These findings underscore both the potential and the limitations of hybrid VAR–NN approaches in energy export forecasting, motivating future research toward more expressive architectures and broader feature incorporation.