Agricultural Price Forecasting Using NARX Model for Soybean Oil

The non-linear, non-stationary and complicated nature of agricultural price series makes their accurate forecasting extremely challenging. In comparison to standard statistical methods, artificial neural networks (ANN) have demonstrated promising results for predicting such series. However, the incorporation of auxiliary information can improve prediction accuracy if it is closely linked to the target series. A dynamical neural architecture called a non-linear autoregressive model with exogenous input (NARX) carefully makes use of the auxiliary information to construct a data-dependent non-linear forecasting model. The study explores the performance of NARX model for the real price series of soybean oil (soybean) using soybean (soybean oil) price as exogenous inputs. NARX models outperform ARIMA, ARIMAX and ANN models in terms of RMSE, MAPE, MASE and directional statistics as evaluation criteria. Further, the Diebold-Mariano test confirms a significant improvement in its predictive accuracy.