EEMD-FCR-TDNN: A Hybrid Model for Forecasting Agricultural Commodity Prices

 This paper aims to develop a hybrid model using ensemble empirical mode decomposition (EEMD) as a decomposition technique and time-delay neural network (TDNN) as a forecasting technique to predict non-stationary and nonlinear agricultural price series. The EEMD first decomposes the agricultural price series into several intrinsic mode functions (IMFs) and a single residual. Further, the resulting IMFs and residual series are grouped into high frequency, low frequency, and a trend component with similar frequency characteristics to capture numerous coexisting hidden factors using the fine-to-coarse reconstruction (FCR) algorithm. After that, a TDNN with a single hidden layer is built to separately forecast each of the three nonlinear components. Finally, the prediction results of all three components are summed up to obtain a final output as the forecast of the original price series. The performance of the proposed hybrid EEMD-FCR-TDNN model is empirically evaluated by comparing it with several benchmark models, including the TDNN model and decomposition-ensemble hybrid models without reconstruction using monthly international maize and soybean oil price series. The results validate that the EEMD-FCR-TDNN model can significantly outperform the other models in terms of both level and directional prediction accuracy with lower computational cost.