Decomposition-based long short-term memory model for price forecasting of agricultural commodities

Accurate and reliable price forecasting of agricultural commodities is one of the research hotspots, as these prices are inherently complex and random. In this paper, we integrate ensemble empirical mode decomposition (EEMD), a data-adaptive decomposition method, with long short-term memory (LSTM), a proficient forecasting technique. The aim is to formulate a novel hybrid EEMD-LSTM model to enhance the precision of predicting agricultural prices. Here, the EEMD decomposes a given price series into distinct and stable subseries called intrinsic mode functions (IMFs). These subseries are subsequently subjected to independent modeling and forecasting using LSTM models. The resultant forecasts from these subseries are ensembled to yield the final prediction for the original series. Notably, EEMD effectively addresses the limitations of empirical mode decomposition (EMD), particularly its challenges linked to mode mixing and end effects. The predictive performance of the proposed hybrid model is rigorously compared against that of individual LSTM and EMD-LSTM models for groundnut oil and soybean prices. The evaluation employs various standard metrics such as root mean square error, mean absolute percentage error, mean absolute error, and directional prediction statistics. Moreover, the performance assessment utilizes the Diebold–Mariano test, a well-established tool, to validate that the suggested hybrid model contributes to a substantial enhancement in prediction accuracy.