Agricultural commodity price forecasting through a hybrid EMD-SSA double decomposition approach

Agricultural commodity price series often exhibit high volatility, nonlinearity, and irregular fluctuations, posing significant challenges for accurate forecasting. To address these complexities, this study proposes a novel hybrid EMD-SSA-TDNN model that combines Empirical Mode Decomposition (EMD), Singular Spectrum Analysis (SSA), and Time-Delay Neural Networks (TDNN). The modelling process begins with EMD, which decomposes the original series into a set of Intrinsic Mode Functions (IMFs) and a residual component. Notably, the first IMF, typically characterized by high-frequency noise, is found to negatively affect predictive performance. Rather than discarding this component, SSA is applied to IMF1 to extract its underlying trend, thereby preserving valuable signal information. The reconstructed trend from IMF1 is then merged with the remaining components and fed into a TDNN to effectively capture temporal dependencies and nonlinear patterns. The proposed model is evaluated on maize and soybean price series using multiple window lengths (12, 36, and 48) and different Trend Rate (TR) thresholds (0%, 85%, 90%, 95%, and 100%) across 1, 3 and 6-step forecasting horizons, with performance assessed through MAE, RMSE, MAPE, and Diebold-Mariano statistical tests. Experimental results demonstrate that the EMD-SSA-TDNN model consistently outperforms benchmark approaches including SVR, EMD-SVR, ARIMA, EMD-ARIMA, TDNN, and EMD-TDNN, achieving significantly lower forecasting errors with statistical significance at the 5% level. The analysis further reveals that TR, window length, and forecasting step jointly influence prediction accuracy, where TR = 85% is generally optimal for short-term forecasting, while TR ≈ 90–95% improves multi-step performance. A window length of 36 often provides a balanced trade-off between noise reduction and trend preservation, whereas very small or very large windows may reduce stability in extended horizons.