Amid intensified volatility in global financial markets, the price trends of soybean futures, as one of the world’s most strategic agricultural commodities, have drawn significant attention from governments, investors, and farmers. Soybean prices are influenced not only by supply-demand dynamics but also by complex factors such as international trade policies, climate change, and financial market sentiment, resulting in highly nonlinear and volatile price behaviors. Developing a precise and robust predictive model to support scientific decision-making has become a shared challenge for academia and industry.
A study led by Professor Hui Liu from Central South University, published in Frontiers of Agricultural Science and Engineering (DOI: 10.15302/J-FASE-2024599), introduces a hybrid deep learning model named “ICEEMDAN-LZC-BVMD-SSA-DELM”, which significantly enhances the prediction accuracy of soybean futures prices. By integrating multi-stage data preprocessing and intelligent optimization algorithms, the model addresses limitations of traditional methods in noise handling, parameter tuning, and generalization capabilities, offering new insights for risk management in agricultural financial markets.
Conventional soybean price prediction models often rely on single algorithms, struggling to handle high-frequency noise and complex volatility in market data. To overcome this, the research team designed an innovative hybrid data preprocessing strategy. First, they applied Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN) to decompose raw price series into multiple sub-signals, effectively isolating fluctuations at different frequencies. High-frequency noise components were then identified through “complexity evaluation” and subjected to secondary decomposition. During this process, the team introduced the Beluga Whale Optimization (BWO) algorithm to automatically adjust key parameters of Variational Mode Decomposition (VMD), eliminating reliance on manual empirical tuning and enabling more precise and efficient signal decomposition.
For the prediction phase, the team adopted a Deep Extreme Learning Machine (DELM) as the core model. Unlike conventional neural networks, DELM drastically reduces training time by randomly generating hidden layer parameters, but its randomness may lead to unstable outputs. To address this, the Sparrow Search Algorithm (SSA) was incorporated to simulate the foraging and predator-avoidance behaviors of sparrow populations, dynamically optimizing DELM’s weight parameters. This “bio-inspired” optimization strategy not only accelerates model convergence but also enhances adaptability in complex market environments.
To validate the model, the team analyzed soybean futures and ETF price data from Chinese, Italian, and U.S. markets, covering nearly 1500 trading days. Experimental results showed that the new model achieved an average Mean Absolute Percentage Error (MAPE) below 0.1% across all three markets, significantly outperforming traditional single models such as LSTM and GRU. Notably, the model demonstrated exceptional performance in predicting prices of Italian soybean ETFs. As financial derivatives tracking commodities, ETFs exhibit high correlation with futures markets but are more susceptible to short-term speculation. The new model effectively filtered such noise through its secondary decomposition strategy, proving its generalizability for complex financial instruments.
This research advances price prediction interpretability through algorithmic innovation and provides data-driven support for governments to regulate agricultural reserves, investors to design trading strategies, and farmers to plan production cycles.

DOI: 10.15302/J-FASE-2024599