Improving trend prediction of agricultural futures price using image encoding and attention mechanisms

Abstract The volatility of agricultural futures prices exerts a profound impact on the global economy, significantly affecting the economic conditions of both agricultural producers and consumers. Accurately predicting the trends of these futures prices is crucial for effective risk management, policy formulation, and investment decision-making. However, these prices are influenced by a multitude of factors, exhibiting complex fluctuation patterns and strong nonlinear characteristics, which make predictions exceptionally complicated and challenging. This paper introduces a novel hybrid deep learning model named ImgEnc-AttNet, which integrates image encoding techniques with attention mechanisms to enhance the model’s capability to interpret agricultural futures price dynamics. By transforming one-dimensional time series data into image representations, the model employs Convolutional Neural Network (CNN) and Long Short-Term Memory Network (LSTM) to extract rich visual patterns and temporal dependencies from the encoded images. It incorporates two types of attention mechanisms to refine feature extraction and capture deeper patterns in price fluctuations. Furthermore, we enhance the model’s feature extraction capabilities by fusing encoded image features with original numerical features, minimizing information loss during data transformation. Experimental results across two datasets demonstrate that ImgEnc-AttNet significantly outperforms benchmark models in both prediction accuracy and robustness.