Enhancing corn industry sustainability through deep learning hybrid models for price volatility forecasting

TL;DR

This paper introduces a new deep learning model to predict corn price changes, aiming to improve market stability and support sustainable farming practices.

Contribution

A novel multi-module wavelet-based model (TLDCF-TSD-BiTCEN-BiLSTM-FECAM) is proposed for accurate short-term corn price volatility forecasting.

Findings

The model outperforms existing methods in predicting corn price volatility with high accuracy.

It achieves MAE values as low as 0.0055 and R2 scores up to 0.9955 across five major corn-producing regions in China.

Abstract

The fluctuations in corn prices not only increase uncertainty in the market but also affect farmers’ planting decisions and income stability, while also impeding crucial investments in sustainable agricultural practices. Collectively, these factors jeopardize the long-term sustainability of the corn sector. In order to address the challenges posed by maize price volatility to the sustainability of the industry, this study proposes a multi-module wavelet transform-based fusion forecasting model: the TLDCF-TSD-BiTCEN-BiLSTM-FECAM (TLDCF-TSD-BBF) model, which is capable of accurately predicting short-term maize price volatility, thereby enhancing the sustainability of the industry. The model integrates a three-layer decomposition combined dual-filter time-series denoising method (TLDCF-TSD), a bidirectional time-convolutional enhancement network (BiTCEN), a bidirectional long- and short-term memory network (BiLSTM), and a frequency-enhanced channel attention mechanism (FECAM) to improve prediction accuracy and robustness. First, TLDCF-TSD is used to decompose the corn price time series into multiple scales, effectively separating the frequency components, extracting the signal details and trend information, and reducing the data complexity and non-stationarity. Secondly, BiTCEN designed in this paper effectively captures the short-term dependencies in the corn price data through the unique bidirectional structure and the special hybrid convolutional structure, and then accurately extracts the local features of the data, while BiLSTM mines the long-term trends and complex dependencies in the data by exploiting its bidirectional processing and long-term memory capabilities. Finally, FECAM enhances the focus on key temporal features in the frequency domain by grouping the input features along the channel dimensions and applying discrete cosine transform to generate attention vectors, improving the prediction accuracy and robustness of the model. The dataset utilized in this study was sourced from the BREC Agricultural Big Data platform, ensuring the reliability and accuracy of the corn price data for our analysis. This study utilizes price data from China’s five major corn-producing regions as a case study to demonstrate the efficacy of the proposed model in corn price forecasting. Through extensive experimentation, it has been established that the model significantly outperforms existing baseline models across various evaluation metrics. To be more specific, when dealing with different datasets, its MAE values are 0.0093, 0.0137, 0.0081, 0.0055, and 0.0101 respectively; the MSE values are 0.0002, 0.0002, 0.0001, 0.0001, and 0.0002 respectively; the MAPE values are 1.3630, 1.7456, 1.1905, 0.8456, and 1.7567 respectively; and the R2 values are 0.9891, 0.9888, 0.9943, 0.9955, and 0.9933 respectively. These data fully demonstrate the excellent performance of this model.

Figures