Prediction of soybean price trend via a synthesis method with multistage model

Abstract

Soybean is an important crop, so it is very important to forecast soybean price trend, which can stabilize the market. This paper presents a synthesis method with multistage model (SMwMM) in order to identify and forecast soybean price trend in China. In the previous work, Toeplitz inverse covariance-based clustering (TICC) has been applied to cluster the prices of four variables. The research has found that there are four patterns in soybean market price, which could be explained by economic theory. This paper considers four patterns as market risk levels. Based on the clustering results, the authors used long short-term memory (LSTM) to forecast the prices of these four variables. Multivariate long short-term memory (MLSTM) is then used to classify soybean price to determine level of risk. Experimental results show that (1) the LSTM model has achieved great fitting effect and high prediction accuracy and (2) the performance of MLSTM-FCN and MALSTM-FCN is better than that of LSTM-FCN and ALSTM-FCN. Furthermore, MALSTM-FCN had a higher accuracy than MLSTM-FCN, which reached 76.39%.