RNN and GNN based prediction of agricultural prices with multivariate time series and its short-term fluctuations smoothing effect

Abstract

In this study, we investigate appropriate machine learning methods for predicting agricultural commodity prices. Since environmental factors including weather affect price fluctuations of agricultural commodities, we constructed a multivariate time series dataset combining wholesale prices of four agricultural commodities in South Korea, six weather variables, and week numbers. We adopted two prominent prediction methods based on recurrent neural networks (RNN) and graph neural networks (GNN): one is the stacked long short-term memory, and the other consists of two GNN-based methods, the spectral temporal graph neural network (StemGNN) and the temporal graph convolutional network. Also, we utilized a univariate prediction model as a control to evaluate the effectiveness of the multivariate approach for predicting agricultural commodity prices. In this investigation, we applied five different smoothing time window lengths to evaluate the effect of mitigating short-term fluctuations on the predictive performance of the models. The experimental results showed that the mitigation of short-term fluctuations had a greater impact on improving the performance of multivariate prediction models compared to the univariate prediction model. Among the multivariate prediction models, the GNN-based network outperformed the RNN-based network. In view of the trained model, we analyzed the main weather variables affecting agricultural commodity prices by utilizing the adjacency weight matrices in the self-attention mechanism of StemGNN.