Multistep prediction for egg prices: An efficient sequence-to-sequence network

Abstract

Egg price has the characteristics of non-stationary, non-linear, and high volatility, which is more difficult to predict accurately. In this paper, we comprehensively consider the multiple factors affecting egg prices and construct a sequence-to-sequence (Seq2seq) model to study the multi-step prediction method of egg prices. Seasonal-trend Decomposition Procedure Based on Loess (STL) is first used to decompose the historical egg price series into trend, seasonal, and residual terms to reduce the interference of sample noise on forecasting performance. Then, Principal Component Analysis (PCA) is used to analyze and downscale the multidimensional factors affecting egg prices, such as feed price, laying hen seedling price, culled chicken price, duck egg price, and consumer index, to eliminate the redundant information in the data. Finally, the above-processed data were introduced into the Seq2seq network for training to establish a multi-step prediction model for egg prices. The experimental results show that the STL-PCA-Seq2seq model proposed in this paper can broadly capture the long-term dependence information of the input series and model the complex nonlinear relationships among the multidimensional factors affecting egg prices with the lowest prediction errors compared to the Long Short-Term Memory (LSTM), Gated Recurrent Units (GRU), the Informer model, the Seq2seq model, and the STL-Seq2seq model. The method proposed in this paper can reach R² of 0.9867, 0.9569, and 0.9106 at prediction steps 6, 12, and 18. With a prediction step size of 6, the RMSE is 0.131, MAE is 0.086, and MAPE is 0.813, respectively, which realizes the accurate prediction of egg price at any number of steps, and the results of the study provide a reference for the multi-step prediction of egg prices.