Egg price forecasting in Beijing market using seasonal-trend decomposition procedures based on seasonal decomposition and long-short term memory

Abstract: Egg price has been attracting public attentions from every community in Beijing market. It is necessary to obtain timely information for the ?uctuation of the future egg price, particularly on the demand and supply of table eggs for human consumption. A lot of efforts have been made to accurately forecast future egg price in short, medium or long terms. However, there are many factors affecting egg prices to make the prediction challenging. In this paper, a hybrid model was proposed to forecast egg price by combining seasonal-trend decomposition procedures based on loess (STL) and long short-term memory (LSTM), denoted as STL-LSTM. In decomposition, a time series can be splitted into three components: seasonality, trends and remainder fluctuation. A more stable variance can be obtained from the non-linear, seasonal and periodic each part of egg price. Then, LSTM can be used to capture appropriate behaviors and predict precisely the trends and remainder parts of egg price, respectively, while the seasonal-na?ve method can be used to predict seasonal trends in a 12-month cycle. The results from three parts were summarized into a total price forecast. The egg price data that used in this study were collected from the China animal husbandry, covering from January 2000 to December 2018 in Beijing egg markets. The monthly data from January 2000 to December 2017 were used as training set, whereas the 12 monthly data in 2018 were used as testing set in the proposed model. The method was evaluated by using the relative error (RE), root mean square error (RMSE) and the mean absolute error percentage (MAPE). The results show that there was an overall upward trend for the egg price in the Beijing market from January 2000 to December 2018, with the seasonal fluctuation of "low spring and high autumn", and random fluctuations. The decomposition indicated that the trend component was the main contributor to egg price fluctuations, where the contribution rate decreased from 71.18% to 56.84% during the test period. The influence of seasonal and remaining components on egg prices increased in recent years, with the contribution rates of 34.24% and 8.92%, respectively. In STL-LSTM model, when the step size was given as 1, 3 and 6, the evaluating indexes were optimum: the relative error of 3.67%, 6.49% and 7.22%, the root mean square errors of 0.19, 0.33, and 0.43, and the average absolute percentage errors of 1.91, 3.53, and 4.58. In terms of the evaluating indexes, the proposed STL-LSTM model demonstrated most efficiency to predict egg prices, compared with the previous models, such as separate LSTM, support vector regression (SVR) and the autoregressive integrated moving average (ARIMA). The proposed model can be expected to extend on price predictions of other similar agricultural product in the future. The findings can provide a great potential to accurately forecast the future egg price for market strategies in animal husbandry.