Long-term forecasting of pork prices is important for production investment and price regulation. However, the pork price series exhibits the characteristics of non-stationary, nonlinear, and pseudo-periodic, which makes the forecasting task challenging. To deal with the intricate temporal patterns, in this paper, we propose a novel hybrid model based on the decomposition-ensemble framework and provide tailored prediction methods for different components. We begin by decomposing the original series into trend and cyclical components using the HP filter. To address the problem of pseudo-cycle, we design a segment-wise neural network, which introduces the attention mechanism to model correlations at the segment level. Then LightGBM is used to implement a dynamic regression model for the trend component. Finally, the predicted series are aggregated. Predictions are carried out in a rolling forward manner to avoid the problem of information leakage. Experiments show that our model outperforms other single or hybrid models.
Pork production has always been topping the list of meat production in China. According to the National Bureau of Statistics, the average domestic pork production in the past five years reached 49.30 million tons, accounting for 59.52% of the total domestic meat production. However, pork prices have been shrouded in frequent fluctuations. Especially in recent years, against the background of transformation and upgrading of the hog industry chain, pork prices have shown great volatility, impacting the nation's daily life, the operation of farmers' enterprises and the smooth operation of the economy. Therefore, studying the pattern of pork price fluctuations and establishing an effective longterm forecasting model are of great significance for the pig industry.
Decomposition-ensemble model is a standard method in time series analysis, and it has been widely
used in dealing with the intricate temporal patterns[
This paper is organized as follows: Section 2 introduces the data and the model we proposed. Section 3 describes the experimental settings, results and analysis. Conclusions are drawn in Section 4.
Weekly pork and hog pricing as well as monthly breeding sow stock are collected from January 2006 to May 2022 from the National Development and Reform Commission, the China Pig website, and the Ministry of Agriculture. For these datasets, a few pertinent statistics indices are computed and displayed in Table 1. (1) (2)
using the HP filter.
It is generally agreed that separating the temporal patterns of entanglement by time series decomposition can highlight the intrinsic properties of the sequence of components. Inspired by the law of market value, we choose HP filter to fit the original sequence to obtain a trend sequence that is easier to predict. The trend series is the solution to the following optimization problem: min ( − ) + ( − ) − ( − ) using the method of least squares. shown in the following formula: where yt is the original series and xt is the trend series. This optimization problem can be solved The periodic series is obtained by the difference between the original series and the trend series, as
= − = , = (3) (4) (5)
Our experiments show that autoregressive time series models relying only on historical information
cannot effectively deal with the huge fluctuations in pork and hog prices, which will be discussed in
Section 3.3, so we do extensive feature engineering to extract effective features from price series and
covariate series. According to the research of relevant literature[
On the other hand, the price series within the lookback window and its statistical characteristics (variation and variance) are used as the input of the time series model to reflect the short-term dependence of the price series. (6)
Gradient Boosting Decision Tree (GBDT) is a member of the Boosting family in the field of
ensemble learning. Once it was proposed, it attracted widespread attention for its outstanding effects.
Although the application of GBDT to time series prediction is rarely found, it does not mean that the
algorithm is uncompetitive in this problem. Shereen2021's research shows that a properly configured
Recent research shows that a properly configured GBDT model can outperform the current SOTA DNN
model in the time series prediction field[
The encoder-decoder model is widely used in the field of NLP, and attention mechanisms are
frequently applied to improve it[
The similar sub-series search algorithm has a better effect on the time series prediction problem, which shows that the correlation between time segments can reflect the correlation of time series better than point-level sequence correlation. Therefore, we design a segment-wise attention model. We use the sliding window to construct the segment list in an overlapping way as the input. Considering that the similarity of the clips is too small if the time interval is too small, we take the step size of the window movement as a hyperparameter. The input sequence is constructed as follows.: where : represents , , … , , s represents the segment length, and step represents the window moving step. Figure 4 shows the difference between our attention model and the traditional attention mechanism. Our motivation is to distinguish the difference in reasoning logic between machine translation and time series prediction. In the machine translation scenario, attention is used to align the vocabulary of the source and target languages, but in the time series prediction scenario, the historical segment similar to the current segment is not our target, but the subsequent segments. During the training phase, the input to the decoder is the ground-truth value at the previous time step, and during the prediction phase, it is the predicted value. The prediction ends when the total segment length reaches the prediction step size.
The disparity between the supply and demand for pig products is largely due to information lag. We
chose the prediction step as 25 weeks, which is the breeding cycle from piglets to killed pigs, in order
to make the prediction results support farmers' predictions of future income before investing. To
eliminate the possibility of experimental findings, we ran 10 experiments on each dataset. The test set
and training set were often dissected simultaneously in earlier studies, although this practice can result
in data leaks and unnaturally high accuracy[
All experiments are implemented in Python 3.6 and 3.8. We use a few popular Python packages. The Statsmodels package provides support for statistical correlation methods, including HP filters, STL decomposition, ARMA, and SARIMA models. SVR modelling is carried out using the LinearSVR class that sklearn package provides. We use darts to implement LightGBM because the library naturally supports historically known covariates and future known covariates. The PyTorch package is used to implement neural network models like the 1D-CNN and LSTM models.
Three statistical indicators, root mean square error (RMSE), mean absolute error (MAE) and mean absolute percentage error (MAPE), are used to evaluate the performance of the model, which are the most commonly used evaluation metrics for regression models. The formulas are as follows: 100% (9) (10) where and represent the ground truth and the predicted sequence respectively, represent the average of , and represents the length of the predicted sequence, which is 25 in this experiment.
To verify the effect of our model on the predictive ability and generalization ability, we conducted
two sets of experiments. The time range of the data used in the first set of experiments is 2006.1-2018.4,
which is the same as the data time range used in researches of Zhu et al. and Liu et al.[
This paper proposes a hybrid model based on a decomposition ensemble framework for predicting hog prices and pork prices in the next hog breeding cycle. Considering the historical dependence information of the time series and the impact of covariates that affect the price of live pigs, the dynamic regression model based on LightGBM is chosen to forecast the trend component of the price. This is done in order to ensure the stability and generalization ability of the model. In order to deal with pseudoperiodic components, we extend the recognition ability of similar subsequence search algorithm with the potent embedding capability of neural networks and propose an encoder-decoder model based on segmental attention for periodic components. Finally, we combine the trend and cyclical components to provide the forecast outcomes. The advantages of the proposed method are reinforced by experiments on real datasets.
Considering the distinct characteristics of Chinese provinces in the pork industry and the transmission of pork price fluctuations between provinces, we will consider introducing geospatial information into the pig price forecast model to achieve more fine-grained price forecasting in the future.