Can online search data improve the forecast accuracy of pork price in China?

• Author: Amin W. Mugera,Shanying Chen,Liwen Ling,Dabin Zhang
• DOI: http://doi.org/10.1002/for.2649
• Date: 01 July 2020
• Published date: 01 July 2020

Received: 9 May 2019 Revised: 4 November 2019 Accepted: 3 January 2020

DOI: 10.1002/for.2649

RESEARCH ARTICLE

Can online search data improve the forecast accuracy of

pork price in China?

Liwen Ling1Dabin Zhang1,2 Shanying Chen1Amin W. Mugera3

1College of Mathematics and Informatics,

South China Agricultural University,

Guangzhou, PR China

2Faculty of Megadata and Computing,

Guangdong Baiyun University,

Guangzhou, PR China

3School of Agriculture and Environment,

University of Western Australia,Perth,

Western Australia,Australia

Correspondence

Dabin Zhang, College of Mathematics and

Informatics, South China Agricultural

University, Guangzhou 510642, PR China.

Email: zdbff@aliyun.com

Funding information

National Natural Science Foundation of

China, Grant/AwardNumber: 91746102,

71971089; Natural Science Foundation of

Guangdong Province, Grant/Award

Number: 2016A030313402

Abstract

Online search data provide us with a new perspective for quantifying public con-

cern about animal diseases, which can be regarded as a major external shock

to price fluctuations. We propose a modeling frameworkfor pork price forecast-

ing that incorporates online search data with support vector regression model.

This novel framework involves three main steps: that is, formulation of the

animal diseases composite indexes (ADCIs) based on online search data; fore-

cast with the original ADCIs; and forecast improvement with the decomposed

ADCIs. Considering that there are some noises within the online search data,

four decomposition techniques are introduced: that is, wavelet decomposition,

empirical mode decomposition, ensemble empirical mode decomposition, and

singular spectrum analysis. The experimental study confirms the superiority of

the proposed framework, which improves both the level and directional predic-

tion accuracy. With the SSA method, the noise within the online search data

can be removed and the performance of the optimal model is further enhanced.

Owing to the long-term effect of diseases outbreak on price volatility, these

improvements are more prominent in the mid- and long-term forecast horizons.

KEYWORDS

animal diseases, Baidu index, online search data, pork price forecasting, SVR

1INTRODUCTION

China is estimated to have produced and consumed

approximate 53 and 55 million metric tons of pork in 2017,

which accounted for 48% and 50% of the world's total

amount (U.S. Department of Agriculture, 2017). Obvi-

ously, hog production is one of the most important sectors

in China's agricultural economy. Meanwhile, pork price is

also an essential component of the consumer price index

(CPI), and the CPI in mainland China is usually regarded

as the “China pork index” (L. Wu, Liu, & Yang, 2016). The

correlation coefficient between pork price and CPI is 0.68,

which is much higher than for other livestock products. In

the past several years, the price of pork has experienced

exceptional volatility due to various influential factors.

These large fluctuations not only increase the risks faced

by hog producers but also have adverse effects on people's

daily lives. In this sense, it is quite necessary to explore the

mechanism of pork price fluctuations as well as to propose

an effective pork price forecasting model.

Although accurate prediction of pork price can be quite

helpful for agricultural participants and authorities, rel-

evant studies are somewhat limited compared with the

rapid development in other fields—for example, finan-

cial and energy prediction (Tang, Wu, & Yu, 2018a;

Zhao, Li, & Yu, 2017). We therefore focus on a broader

realm of agricultural price forecasting and hope to obtain

some workable experiences, so as to further improve

the forecast accuracy of pork price. The models widely

used for agricultural product price forecasting can be

divided into two categories: traditional statistical mod-

els and modern artificial intelligence (AI) models. Tra-

Journal of Forecasting. 2020;39:671–686. wileyonlinelibrary.com/journal/for © 2020 John Wiley & Sons, Ltd. 671

LING ET AL.

ditional models widely adopted for analyzing the agri-

cultural market include autoregressive integrated mov-

ing average model (ARIMA), generalized autoregressive

conditional heteroskedasticity model (GARCH), vector

autoregression (VAR), exponential smoothing, and struc-

tural models (Colino, Irwin, & Garcia, 2011; Zapata &

Garcia, 1990; Zhang, Hu, Revell, & Fu, 2005). In the past 10

years, owing to the increasing volatility and nonlinear pat-

terns emerging in price series, AI models have been widely

adopted and verified to be highly effective for such a com-

plex time series modeling task. For instance, Jha and Sinha

(2013) proposed a time-delay back-propagation neural net-

work to forecast monthly soybean and rapeseed prices.

Xiong, Li, and Bao (2018) combined the seasonal-trend

decomposition model and extreme learning machines to

predict the price of vegetables in China. Y. Liu, Duan,

Wang, Zhang, and Liu (2019) predicted hog prices based

on similar subseries search and support vector regression

(SVR). These studies consistently reported the superiority

of AI models over traditional statistical models. Among all

these models, the regression version of the support vec-

tor machine (SVM), which has the favorable property of

minimizing the structural risk and producing an excellent

generalization performance, has been widely used in aca-

demic and business communities (Yu, Zhang, & Wang,

2017; Zendehboudi, Baseer, & Saidur, 2018).

In general, pork prices are influenced by various factors.

Traditionally, scholars focused on the internal influential

factors that were associated with the hog breeding pro-

cess: for example, the price of piglet (Gjølberg, 1995; Zhou

& Koemle, 2015), the price of corn (Ahumada & Cornejo,

2016; Alexakis, Bagnarosa, & Dowling, 2017) and the num-

ber of sows (Mao, Zeng, & Maeda, 2016). The influences

caused by these factors transmitted through the whole

industry chain and exerted great impacts on the market's

equilibrium. Along with the highly interdependent world

economy,researchers highlighted the fact that agricultural

price movement was more affected by external factors,

including macroeconomic conditions (Y.Chen & Yu, 2018;

Yu, 2014), policy interventions (Arnade, Cooke, & Gale,

2017; Luska, Tonsor, Schroeder, & Hayes, 2018), and ani-

mal diseases (Boni, Galvani, Wickelgren, & Malani, 2013;

Zhou, Turvey, Hu, & Ying, 2016).

Among all the external factors, animal diseases are

regarded as the most challenging and unpredictable risks,

just like the financial crisis and terrorism (Suder &

Inthavong, 2008). Numerous studies have demonstrated

that animal diseases have a great impact on the agricul-

tural market, and pork price is sensitive to these external

shocks (Hassouneh, Radwan, Serra, & Gil, 2012; Shang

& Tonsor, 2017; Taylor, Klaiber, & Kuchler, 2016). For

instance, the porcine reproductive and respiratory syn-

drome virus (PRRSV), a major pathogen that has adversely

influenced the global swine industry for almost 30 years,

was reported to cause almost $600 million loss per year in

the USA (Du, Nan, Xiao, Zhao, & Zhou, 2017). Another

devastating disease of the swine industry globally, the

classical swine fever virus (CSF), is a notifiable disease

to the World Organization for Animal Health (OIE) and

can spread rapidly through all living pigs. With economic

and epidemiological models, the loss induced by CSF was

estimated to be 19.2 million EUR in the Finnish hog pro-

duction sector (Niemi, Lehtonen, Pietola, Lyytikainen, &

Raulo, 2008). With regard to the swine industry in main-

land China, external shocks contributed to roughly 20%

of the price fluctuations, and losses caused by animal dis-

eases were estimated to about 40 billion RMB (Zhang,

Zhang, & Bian, 2012).

Although animal diseases are crucial to price movement,

they are hard to quantify.In previous studies, dummy vari-

ables identifying whether diseases occurred or not and the

number of dead animals were often used as the proxy indi-

cators for measuring the diseases (X. Liu, Shen, & Wang,

2017; Peterson & Chen, 2005; Satoshi & Yuichiro, 2012;

Seo, Jang, Miao, Almanza, & Behnke, 2013). However,

these indicators cannot directly reflect the magnitude of

the diseases. Fortunately, online search data, a recently

emerging idea in the era of big data, provide us with a

new perspective of measuring the effect of disease out-

break on price movement. Based on the analysis of people's

online search behaviors, it is proven that the search vol-

umes of some keywords can finely reflect public concern

about a certain event (Li, Ma, Wang, & Zhang, 2015).

Considering that public concern is an important factor

that influences price movement, we believe that there is a

strong linkage between these two issues, and online search

data should be added to the analytical framework for pork

price prediction. That is to say, the search volumes of the

disease-related keywords can be regarded as an informa-

tive predictor of forthcoming price movement. Also, it is

supposed that the search volumes should have a positive

relationship with the impact of the diseases. A spike in

the time series of online search data is likely to indicate a

possible turning point of the price series in the future.

During the past 5 years, many studies have demon-

strated that the information content extracted from online

search data can effectively improve the forecast accuracy.

Fondeur and Karamé(2013) used Google queries to predict

French unemployment and verified that Google'sreal-time

information could enhance forecast accuracy.Bangwayos-

keete and Skeete (2015) constructed a new indicator

for tourism demand forecasting based on the composite

online search for “hotels and flights.” Fantazzini and Tok-

tamysova (2015) proposed a new multivariate model to

forecast monthly car sales, combining economic variables

and Google online search data. Ghodsi and Yarmoham-

672