Applied Energy 205 (2017) 336–344

Contents lists available at ScienceDirect

Applied Energy
journal homepage: www.elsevier.com/locate/apenergy

How does investor attention affect international crude oil prices? MARK
⁎
Ting Yao, Yue-Jun Zhang , Chao-Qun Ma
Business School, Hunan University, Changsha 410082, PR China
Center for Resource and Environmental Management, Hunan University, Changsha 410082, PR China

H I G H L I G H T S

• We detect the impacting mechanisms of investor attention on crude oil prices.

• We construct a proxy for investor attention in crude oil market based on the GSVI.
• Investor attention has significant negative impact on oil prices during 2004–2016.
• Investor attention contributes 15% to the long-run fluctuation of WTI oil prices.

A R T I C L E I N F O A B S T R A C T

Keywords: In order to investigate the impacting mechanism of investors’ attention and crude oil prices, we construct a
Investor attention direct, timely and unambiguous proxy for investor attention in crude oil market by aggregating the Google
Crude oil prices search volume index (GSVI). Based on the GSVI, we employ the Structural Vector Autoregression (SVAR) model
GSVI to empirically explore the impact of investor attention on WTI crude oil price from January 2004 to November
Principal component analysis
2016. The results indicate that: (1) investor attention does have significant negative impact on WTI crude oil
Structural VAR
price during the sample period; (2) investor attention shocks contributes 15% to the long-run fluctuation of WTI
crude oil price during the sample period, which is second only to that of supply shocks (69%) among the
contributors concerned; and (3) when the business cycle stays in expansion, it has positive influence on both
investor attention and WTI crude oil price. Meanwhile, our robustness check, using Brent crude oil price and a
different construction form of the GSVI, confirms that the central results are reliable.

1. Introduction make the buying or selling decisions based on their belief about the
fundamentals. However, the fundamentals of assets are hard to observe;
Crude oil is a crucial strategic resource, whose price fluctuation has in the end, investors tend to infer the fundamentals by processing re-
significant impact on economic growth [1], financial market [2,3] and lated information. When they process the information related to a
national security [4]. Consequently, as an important energy com- certain asset, they may inevitably pay attention to it. However, atten-
modity, crude oil has the ordinary commodity property, whose prices tion is a scarce resource [12], and the information process will be in-
are initially determined by the supply-demand fundamentals [5–8]; fluenced by attention constraint and attention allocation. Investors may
meanwhile, as a basic strategic resource and a financial product, crude allocate more attention to reduce the uncertainty to the fundamentals
oil also has evident political and financial properties, whose prices tend while the asset price fluctuates as the investors update the belief based
to be sharply influenced by the non-fundamentals (such as geopolitical on the information process. In this way, investors’ attention allocation
events, US dollar exchange rate and market speculation). In particular, has an effect on asset price dynamics [13].
non-fundamentals often lead to the psychological changes of investors In fact, several existing literature on investor attention has con-
in crude oil market, and further cause crude oil price deviating from firmed the significant effect of investor attention on asset prices. For
fundamentals [9–11], which makes crude oil market become a typical instance, Barber and Odean [14] argue that individual investors are net
complex system. attention-driven investors and their attention-grabbed buying behavior
In fact, investor attention proves an important index to reflect the results in a positive pressure to stock prices. Hou et al. [15] examine the
psychological changes of investors, and may significantly influence the effect of investor attention in stock prices, and they find that, on the one
price fluctuations in stock and commodity markets. Investors often hand, limited investor attention is likely to cause potential earnings

⁎
Corresponding author at: Business School, Hunan University, Changsha 410082, PR China.
E-mail address: zyjmis@126.com (Y.-J. Zhang).

http://dx.doi.org/10.1016/j.apenergy.2017.07.131
Received 28 April 2017; Received in revised form 28 July 2017; Accepted 31 July 2017
0306-2619/ © 2017 Elsevier Ltd. All rights reserved.
T. Yao et al. Applied Energy 205 (2017) 336–344

momentum for underreacting to earnings news; on the other hand, the Relevant literature mainly employs the traditional regression ap-
interaction of limited attention and learning biases results in price proaches to investigate the effect of investor attention on asset prices
momentum. Besides, Andrei and Hasler [16] find that both investor [22]. Unfortunately, the traditional regression approaches in current
attention and uncertainty significantly affect asset price changes. literature tend to hardly consider the dynamic relationship among
It should be noted that investor attention is hard to observe and various factors and it is hard for us to identify the endogenous and
measure; consequently, in prior relevant literature, investor attention is exogenous variables in the regression models [26]. In contrast, the
often measured by indirect proxies. The frequently used proxies include SVAR model can avoid these problems and has been widely applied in
extreme returns [14], trading volume [14,15], news and headlines energy economics and policy modeling literature; in particular, it can
[10,17,18], and price limits [19]. These proxies are based on a common analyze the dynamic relationship between crude oil price and its var-
hypothesis that if the asset or stock is of extreme return or trading ious influencing factors and can provide rich quantitative results upon
volume, or mentioned in news headline, investors have paid attention the impact of these factors on crude oil prices [27–30].
to it. However, the abnormal return or trading volume cannot guar- Therefore, in this paper, we first construct a direct, timely and un-
antee investors’ attention to it, and investors even never read the news ambiguous proxy for investor attention in crude oil market by ag-
in The Wall Street related to the asset or stock [20]. Therefore, the gregating the search volume reported by Google. Then, based on the
proxies are usually biased and lagged. Specifically, on the one hand, for GSVI, we employ the SVAR model to explore the dynamic quantitative
the selection of indicators used for investor attention, the proxies are of impact of investor attention on crude oil price. Specifically, we attempt
strong subjectivity. On the other hand, most indicators are calculated to answer the following questions: (1) How does the investor attention
through the statistical data, and they can hardly measure the investor relate to crude oil price fluctuations? (2) Compared with other influ-
attention in time. encing factors, how much does investor attention explain crude oil price
Under these circumstances, in this paper, we develop a direct and fluctuations? (3) Does the business cycle affect the fluctuation me-
unbiased proxy for investor attention in crude oil market by con- chanisms of investor attention and crude oil prices? Besides, we also
structing the Google search volume index (GSVI) and then explore the detect the robustness of results in this paper from two aspects. On the
relationship between investor attention and crude oil prices. Google one hand, we use different global benchmark crude oil prices for the
search volume data are available in Google Trends (https://www. research, including WTI crude oil and Brent crude oil prices. On the
google.com/trends), a service provided by Google. other hand, we consider the effect of different construction forms of
There are two reasons for us to choose the Google search volume as GSVI by using the ten “top searches” related to the primitive words
the measure of investor attention. First, Internet users are inclined to provided by Google Correlate. We find that the empirical results in this
use search engines to collect information, and Google continues to be paper are reliable on the basis of solid robustness tests.
the most popular, especially in developed countries. Thus, the Internet The contribution of this paper consists of two aspects. First, we
search behaviors of general population are most likely to be reflected by construct a new proxy for investor attention in crude oil market by
the search volume provided by Google. Second, the Google index aggregating the search volume reported by Google using the PCA ap-
proves a direct measure of the market attention. When we search oil in proach. Second, based on the proxy, we introduce the SVAR model to
Google, we are undoubtedly concerning about it. Consequently, the investigate how the investor attention may affect crude oil price, as well
aggregated Google search volume is a direct and unbiased measure of as how long the effect can last, so as to quantitatively confirm the
the market attention. In fact, some studies have confirmed this view- contribution of investor attention to crude oil price fluctuations.
point. For instance, Da et al. [20] argue that the search frequency in Overall, by constructing the investor attention index, we not only in-
Google is a timely proxy for measuring the investor attention. Drake vestigate the qualitative effect of investor attention on crude oil price,
et al. [21] claim that investors accept information through the Internet but also quantitatively study the contribution of the influencing factors
and the Google searches can reflect investors’ demand for public in- in crude oil price. This study will be beneficial for us to reveal the in-
formation. Da et al. [22] construct the Financial and Economic Atti- fluencing mechanisms of investor attention on crude oil price, so as to
tudes Revealed by Search (FEARS) index by counting the search volume better understand the complex pricing process of crude oil and forecast
related to household concerns. its changing trends, and provide important reference for policy makers
In order to measure the investor attention in crude oil market by to monitor and control extreme risks in crude oil market.
aggregating the search volume, we need to understand how crude oil The remainder of this paper is organized as follows. Section 2 pre-
price can be searched by investors. To do this, we aggregate the search sents the data definitions and methods used in this paper, Section 3
volumes of a series of key words related to “crude oil price”, which are describes the empirical results and discussions, and Section 4 concludes
likely to be used by users for different search habits, and we proceed as this paper.
follows. We input the primitive words “crude oil price” into Google
Correlate (https://www.google.com/trends/correlate), and retrieve the 2. Data and methods
top related words whose correlation coefficients with “crude oil price”
are larger than 0.9. After filtering out the related words, we can acquire 2.1. Data definitions
the search volumes of these words and construct the composite atten-
tion index by weighting the search volumes. There are many widely In this paper, we mainly consider four variables, including global
used weighting methods, such as the subjective weighting method, the crude oil production, global oil demand, investor attention in crude oil
information entropy theory and the linear combination weighting market and WTI crude oil prices. Specifically, global crude oil pro-
method [23]. However, among these search volumes, some should be duction and WTI crude oil prices are available from EIA (https://www.
correlated with other search volumes, and among the widely used eia.gov/totalenergy/data/monthly/#international); global oil demand
weighting methods, the Principle Component Analysis (PCA) is pow- is represented by the global real economic activity index developed by
erful in reducing the dependence of individual proxies. Meanwhile, the Kilian [31]; and investor attention is denoted by the Google search
index constructed by individual proxies would behave almost the same volume index (GSVI), which is constructed based on the search volume
as the index organized by PCA [24,25]. Therefore, we apply the PCA data provided by Google Trends. Google Trends provides both the
approach to build a linear combination of the selected proxies, so as to weekly and monthly search volume data starting from January 2004.
construct the GSVI. However, the global real economic activity index developed by Kilian
After obtaining the unbiased and direct proxy of the investor at- [31] is formed as monthly frequency data. As a result, for data avail-
tention, we need to further investigate the effect of investor attention ability, we select the monthly data ranging from January 2004 to No-
on international crude oil price applying some quantitative methods. vember 2016 as the sample in this paper. It should be noted that, except

337
T. Yao et al. Applied Energy 205 (2017) 336–344

p
the Kilian index and investor attention variables, all the other variables
X t = A−0 1α + ∑ A−0 1Ai X t − i + A−0 1 H dt + et
in this paper are transformed into logarithmic values. (3)
i=1

where et denotes the vector of the estimated residuals in the reduced

2.2. Methods form of SVAR model, and et = A−0 1εt (see Eq. (4)). The restrictions on
A−0 1 are determined according to Kilian and Lee [40].
(1) Principal component analysis approach
Supply SupplyShock
⎡ et ⎤ a11 0 0 0 ⎤ ⎡ εt ⎤
In order to construct the GSVI, we aggregate the Google search ⎢ e Demand ⎥ ⎡ ⎢ ε DemandShock ⎥
t ⎢ a a 0 0 ⎥ ⎢ t ⎥
et = ⎢ Attention ⎥ = ⎢ 21 22
volumes generated by the top related words applying the principal ⎢ et ⎥ a31 a32 a33 0 ⎥ ⎢ εtAttentionShock ⎥
component analysis (PCA). The PCA technique helps us transform the ⎢ Price ⎥ ⎢ a41 a42 a43 a44 ⎥ ⎢ OtherOilPriceShock ⎥
⎣ ⎦⎢ε
original search volume series into several new and uncorrelated vari- ⎣ et
⎢ ⎥
⎦ ⎣ t ⎥
⎦ (4)
ables, which are called the principal components [32,33]. where aij (i = 1,2,3,4; j = 1,2,3,4) denotes the coefficients of the i th re-
In fact, each principal component is a linear combination of the sponse to the j th shock, and the value 0 indicates that there are no
original search volume series, and the amount of the information con- contemporaneous responses from specific shock.
veyed by each principal component is measured by its variance. All the Then, following Chen et al. [37], we use the impulse response
principal components are arranged by the decreasing value of the function (IRF) of the SVAR model to calculate the response of one
variance, so the first principal component is the most informative and standard structural innovation from investor attention to crude oil price
the last principal component proves the least. changes, and the variance decomposition approach (VDA) of the SVAR
Given that we select the top n correlated words, then we have the model is employed to measure the contribution of investor attention to
vector a(n × 1) consisting in the search volumes generated by the top n crude oil fluctuations.
correlated words. Our aim is to simplify the vector a(n × 1) by reducing
the dimensionality n to k (k < n) , and the new vector can be re- 3. Empirical results and discussions
presented as b(k × 1) . The variable bj (j = 1,…,k ) in the new vector b is
the linear combination of the n variables, and is known as principal 3.1. The construction of the GSVI
components [32]. The jth principal components bj can be represented
by Eq. (1). An important choice concern for constructing the GSVI is the
bj = δj1 a1 + δj2 a2 + ⋯ + δjn an identification of the attention to crude oil price. For the sake of different
(1)
search habits and expressions, an Internet user is likely to search crude
where δjm (m = 1,…,n) is a constant denoting the eigenvalue of the mth oil price using several related key words (such as “oil price” and “crude
principal component. oil”). Hence, we select the primitive words as “crude oil price” (cop) and
In this paper, according to Gaitani et al. [33], we select the number we search for the related words in Google Correlate (https://www.
of principal components by examining the proportion of total variance google.com/trends/correlate). We single out 11 related key words of
explained by each component. which the correlation coefficient with “crude oil price” is larger than
0.9. Specifically, the selected key words related to crude oil prices in-
(2) The SVAR model clude “oil price” (op), “current oil prices” (cops), “price per barrel”
(ppb), “bloomberg energy” (be), “oil price per barrel” (oppb), “current
In order to investigate the effect of investor attention on crude oil oil” (co), “crude oil chart” (coc), “crude oil” (col), “current crude oil”
price fluctuations, we define a vector including four variables as (cco), “current crude oil price” (ccop), and “current crude” (cc).
x t = (Supplyt ,Demandt ,Attentiont ,Pricet ) . Moreover, previous studies After downloading the Google search volume of these proxies, we
have proved that the influence of the phase of business cycle on ag- use the Principle Component Analysis (PCA) approach to exclude the
gregate investment on assets [34,35]. Different from these studies, we idiosyncratic elements in the proxies. Specifically, we build a linear
consider the influence of business cycle1 by introducing a dummy combination of the selected proxies using the PCA approach, so as to
variable dt into the SVAR model. On the one hand, we estimate the construct the GSVI. The eigenvalues and explained variances of each
overall effect of investor attention on crude oil price. On the other hand, principal component are shown in Table 1 and the weights of the raw
it will improve the estimation accuracy since more samples can be variables on each principal component are shown in Table 2.
covered by the SVAR models [36,37]. Consequently, the four-variable In the PCA framework, each principal component is a linear com-
SVAR (Supply, Demand, Attention, Price) model can be defined as Eq. (2). bination, and the explained information of each principal component is
reflected by the value of the variance. According to Wang [41], we can
p
select the principal components with their cumulatively explained
A 0Xt = α + ∑ Ai X t − i + H dt + εt
proportion larger than 80% or with their eigenvalues above 1.00. In this
i=1 (2)
paper, according to the results shown in Table 1, there are 12 principal
where α,H,A 0 and Ai are the unknown vectors and matrices to be es- components in total, and the first principal component explains 87.41%
timated; dt takes the value 0 or 1, and if the business cycle stays in of the sample variance of the variables, which is larger than 80%.
expansion, the value is 1; otherwise, the value is 02; εt represents the Meanwhile, as shown in Table 1, only the eigenvalue of the first prin-
vector of continuous and uncorrelated structural innovations. cipal component is above 1.00. As a result, the first principle compo-
The SVAR model above can be reduced as Eq. (3). nent is selected to be the Google Search Volume Index in crude oil
market and the coefficients of the raw variables are determined ac-
1 cording to the weights shown in Table 2.
The expansions and recessions are announced by the National Bureau of Economic
Research (http://www.nber.org/cycles/cyclesmain.html).
2
In fact, there is not a consensus in literature about how to code the dummy to ex- 3.2. The response of crude oil price to investor attention shocks
pansions or recessions. For example, McLean and Zhao [38] define business expansions as
one, while Halling et al. [39] set the dummy as one in NBER recessions. In this paper, as
In order to estimate the response of WTI crude oil price to investor
previous studies have indicated the influence of economic expansions on aggregate in-
vestment in assets, so we are motivated to clarify the relationship of economic expansions
attention, we estimate a four-variable SVAR model according to Eq. (2).
and crude oil prices. Consequently, we define a dummy value that equals one in NBER Before we develop the SVAR model, we test the stationarity of the
expansions. variables using the Augmented Dickey–Fuller (ADF) [42] and

338
T. Yao et al. Applied Energy 205 (2017) 336–344

Table 1 Meanwhile, the business cycle is selected as a dummy variable in

Eigenvalues and explained variances of the principal components. the SVAR model and the least square method is adopted to estimate the
SVAR model. The lag lengths of the VAR model are determined by
Component Eigenvalue Proportion Cumulative proportion
considering the popular criteria, including Likelihood Ratio (LR) [49],
1 10.4897 0.8741 0.8741 Final Prediction Error (FPE) [50], Akaike Information Criterion (AIC)
2 0.6034 0.0503 0.9244 [51], Schwarz Criterion (SC) [52] and Hannan-Quinn (HQ) [53] cri-
3 0.4131 0.0344 0.9589
teria. Most criteria (i.e., LR, FPE and AIC) indicate that the optimal lag
4 0.1566 0.0130 0.9719
5 0.1347 0.0112 0.9831 length can be 3.
6 0.0603 0.0050 0.9881 Then, based on Eqs. (2)–(4), we estimate the SVAR model, and the
7 0.0443 0.0037 0.9918 impact matrix is shown in Eq. (5).
8 0.0384 0.0032 0.9950

⎡ a11 0 0 0 ⎤ ⎡ 0.006
9 0.0318 0.0026 0.9977 0 0 0 ⎤
10 0.0183 0.0015 0.9992
⎢ a21 a22 0 0 ⎥ = ⎢ − 53.916 11.148 0 0 ⎥
11 0.0081 0.0007 0.9999 ⎢ a31 a32 a33 0 ⎥ ⎢− 111.205 − 0.383 19.823 0 ⎥
12 0.0014 0.0001 1 ⎢ a a a a ⎥ ⎢ − 2.391
⎣ 41 42 43 44 ⎦ ⎣ 0.001 − 0.001 0.075⎥
⎦ (5)

Table 2
The weights of the raw variables on principal components.

Variable Com1 Com2 Com3 Com4 Com5 Com6 Com7 Com8 Com9 Com10 Com11 Com12

cop 0.298 −0.164 0.298 −0.122 −0.093 −0.168 −0.295 0.046 −0.202 −0.076 −0.779 −0.051
op 0.296 −0.200 0.268 −0.247 −0.137 −0.040 −0.388 −0.096 −0.428 −0.172 0.588 0.042
cops 0.300 0.111 −0.239 0.120 −0.015 −0.154 0.436 −0.453 −0.187 −0.606 −0.069 0.016
ppb 0.284 0.453 0.111 0.380 −0.074 0.031 −0.129 0.028 −0.024 0.113 0.091 −0.72
be 0.274 0.403 0.315 −0.647 −0.021 −0.075 0.349 0.142 0.305 0.020 0.055 0.011
oppb 0.283 0.463 0.116 0.405 −0.068 0.040 −0.130 0.089 −0.047 0.123 0.001 0.694
co 0.286 0.072 −0.28 −0.151 0.863 0.124 −0.217 0.031 −0.058 0.023 −0.021 −0.005
coc 0.260 −0.471 0.534 0.358 0.272 0.174 0.393 0.066 0.163 0.025 0.087 −0.002
col 0.291 −0.223 −0.305 0.113 −0.039 −0.752 0.053 0.359 0.1375 0.149 0.141 0.000
cco 0.299 −0.129 −0.244 −0.133 −0.175 0.190 0.294 −0.286 −0.332 0.685 −0.060 0.002
ccop 0.300 −0.169 −0.137 −0.013 −0.174 0.121 −0.353 −0.452 0.698 0.010 0.017 0.034
cc 0.292 −0.133 −0.352 −0.038 −0.284 0.526 0.044 0.5777 0.012 −0.278 −0.025 −0.023

Table 3 According to the values of a41,a42,a43 and a44 , we can find that the
Results of unit root test.a values of a41 and a43 are smaller than zero, which indicates that the
supply and investor attention shocks have negative impact on crude oil
Unit root test Supply Demand Attention Price
price. Meanwhile, the values of a42 and a44 are larger than zero, in-
ADF test Level −0.371 −2.376 −3.027 −2.409 dicating the positive impact of the demand and other shocks on crude
(0.910) (0.150) (0.035) (0.141) oil price.
First −12.918 −9.301 −11.449 −8.754 Based on the estimated results, the impulse response results of WTI
differenced (0.000) (0.000) (0.000) (0.000)
crude oil price to one standard deviation of supply shock, demand
PP test Level −0.089 −2.171 −3.179 −2.243 shock, attention shock are displayed in Fig. 1, and we have several
(0.948) (0.218) (0.023) (0.192)
important findings.
First −13.067 −9.301 −11.417 −8.756
differenced (0.000) (0.000) (0.000) (0.000) For one thing, the investor attention has negative impact on WTI
crude oil price during the sample period beginning from the 1st month
a
The numbers in the table are the t statistic values and the p-values are reported in the and lasting for about 16 months, and after the 17th month, the impact
parentheses. becomes positive, indicating the delayed effect of investor attention on
WTI crude oil price. However, the impact of investor attention is not
Phillips–Perron (PP) [43] tests, and the results are shown in Table 3. We significant from the 10th month. The significant negative impact of
can find that, expect for investor attention, all the level variables are investor attention begins in the 2nd month and lasts for 9 months. The
not stationary because they fail to reject the null hypothesis of a unit finding is in line with the results obtained by the impact matrix in Eq.
root at the significance level of 10%. However, all the first differenced (5), and is basically consistent with Da et al. [20], who find that an
variables are stationary. increased investor attention predicts higher stock prices in the next two
In general, the SVAR models are constructed based on the premise weeks, but the effect turns reversed within the year. In this paper, due
that all the variables are stationary [44]. However, as shown in Table 1, to data availability, the search objective is organized as monthly data,
the variables in levels cannot meet the constraint that all the variables so we identify the long-run negative effect of investor attention on
are stationary. Consequently, we estimate the SVAR model with vari- crude oil prices.
ables in levels using the method of Toda and Yamamoto [45], whose The results above can be interpreted by the candidate explanations
advantage is that the existence of unit roots and cointegration does not as follows. First, the price pressure theory as described in Barber and
influence the final results [46]. According to Clarke and Mirza [47], Odean [14] can explain the negative price pressure in the long run.
pre-tests for unit roots and cointegration may suffer from size distor- They find that the increased investor attention predicts temporary
tions. To avoid these problems, the Toda and Yamamoto approach in- higher stock prices, but the prices get reversed in the long run [14].
troduces a Wald test statistic that asymptotically follows a Chi-square According to their research, individual investors are attention-grabbed
distribution irrespective of the orders of integration or cointegration net buyers and thus stock prices experience a temporary positive price
properties of the variables. Moreover, the long-run information can be pressure as the individual investor attention increases. However, when
captured using the variables in levels, which is often ignored in the the excess returns cause price pressures, the demand of individual in-
model using the first-differenced variables [48]. vestors dissipates; resulting in the reverse in crude oil price.

339
T. Yao et al. Applied Energy 205 (2017) 336–344

Supply Demand
.02
.04
.00
.02
-.02
.00
-.04
-.02

-.06 -.04

-.08 -.06

-.08
-.10
5 10 15 20 25 30 35 5 10 15 20 25 30 35
Month Month
(a) Response of WTI crude oil prices to a supply shock (b) Response of WTI crude oil prices to a demand shock

Attention
.04

.02

.00

-.02

-.04

-.06

-.08

-.10
5 10 15 20 25 30 35
Month
(c) Response of WTI crude oil prices to an investor aƩenƟon shock
Fig. 1. Responses of WTI crude oil prices to different shocks. The blue solid lines and the red dotted lines represent the mean impulse response of WTI crude oil prices and two standard
deviations from the mean, respectively. The response period is 36 months. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of
this article.)

Second, according to Da et al. [20], the investor attention generated impact of oil demand shocks is significant and lasts for about 3 months.
by the search volume reflects the attention of individual investors. This finding is different from Chen et al. [37], who develop a five-
However, the institutional investors are rational and not easily to be variable SVAR model for Brent crude oil price and find that oil demand
driven by psychological fluctuations, and the high market attention shocks have significant positive impact on Brent crude oil price while
often leads rational institutional investors to be more conservative, so the small negative influence of supply shocks on crude oil price is not
they are usually to realize that buying assets with high market attention significant. The difference is mainly sourced from different variables
is not the best choice [54]. They are conscious that the highly con- selected in the SVAR models and different sample periods used. In the
cerned asset is likely to experience abnormal increases in prices de- study of Chen et al. [37], there are five variables in the SVAR model,
viated from fundamentals, and there would be bubble compositions in i.e., the political risk of OPEC countries, oil supply, oil demand, spec-
the asset prices. As a result, they may make decisions contrary to the ulation and international crude oil prices, and the sample period ranges
individual investors and impose a reverse pressure to crude oil price from January 1998 to September 2014. In this paper, we develop the
changes. The negative effect of investor attention on crude oil price is SVAR model based on four variables including oil supply, oil demand,
confirmed by the facts. For example, in December 2008, the investor investor attention and international crude oil prices over the period
attention began to drop, and in February 2009, WTI crude oil price from January 2004 to November 2016.
began to increase. At that time, a large portion of investors were in a Besides, from the estimated coefficients of the dummy variables
panic caused by the financial crisis in 2008, they were cautious and shown in Table 4,3 we can find that the business cycle significantly
unwilling to concern crude oil price, and the attention-driven buying changes the fluctuation mechanism of investor attention and WTI crude
was reduced. Subsequently, crude oil price gradually fluctuated around oil prices during the sample period. Specifically, the estimated coeffi-
the fundamentals. Meanwhile, the rational institutional investors rea- cients of investor attention and oil price are 16.7334 and 0.0751, re-
lized that the bubble compositions were reduced and there was an spectively, which are significantly positive and indicate that when the
opportunity to make profits.
For another, WTI crude oil prices respond to oil supply shocks ne-
gatively, while the impact of demand shocks on crude oil price is po- 3
To save space, in Table 4, we only show the estimated results of the dummy variable.
sitive. Specifically, as shown in Fig. 1, the negative impact of oil supply In fact, there are many other results concerning other variables in the vector auto-
shocks is significant and lasts for about 24 months while the positive regression models and the results can be obtained upon request from authors.

340
T. Yao et al. Applied Energy 205 (2017) 336–344

Table 4 3.4.1. Effect of using different benchmark crude oil prices

Estimated results of the dummy variables in the vector autoregression model.a We re-estimate the SVAR model with Brent crude oil prices, since
Brent crude oil price is also a major benchmark of international crude
Dependent Supply Demand Attention Price
variable oil pricing across the world. The impulse response results of Brent crude
oil prices to supply shock, demand shock and investor attention shock
d 0.0004 4.8354 16.7334** 0.0751*** are shown in Fig. 2, and the variance decomposition results are shown
(0.0023) (4.0849) (8.4753) (0.0289)
in Table 6.
a
** and *** denote the significance at 5% and 1% levels, respectively. The values in
As can be seen from Fig. 2, the impulse response results are in line
parentheses are standard errors of the coefficients to be estimated. with those obtained based on WTI crude oil prices. Specifically, investor
attention shocks do have negative effect on Brent crude oil prices, and
business cycle stays in expansion, it has positive influence on both in- the effect becomes significant in the 2nd month and lasts for 9 months.
vestor attention and WTI crude oil price. In fact, many economists claim That is to say, investor attention shocks also have delayed impact on
that the level of aggregate investment could be a function of the phase Brent crude oil prices.
of business cycle [34,35]. Specifically, when the economy stays in the Meanwhile, from the variance decomposition results shown in
process of expansion, market investors are more likely to have higher Table 6, we can find that, compared to the results in Section 3.3, there is
growth expectations and easier to have capital access [34]. Hence, they a little difference in the values of the four shocks’ contribution, but the
tend to pay more attention to asset prices and invest in assets which main conclusion still holds. For example, the contribution of investor
they are interested in. On the contrary, when the economy stays in a attention shocks to WTI crude oil prices (15.18%) is more than that to
state of recession, investors are more likely to face financing constraints Brent crude oil prices (14.78%). However, similar to the case of WTI
and the access to capital becomes harder [35]. Consequently, they tend crude oil prices, among the four shocks in SVAR model, the contribu-
to pay less attention to asset price fluctuations and underinvest po- tion of investor attention shocks to Brent crude oil price changes is only
tential assets [35]. less than that of supply shocks (71.58%).

3.3. The proportion of WTI crude oil price fluctuations explained by investor 3.4.2. Effect of using different construction forms of the GSVI
attention shocks For the sake of different search habits and expressions, an Internet
user is likely to search crude oil price using several related key words
In order to investigate the contribution of investor attention shocks (such as “oil price” and “crude oil”). However, as for how many related
to WTI crude oil price fluctuations, we employ the variance decom- key words should be considered, there is not a consensus. As a result,
position approach in SVAR model to decompose crude oil price changes we construct two GSVIs using different number of related words to test
into four components and the empirical results are shown in Table 5. the robustness. Specifically, we reconstruct the GSVI by weighting the
We can see that, in the short run, only 1.75% of the variation in crude search volume of the ten “top searches” related to the primitive words
oil prices is attributed to investor attention shocks. As time passes, in- provided by Google Correlate, which is motivated by Da et al. [22].
vestor attention shocks has a good explanatory power of WTI crude oil The eigenvalue, explained proportion and cumulatively explained
price fluctuations in the 6th month, which contributes 23.25% to WTI proportion of the principal components are shown in Table 7, from
crude oil price changes. Then the contribution of investor attention which we find that there are 10 principal components in total, and the
shocks reaches the peak (30.15%) in the 12th month. After that, the first principal component explains 86.78% of the sample variance of the
explanatory power of investor attention gradually decreases, as the variables, which is larger than 80%. Meanwhile, only the eigenvalue of
horizon is lengthened. In the long run, investor attention shocks ac- the first principal component is above 1.00. As a result, the first prin-
count for 15.18% of WTI crude oil price changes, suggesting that shocks ciple component is selected to be the Google Search Volume Index in
in investor attention are an important factor for crude oil market. crude oil market, and the coefficients of the raw variables are de-
As for the shocks of fundamental factors in crude oil markets, we termined according to the weights shown in Table 8.
can see that in the short run, the effect of both supply and demand The impulse response results of WTI crude oil prices to supply
shocks is relatively smaller. Specifically, the supply and demand shocks shock, demand shock and investor attention shock are shown in Fig. 3,
contribute 3.74% and 3.48% to WTI crude oil price fluctuations, re- and the variance decomposition results are shown in Table 9. We do not
spectively. In the long run, the contribution of supply shocks finally have significantly different findings compared to the results from the
increases to 69.49%, while that of demand shocks only reaches 5.88%in former construction of the GSVI. In other words, the central results in
the end. this paper are robust.
Specifically, the delayed negative influence of investor attention on
3.4. Robustness analysis WTI crude oil prices is identified under the new construction of the
GSVI. The investor attention has significant negative impact on WTI
In order to ensure the robustness of empirical results in this paper, crude oil prices during the sample period beginning in the 2nd month
here we also investigate the effect of investor attention on crude oil and lasting for about 9 months. Meanwhile, the investor attention
price changes using different benchmark crude oil prices and different shocks contribute 14.78% to WTI crude oil price changes, which is only
construction forms of the GSVI. less than that of the supply shocks (71.58%) among the contributors
concerned. Therefore, as an important financial commodity, the influ-
Table 5 ence of investor attention on crude oil prices cannot be ignored and it is
Contribution of different shocks to WTI crude oil price fluctuations (%).
of great significance to investigate and quantify the effect of investor
Month Supply shocks Demand shocks Attention shocks Other shocks attention on crude oil prices in risk aversion.

1 3.74 3.48 1.75 91.03

6 12.39 4.51 23.25 59.86 4. Conclusions and future work
12 21.99 9.73 30.15 38.13
18 28.24 14.97 23.90 32.89 In this paper, we develop several SVAR models to investigate the
24 32.87 15.13 20.78 31.20
influence of investor attention on crude oil prices from January 2004 to
30 36.93 14.91 18.76 29.40
∞ 69.49 5.88 15.18 9.44 November 2016 using the GSVI, and several main conclusions can be
safely drawn as follows.

341
T. Yao et al. Applied Energy 205 (2017) 336–344

Supply Demand
.02 .06

.00 .04

-.02 .02

-.04 .00

-.06 -.02

-.08 -.04

-.10 -.06

-.12 -.08

-.14 -.10

5 10 15 20 25 30 35 5 10 15 20 25 30 35
Month Month
(a) Response of Brent crude oil prices to a supply shock (b) Response of Brent crude oil prices to a demand shock

Attention

.04

.02

.00

-.02

-.04

-.06

-.08

-.10
5 10 15 20 25 30 35
Month
(c) Response of Brent crude oil prices to an investor aƩenƟon shock
Fig. 2. Responses of Brent crude oil prices to different shocks. The blue solid lines and the red dotted lines represent the mean impulse response of Brent crude oil prices and two standard
deviations from the mean, respectively. The response period is 36 months. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of
this article.)

Table 6 Table 7
Contribution of different shocks to Brent crude oil price fluctuations (%). Eigenvalues and explained variances of the principal components of different GSVIs.

Month Supply shocks Demand shocks Attention shocks Other shocks Component Eigenvalue Proportion Cumulative

1 2.62 6.94 2.15 88.30 1 8.6776 0.8678 0.8678

6 10.83 5.08 24.19 59.90 2 0.5788 0.0579 0.9256
12 21.04 9.64 29.26 40.06 3 0.3484 0.0348 0.9605
18 28.05 13.84 22.31 35.79 4 0.1564 0.0156 0.9761
24 33.35 14.10 18.40 34.14 5 0.1190 0.0119 0.9880
30 38.04 13.94 15.80 32.22 6 0.0476 0.0048 0.9928
∞ 71.58 5.07 14.78 8.57 7 0.0415 0.0041 0.9969
8 0.0213 0.0021 0.9990
9 0.0081 0.0008 0.9999
(1) During the sample period, the investor attention does have sig- 10 0.0014 0.0001 1

nificant negative impact on crude oil prices, and the significant

negative impact appears in the 2nd month and lasting for about
mechanisms of investor attention and WTI crude oil prices during
9 months. In addition, the oil supply shocks have significant posi-
the sample period. When the business cycle stays in expansion, it
tive impact on crude oil price while the oil demand shocks have
has positive influence on both investor attention and WTI crude oil
significant negative impact.
prices.
(2) During the sample period, the investor attention contributes
(4) The robustness check by using Brent crude oil prices and re-con-
15.18% to WTI crude oil price fluctuations in the long run, which
structing the GSVI from the search volume of the ten “top searches”
ranks the second among the influencing factors considered in this
related to the primitive words provided by Google Correlate con-
paper. Specifically, the oil supply shocks contribute 69.49% to WTI
firms that our empirical results are robust.
crude oil price fluctuations, while the contribution of oil demand
shocks arrives at 5.88%.
Consequently, analyzing and quantifying the response of crude oil
(3) The business cycle significantly changes the fluctuation

342
T. Yao et al. Applied Energy 205 (2017) 336–344

Table 8
The weights of the raw variables on principal components of different GSVIs.

Variable Com1 Com2 Com3 Com4 Com5 Com6 Com7 Com8 Com9 Com10

cop 0.328 0.231 0.229 −0.117 −0.133 0.221 −0.296 −0.120 −0.776 −0.042
op 0.325 0.262 0.197 −0.240 −0.182 0.043 −0.534 −0.237 0.5917 0.045
cops 0.328 −0.117 −0.301 0.111 −0.147 −0.336 0.272 −0.751 −0.0612 0.022
ppb 0.316 −0.416 0.161 0.385 −0.026 0.037 −0.129 0.110 0.0851 −0.7173
be 0.306 −0.339 0.376 −0.640 −0.009 0.051 0.476 0.088 0.0556 0.009
oppb 0.314 −0.426 0.169 0.410 −0.013 0.057 −0.133 0.146 0.0031 0.694
co 0.314 −0.075 −0.400 −0.203 0.805 0.064 −0.205 0.021 −0.0211 −0.004
coc 0.287 0.576 0.397 0.348 0.339 −0.184 0.382 0.075 0.089 0.000
col 0.317 0.210 −0.447 0.101 −0.273 0.654 0.313 0.163 0.139 0.000
cco 0.325 0.118 −0.325 −0.132 −0.299 −0.603 −0.078 0.541 −0.077 −0.002

prices to investor attention is of great significance for us to understand fluctuation mechanism and forecast based on more microscopic
and forecast the fluctuations in crude oil prices, and for regulators and Internet search data by mining data from other search engines or social
policy makers to monitor and avoid the extreme risk in crude oil media. Meanwhile, whether the effect of investor attention on crude oil
market. Specifically, the investors in crude oil market can make profits prices relies on the volatility of oil market dynamics is also of interest.
by taking sufficient consideration to the impact of investor attention,
and regulators and policy makers can evade the extreme risk and fur-
Acknowledgments
ther stabilize the operation of crude oil market by tracing the dynamic
impact of investor attention on crude oil prices.
We gratefully acknowledge the financial support from the National
In the future, we still have some relevant work to do. For instance,
Natural Science Foundation of China (nos. 71273028, 71322103 and
we can further analyze the effect of investor attention on crude oil price
71431008), National Special Support Program for High-Level Personnel

Supply Demand
.02 .06

.04
.00
.02
-.02
.00

-.04 -.02

-.04
-.06
-.06
-.08
-.08

-.10 -.10
5 10 15 20 25 30 35 5 10 15 20 25 30 35
Month Month
(a) Response of WTI crude oil prices to a supply shock (b) Response of WTI crude oil prices to a demand shock

Attention
.04

.02

.00

-.02

-.04

-.06

-.08

-.10
5 10 15 20 25 30 35
Month
(c) Response of WTI crude oil prices to an investor aƩenƟon shock
Fig. 3. Responses of WTI crude oil prices to different shocks under different GSVIs. The blue solid lines and the red dotted lines represent the mean impulse response of WTI crude oil
prices and two standard deviations from the mean, respectively. The response period is 36 months. (For interpretation of the references to colour in this figure legend, the reader is
referred to the web version of this article.)

343
T. Yao et al. Applied Energy 205 (2017) 336–344

Table 9 [22] Da Z, Engelberg J, Gao P. The sum of all fears investor sentiment and asset prices.
Contribution of different shocks to WTI crude oil price fluctuations under different GSVIs Rev Financ Stud 2015;28(1):1–32.
[23] Zhang YJ, Hao JF. Carbon emission quota allocation among China’s industrial
(%).
sectors based on the equity and efficiency principles. Ann Oper Res
2017;255:117–40.
Month Supply shocks Demand shocks Attention shocks Other shocks
[24] Baker M, Wurgler J. Investor sentiment in the stock market. J Econ Perspect
2007;21(2):129–51.
1 3.67 3.45 2.37 90.51 [25] DeFelice M, Alessandri A, Catalano F. Seasonal climate forecasts for medium-term
6 11.80 4.53 24.76 58.92 electricity demand forecasting. Appl Energy 2015;137:435–44.
12 21.82 9.45 31.57 37.76 [26] Kilian L, Murphy DP. The role of inventories and speculative trading in the global
18 27.66 14.32 25.22 32.80 market for crude oil. J Appl Econ 2014;29(3):454–78.
24 32.44 14.43 21.94 31.19 [27] Narayan PK, Narayan S, Prasad A. A structural VAR analysis of electricity con-
30 36.60 14.18 19.78 29.44 sumption and real GDP: evidence from the G7 countries. Energy Policy
∞ 69.49 5.88 15.18 9.44 2008;36(7):2765–9.
[28] Wang Y, Wu C, Yang L. Oil price shocks and agricultural commodity prices. Energy
Econ 2014;44:22–35.
[29] Li Q, Cheng K, Yang X. Response pattern of stock returns to international oil price
from the Central Government of China, Changjiang Scholars Program of shocks: from the perspective of China’s oil industrial chain. Appl Energy
the Ministry of Education of China and Hunan Youth Talent Program. 2017;185:1821–31.
We would also like to thank the editors, two reviewers, and the seminar [30] Ding Z, Liu Z, Zhang Y, Long R. The contagion effect of international crude oil price
fluctuations on Chinese stock market investor sentiment. Appl Energy
participants at the Center for Resource and Environmental Management 2017;187:27–36.
of Hunan University for their insightful comments. [31] Kilian L. Not all oil price shocks are alike: disentangling demand and supply shocks
in the crude oil market. Am Econ Rev 2009;99:1053–69.
[32] Wold S, Esbensen K, Geladi P. Principal component analysis. Chemometr Intell Lab
References 1987;2(1–3):37–52.
[33] Gaitani N, Lehmann C, Santamouris M, Mihalakakou G, Patargias P. Using principal
[1] Yu L, Wang Z, Tang L. A decomposition–ensemble model with data-characteristic- component and cluster analysis in the heating evaluation of the school building
driven reconstruction for crude oil price forecasting. Appl Energy 2015;156:251–67. sector. Appl Energy 2010;87(6):2079–86.
[2] Zhang YJ. Speculative trading and WTI crude oil futures price movement: an em- [34] Schumpeter JA. The theory of economic development: An inquiry into profits, ca-
pirical analysis. Appl Energy 2013;107:394–402. pital, credit, interest, and the business cycle. Transaction Publishers; 1934.
[3] Zhang YJ, Wang ZY. Investigating the price discovery and risk transfer functions in [35] Arif S, Lee CM. Aggregate investment and investor sentiment. Rev Financ Stud
the crude oil and gasoline futures markets: some empirical evidence. Appl Energy 2014;27(11):3241–79.
2013;104:220–8. [36] Kim S. Do monetary policy shocks matter in the G-7 countries? Using common
[4] An H, Gao X, Fang W, Ding Y, Zhong W. Research on patterns in the fluctuation of identifying assumptions about monetary policy across countries. J Int Econ
the co-movement between crude oil futures and spot prices: a complex network 1999;48(2):387–412.
approach. Appl Energy 2014;136:1067–75. [37] Chen H, Liao H, Tang BJ, Wei YM. Impacts of OPEC's political risk on the inter-
[5] Bunn DW, Chevallier J, Le Pen Y, Sevi B. Fundamental and financial influences on national crude oil prices: an empirical analysis based on the SVAR models. Energy
the co-movement of oil and gas prices. Energy J 2017;38:201–28. Econ 2016;57:42–9.
[6] Mi ZF, Wei YM, Tang BJ, Cong RG, Yu H, Cao H, Guan D. Risk assessment of oil [38] McLean RD, Zhao M. The business cycle, investor sentiment, and costly external
price from static and dynamic modelling approaches. Appl Econ finance. J Financ 2014;69(3):1377–409.
2017;49(9):929–39. [39] Halling M, Yu J, Zechner J. Leverage dynamics over the business cycle. J Financ
[7] Zhang XB, Qin P, Chen X. Strategic oil stockpiling for energy security: the case of Econ 2016;122(1):21–41.
China and India. Energy Econ 2017;61:253–60. [40] Kilian L, Lee TK. Quantifying the speculative component in the real price of oil: the
[8] Zhang JL, Zhang YJ, Zhang L. A novel hybrid method for crude oil price forecasting. role of global oil inventories. J Int Money Financ 2014;42:71–87.
Energ Econ 2015;49:649–59. [41] Wang E. Benchmarking whole-building energy performance with multi-criteria
[9] Zhang YJ, Zhang L. Interpreting the crude oil price movements: Evidence from the technique for order preference by similarity to ideal solution using a selective ob-
Markov regime switching model. Appl Energy 2015;143:96–109. jective-weighting approach. Appl Energy 2015;146:92–103.
[10] Narayan PK, Ranjeeni K, Bannigidadmath, D. New evidence of psychological barrier [42] Dickey DA, Fuller WA. Distribution of the estimators for autoregressive time series
from the oil market. J Behav Finance 2016. with a unit root. J Am Stat Assoc 1979;74(366a):427–31.
[11] Zhang YJ, Yao T. Interpreting the movement of oil prices: driven by fundamentals [43] Phillips PC, Perron P. Testing for a unit root in time series regression. Biometrika
or bubbles? Econ Model 2016;55:226–40. 1988;75(2):335–46.
[12] Kahneman D. Attention and effort. Englewood Cliffs, NJ: Prentice-Hall; 1973. p. [44] Sims CA. Macroeconomics and reality. Econometrica 1980;48(1):1–48.
1063. [45] Toda HY, Yamamoto T. Statistical inference in vector autoregressions with possibly
[13] Peng L, Xiong W. Investor attention, overconfidence and category learning. J Financ integrated processes. J Econ 1995;66(1):225–50.
Econ 2006;80(3):563–602. [46] Basher SA, Haug AA, Sadorsky P. Oil price, exchange rates and emerging stock
[14] Barber BM, Odean T. All that glitters: the effect of attention and news on the buying markets. Energy Econ 2012;34(1):227–40.
behavior of individual and institutional investors. Rev Financ Stud [47] Clarke JA, Mirza S. A comparison of some common methods for detecting Granger
2008;21(2):785–818. noncausality. J Stat Comput Sim 2016;76(3):207–31.
[15] Hou K, Xiong W, Peng L. A tale of two anomalies: the implications of investor [48] Wu G, Zhang YJ. Does China factor matter? An econometric analysis of interna-
attention for price and earnings momentum. Working Paper, Ohio State University tional crude oil prices. Energy Policy 2014;72:78–86.
and Princeton University; 2009. [49] Dickey DA, Fuller WA. Likelihood ratio statistics for autoregressive time series with
[16] Andrei D, Hasler M. Investor attention and stock market volatility. Rev Financ Stud a unit root. Econometrica 1981:1057–72.
2015;28(1):33–72. [50] Akaike H. Fitting autoregressive models for prediction. Ann Inst Stat Math
[17] Yuan Y. Market-wide attention, trading, and stock returns. J Financ Econ 1969;21(1):243–7.
2015;116(3):548–64. [51] Parzen E, Tanabe K, Kitagawa G. Information theory and an extension of the
[18] Drake MS, Jennings J, Roulstone DT, Thornock JR. The comovement of investor maximum likelihood principle. In: Selected Papers of Hirotugu Akaike. Springer
attention. Manage Sci 2016. Series in Statistics. New York: Springer; 1998.
[19] Seasholes MS, Wu G. Predictable behavior, profits, and attention. J Empirical [52] Schwarz G. Estimating the dimension of a model. Ann Stat 1978;6(2):461–4.
Financ 2007;14(5):590–610. [53] Hannan EJ, Quinn BG. The determination of the order of an autoregression. J Roy
[20] Da Z, Engelberg J, Gao P. In search of attention. J Financ 2011;66(5):1461–99. Stat Soc B Met 1979:190–5.
[21] Drake MS, Roulstone DT, Thornock JR. Investor information demand: evidence [54] Edelen RM, Ince OS, Kadlec GB. Institutional investors and stock return anomalies.
from Google searches around earnings announcements. J Account Res J Financ Econ 2016;119(3):472–88.
2012;50(4):1001–40.

344