Upload
others
View
7
Download
0
Embed Size (px)
Citation preview
Googling Investor Sentiment around the World
Zhenyu Gao, Haohan Ren, and Bohui Zhang*
Preliminary Version: Oct. 29, 2015
*Zhenyu Gao and Haohan Ren are from the Department of Finance, CUHK Business School, The Chinese University of Hong Kong, Shatin, Hong Kong, and Bohui Zhang is from the School of Banking and Finance, UNSW Business School, UNSW Australia, Sydney, NSW, Australia, 2052. Authors’ contact information: Gao: [email protected], (852) 39431824; Ren: [email protected], (852) 39431824. Zhang: [email protected], (61) 2-93855834. Zhanyu Gao acknowledges the research support from the Early Career Scheme (ECS) (2192091). Bohui Zhang acknowledges the research grants from the ARC discovery grant (DP 120104755) and ARC linkage grant (LP130101050) from the Australian Research Council and the CIFR research grants (E026 and E028) from the Centre for International Finance and Regulation.
Googling Investor Sentiment around the World
Abstract
We study how investor sentiment affects stock markets around the world. Relying on the Google search behavior of households, we construct a weekly search-based measure of sentiment for 40 countries during the 2004–2014 period and provide evidence that the sentiment measure is a contrarian predictor of country-level market returns worldwide. Quasi-natural experiments, based on the implementation of the Markets in Financial Instruments Directive (MiFID) and short-selling bans, support two sentiment channels: the difficult-to-value channel and the limits-to-arbitrage channel. We also identify the fact that sentiment travels across countries using the adoption of International Financial Reporting Standards (IFRS) in the European Union in 2005. Our paper supports the behavioral view that sentiment prevails in stock markets. Keywords: Sentiment; Google search; International markets; Limits to arbitrage; IFRS; Sentiment co-movement JEL Code: G12; G14; G15
1
1. Introduction
Baker and Wurgler (2006, BW hereafter) suggest that investor sentiment plays an important
role in explaining stock price variation. The rationale for the significant return impact of investor
sentiment is that noise traders are subject to sentiment and rational investors are limited by
arbitrage constraints. This explanation is built on both De Long et al.’s (1990) sentiment
conjecture and Shleifer and Vishny’s (1997) limits-to-arbitrage theory. Empirically, BW and
Baker, Wurgler, and Yuan (2012) present affirmative evidence that a composite market-based
sentiment index predicts stock returns in the U.S. and five non-U.S. developed markets.
Despite the intuitive explanation for this intriguing observation, recent studies challenge the
behavioral view that the return prediction of sentiment indexes is driven by sentiment. Critiques
focus on three aspects. First, the power of the sentiment index to predict stock returns is mainly
driven by the business cycle and risk component (e.g., Sibley, Xing, and Zhang, 2012). Second,
many seemingly irrational phenomena and anecdotal evidence of investor sentiment can be
explained in rational models (e.g., Pastor and Veronesi, 2003, 2005, and 2006). Third, noise
traders can only induce large price movements and excess volatility in the short run (Kogan et al.,
2006, 2009).
Given the exploding interests of investor sentiment in finance literature, it is imperative to
resolve the above debate on the market prevalence of sentiment using alternative sentiment
measures and out-of-sample analyses. Thus, in this paper, we measure investor sentiment
through the Google search behavior of households newly used in Da, Engelberg, and Gao (2015),
and explore the sentiment effect at a weekly frequency in a large sample of 40 countries between
2004 and 2014. According to NetMarketShare’s statistics in 2014, Google occupies 67.5%
global market shares in household search. Though censored or banned in some countries like
China, Google dominates most developed and emerging markets such as Australia (93.0%),
2
Brazil (96.9%), India (97.0%), Japan (75.3%), and Poland (97.4%). Serving as the most popular
search engine around the world, Google search provides a unique international platform to our
cross-country study on investor sentiment.
This paper has three objectives. First, we investigate whether the search-based sentiment
measure is a contrarian predictor of country-level market returns worldwide. Second, we test
BW’s theoretical channels on the return prediction of investor sentiment: the difficult-to-value
channel and the limits-to-arbitrage channel using two types of country-level quasi-natural
experiments. Finally, we study how sentiment travels across countries using the adoption of
International Financial Reporting Standards (IFRS) in the European Union in 2005, which
improved accounting comparability and increased foreign investments.
Our results support the prevalence of sentiment in stock markets around the world. All 40
countries in our sample present a negative relationship between sentiment and the next week’s
market returns. Among them, 29 countries display a significant pattern at the 5% level. In terms
of economic significance, a one standard deviation increase in investor sentiment predicts a
weekly return decline of 44 basis points. Also consistent with the sentiment theory, there are
sentiment reversals and contemporaneous co-movement between sentiment and market returns.
Our results are also robust to the exclusion of financial crisis period, local currency returns,
longer time horizons, and the inclusion of additional control variables.
Next, the two country-level experiments offer supporting evidence to BW’s sentiment
channels: the difficult-to-value channel and the limits-to-arbitrage channel. First, we study the
implementation of the Markets in Financial Instruments Directive (MiFID) as an exogenous
shock to the information environment. With the improvement of market transparency and
efficiency, we indeed find that the asset valuation would be less subjective and investors are less
3
likely to react to the market sentiment. This finding supports the difficult-to-value channel. Our
second experiment is to study the short-selling ban as evidence of limits to arbitrage. In response
to the market crashes during the financial crisis in 2007 to 2009, several countries banned the
short-selling of financial stocks or all stocks for a period of time. Consistent with the limits-to-
arbitrage channel, our sentiment measure has a larger impact on the markets experiencing the
short-selling ban because rational investors find it difficult in correcting the mispricing.
Finally, by aggregating country-level sentiment indexes into a global sentiment measure, we
find that global sentiment prevails in the international markets. Specifically, we start by
documenting the commonality of sentiment across countries. In that, global sentiment on average
explains 16.5% sentiment variation for all countries and 21.2% for developed countries.
Moreover, we use the implementations of IFRS across countries and over time, as a country-
level experiment to test the effect of capital market integration on sentiment co-movement. The
result shows that global sentiment has a larger positive impact on market sentiment of countries
that adopted IFRS than countries that did not. Last, we document that 30 countries have global
sentiment that could significantly predict future market returns while only 11 countries have
significant return predicting power of local sentiment.
The existing literature investigating investor sentiment centers on an individual market or
specific sentiment-triggering events. Most of previous research focuses on the U.S. equity
market. Besides WB and Da, Engelberg, and Gao (2015), Tetlock (2007) finds the media
pessimism could predict the market price reversals and trading volume; Das and Chen (2007)
also employ the text based methodology to extract sentiment from stock message boards; Brown
and Cliff (2004), Lemmon and Portniaguina (2006), and Qiu and Welch (2004) use survey-based
sentiment index to examine its relationship with stock market returns. Yu and Yuan (2011) and
4
Stambaugh, Yu, and Yuan (2012, 2014) explore the role of market sentiment based on WB in
explaining well-known market anomalies.
A literature studying the international markets looks into specific events or environment
which could trigger investor sentiment. For example, Hirshleifer and Shumway (2003) find that
morning sunshine would affect stock market indices around the world. Kamstra, Kramer, and
Levi link seasonal variation in the length or daylight and depression, which could in turn
influence risk aversion and stock performances across countries located at different latitudes.
Edmans, Garcia, and Norli (2007) attributes the next day abnormal stock returns to investor
sentiment shock after the international sports.
In this paper, we have made several important contributions to the behavioral finance and
international finance literature. First, we construct high frequency sentiment indices for 40
representative markets around the world using Google search data. To the best of our knowledge,
this is most comprehensive international sentiment. Baker, Wurgler, and Yu (2012) develop the
annual indices for six advanced stock markets based on the approach in Baker and Wurgler
(2006). Our weekly indices have a broader coverage of both major developed and emerging
markets. Researchers hence could use our measures to conduct extensive studies in the
international markets.
Second, we find that our sentiment indices could predict the market returns in most of
countries in our sample. The reversal pattern seems ubiquitous around the world, suggesting
sentiment has a universal impact on the market prices.
Third, we verify our sentiment indices by testing two channels of the sentiment effect
separately using novel experiments in the international setting. International markets provide
unique opportunities in examining the mechanism of sentiment effects.
5
Fourth, we build a global sentiment index and investigate the co-movement between country
specific sentiment and global sentiment. We find that IFRS adoptions intensify the sentiment co-
movement, which is a new feature of financial integration.
The remainder of the paper proceeds as follows. We explain the variable construction for the
search-based measure of investor sentiment and describe sample characteristics in Section 2. In
Section 3, we study the relationship between our sentiment indices and the market returns across
countries. In Section 4, we test BW’s theoretical channels on sentiment: the difficult-to-value
channel and the limits-to-arbitrage channel. In Section 5, we study the commonality of sentiment
across countries. Finally, we provide concluding remarks in Section 6.
2 Data and Methodology
2.1. Google Search
Google search is the largest search engine around the world, capturing 67.5% global market
shares.1 Panel A of Table 1 reports Google search volume per year, month, day, and second
respectively from 2004 to 2014. There were 2.095 trillion search queries through Google in 2014;
66,440 searches queries were performed on Google every second all over the world. This figure
increases by almost 25 times from 2,730 per second in 2004. Moreover, Google search prevails
across a large number of countries including both developed and emerging markets. As an
American company, Google accounts for 80.6% in the U.S. search engine market (the fourth
column of Panel B in Table 1). Google dominates many countries’ search markets with even
higher shares. In our sample of 40 countries, 27 countries have Google market shares over 90%.
In Thailand, for example, Google search reaches 99% market shares. Google search, as the most
1 See https://www.netmarketshare.com/search-engine-market-share.aspx?qprid=4&qpcustomd=0
6
popular search engine in most of countries, provides an ideal platform to track households’
interest and attention, and reveal their sentiment around the world.
[Insert Table 1 Here]
2.2. Sentiment Index Constructions
We follow the approach in Da, Engelberg, and Gao (2015) to develop the weekly sentiment
indices of 40 countries around the world. Following we briefly describe the main construction
steps in Da, Engelberg, and Gao (2015) and highlight the differences in our method applied to
this international setting.
The main data source, Google Trends (https://www.google.com/trends/) provides the Search
Volume Index (SVI) of any search item across various countries, in different languages, and
during specific period of time. Our task is to track the search activity and measure the economy
related sentiment of households in each country of our sample. We use main languages in every
individual country to download the weekly SVI of economics and finance related terms exactly
searched from that country.
We begin with the same primitive word list in Da, Engelberg, and Gao (2015). These
economics and finance related words are collected from the Havard IV-4 Dictionary and the
Lasswell Value Dictionary and all have a positive or negative sentiment tag. To build a list of
search terms related to these primitive words, we translate them into the corresponding languages
using Google Translate (https://translate.google.com/), input these translated words into Google
Trends, and choose the specific country as the search region. We find ten “top searches”
associated with each primitive word in the specific languages of each country.
7
Da, Engelberg, and Gao (2015) download the daily data and keep the search terms with more
than 1,000 observations. Similarly, since we are interested in weekly data for our international
setting, we only keep terms with at least 100 observations. We also remove items without direct
relationship with economics and finance and then have the search list in the specific language for
each country. Next, we download the weekly SVI (covering search volume from Sunday to
Saturday) of these terms of each list searched from its corresponding country from January 2004
to December 2014. We also calculate the SVI changes and winsorize the observations, eliminate
seasonality, and finally standardize each time series to make all records comparable.
Next we employ the methodology in Da, Engelberg, and Gao (2015), and let the market data
speak itself. We use the terms most related to the each country’s market returns to construct our
index in that country. We run expanding backward rolling regressions of adjusted SVI changes
(ΔASVI) on every country’s market returns and identify the relationship between the search and
returns.2 In U.S. data sample, Da, Engelberg, and Gao (2015) find no terms with sufficient
positive t-statistics and thus focus on the top 30 negative terms in constructing their “FEARS”
index (“Financial and Economic Attitudes Revealed by Search” index). Differently, many
countries in our samples do have terms with high positive t-statistics. It suggests that outside of
the U.S., positive terms also play a role in measuring the household sentiment. To fully explore
the useful information, we construct the our sentiment indices by averaging SVI of the top 30
positive and the top 30 negative search terms respectively and calculate the difference as the
measure of sentiment, specifically:
𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑡 = �𝑅+𝑖 (30
𝑖=1
ΔASVI𝑖) −�𝑅−𝑖 (30
𝑖=1
ΔASVI𝑖)
2 Please refer to Da, Engelberg, and Gao (2015) for details about expanding backward rolling regressions.
8
where ∑ 𝑅±𝑖 (30
𝑖=1 ΔASVI𝑖) is the t-statistics weighted average of top 30 positive (negative) search
items.
Finally, we leave the first six month as the initial rolling regression window and form the
weekly sentiment proxy for 40 countries from July, 2004 to December, 2014.
Besides internet usage shares and Google search market shares, Panel B of Table 1 also
displays the languages we employ to build our sentiment indices and their corresponding
population shares in each country. We choose the major and official languages in each country.
If a country has multiple official languages, we download the data in these languages and keep
the series with the most weekly observations.
Table 2 presents summary statistics of our sentiment indices. Each country has a 548 week
long sentiment index series from July 2014 to December, 2014. The sentiment index has a mean
close to zero across countries and there is a large variation over time, suggesting sentiment
swings between the positive and the negative. Moreover, standard deviations of index vary from
country to country. For instance, Norway has a standard deviation close to 1.3 while it’s only
0.28 in Indonesia. The table also reports the quartiles of index distributions. We find that the
median is higher than the mean in most of countries, implying the negative skewness of
distribution. Although we may observe positive sentiment more frequently, the negative
sentiment could be in a larger magnitude than the positive.
[Insert Table 2 Here]
2.3. Other data
Our market return data come from DataStream. We download the country-level daily total
return index RI in U.S. dollars from 2004 to 2015. Because Google Trends records the weekly
9
search queries from Sunday to Saturday, it’s appropriate to look at the market index in the
middle between Sunday and Saturday, and then calculate the weekly return between two
Wednesdays, which are exactly the middle days of two consecutive weeks. We also download
the market index in the country currency to test the robustness of results.
In addition, to control some economic conditions, we collect the variables including Market
Volatility Index (VIX in the Chicago Board Options Exchange), macroeconomic activities (ADS,
constructed by Aruoba, Diebold, and Scoitti, 2009), and economic policy uncertainty (EPU,
developed by Baker, Bloom, and Davis, 2013) which are also used in Da, Engelberg, and Gao
(2015).
3. Sentiment and return predictability
In this section, we study the relationship between our sentiment indices and the market returns
across countries. According to the theory on sentiment, sentiment drives market prices away
from the economic fundamentals if there are limits to arbitrage. If high or low sentiment is not
persisting, the market prices will revert to the fundamental after the sentiment retreats.
We have three basic predictions related to the theory. First, sentiment should be reverse rather
than persistent; Second, there is a positive contemporary relationship between sentiment and
market returns as high (low) sentiment pushes the price to increase (decrease) from the previous
period; last, return should present reversal pattern following the sentiment, i.e., high (low)
sentiment would predict the low (high) return in the following week.
[Insert Table 3 Here]
10
We first examine the autocorrelation of our sentiment index by country. The first part of Panel
A in Table 3 reports the coefficients, t-statistics and R-squared when regressing sentiment index
on last week’s sentiment for each country and Panel B summarizes the average coefficient, t-
statistics, R-squared and the number of countries which display the significant negative
autocorrelation at 5% level. As a result, all countries in our sample report the strong reversal
patterns, consistent with our first prediction that the high or low level of sentiment is temporary
and market sentiment fluctuates frequently over time. Interestingly, the average coefficient of
developed markets is -0.386, lower than the coefficient of -0.304 in emerging markets,
suggesting the reversal pattern is stronger in developed markets. Relative to developed markets,
sentiment in emerging markets could be more persistent.
We then visit the second prediction by examining the contemporary relationship between
sentiment and market returns across countries in the second part of Panel A of Table 3 and
summarize the results in Panel B. All 40 countries in our sample exhibit a positive relationship:
the market prices increase (decrease) at high (low) level of sentiment. On average, the weekly
market returns increase by 1.49% with one standard deviation rise in our sentiment index.
The most important prediction is about the return reversals following the market sentiment.
The third part of Panel A and B in Table 3 verifies the predictability of sentiment for the
following week’s market returns. All 40 countries in our sample present a negative relationship
between sentiment and the next week’s market returns while 29 countries display a significant
pattern at the 5% level. Interestingly, most of insignificant relationships occur in the emerging
markets while only Singapore as a developed market has an insignificant result. The possible
reason is that households in some emerging markets have limited access to the internet (e.g.
India with internet usage percentage 7.9% shown in Panel B of Table 1) or Google search does
11
not account for the major market shares (e.g. China with 23.8% Google market shares from
Panel B of Table 1). The sentiment constructed by Google search could not fully capture the
market sentiment in the country and presents the limited prediction power than other countries.
Specifically, on average, one standard deviation increases in market sentiment predicts a
weekly return decline of 44 basis points. Among all the countries, Thailand has a strongest
reverse pattern as the one standard deviation rise in sentiment predicts the weekly return to
decrease by 92 basis points. In the developed markets, Sweden is associated with the strongest
reversal patterns: 72 basis points decrease in the weekly return after one standard deviation
increase in sentiment.
To examine the robustness of these three basic predictions, we conduct several additional tests
related to our sentiment indices and summarize the main findings in Table 4. First, we remove
the global financial crisis period of September 2008 to August 2009, as one may concern the
pessimism would prevail all over the world during Great Recession and the consequent global
economic turmoil. Indeed we find the stronger sentiment swings during the financial crisis (also
see in Figure 1 for global sentiment movement). Removing this period would lower the level of
reversal pattern (the coefficient of the third regression changes from -0.476 to -0.378). However,
as evident in Panel A, most of countries still display the significant relationship in non-crisis
period, confirming that our sentiment indices could predict the market returns in the following
week.
We then employ the country’s own currency and calculate the market returns again to isolate
the local market responses to sentiment and exclude the potential role played by the movement
of exchange rates. Form Panel B, we find the three tests are still robust while presenting slightly
12
moderate effects than the basic results. Sentiment could also influence the foreign exchange
markets, however it’s beyond the scope of this paper and we leave it for the future research.
Moreover, we look into the longer horizon by extending the one week to two weeks. The
reversal pattern exists in the two-week window and Panel C reports the similar coefficients as
Panel B of Table 3 (one week horizon) does. This implies that the market returns in the second
week less likely respond sentiment this week and supports our choice of one week return in our
main analysis.
We add more controls including ADS, EPU, VIX and five market return lags to test our
results. The controlled regressions display even stronger sentiment predictability as the
coefficient of sentiment index on average become -0.527, in contrast with -0.476 in the basic
return reversal tests.
[Insert Table 4 Here]
In summary, our sentiment measure could predict the future market returns across countries.
Consistent with the sentiment theory, we find sentiment’s transitoriness, the co-movement
between market index and sentiment, and the next week return reversals with respect to
sentiment. To further verify the effect of sentiment on the market index across country, we
consider several natural experiments in the international setting and test key channels separately.
Note that in the following analysis, we exclude 5 countries from our 40 country sample:
Belgium, China, Philippines, Russia, and Switzerland. We do this for two reasons: low market
shares of Google search and the insufficient observations due to multiple languages spoken in a
country. As shown in Panel B of Table 1, Google search accounts for less than 30% market
shares in China (23.8%) and Russia (23.3%). Belgium, Switzerland, and Philippines have
13
multiple speaking and official languages and none of these languages have sufficient
observations to construct reliable sentiment indices. Therefore we focus on the other 35 countries
with high quality of sentiment indices and conduct the rest of tests.
4. Channels of Sentiment Effect
Sentiment reflects households’ belief deviating from the fundamentals. The predictability of
sentiment depends on two key factors: first, to what extent the households are subject to the
sentiment effects; second, markets are under the influence of limits to arbitrage and hence could
not correct the inefficiency (Shleifer, 2000). Given the heterogeneity of institutional environment
and the implementations of regulation, international markets provide us with opportunities to
examine these two distinct channels of behavioral finance. The previous literature mainly studies
the cross section of stocks subject to market sentiment (Baker and Wurgler, 2006, 2007;
Stambaugh, Yu, and Yuan, 2011 among others). Baker and Wurgler (2006) identify the stocks
that are highly subjective to value and costly and risky to arbitrage. They argue that these stocks
happen to be in the same category.
In contrast, our international setting allows us to employ natural experiments to identify these
two channels separately. Our difference-in-difference tests avoid the potentially endogenous
issues which previous studies find difficult to resolve.
We consider two quasi-natural experiments in this section. First, we study the implementation
of the Markets in Financial Instruments Directive (MiFID) as an exogenous shock to the
information environment. With the improvement of market transparency and efficiency, we
expect that the asset valuation would be less subjective and investors are less likely to react to
the market sentiment.
14
Our second experiment is to study the short-selling ban as evidence of limits to arbitrage. In
response to the market crashes during the financial crisis in 2007 to 2009, several countries
banned the short-selling of financial stocks or all stocks for a period of time. Sentiment is
expected to have a larger impact on the markets with more limits to arbitrage as investors find it
difficult in correcting the mispricing.
4.1. The Difficult-to-value Channel and MiFID
The Markets in Financial Instruments Directive (MiFID) is an influential European Union
financial markets directive that became effective for all EU members on November 1 2007. It is
aimed at establishing an integrated, transparent, competitive and efficient European financial
market. It contains a series of harmonized rules that enhance investor protection and market
transparence. Following the implementation of MiFID, trading rules of European exchanges
become more comprehensive and market information environment improves (Cumming, Johan
and Li 2011). We thus make use of the implementation of MiFID to test whether changes in
market information environment would impact the relation between future return and current
sentiment. Specifically, we collect the MiFID implementation information for our 35 countries
and obtain seven indices of exchange trading rules from Cumming, Johan and Li (2011). We use
these seven indices of exchange trading rules to proxy for the information environment of a
country’s stock market. The seven indices of exchange trading rules are Insider Trading Rules
Index (ITI), Market Manipulation Rules Index (MMI), Broker-Agency Conflict Rules Index
(BAI), Price Manipulation Rules Index (PMI), Volume Manipulation Rules Index (VMI),
Spoofing Manipulation Rules Index (SI) and False Disclosure Rules Index (FDI). MMI
encompasses PMI, VMI, SI and FDI. Detailed definitions of these indices are in Appendix. Table
15
5 presents the changes in these indices after November 2007 for each country. Some of the
indices of European countries change significantly after November 2007 while other countries
that are not subject to MiFID experience no change in these indices.
[Insert Table 5 Here]
To test the impact of information environment on relation between return and sentiment, we
run the following regression:
Returni,t+1=α+β1Sentimenti,t+ β2EUi,t+ β3Sentimenti,t* EUi,t+εi,t
where Returni,t+1 is country i’s market return in week t+1 and Sentimenti,t is country i’s market
sentiment in week t. In Model (2) of Table 6, EUi,t equals to 1 if a country i is exposed to MiFID
at time t and 0 otherwise. In Model (3) to Model (9), EUi,t equals to the change in the
corresponding index of exchange trading rules if the date is after November 1, 2007 and 0
otherwise. When stock market information environment improves, stock price would respond
less to changes in market sentiment. We thus expect β3 to be positive and significant. The main
results in Panel A of Table 6 confirm our prediction. Consistent with results in prior section, β1
is always negative and significant in all of the specifications, suggesting that current period
market sentiment is negatively related to next period market return. Most importantly, β3 is
positive and significant in seven out of eight models, indicating that the implementation of
MiFID and the increase in the six indices of exchange rules would weaken the relation between
current market sentiment and future market return.
The effect is also economically significant. For example, according to the result of Model (2),
an increase of one standard deviation in market sentiment of countries that are not affected by
MiFID is associated with a decrease of 28.2% in market return while for countries that are
16
subject to MiFID, the corresponding decrease in market return is 10.8%. In addition, changes in
six indices of exchange rule have different degrees of impacts on the relation between market
sentiment and market return. Specifically, increases in FDI would strongly weaken the negative
relation between current sentiment and future return: an increase of one standard deviation in
FDI would decrease the sensitivity of market return to market sentiment by 35.9%
( β3*std.∆FDI/ β1) while changes in BAI would have smaller impacts: an increase of one
standard deviation in BAI decreases the sensitivity of market return to market sentiment by
23.5%.
Panel B and Panel C report robustness tests using a different sample of countries and a
different sample period, respectively. Specifically, In Panel B, we match 13 countries that are not
impacted by MiFID with the 13 European countries that are subject to MiFID using the 11-year
(2004 to 2014) average of GDP per Capita and repeat the regressions in Panel A with the
matched sample. Panel C reports the results of regressions based on a longer sample period:
November 2006 (one year before the event date) to September 2008. The results in Panel B and
Panel C are quite similar to that in Panel A: β3 is positive and significant in most of the
specifications and the magnitude of the effect is also economically significant. To provide
evidence that the effect we find in Panel A is truly due to MiFID, we conduct a placebo test and
report the results in Panel D. Specifically, we bootstrap 13 event dates between May 2005 and
February 2014 and assign these bootstrapped pseudo MiFID event dates to the 13 European
countries that are subject to MiFID listed in Table 5. We then run the regression in Panel A with
the new MiFID event dates and obtain β3. We repeat the procedures for 1000 times, which leave
us with 1000 β3s. We sort these 1000 β3s from smallest to largest and report the 90th percentile,
95th percentile and 99th percentile in Panel D. In seven out of eight specifications, our β3 based
17
on the true MiFID effective date is larger than the 99th percentile of β3 generated from
bootstrapped pseudo MiFID effective date. The only exception is the model based on Broker-
Agency Rules Index (BAI) where β3 of Panel A is larger than 95th percentile but smaller than the
99th percentile.
[Insert Table 6 Here]
4.2. The Limits to Arbitrage Channel and Short-Selling Bans
There are significant variations in scopes and types of bans on short-selling across countries
(McKenzie et.al (2011)). Even within a country, short-selling policy may change over time.
During the financial crisis period of 2007-2009, many countries that did not forbid short sales
before imposed short-selling bans on either financial stocks or all stocks. Different countries
imposed and then lifted these bans on different dates. There are two types of bans imposed on
short sales during the financial crisis period: bans on naked short sales where the seller does not
first borrow the security and bans on covered short sales where the seller manages to borrow the
security. Some countries only impose the naked short sales while other countries may use a
combination of these two types. According to Beber and Pagano (2013), covered short-selling
bans are more frequently used than naked short-selling bans at the beginning of financial crisis
period while at the end of the financial crisis, only naked bans were in effect.
We collect the characteristics of short-selling bans for each of our 35 countries from Beber
and Pagano (2013) and McKenzie et.al (2011). As shown in Table 7, the inception dates of short-
selling ban range from 19 September 2008 (Canada, United Kingdom and United States) to 30
October 2008 (Japan). The durations of the bans also vary across countries with the longest being
over two years (e.g. Japan) and the shortest being less than one month (e.g. United States). We
18
choose six months before the first ban inception date and six months after the last ban inception
date as the sample period for our main regressions. There are also countries that never imposed
bans during our sample period and other countries that always forbid short-selling on stocks.
[Insert Table 7 Here]
To test the impact of short-selling bans on the relation between market stock return and
market sentiment, we run the following panel regressions:
Returni,t+1=α+β1Sentimenti,t+ β2Bani,t+ β3Sentimenti,t* Bani,t+εi,t
where Returni,t+1 is country i’s market return in week t+1 and Sentimenti,t is country i’s total
market sentiment in week t. Bani,t equals to 1 if a country i has imposed a short-selling ban on
either financial or all stocks at week t and 0 otherwise. Our main interest is the coefficient on the
interaction term. We conjecture that short-selling bans would make it more difficult for
investors to correct the stock mispricing driven by market sentiment, which would lead to a
larger impact of current week market sentiment on next week market return. Given the negative
relation between future return and current sentiment documented in previous section, we expect
β3 to be significantly negative. The results in Panel A of Table 8 are consistent with our
prediction. A negative β3 significant at 5% indicates that relative to countries that do not impose
short-selling bans on stocks, countries with short-selling bans show a larger future return
sensitivity of current market sentiment. In addition, compared with a β1 of -0.529, a β3 of -0.572
suggests that the effect of short-selling ban is also economically significant. One standard
deviation increase in market sentiment of countries with no bans on short sales is associated with
a decrease of 49 basis points in weekly market returns while for countries with bans on short
sales, the corresponding decrease is 1.01%.
19
We conduct several robustness tests in Panel B and Panel C. In Model (1) of Panel B, we use
a different sample period (January 2008-June 2009) and In Model (2) we use the countries in
Beber and Pagano (2013). In Model (3), we match the 15 countries (control) that never imposed
short-selling bans during the sample period with 15 countries (treated) that did based on 11-year
(2004 to 2014) average of GDP per Capita. We then repeat the regressions using the matched
sample of 30 countries. As shown in Panel B, our results in Panel A are robust to using different
countries and different sample periods. We do a placebo test in Panel C. Specifically, we first
bootstrap a 40-day window between January 2005 and April 2014. Then within the 40-day
window, we randomly select (with replacement) 17 event dates as the pseudo inception dates of
short-selling bans and allocate them to the 17 countries that have imposed a short-selling ban
during the sample period. The durations of each short-selling ban for each country remains the
same as those in Table 7. We set the sample period as six months before the first pseudo ban
inception date and six months after the last pseudo ban inception date. We then run the
regressions in Panel A using the bootstrapped event dates and keep β3. We repeat the procedure
1000 times and report the 90th, 95th and 99th percentile of β3 in Panel C. The result in Panel C
shows that β3 in Panel A is beyond the 95th percentile of the 1000 β3s generated from
bootstrapping the event dates. We take this result as supporting evidence that it is indeed the
short-selling bans that are affecting the relation between future return and current market
sentiment.
[Insert Table 8 Here]
5. Global Sentiment and Sentiment Co-movement
20
Does sentiment across countries co-move with each other? Is there global sentiment that
affects sentiment in specific countries? As our sentiment indices is based on Google search, the
most popular search engine around the world, with this international platform, we are able to
study the commonality of sentiment across countries. Moreover, we examine whether capital
market integration has an impact on sentiment co-movement. We use another natural experiment
to test this hypothesis: the implementations of the International Financial Reporting Standards
(IFRS) across countries and over time.
5.1. Global Sentiment and Country Specific Sentiment
In this sub-section, we construct a global sentiment index and examine to what extent global
sentiment would affect local sentiment. We simply average Google sentiment measures of 35
countries in our sample to form the global sentiment index. We have also considered different
weighted averages including the population weighted, GDP per capita weighted, Internet usage
weighted, and Google market share weighted sentiment averages across countries, and all yield
the similar results. To make our construction more transparent, we choose to use simple averages
when reporting our following analysis.
Figure 1 depicts the weekly time series of our global sentiment from 2004 to 2014. Global
sentiment index is immensely volatile. Not surprisingly, we observe several sharp spikes of
global sentiment during the 2007-2008 financial crisis, and 2011-2002 when Eurozone crisis
intensified. Given the vibration characteristic of sentiment, the high frequency data are valuable
to study the sentiment effect.
[Insert Figure 1 Here]
21
We then examine how individual country’s sentiment is correlated with global sentiment.
Table 9 reports the results of regressing country sentiment on global sentiment for 35 countries
in Panel A and summarizes the results in Panel B. All countries display a significant positive
relationship with global sentiment. Integrated into the global economy to a greater extent,
developed countries’ sentiment has a higher correlation with global sentiment than emerging
countries. Specifically, global sentiment could on average explain 16.5% sentiment variation for
all countries and 21.2% for developed countries. The United States’ market sentiment has a
relatively low correlation with global sentiment, suggesting the largest economy is less subject to
global sentiment movement. Austria ranks at the top of the list with 33.4% country specific
variation explained by global sentiment.
[Insert Table 9 Here]
The results above suggest that global market integration could play a potential role in
affecting the co-movement of sentiment across countries. But it could be endogenous from the
simple cross-sectional observations. We thus employ a natural experiment of IFRS adoptions to
examine this causal relationship.
5.2. Market Integration and IFRS
International Financial Reporting Standards (IFRS) has been adopted widely around the
world. According to the website of the IFRS Foundation and International Accounting Standards
Board (IASB), at least 116 jurisdictions over the world have required IFRS for all or most
publicly accountable entities up to 2015. As stated by IFRS Foundation and IASB, their mission
is “to develop IFRS that bring transparency, accountability and efficiency to financial markets
22
around the world…” However, the true effects of IFRS are not as simple as stated above. In this
paper we focus on the effect of IFRS adoption on capital market integration. Prior literature
shows that IFRS would improve information transfer and capital flows across countries, which
would contribute to a higher degree of capital market integration (refer to George, Li and
ShivaKumar (2015) for a good review on IFRS literature). We thus make use of the adoption of
IFRS to test how market integration would affect the co-movement between a country’s market
sentiment and global sentiment.
Table 10 shows the mandatory IFRS adoption years for each of the 35 countries in our sample.
European countries, Hong Kong and South Africa adopted IFRS in 2005, which makes the year
of 2005 a frequently used event year in prior IFRS literature. Several countries adopted IFRS
during 2007 to 2013 while other countries like Japan and India never adopted IFRS during our
sample period. Given the relatively wide distribution of the adoption years of IFRS, we use the
full sample period of 2004 to 2014 for our main regressions. In the robustness tests, we also
follow the prior literature to use event windows around 2005.
[Insert Table 10 Here]
We use the mandatory adoption of IFRS as a proxy for market integration and test whether a
higher level of market integration would enhance the co-movement between an individual
country’s market sentiment and the global sentiment. Specifically, we run the following
regression:
Sentimenti,t=α+β1SentimentG,t+ β2SentimentG,t*IFRSi,t+ β3IFRSi,t+εi,t
where Sentimenti,t is country i’s market sentiment in week t and SentimentG,t is the simple
average of Sentimenti,t across the 35 countries in week t. IFRSi,t equals to 1 when a country
23
adopted IFRS at year t-1 and 0 otherwise (note that subscript t denotes week for sentiment and
year for IFRS). Panel A of Table 11 presents our main results where Model (1) is the baseline
model without the IFRS dummy and Model (2) is our main regression with the IFRS dummy.
The coefficients of SentimentG,t are positive and significant at 1% level in both Model (1) and
Model (2), indicating that countries’ market sentiment is positively correlated with the global
sentiment. Our focus is β2, the interaction of global sentiment and IFRS adoption dummy. As
shown in the table, β2 is positive and significant at 1% level, which suggests that global
sentiment has a larger positive impact on market sentiment of countries that adopted IFRS than
countries that did not. The effect is also economically significant. Specifically, for countries that
do not adopt IFRS, an increase of one standard deviation in global sentiment is related with an
increase of 15.2% in countries’ market sentiment while for countries that adopt IFRS, the
increase in market sentiment associated with one standard deviation increase in global sentiment
is 21.5%. It is consistent with our prediction that market integration increases the co-movement
between countries’ market sentiment and the global sentiment.
Panel B reports the results of two robustness tests. In Model (1), instead of using the full
sample period of 2004 to 2014, we choose 2004 to 2007 as the testing window where 2004 to
2005 act as the pre-IFRS adoption period and 2006 to 2007 act as the post-IFRS adoption period.
In Model (2), we match the 7 countries that did not adopt IFRS before 2014 in Table 10 with 7
countries that did according to 11-year (2004 to 2014) average of GDP per Capita. Consistent
with the results in Panel A, β2 remains positive and significant in both of the models in Panel B.
Panel C presents the results of placebo test. We bootstrap 28 IFRS adoption years from
2005 to 2014 and allocate them to the 28 countries that have adopted IFRS during the sample
period of 2005-2014. Using the newly bootstrapped IFRS adoption years, we then run the
24
regressions in Panel A and keep β2. We repeat the procedure 1000 times and report the 90th, 95th
and 99th percentile of β2s generated from bootstrapping in Panel C. We find that our β2 from the
main regression in Panel A is out of the 99th percentile of the β2s created from bootstrapping
IFRS adoption years.
[Insert Table 11 Here]
5.3. Global Sentiment and Return Predictability
Given the strong co-movement pattern between the country sentiment and global sentiment,
we decompose the specific sentiment of each country into global sentiment and local sentiment.
Local sentiment is the residual of regressing the individual country’s market sentiment on the
global sentiment. To examine the role played by global sentiment in predicting the future returns,
we then regress the market return next week on both global and local sentiment and report the
results in Table 12. Panel A presents the coefficients and t-statistics of global and local sentiment
for each country and Panel B summarize our findings. Out of 35 countries in our sample, 30
countries have global sentiment that could significantly predict the next week market returns. In
contrast, only 11 countries have significant coefficients of local sentiment in predicting the next
week returns.
Moreover, global sentiment has predictability power in all 17 developed countries while only
in 13 out of 18 emerging markets. On the other hand, local sentiment rarely has an impact on
predicting future returns in developed markets (2 out of 17) while local sentiment could be more
important to emerging markets as the coefficients are significant for half of these markets. As
evident, developed markets are more subject to global sentiment and emerging markets are more
subject to local sentiment. These findings again imply that developed countries are well
25
integrated to the global economy and thus their markets are more likely under the influence of
global sentiment.
[Insert Table 12 Here]
6. Conclusion
In this paper, we have made several important contributions to the literature. First, we
construct high frequency sentiment indices for 40 representative markets around the world using
Google search data. To the best of our knowledge, this is most comprehensive international
sentiment. Baker, Wurgler, and Yu (2012) develop the annual indices for six advanced stock
markets based on the approach in Baker and Wurgler (2006). Our weekly indices have a broader
coverage of both major developed and emerging markets. Researchers hence could use our
measures to conduct extensive studies in the international markets.
Second, we find that our sentiment indices could predict the market returns in most of
countries in our sample. The reversal pattern seems ubiquitous around the world, suggesting
sentiment has a universal impact on the market prices.
Third, we verify our sentiment indices by testing two channels of the sentiment effect
separately using novel experiments in the international setting. International markets provide
unique opportunities in examining the mechanism of sentiment effects.
Fourth, we build a global sentiment index and investigate the co-movement between country
specific sentiment and global sentiment. We find that IFRS adoptions intensify the sentiment co-
movement, which is a new feature of financial integration.
There is still much to be done for future research. With international sentiment indices,
researchers could potentially improve the understanding of international markets from behavioral
26
finance prospective. Also the heterogeneity across countries, global integration, and specific
policy and regulation implementations consists of opportunities to test the sentiment effect in a
broad context.
27
Appendix
Index Definitions (Cumming, Johan and Li 2011) Index Definition Market Manipulation Rules Index
Sum of Price Manipulation Rules Index,Volume Manipulation Rules Index, Spoofing Rules Index, and False Disclosure Rules Index.
Price Manipulation Rules Index
Sum of dummy variables for marking the open, marking the close, misleading end of the month/ quarter/year trades, intraday ramping/gouging, market setting, pre-arranged trades, and domination and control.
Volume Manipulation Rules Index
Sum of dummy variables for Churning and Wash trade
Spoofing Rules Index Sum of dummy variables for Giving up priority, Switch, and Layering of bids-
asks. False Disclosure Rules Index
Sum of dummy variables for Dissemination of false and misleading information and Parking or ware housing.
Insider Trading Rules Index
Sum of dummy variables for Front-running, Client precedence, Trading ahead of research reports, Separation of research and trading, Broker ownership limit, Restrictions on affiliation, Restrictions on communications, Investment company securities, Influencing or rewarding the employees of others, and Anti-intimidation/coordination.
Broker–Agency Index Sum of dummy variables for Trade through, Improper execution,Restrictions
on member use of exchange name,Restrictions on sales materials and telemarketing, and Fair dealing with customers.
28
References
Antweiler, W., Frank, M.Z., 2004. Is all that talk just noise? The information content of internet stock message boards. The Journal of Finance 59, 1259-1294
Aruoba, S.B., Diebold, F.X., Scotti, C., 2012. Real-time measurement of business conditions. Journal of
Business & Economic Statistics Baker, M., Wurgler, J., 2006. Investor sentiment and the cross‐section of stock returns. The Journal of
Finance 61, 1645-1680 Baker, M., Wurgler, J., 2007. Investor sentiment in the stock market. National Bureau of Economic
Research Cambridge, Mass., USA Baker, M., Wurgler, J., Yuan, Y., 2012. Global, local, and contagious investor sentiment. Journal of
Financial Economics 104, 272-287 Baker, S.R., Bloom, N., Davis, S.J., 2013. Measuring economic policy uncertainty. Chicago Booth
research paper Beber, A., Pagano, M., 2013. Short‐selling bans around the world: Evidence from the 2007–09 crisis. The
Journal of Finance 68, 343-381 Brown, G.W., Cliff, M.T., 2004. Investor sentiment and the near-term stock market. Journal of Empirical
Finance 11, 1-27 Christensen, H.B., Hail, L., Leuz, C., 2011. Capital-Market Effects of Securities Regulation: The Role of
Prior Regulation, Implementation and Enforcement. SSRN eLibrary Cumming, D., Johan, S., Li, D., 2011. Exchange trading rules and stock market liquidity. Journal of
Financial Economics 99, 651-671 Da, Z., Engelberg, J., Gao, P., 2011. In search of attention. The Journal of Finance 66, 1461-1499 Da, Z., Engelberg, J., Gao, P., 2015. The Sum of All FEARS Investor Sentiment and Asset Prices.
Review of Financial Studies 28, 1-32 Das, S.R., Chen, M.Y., 2007. Yahoo! for Amazon: Sentiment extraction from small talk on the web.
Management Science 53, 1375-1388 De George, E.T., Li, X., Shivakumar, L., 2015. A Review of the IFRS-Adoption Literature. Available at
SSRN 2664475 De Long, J.B., Shleifer, A., Summers, L.H., Waldmann, R.J., 1990. Noise trader risk in financial markets.
Journal of political Economy, 703-738 Edmans, A., Garcia, D., Norli, Ø., 2007. Sports sentiment and stock returns. The Journal of Finance 62,
1967-1998 Gao, P., Parsons, C.A., Shen, J., 2013. The global relation between financial distress and equity returns.
Available at SSRN 2086475
29
Hirshleifer, D., Shumway, T., 2003. Good day sunshine: Stock returns and the weather. Journal of finance
58 Jain, A., Jain, P., McInish, T., McKenzie, M., 2010. Worldwide short selling: Regulations, activity. and
implications, Working Paper Kamstra, M.J., Kramer, L.A., Levi, M.D., 2003. Winter blues: A SAD stock market cycle. American
Economic Review, 324-343 Kogan, L., Ross, S., Wang, J., Westerfield, M.M., 2009. Market selection. National Bureau of Economic
Research Kogan, L., Ross, S.A., Wang, J., Westerfield, M.M., 2006. The price impact and survival of irrational
traders. The Journal of Finance 61, 195-229 Lemmon, M., Portniaguina, E., 2006. Consumer confidence and asset prices: Some empirical evidence.
Review of Financial Studies 19, 1499-1529 Pastor, L., Veronesi, P., 2002. Stock valuation and learning about profitability. National Bureau of
Economic Research Pástor, Ľ., Veronesi, P., 2005. Rational IPO waves. The Journal of Finance 60, 1713-1757 Pástor, Ľ., Veronesi, P., 2006. Was there a Nasdaq bubble in the late 1990s? Journal of Financial
Economics 81, 61-100 Qiu, L., Welch, I., 2004. Investor sentiment measures. National Bureau of Economic Research Saunders, E.M., 1993. Stock prices and Wall Street weather. The American Economic Review, 1337-
1345 Shleifer, A., Vishny, R.W., 1997. The limits of arbitrage. The Journal of Finance 52, 35-55 Sibley, S., Xing, Y., Zhang, X., 2013. Is Sentiment Sentimental? Available at SSRN 2176348 Stambaugh, R.F., Yu, J., Yuan, Y., 2012. The short of it: Investor sentiment and anomalies. Journal of
Financial Economics 104, 288-302 Stambaugh, R.F., Yu, J., Yuan, Y., 2014. The long of it: Odds that investor sentiment spuriously predicts
anomaly returns. Journal of Financial Economics 114, 613-619 Tetlock, P.C., 2007. Giving content to investor sentiment: The role of media in the stock market. The
Journal of Finance 62, 1139-1168 Yu, J., Yuan, Y., 2011. Investor sentiment and the mean–variance relation. Journal of Financial
Economics 100, 367-381 Yuan, Y., 2015. Market-wide attention, trading, and stock returns. Journal of Financial Economics 116,
548-564
30
Figure 1 Weekly global sentiment This figure shows the weekly time series of our global sentiment index from June 2004 to December 2014.
31
Table 1 Summary statistics of Google search This table reports the summary statistics of Google search for 40 countries from 2004 to 2014. Panel A presents total google search volume3 (from annual to per second) around the world during the sample period. Panel B shows the details of the usage of Google for each of the 40 countries. Internet user (ITUsers) from World Bank is measured as the 11-year average of internet users per hundred people and google market share is Google’s market share4 in each country at the end of 2013. Column 5 and Column 6 show the languages used to construct the sentiment measure for each country and their corresponding population shares obtained from The World Factbook of CIA5 and Wikipedia, respectively. In Panel C, we regress the historical market weekly returns from Jun.2004 to Jun.2014 on the adjusted change in weekly search volume of a certain number of finance-related words and sort the words according to the t-statistics of their coefficients. Market weekly return is constructed based on daily total market index from datastream. Panel C lists three words with the largest negative t-statistics and three words with the largest positive t-statistics. DEV, EMG and WRD denote developed, emerging and global markets, respectively. Panel A: Google Search Volume Over the World
Year Annual (1012) Monthly (1011) Daily (109) Second (104) 2004 0.086 0.072 0.236 0.273 2005 0.141 0.118 0.386 0.447 2006 0.231 0.192 0.633 0.732 2007 0.438 0.365 1.200 1.389 2008 0.637 0.531 1.745 2.021 2009 0.954 0.795 2.610 3.024 2010 1.325 1.104 3.627 4.201 2011 1.722 1.435 4.717 5.461 2012 1.874 1.562 5.134 5.942 2013 2.162 1.801 5.922 6.854 2014 2.095 1.746 5.740 6.644
3 Data available at http://www.internetlivestats.com/google-search-statistics/#trend & http://www.statisticbrain.com/ google-searches/ 4 Data available at http://returnonnow.com/internet-marketing-resources/2013-search-engine-market-share-by-country/ & http://www.mvfglobal.com 5 Data available at https://www.cia.gov/library/publications/the-world-factbook/fields/2098.html
32
Panel B: IT Users per 100 People & Google Market Share Country EMG/DEV ITUsers GoogleShare Language LangShare Brazil EMG 39.1% 96.9% Portuguese Official Chile EMG 47.2% 97.8% Spanish Official(99.5%) China EMG 29.3% 23.8% Chinese Official Colombia EMG 33.3% 96.3% Spanish Official(99.2%) Egypt EMG 30.1% 97.0% Arabic Official(97.0%) Hungary EMG 61.1% 97.4% Hungarian Official(99.6%) India EMG 7.9% 97.0% English Subsidiary official Indonesia EMG 9.7% 96.0% Indonesian Official Israel EMG 57.2% 97.7% Hebrew Official Malaysia EMG 58.5% 82.1% Malay Official Mexico EMG 30.0% 92.6% Spanish Official(92.7%) Peru EMG 31.2% 98.0% Spanish Official(84.1%) Philippines EMG 19.7% 81.7% English Official Poland EMG 55.6% 97.4% Polish Official(98.2%) Portugal EMG 50.1% 97.1% Portuguese Official Russia EMG 39.2% 23.3% Russian Official(96.3%) South Africa EMG 23.8% 94.8% English Official South Korea EMG 81.4% 36.9% Korean Official Taiwan EMG 65.5% 47.6% Chinese Official Thailand EMG 22.1% 99.0% Thai Official(90.7%) Turkey EMG 35.4% 98.8% Turkish Official(85.0%) Australia DEV 74.5% 93.0% English Official(76.8%) Austria DEV 73.2% 92.8% German Official(88.6%) Belgium DEV 71.8% 98.5% Dutch Official(60.0%) Canada DEV 79.2% 92.7% English Official(58.7%) Denmark DEV 88.6% 95.0% Danish Official France DEV 69.8% 95.0% French Official(100%) Germany DEV 78.7% 91.7% German Official Hong Kong DEV 68.4% 68.0% English Official Italy DEV 48.8% 86.0% Italian Official(99%) Japan DEV 77.9% 75.3% Japanese Official(99%) Netherlands DEV 89.0% 94.7% Dutch Official New Zealand DEV 76.2% 92.0% English English(89.8%) Norway DEV 90.6% 96.1% Norwegian Official(95%) Singapore DEV 68.8% 84.4% English Official(29.8%) Spain DEV 62.2% 92.9% Spanish Official(74%) Sweden DEV 90.1% 96.8% Swedish Official Switzerland DEV 81.1% 92.0% German Official(64.9%) United Kingdom DEV 81.3% 89.9% English Official(95.0%) United States DEV 74.6% 80.6% English Official(79.2%) Total WRD 56.8% 86.4% N.A. N.A.
33
Table 2 Summary Statistics of Sentiment This table shows the summary statistics of weekly our sentiment measure for 40 countries over the sample period of Jun 2004 to Dec 2014. Every June and December, we regress the historical market weekly returns on the adjusted changes in weekly search volume of a certain number of finance-related search terms and then sort the search terms according to the t-statistics of their coefficients. We use the top 30 search terms with largest negative (positive) t-statistics as our negative (positive) sentiment portfolio. Then for every week in the following 6 months, we calculate the t-value-weight average of the adjusted changes in weekly search volume of these 30 search terms and use this average as our negative (positive) sentiment. The total market sentiment measure is then constructed as positive sentiment minus negative sentiment. DEV, EMG and WRD denote developed, emerging and global markets, respectively.
Panel A: Emerging markets Country EMG/DEV Obs Mean STD Min Q1 Median Q3 Max Brazil EMG 548 0.006 0.322 -1.419 -0.210 0.004 0.209 1.174 Chile EMG 548 0.003 0.397 -1.913 -0.212 0.004 0.246 1.499 China EMG 548 0.000 0.419 -1.610 -0.241 -0.026 0.237 2.020 Colombia EMG 548 -0.006 0.391 -1.467 -0.221 -0.018 0.212 1.632 Egypt EMG 548 0.009 0.529 -3.066 -0.230 0.010 0.259 2.274 Hungary EMG 548 -0.001 0.768 -2.832 -0.430 0.014 0.397 5.653 India EMG 548 0.005 0.317 -0.896 -0.198 0.013 0.186 1.270 Indonesia EMG 548 -0.004 0.280 -1.147 -0.127 0.007 0.137 1.211 Israel EMG 548 -0.004 0.421 -1.386 -0.292 0.016 0.260 1.278 Malaysia EMG 548 0.000 0.331 -1.087 -0.190 0.002 0.188 1.292 Mexico EMG 548 0.006 0.377 -1.450 -0.242 0.010 0.251 1.168 Peru EMG 548 -0.006 0.430 -4.441 -0.245 0.015 0.237 1.592 Philippines EMG 548 -0.003 0.341 -1.045 -0.218 -0.012 0.207 1.162 Poland EMG 548 0.003 0.407 -1.254 -0.253 0.007 0.253 2.330 Portugal EMG 548 0.004 0.488 -1.676 -0.287 0.022 0.309 1.694 Russia EMG 548 0.005 0.463 -2.884 -0.187 -0.005 0.208 2.670 South Africa EMG 548 0.000 0.344 -1.205 -0.209 0.003 0.216 1.057 South Korea EMG 548 0.000 0.500 -2.252 -0.246 0.019 0.257 3.130 Taiwan EMG 548 0.004 0.425 -1.515 -0.239 0.006 0.261 1.486 Thailand EMG 548 0.004 0.449 -2.040 -0.270 0.000 0.256 1.750 Turkey EMG 548 0.007 0.369 -1.556 -0.200 0.016 0.230 1.142
34
Panel B: Developed markets Country EMG/DEV Obs Mean STD Min Q1 Median Q3 Max Australia DEV 548 -0.003 0.379 -1.633 -0.237 0.000 0.227 1.833 Austria DEV 548 0.004 0.646 -3.451 -0.306 0.012 0.312 5.624 Belgium DEV 548 -0.005 0.500 -3.638 -0.293 -0.006 0.297 3.190 Canada DEV 548 0.002 0.387 -1.539 -0.245 -0.004 0.261 1.323 Denmark DEV 548 -0.005 0.556 -2.510 -0.255 0.006 0.260 3.576 France DEV 548 -0.002 0.432 -1.933 -0.265 0.009 0.259 1.668 Germany DEV 548 -0.002 0.361 -1.146 -0.217 -0.003 0.240 1.075 Hong Kong DEV 548 0.000 0.583 -3.129 -0.293 -0.001 0.311 2.127 Italy DEV 548 0.000 0.402 -2.079 -0.222 0.000 0.241 2.019 Japan DEV 548 -0.003 0.340 -1.102 -0.222 0.009 0.221 1.554 Netherlands DEV 548 -0.001 0.424 -1.517 -0.275 0.004 0.256 1.336 New Zealand DEV 548 0.011 0.444 -1.825 -0.262 0.023 0.293 1.468 Norway DEV 548 -0.007 1.297 -5.751 -0.481 0.009 0.466 9.654 Singapore DEV 548 0.003 0.442 -1.823 -0.228 0.012 0.265 2.077 Spain DEV 548 0.007 0.392 -1.082 -0.246 -0.006 0.265 1.590 Sweden DEV 548 0.003 0.449 -1.892 -0.273 0.002 0.280 1.634 Switzerland DEV 548 0.000 0.463 -1.515 -0.274 0.016 0.300 1.318 United Kingdom DEV 548 0.003 0.371 -1.716 -0.248 0.012 0.244 1.644 United States DEV 548 0.003 0.344 -1.221 -0.211 0.005 0.231 1.016 Total WRD 21,920 0.001 0.478 -5.751 -0.244 0.004 0.251 9.654
35
Table 3 Weekly sentiment and returns This table reports the results of the following three regressions for each of our 40 countries over 2004 to 2014: a) Sentimentt+1=a+bSentimentt ; b) Returnt=a+bSentimentt ; c) Returnt+1=a+bSentimentt, where Sentimentt is defined in Table 2. Returnt is the weekly total market index return from Datastream. Panel A presents the regression results of each country and Panel B shows the summary statistics of the 40 regression results. Positive (Negative) Sig. is the number of coefficients on Sentimentt that are significantly different from zero at 5% level. DEV and EMG denote developed and emerging markets, respectively. Panel A: Regression results by country
Sentimentt+1=a+bSentimentt Returnt=a+bSentimentt Returnt+1=a+bSentimentt
Country EMG/DEV b T Stat R2 b T Stat R2 b T Stat R2
Brazil EMG -0.358 -8.96 12.9% 1.922 6.68 7.6% -0.835 -2.80 1.4%
Chile EMG -0.348 -8.64 12.1% 1.561 10.53 16.9% -0.580 -3.60 2.3%
China EMG -0.295 -7.21 8.7% 1.596 9.02 13.0% -0.126 -0.66 0.1%
Colombia EMG -0.095 -2.22 0.9% 1.146 6.14 6.5% -0.119 -0.62 0.1%
Egypt EMG -0.180 -4.26 3.2% 1.465 10.12 15.8% -0.263 -1.66 0.5%
Hungary EMG -0.160 -3.79 2.6% 1.052 7.74 9.9% -0.173 -1.21 0.3%
India EMG -0.308 -7.55 9.5% 2.425 8.83 12.5% -0.107 -0.36 0.0%
Indonesia EMG -0.304 -7.43 9.2% 1.580 5.04 4.5% -0.629 -1.98 0.7%
Israel EMG -0.351 -8.75 12.3% 1.410 10.62 17.1% -0.231 -1.59 0.5%
Malaysia EMG -0.333 -8.26 11.1% 1.367 9.95 15.3% -0.523 -3.55 2.3%
Mexico EMG -0.330 -8.17 10.9% 1.654 8.75 12.3% -0.339 -1.68 0.5%
Peru EMG -0.402 -10.22 16.1% 1.232 9.04 13.0% -0.656 -4.56 3.7%
Philippines EMG -0.258 -6.22 6.6% 1.493 7.64 9.6% -0.396 -1.93 0.7%
Poland EMG -0.324 -7.99 10.5% 2.464 11.54 19.6% -0.988 -4.21 3.2%
Portugal EMG -0.301 -7.37 9.1% 1.745 12.79 23.0% -0.410 -2.65 1.3%
Russia EMG -0.361 -9.00 12.9% 0.693 2.82 1.4% -0.283 -1.14 0.2%
South Africa EMG -0.340 -8.44 11.6% 2.651 11.24 18.8% -0.775 -2.98 1.6%
South Korea EMG -0.393 -9.94 15.3% 1.708 9.46 14.1% -0.444 -2.28 0.9%
Taiwan EMG -0.368 -9.22 13.5% 1.479 9.77 14.9% -0.445 -2.72 1.3%
Thailand EMG -0.366 -9.18 13.4% 1.816 7.95 10.4% -1.061 -4.47 3.5%
Turkey EMG -0.212 -5.07 4.5% 2.785 9.87 15.1% -0.017 -0.06 0.0%
Australia DEV -0.351 -8.71 12.2% 2.166 11.53 19.6% -0.682 -3.27 1.9%
Austria DEV -0.352 -8.79 12.4% 1.511 13.45 24.9% -0.597 -4.70 3.9%
Belgium DEV -0.404 -10.32 16.3% 1.670 13.61 25.3% -0.586 -4.19 3.1%
Canada DEV -0.422 -10.86 17.8% 1.551 9.22 13.5% -0.456 -2.53 1.2%
Denmark DEV -0.358 -8.97 12.9% 1.581 14.07 26.6% -0.402 -3.09 1.7%
France DEV -0.346 -8.61 12.0% 1.656 10.97 18.1% -0.530 -3.20 1.8%
Germany DEV -0.353 -8.80 12.4% 1.814 10.06 15.6% -0.701 -3.61 2.3%
36
Panel A Continued
Country
Hong Kong DEV -0.470 -12.43 22.1% 1.379 13.93 26.2% -0.435 -3.82 2.6%
Italy DEV -0.307 -7.53 9.4% 1.944 10.65 17.2% -0.679 -3.42 2.1%
Japan DEV -0.360 -9.03 13.0% 1.487 9.64 14.6% -0.415 -2.50 1.1%
Netherlands DEV -0.437 -11.32 19.0% 1.783 11.31 19.0% -0.587 -3.38 2.0%
New Zealand DEV -0.393 -9.99 15.5% 1.264 10.66 17.2% -0.447 -3.47 2.2%
Norway DEV -0.504 -13.63 25.4% 0.595 8.30 11.2% -0.279 -3.71 2.5%
Singapore DEV -0.358 -8.99 12.9% 1.389 10.96 18.0% -0.255 -1.82 0.6%
Spain DEV -0.347 -8.65 12.1% 1.987 10.65 17.2% -0.447 -2.19 0.9%
Sweden DEV -0.401 -10.22 16.1% 2.028 12.20 21.4% -0.836 -4.53 3.6%
Switzerland DEV -0.370 -9.29 13.7% 1.271 12.46 22.1% -0.363 -3.17 1.8%
United Kingdom DEV -0.365 -9.14 13.3% 1.352 8.00 10.5% -0.506 -2.85 1.5%
United States DEV -0.432 -11.17 18.6% 1.184 8.31 11.2% -0.456 -3.04 1.7%
Panel B: Summary statistics of the coeficients on sentimentt in Panel A
Sentiment t+1=a+bSentiment t Returnt=a+bSentimentt Returnt+1=a+bSentiment t
EMG/DEV b T Stat Negative Sig. R2 b T Stat Positive Sig. R2 b T Stat Negative Sig. R2
EMG -0.304 -7.52 21 out of 21 9.8% 1.678 8.83 21 out of 21 12.9% -0.448 -2.22 11 out of 21 1.2%
DEV -0.386 -9.81 19 out of 19 15.1% 1.558 11.05 19 out of 19 18.4% -0.508 -3.29 18 out of 19 2.0%
WRD -0.343 -8.61 40 out of 40 12.3% 1.621 9.89 40 out of 40 15.5% -0.476 -2.73 29 out of 40 1.6%
37
Table 4 Robustness tests on weekly return and sentiment This table presents the summary statistics of regressions of market weekly return on market sentiment and regressions of next week market sentiment on current week sentiment based on four different specifications. Panel A to Panel C reports the results of the following regression:
a) Sentimentt+1=a+bSentimentt ; b) Returnt=a+bSentimentt ; c) Returnt+1=a+bSentimentt where Sentimentt is defined in Table 2. Returnt is the weekly total market index return from datastream. We run these regressions by each country from 2004 to 2014 and report the averages of the coefficients (b) and t-statistics (T Stat) on Sentimentt and R2 across the 40 countries. Positive (Negative) Sig. is the number of coefficients on Sentimentt that are significantly different from zero at 5% level. In Panel A, we remove the financial crisis period (Sep.08 to Sep.09) and run the above regressions. Panel B presents the results where we use local currency instead of U.S. dollar for total market return index in Datastream to construct market weekly return. In Panel C, we use two-week return as the market return and the simple average of two-week sentiment as the market sentiment to run the above regressions. In Panel D, we add seven control variables to the above a) to c) regressions. EPU from Baker, Bloom and Davis (2013) is used to capture economic policy uncertainty and we take the simple average of EPU in a week as weekly EPU. ADS is a daily measure of macroeconomic activities obtained from Aruoba, Diebold, and Scotti (2009). VIX is the Chicago Board Options Exchange daily market volatility index and we use the simple average of VIX within a week as weekly VIX. Returnt-1 to Returnt-5 represents the lagged one to lagged five weekly returns. DEV and EMG denote developed and emerging markets, respectively. Panel A: Removing financial crisis period
Sentiment t+1=a+bSentiment t Returnt=a+bSentimentt Returnt+1=a+bSentiment t
EMG/DEV b T Stat Positive Sig. R2 b T Stat Positive Sig. R2 b T Stat Negative Sig. R2
EMG -0.294 -6.90 21 out of 21 9.4% 1.463 8.14 21 out of 21 12.1% -0.353 -1.89 10 out of 21 1.1%
DEV -0.382 -9.24 19 out of 19 14.9% 1.256 9.16 19 out of 19 14.7% -0.405 -2.79 16 out of 19 1.7%
WRD -0.336 -8.01 40 out of 40 12.0% 1.365 8.62 40 out of 40 13.3% -0.378 -2.32 26 out of 40 1.4%
Panel B: Local currency
Sentiment t+1=a+bSentiment t Returnt=a+bSentimentt Returnt+1=a+bSentiment t
EMG/DEV b T Stat Positive Sig. R2 b T Stat Positive Sig. R2 b T Stat Negative Sig. R2
EMG -0.304 -7.34 21 out of 21 9.8% 1.241 7.99 20 out of 21 11.5% -0.315 -1.94 8 out of 21 1.1%
DEV -0.385 -9.56 19 out of 19 15.1% 1.204 9.79 19 out of 19 15.7% -0.357 -2.70 14 out of 19 1.5%
WRD -0.343 -8.40 40 out of 40 12.3% 1.224 8.85 39 out of 40 13.5% -0.335 -2.30 22 out of 40 1.3%
Panel C: Two-week return
Sentiment t+1=a+bSentiment t Returnt=a+bSentimentt Returnt+1=a+bSentiment t
EMG/DEV b T Stat Positive Sig. R2 b T Stat Positive Sig. R2 b T Stat Negative Sig. R2
EMG -0.346 -6.15 21 out of 21 12.7% 1.807 5.61 21 out of 21 10.8% -0.449 -1.38 7 out of 21 1.1%
DEV -0.356 -6.34 18 out of 19 13.4% 1.535 6.07 19 out of 19 12.1% -0.360 -1.30 5 out of 19 1.2%
WRD -0.351 -6.24 39 out of 40 13.0% 1.678 5.83 40 out of 40 11.4% -0.407 -1.34 12 out of 40 1.1%
38
Panel D: Add more controls (EPU,VIX ADS and return_t-1 to return_t-5)
Sentiment t+1=a+bSentiment t+Controlst
Returnt=a+bSentimentt+Controlst
Returnt+1=a+bSentiment t+Controlst
EMG/DEV
b
T Stat
Positive Sig.
R2
b
T Stat
Positive Sig.
R2
b
T Stat
Negative Sig.
R2
EMG
-0.312
-6.93
20 out of 21
11.7%
1.703
8.62
21 out of 21
18.0%
-0.494
-2.20
11 out of 21
4.4%
DEV
-0.412
-9.10
19 out of 19
17.3%
1.529
10.64
19 out of 19
25.0%
-0.563
-3.17
18 out of 19
5.3%
WRD
-0.360
-7.96
39 out of 40
14.4%
1.621
9.58
40 out of 40
21.3%
-0.527
-2.66
29 out of 40
4.8%
39
Table 5 Change in indices of exchange trading rules This table presents changes in seven indices of exchange trading rules (Cumming, Johan and Li (2011)) across 35 countries after a directive on markets in financial instruments (MiFID) became effective for European countries on 1 November 2007. The definitions of the seven indices are in Appendix A. The sample period is January 2007 to September 2008 and the event date is 1 November 2007. Country Sample Period Event Date ∆PMI ∆VMI ∆SI ∆FDI ∆MMI ∆ITI ∆BAI
Austria Jan.07-Sep.08 1Nov.07 6 1 3 1 11 2 0
Denmark Jan.07-Sep.08 1Nov.07 5 0 1 0 6 1 0
France Jan.07-Sep.08 1Nov.07 4 1 2 1 8 2 0
Germany Jan.07-Sep.08 1Nov.07 7 1 2 1 11 1 -1
Hungary Jan.07-Sep.08 1Nov.07 N.A. N.A. N.A. N.A. N.A. N.A. N.A.
Italy Jan.07-Sep.08 1Nov.07 7 1 2 0 10 2 0
Netherlands Jan.07-Sep.08 1Nov.07 N.A. N.A. N.A. N.A. N.A. N.A. N.A.
Norway Jan.07-Sep.08 1Nov.07 5 0 2 1 8 1 0
Poland Jan.07-Sep.08 1Nov.07 N.A. N.A. N.A. N.A. N.A. N.A. N.A.
Portugal Jan.07-Sep.08 1Nov.07 N.A. N.A. N.A. N.A. N.A. N.A. N.A.
Spain Jan.07-Sep.08 1Nov.07 7 1 2 0 10 0 0
Sweden Jan.07-Sep.08 1Nov.07 5 0 1 0 6 1 0
United Kingdom Jan.07-Sep.08 1Nov.07 1 0 0 0 1 1 0
Australia Jan.07-Sep.08 N.A. 0 0 0 0 0 0 0
Brazil Jan.07-Sep.08 N.A. 0 0 0 0 0 0 0
Canada Jan.07-Sep.08 N.A. 0 0 0 0 0 0 0
Chile Jan.07-Sep.08 N.A. 0 0 0 0 0 0 0
Colombia Jan.07-Sep.08 N.A. 0 0 0 0 0 0 0
Egypt Jan.07-Sep.08 N.A. 0 0 0 0 0 0 0
Hong Kong Jan.07-Sep.08 N.A. 0 0 0 0 0 0 0
India Jan.07-Sep.08 N.A. 0 0 0 0 0 0 0
Indonesia Jan.07-Sep.08 N.A. 0 0 0 0 0 0 0
Israel Jan.07-Sep.08 N.A. 0 0 0 0 0 0 0
Japan Jan.07-Sep.08 N.A. 0 0 0 0 0 0 0
Malaysia Jan.07-Sep.08 N.A. 0 0 0 0 0 0 0
Mexico Jan.07-Sep.08 N.A. 0 0 0 0 0 0 0
New Zealand Jan.07-Sep.08 N.A. 0 0 0 0 0 0 0
Peru Jan.07-Sep.08 N.A. 0 0 0 0 0 0 0
Singapore Jan.07-Sep.08 N.A. 0 0 0 0 0 0 0
South Africa Jan.07-Sep.08 N.A. 0 0 0 0 0 0 0
South Korea Jan.07-Sep.08 N.A. 0 0 0 0 0 0 0
Taiwan Jan.07-Sep.08 N.A. 0 0 0 0 0 0 0
Thailand Jan.07-Sep.08 N.A. 0 0 0 0 0 0 0
Turkey Jan.07-Sep.08 N.A. 0 0 0 0 0 0 0
United States Jan.07-Sep.08 N.A. 0 0 0 0 0 0 0
40
Table 6 Sentiment and information environment This table reports the results of the following panel regression:
Returni,t+1=α+β1Sentimenti,t+ β2EUi,t+ β3Sentimenti,t* EUi,t+ε i,t where Returni,t+1 is country i’s market return in week t+1 and Sentimenti,t is country i’s market sentiment (as defined in Table 2) in week t. In Model (2), EUi,t equals to 1 if a country i is exposed to MiFID at time t and 0 otherwise. In Model (3) to Model (9), EUi,t equals to the change in the corresponding index of exchange trading rules if the date is after 1 November 2007 and 0 otherwise. Panel A shows the main results based on 35 countries and the sample period is Jan 2007 to Sep 2008. In Panel B, we match 13 countries that are not impacted by MiFID with the 13 European countries that are subject to MiFID using the 11-year (2004 to 2014) average of GDP per Capita and repeat the regressions in Panel A with the 26 countries. Panel C reports the results of regressions using a longer sample period: Nov.2006 to Sep.2008. In Panel D, we conduct placebo tests. Specifically, we bootstrap 13 event dates between May. 2005 and Feb. 2014 and assign them to the 13 European countries to run the regressions in Panel A. We repeat the procedures for 1000 times. Panel D shows the 90th percentile, 95th percentile and 99th percentile of the 1000 coefficients on the interaction term. We include country fixed effect in all the regressions. The standard errors are adjusted for heteroskedasticity and clustered by country. Panel A: Main results Baseline MiFID ∆PMI ∆VMI ∆SI ∆FDI ∆MMI ∆ITI ∆BAI Variable Model Model Model Model Model Model Model Model Model (1) (2) (3) (4) (5) (6) (7) (8) (9) Sentiment -0.493 -0.590 -0.617 -0.546 -0.620 -0.595 -0.616 -0.609 -0.523 (-7.48) (-7.08) (-6.87) (-6.02) (-6.91) (-6.76) (-6.89) (-6.86) (-5.98) Sentiment×Event 0.364 0.093 0.152 0.279 0.627 0.060 0.347 0.684 (1.90) (1.89) (0.39) (2.33) (1.77) (1.90) (1.81) (7.81) Event -0.760 -0.124 -0.670 -0.354 -0.762 -0.080 -0.449 0.685 (-20.60) (-9.51) (-25.29) (-6.79) (-8.14) (-8.77) (-6.22) (779.89) Obs 3,520 3,520 3,168 3,168 3,168 3,168 3,168 3,168 3,168 AdjRsq 1.8% 3.6% 3.3% 2.6% 3.3% 2.9% 3.3% 3.1% 2.3% Panel B: matched sample Baseline MiFID ∆PMI ∆VMI ∆SI ∆FDI ∆MMI ∆ITI ∆BAI Variable Model Model Model Model Model Model Model Model Model (1) (2) (3) (4) (5) (6) (7) (8) (9) Sentiment -0.467 -0.645 -0.714 -0.556 -0.718 -0.657 -0.712 -0.694 -0.512 (-5.77) (-7.74) (-9.27) (-4.44) (-8.81) (-6.52) (-9.18) (-8.08) (-4.29) Sentiment×Event 0.419 0.111 0.162 0.325 0.689 0.071 0.402 0.695 (2.20) (2.30) (0.40) (2.90) (1.94) (2.32) (2.11) (5.82) Event -0.760 -0.124 -0.671 -0.354 -0.762 -0.0801 -0.450 0.685 (-20.46) (-9.42) (-25.05) (-6.76) (-8.07) (-8.70) (-6.17) (571.98) Obs 2,288 2,288 1,936 1,936 1,936 1,936 1,936 1,936 1,936 AdjRsq 1.6% 4.5% 4.1% 2.9% 4.1% 3.4% 4.1% 3.8% 2.2%
41
Panel C: Different sample period Baseline MiFID ∆PMI ∆VMI ∆SI ∆FDI ∆MMI ∆ITI ∆BAI Variable Model Model Model Model Model Model Model Model Model (1) (2) (3) (4) (5) (6) (7) (8) (9) Sentiment -0.481 -0.575 -0.604 -0.537 -0.607 -0.583 -0.604 -0.597 -0.515 (-7.58) (-6.88) (-6.82) (-5.97) (-6.84) (-6.70) (-6.83) (-6.76) (-5.99) Sentiment×Event 0.349 0.091 0.143 0.274 0.615 0.058 0.339 0.692 (1.83) (1.86) (0.36) (2.29) (1.74) (1.86) (1.77) (8.04) Event -0.764 -0.125 -0.686 -0.358 -0.763 -0.081 -0.458 0.701 (-23.89) (-10.39) (-36.13) (-7.48) (-10.33) (-9.70) (-6.62) (871.09) Obs 3,880 3,880 3,492 3,492 3,492 3,492 3,492 3,492 3,492 AdjRsq 1.7% 3.5% 3.2% 2.6% 3.2% 2.8% 3.2% 3.1% 2.2% Panel D: Placebo tests based on bootstrapping MiFID ∆PMI ∆VMI ∆SI ∆FDI ∆MMI ∆ITI ∆BAI Ours 0.364 0.093 0.152 0.279 0.627 0.060 0.347 0.684 P90 -0.001 -0.032 -0.054 -0.086 -0.137 -0.020 -0.132 0.494 P95 0.069 -0.023 0.019 -0.053 -0.046 -0.014 -0.100 0.521 P99 0.198 -0.007 0.150 -0.002 0.051 -0.003 -0.044 0.922
42
Table 7 Summary of short-selling ban This table summarizes the details of short-selling bans for 35 countries during the financial crisis period. The information is from Beber and Pagano (2013). Country Sample Period Ban Begin Date Ban End Date Scope of Short-selling Ban Australia Mar.08-Apr.09 22Sep.08 25May.09 All Stocks Austria Mar.08-Apr.09 26Oct.08 30Nov.10 Financial Stocks Canada Mar.08-Apr.09 19Sep.08 8Oct.08 Financial Stocks Denmark Mar.08-Apr.09 13Oct.08 After 31Dec.10 Financial Stocks France Mar.08-Apr.09 22Sep.08 After 31Dec.10 Financial Stocks Germany Mar.08-Apr.09 20Sep.08 After 31Dec.10 Financial Stocks Indonesia Mar.08-Apr.09 1Oct.08 1May.09 All Stocks Italy Mar.08-Apr.09 22Sep.08 1Jun.09 Financial, then All Japan Mar.08-Apr.09 30Oct.08 After 31Dec.10 All Stocks Netherlands Mar.08-Apr.09 22Sep.08 1Jun.09 Financial Stocks Norway Mar.08-Apr.09 8Oct.08 After 31Dec.10 Financial Stocks Portugal Mar.08-Apr.09 22Sep.08 After 31Dec.10 Financial Stocks South Korea Mar.08-Apr.09 1Oct.08 After 31Dec.10 All Stocks Spain Mar.08-Apr.09 24Sep.08 After 31Dec.10 All Stocks Taiwan Mar.08-Apr.09 1Oct.08 28Nov.08 All Stocks United Kingdom Mar.08-Apr.09 19Sep.08 16Jan.09 Financial Stocks United States Mar.08-Apr.09 19Sep.08 8Oct.08 Financial Stocks Brazil Mar.08-Apr.09 N.A. N.A. No Ban Chile Mar.08-Apr.09 N.A. N.A. No Ban Colombia Mar.08-Apr.09 N.A. N.A. Always Ban Egypt Mar.08-Apr.09 N.A. N.A. Always Ban Hong Kong Mar.08-Apr.09 N.A. N.A. No Ban Hungary Mar.08-Apr.09 N.A. N.A. No Ban India Mar.08-Apr.09 N.A. N.A. No Ban Israel Mar.08-Apr.09 N.A. N.A. No Ban Malaysia Mar.08-Apr.09 N.A. N.A. No Ban Mexico Mar.08-Apr.09 N.A. N.A. No Ban New Zealand Mar.08-Apr.09 N.A. N.A. No Ban Peru Mar.08-Apr.09 N.A. N.A. Always Ban Poland Mar.08-Apr.09 N.A. N.A. No Ban Singapore Mar.08-Apr.09 N.A. N.A. No Ban South Africa Mar.08-Apr.09 N.A. N.A. No Ban Sweden Mar.08-Apr.09 N.A. N.A. No Ban Thailand Mar.08-Apr.09 N.A. N.A. No Ban Turkey Mar.08-Apr.09 N.A. N.A. No Ban
43
Table 8 Sentiment and short-selling ban This table reports the results of the following panel regressions:
Returni,t+1=α+β1Sentimenti,t+ β2Bani,t+ β3Sentimenti,t* Bani,t+ε i,t where Returni,t+1 is country i’s market return in week t+1 and Sentimenti,t is country i’s market sentiment (as defined in Table 2) in week t. Bani,t equals to 1 if a country i has imposed a short-selling ban at week t and 0 otherwise. Panel A presents the main results where we use the 35 countries listed in Table 7 over the sample period of Mar.2008 to Apr.2009. In Panel B, we report the results using different sample periods and countries. Model (1) uses a sample period of Jan. 2008 to Jun. 2009 and Model (2) uses the countries in Beber and Pagano (2013). In Model (3), using the 11-year (2004 to 2014) average of GDP per Capita, we match 15 countries that have never imposed a ban during the sample period with 15 treated countries. Panel C reports the results of placebo tests. Specifically, we first bootstrap a 40-day window between Jan. 2005 and Apr. 2014. Then within the 40-day window, we randomly select (with replacement) 17 event dates as the pseudo starting date of short-selling ban and allocate them to the 17 countries that have imposed a short-selling ban. The durations of each short-selling ban remains the same as those in Table 7. We run the above regressions using the newly bootstrapped event dates and keep β3. We repeat the procedure for 1000 times and report the 90th, 95th and 99th percentile of β3 in Panel C. We include country fixed effect in all the regressions. The standard errors are adjusted for heteroskedasticity and clustered by country. Panel A: Main Results Variable Model Model (1) (2) Sentiment -0.718 -0.529 (-5.85) (-3.96) Sentiment×Short-selling Ban -0.572 (-2.30) Short-selling Ban 0.291 (1.21) Observations 2,280 2,280 R-squared 2.3% 2.9% Panel B: Different periods and samples Variable Model Model Model (1) (2) (3) Sentiment -0.558 -0.459 -0.581 (-4.61) (-2.80) (-3.72) Sentiment×Short-selling Ban -0.468 -0.623 -0.700 (-2.22) (-1.92) (-2.30) Short-selling Ban 0.410 0.251 0.118 (2.02) (0.99) (0.44) Obs 3,040 1,254 1,995 AdjRsq 2.6% 3.3% 2.4%
44
Panel C: Placebo tests based on bootstrapping
Distribution Ours -0.572 P90 -0.308 P95 -0.463 P99 -0.761
45
Table 9 Global sentiment and total sentiment This table reports the regression of total sentiment on global sentiment for each of the 35 countries. Specifically, we run the following regression:
Sentimentt=a+bSentimentG,t+ε t where Sentimenti,t is country i’s market sentiment (as defined in Table 2) in week t and SentimentG,t is the simple average of Sentimenti,t across the 35 countries in week t. The sample period is Jul. 2004 to Dec. 2014. Panel A reports the coefficients and t-statistics of SentimentG,t and R2 for each country. Panel B shows the simple average of these statistics across the 35 countries. Positive Sig. is the number of coefficients on SentimentG,t that are significantly different from zero at 5% level. DEV, EMG and WRD denote developed, emerging and global markets, respectively. We include year fixed effect and month fixed effect in all the regressions. The t-statistics are based on standard errors adjusted for heteroscedasticity. Panel A:Regressions by country Sentimentt=a+bSentimentG,t
Country EMG/DEV b T Stat R2 Brazil EMG 0.408 4.57 7.1% Chile EMG 0.590 4.94 8.5% Colombia EMG 0.530 5.83 7.0% Egypt EMG 1.063 5.61 14.9% Hungary EMG 1.176 5.72 9.2% India EMG 0.478 6.75 9.7% Indonesia EMG 0.264 4.67 5.5% Israel EMG 0.846 8.51 14.3% Malaysia EMG 0.601 7.15 12.4% Mexico EMG 0.724 8.65 13.3% Peru EMG 0.634 6.21 8.1% Poland EMG 0.993 7.54 20.8% Portugal EMG 1.255 11.37 23.2% South Africa EMG 0.806 9.38 20.4% South Korea EMG 0.788 5.61 8.7% Taiwan EMG 0.683 7.62 9.2% Thailand EMG 0.566 6.93 10.8% Turkey EMG 0.734 7.68 14.1% Australia DEV 0.894 8.16 19.7% Austria DEV 2.013 7.07 33.4% Canada DEV 0.810 8.67 15.3% Denmark DEV 1.596 9.71 28.5% France DEV 1.097 10.49 22.3% Germany DEV 0.892 9.11 21.2% Hong Kong DEV 1.677 10.65 28.5% Italy DEV 1.076 9.04 24.9%
46
Panel A Continue Country EMG/DEV b T Stat R2 Japan DEV 0.624 6.74 12.3% Netherlands DEV 1.194 13.92 27.3% New Zealand DEV 0.898 9.62 14.5% Norway DEV 2.808 8.45 16.5% Singapore DEV 1.251 12.52 27.8% Spain DEV 0.881 11.11 17.7% Sweden DEV 1.083 10.78 20.4% United Kingdom DEV 0.871 9.92 19.3% United States DEV 0.581 6.28 10.1% Panel B: Summary statistics of the coefficients on sentimentG,t
Sentimentt=a+bSentimentG,t
EMG/DEV b T Stat Positive Sig. R2 EMG 0.730 6.93 18 out of 18 12.1% DEV 1.191 9.54 17 out of 17 21.2% WRD 0.954 8.20 35 out of 35 16.5%
47
Table 10 Summary of IFRS adoption This table presents the IFRS adoption years for each of the 35 countries over our sample period of 2004 to 2014. Country Sample Period IFRS Adoption Year Australia Jun.04-Dec.14 2005 Austria Jun.04-Dec.14 2005 Brazil Jun.04-Dec.14 2010 Canada Jun.04-Dec.14 2011 Chile Jun.04-Dec.14 2009 Denmark Jun.04-Dec.14 2005 France Jun.04-Dec.14 2005 Germany Jun.04-Dec.14 2005 Hong Kong Jun.04-Dec.14 2005 Hungary Jun.04-Dec.14 2005 Israel Jun.04-Dec.14 2008 Italy Jun.04-Dec.14 2005 Malaysia Jun.04-Dec.14 2012 Mexico Jun.04-Dec.14 2012 Netherlands Jun.04-Dec.14 2005 New Zealand Jun.04-Dec.14 2007 Norway Jun.04-Dec.14 2005 Peru Jun.04-Dec.14 2012 Poland Jun.04-Dec.14 2005 Portugal Jun.04-Dec.14 2005 Singapore Jun.04-Dec.14 2003 South Africa Jun.04-Dec.14 2005 South Korea Jun.04-Dec.14 2011 Spain Jun.04-Dec.14 2005 Sweden Jun.04-Dec.14 2005 Taiwan Jun.04-Dec.14 2013 Turkey Jun.04-Dec.14 2008 United Kingdom Jun.04-Dec.14 2005 Colombia Jun.04-Dec.14 N.A. Egypt Jun.04-Dec.14 N.A. India Jun.04-Dec.14 N.A. Indonesia Jun.04-Dec.14 N.A. Japan Jun.04-Dec.14 N.A. Thailand Jun.04-Dec.14 N.A. United States Jun.04-Dec.14 N.A.
48
Table 11 IFRS and the relation between total sentiment and global sentiment This table reports the panel regressions of total sentiment on global sentiment, IFRS adoption dummy and the interaction between global sentiment and the IFRS dummy. The regression is as follows:
Sentimenti,t=α+β1SentimentG,t+ β2SentimentG,t*IFRSi,t+ β3IFRSi,t+ε i,t where Sentimenti,t is country i’s market sentiment (as defined in Table 2) in week t and SentimentG,t is the simple average of Sentimenti,t across the 35 countries in week t. IFRSi,t equals to 1 when a country adopted IFRS at time t-1 and 0 otherwise. Panel A reports the results where we use 35 countries over Jul. 2004 to Dec. 2014. Panel B presents results of different samples and periods. Model (1) of Panel B uses the sample period of 2004-2007 where 2004-2005 are the pre-IFRS adoption periods and 2006-2007 are the post-IFRS adoption periods. In Model (2), we use the 11-year (2004 to 2014) average of GDP per Capita to match 7 countries that have never adopted IFRS during the sample period with 7 countries that have. Panel C reports the results of placebo tests. Specifically, we bootstrap 28 IFRS adoption years from 2005 to 2014 and allocate them to the 28 countries that have adopted IFRS during the sample period of 2005-2014. We then run the above regressions and keep β2. We repeat the procedure for 1000 times and report the 90th, 95th and 99th percentile of β2 in Panel C. We include country fixed effect in all the regressions. The standard errors are adjusted for heteroskedasticity and clustered by country. Panel A: Main Results Variable Model Model (1) (2)
SentimentG 1.000 0.845 (41.82) (9.81)
SentimentG×IFRS 0.345 (3.37) IFRS -0.003 (-0.68) Obs 21,920 21,920 AdjRsq 13.3% 13.7% Panel B: Different samples and periods Variable Model Model (1) (2)
SentimentG 0.893 0.589 (6.24) (10.24)
SentimentG×IFRS 0.365 0.355 (2.72) (3.08) IFRS -0.001 -0.001 (-0.12) (-0.18) Observations 7,320 7,672 R-squared 11.4% 11.0%
49
Panel C: Placebo tests based on bootstrapping
β2 Distribution Ours 0.345 P90 0.170 P95 0.211 P99 0.292
50
Table 12 Return and global sentiment This table reports the regressions of next week return on global sentiment and local sentiment for each of the 35 countries. We run the following regression:
Returnt+1=a+ bSentimentG,t+ cSentimentL,t+ε t where global sentiment (SentimentG,t) is the simple average of total market sentiment (Sentimenti,t defined in Table 2) of the 35 countries in week t. Local sentiment (SentimentL,t) of a country is the residual of regressing the country’s total market sentiment (Sentimenti,t) on global sentiment (SentimentG,t). The sample period is Jul. 2004-Dec. 2014. Panel A reports the results of the regressions for each country and Panel B presents the simple averages of coefficients and t-statistics of the 35 countries in Panel A. DEV, EMG and WRD denote developed, emerging and global markets, respectively. We include country fixed effect in all the regressions. The standard errors are adjusted for heteroskedasticity and clustered by country.
Panel A: Regressions by country
Returnt+1=a+bSentimentG,t+cSentimentL,t
Country EMG/DEV b T Stat (b) c T Stat (c) R2 Brazil EMG -3.010 -5.00 -0.452 -1.70 6.5% Chile EMG -1.308 -3.77 -0.444 -2.28 3.8% Colombia EMG -1.622 -3.33 0.078 0.45 2.9% Egypt EMG 0.717 1.26 -0.415 -1.97 1.5% Hungary EMG -2.792 -2.65 0.020 0.12 4.0% India EMG -0.754 -1.15 0.019 0.07 0.4% Indonesia EMG -0.338 -0.46 -0.610 -1.80 0.7% Israel EMG -1.123 -2.57 -0.056 -0.31 2.1% Malaysia EMG -0.228 -0.65 -0.542 -3.35 2.3% Mexico EMG -2.328 -5.13 0.074 0.38 5.9% Peru EMG -1.058 -2.46 -0.575 -2.68 4.4% Poland EMG -2.895 -3.78 -0.498 -1.73 6.2% Portugal EMG -2.224 -5.03 -0.012 -0.07 5.4% South Africa EMG -3.274 -5.70 -0.013 -0.04 8.3% South Korea EMG -1.750 -2.05 -0.278 -1.53 2.4% Taiwan EMG -0.827 -1.75 -0.370 -2.17 1.7% Thailand EMG -0.709 -1.34 -1.041 -3.31 3.6% Turkey EMG -2.204 -2.57 0.451 1.41 2.7% Australia DEV -2.040 -3.16 -0.305 -0.96 4.4% Austria DEV -2.593 -4.39 -0.255 -1.09 6.5% Canada DEV -2.407 -3.89 -0.015 -0.08 7.4% Denmark DEV -2.381 -4.41 0.025 0.09 6.7% France DEV -2.655 -5.24 0.005 0.02 8.5% Germany DEV -2.598 -5.31 -0.119 -0.44 8.5% Hong Kong DEV -0.939 -2.06 -0.386 -2.16 2.7% Italy DEV -2.489 -4.90 -0.147 -0.52 6.0%
51
Panel A Continue Country EMG/DEV b T Stat (b) c T Stat (c) R2 Japan DEV -1.053 -2.98 -0.251 -1.27 2.5% Netherlands DEV -2.422 -4.16 -0.051 -0.28 6.7% New Zealand DEV -1.301 -2.97 -0.284 -2.24 3.9% Norway DEV -3.359 -3.91 -0.104 -1.31 7.5% Singapore DEV -0.965 -2.25 -0.060 -0.32 1.5% Spain DEV -2.367 -4.11 0.018 0.07 5.4% Sweden DEV -3.108 -5.23 -0.334 -1.54 9.0% United Kingdom DEV -2.315 -4.19 -0.011 -0.05 7.6% United States DEV -1.641 -3.56 -0.203 -0.99 6.5% Panel B: Summary statistics of the coefficients in Panel A
Returnt+1=a+bSentimentG,t+cSentimentL,t
EMG/DEV b T Stat (b) Negative Sig. c T Stat (c) Negative Sig. R2 EMG -1.540 -2.67 13 out of 18 -0.259 -1.14 9 out of 18 3.6% DEV -2.155 -3.92 17 out of 17 -0.146 -0.77 2 out of 17 6.0% WRD -1.839 -3.28 30 out of 35 -0.204 -0.96 11 out of 35 4.7%