10.1016 j.jebo.2014.06.004 facebook s daily sentiment and international stock markets

G ModelJ

F

Aa

b

c

a

ARRAA

KBSF

1

Miaamets

ltmp

(

e

bd

0

ARTICLE IN PRESSEBO-3383; No. of Pages 14

Journal of Economic Behavior & Organization xxx (2014) xxx–xxx

Contents lists available at ScienceDirect

Journal of Economic Behavior & Organization

j ourna l h om epa ge: w ww.elsev ier .com/ locate / jebo

acebook’s daily sentiment and international stock markets

ntonios Siganosa, Evangelos Vagenas-Nanosa, Patrick Verwijmerena,b,c,∗

University of Glasgow, Adam Smith Business School, Glasgow, G12 8QQ, Scotland, United KingdomErasmus University Rotterdam, Burgemeester Oudlaan 50, 3000DR Rotterdam, NetherlandsUniversity of Melbourne, 198 Berkeley Street, Victoria 3010, Australia

r t i c l e i n f o

rticle history:eceived 25 February 2013eceived in revised form 30 April 2014ccepted 10 June 2014vailable online xxx

eywords:ehavioral financeentimentacebook’s gross national happiness index

a b s t r a c t

We examine the relation between daily sentiment and trading behavior within 20 inter-national markets by exploiting Facebook’s Gross National Happiness Index. We find thatsentiment has a positive contemporaneous relation to stock returns. Moreover, sentimenton Sunday affects stock returns on Monday, suggesting causality from sentiment to stockmarkets. We observe that the relation between sentiment and returns reverses the follow-ing weeks. We further show that negative sentiments are related to increases in tradingvolume and return volatility. These results highlight the importance of behavioral factorsin stock investing.

© 2014 Elsevier B.V. All rights reserved.

. Introduction

An important part of behavioral finance concerns the relation between investor sentiment and stock market returns.easuring sentiment is, however, not a trivial exercise. The conventional method to obtain measurements of sentiment

s to take surveys of households. In this type of study, researchers typically select a random number of households andsk a small number of questions to identify the level of optimism or pessimism per household. The responses are thenggregated to construct an average sentiment level.1 Although these studies have provided important insights, the surveyethod has some important weaknesses. One weakness is that sample sizes and participation rates are typically low. For

xample, the Michigan Consumer Sentiment survey is sent to only 500 households, and the Consumer Confidence Indexo 5000 households. Another weakness is that the surveys are typically conducted on a monthly frequency. The resultanttudies then typically rely on the assumption that sentiment remains stable from day to day over the survey period.2

We propose to use an alternative measure of sentiment, based on status updates on Facebook, which is the world’s

Please cite this article in press as: Siganos, A., et al., Facebook’s daily sentiment and international stock markets. J. Econ.Behav. Organ. (2014), http://dx.doi.org/10.1016/j.jebo.2014.06.004

argest social network site. Facebook’s Gross National Happiness Index (FGNHI) has been developed by Facebook’s dataeam and offers daily sentiment for twenty international markets. The website investorwords.com defines sentiment as “a

easurement of the mood of a given investor or the overall investing public, either bullish or bearish.” Facebook measureseople’s mood by examining the positive and negative terms used by Facebook participants. The assumption is that happy

∗ Corresponding author at: Erasmus University Rotterdam, Burgemeester Oudlaan 50, 3000DR Rotterdam, Netherlands. Tel.: +31 104081392.E-mail addresses: [email protected] (A. Siganos), [email protected] (E. Vagenas-Nanos), [email protected]

P. Verwijmeren).1 Sentiment indexes based on surveys include the University of Michigan Consumer Sentiment Index, and the Consumer Confidence Index (see for

xample Brown and Cliff, 2004; Lemmon and Portniaguina, 2006; Qiu and Welch, 2006).2 Several other studies have used indirect measures of sentiment. Indirect sentiment measures represent economic and financial variables that are

elieved to capture investors’ state of mind. Examples of indirect proxies are fund flows, trading volume, IPO volume-first day return, and closed-end fundiscounts (see also Lee et al., 1991; Baker and Wurgler, 2007; Brown et al., 2008).

http://dx.doi.org/10.1016/j.jebo.2014.06.004167-2681/© 2014 Elsevier B.V. All rights reserved.

dx.doi.org/10.1016/j.jebo.2014.06.004

dx.doi.org/10.1016/j.jebo.2014.06.004

http://www.sciencedirect.com/science/journal/01672681

http://www.elsevier.com/locate/jebo

mailto:[email protected]



dx.doi.org/10.1016/j.jebo.2014.06.004

G Model
ARTICLE IN PRESSJEBO-3383; No. of Pages 14
2 A. Siganos et al. / Journal of Economic Behavior & Organization xxx (2014) xxx–xxx

participants on average use more positive terms when updating their status and unhappy participants on average use morenegative terms.

Although many participants on Facebook are young, Facebook is no longer the exclusive domain of young people, and asubstantial amount of Facebook users are likely to invest. Appendix A reports the percentage of a nation’s population thathas a Facebook account. Typically, this percentage is close to 50%, which highlights the high participation rates of Facebook.The average age of Facebook users in recent times is about 31 years (Kramer and Chung, 2011), with more than a quarter ofFacebook users being older than 45. It is also important to note that even when investors are underrepresented on Facebook,it is still the case that the underlying factors that make Facebook users optimistic, like their nation’s win in the World Cup,are also likely to make the investors in that country more optimistic.

The data from Facebook provide some important advantages. First, the index has been constructed based on text analysesof the status updates of millions of participants, which stands in contrast to the limited sample sizes of household surveys.3

Second, FGNHI represents sentiment on a daily level, which allows us to test contemporaneous relations between sentimentand stock market returns. Third, status updates on Facebook are undirected by any particular question that may be askedin surveys, but are self-descriptive messages.4 A fourth benefit of our data is the international coverage. Other sentimentindexes are typically only available for the United States (like the University of Michigan Consumer Sentiment Index) orfor a small number of developed markets (the UBS/Gallup Index of Investor Optimism offers monthly sentiment levels forFrance, Germany, Italy, Spain, and the United Kingdom). We obtain a direct measure of sentiment for twenty countries.5

We explore whether Facebook’s Gross National Happiness Index is related to stock market returns for the periodSeptember 2007–March 2012. Our main hypothesis is that positive sentiment leads to positive biases in returns. This hypoth-esis follows from the behavioral finance theory of De Long et al. (1990), who predict that noise trader sentiment affectsfinancial markets when noise traders are plentiful and there are limits to arbitrage. Other studies have mostly focused ona related prediction following from De Long et al. (1990), which is that prices will revert to fundamental values in the longterm. Most notably, Schmeling (2009) and Baker et al. (2012) show that their measures of investor sentiment are relatedto negative returns in the future, when any overly optimistic or pessimistic expectation is corrected. Our daily sentimentmeasure from Facebook allows us to also test behavioral finance’s predictions on the contemporaneous relation betweensentiment and stock returns.

We find a significant positive relation between sentiment and contemporaneous stock market returns, showing thatoptimistic (pessimistic) sentiment is related to gains (losses) in the market index. These results hold for different regions,languages, and religions. Moreover, these results are not solely driven by particular days on which Facebook’s measure ofsentiment reaches extremely high or low levels. In the cross-section, we expect that optimism is especially related to stockreturns for stocks that are disproportionally held by noise traders (Lee et al., 1991). Because small firms might have relativelymore noise traders as compared to institutional traders, Lemmon and Portniaguina (2006) and Baker and Wurgler (2007)argue that behavioral biases are expected to be mostly present in the stock returns of small firms. We exploit MSCI indexesand confirm that our results are strongest for small firms.

Potentially, the relation between sentiment and stock returns is subject to reverse causality, as good market performancecould create positive feelings (Brown and Cliff, 2004). Our data provide substantial research leverage in this regard. As peoplealso update their status in the evening (after the markets close), we expect to find that sentiment on day t affects returnson day t + 1. In line with this expectation, we observe that sentiment is related to the next day’s market returns. In addition,we exploit the availability of sentiment data on Sundays. Any sentiment observed on Sunday is not likely to be the directresult of market returns on Friday, reducing worries of reversed causality when returns are auto-correlated. We find thatsentiment on Sunday is related to market returns on Monday.

To examine causality further, we use models that adjust for lead-lag effects. The results of our analysis with these modelsagain suggest that sentiment affects market returns. Although these results provide new insights into the relation betweensentiment and stock returns, it is important to stress that our results on causality have to be interpreted with appropriatecaution, as several events might affect both sentiment and stock returns. For example, NASA’s successful Mars landing couldat the same time increase people’s sentiment and increase expected future spending on space programs. Still, we considerit unlikely that these types of events happen frequently enough to drive our results across international markets and in thecross-section.6 In addition, we find that controlling for macroeconomic conditions by using the Policy Uncertainty Index (asdeveloped by Baker et al., 2013) does not change our conclusions. Our results are further strengthened as we show that the


relation between sentiment and stock returns reverses over the following weeks, indicating a correction to fundamentalvalues.

3 Kramer (2010) reports that, on average, over 40 million status updates are posted on Facebook per day.4 Facebook users write their status updates in a box that contains an open question, which is typically: “How are you feeling?”, “How are you doing?”,

“What’s on your mind?”, or “How is it going?”5 Using indirect measures of sentiment also allows for an international study. In particular, Baker et al. (2012) construct sentiment indexes within six

developed countries, using indicators like volatility premiums, IPO underpricing, and number of IPOs. Schmeling (2009) uses consumer confidence levelswithin 18 countries as a measure of sentiment.

6 We have checked all the status updates of our Facebook friends over January 2013 and observed that less than one percent of the updates relate to anevent with potentially important effects on the economy.

dx.doi.org/10.1016/j.jebo.2014.06.004

G ModelJ

tssap

eDfiMmsmstwo

iroo

2

atPiaNtaIS

ddLsF

wwueo(

mpt

2so


A. Siganos et al. / Journal of Economic Behavior & Organization xxx (2014) xxx–xxx 3

We further explore whether sentiment on Facebook is related to trading volume and stock price volatility. We findhat sentiment has a significant negative relation to trading volume across international markets, indicating that negativeentiment is associated with higher transaction volume. This result is in line with evidence from psychology, where negativeentiment causes investors to trade more, as they look to overcome their negative sentiment with a positive outcome fromn alternative activity (Erber and Tesser, 1992). Similarly, we find that sentiment on Facebook is negatively related to stockrice volatility, suggesting that negative sentiment is associated with a higher propensity of investors to speculate.

Our study is related to other contemporary papers that use data from social media to examine financial markets. Forxample, Bollen et al. (2011), Zhang et al. (2011) and Yang et al. (2013) examine mood on Twitter.7 In another related study,a et al. (2013) exploit the volume of queries related to household concerns in Google, and conclude that this volume predictsnancial markets in the United States. Most notably, a higher volume of concerns corresponds to lower S&P 500 returns.ost closely related to our study is Karabulut (2013), who also uses sentiment on Facebook. His study focuses on the U.S.arket, and corroborates our findings that sentiment is positively related to stock returns. The main contribution of our

tudy compared to other papers on social media is to exploit Facebook’s daily sentiment proxy across twenty internationalarkets, and to provide insights into potential causality by exploiting sentiment information on non-trading days. The

trength of the Facebook measure in representing sentiment for a specific country on a specific day originates mostly fromhe sheer size of Facebook. Facebook has over a billion users, and in 2010 has passed Google to become the most visitedebsite in the United States, accounting for more than 7% of U.S. web traffic. Importantly, over 80% of Facebook users reside

utside of the United States, which makes Facebook data perfectly suitable for an international study (Wilson et al., 2012).8

The remainder of the paper is structured as follows. Section 2 describes our data and explores the validity of Facebook’sndex. Section 3 discusses the empirical results on the relation between sentiment and stock returns. We examine issueselated to causality in Section 4 and consider stock price reversals and additional tests exploiting the international aspect ofur data in Section 5. Section 6 examines the relation between sentiment and volume and volatility, and Section 7 containsur conclusions.

. Data and validity of FGNHI

We obtain daily sentiment data from Facebook (http://www.facebook.com/gnh/). Facebook refers to its sentiment indexs the ‘Gross National Happiness’ index, inspired by the former king of Bhutan, Jigme Singye Wangchuck, who in 1972 begano construct an index that attempted to capture his nation’s level of happiness more accurately than the Gross Nationalroduct. Bhutan’s index measures happiness within a multidimensional framework by using 33 (in the latest 2010 index)ndicators based on the following nine domains: psychological wellbeing, health, education, time use, cultural diversitynd resilience, good governance, community vitality, ecological diversity and resilience, and living standards. The Grossational Happiness Index developed by Facebook measures happiness based on people’s status updates, which relates to

he dimension of valence. Facebook’s index was first published in 2009. We collect data in March 2012, when sentiment datare available for the following twenty countries: Argentina, Australia, Austria, Belgium, Canada, Chile, Colombia, Germany,ndia, Ireland, Italy, Mexico, the Netherlands, New Zealand, Singapore, South Africa, Spain, the United Kingdom, the Unitedtates, and Venezuela.

FGNHI is estimated by Facebook’s data team based on the status updates of millions of Facebook participants. The proce-ure is explained and validated in Kramer (2010). Based on Text Analysis and Word Count (TAWC) programs, the Facebookata team analyzes the percentage of ‘positive’ and ‘negative’ terms that are used across all participants. They follow theinguistic Inquiry and Word Count (LIWC) dictionary to categorize terms as positive, neutral, or negative. For example, atatus update of ‘What a nice day’ contains one positive term (‘nice’), and all remaining terms are neutral. More specifically,GNHI is estimated by the Facebook data team as follows:

FGNHIi,j = xp,i − xp,all

�p,all− xn,i − xn,all

�n,all(1)

here FGNHIi,j is the sentiment index of country j at day i, xp,i and xn,i show the average positive (p) and negative (n)ords used respectively at day i for the country, and xp,all , xn,all �p,all , �n,all are the average (x) positive and negative wordssed over the duration of the index and the standard deviation (�) of those variables. Facebook’s data team excludes the


xtreme high and low 10% of the days when estimating xp,all , xn,all �p,all , �n,all to minimize the impact of extreme valuesn the estimation of daily sentiment levels. A positive (negative) FGNHI score at day i for a country indicates an optimisticpessimistic) sentiment above (below) which is found on a typical day in that country.

7 Our sample period is substantially larger than the sample periods used in these studies. Bollen et al. (2011) examine tweets in 2008 and find that someood dimensions are related to the Dow Jones index. Zhang et al. (2011) use a randomized sample of tweets over six months in 2009 and find that the

ercentage of emotional tweets is negatively related to U.S. stock market returns. Yang et al. (2013) conclude that sentiment in tweet messages is relatedo the Dow Jones index for one month in 2013.

8 We further relate to studies on the effects of weather and sports results on stock markets (see for example Saunders, 1993; Hirshleifer and Shumway,003; Edmans et al., 2007). In these studies, sentiment cannot be directly observed, but the assumption is that the weather and sport results affectentiment, which in turn affects market outcomes. With Facebook data we observe sentiment more directly, which, for example, overcomes the existencef non-monotonic relations between weather and sentiment, and the fact that different people prefer different types of weather.

dx.doi.org/10.1016/j.jebo.2014.06.004

http://www.facebook.com/gnh/

ARTICLE IN PRESSG ModelJEBO-3383; No. of Pages 14


Table 1Summary statistics.

Number of observations Average St. dev. Max. Min.

Argentina 1487 −0.012 0.022 0.137 −0.089Australia 1492 −0.004 0.017 0.144 −0.046Austria 1497 −0.007 0.021 0.146 −0.072Belgium 1619 −0.005 0.019 0.142 −0.059Canada 1491 −0.009 0.018 0.135 −0.054Chile 1611 −0.013 0.024 0.142 −0.109Colombia 1600 −0.006 0.024 0.142 −0.072Germany 1620 0.000 0.020 0.134 −0.040India 1611 −0.053 0.035 0.144 −0.183Ireland 1492 −0.013 0.020 0.141 −0.072Italy 1617 0.014 0.032 0.145 −0.042Mexico 1617 −0.006 0.020 0.143 −0.061Netherlands 1615 −0.015 0.020 0.138 −0.070New Zealand 1494 −0.010 0.020 0.137 −0.131Singapore 1494 −0.005 0.017 0.146 −0.043South Africa 1489 −0.008 0.018 0.140 −0.051Spain 1618 −0.011 0.021 0.137 −0.106United Kingdom 1615 −0.007 0.017 0.136 −0.062United States 1613 −0.012 0.022 0.133 −0.058Venezuela 1612 −0.010 0.027 0.144 −0.106

All countries 31,304 −0.010 0.025 0.146 −0.183America 11,031 −0.010 0.023 0.144 −0.109Europe 12,693 −0.005 0.023 0.146 −0.106Other countries 7580 −0.017 0.030 0.146 −0.183

This table reports descriptive statistics for Facebook’s Gross National Happiness Index across twenty international markets. Observations during non-trading days are included. In the last four rows of this table we either cluster all countries, all countries in America, all countries in Europe, or all countriesoutside of America and Europe.

We exclude daily FGNHI observations above the 99th percentile, as we observe that these typically relate to messages like“Merry Christmas” and “Happy New Year.”9 These messages might not necessarily be informative about people’s sentiment.Table 1 reports the number of observations and other summary statistics of our Gross National Happiness index acrosscountries. By construction, the averages are close to zero.

As an untabulated descriptive statistic, we have estimated the correlations of FGNHI across countries. We find that thecorrelations tend to be positive and statistically significant, with an average correlation coefficient of 0.589. We observe thehighest correlation between FGNHI in the United States and Canada (0.921).

Similar to most other direct sentiment indexes, FGHNI reflects sentiment of non-investors. Although Facebook wasinitially intended to be used by students (upon its introduction in 2004), the average age has gradually increased throughoutthe years. For a sample period from September 2007 to February 2010, Kramer and Chung (2011) report that the averageage is 33, 32, 30, and 31 within the United States, Canada, the United Kingdom, and Australia, respectively. In fact, morethan a quarter of Facebook users are older than 45, and less than 10% of users are younger than 18. Appendix A showsthe high participation rates of a country’s (online) population in Facebook. As such, many investors are expected to be onFacebook. In addition, the same underlying factors that make a country’s Facebook population happy should also have apositive influence on the mood of most of the investing population of that country. In line with this reasoning, studies haveshown that investors respond to sentiments that would also influence the mood in a country, like the weather and footballresults (e.g., Saunders, 1993; Hirshleifer and Shumway, 2003; Edmans et al., 2007; Kaplanski et al., 2013). We therefore arguethat the demographic characteristics of Facebook users do not generate a major concern regarding our study’s validity.10

To validate FGNHI empirically, Table 2 tests whether FGNHI is correlated with other recently developed daily sentimentindexes. In particular, we compare the U.S. FGNHI measure to the Gallup and Google indexes. The “Gallup Daily” Index isa sentiment index based on phone call interviews in which U.S. participants are asked about their future expectations.11


The Google sentiment index is developed by Da et al. (2013) and is based on the search activity of U.S. households ofthirty negative terms toward the economy in Google.12 As the Google sentiment index measures pessimistic sentiment, wemultiply the Google sentiment index by minus one.

9 Although stock markets are closed on public holidays, they could be open on, for example, Mother’s Day, when similar messages are posted (Kramer,2010). We also report the results when we exclude values of the FGNHI variable above 0.05 and below −0.05, and when no outliers are excluded from thesample.

10 Because personal information is deleted by Facebook’s data team before they construct the sentiment indexes, the exact demographics of Facebookusers in our sample are not available.

11 See www.gallup.com/poll/122840/gallup-daily-economic-indexes.aspx. We divide Gallup’s values by 1000 for comparability.12 These terms include “recession”, “depression”, “bankruptcy”, and “unemployment”. We manually download the search activity of U.S. households in

the thirty terms through Google’s Insight. The sentiment index is the average logarithmic change in search activity.

dx.doi.org/10.1016/j.jebo.2014.06.004

http://www.gallup.com/poll/122840/gallup-daily-economic-indexes.aspx



Table 2Comparison of U.S. daily sentiment indexes.

Panel A: Descriptive statistics

N Average St. dev Max. Min. Median Start date End date

U.S. FGNHI 1613 −0.012 0.022 0.133 −0.058 −0.013 09/09/2007 02/03/2012Gallup Index 1452 −0.037 0.012 −0.014 −0.066 −0.033 02/01/2008 02/03/2012Google Index 1613 −0.001 0.177 0.460 −0.608 0.008 09/09/2007 02/03/2012

Panel B: Correlation coefficients

Gallup Index Google Index

U.S. FGNHI 0.434* 0.167*

(0.000) (0.000)

This table compares U.S. daily sentiment indexes. Panel A shows descriptive statistics for the U.S. FGNHI, Gallup and Google indexes and Panel B shows thePearson correlations of the FGNHI index with the alternative daily sentiment indexes. U.S. FGNHI is Facebook’s U.S. sentiment index. The Gallup index isbased on phone call interviews regarding households’ confidence in the U.S. economy. The Google index is estimated based on thirty negative terms towardthe economy, as identified by Da et al. (2013), by taking the average logarithmic change in search activity across these terms. We multiply the Google Indexby minus one. Observations during non-trading days are included. P-values are shown in parentheses.

* Indicates statistical significance at the 1% level.

Table 3Sentiment and stock market returns.

Stock market returns

N Parameter estimate Standard error

Panel A: Overall sampleAll countries 22,361 0.031*** 0.008America 7888 0.029*** 0.009Europe 9063 0.035*** 0.013Other countries 5410 0.029*** 0.011

Panel B: Sample with potential outliers excludedAll countries 20,870 0.048*** 0.013America 7557 0.035** 0.014Europe 8743 0.066*** 0.021Other countries 4570 0.037** 0.015

This table shows whether sentiment is related to stock market returns. The parameter estimate represents the coefficient of regressing daily stock marketreturns on our daily sentiment measure from Facebook. Our sample period is September 2007–March 2012. All regressions include day-of-the-week fixedeffects and country fixed effects. In each panel we estimate the regression four times: once for all countries, once for all countries in America, once for allcountries in Europe, and once for all countries outside of America and Europe. We report standard errors clustered by date. Panel B excludes sentimentvalues above 0.05 and below −0.05.

** Indicates statistical significance at the 5% level.

tbo

3

3

iartr

*** Indicates statistical significance at the 1% level.

Panel A of Table 2 offers descriptive statistics of the Gallup and Google indexes. Note that Gallup’s coverage is shorterhan FGNHI’s and Google’s coverage. Panel B of Table 2 shows that there are significantly positive correlation coefficientsetween the U.S. sentiment measure from Facebook and the Gallup and Google indexes, with Pearson correlation coefficientsf 0.434 and 0.167 (both significant at the 1% level), respectively.

. Empirical results on stock returns

.1. The relation between sentiment and contemporaneous stock market returns

De Long et al. (1990) predict that optimism (pessimism) of noise traders causes temporary upward (downward) biasesn stock prices. To test the relation between sentiment and stock returns, we start with a relatively simple test. We poolll countries and focus on contemporaneous relations, i.e. we measure sentiment and stock returns on the same day.13 Our


egression analyses include country fixed effects and day-of-the-week fixed effects, and we cluster standard errors by dateo account for the correlation in returns across countries. Panel A of Table 3 shows the results of regressing stock marketeturns on Facebook’s sentiment measure.

13 TOTMK indexes from Datastream are used for countries’ market returns.

dx.doi.org/10.1016/j.jebo.2014.06.004



Table 4The relation between sentiment and MSCI return indexes.

Small Large Premium Value Growth Premium

All countries 0.043*** 0.035*** 0.009* 0.041*** 0.028*** 0.013***

(0.009) (0.011) (0.005) (0.012) (0.010) (0.005)America 0.048*** 0.032** 0.018*** 0.037*** 0.025* 0.012**

(0.012) (0.014) (0.007) (0.014) (0.013) (0.006)Europe 0.042*** 0.037** 0.005 0.050*** 0.028** 0.022**

(0.012) (0.015) (0.007) (0.017) (0.013) (0.010)Other countries 0.040*** 0.035** 0.006 0.028** 0.032** −0.004

(0.011) (0.015) (0.011) (0.013) (0.014) (0.005)

This table shows whether sentiment is related to the returns within alternative MSCI indexes. We distinguish between the MSCI indexes for small, large,value, and growth stocks. The parameter estimate represents the coefficient of regressing daily stock returns on our daily sentiment measure from Facebook.Our sample period is September 2007–March 2012. All regressions include day-of-the-week fixed effects and country fixed effects. For each style category,we estimate the regression four times: once for all countries, once for all countries in America, once for all countries in Europe, and once for all countriesoutside of America and Europe. The premium is estimated by replacing the dependent variable by the difference in returns of the relevant MSCI indexes.We report standard errors clustered by date in parentheses.

* Indicates statistical significance at the 10% level.** Indicates statistical significance at the 5% level.


We find that FGNHI is positively related to stock returns. This relation is statistically significant at the 1% level. Thecoefficient of 0.031 implies that if sentiment changes from zero to 0.1, then, on average, daily contemporaneous returns are31 basis points higher. These results suggest that optimistic sentiment is associated with gains in the aggregate market.

Panel A of Table 3 also shows the results for different regions. Most countries in our sample are from either America orEurope, and we create subsamples based on these regions. Our third subsample pools all remaining countries. The creationof subsamples is likely to be informative about the robustness of our results. It can be seen that the positive relation betweensentiment and stock market returns is present in all three subsamples.

We further examine whether our results are driven by a few days with very high or low sentiment. Although we havealready excluded observations above the 99th percentile, we extend our exclusion to all FGNHI observations above 0.05or below -0.05. Panel B of Table 3 shows that excluding these observations does not change our conclusions. In fact, thecoefficient estimates are increased, with the coefficient for sentiment being 0.048 rather than 0.031 for the estimation thatincludes all countries.14

3.2. Cross-sectional stock returns

Optimism is expected to be especially related to stock returns for stocks that are disproportionally held by noise traders(Lee et al., 1991). Baker and Wurgler (2007) and Schmeling (2009) argue that small firms in particular might be associatedwith many noise traders and could be more subject to behavioral biases. Indeed, Lemmon and Portniaguina (2006) find thatinvestors appear to overvalue small relative to large stocks when consumer confidence is high. To test the conjecture thatsentiment is more important for small firms, we download both the MSCI indexes for small and large firms from Datastream.

Moreover, we differentiate between value and growth stocks by downloading the representative MSCI indexes for thesetwo classifications. Kumar and Lee (2006) argue that noise traders overweight value stocks, but Baker and Wurgler (2006)argue that extreme growth firms are relatively hard to arbitrage, which could also increase the likelihood of behavioralbiases. The latter study finds empirically that the coefficients are similar for both value and growth firms in the UnitedStates. Schmeling (2009) uses an international setting and finds that the relation is stronger for value firms, but observesthat the relation is also present for growth firms.

Table 4 shows our results for small, large, value, and growth stocks. In line with our expectations, we find that our resultsare strongest for small firms. The small size premium is significant at the 10% level for our overall sample and significant atthe 1% level for the American sample. Our results further corroborate the findings of Schmeling (2009) in that the relationbetween sentiment and returns is stronger for value firms, but also present in growth firms.15 Overall, we conclude that ourresults are relevant for different types of firms.


14 We have also estimated the relation between sentiment and stock returns in a sample where no outliers were deleted. We find that the parametercoefficients of the sentiment variable are 0.013 (significant at the 1% level), 0.011 (significant at the 1% level), 0.012 (significant at the 5% level) and 0.016(significant at the 1% level) for all countries, America, Europe, and the “Other countries”, respectively. Hence, the relation between stock returns andsentiment is present both with and without our sample restrictions.

15 In untabulated results, we focus on the U.S. and corroborate the findings of Baker and Wurgler (2006) that the coefficients for value and growth firmsare similar in the United States.

dx.doi.org/10.1016/j.jebo.2014.06.004



Table 5Sentiment and next day’s stock market returns.



Panel A: Stock market returns on the next dayAll countries 22,397 0.021*** 0.007America 7890 0.014* 0.008Europe 9085 0.026** 0.011Other countries 5422 0.024** 0.011

Panel B: Sunday’s sentiment and Monday’s stock market returnsAll countries 4488 0.042** 0.017America 1573 0.023 0.017Europe 1826 0.050* 0.027Other countries 1089 0.063*** 0.023

This table shows whether sentiment is related to stock market returns on the next day. The parameter estimate represents the coefficient of regressingdaily stock market returns on the lagged value of our daily sentiment measure from Facebook. Our sample period is September 2007–March 2012. Allregressions include day-of-the-week fixed effects and country fixed effects. In each panel we estimate the regression four times: once for all countries,once for all countries in America, once for all countries in Europe, and once for all countries outside of America and Europe. We report standard errorsclustered by date. Panel B is a sub-set of Panel A and only includes our sentiment measure on Sunday, with market returns on Monday.

4

4

hc

etopl

ocw

trses

r

4

rau



. Examining the causal relation between sentiment and stock returns

.1. Reversed causality

Brown and Cliff (2004) stress the importance of potential reversed causality. When returns are high, people could becomeappier. Moreover, evidence by Heimer and Simon (2012) suggests that traders with good performance are more likely toommunicate about their trading activity on networking sites.

Our daily data provide substantial research leverage in examining causality. Facebook statuses are also updated in thevening. In fact, a 2012 Oracle white paper reports that Facebook activity is at particularly high levels around 8 pm, althoughhe overall peak occurs at 3 pm.16 Therefore, as our daily sentiment measure captures some of the sentiment after the closef the market, we can test whether today’s sentiment measures are partially reflected in tomorrow’s stock returns. Thisotential relation is unlikely to be explained by reversed causality. Panel A of Table 5 shows the results when we use the

agged value of our sentiment measure.The results in Table 5 suggest a positive relation between sentiment on Facebook on day t and stock market returns

n day t + 1. This relation holds for all our different regions. The magnitude of the relation is lower than for sentiment andontemporaneous returns, as the coefficient estimate for our sentiment measure is reduced to 0.021 in our specificationith all countries included.

A potentially even stronger test to control for reversed causality is to explore whether Sunday’s sentiment is relatedo Monday’s market returns. That is, we exploit the availability of our sentiment measures on non-trading days. Friday’seturns could perhaps affect mood on Saturday, but it is unlikely that stock returns on Friday have a strong effect on Facebookentiment on Sunday. Therefore, any relation between sentiment on Sunday and stock returns on Monday is unlikely to bexplained by reverse causality, even when the returns on Friday and Monday would be auto-correlated. Panel B of Table 5hows the results.

We find that our results remain relatively strong when we focus on the relation between sentiment on Sunday and stocketurns on Monday. As before, sentiment and stock returns are positively related.

.2. Lead-lag relationships

In this section, we use models representing a more sophisticated method of testing whether sentiment affects stocketurns, as they adjust for multiple lead-lag effects. The goal is to examine interactions between sentiment and stock returnsnd establish Granger-type causality. In line with other recent studies in the field (see for example Schmeling, 2009), wese five lags for sentiment and market returns. More specifically, we estimate the following model:


Market returnit = a1 + a2FGNHIit +5∑

j=1

bijFGNHIit−j +5∑

j=1

cijMarket returnit−j + uit (2)

16 See http://www.oracle.com/us/products/managing-your-facebook-community-1840523.pdf

dx.doi.org/10.1016/j.jebo.2014.06.004

http://www.oracle.com/us/products/managing-your-facebook-community-1840523.pdf



Table 6Lead-lag effects.

All countries America Europe Other countries

Constant 0.001 0.001** 0.000 0.000(0.001) (0.000) (0.001) (0.001)

FGNHIt 0.021** 0.016* 0.018 0.033**

(0.009) (0.010) (0.016) (0.015)FGNHIt−1 −0.001 0.005 0.004 −0.013

(0.011) (0.011) (0.020) (0.017)FGNHIt−2 0.026** 0.025 0.016 0.041**

(0.011) (0.015) (0.018) (0.017)FGNHIt−3 −0.001 −0.005 0.007 −0.010

(0.011) (0.012) (0.017) (0.019)FGNHIt−4 0.009 0.012 0.021 −0.013

(0.011) (0.012) (0.020) (0.015)FGNHIt−5 −0.013 −0.015 −0.015 −0.009

(0.009) (0.010) (0.016) (0.015)Returnst−1 0.028 0.017 0.026 0.047

(0.025) (0.030) (0.035) (0.030)Returnst−2 −0.019 −0.014 −0.031 0.002

(0.032) (0.035) (0.040) (0.035)Returnst−3 −0.034 −0.001 −0.050 −0.039

(0.028) (0.031) (0.038) (0.030)Returnst−4 0.025 0.028 0.028 0.013

(0.028) (0.032) (0.037) (0.033)Returnst−5 −0.031 −0.045 −0.037 −0.003

(0.033) (0.036) (0.041) (0.032)N 22,255 7852 9022 5381

This table examines the relation between sentiment and stock market returns when allowing multiple lead-lag effects. We estimate the following model:

Market returnit = a1 + a2FGNHIit +5∑

j=1

bijFGHNHIit−j +5∑

j=1

cijMarket returnit−j + uit

Five lags are used for both sentiment (FGNHI) and the corresponding market return. Our sample period is September 2007–March 2012. All regressionsinclude day-of-the-week fixed effects and country fixed effects. We estimate the regression four times: once for all countries, once for all countries inAmerica, once for all countries in Europe, and once for all countries outside of America and Europe. We report standard errors clustered by date inparentheses.

* Indicates statistical significance at the 10% level.

** Indicates statistical significance at the 5% level.

We again include day-of-the-week and country fixed effects. Such a fixed effects specification allows individual countriesand days of the week to have different regression constants, whereas slope coefficients are restricted to be equal acrosscountries. Table 6 shows the results of estimating the model.

Table 6 shows that the coefficient estimates for the relation between sentiment (FGNHIit) and stock returns are again pos-itive. For our estimation with all countries included, the coefficient estimate is 0.021 and the effect is statistically significantat the 5% level. As such, our conclusions are unchanged and suggest that sentiment positively affects stock returns.

5. Additional tests

5.1. Stock price reversals

Prior studies (e.g., Schmeling, 2009) have tested the relation between sentiment and stock markets by using monthlydata and examining whether there is a reversed relation between sentiment and stock returns in the subsequent month. Therationale is that if sentiment results in a contemporaneous increase in stock prices, returns should move back to fundamentalvalues in the next period. In this subsection, we examine whether patterns of reversals in stock prices are present in ourdata.

We estimate a regression model for explaining stock returns that uses up to 30-day lags of Facebook’s sentiment measure.That is, we use 31 main explanatory variables, which are contemporaneous sentiment, sentiment at day −1, sentiment atday −2, and so on, until sentiment at day −30. We also include day-of-the-week and country fixed effects. Table 7 reports theparameter coefficients. It can be seen that the relation between sentiment and returns tends to weaken for a higher numberof lags. In other words, sentiment does not have a strongly positive relation with stock returns that are measured a few days


later. After nine days, the relation is insignificantly negative. Many of the days do not show significant effects. The strongesteffect after day 0 is on day 16, when the relation is significantly negative. When one would sum all the coefficients for thesentiment measure from day 0 to day t, this sum firstly becomes negative after 20 days. Although we acknowledge that the

dx.doi.org/10.1016/j.jebo.2014.06.004



Table 7Stock market reversals.

Lags Parameter Standard error Lags Parameter Standard error

0 0.027** 0.0111 0.003 0.013 16 −0.029** 0.0122 0.010 0.010 17 0.018* 0.0103 0.003 0.011 18 0.009 0.0114 0.007 0.009 19 −0.027** 0.0135 0.002 0.010 20 −0.022 0.0226 0.012 0.014 21 −0.001 0.0117 0.000 0.012 22 −0.002 0.0108 0.005 0.010 23 0.010 0.0119 −0.012 0.011 24 0.008 0.011

10 0.003 0.010 25 −0.014 0.01211 −0.005 0.011 26 0.010 0.01212 −0.007 0.010 27 0.005 0.01313 −0.015 0.013 28 0.012 0.01614 0.002 0.013 29 −0.009 0.01115 0.010 0.010 30 0.014 0.010

This table examines the relation between sentiment and stock market returns when adding 30 lagged sentiment variables to our basic regression on therelation between sentiment and stock market returns. Our sample period is September 2007–March 2012 and we include all countries. All regressionsinclude day-of-the-week fixed effects and country fixed effects. We report the parameter coefficients for the different lags of sentiment, and also reportstandard errors clustered by date.

*

at

5

asuUfispmKoB

abwmirs

tIdb

i

Indicates statistical significance at the 10% level.** Indicates statistical significance at the 5% level.

nalysis is subject to noise, the results are in line with prior studies on price reversals and hint toward the suggestion thathere is a correction to fundamental values.17

.2. Macroeconomic adjustments

This subsection examines the effect of macroeconomic news releases on our results. This follows for example Kumarnd Lee (2006), who control for macroeconomic variables such as inflation and GDP to show that the relation betweenentiment and returns is robust. We require a macroeconomic variable that is available on a daily basis. Da et al. (2013)se the Policy Uncertainty Index as developed by Baker et al. (2013) as one of the variables to control for changes in daily.S. macroeconomic conditions.18 We examine the FGNHI coefficients after controlling for the Policy Uncertainty Index andnd that our results are robust. More specifically, we find that U.S. FGNHI is positively related to U.S. returns, for a U.S.ample that consists of 1120 observations with all required information: The FGNHI coefficient changes from 0.049 (with a-value of 0.096) without controlling for macroeconomic conditions to 0.050 (with a p-value of 0.080) with controlling foracroeconomic conditions. The Policy Uncertainty Index obtains a coefficient of 0.003, with a p-value of 0.817. A study of

arabulut (2013) focuses on the United States and corroborates the positive relation between stock returns and sentimentn Facebook when controlling for an alternative measure of macroeconomic conditions, which is the Aruoba-Diebold-Scottiusiness Conditions index.

As an alternative test of the robustness of our findings to fundamental news, we exploit variation in the correlationsmong countries. Weeks in which important fundamental news on the state of the world economy is released are likely toe associated with relatively high correlations in the stock returns among the countries in our sample. On the other hand,eeks in which the correlation among returns is relatively low are less likely to be associated with the release of importantacroeconomic news. We therefore split our sample into weeks in which the average correlation between stock markets

s below the median, and weeks in which the average correlation between stock markets exceeds the median. We thene-estimate the relation between sentiment and stock returns for each subsample with data availability. Panel A of Table 8hows the results.

We find that the relation between sentiment and returns is statistically significant within both subsamples, indicatinghat global macroeconomic news releases are unlikely to drive the observed relation between sentiment and stock returns.n Panel B of Table 8 we examine subsamples based on the average correlation in sentiment levels around the world, againistinguishing between weeks with above-median and weeks with below-median correlations. Again, the positive relation


etween sentiment and stock returns is present in both subsamples.

17 We find that our results are similar if we distinguish between the different geographical regions.18 The Policy Uncertainty Index captures uncertainty in economic policy through a news-based measure that counts terms like “uncertain” and “deficit”n newspaper articles (see http://www.policyuncertainty.com/us daily.html).

dx.doi.org/10.1016/j.jebo.2014.06.004

http://www.policyuncertainty.com/us_daily.html



Table 8Results based on subsamples.


N Parameter Standard error

Panel A: Correlation in returns among countriesAbove-median correlations 10,801 0.032* 0.015Below-median correlations 10,482 0.032** 0.010

Panel B: Correlation in sentiment among countriesAbove-median correlations 10,570 0.044*** 0.012Below-median correlations 10,713 0.027* 0.013

Panel C: LanguageChinese 1065 0.029 0.023Dutch 2309 0.049* 0.025English 7626 0.046*** 0.015German 2223 0.048** 0.020Hindi 1154 0.024* 0.014Italian 1159 0.001 0.013Spanish 6825 0.028*** 0.008

Panel D: ReligionCatholic 13,488 0.030*** 0.009Protestants 6654 0.041*** 0.012Hindu 1154 0.024* 0.014Buddhist 1065 0.029 0.023

This table explores subsamples. In Panel A we distinguish between subsamples with above-median and below-median weekly correlations in daily stockreturns among countries. In Panel B we distinguish between subsamples with above-median and below-median weekly correlations in daily sentimentlevels among countries. In Panel C we distinguish between subsamples based on language, and in Panel D we distinguish between subsamples based onreligion. In Panels A and B, we first split into above- and below-median values for the complete dataset, and then estimate regressions for each group withall available data. The parameter estimate represents the coefficient of regressing daily stock returns on our daily sentiment measure from Facebook. Oursample period is September 2007–March 2012. All regressions include day-of-the-week fixed effects and country fixed effects. We report standard errorsclustered by date.



5.3. Language and religion

We further exploit the availability of international data by examining cultural dimensions. Mihalcea et al. (2007) showthat the measurement of sentiment within non-English languages can be challenging due to differing attributions in themeaning of terms. As Facebook’s data team has to use dictionaries in several languages to distinguish positive from negativeterms, we explore the robustness of the observed relation between sentiment and stock returns for alternative languages.Following Stulz and Williamson (2003), we classify languages based on the language of the majority of households withina country.

Panel C of Table 8 shows the results. We find that the positive relation between sentiment and returns is observablefor different languages. The results are not statistically significant for the Italian and Chinese languages, but it should benoted that for these languages the number of observations is relatively low. We do observe statistically significant positiverelations for the languages Dutch, English, German, Hindi and Spanish.

We further examine subsamples based on religion, as prior studies have indicated that individuals adapt their linguisticbehavior due to their religious network (Baker and Bowie, 2010). We follow Stulz and Williamson (2003) in identifying theprimary religion within a country. Panel D of Table 8 shows that the positive relation between sentiment and returns isstatistically significant for the Catholic, Protestant, and Hindu subsamples, which again highlights the broad relevance ofour results.

6. Empirical results on volume and volatility


In this section we examine whether FGNHI is related to trading volume and volatility. In line with evidence from psy-chology (Erber and Tesser, 1992), negative sentiment could cause investors to trade more in an attempt to overcome theirnegative sentiment with a positive outcome from an alternative activity. Indeed, Chang et al. (2008) find that cloudy weatheris related to higher transaction volumes. Sentiment could also affect investors’ propensity to trade. Brown (1999) finds thatunusual levels of sentiment are related to higher volatility in closed-end fund returns. Lee et al. (2002) use the Investors’

dx.doi.org/10.1016/j.jebo.2014.06.004



Table 9Sentiment and trading volume.

Trading volume


Panel A: Overall sampleAll countries 21,514 −5.097*** 0.742America 7488 −4.293*** 0.793Europe 8830 −5.783*** 0.865Other countries 5196 −5.263*** 1.244

Panel B: Trading volume on the next daysAll countries 21,503 −4.306*** 0.561America 7484 −4.732*** 0.696Europe 8824 −4.205*** 0.701Other countries 5195 −3.787*** 1.099

Panel C: Sunday’s sentiment and Monday’s trading volumeAll countries 4168 −2.426** 1.138America 1394 −2.447 1.522Europe 1751 −3.451** 1.419Other countries 1023 −0.816 1.802


Panel D: Lead-lag effectsConstant 0.031 −0.079** 0.106** 0.058

(0.036) (0.033) (0.044) (0.039)FGNHIt −5.043*** −2.894** −6.882*** −5.952***

(1.339) (1.437) (1.728) (1.423)FGNHIt−1 0.204 −1.434 2.354** −0.002

(0.709) (0.980) (1.026) (1.166)FGNHIt−2 1.262** 0.961 0.763 1.809

(0.636) (0.885) (0.900) (1.144)FGNHIt−3 0.220 0.061 0.046 1.059

(0.556) (0.837) (0.794) (1.114)FGNHIt−4 0.188 −0.728 0.607 1.641

(0.592) (0.877) (0.862) (1.067)FGNHIt−5 2.052*** 1.594* 2.383*** 1.308

(0.629) (0.864) (0.833) (0.948)Volumet−1 0.421*** 0.347*** 0.498*** 0.415***

(0.016) (0.022) (0.023) (0.023)Volumet−2 0.143*** 0.147*** 0.126*** 0.151***

(0.013) (0.021) (0.021) (0.021)Volumet−3 0.058*** 0.071*** 0.047** 0.050**

(0.012) (0.017) (0.020) (0.019)Volumet−4 0.064*** 0.061*** 0.062*** 0.071***

(0.012) (0.015) (0.021) (0.020)Volumet−5 0.105*** 0.104*** 0.098*** 0.101***

(0.011) (0.014) (0.019) (0.019)N 18,609 5960 8194 4455

This table shows whether sentiment is related to trading volume. The parameter estimate represents the coefficient of regressing daily standardized tradingvolume on our daily sentiment measure from Facebook. Our sample period is September 2007–March 2012. All regressions include day-of-the-week fixedeffects, country fixed effects, and five day lagged returns. In each panel we estimate the regression four times: once for all countries, once for all countriesin America, once for all countries in Europe, and once for all countries outside of America and Europe. We report standard errors clustered by date. PanelA examines contemporaneous relations, whereas Panel B examines the relation between sentiment on day t and trading volume on day t + 1. Panel C isa sub-set of Panel B and only includes our sentiment measure on Sunday and trading volume on Monday. In Panel D, we estimate the following model:

trading volumeit = a1 + a2FGNHIit +5∑

j=1


j=1

cijtrading volumeit−j + uit

Five lags are used for both sentiment (FGNHI) and the corresponding trading volume.*

It

s

Indicates statistical significance at the 10% level.** Indicates statistical significance at the 5% level.



ntelligence sentiment index19 in the United States and report that ‘bearish’ shifts in sentiment lead to upward revisions inhe volatility of returns. To our knowledge, ours is the first study to examine the relation between sentiment, stock price

19 The Investors’ Intelligence index is based on classifications of advisory services into ‘bullish’ and ‘bearish’. The Investor Intelligence Sentiment Indexcore is the number of investment advisory services that are ‘bullish’ in relation to the total ‘bullish’ and ‘bearish’ advisory services.

dx.doi.org/10.1016/j.jebo.2014.06.004



Table 10Sentiment and volatility.

Volatility


Panel A: Overall sampleAll countries 21,113 −0.108*** 0.011America 6705 −0.080*** 0.014Europe 9023 −0.108*** 0.014Other countries 5385 −0.133*** 0.018

Panel B: Volatility on the next dayAll countries 21,102 −0.115*** 0.012America 6701 −0.083*** 0.014Europe 9017 −0.119*** 0.015Other countries 5384 −0.136*** 0.018

Panel C: Sunday’s sentiment and Monday’s volatilityAll countries 4259 −0.004 0.003America 1344 −0.002 0.004Europe 1826 −0.006 0.005Other countries 1089 −0.001 0.001


Panel D: Lead-lag effectsConstant 0.003* 0.002 0.005* 0.004**

(0.002) (0.001) (0.002) (0.002)FGNHIt −0.026*** −0.012 −0.030** −0.036**

(0.009) (0.010) (0.015) (0.014)FGNHIt−1 −0.022** −0.030*** −0.034** 0.000

(0.011) (0.011) (0.017) (0.016)FGNHIt−2 0.009 0.000 0.022 0.002

(0.009) (0.012) (0.015) (0.016)FGNHIt−3 −0.016 0.006 −0.036** −0.018

(0.011) (0.017) (0.017) (0.016)FGNHIt−4 0.013 0.002 0.021 0.016

(0.010) (0.012) (0.014) (0.021)FGNHIt−5 −0.002 −0.001 0.009 −0.019

(0.009) (0.010) (0.014) (0.014)Volatilityt−1 0.662*** 0.606*** 0.660*** 0.726***

(0.073) (0.078) (0.091) (0.078)Volatilityt−2 0.009 0.050 −0.005 −0.005

(0.060) (0.071) (0.073) (0.068)Volatilityt−3 −0.078 −0.071 −0.092 −0.046

(0.066) (0.068) (0.085) (0.071)Volatilityt−4 −0.108 −0.118 −0.065 −0.212**

(0.069) (0.072) (0.088) (0.100)Volatilityt−5 0.311*** 0.365*** 0.298*** 0.292***

(0.056) (0.068) (0.068) (0.087)N 21,107 6704 9022 5381

This table shows whether sentiment is related to stock price volatility. The parameter estimate represents the coefficient of regressing the volatility asmeasured by GARCH(1,1) on our daily sentiment measure from Facebook. Our sample period is September 2007–March 2012. All regressions include day-of-the-week fixed effects, country fixed effects, and five day lagged returns. In each panel we estimate the regression four times: once for all countries, oncefor all countries in America, once for all countries in Europe, and once for all countries outside of America and Europe. We report standard errors clusteredby date. Panel A examines contemporaneous relations, whereas Panel B examines the relation between sentiment on day t and stock price volatility on dayt + 1. Panel C is a sub-set of Panel B and only includes our sentiment measure on Sunday and stock price volatility on Monday. In Panel D, we estimate the

following model: volatilityit = a1 + a2FGNHIit +5∑

j=1


j=1

cijvolatilityit−j + uit

Five lags are used for both sentiment (FGNHI) and the corresponding volatility.* Indicates statistical significance at the 10% level.

** Indicates statistical significance at the 5% level.*** Indicates statistical significance at the 1% level.

volatility, and trading volume in an international context. We hypothesize that negative sentiment leads to increased tradingvolume and volatility.

We standardize trading volume by subtracting the mean trading volume in a country and dividing by the standard


deviation of trading volume in a country, and use GARCH(1,1) to measure daily volatility (e.g., Bollerslev, 1986). We controlfor five-day lagged returns in all estimations (coefficients are not reported due to space considerations). The empirical resultsare shown in Tables 9 and 10. In line with our hypothesis, the results in Panel A of Table 9 indicate that sentiment and trading

dx.doi.org/10.1016/j.jebo.2014.06.004

G ModelJ

vrPbb

7

vi

iccct

A

Vh

A

a

R

BBBBBBB

ibn



olume have a negative contemporaneous relation. The relation is highly statistically significant and is observed in differentegions. Similarly, Panel A of Table 10 shows evidence of a negative relation between sentiment and stock price volatility.anels B, C, and D of Tables 9 and 10 examine the robustness of these findings by examining lagged sentiment, the relationetween Sunday’s sentiment and Monday’s market characteristics, and by controlling for multiple lead-lag effects. Overall,oth sentiment and trading volume and sentiment and stock price volatility are negatively related.20

. Conclusion

We employ Facebook’s daily sentiment index and examine its relation to stock returns, trading volume, and stock priceolatility across twenty international markets. Facebook is the world’s largest social network site, and their sentiment indexs based on textual analysis of the status updates of millions of participants.

We observe a positive relation between sentiment on Facebook and stock market returns. We further find that sentiments negatively related to trading volume and volatility. The daily frequency of our data allows for some novel tests on theausality of these relations. Most notably, we examine the relation between sentiment on Sunday, when stock markets arelosed, and stock market characteristics on Monday. Our findings suggest that sentiment has a causal effect on stock marketharacteristics in different geographical regions, highlighting the importance of behavioral finance for stock markets aroundhe world.

cknowledgements

We would like to thank the Guest editors, three anonymous referees, Bruce Grundy, Guy Kaplanski, Meir Statman, Chriseld, Vadym Volosovych and seminar participants at the University of Glasgow for valuable suggestions, and Lisa Zhang forer support during data collection.

ppendix A. Facebook coverage

Data are obtained from Socialbakers.com (last updated on 4th January 2013). The percentage of the online population in country that is on Facebook can exceed 100%, as more than one Facebook account is possible per internet connection.

Percentage of population Percentage of online population

Argentina 50.00 142.08Australia 55.28 70.27Austria 36.07 48.50Belgium 47.45 61.44Canada 54.85 66.55Chile 57.77 125.62Colombia 39.69 103.84Germany 30.98 37.56India 5.34 68.19Ireland 49.05 72.63Italy 38.33 71.16Mexico 35.77 114.22Netherlands 45.52 50.27New Zealand 54.00 63.34Singapore 62.39 81.22South Africa 13.19 104.78Spain 37.83 58.03United Kingdom 53.17 62.87United States 54.37 73.44Venezuela 36.31 95.63

eferences

aker, M., Wurgler, J., 2006. Investor sentiment and the cross-section of stock returns. J. Finance 61, 1645–1680.aker, M., Wurgler, J., 2007. Investor sentiment in the stock market. J. Econ. Perspect. 21, 129–151.


aker, W., Bowie, D., 2010. Religious affiliation as a correlate of linguistic behavior. Working paper.aker, M., Wurgler, J., Yuan, Y., 2012. Global, local and contagious investor sentiment. J. Financ. Econ. 104, 272–287.aker, S., Bloom, N., Davis, S., 2013. Measuring economic policy uncertainty. Working paper.ollen, J., Mao, H., Zeng, X., 2011. Twitter mood predicts the stock market. J. Comput. Sci. 2, 1–8.ollerslev, T., 1986. Generalized autoregressive conditional heteroskedasticity. J. Econometrics 31, 307–327.

20 We find that our results are robust for including the Policy Uncertainty Index as an additional control variable. Tetlock (2007) shows that pessimismn the Wall Street Journal predicts increases in trading volume, and that this relation is non-linear. In untabulated analyses, we explore any non-linearityetween sentiment and trading volume. We use the squared term of our sentiment value and find some evidence for non-linear effects. Importantly, theon-squared term remains consistently negative in these analyses.

dx.doi.org/10.1016/j.jebo.2014.06.004

http://refhub.elsevier.com/S0167-2681(14)00173-5/sbref0005













































































































































G Model
ARTICLE IN PRESSJEBO-3383; No. of Pages 14

Brown, G., 1999. Volatility, sentiment and noise traders. Financ. Anal. J. (March/April), 82–90.Brown, G., Cliff, M., 2004. Investor sentiment and the near-term stock market. J. Empir. Finance 11, 1–27.Brown, S., Goetzmann, W., Hiraki, T., Shiraishi, N., Watanabe, M., 2008. Investor sentiment in Japanese and U.S. daily mutual fund flows. Manager. Finance

34, 772–785.Chang, S., Chen, S., Chou, R., Lin, Y., 2008. Weather and intraday patterns in stock returns and trading activity. J. Bank. Finance 32, 1754–1766.Da, Z., Engelberg, J., Gao, P., 2013. The sum of all fears: investor sentiment and asset prices. Working paper.De Long, B., Shleifer, A., Summers, L., Waldmann, R., 1990. Noise trader risk in financial markets. J. Pol. Econ. 98, 703–738.Edmans, A., Garcia, D., Norli, O., 2007. Sport sentiment and stock returns. J. Finance 4, 1967–1998.Erber, R., Tesser, A., 1992. Task effort and the regulation of mood: the absorption hypothesis. J. Exp. Soc. Psychol. 28, 339–359.Heimer, R., Simon, D., 2012. Facebook finance: how social interaction propagates active investing. Working paper.Hirshleifer, D., Shumway, T., 2003. Good day sunshine: stock returns and the weather. J. Finance 58, 1009–1032.Kaplanski, G., Levy, H., Veld, C., Veld-Merkoulova, Y., 2013. Do happy people make optimistic investors? J. Financ. Quant. Anal. (forthcoming).Karabulut, Y., 2013. Can Facebook predict stock market activity? Working paper.Kramer, A., 2010. An Unobtrusive Behavioral Model of Gross National Happiness, Proceedings CHI. ACM Press, New York, pp. 287–290.Kramer, A., Chung, K., 2011. Dimensions of self-expression in Facebook status updates. In: Proceedings of the 5th international AAAI Conference on Weblogs

and Social Media.Kumar, A., Lee, C., 2006. Retail investor sentiment and return comovements. J. Finance 61, 2451–2486.Lee, C., Shleifer, A., Thaler, R., 1991. Investor sentiment and the closed-end fund puzzle. J. Finance 46, 75–109.Lee, W., Jiang, C., Indro, D., 2002. Stock market volatility, excess returns and the role of investor sentiment. J. Bank. Finance 26, 2277–2299.Lemmon, M., Portniaguina, E., 2006. Consumer confidence and asset prices: some empirical evidence. Rev. Financ. Stud. 19, 1499–1529.Mihalcea, R., Banea, C., Wiebe, J., 2007. Learning multilingual subjective language via cross-lingual projections. Working paper.Qiu, L., Welch, I., 2006. Investor sentiment measures. Working paper.Saunders, E., 1993. Stock prices and Wall Street weather. Am. Econ. Rev. 83, 1337–1345.Schmeling, M., 2009. Investor sentiment and stock returns: some international evidence. J. Empir. Finance 16, 394–408.Stulz, R., Williamson, R., 2003. Culture, openness, and finance. J. Financ. Econ. 70, 313–349.


Tetlock, P., 2007. Giving content to investor sentiment: the role of media in the stock market. J. Finance 62, 1139–1168.Wilson, R.E., Goslin, S.D., Graham, L.T., 2012. A review of Facebook research in the social sciences. Perspect. Psychol. Sci. 7, 203–220.Yang, S., Mo, S., Zhu, X., 2013. An empirical study of the financial community network on Twitter. Working paper.Zhang, X., Fuehres, H., Gloor, P., 2011. Predicting stock market indicators through Twitter: I hope it is not as bad as I fear. Procedia – Soc. Behav. Sci. 26,

55–62.

dx.doi.org/10.1016/j.jebo.2014.06.004


































































































































































































































































































































10.1016 j.jebo.2014.06.004 facebook s daily sentiment and international stock markets

Documents

investor sentiment

jebofacebooks daily

dataffers daily sentiment

average sentiment level

stock returns

stock market returns

international markets

university of glasgow