Facebook's daily sentiment and international stock markets

Journal of Economic Behavior & Organization 107 (2014) 730–743

Contents lists available at ScienceDirect

Journal of Economic Behavior & Organization journal homepage: www.elsevier.com/locate/jebo

Facebook’s daily sentiment and international stock markets Antonios Siganos a , Evangelos Vagenas-Nanos a , Patrick Verwijmeren a,b,c,∗ a b c

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

a r t i c l e

i n f o

Article history: Received 25 February 2013 Received in revised form 30 April 2014 Accepted 10 June 2014 Available online 14 June 2014 Keywords: Behavioral finance Sentiment Facebook’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 international markets by exploiting Facebook’s Gross National Happiness Index. We find that sentiment has a positive contemporaneous relation to stock returns. Moreover, sentiment on Sunday affects stock returns on Monday, suggesting causality from sentiment to stock markets. We observe that the relation between sentiment and returns reverses the following weeks. We further show that negative sentiments are related to increases in trading volume and return volatility. These results highlight the importance of behavioral factors in stock investing. © 2014 Elsevier B.V. All rights reserved.

1. Introduction An important part of behavioral finance concerns the relation between investor sentiment and stock market returns. Measuring sentiment is, however, not a trivial exercise. The conventional method to obtain measurements of sentiment is to take surveys of households. In this type of study, researchers typically select a random number of households and ask a small number of questions to identify the level of optimism or pessimism per household. The responses are then aggregated to construct an average sentiment level.1 Although these studies have provided important insights, the survey method has some important weaknesses. One weakness is that sample sizes and participation rates are typically low. For example, the Michigan Consumer Sentiment survey is sent to only 500 households, and the Consumer Confidence Index to 5000 households. Another weakness is that the surveys are typically conducted on a monthly frequency. The resultant studies 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 largest social network site. Facebook’s Gross National Happiness Index (FGNHI) has been developed by Facebook’s data team and offers daily sentiment for twenty international markets. The website investorwords.com defines sentiment as “a measurement of the mood of a given investor or the overall investing public, either bullish or bearish.” Facebook measures people’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 example 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 believed to capture investors’ state of mind. Examples of indirect proxies are fund flows, trading volume, IPO volume-first day return, and closed-end fund discounts (see also Lee et al., 1991; Baker and Wurgler, 2007; Brown et al., 2008). http://dx.doi.org/10.1016/j.jebo.2014.06.004 0167-2681/© 2014 Elsevier B.V. All rights reserved.

A. Siganos et al. / Journal of Economic Behavior & Organization 107 (2014) 730–743

731

participants on average use more positive terms when updating their status and unhappy participants on average use more negative terms. Although many participants on Facebook are young, Facebook is no longer the exclusive domain of young people, and a substantial amount of Facebook users are likely to invest. Appendix A reports the percentage of a nation’s population that has 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 of Facebook 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 analyses of 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 sentiment and stock market returns. Third, status updates on Facebook are undirected by any particular question that may be asked in surveys, but are self-descriptive messages.4 A fourth benefit of our data is the international coverage. Other sentiment indexes are typically only available for the United States (like the University of Michigan Consumer Sentiment Index) or for a small number of developed markets (the UBS/Gallup Index of Investor Optimism offers monthly sentiment levels for France, 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 period September 2007–March 2012. Our main hypothesis is that positive sentiment leads to positive biases in returns. This hypothesis follows from the behavioral finance theory of De Long et al. (1990), who predict that noise trader sentiment affects financial markets when noise traders are plentiful and there are limits to arbitrage. Other studies have mostly focused on a related prediction following from De Long et al. (1990), which is that prices will revert to fundamental values in the long term. Most notably, Schmeling (2009) and Baker et al. (2012) show that their measures of investor sentiment are related to negative returns in the future, when any overly optimistic or pessimistic expectation is corrected. Our daily sentiment measure from Facebook allows us to also test behavioral finance’s predictions on the contemporaneous relation between sentiment and stock returns. We find a significant positive relation between sentiment and contemporaneous stock market returns, showing that optimistic (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 of sentiment reaches extremely high or low levels. In the cross-section, we expect that optimism is especially related to stock returns for stocks that are disproportionally held by noise traders (Lee et al., 1991). Because small firms might have relatively more 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 indexes and 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 performance could create positive feelings (Brown and Cliff, 2004). Our data provide substantial research leverage in this regard. As people also update their status in the evening (after the markets close), we expect to find that sentiment on day t affects returns on 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 direct result of market returns on Friday, reducing worries of reversed causality when returns are auto-correlated. We find that sentiment 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 models again suggest that sentiment affects market returns. Although these results provide new insights into the relation between sentiment and stock returns, it is important to stress that our results on causality have to be interpreted with appropriate caution, as several events might affect both sentiment and stock returns. For example, NASA’s successful Mars landing could at the same time increase people’s sentiment and increase expected future spending on space programs. Still, we consider it unlikely that these types of events happen frequently enough to drive our results across international markets and in the cross-section.6 In addition, we find that controlling for macroeconomic conditions by using the Policy Uncertainty Index (as developed 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 fundamental values.

3

Kramer (2010) reports that, on average, over 40 million status updates are posted on Facebook per day. 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 levels within 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 an event with potentially important effects on the economy. 4

732

A. Siganos et al. / Journal of Economic Behavior & Organization 107 (2014) 730–743

We further explore whether sentiment on Facebook is related to trading volume and stock price volatility. We find that sentiment has a significant negative relation to trading volume across international markets, indicating that negative sentiment is associated with higher transaction volume. This result is in line with evidence from psychology, where negative sentiment causes investors to trade more, as they look to overcome their negative sentiment with a positive outcome from an alternative activity (Erber and Tesser, 1992). Similarly, we find that sentiment on Facebook is negatively related to stock price 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. For example, Bollen et al. (2011), Zhang et al. (2011) and Yang et al. (2013) examine mood on Twitter.7 In another related study, Da et al. (2013) exploit the volume of queries related to household concerns in Google, and conclude that this volume predicts financial markets in the United States. Most notably, a higher volume of concerns corresponds to lower S&P 500 returns. Most closely related to our study is Karabulut (2013), who also uses sentiment on Facebook. His study focuses on the U.S. market, and corroborates our findings that sentiment is positively related to stock returns. The main contribution of our study compared to other papers on social media is to exploit Facebook’s daily sentiment proxy across twenty international markets, and to provide insights into potential causality by exploiting sentiment information on non-trading days. The strength of the Facebook measure in representing sentiment for a specific country on a specific day originates mostly from the sheer size of Facebook. Facebook has over a billion users, and in 2010 has passed Google to become the most visited website in the United States, accounting for more than 7% of U.S. web traffic. Importantly, over 80% of Facebook users reside outside 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’s index. Section 3 discusses the empirical results on the relation between sentiment and stock returns. We examine issues related to causality in Section 4 and consider stock price reversals and additional tests exploiting the international aspect of our data in Section 5. Section 6 examines the relation between sentiment and volume and volatility, and Section 7 contains our conclusions. 2. Data and validity of FGNHI We obtain daily sentiment data from Facebook (http://www.facebook.com/gnh/). Facebook refers to its sentiment index as the ‘Gross National Happiness’ index, inspired by the former king of Bhutan, Jigme Singye Wangchuck, who in 1972 began to construct an index that attempted to capture his nation’s level of happiness more accurately than the Gross National Product. Bhutan’s index measures happiness within a multidimensional framework by using 33 (in the latest 2010 index) indicators based on the following nine domains: psychological wellbeing, health, education, time use, cultural diversity and resilience, good governance, community vitality, ecological diversity and resilience, and living standards. The Gross National Happiness Index developed by Facebook measures happiness based on people’s status updates, which relates to the dimension of valence. Facebook’s index was first published in 2009. We collect data in March 2012, when sentiment data are available for the following twenty countries: Argentina, Australia, Austria, Belgium, Canada, Chile, Colombia, Germany, India, Ireland, Italy, Mexico, the Netherlands, New Zealand, Singapore, South Africa, Spain, the United Kingdom, the United States, and Venezuela. FGNHI is estimated by Facebook’s data team based on the status updates of millions of Facebook participants. The procedure is explained and validated in Kramer (2010). Based on Text Analysis and Word Count (TAWC) programs, the Facebook data team analyzes the percentage of ‘positive’ and ‘negative’ terms that are used across all participants. They follow the Linguistic Inquiry and Word Count (LIWC) dictionary to categorize terms as positive, neutral, or negative. For example, a status update of ‘What a nice day’ contains one positive term (‘nice’), and all remaining terms are neutral. More specifically, FGNHI is estimated by the Facebook data team as follows: FGNHIi,j =

xp,i − xp,all p,all

−

xn,i − xn,all n,all

(1)

where 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) words used respectively at day i for the country, and xp,all , xn,all p,all , n,all are the average (x) positive and negative words used over the duration of the index and the standard deviation () of those variables. Facebook’s data team excludes the extreme high and low 10% of the days when estimating xp,all , xn,all p,all , n,all to minimize the impact of extreme values on the estimation of daily sentiment levels. A positive (negative) FGNHI score at day i for a country indicates an optimistic (pessimistic) 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 some mood 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 percentage of emotional tweets is negatively related to U.S. stock market returns. Yang et al. (2013) conclude that sentiment in tweet messages is related to 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, 2003; Edmans et al., 2007). In these studies, sentiment cannot be directly observed, but the assumption is that the weather and sport results affect sentiment, which in turn affects market outcomes. With Facebook data we observe sentiment more directly, which, for example, overcomes the existence of non-monotonic relations between weather and sentiment, and the fact that different people prefer different types of weather.

A. Siganos et al. / Journal of Economic Behavior & Organization 107 (2014) 730–743

733

Table 1 Summary statistics.

Argentina Australia Austria Belgium Canada Chile Colombia Germany India Ireland Italy Mexico Netherlands New Zealand Singapore South Africa Spain United Kingdom United States Venezuela All countries America Europe Other countries

Number of observations

Average

St. dev.

Max.

Min.

1487 1492 1497 1619 1491 1611 1600 1620 1611 1492 1617 1617 1615 1494 1494 1489 1618 1615 1613 1612

−0.012 −0.004 −0.007 −0.005 −0.009 −0.013 −0.006 0.000 −0.053 −0.013 0.014 −0.006 −0.015 −0.010 −0.005 −0.008 −0.011 −0.007 −0.012 −0.010

0.022 0.017 0.021 0.019 0.018 0.024 0.024 0.020 0.035 0.020 0.032 0.020 0.020 0.020 0.017 0.018 0.021 0.017 0.022 0.027

0.137 0.144 0.146 0.142 0.135 0.142 0.142 0.134 0.144 0.141 0.145 0.143 0.138 0.137 0.146 0.140 0.137 0.136 0.133 0.144

−0.089 −0.046 −0.072 −0.059 −0.054 −0.109 −0.072 −0.040 −0.183 −0.072 −0.042 −0.061 −0.070 −0.131 −0.043 −0.051 −0.106 −0.062 −0.058 −0.106

31,304 11,031 12,693 7580

−0.010 −0.010 −0.005 −0.017

0.025 0.023 0.023 0.030

0.146 0.144 0.146 0.146

−0.183 −0.109 −0.106 −0.183

This table reports descriptive statistics for Facebook’s Gross National Happiness Index across twenty international markets. Observations during nontrading 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 countries outside 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 across countries. 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 the correlations tend to be positive and statistically significant, with an average correlation coefficient of 0.589. We observe the highest 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 was initially intended to be used by students (upon its introduction in 2004), the average age has gradually increased throughout the years. For a sample period from September 2007 to February 2010, Kramer and Chung (2011) report that the average age is 33, 32, 30, and 31 within the United States, Canada, the United Kingdom, and Australia, respectively. In fact, more than a quarter of Facebook users are older than 45, and less than 10% of users are younger than 18. Appendix A shows the high participation rates of a country’s (online) population in Facebook. As such, many investors are expected to be on Facebook. In addition, the same underlying factors that make a country’s Facebook population happy should also have a positive influence on the mood of most of the investing population of that country. In line with this reasoning, studies have shown that investors respond to sentiments that would also influence the mood in a country, like the weather and football results (e.g., Saunders, 1993; Hirshleifer and Shumway, 2003; Edmans et al., 2007; Kaplanski et al., 2013). We therefore argue that 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 sentiment indexes. In particular, we compare the U.S. FGNHI measure to the Gallup and Google indexes. The “Gallup Daily” Index is a 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 of thirty negative terms toward the economy in Google.12 As the Google sentiment index measures pessimistic sentiment, we multiply 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 the sample. 10 Because personal information is deleted by Facebook’s data team before they construct the sentiment indexes, the exact demographics of Facebook users 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.

734

A. Siganos et al. / Journal of Economic Behavior & Organization 107 (2014) 730–743

Table 2 Comparison of U.S. daily sentiment indexes. Panel A: Descriptive statistics

U.S. FGNHI Gallup Index Google Index

N

Average

St. dev

Max.

Min.

Median

Start date

End date

1613 1452 1613

−0.012 −0.037 −0.001

0.022 0.012 0.177

0.133 −0.014 0.460

−0.058 −0.066 −0.608

−0.013 −0.033 0.008

09/09/2007 02/01/2008 09/09/2007

02/03/2012 02/03/2012 02/03/2012

Panel B: Correlation coefficients

U.S. FGNHI

Gallup Index

Google Index

0.434* (0.000)

0.167* (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 the Pearson correlations of the FGNHI index with the alternative daily sentiment indexes. U.S. FGNHI is Facebook’s U.S. sentiment index. The Gallup index is based on phone call interviews regarding households’ confidence in the U.S. economy. The Google index is estimated based on thirty negative terms toward the economy, as identified by Da et al. (2013), by taking the average logarithmic change in search activity across these terms. We multiply the Google Index by minus one. Observations during non-trading days are included. P-values are shown in parentheses. * Indicates statistical significance at the 1% level.

Table 3 Sentiment and stock market returns. Stock market returns N

Parameter estimate

Standard error

22,361 7888 9063 5410

0.031*** 0.029*** 0.035*** 0.029***

0.008 0.009 0.013 0.011

Panel B: Sample with potential outliers excluded 20,870 All countries America 7557 8743 Europe 4570 Other countries

0.048*** 0.035** 0.066*** 0.037**

0.013 0.014 0.021 0.015

Panel A: Overall sample All countries America Europe Other countries

This table shows whether sentiment is related to stock market returns. The parameter estimate represents the coefficient of regressing daily stock market 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. 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 errors clustered by date. Panel B excludes sentiment values above 0.05 and below −0.05. ** Indicates statistical significance at the 5% level. *** 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 shorter than FGNHI’s and Google’s coverage. Panel B of Table 2 shows that there are significantly positive correlation coefficients between the U.S. sentiment measure from Facebook and the Gallup and Google indexes, with Pearson correlation coefficients of 0.434 and 0.167 (both significant at the 1% level), respectively. 3. Empirical results on stock returns 3.1. 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) biases in stock prices. To test the relation between sentiment and stock returns, we start with a relatively simple test. We pool all countries and focus on contemporaneous relations, i.e. we measure sentiment and stock returns on the same day.13 Our regression analyses include country fixed effects and day-of-the-week fixed effects, and we cluster standard errors by date to account for the correlation in returns across countries. Panel A of Table 3 shows the results of regressing stock market returns on Facebook’s sentiment measure.

13

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

A. Siganos et al. / Journal of Economic Behavior & Organization 107 (2014) 730–743

735

Table 4 The relation between sentiment and MSCI return indexes.

All countries America Europe Other countries

Small

Large

Premium

Value

Growth

Premium

0.043*** (0.009) 0.048*** (0.012) 0.042*** (0.012) 0.040*** (0.011)

0.035*** (0.011) 0.032** (0.014) 0.037** (0.015) 0.035** (0.015)

0.009* (0.005) 0.018*** (0.007) 0.005 (0.007) 0.006 (0.011)

0.041*** (0.012) 0.037*** (0.014) 0.050*** (0.017) 0.028** (0.013)

0.028*** (0.010) 0.025* (0.013) 0.028** (0.013) 0.032** (0.014)

0.013*** (0.005) 0.012** (0.006) 0.022** (0.010) −0.004 (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 countries outside 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. *** Indicates statistical significance at the 1% level.

We find that FGNHI is positively related to stock returns. This relation is statistically significant at the 1% level. The coefficient of 0.031 implies that if sentiment changes from zero to 0.1, then, on average, daily contemporaneous returns are 31 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 or Europe, and we create subsamples based on these regions. Our third subsample pools all remaining countries. The creation of subsamples is likely to be informative about the robustness of our results. It can be seen that the positive relation between sentiment 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 have already excluded observations above the 99th percentile, we extend our exclusion to all FGNHI observations above 0.05 or below -0.05. Panel B of Table 3 shows that excluding these observations does not change our conclusions. In fact, the coefficient estimates are increased, with the coefficient for sentiment being 0.048 rather than 0.031 for the estimation that includes 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 associated with many noise traders and could be more subject to behavioral biases. Indeed, Lemmon and Portniaguina (2006) find that investors appear to overvalue small relative to large stocks when consumer confidence is high. To test the conjecture that sentiment 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 these two 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 behavioral biases. The latter study finds empirically that the coefficients are similar for both value and growth firms in the United States. Schmeling (2009) uses an international setting and finds that the relation is stronger for value firms, but observes that 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 results are strongest for small firms. The small size premium is significant at the 10% level for our overall sample and significant at the 1% level for the American sample. Our results further corroborate the findings of Schmeling (2009) in that the relation between sentiment and returns is stronger for value firms, but also present in growth firms.15 Overall, we conclude that our results 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 parameter coefficients 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 and sentiment 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 firms are similar in the United States.

736

A. Siganos et al. / Journal of Economic Behavior & Organization 107 (2014) 730–743

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

Standard error

Panel A: Stock market returns on the next day All countries 22,397 7890 America 9085 Europe 5422 Other countries

N

0.021*** 0.014* 0.026** 0.024**

0.007 0.008 0.011 0.011

Panel B: Sunday’s sentiment and Monday’s stock market returns All countries 4488 1573 America 1826 Europe 1089 Other countries

0.042** 0.023 0.050* 0.063***

0.017 0.017 0.027 0.023

This table shows whether sentiment is related to stock market returns on the next day. The parameter estimate represents the coefficient of regressing daily stock market returns on the lagged value of 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. 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 errors clustered by date. Panel B is a sub-set of Panel A and only includes our sentiment measure on Sunday, with market returns on Monday. * Indicates statistical significance at the 10% level. ** Indicates statistical significance at the 5% level. *** Indicates statistical significance at the 1% level.

4. Examining the causal relation between sentiment and stock returns 4.1. Reversed causality Brown and Cliff (2004) stress the importance of potential reversed causality. When returns are high, people could become happier. Moreover, evidence by Heimer and Simon (2012) suggests that traders with good performance are more likely to communicate about their trading activity on networking sites. Our daily data provide substantial research leverage in examining causality. Facebook statuses are also updated in the evening. In fact, a 2012 Oracle white paper reports that Facebook activity is at particularly high levels around 8 pm, although the overall peak occurs at 3 pm.16 Therefore, as our daily sentiment measure captures some of the sentiment after the close of the market, we can test whether today’s sentiment measures are partially reflected in tomorrow’s stock returns. This potential relation is unlikely to be explained by reversed causality. Panel A of Table 5 shows the results when we use the lagged 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 on day t + 1. This relation holds for all our different regions. The magnitude of the relation is lower than for sentiment and contemporaneous returns, as the coefficient estimate for our sentiment measure is reduced to 0.021 in our specification with all countries included. A potentially even stronger test to control for reversed causality is to explore whether Sunday’s sentiment is related to Monday’s market returns. That is, we exploit the availability of our sentiment measures on non-trading days. Friday’s returns could perhaps affect mood on Saturday, but it is unlikely that stock returns on Friday have a strong effect on Facebook sentiment on Sunday. Therefore, any relation between sentiment on Sunday and stock returns on Monday is unlikely to be explained by reverse causality, even when the returns on Friday and Monday would be auto-correlated. Panel B of Table 5 shows the results. We find that our results remain relatively strong when we focus on the relation between sentiment on Sunday and stock returns on Monday. As before, sentiment and stock returns are positively related. 4.2. Lead-lag relationships In this section, we use models representing a more sophisticated method of testing whether sentiment affects stock returns, as they adjust for multiple lead-lag effects. The goal is to examine interactions between sentiment and stock returns and establish Granger-type causality. In line with other recent studies in the field (see for example Schmeling, 2009), we use five lags for sentiment and market returns. More specifically, we estimate the following model: Market returnit = a1 + a2 FGNHIit +

5  j=1

16

bij FGNHIit−j +

5 

cij Market returnit−j + uit

j=1

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

(2)

A. Siganos et al. / Journal of Economic Behavior & Organization 107 (2014) 730–743

737

Table 6 Lead-lag effects.

Constant FGNHIt FGNHIt −1 FGNHIt−2 FGNHIt−3 FGNHIt−4 FGNHIt−5 Returnst−1 Returnst−2 Returnst−3 Returnst−4 Returnst−5 N

All countries

America

Europe

Other countries

0.001 (0.001) 0.021** (0.009) −0.001 (0.011) 0.026** (0.011) −0.001 (0.011) 0.009 (0.011) −0.013 (0.009) 0.028 (0.025) −0.019 (0.032) −0.034 (0.028) 0.025 (0.028) −0.031 (0.033) 22,255

0.001** (0.000) 0.016* (0.010) 0.005 (0.011) 0.025 (0.015) −0.005 (0.012) 0.012 (0.012) −0.015 (0.010) 0.017 (0.030) −0.014 (0.035) −0.001 (0.031) 0.028 (0.032) −0.045 (0.036) 7852

0.000 (0.001) 0.018 (0.016) 0.004 (0.020) 0.016 (0.018) 0.007 (0.017) 0.021 (0.020) −0.015 (0.016) 0.026 (0.035) −0.031 (0.040) −0.050 (0.038) 0.028 (0.037) −0.037 (0.041) 9022

0.000 (0.001) 0.033** (0.015) −0.013 (0.017) 0.041** (0.017) −0.010 (0.019) −0.013 (0.015) −0.009 (0.015) 0.047 (0.030) 0.002 (0.035) −0.039 (0.030) 0.013 (0.033) −0.003 (0.032) 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 + a2 FGNHIit +

5 

j=1

bij FGHNHIit−j +

5 

cij Market returnit−j + uit

j=1

Five lags are used for both sentiment (FGNHI) and the corresponding market return. Our sample period is September 2007–March 2012. All regressions include day-of-the-week fixed effects and country fixed effects. 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 errors clustered by date in parentheses. * 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 countries and days of the week to have different regression constants, whereas slope coefficients are restricted to be equal across countries. 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 positive. For our estimation with all countries included, the coefficient estimate is 0.021 and the effect is statistically significant at 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 monthly data and examining whether there is a reversed relation between sentiment and stock returns in the subsequent month. The rationale is that if sentiment results in a contemporaneous increase in stock prices, returns should move back to fundamental values in the next period. In this subsection, we examine whether patterns of reversals in stock prices are present in our data. 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 at day −2, and so on, until sentiment at day −30. We also include day-of-the-week and country fixed effects. Table 7 reports the parameter coefficients. It can be seen that the relation between sentiment and returns tends to weaken for a higher number of 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 strongest effect after day 0 is on day 16, when the relation is significantly negative. When one would sum all the coefficients for the sentiment measure from day 0 to day t, this sum firstly becomes negative after 20 days. Although we acknowledge that the

738

A. Siganos et al. / Journal of Economic Behavior & Organization 107 (2014) 730–743

Table 7 Stock market reversals. Lags

Parameter

Standard error

Lags

Parameter

Standard error

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15

0.027** 0.003 0.010 0.003 0.007 0.002 0.012 0.000 0.005 −0.012 0.003 −0.005 −0.007 −0.015 0.002 0.010

0.011 0.013 0.010 0.011 0.009 0.010 0.014 0.012 0.010 0.011 0.010 0.011 0.010 0.013 0.013 0.010

16 17 18 19 20 21 22 23 24 25 26 27 28 29 30

−0.029** 0.018* 0.009 −0.027** −0.022 −0.001 −0.002 0.010 0.008 −0.014 0.010 0.005 0.012 −0.009 0.014

0.012 0.010 0.011 0.013 0.022 0.011 0.010 0.011 0.011 0.012 0.012 0.013 0.016 0.011 0.010

This table examines the relation between sentiment and stock market returns when adding 30 lagged sentiment variables to our basic regression on the relation between sentiment and stock market returns. Our sample period is September 2007–March 2012 and we include all countries. All regressions include day-of-the-week fixed effects and country fixed effects. We report the parameter coefficients for the different lags of sentiment, and also report standard errors clustered by date. * Indicates statistical significance at the 10% level. ** Indicates statistical significance at the 5% level.

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

5.2. Macroeconomic adjustments This subsection examines the effect of macroeconomic news releases on our results. This follows for example Kumar and Lee (2006), who control for macroeconomic variables such as inflation and GDP to show that the relation between sentiment and returns is robust. We require a macroeconomic variable that is available on a daily basis. Da et al. (2013) use the Policy Uncertainty Index as developed by Baker et al. (2013) as one of the variables to control for changes in daily U.S. macroeconomic conditions.18 We examine the FGNHI coefficients after controlling for the Policy Uncertainty Index and find that our results are robust. More specifically, we find that U.S. FGNHI is positively related to U.S. returns, for a U.S. sample that consists of 1120 observations with all required information: The FGNHI coefficient changes from 0.049 (with a p-value of 0.096) without controlling for macroeconomic conditions to 0.050 (with a p-value of 0.080) with controlling for macroeconomic conditions. The Policy Uncertainty Index obtains a coefficient of 0.003, with a p-value of 0.817. A study of Karabulut (2013) focuses on the United States and corroborates the positive relation between stock returns and sentiment on Facebook when controlling for an alternative measure of macroeconomic conditions, which is the Aruoba-Diebold-Scotti Business Conditions index. As an alternative test of the robustness of our findings to fundamental news, we exploit variation in the correlations among countries. Weeks in which important fundamental news on the state of the world economy is released are likely to be associated with relatively high correlations in the stock returns among the countries in our sample. On the other hand, weeks in which the correlation among returns is relatively low are less likely to be associated with the release of important macroeconomic news. We therefore split our sample into weeks in which the average correlation between stock markets is below the median, and weeks in which the average correlation between stock markets exceeds the median. We then re-estimate the relation between sentiment and stock returns for each subsample with data availability. Panel A of Table 8 shows the results. We find that the relation between sentiment and returns is statistically significant within both subsamples, indicating that global macroeconomic news releases are unlikely to drive the observed relation between sentiment and stock returns. In Panel B of Table 8 we examine subsamples based on the average correlation in sentiment levels around the world, again distinguishing between weeks with above-median and weeks with below-median correlations. Again, the positive relation between 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. The Policy Uncertainty Index captures uncertainty in economic policy through a news-based measure that counts terms like “uncertain” and “deficit” in newspaper articles (see http://www.policyuncertainty.com/us daily.html). 18

A. Siganos et al. / Journal of Economic Behavior & Organization 107 (2014) 730–743

739

Table 8 Results based on subsamples. Stock market returns N Panel A: Correlation in returns among countries Above-median correlations Below-median correlations Panel B: Correlation in sentiment among countries Above-median correlations Below-median correlations

Parameter

Standard error

10,801 10,482

0.032* 0.032**

0.015 0.010

10,570 10,713

0.044*** 0.027*

0.012 0.013

1065 2309 7626 2223 1154 1159 6825

0.029 0.049* 0.046*** 0.048** 0.024* 0.001 0.028***

0.023 0.025 0.015 0.020 0.014 0.013 0.008

13,488 6654 1154 1065

0.030*** 0.041*** 0.024* 0.029

0.009 0.012 0.014 0.023

Panel C: Language Chinese Dutch English German Hindi Italian Spanish Panel D: Religion Catholic Protestants Hindu Buddhist

This table explores subsamples. In Panel A we distinguish between subsamples with above-median and below-median weekly correlations in daily stock returns among countries. In Panel B we distinguish between subsamples with above-median and below-median weekly correlations in daily sentiment levels among countries. In Panel C we distinguish between subsamples based on language, and in Panel D we distinguish between subsamples based on religion. 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 with all available data. 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. We report standard errors clustered by date. * Indicates statistical significance at the 10% level. ** Indicates statistical significance at the 5% level. *** Indicates statistical significance at the 1% level.

5.3. Language and religion We further exploit the availability of international data by examining cultural dimensions. Mihalcea et al. (2007) show that the measurement of sentiment within non-English languages can be challenging due to differing attributions in the meaning of terms. As Facebook’s data team has to use dictionaries in several languages to distinguish positive from negative terms, 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 within a country. Panel C of Table 8 shows the results. We find that the positive relation between sentiment and returns is observable for different languages. The results are not statistically significant for the Italian and Chinese languages, but it should be noted that for these languages the number of observations is relatively low. We do observe statistically significant positive relations 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 linguistic behavior due to their religious network (Baker and Bowie, 2010). We follow Stulz and Williamson (2003) in identifying the primary religion within a country. Panel D of Table 8 shows that the positive relation between sentiment and returns is statistically significant for the Catholic, Protestant, and Hindu subsamples, which again highlights the broad relevance of our 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 psychology (Erber and Tesser, 1992), negative sentiment could cause investors to trade more in an attempt to overcome their negative sentiment with a positive outcome from an alternative activity. Indeed, Chang et al. (2008) find that cloudy weather is related to higher transaction volumes. Sentiment could also affect investors’ propensity to trade. Brown (1999) finds that unusual levels of sentiment are related to higher volatility in closed-end fund returns. Lee et al. (2002) use the Investors’

740

A. Siganos et al. / Journal of Economic Behavior & Organization 107 (2014) 730–743

Table 9 Sentiment and trading volume. Trading volume N

Parameter estimate

Standard error

21,514 7488 8830 5196

−5.097*** −4.293*** −5.783*** −5.263***

0.742 0.793 0.865 1.244

Panel B: Trading volume on the next days All countries 21,503 7484 America 8824 Europe 5195 Other countries

−4.306*** −4.732*** −4.205*** −3.787***

0.561 0.696 0.701 1.099

Panel C: Sunday’s sentiment and Monday’s trading volume 4168 All countries 1394 America 1751 Europe Other countries 1023

−2.426** −2.447 −3.451** −0.816

1.138 1.522 1.419 1.802

Panel A: Overall sample All countries America Europe Other countries

Panel D: Lead-lag effects Constant FGNHIt FGNHIt−1 FGNHIt−2 FGNHIt−3 FGNHIt−4 FGNHIt−5 Volumet−1 Volumet−2 Volumet−3 Volumet−4 Volumet−5 N

All countries

America

Europe

Other countries

0.031 (0.036) −5.043*** (1.339) 0.204 (0.709) 1.262** (0.636) 0.220 (0.556) 0.188 (0.592) 2.052*** (0.629) 0.421*** (0.016) 0.143*** (0.013) 0.058*** (0.012) 0.064*** (0.012) 0.105*** (0.011) 18,609

−0.079** (0.033) −2.894** (1.437) −1.434 (0.980) 0.961 (0.885) 0.061 (0.837) −0.728 (0.877) 1.594* (0.864) 0.347*** (0.022) 0.147*** (0.021) 0.071*** (0.017) 0.061*** (0.015) 0.104*** (0.014) 5960

0.106** (0.044) −6.882*** (1.728) 2.354** (1.026) 0.763 (0.900) 0.046 (0.794) 0.607 (0.862) 2.383*** (0.833) 0.498*** (0.023) 0.126*** (0.021) 0.047** (0.020) 0.062*** (0.021) 0.098*** (0.019) 8194

0.058 (0.039) −5.952*** (1.423) −0.002 (1.166) 1.809 (1.144) 1.059 (1.114) 1.641 (1.067) 1.308 (0.948) 0.415*** (0.023) 0.151*** (0.021) 0.050** (0.019) 0.071*** (0.020) 0.101*** (0.019) 4455

This table shows whether sentiment is related to trading volume. The parameter estimate represents the coefficient of regressing daily standardized trading volume 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, 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 errors clustered by date. Panel A examines contemporaneous relations, whereas Panel B examines the relation between sentiment on day t and trading volume on day t + 1. Panel C is a 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 + a2 FGNHIit +

5 

j=1

bij FGNHIit−j +

5 

cij trading volumeit−j + uit

j=1

Five lags are used for both sentiment (FGNHI) and the corresponding trading volume. * Indicates statistical significance at the 10% level. ** Indicates statistical significance at the 5% level. *** Indicates statistical significance at the 1% level.

Intelligence sentiment index19 in the United States and report that ‘bearish’ shifts in sentiment lead to upward revisions in the 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 Index score is the number of investment advisory services that are ‘bullish’ in relation to the total ‘bullish’ and ‘bearish’ advisory services.

A. Siganos et al. / Journal of Economic Behavior & Organization 107 (2014) 730–743

741

Table 10 Sentiment and volatility. Volatility N Panel A: Overall sample All countries America Europe Other countries Panel B: Volatility on the next day All countries America Europe Other countries

Parameter estimate

Standard error

21,113 6705 9023 5385

−0.108*** −0.080*** −0.108*** −0.133***

0.011 0.014 0.014 0.018

21,102 6701 9017 5384

−0.115*** −0.083*** −0.119*** −0.136***

0.012 0.014 0.015 0.018

−0.004 −0.002 −0.006 −0.001

0.003 0.004 0.005 0.001

Panel C: Sunday’s sentiment and Monday’s volatility 4259 All countries 1344 America 1826 Europe Other countries 1089

Panel D: Lead-lag effects Constant FGNHIt FGNHIt−1 FGNHIt−2 FGNHIt−3 FGNHIt−4 FGNHIt−5 Volatilityt−1 Volatilityt−2 Volatilityt−3 Volatilityt−4 Volatilityt−5 N

All countries

America

Europe

Other countries

0.003* (0.002) −0.026*** (0.009) −0.022** (0.011) 0.009 (0.009) −0.016 (0.011) 0.013 (0.010) −0.002 (0.009) 0.662*** (0.073) 0.009 (0.060) −0.078 (0.066) −0.108 (0.069) 0.311*** (0.056) 21,107

0.002 (0.001) −0.012 (0.010) −0.030*** (0.011) 0.000 (0.012) 0.006 (0.017) 0.002 (0.012) −0.001 (0.010) 0.606*** (0.078) 0.050 (0.071) −0.071 (0.068) −0.118 (0.072) 0.365*** (0.068) 6704

0.005* (0.002) −0.030** (0.015) −0.034** (0.017) 0.022 (0.015) −0.036** (0.017) 0.021 (0.014) 0.009 (0.014) 0.660*** (0.091) −0.005 (0.073) −0.092 (0.085) −0.065 (0.088) 0.298*** (0.068) 9022

0.004** (0.002) −0.036** (0.014) 0.000 (0.016) 0.002 (0.016) −0.018 (0.016) 0.016 (0.021) −0.019 (0.014) 0.726*** (0.078) −0.005 (0.068) −0.046 (0.071) −0.212** (0.100) 0.292*** (0.087) 5381

This table shows whether sentiment is related to stock price volatility. The parameter estimate represents the coefficient of regressing the volatility as measured by GARCH(1,1) on our daily sentiment measure from Facebook. Our sample period is September 2007–March 2012. All regressions include dayof-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, 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 errors clustered by date. Panel A examines contemporaneous relations, whereas Panel B examines the relation between sentiment on day t and stock price volatility on day t + 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 + a2 FGNHIit +

5 

j=1

bij FGNHIit−j +

5 

cij volatilityit−j + uit

j=1

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 trading volume 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 control for five-day lagged returns in all estimations (coefficients are not reported due to space considerations). The empirical results are shown in Tables 9 and 10. In line with our hypothesis, the results in Panel A of Table 9 indicate that sentiment and trading

742

A. Siganos et al. / Journal of Economic Behavior & Organization 107 (2014) 730–743

volume have a negative contemporaneous relation. The relation is highly statistically significant and is observed in different regions. Similarly, Panel A of Table 10 shows evidence of a negative relation between sentiment and stock price volatility. Panels B, C, and D of Tables 9 and 10 examine the robustness of these findings by examining lagged sentiment, the relation between Sunday’s sentiment and Monday’s market characteristics, and by controlling for multiple lead-lag effects. Overall, both sentiment and trading volume and sentiment and stock price volatility are negatively related.20 7. Conclusion We employ Facebook’s daily sentiment index and examine its relation to stock returns, trading volume, and stock price volatility across twenty international markets. Facebook is the world’s largest social network site, and their sentiment index is 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 sentiment is negatively related to trading volume and volatility. The daily frequency of our data allows for some novel tests on the causality of these relations. Most notably, we examine the relation between sentiment on Sunday, when stock markets are closed, and stock market characteristics on Monday. Our findings suggest that sentiment has a causal effect on stock market characteristics in different geographical regions, highlighting the importance of behavioral finance for stock markets around the world. Acknowledgements We would like to thank the Guest editors, three anonymous referees, Bruce Grundy, Guy Kaplanski, Meir Statman, Chris Veld, Vadym Volosovych and seminar participants at the University of Glasgow for valuable suggestions, and Lisa Zhang for her support during data collection. Appendix A. Facebook coverage Data are obtained from Socialbakers.com (last updated on 4th January 2013). The percentage of the online population in a country that is on Facebook can exceed 100%, as more than one Facebook account is possible per internet connection.

Argentina Australia Austria Belgium Canada Chile Colombia Germany India Ireland Italy Mexico Netherlands New Zealand Singapore South Africa Spain United Kingdom United States Venezuela

Percentage of population

Percentage of online population

50.00 55.28 36.07 47.45 54.85 57.77 39.69 30.98 5.34 49.05 38.33 35.77 45.52 54.00 62.39 13.19 37.83 53.17 54.37 36.31

142.08 70.27 48.50 61.44 66.55 125.62 103.84 37.56 68.19 72.63 71.16 114.22 50.27 63.34 81.22 104.78 58.03 62.87 73.44 95.63

References Baker, M., Wurgler, J., 2006. Investor sentiment and the cross-section of stock returns. J. Finance 61, 1645–1680. Baker, M., Wurgler, J., 2007. Investor sentiment in the stock market. J. Econ. Perspect. 21, 129–151. Baker, W., Bowie, D., 2010. Religious affiliation as a correlate of linguistic behavior. Working paper. Baker, M., Wurgler, J., Yuan, Y., 2012. Global, local and contagious investor sentiment. J. Financ. Econ. 104, 272–287. Baker, S., Bloom, N., Davis, S., 2013. Measuring economic policy uncertainty. Working paper. Bollen, J., Mao, H., Zeng, X., 2011. Twitter mood predicts the stock market. J. Comput. Sci. 2, 1–8. Bollerslev, 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 pessimism in the Wall Street Journal predicts increases in trading volume, and that this relation is non-linear. In untabulated analyses, we explore any non-linearity between sentiment and trading volume. We use the squared term of our sentiment value and find some evidence for non-linear effects. Importantly, the non-squared term remains consistently negative in these analyses.

A. Siganos et al. / Journal of Economic Behavior & Organization 107 (2014) 730–743

743

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.