PreprintPDF Available

Music Sentiment and Stock Returns Around the World

Authors:
Preprints and early-stage research may not have been peer reviewed yet.

Abstract and Figures

This paper introduces a real-time, continuous measure of national sentiment that is language-free and thus comparable globally: the positivity of songs that individuals choose to listen to. This is a direct measure of mood that does not require us to pre-specify certain mood-affecting events, nor assume the extent of their impact on investors. We validate our music-based sentiment measure by documenting a correlation with mood swings induced by seasonal factors and weather conditions. We find that music sentiment is positively correlated with same-week market returns and negatively correlated with next-week returns, consistent with sentiment-induced temporary mispricing. Results also hold under a daily analysis and are stronger for countries with greater limits to arbitrage. Music sentiment also predicts increases in net mutual fund flows and absolute sentiment precedes a rise in stock market volatility.
Content may be subject to copyright.
1
Music Sentiment and Stock Returns Around the World*
Alex Edmans a
London Business School, CEPR, and ECGI
Adrian Fernandez-Perez b
Auckland University of Technology
Alexandre Garel c
Audencia Business School
Ivan Indriawan d
Auckland University of Technology
Current draft: January 30, 2020
Abstract
This paper introduces a real-time, continuous measure of national sentiment that is language-
free and thus comparable globally: the positivity of songs that individuals choose to listen to.
This is a direct measure of mood that does not require us to pre-specify certain mood-affecting
events, nor assume the extent of their impact on investors. We validate our music-based
sentiment measure by documenting a correlation with mood swings induced by seasonal factors
and weather conditions. We find that music sentiment is positively correlated with same-week
market returns and negatively correlated with next-week returns, consistent with sentiment-
induced temporary mispricing. Results also hold under a daily analysis and are stronger for
countries with greater limits to arbitrage. Music sentiment also predicts increases in net mutual
fund flows and absolute sentiment precedes a rise in stock market volatility.
JEL Classification: G12; G14
Keywords: Investor Sentiment; Investor Mood; Behavioral Finance
______________________________
* We thank David Hirshleifer for helpful comments.
a aedmans@london.edu, London Business School, Regent’s Park, London NW1 4SA.
b adrian.fernandez@aut.ac.nz, Auckland University of Technology, Private Bag 92006, 1142, Auckland, New
Zealand.
c agarel@audencia.com, Audencia Business School, 8 Route de la Jonelière, 44312 Nantes, France.
d ivan.indriawan@aut.ac.nz, Auckland University of Technology, Private Bag 92006, 1142, Auckland, New
Zealand.
2
The behavioral finance literature shows that investor sentiment significantly affects stock
returns, in contradiction to the efficient market hypothesis. This literature has pioneered a range
of sentiment measures that share a common theme they specify an exogenous shock to a
country’s mood, such as international sporting results, aviation disasters, or the weather, and
assume that it affects the sentiment of the marginal investor.
In this paper, we take a different approach. Rather than studying shocks to sentiment,
we wish to measure a country’s actual sentiment at a point in time. Actual sentiment may be
driven by a wide variety of different events and thus does not require us to pre-specify a
particular set of events. In addition, actual sentiment aims to capture the extent to which events
affect investor mood. It may be that a country has lost an international soccer match, but the
effect on mood is muted either because the loss was predictable or soccer is not a popular sport
in that particular country. Thus, rather than using an exogenous shock that is assumed to affect
how people are feeling, we seek an endogenous measure that reflects it. We wish this measure
to be available at high frequency, at a country rather than city level, and globally comparable.
This final requirement means that we desire a proxy that is language-free and thus does not
require a sentiment dictionary, the accuracy of which may vary across languages.
While feelings are unobservable, they manifest in observable actions. However, there
is no dataset on the vast majority of actions that reflect people’s mood, such as aggressive
behavior or language. We thus study the sentiment of songs that a country’s citizens listen to.
This is based on research from the psychology literature that individuals reflect their mood in
their music choices. In particular, a range of studies document “emotion congruity” that
music is used to validate emotion. Cantor and Zillman (1973) induce emotions in subjects by
showing them films and find that they then prefer emotionally congruent music. North and
Hargreaves (1996) show that participants’ preference for music matches their current emotional
states. Saarikallio and Erkkilä (2007) document that subjects who are sad or angry are inclined
3
to listen to sad music to express their emotions or attain closure. Chen, Zhou, and Bryant (2007)
find that the desire to listen to sad music is strongest immediately after experiencing a negative
mood; they are only likely to listen to uplifting music when some time has passed. Hunter,
Schellenberg, and Griffith (2011) find that the typical preference for upbeat music is eliminated
after inducing a sad mood. Van den Tol and Edwards (2013) find that people listen to sad music
after experiencing negative circumstances due to feeling connected with the music.
Listening data is available on a large scale from Spotify, the leading online music
platform worldwide. It had 286 million monthly active users as of the first quarter of 2020,
ensuring that music played on the platform reflects the mood of a sizeable share of a country’s
population. Based on Q4-2017 U.S. data, 74% of the Spotify users were above 24 years old,
while more than 30% of users are older than 45.
1
Hence, financial market participants are likely
to be represented in the sample of Spotify users. Spotify provides daily statistics of the top 200
songs by the total number of streams in a particular country. It also has an algorithm that
classifies a song’s valence, or positivity, trained on ratings of positivity by musical experts.
We use the valence of the daily top-200 songs played on Spotify in 40 countries as a measure
of the mood of its citizens.
Using an endogenous measure of sentiment also has potential disadvantages. The main
concern is that people may choose to listen to songs whose sentiment contrasts their actual
mood to attenuate mood swings caused by exogenous events for example, attenuating
negative sentiment by playing an upbeat song. Such a concern is inconsistent with the above
papers, which find that people listen to music that reflects their emotions rather than attempting
to neutralize it. For example, funerals play sad songs to reflect the mood, rather than happy
songs to affect it. To address this concern directly, we provide a validation test using
1
Source: https://www.businessofapps.com/data/spotify-statistics/
4
established mood proxies. First, we build on prior literature to identify seasonal factors likely
to affect individuals’ moods (e.g., Thaler, 1987; Kamstra et al., 2017; Birru, 2018; Hirshleifer,
Jiang, and DiGiovanni, 2020). We find that periods of declining mood (e.g. September to
October in the Northern Hemisphere) are associated with a significant decrease in our music-
based sentiment measure. Second, prior literature documents evidence that cloud cover
dampens investor mood (e.g., Hirshleifer and Shumway, 2003; Goetzmann et al., 2015); we
find it is similarly associated with music-based sentiment.
Our main analyses investigate the relation between music sentiment and stock market
returns. We find a positive and significant association between music sentiment and
contemporaneous returns, controlling for past returns, the world market return, seasonalities,
weather conditions, and macroeconomic variables. A one standard deviation increase in music
sentiment is associated with a higher weekly return of 8.5 basis points (bps), or 4.5%
annualized. This effect reverses over the next week: a one standard deviation increase in music
sentiment predicts a lower next-week return of 6 bps, or -2.8% annualized. Both results are
consistent with sentiment-induced temporary mispricing, and prior theoretical and empirical
findings that negative investor sentiment causes prices to temporarily fall but subsequently
correct (De Long et al., 1990; Baker and Wurgler, 2006, 2007; Edmans, Garcia, and Norli,
2007). We obtain similar results with a daily analysis music sentiment is associated with
significantly higher next-day stock returns, but lower returns on the following days. Our results
hold for both dollar and local currency returns, and when excluding one country at a time to
attenuate concerns that they may be driven by a specific country.
To further test whether national sentiment is driving our results, we perform a series of
additional analyses. First, the impact of sentiment should be stronger when there are higher
limits to arbitrage (Baker and Wurgler, 2006, 2007). Over our sample period, some countries
implemented bans on short-selling at the beginning of the COVID-19 pandemic, limiting
5
arbitrage opportunities. We conduct difference-in-difference analyses around these plausibly
exogenous shocks and find that the effect of sentiment on current and future returns intensifies.
Second, prior theoretical and empirical literature suggests that investor sentiment and
the resulting noise trading can affect the volatility as well as level of asset prices (e.g. Black,
1986; De Long et al., 1990; Da, Engelberg, and Gao, 2015). We indeed find a significant
contemporaneous correlation between absolute music sentiment and stock market volatility.
Third, as an out-of-sample test, we move from studying equity indices to equity mutual
funds. Prior literature shows that mutual fund flows are affected by investor sentiment (e.g.,
Ben-Raphael, Kandel, and Wohl, 2011, 2012). We indeed find that music sentiment is a
significantly positive predictor of next-week net fund flows.
Our study contributes to the literature on the effect of investor sentiment on the stock
market. Prior studies have introduced a range of investor sentiment measures, each with their
unique strengths, but also with some limitations. Some studies use rare events that capture
sudden changes to investor mood, such as international sporting results (Edmans, Garcia, and
Norli, 2007), aviation disasters (Kaplanski and Levy, 2010), terrorist attacks (Chen et al., 2019),
and clock changes (Kamstra, Kramer, and Levi, 2000). While powerful where available, such
sentiment measures do not exist for most of the year. In addition, since they are discrete, they
show that sudden shocks to sentiment affect asset prices but do not have implications for more
moderate changes. The market-based sentiment index of Baker and Wurgler (2006) and
surveys (used by, e.g., Brown and Cliff, 2005; Lemmon and Portniaguina, 2006) are continuous
measures, but available at a lower frequency and may capture economic forces other than
sentiment. Other studies have used weather variables such as cloud cover (Hirshleifer and
Shumway, 2003; Goetzmann et al., 2015) or daylight hours (Kamstra, Kramer, and Levi, 2003).
These measures are both continuous and available at high-frequency but do not capture the
strength of their effect on investor mood; in addition, weather in the city where the national
6
stock exchange is located may not be shared by the rest of the country. Our contribution is to
develop a continuous, high-frequency, country-level measure that captures direct
manifestations of citizens’ mood.
More closely related is Gao, Rhen, and Zhang (2020), who use textual analysis of
internet searches to measure sentiment, as developed in Da, Engelberg, and Gao (2015).
2
Like
us, they study an endogenous high-frequency measure of investor sentiment available globally.
However, textual analysis requires pre-specifying a set of keywords as being positive or
negative. The accuracy of this set may vary across languages, reducing the global comparability
of the sentiment measure. Loughran and McDonald (2016) review other limitations of textual
analysis, such as disambiguating sentences, which likely also vary across languages. While our
music-based sentiment measure also involves subjectivity in experts’ opinions of song valence,
similar to the subjectivity in choosing a set of words for a particular language, the sentiment
measure applies to songs all over the world, which increases comparability. While the same
word may have multiple meanings in different languages, music is less equivocal: as is often
emphasized, “music is a universal language.” Mehr et al. (2019) study 315 cultures and find
that they use similar kinds of music in a similar context, suggesting there are universal
properties of music that likely reflect commonalities of human cognition throughout the world.
Thus, a measure of song valence is likely to be applicable globally. Moreover, music captures
ineffable emotions that a word-based sentiment measure cannot capture.
This paper substantially expands and updates a preliminary paper by Fernandez-Perez,
Garel, and Indriawan (2020) which documents a correlation between weekly music sentiment
and stock returns in the U.S. Since the music sentiment measure is only available for a short
2
Other papers using textual analysis to construct a sentiment measure include Tetlock (2007), Das and Chen
(2007), Bollen, Mao, and Zeng (2011), and Garcia (2014). Our paper is also related to studies investigating high-
frequency proxies of sentiment using non-textual sources. For instance, Obaid and Pukthuanthong (2019) measure
sentiment through a sample of editorial news photos.
7
time series, our cross-section of 40 countries is particularly important to verify the robustness
of its impact on stock returns, as well as to conduct cross-country analyses exploiting variation
in limits to arbitrage. We also study the impact of sentiment on volatility and mutual fund flows.
The rest of the paper is organized as follows. In Section 1, we discuss and validate the
music sentiment measure. Section 2 reports our main results and Section 3 additional analyses.
Section 4 concludes.
1. Data and Variable Measurement
1.1 Music sentiment
To measure music sentiment, we collect data from Spotify. Starting from January 1,
2017, Spotify releases, per country, daily statistics of the top 200 songs by the total number of
streams. As of August 2020, Spotify provides data for 62 countries. We only select countries
where Spotify data is available since January 1, 2017, and MSCI stock market indices are
available from Refinitiv (formerly Thomson Reuters). This results in a total sample of 40
countries over the sample period from January 1, 2017 to August 28, 2020.
3
We identify over
54,000 unique songs with over 450 billion streams in total. On average, there are 8.4 million
streams daily, with around 42,000 streams per song.
In addition to the top-200 songs, Spotify also has an algorithm that classifies a song’s
valence, which measures the musical positivity conveyed by a song and ranges from 0 to 1.
This algorithm is trained on positivity ratings by musical experts and can be linked to any song
using the Spotify application-programming interface. Songs with high valence sound more
positive (e.g., happy, cheerful, euphoric), while songs with low valence sound more negative
3
We drop Bulgaria, Estonia, India, Israel, Lithuania, Luxembourg, Romania, South Africa, Thailand, and Vietnam
since their Spotify data is only available for less than one year. We also drop Bolivia, Costa Rica, Dominican
Republic, Ecuador, El Salvador, Guatemala, Honduras, Malta, Nicaragua, Paraguay, Slovakia and Uruguay due to
unavailability of MSCI stock market data.
8
(e.g., sad, depressed, angry). Table A1 reports the songs with the highest and lowest non-zero
valance per country in our sample period. We then construct a stream-weighted average
valence (henceforth SWAV) across the top-200 songs for each day d and country i as follows:
 


 

 
where  is the total streams for song j of country i on day d, and  is the
valence of the song j of country i on day d.
Figure 1 shows a chart of the full sample average SWAV across countries. We observe
that South American countries have a higher average SWAV, while Asian countries have a lower
average SWAV.
Insert Figure 1 here
To match our music sentiment with the stock market and macroeconomic data, we
aggregate the information at a weekly level to avoid non-synchronicity between the opening
and closing times of the stock markets and the time of the day that Spotify reports their daily
statistics. Such an issue may lead to instances where the daily measure of SWAV would partially
lead the daily measure of stock market return and other instances where it would lag it. We
define our sentiment measure as the weekly change in sentiment, both to control for country-
level differences in the average level of sentiment, as shown in Figure 1, and also because we
expect the change in sentiment to cause changes in stock prices. Our music-based mood proxy,
labeled Music Sentiment, is thus given by:
  
9
where  is the stream-weighted average valence for week t (taken every Friday). Music
Sentiment is, therefore, the total change in the stream-weighted average valence of the top-200
songs citizens of country i listen to in week t.
1.2 Sample and summary statistics
We obtain country-level MSCI total return indices from Refinitiv. We use dollar
returns, consistent with the literature on international asset pricing (e.g., Griffin, 2002; Fama
and French, 2017). The list of indices used for each country is given in Table A2 in the
Appendix. Table 1 provides summary statistics by country on our music-based sentiment
measure, market index returns and volatility. We winsorize all continuous variables in our
study at the 2.5% and 97.5% levels similar to Da, Engelberg, and Gao (2015). Music Sentiment
ranges from -0.019% (Argentina) to 0.077% (Latvia). Weekly stock market returns range from
-0.05% (Turkey) to 0.39% (Taiwan) and weekly stock market volatility ranges from 0.61%
(Malaysia) to 2.07% (Argentina).
Insert Table 1 here
1.3 Validation of our music-based sentiment measure
We begin our empirical analysis by validating our music-based sentiment measure using
variables that prior research has shown to affect mood and that are also available for our sample
countries over the sample period. We first draw on prior literature to identify seasonal factors
likely to affect individuals’ moods (e.g., Thaler, 1987; Kamstra, Kramer, and Levi, 2017; Birru,
2018; Hirshleifer, Jiang, and DiGiovanni, 2020). January is associated with the improving
mood of the New Year period. For Northern Hemisphere countries, March is associated with
the highest recovery from seasonal affective disorder (SAD). In contrast, the months of
September and October are associated with the highest onset of the SAD effect. Kamstra et al.
10
(2003) show that the SAD effect is observed both in the Northern and in Southern Hemispheres,
except that for the latter, it is six months out of phase.
Another strand of papers relates mood to weather conditions. Prior literature finds that
cloud cover affects mood (see, e.g., Hirshleifer and Shumway, 2003; Goetzmann et al., 2015).
We test whether our music sentiment is related to weather conditions. We collect local
climatological data from the National Oceanic and Atmospheric Administration website, which
contains hourly weather observations from over 20,000 weather stations worldwide. For each
weather station, we can observe the degree of cloud cover, which takes on integer values from
zero (clear sky) to eight (overcast sky). Following Goetzmann et al. (2015), the average daily
cloud cover is calculated per country using hourly values from 6am to 12pm across the
country’s various weather stations.
4
Since daily cloud cover is highly seasonal, we
deseasonalize it by subtracting each week’s mean cloudiness from the time-series mean, similar
to Hirshleifer and Shumway (2003). We call this measure deseasonalized cloud cover (DCC).
Because our sentiment measure captures a change in sentiment, we use the average daily change
in deseasonalized cloud cover within a week in our validation test (
).
5
We use weather-
induced and calendar-related mood swings rather than events such as international sports results
or aviation disasters, due to few such events in our sample period.
To validate our music construct as a proxy for mood, we test how it relates to the above
seasonal mood patterns and weather conditions. More specifically, we estimate the following
panel regression:
4
Goetzmann et al. (2015) explain that the 6am to 12pm window is when investors are most likely to observe
outdoor weather conditions. For robustness, we also calculate the average daily cloud cover from 6am to 4pm,
similar to Hirshleifer and Shumway (2003). Both results are qualitatively similar.
5
Hirshleifer and Shumway (2003) show that both the change and level of cloudiness are related to mispricing.
11
 

 
where  is an indicator variable that equals one for January and March for
Northern Hemisphere countries (January and September for Southern Hemisphere countries
we do not shift January since it remains the New Year in the Southern Hemisphere) and 0
otherwise,  is an indicator variable that equals one in September and October
for Northern Hemisphere countries (March and April for Southern Hemisphere countries) and
0 otherwise
6
, and 
 is the average daily change in deseasonalized cloud cover within
week t. We estimate equations (2) using Ordinary Least Squares (OLS) and report White-
corrected t-statistics, which are robust to heteroscedasticity. Table A3 lists the variable
definitions and sources.
Table 2 reports the regression estimates. Column (1) includes the month dummies and
country and year fixed effects. It shows that increasing mood periods (Positive Months) are
positively but insignificantly associated with an increase in our music-based sentiment measure
while decreasing mood periods (Negative Months) are significantly negatively associated with
music-based sentiment, with a t-statistic exceeding 7. Column (2) includes the change in
cloudiness and country and month fixed effects, and shows that an increase in cloudiness is
associated with a significant decrease in music sentiment (at the 1% level). Column (3) includes
all of the above explanatory variables, and shows that negative months and increases in
cloudiness continue to be associated with a decline in music sentiment. These results suggest
that our music-based sentiment measure captures mood swings of a country’s individuals
6
Kamstra et al. (2003) find that the effect of SAD is more pronounced in higher latitude countries. Therefore, we
consider only mid-latitude countries (N23º26'22'' - N66º33'39'' in the Northern hemisphere and S23º26'22'' -
S66º33'39'' in the Southern hemisphere) where the four seasons are clearly distinguished. The results are similar if
we consider all countries.
12
caused by well-established mood-affecting factors.
7
The stronger results for decreasing mood
periods are consistent with prior research that negative sentiment has greater effects than
positive sentiment (e.g. Edmans, Garcia, and Norli, 2007).
Insert Table 2 here
2. Results
2.1 Music sentiment and stock market returns
In our main analysis, we investigate the relation between music sentiment and stock
market returns. We estimate the following baseline panel regression:
  
where  is the weekly return of the country’s stock market index, and is a
vector of control variables. We control for the one-week-lagged market return to address
autocorrelation, and the change in cloud cover (
) since it is correlated with both music
sentiment (as shown in Table 2) and stock returns (Hirshleifer and Shumway, 2003). If
sentiment affects domestic stock returns, it should do so over and above the effect of global
events on the domestic market. Thus, we include the contemporaneous weekly world return
(World RET) and three macroeconomic variables. Since macroeconomic variables are
unavailable at high frequency for non-U.S. countries, we employ U.S. variables as in Gao,
Rhen, and Zhang (2020); relatedly, Brusa, Savor, and Wilson (2020) show that US
macroeconomic policy has a larger effect on foreign country stock markets than local
macroeconomic policy. Specifically, we control for the weekly change in uncertainty related
to economic policies, using the weekly news-based measure of U.S. economic policy
7
Table A4 confirms the results of Table 2 at a daily frequency. Specifically, music sentiment is lower on
decreasing mood days (Monday and Sunday) and higher on increasing mood days (Friday and Saturday). In
addition, the daily increase in cloud cover remains negatively associated with music sentiment.
13
uncertainty (ΔEPU) developed by Baker, Bloom, and Davis (2016) and taken from Scott
Baker’s website.
8
We also control for the weekly change in weekly U.S. macroeconomic
activity using the Aruoba, Diebold, and Scotti (2009) index (ΔADS) from the Federal Reserve
website.
9
Finally, we control for the implied volatility of the S&P 500 (VIX) (as in Baker and
Wurgler, 2007; Da, Engelberg, and Gao, 2015), obtained from the Chicago Board Options
Exchange website. It captures investors’ expectations about the volatility of the U.S. stock
market over the following 30 days. For all regressions henceforth, we use country fixed effects
to control for other cross-sectional differences that may drive stock returns, and month fixed
effects to control for time-varying global drivers, including season-induced mood swings not
captured by our music-based sentiment measure.
Table 3, Panel A reports the estimation results of equation (3). We find a positive
association between music sentiment and contemporaneous market returns. A one standard
deviation increase in music sentiment is associated with a higher weekly return of 8.5 bps (4.5%
annualized), significant at the 1% level. Panel B reports the estimation results of equation (3)
using one-week lagged music sentiment as the key independent variable and finds evidence of
reversal. A one standard deviation increase in music sentiment is associated with a lower next-
week return of 6 bps (-2.8% annualized), significant at the 5% level. In sum, music sentiment
is positively correlated with same-week returns and negatively correlated with next-week
returns, a price reversal pattern consistent with sentiment-induced temporary mispricing.
8
This measure is constructed by counting the number of U.S. newspaper articles achieved by the NewsBank
Access World News database with at least one term from each of the following three categories: (i) “economic”
or “economy”; (ii) “uncertain” or “uncertainty”; and (iii) “legislation,” “deficit,” “regulation,” “congress,”
“Federal Reserve,” or “White House.” Baker, Bloom, and Davis (2016) provide evidence that EPU captures
perceived economic policy uncertainty.
9
This index extracts the latent state of macroeconomic activity from a large number of macroeconomic variables
(jobless claims, payroll employment, industrial production, personal income less transfer payments, manufacturing
and trade sales, and quarterly real gross domestic product) using a dynamic factor model.
14
Turning to the control variables, we observe a positive association between world and
domestic market returns, significant at the 1% level. This suggests that domestic stock markets
are highly integrated. Results also show that domestic market returns are serially correlated
and negatively related to increases in economic policy uncertainty.
Insert Table 3 here
Table 4 reports the results of robustness tests. Panel A demonstrates that the results are
robust to estimating equation (3) with local currency returns, to address the concern that
sentiment affects the exchange rate. Panel B reports the results of Table 3 when excluding one
country at a time from our sample. It shows that our main results are not driven by a specific
country.
Insert Table 4 here
Our main analysis focuses on contemporaneous weekly returns because of the non-
synchronicity between the valence of songs streamed on Spotify and stock market returns.
However, one potential concern with a contemporaneous analysis is reverse causality. For
example, it might be that negative stock returns induce low mood and cause people to listen to
negative songs. As a result, the association between music sentiment and stock market returns
at a weekly frequency could result from positive (negative) market returns at the start of the
week inducing positive (negative) mood later in the week.
Table 5 thus studies the link between daily music sentiment and next-day stock returns.
In the daily setting, we include up to five lags of music sentiment, the change in cloud cover,
and the domestic market returns. We include contemporaneous, next-day and prior-day world
market returns, as in Edmans, Garcia, and Norli (2007), because some markets may be lagging
while others may be leading the world index. For similar reasons, we include daily leads and
lags for the U.S. macroeconomic variables. In addition to country and month fixed effects, we
15
include day-of-the-week fixed effects since Table A4 shows that they are significantly
correlated with music sentiment. We find that daily music sentiment is positively correlated
with the next-day index return and negatively correlated with the return five days later. Both
coefficients are significant at the 5% level or better. In economic terms, a one standard
deviation increase in daily music sentiment is associated with a higher next-day return of 1.1
bps (2.8% annualized) and a subsequent lower daily return of 1.4 bps five days later (-3.5%
annualized). This result is consistent with the pattern we observe at the weekly frequency and
suggests that mood swings, as reflected in music sentiment, lead changes in stock prices.
Insert Table 5 here
3. Additional Analyses
3.1.1 Limits to arbitrage
Several factors can exacerbate the effect of investor sentiment on asset prices. One of the most
salient ones is limits to arbitrage (Pontiff, 1996; Shleifer and Vishny, 1997; Baker and Wurgler,
2006). We thus conduct difference-in-difference analyses around plausibly exogenous shocks
to limits to arbitrage. Specifically, we exploit the introduction of short-selling bans by some of
our sample countries during the COVID-19 pandemic as a shock that increased limits to
arbitrage. Prior studies support the introduction of short-selling restrictions as hindering
arbitrage. For example, Ofek, Richardson, and Whitelaw (2004) find that short-sale restrictions
lead to greater deviations from put-call parity in options markets. Bris, Goetzmann and Zhu
(2007) document that prices incorporate negative information faster in countries where short
sales are allowed and practiced. Gao, Rhen, and Zhang (2020) show that the effect of sentiment
is stronger in countries with short-selling bans during the global financial crisis.
Table A5 lists the countries that introduced short-selling bans during the COVID-19
pandemic, as well as the start and end dates of the short-selling bans, from the Yale Program
16
on Financial Stability. For instance, in France, the Financial Market Authority announced a
short-selling ban between March 17, 2020 and May 18, 2020, in the light of the outbreak of
the Coronavirus and its consequences on the economy and financial markets.” These bans were
unexpected and country-specific; many countries exposed to COVID-19 did not introduce
them. We estimate the following difference-in-difference regression:
   
 
where BAN equals 1 if a country i’s stock market is subject to a short-selling ban for the full
week t, and 0 otherwise. We expect the stock price to be more responsive to changes in music
sentiment when limits to arbitrage are greater, i.e., β2 to be positive (negative) for current
(lagged) music sentiment.
Panels A and B of Table 6 report the estimation results of equation (4) for current and
one-week lagged music sentiment, respectively. We find that the coefficient of the interaction
term is positive for current returns and negative for future returns. Music sentiment is
associated with greater contemporaneous stock returns and subsequent reversals under short-
selling bans. Specifically, a one standard deviation increase in music sentiment is associated
with a 39 bps greater increase in the contemporaneous return in ban weeks versus non-ban
weeks, and a 98 bps greater decrease in future returns
10
. In sum, the effect of music sentiment
on market returns is markedly stronger when a country’s stock market is subject to limits to
arbitrage.
Insert Table 6 here
10
While the magnitude is large, we also find a similar magnitude when we control for the COVID-19 period, drop
one country at a time, focus on countries implementing short-selling bans only, focus on EU countries as they are
likely to have been exposed to COVID-19 to a similar degree, compare the association in post-ban months to the
one in the same number of pre-ban months, and interact the ban dummy with the other control variables.
17
3.1.2 Stock market volatility
Prior literature suggests that investor sentiment and the resulting noise trading can affect
the volatility as well as level of asset prices (Black 1986; De Long et al., 1990) since sentiment
should cause prices to first deviate from fundamentals and then correct. Our results at a daily
frequency already show that, within a week, music sentiment is first associated with an increase
in stock market returns and then a reversal, consistent with sentiment exacerbating stock market
return variations. We expect weekly stock market volatility to be positively affected by
contemporaneous weekly absolute music sentiment. We study absolute music sentiment
because large changes in sentiment in either direction should lead to trading. We measure
weekly volatility as the standard deviation of daily stock market returns within a week. To test
our conjecture, we estimate the following panel regression:
  
where Controls include the previous control variables, month and country fixed effects, and
one-week lagged stock market volatility. We exclude the VIX since our dependent variable is
market volatility.
Table 7 reports the estimation results of equation (5). We document a strong
contemporaneous correlation between absolute music sentiment and stock market volatility. A
one standard deviation increase in absolute music sentiment corresponds with a
contemporaneous 3 bps increase in stock market volatility, which is 3 % of the average weekly
volatility of 1.042%. Our findings on stock market returns and stock market volatility paint a
consistent picture of sentiment-induced stock price deviations from fundamentals.
Insert Table 7 here
18
3.1.3 Net equity fund flows
If sentiment affects investment decisions, we would expect it to influence trades of mutual
funds, not just individual equities. For example, positive mood should lead investors to be
optimistic and thus buy into funds; indeed Ben-Raphael, Kandel, and Wohl (2011, 2012) find
that individual investor sentiment is significantly positively correlated with mutual fund flows.
We expect music sentiment to be positively related to mutual fund net inflows. We use
one-week lagged music sentiment because it takes several days for flows to be settled and
reported (Da, Engelberg, and Gao, 2015). We collect information on daily net fund flows from
Morningstar, focusing on open-end equity mutual funds denominated in local currency, and
convert these flows to US dollars. We remove duplicates (funds with exactly the same time
series of net flows and size) and funds with less than one observation per week on average (i.e.,
less than 188 observations over our sample period). We also drop funds that started after the
beginning of our sample period (January 1, 2017) and fund-week observations with less than
$15 million of assets under management, following Pastor and Vorsatz (2020). The latter is
because, for small funds, modest dollar flows can translate into extreme percentage flows; the
results are similar when we use alternative cut-off points such as $20 million of assets under
management. This screening process results in 8,340 equity funds from 31 different countries
and around 1,432,000 fund-week observations
11
. For each fund, we aggregate the daily net
fund flows within the week and scale the weekly net fund flows by the fund’s total assets under
management in the previous week. We then estimate the following panel regression:
  
11
The countries we exclude from our analysis as a result of our screening process are: Argentina, Canada,
Colombia, Hungary, Latvia, Panama, Peru, Poland, and Turkey.
19
where  is the weekly size-scaled net flow of fund f, in country i, in week t.
Controls are our previous controls, including month and fund fixed effects, plus one-week-
lagged net equity fund flows to control for potential serial correlation in the fund flows. These
controls are used in Da, Engelberg, and Gao (2015), for instance.
Table 8 reports the results of the estimation of equation (6). We find that music
sentiment is positively related to future equity fund flows. A one standard deviation increase
in music sentiment corresponds to an average increase in net fund flows of 0.0031%. Since the
average fund size is $963 million, a one standard deviation increase in music sentiment
corresponds to a weekly (annual) net flow of about $30,000 ($1.5 million).
12
The former is
comparable with the average weekly net flow in our sample of -$30,672. Our results suggest
significant inflows to the equity market the week following an increase in music sentiment.
This finding is consistent with the argument that music sentiment affects investment decisions.
Insert Table 8 here
4. Conclusion
This study introduces a novel measure of investor sentiment, which captures actual sentiment
rather than shocks to sentiment. It is continuous, available at high-frequency and on a global
scale, and does not require the pre-specification of particular mood-affecting events or words
that capture mood. We provide validation tests and show that seasonal factors, such as mood-
decreasing months and increases in cloud cover, are associated with a significant decrease in
our music-based sentiment measure.
In our main findings, we document a positive and significant relation between music
sentiment and contemporaneous market returns, controlling for world market returns,
12
Wang and Young (2020) find that a one standard deviation increase in the level of terrorism corresponds to an
average decline in fund inflows of $197,000 per month, or $45,500 per week. This is a similar order of magnitude
to our effect, although larger since terrorism likely has a larger effect than sentiment reflected in music.
20
seasonalities, and macroeconomic variables. We also find a significant price reversal the
following week. Hence, our findings are consistent with sentiment-induced temporary
mispricing that subsequently reverses.
We show that the relationship between music sentiment and market returns is stronger
for countries with greater limits to arbitrage, such as those that implement short-selling
restrictions during the COVID-19 pandemic. Music sentiment also predicts increases in net
mutual fund flows and absolute sentiment precedes a rise in stock market volatility. Overall,
our study provides evidence that the actual sentiment of a country’s citizens significantly affects
asset prices.
21
References
Aruoba, S. B., Diebold, F. X., and Scotti, C. (2009). Real-time measurement of business
conditions. Journal of Business and Economic Statistics 27, 417-427.
Baker, M., and Wurgler, J. (2006). Investor sentiment and the cross‐section of stock returns.
Journal of Finance 61, 1645-1680.
Baker, M., and Wurgler, J. (2007). Investor sentiment in the stock market. Journal of Economic
Perspectives 21, 129-152.
Baker, S. R., Bloom, N., and Davis, S. J. (2016). Measuring economic policy uncertainty.
Quarterly Journal of Economics 131, 1593-1636.
Barber, B. M., Odean, T., and Zhu, N. (2009). Systematic noise. Journal of Financial Markets
12, 547-569.
Ben-Rephael, A., Kandel, S., and Wohl, A. (2011). The price pressure of aggregate mutual fund
flows. Journal of Financial and Quantitative Analysis, 46, 585-603.
Ben-Rephael, A., Kandel, S., and Wohl, A. (2012). Measuring investor sentiment with mutual
fund flows. Journal of Financial Economics, 104, 363-382.
Birru, J. (2018). Day of the week and the cross-section of returns. Journal of Financial
Economics 130, 182-214.
Black, F. (1986). Noise. Journal of Finance 41, 529-543.
Bollen, J., Mao, H., and Zeng, X. (2011). Twitter mood predicts the stock market. Journal of
Computational Science 2, 1-8.
Bris, A., Goetzmann, W. N., and Zhu, N. (2007). Efficiency and the bear: Short sales and
markets around the world. Journal of Finance 62, 1029-1079.
Brown, G. W., and Cliff, M. T. (2005). Investor sentiment and asset valuation. Journal of
Business 78, 405-440.
Brown, S. J., Goetzmann, W. N., Hiraki, T., Shirishi, N., and Watanabe, M. (2003). Investor
sentiment in Japanese and US daily mutual fund flows. Working Paper.
Brusa, F., Savor, P., and Wilson, M. (2020). One central bank to rule them all. Review of
Finance 24, 263-304.
Cantor, J. R., and Zillman, D. (1973). Resentment toward victimized protagonists and severity
of misfortunes they suffer as factors in humor appreciation. Journal of Experimental
Research in Personality 6, 321-329.
Chen, Y., Goyal, A., Veeraraghavan, M., and Zolotoy, L. (2019). Terrorist attacks, investor
sentiment, and the pricing of initial public offerings. Working Paper.
Chen, L., Zhou, S., and Bryant, J. (2007). Temporal changes in mood repair through music
consumption: Effects of mood, mood salience, and individual differences. Media
Psychology 9, 695-713.
Da, Z., Engelberg, J., and Gao, P. (2015). The sum of all FEARS investor sentiment and asset
prices. Review of Financial Studies 28, 1-32.
Das, S. R., and Chen, M. Y. (2007). Yahoo! for Amazon: Sentiment extraction from small talk
on the web. Management Science 53, 1375-1388.
22
De Long, J. B., Shleifer, A., Summers, L. H., and Waldmann, R. J. (1990). Noise trader risk in
financial markets. Journal of Political Economy 98, 703-738.
Edmans, A., Garcia, D., and Norli, Ø. (2007). Sports sentiment and stock returns. Journal of
Finance 62, 1967-1998.
Fama, E. F., and French, K. R. (2017). International Tests of a Five-Factor Asset Pricing Model.
Journal of Financial Economics 123, 441463.
Fernandez-Perez, A., Garel, A., and Indriawan, I. (2020). Music sentiment and stock returns.
Economics Letters, 109260.
Garcia, D. (2013). Sentiment during recessions. Journal of Finance 68, 12671300.
Gao, Z., Rhen, H., and Zhang. B., (2020). Googling investor sentiment around the world.
Journal of Financial and Quantitative Analysis 55, 549-580.
Goetzmann, W. N., Kim, D., Kumar, A., and Wang, Q. (2015). Weather-induced mood,
institutional investors, and stock returns. Review of Financial Studies 28, 73-111.
Griffin, J. M. (2002). Are the Fama and French factors global or country specific? Review of
Financial Studies 15, 783803.
Hirshleifer, D., and Shumway, T. (2003). Good day sunshine: Stock returns and the weather.
Journal of Finance 58, 1009-1032.
Hirshleifer, D., Jiang, D., and DiGiovanni, Y. M. (2020). Mood beta and seasonalities in stock
returns. Journal of Financial Economics 137, 272295.
Hunter, P. G., Schellenberg, E. G., and Griffith, A. T. (2011). Misery loves company: Mood-
congruent emotional responding to music. Emotion 11, 1068.
Kamstra, M. J., Kramer, L. A., and Levi, M. D. (2000). Losing sleep at the market: The daylight
saving anomaly. American Economic Review 90, 1005-1011.
Kamstra, M. J., Kramer, L. A., and Levi, M. D. (2003). Winter blues: A SAD stock market
cycle. American Economic Review 93, 324-343.
Kamstra, M. J., Kramer, L. A., Levi, M. D., and Wermers, R. (2017). Seasonal asset allocation:
Evidence from mutual fund flows. Journal of Financial and Quantitative Analysis 52,
71-109.
Kaplanski, G., and Levy, H. (2010). Sentiment and stock prices: The case of aviation disasters.
Journal of Financial Economics 95, 174-201.
Kumar, A., and Lee, C. M. (2006). Retail investor sentiment and return comovements. Journal
of Finance 61, 2451-2486.
Lemmon, M., and Portniaguina, E. (2006). Consumer confidence and asset prices: Some
empirical evidence. Review of Financial Studies 19, 1499-1529.
Loughran, T., and McDonald, B. (2016). Textual analysis in accounting and finance: A survey.
Journal of Accounting Research 54, 1187-1230.
Mehr, S. A., Singh, M., Knox, D., Ketter, D. M., Pickens-Jones, D., Atwood, S., ... and Howard,
R. M. (2019). Universality and diversity in human song. Science 366.
Obaid, K., and Pukthuanthong, K. (2019). A picture is worth a thousand words: Measuring
investor sentiment by combining machine learning and photos from news. Working
Paper.
23
Ofek, E., Richardson, M., and Whitelaw, R. F. (2004). Limited arbitrage and short sales
restrictions: Evidence from the options markets. Journal of Financial Economics 74,
305-342.
Pástor, L., and Vorsatz, M. B. (2020). Mutual fund performance and flows during the COVID-
19 crisis. Review of Asset Pricing Studies 104, 791-833.
Pontiff, J., (1996). Costly arbitrage: Evidence from closed-end funds. Quarterly Journal of
Economics 111, 11351151.
Saarikallio, S., and Erkkilä, J. (2007). The role of music in adolescents' mood regulation.
Psychology of Music 35, 88-109.
Shleifer, A., and Vishny, R. W. (1997). The limits of arbitrage. Journal of Finance 52, 35-55.
Tetlock, P. C. (2007). Giving content to investor sentiment: The role of media in the stock
market. Journal of Finance 62, 1139-1168.
Thaler, R. H. (1987). Anomalies: Weekend, holiday, turn of the month, and intraday effects.
Journal of Economic Perspectives 1, 169-177.
Van den Tol, A. J., and Edwards, J. (2013). Exploring a rationale for choosing to listen to sad
music when feeling sad. Psychology of Music 41, 440-465.
Wang, A. Y., and Young, M. (2020). Terrorist attacks and investor risk preference: Evidence
from mutual fund flows. Journal of Financial Economics 137, 491-514.
24
Table 1. Summary statistics
This table reports summary statistics (full sample average) on our main variables. The sample
period is from January 1, 2017 to August 28, 2020. Music Sentiment is the weekly change in
the stream-weighted average valence of the top-200 songs played on Spotify for a country
(multiplied by 100). RET is the weekly stock market returns. VOL is the standard deviation of
the daily stock market returns within the week.
Country
Music Sentiment
RET (%)
VOL (%)
Argentina
-0.019
0.138
2.075
Australia
0.006
0.246
0.938
Austria
0.029
0.063
1.211
Belgium
0.027
0.054
0.992
Brazil
0.015
0.131
1.660
Canada
0.047
0.195
0.736
Chile
-0.011
-0.002
1.192
Colombia
0.026
0.077
1.217
Czech
0.047
0.217
0.814
Denmark
0.039
0.383
0.894
Finland
0.020
0.346
0.963
France
0.010
0.206
0.871
Germany
-0.004
0.203
0.926
Greece
-0.008
0.028
1.513
Hong Kong
-0.001
0.153
0.898
Hungary
0.034
0.283
1.247
Iceland
0.032
0.140
0.963
Indonesia
-0.018
0.154
1.170
Ireland
0.030
0.250
1.028
Italy
0.011
0.246
1.063
Japan
0.025
0.149
0.823
Latvia
0.077
0.293
0.900
Malaysia
0.043
0.114
0.611
Mexico
0.012
0.062
1.224
Netherlands
0.036
0.350
0.805
New Zealand
0.007
0.389
0.955
Norway
0.033
0.233
1.080
Panama
0.027
-0.001
0.697
Peru
0.014
0.145
1.145
Philippines
0.004
0.082
1.096
Poland
0.039
0.232
1.226
Portugal
0.004
0.284
0.951
Singapore
0.010
0.139
0.801
Spain
0.010
0.103
0.988
Sweden
0.046
0.293
1.031
Switzerland
0.027
0.352
0.698
Taiwan
-0.008
0.387
0.915
Turkey
0.005
-0.051
1.721
UK
0.024
0.077
0.843
US
0.036
0.330
0.782
Whole sample average
0.020
0.187
1.042
Whole sample SD
1.183
2.680
0.694
25
Table 2. Validation of our music-based sentiment measure
This table reports the regression estimates of equation (2) from January 1, 2017 to August 28,
2020. The dependent variable, Music Sentiment, is weekly change in stream-weighted average
valence of the top-200 songs on Spotify. In columns (1), Positive months is an indicator variable
that equals one in January and March (January and September) for the North Hemisphere (South
Hemisphere) countries, and zero otherwise. Negative months is an indicator variable that equals
one in September and October (March and April) for the North Hemisphere (South
Hemisphere) countries, and zero otherwise. In column (2), 
is the average daily change
in deseasonalized cloud cover over the week. Column (3) combines all variables. In columns
(1) and (3) regressions include country and year fixed effects. In column (2), the regression
includes country and month fixed effects. Constants are not reported. White-corrected t-
statistics are in parentheses. *, **, and *** denote significance at the 10%, 5% and 1% level,
respectively. Table A3 provides the variable definitions. All coefficients are multiplied by 100.
Music Sentiment
(1)
(2)
(3)
Positive months
0.013
(0.29)
0.014
(0.33)
Negative months
-0.319***
(-7.35)
-0.318***
(-7.31)

-0.158***
(-2.95)
-0.167***
(-2.99)
Fixed Effects
Country, year
Country, month
Country, year
0.88%
1.81%
1.03%
#Obs.
5,890
7,560
5,857
26
Table 3. Music sentiment and stock market returns
This table reports the regression estimates from equation (3) over the sample period from January 1, 2017 to August 28, 2020. The dependent
variable is the weekly stock market return (RET). In Panel A, the main independent variable is Music Sentiment, the weekly change in the stream-
weighted average valence of the top-200 songs on Spotify for week t in a country i. The control variables are the one-week lagged dependent
variable (RET(t-1)), weekly return of the MSCI World index (World RET), contemporaneous implied volatility (VIX), weekly change in economic
policy uncertainty (ΔEPU), weekly change in macroeconomic activity (ΔADS), and the average daily change in deseasonalized cloud cover over
the week (
. All regressions include country and month fixed effects. White-corrected t-statistics are in parentheses. *, **, and *** denote
significance at the 10%, 5% and 1% level, respectively. Constants are not reported. Table A3 provides the variable definitions.
RET (%)
Panel A: Contemporaneous Music Sentiment
Panel B: One-week lagged Music Sentiment
(1)
(2)
(3)
(4)
Music Sentiment
8.285***
(3.27)
7.180***
(3.70)
-16.553***
(-6.34)
-4.655**
(-2.26)
World RET
0.870***
(55.10)
0.868***
(54.60)
VIX
-0.001
(-0.03)
-0.001
(-0.10)
ΔEPU
-0.003***
(-6.51)
-0.003***
(-6.42)
ΔADS
0.021
(0.43)
0.010
(0.20)

0.047
(0.54)
0.037
(0.43)
RET(t-1)
-0.037***
(-2.57)
-0.036***
(-2.49)
Fixed Effects
Country, month
Country, month
Country, month
Country, month
3.10%
36.71%
3.22%
36.65%
#Obs.
7,560
7,520
7,560
7,520
27
Table 4. Robustness checks
This table reports the regression estimates from equation (3) over the sample period from
January 1, 2017 to August 28, 2020. Panel A reports the results of the estimation of Table 3
using local-currency market returns. Panel B reports the regression estimates dropping one
country at a time. All regressions include country and month fixed effects. White-corrected t-
statistics are in parentheses. *, **, and *** denote significance at the 10%, 5% and 1% level,
respectively. Constants are not reported. Table A3 provides the variable definitions.
Panel A: Local-currency market returns
RET (%)
Contemporaneous
One-week lagged
(1)
(2)
(3)
(4)
Music Sentiment
4.599**
(2.13)
5.059***
(2.69)
-10.579***
(-4.72)
-5.525***
(-2.79)
Fixed Effects
Country, month
Country, month
Country, month
Country, month
Controls
No
Yes
No
Yes
2.29%
20.96%
2.43%
20.97%
#Obs.
7,600
7,520
7,560
7,520
28
Panel B: Excluding one country
Excluded
country
Contemporaneous
One-week lagged
without controls
with controls
without controls
with controls
Argentina
8.431***
(3.39)
7.048***
(3.70)
-16.565***
(-6.39)
-5.043**
(-2.47)
Australia
8.776***
(3.45)
7.287***
(3.70)
-16.906***
(-6.40)
-4.961**
(-2.37)
Austria
7.624***
(3.01)
6.438***
(3.29)
-16.242***
(-6.20)
-4.289**
(-2.06)
Belgium
8.520***
(3.34)
6.985***
(3.56)
-16.379***
(-6.18)
-4.398**
(-2.10)
Brazil
8.169***
(3.27)
6.635***
(3.48)
-15.963***
(-6.16)
-4.095**
(-2.02)
Canada
8.923***
(3.45)
7.656***
(3.81)
-16.072***
(-5.97)
-4.642**
(-2.17)
Chile
8.220***
(3.27)
7.017***
(3.61)
-16.327***
(-6.24)
-4.433**
(-2.14)
Colombia
8.415***
(3.35)
6.950***
(3.59)
-15.964***
(-6.12)
-4.195**
(-2.03)
Czech
8.641***
(3.37)
7.156***
(3.63)
-16.265***
(-6.12)
-4.226**
(-2.02)
Denmark
8.545***
(3.35)
7.376***
(3.74)
-16.855***
(-6.34)
-5.089**
(-2.42)
Finland
8.383***
(3.28)
7.278***
(3.66)
-16.891***
(-6.33)
-5.071**
(-2.39)
France
9.293***
(3.60)
7.583***
(3.77)
-17.474***
(-6.50)
-4.835**
(-2.25)
Germany
8.139***
(3.19)
6.933***
(3.49)
-16.957***
(-6.40)
-4.923**
(-2.33)
Greece
9.262***
(3.72)
7.902***
(4.15)
-17.375***
(-6.83)
-5.539***
(-2.81)
Hong Kong
8.354***
(3.30)
7.135***
(3.65)
-16.426***
(-6.25)
-4.521**
(-2.18)
Hungary
8.484***
(3.35)
6.975***
(3.59)
-16.480***
(-6.27)
-4.709**
(-2.27)
Iceland
8.730***
(3.32)
6.675***
(3.33)
-16.978***
(-6.22)
-4.463**
(-2.11)
Indonesia
8.824***
(3.50)
7.349***
(3.79)
-16.395***
(-6.26)
-4.800**
(-2.32)
Ireland
8.614***
(3.36)
7.180***
(3.61)
-16.530***
(-6.20)
-4.730**
(-2.23)
Italy
8.095***
(3.16)
6.893***
(3.47)
-16.329***
(-6.13)
-4.286**
(-2.03)
Japan
8.787***
(3.47)
7.251***
(3.71)
-16.866***
(-6.41)
-4.709**
(-2.26)
Latvia
9.360***
(3.57)
7.767***
(3.87)
-17.764***
(-6.53)
-5.584***
(-2.62)
Malaysia
8.593***
(3.38)
7.209***
(3.68)
-16.677***
(-6.31)
-4.750**
(-2.27)
Mexico
8.371***
(3.33)
7.137***
(3.68)
-16.111***
(-6.16)
-4.450**
(-2.15)
Netherlands
8.899***
(3.49)
7.154***
(3.62)
-16.657***
(-6.28)
-4.781**
(-2.27)
New Zealand
8.523***
(3.36)
6.925***
(3.54)
-16.096***
(-6.10)
-4.184**
(-2.01)
Norway
8.696***
(3.44)
7.288***
(3.73)
-15.857***
(-6.03)
-4.007*
(-1.93)
Panama
8.621***
(3.40)
7.390***
(3.77)
-16.266***
(-6.17)
-4.433**
(-2.13)
Peru
8.656***
(3.44)
7.342***
(3.78)
-16.432***
(-6.28)
-4.566**
(-2.21)
Philippines
8.485***
(3.37)
7.085***
(3.65)
-16.277***
(-6.22)
-4.403**
(-2.13)
Poland
9.172***
(3.59)
7.347***
(3.74)
-17.420***
(-6.58)
-5.160**
(-2.46)
Portugal
8.535***
(3.36)
7.428***
(3.77)
-16.437***
(-6.23)
-4.679**
(-2.25)
Singapore
8.447***
(3.34)
7.151***
(3.66)
-16.440***
(-6.26)
-4.661**
(-2.24)
Spain
8.384***
(3.32)
6.907***
(3.54)
-16.245***
(-6.20)
-4.343**
(-2.09)
Sweden
8.489***
(3.32)
7.197***
(3.63)
-16.795***
(-6.31)
-4.888**
(-2.32)
Switzerland
8.947***
(3.46)
7.519***
(3.77)
-17.147***
(-6.38)
-4.955**
(-2.33)
Taiwan
8.575***
(3.38)
7.523***
(3.85)
-16.637***
(-6.32)
-4.903**
(-2.36)
Turkey
7.399***
(2.96)
6.258***
(3.26)
-15.531***
(-5.97)
-3.891*
(-1.91)
UK
8.934***
(3.44)
7.254***
(3.60)
-16.689***
(-6.18)
-4.695**
(-2.18)
US
9.063***
(3.48)
7.662***
(3.75)
-16.618***
(-6.11)
-5.025**
(-2.31)
29
Table 5. Music sentiment and stock market returns at daily frequency
This table reports the daily regression estimates from equation (3) over the sample period from January 1, 2017 to August 28, 2020. The dependent
variable is the daily stock market return (RET). The main independent variable is Music Sentiment, the daily change in the stream-weighted average
valence of the top-200 songs on Spotify, lagged by one to five days. The control variables are the one-to-five-day lagged values of the dependent
variable and the change in deseasonalized cloud cover (
), as well as contemporaneous, next-day, and prior-day daily returns of the MSCI
World index (World RET), daily change in economic policy uncertainty (ΔEPU), daily change in macroeconomic activity (ΔADS), and implied
volatility (VIX). All regressions include country, month, and day-of-the-week fixed effects. Column (1) reports the result of a regression including
the five lags of Music Sentiment, Columns (2) to (6) show the regression for individual lagged value. White-corrected t-statistics are in parentheses.
*, **, and *** denote significance at the 10%, 5% and 1% level, respectively. Constants are not reported. Table A3 provides the variable
definitions.
RETd (%)
(1)
(2)
(3)
(4)
(5)
(6)
Music Sentiment(d-1)
1.883**
(2.04)
1.915**
(2.14)
Music Sentiment(d-2)
0.392
(0.41)
0.132
(0.15)
Music Sentiment(d-3)
-0.241
(-0.26)
-0.3707
(-0.41)
Music Sentiment(d-4)
0.798
(0.86)
1.140
(1.29)
Music Sentiment(d-5)
-2.157**
(-2.32)
-2.447***
(-2.72)
World RET (d+1)
-0.022
(-1.63)
-0.022
(-1.59)
-0.022
(-1.61)
-0.022
(-1.61)
-0.022
(-1.61)
-0.022*
(-1.66)
World RET (d)
0.857***
(62.50)
0.858***
(62.52)
0.858***
(62.52)
0.857***
(62.53)
0.857***
(62.49)
0.858***
(62.60)
World RET (d-1)
0.190***
(17.75)
0.190***
(17.71)
0.190***
(17.68)
0.190***
(17.68)
0.190***
(17.69)
0.190***
(17.69)
RET(d-1)
-0.044***
(-5.72)
-0.044***
(-5.72)
-0.044***
(-5.74)
-0.044***
(-5.72)
-0.044***
(-5.71)
-0.044***
(-5.71)
RET(d-2)
-0.027***
(-4.01)
-0.027***
(-4.01)
-0.027***
(-4.03)
-0.027***
(-4.04)
-0.027***
(-4.03)
-0.027***
(-4.07)
RET(d-3)
-0.004
(-0.64)
-0.004
(-0.63)
-0.004
(-0.62)
-0.004
(-0.62)
-0.004
(-0.63)
-0.004
(-0.65)
RET(d-4)
-0.011*
(-1.70)
-0.011*
(-1.69)
-0.011*
(-1.67)
-0.011*
(-1.70)
-0.011*
(-1.72)
-0.011*
(-1.70)
RET(d-5)
0.005
(0.74)
0.005
(0.75)
0.005
(0.77)
0.005
(0.75)
0.005
(0.73)
0.005
(0.76)
VIX controls
d-1, d, d+1
d-1, d, d+1
d-1, d, d+1
d-1, d, d+1
d-1, d, d+1
d-1, d, d+1
ΔEPU controls
d-1, d, d+1
d-1, d, d+1
d-1, d, d+1
d-1, d, d+1
d-1, d, d+1
d-1, d, d+1
ΔADS controls
d-1, d, d+1
d-1, d, d+1
d-1, d, d+1
d-1, d, d+1
d-1, d, d+1
d-1, d, d+1

controls
d-1, …, d-5
d-1, …, d-5
d-1, …, d-5
d-1, …, d-5
d-1, …, d-5
d-1, …, d-5
Fixed Effects
Country, month, day
25.55%
35,945
Country, month, day
25.53%
36,013
Country, month, day
25.51%
36,013
Country, month, day
25.51%
36,013
Country, month, day
25.52%
36,013
Country, month, day
25.52%
36,013
#Obs.
30
Table 6. Effect of music sentiment on stock market returns and limits to arbitrage
This table reports the regression estimates from equation (4) over the sample period from
January 1, 2017 to August 28, 2020. The dependent variable is the weekly stock market return
(RET). In Panel A, the main independent variable is Music Sentiment, the weekly change in the
stream-weighted average valence of the top-200 songs on Spotify for week t in a country i. The
control variables are the one-week lagged dependent variable (RET(t-1)), weekly return of the
MSCI World index (World RET), contemporaneous implied volatility (VIX), weekly change in
economic policy uncertainty (ΔEPU), weekly change in macroeconomic activity (ΔADS), and
the average daily change in deseasonalized cloud cover over the week (
. BAN is a
dummy variable equal to 1 if country i’s stock market is under a short-selling ban for the full
week t, and 0 otherwise. In Panel B, Music Sentiment and BAN are lagged by one week. All
regressions include country and month fixed effects. White-corrected t-statistics are in
parentheses. *, **, and *** denote significance at the 10%, 5% and 1% level, respectively.
Constants are not reported. Table A3 provides the variable definitions. Table A5 provides the
start and end periods of short-sale bans during the COVID-19 pandemic by country.
RET (%)
Panel A: Contemporaneous
Panel B: One-week lagged
(1)
(2)
(3)
(4)
Music Sentiment
6.332***
(2.54)
6.533***
(3.38)
-14.019***
(-5.45)
-3.006
(-1.48)
Music Sentiment x BAN
106.541***
(3.80)
33.116*
(1.71)
-121.953***
(-4.46)
-83.025***
(-3.66)
BAN
0.097
(0.25)
-0.267
(-0.85)
0.445
(1.22)
-0.1065
(-0.34)
World RET
0.867***
(54.74)
0.867***
(54.49)
VIX
0.001
(0.23)
0.001
(0.24)
ΔEPU
-0.003***
(-6.46)
-0.003***
(-6.42)
ΔADS
0.031
(0.64)
0.005
(0.11)

0.045
(0.53)
0.038
(0.44)
RET(t-1)
-0.037***
(-2.57)
-0.032**
(-2.24)
Fixed Effects
Country, month
Country, month
Country, month
Country, month
3.42%
36.76%
3.82%
36.92%
#Obs.
7,600
7,520
7,560
7,520
31
Table 7. Music sentiment and stock market volatility
This table reports the regression estimates from equation (5) over the sample period from
January 1, 2017 to August 28, 2020. The dependent variable is the weekly stock market
volatility (VOL) obtained as the standard deviation of the daily stock market returns within the
week. The main independent variable is the absolute of |Music Sentiment|, the absolute weekly
change in the stream-weighted average valence of the top-200 songs on Spotify for week t in a
country i. The control variables are the one-week lagged dependent variable (VOL(t-1)), one-
week lagged stock market return (RET(t-1)), contemporaneous implied volatility (VIX), weekly
change in macroeconomic activity (ΔADS), weekly change in economic policy uncertainty
(ΔEPU), average daily change in deseasonalized cloud cover over the week (
, and
weekly return of the MSCI World index (World RET). All regressions include country and
month fixed effects. White-corrected t-statistics are in parentheses. *, **, and *** denote
significance at the 10%, 5% and 1% level, respectively. Constants are not reported. Table A3
provides the variable definitions.
VOL (%)
Without Controls
With Controls
(1)
(2)
|Music Sentiment|
3.836***
(3.67)
1.693**
(2.06)
World RET
-0.049***
(-11.13)
ΔEPU
0.000**
(2.37)
ΔADS
0.008
(0.62)

-0.019
(-0.94)
VOL(t-1)
0.464***
(29.00)
RET(t-1)
-0.028***
(-7.74)
Fixed Effects
Country, month
Country, month
21.48%
41.79%
#Obs.
7,600
7,520
32
Table 8. Music sentiment and net equity mutual fund flows
This table reports the regression estimates from equation (6) over the sample period from
January 1, 2017 to August 28, 2020. The dependent variable is Net Flows, the weekly net fund
flows scaled by the fund’s assets under management in the previous week. The main
independent variable is one-week lagged Music Sentiment, the weekly change in the stream-
weighted average valence of the top-200 songs on Spotify for week t in a country i. The control
variables are the one-week lagged dependent variable (Net Flows (t-1)), one-week lagged stock
market return (RET(t-1)), contemporaneous implied volatility (VIX), weekly change in economic
policy uncertainty (ΔEPU), weekly change in macroeconomic activity (ΔADS), average daily
change in deseasonalized cloud cover over the week (
, and weekly return of the MSCI
World index (World RET). All regressions include fund and month fixed effects. White-
corrected t-statistics are in parentheses. *, **, and *** denote significance at the 10%, 5% and
1% level, respectively. Constants are not reported. Table A3 provides the variable definitions.
Net Flows (%)
Without Controls
With Controls
(1)
(2)
Music Sentiment(t-1)
0.186**
(2.08)
0.228***
(2.57)
World RET
0.768***
(10.28)
VIX
-0.005***
(-26.11)
ΔEPU
0.000***
(-3.39)
ΔADS
0.015***
(7.99)

-0.004
(-0.75)
RET(t-1)
0.204***
(3.21)
Net Flows (t-1)
0.170***
(21.47)
Fixed Effects
Fund, month
Fund, month
0.53%
6.15%
#Obs.
1,432,005
1,413,496
33
Figure 1. Stream-weighted average valence of top-200 songs by geographical regions and country
This figure plots the average daily stream-weighted average valence (SWAV) per country over our sample period from January 1, 2017 to August
28, 2020. The 40 countries in our sample are grouped by geographical regions.
34
Appendix
Table A1: Songs with the highest and lowest valence per country from January 1, 2017 to August 28, 2020
Country
Songs with highest Valence
Songs with lowest Valence
Track.Name
Artist
Valence
Track.Name
Artist
Valence
Argentina
Dame Tu Mano
El Dipy
0.979
Delicate
Taylor Swift
0.050
Australia
September
Earth, Wind and Fire
0.982
Legion Inoculant
TOOL
0.026
Austria
September
Earth, Wind and Fire
0.982
The arrival | Die Ankunft
Claudius Vlasak
0.031
Belgium
September
Earth, Wind and Fire
0.982
Legion Inoculant
TOOL
0.026
Brazil
Matuto de Verdade
Mano Walter
0.981
Memories
Vintage Culture
0.039
Canada
September
Earth, Wind and Fire
0.982
Legion Inoculant
TOOL
0.026
Chile
Tus Ojos Moreno Vide
Hermanos Morales
0.980
Malagradecido
Mon Laferte
0.039
Colombia
Vispera de Año Nuevo
Guillermo Buitrago
0.989
Delicate
Taylor Swift
0.050
Czech
September
Earth, Wind and Fire
0.982
v korunach stromov
Samey
0.011
Denmark
September
Earth, Wind and Fire
0.982
The Ricochet
Dizzy Mizz Lizzy
0.034
Finland
Pohjoiskarjala
Leevi and the Leavings
0.978
Legion Inoculant
TOOL
0.026
France
September
Earth, Wind and Fire
0.982
The Plan
Travis Scott
0.036
Germany
September
Earth, Wind and Fire
0.982
Rodeo Depatro
awais
0.033
Greece
Running Over
Justin Bieber
0.977
Legion Inoculant
TOOL
0.026
Hong Kong
Running Over
Justin Bieber
0.977
The Plan
Travis Scott
0.036
Hungary
September
Earth, Wind and Fire
0.982
Falling Down
ARTY
0.036
Iceland
September
Earth, Wind and Fire
0.982
Legion Inoculant
TOOL
0.026
Indonesia
There's Nothing Holdin' Me Back
Shawn Mendes
0.969
Pizza
Martin Garrix
0.038
Ireland
September
Earth, Wind and Fire
0.982
0.00
Childish Gambino
0.034
Italy
I Puffi sanno
Cristina D'Avena
0.972
DM
Vegas Jones
0.034
Japan
HACK
Shuta Sueyoshi
0.978
Reflection
Brian Eno
0.031
Latvia
Here Comes Santa Claus
Gene Autry
0.976
Sunrise
Coldplay
0.034
Malaysia
Running Over
Justin Bieber
0.977
The Plan
Travis Scott
0.036
Mexico
September
Earth, Wind and Fire
0.982
Renacer
Zoé
0.049
Netherlands
Hop, Hop, Hop, Paardje In Galop
Noord-Hollands Kinderkoor
0.989
Sunrise
Coldplay
0.034
New Zealand
September
Earth, Wind and Fire
0.982
Legion Inoculant
TOOL
0.026
Norway
September
Earth, Wind and Fire
0.982
Mountaineers (feat. John Grant)
Susanne Sundfør
0.033
Panama
Vive Tu Vida Contento
Héctor Lavoe
0.979
Jaded
Drake
0.037
Peru
Ya Vienen Los Reyes Magos
Villancicos
0.978
Delicate
Taylor Swift
0.050
Philippines
Loving you is so easy
HONNE
0.973
Midnight
Coldplay
0.035
Poland
September
Earth, Wind and Fire
0.982
Legion Inoculant
TOOL
0.026
Portugal
Sempre Bem
Capitão Fausto
0.982
Legion Inoculant
TOOL
0.026
Singapore
Running Over
Justin Bieber
0.977
Colour Spectrum
Coldplay
0.034
Spain
Desamortil
Arnau Griso
0.980
Pizza
Martin Garrix
0.038
Sweden
September
Earth, Wind and Fire
0.982
Bethlehems Stjärna
Cappella Snöstorp
0.035
Switzerland
September
Earth, Wind and Fire
0.982
Ouverture
Faber
0.030
Taiwan
Running Over
Justin Bieber
0.977
The Papers
John Williams
0.031
Turkey
Johnny B. Goode
Chuck Berry
0.969
All That Is or Ever Was or Ever Will Be
Alan Silvestri
0.034
UK
September
Earth, Wind and Fire
0.982
0.00
Childish Gambino
0.034
US
September
Earth, Wind and Fire
0.982
Legion Inoculant
TOOL
0.026
35
Table A2: MSCI index considered per country
No
Country
MSCI Index (USD)
MSCI Index (local)
No
Country
MSCI Index (USD)
MSCI Index (local)
1
Argentina
MSARGT$
MSARGTL
21
Japan
MSJPAN$
MSJPANL
2
Australia
MSAUST$
MSAUSTL
22
Latvia
RIGSEIN
RIGSEIN
3
Austria
MSASTR$
MSASTRL
23
Malaysia
MSMALF$
MSMALFL
4
Belgium
MSBELG$
MSBELGL
24
Mexico
MSMEXF$
MSMEXFL
5
Brazil
MSBRAZ$
MSBRAZL
25
Netherlands
MSNETH$
MSNETHL
6
Canada
MSCNDA$
MSCNDAL
26
New Zealand
MSNZEA$
MSNZEAL
7
Chile
MSCHIL$
MSCHILL
27
Norway
MSNWAY$
MSNWAYL
8
Colombia
MSCOLM$
MSCOLML
28
Panama
IFFPNM$
IFFMPAL
9
Czech
MSCZCH$
MSCZCHL
29
Peru
MSPERU$
MSPERU$
10
Denmark
MSDNMK$
MSDNMKL
30
Philippines
MSPHLF$
MSPHLFL
11
Finland
MSFIND$
MSFINDL
31
Poland
MSPLND$
MSPLNDL
12
France
MSFRNC$
MSFRNCL
32
Portugal
MSPORD$
MSPORDL
13
Germany
MSGERM$
MSGERML
33
Singapore
MSSING$
MSSINGL
14
Greece
MSGREE$
MSGREEL
34
Spain
MSSPAN$
MSSPANL
15
Hong Kong
MSHGKG$
MSHGKGL
35
Sweden
MSSWDN$
MSSWDNL
16
Hungary
MSHUNG$
MSHUNGL
36
Switzerland
MSSWIT$
MSSWITL
17
Iceland
ICEXALL
ICEXALL
37
Taiwan
MSTAIW$
MSTAIWL
18
Indonesia
MSINDF$
MSINDFL
38
Turkey
MSTURK$
MSTURKL
19
Ireland
MSEIRE$
MSEIREL
39
UK
MSUTDK$
MSUTDKL
20
Italy
MSITAL$
MSITALL
40
US
MSUSAM$
MSUSAML
36
Table A3: Variables definition and sources
Variable
Description
Source
ADS
U.S. macroeconomic activity index.
Aruoba, Diebold, and Scotti (2009)
BAN
Dummy variable equal to 1 if country’s i stock market is under short-selling ban at
week t, and 0 otherwise.
Yale Program on Financial Stability
DCC
Deseasonalized cloud cover
National Oceanic and Atmospheric
Administration
EPU
News-based measure of U.S. economic policy uncertainty.
Baker et al. (2016)
Music Sentiment
Total change in the stream-weighted average valence of the top-200 songs individuals
of country i listen to in week t.
Spotify
Net Flows (%)
Weekly net flows of an open-end equity mutual fund, scaled by the fund’s assets under
management in the previous week.
Morningstar
RET (%)
Weekly return (Friday-end) of the country’s stock market index. Index values are in
US dollars.
Refinitiv
Valence
The musical positivity conveyed by a song ranging from 0 to 1.
Spotify
VIX
Implied volatility of the S&P 500.
Chicago Mercantile Exchange
VOL (%)
Weekly stock market volatility, measured as the standard deviation of the daily stock
market returns within the week.
Refinitiv
World RET (%)
Weekly return of the MSCI World Index, in US dollars.
Refinitiv
37
Table A4: Music sentiment as a mood proxy at daily frequency
This table reports the regression estimates from the following equation over the sample period
from January 1, 2017 to August 28, 2020:

   
The dependent variable, Daily Music Sentiment, is daily change in stream-weighted average
valence of the top-200 songs on Spotify. The controls are the days of the week and the daily
change in deseasonalized cloud cover (. All regressions include country and month fixed
effects. White-corrected t-statistics are in parentheses. *, **, and *** denote significance at the
10%, 5% and 1% level, respectively. Constants are not reported. Table A3 provides the variable
definitions. All coefficients are multiplied by 100.
Daily Music Sentiment(i,d)
(1)
Calendar-based
mood proxy
(2)
Calendar-based
mood proxy
(3)
Calendar-based +
Weather-induced mood
proxy
Monday(d)
-0.324***
(-35.53)
-0.321***
(-35.33)
Tuesday(d)
-0.057***
(-8.49)
-0.057***
(-8.38)
Thursday(d)
0.023***
(3.46)
0.022***
(3.27)
Friday(d)
0.103***
(10.85)
0.102***
(10.40)
Saturday(d)
0.330***
(41.80)
0.330***
(41.55)
Sunday(d)
-0.350***
(-43.44)
-0.343***
(-42.56)
ΔDCC(d)
-0.034***
(-13.25)
-0.035***
(-14.77)
Fixed Effects
Country, month
Country, month
Country, month
13.84%
53,358
0.71%
52,152
14.15%
52,152
#Obs.
38
Table A5: Short-sale bans in the COVID-19 pandemic
Start and end periods of short-sale bans during the COVID-19 pandemic, from the Yale
Program on Financial Stability.
Country
Begin
End
Austria
18/03/2020
18/05/2020
Belgium
16/03/2020
18/05/2020
France
17/03/2020
18/05/2020
Greece
17/03/2020
18/05/2020
Indonesia
02/03/2020
Still in place as of 28/08/2020
Italy
12/03/2020
18/06/2020
Malaysia
23/03/2020
Still in place as of 28/08/2020
Philippines
15/03/2020
16/04/2020
Spain
12/03/2020
18/05/2020
Taiwan
18/03/2020
18/06/2020
Turkey
28/02/2020
Still in place as of 28/08/2020
ResearchGate has not been able to resolve any citations for this publication.
ResearchGate has not been able to resolve any references for this publication.