Content uploaded by Hisam Sabouni
Author content
All content in this area was uploaded by Hisam Sabouni on Mar 05, 2018
Content may be subject to copyright.
The Rhythm of Markets
Abstract
From the Billboard Hot 100 and Official Charts Company top 100 lists, we create music sentiment
indices that measure several dimensions of emotions through natural language processing tools. The music
sentiment indices are found to be in a long-run equilibrium with existing consumer sentiment indices as
well as act as economically and statistically significant predictors of a number of financial indices. More
precisely, we find that the monthly returns of the Dow Jones, Nasdaq, and S&P 500 are influenced by
our music sentiment indices in the short-run, but fundamental’s rule out in the long-run as financial
markets guide long-term trends in our music sentiment indices. Utilizing these results, we are able to
create trading strategies with our music sentiment indices that out-perform traditional buy-and-hold
strategies in terms of reward and risk.
Introduction
The explosion of online text data has led to an increasing amount of researchers building
web-scrapers to collect data on blogs, tweets, and news to gain insights into the emotions
individuals are expressing (Gilbert and Karahalios 2010; Mishne and Glance 2006). As a result,
many institutions have began compiling lexicons, or dictionaries, of words and their emotional
associations to help measure emotions in text data. Examples of these lexicons include
OpenFinder (OF), Google-Profile of Mood States (GPOMS), and the National Resource
Council of Canada’s Word-Emotion Association lexicon (EmoLex). The OF lexicon classifies
text on a single-dimensional negative to positive scale (usually ranging between [0,1]), with
the more negative words appearing, the more negative a text is classified (closer to 0) and the
more positive words appearing, the more positive a text is classified (closer to 1). GPOMS
was created by Bollen, Mao, and Zeng (2010) as an extension of the Profile of Mood States
(POMS) test, a bipolar scale originally developed by Lorr and Douglas (1971, 1981, and 2003)
to “accurately document the effects of drugs on mood state” by surveying cancer patients.
The POMS exam utilizes 72 adjectives that subjects rate on a 5-point Likter scale to describe
their moods in the previous week. GPOMS extends the POMS to a lexicon of approximately
950 words by “analyzing word co-occurrences in a collection of 2.5 billion 4- and 5-grams
computed by Google in 2006 from approximately 1 trillion word tokens observed in publicly
1
accessible Webpages,” where each word is mapped to six dimensions of sentiment (Bollen,
Mao, and Zeng 2010).
GPOMS maps words to the following dimensions of sentiment: calm, alert, sure, vital, kind,
and happy. While GPOMS has been shown to work well with Twitter data for predicting
financial markets, EmoLex provides a finer level decomposition of sentiment and was created
through a less artificial manner by analyzing the opinions of actual people. More specifically,
EmoLex was developed by the National Research Council of Canada to map “English words
and their associations with eight basic emotions: anger, fear, anticipation, trust, surprise,
sadness, joy and disgust” through use of Amazon Mechanical-Turk workers (Mohammad and
Turney 2013).
The decomposition of sentiment into multiple dimensions has been utilized recently by Bollen,
Mao, and Zeng (2010) to analyze measured sentiment of Twitter feeds that were later used
for predicting the closing direction of the Dow Jones Industrial Average (DJIA). Bollen et al.
compared high level sentiment indices (OpinonFinder positive versus negative sentiment) to
finer sentiment indices (GPOMS) and found that certain dimensions of sentiment, calmness
and happiness, tended to be strongly correlated with the future returns of the DJIA, while
more generalized sentiment indices (OF), tended to have little to no correlation with the
future returns of the DJIA.
Given a particular set of social norms, or customs, an individual is accustomed to within their
social media networks, the information individuals may choose to share within their networks
may be systematically biased as individuals may be pressured to share information in a
manner that is not socially distant from their peers (Akerlof 1997). As a result, individuals
may only choose to share with the public certain dimensions of their emotions while keeping
others private. Hence, analysis of content shared on social media as conducted by Bollen et
al. may not be the best measure of national sentiment. In a more robust analysis of linking
sentiment to financial markets, Schmeling (2009) recently found links between sentiment
and financial markets by analyzing the text of the “Abreast the Market” daily column in
the Wall Street Journal from 1984 to 1999. Tetlock utilized of the Harvard psychosocial
dictionary to decompose the text of the Wall Street Journal gossip column into a number of
emotional categories (that is analyzing the frequency of words associated with emotions) and
found through a vector auto regression (VAR) specification that increases in the frequency of
words associated with bad news corresponded to a decrease in approximately 8.1 basis points
2
on the daily returns of the DJIA within short-term lags of one-day. In some sense, Tetlock
took the approach of measuring sentiment by analyzing the goods individuals consume that
are related to sentiment rather than analyzing the information individuals choose to share
over their social networks. We take an approach similar to that of Tetlock by analyzing a
good that is associated with sentiment: music. Music is a good that has been consumed by
individuals throughout history, going back to biblical times, and has been shown to be linked
with sentiment by numerous psychologists. Individually, this is rather simple to comprehend;
music has the ability to inspire us, make us cry, and even invoke patriotism in us.
In the early days of behavioral economics, psychologist Zullow (1991) was the first to analyze
relationships between music and economic indices. Zullow worked with a group of research
assistants to listen to each of the annual Billboard Hot 40 songs from 1955 to 1981 and record
occurrences of pessimistic rumination, that is “a negative description or evaluation of an
event.” Zullow hypothesized that increases in pessimistic rumination in popular culture would
correspond to consumers becoming increasingly pessimistic about the economy, their future
finances, and consumption, which would therefore affect overall GNP. Zullow’s non-automated
methods of analyzing songs found significant results in forecasting GNP two years in advance.
Since Zullow’s original research, there has been little research in the same direction.
Nearly two-decades passed before another researcher looked into analyzing economic rela-
tionships with sentiment as expressed by music. Pettijohn and Sacco (2009) analyzed the
number one songs on the annual Billboard Hot 100 lists over the periods of 1955 to 2003 in
hopes of finding evidence for changes in lyrics during economically stressful times. Through
the use of a sample of primarily undergraduate students, Pettijohn et al. had each student
read through and rank each of the number one Billboard Hot 100 songs lyrics on 7-point
scales in terms of the meaningfulness of content, the degree of comfort, and romanticness.
Pettijohn et al. also analyzed the Billboard number one songs through use of a text analysis
software that analyzed frequency of words. Their undergraduate survey results did not prove
to be statistically correlated with an economic index they coined the general hard times
measure (GHTM) composed of measures for U.S. unemployment, changes in disposable
personal income, changes in the consumer price index, death rates, birth rates, marriage
rates, divorce rates, suicide rates, and homicide rates. However, they found that frequencies
of words associated with increased social affiliations (use of person pronouns) and financial
issues were in fact statistically linked to their GHTM index. Pettijohn’s results indicate that
3
more computational approaches to analyzing the lyrics of songs may prove to be a more
fruitful approach than simply surveying individuals on their interpretations of songs.
More recently, Maymin (2012) found some effects linking financial market volatility to music
analyzed through computational methods. Maymin compared the square root of beat variance,
that is the standard deviation of automatically estimated number of beats per minute across
songs, of the annual Billboard Top 100 list to the annualized volatility of the S&P 500.
1
Maymin showed some degree of forecastablity of volatility based on the choices of music
individuals were choosing to consume. Maymin argued that songs with higher beat variance
were associated with a higher cognitive cost for listeners, where during happy times (eg. low
volatility market states) individuals would be willing to expend higher cognitive costs to
listen to faster paced music and during unhappy times (eg. high volatility market states),
individuals would only be willing to listen to slower paced music. While Maymin (2012) took
a computational based approach to analyze music on an annual basis, he did not look into
the content within the most popular songs, as we do in this paper.
Beyond linking music to economic indices, psychologists have long studied the effects of music
on emotions with experiments, theories, and empirical investigations. Numerous researchers in
the field of musical psychology have found strong relationships between emotional judgement
and psycho-physical dimensions such as tempo, timbre, and loudness (Behrens and Green
1993; Gabrielsson and Juslin 1996; Gerardi and Gerken 1995). Balkwill and Thompson (1999)
theorized that musically expressed emotion is interpreted by listeners through a combination
of culture-specific and perceptual cues, where perceptual cues are measured through tempo
and lyrical ‘complexity’ of songs and culture-specific cues provide a ‘shared understanding
of representational emotion within a tonal system.’ In other words, the underlying rhythm
within songs is interpretable across cultures whereas the lyrics within songs are interpreted
within specific cultures, where each culture may interpret the words in a given song based
upon cultural norms.
In order to gain a deeper understanding of the music individuals are consuming, we examine
the top 100 songs from the weekly Billboard Hot 100 list and the monthly Official Charts
Company top 100 list on a month-to-month basis. Utilizing natural language processing tools
and the Spotify developer API, we analyze the cultural and perceptual cues expressed by the
1
Maymin (2012) extracted data from EchoNest, which was later acquired by Spotify and is part of the data source’s utilized
in this paper.
4
most popular songs over time to measure the sentiment of the general public.
The top 100 lists are composed of songs with the highest number of plays on public radio, the
highest number of plays through online streaming music services, and the highest amount of
dollar sales. We compile our data on the top 100 songs from the Billboard’s Hot 100, to gain
a sense of music consumption in the United States, and from the Official Charts Company,
to gain a sense of music consumption in the United Kingdom, over the period of January
2000 to December 2016, for a total of 204 monthly data points.
2
Out of a possible 88,800
unique weekly songs on the weekly Billboard Hot 100 we have 6,940 unique songs, where
on average, a song on the Billboard Hot 100 list remains on the list for approximately 12.8
weeks. The number of songs that make it to the top of the list is an even smaller subset, with
only 586 unique songs reaching the top five positions and a mere 212 unique songs reaching
the number one position of the Billboard Hot 100. Interestingly, if a song is ever going to
reach any of the top five positions on the weekly Billboard Hot 100 list, the song will take on
average 8 weeks to reach the top five positions from the time the song enters the Hot 100 list.
These summary statistics highlight the amount of persistence over time in the Billboard Hot
100 weekly list and hence to expand our sample of music consumption we also analyze music
preferences within the United Kingdom. Out of a possible 20,400 unique monthly songs on
the monthly Official Charts Company top 100 lists we have 8,734 unique songs from January
2000 to December 2016. Furthermore, the Official Charts Company monthly top 100 list had
859 songs in our sample to reach the top five positions and 193 songs to reach the top of the
list over the course of our 204 month sample period. Similar to the Billboard Hot 100 list, on
average a song lasts approximately 8 weeks on the Official Charts Company monthly top 100
list.
We construct our sentiment indices using the National Resource Council of Canada’s Word-
Emotion Association lexicon (EmoLex) by analyzing the lyrics of the top 100 songs, from
our Billboard and Official Charts Company lists, and we also work with data compiled from
Spotify, an online music streaming service, to extract perceptual cues of music. Given our
music sentiment indices, we model the dynamics of the sentiment indices with respect to
a number of economic and financial indices (Michigan Consumer Sentiment as well as the
DJIA, S&P 500, and NASDAQ). More precisely, we study the dynamics between our series
through a vector error correction model (VECM) to shed light on how music sentiment
2We ag g rega t e the w eek l y B ill b o ard H o t 100 li s t t o a mon thl y f req u e ncy.
5
Figure 1: A histogram of the distribution of the positions of songs on the Billboard Hot 100 list. On average
songs remain on the Billboard Hot 100 list for approximately 13 weeks. The maximum time a song has
remained on the Billboard Hot 100 list from our sample period ranging from January 2000 to December 2016
was 87 weeks, which was Radioactive by Imagine Dragons.
6
amplifies, leads, or reflects economic/financial market behavior. To our knowledge, this is
the first paper of its kind to analyze the finely decomposed sentiment of music through a
highly computational based approach. Our results suggest that consumer sentiment indices
can be compiled from purely analyzing the music individuals choose to consume, acting
as a complementary measure of sentiment to existing monthly surveys. Furthermore, our
VECM results suggest that short-run fluctuations in a number of U.S. financial indices are
partially driven by consumer sentiment as captured by our music indices and long-run trends
in consumer sentiment are driven by movements in financial markets. Utilizing these results,
we find profitable trading strategies using our music sentiment indices to predict monthly
changes in a number of financial indices (DJIA, S&P 500, and NASDAQ).
Data
Spotify, an online music streaming service with a library of over 70 million songs, utilizes
a plethora of natural language processing tools, music psychologists, and machine learning
experts in order to quantify emotional dimensions of songs. These metrics are later used by
Spotify to support their business by recommending songs, artists, and custom playlists to
their subscribers. Of primary interest to our study is the Spotify Developer API that makes
many of the metrics Spotify compiles to support their business publicly accessible. From the
Spotify developer API, we collect the following metrics: danceability, energy, key, loudness,
and valence of songs to provide insights into cultural and perceptual cues of the top 100 songs.
Danceability, as you may guess, measures “how suitable a track is for dancing” by using
a “combination of musical elements including tempo, rhythm stability, beat strength, and
overall regularity.” Energy “represents a perceptual measure of intensity and activity,” with
music such as death metal having high energy and music such as Bach having low energy.
Key represents the key that a track is played in pitch class notation. Loudness represents the
average track loudness in terms of decibels (dB) and lastly, valence describes the musical
positiveness conveyed by a given track, where “tracks with high valence sound more positive
(e.g. happy, cheerful, euphoric), while tracks with low valence sound more negative (e.g. sad,
depressed, angry).”
We searched through the Spotify API for each song with use of the RSpotify Github repository,
where we first search through the API by artist, then through each of the albums we search
7
for a match with the songs on our top 100 lists using regular expressions. Repeating this
process for each of the weekly Billboard Hot 100 tracks and monthly for the Official Charts
Company top 100 tracks, we are able to extract on average 70 of the top 100 songs for each
month, where some of the historical top 100 songs are not readily available on Spotify. With
the metrics for each song in hand, we create monthly time-series of the top 100 songs by
simply averaging across all songs that were on the Billboard Hot 100 within each month as
well as averaging across the Official Charts Company top 100 list. We then average across
the Billboard Hot 100 metrics and the Official Charts Company top 100 lists to get our final
sentiment indices that represent the sentiment of all of the most popular songs in the United
States and United Kingdom.
Our plots in Figure 2 produce some interesting insights into the music individuals in the
United States and the United Kingdom are choosing to consume over our sample period
from January 2000 to December 2016. Recall that valence of songs represents the degree
of positivity of songs that ranges between 0 and 1, with 1 representing songs with a more
positive emotional polarity and 0 representing songs with a more negative emotional polarity.
3
Interestingly, the average valence of the top 100 songs seems to be in a downward trend,
where songs in the early 2000s seem to have been notably more positive. Along with the top
100 songs becoming more ‘negative’ the energy of tracks seems to also be in a downward trend
over the course of our sample. While the energy and valence of tracks has been falling, the
tempo of the most popular songs each week has remained fairly constant around 120 to 130
beats per minute indicating that the top songs tend to be mostly ‘up-beat,’ but on average
not too fast or too slow. The average duration of songs has also been persistently falling from
an average of 4.2 minutes per track to around 3.75 minutes per track and the top 100 songs
seem to have fairly stable amounts of loudness (around -6.0 dB). Rather surprisingly, even
though the valence and energy of tracks on the top 100 lists has been falling, the danceability
of tracks has been increasing from recent lows in 2008. In addition to the data gathered from
the Spotify developer API, we also generate finer level sentiment indices by analyzing the
lyrics of thousands of songs on our monthly lists.
3
For those readers looking for a set of new songs to enjoy we have included in our appendix a list of the top ten most positive
songs in our sample period as well as a list of the top ten most negative songs in our sample period as indicated by our Spotify
valence data.
8
Figure 2: Plots of the data collected from Spotify on the combination of songs in the weekly Billboard Hot
100 list and the monthly Official Charts Company top 100 list.
EmoLex
EmoLex was compiled by Mohammad and Turney (2013) to analyze “how emotions manifest
themselves in language through words.” Following Plutchik (2001) and Ekman (1992),
Mohammad and Turney focus on eight basic emotions for the formation of the lexicon: joy,
sadness, anger, fear, disgust, surprise, trust and anticipation. The eight basic emotions,
psycoevolutionarily theorized by Plutchik and Ekman, are actually four opposing pairs of
emotions: joy–sadness, anger–fear, trust–disgust, and anticipation–surprise, where an even
larger degree of emotions can be considered as middle-points within the pairs as depicted in
Figure 3.
Mohammad and Turney compiled EmoLex by crowd sourcing small pieces of the task of
associating words with emotions through Amazon’s Mechanical Turk. Each word from the
Macquarie Thesaurus is reviewed by five Mechanical Turk workers and each turker is paid
for each review they complete. The lexicon currently includes approximately 14,000 word
associations and is growing each day with an end target of approximately 40,000 word
associations.
9
Figure 3: Plutchik wheel, originally proposed by Plutchik (1957) summarizes how emotions can be depicted
in a similar manner to that of a color wheel, where more complex emotions can be created from mixing eight
basic emotions together.
10
In order to estimate the emotions of the top 100 songs, we compile the lyrics of each song
from the Genius developer API, where Genius is a company specializing in collecting song
lyrics and gathering musical knowledge. The mapping of lyrics to each of the eight core
emotions measured by EmoLex will allow for a further decomposition of Spotify’s valence
metrics and provide a clearer depiction of the types of emotions that affect may financial
markets. We collect each songs lyrics on the Billboard Hot 100 list and the Official Charts
Company’s lists by searching across the Genius developer API, we find matches for again 70
songs on average per month of the top 100, where older songs are more difficult to track down.
The song lyrics is processed by removal of punctuation and common English stop-words
(terms such as the, is, at, which, on, etc.). We iterate through each song’s lyrics and count
the total number of words that co-occur within each emotion’s dictionary for each month,
across the Billboard Hot 100 lists and Official Charts Company top 100 lists, and normalize
the data by the total number of songs we were able to gather lyrics for each month. For
example, our Anticipation index at time tis computed as follows:
Anticipationt=qNt
i=1 qM
j=1 xi,j,t
Nt
Where, xi,j,t =Y
_
_
]
_
_
[
1if Lyricsi,j,t œEmoLexAnticipation
0if Lyricsi,j,t /œEmoLexAnticipation
Here
Nt
represents the number of songs we were able to collect in month
t
from our Billboard
Hot 100 and Official Charts company lists and
M
represents the total number of words in
song
i
after removal of common English stop-words. Figure 4 depicts the natural log of the
each of our eight emotional series extracted from the top 100 song’s lyrics, which we will be
using throughout the duration of the paper, through the EmoLex dictionary.
Notice that in Figure 4 there is substantial variation in each of the series over time. At a high-
level we can see that words associated with anger and disgust tend to increase post-2008, while
words associated with trust tend to decrease post-2008. In order to get a better understanding
of the behavior of these series, we can analyze there behavior’s around nationally stressful
events that take place over a prolonged period of time. A prime example of multi-period
stressful event for consumers was of course the financial meltdown in 2008. To get a clear
picture of the behavior of the sentiment indices throughout the financial crises, we analyze a
11
Figure 4: Plots of the data collected from finely decomposed sentiment indices created from the lyrics of the
combination of songs in the weekly Billboard Hot 100 list and the monthly Official Charts Company top 100
list through the use of Mohammad and Turney’s (2013) Word-Emotion Association lexicon.
two-year event window around the financial crises and analyze the Z-Scores of each index
during this time.
Financial Crisis
We look at an event window from March 2008 to March 2010 (Depicted in Figure 5), to
analyze the behavior of our music sentiment indices constructed from the lyrics of the top
songs in the United States and United Kingdom during the recent financial crises. Exactly in
the center of our event window, the Standard and Poors 500 bottomed out from its free-fall
in February 2009 and began to recover. As a result, one would expect frequencies of words
associated with negative emotions to be increasing before February 2009 and decreasing after
February 2009. In line with expectations, we see that the index for anticipation, disgust,
sadness, fear, and anger all peaked before the bottoming out the S&P 500 and slowly began
to fall afterwards. These plots indicate that individuals are projecting their current states of
mind into the music they choose to listen to. The sentiment indices depict that during this
difficult time for the global economy words associated with more negative emotional polarity
12
Figure 5: Plots of the Z-Score of the sentiment indices over the period of March 2008 to March 2010. Exactly
in the center of this event window for the recent financial crises is February 2009, which is marked in red to
denote the bottoming out of the Standard and Poor’s 500 index.
seemed to occur much more frequently in the most popular songs than words associated with
more positive emotional polarity.
Overall, we see that the EmoLex sentiment indices compiled from the top 100 songs in the
United States seemed to be fairly responsive to the financial crisis indicating that use of these
indices may be fruitful to in analyzing dynamics of consumer sentiment as well as financial
markets. For further robustness beyond this simple eye-ball econometric event study, we
proceed by comparing each of the sentiment indices with well known consumer sentiment
indices, the University of Michigan Consumer Sentiment Index (UMCSENT) with a formal
econometric study.
Consumer Sentiment Indices
Since 1966, the University of Michigan Consumer Sentiment Index (UMCSENT) has been
estimated monthly through a survey of five questions conducted by random-digit dialing
4
4
In recent years, random-digit dialing methods have become less effective as consumers are tending to avoid phone numbers
they are unfamiliar with.
13
(Curtin 1982). The five equally weighted questions, seek to gauge how individuals feel about
their own personal future financial prospects as well as the future financial prospects of the
nation as a whole. The index was designed to measure changes in attitudes and expectations,
acting as a leading indicator of aggregate economic activity (Mack 1953). Given that the
survey conducted by the University of Michigan only includes questions pertaining to current
and future financial prospects, we can imagine that the UMCSENT index captures a piece
of a consumer’s total sentiment, that is the piece of sentiment relating purely to household
finances, which may in fact be reactive to financial markets rather than predictive of financial
markets.
Over our sample period, consumer sentiment fell substantially throughout the 2008-2009
financial crises, where consumer sentiment seemed to bottom-out. Since the financial crises in
2008-2009, consumer sentiment by UMCSENT has continued to slowly rise back to pre-crises
levels.
Given that our constructed sentiment indices are created by analyzing a good individuals
are choosing to consume, our sentiment indices should reflect a larger portion of consumer
sentiment rather than purely financial sentiment. We first check and analyze the dynamic
relationships between overall consumer sentiment as proxied by our music sentiment indices
and purely financial consumer sentiment as proxied by the UMCSENT index. This robustness
check will show that our music sentiment indices are reflecting financial sentiment along with
other dimensions of consumer sentiment.
Table 1:
Variable ADF Test Statistic
Michigan -0.576
Valence -0.850
Energy -1.106
Danceability 0.092
Temp o -0.141
Loudness -0.248
anticipation 0.304
disgust 0.211
joy 0.199
sadness 0.115
surprise 0.204
trust 0.074
anger 0.289
fear 0.420
14
As indicated by Table 1, given that the critical values for the Augmented Dickey Fuller test
statistic at the 1%, 5%, and 10% levels are respectively -3.46, -2.88, -2.57, we clearly do not
have enough information to reject the null hypothesis of non-stationarity, via the presence
of a unit root. Given that the Spotify data is conveniently scaled between 0 and 1, we
keep this data as is, in level-form, however for the Michigan Consumer Sentiment data and
each of our EmoLex indices, we model the relationships in logarithmic form to smooth out
month-to-month fluctuations. Given that music sentiment is expressed between cultural as
well as perceptual cues, we include the metrics for energy, danceability, tempo, and loudness
across each of our specifications as controls for perceptual cues of music. Our measures for
the cultural cues expressed (proxied by language use) by the top 100 lists will be measured
at a high level in our first specification using the valence metrics from Spotify and at a more
detailed level in our second specification using the compiled EmoLex indices. In general, we
are modeling UMCSENT as follows:
UMCSENTt=–+—C ulturalCuest+◊P erceptualC uest+‘t
Here perceptual cues contains our measures of energy, danceability, tempo, and loudness and
cultural cues in our first specification contains only valence from Spotify and in our second
specification contains our EmoLex indices.
Given the presence of non-stationarity, we estimate the appropriate number of lags of our
dependent and independent variables in a VAR model through use of Akaike Information
Criterion (AIC) and run a Johansen trace test in order to test for cointegration between
our variables of interest. The results of our Johansen trace tests indicate the presence of
cointegration between UMCSENT and our Spotify and EmoLex specifications, meaning some
of our variables can be modeled as linear combinations of our other variables. Utilizing this
information, we model the short-run and long-run dynamics between UMCSENT through a
vector error correction model (VECM). Our model specification is as follows:
UMCSENTt=–0+–(L)UMCSENTt≠1+—(L)CulturalCuest≠1+◊(L)P erceptualC uest≠1
+“ECT +÷t
15
Here the short-run changes in UMCSENT are estimated as a function of the historical changes
in UMCSENT, changes in our cultural cues indices (Spotify valance and EmoLex indices),
changes in our perceptual cues indices (Spotify measures of energy, danceability, tempo, and
loudness), and long-run trends are estimated through
“
, which measures the convergence
back to an error correction term (ECT) derived from the previous equation
5
. Again, we select
the optimal number of lags, for our system of equations through use of the AIC6.
Table 2:
Dependent variable:
UMCSENT.d Valence.d Energy.d Danceability.d Tempo.d Loudness.d
(1) (2) (3) (4) (5) (6)
ect1 ≠0.052úúú 0.021úúú 0.002 0.012úúú
≠2.679úúú 0.169úúú
(0.018) (0.005) (0.004) (0.004) (0.550) (0.063)
constant 0.915úúú
≠0.363úúú
≠0.035 ≠0.209úúú 46.996úúú
≠2.955úúú
(0.318) (0.089) (0.072) (0.063) (9.638) (1.102)
UMCSENT.dl1 ≠0.041 0.004 0.012 0.005 ≠1.163 0.275
(0.071) (0.020) (0.016) (0.014) (2.160) (0.247)
Valence.dl1 0.530ú
≠0.295úúú 0.112 ≠0.022 24.405úúú 0.412
(0.309) (0.087) (0.069) (0.061) (9.364) (1.070)
Energy.dl1 ≠0.460 0.166 ≠0.354úúú 0.083 ≠22.729 1.127
(0.464) (0.130) (0.104) (0.092) (14.059) (1.607)
Danceability.dl1 ≠0.373 ≠0.176 0.021 ≠0.347úúú 0.340 0.181
(0.404) (0.114) (0.091) (0.080) (12.243) (1.399)
Tempo.dl1 0.001 ≠0.00003 ≠0.001 0.0004 ≠0.340úúú
≠0.004
(0.002) (0.001) (0.001) (0.0005) (0.074) (0.008)
Loudness.dl1 0.001 ≠0.011 ≠0.001 ≠0.004 1.353 ≠0.475úúú
(0.029) (0.008) (0.006) (0.006) (0.868) (0.099)
Observations 202 202 202 202 202 202
R20.065 0.162 0.159 0.148 0.159 0.210
Adjusted R20.027 0.128 0.125 0.113 0.125 0.178
Residual Std. Error (df = 194) 0.053 0.015 0.012 0.011 1.604 0.183
F Statistic (df = 8; 194) 1.693 4.691úúú 4.600úúú 4.217úúú 4.601úúú 6.454úúú
Note: VECM, Spotify Data úp<0.1; úúp<0.05; úúú p<0.01
5
The error correction term can be thought of as
ECT
=
UMCSENTt≠
(
ˆ–
+
ˆ
—CulturalCuest
+
ˆ
◊P erceptualCuest
), where
ˆ
·represents the OLS estimate of our coefficients.
6Two lags are selected for our first specification and ten lags are selected for our second specification.
16
Table 3:
Dependent variable:
UMCSENT.d anticipation.d disgust.d joy.d sadness.d surprise.d trust.d anger.d fear.d
(1) (2) (3) (4) (5) (6) (7) (8) (9)
ect1 ≠0.075úúú
≠0.015 ≠0.051 0.025 ≠0.029 ≠0.033 0.018 ≠0.043 ≠0.035
(0.024) (0.039) (0.046) (0.037) (0.043) (0.042) (0.041) (0.044) (0.041)
constant 1.363úúú 0.270 0.923 ≠0.449 0.526 0.594 ≠0.327 0.783 0.641
(0.437) (0.714) (0.838) (0.669) (0.778) (0.770) (0.738) (0.796) (0.748)
UMCSENT.dl1 ≠0.022 0.091 0.020 0.038 ≠0.053 0.068 0.047 0.002 ≠0.007
(0.101) (0.164) (0.193) (0.154) (0.179) (0.177) (0.170) (0.183) (0.172)
anticipation.dl1 ≠0.114 0.386 0.690 0.397 0.455 0.570 0.409 0.550 0.527
(0.261) (0.426) (0.501) (0.399) (0.465) (0.460) (0.440) (0.475) (0.446)
disgust.dl1 0.028 0.043 0.361 0.152 0.284 0.197 0.097 0.329 0.413
(0.213) (0.347) (0.408) (0.325) (0.379) (0.375) (0.359) (0.387) (0.364)
joy.dl1 ≠0.297 ≠0.003 0.087 ≠0.086 ≠0.022 0.112 ≠0.004 0.126 ≠0.056
(0.201) (0.328) (0.385) (0.307) (0.357) (0.354) (0.339) (0.365) (0.343)
sadness.dl1 0.600úú
≠0.528 ≠0.623 ≠0.772ú
≠0.807 ≠0.684 ≠0.515 ≠0.679 ≠0.825ú
(0.273) (0.446) (0.524) (0.418) (0.486) (0.481) (0.461) (0.497) (0.467)
surprise.dl1 ≠0.376 ≠1.030úúú
≠1.770úúú
≠0.831úú
≠1.585úúú
≠1.292úúú
≠1.143úúú
≠1.685úúú
≠1.380úúú
(0.227) (0.370) (0.435) (0.347) (0.404) (0.400) (0.383) (0.413) (0.388)
trust.dl1 0.601úú
≠0.371 0.046 ≠0.338 ≠0.116 ≠0.345 ≠0.327 0.031 ≠0.260
(0.252) (0.411) (0.483) (0.385) (0.448) (0.443) (0.425) (0.458) (0.430)
anger.dl1 ≠0.257 ≠0.374 ≠0.528 ≠0.411 ≠0.206 ≠0.406 ≠0.439 ≠0.406 ≠0.353
(0.245) (0.399) (0.469) (0.374) (0.435) (0.431) (0.413) (0.445) (0.418)
fear.dl1 ≠0.116 0.907ú0.850 1.019úú 1.035úú 0.908ú0.948ú0.873ú0.937ú
(0.285) (0.465) (0.547) (0.436) (0.507) (0.502) (0.481) (0.519) (0.487)
Energy.dl1 ≠1.185 ≠1.050 ≠1.062 ≠1.062 ≠1.613 ≠0.531 ≠0.951 ≠1.242 ≠1.731
(0.758) (1.236) (1.452) (1.158) (1.348) (1.334) (1.278) (1.379) (1.295)
Danceability.dl1 ≠2.037úúú 0.668 ≠0.159 1.399 0.284 0.344 1.244 0.289 0.398
(0.697) (1.137) (1.336) (1.066) (1.240) (1.228) (1.176) (1.268) (1.192)
Tempo.dl1 0.005 0.002 0.0002 ≠0.002 0.004 0.006 0.002 0.003 0.005
(0.004) (0.006) (0.007) (0.006) (0.007) (0.007) (0.006) (0.007) (0.006)
Loudness.dl1 0.062 0.173úú 0.131 0.162úú 0.132 0.113 0.170ú0.113 0.173úú
(0.051) (0.083) (0.097) (0.077) (0.090) (0.089) (0.085) (0.092) (0.086)
Observations 194 194 194 194 194 194 194 194 194
R20.690 0.901 0.880 0.889 0.889 0.894 0.901 0.890 0.902
Adjusted R20.199 0.744 0.691 0.713 0.712 0.725 0.743 0.715 0.748
Residual Std. Error (df = 75) 0.049 0.080 0.094 0.075 0.087 0.086 0.082 0.089 0.083
F Statistic (df = 119; 75) 1.406ú5.739úúú 4.642úúú 5.055úúú 5.023úúú 5.297úúú 5.717úúú 5.096úúú 5.833úúú
Note: VECM, EmoLex Data. úp<0.1; úúp<0.05; úúú p<0.01
Only first lag of variables reported.
Model is actually fit with 10 lags of each variable as selected by Akaike Information Criterion.
17
Table 4:
ECT: Spotify ECT: EmoLex
UMCSENT.l2 11
Valence.l2 -9.790
Energy.l2 6.524
Danceability.l2 -1.604
Tempo.l2 0.100
Loudness.l2 -0.419
anticipation.l10 6.409
disgust.l10 -1.589
joy.l10 -4.390
sadness.l10 3.326
surprise.l10 5.843
trust.l10 -1.604
anger.l10 -4.630
fear.l10 2.525
Energy.l10 -11.687
Danceability.l10 30.885
Tempo.l10 0.0004
Loudness.l10 1.344
The Granger-representation for the first difference of UMCSENT, UMCSENT.d, has a highly
significant and negative coefficient on our error correction term (ECT), indicating the existence
of a non-explosive long-run relationship that UMCSENT reverts back to over time in both
our Spotify and EmoLex specifications as shown in the first columns of Tables 2 and 3 (The
estimated coefficients used to form the ECT by Ordinary Least Squares (OLS) in both the
Spotify and EmoLex specifications are reported in Table 4). In the Spotify specification we
have that if UMCSENT moves away from our ECT, we expect convergence back to the ECT
at a rate of 5
.
2% per month and in the EmoLex specification we have that if UMCSENT
moves away from our ECT, we expect convergence back to the ECT at a rate of 7
.
5% per
month. Hence, a shock to UMCSENT away from the ECT will take approximately 19.2
months to revert back to the ECT in the Spotify specification and approximately 13.3 months
to revert back to the ECT in the EmoLex specification. Notice the significant improvement
in the strength of the equilibrium force by further decomposing the Spotify estimates of
sentiment with the EmoLex indices created directly from the lyrics of the top 100 songs,
in-line with the results of Bollen, Mao, and Zeng (2010). We also see large improvements
in terms of Adj-
R2
from 2
.
7% in the high level Spotify specification to 19
.
9% in the finer
level EmoLex sentiment indices. Furthermore, notice that in the EmoLex specification the
ECT is only statistically significant for UMCSENT in our VECM, indicating that each of
18
the EmoLex indices are not in a long-run equilibrium with the ECT, while UMCSENT is in
fact in a long-run equilibrium with estimated ECT.
Given the presence of cointegration as indicated by the Johansen trace test and our VECM
results, we have some form of Granger causality, which we test for on the first-differences
of our data sets to gain a more detailed level understanding on the directional relationship
between UMCSENT and each of our constructed indices in Tables 5-6.
Table 5:
Valence.d Energy.d Danceability.d Tempo.d Loudness.d
1 Month 0.859 0.589 0.396 0.09740.378
2 Months 0.577 0.898 0.382 0.2542 0.629
3 Months 0.664 0.499 0.570 0.4338 0.575
4 Months 0.696 0.559 0.727 0.6169 0.709
5 Months 0.220 0.673 0.549 0.6218 0.771
6 Months 0.255 0.780 0.555 0.7073 0.890
Granger causality tests for UMCSENT.d on Spotify.d
Pr( > F): p-value < 0.10:*,p-value < 0.05:**,p-value < 0.01:***
Table 6:
Valence.d Energy.d Danceability.d Tempo.d Loudness.d
1 Month 0.905 0.57 0.951 0.898 0.4514
2 Months 0.922 0.2783 0.762 0.826 0.1931
3 Months 0.969 0.1623 0.872 0.887 0.0984
4 Months 0.905 0.09420.963 0.955 0.074
5 Months 0.950 0.2446 0.676 0.920 0.1761
6 Months 0.990 0.2546 0.555 0.556 0.1605
Granger causality tests for Spotify.d on UMCSENT.d
Pr( > F): p-value < 0.10:*,p-value < 0.05:**,p-value < 0.01:***
In line with the VECM results, we find a link between the degree of tempo of songs, which acts
as a form of proxy for more positive songs, and UMCSENT with one lag of tempo Granger
causing UMCSENT. In terms of reverse Granger causality, we have that the UMCSENT
index Granger causes a few lags of the energy and loudness indices as shown in Tables 5-6.
The Granger causality tests utilizing the EmoLex indices indicate that the anticipation,
disgust, joy, sadness, surprise, trust, anger, and fear indices all Granger cause UMCSENT
in first-differences form, while UMCSENT Granger causes a few lags of the trust and anger
sentiment indices in first-differences form (Tables 7-8). In total, these results suggest that
the sentiment of the top 100 songs lyrics precedes the University of Michigan Consumer
Sentiment Index. Given that each of our EmoLex sentiment indices seems to align well with
19
Table 7:
Anticipation.d Disgust.d Joy.d Sadness.d Surprise.d Trust.d Anger.d Fear.d
1 Month 0.06990.05110.1361 0.06450.1591 0.1054 0.07360.0582
2 Months 0.036 0.05070.08510.05150.08490.05180.0449 0.0362
3 Months 0.06790.08080.1527 0.1095 0.1401 0.09740.09130.0717
4 Months 0.0231 0.0291 0.1307 0.0870.0484 0.1222 0.1411 0.0349
5 Months 0.0382 0.0324 0.1884 0.1058 0.08030.1773 0.143 0.0472
6 Months 0.05940.05990.245 0.1563 0.1211 0.2265 0.1719 0.0747
Granger causality tests for UMCSENT.d on EmoLex.d
Pr( > F): p-value < 0.10:*,p-value < 0.05:**,p-value < 0.01:***
Table 8:
Anticipation.d Disgust.d Joy.d Sadness.d Surprise.d Trust.d Anger.d Fear.d
1 Month 0.925 0.842 0.874 0.829 0.770 0.9373 0.869 0.907
2 Months 0.386 0.301 0.604 0.302 0.560 0.4619 0.3655 0.318
3 Months 0.155 0.257 0.232 0.101 0.344 0.09840.08150.144
4 Months 0.273 0.131 0.397 0.102 0.430 0.171 0.060.154
5 Months 0.609 0.110 0.554 0.168 0.727 0.3612 0.07890.310
6 Months 0.721 0.162 0.608 0.295 0.825 0.5098 0.09440.438
Granger causality tests for EmoLex.d on UMCSENT.d
Pr( > F): p-value < 0.10:*,p-value < 0.05:**,p-value < 0.01:***
the Michigan Consumer Sentiment Index, we now proceed to model the relationship between
our EmoLex indices and a number of financial indices.
Financial Markets
We focus on some of the most well known indices in the United States: the DJIA, the S&P
500, and the NASDAQ over our sample period from January 2000 to December 2016.
Given that the causal effects between sentiment and financial markets are rather ambiguous
a priori and we have non-stationary data, we proceed to model the dynamics between our
constructed EmoLex sentiment indices and financial indices through use of a VECM (as
was done for the University of Michigan Consumer Sentiment index)
7
. We utilize AIC to
select the appropriate number of lags for our VECM specification and find that for the
DJIA, NASDAQ, and S&P 500 a VECM of order one is selected. Interestingly, we find that
after correcting our standard-errors for heteroskedasticity and auto-correlation for up to
six-lags through Newey-West (1987) the DJIA, NASDAQ, and S&P 500 have no statistically
7We re p o r t plo t s of hi s tori c al da t a alo n g with G r ang e r cau s a lit y tes t s in th e A p p end i x.
20
significant long-run relationships with our constructed sentiment indices (Results reported in
the first column of Tables 9-11). However, we find significant short-run effects where changes
of frequencies of words associated with anticipation and joy affect the NASDAQ’s returns
as well as the S&P 500’s returns, whereas none of the constructed sentiment indices which
measure cultural cues of music are found to affect the short-run changes of the DJIA’s returns.
An increase of one-percent in the occurrence of words associated with anticipation in the top
100 songs correspond to a 0.444% drop in the NASDAQ and 0.315% drop in the S&P 500.
Furthermore, an increase in the occurrence of words associated with joy by one-percent in
the top 100 songs corresponds to a 0.244% increase in the NASDAQ and 0.165% increase in
the S&P 500. Along with changes in frequency of words in the top 100 songs, we have that
the average danceability of songs affects each of our financial indices in the short-run, where a
one-percent increase in the danceability of songs on the top 100 lists corresponds to a 0.542%
increase in the DJIA, 0.973% increase in the NASDAQ, and 0.576% increase in the S&P 500.
Our estimates of the dynamics between our constructed music sentiment indices and financial
markets demonstrate that fundamental’s rule out in the long-run for financial markets as
sentiment is found to have no significant long-run effects, but short-run movements are in fact
affected by investor sentiment, which we capture through our constructed EmoLex indices.
Furthermore, we find a long-run equilibrium between each of our constructed sentiment
indices and financial markets, that is general trends in financial markets drive the general
trends in our constructed music sentiment indices.
Leveraging our findings of short-run predictability of financial markets through a select few of
our music sentiment indices (Anticipation, Joy, Trust, and Danceability) we develop a trading
strategy using the indices to forecast the DJIA, S&P 500 and the NASDAQ. To simulate
out-of-sample profitability, we adapt the following trading rule on a month-to-month basis
for each of our financial indices:
Buytif —1,tAnticipation +—2,tJoy +—3,tTrust +—4,tDanceability >0otherwise Sellt
Here if a security is purchased at month
t
the security is simulated to be sold at month
t
+1(when deployed in the real-world if the signal received at time
t
+1is the same as
the signal received at time
t
the the position can simply be held to reduce trading costs).
21
Notice the time index on our coefficients that are in place to denote that each time a new
monthly data point arrives (financial index monthly closing price) the estimated coefficients
are re-estimated with the new information through OLS. More precisely, each month we
regress the log-changes of our financial indices at time
t
on the log-changes of the EmoLex
Indices at time
t≠
1and extract the coefficients on Anticipation, Joy, Trust, and Danceability
to form our buy or sell signals (We begin this procedure after the first 100 months of our data
set from January 2000 to October 2008 so that our regressions do not suffer from sample
size issues). We compare our model to a naive trading strategy of buying and holding and
to trading based on the changes of the University of Michigan Consumer Sentiment Index
(UMCSENT), where:
Buytif UMCSENT >0otherwise Sellt
In our simulations, a buy-and-hold strategy for the DJIA over our sample period would have
yielded investors an average annualized return of 3.87%, a UMCSENT trading strategy for
the DJIA would have yielded an average annualized return of 0.14%, and a music sentiment
trading strategy for the DJIA would have yielded an average annualized return of 4.37%.
Assuming an average annualized risk-free rate of 1% across our sample period the three
strategies would have annualized Sharpe ratios of 0.20, -0.05, and 0.30 respectively. Similar
performance is found for the S&P 500 with average annualized returns of 2.87% for buy-
and-hold investors, 0.133% for a UMCSENT trading strategy, and 3.02% for trading with
the music sentiment indices (Sharpe ratios of 0.12, -0.06, and 0.19 respectively). The music
indices trading strategy is also found to out-perform while deployed for the NASDAQ with
average annualized returns of buy-and-hold investors of 3.04%, UMCSENT investors 0.91%,
and music sentiment investors 3.05% (Sharpe ratios of 0.09, -0.006, and 0.13 respectively).
While the music sentiment indices slightly out-perform traditional buy-and-hold methods in
terms of average annualized returns, the ability for the music sentiment indices to predict
short-run fluctuations in financial markets leads to a reduction in volatility of returns.8
8Plots of the cumulative returns of each of the strategies is provided in the appendix.
22
Table 9:
Dependent variable:
LnDJIA.d LnAnticipation.d LnDisgust.d LnJoy.d LnSadness.d LnSurprise.d LnTrust.d LnAnger.d LnFear.d LnUMCSENT.d
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
ect1 ≠0.001 ≠0.037úúú
≠0.037úúú
≠0.034úúú
≠0.038úúú
≠0.038úúú
≠0.038úúú
≠0.037úúú
≠0.038úúú 0.177
(0.001) (0.003) (0.003) (0.002) (0.003) (0.003) (0.003) (0.003) (0.003) (0.114)
constant 0.097 4.729úúú 4.746úúú 4.326úúú 4.917úúú 4.872úúú 4.929úúú 4.821úúú 4.941úúú
≠22.993
(0.151) (0.351) (0.444) (0.300) (0.379) (0.376) (0.360) (0.431) (0.428) (14.588)
LnDJIA.dl1 0.075 0.080 0.109 0.001 0.023 0.062 0.045 0.071 0.041 20.363úúú
(0.069) (0.173) (0.155) (0.149) (0.146) (0.173) (0.174) (0.156) (0.152) (7.533)
LnAnticipation.dl1 ≠0.250 ≠0.144 ≠0.054 ≠0.198 ≠0.173 ≠0.005 ≠0.238 ≠0.171 ≠0.072 25.349úú
(0.153) (0.302) (0.333) (0.250) (0.306) (0.300) (0.333) (0.301) (0.326) (11.213)
LnDisgust.dl1 ≠0.002 0.484úú 0.613úúú 0.477úú 0.555úú 0.592úúú 0.498úú 0.594úú 0.691úúú 13.779
(0.144) (0.241) (0.236) (0.205) (0.232) (0.224) (0.246) (0.235) (0.251) (9.893)
LnJoy.dl1 0.098 ≠0.231 ≠0.290 ≠0.223 ≠0.348 ≠0.287 ≠0.219 ≠0.298 ≠0.494úú
≠8.814
(0.086) (0.233) (0.251) (0.216) (0.228) (0.227) (0.236) (0.237) (0.231) (8.662)
LnSadness.dl1 0.046 ≠0.966úúú
≠1.096úúú
≠0.840úúú
≠1.030úúú
≠0.852úúú
≠0.882úú
≠1.084úúú
≠1.144úúú 0.915
(0.148) (0.325) (0.387) (0.268) (0.338) (0.329) (0.343) (0.350) (0.337) (12.852)
LnSurprise.dl1 ≠0.034 ≠0.316 ≠0.636úú
≠0.239 ≠0.516ú
≠0.358 ≠0.353 ≠0.552ú
≠0.465 ≠18.001úú
(0.122) (0.243) (0.299) (0.199) (0.279) (0.249) (0.241) (0.282) (0.287) (7.789)
LnTrust.dl1 0.188 0.109 0.306 0.152 0.206 0.064 0.195 0.309 0.234 ≠5.208
(0.121) (0.261) (0.265) (0.242) (0.246) (0.256) (0.265) (0.250) (0.253) (11.822)
LnAnger.dl1 ≠0.013 ≠0.382 ≠0.447 ≠0.531úú
≠0.379 ≠0.443 ≠0.421 ≠0.272 ≠0.508 ≠7.006
(0.115) (0.278) (0.303) (0.264) (0.311) (0.303) (0.296) (0.306) (0.311) (12.111)
LnFear.dl1 ≠0.031 0.211 0.351 0.294 0.426 ≠0.002 0.132 0.198 0.445 4.464
(0.155) (0.294) (0.343) (0.251) (0.311) (0.320) (0.285) (0.317) (0.303) (12.653)
LnUMCSENT.dl1 0.0001 0.001 0.001 0.001 ≠0.00004 0.002 0.001 0.00003 0.0004 ≠0.053
(0.001) (0.002) (0.002) (0.001) (0.002) (0.002) (0.002) (0.002) (0.002) (0.078)
Energy.dl1 ≠0.187 0.661 0.343 0.485 0.386 0.732 0.388 0.147 0.016 ≠20.357
(0.280) (0.943) (0.958) (0.805) (0.939) (0.863) (0.862) (0.912) (0.924) (30.804)
Danceability.dl1 0.542úú
≠1.485úúú
≠0.910 ≠1.332úúú
≠1.213úú
≠1.403úú
≠1.233úú
≠1.039ú
≠1.128ú
≠1.720
(0.261) (0.564) (0.634) (0.515) (0.590) (0.588) (0.509) (0.613) (0.587) (23.838)
Tempo.dl1 ≠0.001 ≠0.012úúú
≠0.014úúú
≠0.012úúú
≠0.012úúú
≠0.009úú
≠0.012úúú
≠0.013úúú
≠0.013úúú 0.221
(0.002) (0.004) (0.004) (0.003) (0.004) (0.004) (0.003) (0.004) (0.004) (0.175)
Loudness.dl1 0.001 ≠0.101ú
≠0.129úú
≠0.084ú
≠0.126úú
≠0.106úú
≠0.106ú
≠0.122úú
≠0.098ú
≠0.163
(0.022) (0.061) (0.061) (0.047) (0.059) (0.054) (0.055) (0.059) (0.058) (1.596)
Observations 202 202 202 202 202 202 202 202 202 202
R20.068 0.635 0.596 0.643 0.614 0.639 0.651 0.618 0.620 0.112
Adjusted R2
≠0.012 0.603 0.561 0.612 0.581 0.608 0.621 0.585 0.588 0.036
Residual Std. Error (df = 186) 0.041 0.100 0.112 0.088 0.104 0.103 0.101 0.107 0.106 4.164
F Statistic (df = 16; 186) 0.848 20.184
úúú 17.137úúú 20.927úúú 18.519úúú 20.589úúú 21.693úúú 18.788úúú 18.995úúú 1.469
Note: úp<0.1; úúp<0.05; úúú p<0.01
Newey-West (1987) HAC 6-lag standard errors reported in parentheses.
23
Table 10:
Dependent variable:
LnNASDAQ.d LnAnticipation.d LnDisgust.d LnJoy.d LnSadness.d LnSurprise.d LnTrust.d LnAnger.d LnFear.d LnUMCSENT.d
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
ect1 0.00004 ≠0.004úúú
≠0.004úúú
≠0.004úúú
≠0.004úúú
≠0.004úúú
≠0.004úúú
≠0.004úúú
≠0.004úúú 0.020
(0.0002) (0.0003) (0.0004) (0.0003) (0.0003) (0.0003) (0.0003) (0.0004) (0.0004) (0.013)
constant ≠0.035 3.961úúú 3.959úúú 3.627úúú 4.109úúú 4.088úúú 4.128úúú 4.035úúú 4.139úúú
≠18.938
(0.232) (0.300) (0.382) (0.257) (0.326) (0.315) (0.306) (0.364) (0.364) (12.467)
LnNASDAQ.dl1 0.108úú 0.015 ≠0.012 ≠0.012 ≠0.051 0.015 0.012 ≠0.008 ≠0.028 11.953úú
(0.044) (0.101) (0.112) (0.089) (0.102) (0.106) (0.107) (0.101) (0.104) (5.699)
LnAnticipation.dl1 ≠0.444úú
≠0.161 ≠0.083 ≠0.199 ≠0.186 ≠0.017 ≠0.246 ≠0.190 ≠0.086 22.080ú
(0.218) (0.294) (0.331) (0.251) (0.302) (0.298) (0.330) (0.299) (0.323) (11.494)
LnDisgust.dl1 ≠0.128 0.479úú 0.588úú 0.478úú 0.535úú 0.591úúú 0.499úú 0.581úú 0.677úúú 14.025
(0.210) (0.237) (0.238) (0.206) (0.229) (0.222) (0.245) (0.238) (0.251) (10.182)
LnJoy.dl1 0.244ú
≠0.232 ≠0.288 ≠0.220 ≠0.340 ≠0.287 ≠0.218 ≠0.296 ≠0.489úú
≠10.476
(0.136) (0.237) (0.255) (0.218) (0.230) (0.229) (0.240) (0.242) (0.233) (8.514)
LnSadness.dl1 0.105 ≠0.966úúú
≠1.098úúú
≠0.835úúú
≠1.025úúú
≠0.851úúú
≠0.880úú
≠1.083úúú
≠1.140úúú
≠0.558
(0.175) (0.323) (0.391) (0.267) (0.339) (0.328) (0.343) (0.353) (0.339) (12.810)
LnSurprise.dl1 ≠0.063 ≠0.317 ≠0.630úú
≠0.253 ≠0.525ú
≠0.363 ≠0.362 ≠0.554úú
≠0.473ú
≠14.617úú
(0.152) (0.235) (0.298) (0.195) (0.275) (0.247) (0.240) (0.280) (0.282) (7.439)
LnTrust.dl1 0.322 0.140 0.347 0.171 0.238 0.093 0.222 0.343 0.265 ≠3.451
(0.205) (0.262) (0.269) (0.242) (0.249) (0.257) (0.265) (0.255) (0.256) (12.492)
LnAnger.dl1 0.169 ≠0.389 ≠0.439 ≠0.534úú
≠0.364 ≠0.452 ≠0.430 ≠0.270 ≠0.501ú
≠10.075
(0.155) (0.274) (0.301) (0.267) (0.299) (0.294) (0.293) (0.301) (0.303) (11.248)
LnFear.dl1 ≠0.145 0.212 0.353 0.284 0.414 ≠0.004 0.129 0.195 0.437 8.090
(0.201) (0.296) (0.344) (0.253) (0.311) (0.325) (0.288) (0.320) (0.305) (12.361)
LnUMCSENT.dl1 0.0001 0.001 0.0004 0.001 ≠0.0001 0.002 0.001 ≠0.0001 0.0004 ≠0.052
(0.001) (0.002) (0.002) (0.001) (0.002) (0.002) (0.002) (0.002) (0.002) (0.075)
Energy.dl1 0.533 0.847 0.522 0.635 0.545 0.922 0.575 0.326 0.188 ≠13.677
(0.396) (0.926) (0.951) (0.795) (0.924) (0.848) (0.853) (0.900) (0.911) (31.718)
Danceability.dl1 0.973úúú
≠1.478úúú
≠0.947 ≠1.318úú
≠1.249úú
≠1.388úú
≠1.209úú
≠1.054ú
≠1.150úú 0.987
(0.345) (0.564) (0.616) (0.525) (0.577) (0.581) (0.524) (0.598) (0.574) (22.313)
Tempo.dl1 ≠0.003 ≠0.012úúú
≠0.014úúú
≠0.012úúú
≠0.012úúú
≠0.009úú
≠0.012úúú
≠0.013úúú
≠0.012úúú 0.226
(0.003) (0.004) (0.004) (0.003) (0.004) (0.004) (0.003) (0.004) (0.004) (0.167)
Loudness.dl1 ≠0.077ú
≠0.110ú
≠0.138úú
≠0.091úú
≠0.134úú
≠0.115úú
≠0.115úú
≠0.130úú
≠0.106ú
≠0.532
(0.041) (0.060) (0.060) (0.046) (0.058) (0.053) (0.055) (0.058) (0.057) (1.735)
Observations 202 202 202 202 202 202 202 202 202 202
R20.109 0.634 0.595 0.643 0.615 0.640 0.651 0.618 0.622 0.108
Adjusted R20.033 0.603 0.560 0.612 0.582 0.609 0.621 0.585 0.589 0.031
Residual Std. Error (df = 186) 0.065 0.100 0.112 0.088 0.104 0.103 0.101 0.107 0.106 4.174
F Statistic (df = 16; 186) 1.428 20.169
úúú 17.079úúú 20.953úúú 18.585úúú 20.636úúú 21.646úúú 18.808úúú 19.090úúú 1.410
Note: úp<0.1; úúp<0.05; úúú p<0.01
Newey-West (1987) HAC 6-lag standard errors reported in parentheses.
24
Table 11:
Dependent variable:
LnSP500.d LnAnticipation.d LnDisgust.d LnJoy.d LnSadness.d LnSurprise.d LnTrust.d LnAnger.d LnFear.d LnUMCSENT.d
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
ect1 ≠0.0004 ≠0.038úúú
≠0.038úúú
≠0.035úúú
≠0.040úúú
≠0.039úúú
≠0.040úúú
≠0.039úúú
≠0.040úúú 0.192
(0.001) (0.003) (0.004) (0.002) (0.003) (0.003) (0.003) (0.004) (0.003) (0.121)
constant 0.047 4.634úúú 4.654úúú 4.243úúú 4.823úúú 4.776úúú 4.832úúú 4.728úúú 4.850úúú
≠23.371
(0.167) (0.346) (0.437) (0.296) (0.372) (0.370) (0.357) (0.425) (0.420) (14.541)
LnSP500.dl1 0.120ú0.044 0.059 ≠0.015 ≠0.030 0.031 0.001 0.040 0.005 21.975úúú
(0.071) (0.171) (0.164) (0.151) (0.158) (0.178) (0.177) (0.164) (0.165) (7.015)
LnAnticipation.dl1 ≠0.315ú
≠0.155 ≠0.070 ≠0.201 ≠0.184 ≠0.015 ≠0.249 ≠0.181 ≠0.081 24.345úú
(0.164) (0.299) (0.335) (0.249) (0.306) (0.299) (0.331) (0.303) (0.327) (11.187)
LnDisgust.dl1 ≠0.020 0.471úú 0.597úú 0.466úú 0.533úú 0.579úúú 0.480úú 0.582úú 0.676úúú 15.403
(0.149) (0.238) (0.236) (0.205) (0.230) (0.221) (0.243) (0.235) (0.250) (9.986)
LnJoy.dl1 0.165ú
≠0.223 ≠0.282 ≠0.216 ≠0.341 ≠0.279 ≠0.211 ≠0.290 ≠0.486úú
≠8.584
(0.093) (0.234) (0.251) (0.216) (0.228) (0.227) (0.236) (0.237) (0.231) (8.697)
LnSadness.dl1 0.035 ≠0.966úúú
≠1.098úúú
≠0.836úúú
≠1.027úúú
≠0.851úúú
≠0.880úú
≠1.083úúú
≠1.142úúú
≠0.073
(0.147) (0.325) (0.389) (0.268) (0.339) (0.329) (0.344) (0.351) (0.338) (12.970)
LnSurprise.dl1 ≠0.001 ≠0.309 ≠0.626úú
≠0.238 ≠0.508ú
≠0.352 ≠0.346 ≠0.546ú
≠0.460 ≠17.568úú
(0.128) (0.241) (0.302) (0.198) (0.280) (0.249) (0.241) (0.283) (0.287) (7.612)
LnTrust.dl1 0.177 0.110 0.309 0.151 0.211 0.065 0.198 0.309 0.235 ≠5.691
(0.135) (0.263) (0.269) (0.243) (0.249) (0.259) (0.267) (0.254) (0.255) (12.299)
LnAnger.dl1 0.029 ≠0.375 ≠0.440 ≠0.522úú
≠0.365 ≠0.435 ≠0.410 ≠0.265 ≠0.498 ≠8.762
(0.111) (0.277) (0.302) (0.266) (0.308) (0.301) (0.293) (0.304) (0.309) (11.589)
LnFear.dl1 ≠0.053 0.213 0.355 0.288 0.422 ≠0.003 0.130 0.198 0.442 6.543
(0.155) (0.295) (0.344) (0.251) (0.311) (0.323) (0.287) (0.319) (0.305) (12.649)
LnUMCSENT.dl1 0.00002 0.001 0.001 0.001 0.00000 0.002 0.001 0.0001 0.0005 ≠0.057
(0.001) (0.002) (0.002) (0.001) (0.002) (0.002) (0.002) (0.002) (0.002) (0.077)
Energy.dl1 0.076 0.659 0.346 0.476 0.380 0.728 0.383 0.144 0.011 ≠18.571
(0.294) (0.929) (0.948) (0.797) (0.928) (0.852) (0.850) (0.903) (0.915) (31.172)
Danceability.dl1 0.576úú
≠1.484úúú
≠0.918 ≠1.322úú
≠1.221úú
≠1.399úú
≠1.235úú
≠1.036ú
≠1.127ú
≠0.953
(0.261) (0.558) (0.626) (0.514) (0.583) (0.582) (0.508) (0.606) (0.581) (23.550)
Tempo.dl1 ≠0.001 ≠0.012úúú
≠0.014úúú
≠0.012úúú
≠0.012úúú
≠0.009úú
≠0.012úúú
≠0.013úúú
≠0.013úúú 0.242
(0.002) (0.004) (0.004) (0.003) (0.004) (0.004) (0.003) (0.004) (0.004) (0.173)
Loudness.dl1 ≠0.019 ≠0.101ú
≠0.130úú
≠0.084ú
≠0.126úú
≠0.107úú
≠0.107ú
≠0.122úú
≠0.098ú
≠0.151
(0.025) (0.060) (0.060) (0.047) (0.059) (0.054) (0.055) (0.058) (0.058) (1.647)
Observations 202 202 202 202 202 202 202 202 202 202
R20.082 0.635 0.596 0.643 0.615 0.640 0.652 0.618 0.621 0.122
Adjusted R20.004 0.604 0.562 0.613 0.582 0.609 0.622 0.586 0.589 0.047
Residual Std. Error (df = 186) 0.043 0.100 0.111 0.088 0.104 0.103 0.100 0.107 0.106 4.140
F Statistic (df = 16; 186) 1.045 20.217
úúú 17.183úúú 20.959úúú 18.597úúú 20.630úúú 21.745úúú 18.834úúú 19.082úúú 1.623ú
Note: úp<0.1; úúp<0.05; úúú p<0.01
Newey-West (1987) HAC 6-lag standard errors reported in parentheses.
25
Conclusion
Throughout this paper we have taken a consumption based approach towards measuring
consumer sentiment by analyzing the sentiment expressed in the top 100 songs in the United
States. We gathered data from numerous sources to construct sentiment indices that model
cultural and perceptual cues expressed by music over time. We found that our constructed
indices were in a statistically significant long-run equilibrium with the University of Michigan
Consumer Sentiment Index indicating that the constructed indices are a reasonable proxy for
consumer sentiment. Given that the indices were reasonable proxies for consumer sentiment,
we proceeded to model the dynamics of the music sentiment indices with financial markets
were we also found economically and statistically significant effects, from which we were
able to develop trading strategies that were found to out perform conventional buy-and-hold
strategies. However, our results for financial markets indicate that fundamental’s rule out in
the long-run, as sentiment was found to be a statistically significant predictor for financial
markets only in the short-run. Future research along these lines should be extended by
analyzing other goods that are widely consumed that have emotional attachments to them,
such as movie scripts, television show scripts and perhaps even popular novels.
Bibliography
Akerlof, George A. 1997. “Social Distance and Social Decisions.” Econometrica 65 (5). [Wiley,
Econometric Society]: 1005–27. http://www.jstor.org/stable/2171877.
Balkwill, Laura-Lee, and William Forde Thompson. 1999. “A Cross-Cultural Investigation
of the Perception of Emotion in Music: Psychophysical and Cultural Cues.” Music Percep-
tion: An Interdisciplinary Journal 17 (1). University of California Press Journals: 43–64.
doi:10.2307/40285811.
Behrens, Gene Ann, and Samuel B. Green. 1993. “The Ability to Identify Emotional Content
of Solo Improvisations Performed Vocally and on Three Different Instruments.” Psychology
of Music 21 (1). SAGE Publications Ltd: 20–33. doi:10.1177/030573569302100102.
Bollen, J., H. Mao, and X.-J. Zeng. 2010. “Twitter mood predicts the stock market.” ArXiv
E-Prints,October.
Curtin, Richard T. 1982. “Indicators of Consumer Behavior: The University of Michigan
26
Surveys of Consumers.” Public Opinion Quarterly 46 (3): 340. doi:10.1086/268731.
Ekman, Paul. 1992. “An Argument for Basic Emotions.” Cognition and Emotion 6(3):
169–200.
Gabrielsson, Alf, and Patrik N. Juslin. 1996. “Emotional Expression in Music Performance:
Between the Performer’s Intention and the Listener’s Experience.” Psychology of Music 24
(1): 68–91. doi:10.1177/0305735696241007.
Gerardi, Gina M., and Louann Gerken. 1995. “The Development of Affective Responses
to Modality and Melodic Contour.” Music Perception: An Interdisciplinary Journal 12 (3).
University of California Press Journals: 279–90. doi:10.2307/40286184.
Gilbert, Eric, and Karrie Karahalios. 2010. “Widespread Worry and the Stock Market.”
https://www.aaai.org/ocs/index.php/ICWSM/ICWSM10/paper/view/1513.
Mack, Ruth P. 1953. The American Economic Review 43 (5). American Economic Association:
931–33. http://www.jstor.org/stable/1808745.
Maymin, Philip. 2012. “Music and the Market: Song and Stock Volatility.” The North Ameri-
can Journal of Economics and Finance 23 (1): 70–85. doi:http://dx.doi.org/10.1016/j.najef.2011.11.004.
Mishne, Gilad, and Natalie Glance. 2006. “Predicting Movie Sales from Blogger Sentiment.”
In AAAI Symposium on Computational Approaches to Analysing Weblogs (Aaai-Caaw),
155–58.
Mohammad, Saif M., and Peter D. Turney. 2013. “Crowdsourcing a Word-Emotion Associa-
tion Lexicon” 29 (3): 436–65.
Plutchik, R. 2001. “The Nature of Emotions.” American Scientist 89 (July): 344.
doi:10.1511/2001.4.344.
Schmeling, Maik. 2009. “Investor Sentiment and Stock Returns: Some International Evidence.”
Journal of Empirical Finance 16 (3): 394–408. http://EconPapers.repec.org/RePEc:eee:
empfin:v:16:y:2009:i:3:p:394-408.
Zullow, Harold M. 1991. “Pessimistic Rumination in Popular Songs and Newsmagazines
Predict Economic Recession via Decreased Consumer Optimism and Spending.” Journal of
Economic Psychology 12 (3): 501–26. http://EconPapers.repec.org/RePEc:eee:joepsy:v:12:y:
1991:i:3:p:501-526.
27
Appendix
This section of the appendix is dedicated to additional robustness checks for the paper.
Granger causality tests are reported between the DJIA, NASDAQ, S&P500, and the EmoLex
music sentiment indicies. Furthermore, we decompose the returns of the S&P500 into nine
different sectors to analyze which sectors returns are most influenced by consumer sentiment.
28
Table 12:
LnAnticipation.d LnDisgust.d LnJoy.d LnSadness.d LnSurprise.d LnTrust.d LnAnger.d LnFear.d UMCSENT.d
1 Month 0.709 0.605 0.486 0.556 0.624 0.464 0.600 0.621 0.912
2 Months 0.532 0.496 0.589 0.518 0.404 0.519 0.575 0.456 0.688
3 Months 0.659 0.550 0.476 0.665 0.567 0.653 0.703 0.552 0.924
4 Months 0.780 0.215 0.572 0.434 0.702 0.686 0.696 0.404 0.994
5 Months 0.852 0.291 0.716 0.554 0.746 0.826 0.830 0.546 0.997
6 Months 0.900 0.499 0.881 0.762 0.876 0.944 0.921 0.753 0.995
Granger causality tests for LnDJIA.d on LnEmoLex.d and UMCSENT.d
Pr( > F): p-value < 0.10:*,p-value < 0.05:**,p-value < 0.01:***
Table 13:
LnAnticipation.d LnDisgust.d LnJoy.d LnSadness.d LnSurprise.d LnTrust.d LnAnger.d LnFear.d UMCSENT.d
1 Month 0.988 0.7561 0.783 0.631 0.98 0.9501 0.8033 0.716 0.0148
2 Months 0.121 0.1673 0.499 0.284 0.1354 0.08420.19 0.158 0
3 Months 0.199 0.2293 0.239 0.392 0.1361 0.1088 0.2134 0.293 0
4 Months 0.335 0.3502 0.332 0.568 0.1809 0.2018 0.2908 0.390 0
5 Months 0.148 0.3238 0.127 0.394 0.0670.0249 0.1369 0.354 0
6 Months 0.143 0.07550.149 0.128 0.07070.0399 0.0487 0.154 0
Granger causality tests for LnEmoLex.d and UMCSENT.d on LnDJIA.d
Pr( > F): p-value < 0.10:*,p-value < 0.05:**,p-value < 0.01:***
29
Figure 6: The National Association of Securities Dealers Automated Quotations (NASDAQ) historical price
history from January 2000 to December 2016.
30
Table 14:
LnAnticipation.d LnDisgust.d LnJoy.d LnSadness.d LnSurprise.d LnTrust.d LnAnger.d LnFear.d UMCSENT.d
1 Month 0.636 0.592 0.399 0.517 0.597 0.407 0.501 0.620 0.898
2 Months 0.662 0.457 0.799 0.630 0.637 0.600 0.589 0.525 0.831
3 Months 0.700 0.284 0.615 0.620 0.731 0.684 0.611 0.392 0.905
4 Months 0.717 0.202 0.705 0.565 0.734 0.631 0.562 0.386 0.747
5 Months 0.731 0.233 0.780 0.648 0.677 0.772 0.682 0.429 0.846
6 Months 0.844 0.343 0.856 0.790 0.826 0.881 0.780 0.562 0.850
Granger causality tests for LnNASDAQ.d on LnEmoLex.d and UMCSENT.d
Pr( > F): p-value < 0.10:*,p-value < 0.05:**,p-value < 0.01:***
Table 15:
LnAnticipation.d LnDisgust.d LnJoy.d LnSadness.d LnSurprise.d LnTrust.d LnAnger.d LnFear.d UMCSENT.d
1 Month 0.646 0.2873 0.534 0.232 0.593 0.636 0.456 0.280 0.0137
2 Months 0.430 0.2124 0.856 0.466 0.655 0.287 0.362 0.316 0.0002
3 Months 0.631 0.398 0.706 0.686 0.681 0.462 0.605 0.552 0
4 Months 0.804 0.3483 0.877 0.670 0.791 0.698 0.636 0.481 0
5 Months 0.611 0.2553 0.390 0.486 0.492 0.222 0.362 0.480 0
6 Months 0.442 0.09020.414 0.117 0.308 0.256 0.248 0.198 0.0001
Granger causality tests for LnEmoLex.d and UMCSENT.d on LnNASDAQ.d
Pr( > F): p-value < 0.10:*,p-value < 0.05:**,p-value < 0.01:***
31
Figure 7: The Standard and Poors 500 historical price history from January 2000 to December 2016.
32
Table 16:
LnAnticipation.d LnDisgust.d LnJoy.d LnSadness.d LnSurprise.d LnTrust.d LnAnger.d LnFear.d UMCSENT.d
1 Month 0.824 0.706 0.515 0.670 0.691 0.555 0.675 0.754 0.856
2 Months 0.743 0.569 0.782 0.662 0.642 0.702 0.676 0.574 0.746
3 Months 0.805 0.521 0.616 0.743 0.772 0.826 0.777 0.550 0.932
4 Months 0.872 0.144 0.734 0.438 0.811 0.709 0.651 0.337 0.966
5 Months 0.883 0.186 0.815 0.562 0.815 0.845 0.769 0.429 0.990
6 Months 0.938 0.355 0.944 0.749 0.912 0.931 0.874 0.626 0.964
Granger causality tests for LnSP500.d on LnEmoLex.d and UMCSENT.d
Pr( > F): p-value < 0.10:*,p-value < 0.05:**,p-value < 0.01:***
Table 17:
LnAnticipation.d LnDisgust.d LnJoy.d LnSadness.d LnSurprise.d LnTrust.d LnAnger.d LnFear.d UMCSENT.d
1 Month 0.888 0.759 0.920 0.615 0.8864 0.954 0.9175 0.726 0.0074
2 Months 0.239 0.3229 0.630 0.5585 0.2673 0.1882 0.3379 0.344 0
3 Months 0.330 0.401 0.299 0.6144 0.2467 0.2281 0.3868 0.524 0
4 Months 0.547 0.4957 0.457 0.7625 0.3618 0.4438 0.503 0.552 0
5 Months 0.243 0.3214 0.136 0.3882 0.1114 0.05950.2122 0.384 0
6 Months 0.159 0.05020.142 0.07120.08720.06680.05660.116 0.0001
Granger causality tests for LnEmoLex.d and UMCSENT.d on LnSP500.d
Pr( > F): p-value < 0.10:*,p-value < 0.05:**,p-value < 0.01:***
33
Table 18: Top 10 Songs with Highest Valence from Spotify
Artist Title danceability energy key loudness valence tempo
Luke Bryan All My Friends Say 0.639 0.813 9 -6.877 0.987 117.010
Austin Mahone Featuring Pitbull Mmm Yeah 0.712 0.922 6 -3.902 0.978 125.984
Bloodhound Gang The Bad Touch 0.839 0.720 0 -6.905 0.971 122.969
Janet Jackson Just A Little While 0.773 0.775 0 -3.586 0.970 135.104
The Pussycat Dolls Featuring Timbaland Wait A Minute 0.870 0.853 8 -5.134 0.970 136.467
Juanes Lo Que Me Gusta A Mi 0.688 0.891 4 -5.261 0.968 94.993
Hilary DuffWith Love 0.860 0.938 1 -4.196 0.968 121.993
Olly Murs Featuring Flo Rida Troublemaker 0.761 0.856 0 -3.650 0.968 106.009
Lou Bega Tricky, Tricky 0.858 0.947 0 -3.557 0.967 151.037
Third Eye Blind Never Let You Go 0.735 0.928 4 -6.115 0.966 113.760
Table 19: Bottom 10 Songs with Lowest Valence from Spotify
Artist Title danceability energy key loudness valence tempo
Miguel How Many Drinks? 0.579 0.803 4 -18.812 0.035 86.158
Glee Cast Somewhere 0.274 0.358 1 -9.285 0.036 123.960
My Chemical Romance The Ghost Of You 0.181 0.752 9 -5.519 0.037 144.766
Drake 9 0.700 0.721 2 -7.296 0.038 100.003
T-Pain Featuring Lil Wayne Can’t Believe It 0.685 0.254 10 -12.259 0.038 90.015
Evanescence Featuring Paul McCoy Bring Me To Life 0.277 0.812 4 -7.240 0.039 94.285
The-Dream I Luv Your Girl 0.686 0.322 8 -9.829 0.040 89.958
PnB Rock Selfish 0.638 0.600 1 -5.840 0.040 101.981
Kanye West Waves 0.386 0.587 10 -4.454 0.042 96.072
Pearl Jam Nothing As It Seems 0.241 0.508 9 -9.791 0.046 140.627
34
Figure 8: Comparison of trading profitability strategies used for the DJIA. Looking from the right of the plot,
the top redline is the portfolio value yielded by trading using our music sentiment indices, the middle black
line is the portfolio value of simply buying-and-holding the DJIA, and the bottom blue line is the portfolio
value yielded by trading with the University of Michigan Consumer Sentiment Index. The black vertical line
indicates the beginning of the out-of-sample testing period, where initial data was utilized to calibrate the
trading methods.
36
Figure 9: Comparison of trading profitability strategies used for the Standard and Poors 500 Looking from
the right of the plot, the top redline is the portfolio value yielded by trading using our music sentiment
indices, the middle black line is the portfolio value of simply buying-and-holding the Standard and Poors 500,
and the bottom blue line is the portfolio value yielded by trading with the University of Michigan Consumer
Sentiment Index. The black vertical line indicates the beginning of the out-of-sample testing period, where
initial data was utilized to calibrate the trading methods.
37
Figure 10: Comparison of trading profitability strategies used for the NASDAQ Looking from the right of the
plot, the top redline is the portfolio value yielded by trading using our music sentiment indices, the middle
black line is the portfolio value of simply buying-and-holding the NASDAQ, and the bottom blue line is the
portfolio value yielded by trading with the University of Michigan Consumer Sentiment Index. The black
vertical line indicates the beginning of the out-of-sample testing period, where initial data was utilized to
calibrate the trading methods.
38
Table 20:
Dependent variable:
LnXLB.d LnXLE.d LnXLF.d
(1) (2) (3)
LnAnticipation.dl1 0.928 ≠0.556 1.639
(2.394) (3.056) (2.506)
LnDisgust.dl1 ≠2.945 ≠2.666 ≠2.083
(1.957) (1.724) (1.877)
LnJoy.dl1 0.017 0.449 ≠4.191
(2.392) (2.748) (2.843)
LnSadness.dl1 0.162 ≠5.925úú 3.145
(3.129) (2.860) (2.113)
LnSurprise.dl1 ≠0.597 ≠0.148 ≠2.913ú
(1.483) (2.023) (1.620)
LnTrust.dl1 1.870 0.169 3.225
(2.925) (2.600) (3.161)
LnAnger.dl1 4.005úú 6.465úú 3.830
(1.976) (2.923) (2.564)
LnFear.dl1 ≠1.253 3.461 ≠2.153
(2.308) (2.633) (2.065)
Energy.dl1 ≠4.230ú
≠2.869 1.679
(2.196) (2.069) (1.896)
Danceability.dl1 ≠2.556 ≠0.150 0.242
(1.905) (1.739) (1.747)
Tempo.dl1 1.337 1.064 0.070
(2.541) (1.683) (1.681)
Loudness.dl1 ≠1.779 ≠1.157 0.898
(1.143) (1.152) (1.480)
Chi-Squared Statistic (df = 12; 188) 19.53* 19.017* 20.232*
Observations 200 200 200
R20.228 0.194 0.164
Adjusted R20.160 0.124 0.091
Residual Std. Error (df = 183) 0.399 0.351 0.390
úp<0.1; úúp<0.05; úúú p<0.01
Note: Each column represents a sector of the Standard and Poor’s 500 (Sector
SPDR ETFs), where XLB represents the building sector, XLE represents the
energy sector, and XLF represents the financial sector. Newey-West (1987)
HAC 6-lag standard errors reported in parentheses. Three lags of the dependent
variable are included as controls for within quarter variation. Furthermore, a
dummy variable for the month of January is included.
35
Table 21:
Dependent variable:
LnXLK.d LnXLU.d LnXLP.d
(1) (2) (3)
LnAnticipation.dl1 ≠0.080 2.876 0.254
(2.564) (2.628) (2.419)
LnDisgust.dl1 0.714 1.168 ≠1.312
(1.591) (1.752) (1.803)
LnJoy.dl1 1.894 4.167 ≠0.329
(2.003) (3.281) (2.933)
LnSadness.dl1 ≠0.608 0.513 ≠1.938
(2.439) (3.178) (2.984)
LnSurprise.dl1 ≠0.152 ≠3.436ú
≠0.740
(1.516) (1.756) (1.353)
LnTrust.dl1 ≠0.613 ≠4.398 0.011
(2.320) (3.179) (2.730)
LnAnger.dl1 ≠0.634 ≠0.974 2.386
(2.044) (2.144) (2.289)
LnFear.dl1 0.684 1.859 2.968
(2.162) (2.826) (2.439)
Energy.dl1 ≠1.281 ≠0.393 ≠1.915
(1.792) (2.115) (2.615)
Danceability.dl1 ≠1.657 ≠0.030 ≠2.553
(1.124) (1.416) (1.808)
Tempo.dl1 ≠1.404 0.887 1.033
(1.384) (2.137) (2.316)
Loudness.dl1 ≠0.539 0.249 ≠1.950
(0.778) (1.476) (1.498)
Chi-Squared Statistic (df = 12; 188) 14.881 18.238 24.29**
Observations 200 200 200
R20.198 0.226 0.319
Adjusted R20.128 0.158 0.259
Residual Std. Error (df = 183) 0.333 0.406 0.366
úp<0.1; úúp<0.05; úúú p<0.01
Note: Each column represents a sector of the Standard and Poor’s 500 (Sector
SPDR ETFs), where XLK represents the technology sector, XLU represents
the utilities sector, and XLP represents the consumer staples sector. Newey-
West (1987) HAC 6-lag standard errors reported in parentheses. Three lags of
the dependent variable are included as controls for within quarter variation.
Furthermore, a dummy variable for the month of January is included.
39
Table 22:
Dependent variable:
LnXLI.d LnXLV.d LnXLY.d
(1) (2) (3)
LnAnticipation.dl1 ≠1.143 ≠1.999 ≠2.492
(3.196) (2.476) (3.556)
LnDisgust.dl1 ≠4.169 1.502 ≠2.577ú
(2.749) (2.435) (1.469)
LnJoy.dl1 ≠0.666 ≠3.825 1.983
(2.628) (3.216) (3.564)
LnSadness.dl1 4.201 0.472 ≠1.012
(3.336) (3.006) (3.195)
LnSurprise.dl1 0.079 ≠1.263 1.257
(1.777) (2.141) (2.177)
LnTrust.dl1 2.167 3.409 ≠3.784
(3.077) (3.561) (3.540)
LnAnger.dl1 4.032úú
≠1.707 5.565
(1.636) (3.546) (3.744)
LnFear.dl1 ≠3.042 3.176 2.104
(2.253) (2.204) (2.467)
Energy.dl1 ≠1.573 ≠5.108ú
≠1.437
(3.178) (2.982) (3.697)
Danceability.dl1 ≠2.388 ≠2.432 ≠3.445
(1.679) (1.778) (2.205)
Tempo.dl1 ≠2.197 ≠1.949 6.265
(2.880) (2.948) (4.081)
Loudness.dl1 ≠1.707 ≠2.207ú0.284
(1.252) (1.129) (1.283)
Chi-Squared Statistic (df = 12; 188) 27.153*** 23.701** 31.196***
Observations 200 200 200
R20.206 0.403 0.309
Adjusted R20.137 0.350 0.249
Residual Std. Error (df = 183) 0.417 0.470 0.455
úp<0.1; úúp<0.05; úúú p<0.01
Note: Each column represents a sector of the Standard and Poor’s 500 (Sector
SPDR ETFs), where XLI represents the industrial sector, XLV represents
the health care sector, and XLY represents the consumer discretionary sector.
Newey-West (1987) HAC 6-lag standard errors reported in parentheses. Three
lags of the dependent variable are included as controls for within quarter
variation. Furthermore, a dummy variable for the month of January is included.
40
