1
Section 1: Foundation of the Study
The efficient market hypothesis (EMH) is not a new theory for economic or
financial scholars. Ren, Wu, and Liu (2019) suggested that the EMH has flaws when
measuring information sentiment and how it may help investors gain significant profits.
However, many articles testing EMH validity have suggested that there is truth to the
theory for instances that depend on the variables a researcher may use for analysis
(Leekha, Wadhwa, Jain, & Wadhwa, 2018). The purpose of this study, therefore, was to
examine the relationship between customer Twitter comments, employee Glassdoor
feedback, and successful investing decisions measured by stock prices.
Background of the Problem
The stock market is a fickle instrument for exchanging property for money. The
stock market consists of algorithms and people that inhibit the flow of trends based on
new information. Investors use new information in various forms, from social media
posts to news articles and insider trading in secret places. The effect of this information,
whether it be a positive or negative sentiment, is a broad research subject for academics
and financial practitioners alike (Ren et al., 2019). The creation of the world wide web
brought about a tremendous amount of information flow known today as big data. A
large amount of information growth brought new and more significant social platforms
every day (Pourkhani, Abdipour, Baher, & Moslehpour, 2019). The information from
these social platforms could hold much value for investors to find and use these data for
profit. There are current investment tools that use sentiment analysis on large amounts of
text from social media platforms to increase accurate investment decision-making by
2
18% (Cwynar, Cwynar, & Pater, 2017). However, only 31% of professional financial
investors use sentiment analysis as an investment decision-making tool (Daniel, Neves, &
Horta, 2017). The stock market is an evolving entity, and with it, the analytics necessary
to help with safer investment decision-making.
The social media platform that has the most attention from sentiment analysis
researchers is Twitter. Twitter posts can come from random consumers, news outlets,
influencers, and professional financial advisors (Ramon Saura, Reyes-Menendez, &
Alvarez-Alonso, 2018; Xun & Gou, 2017). The accuracy of sentiment from the different
types of post-relationship to business stock prices varies widely. Zaeem and Barber's
(2019) study sentiment analyzed Twitter posts and company stock price fluctuations with
an 80% accuracy. However, Nisar and Yeung (2017) found no significant correlation
between Twitter posts' sentiment and forecasting different company stock prices.
Other social platforms may offer safer information for investment decision-
making, such as Glassdoor. Chamberlain (2015) is an economist who produced a study
with findings that Glassdoor's 100 best companies to work for portfolio outperformed the
S&P 500 for 5 straight years with an average of 115% higher returns. However, Sheng
(2019) found that using Glassdoor data increased investors' decision-making by 7% to
9% on average. However, sentiment analysis tools may be too risky for potential
enhancement of investment decision-making (Pourkhani et al., 2019; U.S. Securities and
Exchange Commission, 2019). Investors may be losing out on using sentiment analysis,
or they are preventing potential losses.
3
Problem Statement
The U.S. Securities and Exchange Commission (2019) warns investors that
sentiment analysis tools may lead to risky, emotional, or impulsive investment decision-
making. This warning comes despite evidence showing that aided investment decision-
tools such as those used to measure social sentiment can have an accuracy ranging from
70% to 90% for investment decision-making (Ren et al., 2019). The general business
problem was that investors are unsure whether the use of sentiment analysis leads to
better investment decisions. The specific business problem was that some investors may
not know the relationship between customer Twitter comments, employee Glassdoor
feedback, and successful investing decisions measured by business stock prices.
Purpose Statement
The purpose of this quantitative correlational study was to examine the
relationship between customer Twitter comments, employee Glassdoor feedback, and
successful investing decisions measured by business stock prices. The independent
variables were customer comments from Twitter on company profiles, Glassdoor Likert
scale ratings, and employee comments on company profiles. The dependent variable was
successful investing decisions measured by business stock prices. The target population
was archival data from U.S. companies listed on the NASDAQ index. The implications
for social change might include more effective investment decision-making, which may
increase funds for retirements, children's education, or other socially desirable uses.
4
Nature of the Study
I chose a quantitative methodology for this study. A quantitative method is useful
for understanding patterns of connections mathematically between dependent and
independent variables (Bernard, 2013). This quantitative study's foundational data was
from public sources used to predict the relationships between the independent and
dependent variables. A quantitative methodology was appropriate for this study because
parametric measurements are useful for identifying generalized relationships (Allwood,
2012). Researchers using a qualitative methodology usually use inductive analysis and
interviews to uncover experiences and meaning (Castellan, 2010). This study was not
about uncovering experiences or meaning from the variables. Therefore, as this study
aimed to discover relationships between variables, the qualitative or mixed-method
approaches were not appropriate.
Specifically, the correlation design via multiple regression was appropriate for
this study. Researchers often use a correlational design for testing relationships without
assuming cause and effect (Bernard, 2013). A correlation research design focuses on
independent predictor variables and a target-dependent variable, all measured for a
positive or negative trend to determine inference (Tabachnick & Fidell, 2007). If a study
has only two independent variables and measuring the difference of the means is
necessary, then a t test could be employed (Schneider, 2013). However, this study has
two independent variables and a single dependent variable, making a t test unnecessary.
Researchers using an experimental design mostly use ANOVA, ANCOVA, or
MANOVA techniques, and assess for causality; in contrast, this study calls for assessing
5
correlation (Bernard, 2013). The purpose of this study was to measure multiple
independent variables with one dependent variable. Thus, a correlation design used to
measure the relationships between independent variables and a single dependent variable
was appropriate for this study.
Research Question
The research question for this study was: What is the relationship between
customer Twitter comments, employee Glassdoor feedback, and successful investing
decisions measured by business stock prices?
Hypotheses
H0: There is no statistically significant relationship between customer Twitter
comments, employee Glassdoor feedback, and successful investing decisions measured
by business stock prices.
H1: There is a statistically significant relationship between customer Twitter
comments, employee Glassdoor feedback, and successful investing decisions measured
by business stock prices.
Theoretical or Conceptual Framework
I chose the EMH as the theoretical framework for this study. Fama, who
conceptualized this theory in 1965, stipulated that relevant information inherently applies
to stock prices, making them at fair value at all times (Fama, 1965). The EMH’s
development from its debut produced naysayers and hypothesis testing with different
statistical methods to see if all information is genuinely apparent in the current stock
price (Leekha et al., 2018). Fama also proposed that new public information is
6
immediately calculable in the stock price, and beating the market is more about luck than
skill (Leekha et al., 2018). However, if one investor undervalues a stock while another
investor looks at the potential of growth from the same stock, the fair market value is
determined differently (Ding, Mazouz, & Wang, 2019). One investor may be interpreting
or not using specific information to account for a stock's market value, which brings the
validity of the EMH into question (Ding et al., 2019). With a larger quantity and a
different form of measuring information and an ability to influence decision sentiment
from social media could impact investor decisions (Cakra & Distiawan-Trisedya, 2015).
Social media produces a higher quantity of information than was possible in the past.
Depending on the investors' perception of sentiment, it could drive the stock's market
value or price (Pourkhani et al., 2019). If the EMH theory were holding, then the use of
sentiment analysis would not produce better-investing decisions. However, if sentiment
analysis helped predict future stock prices, it would indicate the EMH might not apply in
the cases studied.
Operational Definitions
NASDAQ: The National Association of Securities Dealers Automated Quotations
(NASDAQ) is a trading platform of over 3,100 securities that investors use to buy and
sell electronically (Li, Chan, Ou, & Ruifeng, 2017).
Investment decision-making: Investment decision-making is the choice to buy,
sell, or hold assets (Cao, Ouyang, Li, Li, & Chen, 2018).
7
Artificial intelligence (AI): AI is computerized algorithmic platforms that try to
mimic and enhance human intelligence from decision-making to visual perception
(Ibrahim & Wang, 2019).
Assumptions, Limitations, and Delimitations
Assumptions
According to Sale, Lohfeld, and Brazil (2015), assumptions are prevalent and
necessary for studies involving any trace of data gathering. An assumption in a study
thought to be true even without evidence to prove the assertion is apparent (Castellan,
2010). The primary assumption of this study was that the sentiment data gathered was
from humans and not bots. Another assumption was that the data obtained would be large
enough to provide some inference and validity to the outcome of the study. Finally, I
assumed that the Glassdoor reviews are by people that worked at the company, and
Twitter users understood what company they referred to when providing public
sentiment.
Limitations
The limitations in studies are beyond a researcher's control (Polit & Beck, 2010).
Limitations may affect the study creating a different outcome or misinterpretations of the
findings. The primary limitation of this study was the use of secondary data that may not
be accurate. Another limitation was from using quantitative research that may be too
generalized and misses the potential depth of understanding when using a qualitative
method. The quantitative study can only show correlation and not causation, which was a
limitation. The timeframe and amount of data was another limitation.
8
Delimitations
The researcher chooses delimitations as a form of boundaries to narrow the study
focus (O'Leary, 2017). A delimitation of this study was the focus on only large capital
NASDAQ companies. The readers should account for the delimitation when interpreting
the findings. However, stock prices and sentiment are elements used in many different
country stock market indexes.
Significance of the Study
Value to the Business
The value of this study to investors is that it may help increase knowledge of
sentiment analysis and investing decisions from customers and employees' posts and
feedback. Many previous studies used sentiment analysis, looking at the correlation
between positive or negative sentiment expressed on different social media platforms and
stock prices. Still, the results have no direct conclusive correlation between sentiment and
fluctuations in stock price (Pourkhani et al., 2019). The value to the investor's knowledge
about AI sentiment usage may help with gaining profit or preventing losses.
Contribution to Business Practice
Identifying the importance of sentiment analysis for investor decision-making
could increase optimism and risk-taking. Paraboni, Righi, Vieira, and Silveira (2018)
identified that significant markets from the United States and Germany view sentiment as
a drive for optimism and risk barring investment decision-making. Yadav, Kumar, and
Kumar (2019b) suggested that investors who can decipher sentiment faster than other
investors in the active market when making investment decisions can make a more
9
substantial profit than others. Investors wary of using sentiment analysis AI as a business
practice could prevent losses or gain more profits by having a relationship identifiable
between them.
Implications for Social Change
This study's potential positive social change may improve investment decision-
making, leading to profit, philanthropy, and increasing the cash flow to a local economy.
Increasing knowledge of sentiment analysis effectiveness may improve investors' success
rate and increase shareholder value. The change could lead to the preservation of
retirement savings from IRAs and 401ks that would benefit many people.
A Review of the Professional and Academic Literature
The purpose of this quantitative correlational study was to examine the
relationship between customer Twitter comments, employee Glassdoor feedback, and
successful investing decisions. The independent variables were customer comments from
Twitter on company profiles, Glassdoor Likert scale ratings, and employee comments on
company profiles. The dependent variable was the outcome of investing decisions
measured by stock prices. I reviewed the EMH, which was the theoretical framework of
my study. The theory has multiple lenses from a weak to a strong form, depending on the
different types of information and its potential effect on stock price variances (Jovanovic,
2018). The EMH maintains that a stock price is always currently allocated correctly at
any point in time. No new information will give any investor the ability to gain a
significant profit (Leekha et al., 2018). A market's efficiency as being emerging or mature
plays a part in Fama's EMH theory validity.
10
U.S. Stock Market
The U.S. market is a capital trading platform for new and existing stocks to be
bought and sold (Machlup, 1940; Worthington, 2013). The buying and selling entities are
called stock exchanges or stock markets, where those who have ownership of stocks and
bonds and those who wish to buy stocks can exchange currency for ownership
(Worthington, 2013). The stock market's primary purpose is to allow firms to gain
financing from their stock offering. The firms can eventually gain more capital when
selling additional stocks or splitting the stocks to create a larger quantity of available
stocks to trade (Machlup, 1940). The stock market serves as a resource for businesses by
providing investment to the firms and what to expect in terms of capital cost when
buying, selling, or splitting stocks (Machlup, 1940). Furthermore, stock prices will
depend on the investors' forecast of a company's performance, with the added effect of
supply and demand driving the trading price.
The foresight of investor decision-making is driven by emotional bias at times,
which is subjectively a motivator to expect a company to do well or not (Chinas, 2018).
The stock will then either gain or fall in price, leaving the investor with a potential profit
or loss (Chinas, 2018). The loss and profit potential are emotional drivers, and AI tools
may help or hinder the investment decision process for investors.
This literature review consists of six sections: Purpose statement; hypotheses;
EMH theory; Twitter comments; and Glassdoor Feedback as independent variables,
investment decision-making measured by stock prices as a dependent variable, and the
measurement of variables. The first section addresses the purpose of this ex post facto
11
correlational study, followed by a restatement of the hypotheses. The next section
provides an in-depth analysis and synthesis of EMH theory literature, including other
supporting and contrasting theories. The following section addresses the relevance and
contrasting nature of EMH theory through critical analysis and synthesis of the
independent and dependent variables. A review of the independent and dependent
variables' measurement follows an examination of the independent and dependent
variables. The final section compares and contrasts differing points of view and their
relationship to previous research and findings.
Literature Review Organization and Strategy
The literature review's organization and strategy may help provide readers with an
understanding of the structure of the review content. The next section includes an in-
depth review of the theoretical framework, followed by a review comparing and
contrasting rival theories and measurement methods. The literature review concludes
with a review and explanation of the predictor variables I used in the study.
I used multiple electronic databases for online peer-reviewed journal articles and
other materials on the study's nature. I searched using ProQuest, EBSCO, IEEE, Walden
University Library, Cornell's Arvix.com, and Google Scholar. I searched using Boolean
parameters and peer-review setting for articles that included the terms: sentiment
analysis, EMH, SVM, Naïve Bayes, stock market, stocks, stock price, neural networks,
convolutional neural networks, Monkeylearn, MITSMR Culture 500, sentiment
algorithms, Twitter, Glassdoor.com, political news, news, and investment decision-
making. The literature search output ranged from eight to 980. I found 112 relevant
12
articles published between 2015-2020, and 107 of those articles are peer reviewed. Table
1 provides a depiction of the content from the citations found in the literature review.
Table 1
Literature Review Citations
Type
Citations
Peer-reviewed
% Peer
review
Published
within 5
years
Percent
within 5
years
Journal article
110
107
97%
94
85%
Book
1
0
0%
1
100%
Dissertation
2
0
0%
2
100%
Total
113
107
97%
97
95%
Purpose Statement
The purpose of this quantitative correlational study was to examine the
relationship between customer Twitter comments, employee Glassdoor feedback, and
successful investing decisions measured by business stock prices. The independent
variables were customer comments from Twitter on company profiles, Glassdoor Likert
scale ratings, and employee comments on company profiles. The dependent variable was
successful investing decisions measured by business stock prices. The target population
was archival data from U.S. companies listed on the NASDAQ index. The implications
for social change might include more effective investment decision-making, which may
increase funds for retirements, children's education, or other socially desirable uses.
13
Hypotheses
H0: There is no statistically significant relationship between customer Twitter
comments, employee Glassdoor feedback, and successful investing decisions measured
by business stock prices.
H1: There is a statistically significant relationship between customer Twitter
comments, employee Glassdoor feedback, and successful investing decisions measured
by business stock prices.
Efficient Market Hypothesis
I chose the EMH as the theoretical framework for this study. Fama conceptualized
the theory in his 1965 PhD dissertation, stipulating that relevant information inherently
applies to stock prices, making them fair at all times in a perfectly efficient market. The
development of the theory came from economists trying to explain the randomness of
stock market prices (Jovanovic, 2018). The ideas of stock price movement having a
predictable pattern and how information influences prices were first discussed in 1863 by
a Frenchmen named Regnault and a mathematician in the 1900s who created a
mathematical model to test this idea (Jovanovic, 2018). The model was called Brownian
motion, which led to the formulation of two other models called the random walk and
martingale in the 1960s. These models began the race to theorize stock price changes and
how information may affect such price movements. Fama (1970) used the martingale
model to test a hypothesis that all available information used would produce a zero-sum
profit. Fama's theory has been widely tested for accuracy and debate. The EMH theory
includes new public information, as well. If someone makes an abnormal return from
14
public information, it is luck rather than skill (Leekha et al., 2018). However, many
economists have challenged the EMH and Fama's definition and choice of mathematical
models.
There are extensive debates pertaining to the validity of EMH. Economists Leroy,
Lucas, and Samuelson disputed EMH in the 1970s by analyzing the theoretical proof
showing stock price tendencies to be equal to the randomly fluctuating intrinsic values of
individuals (Jovanovic, 2018). Grossman and Stiglitz in the 1980s postulated that
informationally efficient markets are impossible because all available information cannot
correctly reflect prices (Gillet & Renault, 2019). The perceivable scope of EMH began to
be more debatable, depending on a given economist's data and analysis method, including
the notion that only specific sets of information testable with EMH provides that a set of
information may prove to be true or false. Not all public information is accountable at
any one time to provide inferential validity.
As information comes in droves in the 21st century, the past economist assertions
may have less validity today. Malkiel and Fama's (1970) EMH study indicated that
securities prices fully reflect all publicly available information. However, in 2003, Shiller
found that no study at the time ever supported the EMH linking stock market fluctuations
for the aggregate stock market. The validity and applicability of the theory became a
debatable subject for many years. In 2003, Malkiel and other economists challenged the
EMH to explain diverging market sentiment, including how psychological and behavioral
elements impact stock prices. The evolution of the EMH theory definition and
applicability became a common test ground for economic researchers.
15
The EMH has three differential tiers for the different types of perceivable market
environments: weak, semi-strong, and strong. The weak form of efficiency is the notion
that the past prices of any stock will never predict future prices (Chinas, 2018; Hawaldar,
Rohit, & Pinto, 2017; Nwachukqu & Shitta, 2015). Semi-strong efficiency refers to any
form of public information already allocated in the stock price (Chinas, 2018; de Aragão
Batista, Maia, & Romero, 2018; Nwachukqu & Shitta, 2015). No one can use such public
information to make an egregious profit (Chinas, 2018; de Aragão Batista et al., 2018;
Nwachukqu & Shitta, 2015). Strong form efficiency is when all public and insider trading
information is accountable in the current stock prices, which means there is no way of
gaining a substantial profit because it already reflects all available information
(Alekneviciene, Kviedaraitiene, & Alekneviciute, 2018; Chinas, 2018; Nwachukqu &
Shitta, 2015). The three different EMH elements cover the majority of potential markets
from the formulation of a market to its most mature point in time.
Weak form. The weak form of efficiency is that no past prices of any stock will
ever predict the future price. Hawaldar et al. (2017) tested the weak form of EMH with a
Bahrain stock portfolio of 43 companies and found that seven of those companies' stock
prices were predictable based on their past prices for up to 4 years. Moreover, Kršikapa-
Rašajski and Rankov-Siniša's (2016) study on the Baltic stock exchange, using 1 year of
stock price data, provided similar predictability with past stock prices that also rejects the
weak EMH theory. However, there is a suggestion that most emerging markets may start
with no efficiency. With time and successful growth, however, the market will attain a
present weak form of EMH. (Fama, 1970). Andrianto and Mirza's (2016) correlation
16
study of the Indonesian major stock exchanges with 1 year of stock price data from 2013-
2014 and a random walk test for the weak form of EMH found a strong relationship with
the market having a weak form of efficiency. Andrianto and Mirza suggested that the
market slowly grew more efficient by maturing market practices and a greater amount of
capital investment. Nwachukqu and Shitta (2015) looked at 24 emerging markets stock
prices for 10 years from 2000-2010 from all regions of the world with a random walk
analysis to test the weak form of EMH. Nwachukqu and Shitta found that 18 of the
markets do not have a weak form of efficiency.
Furthermore, these studies were on markets believed to be newer and more
volatile compared to the mature markets of the United States, Japan, Germany, and the
United Kingdom, among others. Nwachukqu and Shitta also measured nine mature
economies from 2000-2010 with a random walk analysis to the weak form of EMH. The
study findings suggested that all nine mature industries have a weak form efficiency, with
at least three of them exhibiting short periods of stock price predictability for a random
number of days that collectively equates to six months out of the ten years of data. The
evidence that a weak form of efficiency can exist in mature and emerging markets brings
the validity of semi-strong market efficacy into question.
Semi-strong. A semi-strong efficiency is that all public information will be
reflected in the real-time stock price, meaning gaining a significant profit is not possible
when using public information. Alekneviciene et al.'s (2018) study on the Baltic
exchange using over 16 years of data to test the semi-strong EMH model via measuring
abnormal returns found that the exchange was inefficient. Alekneviciene et al. (2018)
17
suggested that many investors could make a significant return from the stocks on the
Baltic exchange. de Aragão Batista et al.'s (2018) study of the Brazilian stock exchange
and the semi-strong EMH theory for 2016 during a presidential impeachment found a
correlation with a semi-strong efficient stock market. The efficiency of a stock market
may have different correlations depending on many factors of information and the
market's maturity. Gyamfi, Kyei, and Gill's (2017) study of the African stock market
semi-strong EMH correlation studies from 1974-2016 findings suggested no strong
correlation with semi-strong efficiency in any of the markets. Gymafi et al. also stated
that potential publication bias may be present in the body of research. With the symptom
of potential study bias, present validation, and reliability lessen.
In contrast, Manasseh, Ozuzu, and Ogbuabor's (2016) study of the Nigerian stock
market from 2002-2006 via event study methodology findings on 121 information
releases of events had a relationship to supporting strong efficiency only in 2002.
Moreover, the small capital stocks or penny stocks were affected the most. In contrast,
medium-sized and large capital stocks were not affected (Manasseh et al., 2016). The
different market capitalizations of companies' stocks may be a dividing factor to EMH
validity. However, Ayodele, Anthony, and Adekemi's (2016) study of the Nigerian stock
market with randomly sampled data from 2005-2013 and all small-, medium-, and large-
capitalization stocks in the population found that public information has a relationship to
all sizes of stock capitalization prices and suggested the market has semi-strong
efficiency. The size of stock capitalization and even the overall market's maturity may
have a relationship to EMH validity.
18
Strong form. The strong form of EMH is testing private information with all
other available information. de Aragão Batista et al.'s (2018) study of the Brazilian stock
exchange also tested for the strong form of EMH with findings rejecting the hypothesis
because the information known by insider traders was the same as the public information.
Rani, Kaushal, and Shakir (2019) suggested that Twitter influences business and
corporate information leakage among the Fortune 500 companies because 77% of the
employees are actively tweeting throughout the day.
Evaluating internal information and testing the strong form of EMH is quite
difficult for many researchers to publish findings because of insider trading implications
(Lekovic, 2018). However, Kumari (2019) found a way to measure insider trading
knowledge through bonus announcements from 2014-2019 in 9 large banking companies
in India. The study findings suggested that even with bonus information staying internal
and later being accessible to the public, it has no relationship to stock price variances,
suggesting that the nine Indian businesses' stock prices may have strong efficiency.
Private and public information may not have a different relationship when being used for
investment decision-making.
The EMH is the framework for several correlation studies relating to measuring
investment decisions making by information usage and stock prices. The following
section delves into the relationships between EMH and stock price predictability
measured by sentiment analysis. The review of the EMH framework concludes with a
discussion of the relationship to investment decision-making.
19
Sentiment Theory
The sentiment itself refers to the negative, neutral, or positive belief in context,
either heard or written. According to sentiment theory, many subjective written instances,
either publicly available or privately shared, may affect decision-making (Gao, Ren, &
Zhang, 2020). The theory specific to economics concerns the presence of information
having a negative, neutral, or positive context affecting stock prices. If there is no
information present, the stock prices will revert to the original price (Gao et al., 2020;
Zhou, Xu, & Zhao, 2018). The most recent literature uses sentiment theory to measure
sentiment with a popular social media app called Twitter and measuring the relationship
to stock prices and customer decision-making around the world.
Twitter is a social application for posting up to 280-character messages to the 250
million+ users of the platform (Ibrahim, Wang, Bourne, 2017; Sheng, Amankwah-
Amoah, & Wang, 2017). The use of sentiment classification on social media posts has no
perfect generalized criteria for evaluating subjective sentiment regarding positive,
neutral, or negative expressions (Ibrahim & Wang, 2019). However, the first
computerized sentiment analysis tool debuted in the early 2000s and began the innovation
to generalize the practice of accurately analyzing sentiment.
Sentiment Analysis
Humans have used sentiment analysis when reading or listening to conversations
by mapping words, tones, and contexts to understand sentiment output. Sentiment
analysis methods, techniques, and tools can detect and extract subjective information
such as opinions and attitudes from language (Mäntylä, Graziotin, & Kuutila, 2018). One
20
of the earliest studies of sentiment analysis came from a pre-World War II public opinion
collection on socialism (Mäntylä et al., 2018). However, today the innovations of
artificial intelligence (AI) have brought about algorithm-driven sentiment tools that
involve software development languages to develop and build accurate predictions. An
AI sentiment tool can recognize what is known as unstructured text meaning from typed
human sentences and analyze this type of data on a giant scale (Daniel et al., 2017; Kim,
Jeong, Kim, Kang, & Song, 2016). The AI is trainable when matching analyzed
unstructured text to a list of predefined words determined as being positive or negative
(Ibrahim & Wang, 2019). Once the computer evaluates the information with a predefined
library of positive and negative words, a human is still necessary to train the computer to
distinguish potential sarcasm from negative words that could mean something positive
(Gao et al., 2020; Zhou et al., 2018). The most recent use of sentiment analysis AI is
testing on 280 characters unstructured text from many Twitter posts.
The most recent advancement for sentiment analysis algorithms is significantly
impacting accuracy for depicting human sentiment. A type of sentiment analysis lexicon
technique helps to better evaluate the Twitter post sentiment from many areas such as
retail, finance, and airlines, producing a higher accuracy of sentiment scoring output
(Daniel et al., 2017; Guo, Vargo, Pan, Ding, & Ishwar, 2016; Kim et al., 2016; A.
Yildirim, Sküdarlı, & Zgür, 2016). Kim et al.'s (2016) study showed that Twitter users
express their interests and attitudes about news updates, and 19% of the tweets contain
brand sentiment, both positive and negative. Researchers even used sentiment analysis to
derive relationships between financial market predictions and customer decision-making.
21
Bollen, Mao, and Zeng's (2011) study measuring sentiment from randomly
sampled Dow Jones Industrial Average company Twitter posts and time-series stock
price variances findings suggested they have a significant relationship (Bollen et al.,
2011). A study by Liau and Tan (2014) of 10,895 Tweets about airlines in Malaysia
revealed four main topics: ticket promotions, customer service, flight cancellations, and
pre-booking management as the most sentiment-derivable posts for decision-making.
Liau and Tan's study produced business intelligence on the relationship of these topics to
customer decision-making. Chu and Sung (2015) found a correlation with sentiment
influencing retweeting behaviors. Researchers have even used sentiment analysis output
on different publicly traded companies' profiles to discern a potential relationship
between stock prices and index valuations.
Investors that use professional financial advice can come from television to social
media posts. Daniel et al.'s (2017) study on financial tweet sentiments from experienced
personnel in the financial industry correlated to successful financial decisions for
investors such as profit gains and preventing losses. Daniel et al. also found that 80% of
financial professionals use social media, with only 31% using social media sentiment
information for investment decisions. Even with forecasting projections showing an
18.6% accuracy increase for successful decision-making when sentiment is incorporated
(Cwynar et al., 2017). The professional financial industry may have a cautionary view of
sentiment analysis tools and its potential ability to help with investment decision-making.
The information and sentiment on different companies can have a relationship
with its stock price fluctuations. Public information that directly impacts a company may
22
influence people's decisions to buy and sell stocks (Heston & Sinha, 2017; Nisar &
Yeung, 2017; Yoosin, Rahul, Jie, Seung, 2013). Moreover, people tend to buy company
stocks with a good reputation (Yoosin et al., 2013). One way to know a company's
reputation is by seeing the relationship of sentiment between the company and its
customers. One way to derive such information may be from social media posts that
pertain to the moods that affect customer decision-making.
With enough social media posts projecting a specific sentiment, the probability of
a relationship to stock market performance increases. Bollen et al. (2011) found that
certain moods from Twitter posts could predict the Dow Jones Industrial Average.
However, Nisar and Yeung (2017) found no significant correlation between the sentiment
from Twitter posts and forecasting different company stock prices. The findings of
different studies suggest that sentiment may or may not have a relationship to stock price
fluctuations. Moreover, whether negative, positive, or neutral, the quantity and type of
information may correlate with specific company stock prices (Leekha et al., 2018).
Whether information immediately created by customers on Twitter reflects the current
price has wavering correlations, and no one definitive answer (Nisar & Yeung, 2017). If
there is a relationship from information, it seems to have no or a delayed instance of
predicting stock prices (Heston & Sinha, 2017). The use of sentiment information from
social media platforms is still in the wavering phases of validation and lack of reliability.
Review of Alternative Theories
Noise trader theory. A quick and emotionally driven decision maker is what
some consider to be a noise trader. The opposing theory to EMH is the noise trader
23
theory because of quick asset price changes based on investor behaviors affected by
emotional and knee-jerk reaction trades from new investors (Ding et al., 2019).
Moreover, the noise trader theory is measurable with a positive and negative volume
index that is useful to traders (Ding et al., 2019). These indexes are used by professional
traders as a compass to verify if noise traders are manipulating prices via the volume of
stock movement over a day or longer (Ding et al., 2019). Thus, this theory begins with a
similar approach of using sentiment to aid decision-making.
Noise trading and stock price. Noise traders and stock price fluctuations are a
phenomenon that professional investors believe to be true. Bergsma and Tayal's (2019)
study measures noise trader theory with stock price changes and the potential to gain
large profits. Bergsma and Tayal stated that noise trading has a short effect on the
overvaluing and undervaluing of individual risky stocks. Sinha's (2015) study of the India
NIFTY 50 index and the effects of noise trading from one to five days showed a positive
relationship between noise trading and stock price fluctuations. However, once a specific
time has passed, the stock prices begin to show lower volatility (Bouteska, 2019). The
noise trader theory is another lens for viewing investment decision-making based on
stock price variances. However, the theory is not concerned with the sentiment of the
information, but rather on the specific body of investors driven by emotional responses to
buy and sell stocks.
Prospect theory. The potential negative information of companies may have a
relationship with investors' decision-making. Kahneman and Tversky's (1979) economic
prospect theory suggested that negative news has a prolonged and more significant effect
24
on investment decisions than positive news (Chen, Luo, Liu, & Zhang, 2018). Chen et al.
(2018) used prospect theory as a theoretical lens and sentiment analysis to evaluate the
impacts of negative and positive information on investors' stock purchasing decisions.
Paraboni et al. (2018) completed a study using sentiment analysis of company
information, stock prices, and risk aversion for investor decision-making with a positive
relationship founded in prospect theory. Cao et al., (2018) used prospect theory as a lens
to evaluate stock price jumps by simulating different traders such as noise, day, and long-
term traders and their investment decision-making. Cao et al.'s simulation findings
suggested a relationship between price jumps from the different types of traders and
information they all use to make decisions. Rahayu (2017) focused on 192 students'
investment decision-making when given positive and negative dividend information in
different market environments. The study's findings suggested that in a bull market, both
negative and positive information have a greater relationship to changes in investment
decisions. The prospect theory is not an applicable theory to this study because the
research question does not pertain to the length of time for negative or positive effects
from information on investing decisions.
Adaptive market hypothesis. A newer proposed theory called adaptive market
hypothesis (AMH) is the leading theory that expands on the EMH perimeters. Lo (2004)
developed an alternative theory to the traditional EMH (Lekhal & Ahmed, 2020; Shahid,
Jehanzeb, Abbas, Zubair, & Akbar, 2019). The AMH framework reconciles market
efficiency with principles of evolution, adaptation, competition, and natural selection in
terms of investment decision-making (Lekhal & Ahmed, 2020; Shahid et al., 2019). Lo
25
suggested that investors are rational and choose good or bad investments from market
conditions such as market evolution, financial technologies, and survival of the fittest
practices (Lekhal & Ahmed, 2020; Shahid et al., 2019). When market conditions change
and information technology advances can cause fluctuations in prices, this is an indicator
of the AMH theory in practice (Lekhal & Ahmed, 2020). AMH theory is beyond the
scope of this study because there is a lack of mathematical models available to justify
relationships or causation.
Measurement of Variables
Naïve Bayes
The Bayesian Theorem is a decision algorithm used in many different studies
from statistical theories in biological studies to evaluate sentiment from text. The unique
characteristics of this theorem are that it analyzes each word in a sentence as an
individual variable and uses predefined values for each word to produce an average mean
for the sentences as a whole (Song, Kim, Lee, Kim, & Youn, 2017; Jadon, Bhatia, &
Mishra, 2019). The Bayesian approach has some weaknesses because of stop words such
as and, it, the, and is that affect the accuracy of sentiment analysis output. Studies that use
only Naïve Bayes for sentiment analysis widely varied in accuracy from 50% to 80%
(Song et al., 2017; Jadon et al., 2019). Dey et al. (2020) believed that Naïve Bayes is not
as accurate as other decision-making algorithms, even with enhanced data cleansing and
using AI synergy techniques.
Bayesian Theorem:
𝒫
(
𝐴
|
𝐵
)
=𝒫
(
𝐵
|
𝐴
)
𝑥𝒫
(
𝐴
)
𝒫
(
𝐵
)
26
Support Vector Machine
A support vector machine (SVM) is a decision-making algorithm that can help
predict the meaning and context of a sentence, paragraph, and multiple paragraphs. Ren
(2019) suggested using an SVM to help increase the accuracy of tweets sentiment by
considering the day-of-week effect, which constructs a more reliable and realistic test
with stock prices. The day-of-week effect is one of the most well-known financial
anomalies; it means that the average return on a Monday is much lower than on the other
days of the week. Furthermore, when combined with a stop-loss for Monday's order, the
SVM decision algorithm can yield 89% accuracy when predicting stock prices from
sentiment analysis of written text about companies on the SSE 50 Index in China (Ren et
al., 2019). Attarwala, Dimitrov, and Obeidi (2017) used an SVM to predict the 2012 U.S.
presidential election via 100,000 tweet data against the Iowa Electronic Markets (IEM).
The IEM is an online futures market for practicing market fluctuations resulting from
real-world events such as political outcomes, companies' earnings per share (EPS), and
stock returns (Attarwala et al., 2017). Attarwala et al. use of the SVM model correlation
output predicted ex-President Obama winning reelection based on Twitter comment
sentiments. The study's implied outcome was that Twitter is a valid predictor source for
U.S. presidential election outcomes (Attarwala et al., 2017). Zhang, Li, Ye, Li, and Ngai
(2015) used an SVM to evaluate the sentiment of two different firms' Twitter feeds and
their company stock prices per day and produced a 78% predictive accuracy. Mankar,
Hotchandani, Madhawani, Chidrawar, and Lifna (2018) suggested that SVMs are the
most efficient and feasible model for predicting stock price movement when it comes to
27
tweet sentiments. However, Akba, Medeni, Guzel, and Askerzade's (2020) study showed
that SVMs are not the most effective for predicting cryptocurrency market price
fluctuations from sentiment-analyzed tweets. Gunay (2019) found that cryptocurrency
during a bull market directly correlates with positive information; however, it is the
opposite in a bear market. The use of different methods of algorithms and AI memory
and preprocessors is an attempt to enhance SVMs and others when their accuracy is low
(Sunitha, Joseph, & Akhil, 2019). An SVM is not a perfect decision-making algorithm.
However, it is the best option available for those using sentiment analysis tools.
Neural Network
The development of machines that comprehend and understand emotions is one of
the accomplishments of neural networks. The neural network mimics the human brain's
biology, the neural cells that use the synopsis to enact brain function (Attarwala et al.,
2017). Even today, computer neural networks do not stop with written text
comprehension; they can compute hearing and even visual emotions. Furthermore, neural
network logic employs a long short-term memory (LSTM) that retains already learned
data and continues learning based on what it has in its memory (Sreevidya, Murthy, &
Veni, 2020; S. Yildirim, Jothimani, Kavaklioglu, & Basar, 2019). Briefly, recurrent
neural networks (RNN), LSTM, and convolutional neural networks (CNN)s are
implementations within this study's analysis tool for comprehending sentiment drawn
from Twitter and Glassdoor comments. At the root of these machines is a training set
with a body of words having a predetermined key-value pair showing whether their
meanings are negative, positive, or neutral, along with their synonyms or nearest
28
neighbors to these predefined words mimicking a value closest to a designated sentiment.
Neural network algorithms are useful for enhancing predictive modeling for other
algorithms such as SVMs (Sreevidya et al., 2020; S. Yildirim et al., 2019). The researcher
uses an AI tool to measure sentiment from written text and uses all of the techniques
found to produce higher accuracy.
Term frequency-inverse document frequency. How a computer analysis typed
text is a complex process of cleaning and analyzing data. A term frequency-inverse
document frequency (TF-IDF) produces a key-value pair for each word and its
importance in a document (SIET, 2017). The TF-IDF is the beginning of creating a value
used for comprehending negative or positive connotations (SIET, 2017). The AI is
trained based on a human or, in this case, the researchers showing the program what is
recognizable as a negative or positive comment. The AI will also use icons to determine
if a sentence is positive or negative; for example, a smiley face has a high probability of
positive value (SIET, 2017). The TF-IDF is a cleaning and pre-processing algorithm that
helps enhance decision algorithms.
Lemmatization. Cleaning data with multiple processes is a must for accurate
computational output. An AI algorithm will need as much information from a sentence as
possible to produce an accurate prediction based on the words in a sentence, paragraph,
or an entire page and what they mean in context (Akba et al., 2020; Attarwala et al.,
2017; Jadon et al., 2019). The AI algorithm pre-processing enhancement tool identifies
the part of speech intent for each word, sentence, and perceivable context as negative,
neutral, or positive sentiment (Akba et al., 2020; Attarwala et al., 2017; Jadon et al.,
29
2019). Lemmatization is a useful practice for increasing the accuracy of decision
algorithms such as SVM output, but it is not entirely accurate at all times (Akba et al.,
2020; Attarwala et al., 2017; Jadon et al., 2019). Without the use of lemmatization,
however, the AI may confuse the meaning of sentences and provide incorrect output.
Stop words. Some words can confuse a computer and make them mispredict the
meaning of a sentence. The algorithms for analyzing parts of speech (POS) use massive
computer resources and often miscalculate sentence meaning because of different writing
styles (Akba et al., 2020; Attarwala et al., 2017; Jadon et al., 2019). For the algorithm to
break sentences into useful uniformity, deleting all stop words could increase analysis
and output accuracy (Akba et al., 2020; Attarwala et al., 2017; Jadon et al., 2019). Stop
words such as the, is, at, or which, among many others, are taken out of the analysis
(Akba et al., 2020; Attarwala et al., 2017; Jadon et al., 2019). However, taking out stop
words has the limitation of potentially missing specific context that may lead to
inaccurate analysis and context (Akba et al., 2020; Attarwala et al., 2017; Jadon et al.,
2019). A potential miscalculation can occur when leaving or deleting stops words from a
sentence, which is a conundrum combatted by a process called N-Grams.
N-grams. N-grams can increase the accuracy of a computer's ability to
comprehend the context of a sentence. N-grams break sentences down into individual
words, two-word group grams, three-word group grams, and sentence-grouping grams,
with a particular value calculated for each gram (Akba et al., 2020; Attarwala et al., 2017;
Jadon et al., 2019). The stop words are usually taking out before this process, and the N-
Gram values predict the meaning without needing the stop words. Each grams' value is
30
comparable to a set of predefined values offering a means of analysis for determining
people, places, sentiment properties, and similarities of context, among other aspects
(Akba et al., 2020; Attarwala et al., 2017; Jadon et al., 2019). The N-gram process is an
analysis that benefits decision algorithms based on increasing output accuracy (Akba et
al., 2020; Attarwala et al., 2017; Jadon et al., 2019). However, N-grams can potentially
fall into the trap of miscalculations, causing a decision algorithm to predict with bad data
at times (Akba et al., 2020; Jadon et al., 2019). These N-gram algorithms still increase the
average accuracy of decision algorithms by 10% or more (Akba et al., 2020; Attarwala et
al., 2017). The N-gram pre-processing is an important step for analysis and ensuring the
best accuracy is probable for sentiment prediction.
Stock prices. Stock prices are an integral measurement for investment decision-
making. The measurement of stock prices is done by evaluating each stock's starting day
value and comparing the price at the end of the day as higher or lower (Derakhshan &
Beigy, 2019; Darena, Petrovsky, Prichystal, & Zizka, 2018; Kumar, Kumar, & Yadav,
2017; Teng-Chieh, Zaeem, & Barber, 2019). Time-series analysis is a valuable measure
for stock prices when evaluating minute increments in variances in the price (Darena et
al., 2018). The use of time-series analysis is beneficial in forecasting stock prices
(Derakhshan & Beigy, 2019; Darena et al., 2018; Kumar et al., 2017; Teng-Chieh et al.,
2019). However, this is only necessary for forecasting, which does not fall within the
scope of this study. For this study, a simple evaluation of the stock price's initial starting
day amount and the end-of-day price will be justifiable through a binary process of price
increase or decrease.
31
Independent Variable: Twitter
The beginning of social media. From the beginning of word of mouth to the
creation of the internet, social platforms carry on the necessary marketing process for
many businesses. Social media is a prevalent tool that led to the rise of software
platforms like Twitter, Facebook, and LinkedIn that hold many different opinions from
their 300 million-plus monthly users on products, services, and companies (Asghar,
Khan, Ahmad, Qasim, & Khan, 2017; Pak & Paroubek, 2010). Iankova, Davies, Archer-
Brown, Marder, and Yau (2018) discussed the different uses of social media from
business-to-business (B2B) to business-to-consumer (B2C) marketing models. The
outreach of companies advertising goods like commercials bloomed on Facebook and
Twitter. The result of this form of marketing was instant customer feedback in real-time,
seen by a broad audience that created both a significant opportunity and a problem (Nisar
& Yeung, 2018). Even traditional television news outlets began to change because social
media platforms gave anyone with a smartphone the ability to publish videos, articles,
and opinions viewable on a global scale.
News. The use of news and its psychological effect on traders is an essential topic
for behavioral finance researchers. Among the many studies on news sentiment, Kumar et
al.'s (2017) paper on Bloomberg news stories and their effect on stock price increases and
decreases found a positive correlation. A stock researcher can analyze these news stories
for sentiment and relevance to the company using machine-learning algorithms to detect
negative or positive content. A conventional method of calculating positive or negative
sentiment from findings industry, company, or product-specific news stories and count
32
the number of positive or negative sentiment words in the text (Asghar et al., 2017;
Leekha et al., 2017). The effects of these negative and positive words on the investor's
decision-making are potential evidence of price spikes from days with news about the
company, however less so when the company is not mentioned (Kumar et al., 2017). The
amount of news and sentiment seems to have a relationship with investor decision-
making.
The amount of time information and its sentiment is saturating in the market, the
greater the impact. The length of the news story's effect on the trader's investment
decision-making can depend on the amount and type of sentiment (Leekha et al., 2018).
The news sentiment effect from positive news can affect stock prices for up to one week
(Heston & Sinha, 2017). However, negative news can affect stock returns for up to one
quarter (Heston & Sinha, 2017). People and news authors' expression may positively or
negatively affect the stock market, which is also called bearish or bullish behavior in
stock market terminology (Van de Kauter, Breesh, & Hoste, 2015). Yadav, Kumar, and
Kumar (2019a) study on Twitter sentiment analysis on only objective or fact content in
news sources proved to have inaccuracies and limitations that led to no relationship and
predictability of the Indian NIFTY 50 index price fluctuations.
In contrast, Rani et al.'s (2019) study on the NIFTY 50 index ten-day moving
average predictions based on sentiment analysis with one year of 154,000 tweets found a
significant correlation. News stories from different outlets may influence the inertia of
the Twitter feed. Moreover, the news's effect on a social media platform could enhance
the stock prices' sentiment and effect (Zhang et al., 2015). With more social platforms,
33
housing news and other information could lead to increasing the relationship of sentiment
and stock price fluctuations.
Social media and breaking news. Social media platforms have enhanced
informed decision-making through feedback from many users in the form of negative and
positive sentiments. Bollen et al. (2011) mentioned that 19% of their randomly sampled
tweets contained positive and negative sentiments that affected specific company stock
prices. However, Tamrakar (2016) debunked the theory that stock price fluctuation
prediction could come from Twitter, Facebook, Reddit, or YouTube sentiment by
analyzing two years of data with 180 firms' stock prices and found no significant
relationship. Studying Twitter feedback as an external correlative medium of sentiment is
intensely debatable as a useful tool for investor decision-making (Derakhshan & Beigy,
2019). Moreover, investigating internal employee sentiment simultaneously from
websites like Glassdoor is not overtly dismissed (Derakhshan & Beigy, 2019). The
general reliability of internal and external social feedback sentiments from customers and
workers of companies' relationship vis-a-vis enhancing investor decision-making is in
question. The investor's use of sentiment information may well sway their decision-
making (Asghar et al., 2017). The social media platforms are beneficial for extracting
sentiment and using it for investment decision-making.
Social media and investor sentiment. Investor sentiment on social media comes
from investors speaking specifically about the stock or performance of any industry or
company. Barberis, Shleifer, and Vishny (1998) found investor sentiment to be unreliable
when predicting stock prices. The 1998 study instead found that investors underreact to
34
factual information that involved dividend and earning reports and instead reacted to the
long history of the firms' exceptional or underwhelming performance. Kaplanski, Levy,
Veld, and Veld-Merkoulova (2014) characterized positive sentiment as a relationship of
market returns in the private markets. Babu and Kumar (2015) suggested that positive
sentiments have a more substantial effect on the NSE index, while negative sentiments
had more robust findings in other studies. However, McGurk, Nowak, and Hall (2020)
found a correlation between investor sentiment from social media posts and abnormal
stock returns, using 3000 random sampled tweets for small capital stocks on the NYSE
index, which suggests that small firms have a higher relationship to sentiment than larger
firms. The potential power of social media sentiment may be the new word-of-mouth
marketing technique that has a relationship to customer decision-making, which affects
revenue and perhaps also stock price (Teng-Chieh et al., 2019). Investor sentiment and
the sentiment information derivable may or may not be beneficial for investment
decision-making.
Trust filters. The influencers of people on social media may be one of the best
places to gather information for investment decision-making. Teng-Chieh et al. (2019)
discussed the trustworthiness of social media users by filtering their expertise,
experience, reputation, and authority. The number of discussions on a topic by such
social media users is measurable by their prestige and the number of followers (Teng-
Chieh et al., 2019). A strong following or connection to include the number of networks
structured around the user is how reputation and authority measurements are derived.
Teng-Chieh et al. concluded that influencer posts positively correlate to product sales and
35
stock prices. Teng-Chieh et al. trimmed the users down by 90% in this particular study to
only having influencers that have potentially more value to their users (Zaeem & Barber,
2019). The pre-cleaning of users from those who may hold little value to the overall test
of sentiment results in a more reliable analysis of sentiment and stock buying and selling
decisions. With only 1% of the study data considered reliable for six months from stock
prices and influencers, predicting influencer sentiment with stock prices was at 80%
(Zaeem & Barber, 2019). An investor may be able to use influencer sentiment
information to enhance their investment decision-making.
Individual and Group Accuracy
A single stock relationship to sentiment analysis and its stock price fluctuations
compared to a whole index fluctuation is not equivocal. Garcia-Lopez, Batyrshin, and
Gelbukh (2018) insisted that sentiment analysis is a useful tool for predicting a stock
index price fluctuation, but not a practical choice for individual stocks. Such evidence
from Bollen et al.’s (2011) study found that mood states on Twitter data yield 87.6%
accuracy when predicting the Dow Jones Industrial Average point fluctuations.
Birbeck and Cliff (2017) found that Twitter sentiment feedback on four separate
large technology companies' profiles and their stock price buying and selling automation
produced a 54% accuracy rate. The accuracy drops considerably when only evaluating
single stock prices. Pagolu, Challa, Panda, and Majhi (2016) conducted a similar analysis
with one year's worth of Microsoft tweets and stock prices. Pagolu et al. found an
accuracy of 69% in predicting the Microsoft stock price movement. Groß-Klußmann,
König, and Ebner (2019) found 190 finance experts' tweets with nine years of data
36
evaluated with sentiment analysis on a few individual stocks. They found a 51% accuracy
for predicting future stock prices and investment decision-making. There is a trend from
large datasets and predictions averaging around 50% or a little higher, showing signs of
the probability of profit having the same chances of something like flipping a fair two-
sided coin.
The positive influence on a firm's stock value may also have to do with the
current economic state of a country and the perception of the company on social media.
Social media affects business performance through four channels: social capital, revealed
preferences, social marketing, and corporate social networking (Paniague & Sepana,
2014; Zhang et al., 2015). Many areas from destination planning, consumer confidence,
unemployment claims, and more are all derivable and forecasted from social media posts
(Paniague & Sepana, 2014; Zhang et al., 2015). There is also evaluating the sentiment
from consumers on the different company's social media profiles. Nisar and Yeung
(2018) and Usher, Morales, and Dondio (2019) studies found a relationship between
consumer sentiment and stock market movements in terms of the relationship between
sentiment and the daily opening points with the daily closing points. Zhang et al. (2015)
evaluated the sentiment of Twitter feeds from two different firms (CVS and Walgreens)
that correlate with variance in the stock's prices each day for six months. The Zhang
study findings suggested that there is a stronger correlation from using Twitter feeds than
other after-market predictors for stock prices such as Sec 10k evaluations. The amount of
information an investor can use to help with investment decision-making is vast;
however, the best information is still a well sought-after.
37
At an international conference on sustainable information engineering technology,
the main topic was the use of sentiment analysis on consumer engagement with social
media (Saragih, & Girsang, 2017). The conference derived value from sentiment analysis
on Facebook and Twitter comments from transportation companies like Uber to see
where customer engagements are taking place. It appears that there are more negative
comments to include more coming from employees in this industry, and different
companies have different values of engagement from different social media platforms
(Saragih, & Girsang, 2017). There are areas of the Saragih, and Girsang (2017) study that
would create a more substantial quantitative value by adding a Likert-scale to measure
the relationship of these companies with customers' and employees' points of view. A
combination of consumer and employee analysis could be beneficial information for
investor decision-making.
Political Sentiment
Social media has a strong presence in political campaign outcomes and the
relationship to stock and index fluctuations. Nisar and Yeung (2018) explored the
relationship between related political sentiment from over 60,000 Twitter comments from
a U.K. 2016 local election and the FTSE 100 index performance. The study involves
measuring FTSE 100 price movements and evaluating correlation via multiple regression
analysis by comparing daily moods with daily stock price changes. The findings
suggested that there is a correlation between the mood of the public and investment
behavior in the short term; however, the relationship was determined to be statistically
insignificant (Nisar & Yeung, 2018). Usher et al. (2019) study looking at BREXIT and
38
political sentiment relationship on the FTSE 100 fluctuations found that certain words
had a short-term causal effect on index performance.
Twitter post data and sentiment analysis may offer promising analysis results.
Oyebode and Orji (2019) analyzed 118,000 Twitter posts about the Nigerian presidential
election, trying to predict the outcome one month before the last votes were submitted.
The study findings from sentiment analysis found that a single candidate had an 81%
chance of winning the election, and the study's prediction was correct. Moreover, Salari,
Sedighpour, Vaezinia, and Momtazi (2018) used a similar approach of sentiment analysis
from a different social media platform in Iran that found a 97.3% chance of a specific
single candidate winning the election, which was later correct as well. The use of
sentiment analysis in these studies may not pertain specifically to the stock market;
however, the utility of sentiment analysis to predicting relationships is applicable.
Overall, these results show promise for using sentiment analysis on Twitter posts and
potentially other social media platforms.
Electronic Word of Mouth
The marketing term word-of-mouth meaning a person speaking about a product,
company, among others, could intentionally or unintentionally influence another people's
decision-making. A traditional form of marketing communication between friends and
families has evolved into a broad audience around the world because of the Internet
(Feng & Liu, 2017). Electronic word of mouth (eWOM) online communication has come
in the form of blogs, social media, review platforms, and others (Ramon Saura et al.,
2018). The ability to have a more considerable impact through word-of-mouth marketing
39
has evolved with changes in communication platforms to becoming accessible
worldwide.
Twitter is such a primary platform for eWOM and a useful evaluative population
for specific research questions regarding how comments affect the purchasing of
products, services, investments, and even the company's stock price. Xun and Gou (2017)
looked into one year of Twitter eWOM feedback sentiment and companies' time-series
stock price fluctuation and found a positive correlation. The Xun and Gou data also
showed that eWOM has a shorter effect on decision-making than traditional word-of-
mouth practices. Large investment firms use eWOM from crowdfunding sites and social
media to evaluate seed or any series of funding (Bi, Liu, & Usman, 2018). Bi et al. (2018)
suggested that eWOM insights have a relationship to investment decision-making for
billions of issued funds. The social media platforms seem to offer more than just
entertainment, but also an insight into the possibility of companies' evaluation and
success.
The use of sentiment is prevalent for different practitioners, from the consumers
to investors using eWOM to garner insight for decision-making. Chu and Sung (2015)
studied eWOM on consumers' influence of brands from Twitter posts. Chu and Sung
found that customers are more likely to engage in brand discussions when their peers
have positive brand attitudes towards a company. Zhao, Wang, Tang, and Zhang (2020)
found that peer communication on social media impacts purchasing decisions. Thus, one
peer follower's post on a company or product can strongly influence purchasing
40
intentions for many products (Zhao et al., 2020). The influence of peer sentiment may
have a strong relationship to decision-making for consumers and investors.
Cryptocurrencies
The new electronic currencies have a place in the investor market. A significant
growth cryptocurrency called Bitcoin, reaching the price of $20,000 a coin in 2017,
began the shift to a new form of investment currency (Norman & Uulu, 2020). With the
rise of Bitcoin came studies testing for sentiment relationships to find ways to predict
price changes. Valencia, Gómez-Espinosa, and Valdés-Aguirre (2019) found that
sentiment analysis from Twitter comments correlates with predicting the price
fluctuations of the major cryptocurrencies. Moreover, Jain, Tripathi, Dwivedi, and
Saxena (2018) conducted multiple regression studies on two cryptocurrencies, and the
relationship from social media sentiment predicted price change with 59% accuracy.
Narman and Uulu (2020) suggested that the relationship between social media sentiment
and cryptocurrencies only impacts price change sometimes. However, Inamdar, Bhagtani,
Bhatt, and Shetty (2019) found that there is no significant relationship between social
media sentiment and the price for any cryptocurrencies. The use of sentiment analysis in
the cryptocurrency market is not definitive, just like the stock market. Perhaps other
significant investment opportunities such as real estate may fare better.
Real Estate Investments
Sentiment analysis has even reached the commercial and securitized market for
real estate decision-making. Ruscheinsky, Lang, and Schäfers (2018) found a constant
relationship between news media from 2005 to 2015 by measuring 125,000 news articles
41
and real estate investment trust (REIT) returns and predicting prices for three to four
months. Similarly, Hausler, Ruscheinsky, and Lang (2018) found a correlation between
news media and individual decision-making based on the perception of sentiment and
real estate market movements. However, Braun, Hausler, and Schäfers (2020) suggested
that real estate liquidity has a lagging effect after news sentiment is accessible, and
investors should use the lag to invest or divest their estate holdings.
Glassdoor
Glassdoor is a software company founded in 2008 that offers employees the
ability to have anonymity with ranking and writing opinions about the companies they
are currently or have worked for in the past. The companies' ranking is a Likert-scale
format allowing the employees to rate their career opportunities, compensation, benefits,
culture, values, and the CEO (Chamberlain, 2015). There is also the option to give salary
information for the positions held by the employees, which offers other users a sense of
the range of potential compensation for the geographical area and title (Chamberlain,
2015). The users can also write positive and negative reviews of the companies with
unstructured text that other users view. The large pool of public information from
Glassdoor is an excellent database for different hypotheses testing in terms of leadership,
employee satisfaction, and benefits' relationships to sentiment analysis.
The chief economist at Glassdoor, Dr. Andrew Chamberlain, publishes many
different research papers using the Glassdoor platform data and testing hypotheses. One
such research paper published in 2015 analyzed Glassdoor's Best Place to Work
companies based on employee feedback and the stock performance of those firms
42
(Chamberlain, 2015). Chamberlain even used a third-party set of data from Fortune's 100
Best Companies to Work For as an independent variable. The control and comparison
variable is the S&P Index 500 performance. The period of measurement was from 2009
to 2014 stock prices (Chamberlain, 2015). The idea that positive feedback from
Glassdoor and Fortune's methods of feedback-populated individual companies on their
list shows as a portfolio that they outperformed the S&P 500 index by 84% (Fortune) and
115% (Glassdoor; Chamberlain, 2015). The use of the 100 best companies from either
source may be a place to help with investment decision-making.
The use of Glassdoor data in the only couple of peer-reviewed articles measures
the crowdsourcing reviews and how they affect stock returns. Green, Huang, Wen, and
Zhou (2019) suggested that Glassdoor reviews can predict up to three months of stock
returns. The study's conclusion is not particularly strong, and the limitation of a single
source of data shows a weakness in the lack of independent variables. Sheng (2019)
suggested that crowdsourced platforms like Glassdoor can enhance investment decision-
making better than traditional sources, with 7 to 9% higher returns. The potential increase
for investors looking to enhance their decision-making may need to analyze the sentiment
from platforms like Glassdoor.
Internal employee impact. There are many studies on employees receiving stock
options and a relationship to changes in business culture, stock prices, and production
drive. Aziz, Ansari, and Alam (2019) evaluated the Indian 100 best companies to work
with in terms of the employees' view of business culture and relationship to stock prices.
The study comes after peer-reviewed studies of United States companies with a similar
43
survey and outcomes of outperforming the market. The best companies to work with
surveys measure different business cultural perspectives, including employee turnover
and benefits (Aziz et al., 2019). Aziz and Chamberlain suggested that employee feedback
may relate to business stock price fluctuations. The scale of the surveys on informational
websites like Glassdoor offers a more extensive array of feedback as an opportunity for
studying a potential relationship with investment decision-making.
Review of Dependent Variable
Investment decision-making. The types of information an investor uses to
execute a decision can be vast or very simple. Individual investors tend to buy or sell
based on news, whether positive or negative (Rajdeep, 2020). Moreover, Rubaltelli,
Agnoli, and Franchin (2016) suggested that the more a company has a media presence,
the more likely the company's stock will yield one-day extreme returns. The type of
information and the source of media can be a factor in investment decision-making. The
influencers of decision-making come in many different areas from marketing and
behavioral finance perspectives from word-of-mouth, professional investment advice, and
emotions (Lynderber, 2016; Rubaltelli et al., 2016; Zhou et al., 2018). The use of
investment decision-making from AI to insider trading is an extensive program of
evaluation in economics and financial services of how these platforms affect a stock
market (Rajdeep, 2020). After completing a preliminary investigation of different
company information, the use of stock price as an indicator for investment decision-
making is a pragmatic approach to buying, selling, or holding stocks (Daniel et al., 2017;
Derakhshan & Beigy, 2019; Nisar & Yeung, 2018). Whether small or large, an
44
investment decision is a conclusive commitment, and having the best information to help
with the decision is paramount.
Enhancing Investment Decision-making
The breadth of sentiment research from using different data and hypotheses
approaches is quite vast. Gao et al. (2020) looked into Google search sentiment and the
relationship between 38 different country stock market fluctuations. Gao et al. (2020)
found a relationship in predicting all 38 stock market future returns. Che, Zhu, and Li
(2020) focused on CEO letters about corporate social responsibility sentiment and their
correlation with the company's future financial performance. Li et al. (2017) suggested
that some economists believe the information published about a company affects their
stock price. Li et al. used 30 large-cap NASDAQ companies' 200 million tweets'
sentiment related to the companies' stock price. Li, et al. found 70% accuracy when
predicting the future price fluctuations of the 30 NASDAQ companies.
The possibility that a crowd's wisdom can be more influential and correct than
single financial professionals' advice is possible. Wu, Ye, Hong, and Li (2020) focused
on crowd wisdom and professional analyst recommendation sentiments for investment
decision-making to provide relational insights. The professional analyst career field for
investment decisions has a lower probability of being correct than a crowd suggestion
from social media posts (Wu et al., 2020). The sentiment evaluated in Wu et al. found
that both professional and crowd investment suggestions could predict future prices;
however, crowd sentiments have a more significant relationship to successful investment
decision-making. Moreover, investors only using news article sentiments could be too
45
late when making an investment decision. Hwang and Kim (2019) found that three
months of news articles' effect on stock prices have a greater effect on news content than
news affecting a stock's price. Zhou et al. (2018) suggested that emotions from sentiment
drive 98% of investor decision-making in the Chinese stock market. Zhou et al. evaluated
ten million tweets sentiment to the Shanghai Stock Exchange price fluctuations with
correlation and Granger causation analysis, producing 64% prediction accuracy. Kesavan,
Karthiraman, Ebenezer, and Adhithvan's (2020) sentiment analysis and stock price
predictions suggested that sentiment added to an investor's decision-making process is
wiser and more profitable. This study should add to the body of research on the subjects
of sentiment analysis and investment decision-making.
The United States leads the world in research into how social media affects stock
prices and purchasing decisions (Teng-Chieh et al., 2019). The effect of each social
media platform and its effect on business performance from many areas such as brand
awareness, stock prices, and customer engagement will continue to be an area of future
research (Pourkhani et al., 2019). The findings suggest that there are fluctuating
relationships and no substantial causation from the studies presented in the literature to
date.
Transition
In Section 1, I presented the foundation of the study of a correlation ex post facto
study between EMH and successful decision-making measured by stock prices. This
section covered the background and statement of the business problem, purpose, nature,
and theoretical framework to bound the research. I also state the assumptions and
46
limitations of the study and a review of the possible significance of the study to the
practice of business. I concluded with a critical analysis and comprehensive literature
review of the EMH theory, measurement tool: sentiment analysis, and the variables of the
study: Twitter, Glassdoor, investment decision-making as measured by stock prices.
There is also a critical analysis of the EMH theory, independent variables, dependent
variables, and alternative theories such as noise trader theory and prospect theory. The
section includes the role of the researcher and identification of participants, a discussion
of research methodology and design, the role of, and adherence to, the requirements of
ethical research, and a review of data collection and analysis techniques.
The data I present in Section 2 includes an overview of the project, the purpose
statement, the researcher's role, the participants, and an outline of the research method
and design. The further details are the population and sampling method, ethical research,
the data collection instruments, data collection techniques, data analysis methods, study
validity, and how they pertain to this study. Section 3 includes information about the
presentation of the study findings, the application of research findings to professional
practice, implications of the study for social change, recommendations for action and
future research, reflections, a summary, and conclusions.
47
Section 2: The Project
Section 2 includes a discussion of the processes used for the execution of the
study. I begin with a restatement of the purpose, a discussion of the role of the researcher,
and identification of the archival data for this study. I continue with an in-depth
discussion of research methodology and design, population and sampling techniques, and
ethical research requirements. I close this section with a review of data collection,
analysis techniques, and study validity.
Purpose Statement
The purpose of this quantitative ex post facto correlational study was to examine
the relationship between customer Twitter comments, employee Glassdoor feedback, and
successful investing decisions measured by business stock prices. The independent
variables were customer comments from Twitter on company profiles, Glassdoor Likert
scale ratings, and employee comments on company profiles. The dependent variable was
the outcome of investing decisions measured by stock prices. There is no target
population of people because I used archival data from U.S. companies listed on the
NASDAQ 100 index. The implications for social change might include more effective
investment decision-making that may increase funds for retirements, children's education,
or other socially desirable uses.
Role of the Researcher
As the researcher, my objective was to answer the research question: Do customer
Twitter comments and employee Glassdoor feedback have a relationship with successful
investing decisions measured by business stock prices? As an objective observer of a
48
secondary dataset, I have maintained the integrity of the data and have openly discussed
any cleaning or manipulation of the data. I was a certified day trader for one year, which
sparked an interest in doing this study, including learning about machine learning and
data science as a hobby. Prior to the collection of the data, I had obtained Walden
University's institutional review board (IRB) approval.
Participants
There were no participants in this study, only archival data with the NASDAQ
100 index companies obtained by secondary dataset from Followthehashtag
(http://www.followthehashtag.com/). The companies are large capital firms with active
Twitter hashtags and accounts on Glassdoor with over 100 comments on each platform.
All of the participants' data were publicly available.
Research Method and Design
The purpose of this quantitative ex post facto correlational study was to
examine the relationship between customer Twitter comments, employee
Glassdoor feedback, and successful investing decisions. The study involved
secondary data from Followthehashtag’s NASDAQ 100 Companies dataset and
MIT SMR Culture 500 data that collects and analyzes from publicly accessible
Glassdoor employee reviews. The NASDAQ 100 Companies dataset had
sentiment analysis processes executed from a third-party machine learning
algorithm. The MIT SMR Culture 500 data were already cleaned and analyzed via
sentiment analysis from MIT AI algorithms.
49
Research Method
I have a positivistic mindset as a researcher that influences the method and design
for this study. Positivistic researchers seek empirical evidence to suggest truths (Hawang,
2019). I chose a quantitative method for this study because of empirical, publicly
available secondary data that is useful for analyzing and answering the research question.
The importance of quantitative data is the generalization and capability of testing
relationships with independent and dependent variables (Wienclaw, 2013). The
understanding of predictors for dependent variables is achievable when using quantitative
methods (O'Leary, 2017; Palinkas et al., 2015). If this study had needed interviews and
precise nongeneralizable data, a qualitative study could have been more applicable.
Qualitative studies have strengths when specific humanistic areas of data and
triangulation approaches are necessary to gain quality results (O'Leary, 2017). If the
objective of a study is to gather data from participants with and produce a thematic output
for measuring interpretations, then a qualitative study would be useful. However, this is
not the focus of this study. The same qualitative processes come with using a mixed-
method approach.
A mixed-method approach is the combination of qualitative and quantitative data
analysis researchers use to bring both strengths of the methods to a study. Different data
types are analyzed depending on the population and available data for a researcher to
consider using a mixed-method (Alavi, Archibald, McMaster, Lopez, & Cleary, 2018).
The availability of time and money is a factor when using a mixed-method approach that
could increase both as a necessity, whereas choosing a single method could be less time-
50
consuming and more cost-effective. Quantitative data collection is the best option for
studying relationships with independent and dependent variables (Chirkov & Anderson,
2018; Hwang, 2019). This study encompassed unbiased numerical data that meets the
positivist worldview and reasoning for using a quantitative research method.
Research Design
I chose correlation as the design because of the positivistic view and approach to
research. A positivistic researcher uses correlation to measure relationships and not
causality (Jackson, 2017). In contrast, an experimental design is appropriate for
establishing a causation experiment (Bryman & Bell, 2015). The correlation design
worked well for this study exploring the relationship between the predictor and
independent variables. A researcher using a correlational design can also use longitudinal
measurements to analyze data at different times, which was a component of this study
(Leedy & Ormrod, 2015). A correlational design and a positivistic view have provided
evidence to determine relationships from the study findings.
There are other quantitative designs, such as experimental and quasi-
experimental. Quasi-experimental and experimental studies are beneficial for examining
the difference between variables when researchers seek cause and effect (Abutabenjeh &
Jaradat, 2018). Quasi-experimental and experimental designs also need controllable
dependent variables. Since I had no control over the variables, an experimental or quasi-
experimental design was not appropriate.
There are three main types of correlational designs, observational, survey, and
archival research. The researchers using observational design gather data in a
51
noncontrolled setting from live participants (Boyd, Gove, & Solarino, 2017; Wyman et
al., 2015). Researchers using survey design have a preset questionnaire that is later
measurable and can add internal validity to the study (Boyd et al., 2017; Wyman et al.,
2015). An archival design researcher uses past or secondary data when measuring
relationships via data analysis (Boyd et al., 2017; Wyman et al., 2015). All correlation
designs have their appropriate need for different research questions and availability of
time, cost, and population. I used public secondary data for my study that was already
open source and relevant to my research question. Researchers using the archival design
or ex post facto design can balance the need for internal and external validity because the
data are available to other researchers (Boyd et al., 2017; Handley, Lyles, McCulloch, &
Cattamanchi, 2018). The ex post facto design properties can be useful in a quasi-
experimental study when evaluating cause and effect, not relationships (Abutabenjeh &
Jaradat, 2018). The ex post facto design is beneficial when using publicly available
secondary data and effective for correlational analysis and validation testing (Boyd et al.,
2017). The data gathered from the secondary dataset were randomly selected; the original
researcher gathered the data from NASDAQ 100 companies’ Twitter feeds, specifically
for analysis.
Population and Sampling
Population
There was no population for the study. The archival data came from large capital
NASDAQ 100 index companies in a publicly available secondary dataset from the
Followthehastag website. The companies were active businesses throughout the chosen
52
timeframe of the study. The firms were publicly owned businesses, not privately owned,
with a presence on the NASDAQ stock exchange. The companies' stocks were past the
initial public offering stage, which means enough stock pricing data were gatherable to
meet the desired timeframe of analysis. I completed a probabilistic random sampling for
this study using Microsoft Excel randomizing algorithms to produce such results for
analysis.
Sampling
Quantitative studies involve gathering samples in a probabilistic or
nonprobabilistic manner. Probabilistic sampling is grabbing participants in a population
with the same chances of being selected (Barker & Pistrang, 2016; Palinkas et al., 2015).
A researcher determining a probabilistic or nonprobabilistic sample for a study is
derivable by time, cost, and population availability (Barker & Pistrang, 2016; Palinkas et
al., 2015). The population determination also involves whether the data or population is
current and willing to be involved in a study. A nonprobabilistic sampling is appropriate
for nonrandomized data, and generalization is not probable from the available data
(Barker & Pistrang, 2016; Palinkas et al., 2015). Since this study was a correlational
design and provides generalizability, I did not use nonprobabilistic sampling.
There are different probabilistic samples, including simple random sampling,
systematic sampling, stratified sampling, and cluster sampling. A simple random sample
is having a whole population with an equal chance of being chosen and then using it to
analyze (Barker & Pistrang, 2016; Palinkas et al., 2015). A researcher can use a
systematic sampling process by choosing participants from a population in a preplanned
53
manner until a sufficient amount of data is obtainable for the study (Barker & Pistrang,
2016; Palinkas et al., 2015). If the sample was to be split up into different groups and
chosen randomly, then this method is a stratified sample (Barker & Pistrang, 2016;
Palinkas et al., 2015). A cluster sample is selecting random participants in population
clusters (Barker & Pistrang, 2016; Palinkas et al., 2015). The sampling method for this
quantitative study was probabilistic in the form of simple random sampling.
The secondary dataset in this study consisted of 100 NASDAQ companies, while
the full NASDAQ index consists of 3,300 companies' stocks. Simple random sampling
from a larger pool of participants and representing a population with a smaller set of data
to analyze and generalize is valuable for a study and the researcher's time constraints
(Kessler et al., 2015). The value of simple random sampling is the ease of gathering data
and representing a larger population (Kessler et al., 2015). A weakness from this
sampling method is the requirement to have a full list and equality of all participants in
the population, or the generalization factor fails (Barker & Pistrang, 2016; Kessler et al.,
2015). Moreover, the necessary amount of data was paramount to study findings and
generalizability.
The number of companies needed was anything over 71, including comments
from Twitter and Glassdoor reviews for each company. Using G*Power version 3.1.9
software, a power analysis is appropriate to determine the study's appropriate sample
size. An a priori power analysis, assuming a medium effect size (f 2= .15), α = .05, and
two predictor variables, identified that a minimum sample size of 72 participants is
required to achieve a power of .80. Increasing the sample size to 146 will increase power
54
to .99. Therefore, I used the secondary dataset NASDAQ 100 companies that encompass
the target population (see Figure 1). I was not seeking or using small capital or private
companies from the NASDAQ dataset or index.
Figure 1. Power as a function of sample size.
Ethical Research
I collected the secondary data for this study that was open to the public for
analysis and an open license to use for any purpose. I had completed all necessary IRB
form B documentation to enforce the ethical use of the data. The data does not need
traditional methods of encryption or protection of personal identifiable information (PII).
There was no need to offer incentives or withdrawal procedures, as there were no
individuals to provide data for the study. The data are accessible without security
firewalls such as a login or password. I protected the company's and Twitter user's
identities by not identifying them in this study. The company information is all publicly
available information; regardless, I maintained confidentiality by encrypting the data. I
will keep the data collected from this study locked in a safe for five years in a thumb
0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95
0
20
40
60
80
100
120
140
Total sample size
= 0.15
Effect size f²
F tests - Linear multiple regression: Fixed model. R² increase
Number of tested predictors = 2. Total number of predictors = 3. α err prob = 0.05. Effect size f²
= 0.15
Power (1-β err prob)
55
drive, and then I will completely wipe the memory destroy the thumb drive with a
hammer after five years. The Walden University IRB approval number for this doctoral
study is 10-02-20-0759876.
Data Collection Instruments
I used archival data as the data source for this study, and no instruments were
necessary. I gathered the archival data from three separate sites Followthehashtag
(followthehashtag.com), MIT SMR Culture 500 (sloanreview.mit.edu/culture500), and
Yahoo finance (finance.yahoo.com). Followthehashtag has the NASDAQ 100 companies
tweets archival dataset available for public use. MIT SMR Culture 500 is an archive,
sentiment analysis, and data visualization website for 550 different companies’ Glassdoor
post. Yahoo finance website has archives of past stock prices from many different
companies.
MIT SMR Culture 500
The MIT SMR Culture 500 platform was a new online data visualization software
created with MIT, Glassdoor, and a company called CultureX. The most recent
researcher to use this platform is from a master's thesis from Claremont McKenna
College on using Glassdoor data to measure the impact of culture and employee
satisfaction (Uyeno, 2020). The Glassdoor review data was from 531 organizations from
January 1, 2014, to March 31, 2019. The number of average reviews for all companies
was around 2,182, with a median of 1,221 (Sull, Sull, & Chamberlain, 2019). The entire
approach for the MIT platform was to evaluate the typed text found on Glassdoor
reviews, including the Likert-scale feedback. The type of AI MIT uses is Natural
56
Language Processing (NLP). MIT SMR provides different evaluation values from the
typed text with sentiment analysis and frequency scores in the following nine categories:
• Agility: Employees have the environment to be flexible, nimble, and fast to
seize new opportunities or market changes effectively (Sull et al., 2019).
• Collaboration: Employees have success with teamwork practices from
different areas of the company.
• Customer: The employees have the customer as the main focus of their
everyday practices (Graham, Harvey, Popadak, & Rajgopal, 2017; Sull et al.,
2019).
• Diversity: The company promotes and executes diversity and includes it in the
workplace without any person's creed at a disadvantage (Sull et al., 2019).
• Execution: Projects managed well, and operational excellence is seen through
employee empowerment and adhering to accountable conclusions (Graham et
al., 2017; Sull et al., 2019).
• Innovation: The company is a leader in creating cutting edge services or
products and ways of working (Graham et al., 2017; Sull et al., 2019).
• Integrity: Employees have a steady presence of honesty and ethical behavior
(Graham et al., 2017; Sull et al., 2019).
• Performance: A company rewards employees with many different avenues of
appreciation and helps underperforming employees.
• Respect: Employees treat each other with respect (Graham et al., 2017; Sull et
al., 2019).
57
All nine culture 500 categories measured for sentiment are considered a single
independent variable for this study. The average mean of the sentiment scores reflects the
company's overall sentiment as a whole from Glassdoor reviews.
The data’s reliability and validity are reviewed by the MIT Sloan Management
peer review committee and created by staff in the MIT and the private industry (Sull et
al., 2019). The MIT SMR Culture 500 used in a 2020 master's thesis at Claremont
Mckenna College was proof of academic attention from the data. It was openly accessible
for anyone to view.
Monkeylearn
The followthehastag.com NASDAQ 100 dataset was not pre-analyzed for
sentiment like the MIT SMR Culture 500 data. The NASDAQ 100 dataset needed the
same sentiment analysis to provide the input necessary for multiple regression analysis.
Monkeylearn was an AI online software for early developers or anyone new to AI that
provides text sentiment analysis by uploading data to their site (Ramon Saura et al.,
2018). The software allows users to select the types of models from Twitter sentiment
analysis for financial specific context to conversational context (Rakshitha &
Gowrishankar, 2018). The software has an added feature of having the user train the
algorithm by reading and manually labeling text with a positive, neutral, or negative
sentiment. The feature then uses the user input to output millions of individual Twitter
posts sentiment in seconds.
The use of an AI called Monkeylearn for analyzing Twitter data sentiment is
found in peer-reviewed studies from the social sciences of evaluating well-being to
58
environmental management sustainability of hotels (Ramon Saura et al., 2018). The use
of AI tools was beneficial for those that wish to employ SVM and naïve Bayesian
algorithms that are the latest and most accurate AI available to the public. Ramon Saura
et al. (2018) used Monkeylearn with online Twitter comments producing a 98.5%
accuracy for evaluating correct sentiment from 500,000 reviews. Rakshitha and
Gowrishankar's (2018) use of Monkeylearn for determining a person's mental health
well-being via 65,994 Twitter tweets came with 86.4% accuracy. The most recent study
using Monkeylearn came from a published dissertation about the impact of positive and
negative news on the stock performance of sector and market performance, with a
rejection of sector and market performance correlating to news and stock performance
(Bernardo, 2020).
The Monkeylearn platform was used in peer-reviewed journals for sentiment
analysis and used in a 2020 Ph.D. dissertation from Capella University (Bernardo, 2020).
A standardized sentiment analysis tool mitigates the biases found in human participation
studies (Babu & Kumar, 2015). While using the analysis tool to read emotional responses
with standardized words and context of phrasing reduces the need for people to read or
synthesizes writings for the sentiment. The analysis tool treats every sentence with the
same probability as the next with a constant state of emotional standardization (Babu &
Kumar, 2015). The Monkeylearn sentiment analysis tool evaluates large sets of data in
seconds for the sentiment. The sentiment output is then dispensed back to the user in a
CSV format with each tweet comment and sentiment determination and validation
(Ramon Saura et al., 2018; Gowrishankar, 2018). The tool was open source and does not
59
need any special requirements outside of uploading data as a CSV and choose the
columns that the system should analyze for the sentiment.
Yahoo Finance
The Yahoo finance intraday data was a valuable asset in peer-reviewed research
to include studies involving sentiment analysis. Ranco et al. (2016) used Yahoo Finance
data for 100 companies spanning a year-long stock price dataset for each company. Ito,
Sakaji, Izumi, Tsubouchi, and Yamashita (2017) used stock market data and board
textual data to enhance their machine learning algorithm for rating financial statement
sentiment. Moreover, Ho, Damien, Gu, and Konana (2017) used Yahoo Finance message
boards content to evaluate sentiment and potential stock market price predictions. The
use of Yahoo finance data was prevalent in different types of studies that include stock
price metrics and Twitter comments. Thus, providing sufficient evidence for using Yahoo
finance as a dependent variable data source for this study.
Data Collection Technique
Is there a statistically significant relationship between customer Twitter
comments, employee Glassdoor feedback, and successful investing decisions measured
by business stock prices? The Twitter comments are gatherable from secondary data
sources from followthehashtag.com. The secondary dataset was collected from the
Twitter API from all NASDAQ 100 companies from March 28, 2016, to June 15, 2016
(79 days), consisting of about 1 million tweets. The dates for the secondary data are not
significant other than being randomly gathered by followthehashtag.com. The Glassdoor
feedback came from MIT SMR Culture 500, which has over 513 companies with over 1.2
60
million comments already analyzed for the sentiment from March 2016 to March 31,
2019. The MIT SMR analyzed data was an excellent source found to be integral from the
MIT staff. The only data gathered from the MIT SMR platform came from the NASDAQ
100 companies' profiles in the secondary dataset. When gathering the data from the MIT
SMR Culture 500 website, I used the homepage search function to aggregate the company
data from the site into an excel spreadsheet. The spreadsheet consists of company names,
the number of comments analyzed, and the average sentiment score. The data from
followthehastag.com was already formatted with the date, time, context, and added
sentiment scores when analyzed by Monkeylearn.com. There is also no need for pilot
testing since it was archival data.
When using secondary data or ex post facto, the study’s data are available for
other researchers to validate and continue future research with the same data (Boyd et al.,
2017). The issue with some secondary data is the relevance of the data and how it may
not be useful for future researchers with different hypotheses and newer data (Boyd et al.,
2017). The secondary data may come from sound sources; however, when using a
questionnaire or observational data gathering technique, the researcher controls the data
consumption from the beginning, unlike secondary data held first by someone else. There
is a level of assumption from secondary data that is not a particular problem from being
the first person to gather any data. An advantage with secondary data is the public
availability of the data, so other researchers can use it for reliability and validity testing.
61
Data Analysis
To reaffirm the research question of this quantitative study: What is the
relationship between customer Twitter comments, employee Glassdoor feedback, and
successful investing decisions measured by business stock prices?
Hypotheses
H0: There is no statistically significant relationship between customer Twitter
comments, employee Glassdoor feedback, and successful investing decisions measured
by business stock prices.
H1: There is a statistically significant relationship between customer Twitter
comments, employee Glassdoor feedback, and successful investing decisions measured
by business stock prices.
As a researcher, I have to consider the types of data and variables when
suggesting and using statistical analysis for hypothesis testing (O'Leary, 2017). This
study had two independent variables and one dependent variable. The different types of
quantitative variables are found in two different general categories continuous and
discrete (Bezzina & Saunders, 2014; Mertler & Reinhart, 2017). A continuous variable is
a numerical variable between two parameter values: 0 and 1, with no infinite value from
fractions, decimals, and whole numbers (Bezzina & Saunders, 2014; Mertler & Reinhart,
2017). Discrete variables have only finite possibilities from two set parameters (Bezzina
& Saunders, 2014; Mertler & Reinhart, 2017). This study's two independent variables are
continuous variables with the infinite possible values ranging from -1 to 1, including all
fractions and decimals. If any score was greater than 0.5 maximum, then the data was
62
represented with a value of 1 for positive sentiment. Any score less than 0.5 maximum is
represented as a value of 0, meaning neutral; any score less than 0 is represented as a
value of -1, meaning the sentiment is negative (Akba et al., 2020; Attarwala et al., 2017;
Birbeck & Cliff, 2017; Derakhshan, & Beigy, 2019). At the same time, this study's
dependent variable was also continuous because it is stock prices. The stock prices have a
wide range of possible values; however, the starting-day stock price and end-of-day stock
price were measurable with a binary construct for this study. An increase of only 1 cent
for the stock price means a profit, while a decrease by one cent is a loss represented in the
data as 1 being a profit and -1 as a loss (Akba et al., 2020; Attarwala et al., 2017).
Multiple regression analysis and structural equation modeling (SEM) are suitable
for studies containing multiple continuous independent variables with a single dependent
variable (Henseler, 2017; Kline, 2016). Multiple regression is analyzing all of the
independent variables at the same time that enlightens a researcher with the best
predictors from the body of data. An SEM is a layer of involved mediators for multiple
independent variables, and it tests multiple equations in the same breath (Henseler, 2017;
Kline, 2016). The amount of interpretation and deployment setup for the SEM is quite
complicated and time-consuming (Henseler, 2017; Kline, 2016). The multiple regression
analysis was the most appropriate for this study because there was only one set of
independent variables and a single equation derivable for adequate analysis.
The scale of measurements found in quantitative studies is nominal, ordinal,
interval, and ratio (Mishra, Pandey, Singh, & Gupta, 2018). A nominal scale is
appropriate for assigning numbers to groups such as males equals 1, and females equal
63
two; this is useful when the variable has no fundamental numerical properties (Mishra et
al., 2018). An ordinal scale is useful for placement of magnitude, such as rankings from
first to last (Bishop & Herron, 2015). The interval scale is an equitable variable populated
between two integers and is addressable for quantifying (Mishra et al., 2018). The ratio
scale has a definitive definition for zero; the pulse of a human being can is measurable as
a ratio (Mishra et al., 2018). The scale of measurement for this study's independent and
dependent variables was interval numbers. The use of sentiment analysis outputs a wide
range of number properties from -1 to 1 with any possible values between these numbers
(Birbeck & Cliff, 2017; Derakhshan, & Beigy, 2019). The dataset had multiple missing
values that were preventing the sample from reaching the minimum. I had to input the
missing data with bootstrapping, which was the most reliable approach (Ales & Marko,
2017; Sessa & Syed, 2016). All of the raw data in this study is available upon request.
Statistical Analysis Assumptions
When using multiple regression analysis, there are some assumption requirements
to consider. The main assumptions are outliers, normality, linearity, heteroscedasticity,
heteroscedasticity, and independence of residuals. I used a normal Probability Plot (P-P)
of the Regression Standardized Residuals to test the normality, linearity,
multicollinearity, independence of residuals, and a scatter plot to test and view potential
outliers. The normal P-P is an affective test because of the precision output that is not
found in histograms when too little, or too much data is involved (Keith, 2014;
Krzywinski & Altman, 2015; Yang & Yuan, 2016). With the main assumptions of
64
multiple regressions discussion, there was a way to inspect for anomalies known as
outliers by inspecting the output in a graph.
Outliers can manipulate a relationship test because of the variance imposed on
data analysis output (Keith, 2014; Krzywinski & Altman, 2015). Outliers come in many
forms: errors, population crossing, typographical mistakes, purposeful and random
anomalies (Keith, 2014; Krzywinski & Altman, 2015). I used the normal P-P and
scatterplot to review and deal with any outliers by cleaning the outliers out of the data to
prevent Type I and Type II errors.
The assumption of normality means all variables are present, and normally
distributed data points are sought-after (Keith, 2014; Krzywinski & Altman, 2015).
Multiple regressions work best with the normality of data, and to make sure that
normality exists, it can be viewed and analyzed with a normal P-P. I also examined for
skewness, meaning symmetry of the data, and kurtosis meaning long or short tails scene
in the normal P-P output.
The linearity assumption is necessary for multiple regression analysis. If an
outcome of multiple regression analysis shows a nonlinear output, the relationship can be
skewed and often untrustworthy (Keith, 2014; Krzywinski & Altman, 2015). When a
linear relationship is derivable, evaluating the correlation coefficients for random
distribution can provide evidence to evaluate homoscedasticity integrity (Keith, 2014;
Krzywinski & Altman, 2015). When the heteroscedasticity is profound, the probability of
Type 1 errors increases and distorts the study findings (Keith, 2014; Krzywinski &
65
Altman, 2015; Yang & Yuan, 2016). The multiple regression analysis was essential for
evaluating linearity with the independent and dependent variables.
Multicollinearity is the process of testing independent predictor variables for
correlation. If a correlation exists between independent variables can make insignificant
and unstable estimations of coefficients (Yang & Yuan, 2016; Yoo et al., 2014). When
multicollinearity inflates the analysis, the ability to imply the inference or validity of the
testing practices is questionable (Yang & Yuan, 2016; Yoo et al., 2014). I tested for
multicollinearity by using the variance inflation factor (VIF) to measure the predictors.
The independence of residuals is estimated values from the dependent variable
output. A residual is measuring the difference between the estimated and actual
dependent value (Keith, 2014; Krzywinski & Altman, 2015). I tested for the
independence of residuals by using the normal P-P plot to evaluate random dispersal and
constant variance from the residuals (Keith, 2014; Krzywinski & Altman, 2015). The
results should be randomly numerical values in ranges of negative and positive -1 < 0 > 1
(Keith, 2014; Krzywinski & Altman, 2015). The residuals may present as independent
because of potential study design, and understanding the distribution of the data is not
necessarily random but conditioned (Ernst & Albers, 2017; Yang & Yuan, 2016).
Multiple Regression Analysis
Analyzing the multiple independent variables allows researchers to view the
predictors at the same time. Researchers using a multiple regression need two or more
independent variables and at least one dependent variable (Mertler & Reinhart, 2017). I
used the multiple regression algorithm with IBM SPSS 25. I imported my CSV or XLS
66
data from the output of the Monkeylearn sentiment analysis. The data cleaning
progressed, meaning all of the 1 million tweets had the sentiment signifiers changed from
a scale of positive, neutral, and negative sentiment to signifying their interval value: 1
(positive), 0 (neutral), and -1 (negative). The Glassdoor data analyzed from MIT SMR
Culture 500 was labelable with the same scoring of 1 (positive), 0 (neutral), and -1
(negative). The stock market data for each company during the 79 days was evaluated on
a daily price basis from the beginning of trading to the end of the trading day. The
evaluation of price is accomplishable by noticing if the end of day price is higher than the
beginning day price; in this case, the signifier was labelable with a 1, or if there is a lower
price, then there is a -1 labeled for the day. The CSV Yahoo finance tables had all
available data such as company stock ticker, date, and total day price signifier labeled in
each column, and the analysis was commenced, which produced the multiple regression
output.
Interpretation of Multiple Regression Analysis
I had tested the multiple regression model to my data for goodness-of-fit via
adjusted R2, F- statistic, and p value with SPSS. The use of R2 is determining the
percentage of variation between models and the data (Keith, 2014; Krzywinski &
Altman, 2015). The higher the adjusted R2, the better the fit is between a model and data
(Keith, 2014; Krzywinski & Altman, 2015). To reinforce the R2 output, I used the p value
and f-statistic to evaluate the data and model fit assumptions. The p value output
definitions come from the output of measurements greater than or less than 0.5. A value
that is < 0.5 suggests a model is statistically significant in estimating a dependent
67
variable. A value > 0.5 suggests that changes in the independent variable do not signify a
correlational change for the dependent variable (Keith, 2014; Krzywinski & Altman,
2015). A beta (β) is a measurement level for the probability of a Type II error being
present. An excellent value a researcher is looking for is 0.20 (Keith, 2014; Krzywinski
& Altman, 2015). The outputs of multiple regressions are the beta coefficient, t-statistic,
and the p value (Keith, 2014; Krzywinski & Altman, 2015). The testing for the
relationship between the independent variables and the dependent variable, the p value
tests for the null hypothesis. A p value < 0.05 indicates that the null hypothesis is
rejectable, while a value > 0.05 is accepting the null hypothesis (Keith, 2014; Krzywinski
& Altman, 2015).
The β coefficients represent the strength between an independent variable and the
dependent variable. The β are standardized for each independent variable so that a
researcher can differentiate each independent variable strength of relationship to the
dependent variable (Keith, 2014; Krzywinski & Altman, 2015). An unstandardized
coefficient (B) is an analysis of the dependent variable when independent variables
change (Keith, 2014; Krzywinski & Altman, 2015). At the same time, the original units
are comparable to the normalized units of the study (Keith, 2014; Krzywinski & Altman,
2015). The standard errors (S.E.) from the standardized and unstandardized coefficients
are useful for testing the confidence intervals (CI) and hypothesis. I obtained the output
expected with the 95% CI for B, and I am 95% confident the interval contains the
population mean.
68
Study Validity
This quantitative correlation study was a nonexperimental design, and threats to
external and internal validity were not applicable. Researchers seek external and internal
validity when exploring causal relationships in experimental or quasi-experimental
designs. However, threats to statistical conclusion validity inflate the type I error and
Type II error rates and are of concern in nonexperimental designs. To reduce the threats
of statistical conclusion validity, the analysis instrument selected for this study was
reliable. I conducted a pre-study power analysis with G*Power 3 to determine the
appropriate sample size and avoided violation of the assumptions for multiple regression
analysis.
Statistical Conclusion Validity
A researcher ensuring statistical conclusion validity is obtainable starts with
ensuring the sampling methods, statistical tests, and instrument reliability are all correct
before executing the study (Anestis, Anestis, Zawilinski, Hopkins, & Lilienfeld, 2014;
Stantchev, Colomo-Palacios, Soto-Acosta, & Misra, 2014). If an observable sample or
population happens, then the collected data could all be uniform and potentially increase
the chances of Type 1 and Type II errors (Anestis et al., 2014; Stantchev, 2014). The use
of heterogeneity or different types of samples is a big part of successful regression
analysis; however, there are cases of high standard errors and small effect sizes that can
impact the test output (Anestis et al., 2014; Stantchev, 2014). The three subject areas I
covered in this study are: reliability of the instrument, data assumptions, and controlling
Type I and Type II errors.
69
Reliability of the instrument. I conducted an internal reliability check for the
data collection instrument via measurement from the Cronbach's alpha reliability
coefficients. Publishers and researchers seek a 0.70 coefficient or higher validation (Da
Silva et al., 2013). The completed validation was with the IBM SPSS version 25
software's function for analysis/scale/reliability analysis that produces Cronbach's alpha
coefficients. The presented results are in section 3 of this study.
In this study, the instrument for collection and pre-data analysis comes from
instruments used in academic institutions and private businesses. The availability of the
validity scores is not available at the moment. The prestigious top-level entities that
provide context about the tool are Harvard Business Review, The Economist, MIT Sloan,
and backed research by the top universities in the country and world (Sull et al., 2019).
Data assumptions. The primary data assumptions are normality, homogeneity of
variances, and statistical independence. The normality of data refers to the potential of a
normal distribution from the data points gathered (Ernst & Albers, 2017). When using
multiple regression, normality is essential when testing assumptions. Homogeneity of
variances is having the exact number of variances spread across groups. If there is an
imbalance in any group, then false positives can persist and cause incorrect data
assumptions. When using an F-test, a researcher can compare variation between the data
sets (Ernst & Albers, 2017). Statistical independence is having no relationship between
the data sets. I performed a Chi-Square test for evaluating data for statistical
independence.
70
Controlling Type I and Type II errors. To curb bias and increase accuracy, the
sample size for collecting and gaining data needs to be appropriate for the study
(Hoffman, 2013; Williams, Gomez Grajales, & Kurkiewicz, 2013). Achieving the correct
power, confidence levels, and effect size is the next step toward a reliable sample size
(Hawkins, Gallacher, & Gammell, 2013; Williams et al., 2013). As a researcher having a
larger sample size past the minimum 85%, the study can achieve a better generalization
in different populations (Tarhan & Yilmaz, 2014). I had ensured that the
followthehashtag.com archive data had the appropriate amount of samples. However,
when gathering the data from the other databases, I could not achieve 85% confidence,
with only 53 companies data usable.
Statistical Test
A researcher must ensure statistical validity by using an approved statistical test.
Assumptions are part of the researcher's process when doing multiple regressions
(Mertler & Reinhart, 2017). The data analysis paragraph reviews some of the
assumptions addressed when using multiple regression.
The use of bootstrapping is a technique to help with resampling the original
dataset encase the first sampling falls short of the requirements (Elvarsson, Taylor,
Trenkel, Kupca, & Stefansson, 2014). Bootstrapping is also an excellent substitute for the
use of inferential statistics (Elvarsson et al., 2014). The act of bootstrapping uses a
randomizer algorithm that duplicates and even forecasts data points from existing data
until the desired sample amount is achieved (Elvarsson et al., 2014). I used the
71
bootstrapping resample technique since my sample size was 53 out of the required
minimum of 72, as provided by the power analysis found in the sampling paragraph.
Multiple regression assumptions. When conducting a multiple linear regression,
the assumptions of normally distributable variables and a linear relationship between
independent and dependent variables, measurement reliability, and the variances of errors
is homoscedastic is necessary (Casson & Farmer, 2014; Cohen, & Abedallah, 2015).
Researchers have several methods for testing assumptions and handling violations. Data
cleaning, acceptability of results, exclusion of results, and changing variables are tools
available to handle assumption violations (Cohen, & Abedallah, 2015). Researchers that
choose not to address violations lessen their ability to infer statistical results in a broader
population (Cohen, & Abedallah, 2015).
Data cleaning is a helpful method to improve the distribution of variables by
making corrections or mitigating the impact of normal distribution violations when doing
a regression analysis (Osborne & Waters, 2002). Using visual plotting graphs is helpful
when identifying problems with variable distributions (Osborn & Waters, 2002). I used
the IBM SPSS version 25 software to develop and inspect the distribution of plots from
the archival data.
Researchers can identify linearity by plotting the dependent variable against
predictor variables (Casson & Farmer, 2014). Pearson's correlation coefficient identifies a
linear relationship between having a positive relationship when being closer or equal to
+1 the opposite when closer or equal to -1 (Mertler & Reinhart, 2017). I used the IBM
SPSS version 25 software to measure linearity between successful investment decision-
72
making measured by stock price and each of the predictor variables. Linearity with
violations can indicate the regression analysis underestimates the relationship (Osborn &
Waters, 2002). All of the nonlinear possibilities from the relationship between variables
must be accounted for and considered when stating conclusions (Osborne & Waters,
2002). As a researcher, I accounted for nonlinear possibilities and have found none with
the study data.
Overstated or understated results are possible when violating the assumptions
regarding the reliability of the measurement (Osborne & Waters, 2002; Cohen, &
Abedallah, 2015). Adding additional variables can further complicate reliability and
result in less accurate conclusions (Osborne & Waters, 2002). I verified the reliability of
measurements by using Cronbach's alpha as a measurement of scale reliability calculated
through IBM SPSS software.
Homoscedasticity assumption violations can result in errors and not qual
statements across all of the predictor variables in SPSS (Osborne & Waters, 2002). These
types of assumptions are observable while reviewing a scatterplot in SPSS. If there is a
significant amount of heteroscedasticity, then replacement variables may be necessary to
prevent distorted findings (Osborne & Waters, 2002). When completing the test, there
were no heteroscedasticity instances found when reviewing the scatterplot.
The external validity of the general population is vital once having an appropriate
sample size confirmed by G* Power analysis. With the generalizability of archival data,
once gathered, proper external validation is possible for the large population of stock
market investors. At the same time, I rely on the population and study findings for
73
external validity (Boyd et al., 2017; Handley et al., 2018). I used the G*power analysis to
find an appropriate sample has provided greater reliability and external validity in my
study.
Transition and Summary
Section 2 outlined a methodical process for participant selection, research
collection, research methods, design, and data analysis. Additional material presented in
Section 2 includes a detailed view of the data collection techniques, the purpose
statement, the role of the researcher, the population and sampling method, ethical
research, data collection techniques, research analysis methods, research reliability, and
study validity. The information included in Section 3 contains a presentation of findings,
the application of research findings to professional practice, implications of the study for
social change, recommendations for action and future research, reflections, a summary,
and study conclusions.
74
Section 3: Application to Professional Practice and Implications for Change
Introduction
The purpose of this quantitative correlational study was to examine the
relationship between customer Twitter comments, employee Glassdoor feedback, and
successful investing decisions measured by business stock prices. The independent
variables were customer comments from Twitter on company profiles, Glassdoor Likert
scale ratings, and employee comments on company profiles. The dependent variable was
successful investing decisions measured by business stock prices. The model could not
predict successful investing decisions measured by business stock prices, F(10, 2993) =
.295, p = .982, R2 = .001, meaning the null hypothesis was accepted the alternative
hypothesis rejected. The R2 (.001) value indicated approximately 1/1000% of variations
in successful investing decisions measured by business stock prices were accounted for
by combined predictor variables. The study findings suggest that there are no significant
predictor variables. However, readers should judge the results with caution due to the
small sample size and instrument alteration.
Presentation of the Findings
In this section, I discuss the variables' reliability, testing of the assumptions,
descriptive statistics, and interpretation of the findings. I conclude with a concise
summary. I used the bootstrapping technique to resample the data to address assumption
violations because my sample size resulted in 53 complete company data instead of the
77 necessaries for an 85% confidence interval. Small sample sizes often result in low
statistical power and imply that statistically significant findings will not reflect a true
75
effect (Button et al., 2013). Where appropriate, I employed bootstrapping 95%
confidence intervals using 1,000 samples to address the possible impact of assumption
violations.
Reliability of the Variables
The variables used in this study were not instruments of scale. A Cronbach alpha
analysis was not necessary for nonscale archival data. A correlation determines the
reliability of the data used in this study on the research question and predictor variables
reliability. There were no correlations found upon completing the multiple regression
analysis. This specific study determines that the preprocessed sentiment analyzed data's
reliability may not be reliable enough to predict a relationship between the independent
and dependent variables.
Test of Assumptions
I evaluated the assumptions of outliers, normality, linearity, multicollinearity, and
independence of residuals. I used bootstrapping with 1,000 samples to focused on the
impact of possible assumption violations.
Multicollinearity. I evaluated multicollinearity by viewing the correlation
coefficients among the predictor variables. All bivariate correlations were small to
medium (see Table 2); therefore, the violations of multicollinearity assumptions were not
evident.
76
Table 2
Correlation Coefficients Among Study Predictor Variables
P/L
T/Sent
.
Agil.
Coll.
Cust.
Div.
Exec.
Inno.
Integ.
Per.
Res.
P/L
1.00
.019
-.006
.014
.004
.006
.004
-.006
-.002
-.010
-.001
T/Sent.
.019
1.00
.142
.051
.047
-.212
.159
-.092
-.113
.067
.038
Agil.
-.006
.142
1.00
-.021
-.150
-.119
.060
.135
-.378
.323
-.518
Coll.
.014
.051
-.021
1.00
.014
.164
-.148
-.641
-.111
.031
-.006
Cust.
.004
.047
-.150
.014
1.00
-.323
-.304
-.166
.028
-.085
-.176
Div.
.006
-.212
-.119
.164
-.323
1.00
-.202
-.305
-.156
-0.35
-.148
Exec.
.004
.159
.060
-.148
-.304
-.202
1.00
.141
-.321
-.127
.109
Inno.
-.006
-.092
.135
-.641
-.166
.305
.141
1.00
-.126
-.200
-.121
Integ.
-.002
.113
-.378
-.111
.028
-.156
-.321
-.126
1.00
-.253
.457
Per.
-.010
.067
.323
.031
-.085
-.035
-.127
-.200
-.253
1.00
-.497
Res.
-.001
.038
-.518
-.006
-.176
-.148
.109
-.121
.457
-.497
1.00
Note. N = 53. Twitter sentiment (T/Sent); Agility (Agil.); Collaboration (Coll.);
Customer (Cust.); Diversity (Div.); Executive (Exec.); Innovation (Inno.); Integrity
(Integ.); Perception (Per.); Respect (Res.).
Outliers, normality, linearity, homoscedasticity, and independence of
residuals. I evaluated outliers, normality, homoscedasticity, and independence of
residuals by examining the normal probability plot (P-P) of regression standardized
residuals (see Figure 2), a scatterplot of the standardized residuals (see Figure 3), and the
histogram of standardized residuals (see Figure 4). I examined linearity with bivariate
scatterplots for each independent variable in Figures 5-14). My examination revealed no
major violations of outliers, linearity, homoscedasticity, and independence of residuals.
However, there was a violation of the assumption of normality.
Figure 2 depicts the normality results of the distribution around the fitted line. My
study has a positive skew; however, data distribution showed that most of the points
deviated from the line. The study data points' distribution had a wavering deviation from
the probability line, which provided evidence that there was a violation of normality
77
assumptions when reviewing Figure 2. I examined the residuals and found there was a
pattern to support the assumptions. I used SPSS to compute 1,000 bootstrapping samples
to influence violations' possible assumptions, and I reported 95% confidence intervals
from the bootstrap samples.
Figure 2. Normal P-P plot of the regression standardized residuals.
78
Figure 3. Scatterplot of the standardized residuals.
Figure 4. Histogram of the standardized residuals.
79
Figure 5. Scatterplot of Twitter sentiment and business stock price profit/loss.
Figure 6. Scatterplot of Glassdoor agility sentiment and business stock price profit/loss.
80
Figure 7. Scatterplot of Glassdoor collaboration sentiment and business stock price
profit/loss.
Figure 8. Scatterplot of Glassdoor customer sentiment and business stock price
profit/loss.
81
Figure 9. Scatterplot of Glassdoor diversity sentiment and business stock price
profit/loss.
Figure 10. Scatterplot of Glassdoor execution sentiment and business stock price
profit/loss.
82
Figure 11. Scatterplot of Glassdoor innovation sentiment and business stock price
profit/loss.
Figure 12. Scatterplot of Glassdoor integrity sentiment and business stock price
profit/loss.
83
Figure 13. Scatterplot of Glassdoor performance sentiment and business stock price
profit/loss.
Figure 14. Scatterplot of Glassdoor respect sentiment and business stock price profit/loss.
84
I found only 53 of the 100 companies on all three of the archival databases. I
eliminated 47 companies because their information was not available on either the
MITSMR culture 500 or Yahoo finance websites. The small sample size resulted in
potential bias in the study. Potential data bias is an error that results from an insufficient
number of data available by the target population (Fincham, 2008). Because of the small
amount of data present in the data, the reader should judge the results with caution. Table
3 presents descriptive statistics of the independent and dependent variables.
Table 3
Descriptive Statistics for Independent and Dependent Variables
Variable
M
SD
ProfitLoss
.06
.988
TwitterSent
.47
.375
Agility
-.80
.622
Collaboration
.54
.619
Customer
.42
.695
Diversity
.55
.750
Execution
-.42
.553
Innovation
1.64
.779
Integrity
-.65
.556
Performance
-.41
.538
Respect
-.75
.508
Note. N = 53. The MIT SMR Culture 500 Glassdoor sentiment was measured based on
Agility, Collaboration, Customer, Diversity, Execution, Innovation, integrity,
Performance, and Respect.
Inferential Results
Of the expected 100 company archival data, I could only use 53 based on the
data's availability from all three archival databases. The sample size resulted in
significant underpowering of the results. I still used bootstrapping of 1000 samples to
85
evaluate the inferential output. As the number of resampled data sets decreases, the result
introduces more variability into the confidence estimation (Haukoos & Lewis, 2005).
Having a smaller dataset may have had an inverse effect on the measurement of accuracy.
I used a standard multiple regression analysis to examine the relationship between
customer Twitter comments, employee Glassdoor feedback, and successful investing
decisions measured by business stock prices. The model was not significantly able to
predict successful investing decisions measured by business stock prices, F(10, 2993) =
.295, p = .982, R2 = .001, meaning a rejection of the alternative hypothesis. Table 4
presents a summary of the regression analysis.
Table 4
Multiple Regression Analysis Summary for Predictor Variables
Variable
B
SE B
B 95% Bootstrap
CI
t
p
T/Sent.
.608
.024
[-.367, 1.583]
1.043
.221
Agil.
.004
.001
[-.222, .230]
.036
.971
Coll.
.112
.033
[-.170, .414]
.821
.412
Cust.
.084
.025
[-.237, .404]
.511
.609
Div.
.098
.032
[-.206, .403]
.633
.527
Exec.
.091
.022
[-.205, .387]
.600
.548
Inno.
.095
.032
[-.238, .427]
.558
.577
Integ.
.083
.020
[-.212, .377]
.549
.583
Per.
.026
.006
[-.333, .386]
.144
.886
Res.
.016
.004
[-.375, .408]
.083
.934
Note. N = 53. Bootstrap results are based on 1000 bootstrap samples.
Analysis summary. The purpose of this quantitative correlational study was to
examine the relationship between customer Twitter comments, employee Glassdoor
feedback, and successful investing decisions measured by business stock prices. I used
86
multiple linear regression to examine the customer Twitter comments, employee
Glassdoor feedback to predict successful investing decisions measured by business stock
prices. I assessed the assumptions surrounding multiple regression and noted a violation
of the assumption of normality. The model, including Twitter comments, employee
Glassdoor feedback was not able to significantly predict successful investing decisions
measured by business stock prices, F(10, 2993) = .295, p = .982, R2 = .001. Glassdoor
feedback provided useful predictive information about investing decisions measured by
business stock prices. The bootstrap computation relies on the sampling distribution when
calculating confidence intervals but using extremely small samples will interfere with the
computation (Elvarsson et al., 2014). The small original data set was less likely to
represent the intended population. Extremely small data sets make it difficult to compute
valid confidence intervals (Meeker & Escobar, 2014). This analysis concluded that the
predictor variables were not significantly associated with successful investing decisions
measured by business stock prices.
Theoretical discussion of findings. A semi-strong EMH is that all public
information will reflect in real-time stock prices, meaning gaining a significant profit is
not possible when using public information. My findings supported the definition of
semi-strong EMH along with the findings of (Nisar & Yeung, 2017; Heston & Sinha,
2017; Tamrakar, 2016, Yadav et al., 2019b), all of whom found that Twitter post
sentiment did not correlate with stock price changes nor successful decision-making.
However, other studies refute my findings (Bollen et al., 2011, Cwynar et al., 2017,
Daniel et al., 2017, Green et al., 2019, McGurk et al., 2020; Usher et al., 2019, 2020
87
Zhang et al., 2015), and suggest that Twitter post and Glassdoor feedback correlate with
firm performance, stock prices changes and successful decision-making. Moreover, this
study extended the body of research from using the new MIT SMR culture 500 sentiment
data in the context of hypothesis testing with EMH and investment decision-making. The
past peer-reviewed studies that suggest the full conclusion of EMH theory validity are
still questionable and an area of future research.
Applications to Professional Practice
This study's findings have applicability to the financial investment professional
practice by providing investors with information better to understand sentiment analysis
tools' potential misguiding output. The findings of the research, with its limitations,
suggest that sentiment analysis tools may not be effective for enhancing investment
decision-making. The U.S. Securities and Exchange Commission (SEC) suggests that
sentiment tools come with a warning that such tools may cause emotional or risky
investment decision-making (U.S. SEC, 2019). The continued research on this subject is
something to consider as well; while AI is still enhancing, investors may want to try
using sentiment tools sparingly once they are consistently proven to be more useful than
traditional investment decision-making practices.
Implications for Social Change
Improving investment decision-making could increase funds for retirements,
children's education, or other socially desirable uses. The research may help prevent an
investor from using a risky sentiment tool for investment decision-making that may help
them prevent loss of capital. Future warnings of new AI investment tools may bring
88
investors to ask for empirical research to help provide evidence of such tools as beneficial
and profitable.
Warnings from the SEC about sentiment analysis tools can also reevaluate each
year to include the governmental department’s most recent stance. The benefits of
sentiment analysis to traders may be beneficial in other areas of investment decision-
making for evaluating employees' beliefs of diversity, which is a metric found on the
MIT SMR 500 culture website. Such cultural sentiment from employees may also help
investors change their buying behaviors that may change companies' programs to boost
their diversity efforts.
Recommendations for Action
AI sentiment analysis tools relationship to investment decision-making is an on-
going debate for the validity of findings. I recommend investors (a) request empirical
research before using sentiment analysis investment tools, (b) seek professional financial
advice from an accredited institution, (c) and heed warnings from the SEC. The
dissemination and acquiring information for sentiment analysis tools may come from the
Institute of Electrical and Electronics Engineers (IEEE) articles/conferences and SEC
investor bulletins.
Recommendations for Further Research
A limitation of this study was the use of secondary data that may not be accurate.
The data used in this study lacked all of the data available for the year 2016 as well. The
data was accessible through a Twitter API that is the closest a researcher can get to be the
89
data's first handler. Having the data gathered from the source will decrease the chances of
secondary data manipulation.
The next limitation is the amount of data and timeframe of only 79 days in 2016.
The amount of data could be increased to a whole year, or longer and account for
decision-making based on a week, month, year, or decade from stock prices. The
beginning of the data gathering may also start with only the MIT SMR culture 500
database that may ensure more companies having information on all three databases from
Twitter to Yahoo Finance. There are also different types of AI sentiment tools that may
help derive more accurate output that may correlate differently with the stock price data.
Another limitation was from using quantitative research that may be too
generalized and misses the potential depth of understanding when using a qualitative
method. Qualitative studies have strengths when specific humanistic areas of data and
triangulation approaches are necessary to gain quality results (O'Leary, 2017). A future
research area could be to use qualitative interviews with past and current employees,
customers, and professional investors to derive thematic results.
In addition, I would recommend that the data come from other markets and
indexes such as the NYSE, DOW, and any publicly accessible market indexes. I had only
used NASDAQ 100 companies. Many other companies on the index could be used in a
study, or mixing the indexes as a form of random population samples may be useful. I
encourage a researcher to gather the appropriate amount of data to fulfill the power
analysis range to gain a stronger study output.
90
Reflections
My Walden University DBA experience was full of ambiguities from committee
subjective study input to the publication process's frustrations. A professor told me that
my experience is not profound in any way. The road to becoming a doctor is hardly
paved and, at times, directionless. However, Walden University's program does offer
great information that may, at times, be hidden deep in their databases. Moreover, all of
the guidance was there to help me be successful. I also learned that if you do not know
what to ask when needing helping with a study, it increases the chances of receiving
incorrect information.
However, I have met with more faculty members having a positive effect on me
as a student and future doctor. I also used an economic theory that was very interesting
that brought to question a lot about what kind of impact or furthering the subject's
research would even come from my study. My answer was simple: no one knows if a
stock price will go up or down, the past does not necessarily mean predicting the future
when looking at the companies past performances. Moreover, the research completed
here is a small step to understanding that we may not fully understand if social media
sentiment relates to stock prices. However, the AI sentiment tools are constantly being
advanced. One day, they may be able to predict stock price outcomes.
Conclusion
The successful decision-making of an investor affects their ability to gain profit
leading to philanthropy and increasing the cash flow to a local economy. This study's
value to investors was to increase the knowledge of sentiment analysis and investing
91
decisions from customers and employees' posts and feedback. Yadav et al. (2019b)
suggested that investors who can decipher sentiment faster than other investors in the
active market when making investment decisions can make a more substantial profit than
others. However, this study suggests that there is no significant correlation between
sentiment analyzed social media content for only 79 days from 2016. Investors must
focus on factors that contribute to successful investment decisions and find my studies
knowledge to be only a continuation of a conversation, not the end.
92
References
Abutabenjeh, S., & Jaradat, R. (2018). Clarification of research design, research methods,
and research methodology. Teaching Public Administration, 36, 237-258.
doi:10.1177/0144739418775787
Akba, F., Medeni, I., Guzel, M., & Askerzade, I. (2020). Assessment of iterative semi-
supervised feature selection learning for sentiment analyses: Digital currency
markets. 2020 IEEE 14th International Conference on Semantic Computing
(ICSC), 459-463. doi:10.1109/ICSC.2020.00088
Alavi, M., Archibald, M., McMaster, R., Lopez, V., & Cleary, M. (2018). Aligning
theory and methodology in mixed methods research: Before design theoretical
placement. International Journal of Social Research Methodology, 21, 527-540.
doi:10.1080/13645579.2018.1435016
Alekneviciene, V., Kviedaraitiene, L., & Alekneviciute, E. (2018). Semi-strong form
efficiency in the Baltic stock markets under changing economic situation.
Engineering Economics, 29(5), 495-506. doi:10.5755/j01.ee.29.5.19083
Allwood, C. (2012). The distinction between qualitative and quantitative research
methods is problematic. Quality and Quantity, 46, 1417-1429.
doi:10.1007/s11135-011-9455-8
Alostad, H., & Davulcu, H. (2017). Directional prediction of stock prices using breaking
news on Twitter. Web Intelligence (2405-6456), 15(1), 1-17. doi:10.3233/WEB-
170349
Andrianto, Y., & Mirza, A. (2016). A testing of efficient markets hypothesis in Indonesia
93
stock market. Procedia - Social and Behavioral Sciences, 219, 99-103.
doi:10.1016/j.sbspro.2016.04.048
Anestis, M., Anestis, J., Zawilinski, L., Hopkins, T., & Lilienfeld, S. (2014). Equine‐
related treatments for mental disorders lack empirical support: A systematic
review of empirical investigations. Journal of Clinical Psychology, 70, 1115-
1132. doi:10.1002/jclp.22113
Asghar, M., Khan, A., Ahmad, S., Qasim, M., & Khan, I. (2017). Lexicon-enhanced
sentiment analysis framework using a rule-based classification scheme. PLOS
ONE, 12(2), 1-22. doi:10.1371/journal.pone.0171649
Attarwala, A., Dimitrov, S., & Obeidi, A. (2017). How efficient is Twitter: Predicting
2012 U.S. presidential elections using support vector machine via Twitter and
comparing against Iowa Electronic Markets. 2017 Intelligent Systems Conference
(IntelliSys), 646-652. doi:10.1109/IntelliSys.2017.8324363
Ayodele, A., Anthony, O., & Adekemi, O. (2018). Semi-strong form of efficiency of
Nigerian stock market: An empirical test in the context of input and output index.
International Journal of Financial Research, 9(1), 115-120.
doi:10.5430/ijfr.v9n1p115
Aziz, T., Ansari, V., & Alam, M. (2019). Corporate culture and stock returns: evidence
from Indian companies. Journal of Strategy & Management, 12(4), 464.
doi:10.1108/JSMA-01 2019-0009
Babu, A., & Kumar, R. (2015). The impact of sentiments on stock market: A fuzzy logic
approach. The IUP Journal of Applied Finance, 21(2), 22-33.
94
Barberis, N., Shleifer, A., & Vishny, R. (1998). A model of investor sentiment. Journal
of Financial Economics, 49, 307-343. doi:10.1016/s0304-405x(98)00027-0
Barker, C., & Pistrang, N. (2016). Research methods in clinical psychology: An
introduction for students and practitioners (3rd ed.) Wiley Online Library.
doi:10.1002/9781119154082
Button, K., Ioannidis, J., Mokrysz, C., Nosek, B., Flint, J., Robinson, E., & Munafo, M.
(2013). Power failure: Why small sample size undermines the reliability of
neuroscience. Nature Reviews Neuroscience, 14, 365-376. doi:10.1038/nrn3475
Bergsma, K., & Tayal, J. (2019). Short interest and lottery stocks. Financial Management
Wiley-Blackwell, 48(1), 187-227. doi:10.1111/fima.12220
Bernardo, M. (2020). The four-week impact of positive and negative news on stock
performance in correlation to sector and market performance (Doctoral
dissertation). Retrieved from ProQuest Dissertations & Theses Global. (Order No.
27741062)
Bi, S., Liu, Z., & Usman, K. (2017). The influence of online information on investing
decisions of reward-based crowdfunding. Journal of Business Research, 71, 10-
18. doi:10.1016/j.jbusres.2016.10.001
Birbeck, E. & Cliff, D. (2017). Using stock prices as ground truth in Sentiment analysis
to generate profitable trading signals. 2018 IEEE Symposium Series on
Computational Intelligence (SSCI). doi:10.1109/SSCI.2018.8628841
Bishop, P., & Herron, R. (2015). Use and misuse of the Likert item responses and other
ordinal measures. International Journal of Exercise Science, 8, 297-302.
95
Retrieved from https://digitalcommons.wku.edu/ijes
Bollen, J., Mao, H., & Zeng, X. (2011) Twitter mood predicts the stock market. IEEE
Computer Society. 2(1) 1-8. doi:10.1109/MIS.2015.25
Bouteska, A. (2019). Some evidence from a principal component approach to measure a
new investor sentiment index in the Tunisian stock market. Managerial Finance,
46(3), 401-420. doi:10.1108/MF-11-2018-0570
Boyd, B., Gove, S., & Solarino, A. (2017). Methodological rigor of corporate governance
studies: A review and recommendations for future studies. Corporate
Governance: An International Review, 25(6), 384 -396. doi:10.1111/corg.12208
Braun, J., Hausler, J., & Schäfers, W. (2020). Artificial intelligence, news sentiment, and
property market liquidity. Journal of Property Investment & Finance, 38(4), 309-
325. doi:10.1108/JPIF-08-2019-0100
Brutus, S., Aguinis, H., & Wassmer, U. (2013). Self-reported limitations and future
directions in scholarly reports: Analysis and recommendations. Journal of
Management, 39(1), 48-75. doi:10.1177/0149206312455245
Bryman, A., & Bell, E. (2015). Business Research Methods. North York, ON: Oxford
University Press.
Cao, H., Ouyang, H., Li, Y., Li, X., & Chen, Y. (2018). The power law characteristics of
stock price jump intervals: An Empirical and Computational Experimental Study.
ENTROPY, 20(4). doi:10.3390/e20040304
Castellan, C. (2010). Quantitative and qualitative research: A view for clarity.
International Journal of Education, 2(2), 1-14. doi:10.5296/ije.v2i2.446
96
Chamberlain, A. (2015, March 11). Does company culture pay off? Glassdoor. Retrieved
from https://www.glassdoor.com/research/does-company-culture-pay-off-
analyzing-stock-performance-of-best-places-to-work-companies/
Che, S., Zhu, W., & Li, X. (2020). Anticipating corporate financial performance from
CEO letters utilizing sentiment analysis. Mathematical Problems in Engineering,
1-17. doi:10.1155/2020/5609272
Chen, K., Luo, P., Liu, L., & Zhang, W. (2018). News, search and stock co-movement:
Investigating information diffusion in the financial market. Electronic Commerce
Research and Applications, 28, 159-171. doi:10.1016/j.elerap.2018.01.015
Chinas, M. (2018) Efficient Market Hypothesis: Weak form efficiency: An examination of
weak form efficiency [Kindle edition]. Retrieved from
https://read.amazon.com/ref=kcr_app_surl_cloudreader
Chirkov, V., & Anderson, J. (2018). Statistical positivism versus critical scientific
realism A comparison of two paradigms for motivation research: Part 1 A
philosophical and empirical analysis of statistical positivism. Theory &
Psychology, 28(6), 712-736. doi:10.1177/0959354318804670
Cohen, A., & Abedallah, M. (2015). The mediating role of burnout on the relationship of
emotional intelligence and self-efficacy with OCB and performance. Management
Research Review, 38, 2-28. doi:10.1108/mrr-10-2013-0238
Chu, S., Sung, Y. (2015). Using a consumer socialization framework to understand
electronic word-of-mouth (eWOM) group membership among brand followers on
Twitter. Electronic Commerce Research and Applications, 14(4) (2015) 251-260.
97
doi:10.1016/j.elerap.2015.04.002
Cakra, Y., & Distiawan Trisedya, B. (2015). Stock price prediction using linear
regression based on sentiment analysis. 2015 International Conference on
Advanced Computer Science and Information Systems (ICACSIS),147–154.
doi:1109/ICACSIS.2015.7415179
Cwynar, A., Cwynar, W., & Pater, R. (2017). Can social media content increase financial
market returns? Organizacija, 2, 97. doi:10.1515/orga-2017-0007
Kesavan, M., Karthiraman, J., Ebenezer, T., & Adhithvan, S. (2020). Stock market
prediction with historical time series data and sentimental analysis of social media
data. 2020 4th International Conference on Intelligent Computing and Control
Systems (ICICCS), 477-482. doi:10.1109/ICICCS48265.2020.9121121
Krzywinski, M., & Altman, N. (2015). Points of significance: Multiple linear regression.
Nature Methods, 12, 1103-1104. doi:10.1038/nmeth.3665
Kumari, P. (2019). An empirical analysis of stock price behavior around bonus issue
announcement in India. Journal of Management, 6(1) 186-196.
doi:10.34218/JOM.6.1.2019.020
Da Silva, F., França, A., Suassuna, M., de Sousa Mariz, L., Rossiley, I., de Miranda, R.,
Espindola, E. (2013). Team building criteria in software projects: A mix-method
replicated study. Information and Software Technology, 55, 1316-1340.
doi:10.1016/j.infsof.2012.11.006
Daniel, M., Neves, R., & Horta, N. (2017). Company event popularity for financial
markets using Twitter and sentiment analysis. Expert Systems with Applications,
98
71, 111-124. doi:10.1016/j.eswa.2016.11.022
Darena, F., Petrovsky, J., Prichystal, J., & Zizka, J. (2018). Machine learning-based
analysis of the association between online texts and stock price movements.
Inteligencia Artificial: Revista Iberoamericana de Inteligencia Artificial, 21(61),
95-110. doi:10.4114/intartif.vol21iss61pp95-110
de Aragão Batista, A., Maia, U., & Romero, A. (2018). Stock market under the 2016
Brazilian presidential impeachment: a test in the semi-strong form of the efficient
market hypothesis. Revista Contabilidade & Finanças - USP, 29(78), 405-417.
doi:10.1590/1808-057x201805560
Derakhshan, A., & Beigy, H. (2019). Sentiment analysis on stock social media for stock
price movement prediction. Engineering Applications of Artificial Intelligence,
85, 569-578. doi:10.1016/j.engappai.2019.07.002
Dey, S., Wasif, S., Tonmoy, D., Sultana, S., Sarkar, J., & Dey, M. (2020). A comparative
study of support vector machine and Naive Bayes classifier for sentiment analysis
on Amazon product reviews. 2020 International Conference on Contemporary
Computing and Applications (IC3A), 217-220.
doi:10.1109/IC3A48958.2020.233300
Ding, W., Mazouz, K., & Wang, Q. (2019). Investor sentiment and the cross-section of
stock returns: new theory and evidence. Review of Quantitative Finance and
Accounting, 53(2), 493-525. doi:10.1007/s11156-018-0756-z
Elvarsson, B., Taylor, L., Trenkel, V., Kupca, V., & Stefansson, G. (2014). A bootstrap
method for estimating bias and variance in statistical fisheries modelling
99
frameworks using highly disparate datasets. African Journal of Marine Science,
36, 99-110. doi:10.2989/1814232x.2014.897253
Ernst, A., & Albers, C. (2017). Regression assumptions in clinical psychology research
practice-A systematic review of common misconceptions. PeerJ, 5, Article
e3323. doi:10.7717/peerj.3323
Fama, E. (1965). Tomorrow on the New York Stock Exchange. Journal of Business,
38(3), 285-299. doi:10.1086/294788
Fama, E. (1970). Efficient capital markets: A review of theory and empirical work. The
Journal of Finance, 25(2), 383-417.
Feng, J., & Liu, B. (2018). Dynamic impact of online word-of-mouth and advertising on
supply chain performance. International Journal of Environmental Research and
Public Health, 15(1), 69-75. doi:10.3390/ijerph15010069
Fincham, J. (2008). Response rates and responsiveness for surveys, standards, and the
journal. American Journal of Pharmaceutical Education, 72, 43.
doi:10.5688/aj720243
Gao, Z., Ren, H., & Zhang, B. (2020). Googling investor sentiment around the world.
Journal of Financial & Quantitative Analysis, 55(2), 549-580.
doi:10.1017/S0022109019000061
Garcia-Lopez, F., Batyrshin, I., & Gelbukh, A. (2018). Analysis of relationships between
tweets and stock market trends. Journal of Intelligent & Fuzzy Systems, 34, 3337-
3347. doi:10.3233/JIFS-169515
Gillet, R., & Renault, T. (2019). When machines read the web: market efficiency and
100
costly information acquisition at the intraday level. Finance & Bien Commun,
40(2), 7-49. doi:10.3917/fina.402.0007
Graham, J., Harvey, C., Popadak, J., & Rajgopal, S. (2017). Corporate culture: Evidence
from the field (No. w23255). National Bureau of Economic Research.
doi:10.3386/w23255
Green, T., Huang, R., Wen, Q., & Zhou, D. (2019). Crowdsourced employer reviews and
stock returns. Journal of Financial Economics, 134(1), 236-251.
doi:10.1016/j.jfineco.2019.03.012
Groß-Klußmann, A., König, S., & Ebner, M. (2019). Buzzwords build momentum:
Global financial Twitter sentiment and the aggregate stock market. Expert
Systems with Applications, 136, 171-186. doi:10.1016/j.eswa.2019.06.027
Gunay, S. (2019). Impact of public information arrivals on cryptocurrency market: A
case of Twitter posts on ripple. East Asian Economic Review (EAER), 23(2), 149-
168. doi:10.11644/KIEP.EAER.2019.23.2.359
Guo, L., Vargo, C., Pan, Z., Ding, W., & Ishwar, P. (2016) Big social data analytics in
journalism and mass communication. Journalism & Mass Communication
Quarterly, 93(2) 332-359. doi:10.1177/1077699016639231
Gyamfi, E., Kyei, K., & Gill, R. (2017). Market efficiency of African stock markets: A
meta-analysis. Journal of Developing Areas, 51(4), 69-80.
doi:10.1353/jda.2017.0088
Handley, M., Lyles, C., McCulloch, C., & Cattamanchi, A. (2018). Selecting and
improving quasi-experimental designs in effectiveness and implementation
101
research. Annual Review of Public Health, 39, 5-25. doi:10.1146/annurev-
publhealth-040617-014128
Hausler, J., Ruscheinsky, J., & Lang, M. (2018). News-based sentiment analysis in real
estate: a machine learning approach. Journal of Property Research, 35(4), 344.
doi:10.1080/09599916.2018.1551923
Hawaldar, I., Rohit, B., & Pinto, P. (2017). Testing of weak form of efficient market
hypothesis: evidence from the Bahrain Bourse. Investment Management &
Financial Innovations, 14(2), 376-385. doi:10.21511/imfi.14(2-2).2017.09
Hawkins, D., Gallacher, E., & Gammell, M. (2013). Statistical power, effect size and
animal welfare: Recommendations for good practice. Animal Welfare, 22, 339-
344. doi:10.7120/09627286.22.3.339
Hwang, K. (2019). Positivism versus realism: Two approaches of indigenous
psychologies. Journal of Theoretical and Philosophical Psychology, 39(2), 127-
129. doi:10.1037/teo0000117
Heston, S., & Sinha, N. (2017). News vs. sentiment: predicting stock returns from news
stories. Financial Analysts Journal, 73(3), 67-83. doi:10.2469/faj.v73.n3.3
Ho, C., Damien, P., Gu, B., & Konana, P. (2017). The time-varying nature of social
media sentiments in modeling stock returns. Decision Support Systems, 101, 69-
81. doi:10.1016/j.dss.2017.06.001
Hwang, E., & Kim, Y. (2019). Interdependency between the stock market and financial
news. 2019 IEEE International Conference on Big Data, 6074-6076.
doi:10.1109/BigData47090.2019.9006533
102
Iankova, S., Davies, I., Archer-Brown, C., Marder, B., & Yau, A. (2019). A comparison
of social media marketing between B2B, B2C and mixed business models.
Industrial Marketing Management, 81, 169–179.
doi:10.1016/j.indmarman.2018.01.001
Ibrahim, N., Wang, X., Bourne, H., (2017). Exploring the effect of user engagement in
online brand communities: evidence from Twitter. Computers in Human
Behavior, 72, 321-338, doi:10.1016/j.chb.2017.03.005
Ibrahim, N., & Wang, X. (2019). A text analytics approach for online retailing service
improvement: Evidence from Twitter. Decision Support Systems, 121, 37-50.
doi:10.1016/j.dss.2019.03.002
Inamdar, A., Bhagtani, A., Bhatt, S., & Shetty, P. (2019). Predicting cryptocurrency value
using sentiment analysis. 2019 International Conference on Intelligent Computing
and Control Systems (ICCS), 932-934. doi:10.1109/ICCS45141.2019.9065838
Ito, T., Sakaji, H., Izumi, K., Tsubouchi, K., & Yamashita, T. (2017). Development of
sentiment indicators using both unlabeled and labeled posts. 2017 IEEE
Symposium Series on Computational Intelligence (SSCI), 1-8.
doi:10.1109/SSCI.2017.8280918
Jackson, P. (2017). Causal claims and causal explanation in international studies. Journal
of International Relations and Development, 20(4), 689-716.
doi:10.1057/jird.2016.13
Jadon, P., Bhatia, D., & Mishra, D. (2019). A bigdata approach for sentiment analysis of
Twitter data using Naive Bayes and SVM Algorithm. 2019 Sixteenth
103
International Conference on Wireless and Optical Communication Networks
(WOCN), 1-6. doi:10.1109/WOCN45266.2019.8995109
Jain, A., Tripathi, S., Dwivedi, H., & Saxena, P. (2018). Forecasting price of
cryptocurrencies using tweets sentiment analysis. 2018 Eleventh International
Conference on Contemporary Computing (IC3), 1-7.
doi:10.1109/IC3.2018.8530659
Jovanovic, F. (2018). A comparison between qualitative and quantitative histories: the
example of the efficient market hypothesis. Journal of Economic Methodology,
25(4), 291-310. doi:10.1080/1350178X.2018.1529135
Kaplanski, G., Levy, H., Veld, C. & Veld-Merkoulova, Y. (2014). Do happy people make
optimistic investors? Journal of Financial and Quantitative Analysis, 50(1-2),
145-68.
Kessler, L., Barnhart, H., Buckler, A., Choudhury, K., Kondratovich, M.,
Toledano, A., ... & Sullivan, D. (2015). The emerging science of quantitative
imaging biomarkers terminology and definitions for scientific studies and
99 regulatory submissions. Statistical Methods in Medical Research, 24, 9-26.
doi:10.1177/0962280214537333
Kim, E., Jeong, Y., Kim, Y., Kang, K., Song, M. (2016). Topic-based content and
sentiment analysis of Ebola virus on Twitter and in the news. Journal of
Information Science, 42(6) 763-781. doi:/10.1177/0165551515608733
Kršikapa-Rašajski J., & Rankov-Siniša G. (2016). Testing weak form efficiency on the
capital markets in Serbia. Megatrend Revija, 13(1), 265-278.
104
doi:10.5937/MegRev1601265K
Kumar, A., Kumar, A., & Yadav, R. (2017). Stock price jumps and news sentiment: A
case of investor overreaction. Journal of Prediction Markets, 11(2), 20-37.
doi:10.5750/jpm.v11i2.1355
Liau, B., Tan, P. (2014). Gaining customer knowledge in low cost airlines through text
mining. Industrial Management & Data Systems, 114(9) 1344-1359.
doi:10.1108/imds-07-2014-0225
Leedy, P., & Ormrod, J. (2015). Practical research: Planning and design (11th ed).
Essex, England: Pearson Education Unlimited.
Leekha, A., Wadhwa, A., Jain, N., & Wadhwa, M. (2018). Understanding the impact of
news on stock market trends using natural language processing and machine
learning algorithms. International Journal of Knowledge-Based Computer
Systems, 6(2), 23-30. Retrieved from
https://www.academia.edu/40508550/Understanding_the_Impact_of_News_on_S
tock_Market_Trends_Using_Natural_Language_Processing_and_Machine_Learn
ing_Algorithms
Lekhal, M. & Ahmed, O. (2020). Does the adaptive market hypothesis explain the
evolution of emerging markets efficiency? Evidence from the Moroccan financial
market. Heliyon, 6(7), 1-12. doi:10.1016/j.heliyon.2020.e04429
Leković, M. (2018). Evidence for and against the validity of efficient market hypothesis.
Economic Themes, 56(3), 369-387. doi:10.2478/ethemes-2018-0022
Li, B., Chan, K., Ou, C., & Ruifeng, S. (2017). Discovering public sentiment in social
105
media for predicting stock movement of publicly listed companies. Information
Systems, 69, 81-92. doi:10.1016/j.is.2016.10.001
Lydenberg, S. (2016). Integrating systemic risk into modern portfolio theory and practice.
Journal of Applied Corporate Finance, 28(2), 56-61. doi:10.1111/jacf.12175
Malkiel, B. (2003). The efficient market hypothesis and its critics. Journal of Economic
Perspectives, 17(1), 59-82. doi:10.1257/089533003321164958
Malkiel, B., & Fama. (1970). Efficient capital markets: A review of theory and empirical
work. The Journal of Finance, 25(2), 383-417. doi:10.1111/j.1540-
6261.1970.tb00518
Manasseh, C., Ozuzu, C., & Ogbuabor, J. (2016). Semi strong form efficiency test of the
Nigerian stock market: Evidence from event study analysis of bonus issues.
International Journal of Economics and Financial Issues, 6(4), 1474-1490.
doi:10.13140/2.1.4941.3448
Mankar, T., Hotchandani, T., Madhawani, M., Chidrawar, A., & Lifna, C. (2018). Stock
market prediction based on social sentiments using Machine Learning. 2018
International Conference on Smart City and Emerging Technology (ICSCET), 1.
doi:10.1109/ICSCET.2018.8537242
Mäntylä, M., Graziotin, D., & Kuutila, M. (2018). The evolution of sentiment analysis: A
review of research topics, venues, and top cited papers. Computer Science
Review, 27, 16-32. doi:10.1016/j.cosrev.2017.10.002
McGurk, Z., Nowak, A., & Hall, J. (2020). Stock returns and investor sentiment: textual
analysis and social media. Journal of Economics & Finance, 44(3), 458-485.
106
doi:10.1007/s12197-019-09494-4
Meeker, W., & Escobar, L. (2014). Statistical methods for reliability data. Hoboken, NJ:
John Wiley & Sons.
Mertler, C., & Reinhart, R. (2017). Advanced and multivariate statistical methods:
Practical application and interpretation (6th ed.). New York, NY: Routledge;
Taylor & Francis Group
Mishra, P., Pandey, C., Singh. U. & Gupta, A. (2018) Scales of measurement and
presentation of statistical data. National Institute of Health, 21(4), 419-422.
doi:10.4103/aca.ACA_131_18
Narman, H., & Uulu, A. (2020). Impacts of positive and negative comments of social
media users to cryptocurrency. 2020 International Conference on Computing,
Networking and Communications (ICNC), 187-192.
doi:10.1109/ICNC47757.2020.9049693
Nwachukqu, J., & Shitta, O. (2015). Testing the weak-form efficiency of stock markets:
A comparative study of emerging and industrialized economies. International
Journal of Emerging Markets, 10(3), 409-426. doi:10.1108/IJoEM-07-2013-0115
O’Leary, Z. (2017). The essential guide to doing your research project (3rd ed.). London,
UK: SAGE Publications.
Oyebode, O., & Orji, R. (2019). Social media and sentiment analysis: The Nigeria
presidential election 2019. 2019 IEEE 10th Annual Information Technology,
Electronics and Mobile Communication Conference (IEMCON), 0140-0146.
doi:10.1109/IEMCON.2019.8936139
