See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/368507953

The Impact of Twitter Sentiments on Stock Market Trends

Preprint · February 2023

DOI: 10.48550/arXiv.2302.07244

CITATIONS READS
0 2,448

4 authors, including:

Melvin Mokhtari
Isfahan University of Technology
2 PUBLICATIONS 0 CITATIONS

SEE PROFILE

All content following this page was uploaded by Melvin Mokhtari on 18 February 2023.

The user has requested enhancement of the downloaded file.

 ISFAHAN UNIVERSITY OF TECHNOLOGY
DEPARTMENT OF ELECTRICAL AND COMPUTER ENGINEERING

The Impact of Twitter Sentiments on

Stock Market Trends

Authors:
Melvin Mokhtari (melvin.mokhtari@ec.iut.ac.ir)
Ali Seraj
Niloufar Saeedi
Adel Karshenas
 Authors’ Declaration

We declare that the work included in the above paper is original and is an outcome of the
research carried out by the authors indicated in it, and that we are the sole authors of this
work. We would like to certify that there are no known conflicts of interest linked with this
publication and that no financial assistance has been provided for this study that might have
impacted its conclusion.

We also confirm that the work does not infringe on any copyrights or property rights of oth-
ers, including licenses, and it is free from plagiarism. For the inclusion of figures, tables,
animations, or text quotations from copyrighted works (including websites), necessary per-
mission has been obtained from the copyright holders. By doing so, we guarantee that we
have followed the regulations concerning intellectual property.

iii
 Abstract

The Web is a vast virtual space where people can share their opinions, impacting all aspects
of life and having implications for marketing and communication. The most up-to-date and
comprehensive information can be found on social media because of how widespread and
straightforward it is to post a message. Proportionately, they are regarded as a valuable
resource for making precise market predictions. In particular, Twitter has developed into a
potent tool for understanding user sentiment.

This dissertation examines how well tweets can influence stock symbol trends. We analyze
the volume, sentiment, and mentions of the top five stock symbols in the S&P 500 index
on Twitter over three months. Long Short-Term Memory, Bernoulli Naïve Bayes, and Ran-
dom Forest were the three algorithms implemented in this process. Our study revealed a
significant correlation between stock prices and Twitter sentiment.

Keywords: Natural Language Processing, Machine Learning, Sentiment Analysis, Data

Mining, Social Media, Twitter, Stock Market

v
Table of Contents

Authors’ Declaration iii

Abstract & Keywords v

Table of Contents vii

List of Figures xi

xiii

List of Tables xiii

1 Introduction 1
1.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Motivation and Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.4 Document Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2 Related Work 5
2.1 The Analogy Between Emotions and Sentiments . . . . . . . . . . . . . . . 5
2.2 Sentiment Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2.1 Sentiment Analysis Levels . . . . . . . . . . . . . . . . . . . . . . 7
2.2.1.1 Document-Level Sentiment Classification . . . . . . . . 7
2.2.1.2 Sentence-Level Sentiment Classification . . . . . . . . . 7
2.2.1.3 Entity and Aspect Level Sentiment Classification . . . . . 8
2.2.2 Sentiment Classification Techniques . . . . . . . . . . . . . . . . . 8
2.2.2.1 Lexicon-Based Approaches . . . . . . . . . . . . . . . . 8
2.2.2.2 Machine Learning Approaches . . . . . . . . . . . . . . 9

vii
 The Impact of Twitter Sentiments on Stock Market Trends

Supervised Learning . . . . . . . . . . . . . . . . . . . . . 10
Regression . . . . . . . . . . . . . . . . . . . . . . . . 10
Classification . . . . . . . . . . . . . . . . . . . . . . 10
Unsupervised Learning . . . . . . . . . . . . . . . . . . . . 10
Clustering . . . . . . . . . . . . . . . . . . . . . . . . 10
Association . . . . . . . . . . . . . . . . . . . . . . . 10
2.2.2.3 The Hybrid Approach . . . . . . . . . . . . . . . . . . . 10
2.2.3 Algorithmic Approaches . . . . . . . . . . . . . . . . . . . . . . . 11
2.2.3.1 Naïve Bayes Classifier . . . . . . . . . . . . . . . . . . . 11
Advantages of Naïve Bayes Classifier . . . . . . . . . . . . 12
Disadvantages of Naïve Bayes Classifier . . . . . . . . . . . 12
2.2.3.2 Random Forest Classifier . . . . . . . . . . . . . . . . . 13
Advantages of Random Forest Classifier . . . . . . . . . . . 13
Disadvantages of Random Forest Classifier . . . . . . . . . 14
2.2.3.3 Long Short Term Memory Model (LSTM) . . . . . . . . 14
2.3 Applications of Sentiment Analysis . . . . . . . . . . . . . . . . . . . . . 15
Stock Market . . . . . . . . . . . . . . . . . . . . . . . . . 15
E-commerce . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.3.1 Twitter Sentiment Analysis . . . . . . . . . . . . . . . . . . . . . . 16
Politics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
Business . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.4 Data Collection and Preparation . . . . . . . . . . . . . . . . . . . . . . . 16
2.4.1 Twitter Data Collection Methods . . . . . . . . . . . . . . . . . . . 17
2.4.1.1 Twitter API . . . . . . . . . . . . . . . . . . . . . . . . 17
2.4.1.2 Tweepy . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.4.1.3 Twint . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.4.1.4 Preprocessed Datasets . . . . . . . . . . . . . . . . . . . 18

3 Methodological Approaches and Technical Frameworks 19

3.1 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.1.1 Twitter Data Collection . . . . . . . . . . . . . . . . . . . . . . . . 19
3.1.2 Financial Data Collection . . . . . . . . . . . . . . . . . . . . . . 20
3.1.3 Data Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.1.4 Labelling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.2 Algorithmic Implementation . . . . . . . . . . . . . . . . . . . . . . . . . 23

viii
 Table of Contents

3.2.1 Random Forest Classifier . . . . . . . . . . . . . . . . . . . . . . . 24

3.2.2 Naïve Bayes (Bernoulli) Classifier . . . . . . . . . . . . . . . . . . 25
3.2.3 Long short-term memory Model(LSTM) . . . . . . . . . . . . . . 25

4 Performance Evaluation 29
4.1 A Diagrammatic Structure for Performance Evaluation . . . . . . . . . . . 29
4.2 Detailed evaluation of the results . . . . . . . . . . . . . . . . . . . . . . . 30
4.2.1 Evaluation of The Apple Stock ($AAPL) . . . . . . . . . . . . . . 30
4.2.2 Evaluation of the Facebook Stock ($FB) . . . . . . . . . . . . . . . 36
4.2.3 Evaluation of the Microsoft Stock ($MSFT) . . . . . . . . . . . . . 39
4.2.4 Evaluation of the Tesla Stock ($TSLA) . . . . . . . . . . . . . . . 42
4.2.5 Evaluation of the Amazon Stock ($AMZN) . . . . . . . . . . . . . 45
4.2.6 Overall Comparison . . . . . . . . . . . . . . . . . . . . . . . . . 47

5 Conclusion 49
5.1 Results and Discussions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
5.2 Prospects and Future work . . . . . . . . . . . . . . . . . . . . . . . . . . 50

References 51

ix
List of Figures

2.1 The two-dimensional structure of emotions . . . . . . . . . . . . . . . . . 6

2.2 Sentiment Analysis Different Levels . . . . . . . . . . . . . . . . . . . . . 8
2.3 Different Sentiment Classification Techniques . . . . . . . . . . . . . . . . 9
2.4 K-means’ propensity to form clusters of the same size . . . . . . . . . . . . 11
2.5 The assumption of conditional independence in the Naïve Bayes classifier . 12
2.6 An example of how the Random Forest classifier functions . . . . . . . . . 13
2.7 The architecture of the LSTM model . . . . . . . . . . . . . . . . . . . . . 15

3.1 An example of one-hot encoding . . . . . . . . . . . . . . . . . . . . . . . 22

3.2 Sigmoid and ReLU Functions . . . . . . . . . . . . . . . . . . . . . . . . . 26

4.1 The Graph of AAPL Bullishness . . . . . . . . . . . . . . . . . . . . . . . 31

4.2 The Graph of the LSTM Model for AAPL . . . . . . . . . . . . . . . . . . 33
4.3 The Graph of the Naïve Bayes Classifier for AAPL . . . . . . . . . . . . . 34
4.4 The Graph of the Random Forest Classifier for AAPL . . . . . . . . . . . . 35
4.5 The Graph of FB Bullishness . . . . . . . . . . . . . . . . . . . . . . . . . 36
4.6 The Graph of the LSTM Model for FB . . . . . . . . . . . . . . . . . . . . 37
4.7 The Graph of the Naïve Bayes Classifier for FB . . . . . . . . . . . . . . . 37
4.8 The Graph of the Random Forest Classifier for FB . . . . . . . . . . . . . . 38
4.9 The Graph of MSFT Bullishness . . . . . . . . . . . . . . . . . . . . . . . 39
4.10 The Graph of the LSTM Model for MSFT . . . . . . . . . . . . . . . . . . 40
4.11 The Graph of the Naïve Bayes Classifier for MSFT . . . . . . . . . . . . . 40
4.12 The Graph of the Random Forest Classifier for MSFT . . . . . . . . . . . . 41
4.13 The Graph of TSLA Bullishness . . . . . . . . . . . . . . . . . . . . . . . 42
4.14 The Graph of the LSTM Model for TSLA . . . . . . . . . . . . . . . . . . 43
4.15 The Graph of the Naïve Bayes Classifier for TSLA . . . . . . . . . . . . . 43
4.16 The Graph of the Random Forest Classifier for TSLA . . . . . . . . . . . . 44

xi
 The Impact of Twitter Sentiments on Stock Market Trends

4.17 The Graph of AMZN Bullishness . . . . . . . . . . . . . . . . . . . . . . . 45

4.18 The Graph of LSTM Model for AMZN . . . . . . . . . . . . . . . . . . . 46
4.19 The Graph of the Naïve Bayes Classifier for AMZN . . . . . . . . . . . . . 46
4.20 The Graph of the Random Forest Classifier for AMZN . . . . . . . . . . . 47

xii
List of Tables

2.1 The distinctions between emotion and sentiment . . . . . . . . . . . . . . . 6

xiii
Chapter 1

Introduction

This chapter introduces the research’s motivation, describes the problem, and defines the
scope and theoretical approaches. The goals of the research are then presented. Finally, it
highlights this dissertation’s main contributions and outlines the document’s structure.

1.1 Overview
The recent technological revolution has resulted in an unprecedented data deluge that has
fundamentally altered how we view social and economic sciences. Computers and the In-
ternet are now a part of almost everyone’s daily lives. Similar increases in online activity
have been sparked by the continuously growing Internet usage, such as for business or news.
Due to this issue, social media’s emergence has provided a platform for web users to express
their opinions on various topics. Therefore, events and the creation of textual documents
have grown significantly over the past few years on these platforms, and consequently, mas-
sive datasets are being produced by the interaction of technological systems, documenting
collective behavior in previously unthinkable ways. Finally, we can discover the global
population’s interests, worries, and intentions regarding various economic, political, and
cultural phenomena in this massive database of Internet activities.
Twitter is a widely used microblogging and social networking platform where people can
read, write, publish, and update short text messages called tweets [1–4]. In 2021, Twitter’s
global user base was approximately 429.79 million. Furthermore, it is anticipated that by
2025, there will be 497.48 million users worldwide [5]. As a result, this microblogging
and social networking service piqued the interest of researchers from a variety of fields,
including politics, health care, and finance.
Sentiment analysis is the name given to the scientific study of the semantic content of
these tweets. In general, sentiment analysis is a technique for identifying and classifying

1
 The Impact of Twitter Sentiments on Stock Market Trends

a text’s polarity, aiming to determine whether a specific document has a positive or neg-
ative value in accordance with a standard categorization. According to some researchers,
processing Twitter user opinions in the context of financial markets makes it possible for ev-
eryone to gather pertinent information about the stock market and use it to forecast changes
in stock prices. Indeed, since the very beginning of stock markets, investors have always
sought to gain a competitive advantage. Not illegally, such as through insider trading, but
by wisely and sparingly using information.
Here, we present a case study on financial markets, one of the numerous fields in which
data collection, analysis, and modeling are applied. According to our analysis, social net-
work metrics on social aspects are particularly helpful in understanding financial turnovers.
Indeed, collective phenomena such as herd behavior frequently precipitate financial crises
and demonstrate the financial system’s complex nature. As a result, policymakers are ea-
ger to anticipate unusual collective behavior among investors to be able to intervene sooner
when necessary.
The inherent risks and complications of investing in the financial market have led to the
development of techniques that minimize investor losses and maximize returns. These tech-
niques use machine learning algorithms to find connections between individuals’ thoughts
about the financial market and the movements of stock prices. This will make it possible to
predict how the market will change.

1.2 Motivation and Objectives

The impact of Twitter on financial markets is the focus of this dissertation. The research
will particularly examine how tweets about a company may influence its share price trends.
Early attempts to predict the stock market aimed to determine if stock prices could be
predicted. There are two approaches to this: the random walk theory and the efficient market
hypothesis (EMH) [6]. According to the efficient market hypothesis, rather than reflecting
current and historical prices, stock indices primarily reflect news from existing investors.
However, the random walk theory contends that because news breaks at unpredictable times,
predictions can never be trusted to be accurate.
According to research by Qian et al., EMH and the random walk model’s basic assump-
tions have been identified and explained [7]. These theories have demonstrated that some de-
gree of forecasting is feasible using different economic and business indicators. According
to the widespread use of the powerful version of EMH, prices aggregate publicly available
information and immediately reflect the most recent version [8]. It is widely acknowledged
that news influences the macroeconomic movement of the markets. However, research in-
dicates that social media buzz significantly impacts the microeconomic level, especially in
the major indices like the S&P 500 and DJIA [9–12]. It has been found through earlier

2
 Chapter 1. Introduction

research that investor sentiments and decisions ultimately drive the market [7]. Therefore,
a model that considers feelings will provide better microeconomic predictions.
Traders take positions in the market that are held for only a short time, sometimes for
just a few moments, so many investors may quickly join and leave those positions daily.
Consequently, many investors will look for anything that will give them an advantage when
placing market bets. Several great traders have prepared bots based on artificial intelligence
to identify trends in social media without understanding the implications of the more in-
depth knowledge available there.
If we collect tweets and financial data associated with our target stock symbols, process
the data using frameworks to deduce a tweet’s sentiment score, and then compare the results
to changes in the stock price, we can create a useful momentum marker for the share market.
The frameworks will classify tweets by polarity to resolve these concerns. A tweet’s
sentiment can be classified as positive, neutral, or negative, depending on its polarity. After
tweets are categorized, a daily score is computed.
We are particularly interested in the following main research questions:

• What potential effects could tweets about a particular company have on the future
trends of its share?

• Is it possible for the sentiment analysis results to be used to determine whether a

stock’s price is strong or weak in terms of momentum?

Hence, understanding whether sentiment can be used to predict the direction of trends for a
share price is a crucial part of the research.

1.3 Methodology
As implied by the preceding sections, this study’s methodological approach entails gather-
ing data from tweets and then categorizing them based on their sentiment polarity. However,
because the scope of the world’s stock markets is immense, this document will rely on some
of the most frequently occurring companies in the S&P 500 index. This decision was made
for many reasons, including that the S&P 500, one of the world’s most widely used and rel-
evant indices, is recognized as the best measure of US big-cap equities, and its companies
are popular on social media platforms.
The study was performed using the Python programming language. Additionally, Jupyter
Notebook and the PyCharm IDE were used to implement the scripts created throughout the
various modules.
The official Twitter API can be used to gather tweets. However, given the vast amount of
data we seek, this has several limitations and is inappropriate for our case. The fact that there

3
 The Impact of Twitter Sentiments on Stock Market Trends

are few options makes finding Twitter data one of the study’s biggest challenges. Although
there were other options, such as Twint, a comprehensive and reliable Python web scraper,
or Tweepy, a Python library for extracting tweets, we used a preprocessed dataset from the
Kaggle website [13] to get around these constraints.
We used the top 5 index companies of the S&P 500, the hashtag ”# STOCK,” the cahsh-
tag ”$ STOCK,” and a carefully curated and annotated dataset of tweets from April 9th
to July 16th, 2020, to explain the relationship between stock price returns and sentiment
reflected in tweets.

1.4 Document Structure

Five chapters are comprised in this document, each describing the steps required to achieve
the objectives outlined in previous sections. Aside from Chapter 1, which contains some
introductions, the rest of this dissertation is as described below.
The second chapter describes several concepts in the field of sentiment analysis and
provides details on different levels and types of these concepts. The key components are
its associated ideas, the application of Twitter in this setting, and the various algorithmic
working methods, specifically machine learning methods. This chapter has described the
methodology, information on data collection, dataset preparation for sentiment analysis, and
a brief overview of the simulation setup.
The foundational theories and approaches that support the practice of sentiment analysis
on stock market data are presented in Chapter 3. In addition, it provides many concepts,
techniques, and details about two blocks of algorithm implementation and data acquisition.
Chapter 4 provides a detailed evaluation of the trials carried out on the selected five
stocks. Furthermore, it also discusses the relationship between financial facts and senti-
mental bullishness via diagrams to clarify the study’s findings.
Finally, overall conclusions and a summary of this dissertation, as well as possible future
research paths for further development, are presented in Chapter 5.

4
Chapter 2

Related Work

We will discuss sentiment analysis’s state of the art in this chapter. The background and its
underlying ideas are presented in the beginning. Then we take a look at how to use Twitter
as a valuable tool in sentiment analysis. After that, we take a quick look at various algo-
rithmic strategies, specifically machine learning methods. Finally, we will discuss various
approaches to Twitter data collection, including their advantages and disadvantages, and our
chosen approach.

2.1 The Analogy Between Emotions and Sentiments

In daily life, the terms ”emotion” and ”sentiment” are frequently used interchangeably.
Nevertheless, they are treated as two distinct ideas in psychology. The well-known psy-
chological model by Watson and Tellegen depicts the two-dimensional model of emotions
in Figure 2.1 [14]. Psychologically speaking, emotions play a key role in motivating be-
havior. Paul Ekman, a renowned psychologist, proposed six fundamental emotions: happi-
ness, surprise, sadness, anger, fear, and disgust [15]. Later, other psychologists expanded
these fundamentals by including excitement, contempt, pride, embarrassment, and shame.
In contrast, sentiments describe mental states formed as a result of emotion. The defini-
tion emphasizes how sentiment expresses a person’s thoughts that are influenced by their
emotions. Sentiment, as opposed to emotion, is more expansive. The sentiment, which is
exceptionally well organized, generally links actions to an emotion.
In summary, four factors can distinguish emotion from sentiment, as shown in Table 2.1.
Furthermore, all emotions can be viewed as a synthesis of these four sentiment-influencing
variables. Thus, analyzing opinions and identifying sentiments is critical.

5
 The Impact of Twitter Sentiments on Stock Market Trends

Figure 2.1 The two-dimensional structure of emotions

[16]

Definition Connection Dimension Nature

Emotion Psychological states — Psychological Natural
Sentiment Mental states Expression of emotions Social Organized
Table 2.1 The distinctions between emotion and sentiment
[16]

6
 Chapter 2. Related Work

2.2 Sentiment Analysis

Sentiment describes how someone feels, either positively or negatively, about a topic. Opin-
ion mining is another name for sentiment analysis. This is a natural language processing
technique and is the study of a person’s sentiments, opinions, and attitudes towards an en-
tity such as a social event, product, or service. Text mining is the process of converting raw
text with no structure to structured data in order to extract useful information. That is why
studying algorithms that scan through the text and classify that text into different categories
has received much attention.
Because of the popularity of social media platforms such as Twitter, Facebook, and
YouTube, the number of comments containing individuals’ opinions and information has
increased dramatically; as a result, sentiment analysis has emerged as a hot and trending
topic among academics and professionals. Sentiment analysis helps us make decisions about
various subjects based on other people’s opinions. Many companies use sentiment analysis
to offer better products and services, run campaigns, and share promotional deals.

2.2.1 Sentiment Analysis Levels

As shown in Figure 2.2, sentiment analysis consists of three levels of granularity:

• Document Level

• Sentence Level

• Entity and Aspect Level

2.2.1.1 Document-Level Sentiment Classification

This level identifies whether the concept of the entire document is positive or negative at
this point by treating it as a whole. This approach presumes that a document discusses a
single subject. For example, it determines whether a movie or book review is positive or
negative overall.

2.2.1.2 Sentence-Level Sentiment Classification

At this stage, it separates each sentence and classifies its opinion as positive, negative, or
neutral. Unlike the document level, each document can have different opinions about mul-
tiple entities. Each sentence is classified into an objective or subjective type [18]. Subjec-
tive sentences are based on personal perspective and opinion, like ”She has beautiful hair.”
Whereas the objective sentences do not consist of any people’s opinions and are based on
facts like the sentence ”I have black hair.”

7
 The Impact of Twitter Sentiments on Stock Market Trends

Figure 2.2 Sentiment Analysis Different Levels

[17]

2.2.1.3 Entity and Aspect Level Sentiment Classification

This level is also known as the feature-based level and, unlike document and sentence lev-
els, approaches opinions directly. Its goal is to discover the opinions and determine their
sentiments that could be positive, negative, or neutral about a subject or an entity.

2.2.2 Sentiment Classification Techniques

We got familiar with the sentiment analysis before and understand that the classification of
opinions regarding entities is called sentiment analysis. Different strategies are used to col-
lect opinions and associated sentiments; some are pretty sophisticated and time-consuming,
and some are simple and fast.
As shown in Figure 2.3, we will look at them all:

• Lexicon-based approaches

• Machine learning approaches

• The Hybrid approach

2.2.2.1 Lexicon-Based Approaches

Lexicon-based techniques were among the first techniques used for sentiment analysis. Each
word is given a polarity score in lexicon-based systems, and after going through the text,

8
 Chapter 2. Related Work

Figure 2.3 Different Sentiment Classification Techniques

[19]

the system gives a sentiment score to that text. The advantage of this method is that it does
not require data automation [20].
The lexicon-based technique is divided into two sub techniques which are dictionary-
based and corpus-based. Dictionary-based is done using a dictionary of terms like WordNet
[21]. On the other hand, corpus-based does not need a predefined dictionary; instead, it uses
statistical analysis approaches such as K-nearest neighbors.

2.2.2.2 Machine Learning Approaches

Machine learning is a branch of artificial intelligence that uses data as input to make better
judgments rather than being explicitly instructed on what to do.
Machine learning approaches in sentiment analysis [22] can derive opinions and their
sentiment towards an entity. For this purpose, various methods are employed, including
support vector machines (SVM) and Naïve Bayes.
Generally, machine learning techniques are divided into two categories:

• Supervised learning

• Unsupervised learning

9
 The Impact of Twitter Sentiments on Stock Market Trends

Supervised Learning The goal of this technique is to predict outcomes in uncertain sit-
uations.
The model is first trained using data with a known output and label. Then, the system
will learn from this data to make the best decision.
There are two types of supervised learning techniques:

Regression The goal is to predict a real value based on different independent inputs.
For example, determining a budget based on sales.

Classification Placing various items in different categories. This is the approcehe that
we use for sentiment analysis.

Serval algorithms are used to classify objects. For example, a support vector machine
separates every object into two different classes, called binary classification. It tries to find
a line to separate the members of the classes. Then it updates the values of the slope and
parameters of that line based on each piece of data to find the best line for this classification.

Unsupervised Learning Unsupervised learning differs from supervised learning since it

does not base its decisions on a pre-trained model. Instead, it analyzes raw data to make
the optimal judgments by attempting to extract patterns and similarities between the data
without prior knowledge.
Unsupervised learning techniques are also divided into two sub categories:

Clustering This Technique divides unlabeled data into groups with structures based
on similarities and contrasts between the items. Clustering algorithms might be hierarchical,
overlapping, or probabilistic.
K-means clustering is an example of clustering where the data is separated into K groups,
and each set of data closest to each centroid is categorized under that category. Each group’s
size and level of detail can be altered by altering K, like in Figure 2.4.

Association It is the process of identifying patterns and connections among many

elements of a vast body of data.

2.2.2.3 The Hybrid Approach

The hybrid sentiment analysis approach uses both statistical and knowledge-based method-
ologies for polarity detection. As a result, it inherits both stability and excellent accuracy
from the lexicon-based method and machine learning [24].

10
 Chapter 2. Related Work

Figure 2.4 K-means’ propensity to form clusters of the same size

[23]

2.2.3 Algorithmic Approaches

Numerous algorithms have been proposed for extracting sentiments regarding an entity or
topic. Among the most popular algorithms for sentiment analysis are Naïve Bayes, Random
Forest, and Long Short-Term Memory. We shall quickly review these algorithms in detail.

2.2.3.1 Naïve Bayes Classifier

One of the most straightforward and least time-complex supervised algorithms for datasets
containing big data with millions of records is Naïve Bayes. According to Naïve Bayes,
there is no correlation between the incidence of one character in a class and the occurrence
of another feature. Because of this, it is known as a ”Naïve” algorithm.
The Bayes theorem, shown in Figure 2.6, on which the Naïve Bayes algorithm is based,
enables us to calculate the posterior probability, for instance, 𝑃(𝐻|𝐸), using 𝑃(𝐸 |𝐻), 𝑃(𝐻),
and 𝑃(𝐸) where 𝑃(𝐻|𝐸) = 𝑃(𝑒 1 |𝐻) ∗ 𝑃(𝑒 2 |𝐻) ∗ ... ∗ 𝑃(𝑒 𝑛 |𝐻) ∗ 𝑃(𝐻).
To better understand how the Naïve Bayes algorithm works, consider the scenario of
differentiating spam and genuine emails by scanning each email and extracting words and

11
 The Impact of Twitter Sentiments on Stock Market Trends

Figure 2.5 The assumption of conditional independence in the Naïve Bayes classifier
[25]

phrases. In order to determine if an email is spam or not, we create a table that contains
the likelihood of discovering a word inside the message. For example, given that the email
is spam, we compute the likelihood of detecting the term ”offer.” Additionally, we have to
consider the possibility that the email was not a spam email. Remember that we used the
training dataset to calculate these probabilities.
The followings are the advantages and disadvantages of the Naïve Bayes algorithm:

Advantages of Naïve Bayes Classifier

• It has a lower time complexity than other machine learning models.

• It performs better in classification problems.

• It needs less training data.

Disadvantages of Naïve Bayes Classifier

• The zero-frequency problem causes predictions to be ignored for categories with zero
occurrences in the training data and zero probability.

• It is a terrible model for making numerical forecasts and estimates.

• This classifier tends to create assumptions that are mostly not valid in reality.

12
 Chapter 2. Related Work

2.2.3.2 Random Forest Classifier

Random forest is another standard classification algorithm. The moniker ”Random Forest”
refers to a system that builds a lot of random decision trees and uses most of those predictions
to determine the outcome. Each decision tree’s features and data are chosen at random.
As seen in Figure 2.6, the dataset is divided into separate, entirely random decision trees,
and when attempting to forecast an item, it employs a majority vote as the final prediction.
In addition, it could be used to solve regression problems by calculating the mean or average
of each decision tree prediction.

Figure 2.6 An example of how the Random Forest classifier functions

[26]

This algorithm has several advantages and disadvantages like other algorithms:

Advantages of Random Forest Classifier

• Making reasonable forecasts

• Working with enormous datasets

• Superior to decision trees in accuracy predictions

• Handling missing data better than other algorithms

• Being less problematic than a decision tree model in terms of overfitting

13
 The Impact of Twitter Sentiments on Stock Market Trends

Disadvantages of Random Forest Classifier

• High time complexity

• High space complexity

2.2.3.3 Long Short Term Memory Model (LSTM)

Long short-term memory models are a kind of recurrent neural network that is very effective
in handling a large amount of sequential data. New recurrent neural networks differ from
traditional ones because they can take feedback from previous outputs to generate their
current output. The main feature of LSTMs is storing information for an extended period of
time, and they do that using a cell state. The cell states’ primary role is to store information
from the past efficiently [27].
In each iteration, the LSTM must decide:

• What information to add to the cell state

• Delete useless information from the state.

• Produce the output based on the input and report it to the cell.

The LSTM model will not be affected by outliers in making its predictions and can train its
parameters. Figure 2.7 depicts a simple mechanism of this model.

14
 Chapter 2. Related Work

Figure 2.7 The architecture of the LSTM model

[28]

2.3 Applications of Sentiment Analysis

The number of people using social media is increasing every year. That is why millions
of opinions are being shared through applications such as Twitter, so the amount of data
containing opinions about various topics is growing rapidly. Therefore, employing senti-
ment analysis to extract opinions and their polarity has emerged as a trending study with
numerous applications.
Here are some of the applications of sentimental analysis:

Stock Market The stock market is crucial to every country’s economy. Predicting how
a stock will act in the upcoming future helps us make a better investment with our money.
Sentimental analysis extracts opinions about stocks and tries to find the correlation between
the opinions and the stock by gathering different datasets and processing them. With the
help of sentimental analysis, we could try to predict a stock’s trajectory after data collection
and preparation.

E-commerce Various e-commerce businesses use different types of feedback, such as

reviews and comments about their products, to improve the quality of their products and
customer service. Recommender systems also use sentiment analysis to analyze people’s
feelings toward entities and improve their recommendations. For example, Amazon has

15
 The Impact of Twitter Sentiments on Stock Market Trends

used sentiment analysis to improve the quality of the different services it provides.
So, the popularity of sentiment analysis is due to the great interest shown by these ap-
plications.

2.3.1 Twitter Sentiment Analysis

Twitter is one of the most used social media applications worldwide; launched in 2006, it has
more than 400 million active users monthly. The main reason behind twitters creation was to
allow its users to express their opinions about different topics in a tweet which is a text with
less than 280 characters. As a result, Twitter users vary significantly from ordinary people
to highly influential characters like Jeff Bezos and political figures like Barack Obama.
The main difference between Twitter and other social media applications is that every-
thing you post on Twitter is available to the public even if they do not have Twitter; this
feature makes sentiment analysis very effective on Twitter. Health care and politics are
among the applications of sentimental analysis on Twitter.

Politics With the emergence of Twitter, many studies have been conducted to replace tra-
ditional polling methods like telephone poles and internet polls with Twitter sentiment anal-
ysis. Studies showed that algorithms like SVM classifier could help us predict the polls with
high accuracy. Detecting sarcasm about a candidate is the most challenging part. In this
case, positive sentiments are viewed negatively.

Business Many companies use Tweets to extract sentiment and feelings toward their prod-
uct directly or indirectly. This helps companies to improve their product and offer better
services.

2.4 Data Collection and Preparation

Data collection and Preparation is gathering raw data from multiple sources and coveting
that into structured data to use for sentimental analysis. The main job of data preparation is
to clean up into a single format.
In the case of Twitter, there are five steps to data preparation [29–31]:

1. Removing pictures and URLs

2. Removing usernames

16
 Chapter 2. Related Work

3. Removing special characters: Here, punctuation marks must be removed. Another

type of special character that is removed and replaced with space is emoticons. Even
though it may seem like emoticons help discover a tweet’s sentiment, they are too
complex to interpret and be used in sentimental analysis.

4. Removing letters that contain no emotion, such as ”a”, and ”the,” which are also
known as stop characters

5. Converting everything to lower case

2.4.1 Twitter Data Collection Methods

Several methods are available to collect tweets for use by sentiment analysis models. This
study will examine the pros and cons of the different data collation methods such as Twitter
API, Twint web scraper, Tweepy, and preprocessed datasets.

2.4.1.1 Twitter API

We could use the official Twitter API for collecting data. The Twitter API comes in 3 sub-
scriptions, standard, premium, and enterprise, where the standard is free to use. The main
limitation of the Twitter API is that you can only fetch 3200 tweets from the past seven
days, and it also needs an account to use. To overcome these limitations, we can subscribe
to premium services, but they are not free. That is why we prefer to use libraries and tools
such as Twint and Tweepy.

2.4.1.2 Tweepy

Tweepy is a Python library used for collecting Twitter data. It uses the Twitter API for
collecting and gathering tweets. Tweepy has complete documentation from beginner level
to advanced. Moreover, it allows its users to post tweets and retweets and even has the ability
to like tweets.
Despite this, it is a paid application, and its free version also has the same limitations as
the Twitter API.

2.4.1.3 Twint

Twint uses scraping to collect tweets and does not depend on the official Twitter API. With
Twint, you could collect unlimited tweets with no limitations. Nevertheless, the documen-
tation on Twint is incomplete and outdated compared to the recent updates on Twitter and
cannot be used for posting or liking tweets.
The main differences between Twint and Tweepy are as follows:

17
 The Impact of Twitter Sentiments on Stock Market Trends

• Tweepy uses the Twitter API to collect tweets, while Twint uses web scraping to gather
tweets.

• Tweepy has the same limitations as the Twitter API in the number of tweets it can
collect, while on the other hand, Twint has no limitations in the number of tweets it
can gather.

• Due to its excellent documentation, Tweepy is much easier to learn and use. On the
other hand, in the case of Twint, because of its weak documentation, it can be a hassle
to learn and use its features.

• Twint is limited to only collecting tweets and gathering information, while Tweepy
can be used for posting and liking tweets.

2.4.1.4 Preprocessed Datasets

Using preprocessed datasets is the alternative to the Twitter API and other third-party appli-
cations. Employing preprocessed datasets removes the trouble of trying to collect the tweets
but has less flexibility than tweet collecting tools.
For example, the Kaggle website offers many preprocessed datasets that can be used for
different studies.
We have chosen this method for our data and used a semi-preprocessed dataset from the
Kaggle website.

18
 Chapter 3

Methodological Approaches and

Technical Frameworks

3.1 Data
In this section, we explain the details of the datasets that we have used, the purpose of
those specific datasets, the operations that we performed to make data ready for use, and the
procedure of labeling the data.

3.1.1 Twitter Data Collection

We used two datasets collected from twitter.
The first dataset consists of 1.6 million labeled records, containing tweets that are not
necessarily related to a specific topic.
We obtained this dataset from deepnote.com to train our natural language processing
models. Half of the records are labeled as 0 and half as 4. The labels represent the impres-
sion of the tweets. Label 0 indicates a negative vibe and 4 represents a positive attitude.
After building and testing the classifier models, We used the models to label a second
dataset from the Kaggle website. This dataset consists of around 90,000 records of tweets
relevant to stock markets and three columns comprising an id, created date, and full text of
the tweets. This data was collected during a period of three months, from April 9th to July
16th, 2020.
The tweets from this dataset contain the names of more than 13000 stocks. We extracted
the data related to only five stock markets: Facebook (which has recently changed to Meta),
Microsoft, Tesla, Apple, and Amazon, which are the most popular and highly traded ones.
1 top5_dataframe_noLable = data[data['full_text ']. str. contains ('|'.join(
top5_tickers ))]

19
 The Impact of Twitter Sentiments on Stock Market Trends

Listing 1 Extracting data of the top 5 stocks

Then we preprocessed the data and created a separate data frame for each stock:
AAPL_DF = stockDF [ stockDF ['full_text ']. str. contains ('aapl ')]
2 AMZN_DF = stockDF [ stockDF ['full_text ']. str. contains ('amzn ')]
FB_DF = stockDF [ stockDF ['full_text ']. str. contains ('fb ')]
4 MSFT_DF = stockDF [ stockDF ['full_text ']. str. contains ('msft ')]
TSLA_DF = stockDF [ stockDF ['full_text ']. str. contains ('tsla ')]

Listing 2 Separating the data of each stock

3.1.2 Financial Data Collection

We collected the market data from the Yahoo Finance API for five stocks: Meta, Microsoft,
Tesla, Apple, and Amazon.
Yahoo Finance API is a free API with an impressive amount of information and news
on different markets. Extracting data from it is also pretty simple.
1 import yfinance as yf
start = AAPL_DF ['Date ']. iloc [0]
3 end = AAPL_DF ['Date ']. iloc [ -1]
apple_data = yf. download ("AAPL", start =start , end=end)
5 meta_data = yf. download ("META", start =start , end=end)
micro_data = yf. download ("MSFT", start =start , end=end)
7 tesla_data = yf. download ("TSLA", start =start , end=end)
amazon_data = yf. download ("AMZN", start =start , end=end)

Listing 3 Downloading data from Yahoo Finance

This data contains six attributes: Open, High, Low, Close, Adj Close, and Volume. We
only Close, to calculate the amount of 𝑅𝑒𝑡𝑢𝑟𝑛 and check the correlation with twitter data,
respectively.

3.1.3 Data Preprocessing

Data needs preprocessing, both before being used to build a supervised learning model, and
before being labeled by that model.
We explain what we did to preprocess both twitter datasets.
The first important prerequisite to text preprocessing is to clean the text of the tweets
from pronouns and stative verbs, URLs, emojis, usernames, numbers, and punctuations.
We put pronouns and stative verbs in a list called stopwordslist and cleaned the text of the
tweets from them:

20
 Chapter 3. Methodological Approaches and Technical Frameworks

STOPWORDS = set( stopwordlist )

2 def cleaning_stopwords (text):
return " ".join ([ word for word in str(text). split () if word not in
STOPWORDS ])

Listing 4 Cleaning data from redundant words

We removed punctuations:
1 english_punctuations = string . punctuation
punctuations_list = english_punctuations
3 def cleaning_punctuations (text):
translator = str. maketrans ('', '', punctuations_list )
5 return text. translate ( translator )

Listing 5 Removing punctuations

Removed repeating characters, for example converts hiiiii to hi:

1 def cleaning_repeating_char (text):
return re.sub(r'(.) 1+ ', r'1', text)

Listing 6 Cleaning repeating characters

Removed URLs and numbers:

def cleaning_URLs (data):
2 return re.sub('(( www .[^s]+) |( https ?://[^ s]+))',' ',data)
def cleaning_numbers (data):
4 return re.sub(r'[\ W_]+ ',' ',data)

The next step is to tokenize the words in the text so that it will be easier to go through
and process. We changed the text of each tweet into a vector of words.
tokenizer = RegexpTokenizer (r'\w+')

We also stem the words. It means converting the different words with the same root to
their single root. e.g., the root of all the words payment, pay, and paying is the word pay.
Having the stem word instead of these different forms helps the procedure of processing and
learning easier while maintaining the main meaning and context.
1 import nltk
st = nltk. PorterStemmer ()
3 def stemming_on_text (data):
text = [st.stem(word) for word in data]

21
 The Impact of Twitter Sentiments on Stock Market Trends

5 return data

Listing 7 Stemming

Another necessity is to lemmatize the words, meaning we omit all different shapes of one
word using the nltk built-in function WordNetLemmatizer().
1 nltk. download ('wordnet ')
lm = nltk. WordNetLemmatizer ()
3 def lemmatizer_on_text (data):
text = [lm. lemmatize (word) for word in data]
5 return data

Listing 8 Lemmatizing

We applied all the created functions using lambda expression sing the following syntax:
data = data. apply ( lambda x: function (x))

Listing 9 Applying the preprocessing functions

And further did a one-hot encoding on each word, converting each word to a vector to
be used for processing.
One-hot encoding (Figure 3.1) represents each word with a vector containing N values,
N being the total number of words in the database. All the values in this vector are 0, except
the one for which the index belongs to the target word. As the number of words in the
database is enormous, we omitted the least repeated words and kept the most repeated ones
to make processing workable for an extensive database.

Figure 3.1 An example of one-hot encoding

[32]

Once preprocessing is done, we are good to go to fit our model into the data. For this
purpose, we split the data into train and test sections and built our classifier models.

22
 Chapter 3. Methodological Approaches and Technical Frameworks

3.1.4 Labelling
We labelled the data extracted from our second dataset, by the classifier models built using
the first dataset. As mentioned before, the labels of our training data were 0 and 4. For the
sake of simplicity, we replaced label 4 with 1.
Labelling using the Random Forest model:
rfPred = np. where ( rfPred ==4 ,1 , rfPred )
2 stockDF = dup_top5_noLable .copy ()
stockDF ['Label_rf ']= rfPred

Listing 10 Labelling using Random Forest model

Labelling using the Naïve Bayes model:

1 nbPred = BNBmodel . predict ( X_test )
nbPred = np. where ( nbPred ==4 ,1 , nbPred )
3 stockDF ['Label_nb ']= nbPred

Listing 11 Labelling using Naïve Bayes model

Labelling using the LSTM algorithm:

1 test_sequences = tokenizer . texts_to_sequences ( dup_top5_noLable ['
full_text '])
test_padded = pad_sequences ( test_sequences , padding ='post ', maxlen =
max_length )
3 pred_labels = []
prediction = model . predict ( test_padded )
5 for i in prediction :
if i >= 0.5:
7 pred_labels . append (1)
else:
9 pred_labels . append (0)
pred_labels
11 stockDF ['Label_lstm '] = pred_labels

Listing 12 Labelling using LSTM model

3.2 Algorithmic Implementation

In this section, we explain the details of building the classifier models, including the code
and parameters, hyper-parameter tuning considerations, the structure of the implemented
models, checking the accuracy, and the concepts that are necessary to comprehend the code.

23
 The Impact of Twitter Sentiments on Stock Market Trends

3.2.1 Random Forest Classifier

We used random forest classifier model from sklearn. There are two parameters passed to
the classifier model that we wish to explain: the number of estimators and the maximum
depth.
The number of estimators represents the number of decision trees we wish to generate.
A higher number of trees guarantees a more accurate prediction, but increases the amount
of calculations and slows down the processing in return. Therefore, the optimal choice for
this parameter is the largest that our processor can handle.
Max depth represents the maximum depth of each tree. If we do not set a maximum
depth, then the algorithm continues to generate leaves until it runs out of samples, or all of
the leaves are pure.
There are also other critical parameters that can be adjusted such as max_features and
min_sample_leaf which modify the performance and speed inseparably like n_estimators
did. There are also some parameters that can directly modify the speed of training including
n_jobs, random_state, and oob_score.
1

model = RandomForestClassifier ( n_estimators =200 , max_depth =60)

3 model .fit(X_train , y_train )
predictions = model . predict ( X_test )

Listing 13 Buliding Random Forest model

We further use a confusion matrix to observe the result. A confusion matrix is a table
containing four cells: TP (True Positive), TN (True Negative), FP (False Positive), and FN
(False Negative).

• TP indicates the number of accurately predicted values that are positive.

• TN shows the number of values that are predicted as negative and are actually nega-
tive.

• FP keeps count of the negative values that are predicted as positive.

• FN represents the number of positive values that are falsely predicted as negative.

from sklearn . metrics import confusion_matrix

2 confusion_matrix (y_test , predictions )

Listing 14 Showing Confusion Matrix

And checked the accuracy score, representing the percentage of records the label of
which our model has predicted accurately.

24
 Chapter 3. Methodological Approaches and Technical Frameworks

from sklearn . metrics import accuracy_score

2 accuracy_score (y_test , predictions )*100

Listing 15 Accuracy of predictions of the Random Forest model

3.2.2 Naïve Bayes (Bernoulli) Classifier

The implementation can be simply done by calling the Bernoulli model’s built-in fit function
on the train and test data and predicting and checking the accuracy.
The most critical optimization parameters of this algorithm are alpha, fit_prior, and
class_prior.
from sklearn . Naïve_bayes import BernoulliNB
2 BNBmodel = BernoulliNB ()
BNBmodel .fit(X_train , y_train )
4 y_pred1 = BNBmodel . predict ( X_test )
accuracy_score (y_test , y_pred1 )*100

Listing 16 Building Naïve Bayes model

3.2.3 Long short-term memory Model(LSTM)

We built a sequential 4-layer model, starting with 200 features and ending with 1. The first
layer is a neural network with 200 starting nodes, whose values are passed into the middle
nodes to get the values of the 100 ending nodes, whose values are returned as the output.
Obviously, we need the model to return one value only, and that is the predicted value.
The embedding layer works like a lookup table that helps embed each word in the vector
of the total words existing in the dataset. It specifies an integer value for each word and maps
it to the dense vectors. The embedding layer helps embed high-dimensional data into lower-
dimension space.
A bidirectional layer is a combination of two RNNs, one of which data passes through
beginning from the start of the data sequence, and the other through which data moves in
the opposite direction. The need for this kind of traversal can be perceived by the fact that
in text processing, the meaning of a word is not only dependent on the previous words but
also on the following ones.
We use activation functions in the dense layers.
1 import keras
model = keras . Sequential ([
3 keras . layers . Embedding ( vocab_size , embedding_dim , input_length =
max_length ),

25
 The Impact of Twitter Sentiments on Stock Market Trends

keras . layers . Bidirectional ( keras . layers .LSTM (64)),

5 keras . layers . Dense (24 , activation ='relu '),
keras . layers . Dense (1, activation ='sigmoid ')
7 ])

Listing 17 Sequential LSTM layers

Activation functions are necessary for a deep model to learn appropriately as they fix the
neuron values to prevent them from becoming so big or small that they become useless.
Popular activation functions are Relu, sigmoid, and tangent.
ReLU function is considered to be the most reliable activation function to be used in
between the layers. It returns the input itself unless the input is negative, in which case it
returns zero. The sigmoid and tangent activation functions are used in classification prob-
lems.

Figure 3.2 Sigmoid and ReLU Functions

[33]

In order to proceed to the explanation of the code, we need to define the loss function
and optimizer.
The loss function compares the prediction output of the model to the expected values,
which are the actual values, and returns a number, or loss, for each observation. The model
is trained best when the amount of loss is minimized.
The most popular loss functions are mean squared error (MSE), categorical cross-entropy,
and binary cross-entropy. Each is used for a different kind of problem.
MSE works best for regression problems and cross-entropy for classification ones. As
we are expecting our input to have only two values, we use binary cross-entropy.
The optimizer is the algorithm used to update the weights of the neurons using the gra-
dient of the loss function. GSD and Adam are two of the most stable optimizers. The only

26
 Chapter 3. Methodological Approaches and Technical Frameworks

parameter of the optimizer that may require adjustments is the learning rate. If the learning
rate is too small, it searches for the minima of the loss function in very tiny steps, resulting
in a super-slow process of learning. Otherwise, if the learning rate is too large, it may learn
quickly initially but may not find the minima due to the large steps it takes while searching,
resulting in an unstable training process.
model . compile (loss='binary_crossentropy ',
2 optimizer ='adam ',
metrics =[ 'accuracy '])
4 model . summary ()

Listing 18 Compiling the LSTM model

The batch size specifies the number of observations passed to the network at each train-
ing step. The number of epochs assesses the number of times that the full training data is
shown to the neural network.
num_epochs = 2
2 history = model .fit( train_padded , lstm_y_train ,
epochs = num_epochs , verbose =1,
4 validation_split =0.1)

Listing 19 Fitting the LSTM model to the data

Setting the 𝑣𝑒𝑟𝑏𝑜𝑠𝑒 attribute results in seeing the progress bar while running, and
𝑣𝑎𝑙𝑖𝑑𝑎𝑡𝑖𝑜𝑛_𝑠𝑝𝑙𝑖𝑡 represents the fraction of data used for validation to the total amount of
training data.
Eventually, we tested the accuracy of the model along with replacing the label 4 with 1.
prediction = model . predict ( test_padded )
2 lstm_y_test = np.where ( lstm_y_test ==4 ,1 , lstm_y_test )
pred_labels = []
4 for i in prediction :
if i >= 0.5:
6 pred_labels . append (1)
else:
8 pred_labels . append (0)
print (" Accuracy of prediction on test set : ", accuracy_score (
lstm_y_test , pred_labels ))

Listing 20 Checking the accuracy of LSTM model

27
 Chapter 4

Performance Evaluation

4.1 A Diagrammatic Structure for Performance Evalua-

tion
We defined and calculated a variable bullishness (𝐵𝑡 ) for each day, extracted from Twitter
data features [34], as:
1 + 𝑀𝑡 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒
𝐵𝑡 = ln( )
1 + 𝑀𝑡 𝑁𝑒𝑔𝑎𝑡𝑖𝑣𝑒
1 def bullishness (data):
newData = pd. DataFrame ( columns =[ 'Date ','lstm ','rf ','nb'])
3 newData .Date = data. created_at . unique ()
grouped = data. groupby (data. created_at )
5 for date in newData .Date:
dfTemp = grouped . get_group (date)
7 lstmPos = ( dfTemp . Label_lstm ==1).sum ()
lstmNeg = ( dfTemp . Label_lstm ==0).sum ()
9 newData .loc[ newData .Date == dfTemp . created_at . values [0] ,"lstm"]= math.
log ((1+ lstmPos )/(1+ lstmNeg ))
rfPos = ( dfTemp . Label_rf ==1).sum ()
11 rfNeg = ( dfTemp . Label_rf ==0).sum ()
newData .loc[ newData .Date == dfTemp . created_at . values [0] ,"rf"]= math.log
((1+ rfPos )/(1+ rfNeg ))
13 nbPos = ( dfTemp . Label_nb ==1).sum ()
nbNeg = ( dfTemp . Label_nb ==0).sum ()
15 newData .loc[ newData .Date == dfTemp . created_at . values [0] ,"nb"]= math.log
((1+ nbPos )/(1+ nbNeg ))
return newData

Listing 21 Calculating the Bullishness

And extracted a variable 𝑅𝑒𝑡𝑢𝑟𝑛 from the 𝐶𝑙𝑜𝑠𝑒 feature of financial data.

29
 The Impact of Twitter Sentiments on Stock Market Trends

The Opening value of a stock is the value for which the stock is traded at the beginning
of the day, and the Closing value is the price of a stock that has been agreed upon after all
transactions of that day, or simply the price for the last transaction of the day for that stock.
To see how much the value of a stock has grown or dropped in one day relative to the
day before, we calculate a variable Return for each day 𝑡 [34].

𝑅𝑒𝑡𝑢𝑟𝑛𝑡 = (ln 𝐶𝑙𝑜𝑠𝑒 𝑡 − ln 𝐶𝑙𝑜𝑠𝑒 𝑡−1 ) × 100

def calc_Return (data):

2 newData = pd. DataFrame ( columns =[ 'Date ','Values '])
for i in range (1, len(data. index )):
4 newData .loc[i -1] = [data. index [i],( math.log(data. Close .iloc[i]) -math.
log(data. Close .iloc[i -1])) *100]
return newData

Listing 22 Calculating the Return value

Finally, we analyzed the correlation between the financial feature ”Return” and the Twit-
ter data feature ”Bullishness”.

4.2 Detailed evaluation of the results

In this section, we will depict the analysis and look for correlations between the financial
and the Twitter data.

4.2.1 Evaluation of The Apple Stock ($AAPL)

We visualized the difference between the calculated bullishness and the output of each clas-
sification model for the period of three months.
1 AAPL_B = bullishness ( AAPL_DF )
import matplotlib . pyplot as plt
3 plt.plot( AAPL_B .Date , AAPL_B .lstm)
plt.plot( AAPL_B .Date , AAPL_B .nb)
5 plt.plot( AAPL_B .Date , AAPL_B .rf)
plt. legend (["LSTM"," Naïve base"," Random forest "])

Listing 23 Plotting three predictions of one stock

30
 Chapter 4. Performance Evaluation

Figure 4.1 The Graph of AAPL Bullishness

As we can see in Figure 4.1, Naïve Bayes and LSTM models have pretty similar results,
but far from the predicted bullishness of Random Forest model. If we look more precisely,
we can recognize a similar pattern for all classifiers for most of the time intervals.
Here is the diagram of calculated return on the data extracted from Yahoo Finance statis-
tics (graphical results are available in Figure ??):
1 AAPL_R = calc_Return ( apple_data )
plt.plot( AAPL_R .Date , AAPL_R . Values )
3 plt.show

Listing 24 Plotting the amount of Return of one stock

Then we merged the financial and Twitter results into one data frame using 𝐹𝑖𝑛𝑅𝑒𝑠:
1 import time
import datetime
3 def finalRes ( return_data , bullishness_data ):
fin = pd. DataFrame ( columns =[ 'Date ','lstm ','rf','nb ','Values '])
5 l =[]
for i in range (1, len( return_data . index )):
7 for j in range (0, len( bullishness_data . index )):
if datetime . datetime . strptime ( bullishness_data .Date.iloc[j],"%Y -%m -%d"
) >= return_data .Date.iloc[i]:
9 l. append ([ bullishness_data .Date.iloc[j -1] , bullishness_data .lstm.iloc[j
-1], bullishness_data .rf.iloc[j -1] , bullishness_data .nb.iloc[j -1] ,
return_data . Values .iloc[i]])

31
 The Impact of Twitter Sentiments on Stock Market Trends

break
11 fin = pd. DataFrame (l, columns =[ 'Date ','lstm ','rf ','nb ','Values '])
return fin
13 res = finalRes (AAPL_R , AAPL_B )

Listing 25 Merging the prediction and Return value into one dataset

We further used the merged data to compare each classifier output to the real market
events by visualization.
1 def drawAllForStock (data ,str):
bull = bullishness (data)
3 start = bull['Date ']. iloc [0]
end = bull['Date ']. iloc [ -1]
5 stock_data = yf. download (str , start =start , end=end)
stock_data .drop( apple_data . columns . difference ([ 'Close ']) , 1, inplace =
True)
7 ret = calc_Return ( stock_data )
res = finalRes (ret ,bull)
9 drawPlot (res.Date ,res.lstm ,res.Values ,"lstm model ")
drawPlot (res.Date ,res.rf ,res.Values ," Random forest model ")
11 drawPlot (res.Date ,res.nb ,res.Values ," Naïve Bayes model ")

Listing 26 Draw the comparison plots of Return and all predictions

And drew and saved the diagrams in one block:

1 drawAllForStock (AAPL_DF ,'AAPL ')
figs = [plt. figure (n) for n in plt. get_fignums ()]
3 i=0
for fig in figs:
5 fig. savefig (f" apple_RB {i}. png", dpi =300)
i = i+1

Listing 27 Drawing and saving the open figures

32
 Chapter 4. Performance Evaluation

Figure 4.2 The Graph of the LSTM Model for AAPL

Considering Figure 4.2, for most days, the positivity or negativity of the Return approves
of the predicted bullishness of the market, and there are close predictions on some days, such
as June 25th to July 2nd. The gap between the prediction and the actual return after this date
is quite understandable. As the stock has been growing rapidly, people would expect it to
continue likewise, resulting in many positive comments. Nevertheless, the amount of the
Return has dropped unexpectedly. The predictions are also pretty close after this gap to July
15th.

33
 The Impact of Twitter Sentiments on Stock Market Trends

Figure 4.3 The Graph of the Naïve Bayes Classifier for AAPL

Figure 4.3 depicts that the Naïve Bayes classifier has probably predicted the worst.
Though the predictions of this model are pretty similar to the Naïve Bayes’, the concur-
rent gaps are larger for this classifier. One interesting point to take into consideration is that
on some days, the amount of Return is pretty close to the prediction for tomorrow, meaning
that there has been a closer relationship between the positivity of people and the amount of
Return on the same day than the day after.

34
 Chapter 4. Performance Evaluation

Figure 4.4 The Graph of the Random Forest Classifier for AAPL

Random Forest has predicted the most satisfactory, especially from June 18th on. (Fig-
ure 4.4)

35
 The Impact of Twitter Sentiments on Stock Market Trends

4.2.2 Evaluation of the Facebook Stock ($FB)

Figure 4.5 The Graph of FB Bullishness

As shown in Figure 4.5, like the Apple stock, Naïve Bayes and LSTM outputs are more
similar compared to Random Forest. We can observe some similarities between the patterns.

36
 Chapter 4. Performance Evaluation

Figure 4.6 The Graph of the LSTM Model for FB

In Figure 4.6, we can see that there are some lucky days that predictions are close, but
in most cases, the predictions are quite unreliable. The closest shot time interval is similar
to that of Apple’s, from June 23th to 29th.

Figure 4.7 The Graph of the Naïve Bayes Classifier for FB

37
 The Impact of Twitter Sentiments on Stock Market Trends

Naïve Bayes’ predictions (Figure 4.7) are slightly closer than those of LSTM. The results
are the most satisfactory from May 1st to 5th and June 24th to July 7th.

Figure 4.8 The Graph of the Random Forest Classifier for FB

As it is visible in Figure 4.8, many of the predictions of Random Forest are far away
from the actual stock Returns, and a few are pretty close, making it difficult to evaluate the
job. The closest predictions belong to the first days of July, in which the predictions of the
other classifiers, even the ones belonging to the other stocks, have also been close.

38
 Chapter 4. Performance Evaluation

4.2.3 Evaluation of the Microsoft Stock ($MSFT)

Figure 4.9 The Graph of MSFT Bullishness

Naïve Bayes and LSTM outputs are more similar compared with Random Forest (Fig-
ure 4.9). Many similarities between the patterns can be observed in all three. All classifiers
agree on the peak points.

39
 The Impact of Twitter Sentiments on Stock Market Trends

Figure 4.10 The Graph of the LSTM Model for MSFT

Figure 4.10 depicts that there are some patterns that are predicted closely in LSTM
Model. However, some predictions are far away from the actual values.

Figure 4.11 The Graph of the Naïve Bayes Classifier for MSFT

As described through Figure 4.11, most predictions of the Naïve Bayes Classifier are

40
 Chapter 4. Performance Evaluation

wrong, even concerning the positivity or negativity of the Return.

Figure 4.12 The Graph of the Random Forest Classifier for MSFT

The predictions of the Random Forest classifier are not that satisfactory Figure 4.12, but
closer than the other two.

41
 The Impact of Twitter Sentiments on Stock Market Trends

4.2.4 Evaluation of the Tesla Stock ($TSLA)

Figure 4.13 The Graph of TSLA Bullishness

Considering Figure 4.13, Naïve Bayes and LSTM outputs are more similar compared with
Random Forest. Many similarities between the patterns can be observed in all three. All
classifiers agree on the peak points.

42
 Chapter 4. Performance Evaluation

Figure 4.14 The Graph of the LSTM Model for TSLA

There are a few patterns that are predicted closely in LSTM model (Figure 4.14). How-
ever, most predictions are far from the actual values.

Figure 4.15 The Graph of the Naïve Bayes Classifier for TSLA

By analyzing (Figure 4.15), we can understand that the Naïve Bayes has done better than

43
 The Impact of Twitter Sentiments on Stock Market Trends

LSTM in this case, most of the peaks and minimums are predicted accurately.

Figure 4.16 The Graph of the Random Forest Classifier for TSLA

Random Forest predictions are the closest (Figure 4.16), and this diagram has the closest
predictions of all, with only a few errors.

44
 Chapter 4. Performance Evaluation

4.2.5 Evaluation of the Amazon Stock ($AMZN)

Figure 4.17 The Graph of AMZN Bullishness

The three models agree on most of the rising or falling patterns as shown in Figure 4.17.
Random Forest has predicted the most optimistic and Naïve Bayes the most pessimistic.
LSTM is somewhere in between.

45
 The Impact of Twitter Sentiments on Stock Market Trends

Figure 4.18 The Graph of LSTM Model for AMZN

There are some patterns that are predicted very closely in LSTM model (Figure 4.18).
However, some predictions are far away from the actual values.

Figure 4.19 The Graph of the Naïve Bayes Classifier for AMZN

The predictions of Naïve Bayes are not that satisfactory as it seems in Figure 4.19, yet

46
 Chapter 4. Performance Evaluation

still better than LSTM’s.

Figure 4.20 The Graph of the Random Forest Classifier for AMZN

Figure 4.20 depicts that the Random Forest’s performance is similar to Naïve Bayes, on
some days Naïve Bayes’ results are closer, and on some Random Forest’s.

4.2.6 Overall Comparison

Random Forest predictions are the closest of all. LSTM and Naïve Bayes performances were
quite similar. Sometimes LSTM did better, and in some cases, Naïve Bayes outperformed.
Considering all cases, Naïve Bayes did slightly better. We must take into account that our
parameters were not set with exact care in building the models, along with other possible
errors in data or parameter calculation.

47
Chapter 5

Conclusion

The key results of this dissertation, along with potential future research directions, are out-
lined in this chapter.

5.1 Results and Discussions

Many outstanding pieces of research have been reviewed in the increasingly popular field of
sentiment analysis, and a variety of cutting-edge works have been studied in this document.
Here, we predict how the US stock market will move in the future by analyzing the
polarity of Twitter posts about the market. We presented various stages and strategies in
each sentiment analysis technique. Additionally, we looked into the several forms of data
and tools that might have been used in the sentiment analysis study and offered suggestions
for their advantages and disadvantages. In general, this study seeks to determine whether
there is any connection between public opinion ratings and stock price changes.
This study also discusses identifying the most appropriate sentiment methodologies to
utilize in such a scenario and determining the most relevant variables for predicting stock
price fluctuations. However, filtering and processing the dataset in the data portion proved
difficult and time-consuming. For instance, a significant portion of tweets occasionally
contained spam. Therefore, we also spent considerable time labeling and tokenizing tweets.
After the data had been correctly organized and cleaned, various sentiment analysis
models and algorithms were used, and their evaluation was conducted.
Specifically, we examined the average sentiment values of this Twitter dataset using
Long Short-Term Memory, Bernoulli Naïve Bayes, and Random Forest classifiers. Then,
we compared the results to see which model was the most accurate, showing that all these
algorithms correlate.
Finally, the primary assessment initiatives were summarized in chapter three, along with
a thorough analysis of our model’s outcomes for the five stock symbols.

49
 The Impact of Twitter Sentiments on Stock Market Trends

According to the outcome of stocks’ evaluation in chapter 4, the sentiments of tweets

significantly impact share prices, and the results of the algorithms have demonstrated a
meaningful influence on the trends in the financial markets.
In other words, we concluded that tweets about a specific company might affect its share
price performance. Of course, by stating this, we are not claiming an always direct connec-
tion, but data has shown that tweet sentiments often cause fluctuations in stock prices and
market cap.

5.2 Prospects and Future work

Although the overall goals of the dissertation were met, there are still a few areas that may
be improved. It would be intriguing to see this research expanded upon in order to enhance
a sentiment indicator used as a momentum signal for the stock market.
In other words, we think there is room for improvement in the precision of these machine
learning methods. Here is a list of suggestions that we believe could be helpful in this regard.

• Although Twitter has many international users, we solely concentrate on English sen-
tences. Our method should work to categorize sentiment in other languages as well.

• Neutral tweets cannot be disregarded; neutral sentiment requires adequate consider-

ation. To obtain more accurate results, the same criteria must be taken into account
for audio files, movies, and photos.

• Current algorithms categorize the general tone of a tweet. The polarity of a tweet
can vary depending on your interpretation of the tweet. For example, the sentiment
in the tweet ”England beats Germany” is favorable for England and unfavorable for
Belgium. Semantics could be useful here. The classification would be done in accor-
dance with which noun is primarily connected with the verb based on the results of
a semantic role labeler. This may allow ”Germany beats England” to be categorized
differently as ”England beats Germany.”

• It is important to extend the study period to capture a wider variety of tweets and
events, overcoming the limitations of specific time events like the COVID-19 pan-
demic. The productivity of the models may also be better understood by examining
them in the context of various economic conditions, such as booms and recessions.

• An artificial neural network can be used to analyze sentimental tweets more accu-
rately.

50
References

[1] M. Graham, S. A. Hale, D. Gaffney, Where in the world are you? geolocation and language
identification in twitter, The Professional Geographer 66 (2014) 568–578.
[2] V. D. Nguyen, B. Varghese, A. Barker, The royal birth of 2013: Analysing and visualising
public sentiment in the uk using twitter, in: 2013 IEEE International Conference on Big Data,
IEEE, 2013, pp. 46–54.
[3] A. Java, X. Song, T. Finin, B. Tseng, Why we twitter: understanding microblogging usage and
communities, in: Proceedings of the 9th WebKDD and 1st SNA-KDD 2007 workshop on Web
mining and social network analysis, 2007, pp. 56–65.
[4] G. Ranco, D. Aleksovski, G. Caldarelli, M. Grčar, I. Mozetič, The effects of twitter sentiment
on stock price returns, PloS one 10 (2015) e0138441.
[5] J. Degenhard, Forecast of the number of twitter users in the world from 2017 to 2025, 2021.
URL: https://www.statista.com/forecasts/1146722/twitterusersintheworld.
[6] H. Miao, S. Ramchander, J. K. Zumwalt, Information driven price jumps and trading strategy:
Evidence from stock index futures, SSRN eLibrary (2011).
[7] B. Qian, K. Rasheed, Stock market prediction with multiple classifiers, Applied Intelligence
26 (2007) 25–33.
[8] B. G. Malkiel, The efficient market hypothesis and its critics, Journal of economic perspectives
17 (2003) 59–82.
[9] H. Mao, S. Counts, J. Bollen, Predicting financial markets: Comparing survey, news, twitter
and search engine data, arXiv preprint arXiv:1112.1051 (2011).
[10] T. O. Sprenger, A. Tumasjan, P. G. Sandner, I. M. Welpe, Tweets and trades: The information
content of stock microblogs, European Financial Management 20 (2014) 926–957.
[11] E. Gilbert, K. Karahalios, Widespread worry and the stock market, in: Fourth International
AAAI Conference on Weblogs and Social Media, 2010.
[12] J. Bollen, H. Mao, X. Zeng, Twitter mood predicts the stock market, Journal of computational
science 2 (2011) 1–8.
[13] B. Taborda, A. De Almeida, J. C. Dias, F. Batista, R. Ribeiro, Stock market
tweets data, 2021. URL: https://www.kaggle.com/datasets/utkarshxy/
stock-markettweets-lexicon-data.
[14] D. Watson, A. Tellegen, Toward a consensual structure of mood., Psychological bulletin 98
(1985) 219.
[15] P. Ekman, Are there basic emotions? (1992).

51
 The Impact of Twitter Sentiments on Stock Market Trends

[16] L. Yue, W. Chen, X. Li, W. Zuo, M. Yin, A survey of sentiment analysis in social media,
Knowledge and Information Systems 60 (2019) 617–663.
[17] J. M. G. F. F. Sacramento, Sentiment Analysis in the Stock Market Based on Twitter Data,
Ph.D. thesis, ISCTE-Instituto Universitario de Lisboa (Portugal), 2021.
[18] A. Yadav, D. K. Vishwakarma, Sentiment analysis using deep learning architectures: a review,
Artificial Intelligence Review 53 (2020) 4335–4385.
[19] W. Medhat, A. Hassan, H. Korashy, Sentiment analysis algorithms and applications: A survey,
Ain Shams engineering journal 5 (2014) 1093–1113.
[20] Z. Drus, H. Khalid, Sentiment analysis in social media and its application: Systematic literature
review, Procedia Computer Science 161 (2019) 707–714.
[21] A. Agarwal, B. Xie, I. Vovsha, O. Rambow, R. J. Passonneau, Sentiment analysis of twitter
data, in: Proceedings of the workshop on language in social media (LSM 2011), 2011, pp.
30–38.
[22] N. C. Dang, M. N. Moreno-García, F. De la Prieta, Sentiment analysis based on deep learning:
A comparative study, Electronics 9 (2020) 483.
[23] A. Teixeira-Pinto, J. Harezlak, 2021. URL: https://bookdown.org/tpinto_home/
Unsupervised-learning/k-means-clustering.html.
[24] I. El Alaoui, Y. Gahi, R. Messoussi, Y. Chaabi, A. Todoskoff, A. Kobi, A novel adaptable
approach for sentiment analysis on big social data, Journal of Big Data 5 (2018) 1–18.
[25] B. Gamal, Naïve bayes algorithm, 2021. URL: https://medium.com/analytics-vidhya/
na%C3%AFve-bayes-algorithm-5bf31e9032a2.
[26] O. Mbaabu, 2020. URL: https://www.section.io/engineering-education/
introduction-to-random-forest-in-machine-learning/.
[27] K. Gopalakrishnan, F. M. Salem, Sentiment analysis using simplified long short-term memory
recurrent neural networks, arXiv preprint arXiv:2005.03993 (2020).
[28] MATLAB, YouTube, 2018. URL: https://www.youtube.com/watch?v=5dMXyiWddYs.
[29] A. Go, R. Bhayani, L. Huang, Twitter sentiment classification using distant supervision,
CS224N project report, Stanford 1 (2009) 2009.
[30] S. V. Kolasani, R. Assaf, Predicting stock movement using sentiment analysis of twitter feed
with neural networks, Journal of Data Analysis and Information Processing 8 (2020) 309–319.
[31] G. A. Ruz, P. A. Henríquez, A. Mascareño, Sentiment analysis of twitter data during critical
events through bayesian networks classifiers, Future Generation Computer Systems 106 (2020)
92–104.
[32] G. Novack, Building a one hot encoding layer with ten-
sorflow, 2020. URL: https://towardsdatascience.com/
building-a-one-hot-encoding-layer-with-tensorflow-f907d686bf39.
[33] S. Sharma, Activation functions in neural networks, 2021. URL: https://
towardsdatascience.com/activation-functions-neural-networks-1cbd9f8d91d6.
[34] T. Rao, S. Srivastava, Twitter sentiment analysis: How to hedge your bets in the stock markets,
in: State of the art applications of social network analysis, Springer, 2014, pp. 227–247.

52

View publication stats