Politics, Artificial Intelligence, Twitter and Stock Return : An Interdisciplinary Test for Stock Price Prediction Based on Political Tweets

Politics, Artificial Intelligence,

Twitter and Stock Return

MASTER THESIS WITHIN: Business Administration NUMBER OF CREDITS: 15 ECTS

PROGRAMME OF STUDY: International Financial Analysis AUTHOR: Reamflar Elvio Estebano Troeman & Lisa Fischer JÖNKÖPING May 2020

An Interdisciplinary Test for Stock Price Prediction

Based on Political Tweets

i

Acknowledgements

We want to express our thankfulness to our tutor Haoyong Zhou, PhD for his input and valuable shared knowledge and Toni Duras, PhD for his unquenchable desire to help and pedagogical input with respect to the understanding of STATA coding. We also would like to thank Agostino Manduchi, PhD for his valuable conversations and inspiration through suggested articles. Moreover, we would like to thank Tweet Binder for their excellent cooperation and finally our seminar group, especially our opponents Chaozhong Ma and Jie Qin for their harmonious cooperation and elevated feedbacks.

Reamflar Elvio Estebano Troeman Lisa Fischer

ii

Master Thesis in Business Administration

Date: 2020-05-18

Key terms: Stock Price Prediction, Politics, Efficient Market Hypothesis, Twitter, Artificial Intelligence, Sentiment Analysis, Event Study

Abstract

As the world is gravitating toward an information economy, it has become more and more critical for an investor to understand the impact of data and information. One of the sources of data that can be converted into information are texts from microblogging platforms, such as Twitter. The user of such a microblogging account can filtrate opinion and information to millions of people. Depending on the account holder, the opinion or information originated from the designated account may lead to different societal impact. The microblogging scope of this investigation are politicians holding a Twitter account. This investigation will look into the relationship between political tweets' sentiment and market movement and the subsequent longevity of such an effect. The classified sentiments are positive or negative. The presence of artificial intelligence is vital for a data-driven investigation; in the context of this investigation, artificial intelligence will be used to classify the sentiment of the political tweet. The methods chose to assess the impact of a political tweet and market movement is event-study. The impact is expressed in either a positive or a negative cumulative abnormal return subsequent to the political tweet. The findings of the investigation indicate that on average, there is no statistical evidence that a political tweets' sentiment leads to an abnormal return. However, in specific cases, political tweet leads to abnormal return. Moreover, it has been determined that the longevity of the effect is rather short. This is an interdisciplinary approach that can be applied by individual and institutional investors and financial institutions.

iii

Table of Contents

1.

Introduction ... 1

1.1 Background ... 1

1.2 Course of Investigation ... 3

1.3 Aim of the Research ... 3

2.

Literature Review ... 5

2.1 Twitter Use in Politics ... 5

2.2 Stock Market and Politics ... 6

2.3 Stock Price Prediction... 8

2.3.1 Efficient Market Hypothesis ... 8

2.3.2 Information and the Stock Market ... 9

2.4 Twitter and Stock Price Forecasting ... 10

2.5 Twitter Sentiment Analysis ... 12

2.6 Hypothesis Development ... 13

3.

Data ... 16

4.

Methodology ... 22

4.1.1 Definition of Event and Event Window ... 22

4.1.2 Normal and Abnormal Returns ... 23

5.

Results and Robustness Test ... 27

5.1 Hypothesis Testing ... 27

5.2 Results for Sub-Dataset ... 29

5.3 Robustness Test ... 35

6.

Tweet Analysis... 39

6.1 Tweet and Market Noise ... 39

6.2 NLP and Market Behaviour ... 39

6.3 Time and Contra Direction ... 41

iv

6.5 Tweet Time and Market Activity ... 44

6.6 Relativity Theory in Political Microblogging ... 45

6.7 Democracy and Microtargeting ... 46

7.

Conclusion ... 47

v

Figures

Figure 4.1 Timeline for event study. ... 24

Figure 6.1 Stakeholder groups. ... 41

Figure 6.2 Example of politicians’ tweet with contra-direction effect. ... 42

Figure 6.3 Example of politicians’ tweet with no contra-direction effect... 42

Figure 6.4 Example of politicians’ tweet with contra-direction effect with a negative initial sentiment. ... 43

Figure 6.5 UTC time zones and international trading hours. ... 45

Tables

Table 3.2 Summary of results sentiment analysis. ... 19

Table 3.3 Descriptive statistics of companies’ stock returns... 20

Table 3.4 Descriptive statistics of the indices. ... 21

Table 5.1 CAAR for all events. ... 27

Table 5.2 CAAR for positive sentiment. ... 28

Table 5.3 CAAR for negative sentiment. ... 28

Table 5.4 Results event study for single events. ... 29

Table 5.5 Number of significant observations for each politician. ... 35

Table 5.6 CAAR for all events robustness test. ... 36

Table 5.7 CAAR positive sentiment robustness test. ... 36

Table 5.8 CAAR negative sentiment robustness test. ... 37

Table 5.9 Number of significant observations robustness test. ... 38

Appendix

Appendix 2 – Sample Tweets ... 59

Appendix 3 – Stata Codes ... 68

Appendix 4 – Output Stata: Test Across All Events ... 72

1

1. Introduction

______________________________________________________________________ This chapter covers the background of the investigation and the research question will be presented as well as the course and aim of the study.

______________________________________________________________________

1.1 Background

Humanity is a creature of continuous evolution, and this is also reflected in all life aspects that human beings are related to, from flying with a propeller to a jet engine and sending a post to text messages. In the last four decades, societies around the world have been reshaped by the so-called information technology (IT). The dawn of IT evolution started in the '70s in the four categories evolution of computers, storage and displays technologies as well as software (Kumar, 2014) and has led to a revolution in the pace of every economic activity around the world.

Nowadays, the society has become enormously dependent on the technology and the data produced by it. In value terms, it has been determined that information (data) is not solely data, but it has become a commodity and is by now more valuable than one of the most precious commodities, oil (The economist, 2017). This is why data has become very important in a wide variety of sectors, such as the financial industry. Furthermore, this sector has by now become very dependent on data. However, what kind of data are we talking about? The natural ways to extract data nowadays are newspapers, social media, traditional media such as TV and Radio as well as search engines such as Google or Bing. In these data streams, the data is, in most cases, either an interpretation of a report from an organisation, a speech or interview with a person or an opinion of a writer or journalist. Depending on the content, these data can influence the decision of an investor or a household which subsequently can lead, for example, to an impact on the stock performance of a certain company.

The knowledge circulating in a society is considered as information, because it comes from a perspective and it is organised, therefore it is not considered as data. Data is raw and free of perspective. As mentioned before, there are multiple streams providing information, and one of these is Twitter. In recent history, Twitter has become an

2

authoritative source of information, coming from any individual, celebrity, company or politician.

In this research, we will focus on political tweets, since politicians are potent individuals in society. Besides traditional media, such as TV stations or an official press release, politicians choose to give their opinion on Twitter. One example is the tweet of President Hassan Rouhani of Iran in January 2020, stating that it was Iranian arm forces that have made a mistake to take the Ukrainian Boeing 737 down (Rouhani, 2020). Subsequent the tweet of President Hassan Rouhani, Boeing stock prices have been corrected. This reaction raises the question, whether the tweet had a direct impact on Boeing price and thus if political tweets affect indices and commodity prices in general, and what is the lifespan of the impact? If so, can a subset of commodity, in our case the tweet as data, more generally help predict stock returns? Can the companies addressed in the tweets take advantage of it? These considerations, therefore, lead us to our main research questions for this investigation: “Can a stock return be predicted by political tweets' sentiment and, what is the lifespan of theimpact?”.

In academic institutions, there are multiple theories and mechanism on how to predict market returns. Also, the study of the relationship between Twitter and stock returns is not a new topic and has been investigated frequently in the past. The focus so far, however, has been mainly on specific events, such as discussions on Twitter around earnings announcements (Bartov, Faurel & Mohnram, 2018) and political elections and their influence on stock prices (Nisar and Yeung, 2018), or how economic or financial themed tweets of a single politician influence companies' performance (Juma'h & Alnsour, 2018). However, we recognised that there had been little research to date on whether the growing popularity of Twitter use by top politicians has an impact on the stock performance of companies named in the tweets. Investors or organisations may experience inflation or deflation of their stock as a consequence of a tweet from a politician referring to their company. We believe that the sentiment of the tweet is the determinant factor whether the yielding return will be positive or negative. Therefore, we are convinced that this phenomenon is an investigative opportunity; hence a tweet from a politician may help financial institutions to predict the direction of market movement.

3

Tweets may be purposely filtrated to the audience in order to manifest the desired negative or positive returns, which is also known as microtargeting, a sub-component of predictive analytics. Predictive analytics is a category of data analytics which aims at making predictions about future outcomes based on historical data and analytics techniques (Edwards, 2019). Microtargeting can put stock return unfairly in jeopardy, which is considered a threat to the financial market; hence this may imply it is not free of direct manipulation. Can a politician tweet be a medium of microtargeting to investors and therefore, yield a manipulated return? Inversely, can investors predict market return based on a political tweet? The possibility of microtargeting by politician is believed to be a problem and requires attention; hence politicians use Twitter, have a large audience and the practice of microtargeting by politician towards the financial industry cannot be ruled out.

1.2 Course of Investigation

Chapter 2 is devoted to theoretical framework pertaining to Twitter use in politics, stock market and politics, stock price prediction based on the efficient market hypothesis (EMH), Twitter and stock price forecasting and Twitter sentiment analysis. Subsequent, the hypotheses will be presented on the basis of both the existing literature and own assumptions. Chapter 3 will present the data used and the required data adjustment for the investigation, while Chapter 4 will present the utilised methodology for hypothesis testing. The abnormal return test will be conducted in Chapter 5, which will be accompanied by a thorough analysis and robustness check, which consequently will culminate with a link to the literature in chapter 6. Finally, in chapter 7, a summary and conclusion of our research will be presented as well as the limitations and prospects for future studies in this area.

1.3 Aim of the Research

The aim of this investigation is to analyse whether there is a relationship between political tweets and abnormal stock return. Consequently, the investigation will be looking into the relationship between political tweet sentiment and market movement. Moreover, we also aim to estimate the lifespan of the effect of the political tweets on market return. The investigation will be conducted through event studies and based on tweets from three presidents from three different countries Brazil, United States of America and France.

4

The observation period of the collected tweets is from the last three years (2017-2020). The estimation metrics for the life span will be expressed in days. Ultimately, the aim of the investigation is encapsulated in the main research question “can stock return be predicted by political tweets’ sentiment and, what is the lifespan of the impact”.

5

2. Literature Review

______________________________________________________________________ In this chapter the theoretical background of the thesis is presented. It covers the relationship between politics and Twitter, the EMH and the relationship between Twitter and stock price forecasting. Based on existing literature and own assumptions, the hypotheses of the thesis are formulated.

______________________________________________________________________

2.1 Twitter Use in Politics

In recent years, Twitter has been increasingly used by public figures, such as politicians, to communicate with the public (Aharony, 2012). For this reason, several researchers have been investigating how the channel is used by politicians.

Conway, Kenski and Wang (2013), for example, found out that a few years ago, the platform was used very rarely by the politicians and that they posted irregularly during election campaigns. Increased Twitter activity could only be detected before important events related to the election. They also found that the election results do not necessarily depend on how often a politician has been active on Twitter, but rather on how popular the person is outside the platform and therefore has a higher number of followers. The intensity of Twitter use also plays an important role in how much social media is used in general, according to Graham, Jackson and Broersma (2014). In their study, they examined, that in election campaigns in 2010, Dutch politicians used the platform much more than the British politicians. This can be traced back to the fact that in the Netherlands, the use of social media became more popular earlier. They also came to the same conclusion as Conway et al. (2013), that in particular before or directly related to essential election dates, Twitter activity increased significantly (Graham et al., 2014).

Other researchers, such as Aharony (2012) or Jackson and Lilleker (2011), show that the politicians also use the platform to inform the public about their activities as politicians, for example about international meetings, interviews or speeches they held. Jackson and Lilleker (2011) also found out, that members of UK parliament use Twitter most likely to describe their everyday life, to share thoughts about their work or to present themselves and their projects in a good light.

6

In interviews with French politicians in Frame and Brachotte's (2015) article as well as in the article of Aharony (2012), the authors show that politicians on the other hand often use Twitter to express their opinion about current events, as this is easier and faster than for example via press release. According to Aharony (2012), politicians understand, that Twitter nowadays is a useful tool to communicate with the public, and they mainly share information about current issues such as general information or economics. Nevertheless, some politicians such as Safia Otokoré, former Finance Minister's head of communications, are cautious about what she publishes on Twitter, especially when it comes to the economy, as she fears that this could have a negative impact on the financial market. Since the press is following the accounts of politicians, the new information usually spreads very quickly (Frame & Brachotte, 2015).

Conway et al. (2013) further mention that since Twitter is an unfiltered platform and politicians can express their own opinions, they could use it more to their advantage by posting about events that are not yet known, in order to attract the attention of news agencies, and thus the broad masses. Recently, some politicians have taken advantage of this development, such as Donald J. Trump, 45. President of the USA. He is an influential, powerful man who has access to much information and as President has executive power, which is why Juma'h and Alnsour (2018) are convinced that he can influence the financial market through his tweets.

2.2 Stock Market and Politics

The stock market consists of various elements and a framework dictated by the government on how to operate responsibly and without jeopardising the economy. In this regard, the financial market is considered to have a certain totalitarian characteristic, which is why the financial markets must strictly operate by the rules stipulated by the government.

Day and Harvey (2013) wanted to investigate the question of whether there is a connection between politics and the stock market in the USA in particular. They believe that this is a phenomenon that has existed there for a long time. Economic growth is directly linked to what decisions are made in politics and the people who represent politics. So, does the political structure have an impact on the performance of market indices such as the S&P 500? The authors are convinced that politics affect directly fiscal

7

and monetary policy and thus, the financial market. Consequently, it can be assumed that politicians influence the stock market, which raises the question of whether certain parties or ideologies, such as Republicans or Democrats, are favoured or whether this is irrelevant. As a result, although it was not possible to establish clearly whether the Republican Party or the Democratic Party is more advantageous, it was possible to demonstrate overall that politics influences the stock market. While a Democratic president was responsible for a better development, a Republican Congress had more influence on the stock market. Thus, a combination of these two constellations has in the past led to the greatest impact on the stock markets. However, the political stalemate has an even more substantial influence, as it reduces, for example, changes in fiscal policy and thus a conscience of permanence can be assumed, which is also what markets prefer (Day & Harvey, 2013). In addition, while the Fed is also seen as a political force, it has been observed that a combination of increasing money supply and low-interest rates also leads to a positive development of the stock market (Day & Harvey, 2013 and Paulson, as cited in Day and Harvey, 2013). The conclusions of the research are partly consistent with the investigation carried out by Simay Yardim (n.d.), where the effects of political factors on financial development were examined. It has been determined that presidential countries show a lower level of financial development with respect to parliamentary countries. The financial development is divided into two subsets, which are banking sector and stock market development, but the focus lies on the latter. The research has furthermore proven that plurality systems show higher stock market capitalisation levels than proportional representation system (Yardim, n.d.). The opposite poles of democracy, namely left or right-oriented, were also included in the investigation and it was determined that left-oriented government show higher stock market capitalisation compared to the right-oriented government (Yardim, n.d.). Altogether, it can be concluded from the research that the stock market and its accompanied government is a constellation on itself, where the market lies of the rules and operation based on the the associated government.

Döpke and Pierdzioch (2006), on the other hand, could not find evidence, that in Germany, the stock market is influenced by political structure so that it reacts differently during right-wing or left-wing governments. This contrast to the USA could consist in the fact that political power changes very frequently there, whereas this is less common in

8

Germany. Nevertheless, it also could be determined that political actions that lead to positive developments in the stock markets give the government a better approval rate and vice versa. Overall, we can therefore assume that there is a fundamental relationship between the stock market and politics. However, the strength may vary depending on the country and its political system.

2.3 Stock Price Prediction

Stock price prediction is a popular topic and of great importance, especially for the financial world and science, as it would be of great benefit to investors if the prices were foreseeable. In general, stock prices are subject to fluctuations caused by various factors such as supply and demand, corporate profits, trends or market sentiment. Until today there are different opinions on whether or not it is possible to predict the development of stock prices.

2.3.1 Efficient Market Hypothesis

The first attempts to explain market movements date back to the beginning of the last century, when the predominant opinion was that stock price movements are not predictable and follow a random walk instead (Yen & Lee, 2008). In 1970, Eugene Fama argued that financial markets are efficient, in which securities prices fully reflect all available, known information at all times (Fama, 1970). The theory is also known as the efficient market hypothesis (EMH). Stock prices can therefore only be changed by new information, as current and past information is already included in the prices. Since new information occurs randomly and thus not foreseeable, according to the EMH, it is not possible to predict stock prices (Fama, 1970 and Bollen, Mao & Zeng, 2011). In his article, Fama (1970) defines three categories of market efficiency, namely a weak, semi-strong and semi-strong form. While the weak form of market efficiency assumes that the current price already contains all historical market-relevant information, the semi-strong form additionally includes all publicly available market-relevant information. Furthermore, in the strong efficiency form, all non-public market-relevant information is included in the current stock price. However, these hypotheses require that there are no limits on arbitrage, that uncorrelated errors are anticipated and investors behave rationally (Loredana, 2019). It is precisely the behaviour of investors in response to new information that can be seen as a weakness of the EMH, as they do not necessarily act rationally. The

9

theory entirely ignores that investor may over- or underreact to new information. Underreaction occurs when stock prices do not move up as much in consequence of good news or do not fall as much in the case of bad news. Overreaction, on the other hand, means that stock prices move much more in the respective direction than expected.

2.3.2 Information and the Stock Market

Since it is assumed that new information leads to stock price movements, this correlation has been investigated in numerous studies. According to Vanstone, Grepp and Harris (2018), market prices reflect investors' views on how the value of a company is expected to be in the future. Thus, if these expectations are changed by new information about a company, this could lead to a short-term impact on stock prices, since the investors may not act rationally. The authors attribute the phenomenon to the fact that there are different types of investors. On the one hand, there are individuals who behave sober and thoughtful in their decisions and are guided by fundamental Bayesian beliefs. Other investors, on the other hand, tend to trade according to their own random beliefs, also known as noise traders and are therefore more susceptible to news regarding investments. In this regard, the sentiment of the content and how investors process it is of particular importance. Ab. Rahman, Abdul-Rahman and Mutalib (2017) are convinced that changes in stock prices can be traced back to changes in investor sentiment based on financial news. According to them, there are two ways of analysing the direction in which stock prices move, namely fundamental analysis, which examines markets, business and competitors, and technical analysis, which focuses on the analysis of historical prices. In their research, they conduct a fundamental analysis of stock price prediction, focusing on machine learning algorithms on the basis of financial news about specific companies. As a result, they found that positive attuned articles resulted in the closing price of the respective company being higher than the opening price and vice versa in case of negative news.

Muhammad Tahir Suleman (2012) examined the relationship between political news and the KSE-100 index movements. He divided the collected political news further in good and bad news and used asymmetric GARCH model to measure the relationship between the two variables. The political news included topics such as statements and interviews of key political figures about the future politics or information about agreements between political parties. The author was able to show that political news with positive sentiment

10

led to a positive development in stock returns and caused less volatility overall. Bad news, on the other hand, had a negative impact on stock prices and caused increased market fluctuations. Al-Maadid, Caporale, F. Spagnolo and N. Spagnolo (2020) in contrast, could not find a significant relationship between political news and stock market returns. In their study, they focused on political and business news from Bloomberg and investigated whether these have an impact on stock returns in the Gulf Cooperation Council (GCC) countries. For negative business news, however, a significant correlation with stock returns could be identified.

According to these findings, some researchers such as Bollen et al. (2011) are therefore convinced that not only classic news articles provide significant results, but that news published via social media like Twitter can also be used to predict stock market movements.

2.4 Twitter and Stock Price Forecasting

The increasing popularity of the microblogging platform Twitter (e.g. Sul, Dennis & Yuan, 2017) leads to the question whether the information contained in tweets can be used to explain stock price movements and thus to predict stock returns. Several studies on this topic have already been published but, however, they mainly use individual events that directly affect the respective companies, such as earnings announcements. The results of the studies are mixed, with some researchers failing to find significant relationships between tweets and stock returns (e.g. Nisar & Yeung, 2018 or Juma'h & Alnsour, 2018), while other authors, such as Bartov et al. (2018) or Sul et al. (2017), are able to prove such correlations.

Nisar and Yeung (2018) examined the relationship between UK 2016 local elections and FTSE 100 stock price movements. They have filtered out tweets from Twitter that are related to this election and have been posted by users based in England, and then conducted a sentiment analysis for each of these tweets. In order to investigate whether a relationship exists between Twitter discussion and stock price movement on the days around the election, various correlation and regression analyses were carried out. As a result, no significant correlation could be found. Only trends that may indicate a relationship could be observed in the form of a correlation between sentiment and FTSE 100 closing prices. As possible reasons for the insignificant results, they mention that

11

there might have been too small sample tweets due to short period window and too many neutral tweets.

Juma'h and Alnsour (2018), in contrast, have investigated whether Donald Trump's tweets about individual companies have an impact on their financial performance. They also investigated whether presidential tweets on topics such as economy or finance have a substantial impact on major American indices. The method used was an event study and some correlation test. The authors see a possible reason for non-significant relationships between Trump tweets and stock price movements in the fact that the number of tweets collected was simply not large enough or that tweets might not influence the stock prices of the companies spontaneously since they only used an event window two days before and two days after the event. Another reason could have been that the information Trump publishes on Twitter was already known in advance and had therefore already been included in the companies' stock prices.

Other studies, on the other hand, could prove that there is a significant relationship between tweets and stock price movement. Bartov et al. (2018) have conducted a study on whether company-specific information posted on twitter helps investors better to predict stock returns before earnings announcements. The authors performed a textual analysis of the tweets to classify their mood and then performed correlation tests on this basis. As a result, they found that aggregated Twitter opinions around earnings announcements help to predict the actual quarterly earnings of firms, especially when the companies are placed in weak information environments. They also observed that the aggregated sample tweets led to abnormal returns around the EA. Ranco, Aleksovski, Caldarelli, Grcar and Mozetic (2015) could prove, too, that, when both the volume and the sentiment of tweets are taken in consideration, the tweets about companies' earnings announcements and other firm related information actually do have a statistically significant impact on stock returns. These results were obtained for the study of tweets about 30 listed companies from the Dow Jones Industrial Average index and the relationship to their share performance. In addition to the Pearson correlation and Granger causality test, the authors also conducted an event study to find out if the mood of tweets led to abnormal returns a few days after the tweets.

Sul et al. (2017) have investigated how the sentiment of tweets about companies listed in the S&P 500 affects the stock performance of these firms. Based on regression analysis,

12

the authors come to a conclusion, that positive or negative sentiment of a tweet in which the selected companies occur have an impact on the stock price and thus on the return. However, it is interesting to note that significant results are available, especially for tweets from users with comparatively few followers and less for users with many followers. In addition, tweets that were not retweeted had more significant correlations with the stock market than tweets that were often further shared.

Since Twitter, as already noted in the previous section, is a platform where anyone can express their opinion unfiltered, whether true or not, Huang, Zaeem and Barber (2019) have investigated in their research, whether or not this phenomenon is a problem in predicting the share price. As a result, they were able to demonstrate that investment strategies based on trusted tweets perform better than other strategies and conclude that stock price prediction can therefore be better performed with trusted users on twitter. As we focus on high-ranking politicians in our study, we assume that they are considered trustworthy, due to the fact that their accounts are often verified. This is a tool on Twitter where accounts of public interest are checked for authenticity and then receive a special icon (Twitter, 2020).

2.5 Twitter Sentiment Analysis

When it comes to research on the relationship between social media and stock price prediction, it is often used sentiment analysis approach (e.g. Ranco et al., 2015, Nisar & Yeung, 2018 or Bollen et al. 2019). This can be traced back to the development of publishing news via social media. Investors are becoming tremendously overloaded with information, making it difficult to predict how the market will develop. Therefore, automated analysis methods, such as sentiment analysis, are becoming more and more common to support stock price prediction.

Sentiment analysis is a subfield of data mining and used to identify and classify the opinions of texts, posts or sentences (Ghiassi & Lee, 2018). The texts can be classified manually according to content or automatically using machine learning or natural language processing, for example. However, automatic classification is much more common for scientific studies, as these may involve large amounts of data, and the manual approach would not be profitable in terms of time. The Twitter platform has an overall advantage over long newspaper articles, for example, because the length of characters per

13

message is limited to 140 characters. This means that users have to limit themselves considerably when posting tweets and publish their opinions very concisely and briefly, which could simplify the classification of the mood (Sul et al., 2017).

In general, two different methods exist for the automated text classification process: unsupervised and supervised machine learning (Oliveira, Cortez and Areal, 2017). Supervised machine learning, which includes Support Vector Machine (SVM) or Naive Bayes, is a machine learning algorithm which is trained to correctly classify texts in several runs using examples and test data. However, since platforms such as Twitter make it difficult and expensive to obtain the necessary data, this method is often not used in studies. According to Oliveira et al. (2017), it is more common to classify texts using the unsupervised approach, in which, for example, specific keywords or lexicons are used as a basis for analysing mood. Although this method is more comfortable to carry out, it is also much more time-consuming and possibly less accurate.

In general, there are many different possibilities for the classification of text sentiment. Especially in the area of news and stock market relations, two ways have become established: the emotional model and the dimensional model (Sul et al., 2017). The emotional model is used to capture a wide range of emotional moods, such as in the study of Bollen et al. (2019), in which they have investigated a total of six mood dimensions, namely Happy, Kind, Calm, Alert, Sure and Vital. In dimensional analysis, on the other hand, usually only two or three sentiments are measured, namely positive and negative as well as occasionally neutral, such as in the studies of Nisar and Yeung (2018) or Ranko et al. (2015). According to Sul et al. (2017), the emotional method is primarily used in studies of natural language processing, because here, a much more detailed insight is needed. The dimensional method of sentiment analysis, on the other hand, is mainly used in psychological studies, such as research on the connection between Twitter messages and stock price movement.

2.6 Hypothesis Development

In the modern dynamics of the stock market, where social media plays a considerable role, we believe it is essential to have a rather molecular view of specific elements that may conduce to a particular behaviour on the stock market. In our view, this entails an

14

understanding of how one particle within a spectrum can influence the other, or more precisely, how one politician’s opinion may have an impact on the entire stock market. The theory behind it can be compared with the theory of relativity, where the energy (E) is determined by the mass (M), times quantity (C) squared. One politician within the constellation may have a massive reach due to his or her tweets and followers, which subsequently may lead to more substantial influence on the stock market. The politician is still part of the spectrum within the constellation, and yet we believe that this individual alone can exercise influence on the stock market. Applied to the relativity theory, this would mean that the energy or influence produced by one politician, the E, can have more significant influence, the M, than the entire spectrum within the constellation, where C is the numbers of followers on Twitter. Based on these thoughts, we believe that it is essential to investigate whether there is a relationship between tweets from politicians and stock returns of companies.

A look at existing literature has shown that several studies have already been carried out on the subject of politics and the stock market with a wide variety of results. However, these studies have mainly focused on specific events, such as a single election (Nisar & Yeung, 2018) or a single politician (Juma'h & Alnsour, 2018). Since we have also seen that Twitter has become more and more important for politicians over the years and that they share their personal opinions or thoughts on a wide range of topics (e.g. Jackson & Lilleker, 2011), we will go one step further in our research and include this development in our analysis.

We will investigate whether any opinions about companies published on Twitter by leading politicians from several countries over the last three years can influence the behaviour of investors and lead to changes in the stock prices of these companies. The companies will be selected from the most extensive market index in the country and the top 30 Forbes worldwide, so that we can use a large number of events for more reliable results. We agree with the views of, for example, Ab. Rahman et al. (2017) or Suleman (2012) and believe that political tweets about the companies will cause a development in the respective direction depending on the sentiment, either positive or negative. Neutral tweets cause, in our opinion, no changes for the stock prices. Based on this, we formulate the following hypotheses:

15

H1: Tweets from politicians with positive sentiment have a positive impact on a company's stock returns.

H2: Tweets from politicians with negative sentiment have a negative impact on a company’s stock returns.

Another aspect that we want to investigate in our paper is the question of whether the influence of tweets is rather temporary or lasts for several days. One reason for our considerations is the results of the study by Ranco et al. (2015), in which significant abnormal returns were observed over several days. However, this study dealt with specific events such as earnings announcements, which directly influence the returns of companies. Since politicians' tweets can contain all sorts of topics, some sources have claimed that the tweets of politicians are more likely to cause temporary fluctuations in stock prices, mostly visible only on the same day (e.g. Valetkevitch and Mikolajczak, 2017). We support this theory and think that the effects are rather temporary, which is why further the following hypothesis is formulated:

H3: The impact of politicians' tweets on a company's stock price is more likely to be temporary than visible over several days.

16

3. Data

______________________________________________________________________ This section contains the description of the data used in this research. It includes the collection of political tweets, sentiment analysis and the compilation of market data. ______________________________________________________________________

3.1 Political Tweets

In order to obtain a broad spectrum of data, we have selected leading politicians from three different countries for our study, who are very active on the platform and have a large number of followers. More precisely, we will be looking at tweets of Donald Trump, current President of the United States of America, Jair Bolsonaro, current President of Brazil and Emmanuel Macron, current President of France.

The focus lies on tweets in which the politicians mention companies listed in the respective indices of the three countries, namely S&P 500, Bovespa and CAC 40, and additionally the 30 top public companies worldwide (Forbes, 2019). The merit of using political events as the proxy is that Twitter has become a medium of preference for politicians to make their statements. However, since the platform has only been used more intensively by politicians in recent years, as shown in the research of Juma'h and Alnsour (2018), we focus on tweets in the period from March 2017 to March 2020.

Since politicians are free to express their opinions on Twitter, a tweet may not reflect the truth in some cases. However, the goal of the investigation is not to examine tweets for evidence of actual actions. The focus of our research is on what happens in the virtual world, the tweets, and not the real world. Therefore, a "fake political tweet" could also be considered, as we believe that such tweets can also have an impact on stock returns.

The tweets of the three politicians were collected with the service of the platform tweetbinder.com. Tweet Binder is a fee-based portal, where reports on certain Twitter users, hashtags or keywords can be created (Tweet Binder, 2020). We requested the company to provide us reports with all original tweets for the users @realDonaldTrump, @jairbolsonaro and @EmmanuelMacron in the period from 09 March 2017 to 09 March 2020.

17

The reports of Tweet Binder contain the listing of all posted tweets in the selected time period as well as additional statistics on the number of total tweets and followers per contributor, the average amount of follower per contributor, the potential impacts, which implies the potential numbers of times an individual could have seen the hashtag, and the number of replies and retweets. Furthermore, the reports cover statistics about the economic value, which refers to the monetary value of the hashtag in the market, and the average tweet value. However, these statistics will not be part of the event study, but will be used as additional data for the purpose of analysis and will be referred to as external data. In Table 3.1, the most important data from the reports are presented.

Table 3.1 Summary of data provided by Twitter Binder.

Donald Trump Jair Bolsonaro Emmanuel Macron

Number of Tweets 8,344 4,328 2,990 2017 1 0 0 2018 3,001 1,615 2,188 2019 4,592 2,303 698 2020 750 410 104 Number of Followers 73,532,079 3,305,765 4,525,861 Potential impacts 545,699,890,251 26,368,460,319 13,532,149,549

Economic value (in $) 833,819,848 44,556,854 26,580,149

Average tweets value (in $) 99,930 6,240 8,890

Source: Tweet Binder. Data from 09 March 2020.

It can be seen that Donald Trump has by far the highest number of tweets and followers, while Emmanuel Macron and Jair Bolsonaro are quite close to each other, with an evident distance to Trump. A similar observation can be made for the additional statistics.

Since the reports contain all tweets of the presidents in the period under investigation, in a first step we manually examined the tweets to see if the presidents mention companies in their tweets that are listed in the S&P 500, Bovesta or CAC 40, as well as the top 30

18

global public companies. With this approach, we obtain a data set of national and international companies mentioned by the presidents. Ultimately, our sample includes companies such as Apple, Boeing, Comcast, Facebook and General Motors from the S&P 500 as well as Embraer and Petrobas from Bovespa. From the CAC 40, for example, Airbus, Renault or Safran are included, and Samsung (Kospi), Toyota (Nikkei), Volkswagen and Allianz (DAX) are listed on the Forbes top 30.

In a further step, we have removed such tweets, that mention the specific company within -1 and +5 days several times, according to the event windows, which will be explained more in detail in the following chapter. The tweets mentioning the same company within this time window would possibly affect the result of the abnormal returns and lead to biased results.

Since in some cases, the politicians mentioned a company in several tweets on the same day, we also had to make sure that we included each event date only once in our data set. In such cases, the NET sentiment of the tweets is used, which means that on average we have chosen the most represented sentiment for this event date, such as in the case of a negative and neutral tweet, the sentiment finally chosen is negative.

After all adjustments, a total of 91 event dates for 34 companies remain for our investigation. While for some companies there are several events in the data set, for other companies there is only one event. Appendix 2 contains a detailed list of the chosen tweets.

3.2 Sentiment Analysis

In order to test our hypotheses, the selected tweets need to be further classified according to their sentiment, depending on whether the content of the tweets has a positive, negative or neutral touch. For this purpose, we have used the platform MonkeyLearn (https://www.monkeylearn.com), which is a machine learning platform that offers a pre-trained sentiment analysis model (Roldós, 2019). The service uses a well-developed system of supervised and unsupervised systems, more precisely a combination of machine learning algorithm and lexicon-based classification.

19

We have performed the classification once with the original tweets and a second time with customised tweets without links because the links only consist of number and letter combinations. When comparing the two outcomes, we noticed that the classification of the tweets without links was more accurate because the system could not misinterpret the content. After that, only a few manual changes to the classification were necessary, because in some tweets, for example, politicians mention different topics that affect different companies or people, which may have led to a wrong classification of the sentiment for the specific company we are investigating. However, in most cases, the mood of the tweets was correctly classified. Table 3.2 summarises how many tweets were rated positive, negative or neutral, both at a total level and for individual politicians. It can be seen that we mostly have tweets with either positive or negative sentiment and only a few are classified as neutral.

Table 3.2 Summary of results sentiment analysis.

Positive Tweets Negative Tweets Neutral Tweets Donald Trump 29 30 5 Jair Bolsonaro 7 5 4 Emmanuel Macron 8 0 3 Total 44 35 12 3.3 Market Data

For market data, the daily stock prices for all companies used in the event study, as well as the corresponding indices in which they are listed, were obtained from Thomson Reuters DataStream and Yahoo! Finance for the period from 16 March 2016 to 16 March 2020.

Since event study is not conducted with daily prices, but on the basis of daily returns, the share prices for the companies and indices were transformed into natural logarithm returns (Elad & Bongbee, 2017) using the following formula:

𝑅_𝑡 = 𝐿𝑁 ( 𝑃𝑡

20

where 𝑅_𝑡 refers to the return at time t, 𝑃_𝑡 is the current share price and 𝑃_𝑡−1is the share price of the previous trading day.

In Table 3.3, the market data is summarised for all 34 companies which are mentioned in the sample tweets from the politicians, including the mean, median and standard deviation of daily log returns. The number of observations varies depending on the data available and the corresponding trading days of the different countries.

Table 3.3 Descriptive statistics of companies’ stock returns.

Company Obs. Mean Median Std.Dev. Max. Min. Range

Airbus 1023 0.0001 0.0004 0.0172 0.0978 -0.1831 -0.0853 Allianz 1010 -0.0001 0.0007 0.0138 0.0391 -0.1664 -0.1273 Amazon 1006 0.0011 0.0014 0.0174 0.1241 -0.0825 0.0416 Apple 1006 0.0008 0.0009 0.0174 0.1132 -0.1377 -0.0246 Aptiv 1006 -0.0002 0.0008 0.0216 0.1125 -0.2626 -0.1501 Boeing 1006 0.0000 0.0009 0.0213 0.0946 -0.2724 -0.1779 Caterpillar 1006 0.0002 0.0004 0.0180 0.0770 -0.1541 -0.0772 CBS 1006 -0.0014 0.0004 0.0196 0.0870 -0.2044 -0.1174 Comcast 1006 0.0002 0.0003 0.0147 0.1184 -0.0874 0.0310 Embraer 989 -0.0008 -0.0005 0.0242 0.2029 -0.3071 -0.1042 Exxonmobil 1006 -0.0009 -0.0001 0.0136 0.0467 -0.1304 -0.0837 Facebook 1006 0.0003 0.0010 0.0191 0.1027 -0.2102 -0.1075 Ford Motor 1006 -0.0010 0.0000 0.0172 0.1021 -0.1167 -0.0146 General Motors 1006 -0.0004 0.0008 0.0187 0.1211 -0.1627 -0.0416 Gilead 1006 -0.0003 0.0002 0.0162 0.0835 -0.0950 -0.0115 Harley-Davidson 1006 -0.0009 0.0002 0.0211 0.1803 -0.1401 0.0402 Home Depot 1006 0.0002 0.0007 0.0147 0.0761 -0.2206 -0.1444 Lockheed Martin 1006 0.0003 0.0010 0.0136 0.0711 -0.1365 -0.0655 Michelin 1023 -0.0002 0.0002 0.0158 0.1226 -0.1567 -0.0341 Microsoft 1006 0.0009 0.0010 0.0162 0.1329 -0.1595 -0.0265 Netflix 1006 0.0011 0.0004 0.0245 0.1742 -0.1407 0.0335 Nike 1006 0.0001 0.0003 0.0160 0.1055 -0.1241 -0.0185 Petrobras 1006 0.0000 0.0016 0.0335 0.1884 -0.3709 -0.1825 Pfizer 1006 0.0000 0.0005 0.0125 0.0858 -0.0805 0.0053 Renault 1023 -0.0017 -0.0010 0.0208 0.1142 -0.2477 -0.1335 Safran 1023 0.0002 0.0002 0.0156 0.0618 -0.1845 -0.1228

21 Samsung 979 0.0006 0.0011 0.0162 0.0464 -0.0838 -0.0374 Sempra Energy 1006 -0.0001 0.0007 0.0144 0.1444 -0.1877 -0.0433 Southwest Airlines 1006 -0.0002 0.0005 0.0186 0.1349 -0.1638 -0.0289 Starbucks 1006 0.0000 0.0004 0.0151 0.1186 -0.1768 -0.0582 Target 1006 0.0001 0.0007 0.0191 0.1859 -0.1310 0.0549 Toyota 976 0.0001 0.0001 0.0141 0.0682 -0.0906 -0.0225 Volkswagen 1010 -0.0002 -0.0004 0.0185 0.0640 -0.1650 -0.1010 Walmart 1006 0.0004 0.0008 0.0131 0.1034 -0.1074 -0.0040

Source: Thomson Reuters DataStream as of 17 March 2020.

In Table 3.4, the descriptive data for the respective indices are presented, including mean, median and standard deviation for the daily log returns. Besides the three national indices S&P 500, Bovespa and C.A.C. 40 we further use the international indices DAX, KOSPI and Nikkei for such companies that are mentioned in the politicians’ tweets and are not listed on the three main indices. This enables us to determine the respective expected returns in the event study more precisely using the corresponding comparative indices.

Table 3.4 Descriptive statistics of the indices.

Index Obs. Mean Median Std.Dev. Max. Min. Range

Bovespa 989 0.0004 0.0011 0.0166 0.1302 -0.1599 -0.0297 CAC 40 1023 -0.0001 0.0004 0.0106 0.0406 -0.1310 -0.0904 DAX 1010 -0.0001 0.0007 0.0109 0.0338 -0.1305 -0.0968 KOSPI 982 -0.0002 0.0005 0.0084 0.0347 -0.0454 -0.0107 Nikkei 998 0.0000 0.0002 0.0112 0.0651 -0.0825 -0.0174 S&P 500 1006 0.0002 0.0006 0.0107 0.0888 -0.1277 -0.0388

22

4. Methodology

______________________________________________________________________ This chapter contains the definition and theoretical foundations of the event study approach. It also describes the specifications chosen for this study.

______________________________________________________________________

4.1 Event Study

To measure the impact of a political tweet on a company’s stock return, an event study is conducted. By definition, an event study is used to measure the impact of a special event on the firm value (McKinlay, 1997). The method is applied to examine whether certain events lead to abnormal returns for companies associated with the events and typically, the approach is conducted to investigate for example the impact of earnings announcements, mergers and acquisitions or macroeconomic variables on company performance (McKinlay, 1997).

While the first approach of event study was published in 1933 by James Dolley, it has been modified extensively since then, so that today’s studies broadly follow the methods of Ray Ball and Philip Brown or Fama et al. from 1968/1969 (Ball & Brown, 1968 and Fama, 1969 as cited in McKinlay, 1997).

However, since social media is playing an increasingly important role nowadays and we believe that information posted on Twitter also has an effect on company performance, we will use the event study approach to measure the impact of political tweets on stock returns. In theory, we will follow the approach of McKinlay (1997) from his article Event “Studies in Economics and Finance”, while in practice, we will use the software Stata to conduct the study. The codes used in this study are adapted from the example of Princeton University (2008). A detailed code list is provided in Appendix 3.

4.1.1 Definition of Event and Event Window

Event Study is generally structured in such a way that the events need to be defined first. Furthermore, it is necessary to specify the period over which the stock returns of the respective company associated with the event are to be examined, also known as the event window (MacKinlay, 1997).

23

In our study, we define as an event the date, on which a tweet was posted by the three selected politicians about a company listed on the chosen indices S&P 500, Bovespa and CAC40 as well as out of the top 30 companies worldwide, denoted as t0 = 0.

Since no general specifications are available for the determination of the event window, we have based our approach on existing literature (e.g. Ranco et al., 2015 or MacKinlay, 1997). For the length of the event window 𝐿₂, we define two different windows to validate H3, that the impact of politicians’ tweets is rather short-term than long-term. The first window includes 2 days and is defined with t₁= -1 and t₂ = 0 while the second window includes 7 days with t₁= -1 and t₂ = 5. These periods are used to measure whether the actual event leads to abnormal returns for the companies compared to the expected returns. The day before the actual event is included, as we cannot ignore the possibility that politicians may share the information or opinions, they post on Twitter, offline beforehand, which may already affect the stock prices (MacKinlay, 1997).

We further note that our study is concentrated on single events. It could be argued that multiple events should have been the chosen method for conducting this research, as there were some tweets that would have resulted in overlapping time intervals. In that case, several corrections would have been necessary (MacKinlay, 1997). To avoid misleading results, we have completely removed such events from our data to present more accurate results.

4.1.2 Normal and Abnormal Returns

Abnormal returns caused by the event are calculated by subtracting normal returns from actual returns in the event window. The normal return is the return that was expected if the event had not occurred (MacKinlay, 1997). For each stock i and time period t we calculate the abnormal return with:

𝐴𝑅_𝑖𝑡 = 𝑅_𝑖𝑡− 𝐸[𝑅_𝑖𝑡] (2)

where 𝐴𝑅_𝑖𝑡 are the abnormal returns, 𝑅_𝑖𝑡 the actual returns and 𝐸[𝑅_𝑖𝑡] the expected normal returns.

24

To estimate the normal returns, we define a 120-day estimation window L₁ with a 30-day gap prior to the actual event, defined as T₀= -150 and T₁= -30. The gap is necessary to ensure that normal returns are not affected by the event itself and therefore, not measured correctly (MacKinley, 1997). Figure 4.1 provides an overview of the defined estimation and event window.

Figure 4.1 Timeline for event study.

According to MacKinlay (1997), the normal returns can be measured with various methods, but the most common methods are the constant mean return and the market model. The constant mean return model is the simplest model and implies that the mean return of a stock is constant over time. The market model, on the other hand, assumes a linear relationship between stock return and return of the market portfolio. In comparison to the constant mean model, the part of the return that is related to the variations in the market return is eliminated, which may reduce the variance of the abnormal return and lead to a higher possibility to detect the effects of an event.

Additional methods include the multifactor model, in which variations in the normal return are examined more in detail using industry classification, and the Capital Asset Pricing Model (CAPM). It is a theory for capital markets pricing in a state of equilibrium and implies that the expected return on holding a stock is related to the risk of that stock. However, since the CAPM has some restrictions, MacKinlay (1997) believes that there is not a substantial benefit in using this method instead of the market model.

Since it is common in the existing literature for our research area to use a market model (e.g. Ranco et al., 2015), we will follow this approach and also use the market model to measure normal returns by applying the following formula:

25

𝑅_𝑖𝑡 = 𝛼_𝑖+ 𝛽_𝑖𝑅_𝑚𝑡 + 𝜀_𝑖𝑡 (3)

where 𝑅_𝑖𝑡 is the return of the stock, 𝑅_𝑚𝑡 the return of the market portfolio and 𝜀_𝑖𝑡 zero mean disturbance term. For the market return, we always use the corresponding index in which the company is listed, e.g. S&P 500 or CAC 40.

Since the abnormal return is the difference between the actual and the normal return, it is simply the error term of equation (3):

𝐴𝑅_𝑖𝑡 = 𝑅_𝑖𝑡− 𝛼̂_𝑖− 𝛽̂ 𝑅_{𝑖 𝑚𝑡} (4)

where 𝛼̂_𝑖 and 𝛽̂ are obtained using O.L.S. estimation to derive a regression of 𝑅_{𝑖 𝑖𝑡} and 𝑅𝑚𝑡 under the estimation period.

After the abnormal returns have been calculated, they will be aggregated over the event window as a cumulative abnormal return for each stock i with:

𝐶𝐴𝑅_𝑖(𝑡₁, 𝑡₂) = ∑𝑡2 𝐴𝑅_𝑡

𝑡=𝑡1 (5)

To test if the aggregated abnormal returns are statistically significant at 5% or 10%, we use the following equation:

(6)

The cumulative abnormal returns are statistically significant at 5% level when we have an absolute value equal or greater than 1.96. For a 10% significant level, the absolute value is equal to or more than 1.645. The null hypothesis states that the cumulative abnormal returns in the event window are not significantly different from zero.

To test whether positive or negative tweets in total can have the expected influence on the stock return, the mean value of the cumulative abnormal returns is calculated. With N

𝜃₁= 𝐶𝐴𝑅(t1,𝑡2) 𝑣𝑎𝑟(𝐶𝐴𝑅 (t1,𝑡2))/√𝑁

26

events for stock i, the cumulative average abnormal returns for period t are estimated using the following formula:

CAAR(t₁, 𝑡₂) = 1

𝑁 ∑ 𝐶𝐴𝑅𝑖(𝑡1, 𝑡2) 𝑁

𝑖=1 (7)

The statistical significance of the CAARs at a 5% and 10% level will be tested with the following formula:

𝜃₁= 𝐶𝐴𝐴𝑅(t1,𝑡2) 𝑣𝑎𝑟(𝐶𝐴𝐴𝑅 (t1,𝑡2))/√𝑁

(8)

Statistical significance at 5% is given when the absolute value equals or is greater than 1.96 and at 10% level 1.645. The null hypothesis states that the cumulative average abnormal returns are not significantly different from zero.

27

5. Results and Robustness Test

______________________________________________________________________ This section contains the presentation of the results obtained by the event study and the examination whether the hypotheses that have been formulated can be supported. Furthermore, a robustness test is performed.

______________________________________________________________________

5.1 Hypothesis Testing

At the beginning of this research, there was a clear understanding of how tweets could benefit politicians in terms of their campaigns by reaching broad masses. Based on this, we were convinced that political tweets have an impact on stock market performance depending on the classified sentiment.

In Table 5.1 the results of event study are summarised, where the mean value of the cumulative abnormal returns per time window and event is shown as CAAR for the time intervals (-1, 0) and (-1, 5) as well as the standard deviation and t-test results.

Table 5.1 CAAR for all events.

2-day window 7-day window

CAAR -0.0022 -0.0083

SD CAAR 0.0311 0.0551

t-stat CAAR -0.6693 -1.4441

* significant at 5% significance level ** significant at 10% significance level

In terms of hypothesis confirmation, the t-statistics of the CAAR indicate that in the majority of the cases, there was no statistical evidence that tweets' sentiment has played a role in market movement.

In order to test H1 and H2, the CAARs and the respective t-value were further analysed for each sentiment. Table 5.2 shows the results for the positively classified tweets. At a

28

5% significance level, the CAARs are not significant in both, the 2-day and 7-day event window. Only at a 10% significance level, the 7-day CAAR is significant. Therefore, we cannot reject the null hypothesis for H1 based on 5% significance level, which means that there is no statistical evidence that political tweets with positive sentiment yields a positive return on companies' stocks on average.

Table 5.2 CAAR for positive sentiment.

2-day window 7-day window

CAAR 0.0016 -0.0132

SD CAAR 0.0153 0.0521

t-stat CAAR 0.7061 -1.6764 **

* significant at 5% significance level ** significant at 10% significance level

It can be further noticed, that only in the 2-day window positive CAARs occurred, while in the 7-day window the positive tweets led to a negative CAAR of -1.32%.

The results for tweets with negative sentiment are shown in Table 5.3. Again, no statistically significant CAARs for both event windows could be found, neither on a 5% nor on a 10% level. Although it is apparent that negative CAARs with -0.84% and -0.85% are present in the 2-day and 7-day window, we cannot reject the null hypothesis for H2. Thus, there is also no statistical evidence that tweets from politicians with negative sentiments lead to a negative slope of stock returns of the mentioned companies.

Table 5.3 CAAR for negative sentiment.

2-day window 7-day window

CAAR -0.0084 -0.0085

SD CAAR 0.0457 0.0643

t-stat CAAR -1.0913 -0.7861

* significant at 5% significance level ** significant at 10% significance level

29

For H3, it is not yet possible to make statistical founded statements for the entire dataset at this point due to the fact that there is lack of overall evidence of the influence of tweets’ sentiment on market movement. The lifespan of tweets’ sentiment effect is perceived to be a derivative of the tweets’ impact on the market movement. In the absence of statistical evidence pertaining to tweets' sentiment and market movement, the lifespan of the effect is non-existent, without a possibility of inverse relationship. In order to be able to analyse H3 and examine more closely how tweets’ sentiment affects the stock price, we will take a closer look at the individual events in the following section.

5.2 Results for Sub-Dataset

On the level of individual tweets, the cumulative abnormal returns for the majority of tweets are statistically insignificant, too. However, there are 11 political tweets that cause statistically significant CAR at a 5% level for the 7-day window and 24 for the 2-day window, which can be seen in Table 5.4. These are in particular tweets that transmits positive or negative sentiment about, for example, employability or national interests.

Table 5.4 Results event study for single events.

Event President Company Index Sentiment 2-day window 7-day window

CAR t-stat CAR t-stat

1 Trump Amazon S&P 500 Positive 0.0010 0.0530 -0.0253 -1.4043

2 Trump Apple S&P 500 Positive 0.0014 2.2340 * -0.0475 -1.8957 **

3 Trump Apple S&P 500 Positive 0.0027 1.6000 0.0794 1.3303

4 Trump Apple S&P 500 Positive 0.0090 1.0700 0.0472 2.8168 *

5 Trump Apple S&P 500 Positive -0.0240 -2.4980 * -0.0370 -0.8286

6 Trump Apple S&P 500 Negative 0.0206 3.2520 * 0.0511 3.1496 *

7 Trump Apple S&P 500 Negative -0.0071 -0.7880 0.0276 0.6183

8 Trump Apple S&P 500 Positive -0.0087 -0.8670 -0.0078 -0.4332

9 Trump Apple S&P 500 Positive 0.0168 1.2530 0.0437 1.9749 *

10 Trump Apple S&P 500 Neutral 0.0004 0.0380 -0.0114 -0.5991

11 Trump Aptiv S&P 500 Positive -0.0016 -3.7020 * -0.0195 -0.8975

12 Trump Boeing S&P 500 Positive -0.0136 -0.4450 -0.0733 -2.1418 *

13 Trump Boeing S&P 500 Negative 0.0018 0.0680 -0.0259 -0.7160

14 Trump Boeing S&P 500 Negative -0.0139 -2.1100 * 0.0075 0.2065

15 Trump Boeing S&P 500 Positive -0.0003 -0.0710 -0.1263 -1.7938 **

30

17 Trump CBS S&P 500 Negative 0.0389 1.3220 0.0215 0.3962

18 Trump CBS S&P 500 Negative 0.0026 0.2240 0.0024 0.0937

19 Trump CBS S&P 500 Negative 0.0233 0.7610 -0.0128 -0.2650

20 Trump CBS S&P 500 Negative -0.0003 -48.6480 * -0.0246 -0.9591

21 Trump Comcast S&P 500 Negative 0.0200 1603.5980 * 0.0387 1.5013

22 Trump Comcast S&P 500 Negative -0.0078 -0.8940 0.0188 0.9702

23 Trump Comcast S&P 500 Negative -0.0129 -1.0580 0.0231 1.0090

24 Trump Comcast S&P 500 Negative -0.0136 -3.9340 * -0.0377 -3.0416 *

25 Trump Comcast S&P 500 Negative -0.0125 -1.3830 -0.0142 -0.9094

26 Trump Comcast S&P 500 Negative 0.0125 153.1620 * 0.0085 0.2854

27 Trump Comcast S&P 500 Negative -0.0225 -3.3180 * 0.0163 0.4281

28 Trump ExxonMobil S&P 500 Negative 0.0145 8.7750 * 0.0077 0.3022

29 Trump Facebook S&P 500 Positive -0.0137 -0.8380 0.0018 0.0643

30 Trump Facebook S&P 500 Positive 0.0080 0.5020 -0.0115 -0.4229

31 Trump Facebook S&P 500 Negative -0.0170 -0.8220 -0.0535 -1.6771 **

32 Trump Facebook S&P 500 Negative 0.0097 0.2010 -0.0078 -0.1026

33 Trump Facebook S&P 500 Negative -0.0137 -0.8380 0.0018 0.0643

34 Trump Facebook S&P 500 Negative 0.0153 1.3360 0.1273 1.3178

35 Trump Facebook S&P 500 Negative -0.0311 -0.6810 0.0138 0.2445

36 Trump Facebook S&P 500 Negative -0.0024 -0.1060 -0.0300 -1.4359

37 Trump Facebook S&P 500 Positive 0.0137 2.1810 * -0.0479 -1.5414

38 Trump Ford Motor S&P 500 Negative 0.0011 0.4300 0.0126 0.5685

39 Trump General Motor S&P 500 Negative -0.0044 -1.0160 -0.0018 -0.1182

40 Trump General Motors S&P 500 Negative 0.0095 14.6770 * 0.0543 1.9892 *

41 Trump General Motors S&P 500 Neutral -0.0432 -1.2280 -0.0373 -0.7411

42 Trump General Motors S&P 500 Positive -0.0050 -0.1850 0.0056 0.2425

43 Trump Genral Motors S&P 500 Positive 0.0040 8.0250 * -0.0066 -0.3559

44 Trump Genral Motors S&P 500 Positive 0.0178 0.5290 -0.0006 -0.0152

45 Trump Gilead S&P 500 Positive 0.0024 0.3460 0.0023 0.1489

46 Trump Harley-Davidson S&P 500 Negative 0.0087 0.4830 0.0123 0.4417

47 Trump Home Depot S&P 500 Positive -0.0167 -3.5620 * 0.0060 0.2390

48 Trump Lockheed Martin S&P 500 Neutral 0.0106 4.5780 * 0.0164 0.8944

49 Trump Lockheed Martin S&P 500 Positive -0.0118 -1.8430 ** -0.0479 -4.0241 *

50 Trump Netflix S&P 500 Negative 0.0274 1.9130 ** -0.0626 -0.9891

51 Trump Nike S&P 500 Neutral 0.0085 0.9780 0.0310 1.5975

52 Trump Pfizer S&P 500 Positive -0.0058 -1.9670 * 0.0062 0.4334

53 Trump Sempra Energy S&P 500 Positive 0.0049 0.4390 0.0137 1.2269

54 Trump SW Airlines S&P 500 Positive -0.0061 -1.1470 -0.0243 -1.1465