The effect of president Trump's Twitter messages on stock market returns

(1)

The Effect of President Trump’s Twitter Messages on Stock Market

Returns

Abstract: Social media, and especially Twitter, has become an important source of financial, political

and sentimental information. President Donald J. Trump uses Twitter as his main communication channel for sharing his opinions and policy changes. These tweets can be considered as new financial information to the markets, and may influence companies and stake holders. Especially tweets that refer to single corporations may have an effect on the stock returns. In this thesis, tweets from the President’s official Twitter account, which target specific companies, are examined in relation to financial, economic or political subjects. Two forms of Event Study Methodology are used to research the effect of targeted tweets on the company’s abnormal returns. The results suggest a significant effect on some levels on the targeted company’s stock returns.

Keywords: Stock Price Returns, Tweet, Event Study, Investor Sentiment, Market Model JEL Classification: G12, G14, G18, G41

UNIVERSITY OF AMSTERDAM

ECONOMICS AND BUSINESS ECONOMICS

BSc Economics & Business

Bachelor Economics and Finance

Author:

A.L. Dammers

Student number:

10528415

Thesis supervisor: Dhr. Dr. Johannes Lemmen

Finish date:

June 2018

(2)

Statement of originality:

I hereby declare that all sources used for this study are properly noted in the references and that this thesis is an original study that is not based upon material that has not been referenced.

(3)

Chapter 1 Introduction

Since the United States of America’s 2016 Presidential Elections, the financial world has been introduced to a new form of spreading presidential policy information. Donald Trump was known to use Twitter messages, also known as tweets, for spreading his ideas and comments even before he was elected as 45th_{President of the United States. Since his election, one of his main methods of} communication continues to be sending 140-characters-long messages into the world (although, as of November 2017 the character limit has been raised to 280). Research has shown that changes in Government Policy, Presidential Elections, or Central Bank board, -policy and -monetary changes have an effect on the return and volatility of international stock markets (Thorbecke, 1997). This may also be the case with messages provided on an hour to hour basis on an online platform, such as Twitter. So the question is: Do Donald Trump’s tweets actually affect the return of stock markets and specific company stocks?

Donald Trump on March 29, 2018

‘I have stated my concerns with Amazon long before the Election. Unlike others, they pay little or no taxes to state & local governments, use our Postal System as their Delivery Boy (causing tremendous loss to the U.S.), and are putting many thousands of retailers out of business!’

There are substantial signs that Twitter messages have a profound influence on the financial market, especially the specific companies being mentioned in the tweets. A small example is a recent study, which proved to have a significant effect on stock price returns based on the volume and sentiment measured on Twitter (Ranco et al., 2015). What exactly is the effect, however, and is the market shock temporary and superficial or do Trump’s Twitter Economics have more permanent

consequences? For an investor with ownership of company or market shares, understanding the short as well as the long term effect of a tweet from Donald Trump is significant. A negative tweet about a specific company will arguably result in a negative stock return in the short run, although it might restore itself eventually. Whatever may be, knowledge of the short run effect of a tweet by President Trump could obviously provide considerable opportunities for stakeholders and companies.

This thesis will focus on how, as of 16 June 2015, Donald Trump’s tweets have affected stock returns of certain US listed corporations, which were specifically named in his tweets. On this date Trump announced his candidacy for the presidential elections, and extended his influence using Twitter as his main communication channel to his voters. The objective of this thesis is to answer the research question:

With Presidents Trumps significant use of Twitter to proclaim future political and economic policy, do specific tweets directed to corporations effect their stock returns?

(5)

While researching the effect of Trump’s tweets on stock returns, the significance of the abnormal returns are measured by conducting an Event Study Methodology (MacKinlay, 1997). The event study will make use of two estimating models to calculate the normal returns. The first estimated expected return is based on an estimation window prior to the tweet being put online and calculates the expected stock return, using the Constant Mean Model. The expected return is then deducted from the actual return in the event window to result in the abnormal stock return. To calculate the second estimated expected return, a Market Model will be used. To check for omitted variables, the stock returns of the index on which the targeted company is listed in the US, will be used in the Market Model to calculate the expected estimation returns.

The share price data of the targeted companies in this thesis will be used with a daily interval. To measure the affected return on the stock price, a time frame of two business days after the President’s tweet will serve as the first event window, and a time frame of two business days before and two business days after the event as the second event window. This second event window is chosen to research the possibility that it is not the President’s tweet, but new information around the

performance of the company that the market is reacting to. In fact, the President might himself just be reacting to this new information. Wesley S. Chan concluded in his research in 2003 that stocks that experienced negative returns concurrent with the incidence of a news story, continued to

underperform in comparison to their peers. Does this also apply to the possible negative returns caused by a negative tweet that Trump sent about a certain corporation? Are Donald Trump’s tweets a new form of spreading economical and financial news into this world? An additional third event window, with a time frame of only the day of the actual tweet, will measure the potential shock effect of the tweet on the stock returns.

The present thesis consists of five chapters. Chapter 2 will review the relevant literature on earlier research studies and their results. Chapter 3 will describe the methodology of the research as well as the use of the data. Chapter 4 will provide the results of the conducted studies and elaborate on them. Finally, chapter 5 will present a conclusion based upon the results of the research, in order to answer the research question. Lastly, an analysis will be provided, debating the limitations of this study and providing steps towards further research.

(6)

Chapter 2 Literature Review

This chapter will give a short introduction on some of de recent studies and findings on the sentiment driven effect on stock returns. Information, financial or non-financial, can have an effect on the investor sentiment, which consequently affects investments in financial markets. As Eugene Fama enlightened in his theory ‘Efficient Markets Hypothesis’ (1970) all available information is already incorporated in the prices of the stocks if the market is efficient or at least semi-efficient. In the case of a perfect capital market, there is no asymmetric information, all participants in the market are price takers and there is no transaction cost. Therefore, all the information is freely available and included in the spot price of a share. Fama explains that there are three forms of market efficiency; weak, semi-strong and semi-strong (Fama, 1970). In the weak market efficiency form, it is impossible to predict future prices by analysing the share prices from the past. Future price movements are not influenced by information in the price series and future share prices follow a random course. In the semi-strong form, share prices adjust very quickly to new information, so that the prices reflect all public information available at that moment. In the strong efficiency form, all information, public and private, are incorporated in the share price. Fama concluded that in all three market efficiency forms, excess returns cannot be earned without inside information or luck. If it is true that tweets influence the stock returns, an explanation may be that online messages like twitter feeds may provide crucial information which was not yet available, even before an official press release or news coverage about the matter. Another explanation may be that it is not the information, but the online investor sentiment that causes the abnormal stock returns.

Classic finance theories do not provide a role for investor sentiment in calculating stock prices or expected returns. As formulated in the widely used Capital Asset Pricing Model (CAPM) introduced by William Sharpe (1964), there is no measurement for investor sentiment. More recent studies, however, use additional empirical data to test the effect of investor sentiment on stock returns. Baker and Wurgler (2006) concluded in their paper that the cross‐section of future stock returns depends on proxies of sentiment of a period before the stock returns. A year later Baker and Wurgler (2007) approached the origin of investor sentiment as exogenous and measured its empirical effects. They concluded that waves of sentiment clearly have important and regular effects on individual firms and on the stock market as a whole. In both papers they pointed out that there is still much unknown about the effects of investor sentiment. In most standard empirical estimation models there is no account for investor sentiment and this needs to be researched more to obtain further conclusions.

Financial news is important for stakeholders when evaluating the performance of corporations. This kind of news can be separated in two groups: 1. Obligated information, which is official information

(7)

made public by the company, which contains financial statements and notes, and 2. Non-obligated information, which is publicised through various (on- and offline) media sources, such as Facebook or Twitter. A relevant study correlated measurements of collective mood states derived from large-scale Twitter feeds to the value of the Dow Jones Industrial Average (DJIA), and found an accuracy of 86.7% in predicting the daily up and down changes in the DJIA in relation to the mood predictors (Bollen et al., 2011). Another research paper concluded that, in a narrow range of popular companies, social media message sentiments are significantly more important for the prediction of future market performance than social media message volume (Zheludev et al., 2014). This conclusion finds similarities with the results presented in a more recent paper, which showed that the sentiment of tweets had an especially strong effect if the company had high investor attention (Guo et al., 2017).

An empirical investigation in 2011 of the ability of online ticker searches to forecast abnormal stock returns concluded that online search intensity can be expected to forecast abnormal returns and increase trading volumes in the current period (Joseph et al., 2011). By simply tracking the search intensity of investors they found that search intensity for ticker symbols served as a proxy for investor sentiment. If the number of searches can act as a proxy for abnormal returns, this may signify that the number of tweets or retweets can also act as a proxy for abnormal return and trading volumes.

In addition to this conclusion, investor sentiment proved to be measurable in the content of tweets referring to specific companies (Corea, 2016). This use of information technology gives a new dimension to the search for information about the valuation of a company. More and more investors use the internet to gather information for financial investment choices, and also react to and discuss their performances online. Curme et al. determined in 2013 that between 2004 and 2012 the increasing searches for information on financial and political issues tended to be followed by lower stock prices. The internet, with its various Social Media Platforms, has presented a relatively new source for stakeholders to find information on sentiment, opinions, and financial data. It also

influences financial and behavioural studies conducted on the psychology related to decision process of investors and the effect of these decisions on stock prices and stock returns. The study conducted in this thesis presents valuable information on the emotional choices made by investors and the effect of their sentiments on the market which has already been proven to be useful in predicting stock returns (Baker & Wurgler, 2006;2007).

As already mentioned above, classic finance theories lack the consideration of the effects of

sentiments in measuring financial performance, and therefore have difficulty explaining, for example, the financial crisis (Zouaoui et al., 2011). Recently, social media have become an important

communication channel for spreading these emotions, thereby strengthening their effect on investors and eventually on stock price returns. These effects can provide challenges to companies, and

(8)

Alexander & Gentry concluded in their paper in 2014 that, with guidance from The Security and Exchange Commission, the number of firms enhancing corporate information sharing through Twitter, Facebook, LinkedIn will only grow. Besides the rising use of social media by companies, a recent study concluded that investor sentiment can not only be found in tweets, but the researchers also used Twitter sentiment to investigate the investor sentiment effect on stock price returns (Ranco et al., 2015). Their conclusion that the cumulative abnormal returns are significantly affected by sentiment expressed on Twitter suggest that tweets from an important Twitter account might have an effect on the stock returns by altering the sentiment. The study found statistical significance at a 1% level for several days after the event of that tweet; four days after a positive event and eight days after a negative one. This indicates a direct relation between Twitter content and market behaviour, and the detection of negative or positive investor sentiment through Twitter can be useful in deciding to sell or buy. As in the previously discussed studies, however, Ranco et al. suggest that extended research is required.

Online information is available in plenitude, and the use of social media especially gives a new dimension to all classic financial theories for stock pricing. Big data use by gathering tweets or search intensity on Google generates a source of information which can be used in extracting the sentiment of investors about certain companies. As mentioned above, a study performed in China concluded that investor sentiment influenced the share price, but only if that specific company had high investor attention (Guo et al, 2017). In fact, this just might be the case when the President of the United States gives that company attention by tweeting about it. This development presents opportunities, but also provides challenges for companies. The internet also provides the space to broadly disperse fake news, which can influence the share prices of companies (Carvalho et al., 2011). Moreover, the use of social media is not without risk, and the effect of real time sharing of financial information or policy changes through Twitter can contribute to the spread of errors and fake news. In the case of Twitter, however, the management of a company can resolve some issues by (pro)actively using an official Twitter account that provides real time communication.

(9)

Chapter 3 Data and Methodology

The Event Study Methodology was introduced by Fama et al. in 1969 and further researched and edited by MacKinlay in 1997. For this thesis, the effect of Donald Trump’s tweets on the return of stocks is measured by using MacKinlay’s method. The Event Study Methodology examines a specific event and measures its effects; in this case, the event of a tweet and its effect on the stock returns. MacKinlay proposes that in order to do this, the abnormal returns should be calculated, and then the cumulative abnormal returns of the stocks. The cumulative abnormal returns are the market’s reaction to the Twitter message. As discussed above, the assumption that the market handles all information efficiently would mean that the price before the tweet incorporates all available information (Fama, 1970) and that any abnormal returns should come from additional, previously unknown, information. The Event Study Methodology therefore tests for the semi-strong form of market efficiency. The stock returns are measured by computing returns of market indexes and the targeted companies.

3.1 Obtaining Data

The Twitter data is collected from an online database1_{containing all tweets of Donald Trump from} 2009. The tweets were extracted directly from Trump’s official Twitter account @realDonaldTrump. For this study, however, only the tweets from the moment Trump announced his candidacy on 16 June 2015 to 31 May 2018 are used. The original sample from this time window contains 11629 tweets, from which86 tweets have been collected in which Donald Trump directly tweets about a certain company. Retweets posted by Trump via @realDonaldTrump are excluded in the study. To avoid a selection bias, the companies listed on the NASDAQ-100 and NYSE-100 were selected from the Tweet database, as well as Toyota and Mazda, which were listed in an online article by Ethan Wolff-Mann in Yahoo Finance (2017) about businesses which have been directly targeted by Donald Trump2_{. All these companies are listed on the NASDAQ or New York Stock Exchange and their} shares are public. The tweets have some unusual characteristics; they contain new information in real time, but it is unclear how new it is exactly. The information could already be known by the investors, and even be copied from other media sources. Tweets from the President are not closely controlled, in contrast to other financially sensitive information.

1_{http://www.trumptwitterarchive.com/archive}

(10)

Table 1: List of targeted companies

Targeted company Number of tweets

Amazon 15 Apple 4 Boeing 3 Chrysler 7 Delta 1 Exxon 6 Facebook 7 Ford 7 GM 3 Intel 1 Lockheed 2 Mazda 3 Merck 3

New York Times 10

Nordstrom 1 Rexnord 2 Time Warner 1 Toyota 4 United Technologies 5 Walmart 1 Total 86

Some Tweets targeted multiple companies, these repeated tweets are used as independent events in this study.

In order to obtain the expected normal return, two estimation methods are used. The first method requires the Constant Mean Return Model, in which the returns of the estimation window are used to measure the return on the stocks. The average real return of the estimation window is used as normal return and subtracted from the actual measured return to obtain the abnormal stock return. The second estimation model is the Market Model, in which Ordinary Least Squares (OLS) regressions are used to estimate the normal returns, which will then be subtracted from the actual return to obtain the abnormal return. Both models are explained more thoroughly in Chapter 3.2.

The daily share prices at the closing of the stock market are gathered from DataStream3_{. With this} model the historical prices are used, so the daily return is given by:

𝑃𝑡−1 = (𝑃𝑡) 𝑅𝑡+ 1 𝑅𝑡= 𝑃𝑡− 𝑃𝑡−1 𝑃𝑡−1

3_{Datastream. (2018) Thomson Reuters Datastream. [Online]. Available at: Subscription Service (Accessed:} June 2018)

(11)

In this formula 𝑃𝑡 is the price of the indices on day t, 𝑃𝑡−1 the price of the indices on day t -1, and 𝑅𝑡

the daily return of the indices on day t. These historical prices are used with the Constant Mean Model and the Market Model to obtain the abnormal returns, with which the cumulative abnormal returns of the event windows are then calculated. However, investor sentiment can be both positive and

negative, and Trump refers both positively and negatively to the specific companies in his tweets. Therefore, it is expected that when the average cumulative abnormal return is calculated, the negative and positive abnormal returns affect each other, creating a downwards bias for any positive abnormal returns and an upwards bias for any negative abnormal return. To remedy this, the study is conducted not only on the entire group of 86 tweets, but also in a second and third group, in which the tweets are split according to their positive and negative content.

3.2 Methodology

To measure the effect of the tweet, the research in this thesis was conducted using three event windows. The first event window consists of a period before and after the tweet, and incorporates the possibility that the market may have already been aware of the information in Trump’s tweet. This may not only be the result of Trump repeating comments that he had previously expressed on other communication channels, but mostly because he is simply reacting to certain events, which have probably already had some sort of effect on the stock price of the targeted company. The second event window contains a period just after the tweet. This allows us to determine if the market is reacting specifically to information in the tweet. As discussed above, some studies concluded that investor sentiment only has effect if the targeted company has high investor attention or is popular (Guo et al., 2017 & Zheludev et al., 2014). If the event window is directly after the tweet, we can determine if the tweet is treated like new information, which then affected the stock return. As already mentioned above, two recent studies used a related form of the Event Study Methodology on the effects of Twitter sentiment on stock price returns (Ranco et al., 2015 & Sprenger et al., 2014). Both concluded that the volume as well as the sentiment of tweets really do have a statistically significant impact on stock returns. Finally, a third event window contains the day of the tweet only.

Hypothesis 1:

The null hypothesis for the effect of Trump’s tweets on the cumulative abnormal return (CAR) of a single company is that it will not be significantly different from zero. The alternative hypothesis is that the abnormal returns will be different from zero. The hypothesis can be expressed as:

(12)

Where CAR represents the cumulative abnormal returns, or the added up abnormal returns of the event window of the targeted company and i corresponds with the targeted tweets.

Hypothesis 2:

The null hypothesis for the effect of Trump’s tweet on the average cumulative abnormal returns of all targeted companies is not significant different from zero. The alternative hypothesis is that the average cumulative abnormal return is different from zero. The hypothesis can be expressed as:

𝐻0: 𝐶𝐴𝑅̅̅̅̅̅̅𝑖 = 0, 𝐻1: 𝐶𝐴𝑅̅̅̅̅̅̅𝑖 ≠ 0

Where 𝐶𝐴𝑅̅̅̅̅̅̅ is the average cumulative abnormal return of the companies and i corresponds with the targeted tweets.

In order to determine if the abnormal return has been affected, an estimation period is chosen before the event has occurred (T0, T1) to measure the normal returns. Mackinlay used an estimation window of 250 business days to conduct the original Event Study. And there is no reason to deviate from that, so the estimation window (T0 = -252, T1= -2) contains 250 days. In addition, multiple event periods – time frames around the event - will be set. The first event window will test a period window just after the tweet (T2, T3). The second event window will test the period surrounding the tweet (T1, T3). And a third event window, the day the tweet was sent.

Sometimes, however, the President’s tweets target individual companies multiple times. This can create a bias CAR or CAR̅̅̅̅̅̅ when the targeted tweet happens in the estimation window of another tweet being measured. To test the robustness of the Constant Mean Model, an estimation window of 3 months prior to the tweet is conducted, when no other tweet targeting the same company was sent by Trump. The second estimation window contains 75 business days, and the results are found in chapter 4.3.

For measuring the effect of the tweet three event windows surrounding the tweet are used. The first event window (t, t + 2) contains three days, the second event window (t - 2, t + 2) contains five days, and the third event window (t) contains only the day the tweet (time t) was sent. Or in some cases if the tweet was sent when the stock market was closed, the day after. By restricting the time window around the event to just shortly after the tweet, the study examines a statistically significant relation between the effect of the tweet on the Twitter sentiment. As existing studies have already proven; the

(13)

sentiment on twitter will have a significant effect on the stock returns (Ranco et al., 2015 & Sprenger et al., 2014).

𝑇0 Estimation window 𝑇1 𝑇2 𝑇3

Event

Mackinlay suggested four models for measuring the normal return of the security, of which two were used for this event study of Trump’s tweets; The Constant Mean Return Model and the statistical Market Model. In the Constant Mean Return Model the disturbance term for the security has an expectation of zero and variance 𝜎𝜀2. This is the most simple model, although Brown and Warner

(1980, 1985) concluded that it often results in similar abnormal returns results as the more sophisticated economical models.

𝑅𝑖𝑡 = 𝜇𝑖+ 𝜀𝑖𝑡 Where: 𝐸(𝜀𝑖𝑡) = 0 𝑣𝑎𝑟(𝜀𝑖𝑡) = 𝜎𝜀2 𝐸(𝑅𝑡+1) = 1 𝑇1 ∑ 𝑅𝑡 𝑇1 𝑡=𝑇0

The second suggested estimation model is the statistical Market Model. To analyse the effect of Trump’s tweets an OLS regression are used, consistent with event studies in similar research in which popular market indexes are used:

𝑅𝑖𝑡 = 𝛼𝑖+ 𝛽𝑖𝑅𝑚𝑡+ 𝜀𝑖𝑡

Where: 𝐸(𝜀𝑖𝑡) = 0 & 𝑣𝑎𝑟(𝜀𝑖𝑡) = 𝜎𝜀2_𝑡

The dependent variable 𝑅𝑖𝑡 is the return of the targeted company at time t and the independent

variable 𝑅𝑚𝑡 is the market return at time t. The constant term is represented by 𝛼𝑖 while the error term

is 𝜀𝑖𝑡 at time t. Finally, 𝛽𝑖 is the systematic risk, or volatility of the targeted company and calculated

by:

𝛽𝑖 =

𝐶𝑜𝑣(𝑅𝑖,𝑅𝑚)

𝑉𝑎𝑟(𝑅𝑚)

For using daily stock returns in event studies, the OLS Market Model does outperform the Constant Mean Return Model (Brown & Warner, 1985). In both cases, however, the non-normality of daily

(14)

returns measured no obvious impact on the Event Study Methodology, and the same was concluded in reference to the biases in estimating the Market Model in testing for abnormal returns. So, while both biases in the Market Model and non-normality are of no concern for the abnormal performance, Brown & Warner do express their concern about the choice of variance estimator to be used in the hypothesis test. Some evidence exists that the test statistics are improved by adjusting the estimating variance to reflect autocorrelation in the mean daily excess returns. However, they point out the improvements are small and only occur in special cases.

Estimating the impact of a tweet on the return requires measurement of the abnormal returns. To do this we subtract the expected return form the actual measured return.

𝐴𝑅𝑖𝑡 = 𝑅𝑖𝑡 − 𝐸(𝑅𝑖𝑡)

The former formula is rewritten as a dependent function of the estimated parameters 𝛼̂ & 𝛽̂, which are estimators of the population parameters α & 𝛽, and used to estimate the abnormal returns 𝐴𝑅𝑖𝑡 for the

selected event window.

𝐴𝑅𝑖𝑡 = 𝑅𝑖𝑡− 𝛼̂ − 𝛽̂𝑅(𝑀𝐾𝑇)𝑖𝑡

Next, the cumulative abnormal return is calculated, where i is represented by the tweet. The

summation of abnormal returns are measured during the event window, in order to obtain the response after the tweet. Two versions are used. The first event window contains two days before the tweet and two days after the tweet.

𝐶𝐴𝑅𝑖(𝑇1, 𝑇3) = ∑ 𝐴𝑅𝑖𝑡 𝑇3

𝑡=𝑇1

In the second event window, only the two days after the tweet are calculated.

𝐶𝐴𝑅𝑖(𝑇2, 𝑇3) = ∑ 𝐴𝑅𝑖𝑡 𝑇3

𝑡=𝑇2

To test the first hypothesis, if the cumulative abnormal return of a single company is significantly different from zero, a simple t student test is conducted.

𝑡 = 𝐶𝐴𝑅𝑖 𝑆𝐶𝐴𝑅

The variation of the cumulative abnormal return is calculated by: 𝑆𝐶𝐴𝑅2 = (𝑇2 − 𝑇1)𝑆𝐴𝑅𝑖

(15)

To test the second hypothesis, the average effect of Trump’s tweets on all the targeted companies must be determined to be significantly different then zero. An analysis of the company’s performance is determined and defined as the average cumulative abnormal returns. Where N tweets at time t will be calculated by: 𝐶̅𝐴𝑅̅̅̅̅𝑡 = 1 𝑁 ∑ 𝐶𝐴𝑅𝑖𝑡 𝑁 𝑖=1

The null hypothesis states that the average cumulative abnormal return is equal to zero, and the alternative hypothese states that average cumulative abnormal return is different from zero. To test the null hypothese for the event study using the Constant Mean Model, a student t-test is conducted, similar to the study conducted by Shin et al. (2013) in the form:

𝑡 =_𝑆𝐶𝐴𝑅̅̅̅̅̅̅ √𝑁

⁄ ~ 𝑡𝑁−1

Where N is the number of tweets from the sample and the standard deviation is calculated by:

𝑆 = √∑ (𝐶𝐴𝑅𝑖

𝑁

𝑖=1 − 𝐶𝐴𝑅̅̅̅̅̅̅)2

𝑁 − 1

Using the Market Model in combination with the event study a different test is conducted. To test the null hypothesis, the significance test similar to the one used by MacKinlay in his study of 1997 is used.

𝜃1=

𝐶𝐴𝑅

̅̅̅̅̅̅(𝑇1, 𝑇2)

𝑣𝑎𝑟(𝐶𝐴𝑅̅̅̅̅̅̅(𝑇1, 𝑇2))1/2 ~ 𝑁(0,1)

The average cumulative abnormal variance is calculated by the following calculations:

𝜎𝜀𝑖 2 ₌ 1 (𝑇1 − 𝑇0) − 2 ∑ (𝐴𝑅𝑖𝑡) 2 𝑇1 𝑡=𝑇0 + 1 𝑣𝑎𝑟(𝐴𝑅̅̅̅̅𝑡) = 1 𝑁2 ∑ 𝜎𝜀𝑖 2 𝑁 𝑖=1 𝑣𝑎𝑟(𝐶𝐴𝑅(𝑇1, 𝑇2)) = ∑ 𝑣𝑎𝑟(𝐴𝑅̅̅̅̅𝑡) 𝑇2 𝑡=𝑇1

(16)

After obtaining the CAR values and testing their significance, an OLS regression is conducted on the event window with the most significant CAR̅̅̅̅̅̅. Factors like high debt leverage or Market-to-Book ratios may be of influence on the effect of Trump’s tweets on the stock value of the mentioned businesses. Companies that are already in trouble are possibly more vulnerable to a reaction of investors on a tweet. To determine the effect of certain independent variables on the CARs, the following regression model will be used:

𝐶𝐴𝑅(𝑐𝑜𝑚𝑝𝑎𝑛𝑦)𝑡

= 𝛼 + 𝛽𝑅𝑂𝐴𝑡+ 𝛾

𝐷

𝐸𝑟𝑎𝑡𝑖𝑜𝑡+ 𝛿𝑀𝑎𝑟𝑘𝑒𝑡 𝐶𝑎𝑝𝑡+ 𝜃𝑆ℎ𝑎𝑟𝑒𝑠𝐻𝑒𝑙𝑑𝑏𝑦𝐼𝑛𝑣𝑒𝑠𝑡𝑜𝑟𝑡 + 𝜇𝐶𝑢𝑟𝑟𝑒𝑛𝑡 𝑟𝑎𝑡𝑖𝑜𝑡+ 𝜋𝑀𝑎𝑟𝑘𝑒𝑡𝑡𝑜𝐵𝑜𝑜𝑘𝑡+ 𝜀𝑡

The dependent variable is the cumulative abnormal return (CAR) of a specific company. The first independent variable 𝑅𝑂𝐴𝑡 represents the return on assets ratio (ROA) of the company. The second

independent variable 𝐷

𝐸𝑟𝑎𝑡𝑖𝑜𝑡 represents the Debt/Equity ratio of the company. 𝑀𝑎𝑟𝑘𝑒𝑡 𝐶𝑎𝑝𝑡

represents the size of the firm by Market Capitalization, and 𝑆ℎ𝑎𝑟𝑒𝑠𝐻𝑒𝑙𝑑𝑏𝑦𝐼𝑛𝑣𝑒𝑠𝑡𝑜𝑟𝑡 is the

percentage of shares held by private investors who can react to the tweets. 𝐶𝑢𝑟𝑟𝑒𝑛𝑡 𝑟𝑎𝑡𝑖𝑜𝑡 is a

liquidity ratio calculated by Current Assets/ Current Liabilities and 𝑀𝑎𝑟𝑘𝑒𝑡𝑡𝑜𝐵𝑜𝑜𝑘𝑡 represents the

ratio of the book value of the firm to its market value. Finally, the constant term is 𝛼, and the error term is represented by 𝜀𝑡. The OLS regression will be conducted on the most significant event

(17)

CHAPTER 4 Results

During the first stage of this thesis we researched two ways of calculating the expected normal returns of targeted companies. This was necessary to conduct the calculation of the abnormal returns for the event studies. Two models were suggested, the Constant Mean Model and the Market Model.

4.1 Constant Mean Model

The first hypothesis was tested on every single tweet using three different event windows with normal returns calculated with the Constant Mean Model. The null hypothesis stated that the effect of

Trump’s tweets on cumulative abnormal returns of a single company is not different from zero. As expected, several tweets measured a significant CAR as seen in table 2. However, for most single tweets the H0 hypothesis was not rejected, as the CAR value was not significant. In the first event window (t, t + 2) eight Tweets of the sample were found to have a significant CAR. The second event window (t - 2, t + 2) measured the CAR of five tweets and the third event widow (t) measured

seventeen of the 86 to be significantly different than zero at the level of 95% significance. With 90% significance we find a slightly higher amount of significant tweets. Appendix A contains all the tweets with measured cumulative abnormal returns for all event windows using the Constant Mean Model. Table 2: Constant Mean Model significant single Cumulative Returns during event windows.

Event Window

Events with significant CARs Positive significant CARs Negative sign. CARs α = 5% α = 10% α = 5% α = 10% α = 5% α = 10% Event [0, 2] ₈ ₁₇ ₃ ₅ ₅ ₁₂ Event [-2, 2] ₅ ₁₀ ₂ ₄ ₃ ₆ Event [0] ₁₇ ₂₅ ₇ ₁₄ ₁₀ ₁₁

The results of the expected returns and cumulative abnormal returns of all companies during the estimation windows and event windows are provided in tables 3 & 4. The mean of the cumulative abnormal return provides the necessary average abnormal cumulative 𝐶̅𝐴𝑅̅̅̅̅𝑡 for testing the second

hypothesis.

Table 3: Constant Mean Model summary of values of event windows 1,2,3

Total Minimum Maximum Mean Standard deviation Observations ER 0.0811 -0.0006 0.0025 0.0010 0.0022 86 CAR1 [0, 2] -0.1860 -0.0601 0.1411 -0.0022 0.0277 86 CAR2 [-2, 2] -0.0718 -0.1348 0.1332 -0.0008 0.0339 86 CAR3 [0] -0.0769 -0.0701 0.0549 -0.0009 0.0203 86

(18)

The second null hypothesis states that the cumulative average abnormal return of all targeted

companies is not different from zero. Results of testing the total effect of Trumps tweets on the stock returns are shown in table 4. As seen in the table, the CAR on average is examined to be negative, but the results are not significant with p < 0.05 or p < 0.1. So, with 95% or 90% confidence interval the overall results suggest that the tweets have no effect on the stock returns of the targeted companies. Table 4: Constant Mean Model results Average Cumulative Abnormal returns of all Tweets

Total CAR 𝐂𝐀𝐑̅̅̅̅̅̅ Standard

Deviation two sided t-value p-value Tweets Event [0, 2] -0.1860 -0.00225 0.0276 -0.7551 0.4525 86 Event [-2, 2] -0.0718 -0.00077 0.0356 -0.2011 0.8412 86 Event [0] -0.0769 -0.00089 0.0203 -0.4077 0.6846 86

However, when the tweets are separated in negative and positive tweets, the results are more significant, seen in tables 5 & 6. As predicted, the negative abnormal returns and the positive

abnormal returns cancelled each other out, which resulted in lower significance for the effect that the combined tweets are having on the stock returns.

The measured CAR̅̅̅̅̅̅ in the first event window (0, t + 2) does not differ significantly from zero for p < 0.05 for both the positive and negative tweets, although the negative CAR̅̅̅̅̅̅ measured is significant for p < 0.1. In the second event window (t - 2, t + 2) the CAR̅̅̅̅̅̅ for both the positive and negative tweets are significant for p < 0.05. In accordance with Fama’s Market Efficiency Theory, it seems that the market does not react to the new information provided by Trump, but to the high investor attention the tweets create for the company. This shows that the sentiment created by the tweets can be significant to the stock returns (Guo et al., 2017).

Table 5: Results Average Cumulative Abnormal Returns of the Positive Tweets

Positive Tweets CAR 𝐂𝐀𝐑̅̅̅̅̅̅ Standard Deviation One sided t-value p-value Tweets Event [0, 2] 0.0435 0.00106 0.0311 0.2185 0.4141 41 Event [-2, 2] 0.4186 0.01021 0.0364 1.7987 0.0399 41 Event [0] -0.0364 -0.00089 0.0199 -0.2858 0.6117 41

Table 6: Results Average Cumulative Abnormal Returns of the Negative Tweets

Negative Tweets CAR 𝐂𝐀𝐑̅̅̅̅̅̅ Standard Deviation One sided t-value p-value Tweets Event [0, 2] -0.2295 -0.00510 0.0246 -1.3904 0.0858 45 Event [-2, 2] -0.4904 -0.01090 0.0318 -2.2947 0.0132 45 Event [0] -0.0405 -0.00090 0.0209 -0.2889 0.3870 45

(19)

As seen in table 5 and 6, the most significant CAR̅̅̅̅̅̅ are measured in the second event window (t -2, t + 2). In order to determine if certain independent factors have an effect on the measured CAR̅̅̅̅̅̅, an OLS regression is conducted on the CAR measured using the Constant Mean Model of the specific companies.

Table 7: OLS regression with dependent variables with the Constant Mean Model CAR(-2,2)

Dependent variable All Tweets

Positive

Tweets Negative Tweets

CAR(-2,2) CAR(-2,2) CAR(-2,2)

Return on Assets -0.0367 -0.0534 -0.0384 (0.156) (0.350) (0.282) Debt/Equity ratio -0.0008 -0.0025 0.0147 (0.003) (0.003) (0.013) Market Capitalization 0.0000 0.0000 0.0000 (0.000) (0.000) (0.000)

Shares held by investors 0.0242* 0.0310* 0.0222

(0.013) (0.015) (0.044) Current ratio 0.0010 0.0019 0.0095 (0.010) (0.023) (0.023) Market-to-Book ratio 0.0001 0.0001 -0.0002 (0.000) (0.000) (0.000) α -0.0084 -0.0106 -0.0320 (0.018) (0.017) (0.049) 0.1438 0.3394 0.0705 F(k, n-k) 1.74 2.23 0.37 No. Observations, n 69 33 36

Note: *** significant at 1%, ** significant at 5% and * significant at 10% according to two-sided test. Note: P-values between parentheses

Whether the data sample contains all the tweets or only the positive tweets when conducting the OLS regression, the results are the same: In both cases the percentage of private investors’ shares are p < 0.1, which is significant explanation for the CAR’s of the companies. The remaining independent variables do not supply a significant explanation for the dependent CAR.

(20)

4.2 Market Estimation Model

The second part of this event study, a Market Model was used for measuring the abnormal returns. The First Hypothesis was tested on single firm cumulative abnormal returns. As suspected, most significant cumulative abnormal returns are measured in the third event window. Appendix C contains all the tweets with measured cumulative abnormal returns for all event windows using the Market Model.

Table 8: Market Model significant single Cumulative Returns during event windows.

Events with significant CARs

Positive significant CARs Negative sign. CARs

a = 5% a = 10% a = 5% a = 10% a = 5% a = 10%

[0, 2] 6 14 4 7 2 7

[-2, 2] 11 15 4 6 7 9

[0] 13 21 6 10 7 11

The second hypothesis, that the cumulative average abnormal return is not significant from zero, has been tested in three forms. First, the CAR̅̅̅̅̅̅ of all the tweets was measured for every event window. As found in table 9, none of the CAR̅̅̅̅̅̅ was significantly different from zero in either of the three event windows. This was expected as the positive and negative stock returns cancel each other out measuring the CAR̅̅̅̅̅̅.

Table 9: Results Average Cumulative Abnormal returns of all tweets

Deviation Two sided t-test p-value Tweets Event [0, 2] 0.0092 0.00011 0.0035 0.0304 0.9759 86 Event [-2, 2] -0.1454 -0.00169 0.0045 -0.3738 0.7095 86 Event [0] -0.0184 -0.00021 0.0020 -0.1055 0.9162 86

Second, all the CARs were measured from the positive tweets, and a t-test was conducted on the measured CAR̅̅̅̅̅̅. Event window 1 and 3 did not measure a significant result difference from zero in the average stock returns as seen in table 10. However, the second event window (t - 2, t + 2) did return a significant result with significance 95%.

Table 10: Results Average Cumulative Abnormal returns of the positive tweets

Positive CAR 𝐂𝐀𝐑̅̅̅̅̅̅ Standard Deviation Single sided t-test p-value Tweets Event [0, 2] 0.1416 0.00345 0.0036 0.9680 0.1694 41 Event [-2, 2] 0.3806 0.00928 0.0046 2.0151 0.0253 41 Event [0] 0.0569 0.00139 0.0021 0.6732 0.2523 41

(21)

Last, all the CARs were measured from the negative tweets, and a t-test was conducted on the measured CAR̅̅̅̅̅̅. Similar to testing the positive tweets only the result of the second event window measured significant for 90%.

Table 11: Results Average Cumulative Abnormal returns of the negative tweets

Negative CAR 𝐂𝐀𝐑̅̅̅̅̅̅ Standard Deviation Single sided t-test p-value Tweets Event [0, 2] -0.1325 -0.00288 0.0059 -0.4919 0.3126 45 Event [-2, 2] -0.5261 -0.01144 0.0076 -1.5130 0.0686 45 Event [0] -0.0752 -0.00164 0.0034 -0.4838 0.3155 45

In accordance with the results of the Constant Mean Model, the event study conducted in combination with the Market Model, the second event window returns the most significant CAR̅̅̅̅̅̅ . For further research on the variables affecting the CAR of a company, an OLS regression is conducted on the cumulative abnormal returns in the second window (t - 2, t + 2) with the results in table 12.

(22)

Table 12: OLS regressions with dependent variables with the Market Model CAR(-2,2)

Dependent variable All Tweets

Positive Tweets Negative Tweets

Market Model CAR(-2,2) CAR(-2,2) CAR(-2,2)

Return on Assets -0.0953 -0.6063* -0.0623 (0.175) (0.331) (0.328) Debt/Equity ratio -0.0076** -0.0099** -0.0086 (0.003) (0.003) (0.015) Market Capitalization 0.0000* 0.0000 0.0000 (0.000) (0.000) (0.000)

Shares held by investors 0.0234 0.0176 -0.0122

(0.015) (0.015) (0.052) Current ratio -0.0166 0.0145 -0.0321 (0.011) (0.022) (0.027) Market-to-Book ratio 0.0001** 0.0002*** 0.0001 (0.000) (0.000) (0.000) α 0.0276 0.0077 0.0601 (0.020) (0.017) (0.057) 0.1819 0.5102 0.0838 F(k, n-k) 2.3 4.51 0.44 No. Observations, n 69 33 36

Note: *** significant at 1%, ** significant at 5% and * significant at 10% according to two-sided test. Note: P-values between parentheses

Similar to the OLS regression conducted on the results of the Constant Mean Model in table 7, with the Market Model there are no significant independent variables if only the negative tweets are used. Conducting the regression on all the tweets, however, the results are significant values of p < 0.05 for the Debt/Equity ratio and Market-to-Book ratio, as well as of p < 0.1 for Market Capitalization. The OLS regression on the positive tweets have a high significance of p < 0.01 for Market-to-Book ratio. Furthermore, the Debt/Equity ratio is significant for p < 0.05 and the ROA for p < 0.1 for the positive tweets.

(23)

4.3 Robustness

Multiple tweets from the President about one company can create a biased abnormal return, influencing the robustness of the tests this study conducts. To check for robustness a second

estimation window is used in measuring the abnormal returns of the Constant Mean Model. Appendix B contains all the tweets with measured cumulative abnormal returns for all event windows using the Constant Mean Model with 75 business days as estimation window.

Table 13: Results of estimation window 75 business days Constant Mean Model.

Deviation Single sided t-value p-value Tweets Event [0, 2] -0.2458 -0.00286 0.0278 -0.9534 0.3431 86 Event [-2, 2] -0.2453 -0.00285 0.0347 -0.7617 0.4484 86 Event [0] -0.1359 -0.00158 0.0201 -0.7290 0.4680 86

Positive CAR 𝐂𝐀𝐑̅̅̅̅̅̅ Standard

Deviation Single sided t-value p-value Tweets Event [0, 2] 0.0475 0.00116 0.0310 0.2392 0.4061 41 Event [-2, 2] 0.4100 0.01000 0.0354 1.8064 0.0392 41 Event [0] -0.0400 -0.00098 0.0197 -0.3171 0.6236 41

Negative CAR 𝐂𝐀𝐑̅̅̅̅̅̅ Standard

Deviation Single sided t-value p-value Tweets Event [0, 2] -0.2933 -0.00652 0.0252 -1.7344 0.0449 45 Event [-2, 2] -0.6552 -0.01456 0.0349 -2.8026 0.0038 45 Event [0] -0.0959 -0.00213 0.0448 -0.3194 0.6246 45

The results suggest a slightly higher significance in the cumulative average abnormal returns of the negative tweets. This suggests that using an estimation window of 250 business days, in combination with the President’s tweets during that estimation window, creates a small upwards bias.

Table 16: Combination of Constant Mean Model p-values measured using two estimation windows.

p-values all p-values positive p-values negative Estimation 75 days 250 days 75 days 250 days 75 days 250 days Event [0, 2] 0.3431 0.4525 0.4061 0.4141 0.0449 0.0858

Event [-2, 2] 0.4484 0.8412 0.0392 0.0399 0.0038 0.0132

(24)

CHAPTER 5 Conclusion & Discussion

5.1 Conclusion

President Donald Trump may have become a source of influence for the stock market, international and domestic, in the US. Moreover, social media is a new source of financial performance reporting and measuring, and this trend creates challenges and opportunities for stakeholders and companies. Trump’s tweets may alter the financial sentiment online about certain companies, and therefore impact the targeted companies. This thesis aims to examine the relationship between Trump’s tweets and the targeted companies, by measuring the financial performance of their stock returns. With Presidents Trumps significant use of Twitter to proclaim future political and economic policy, do specific tweets directed to corporations effect their stock returns?

After estimating the Constant Mean Model and Market Model, an event study was conducted. The expected result was that the tweets would cause an abnormal cumulative return. To test this, a null hypothesis was used: The effect of Trump’s tweets on the cumulative abnormal returns of a single company will not be significantly different from zero. The main results indicate that some tweets had a significant effect on the stock market. For both the Constant Mean Model and Market Model, the most significant tweets were found using the shortest event window of only the day the tweet was sent. As expected, online communication like tweets are incorporated into the stock price almost directly.

A second null hypothesis was then tested: The effect of Trump’s tweet on the average cumulative abnormal returns of all targeted companies is not significantly different from zero. The final results differ in the various event windows. However, when the tweets are separated by their positive and negative content, the second event window (t - 2, t + 2) measures significant CAR̅̅̅̅̅̅ with p < 0.05 for the tests on all three event windows, in which the Constant Mean Model was used for estimating the expected returns (table 5 & 6). The event study conducted in combination with the Market Model show similar results, although the measured CAR̅̅̅̅̅̅ for negative tweets was only significant for p < 0.1 (table 8).

The conclusion of the event study using the Constant Mean Model provides a result that the tweets have a significant effect on the stock returns with p < 0.05 in the second event window (t - 2, t + 2). Using the Market Model the result is that the tweets have significant effect on the stock returns with p < 0.1. Using both the CAR of the second event window of the Constant Mean Model and the Market Model, OLS regressions were conducted. The results found that multiple independent variables significantly affect the CAR of the companies. Conclusively, there is significant abnormal return in

(25)

the timeframe of two days before and two days after the tweet. This suggests that there is an amplified effect of another event before that particular tweet, which might be immeasurable. The tweet itself does not provide significant abnormal stock returns, as in accordance with the semi-strong Market Efficiency Theory, and no new information, previously unknown to the market, is provided. Some may argue that the effect Trump is creating is biased towards coverage of sensational events (Solomon & Soltes, 2011). A possible explanation might be that the tweets highlight the current Book-to-Market or Debt/Equity ratios of the targeted company, which in the OLS regressions are of significant effect on explaining the CARs. The market is merely reacting to that highlighted financial information. The results in the present study, however, are in line with those of Guo et al. (2017); that there is a greater effect when there is a high attention for a certain event.

5.2 Discussion

As stated above, there are some reservations about the results of this study and its possible conclusions. Already, two independent studies concluded that twitter sentiment only affects stock returns when there is high investor attention and the company is well known (Guo et al., 2017 & Zheludev et al., 2014). The cumulative abnormal returns might be caused by another event other than the tweet itself. The tweet, however, highlights the event and therefore creates the sentiment effect on the stock returns, which should not be considered a limitation, but rather gives a different perspective on the effect of the tweets.

Unfortunately, the number of significant tweets usable in this study was small. For future research, assuming the President continues to send tweets targeting companies, accumulating a larger usable sample, I suggest dividing those tweets into various subgroups. Dividing the targeted companies according to size, industry, performance and other macroeconomic factors could contribute to fine-tuning the results of the tests on the stock returns. In addition, a broader approach should be taken to research the effect of President Trump’s tweets on the stock returns of specific companies. For example, tweets targeting oil producing nations may indirectly affect the stock returns of US-listed oil companies.

The dataset this thesis uses contains some research limitations. The first limitation is subjective judgement on separation of positive versus negative tweets. Because the sample was small, a objective split between negative and positive tweets based on keywords indicating the sentiment of Trump’s tweets was not possible, resulting in a selection bias. Furthermore, the fact that Trump sends several similar tweets targeting the same company results in a sample bias, with the data sample being too dependent on the return of stocks of certain companies.

(26)

Besides testing for abnormal returns, further research on abnormal volatility would contribute to a full picture of the effect of Trump’s tweets on the trading of targeted stocks. Policies of U.S. presidents often tend to have an effect on stock prices, but this thesis concludes that Trump’s tweets can have a significant effect on the return of stocks. So, for further investigation in the direction of corporate governance, it should be debated if regulations of some sort are necessary, since it is clear that the tweets have such a manipulative effect on stock returns.

This study concludes that Trump’s tweets have a significant effect, in the measured timeframe, on the stock returns of the targeted companies. It would be premature, however, to conclude that this effect is a direct cause of new information provided by President Trump. Further research is needed that excludes variables that might affect the stock return in the measured event window.

(27)

References

Alexander R. M., Gentry J. K. (2014). Using social media to report financial results, Business Horizons, 57(2), 161-167.

Baker M., Wurgler J. (2006). Investor Sentiment and the Cross‐Section of Stock Returns. The Journal of Finance, 61 (4), 1645-1680.

Baker M., Wurgler J. (2007). Investor Sentiment in the Stock Market, Journal of Economic Perspectives, American Economic Association, 21(2), 129-152.

Bollen J., Mao H., Zeng X. (2011). Twitter mood predicts the stock market, Journal of Computational Science, 2 (1), 1-8.

Brown S. J., Stephen J., Warner J.B. (1984). Measuring Security Price Performance, Journal of Financial Economics, 8(3), 205-58.

Brown S.J. (1985). Using Daily Stock Returns: The Case of Event Studies, Journal of Financial Economics, 14(1), 3-31.

Carvalho C., Klagge N., Moench E. (2011). The persistent effects of a false news shock, Journal of Empirical Finance, 18(4), 597-615

Chan W.S. (2003). Stock price reaction to news and no-news: drift and reversal after headlines, Journal of Financial Economics, 70 (2), 223–260

Corea, F. (2016). Can Twitter Proxy the Investors Sentiment? The Case for the Technology Sector, Big Data Research, 4, 70-74

Curme C., Preis T., Stanley H.E., Moat H.S. (2014). Quantifying the semantics of search behavior before stock market moves, Proc National Academy of Sciences, 111, 11600–11605 Fama E.F. (1970). Efficient Capital Markets: A Review of Theory and Empirical Work, The Journal of Finance, 25(2), 383-417.

Fama, E.F., Fisher L., Jensen M.C., Roll R. (1969). The Adjustment of Stock Prices to New

Information, International Economic Review, 10(1), 1–21.

Guo K., Sun Y., Qian X. (2017). Can investor sentiment be used to predict the stock price? Dynamic analysis based on China stock market. Physica A: Statistical Mechanics and its Applications, 469, 390-396.

Joseph K., Wintoki M.B., Zhang Z. (2011). Forecasting abnormal stock returns and trading volume using investor sentiment: Evidence from online search, International Journal of Forecasting, 27(4), 1116–1127

MacKinlay A. (1997). Event Studies in Economics and Finance. Journal of Economic Literature, 35(1), 13-39.

Ranco G., Aleksovski D., Caldarelli G., Grčar M., Mozetič I. (2015). The Effects of Twitter Sentiment on Stock Price Returns. PLOS ONE, 10(9), e0138441

Sharpe W.F. (1964). Capital Asset Prices – A Theory of Market Equilibrium Under Conditions of Risk, Journal of Finance, 19(3), 425–442

(28)

Shin H., Richardson R., Soluade O. (2013). Review of J. D. Power Quality Measurement of U. S. and Japanese Automotive Industry, Communications of the IIMA, 13(1), Article 3. Solomon D., Soltes E. (2011). The determinants of coverage in the business press. Working paper, University of Southern California.

Sprenger T.O., Sandner P.G., Tumasjan A., Welpe I.M. (2014). News or Noise? Using Twitter to Identify and Understand Company-specific News Flow. Journal of Business Finance & Accounting, 41(7– 8), 791–830

Thorbecke W. (1997). On Stock Market Returns and Monetary Policy. The Journal of Finance, 52(2) , 635-654.

Zheludev I., Smith R., Aste T. (2014). When Can Social Media Lead Financial Markets? Scientific Reports, 4, 4213.

Zouaoui M., Nouyrigat G., Beer F. (2011). How Does Investor Sentiment Affect Stock Market Crises? Evidence from Panel Data, Financial Review, 46(4), 723-747.

(29)

APPENDIX A: Results Constant Mean Model, estimation window 250 business days

CODE CAR1 t-test 1 Sign 1 CAR2 t-test 2 Sign 2 CAR3 t-test 3 Sign 3 Standard

Deviation AMZ1 _-0.0079 _-0.2861 _0.3878 _0.0163 _0.4597 _0.3235 _0.0273 _1.7212 _0.0444 _0.0159 AMZ2 _0.0510 _1.8973 _0.0306 _-0.0051 _-0.1465 _0.4419 _0.0126 _0.8109 _0.2098 _0.0155 AMZ3 _-0.0300 _-1.1157 _0.1338 _-0.0228 _-0.6567 _0.2566 _-0.0540 _-3.4787 _0.0004 _0.0155 AMZ4 _-0.0300 _-1.1157 _0.1338 _-0.0228 _-0.6567 _0.2566 _-0.0540 _-3.4787 _0.0004 _0.0155 AMZ4 _-0.0300 _-1.1157 _0.1338 _-0.0228 _-0.6567 _0.2566 _-0.0540 _-3.4787 _0.0004 _0.0155 AMZ5 _-0.0483 _-1.8702 _0.0324 _-0.1348 _-4.0451 _0.0001 _0.0087 _0.5820 _0.2810 _0.0149 AMZ6 _-0.0023 _-0.1018 _0.4596 _0.0023 _0.0797 _0.4684 _0.0126 _0.9694 _0.1675 _0.0130 AMZ7 _-0.0280 _-1.3401 _0.0919 _-0.0148 _-0.5491 _0.2922 _-0.0057 _-0.4717 _0.3192 _0.0120 AMZ8 _-0.0392 _-1.8717 _0.0323 _-0.0360 _-1.3315 _0.0933 _-0.0036 _-0.2939 _0.3848 _0.0121 AMZ9 _0.0030 _0.1475 _0.4415 _0.0104 _0.4004 _0.3449 _-0.0004 _-0.0360 _0.4857 _0.0116 AMZ9 _0.0030 _0.1475 _0.4415 _0.0104 _0.4004 _0.3449 _-0.0004 _-0.0360 _0.4857 _0.0116 AMZ9 _0.0030 _0.1475 _0.4415 _0.0104 _0.4004 _0.3449 _-0.0004 _-0.0360 _0.4857 _0.0116 AMZ10 _0.0071 _0.3538 _0.3622 _0.0087 _0.3365 _0.3686 _-0.0042 _-0.3635 _0.3586 _0.0116 AMZ10 _0.0071 _0.3538 _0.3622 _0.0087 _0.3365 _0.3686 _-0.0042 _-0.3635 _0.3586 _0.0116 AMZ11 _-0.0127 _-0.6434 _0.2608 _-0.0423 _-1.6620 _0.0501 _0.0126 _1.1043 _0.1363 _0.0114 APL1 _0.0083 _0.4302 _0.3341 _0.0001 _0.0048 _0.4981 _0.0150 _1.3437 _0.0913 _0.0111 APL2 _-0.0055 _-0.2334 _0.4080 _-0.0235 _-0.7747 _0.2203 _0.0038 _0.2789 _0.3905 _0.0135 APL3 _-0.0132 _-0.4383 _0.3311 _-0.0148 _-0.3821 _0.3517 _-0.0013 _-0.0744 _0.4704 _0.0173 APL3 _-0.0132 _-0.4383 _0.3311 _-0.0148 _-0.3821 _0.3517 _-0.0013 _-0.0744 _0.4704 _0.0173 BNG1 _0.0202 _0.7536 _0.2266 _0.0181 _0.5228 _0.3012 _0.0002 _0.0146 _0.4942 _0.0155 BNG2 _0.0008 _0.0312 _0.4876 _0.0083 _0.2435 _0.4041 _-0.0006 _-0.0362 _0.4856 _0.0152 BNG3 _0.0232 _1.1575 _0.1251 _0.0339 _1.3106 _0.0967 _0.0096 _0.8339 _0.2033 _0.0116 CHR1 _0.0598 _1.1624 _0.1242 _0.1332 _2.0068 _0.0240 _0.0133 _0.4486 _0.3274 _0.0297 CHR2 _0.1411 _3.6465 _0.0002 _0.1060 _2.1221 _0.0184 _0.0300 _1.3417 _0.0916 _0.0223 CHR3 _-0.0078 _-0.1975 _0.4219 _0.0569 _1.1162 _0.1337 _-0.0320 _-1.4063 _0.0816 _0.0228 CHR3 _-0.0078 _-0.1975 _0.4219 _0.0569 _1.1162 _0.1337 _-0.0320 _-1.4063 _0.0816 _0.0228 CHR3 _-0.0078 _-0.1975 _0.4219 _0.0569 _1.1162 _0.1337 _-0.0320 _-1.4063 _0.0816 _0.0228 CHR4 _0.0418 _1.0884 _0.1397 _0.0042 _0.0841 _0.4666 _0.0144 _0.6486 _0.2592 _0.0222 CHR5 _-0.0361 _-0.9554 _0.1710 _-0.0157 _-0.3216 _0.3743 _-0.0175 _-0.8010 _0.2127 _0.0218 DLT1 _-0.0480 _-1.3456 _0.0910 _-0.0564 _-1.2257 _0.1118 _-0.0404 _-1.9610 _0.0266 _0.0206 EX1 _0.0165 _0.7826 _0.2180 _0.0259 _0.9510 _0.1721 _0.0217 _1.7799 _0.0393 _0.0122 EX2 _-0.0020 _-0.0944 _0.4625 _0.0271 _1.0084 _0.1580 _0.0171 _1.4269 _0.0786 _0.0120 EX3 _-0.0175 _-1.0536 _0.1475 _-0.0243 _-1.1347 _0.1298 _0.0044 _0.4627 _0.3224 _0.0096 EX3 _-0.0175 _-1.0536 _0.1475 _-0.0243 _-1.1347 _0.1298 _0.0044 _0.4627 _0.3224 _0.0096 EX4 _-0.0141 _-0.8498 _0.1989 _-0.0198 _-0.9257 _0.1786 _-0.0038 _-0.3976 _0.3460 _0.0096 EX4 _-0.0141 _-0.8498 _0.1989 _-0.0198 _-0.9257 _0.1786 _-0.0038 _-0.3976 _0.3460 _0.0096 FB1 _-0.0004 _-0.0191 _0.4924 _-0.0148 _-0.5248 _0.3005 _-0.0012 _-0.0951 _0.4622 _0.0126 FB1 _-0.0004 _-0.0191 _0.4924 _-0.0148 _-0.5248 _0.3005 _-0.0012 _-0.0951 _0.4622 _0.0126 FB2 _-0.0290 _-1.4719 _0.0723 _-0.0375 _-1.4757 _0.0718 _-0.0225 _-1.9801 _0.0254 _0.0114 FB2 _-0.0290 _-1.4719 _0.0723 _-0.0375 _-1.4757 _0.0718 _-0.0225 _-1.9801 _0.0254 _0.0114 FB3 _0.0365 _1.9167 _0.0293 _-0.0026 _-0.1048 _0.4584 _0.0200 _1.8129 _0.0367 _0.0110 FB4 _-0.0438 _-2.2977 _0.0120 _-0.0544 _-2.2125 _0.0148 _-0.0046 _-0.4145 _0.3398 _0.0110 FB5 _0.0112 _0.5878 _0.2791 _0.0047 _0.1900 _0.4249 _-0.0006 _-0.0540 _0.4785 _0.0110 FRD1 _0.0138 _0.4823 _0.3154 _-0.0051 _-0.1372 _0.4456 _0.0032 _0.1931 _0.4237 _0.0165 FRD1 _0.0138 _0.4823 _0.3154 _-0.0051 _-0.1372 _0.4456 _0.0032 _0.1931 _0.4237 _0.0165

The effect of president Trump's Twitter messages on stock market returns