政治人物，推特與金融市場：來自川普推特的證據 - 政大學術集成

全文

(1)國立政治大學應用經濟與社會發展英語碩士學位學程 International Master’s Program of Applied Economics and Social Development College of Social Sciences National Chengchi University. 治政碩士論文大. 立 Master’s Thesis ‧. ‧ 國. 學. 政治人物，推特與金融市場：來自川普推特的證據. y. Nat. sit. n. al. er. io. Politicians, Twitter, and Financial Markets: Evidence from Donald Trump’s tweets. Ch. engchi. i n U. v. Student: Margaret A. Consunji (孔美玲) Advisor: Yang, Tzu-Ting (楊子霆) Co-Advisor: Lo, Kuang-Ta (羅光達). 中華民國 108 年 06 月 June 2019. DOI:10.6814/NCCU201900722.

(2) Acknowledgements I would first like to extend my utmost gratitude to my main advisor, Professor Tzu-Ting Yang, whose endless patience, guidance, and encouragement have made this thesis possible. I would also like to express my great appreciation and gratitude for my cosupervisor, Professor Kuang-Ta Lo for his invaluable comments and suggestions on this thesis.. Special thanks should also be given to the Applied Economics and Social Development Master’s Program (IMES) staff who have been incredibly patient and helpful all throughout the thesis process.. 立. 政治大. Finally, I would also like to thank my friends and family for their unfailing support and. ‧ 國. 學. encouragement throughout my study.. ‧. n. er. io. sit. y. Nat. al. Ch. engchi. i. i n U. v. DOI:10.6814/NCCU201900722.

(3) Abstract The use of Twitter as a key political communication tool has become synonymous with U.S. President Donald Trump’s regime. However, Trump’s tweets can also tend to be unabashedly critical of companies, states, and other political figures. Whether or not these negative tweets have an impact on financial markets is debatable. This paper uses the synthetic control method (SCM) to examine the effects of Trump’s negative trade- and business-related tweets on financial markets, particularly stock prices and exchange rates. Three publicly traded U.S. companies (Boeing, Amazon, Harley-Davidson) and three currencies (Euro, Canadian dollar, Mexican peso) were chosen, while 1-2 tweets were. 治政大each treatment unit, two synthetic extensive control unit data was also collected. Then, for 立 control models were created, with one model containing all outcome lags and all covariates collected for each treatment unit. To create the synthetic control for each treatment unit,. ‧ 國. 學. whilst the other contained all outcome lags and some covariates. We found that for each treatment unit, the two models were similar, indicating that the results were robust. Overall,. ‧. we found that results varied depending on the “target” of Trump’s tweets, with the causal effect being most significant for Amazon and Mexico, likely due to the fact that traders or. Nat. sit. y. investors may react differently to Trump’s tweets and may base their decisions on certain. io. er. company- or country-specific characteristics or features, such as type of industry, trade ties, and geographical proximity, among others.. n. al. Ch. engchi. i n U. v. Keywords: Trump, Synthetic Control Method, Twitter, Stock Prices, Exchange Rates. ii. DOI:10.6814/NCCU201900722.

(4) Table of Contents LIST OF FIGURES ............................................................................................................ v LIST OF TABLES ............................................................................................................ vii I. INTRODUCTION ........................................................................................................... 1 1.1. Background .............................................................................................................. 1 1.2 Objectives and significance ...................................................................................... 3 1.3 Methods..................................................................................................................... 3. 政治大 II. LITERATURE REVIEW 立 ............................................................................................... 6 1.4 Structure .................................................................................................................... 4. ‧ 國. 學. 2.1 The Effect of Political and Macroeconomic News on Financial Markets ................ 6 2.2 Twitter, Tweets, and Influence ................................................................................. 8. ‧. 2.3 Connecting the Dots between Tweets and Financial Markets ................................ 10. sit. y. Nat. 2.4 President Trump’s Tweets and Financial Markets.................................................. 11. io. er. III. METHODOLOGY ..................................................................................................... 15 3.1 Data Collection ....................................................................................................... 15. n. al. Ch. i n U. v. 3.1.1 Tweets ...............................................................................................................15. engchi. 3.1.2 Treatment Companies and Countries ................................................................18 3.1.3 Dependent and Independent Variables .............................................................19 3.2 Theoretical Background .......................................................................................... 19 IV. RESULTS AND DISCUSSION ................................................................................. 27 4.1 Stock Price Models ................................................................................................. 27 4.1.1 Boeing ...............................................................................................................29 4.1.2 Amazon .............................................................................................................32 4.1.3 Harley-Davidson ...............................................................................................35 4.2 Exchange Rate Models ........................................................................................... 37. iii. DOI:10.6814/NCCU201900722.

(5) 4.2.1 Euro ...................................................................................................................40 4.2.2 Canadian dollar .................................................................................................44 4.2.3 Mexican peso ....................................................................................................46 V. CONCLUSION ............................................................................................................ 50 REFERENCES ................................................................................................................. 52 Appendices ........................................................................................................................ 59 Appendix 1 .................................................................................................................... 59 Appendix 2 .................................................................................................................... 61. 政治大 Appendix 4 .................................................................................................................... 64 立 Appendix 3 .................................................................................................................... 63. ‧ 國. 學. Appendix 5 .................................................................................................................... 64 Appendix 6 .................................................................................................................... 65. ‧. Appendix 7 .................................................................................................................... 65. sit. y. Nat. Appendix 8 .................................................................................................................... 66. io. er. Appendix 9 .................................................................................................................... 66 Appendix 10 .................................................................................................................. 67. n. al. Ch. i n U. v. Appendix 11 .................................................................................................................. 67. engchi. Appendix 12 .................................................................................................................. 68 Appendix 13 .................................................................................................................. 68 Appendix 14 .................................................................................................................. 69 Appendix 15 .................................................................................................................. 70. iv. DOI:10.6814/NCCU201900722.

(6) LIST OF FIGURES Figure 1: Boeing Synthetic Control Graph (Model 1) ...................................................... 30 Figure 2: Boeing Placebo Effects Graph (Model 1) ......................................................... 30 Figure 3: Boeing Synthetic Control Graph 1 (Model 2) ................................................... 31 Figure 4: Boeing Placebo Effects Graph (Model 2) ........................................................ 31 Figure 5: Amazon Synthetic Control Graph (Model 1) .................................................... 32 Figure 6: Amazon Placebo Effects Graph (Model 1) ...................................................... 33. 政治大. Figure 7: Amazon Synthetic Control Graph 1(Model 2) .................................................. 33. 立. Figure 8: Amazon Placebo Effects Graph (Model 2) ..................................................... 34. ‧ 國. 學. Figure 9: Harley-Davidson Synthetic Control Graph (Model 1) ...................................... 35. ‧. Figure 10: Harley-Davidson Placebo Effects Graph (Model 1) ....................................... 36 Figure 11: Harley-Davidson Synthetic Control Graph (Model 2) .................................... 36. y. Nat. io. sit. Figure 12: Harley-Davidson Placebo Effects Graph (Model 2) ....................................... 37. n. al. er. Figure 13: Euro Synthetic Control Graph (Model 1) ....................................................... 42. Ch. i n U. v. Figure 14: Euro Placebo Effects Graph (Model 1) ........................................................... 42. engchi. Figure 15: Euro Synthetic Control Graph (Model 1) ....................................................... 43 Figure 16: Euro Placebo Effects Graph (Model 2) .......................................................... 43 Figure 17: Canadian dollar Synthetic Control Graph (Model 1) ...................................... 44 Figure 18: Canadian dollar Placebo Effects Graph (Model 1) ......................................... 45 Figure 19: Canadian dollar Synthetic Control Graph (Model 2) ...................................... 46 Figure 20: Canadian dollar Placebo Effects Graph (Model 2) ......................................... 46 Figure 21: Mexican peso Synthetic Control Graph (Model 1) ......................................... 47 Figure 23: Mexican peso Synthetic Control Graph (Model 2) ......................................... 48 v. DOI:10.6814/NCCU201900722.

(7) Figure 24: Mexican peso Placebo Effects Graph (Model 2)............................................. 49. 立. 政治大. ‧. ‧ 國. 學. n. er. io. sit. y. Nat. al. Ch. engchi. vi. i n U. v. DOI:10.6814/NCCU201900722.

(8) LIST OF TABLES Table 1: Trump tweets targeting publicly traded companies in the U.S. .......................... 16 Table 2: Trump tweets targeting other countries .............................................................. 17 Table 3: Summary of Synthetic Control Models for Stock Prices.................................... 28 Table 4: Summary of Synthetic Control Models for Exchange Rates .............................. 39. 立. 政治大. ‧. ‧ 國. 學. n. er. io. sit. y. Nat. al. Ch. engchi. vii. i n U. v. DOI:10.6814/NCCU201900722.

(9) I. INTRODUCTION 1.1. Background Rapid technological development has allowed more and more people to gain access to the internet, reshaping connectivity and the dispersal of information. It is estimated that from 2018 to 2019, the number of internet users around the world grew about 9.1%, or about 367 million new users, due to improved accessibility brought on by new developments in mobile phone technology (Kemp, 2019). In effect, increased digital connectivity has also. 治政大 have not only revolutionized and Snapchat, among others. These websites and applications 立 social networking; they have also provided a means for users to create and share content. ushered in the development of social media platforms such as Twitter, Facebook, Instagram,. ‧ 國. 學. with a much wider audience. Social media usage, like internet usage, has likewise grown considerably, rising by about 9.0% or by 288 million new users from 2018 to 2019 (Kemp,. ‧. 2019). And recently, social media has shown itself to be an extremely popular communication tool not just for the average user, but for politicians and highly influential. er. io. sit. y. Nat. figures as well.. With regards to social media and politics, one particular social media platform stands out. n. al. i n U. v. — Twitter. With a large and growing user base, Twitter has become an indispensable tool. Ch. engchi. for real-time political communication. Users communicate through “tweets” which are 280-character messages which can include text, images, videos, or links. The service is free to use and users are free to follow other Twitter users to subscribe to more updates. In addition, Twitter features a “retweet” function which allows users to repost another user’s tweet. And despite Twitter’s relatively restrictive character limit, politicians have learned how to fully utilize Twitter to broadcast their political views, boost their campaigns, and engage with the public in real time, essentially extending a politicians’ reach or influence across the public sphere. Indeed, this seems to be the case with individuals such as U.S. President Donald Trump, who has used Twitter as both a key campaigning tool during the electoral race, as well as a major communication tool during his presidency.. 1. DOI:10.6814/NCCU201900722.

(10) Since formally announcing his presidential campaign in 2015, all eyes seem to be on Trump and at the time of writing this, Trump’s follower count has breached past 61 million people (Trump, n.d.). His rather bombastic statements regarding race, immigration, trade, and politics, have placed him in the media spotlight for quite some time. And this focus on Trump has perhaps, been further amplified by the frequency at which he tweets. Brueninger (2018) noted that in 2018, Trump published over 3,400 tweets, which puts his frequency at about 10 tweets per day and effectively making Twitter the president’s main means of communication.. 政治大 slightly more than half of Trump’s tweets were negative, while about 40% were positive 立 In addition, a study by Clemson University’s Social Media Listening Center found that. and about 9-10% were neutral (Mak, 2019). In the same study, researchers found that. ‧ 國. 學. Trump’s negative tweets were more likely to be recirculated (through the retweet function on Twitter), which entails that Trump’s negative tweets do indeed get more attention (Mak,. ‧. 2019). It is not surprising then, that the American media portrays Trump in a more negative light, often emphasizing his more accusatory and negative tweets (Kurtzleben, 2017).. sit. y. Nat. al. er. io. To some extent, these tweets can be interpreted as official presidential statements, which. v. n. would make them points of interest for traders, in the same way, they would pay attention. Ch. i n U. to central bank announcements, for instance. Moreover, the media often tends to highlight. engchi. the market-moving nature of Trump’s tweets, and have released headlines such as: “Trump’s tweets swing stock market amid trade deal uncertainty” (DiChristopher, 2019); “With Two Tweets, Trump Shatters Historic Calm in Global Markets” (Ossinger & Patterson, 2019), and “Trump’s latest tweet takes down Amazon stock and the Nasdaq” (Goldman, 2018), to name a few. But how influential are these tweets and are these tweets sufficient enough to trade on? Does the media tend to overplay this effect?. 2. DOI:10.6814/NCCU201900722.

(11) 1.2 Objectives and significance To answer these questions, this research has selected six negative tweets pertaining to companies and countries to examine whether or not these tweets affect financial markets, particularly equity markets and foreign exchange markets. This research adds to the growing literature in the field of social media research by analyzing the causal relationship between Trump’s tweets and stock prices, as well as exchange rates. This thesis specifically focuses on Trump’s more negative tweets pertaining to trade and business — themes that tend align with Trump’s political credo about “Making America Great Again”. Generally,. 政治大. we would expect these tweets to have a negative impact on stock prices and exchange rates, although as seen in the summary of results, this may not always seem to be the case.. 立. ‧ 國. 學. With the way the media emphasizes the market-moving aspect of Trump’s tweets, this could possibly lead to more attention-based trading, as opposed to trading based on all available information (Rayarel, 2018). This in effect, could create more volatility and. ‧. inefficiency in financial markets. This thesis posits that the effects of Trump’s tweets tend. y. Nat. to be overplayed and that the causal relationship between Trump’s tweets and financial. sit. markets must be further explored. In doing so, this should provide a better understanding. al. er. io. of these linkages and eventually help traders and investors make more informed and. v. n. rational decisions, as opposed to immediately reacting to Trump’s tweets.. 1.3 Methods. Ch. engchi. i n U. To test the causal relationship between Trump’s tweets and financial markets, this thesis utilizes the synthetic control (SC) method, which is a statistical method that evaluates the effect of an event or intervention. The SC method has been most commonly used for state policy evaluation to analyze the effects of reforms or the implementation of laws and policies (Abadie, Diamond, & Hainmueller, 2010; Abadie, Diamond, & Hainmueller, 2015). However, use cases have expanded to other fields such as advertising (Tirunillai & Tellis, 2017), finance (Opatrny, 2017), educational assessment (Johnson, 2013), and health. 3. DOI:10.6814/NCCU201900722.

(12) policies (Kreif et al., 2016), among others. In contrast, the use of the SC method in social media analytics has not been as thoroughly explored yet. Moreover, the existing body of research on this particular topic -- politicians’ tweets and financial markets -- is primarily dominated by machine-learning studies and sentiment analysis. These studies often focus more on or topic-specific tweets, as opposed to userspecific tweets (Mittal & Goel, 2011; Ozturk & Ciftci, 2014; Ranco, Aleksovski, Caldarelli, Grčar, & Mozetič, 2015; Pagolu, Reddy, Panda, & Majhi, 2016; Nisar & Yeung, 2018). And while these studies are particularly helpful for those interested in the predictive capabilities of social media data, these studies often gloss over the causal relationship. 政治大 there are also papers that do in fact analyze the impact of Trump’s tweets on financial 立. between these tweets and movements in stock prices or exchange rates. Though limited,. markets, although the methodology used in these studies differ from the one presented in. ‧ 國. 學. this thesis (Born, Myers, & Clark, 2017; Colonescu, 2018; Ge, Kurov, & Wolfe, 2019; Rayarel, 2018). The existing literature in this particular topic is further discussed in the. ‧. methodology chapter of this thesis.. Nat. al. er. io. sit. y. 1.4 Structure. n. The arrangement of chapters in this study is as follows. The literature review forms the. Ch. i n U. v. second chapter of this thesis and follows right after this introduction. It provides a brief. engchi. explanation of the theoretical framework of this study and discusses the role of social media in politics, as well as the research that has emerged from this. More importantly, the literature review touches on relevant existing literature, particularly studies that focus on how Twitter impacts financial markets. It then provides more details on the use of Twitter among politicians, particularly U.S. President Donald Trump, and how this might affect financial markets as well. In the third chapter, the methodology of this study elaborates on the main statistical procedure used in this study -- the synthetic control method. This chapter also describes the data collection procedure, including the sources used and variables chosen. The fourth chapter presents the results for each event or tweet studied and the interpretation of these results. Finally, the conclusion ties everything together and. 4. DOI:10.6814/NCCU201900722.

(13) summarizes the main points and findings of this thesis. The final chapter also discusses limitations and suggestions to improve on this study.. 立. 政治大. ‧. ‧ 國. 學. n. er. io. sit. y. Nat. al. Ch. engchi. 5. i n U. v. DOI:10.6814/NCCU201900722.

(14) II. LITERATURE REVIEW This section first elaborates on the effect of political and macroeconomic news on financial markets. This relates to the main objective of this thesis, which is to examine whether or not Trump’s tweets have any effect on stock prices and exchange rates. This chapter also includes a brief discussion about Twitter, influence, and how this ties in with the use of Twitter in the political sphere. Following this is a brief review of the existing studies that study the effects of tweets on financial markets. Finally, we then discuss some of the literature surrounding the study of Trump’s tweets and how these tweets might affect financial markets.. 立. 政治大. 2.1 The Effect of Political and Macroeconomic News on. ‧ 國. 學. Financial Markets. ‧. What shapes a trader or an investor’s buy or sell decisions? Most economic research would. sit. y. Nat. suggest that the average person is a rational agent who aims to maximize wealth and minimize risk (Barber & Odean, 2013). In the real world, this may not always hold, as. er. io. investors’ and traders’ decisions may be influenced by where they live, where they work,. al. n. iv n C objective and trading based on all h eavailable h i U these factors and the emotions n g cinformation,. and what they hear or see on the news (Barber & Odean, 2013). Instead of being purely. stemming from them could arguably determine the way traders or investors make decisions, which could then influence asset pricing. This brings us to the concept of attention-based trading which is closely related to behavioral economic theory and the idea that emotional reactions induced by news items could have an impact on decision marking and could have some bearing on financial asset movements (Nisar & Yeung, 2018).. Now, politics and economics are often closely intertwined, and major political events such as regime changes, government policies, and political stability can often have economic repercussions as well (Batista, Maia, & Romero, 2018). Parker (2007, as cited in Nisar & Yeung, 2018) discovered that public mood and trust in the government was a vital part of. 6. DOI:10.6814/NCCU201900722.

(15) investment decision making. Moreover, Niederhoffer (1971, as cited in Nisar & Yeung, 2018) found that world news had a ‘discernible influence’ on stock market movements, specifically the S&P 500 index. Similarly, Chan and Wei (1996, as cited in Nisar & Yeung, 2018) noted that favorable political news in Hong Kong had a positive impact on stock returns and vice versa, although this effect was only observable among blue-chip stocks. On the other hand, the news may have no bearing on market movement at all. In contrast with Niederhoffer’s (1971, as cited in Nisar & Yeung, 2018) findings, Cutler, Poterba, and Summers (1989, as cited in Nisar & Yeung, 2018) found no significant links between political news and movements in the S&P 500 index. Zach (2003, as cited in Nisar & Yeung) has also argued that there is limited evidence that the news has a direct impact on market performance.. 立. 政治大. As this thesis also covers foreign exchange markets, it would be worthwhile to discuss the. ‧ 國. 學. impact of political news on exchange rate movements as well. A country’s currency may reflect the health of the economy, therefore political events could have an impact on. ‧. exchange rate movements. Moreover, currency movements following a political event may be caused by human expectations as investors and traders react to these events and adjust. y. Nat. sit. their decisions accordingly (Mohammed, 2017). Major events in the past have had some. al. er. io. effect on currencies. For instance, Brexit in 2016 caused the British pound to fall to a 31-. v. n. year low (Chu, 2016), while the U.S. dollar hit an almost 11-month high after U.S.. Ch. i n U. President Trump’s electoral win (Reuters, 2016). Hardouvelis (1988) explored how. engchi. macroeconomic news in the U.S. affected exchange rates, particularly the U.S. dollar, and found that monetary news — bank reserves announcements, Fed’s discount and surcharge rates — had significant effects on interest rates and exchange rates. Similarly, Conrad and Lamla (2010, as cited in Uhl, 2017) found that monetary policy news had an impact on future price developments of the EUR/USD currency pair. Ehrmann and Fratzscher (2005, as cited in Uhl, 2017) discovered that economic news in the U.S. and in the Euro area had affected daily EUR/USD movement.. Furthermore, Galati and Ho (2003) found that macroeconomic news affected EUR/USD exchange rate movements and also found that the market tended to be more fixated on negative news as opposed to positive news. Uhl (2017), however, argued that Galati and Ho’s (2003) results may be lacking in a sense that. 7. DOI:10.6814/NCCU201900722.

(16) their interpretation of news had been limited to just macroeconomic news, as opposed to monetary policy outlook, quantitative easing, geopolitical events, and other nonmacroeconomic information. Based on news sentiment data obtained from Thomson Reuters Datastream, Uhl (2017) found no correlation between overall news sentiment and price momentum.. 2.2 Twitter, Tweets, and Influence Launched in 2006, Twitter is a social media platform and an online microblogging service,. 政治大 include images, videos, or links. On Twitter, users are able to send and read tweets for free 立 and can also follow other Twitter users, be it friends, family, celebrities, news outlets,. that allows users to publish and repost tweets, which are 280-character messages that may. ‧ 國. 學. politicians, or other famous personalities. On average, about 6,000 tweets are tweeted every second on Twitter, which comes up to 500 million tweets daily and approximately 200. ‧. billion tweets annually (InternetLiveStats.com, 2019). With about 126 million daily active users in the fourth quarter of 2018, the sheer volume of tweets and the extent of Twitter’s. y. Nat. sit. reach have made it a popular platform for both firms and individuals to communicate and. n. al. er. io. engage with the public (Wagner, 2019).. Ch. i n U. v. With such a wide reach, can tweets become particularly influential and do tweets from. engchi. highly influential individuals have any market-moving potential? For instance, in a Bloomberg report, Vasquez (2018) stated that one tweet from Kylie Jenner — a famous American celebrity — had caused Snapchat shares to sink 6.1% the day after the tweet was posted. Similarly, tweets from Elon Musk (CEO of Tesla) have reportedly affected Tesla stock prices as well. An Ogilvy report by Kornblum (2018) noted that a tweet in 2018: “Am considering taking Tesla private at $420. Funding secured.” (Musk, 2018), caused Tesla shares to increase 11%, while a tweet in 2013: “Really exciting @TeslaMotors announcement coming on Thursday. Am going to put my money where my mouth is in v major way.” (Musk, 2013), caused Tesla shares to rise by 2.4%. In another incident, after the Associated Press’s main Twitter had been hacked and a fake tweet about explosions in. 8. DOI:10.6814/NCCU201900722.

(17) the White House had been posted, the Dow Jones Industrial Average (DJIA) momentarily dropped 150 points before stabilizing (Fisher, 2013).. From electoral campaigns to revolutions, Twitter has inevitably reshaped the political sphere. A large user base, as well as the ease of use and access, make it possible for tweets to go viral online, making it an ideal medium for politicians looking to promote their policies and campaigns. It has also provided greater opportunities for two-way dialogue between the media, political leaders, and the public (Alonso-Muñoz, Marcos-García, and Casero-Ripollés, 2016). Additionally, Lopez-Meri (2016, as cited in Alonso-Muñoz et al., 2016) noted that Twitter has also become a means for political actors to “influence media. 政治大. coverage and avoid the filter of journalists”.. 立. Twitter played a key role in the 2016 U.S. presidential election. From the beginning of the. ‧ 國. 學. primary debates in August 2015 to the end of the elections in November 2016, people in the U.S. sent about 1 billion tweets pertaining to the elections (Coyne, 2016). During the. ‧. election period, the world witnessed candidates on both sides of the political spectrum engaging with their opponents as well as their constituents. Aside from promoting their. y. Nat. sit. own campaigns, presidential candidates also used Twitter to fire back at their opponents as. al. er. io. well. On June 9, 2016, Trump tweeted: “Obama just endorsed Crooked Hillary. He wants. n. four more years of Obama — but nobody else does!”, to which Hillary shot back, reposting. Ch. i n U. v. the tweet and commenting: “Delete your account” (Trump, 2016; Clinton, 2016). Trump’s. engchi. tweet received 74,487 likes and was retweeted 31,289 times, while Clinton’s tweet was liked 681,000 times and retweeted 520,934 times; these figures highlight a high degree of reach and engagement with other users on Twitter (Trump, 2016; Clinton, 2016). After the elections, Twitter continues to be a prominent part of politics and a central part of Trump’s presidency to a point where the White House Press Secretary has even stated that Trump’s tweets should be seen as official statements, despite what other White House officials have said about the official nature of said tweets (Vitali, 2017).. 9. DOI:10.6814/NCCU201900722.

(18) 2.3 Connecting the Dots between Tweets and Financial Markets How then do tweets affect financial markets? To answer this question, existing research has mostly utilized sentiment analysis-based methodologies in order to study links between the content of tweets and financial market movements. Some papers have opted to focus on general public sentiment, as opposed to more topic-specific or user-specific tweets. Bollen, Mao, & Zeng (2011) utilized Granger causality analysis and neural networks and found that public sentiment on Twitter had some capability to predict stock market movements in the DJIA. However, they were not able to find any causative mechanisms. 政治大. that connect stock market index values with public sentiment (Bollen et al., 2011). Moreover, the dataset used in Bollen et al.’s (2011) work had focused on public tweets. 立. with explicit indications of mood that matched expressions such as: “I feel”, “I am feeling”,. ‧ 國. 學. “I don’t feel”, “I’m”, “I am”, and so on. Using a similar method, Mittal and Goel (2011) likewise found that public Twitter sentiment had some capacity to predict DJIA movement.. ‧. In contrast, Ranco et al. (2015) used both tweet volume and sentiment data from public tweets to test for causality between Twitter sentiment on stock price returns and found that. sit. y. Nat. although sentiment — in particular, polarity in sentiment — was not a good predictor of. io. volatility.. er. stock returns, the number of tweets for a company had proved useful for predicting price. al. n. iv n C Researchers may also opt to focus h eonn topic-specific g c h i U tweets by looking for tweets that possess certain keywords. Ozturk and Ciftci (2014) specifically collected tweets with terms. or hashtags pertaining to the US dollar (USD) and Turkish lira (TRY) currency pair, such as “USD/TRY”, “#USD/TRY”, “Dollar”, “#Dollar”. After classifying tweets based on sentiment and using logit regressions, they found that Twitter sentiment had a significant effect on USD/TRY (Ozturk & Ciftci, 2014). A similar study conducted by Pagolu, et al. (2016) collected tweets and filtered them using keywords pertaining to publicly traded stocks in the U.S. (i.e. Microsoft, MSFT, Windows, etc.). Sentiment analysis conducted on these tweets showed a strong correlation between stock price movement and public opinions or emotions regarding a specific publicly traded company (Pagolu et al., 2016). On the other hand, Nisar and Yeung (2018) analyzed political sentiment from Brexit related 10. DOI:10.6814/NCCU201900722.

(19) tweets and hashtags in 2016 and found that there was some evidence of causation between public sentiment and stock market movement in the Financial Times Stock Exchange (FTSE) 100 index, yet this relationship was not deemed statistically significant.. Other works, such as this thesis, have focused on user-specific tweets -- often tweets from high-profile, highly influential users. Jermann (2017) carried out sentiment analysis and utilized logistic regressions and Naive Bayes models using tweets from company executives to examine if these tweets were capable of predicting stock market movement. The results showed that while these tweets had some predictive capacity, only about 52.0% classification accuracy was achieved (Jermann, 2017). In general, academic research. 政治大 Instead, most claims appear to be mostly anecdotal, such as through media headlines. For 立 pertaining to a special individual’s tweet’s effects on financial markets are quite scarce.. example, in an Ogilvy report, Kornblum (2018) cited tweets written by celebrities, CEOs,. ‧ 國. 學. and even news agencies, that had purportedly led to increases or declines in certain company stock prices. These headlines highlight possible causal links between tweets. ‧. posted by influential individuals and stock prices, however, this has not been thoroughly examined by academic researchers and remains open to discussion. This is the gap in the. y. Nat. sit. literature that this thesis aims to examine. Though this thesis is not built sentimental. al. er. io. analysis as its core methodology, it follows the same idea that tweets induce emotional. n. responses which may or may not affect financial markets.. Ch. engchi. i n U. v. 2.4 President Trump’s Tweets and Financial Markets There has been growing interest in studying the causal relationship between Trump’s tweets and financial markets. Trump is known for his candor in public, both offline and online. From disparaging statements about minorities to discrediting former President Barack Obama’s birth certificate, Trump has been notorious for essentially speaking without a filter. Moreover, Trump has not shied away from openly dropping names in his tweets, ranging from political opponents (Hillary Clinton, Joe Biden, and Nancy Pelosi, among others), other sovereign states (Turkey, Mexico, and Canada, for instance), and. 11. DOI:10.6814/NCCU201900722.

(20) publicly traded corporations (Amazon, Boeing, Lockheed Martin, and so forth). And while Trump has been active on Twitter since even before his official declaration of candidacy in 2015, the role of Twitter in both Trump’s presidential campaign and his current presidency has sparked renewed interest in his Twitter activity. Based on tweets collected from the Trump Twitter Archive (Brown, n.d.), Trump’s first tweets were created in May 2009, and as Carr (2018) has noted, Trump’s initial use of Twitter was primarily for marketing himself and his businesses. Since creating his account, Trump has amassed a total of 42,000 tweets and a follower base of over 61 million users (Trump, n.d.). Gallup research found that while only 8% of Americans follow Trump directly on Twitter, 76% of Americans have seen, heard, or read about Trump’s tweets (Newport, 2018). This entails. 政治大 other news outlets, as well as other social media websites or applications. 立. that a majority of Americans learn about Trump’s tweets from secondary sources such as. ‧ 國. 學. Only a small percentage of people follow Trump directly on Twitter, and yet, why are people fixated on Trump’s tweets? For one, at the time of writing this, Trump is the current. ‧. president of the U.S., which makes him a highly influential and powerful individual with a significant amount of sway. Thus, Trump’s tweets can be regarded as public information. y. Nat. sit. which may affect asset prices. And indeed, many media headlines appear to confirm this.. al. er. io. Thielman (2016) noted that a Trump tweet from 2016 scolding military jet manufacturer. n. Lockheed Martin for its “out of control” costs caused the company’s stock price to fall. Ch. i n U. v. from US$259.54 on Friday to US$246 on Monday, the day of the tweet. Similarly, Trump’s. engchi. tweet about supporting a national boycott against American motorcycle maker, HarleyDavidson, saw shares falling by 4% within the week (Farley, 2018). Trump’s tweets have also purportedly affected exchange rates. In 2018, Trump tweeted that there was “no political necessity to keep Canada in the new NAFTA deal” (Trump, 2018b), which caused the U.S. dollar (USD) to rise to 1.3192 against the Canadian dollar (CAD) a few days after the tweet (Imbert, 2018). Interestingly, more recent news items have stated that interest in Trump’s tweets have softened, as interaction on his tweets has declined even though Trump’s tweeting frequency has increased (Evans, 2019). Both investors and companies may have also started to ignore Trump’s tweets. Cancel (2018) noted that Mexican peso traders have also begun to move on from Trump’s tweets pertaining to the border wall. 12. DOI:10.6814/NCCU201900722.

(21) between the U.S. and Mexico and the North American Free Trade Agreement (NAFTA). And even after being on the receiving end, General Motors did nothing to respond to Trump’s angry tweets, opting instead to release a statement regarding in the relocation of workers (Rappeport, 2019).. While these media headlines suggest that these tweets do indeed affect stock prices and exchange rates, the causal effect of Trump’s tweets continues to be widely debated. Trump’s opinions are shaped by his economic views on economic nationalism, which as Colonescu (2018) noted, highlights mercantilism and protectionism in international trade and relations. It comes as no surprise then that this view coincides with Trump’s tagline:. 政治大 and has talked disfavorably of American companies looking to move production offshore. 立. “Make America Great Again”. Indeed, on Twitter, Trump has threatened to impose tariffs Colonescu (2018) analyzed Trump’s tweets from January 2017 until May 10, 2018, to. ‧ 國. 學. analyze the relationship between Trump’s tweet sentiment and daily DJIA returns. In addition, Colonescu (2018) also examined the possible effect of Trump’s tweets on. ‧. exchange rates, particularly the USD/CAD and USD/EUR exchange rate pairs. In the stock market portion, Colonescu (2018) found that Trump’s tweets had statistically significant. y. Nat. sit. short-term effects on DJIA returns. On the other hand, exchange rate effects differed. al. er. io. between the two pairs, such that while there were some significant short-term effects on. v. n. USD/CAD, there was no evidence of any such effect on the USD/EUR pair (Colonescu,. Ch. i n U. 2018). On the contrary, Born et al. (2017) found that Trump’s tweets had transitory price. engchi. impacts and also led to an unexpected increase in trading volume. Moreover, Ge, Kurov, and Wolfe (2019) found that Trump’s company-specific tweets affected stock prices, trading volume, volatility, and institutional investor attention. In general, studies that have found a statistically significant relationship between Trump’s tweets and financial markets have also noticed that these effects were often short-term in duration, with the effect petering out over the course of a few days. However, Juma’h and Alnsour (2018) found that Trump’s tweets had no significant effect on stock market indices or a targeted company’s share prices. It is possible that changes in stock prices or exchange rates may not be specifically caused by Trump tweets, but instead, but other coexisting financial, economic, or political events. In addition, Juma’h and Alnsour’s (2018) results may also. 13. DOI:10.6814/NCCU201900722.

(22) indicate that the effect of Trump’s tweets may be too short-lived to have any significant impact.. 立. 政治大. ‧. ‧ 國. 學. n. er. io. sit. y. Nat. al. Ch. engchi. 14. i n U. v. DOI:10.6814/NCCU201900722.

(23) III. METHODOLOGY The main objective of this thesis is to examine possible causal links between Trump’s tweets and movement in financial markets, namely in the equity market (stock prices) and the foreign exchange market (exchange rates). More specifically, we assess the effects of Trump’s tweets pertaining to trade and business which are more negative in tone, as opposed to more neutral-toned or positive-toned tweets. In doing so, we propose the use of the synthetic control (SC) method, which is a type of statistical method commonly used for comparative case studies. This chapter expounds on the data collection process for this. 政治大. thesis, the theoretical background of the SC method, its advantages and disadvantages, and the choice of variables. For this thesis, the Python programming language was used for. 立. data collection and processing, while Stata was used for the SC models and placebo effect. ‧ 國. 學. graphs.. 3.1 Data Collection. ‧. io. sit. y. Nat. 3.1.1 Tweets. er. Tweets were collected from the Trump Twitter Archive (Brown, n.d.), which is a collection. al. n. iv n C a select few tweets, as opposed h to e analyzing a larger n g c h i Uselection or all of Trump’s tweets,. of Trump’s tweets from as early as 2009 to his most recent tweets. This thesis focused on. which is what is commonly done in most sentiment analysis-based studies. As noted in the previous chapter, it is suggested — often by major news outlets — that Trump’s Twitter tirades may have an impact on financial markets, particularly the stock market and the foreign exchange market. To explore possible cause-and-effect links, we selected certain tweets, which contained statements that were explicitly critical and directed at a certain company (Amazon, Boeing, and Harley-Davidson) or country (Canada, the European Union (EU), and Mexico).. The selected tweets are displayed in the tables below (Tables 1 and 2), showing the date tweeted, tweet content, the number of retweets, and the target of said tweet. The inclusion 15. DOI:10.6814/NCCU201900722.

(24) of the number of retweets should give readers an idea of the extent of a tweet’s reach or engagement. Table 1 shows the tweets chosen to analyze the treatment effect on certain companies, while Table 2 shows the tweets chosen to analyze the effect on certain countries. The ‘Target’ columns also include the ticker symbols for companies or currency code for foreign currencies.. Table 1: Trump tweets targeting publicly traded companies in the U.S. Date. Target. December. Boeing. 6, 2016. Number of Tweet Content Retweets 37,728. (BA). control, more than $4 billion. Cancel order!”. 24,604. (AMZN). losing many billions of dollars a year, while charging Amazon and others so little to deliver. Nat. sit. y. their packages, making Amazon richer and the Post Office dumber and poorer? Should be charging MUCH MORE!”. n. al. er. io August 12, 2018. “Why is the United States Post Office, which is. ‧. 29, 2017. Amazon. 政 One治 for future presidents, but costs are out of 大. 學. December. ‧ 國. 立. “Boeing is building a brand new 747 Air Force. Harley-. Ch 16,701. engchi. i n U. v. “Many @harleydavidson owners plan to boycott. Davidson. the company if manufacturing moves overseas.. (HOG). Great! Most other companies are coming in our direction, including Harley competitors. A really bad move! U.S. will soon have a level playing field, or better”. 16. DOI:10.6814/NCCU201900722.

(25) Table 2: Trump tweets targeting other countries Date. Target. September. Canada. 1, 2018. (CAD). September. Canada. 1, 2018*. (CAD). Number of Tweet Content Retweets 15,687. “I love Canada, but they’ve taken advantage of our Country for many years!”. 23,515. “There is no political necessity to keep Canada in the new NAFTA deal. If we don’t make a fair deal for the U.S. after decade of abuse, Canada will be out. Congress should not interfere with these. 立. 17,832. ‧ 國. “Based on the Tariffs and Trade Barriers long. Union. placed on the U.S. & its great companies and. (EUR). workers by the European Union, if these Tariffs. ‧. and Barriers are not soon broken down and. io. of their cars coming into the U.S. Build them. al. here!”. n. May 30,. Mexico. 2019. (MXN). sit. y. Nat. removed, we will be placing a 20% Tariff on all. Ch. 39,267. er. 2018. European. entirely & we will be far better off..”. 學. June 22,. 治 or I will simply terminate NAFTA 政 negotiations 大. i n U. v. “On June 10th, the United States will impose a 5%. i goods coming into our Country from e nTariff g conh all Mexico, until such time as illegal migrants coming through Mexico, and into our Country, STOP. The Tariff will gradually increase until the Illegal Immigration problem is remedied,..”. May 30,. Mexico. 2019. (MXN). 19,809. “....at which time the Tariffs will be removed. Details from the White House to follow.”. *Note: This tweet was tweeted twice on the same day, but the first instance of the tweet was deleted. 17. DOI:10.6814/NCCU201900722.

(26) 3.1.2 Treatment Companies and Countries This thesis can be divided into two parts of focus: stock prices and exchange rates. Different data sources were utilized and data for both treatment and control companies and countries. Three affected companies were chosen: Harley-Davidson, Amazon, and Boeing. Their ticker symbols are as follows, HOG, AMZN, and BA, respectively. Although Trump has also targeted foreign companies in his tweets, the scope of this thesis was limited to studying the effects of Trump’s tweets on American companies, thereby restricting the choice of control companies to publicly traded American corporations. For each treatment company, we collected around twenty control companies in similar industries and sectors,. 治政大 in the U.S. The full list of these were chosen, all of which are publicly traded companies 立 companies can be found in Appendix 1.. and then pooled them together to create a larger donor pool. A total of 74 control companies. ‧ 國. 學. As for the exchange rate part of this thesis, three exchange rate pairs were chosen as. ‧. treatment groups, namely: the Canadian dollar (CAD), the Euro (EUR), and the Mexican peso (MXN). Canada, the European Union, and Mexico are among America’s top trading. Nat. sit. y. partners (Schwarzenberg, 2018). In addition, along with the U.S., Canada and Mexico are. io. er. also members of the NAFTA trading bloc. Geographically, the two countries border the U.S. on the north (Canada) and the south (Mexico). Trump continues to be highly critical. n. al. i n U. v. of illegal immigration stemming from south of the U.S. border. Since even before officially. Ch. engchi. announcing his candidacy for president in June 2015, Trump had already tweeted about building a wall around the U.S.-Mexico border (Trump, 2014). In addition, Trump has also accused both Canada and the EU of taking advantage of the U.S., usually within the scope of trade relations. A donor pool of control countries was also collected and spans a total of 92 countries — the full list can be found in Appendix 2. A wide variety of countries was selected, although some smaller economies such as smaller Pacific island states and states in active conflict were left out, mostly due to the lack of either currency data or covariate data.. 18. DOI:10.6814/NCCU201900722.

(27) 3.1.3 Dependent and Independent Variables To analyze the effects on equity markets, the daily closing price was set as the key dependent variable. A number of covariates were also chosen, as follows: market capitalization, price-to-earnings (P/E) ratio, gross profit, total revenue, book value, and dividend rate. For both dependent and covariate variables, data was collected using the Python module, yahoofinancials, which allows users to pull data from the now-defunct Yahoo Finance application programming interface (API) (Sanders, 2019). Closing stock price data was collected on a daily frequency, while covariate data is time-invariant and represents only the most recent data for said variables.. 政治大 As for the foreign exchange 立 market part of this study, exchange rates were used as the dependent variable. Using a simple Python program, daily closing exchange rates were. ‧ 國. 學. programmatically collected from the currencylayer API (apilayer, 2019). Exchange rate pairs used in this study have the U.S. dollar (USD) set as the base currency, therefore the. ‧. format should be as follows: USD/country currency. For covariates, the following variables. y. Nat. were used: consumer price index (CPI), population, Gross Domestic Capital (GDP) per. sit. capita (purchasing power parity or PPP) in current international dollars, and current. al. er. io. account in current international dollars. For these variables, data was once again. v. n. programmatically collected, but this time from the World Bank API through the wbdata. Ch. i n U. module for Python (Sherouse, 2014). For some missing data, mostly for current account. engchi. values, data was collected from the American Central Intelligence Agency’s (CIA) World Factbook (CIA, 2019).. 3.2 Theoretical Background As noted earlier, the main objective of this research is to study the causal relationship between Trump’s tweets directed at specific countries or companies, and movements in stock prices and exchange rates. This differs from some of the existing literature which tend to be more focused on studying the predictive power of Trump’s tweets, and thus, are. 19. DOI:10.6814/NCCU201900722.

(28) built on more sentiment analysis-focused, machine-related methodology. Perhaps one downside to these studies is that they may sometimes gloss over ascertaining the actual cause-and-effect relationship that these tweets have with financial markets. On the other hand, by using the SC method, this allows for the estimation of causal relationships in a panel data setting. For this thesis, this statistical procedure was performed using the synth and synth_runner packages for Stata (Abadie et al., 2011; Galiani & Quistorff, 2017).. How does one estimate the effect of a certain intervention, policy, or law? In this case, how does one estimate the effect of a tweet on stock prices or exchange rates? To answer these questions, researchers have utilized comparative case studies to develop causal. 政治大 the event to one or more unexposed units (Abadie et al., 2010). However, a major pitfall 立 explanations for these events. Doing so involves comparing one or more units exposed to. of this method is that even if aggregate data were to be used, there continues to be some. ‧ 國. 學. uncertainty about producing a suitable counterfactual outcome that illustrates how the affected group would have developed or changed in the absence of the intervention or event. ‧. (Abadie et al., 2010). Thus, researchers are left to speculate about what could have happened had the event never occurred.. sit. y. Nat. al. er. io. The synthetic control method aims to address this issue by constructing a weighted. n. combination of control units to model the counterfactual -- the “synthetic control”, so to. Ch. i n U. v. speak (Cunningham, 2018). Doing so allows researchers to estimate treatment effects by. engchi. comparing treated unit outcomes with control unit outcomes. In a traditional comparative case study, the selection of comparison units may lead to erroneous conclusions, especially if the comparison units are not sufficiently similar to the treatment units (Abadie et al., 2015). As a solution to this issue, the SC method offers a more systematic way to choose comparison units by assigning varying weights to different control units. Abadie et al. (2015) have likewise argued that a set of control units provides a better representation of the treatment unit as opposed to using a single comparison unit alone. Through the SC method, the counterfactual is “selected as the weighted average of all potential comparison units that best resembles the characteristics of the case of interest” (Abadie et al., 2015).. 20. DOI:10.6814/NCCU201900722.

(29) First, let. be the outcome observed for unit i, where i = 1, …, J + 1, and at time period. t, where t = 1, …, T, in the absence of the intervention. The interaction or treatment occurs at period. where. , with unit. being affected by the. intervention, while units. remain unaffected by the intervention (Yang,. 2019). Knowing this, let. be the observed outcome for region i at time t if unit i received. the treatment, and let. represent the observed outcome for unit i at time t if unit i had. never received the treatment (Yang, 2019).. Before the event, assume that the intervention has no effect on the outcome before the treatment period, thus, for periods. and all units , then 治政let (Abadie et al., 2010). Next, be the unit ( ) that receives treatment 大 from periods until be units that do not receive treatment. From 立 , and let. ‧ 國. 學. this, we can derive the effect of the intervention for unit i at time t, as follows. Let the observed outcome. be. . Before intervention, the observed. outcomes are. . Therefore, after the intervention (after period. ‧. observed outcome can be written as. y. sit er. io. (1). . Given this, this can be further. Nat. simplified to:. al. iv n C is the causalheffect i U on unit i at time t, or simply, the e n ofg the c htreatment n. where. ), the. difference between the counterfactual and the actual trend of the treated unit (Yang, 2019;. Bouttell, Craig, Lewsey, Robinson, & Popham, 2018). And since only unit 1 (i=1) received treatment, then we estimate the causal effect for this unit over periods therefore,. where for. represents the observed outcome and. , then. , , where. represents the counterfactual (Yang, 2019).. Based on Abadie et al. (2010), the SC method in its simplest form be written as follows:. (2). 21. DOI:10.6814/NCCU201900722.

(30) Where. represents time effects, while. are not affected by the intervention,. is a (r x 1) vector of observed covariates that. is a (1 x r) vector of unknown parameters,. x F) vector of unobserved common factors, variables, while. is a (1. is an (F x 1) vector of permanent unobserved. represents the unobserved transitory shocks at the unit level with zero. mean.. To implement SCM, we choose a vector of positive weights that sum to one. Let this be: , wherein each value of W represents a weighted average of the available control units, thus representing a synthetic control (Yang, 2019; Abadie et al., 2010). We want to choose w* such that the following conditions are met so that the treatment effect (. 政治大. ) is unbiased.. 立. y. Nat. is unobserved, we must choose w* that satisfies:. io. sit. But because. n. al. er. (4). ‧. ‧ 國. 學. (3). (5). Ch. Thus, the unbiased estimator,. engchi. i n U. v. , should be (Yang, 2019):. (6). Fitting. and. is sufficient to match. pre-intervention outcomes,. so long as the synthetic controls can fit. and. (Abadie et al., 2010).. Ideally, the control group’s pre-treatment path must be parallel to that of the treated unit’s. This allows for a better understanding of the treatment’s effect on the affected unit. That is, if the synthetic and treated paths diverge, then it can be said that the treatment does. 22. DOI:10.6814/NCCU201900722.

(31) indeed have an effect and vice versa. But does divergence always point to a statistically significant effect? To assess the significance of the estimates, Abadie et al. (2010) suggest performing a series of placebo studies by iteratively applying the SC method to the control units. This process then returns a set of root mean squared prediction error (RMSPE) values which are calculated for both pre- and post-treatment periods (Cunningham, 2018). When the pre-RMSPE values are too large, this could skew the placebo effects and make p-values too conservative. It is thus suggested that these observations be omitted (Cunningham, 2018). In the model used for this thesis, we have chosen to skip the omission of values as the dependent variables have already been standardized.. 政治大 limitations as well. First, the key identifying assumption is somewhat ambiguous (Yang, 立. Although the SC method certainly has its benefits, it is important to be mindful of its. 2019). Moreover, the method may be too restrictive for some cases, as only the treatment. ‧ 國. 學. unit can be affected by the intervention, and it is assumed that there are no spillovers in the donor pool.. ‧. 3.3 Procedure. y. Nat. sit. We first start with choosing appropriate variables for the dependent variables and. er. io. covariates, for both the stock price and exchange rate models. Again, the main objective of. al. iv n C and stock prices, as well as exchange We choseU h erates. i treatment events which are individual h n c g tweets published by Donald Trump wherein he has threatened a corporation or country. n. this paper is to examine causal relationships between Trump’s more negative-toned tweets. Because of the nature of the synthetic control model, we needed to ensure that the tweets were spaced out enough so that treatment effects were not present both within the 20trading day pre-treatment period and the 5-trading day post-treatment period. Note that days are measured as transaction or trading days and exclude weekends and some national holidays in the U.S. — New Year’s and Christmas, for example. In short, for both periods, we needed to ensure that there were no negative trade or business related tweets, such as tweets threatening to impose tariffs or boycott companies, for instance. This ensures that the before-and-after distinction is maintained, as suggested by Dube and Zipperer (2015). As a result, however, this has somewhat restricted the scope of this thesis to just a handful. 23. DOI:10.6814/NCCU201900722.

(32) of events. After the tweets are collected, to avoid contamination in the donor pool for each event, control countries or companies that Trump tweeted about in either the post-treatment and pre-treatment periods of each tweet event are omitted.. Data for both the dependent and independent variables for the stock price-based and exchange rate-based SC method models were then collected. We first collected two years’ worth of data in order to standardize the dependent variable data, as values for both stock prices and exchange rates varied considerably across corporations and countries. For instance, certain currency pairs such as U.S. dollar/Indonesian rupiah (USD/IDR) or U.S. dollar/Laotian kip (LAK) can easily breach over 1,000 in value, whereas other currency. 政治大 often fall below 1.0 in value (apilayer, 2019). Likewise, for stock prices, for the sake of 立. pairs such as U.S. dollar/Euro (USD/EUR) or U.S. dollar/British pound (USD/GBP) may comparison, on June 6, 2019, Amazon’s closing stock price stood at USD1,734.56,. ‧ 國. 學. whereas Harley-Davidson’s stock price was a mere USD34.16 (NASDAQ, 2019a; NASDAQ, 2019b). Thus, because the range varies considerably it is important to first. ‧. standardize the dependent variable data. To do so, first, both the standard deviation and mean were calculated per country or company. After which, standardized closing stock. y. Nat. sit. prices and standardized closing exchange rates were calculated by subtracting the mean. al. er. io. stock price or exchange rate from the original stock price or exchange rate and then. v. n. dividing by the value of the standard deviation for stock prices or exchange rates. A. Ch. i n U. simplified version of this formula (Formula 7) can be seen below, where A represents either. engchi. closing stock price or closing exchange rates, μ𝐴 represents the mean of A and σ𝐴 represents the standard deviation of A.. (7). Aside from the selected covariates, another potentially important predictor are the lagged outcome variables (McClelland & Gault, 2017). As Botosaru and Ferman (2017) have noted, there is “little guidance” on which variables are to be used as covariates. In selecting covariates for this model, we chose predictors that would have some bearing on the. 24. DOI:10.6814/NCCU201900722.

(33) predicted outcome. Ultimately, these will determine the selection of donor units and weights (McClelland & Gault, 2017). In addition to these predictors, the pre-intervention lagged outcomes were also included in the models. In doing so, matching on preintervention lagged outcomes may help control for unobserved factors and the heterogeneity of effects on unobserved and observed factors (Abadie et al., 2015).. Two models were produced for each event, with the first model (Model 1) containing all of the outcome lags and all of the chosen covariates, while the second model (Model 2) kept all of the outcome lags and some covariates. The choice of covariates for the second model was based on the difference in synthetic and treated predictor values.. 政治大 Using Stata, we first used the synth command to identify weights for donor pool units and 立. to calculate the pre-RMSPE value and covariate balance. Ideally, the optimal set of weights. ‧ 國. 學. should produce a nearly identical pre-treatment trend for the synthetic unit (Cunningham, 2018). In order to see this in action, we also produce graphs to see the goodness-of-fit of. ‧. the pre-treatment synthetic treatment unit with the actual treatment unit data. Ideally, the pre-treatment synthetic treatment series should be similar to the actual treatment series, and. y. Nat. sit. if the intervention does have an effect, then we should expect these two series to diverge. al. er. io. after the event. The synth command also produces a table with the covariate balance values. v. n. for both the actual treated unit and the synthetic treated unit. As Cunningham (2018) has. Ch. i n U. stated, this table is not a technical test, however, if the predictors are more or less balanced,. engchi. then we would expect the synthetic unit to be a suitable approximation of the real treated unit assuming the event had never occurred in the first place (Abadie et al., 2010). However, predictors, arguably, do not have to be perfectly balanced (Botosaru & Ferman, 2019).. Usually, the SC process requires researchers to drop observations with pre-RMSPE values that are deemed to be too large — this scales depending on the dependent variable, for instance, one can opt to drop observations with pre-RMSPE values that are twice as large as the treatment units’. However, since the dependent variables have already been standardized, we posit that this already takes matching quality into account, making it unnecessary to drop observations at this point. In order to obtain the models p-values, we. 25. DOI:10.6814/NCCU201900722.

(34) utilized synth_runner’s pvals1s option which produces a table of standardized two-sided p-values to show whether the treatment effects were significant on the day of treatment up to five days after treatment. Additional commands were then utilized as well to create placebo graphs.. 立. 政治大. ‧. ‧ 國. 學. n. er. io. sit. y. Nat. al. Ch. engchi. 26. i n U. v. DOI:10.6814/NCCU201900722.

(35) IV. RESULTS AND DISCUSSION To determine the extent of the causal link between the selected Trump tweets and financial markets, SC method-based models were produced for each selected tweet and the results are presented here. To further organize the results, this chapter has been split into two parts, with one section discussing the results for the stock price models, and another section detailing the results for the exchange rate models. Each section is further divided into smaller subsections for each tweet, with each subsection containing the SC graphs and placebo effect graphs for each event. At the beginning of each section, a summary table is. 政治大. provided. For the sake of organization and structure, predictor balance and control unit weight tables can all be found in the Appendix section.. 立. ‧ 國. 學. 4.1 Stock Price Models. ‧. A total of three tweets were selected, covering three companies: Boeing (BA), Amazon (AMZN), and Harley-Davidson (HOG). The summary table is provided below, while the. Nat. sit. y. SC graphs for each tweet follows. The columns show the estimates and the standardized p-. io. er. values for the day of the tweet and days after the tweet. Columns with a (1) header show estimates for models that utilize all the outcome lags as well as all the covariates, whereas. al. n. iv n C U total revenue. Figures shown in the h e nbook some covariates, namely: gross profit, h i and g cvalue, columns with a (2) header show estimates for models that utilize all the outcome lags and. table show that estimates and p-values for both Model 1 and Model 2 to be identical, indicating that changing the choice in covariates makes little if no difference at all. This indicates that the estimates are generally robust.. 27. DOI:10.6814/NCCU201900722.

(36) Table 3: Summary of Synthetic Control Models for Stock Prices. Time Day of tweet 1 day after. 立. Amazon December 29, 2017 (1) (2) -0.030 -0.030 (0.000) (0.000) -0.000 -0.000 (1.000) (1.000) 0.028 0.028 (0.800) (0.800) 0.027 0.027 (1.000) (1.000) 0.044 0.044 (1.000) (1.000) 0.061 0.061 (1.000) (1.000). Harley-Davidson August 12, 2018 (1) (2) -0.414 -0.414 (0.154) (0.154) -0.498 -0.498 (0.246) (0.246) -0.428 -0.428 (0.338) (0.338) -0.705 -0.705 (0.246) (0.246) -0.494 -0.494 (0.492) (0.492) -0.4946 -0.4946 (0.492) (0.492). 政治大. 學. ‧ 國. 2 days after 3 days after 4 days after 5 days after. Boeing December 6, 2016 (1) (2) -0.011 -0.011 (0.667) (0.667) -0.087 -0.087 (0.333) (0.333) -0.037 -0.037 (0.667) (0.667) -0.019 -0.019 (1.000) (1.000) 0.030 0.030 (1.000) (1.000) -0.002 -0.002 (1.000) (1.000). On the day of publishing, Trump’s tweets towards Boeing, Amazon, and Harley-Davidson. ‧. appeared to have a negative effect on these stocks’ stock prices. For Boeing stock prices, the effect remained mostly negative, albeit insignificant as the days went by. Similarly,. sit. y. Nat. Trump’s tweet directed at Harley-Davidson maintained a negative impact as long as five days after tweeting. Whereas for Amazon, this negative effect dissipates 2 days onwards.. io. er. Overall, Trump’s tweets about Amazon are the only tweets shown to have a statistically. al. n. iv n C h e n gsignificant tweet’s effects are only really statistically c h i Uon the day of publishing. However,. significant impact on stock prices — Amazon’s stock prices, in particular. Although the this is also likely due to the fact that Trump had tweeted about Amazon on the 29 th of December, the Friday right before the New Year’s trading holiday.. From Table 1, based on the number of retweets, we see that engagement with these tweets varies considerably between the selected tweets. Boeing received the most number of retweets at about 37,728, followed by Amazon (26,604), and Harley-Davidson (16,701). It would appear that engagement may not necessarily be a determinant of whether or not Trump’s tweet may have statistically significant effects on a company’s stock price. Worth noting, perhaps, is not just engagement but the frequency at which Trump tweets about these companies. A quick search on the Trump Twitter Archive (Brown, n.d.) using the 28. DOI:10.6814/NCCU201900722.

(37) keyword “Boeing” shows that the tweet we selected for this study was the first Boeingrelated tweet that Trump had posted since announcing his campaign in June 2015. Similarly, Trump had only started tweeting about Harley-Davidson in 2018, with the tweet selected for this study being the first among a string of negative tweets directed towards the company (Brown, n.d.). In contrast, Trump has tweeted about other companies much more frequently, such as Amazon. Not only has Trump tweeted more frequently about Amazon, but his animosity towards the company and its CEO appears to be more publicized (Kim, 2016; Mack, 2016; Johnson & Brice, 2017). What this could imply is that the frequency at which Trump tweets about a certain company could affect the way traders and investors perceive these company-directed tweets, as a higher frequency of tweets could make. 政治大 this case, Trump tweeted more frequently about Amazon, which in effect, could pull more 立. Trump’s hostility towards a certain company more established or more publicly known. In attention from traders and investors, thus leading to a significant effect on Amazon’s stock. ‧ 國. 學. price.. ‧. 4.1.1 Boeing. y. Nat. For this stock, the tweet chosen shows Trump threatening to cancel an order for Boeing’s. sit. new 747 Air Force One. Boeing continues to be one of America’s largest aerospace. al. er. io. companies and is a leading manufacturer of commercial jetliners and other defense-related. v. n. products. In addition, the U.S. military and the federal government both make up a. Ch. i n U. considerable share of the company’s contracts (Shalal & Tennery, 2016). As such, we. engchi. would expect this tweet to have a negative effect on Boeing’s closing stock price.. 29. DOI:10.6814/NCCU201900722.

(38) Figure 1: Boeing Synthetic Control Graph (Model 1). 學. ‧ 國. 立. 政治大. Figure 2: Boeing Placebo Effects Graph (Model 1). ‧. n. er. io. sit. y. Nat. al. Ch. engchi. 30. i n U. v. DOI:10.6814/NCCU201900722.

(39) Figure 3: Boeing Synthetic Control Graph 1 (Model 2). 學. ‧ 國. 立. 政治大. Figure 4: Boeing Placebo Effects Graph (Model 2) 1. ‧. n. er. io. sit. y. Nat. al. Ch. engchi. i n U. v. Despite omitting some covariates in the second model, Figures 1 and 2 show almost no differences, which entails that the results are robust despite changing the covariates used. We see that the pre-treatment fit is generally quite good, and the synthetic and treatment paths diverge after the event on December 6. However, from Table 3, the standardized pvalue shows this gap to be insignificant at the 5% level. This is further reflected by the placebo effect graphs (Figures 3 and 4) that both show no difference between the donor. 31. DOI:10.6814/NCCU201900722.

(40) pool and treated unit’s post-tweet trajectories, as the gap in closing stock price for the treated unit does not differ significantly from the other control units in the donor pool.. 4.1.2 Amazon Figures 5 and 6, as well as Figures 7 and 8, are virtually identical despite the omission of certain covariates in Model 2. Figures 5 and 6 also show that the pre-treatment fit between the synthetic and treated series are parallel to each other, indicating sufficient goodnessof-fit. Creating a counterfactual for Amazon can be relatively complicated, as the company falls into many categories, ranging from retail to logistics and cloud computing. Primarily,. 政治大 regression is generally satisfactory but can be improved. From Appendix 6, we see that the 立 donor pool is mostly skewed towards Costco, with a weight of 0.285, this is not too however, it can be described as a retail company. The counterfactual created in this. ‧ 國. 學. surprising as, like Amazon, Costco is a major retail player in the U.S. market. Quite surprising, however, is the inclusion of defense companies such as Axon Enterprise (0.133). ‧. and the TransDigm Group (0.165).. y. Nat. n. al. er. io. sit. Figure 5: Amazon Synthetic Control Graph (Model 1). Ch. engchi. 32. i n U. v. DOI:10.6814/NCCU201900722.

(41) Figure 6: Amazon Placebo Effects Graph (Model 1). 學. ‧ 國. 立. 政治大. Figure 7: Amazon Synthetic Control Graph 1(Model 2). ‧. n. er. io. sit. y. Nat. al. Ch. engchi. 33. i n U. v. DOI:10.6814/NCCU201900722.

(42) Figure 8: Amazon Placebo Effects Graph (Model 2) 1. 立. 政治大. ‧ 國. 學. In Figures 5 and 6, we see that the treated unit and control unit series seem to diverge after the tweet. Although the gap between the two series does not seem as pronounced, it was. ‧. found to be statistically significant (see Table 3), which suggests that Trump’s tweet directed at Amazon did, in fact, have a significant negative effect on the company’s stock. sit. y. Nat. price on the day the tweet was published, even though based on the placebo effect graphs (Figures 7 and 8), the gap in closing stock price is not as clear. However, the significant. io. n. al. er. effect peters out just one day after the tweet onwards. Again, this is likely due to the timing. i n U. v. of the tweet as it was published on December 29 ahead of the New Year’s holiday.. Ch. engchi. 34. DOI:10.6814/NCCU201900722.

(43) 4.1.3 Harley-Davidson In the tweet selected for this event, Trump appeared to support a boycott of Wisconsinbased motorcycle manufacturer, Harley-Davidson (see Table 1). Trump’s disdain towards Harley-Davidson aligns with his view of keeping American companies’ production lines within U.S. boundaries.. Figure 9: Harley-Davidson Synthetic Control Graph (Model 1). 立. 政治大. ‧. ‧ 國. 學. n. er. io. sit. y. Nat. al. Ch. engchi. 35. i n U. v. DOI:10.6814/NCCU201900722.

(44) Figure 10: Harley-Davidson Placebo Effects Graph (Model 1). 立. 政治大. ‧ 國. 學. Figure 11: Harley-Davidson Synthetic Control Graph (Model 2)1. ‧. n. er. io. sit. y. Nat. al. Ch. engchi. 36. i n U. v. DOI:10.6814/NCCU201900722.

政治人物，推特與金融市場： 來自川普推特的證據 - 政大學術集成

政治人物，推特與金融市場：來自川普推特的證據 - 政大學術集成