You’re Cancelled: The Power of Online Storms on the Stock Performance of Firms

(1)

1

You’re Cancelled: The Power of Online Storms

on the Stock Performance of Firms

Based on Twitter data from 2015-2020

(2)

2

You’re Cancelled: The Power of Online Storms

on the Stock Performance of Firms

Based on Twitter data from 2015-2020

Master Thesis

Msc. Marketing Intelligence & Marketing Management

Faculty of Economics and Business

University of Groningen

Supervisor: Dr. Evert de Haan

Second Supervisor: MSc. Hidde Smit

Annebeth Seraphine Sterre Meulenberg

(3)

3

Abstract

The rise of the Internet has offered a new platform for consumers to share their knowledge and opinions about certain products or brands with other consumers. This is also known as electronic word of mouth (eWOM), which can spread like a wildfire- fast and abundantly, resulting in social media firestorms. Earlier research on firestorms has found that strong online firestorms relate to negative changes of consumers’ short-term brand perceptions. Besides, firms can mitigate firestorms targeted at them by specific responses over time, ensuring communication and interaction with consumers. This study widens the scope to online storm, including both positive and negative eWOM. In particular, this study will focus on a specific type of online storm created by a new phenomenon called cancel culture. This is known as the fast circulation of information through digital platforms that happens on a large-scale, where certain companies or people are either boycotted or supported. As both topics are rather novel and research has not associated them with financial metrics yet, this study will fill that gap by investigating whether cancel culture can lead to an online storm on Twitter and if this has an effect on the financial performance of firms. To achieve this aim, this study uses panel data consisting of of 73,625 tweets and 977 newspaper articles for 29 firms over a period of 5,5 years (2015-2020). After conducting multiple panel data regressions for

different models, I find that a higher level of virality of an online storm negatively affects the abnormal stock returns of a firm. This effect diminishes two days after the online storm has taken place. Furthermore, whenever an online storm leads to a diversely opiniated discussion, this positively affects the abnormal stock returns of a firm. However, when combined with a higher virality level, these diverse opinions result in a negative effect of virality on abnormal stock returns even four days after the storm occurred. Moreover, a weak negative relationship between articles and abnormal stock returns indicates that traditional media may have become less important, but its effect is still visible. Conclusively, these findings will contribute to extant literature by demonstrating a relationship between online storms and financial performance of firms while using the topical phenomenon of cancel culture.

(4)

4

Preface

Writing this thesis has taught me a few things. I have gained a lot of knowledge on the power of social media and how it can be quantified into the effect on financial performance of firms. Furthermore, to be able to test the hypotheses, I was forced to dive into the field of finance, which was entirely new for me. Therefore, I guess I can say that I have learned quite a bit about the stock market as well as boosting my R and Python skills to be able to scrape and organize all my data.

(5)

5

Abnormal Returns (t+4) and Abnormal Returns (t-1) ... 60 Table 13: Regression results of the interaction between average polarity and virality for Abnormal Returns (t) and Abnormal Returns (t+1) ... 61 Table 14: Regression results of the interaction between average polarity and virality for Abnormal Returns (t+2) and Abnormal Returns (t+3) ... 62 Table 15: Regression results of the interaction between average polarity and virality for Abnormal Returns (t+4) and Abnormal Returns (t-1) ... 63 Table 16: Regression results of the interaction between switching costs and virality for Abnormal Returns (t) and Abnormal Returns (t+1) ... 64 Table 17: Regression results of the interaction between switching costs and virality for Abnormal Returns (t+2) and Abnormal Returns (t+3) ... 65 Table 18: Regression results of the interaction between switching costs and virality for Abnormal Returns (t+4) and Abnormal Returns (t-1) ... 66 Table 19: Regression results of the interaction between article volume and virality for Abnormal Returns (t) and Abnormal Returns (t+1) ... 67 Table 20: Regression results of the interaction between article volume and virality for Abnormal Returns (t+2) and Abnormal Returns (t+3) ... 68 Table 21: Regression results of the interaction between article volume and virality for Abnormal Returns (t+4) and Abnormal Returns (t-1) ... 69 Table 22: Regression results of the interaction between the average polarity of articles and virality for Abnormal Returns (t) and Abnormal Returns (t+1) ... 70 Table 23: Regression results of the interaction between the average polarity of articles and virality for Abnormal Returns (t+2) and Abnormal Returns (t+3)... 71 Table 24: Regression results of the interaction between the average polarity of articles and virality for Abnormal Returns (t+4) and Abnormal Returns (t-1) ... 72 Table 25: Regression results of the interaction between like count and virality for

Abnormal Returns (t) and Abnormal Returns (t+1) ... 73 Table 26: Regression results of the interaction between like count and virality for

Abnormal Returns (t+2) and Abnormal Returns (t+3) ... 74 Table 27: Regression results of the interaction between like count and virality for

(7)

7

1. INTRODUCTION

"Twitter is not a technology. It's a conversation. And it's happening with or without you." ― Charlene Li, author, and analyst at Altimeter

Consumers are not independent decision makers. They are constantly influenced by advertisements, friends and family or other consumers (Payne, Bettman & Johnson, 1991). Traditional word-of-mouth has proved to play a major role in a consumer’s buying behaviour (Richins & Root-Shaffer, 1988). The rise of the Internet has offered a new platform for consumers to share their knowledge and opinions about certain products or brands with other consumers (Hansen et al., 2018). This is also known as electronic word of mouth (eWOM) and can be defined as “the positive or negative statements made about a product, company, or media personality that are made widely available on the Internet” (Hennig-Thurau, Gwinner, Walsh & Gremler, 2004 p.39). The increase of eWOM, however, is a double-edged sword. On the one hand, positive eWOM can be favourable for a brand, as user-generated reviews, recommendations and journals can positively influence a firm’s value such as stock price (Luo & Zhang, 2013). On the other hand, when eWOM is rather negative, it can affect a brand in an unfavourable way. For instance, it can lead to a denigrated corporate image and brand value (Chang, Hsieh & Tseng, 2013), leading to reduced future incomes and lower profits (Luo & Homburg, 2008). This is especially true for negative eWOM that spreads like a wildfire–fast and abundantly. This phenomenon has only been labelled recently by Pfeffer, Zorbach and Carley (2014), referring to it as a social media firestorm, implying that it portrays rather new and unfamiliar behaviour of consumers for firms. There have been studies that focused on the consequences of these waves of outrages on marketing communications of a firm (Pfeffer et al., 2014), the way in which it affects how customers perceive a brand in the short and long term (Hansen, Kupfer & Hennig-Thurau, 2018) and how firms can detect and reduce its virality (Herhausen, Ludwig, Grewal, Wulf and Schoegel, 2019).

(8)

8 information through digital platforms that happens on a large-scale, where certain companies or people are cancelled, or in other words: boycotted (Ng, 2020). A consumer boycott is defined as the “attempt by one or more parties to achieve certain objectives by urging individual consumers to refrain from making selected purchases in the marketplace.”

(Friedman, 1985, p. 97). One can also argue that a cancel culture represents the minorities or the oppressed who are now able to have a voice, reclaim power and fight back towards figures with more visibility and power (Butler, 2018; Semíramis, 2019). Wang & Siegel (2018) give an example where Nike Inc. chose to feature quarterback Colin Kaepernick (who started a discussion when he kneeled during the national anthem, as a protest against racism and police brutality) in its ads and people turned to Twitter using the hashtag #boycottnike. Additionally, after the campagin was launched, it was Donald Trump, who triggered the cancel culture, tweeting the following: “Just like the NFL, whose ratings have gone WAY DOWN, Nike is getting absolutely killed with anger and boycotts…” (Wang & Siegel, 2018). However, 44% among people aged 18-34 showed support for Nike’s decision to include Kaepernick in the ad (Meyersohn, 2018). This is an illustration of how this can lead to a discussion where two parties are involved: #teamboycott and #teamsupport. Additionally, it also shows how waves of eWOM can create group polarity and even a different brand perception amongst

consumers. However, it is still unclear whether online storms have an effect on the financial performance of a firm in the short or long term. This is not only relevant for investors, but it offers managerial insights as well. Investigating this relationship enables the marketing department to link financial output to marketing purposes. Consequently, the marketing department will be more accountable, and its influence will be larger within a firm (de Haan, Verhoef and Wiesel, 2020). Additionally, by differentiating between the effect on the

(9)

9 next to the data retrieved from Twitter. In this way, this research is able to determine whether traditional and social media echo each other (Hewett, Rand, Rust & van Heerde, 2016) and if cross-media coverage enhances the effect of an online storm. Lastly, this study will combine the quantitative and qualitative content of the online storm, through analyzing its volume and sentiment (de Haan, 2020).

The main purpose of this study is to determine if an online storm can actually affect a firm’s financial performance and simultaneously judge whether cancel culture has an effect.

To achieve these aims, this study uses panel data consisting of of 73,625 tweets and 977 articles for 29 firms over a period of 5,5 years (2015-2020). I conduct multiple panel data regressions for different models, enabling the identification of wear-in and wear-out effects of online storms. I find a significant and negative relationship between a higher level of virality of an online storm and the abnormal stock returns of a firm. This effect diminishes two days after the online storm has taken place. Moreover, I find that whenever an online storm leads to a discussion on Twitter with diverse opinions, this positively affects the abnormal stock returns of a firm. However, when combined with a higher virality level, these diverse opinions ensure a negative effect of virality on abnormal stock returns even four days after the storm occurred. Lastly, I conclude that Twitter is a very accurate and fast medium to assess stock returns and that it is likely that the impact of newspaper coverage is delayed. Nevertheless, a weak negative relationship between articles and abnormal stock returns indicates that even though traditional media may have become less important, its effect is still visible. Conclusively, these findings will contribute to extant literature by demonstrating a relationship between online storms and financial performance of firms while using the topical phenomenon of cancel culture.

(10)

10

2. LITERATURE REVIEW

The literature review will further specify earlier research in eWOM, its development into a social media firestorm, how cancel culture plays a role in this and why it is relevant to link this to the financial performance of a firm.

2.1 Interaction on social media

As the marketing landscape develops, customer firm interactions change with it. Nowadays, interaction between firm and consumer has become the center for value creation and value deliberation, resulting in an increase of customer to firm (C2F) and customer to customer (C2C) interactions, especially online (Prahalad & Ramaswamy, 2004). The consumers’ need for social interaction and economic incentives, their involvement with other consumers and the ability to enhance their own self-worth are the main factors leading to eWOM behavior and C2C interactions online (Hennig-Thurau, Gwinner, Walsh, Gremler, 2004). Prior research by Richins and Root-Shaffer (1988) has extensively studied the impact of traditional WOM, which refers to the offline consumer-generated, peer-to-peer sharing of know-how. In their study, they argue that offline C2C interactions usually results in an exchange of consumption-related information, offering a guidance for other customers to make informed purchase decisions. Opposed to WOM, eWOM has the power to spread a message within a much larger community or social network, which leads to impact on a global scale. Research has shown that C2C interactions are usually facilitated by eWOM in customer generated data sources (Libai, Bolton, Bügel, De Ruyter, Götz, Risselada & Stephen, 2010). Thus, eWOM spread within social media networks seems to be an important tool to get valuable insights for customer management or value creation for firms. Not only does it contain a great amount of information and textual-based content, it also has the power to influence consumers’ opinions and make an impact on firms (Xun and Guo, 2017). For example, de Haan (2019) illustrates that eWOM is a good predictor of future firm performance. This indicates how eWOM cannot only serve as a source of valuable information for customer management and value creation, but it can also be used for prediction purposes.

(11)

11 between consumers’ perceived integrity and attitude, suggesting that positive eWOM can lead to an increase in sales. Kudeshia and Kumar (2017) researched the specific context of consumer electronics, where they find that user-generated positive eWOM on social networking sites (in this case Facebook) influences the brand attitude and purchase intention of customers. According to Luo and Zhang (2013), user-generated reviews, recommendations and blogs show that the attitude of the customer can positively influence a firm’s value such as stock price. By contrast, negative eWOM can lead to reduced future incomes, lower profits (Luo and Homburg, 2008) and a denigrated corporate image and brand value (Chang, Hsieh and Tseng, 2013). In his article, de Haan (2020) finds that especially negative eWOM is a good predictor of firm performance. When firms experience high volumes of negative eWOM, it can lead to less purchases, the loss of current customers and keep new customers from buying. Ultimately, this could result in diminishing returns. Therefore, it seems very important for firms to pay close attention to both positive and negative eWOM, especially in the case of high volumes.

2.2 Social media firestorms

Research in the field of eWOM and sentiment has further expanded over the years. For example, combining eWOM and sentiment with interactions in online brand communities (Homburg, Ehm, and Artz 2015). In those communities, social media users have been able to cause waves of outrage as well as hypes about a certain brand or product. These waves of outrage are also known as online firestorms and are said to be the sudden release of large quantities of messages containing negative eWOM against a person, company or group (Pfeffer et al., 2014, p.118). An example of such a firestorm is the one following McDonald’s campaign in 2012. The popular fast-food chain wanted to make its followers aware of the heritage of the company’s food by launching the hashtag #meetthefarmers. After receiving positive feedback, McDonald’s decided to change the hashtag to #McDStories. However, the new hashtag was now used to share negative or funny stories about the company, resulting in more than 1000 negative tweets within the short time span of only two hours (Pfeffer et al., 2014).

(12)

12 the case of social media, consumers trust the credibility of information spread amongst their connections and choose to act accordingly. As a result, an online firestorm can influence consumers in a negative manner, ultimately affecting the performance of a firm itself. Thus, when it comes to marketing communications, it is important that companies that face an online firestorm retain their composure and continue to communicate and interact with its consumers (Pfeffer et al., 2014). Moreover, Hansen et al. (2018) were able to distinguish between the short- and long-term effects of firestorms on brand perceptions. Their findings show that strong online firestorms relate to negative changes of consumers’ short-term brand perceptions. Furthermore, they show that the impact of a firestorm that is triggered by a defective product or service is stronger than a firestorm resulting from inappropriate communication of the company. In addition, Herhausen et al. (2019) found a measure on how to detect and reduce the virality of online firestorms. In their article, they used brand communities on Facebook to identify potential online firestorms and focused on negative eWOM. Their main finding is that firestorms can be mitigated by specific firm responses over time. However, when it comes to virality on social media platforms, Pfeffer et al. (2014) stress that Twitter stands out when compared to other platforms. Due to the short message length communication is short and quick and results in a critical role in the propagation of online firestorms.

2.3 Cancel culture

(13)

13 Trump did not have to wear a mask during COVID while visiting one of Ford Motor Company’s plants, people brought the issue to Twitter using the hashtag #boycottford. On the other hand, there was a support group mainly consisting of Trump followers who claimed that the president and Ford CEO Bill Ford did the right thing, who used the hashtag #supportford in their tweets (Mangan, 2020). This illustrates how there are usually two sides to a discussion and how both positive and negative eWOM can be captured in an online storm characterized by cancel culture. In other words, this is considered to be a demonstration of group polarization where like-minded individuals boost in-group and out-group affiliation (Yardi & Boyd, 2010). When it comes to user-generated comments in such a setting, individuals who feel isolated tend to be more influenceable (Neubaum & Krämer, 2017). This is because believing that others will also boycott or support a company might lead to consumers feeling effective as part of a larger, collective effort (Neilson, 2010). Thus, eventually leading to an online storm, where more consumers feel the need to join and add to the conversation.

2.4 Stock market performance

(14)

14 2.5 Firestorms, cancel culture and stock market performance

(15)

15

3. HYPOTHESES

Using the information that has been provided in the literature, I am able to develop the conceptual model and the corresponding hypotheses illustrated in Figure 1, where H1 represents the main relationship and H2, H3, H4 and H5 are the moderator effects on this relationship. In this chapter I will further elaborate on this model and the corresponding hypotheses.

Figure 1: The conceptual model

3.1 Virality

(16)

16 indicating that social media firestorms can experience exponential growth in a very short time frame. According to Pfeffer et al. (2014), speed and volume are the first factors that are determined as relevant for opinion spreading in social media. Therefore, it can be said that these factors, contribute to the ultimate impact of the online storm. Because the concept of speed is hard to incorporate in the actual virality score, it will be further explained in the data section as the wear-in and wear-out effect. It is expected that once the consumer perception of a firm has been harmed or bettered by an online storm, that this will also have an effect on the financial performance of a firm. As De Haan, Verhoef and Wiesel (2015) find that customer feedback metrics can be used to predict retention, it can be argued that metrics such as consumer perception and satisfaction are ultimately related to financial performance. This is confirmed by Gupta and Zeithaml (2006), who argue that customer satisfaction is strongly correlated with behavioral outcome and financial performance. Referring back to the findings that viral eWOM can affect brand perceptions and ultimately the brand equity of a firm (Hansen et al., 2018; Pfeffer et al., 2014), this can conjointly affect financial performance. This results in the first and following hypothesis:

H1: A higher (lower) virality level of an online storm has a stronger (weaker) effect on the abnormal stock return of a firm.

3.2 Sentiment

(17)

17 sentiment from #support will weaken this negative effect or even have a positive effect. This leads to the second hypothesis, where the effect of the virality of the online storm on the abnormal stock return is moderated by the online sentiment score:

H2: The effect of the virality level of an online storm on the abnormal stock return is moderated by the sentiment score. That is, a lower (higher) online sentiment has a negative

(positive) effect.

According to Hewett et al. (2016), negative eWOM is transmitted more often and is considered to be more influential. This is confirmed by Blaine and Boyer (2018), who argue that negative emotional messages are shared more frequently. Additionally, research by de Haan (2020) and Tirunillai and Tellis (2012) shows that negative sentiment is strongly correlated with firm performance and negative reviews have a much stronger impact than positive reviews on stock market performance. Combining this with the fact that #teamboycott is expected to be bigger than #teamsupport in cancel culture, the effect of negative sentiment is expected to be stronger than that of positive sentiment, leading to the following sub-hypothesis:

H2a: The moderating effect of a lower online sentiment score is stronger than that of the higher online sentiment score.

3.3 Switching costs

(18)

18 they are currently involved with, or because they might be drawn to the products or services offered by another firm. In other words, switching costs become important whenever a firm receives bad or good publicity. Thus, whenever an online storm occurs, its effect on firm performance depends on whether the switching costs of that industry are low or high. It is expected that the effect of the virality of an online storm on a firm’s performance is stronger for industries that are characterized by low switching costs. This leads to the following hypothesis:

H3: The effect of the virality level of an online storm on the abnormal stock return is moderated by the switching costs. That is, an industry with lower (higher) switching costs

strengthens (weakens) the effect.

3.4 Cross-media coverage

Cross-media coverage is inspired by the concept of echo chambers (Key, 1966). In the modern context, traditional and social media echo each other (Hewett et al., 2016). This is because traditional media canals such as newspapers and television stations frequently pick up stories from Twitter at an early stage, confirming that (1) Twitter is a source of valuable information, but also (2) that the breadth of news coverage can further enhance the effect of the online storm (Hansen et al., 2018). In other words, whenever a newspaper covers the topic as well as a large number of tweets, the effect of an online storm on a firm’s performance might be even larger. This is because it will lead to additional repetition effects and provide consumers with increased opportunities to process the message (Pfeffer et al., 2014). Therefore, it is expected that an online storm that is characterized by a stronger cross-media effect will have a bigger impact on the abnormal stock return of a firm. This leads to the following hypothesis:

H4: The effect of the virality level of an online storm on the abnormal stock return is moderated by the cross-media coverage. That is, a higher (lower) cross-media coverage

strengthens (weakens) the effect.

(19)

19 have on abnormal stock returns in the first place. Therefore, the following hypothesis is developed to account for this effect:

H4a: The effect of the virality level of an online storm on the abnormal stock return is moderated by the sentiment score of the cross-media coverage. That is, a lower (higher)

online sentiment has a negative (positive) effect.

3.5 Like count

Social media users generate their affective responses to online messages in ways that are visible to others. For example, one of the ways in which users can express their feelings when it comes to Twitter and Facebook is by “liking” the tweet or post. This affective evaluation creates an emotional response and an attitudinal evaluation of the tweet (Alhabash & McAlister, 2015). In his article, Tucker (2011) argues that liking and disliking qualify as an engagement measure of online performance of people or businesses and that they can actually lead to virality of the content. This provides reason to assume that the amount of likes for the tweets that create the virality of an online storm enhances the effect of the online storm on stock performance. Hence, the following hypothesis has been developed, taking into account the effect of likes on the effect of the online storm:

(20)

20

4. DATA

To test the hypotheses of this study, data is collected from multiple sources that are combined to capture the characteristics and effects of the virality of the online storm on the abnormal stock return of the firms in the sample. This study uses three different data sources: (i) the virality, sentiment and tweet specific data used have been scraped from Twitter, (ii) the cross-media effects were obtained from articles downloaded from Nexis Uni, and (iii) the stock performance data was collected from Yahoo Finance and Fama French. In this section, I will clarify the selection procedure for the sample and further elaborate on the data sources and measures used in this study.

4.1 Sample selection

The aim for this study is to be able to capture positive and negative eWOM effects of the online storms centered around the two sides of cancel culture: #teamsupport and #teamboycott. Thus, the selection procedure for the sample of firms must include businesses that are involved in controversial issues creating either supporters or opponents of the firm. The Corporate Research Project (Corporate Rearch Project, 2020) is a platform that aims to assist the community, environmental and labour organizations in researching companies and industries. It has provided a Corporate Rap Sheet that includes the largest firms that have been involved in controversial cases. For this study, this platform has been used as a source to identify the firms that would most likely be involved in an online storm. The following criteria have been imposed on the sample of firms in this study: (i) The firm should be reported by the Corporate Research Project as a firm that has been involved in controversial activities over the past years. (ii) The firms should be active in the U.S. or have headquarters located in the U.S. This is due to the fact that the tweets and articles that will be analysed will be filtered according to the English language. (iii) The firm should trade on public markets, since the measurement of the abnormal stock return data requires the firms’ stock return data to be publicly available.

(21)

21 4.2 Virality of an online storm

In this research, the main investigated relationship is the effect of virality of an online storm on Twitter on the abnormal stock return of a firm. Twitter has been one of the most popular social media platforms that people use for networking. In terms of social media platforms, all social media networks have a relatively high turnover of information, but Twitter is found to be the fastest one (Pfeffer et al., 2014). It has a big impact because of its volume and speed of information, wide accessibility and significant reach. The short message length of 140 characters leads to the fact that communication is short and quick. Furthermore, the default setting for tweets is public, which allows almost everyone to follow and read other’s tweets without having to give mutual permission (Pfeffer et al., 2014). Not only individual consumers use this platform, but also investors and firms can have Twitter accounts, creating a place for customer-to-customer interactions, but also for customer to firm interactions. Therefore, Twitter seems to play an important role in the creation of online storms. Given the valuable information that is shared on the platform, as well as its wide adoption and active users, Twitter has a huge social and economic impact with multiple dynamics that still have to be explored (Jansen, Zhang, Sobel, Chowdury, 2009). When it comes to using Twitter for research, it is relatively easy to scrape tweets in comparison with other online platforms such as Facebook or Instagram. Moreover, Twitter does not remove tweets after a certain amount of time, resulting in a broad availability of data for many firms over a longer period of time. Therefore, this study uses Twitter as the main source of information when investigating online storms.

Initially, the aim is to use Twitterscraper, which is a Python package developed by Taspinar (https://github.com/taspinar/twitterscraper), to scrape the necessary Twitter data. However, due to new rules imposed by Twitter, there is a maximum set to the number of tweets that can be scraped using this package or others such as GetOldTweets3. Therefore, to scrape the tweets necessary for this study the Snscrape package

(22)

22 used to specify that solely English tweets should be included in the sample, , as this is necessary for the sentiment analysis done afterwards. Additionally, after removing the duplicated tweets, all tweets are unique and occur only once. Once again, this increases the accuracy of the sample.

This results in a total of 73,625 tweets for all the firms associated with these hashtags during the 2015-2020 time period. To determine whether an online storm has taken place, the number of tweets and reweets are summed up per day and per firm and labelled as the “Virality”. Thereafter, the virality score will be categorized by their virality levels from lower to higher. This will enable the possibility to test the main effect, thus whether a higher virality score has a bigger impact on a firm’s stock performance. According to Hansen et al. (2018), an online storm takes place whenever at least 10 tweets have covered the storm topic, so it is expected that an effect on stock performance will become visible with a virality category with tweets above 10. Nevertheless, it is interesting to know whether a number of tweets below this threshold will already have an impact, thus these are included as well, placed under a lower category.

4.3 Sentiment

(23)

23 and negative words and corresponding context clusters and dividing them by the square root number of words in the tweet.

After calculating the sentiment scores of all 399,894 tweets (incl. retweets), I will create several daily scores for each of the 29 firms. In line with De Haan (2020), I will create (i) average sentiment scores illustrating the average sentiment, (ii) the share of positive tweets and (iii) the share of negative tweets and lastly, (iiii) the standard deviation of the sentiment will be used to assess the heterogeneity of the sentiment.

4.4 Switching costs

When defining the switching costs per firm, the industry has to be considered. According to Gremler and Brown (1996), switching costs for goods and food tend to be lower than for services. For example, for some services such as fast-food restaurants or retailers, switching costs are low because customers are mostly anonymous. Another factor that plays a role is the amount of competition that is present in an industry. A “loose monopoly” (Hirschman, 1970) is characterized by a market in which there is a minimum amount of competition and thus there is a presence of near-monopoly. When a firm has a loose monopoly in a market, it is difficult for consumers to choose another firm, resulting in high switching costs. The firms included in the sample are characterized by eight industries: Food/Beverage Processing, Retailing, Automobile, Media/Entertainment, Banking and Finance, Aerospace and Petroleum. The switching costs will be assigned to each industry by two levels, where each industry belongs to the low (0) or high (1) category. Since research is limited in this field, I will use common sense and theory to determine which industries are placed in the low or high switching costs category. The firms included in the Food/Beverage Processing, Retailing and Media/Entertainment industries are placed in the low category and the firms included in the Automobile, Banking and Finance, Aerospace and Petroleum are placed in the high category.

4.5 Cross-media coverage

(24)

24 articles have to include the firm’s name and the word boycott and (iii) the articles have to be written in English. Thereafter, I manually scan the articles to filter out duplicate articles or articles that cover unrelated topics such as governments boycotting other governments. This leaves me with a sample of 3,967 articles that can be used for measuring volume and sentiment per day.

After downloading the articles for all 29 firms, I use a loop that I have developed in Python to transform the articles per firm into useful datasets with the date of publication and the content of the article. Furthermore, the sentiment per article will also be analysed. This was done using the qdap package in R which has also been used for the Twitter data. After cleaning the articles to ensure that punctuation marks, numbers or company names do not bias the results, the sentiment analysis can be performed. Similar to the twitter dataset, the articles receive (i) an average sentiment scores illustrating the average sentiment, (ii) scores for the share of positive articles and (iii) the share of negative articles and lastly, (iiii) a score for the standard deviation of the sentiment, which will be used to assess the heterogeneity of the sentiment. After obtaining these scores per article, the volume of the articles per day is determined through summing the total number of articles published per day that a Tweet occurred. Since the articles have to be matched to the Twitter data to be able to test the hypotheses involving cross-media coverage, this leaves me with less articles (977) than I have initially collected.

Table 1 in the appendix provides an overview of all firms with their corresponding amount of tweets and articles used in the study.

4.6 Abnormal stock returns

(25)

25 Kenneth French (https://mba.tuck.dartmouth.edu/pages/faculty/ken.french/data_library.html) and the stock return data from Yahoo Finance (Finance.yahoo.com).

First, the Capital Asset Pricing Model (CAPM), created by William Sharpe (1964), is able to estimate the return of an asset based on the market returns and the asset’s linear relationship to the market returns, which is also known as the beta. The beta is a measure of the systematic risk of an asset compared to the market as a whole. It is able to describe the relationship between systematic risk and the expected return for assets. I have measured this beta on a yearly basis for each of the 29 firms in the sample, using a rolling period of 24 months. To specify, I have measured the beta’s annually for each company on the stock return and market returns obtained from Fama and French, 24 months before the start of the year. To illustrate, for the beta of 2015, I have used stock return data and market returns data for all firms from 01-01-2013 until 31-12-2014. The yearly beta’s per firm have then been assigned as the beta per day per firm in the dataset. After having done this, I am able to calculate the Capital Asset Expected Return using the CAPM model (1):

𝐸𝐸(𝑅𝑅

𝑖𝑖

) = 𝑅𝑅

𝑓𝑓

+ 𝛽𝛽

𝑖𝑖

(𝐸𝐸(𝑅𝑅

𝑚𝑚

) − 𝑅𝑅

𝑓𝑓

) (1)

Where E(Ri) is the capital asset expected return, Rf is the risk-free rate of interest, βi is

sensitivity and E(Rm) is the expected return of the market (together they represent the Beta).

Thereafter, I am able to calculate the abnormal stock returns per day per firm using the following formula (2):

𝐴𝐴𝑅𝑅

𝑖𝑖𝑖𝑖

= 𝑅𝑅

𝑖𝑖𝑖𝑖

− 𝐸𝐸(𝑅𝑅

𝑖𝑖𝑖𝑖

) (2)

Where ARit is the abnormal stock returns per firm per day, Rit is the actual return per firm per

day and E(Rit) is the expected return per firm per day, which has been calculated using the

(26)

26 4.6.1 Dynamic effects

(27)

27

5. METHODOLOGY

5.1 Descriptives

(28)

28

Notes: N= total observations, St.Dv= standard deviation

Table 2:

Descriptive Statistics

Overview of the descriptive statistics for each of the variables used in this research, the 17 different independent variables and the six different dependent variables for abnormal stock returns. Each of the variables has data on a daily basis for the period of

01-01-2015 until 01-06-2020. I have measured all variables using Twitter data for the twitter variables and newspaper data for the article variables. Moreover, the abnormal stock returns are obtained from Yahoo Finance and the website of Kenneth French.

(29)

29 Table 3: Correlation matrix (to be continued on the next page)

Notes: ****p < .001, ***p < .01, **p < .05, *p < .10.

(30)

30 Table 3: Correlation matrix

*

Articlesavgpolarity = Average polarity articles Articlessdpolarity = SD polarity articles Articlespospolarity = Positive polarity articles Articlesnegpolarity = Negative polarity articles abnormallead1 = Abnormal Returns (t+1) abnormallead2 = Abnormal Returns (t+2) abnormallead3 = Abnormal Returns (t+3) abnormallead4 = Abnormal Returns (t+4) abnormallead1 = Abnormal Returns (t-1)

Notes: ****p < .001, ***p < .01, **p < .05, *p < .10.

AbnormalReturns abnormallead1 abnormallead2 abnormallead3 abnormallead4 abnormallag1 Articledummy

(31)

31 5.2 Model specification

In this section, the models used to analyse the relationships conceptualized in Figure 1 will be introduced, but first the type of data must be specified. The data included in the study can be characterized as multi-dimensional panel data. This is because the data in this study covers multiple firms on multiple points in time, indicating that there is a time and an individual dimension (Markus, 1979). The period in which the data will be studied covers a total of 5,5 years and the sample consists of 29 different firms. Moreover, the dependent variable (abnormal stock return) has an interval scale. Because of this, the most suitable regression analysis is a panel regression. This is a statistical method that is often used to deal with multi-dimensional panel data which includes a time and individual dimension (Xun and Guo, 2017). It is especially useful when regression residuals are correlated across individuals (Davies and Lahiri, 1995) or in the case of endogeneity (Wooldridge, 2011). In this study, there may be unique attributes of firms that are not the result of a random variation and may not change often across time. To illustrate, one firm might have a much larger consumer base than another, resulting in much more Twitter activity and a bigger impact. This type of regression ensures that this fixed effect is controlled for and biased results are avoided.

Before I am able to develop separate models for every hypothesis and dependent variable, I have to determine the type of model that is best suitable for data. There are three types of models suited for panel data: the pooled model, the fixed model and the random effects model (Arellano, 2003). The pooled model is the most restrictive type of panel data model because it specifies constant coefficients. The fixed effects model allows for individual-specific effects αi to be correlated with regressors x, in which αi is included as intercepts. Every firm would have a different intercept term and the same slope parameters. Lastly, the random effects model allows for individual-specific effects αi to be distributed independently of the regressors x. Here, every firm has the same slope parameters and a composite error term and αi is included in the error term (Econometrics Academy, 2013).

(32)

32 Next, I must perform a Breusch-Pagan Lagrange Multiplier test to detect whether I must continue with the random effects regression or the pooled OLS regression. As the results show that there is no significant difference across brands (p-value>.001), I am able to run a simple OLS regression. Finally, I have to include time fixed effects in all of the models.

Now that I have determined what type of model is most suitable for the data, I am able to develop the different models necessary for testing the hypotheses. To incorporate the dynamic effects in this study I have developed six models where each dependent variable is different. The first being (i) the Abnormal Stock Returns on the day of the online storm, (ii) the Abnormal Stock Returns of the day after the online storm, (iii) the Abnormal Stock Returns two days after the online storm, (iiii) the Abnormal Stock Returns three days after the online storm, (v) the Abnormal Stock Returns four days after the online storm and (vi) the Abnormal Stock Returns one day prior to the online storm.

5.3 Assumptions

(33)

33 Therefore, while omitting variables with high VIF scores I will also evaluate the models on the size, sign and significance of the coefficients, the adjusted R-squared and the Akaike Information Criteria (AIC) and Bayesian Information Criteria (BIC), which is in line with De Haan, Verhoef, and Wiesel (2015). Both the AIC and the BIC measure whether a variable increases or decreases the efficiency of the model. The BIC is a more restrictive version of the AIC because it uses the number of observations in its calculation. Therefore, I look at both, but the BIC is the most decisive in the model decision (Leeflang et al., 2015). The equations below represent the formulas behind the calculation of AIC (3) and BIC (4), where L is the value of the likelihood and K is the number of independent variables in the model and T is the amount of observations in the model.

𝐴𝐴𝐴𝐴𝐴𝐴 = −2 × ln(𝐿𝐿) + 2 × (𝐾𝐾 + 1) (3)

𝐵𝐵𝐴𝐴𝐴𝐴 = −2 × ln(𝐿𝐿) + ln(𝑇𝑇) × (𝐾𝐾 + 1) (4)

While omitting and factoring variables and creating new models simultaneously, I check both the VIF scores and the AIC/BIC scores of each model. I do this until removing or factoring a variable significantly worsens the model, which can be seen by an increasing AIC/BIC score. This is because a score that is closer to zero means that the extra variable is beneficial to the model compared to a version of the model without that particular variable. In this way, I am able to create a model with the lowest AIC and VIF scores that ultimately is the best predictor for this study. The final Log Likelihood, AIC and BIC scores per model as well as the VIF scores per variable have been included in Table 4 and 5 in the Appendix.

(34)

34 is the direct abnormal stock return. Before doing so, I find that I cannot transform categorical variables. Thus, I have to replace the variable ‘Viralitycat’ by three dummy variables for the different virality levels and the fixed effects factor Year variable by dummy variables for each year. After transforming the variables, where the current value is subtracted with a multiplication of the correlation value and previous value, I perform another Durbin-Watson test, which turns out to be insignificant. Now all models are free of autocorrelation.

Secondly, a presence of heteroskedasticity in the models must be identified. This exists whenever not all observations have equal variances of the disturbance term (Leeflang et al., 2015). Thus, if differences across firms in the sample exist, these should be accounted for in the model and not be driven into the disturbance term. To investigate whether the prediction of stock performance has changed over the 5,5 years in the dataset and to find out how it is reflected in the error term, I have split the dataset in two periods (1: 2015, 2016, 2017 and 2: 2018, 2019, 2020) consisting of +/- the same number of observations. Thereafter, I have performed the Goldfield-Quandttest for heteroskedasticity, which tests whether the variances are similar in both periods. The results show significance for all six models. This means that the prediction of stock performance in the period 2015-2017 significantly differs from those in the period 2018-2020. Thus, another GLS transformation on the variables will be performed, which is done through standardizing the observations from the first period using their own standard deviation and the observations from the second period using their own standard deviation. Additionally, I have added a column that replaces the constant with a value of 1 that is divided by the standard deviation of the first period and the standard deviation of the second period. Thereafter, the new models will be estimated without a constant. Now all models are free of heteroskedasticity.

(35)

35 manually change the p-values for all models, resulting in six final models that can be classified as normal.

Finally, after having satisfied all the assumptions and selecting the models with the lowest AIC and BIC scores, I am left with six final models that are suitable for the six different dependent variables. The dependent variable Yi,t represents the abnormal stock returns on day t. In the first

model, where abnormal stock returns (t) is the dependent variable, I have added all other dependent variables for robustness. Besides, in all models I have only included virality level 2 and 3, as virality level 1 is not classified as an online storm. Moreover, all models include a time specific dummy to account for the time fixed effects in those models (De Haan, 2020), the moderators explained in this study, an intercept and an error term. Below, the equations for Yi,t

(5) and Yi,t+1 (6) have been provided. These are the equations for the models where Abnormal

Returns (t) and Abnormal Returns (t+1) are the dependent variabels. Besides these, I will test four more models where the dependent variables are: Yi,t+2, Yi,t+3, Yi,t+4 and Yi,t-1.

𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴 𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝐴𝐴𝐴𝐴𝑅𝑅𝑖𝑖,𝑖𝑖 = 𝛽𝛽0+ 𝛽𝛽1𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝐴𝐴𝐴𝐴𝑅𝑅𝑖𝑖,𝑖𝑖+1 1 + 𝛽𝛽2𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝐴𝐴𝐴𝐴𝑅𝑅𝑖𝑖,𝑖𝑖+22 + 𝛽𝛽3𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝐴𝐴𝐴𝐴𝑅𝑅𝑖𝑖,𝑖𝑖+33 + 𝛽𝛽4𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝐴𝐴𝐴𝐴𝑅𝑅𝑖𝑖,𝑖𝑖+44 + 𝛽𝛽5𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝐴𝐴𝐴𝐴𝑅𝑅𝑖𝑖,𝑖𝑖−15 + 𝛽𝛽6𝑉𝑉𝑉𝑉𝐴𝐴𝐴𝐴𝐴𝐴𝑉𝑉𝑅𝑅𝑉𝑉𝐴𝐴𝑅𝑅𝑉𝑉𝑅𝑅𝐴𝐴26+ 𝛽𝛽7𝑉𝑉𝑉𝑉𝐴𝐴𝐴𝐴𝐴𝐴𝑉𝑉𝑅𝑅𝑉𝑉𝐴𝐴𝑅𝑅𝑉𝑉𝑅𝑅𝐴𝐴37+ 𝛽𝛽8𝐴𝐴𝑉𝑉𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝑉𝑉𝑅𝑅𝑉𝑉8+ 𝛽𝛽9𝑆𝑆𝑆𝑆𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝑉𝑉𝑅𝑅𝑉𝑉9+ 𝛽𝛽10𝐿𝐿𝑉𝑉𝐿𝐿𝑅𝑅𝐿𝐿𝐴𝐴𝑅𝑅𝐴𝐴𝑅𝑅10+ 𝛽𝛽11 𝑉𝑉𝐴𝐴𝐴𝐴𝑅𝑅𝐴𝐴𝑅𝑅_𝐴𝐴𝐴𝐴𝑅𝑅𝑉𝑉𝐿𝐿𝐴𝐴𝑅𝑅𝑅𝑅11+ 𝛽𝛽12𝐴𝐴𝑉𝑉𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝑉𝑉𝑅𝑅𝑉𝑉_𝐴𝐴𝐴𝐴𝑅𝑅𝑉𝑉𝐿𝐿𝐴𝐴𝑅𝑅𝑅𝑅12+ 𝛽𝛽13𝑆𝑆𝑆𝑆𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝑉𝑉𝑅𝑅𝑉𝑉_𝐴𝐴𝐴𝐴𝑅𝑅𝑉𝑉𝐿𝐿𝐴𝐴𝑅𝑅𝑅𝑅13+ 𝛽𝛽14𝑆𝑆𝑆𝑆𝑉𝑉𝑅𝑅𝐿𝐿ℎ𝑉𝑉𝐴𝐴𝑖𝑖𝐿𝐿𝐴𝐴𝑅𝑅𝑅𝑅𝑅𝑅14+ ∑2020𝑖𝑖=2015𝛽𝛽15𝑌𝑌𝑅𝑅𝐴𝐴𝐴𝐴𝑖𝑖+ 𝜀𝜀𝑖𝑖 (5) 𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴 𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝐴𝐴𝐴𝐴𝑅𝑅𝑖𝑖,𝑖𝑖+1 = 𝛽𝛽0+ 𝛽𝛽1𝑉𝑉𝑉𝑉𝐴𝐴𝐴𝐴𝐴𝐴𝑉𝑉𝑅𝑅𝑉𝑉𝐴𝐴𝑅𝑅𝑉𝑉𝑅𝑅𝐴𝐴21+ 𝛽𝛽2𝑉𝑉𝑉𝑉𝐴𝐴𝐴𝐴𝐴𝐴𝑉𝑉𝑅𝑅𝑉𝑉𝐴𝐴𝑅𝑅𝑉𝑉𝑅𝑅𝐴𝐴32+ 𝛽𝛽3𝐴𝐴𝑉𝑉𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝑉𝑉𝑅𝑅𝑉𝑉3+ 𝛽𝛽4𝑆𝑆𝑆𝑆𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝑉𝑉𝑅𝑅𝑉𝑉4+ 𝛽𝛽5𝐿𝐿𝑉𝑉𝐿𝐿𝑅𝑅𝐿𝐿𝐴𝐴𝑅𝑅𝐴𝐴𝑅𝑅5+ 𝛽𝛽6 𝑉𝑉𝐴𝐴𝐴𝐴𝑅𝑅𝐴𝐴𝑅𝑅_𝐴𝐴𝐴𝐴𝑅𝑅𝑉𝑉𝐿𝐿𝐴𝐴𝑅𝑅𝑅𝑅6+ 𝛽𝛽7𝐴𝐴𝑉𝑉𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝑉𝑉𝑅𝑅𝑉𝑉_𝐴𝐴𝐴𝐴𝑅𝑅𝑉𝑉𝐿𝐿𝐴𝐴𝑅𝑅𝑅𝑅7+ 𝛽𝛽8𝑆𝑆𝑆𝑆𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝑉𝑉𝑅𝑅𝑉𝑉_𝐴𝐴𝐴𝐴𝑅𝑅𝑉𝑉𝐿𝐿𝐴𝐴𝑅𝑅𝑅𝑅8+ 𝛽𝛽9𝑆𝑆𝑆𝑆𝑉𝑉𝑅𝑅𝐿𝐿ℎ𝑉𝑉𝐴𝐴𝑖𝑖𝐿𝐿𝐴𝐴𝑅𝑅𝑅𝑅𝑅𝑅9+ ∑2020𝑖𝑖=2015𝛽𝛽9𝑌𝑌𝑅𝑅𝐴𝐴𝐴𝐴𝑖𝑖 + 𝜀𝜀𝑖𝑖 (6)

(36)

36

6. RESULTS

In this section, the regression results will be discussed. I will start by discussing the main effects (H1) for the six different models, of which the results have been provided in table 6, 7 and 8. Thereafter the models that include the moderating effects (H2-H6) will be analysed, of which the regression tables can be found in the Appendix.

6.1 Main effects

Tables 6,7 and 8 provide the estimates and p-values for the main effects of the models with the six different dependent variables ((i) Abnormal Returns (t), (ii) Abnormal Returns (t+1), (iii) Abnormal Returns (t+2), (iiii) Abnormal Returns (t+3), (v) Abnormal Returns (t+4) and (vi) Abnormal Returns (t-1)). The main effect can also be captured by the first hypothesis, namely the relationship between virality and abnormal stock returns. Other than that, interesting effects of other variables included in the models will be discussed.

(37)

37 characterized by a U-shape, where it starts negative and becomes more negative until it is absent and becomes positive. Additionally, it is only virality level 2 that corrects itself, indicating that virality level 3 most likely has a bigger impact in the long term. Moreover, a day prior (t+1) to the start of the online storm, there is no significant relationship between virality and abnormal stock returns, indicating that there is no wear-in effect. That is to say, there was probably no prior data on the cancel culture topic of the online storm before it actually took place on Twitter, indicating that Twitter is an accurate and up to date medium when it comes to assessing stock returns.

Drawing from this, H1 can be accepted, indicating that a higher level of virality of an online storm does have an effect on the abnormal stock returns and thus on how a firm performs on the stock market compared to how they are expected to perform. Moreover, the negative and significant effect lasts no longer than two days, being even stronger on the day after the online storm for a higher virality level. This indicates that there is a presence of a wear-out effect of two days after which the effect stabilizes again for lower virality.

Considering the other effects of variables included in the models, it becomes evident that the standard deviation of sentiment is significantly related to abnormal stock returns. To specify, a day after the online storm has taken place, the standard deviation of sentiment shows a positive and significant effect (b= .31, p<.05). It could be that the online storm is indeed strongly characterized by #teamboycott and #teamsupport, creating a discussion with more diverse opinions and leading to a higher standard deviation of sentiment. Consequently, this has a positive effect on the abnormal stock returns, not only a day after but also two days (b= .25, p < .05) and three days (b= .20, p<.10) after the online storm. From it can be concluded that whenever the discussion of cancel culture on Twitter is more diverse, it can weaken the negative effect of the online storm in the short term, or lead to a positive effect on the abnormal stock return of the firm.

Moreover, it becomes evident that the volume of articles has a weak and significant effect (b= .06, p<.10) on abnormal stock returns the day prior to the online storm. This gives a slight reason to assume that newspapers will only report on the topic of the storm a day later, once it visibly affects the abnormal stock returns.

(38)

(39)

39

Table 6: Regression results of main effects for Abnormal Returns (t) and Abnormal Returns (t+1)

(40)

40

Table 7: Regression results of main effects for Abnormal Returns (t+2) and Abnormal Returns (t+3)

(41)

41

Table 8: Regression results of main effects for Abnormal Returns (t+4) and Abnormal Returns (t-1)

Notes: SD = standard deviation, polarity = sentiment. ****p < .001, ***p < .01, **p < .05, *p < .10. Abnormal Returns (t+4) Predictors Estimates p (Intercept) -0.09** 0.012 Virality level 2 0.11*** 0.009 Virality level 3 0.05 0.172 Average polarity -0.01 0.394 SD polarity -0.07 0.316 Like Count -0.00 0.337 Volume articles 0.01 0.388 Average polarity articles 0.01 0.398 SD polarity articles 0.04 0.395 Switching costs -0.06 0.225 2016 0.09* 0.058 2017 0.13**** 0.001 2018 0.11* 0.069 2019 0.16*** 0.005 2020 -0.01 0.398 Observations 8493 R2_{/ R}2_adjusted _{0.004 / 0.002} Abnormal Returns (t-1) Predictors Estimates p (Intercept) -0.03 0.240 Virality level 2 0.05 0.247 Virality level 3 0.03 0.323 Average polarity 0.12 0.157 SD polarity -0.17 0.160 Like Count -0.00 0.365 Volume articles -0.06* 0.088

(42)

42 6.2 Moderating effects

This section will shed light on the models that include the moderating effects that illustrate H2-H6. The corresponding tables are provided in the appendix.

First of all, tables 10-15 in the appendix show that there is no evidence to support the second hypothesese (H2 and H2a), namely that negative (positive) sentiment strengthens the negative (positive) effect of virality on abnormal stock returns and that this effect is even stronger for negative sentiment. Solely the interaction between the standard deviation of sentiment and virality on the fourth day after the online storm turns out to be slightly negative significant for virality level 2 (b=-.5, p<0.10) and very negative significant for virality level 3 (b=-.94, p<0.01). This means that a higher virality level combined with mixed sentiment of the tweets characterized by #teamsupport and #teamboycott leads to a stronger negative effect on the abnormal stock returns four days after the online storm has taken place. It could be that once the effect of virality on abnormal stock returns has stabilized after four days but there is still a discussion going on, it is harder for the negative effect of the storm to diminish.

Secondly, tables 16, 17 and 18 in the appendix show that H3 is also not supported as the interaction effect between switching costs and the virality levels turns out to be insignificant for all models. This might suggest that switching costs do not matter in this relationship or that there is too little variation between the switching costs of the firms included in this sample to be able to cause a significant moderating effect.

Thirdly, tables 19-24 in the appendix illustrate that the hypotheses regarding cross media coverage (H4 and H4a) are also not supported. Nevertheless, there is a very weak significant effect between the interaction of virality level 2 and the average polarity of the articles and the abnormal stock returns the day after (b=.60, p<.10) and three days after (b= -.57, p<.10). A possible explanation for this could be that the whenever virality level 2 is combined with positive polarity of articles that this will only be picked up by investors a day after, resulting in a positive effect on abnormal stock returns. Moreover, this effect can turn out to become negative three days after, whenever this effect diminishes and stabilizes again.

(43)

43 Table 9 below includes an overview of the hypotheses and whether they have been rejected or accepted.

Table 9: Results of the hypotheses

Hypothesis Result

H1: A higher (lower) virality level of an online storm has a stronger (weaker)

effect on the abnormal stock return of a firm.

Accepted

H2: The effect of the virality level of an online storm on the abnormal stock

return is moderated by the sentiment score. That is, a lower (higher) online sentiment has a negative (positive) effect.

H2a: The moderating effect of a lower online sentiment score is stronger

than that of the higher online sentiment score.

Rejected

return is moderated by the switching costs. That is, an industry with lower (higher) switching costs strengthens (weakens) the effect.

Rejected

return is moderated by the cross-media coverage. That is, a higher (lower) cross-media coverage strengthens (weakens) the effect.

H4a: The effect of the virality level of an online storm on the abnormal stock

return is moderated by the sentiment score of the cross-media coverage. That is, a lower (higher) online sentiment has a negative (positive) effect.

Rejected

return is moderated by the amount of likes. That is, more (less) likes have a positive (negative) effect.

(44)

44

7. DISCUSSION

The main objective of this research has been to investigate the effect of online storms characterized by a cancel culture on financial performance of firms. In this study, I collected panel data from multiple sources (Twitter, Newspapers, FamaFrench and YahooFinance) for 29 firms across eight different industries for the period 2015-2020. To the best of my knowledge, this study is the first to find a relationship between an online storm specified by the highly topical phenomena cancel culture and financial performance and is able to make a few important contributions to extant literature.

(45)

45 itself diminishes after two days, one can say that the interest in the topic also diminishes. However once diverse discussion and thus interest stays alive, virality can still have a negative effect four days later.

Secondly, Herhausen et al. (2019) find that the use of more negative eWOM increases virality and makes it more contagious than in the case of less negative eWOM. However, I find no significant results for the interaction effect between sentiment and virality on abnormal stock returns. I did expect the negative sentiment used in the online storm to strengthen the effect of the online storm on financial performance itself. This is in line with de Haan (2019) Xun and Guo (2017) who state that negative eWOM does exhibit strong associations with firm’s stock market performances. Besides, Ranco et al. (2015) find that Twitter sentiment and stock return are significantly related to one another during peaks of Twitter volume. As mentioned, the diversity of sentiment plays a significant role in the effect on financial performance, not the actual sentiment itself. A reason for this could be that the majority of the tweets in the online storm contained the hashtag #boycott compared to the other smaller group that tweeted #support, yet both parties can still create a diverse discussion. In this way, this research is not really able to spot a significant difference between both sentiment groups. I do believe that since virality negatively affects financial performance, that the majority of the tweets in these storms come from #teamboycott. Thus, one can assume that the emotion of the tweets does play a role, yet there is no statistical proof to be found within the sample.

(46)

(47)

47

8. LIMTATIONS AND FURTHER RESEARCH

As with any research, this study has several limitations. These limitations also provide possibilities for further research. First of all, Twitter is the sole source used in this research to capture online storms regarding cancel culture. However, there are many other platforms where an online storm can take place, such as Facebook or Instagram. Previous research has already shown that online brand communities on Facebook can create online storms (Herhausen et al., 2019), thus it is likely that this platform also plays a role in stimulating a cancel culture. Adding data from other platforms can make the results of this study more generalizable and robust. Additionally, when downloading the newspapers, I have used a filter boycott to generate a manageable number of articles. However, despite #boycott being a typical hashtag used in cancel culture on online platforms, it is likely that it is not always included in the articles that cover the topic of the online storm. In this way, I was left with a relatively small sample of newspapers (3967), wondering if a higher number of articles would have created more extensive

(48)

(49)

49

REFERENCES

Alhabash, S., & McAlister, A. R. (2015). Redefining virality in less broad strokes: Predicting viral behavioral intentions from motivations and uses of Facebook and Twitter. New Media & Society, 17(8), 1317–1339.

Arellano, M. (2003). Panel data econometrics. Oxford university press.

Berger, J., Humphreys, A., Ludwig, S., Moe, W. W., Netzer, O., & Schweidel, D. A. (2020). Uniting the tribes: Using text for marketing insight. Journal of Marketing, 84(1), 1-25.

Blaine, T., & Boyer, P. (2018). Origins of sinister rumors: A preference for threat-related material in the supply and demand of information. Evolution and Human Behavior, 39(1), 67-75.

Butler, D. (2018, October 23). “The Misplaced Hysteria About a ‘Cancel Culture’ That Doesn’t Actually Exist.” Retrieved January 07, 2021, from https://verysmartbrothas.theroot.com/the-misplaced-hysteria-about-a-cancel-culture-that-do-1829563238

Chang, A., Hsieh, S.H. & Tseng, T.H. (2013). Online brand community response to negative brand events: the role of group eWOM. Internet Research, Vol. 23 No. 4, pp. 486-506.

Cheung, C. M., Lee, M. K., & Thadani, D. R. (2009, September). The impact of positive electronic word-of-mouth on consumer online purchasing decision. In World Summit on Knowledge Society (pp. 501-510). Springer, Berlin, Heidelberg.

Chih, W. H., Wang, K. Y., Hsu, L. C., & Cheng, I. S. (2012). From disconfirmation to switching: an empirical investigation of switching intentions after service failure and recovery. The Service Industries Journal, 32(8), 1305-1321

(50)

50 Corporate Research Project. (2020). CORPORATE RAP SHEETS. Retrieved October 26, 2020 from https://www.corp-research.org/corporaterapsheets.

Davies, A. and Lahiri, K. (1995). A new framework for analyzing survey forecasts using three-dimensional panel data. Journal of Econometrics. Vol. 68 No. 1, pp. 205-227.

Deephouse, D. L. (2000). Media reputation as a strategic resource: An integration of mass communication and resource-based theories. Journal of management, 26(6), 1091-1112.

De Haan, E.. (2020). Satisfaction Surveys or Online Sentiment: Which Best Predicts Firm Performance? Marketing Science Institute Working Paper Series, 20(101), 1-46.

De Haan, E., Verhoef, P. C., & Wiesel, T. (2015). The predictive ability of different customer feedback metrics for retention. International Journal of Research in Marketing, 32(2), 195-206.

De Haan, E., Verhoef, P.C., Wiesel, T. (2020). Customer Feedback Metrics for Marketing Accountability.

Duan, W., Gu, B., & Whinston, A. B. (2008). The dynamics of online word-of-mouth and product sales—An empirical investigation of the movie industry. Journal of retailing, 84(2), 233-242.

Fama, E., French, K. (1992). The Cross-Section of Expected Stock Returns. The Journal of Finance, 47(2), 427-465.

Friedman, M. (1985). Consumer Boycotts in the United States, 1970–1980: Contemporary Events in Historical Perspective. Journal of Consumer Affairs, 19 (1), 96–117.

Gonçalves, P., Araújo, M., Benevenuto, F., & Cha, M. (2013). Comparing and combining sentiment analysis methods. In Proceedings of the first ACM conference on Online social networks (pp. 27-38)

(51)

51 Hansen, L. K., Arvidsson, A., Nielsen, F. Å., Colleoni, E., & Etter, M. (2011). Good friends, bad news-affect and virality in twitter. In Future information technology (pp. 34-43). Springer, Berlin, Heidelberg.

Hansen, N., Kupfer, A. K., & Hennig-Thurau, T. (2018). Brand crises in the digital age: The short-and long-term effects of social media firestorms on consumers and brands. International Journal of Research in Marketing, 35(4), 557-574.

Hennig-Thurau, T., Gwinner, K. P., Walsh, G., & Gremler, D. D. (2004). Electronic word-of-mouth via consumer-opinion platforms: what motivates consumers to articulate themselves on the internet?. Journal of interactive marketing, 18(1), 38-52.

Herhausen, D., Ludwig, S., Grewal, D., Wulf, J., & Schoegel, M. (2019). Detecting, preventing, and mitigating online firestorms in brand communities. Journal of Marketing, 83(3), 1-21.

Hewett, K., Rand, W., Rust, R. T., & Van Heerde, H. J. (2016). Brand buzz in the echoverse. Journal of Marketing, 80(3), 1-24.

Hirschman, A.O. (1970). Exit, Voice and Loyalty. Harvard University Press, Cambridge, MA.

Homburg, C., Ehm, L., & Artz, M. (2015). Measuring and managing consumer sentiment in an online community environment. Journal of Marketing Research, 52(5), 629-641.

Jansen, B.J., Zhang, M., Sobel, K. & Chowdury, A. (2009). Twitter power: tweets as electronic word of mouth. Journal of the American Society for Information Science and Technology, 60(11), pp. 2169-2188.

Jarque, C. M., & Bera, A. K. (1987). A test for normality of observations and regression residuals. International Statistical Review/Revue Internationale de Statistique, 163-172.

(52)

52 Klemperer, P. (1995). Competition when consumers have switching costs: An overview with applications to industrial organization, macro-economics, and international trade. Review of Economic Studies, 62(4), 515–539.

Kudeshia, C., & Kumar, A. (2017). Social eWOM: does it affect the brand attitude and purchase intention of brands?. Management Research Review.

Leeflang, P.S.H., Wieringa, J.E., Bijmolt, T.H.A. & Pauwels, K.H. (2015). Modeling Markets. Springer.

Libai, B., Bolton, R., Bügel, M. S., De Ruyter, K., Götz, O., Risselada, H., & Stephen, A. T. (2010). Customer-to-customer interactions: broadening the scope of word of mouth research. Journal of service research, 13(3), 267-282.

Luo, X. & Homburg, C. (2008). Satisfaction, complaint, and the stock value gap. Journal of Marketing, 72(4), 29-43.

Luo, X. & Zhang, J. (2013). How do consumer buzz and traffic in social media marketing predict the value of the firm? Journal of Management Information Systems, Vol. 30 No. 2, pp. 213-238.

Mangan, D. (2020, May 22). Trump doesn’t wear coronavirus mask in public at Ford plant. Retreived January 07, 2021, from https://www.cnbc.com/2020/05/21/trump-doesnt-wear-coronavirus-mask-to-ford-plant.html

Markus, G. B. (1979). Analyzing panel data (No. 18). Sage.

Meyersohn, N. (2018, September 13). Young people support Nike’s bet on Kaepernick, poll shows. Retrieved January 08, 2021, from

You’re Cancelled: The Power of Online Storms on the Stock Performance of Firms

You’re Cancelled: The Power of Online Storms

on the Stock Performance of Firms

Based on Twitter data from 2015-2020

You’re Cancelled: The Power of Online Storms

on the Stock Performance of Firms

Based on Twitter data from 2015-2020

Master Thesis

Msc. Marketing Intelligence & Marketing Management

Faculty of Economics and Business

University of Groningen

Supervisor: Dr. Evert de Haan

Second Supervisor: MSc. Hidde Smit

Annebeth Seraphine Sterre Meulenberg

Abstract

Preface

Contents

1. INTRODUCTION

2. LITERATURE REVIEW

3. HYPOTHESES

4. DATA

𝐸𝐸(𝑅𝑅

) = 𝑅𝑅

+ 𝛽𝛽

(𝐸𝐸(𝑅𝑅

) − 𝑅𝑅

) (1)

𝐴𝐴𝑅𝑅

= 𝑅𝑅

− 𝐸𝐸(𝑅𝑅

) (2)

5. METHODOLOGY

6. RESULTS

7. DISCUSSION

8. LIMTATIONS AND FURTHER RESEARCH

REFERENCES