Does Social Media Moderate Information Bias in China? Evidence from the Tone of Social and Traditional Media

Eric Wang

The Chinese University of Hong Kong, Shenzhen

T.J. Wong

University of Southern California

Tianyu Zhang

The Chinese University of Hong Kong

September 2019

Abstract

This paper examines whether social media moderates the information bias in the market by generating less optimistic information when the state-controlled traditional media is positively biased. Using a comprehensive sample of corporate news articles of Chinese newspapers and posts of an online stock forum, East Guba, from 2009 to 2016, we find that East Guba’s tone is less positively associated with that of the newspapers for the same firm on the same day when the newspapers are expected to be more optimistically biased. This decline in association in tone is significantly larger since the 2015 political shock which increased the suppression of negative corporate news by traditional media. Finally, when the tone of the newspapers deviates positively from that of East Guba, the newspaper articles are perceived by the market to be less credible as reflected in the significantly attenuated stock return response.

Eric Wang is assistant professor of Accounting at the Chinese University of Hong Kong, Shenzhen; T.J. Wong is professor of Accounting at the University of Southern California; Tianyu Zhang is professor of accounting at the Chinese University of Hong Kong. We appreciate comments from conference participants of “Text Analysis for Asia and Beyond” Conference at USC, ABFER annual conference of 2019, and MIT-Asia conference of 2019.

37

1. Introduction

Traditional media in autocratic regimes are optimistically biased. The autocrats control the media and bias the news they report as a way to strengthen political power (Lipperman, 1953; Noelle-Neumann, 1984; Djankov et al., 2003; Enikolopov et al., 2011). This optimistic bias in reporting is found in corporate news as well since negative news can potentially destabilize the economy and weaken the perceived competency of the government (Stockmann, 2013; Piotroski et al., 2017). However, in these autocratic regimes, many newly developed social media such as online stock forums have started to supply information to the markets. In contrast with traditional media that typically have state-controlled editorial teams, social media are typically private, and they rely on the wisdom of crowds to generate information in anonymity. Users that contribute corporate information to these online platforms are expected to be independent, as they are more likely to be free from direct censorship by the state (Foucault, 1977; Spears and Lea, 1994).

The objective of the paper is to examine whether social media moderates the information bias in the Chinese market by generating less positively biased information when state-controlled traditional media is optimistically biased. We posit that by generating less biased information, the newly established social media can balance the tone bias of corporate news in the market introduced by traditional media. In our tests, we compare the tone of social media and traditional media for the same firm on the same day. This allows us to restrict the information generated by the two types of media for each firm to be based on the same underlying events of the day, while enabling us to capture any differences in tone due to the media’s linguistic bias (e.g., words or phrases) and/or coverage bias (e.g. events or aspects of the events of the day).[footnoteRef:1] [1: A newspaper can engage in coverage bias by choosing to report positive events or the positive aspects of an event. Thus, the tone of the news report can misrepresent the true impact of the underlying event(s) to the firm. ]

For the main analysis, we regress the tone of social media on the tone of traditional media. The coefficient on the tone of traditional media will capture the relative tone of the two types of media for the same firm on the same day. A larger (smaller) positive coefficient will suggest a narrower (wider) deviation in tone between the two types of media. We posit that when traditional media is expected to be more optimistically biased, the tone of social media will be less positively associated with that of traditional media, as reflected by a significant decrease in the coefficient on the tone of traditional media.[footnoteRef:2] This decline in positive association in tone indicates that social media offers a more independent source of information that moderates the positive bias introduced by traditional media. [2: We use whether the tone of traditional media is positive or negative to proxy for their optimistic bias. As discussed in subsection 2.3, the media’s tone is expected to be more optimistically biased when it is positive than when it is neutral or negative. ]

However, finding a decline in association in the first analysis can also be interpreted as social media reporting noise rather than moderating the bias in the market. That is, the weakened association can result from noise in social media and not from optimistic bias in traditional media. To test whether our interpretation or this alternative explanation is valid, we use the stock market response to the two types of media as our second analysis. First, we posit that if social media generates less optimistic information to moderate the positive bias in the market, its posts can also serve as a benchmark to delineate the bias of traditional media. That is, the difference in tone between traditional and social media for the same firm on the same day can capture the optimistic bias of traditional media. Second, we use the stock return response to the tone of traditional media as a gauge for the market perception of the tone bias. We posit if social media is indeed providing less biased information rather than noise, the stock return response to traditional media’s tone will significantly decrease when it deviates more positively from that of social media. Thus, we use the stock returns to validate whether social media can serve to balance the information bias in the market.

We expect that social media can moderate information bias in an autocratic regime. The state can control public opinion by controlling traditional media. In addition to direct censorship, government control will discourage the public from voicing their views, giving rise to the spiral of the silence of the majority (Noelle-Neumann, 1984). Contrasting with traditional media, computer-mediated communication, such as social media, can free their users from feeling pressured to conform and give in to political or social power, and allows them to express their true opinions anonymously (Spears and Lea, 1994). As compared to editors and journalists in a newspaper, it is much harder to control the crowd on social media, who remain anonymous. This gives rise to the business model of many online platforms that rely on the wisdom of crowds to provide valuable information to the community (Chen et al., 2014; Bartov et al., 2018). Though social media sites can be censored, the government tends not to police them as much because it regards them as a source of grass-root information (Qin et al., 2017) Also, stock opinions are seldom considered threatening since the government is mainly concerned about instigation of collective actions (King et al., 2013, 2014).

On the other hand, social media may not serve the role in moderating information bias. Computer-mediated communication may give rise to uninhibited behavior, leading to polarized and extreme views that are uninformative (Spears and Lea, 1994). Thus, social media posts are primarily noise, despite being relatively free from positive bias. It is also possible that social media sites are subject to heavy censorship. The government may hire agents to post biased messages or delete critical messages on social media sites. This is particularly true in politically sensitive periods, during which the government is concerned about stability and is extremely sensitive to negative news.[footnoteRef:3] [3: The government intervened the 2015 market crash in China by requesting state-owned brokerage firms and mutual funds to prop up the market (http://www.xinhuanet.com/fortune/2015-08/06/c_128100347.htm). ]

China offers a great setting to study whether the wisdom of crowds of social media can moderate media bias in corporate news. Although the equity market was only established in the early 1990s, it has grown to have more than 3,000 firms listed in its two domestic stock exchanges and become the world’s second-largest equity market by the total capitalization as of 2018. Our sample of corporate news from Wisenews database provides comprehensive coverage of 74 state-owned newspapers spanning across most of the provinces in China. For social media information, we use the East Guba internet platform, which is an online forum for stock opinions and analyses. During our sample period from 2009 to 2016, users of East Guba have posted 146 million posts (about 83,000 posts per trading day), covering around 3,000 listed firms in China. Our final sample comprises more than 721 thousand firm-days with at least one news article from newspapers and three posts from East Guba for each firm on the same day.

Following the machine learning method in Piotroski et al. (2017), we compute the tone of each traditional media article and social media post. We then aggregate the tone of each type of media for each firm on each day for comparison. Our evidence shows that the tone of traditional media is generally positive (77.4% positive) with a mean of 0.3756, confirming prior research that traditional media is optimistically biased, while the tone of social media is mostly negative (13% positive) with a mean of -0.2038.[footnoteRef:4] [4: This is comparable to the measurements using “StockTwits”, a US social media. For example, Byard and Wang (2018) report a mean (median) tone of -0.326 (-0.405) when applying similar textual analysis method to StockTwits dataset. ]

To examine whether social media provides less optimistically biased information to the market than traditional media in China, we regress the tone of social media posts on the tone of traditional media’s articles of the same firm on the same day. We find that the coefficient on the tone of traditional media is significantly positive at 1% level, suggesting that the tone of the two sources of corporate news is positively correlated. We also find that the coefficient on the tone of traditional media is significantly reduced by 0.0269 when the tone of traditional media is positive, which is 54% of the coefficient when the tone of traditional media is zero or negative. This is consistent with our prediction that when traditional media are likely to be more positively biased (i.e., having a positive tone), their tone will be less positively associated with that of social media.

Next, we exploit two political events to test if political incentives affect traditional media’s optimistic bias and whether and how social media can moderate the bias. The first political event is an exogenous shock to the traditional media since the government made a strong push to suppress the press from reporting negative news during the 2015 stock market crash.[footnoteRef:5] We find that after the shock, there is a decline in the association of tone between traditional and social media when the former is expected to be more optimistically biased. This suggests that the political shock has a significant effect on traditional media’s incentives to bias the tone upwards while social media remains to be less influenced by political pressure. [5: There is anecdotal evidence that the government pressures the traditional media to be optimistic in its reporting during the stock market intervention (http://www.chinadaily.com.cn/micro-reading/interface_yidian/2015-06-26/13895012.html). ]

The second political event is the National Congress Meeting, which is held every five years. We find that social media moderates the information bias only during the non-National Congress Meeting period and not during the National Congress Meeting period. Additional analysis shows that during the period of the National Congress Meeting (from 45 days before to 45 days after), the number of posts drops significantly while the number of traditional media articles increases significantly. One possible interpretation is that during politically sensitive periods, even social media refrain from posting critical messages, while traditional media are pressured to report more positive news.[footnoteRef:6] [6: This is consistent with the finding in Piotroski et al. (2015) that Chinese firms suppress the release of bad news in the year around the National Congress meetings. ]

Finally, we study whether traditional media with a higher optimistic bias as delineated by social media is associated with a weaker stock return response. First, we partition our sample into terciles based on the divergence in tone between traditional and social media. We find significantly positive stock return responses to the tone of the traditional media articles at five-day, ten-day and twenty-day CAR windows in the low tercile (when the divergence is low). These findings suggest that traditional media articles contain information valuable to the market when there is less optimistic bias.[footnoteRef:7] Second, our evidence shows that the positive stock return response to the traditional media is monotonically decreasing in the divergence in tone between tradtional and social media, with a lower response in the middle tercile and a statistically insignficant response in the top tercile. Third, in stark contrast to traditional media, stock return responses to social media tone remain positive and significant across the three terciles in all the CAR windows. Most importantly, the stock response to social media is significantly positive even when the divergence in tone between traditional and social media is the largest (top tercile), refuting the concern that social media may be just reporting noise when traditional media is optimistically biased. These results provide external validations that social media can help to reveal traditional media’s optimistic bias, thereby serving the role in moderating the information bias in the market. [7: We cumulate the stock returns from the day after (day +1) the news/opinions dissemination by the traditional or social media (day 0) to ensure that our result is not capturing the media’s reaction to the stock price movement on day 0 due to reverse causality. ]

Our study contributes to the literature in the following ways. First, prior research focuses mainly on the information generation and dissemination roles of social media in democratic regimes (e.g., Blankespoor et al., 2013; Chen et al., 2014; Bartov et al., 2018). This is the first paper we know of that studies whether, in an autocratic government, social media moderates information bias in the market by generating less biased information. Our results show that in China, social media’s reliance on the wisdom of crowds can shelter themselves from government’s direct intervention and provide corporate news that can serve as a benchmark against the positive bias of traditional media.

Second, we provide new evidence that in China the market can discount at least partially the positive bias of traditional media and will discount the information when social media reveals traditional media’s bias or moderates the information bias in the market. This extends prior research that newspapers are optimistically biased for economic or political reasons, in either democratic or autocratic countries (Gurun and Butler, 2012; Solomon, 2012; Stockmann, 2013; Piotroski et al., 2015). Our paper provides evidence that the market will make adjustments to the bias when responding to the news of traditional media.

Third, we are one of the few studies that find that social media posts have information content (see also Tumarkin and Whitelaw, 2001; Antweiler and Frank, 2004; Tetlock et al.2008, Chen et al., 2014). Different from the results in Chen et al. (2014) that are based on U.S. data, we find that the social media posts have strong positive stock return response in both the shorter (five-day) as well as, the longer (twenty-day) windows, suggesting that they contain more information in an emerging market.

The rest of the paper is organized as follows. Section two provides a discussion of the institutional background and hypothesis development. Data and sample are presented in section three, and our results are presented in section four. We provide a number of additional tests and robustness tests in section five and conclude the paper in section six.

2. Institutional Background and Hypothesis Development

2.1 Development of the newspaper industry in China

Since the economic reforms in 1978, China’s newspapers have had to balance between two conflicting objectives. While continuing to serve as a government’s mouthpiece to maintain political stability, newspapers also play the role of a key information institution to support market reforms. In addition to serving primarily as a propaganda machine for the state, China began to set up commercialized newspapers that depend on advertising and subscription revenues for funding and respond to market demands for news. This commercialization reform increased the number of newspapers from 186 to 1,943 between 1978 and 2009 (Stockmann, 2012).

Despite such large-scale news media reform, the government retains tight control of the newspapers. None of the newspapers can be majority-owned by non-state entities. There is no press freedom in China as the government continues to control the country’s newspapers through direct ownership control and the appointment and dismissal of the senior editorial staff. The Propaganda Departments of the Chinese Communist Party (CCP) committees (at every level of government) also exerts influence on the reporting behavior of newspapers. To further signal the commitment to control news media, the government elevated the news media’s regulatory and licensing agency, the General Administration of Press and Publication (GAPP), to ministerial-level status.

2.2 Development of social media in China

China has a population of more than 700 million netizens by 2016 (All-China Journalists Association, 2017). This deep internet penetration facilitates the development of social media, as a main source of news for netizens. Social media platforms are also set up by a number of financial service firms to provide financial and investment information to their clients. Many of them are in the form of bulletin board system (BBS), which allows clients to access user-generated content (UGC) and interact with each other (e.g., Hexun (http://www.hexun.com/index.htm), Jinrongjia (http://www.jrj.com.cn), and StockStar (http://www.stockstar.com)).

East Guba, which is a BBS platform launched by EastMoney in 2006, has become the top-ranking online stock platform in China. For example, the total number of posts on Wanke, a real estate company listed in Shenzhen Stock Exchange, has reached more than 360,000 on East Guba as of 2016, more than 10 times higher than Hexun, the second-most-popular social media for stock information. Similar to many other social media platforms, East Guba is privately owned. Unlike Twitter, however, East Guba does not impose any limitation on the post length. They are also different from Seeking Alpha, a popular U.S. online stock forum, in that their users’ posts are not screened by editors like those in Seeking Alpha.

2.3 Hypothesis Development

The Chinese government has the ability and incentives to bias corporate news of traditional newspapers. As an autocratic regime, the government has never granted press freedom to newspapers. Also, local government leaders’ performance is evaluated based on the economic performance of the regions under their control (Li and Zhou, 2005). Thus, there are strong political incentives to influence corporate news because listed firms’ performance in a region can be used as a signal to shape public opinions about local leaders’ achievements. Piotroski et al. (2015) find that listed firms suppress the release of bad news prior to the promotion of the political leaders of their provinces. Likewise, they find that listed firms refrain from reporting bad news around the National Congress Meetings in order to avoid any embarrassment to the central government during this momentous political event. Optimistic bias is also found in corporate news of newspapers by Piotroski et al. (2017).

It is unclear whether corporate information generated by social media such as online stock forums shares newspapers’ optimistic bias in China. On the one hand, private ownership, user-generated content, and user anonymity can help to shelter social media’s information from government control. On the other hand, China is famous for its unsophisticated retail investors dominating the stock market. The information produced by amateur investors in these media platforms can be noisy. In addition, the Chinese government has increased its control over the internet in recent years.[footnoteRef:8] The 2015 stock market intervention has led to a tight control against any public release of negative information that destabilizes the market. The government’s tightening of social media censorship may increase online stock forums’ optimistic bias. [8: Two recent news articles provide discussion on how the government has increased its control of the internet: “The great firewall of China: Xi Jinping’s crackdown,” The Guardian, June 29, 2018 (https://www.theguardian.com/news/2018/jun/29/the-great-firewall-of-china-xi-jinpings-internet-shutdown); “China has launched another crackdown on the internet – but it’s different this time,” CNBC, October 26, 2017 (https://www.cnbc.com/2017/10/26/china-internet-censorship-new-crackdowns-and-rules-are-here-to-stay.html). ]

Thus, whether social media can moderate the media bias in China remains an empirical issue. To test this, we first examine if social media will provide less biased information to the market when traditional media is expected to be more optimistic. Specifically, we examine the tone of Chinese newspapers’ corporate articles relative to that of East Guba’s posts. We posit that when traditional media are more positively biased, the tone of the two media will become less correlated. Our first hypothesis is as follows:

H1: The tone of social media has a weaker association with that of traditional media for the same firm on the same day when the tone of traditional media is positive than when it is neutral or negative.

In this hypothesis, we use whether the tone of traditional media is positive vs. neutral or negative to proxy for its optimistic bias. That is, we expect that the newspaper articles are more optimistically biased when their tone is positive than when it is negative. We make this assumption because China’s state-owned newspapers typically have a strong optimistic bias that can significantly affect the overall tone of the articles.[footnoteRef:9] Our sample shows that 77.4% of the newspaper articles as compared to only 13% of East Guba’s posts have a positive tone, which suggests that newspapers’ optimistic bias is large and is likely to have shifted the tone of the articles to be mostly positive. With such a significant slant in tone, we expect that a large portion of the bias comes from coverage bias (i.e., primarily reporting positive news topics and dropping negative news topics) and not just from linguistic bias. Notwithstanding this explanation, it remains a joint hypothesis that traditional media’s tone is a good proxy for optimistic bias and social media plays a role in moderating media’s information bias. [9: A validation test for this assumption is discussed in subsection 4.1 and Appendix III. ]

Next, we use the stock return response to traditional and social media to validate the social media’s role in moderating the information bias in the market. To test this, we posit that the market’s response to the tone of traditional media will be attenuated when the tone of the traditional media deviates more positively from that of social media. We use the traditional media’s perceived credibility, captured by the stock return response, to validate social media’s ability to delineate traditional media’s positive bias. This is a joint hypothesis that social media can serve as a benchmark to identify the tone bias of traditional media, and the market discounts its response to the tone of traditional media articles when it is positively biased. Our second hypothesis is as follows:

H2: The stock return response to the tone of traditional media is significantly reduced when it deviates positively from the tone of social media.

3. Data and Sample

3.1 Data from traditional media

Following Piotroski et al. (2017), we crawl all available news articles of the listed companies published from 2009 to 2016 in Newswire with an automatic crawler using the company names. We improve the textual analysis techniques used in Piotroski et al. (2017) by cleaning up ambiguous firm names and improving the measurement of the tone. First, we take additional steps to avoid crawling the wrong articles for companies with ambiguous names. For example, a large number of articles are crawled for a company named Laobaixing (老百姓). However, Laobaixing also means “ordinary people” in Chinese, and many articles that contain these words are not about this particular company. We identify a list of companies with ambiguous names and then manually read a set of articles for each name as a training set for identifying whether the articles are talking about that particular company. We then apply the machine learning algorithm to identify the articles for companies with ambiguous names.

Second, in order to further improve the accuracy of the tone classification algorithm, we increase the number of training articles from 10,000 used in Piotroski et al. (2017) to 50,000 for the updated sample, from 10,000 used in Piotroski et al. (2017). That is, we hire two batches of research assistants to label the tone of each sentence of 50,000 articles randomly picked from our sample as negative, positive, and neutral. Their labels of the tone are later cross-checked among different research assistants to verify their validity[footnoteRef:10]. Using these manually labeled training materials, we train a support-vector-machine (SVM) model to classify each sentence into positive, neutral or negative and check the out-of-sample classification accuracy using a subset of manually labeled sentences that the model has not seen. The out-of-sample validation using 10,000 randomly-selected sentences shows that the accuracy rate of our model is above 90%. [10: Our manual coding method follows Piotroski et al. (2017). We randomly assigned nine research assistants to three groups. Each group was assigned 50,000 news articles randomly sampled from our full sample of 1.77 million corporate news articles. For each group, the sampled news articles were uploaded to a specially designed website on which the research assistants were required to read each article sentence by sentence and independently record a judgment about the tone of every sentence. We applied a majority principle to determine the sentiment of each sentence according to the judgments of the three research assistants in the group, and the recorded key words were labeled with the same sentiment as the sentence.]

Next, we aggregate sentence-level tone to form measures of article-level tone. The tone of the article is measured by the relative weight of positive sentences to negative sentences in the article. In addition, we also consider the importance of sentences from different positions within an article. That is, we weigh the sentences from the first and last paragraphs as 2, the first and last sentences of the first and last paragraph as 3, and other sentences from the article as 1. The tone of the body of the article equals (#of positive sentences-# of negative sentences)/(#of positive sentences+# of negative sentences+1). The overall tone of the article, in the end, is defined as (tone of text body*0.7+tone of title*0.3).

Our final traditional media news dataset includes about 1.92 million news articles published over the period of 2009 to 2016, by 74 unique newspapers located in 23 provinces. We removed newspapers with a focus on leisure, life, and entertainment because they are less relevant to investors or companies. Appendix I reports a detailed list of newspapers in our sample.

3.2 Data from Social Media

The Guba platform (aka East Guba) on East Money (http://guba.eastmoney.com) provides a separate discussion board for each listed company. East Guba is one of the oldest and most influential social media platforms with a focus on the capital market.[footnoteRef:11] Users can enter and post on the board by searching the stock code or the name of the company. They can also build up a self-defined list of companies by following them and entering the discussion board of the companies on the list by a direct click. Thus, the posts on this platform are well-matched with the related companies. We develop a web crawler to download all the main posts on the discussion board of each company.[footnoteRef:12] Because the posts on social media are usually short; we thus label the tone at the post level rather than sentence level. Also, we use a different machine learning training set for social media posts due to their different writing style from that of traditional media. Since Emoji is often employed by users to express their opinions, we incorporate emoji as part of the training set. [11: Weibo is the most influential social media platform but it lacks a focus on capital markets. Snowball, another social media platform, was established much later than East Guba and is likely to have more active professionals such as analysts and institutional investors on it. ] [12: We ignore the reply comments on the platform because most of these comments are short and may add noise to our analysis.]

We define the tone of social media for a firm as the relative weight of the number of positive posts and negative posts, (# of positive posts-# of negative post)/ (# of positive posts+# of negative post+1), which is analogous to using the tone of sentences in the body of the text to compute the tone of traditional media articles. Our final social media dataset includes about 31.1 million posts, covering 3,039 firms from 2009 to 2016.

3.3 Sample Selection Process and Descriptive Analyses

We start by obtaining all available firm-day daily stock return observations from 2009 to 2016, for companies listed on China’s Shanghai and Shenzhen Stock exchanges from the China Stock Market & Accounting Research (CSMAR) database. We match the firm-day sample with our traditional media tone and social media tone as discussed previously and then match firm-day stock return observations with firm fundamentals in their last year’s annual reports. Because our empirical analyses focus on the relationship between social media and traditional media on a firm-day level, we include only firm-day observations with at least one traditional media news articles and at least three social media posts. We require a firm-day to have at least three social media posts in order to avoid measurement errors of the social media tone caused by insufficient posts per day.[footnoteRef:13] After merging our social media and traditional media file with CSMAR dataset, our final sample consists of 721,213 firm-day observations, covering 2,942 unique firms. Detailed sample selection process is reported in Table 1, Panel A. [13: Results are robust when we increase the threshold to five posts per day and ten posts per day, despite a reduced sample size due to the increased restrictions. ]

Table 1, Panel B provides descriptive statistics of traditional media tone and social media tone. On average, a firm in our sample is covered by 2.66 traditional media news articles and 43.18 social media posts per day. The tone of traditional media has a mean (median) of 0.3756 (0.4892), and the tone of social media has a mean (median) of -0.2038 (-0.1840). Also, traditional media tone is positive in 77.4% of the sample, while social media tone is positive in only 13% of the sample. Taken together, the descriptive evidence suggests that on average the tone of traditional media is more positive than the tone of social media.

The correlation between the tone of traditional and social media is shown in Table 1, Panel C. The tone of the two types of media is positively correlated at 0.0712, significant at the 5% level. The correlation between the tone and CARs, which measure the stock return responses to traditional and social media (see Appendix II for definitions), is also significantly positive for both types of media. However, the magnitude of CARs’ correlations with social media is greater than that with traditional media, suggesting that the former has more useful information content and is likely to be less biased. The sample distribution of the firm-day observations by year is reported in Table 2, Panel A. The number of firm-days are slightly smaller in 2015 and 2016, but the distribution is generally even over the sample period. The sample distribution by industry, presented in Table 2, Panel B, shows that most of the industries are well-represented in our sample.

4. Empirical Analyses

4.1 Traditional Media Bias and Social Media’s Moderating Role

Our primary empirical analyses examine the relationship between the tone of social media and the tone of traditional media covering the same firm on the same day, and how this relationship varies when traditional media tone is positive than neutral or negative. We estimate the following model with OLS regression using firm-day data:

(1)

In these estimations, is a dummy variable equal to one (and zero otherwise) if is greater than zero. Coefficient captures the unconditional association between the tone of social media and the tone of traditional media. Coefficient of the interaction term captures the change in association between the tone of social media and traditional media when the tone of traditional media is positive. Because both social media and traditional media provide relevant information about a company’s fundamentals (Bartov et al. 2017), we expect to be positive and significant. We expect to be negative and significant because when traditional media are more optimistically biased (as proxied by the positive tone), the tone of social media will be less positively biased relative to traditional media.

We estimate the model with a group of firm-level control variables including Size, ROA, Market to Book and Leverage to capture the variation in firm characteristics that could affect the bias in traditional media, and the relationship between the tone of social media and the tone of traditional media. All variable definitions are presented in Appendix II. However, even with these control variables, there is still concern that our results could be driven by correlated omitted variables. To alleviate this concern, we include firm-fixed effect and year-month fixed effect to absorb further unobservable firm characteristics and time-variant characteristics that potentially affect the tone of both social media tone and traditional media. In all the regressions, we report in parentheses t-statistics with robust standard errors two-way clustered at the firm level and year-month level.

The regression results of model 1 are reported in Table 3. Columns 1 shows a positive association between the tone of traditional media and social media, indicating that on average traditional media tone is positively associated with social media tone. In columns 2 and 3, the coefficient estimates on traditional media tone continue to be positive and significant, and the coefficient estimates on are both negative and significant across both model specifications. The negative coefficients on the interaction term suggest that the positive association between traditional media tone and social media tone is reduced when traditional media is positive. In terms of economic significance, in column 3 the coefficient on the tone of traditional media is significantly reduced by 0.0269 when the tone of the traditional media is positive, which is 53.8% of the coefficient when the tone of the traditional media is zero or negative.[footnoteRef:14] [14: The coefficient on Traditional Media Tone is low because on a typical firm-day, there may not be any important underlying event to anchor the news and the two types of media can be reporting on different topics. When we increase the requirement of having more traditional media articles on each firm-day, it increases the chances that the firm-day has important underlying event(s) and that the two types of media are reporting on the same topics. We find that the coefficient on Traditional Media Tone (untabulated) increases to 0.134 (0.156) when we require more than three (five) articles per firm-day. ]

As a validation test for our assumption that the newspapers’ bias is more optimistic when their overall tone is positive than when it is negative, we examine the relationship between the newspapers’ tone and the optimistic bias. Since our results in Table 3 suggest that the tone of social media can serve the role in moderating the positive bias of traditional media, we use the positive deviation in tone between the newspapers and East Guba to estimate newspaper articles’ optimistic bias. Using this measure, we show in Appendix III that when we rank the tone of newspaper articles into deciles, the optimistic tone bias of the newspapers is monotonically increasing in the tone of the newspaper articles. This supports our assumption that traditional media with a more positive tone is more optimistically biased.

Overall, our baseline regression results show a positive association between traditional tone and social media tone of the same firm on the same day, and this association is attenuated when traditional media tone is positive, consistent with our first hypothesis that social media serves as a benchmark to moderate the optimistic bias of traditional media.[footnoteRef:15] [15: A possible alternative explanation is that there are fewer important underlying events on days when traditional media tone is positive. Thus, tradition and social media are reporting on different topics, which causes the decline in association and the negative coefficient on Traditional Media Toneit x TM Positiveit. However, when we increase the chances of important underlying events by requiring each firm-day to have more than three (five) articles (see footnote 14) or the firm-day to be around earnings announcement and earnings guidance dates (subsection 5.2), the coefficient on Traditional Media Toneit x TM Positiveit remains significantly negative. ]

4.2 Influence of Political Incentives on the Bias Moderating Role of Social Media

Prior literature shows that traditional media bias varies with political incentives of the government (Piotroski et al. 2015; Gentzkow et al. 2006). In this subsection, we exploit China’s recent political events to test if changes in the association of the tone of traditional and social media reflect how the two media respond to changes in political incentives.

4.2.1 Media Intervention in 2015 Stock Market Crash

First, we use an open and sudden media intervention in 2015 as an exogenous shock for our tests of social media’s moderating role of the media bias. In 2015, the Chinese stock market experienced a historic crash with a third of the market value in the Shanghai Stock Exchange being destroyed in the month of June. This was followed by 1,400 companies (more than half of all the listed firms) filed for a trading halt in early July in an attempt to avoid further losses. To stabilize the turbulence in the stock market, the National Bureau of Television, Broadcast, and Newspaper (a bureau under GAPP) issued an authoritative order to all press in China, demanding them to decrease coverage of the stock market, and stop using negative words such as “tumble” or “crash” in the news. The Chinese press responded to this authoritative order by reducing the reports of negative news, resulting in an increase in a positive tone of corporate news. We expect social media’s moderating role to be stronger in the period following this intervention as the positive bias of traditional media is further amplified.

To formally test the impact of such intervention, we partition the full sample into the pre-intervention period which includes all the sample data on or before July 23rd, 2015, and the post-intervention period after this date. We estimate the baseline regression (Model 1) separately in these two sub-samples and test for the difference in regression coefficients using the Chow test. Results are reported in Table 4. The coefficient on Traditional Media Tone association between traditional and social media tone is not statistically different in the pre-intervention period and post-intervention period, but the reduction in the association when traditional media is positivecoefficient on Traditional Media Tone x TM_Postive ishas significantly increased more negative in the post-intervention period. Specifically, in the pre-intervention period, the reduction in association between traditional media tone and social media tone is about 50% (0.0254/0.0508) when traditional media tone is positive, compared to a reduction of about 87% （0.0322/0.0370） in the post-intervention period. Taken together, the results suggest that the 2015 media intervention, as an exogenous political shock, has increased the positive bias in traditional media’s coverage of corporate news. Also, social media can maintain its role in balancing traditional media’s bias, suggesting that they are sheltered from the government’s increased intervention after July 2015.

4.2.2 National Congress Meeting Period versus Non-National Congress Meeting Period

Another political event we exploit is the 18th (2012) CCP National Congress Meeting, which is an important political event held every five years in Beijing. During this period, the government intensifies its monitoring over traditional media to maintain a stable social environment. Our untabulated evidence shows that around the 18th CCP National Congress Meeting, traditional media tone has a positive deviation from that of social media, with a range of 0.4455 to 0.5316, from the three months before the meeting, during the month of the meeting, and to three months after the meeting. Although this is an expected (recurrent) event, we want to examine if this heightened political intervention will affect social media’s role in moderating information bias.

We estimate the same regression model in our baseline analysis (Model 1) separately in the two sample partitions: Congress Meeting sample which includes a window of -45 to +45 days relative to the opening day of the 18th CCP National Congress Meeting, and Non-Congress Meeting sample which includes the rest of the firm-day observations. Column 1 of Table 5 Panel A shows that there is no significant association between the tone of social media and the tone of traditional media around the Congress Meeting sample. Also, the attenuation in the association between the tone of social media and traditional media disappears during this same period. However, column 2 shows similar results to our main finding in Table 3, suggesting our baseline results mainly concentrate in the Non-Congress Meeting period. One possible interpretation of the results is that social media can moderate media bias during Non-Congress Meeting periods but not during Congress Meeting periods.

Next, we explore the potential causes for the loss of social media’s ability to moderate media bias during the Congress Meeting period. In Figure 12, we show univariate evidence that the number of social media posts experienced a gradual decline starting from about six months ahead of the Congress Meeting, reached the lowest level during the meeting and started a slow recovery to reach its normal level in about three months following the meeting. However, during the entire time, the number of news articles published by traditional media for a firm-day remains mostly unchanged.

Table 5, Panel B provides results on the level of activities for traditional media and social media during the Congress Meeting period. We estimate a pooled OLS regression model to determine the impact of the Congress Meeting on the level of activities proxied by the number of social media posts and the number of news articles. The model controls for firm characteristics (Size, ROA, Market to Book, Leverage), daily returns (stock returns and index returns), news from other channels (number of news, and number of social media posts as control variables) and a firm dummy to capture unobserved firm-level characteristics. Our results show that the number of corporate news coverage on a firm-day by traditional media increases by an average of 0.144 (5.5% increase) articles during Congress Meeting period, while the number of social media posts decreases by an average of 21.34 posts (49.4% decrease) during the same period. This finding provides possible explanations for the disappearance of social media’s ability in moderating the bias. That is, during politically sensitive periods such as the National Congress Meetings, even social media users refrain from posting critical (negative) messages while the traditional media is pressured to report more positive news.

4.3 Stock Return Responses to the Tone of Traditional Media and Social Media

Prior studies have investigated the stock return responses to traditional media tone and social media tone: Tetlock et al. (2008) focus on traditional media and find that the fraction of negative words in firm-specific news articles is associated with the same day and following day abnormal returns. Chen et al. (2014) find that opinions transmitted through Seeking Alpha are associated with abnormal stock returns from 3 to 60 days after the posts. Bartov et al. (2018) show that Twitter contains information that is useful in predicting earnings announcement period returns. However, to the best of our knowledge, there is little evidence on whether there are significant stock return responses to traditional media and social media in China. Given China’s unique media system (Piotroski et al. 2017; Qin et al. 2018) and institutions, empirical findings using U.S. data may not be generalizable to China’s setting.

In addition, existing literature examines the information content of social media and traditional media independently. We add to the existing literature by jointly analyzing the relative association between the tone of traditional media articles and social media posts and their corresponding future abnormal stock returns. More specifically, we want to use the relative stock return response to traditional and social media as a way to gauge if social media can serve as a benchmark against the positive tone bias of traditional media. We expect that the market will discount its response to the tone of traditional media when it deviates positively from that of social media. This provides support to our conjecture that social media can delineate the positive bias of traditional media, thereby moderating the information bias in the market.

Specifically, we use regression analyses to test 1) whether the tone of traditional media and tone of social media are associated with future abnormal returns at the firm level, and 2) when the tone of traditional media deviates from that of social media (i.e., more positively biased), whether the association between the tone of traditional media and future abnormal returns is attenuated. To capture the level of traditional media bias, we partition the full sample based on the difference between traditional media tone and social media tone (Traditional Media Tone – Social Media Tone) into terciles: TM Bias LOW, TM Bias MEDIUM, and TM Bias HIGH.

We estimate the following Model 2 with year-month fixed effect separately in the three subsamples. To reduce the likelihood of reverse causality that news articles or social media posts are reacting to extreme abnormal returns, all three CAR windows start from day +1 relative to day 0, which is the day the news articles or social media posts are reported. CAR is the equal-weighted market-adjusted cumulative abnormal returns computed over five days, ten days and twenty days starting with day +1 (CAR(1,5), CAR(1,10) and CAR(1,20)), capturing the short-term and longer-term impact of the media. Following Tetlock et al. (2008), we include CAR of the last week to control for momentum in price movement, and standard errors are double clustered at the firm and year-month level. Our model is as follows:

(2)

Table 6, Panel A reports the regression result for traditional media for CAR(1,5) in columns 1-3, CAR(1,10) in columns 4-6 and CAR(1,20) in columns 7-9. Within each CAR window, we report the regression results for each of the terciles. In all the three CAR windows, the tone of traditional media is positively correlated with future CAR in TM Bias LOW, suggesting that the market reacts strongly to the tone of traditional media when its deviation from social media is low (i.e., less biased). In terms of economic significance, an increase in the tone of traditional media tone from zero (neutral) to one (most positive) is associated with an average increase of 15 basis points in the cumulative abnormal returns over the five trading days after day 0 (column 1). For the TM Bias MEDIUM tercile, the correlation between traditional media tone and future CAR becomes lower for CAR(1,5) and CAR(1,10) and insignificant for CAR(1,20). Traditional media tone is no longer significantly correlated with future CAR in all three windows for the TM Bias HIGH tercile. The F-test results show that the coefficient on traditional media tone is significantly lower in TM BIAS HIGH than in TM BIAS LOW, for CAR(1,5) with p=0.07, and CAR(1,10) with p=0.01. The significant decrease in the correlation between traditional media tone and CAR when traditional media tone deviates from social media tone supports our hypothesis that social media can delineate traditional media’s bias.

We further examine whether the stock return response to the tone of social media is also affected by its deviation from the tone of traditional media.[footnoteRef:16] In Table 6, Panel B, results show that the stock return response to social media is significantly positive for all terciles across the three CAR windows. More importantly, in contrast with the CAR response to traditional media in Panel A, the CAR response to social media is significantly positive for the TM Bias HIGH, suggesting that when social media deviates significantly negatively from traditional media, the market continues to find there is information content in social media but not in tradition media. This evidence provides further support that social media’s information can serve as a benchmark to delineate traditional media’s positive bias. It also rules out an alternative explanation for the finding in Table 3 that the decline in correlation between the tone of traditional media and social media tone is caused by social media is simply capturing noise. [16: We run the regression of CAR on social media tone in a separate regression because as we partition the sample into terciles based on the difference between traditional media and social media tone, social media tone and traditional media tone become highly correlated (between 0.3 to 0.6 in the three partitions), causing multicollinearity problem. ]

However, our F-tests show that for CAR (1,10) and CAR (1,20), the coefficient on Social Media Tone of TM bias HIGH is significantly lower than that of TM bias LOW. This reduction in the coefficient suggests that social media is not entirely bias-free. In the TM bias HIGH tercile, the high Traditional Media Tone – Social Media Tone can be capturing very negative Social Media Tone. Therefore, the decline in stock return response can be due to social media’s negative tone bias. Nevertheless, the percentage difference in stock return response between TM bias High and TM bias LOW is significantly smaller for social media, which ranges from 6.25% to 52%, than that of traditional media, which ranges from 140% to 186%. Notwithstanding that negative bias is found in social media’s posts, the magnitude of its negative bias is small in comparison to that of traditional media’s positive bias.

In summary, the regression analyses provide evidence that the stock reaction to traditional media is attenuated when its tone deviates positively from social media’s tone, supporting our conjecture that social media serves a moderating role in the information bias of traditional media.[footnoteRef:17] [17: We conclude that the stock return results support our conjecture that social media serves as a benchmark for delineating traditional media’s optimistic bias. However, we do not claim that it is the social media information that causes the market to discount its response to the biased information. The discount could simply be reflecting the market’s response to the unbiased underlying information that is captured by the social media. ]

5. Additional Analyses and Robustness Tests

5.1 Social Media’s moderation of information bias for State-owned Firms versus Non-State-Owned Firms

In this subsection, we focus on whether social media’s ability to moderate media bias is different between state-owned enterprises (SOEs) and non-SOEs. Piotroski et al. (2017) document that Chinese newspapers are more likely to optimistically bias the news of SOEs because negative news of SOEs will impose a higher cost on politicians than non-SOEs. If social media maintain their independence and not bias one type of firms more than the other, we expect social media’s posts are able to delineate traditional media’s bias for SOEs than non-SOEs. We use the following cross-sectional model to examine whether social media can delineate the difference in traditional media’s optimistic bias between SOEs and non-SOEs:

(3)

Table 7 presents the result of OLS regression with industry and year-month fixed effects to control for unobservable correlated omitted variables. The results in column 1 suggest that social media can delineate a stronger traditional media’s optimistic bias for SOEs than non-SOEs, but only in the period prior to the 2015 media intervention (pre-intervention period). During this period, the reduction in coefficients between traditional media tone and social media tone when traditional media tone is positive is about 79% for SOEs and 48% for non-SOEs.[footnoteRef:18] We do not find similar results in the post-intervention sample or the full-sample. This is likely because as a response to the strict media intervention order from the government, traditional media increases the bias for all firms regardless of their ownership structure in the post-intervention period. The results suggest that the traditional media bias for SOEs and non-SOEs has been leveled following the intervention.[footnoteRef:19] [18: The reduction in association between traditional media tone and social media tone when traditional media tone is positive (i.e. Social media’s delineation of Traditional Media Bias) is calculated as follow: for SOEs , for non-SOEs /] [19: We find that in the pre-intervention period, tone of traditional media is significantly more positive for SOEs than for no-SOEs, but this difference has disappeared in the post-intervention period. The results (untabulated) are robust when controlling for firm characteristics and including industry fixed effect and year-month fixed effect. ]

5.2 Social Media’s Moderation of Traditional Media Bias around Earnings-Event Windows

Our first main analysis (Tables 3) aims to compare how truthfully traditional and social media capture the tone of the underlying events on the same day. However, if the firm has no significant underlying event(s) on that day to anchor what traditional and social media should report, they could be covering different topics which causes the correlation of their tone to attenuate even in the absence of any intentional bias.

In this subsection, we repeat our main test in Table 3 using firm-days when we can identify significant underlying events for the media coverage. We pick firms’ announcements of earnings and management earnings guidance as two underlying events since these are two of the most critical public information disclosure events. In short windows around these events, we expect that traditional media and social media are anchored to report the same underlying events. Using only the day of and the day subsequent to these earnings events, we can ensure that the news articles and social media posts are covering the same underlying events.[footnoteRef:20] [20: Given an important underlying event such as earnings forecast or announcement, if traditional media chooses to cover different topics from those of social media and causes a significant deviation in tone, we will consider this to be a coverage bias. ]

In Table 8, we estimate Model 1 with only firm-day observations on day 0 (column 1) and on day 0 to day +1 (column 2) relative to the quarterly earnings announcements and the earnings guidance announcements. In both samples, we find significant and negative coefficients on Traditional Media Tone x TM_Positive. These findings suggest that social media’s moderation role exist even when we restrict the firm-days to have significant underlying events related to earnings forecasts and actual earnings announcements.

5.3 Social Media’s Moderation of Domestic Traditional Media vs. Foreign Traditional Media

In our main test, we hypothesize that Chinese traditional media is optimistically biased in reporting corporate news since it is subject to political incentives (Stockmann, 2013; Piotroski et al., 2017). However, traditional media outside mainland China should be free from the control of the Chinese authority and should have no incentives to positively bias its news. Thus, we should only observe the Chinese social media delineating the positive bias of domestic traditional media but not foreign traditional media.

To test this conjecture, we exploit news articles reported by 58 foreign newspapers outside mainland China, covering 114 companies dual-listed in both Hong Kong stock exchange and mainland China.[footnoteRef:21] We only include the firm-day observations in our analysis when there is at least one news article from a foreign media and one from a domestic media. Table 9 reports the regression results of Model 1 separately for foreign traditional media (column 1) and domestic traditional media (column 2). The coefficient on Traditional Media Tone x TM_domestic_Positive in column 2 is significantly negative, which is similar to our main results in Table 3. However, we do not find a significant coefficient on Traditional Media Tone x TM_foreign_Positive for foreign media. These findings further support to our hypothesis that the decline in correlation between the tone of traditional and social media is caused by traditional media’s politically induced reporting bias. For foreign traditional media that is not subject to such political incentives, we do not find anythe decline in the correlation of tone. [21: The geographic distribution of the 58 newspapers: 33 - Hong Kong, 9 - Taiwan, 6 - Singapore, 4 - Macau, 2 -Malaysia, 2 - US, 2 -Canada. See Appendix I for a detailed list. We limit our sample to dual-listed companies because foreign media rarely covers companies that are only listed domestically. ]

5.4 Robustness Tests

5.4.1 Alternative Measures of Traditional Media Bias

Our results of social media’s benchmarking role primarily rely on using an indicator variable whether traditional media tone is positive (i.e., Traditional Media Tone > 0). From the descriptive statistics, we show that traditional media is positive in about 77.4% of the firm-day observations. One may be concerned that this indicator variable has insufficient variation in capturing the bias in tone. As a robustness check, we replace the TM Positive indicator variable with TM Above Median, which is an indicator variable set to one if the tone of traditional media is above the median tone of all firms on a given day. We repeat our baseline regressions and find that our results (untabulated) are statistically and economically robust to this alternative measure.

5.4.2 Alternative Measures of the Dependent Variable - Social Media Tone

In our sample, we require at least three social media posts about the same company for a firm-day to be included in our sample, in order to avoid measurement errors due to a few extreme and unrepresented posts. One may be concerned that limiting the sample to three posts minimum would be insufficient in addressing this issue. As a robustness check, we tried the following restrictions in constructing our dependent variable – Ssocial Mmedia Ttone: 1) include only samples with a minimum of five posts per day 2) include only samples with a minimum of ten posts per day 3) winsorize the most positive and most negative social media posts in 1) & 2). Our results (untabulated) show that the baseline results largely remain unchanged in both sign and significance after applying these three alternative measures of ssocial media tone to our main results.

6. Conclusions

This paper examines whether China’s social media can moderate the information environment by supplying less positively biased information to the market than traditional media. Using a comprehensive sample from 2009 to 2016 of corporate news of newspapers and the users’ posts in East Guba, an online stock forum, we document that there is a positive association in the tone of the newspapers and East Guba for the same firm on the same day. However, we find that the tone of East Guba is less positively associated with that of the newspapers when the tone of the newspapers is positive. Consistent with the conjecture that East Guba plays the moderating role in delineating the bias of state-owned media, these results suggest that the tone of East Guba’s posts is less optimistic than that of the newspapers when the latter are more likely to be positively biased.

Exploiting the 2015 media intervention as an exogenous political shock, we find that this political intervention has increased the newspapers’ optimistic bias and the deviation in tone between the newspapers and East Guba. This suggests the political influence on East Guba is smaller, which allows it to continue its moderating role in delineating the newspapers’ optimistic bias when political pressures heighten. Finally, the positive stock return response to the tone of the newspapers’ articles is significantly attenuated when the tone of newspapers deviates positively from that of East Guba, further supporting that East Guba can serve as a benchmark against newspapers’ reporting bias.

Reference

Antweiler, W., & Frank, M. Z. (2004). Is all that talk just noise? The information content of internet stock message boards. The Journal of finance, 59(3), 1259-1294.

Bartov, E., Faurel, L., & Mohanram, P. S. (2017). Can Twitter help predict firm-level earnings and stock returns?. The Accounting Review, 93(3), 25-57.

Baumeister, Roy F., et al. "Bad is stronger than good." Review of general psychology 5.4 (2001): 323.

Blankespoor, E., Miller, G. S., & White, H. D. (2013). The role of dissemination in market liquidity: Evidence from firms' use of Twitter™. The Accounting Review, 89(1), 79-112.

Byard, Donal, and Eric Wang. "The Impact of Public Disclosure on Information Asymmetry between Sophisticated and Unsophisticated Investors: Evidence from an Investor Social Media Network." (2016).

Cacioppo, John T., and Wendi L. Gardner. "Emotion." Annual review of psychology 50.1 (1999): 191-214.

Chen, H., De, P., Hu, Y. J., & Hwang, B. H. (2014). Wisdom of crowds: The value of stock opinions transmitted through social media. The Review of Financial Studies, 27(5), 1367-1403.

Djankov, S., McLiesh, C., Nenova, T., & Shleifer, A. (2003). Who owns the media?. The Journal of Law and Economics, 46(2), 341-382.

Enikolopov, R., Petrova, M., & Zhuravskaya, E. (2011). Media and political persuasion: Evidence from Russia. American Economic Review, 101(7), 3253-85.

Foucault, M.. 1977. Discipline and punish. Harmondsworth: Penguin.

Gentzkow, M., & Shapiro, J. M. (2006). Media bias and reputation. Journal of political Economy, 114(2), 280-316.

Gurun, Umit G., and Alexander W. Butler. 2012. Don’t Believe the Hype: Local Media Slant, Local Advertising, and Firm Value. Journal of Finance 67:561–97.

Kahneman, Daniel. "Prospect theory: An analysis of decisions under risk." Econometrica 47 (1979): 278.

King, G., Pan, J., & Roberts, M. E. (2013). How censorship in China allows government criticism but silences collective expression. American Political Science Review, 107(2), 326-343.

King, G., Pan, J., & Roberts, M. E. (2014). Reverse-engineering censorship in China: Randomized experimentation and participant observation. Science, 345(6199), 1251722.

Li, H., & Zhou, L. A. (2005). Political turnover and economic performance: the incentive role of personnel control in China. Journal of public economics, 89(9-10), 1743-1762.

Noelle-Neumann, Elisabeth, 1984. The Spiral of Silence: Public Opinion-Our Social Skin. The University of Chicago Press.

Piotroski, J. D., Wong, T. J., & Zhang, T. (2015). Political incentives to suppress negative information: evidence from Chinese listed firms. Journal of Accounting Research, 53(2), 405-459.

Piotroski, J. D., Wong, T. J., & Zhang, T. (2017). Political Bias in Corporate News: The Role of Conglomeration Reform in China. The Journal of Law and Economics, 60(1), 173-207.

Qin, B., Strömberg, D. and Wu, Y., 2017. Why does China allow freer social media? Protests versus surveillance and propaganda. Journal of Economic Perspectives, 31(1), pp.117-40.

Solomon, D. H. (2012). Selective publicity and stock prices. The Journal of Finance, 67(2), 599-638.

Spears R. and Lea M.. 1994. Panacea or Panopticon? The hidden power in computer-mediated communication. Communication Research 21(4), 247-459.

Stockmann, D. (2013). Media commercialization and authoritarian rule in China. Cambridge University Press.

Tetlock, P. C., Saar‐Tsechansky, M., & Macskassy, S. (2008). More than words: Quantifying language to measure firms' fundamentals. The Journal of Finance, 63(3), 1437-1467.

Tumarkin, R., & Whitelaw, R. F. (2001). News or noise? Internet postings and stock prices. Financial Analysts Journal, 57(3), 41-51.

Watts, R.L., 2003a. Conservatism in accounting part I: Explanations and implications. Accounting horizons, 17(3), pp.207-221.

Watts, R.L., 2003b. Conservatism in accounting part II: Evidence and research opportunities. Accounting horizons, 17(4), pp.287-301.

Table 1- Panel A Sample Selection Process

Firms

Firm-day

CSMAR-Financial Statement File Merged with Stock Price File

3,171

4,421,222

Less firm-days with less than 3 posts

(132)

(136,019)

Social Media with no less than 3 posts/day

3,039

4,285,203

Less firm-days with no news articles

(68)

(3,524,454)

Traditional Media with no less than 1 news/day

2,971

760,749

Less firm-days with missing control variables

(29)

(39,536)

Final Sample

2,942

721,213

This table presents the sample selection process of our firm-day observations used in the main tests. We require each firm-day observations to have at least three social media posts to avoid measurement errors of the social media tone caused by insufficient posts per day. To assess the tone of traditional media on the same day, we require at least one traditional media news article for a firm-day to be included in our sample. Results are robust when we increase the threshold to five social media posts per day, despite a reduced sample size. Results are also robust when we increase the threshold on the daily traditional media news articles to three per day.

Table 1- Panel B Descriptive Statistics

Variable

N

Mean

SD

P25

Median

P75

Traditional Media

Number of news

721,213

2.6597

4.7954

1.0000

1.0000

2.0000

Traditional media tone

721,213

0.3756

0.5104

0.0652

0.4892

0.8250

Traditional media positive

721,213

0.7741

0.4182

1.0000

1.0000

1.0000

Social Media

Social media tone

721,213

-0.2038

0.3108

-0.4095

-0.1840

0.0000

Number of posts

721,213

43.1760

104.9907

3.0000

17.0000

47.0000

Stock Returns

CAR[1,5]Raw return

721,213 721,213

0.00050.0016

0.06080.0342

-0.0303-0.0157

-0.00420.0008

0.02520.0171

CAR[1,10]Abnormal return

721,213 721,213

0.00140.0013

0.08290.0279

-0.0418-0.0129

-0.0056-0.0016

0.03660.0118

CAR[1,20]CAR[1,5]

721,213 721,213

0.00320.0005

0.11290.0608

-0.0592-0.0303

-0.0065-0.0042

0.05460.0252

Firm Fundamentals

SOE

18,446

0.4285

0.4949

0.0000

0.0000

1.0000

Size

18,446

21.9314

1.3775

20.9627

21.7425

22.6732

ROA

18,446

0.0539

0.0809

0.0142

0.0423

0.0834

Market to Book

18,446

3.8208

3.8739

1.7208

2.7526

4.5552

Leverage

18,446

0.2126

0.3828

0.0000

0.0341

0.2607

This table presents descriptive statistics on the measures of social media, traditional media and daily stock returns for the period of July 2009 to 2016. Each firm-day observation must have 1) no fewerless than three social media posts and 2) no less than one traditional media news article. Firm fundamentals are firm-year measures of firm characteristics at each fiscal year-end. All variables are defined in Appendix II.

Table 1-Panel C: Correlation Table

1

2

3

4

5

68

79

1

number of news

2

traditional media tone

-0.0441

3

traditional media positiveTM_Positive

0.0649

0.7404

4

social media tone

-0.0169

0.0712

0.0644

5

number of posts

0.1038

-0.0824

-0.0583

-0.1079

68

CAR[1,5]

-0.0067

0.0188

0.0178

0.0267

-0.0533

79

CAR[1,10]

-0.0086

0.0144

0.0141

0.0249

-0.0591

0.7298

810

CAR[1,20]

-0.0105

0.0125

0.0089

0.0205

-0.0573

0.5375

0.7312

This table presents the Pearson’s correlation coefficients of variables used in the regressions.

Correlation coefficients that are significant at 0.05 level are reported in the table. All variables are defined in Appendix II.

This table presents the Pearson’s correlation coefficients of variables used in the regressions. Correlation coefficients that are significant at 0.05 level are reported in the table.

Table 2: Sample Distribution

Panel A- by Year

Sample Distribution - By Year

Year

Freq.

Percent

Cum.

2009

91,610

12.70%

12.70%

2010

85,855

11.90%

24.61%

2011

94,428

13.09%

37.70%

2012

105,429

14.62%

52.32%

2013

103,257

14.32%

66.63%

2014

87,464

12.13%

78.76%

2015

75,641

10.49%

89.25%

2016

77,529

10.75%

100.00%

Total

721,213

100%

100%

Table 2: Sample Distribution

Panel B- by Industry

Industry

Frequency

Percentage%

Cumulative %

Computer and Communications

39,355

5.46

5.46

Real Estate

36,339

5.04

10.50

Pharmaceutical

35,410

4.91

15.41

Electrical Manufacture

32,346

4.48

19.89

Automotive

31,633

4.39

24.28

Chemical Products

31,616

4.38

28.66

Retails

29,569

4.10

32.76

Financial Service

28,475

3.95

36.71

Specialized Equipment Manufacture

26,133

3.62

40.33

Business Service

24,557

3.40

43.74

Software and Information Technology

23,576

3.27

47.01

Alcoholic Beverage, Non-alcoholic Beverage and Tea

21,670

3.00

50.01

Wholesale

17,912

2.48

52.49

Construction

17,847

2.47

54.97

Electricity and Heat Supply

16,827

2.33

57.30

Non-metallic Mineral

15,792

2.19

59.49

General Equipment Manufacture

15,201

2.11

61.60

Ferrous Metal Smelting

14,888

2.06

63.66

Non-Ferrous Metal Smelting

14,779

2.05

65.71

Water Transportation

14,153

1.96

67.67

Aero Transportation

12,781

1.77

69.45

Transportation Equipment Manufacture

11,810

1.64

71.08

Business Service

10,039

1.39

72.48

Road Transportation

9,046

1.25

73.73

Coal Mining and Washing

8,787

1.22

74.95

Internet Service

8,543

1.18

76.13

Food Manufacture

7,914

1.10

77.23

Metallic Product Manufacture

7,587

1.05

78.28

Agriculture

7,129

0.99

79.27

Others

149,499

20.73

100

Total

721,213

100

Table 3: Traditional Media Bias and Social Media’s Bias-Moderating Role

Dependent Variable: Social Media Tone

Independent Variables

1

2

3

Traditional Media Tone

0.0343***

0.0488***

0.0500***

(44.43)

(17.38)

(17.83)

Traditional Media Tone x TM_Positive

-0.0255***

-0.0269***

(-7.91)

(-8.36)

TM_Positive

0.0032*

0.0032**

(1.94)

(1.99)

Size

-0.0326***

(-22.78)

ROA

0.0452***

(5.61)

Market to Book

-0.0024***

(-12.13)

Leverage

0.0277***

(12.04)

Adj-R2

0.096

0.096

0.097

N

721,213

721,213

721,213

Firm-Fixed Effect

YES

YES

YES

Year-Month-Fixed Effect

YES

YES

YES

This table presents the results of OLS regressions of the tone of social media on the tone of traditional media of the same firm on the same day using Model (1) – Section 4.1. The dependent variable is the daily Social Media Tone, which is measured using textual analysis of the posts on East Money Guba, a major social media platform dedicated to the discussion of financial news related to a firm. Traditional Media Tone is the mean tone of traditional media news articles extracted using textual analysis. TM_Positive is a dummy variable set to 1 if Traditional Media Tone is >0. All other variables are defined in Appendix II. Standard errors are estimated by two-way clustering at firm and year-month level with t-statistics presented in parentheses. ***, **, and * denote statistical significance at the 1%, 5%, and 10% levels, respectively.

Table 4: Influence of Political Intervention on the Bias-Moderating Role of Social Media

Dependent Variable: Social Media Tone

(1)

(2)

Difference

Independent Variables

Pre-Intervention

Post-Intervention

(2)-(1)

Traditional Media Tone

0.0508***

0.0370***

-0.0138

(16.56)

(5.65)

(-1.02)

Traditional Media Tone x TM_Positive

-0.0254***

-0.0322***

-0.0068**

(-7.22)

(-4.26)

(-3.94)

TM_Positive

0.0031*

0.0073*

(1.72)

(1.88)

Size

-0.0200***

-0.0812***

(-11.12)

(-10.45)

ROA

0.0092

0.0401

(1.03)

(1.09)

Market to Book

-0.0010***

-0.0059***

(-4.00)

(-9.44)

Leverage

0.0249***

0.0564***

(9.41)

(5.91)

Adj-R2

0.091

0.123

N

617,230

103,983

Firm-Fixed Effect

YES

YES

Year-Month-Fixed Effect

YES

YES

This table presents the results of OLS regressions of the tone of social media on the tone of traditional media of the same firm on the same day in two sub-sample period: pre- and post- media intervention in which traditional media are required to decrease the coverage of stock market-related news, and stop using negative words such as “tumble” or “crash” in the news. Pre-intervention period is before or on July 23rd, 2015, and post-intervention period after this date. All other variables are defined in Appendix II. We use Chow Test and reportwith Chi-square Statistics to test the equality of two coefficients across sample. Standard errors are estimated by two-way clustering at firm and year-month level with t-statistics presented in parentheses. ***, **, and * denote statistical significance at the 1%, 5%, and 10% levels, respectively.

Table 5-Panel A: Social Media’s Bias-Moderating Role during National Congress Meeting Period

Dependent Variable: Social Media Tone

1

2

Independent Variables

Congress Meeting

Non-Congress Meeting

Traditional Media Tone

0.0182

0.0509***

(1.25)

(17.82)

Traditional Media Tone x TM_Positive

0.0071

-0.0284***

(0.44)

(-8.66)

TM_Positive

0.0063

0.0031*

(0.74)

(1.90)

Size

-0.0133

-0.0329***

(-0.22)

(-22.53)

ROA

0.0435

0.0502***

(1.09)

(6.12)

Market to Book

-0.2064

-0.0024***

(-0.77)

(-11.92)

Leverage

0.3121***

0.0283***

(4.20)

(12.19)

Adj-R2

0.214

0.096

N

31,303

689,910

Firm-Fixed Effect

YES

YES

Year-Month-Fixed Effect

YES

YES

This table presents the results of OLS regressions of the tone of social media on the tone of traditional media of the same firm on the same day in two sub-sample periods: Congress Meeting Period and Non-Congress Meeting Period. The Congress Mmeeting sample includes firm-days in a window of -45 to +45 days relative to the opening day of the 18th CCP National Congress Meeting (Nov.8th, 2012), and Nnon-Ccongress Mmeeting sample includes the rest of the firm-day observations. All variables are defined in Appendix II. Standard errors are estimated by two-way clustering at firm and year-month level with t-statistics presented in parentheses. ***, **, and * denote statistical significance at the 1%, 5%, and 10% levels, respectively.

Table 5- Panel B: Number of Social Media Posts / Traditional Media News Articles during Congress Meeting Period

1

2

Dependent Variables

Independent Variables

Number of Social Media Posts

Number of News Articles

Congress Meeting

-21.3418***

0.1442***

(-17.87)

(9.00)

Number of News

3.4990***

(15.19)

Number of Social Media Posts

0.0029***

(15.74)

Social Media Tone

3.2437

0.1771***

(0.31)

(16.08)

Traditional Media Tone

-5.5268***

-0.1157***

(-5.67)

(-7.36)

Stock Return

19.1188***

0.7070***

(5.99)

(7.29)

Index Return

12.4363

0.1961

(0.31)

(0.97)

Size

22.1478***

0.2306***

(15.89)

(11.10)

ROA

7.3900

1.2491**

(0.67)

(2.08)

Market to Book

-0.9025

0.0371**

(-1.32)

(1.91)

Leverage

-7.8183***

-0.4036***

(-14.09)

(15.26)

Adj-R2

0.312

0.675

N

721,213

721,213

Firm-Fixed Effect

YES

YES

This table presents the result of regression results of the relationship between Congress Meeting and the level of activities for traditional media and social media and congress meeting. The dependent variable is the number of social media posts/ traditional media news articles. We estimate a pooled OLS regression model to determine the impact of the Congress Meeting on the level of activities proxied by the number of social media posts and the number of new articles. All variables are defined in Appendix II. Standard errors clustered at firm with t-statistics presented in parentheses. ***, **, and * denote statistical significance at the 1%, 5%, and 10% levels, respectively.

Table 6 - Panel A: Market Responses to Traditional Media: Social Media’s Bias-Moderation Role

CAR(1,5)

CAR(1,10)

CAR(1,20)

TM Bias

LOW

TM Bias MEDIUM

TM Bias

HIGH

TM Bias

LOW

TM Bias MEDIUM

TM Bias

HIGH

TM Bias

LOW

TM Bias MEDIUM

TM Bias

HIGH

Traditional Media Tone

0.0015***

0.0013**

-0.0006

0.0022***

0.0017*

-0.0019

0.0018**

0.0022

-0.0012

(4.24)

(1.98)

(-0.53)

(4.32)

(1.83)

(-1.17)

(2.45)

(1.49)

(-0.72)

Size

0.0007***

0.0006***

0.0003***

0.0005***

0.0005***

0.0001

-0.0002

-0.0005**

0.0001

(7.23)

(7.30)

(2.98)

(3.07)

(3.19)

(0.91)

(-0.85)

(-2.16)

(0.38)

ROA

0.0064***

0.0008

-0.0012

0.0083***

-0.0004

-0.0026

0.0037

-0.0054

-0.0024

(3.15)

(0.40)

(-0.56)

(2.62)

(-0.14)

(-0.85)

(0.72)

(-1.08)

(-0.77)

Market to Book

0.0001

0.0002***

0.0002***

0.0001

0.0001

0.0001

-0.0004***

-0.0006***

-0.0002**

(1.21)

(3.49)

(2.78)

(1.14)

(1.42)

(1.33)

(-2.65)

(-4.61)

(-2.54)

Leverage

-0.0003

-0.0008**

-0.0006*

-0.0001

-0.0013**

-0.0009

0.0006

-0.0017*

-0.0009

(-0.87)

(-2.20)

(-1.67)

(-0.24)

(-2.33)

(-1.45)

(0.59)

(-1.70)

(-1.57)

CAR [-5,-1]

0.3136***

0.3396***

0.2914***

0.4295***

0.4449***

0.4143***

0.5403***

0.5608***

0.4241***

(86.09)

(93.97)

(84.54)

(83.86)

(85.06)

(82.77)

(84.69)

(86.09)

(83.31)

Adj- R2

0.017

0.020

0.015

0.018

0.019

0.016

0.015

0.015

0.014

N

240,174

241,133

239,906

240,174

241,133

239,906

240,174

241,133

239,906

F-Test of Coefficients

H0: β1-TM_Bias_Low =β1-TM_Bias_High

Chi2=3.25, Pr = 0.0714

H0: β1-TM_Bias_Low =β1-TM_Bias_High

Chi2=6.49, Pr = 0.0108

H0: β1-TM_Bias_Low =β1-TM_Bias_High

Chi2=1.32, Pr = 0.2511

This table reports the OLS regression estimates of traditional media tone of Day 0 on the cumulative abnormal return (CAR) of three windows: CAR(1,5) CAR(1,10) and CAR(1,20). Traditional Media Tone is the mean tone of all traditional media news articles covering the same firm on day 0. S; Sample is partitioned into three tercile groups based on the difference of tones between traditional media and social media calculated as (Traditional Media Tone – Social Media Tone), and are titled as “TM Bias LOW, TM Bias MEDIUM and TM Bias HIGH” from the 1st to the 3rd tercile. All variables are defined in Appendix II. F-test reports the Chi2 and corresponding P-value for the test of equality between the coefficients of traditional media tone in TM-Bias Low group and TM-Bias High group. All regressions include a year-month fixed effect. Standard errors are estimated by two-way clustering at firm and year-month level with t-statistics presented in parentheses. ***, **, and * denote statistical significance at the 1%, 5%, and 10% levels, respectively.

Table 6 – Panel B: Market Responses to Social Media: Social Media’s Bias-Moderation Role

CAR(1,5)

CAR(1,10)

CAR(1,20)

TM Bias

LOW

TM Bias MEDIUM

TM Bias

HIGH

TM Bias

LOW

TM Bias MEDIUM

TM Bias

HIGH

TM Bias

LOW

TM Bias MEDIUM

TM Bias

HIGH

Social Media Tone

0.0048***

0.0046***

0.0045***

0.0067***

0.0053***

0.0046***

0.0086***

0.0052***

0.0041***

(12.98)

(11.62)

(11.34)

(12.56)

(9.16)

(8.21)

(10.94)

(5.97)

(5.10)

Size

0.0008***

0.0007***

0.0004***

0.0007***

0.0005***

0.0002

-0.0000

-0.0005**

-0.0007***

(8.79)

(8.07)

(3.73)

(4.34)

(3.69)

(1.44)

(-0.06)

(-2.03)

(-2.58)

ROA

0.0054***

-0.0001

-0.0019

0.0069**

-0.0015

-0.0034

0.0022

-0.0072

-0.0096**

(2.64)

(-0.05)

(-0.91)

(2.17)

(-0.46)

(-1.10)

(0.43)

(-1.41)

(-2.01)

Market to Book

0.0001

0.0002***

0.0002***

0.0001

0.0001

0.0001

-0.0001

-0.0002*

-0.0002

(1.59)

(3.73)

(3.11)

(1.48)

(1.60)

(1.57)

(-0.70)

(-1.74)

(-1.59)

Leverage

-0.0004

-0.0007**

-0.0006

-0.0002

-0.0012**

-0.0008

0.0004

-0.0016

-0.0006

(-1.01)

(-2.07)

(-1.60)

(-0.36)

(-2.24)

(-1.42)

(0.40)

(-1.58)

(-0.67)

CAR [-5,-1]

0.3134***

0.3392***

0.2911***

0.4291***

0.4444***

0.4140***

0.5330***

0.5378***

0.5177***

(86.15)

(94.06)

(84.69)

(83.93)

(85.15)

(82.89)

(85.17)

(83.19)

(85.83)

Adj- R2

0.019

0.023

0.018

0.018

0.019

0.017

0.015

0.014

0.016

N

240,174

241,133

239,906

240,174

241,133

239,906

240,174

241,133

239,906

F-Test of Coefficients

H0: β1-SM_Bias_Low =β1-SM_Bias_High

Chi2=0.34, Pr = 0.5598

H0: β1-SM_Bias_Low =β1-SM_Bias_High

Chi2=9.84, Pr = 0.0017

H0: β1-SM_Bias_Low =β1-SM_Bias_High

Chi2=22.79, Pr = 0.0000

This table reports the OLS regression estimates of social media tone of Day 0 on the cumulative abnormal return (CAR) of three windows: CAR(1,5) CAR(1,10) and CAR(1,20). Social Media Tone is the mean tone of all social media posts covering the same firm on day 0.; Sample is partitioned into three tercile groups based on the difference of tones between traditional media and social media calculated as (Tradit