News or Noise? Using Twitter to Identify and Understand Company‐specific News Flow

• Published date: 01 September 2014
• Author: Andranik Tumasjan,Timm O. Sprenger,Isabell M. Welpe,Philipp G. Sandner
• Date: 01 September 2014
• DOI: http://doi.org/10.1111/jbfa.12086

Journal of Business Finance & Accounting

Journal of Business Finance & Accounting, 41(7) & (8), 791–830, September/October 2014, 0306-686X

doi: 10.1111/jbfa.12086

News or Noise? Using Twitter to Identify

and Understand Company-specific News

Flow

TIMM O. SPRENGER,PHILIPP G. SANDNER,ANDRANIK TUMASJAN

AND ISABELL M. WELPE∗

Abstract: This study presents a methodology for identifying a broad range of real-world

news events based on microblogging messages. Applying computational linguistics to a unique

dataset of more than 400,000 S&P 500 stock-related Twitter messages, we distinguish between

good and bad news and demonstrate that the returns prior to good news events are more

pronounced than for bad news events. We show that the stock market impact of news events

differs substantially across different categories.

Keywords: event study, news events, information leakage, market reaction, computational

linguistics

1. INTRODUCTION

Event studies are widely used in finance and accounting research (Kothari and Warner,

2007) to “measure the impact of a specific event on the value of a firm” (MacKinlay,

1997, p. 13). Since the publication of Fama et al.’s (1998) seminal paper, which

established the event study methodology, many event studies have been conducted,

and the methodology has become a standard approach to assess the impact of event

news on stock prices (Binder, 1998). While event studies continue to be a popular

method across a variety of fields (Corrado, 2011), researchers have noted several

methodological challenges associated with event studies (e.g., MacKinlay, 1997). The

purpose of this paper is to develop a methodology based on a data source (i.e., Twitter)

that has recently been introduced to the literature (Sprenger et al., 2013). Using this

data source, we focus on and address three main challenges when conducting event

studies, thereby offering several advantages to event study researchers.

∗The authors are from the TechnischeUniversit ¨

at M¨

unchen (TUM), TUM School of Management, Arcisstr.

21, 80333 Munich, Germany. The authors thank Peter Pope (Editor of this journal) and an anonymous

referee for their constructive suggestions that have helped to improve this article. The authors further thank

Simon Pickert for his support during data processing. This article is based on the dissertation of the first

author.

Address for correspondence: Philipp G. Sandner, Technische Universit¨

at M¨

unchen (TUM), TUM School of

Management, Arcisstr. 21, 80333 Munich, Germany.

e-mail: philipp.sandner@tum.de

C

2014 John Wiley & Sons Ltd 791

792 SPRENGER, SANDNER, TUMASJAN AND WELPE

First, using the business press as a news source, researchers often struggle to identify

the accurate timing of events (MacKinlay, 1997). Determining the exact event day

is difficult because the press may cover an event from the previous day (i.e., the

market may already have been informed), or alternatively, the publication day may

be the event day (MacKinlay, 1997). A second challenge that event study researchers

are confronted with concerns the identification of the sentiment of the news, i.e.,

distinguishing between good and bad news signals. Very few event studies distinguish

between good and bad news even though taking news sentiment into account is

essential (Ryan and Taffler, 2004), given that good and bad news entail different stock

market reactions (e.g., Schmitz, 2007). A third challenge concerns the concurrent

analysis of multiple types of news events. Because most event studies concentrate on

one particular type of event (e.g., stock splits, dividend initiations and omissions,

acquisitions, or initial public offerings; Agrawal et al., 1992; Michaely et al., 1995;

Ritter, 1991; Ikenberry and Ramnath, 2002), our current knowledge on the relative

market reaction of the various different news events types is limited. Accordingly,

scholars have noted that it would be desirable to investigate many different types

of events simultaneously. For instance, Tetlock et al. (2008, p. 1,438) noted that

“analyzing a more complete set of events that affect firms’ fundamental values allows

researchers to identify common patterns in firm responses and market reactions to

events”. In line with other researchers (e.g., Ryan and Taffler, 2004; Antweiler and

Frank, 2006), Thompson et al. (1987) concluded that “additional research is needed

to determine which news release subcategories [...] are associated withthe largest

contemporaneous security price changes” (p. 268).

Using data from the microblogging service Twitter, we introduce a method that

addresses these challenges that event study researchers face and that provides the

following main advantages to researchers. First, our approach allows event study

researchers to accurately identify the timing of events because it uses a real-time

stream of microblogging messages rather than the business press. Second, our

method explicitly distinguishes between good and bad news, offering researchers the

advantage of considering the sentiment of the news. Third, the extensive database of

Twitter messages covering a variety of news makes it possible to investigate the market

impact of a comprehensive set of several company-specific news events rather than

being confined to one specific type of news event. Thereby, our method makes it

possible to examine which types of news move the market.

To this end, our study identifies news events (related to, for example, Corporate

Governance, Financial Issues, Operations, Restructuring Issues, Legal Issues and Technical

Trading) from more than 400,000 stock-related Twitter messages and examines their

impact on S&P 500 stock prices. Overall, our results show that using data from a

real-time online stock forum enables researchers to reliably identify company-relevant

news events and investigate their impact on stock prices while concurrently taking

into account the sentiment of the news. To validate our approach, we also employed

earnings announcements as event dates to compare both the stock price reactions and

trading volume of these events with those events identified through the massive real-

time stream of Twitter messages. Our results in this respect underpin the assumption

that investor discussion in an online stock forum meaningfully reflects real-world news

events.

The remainder of this article is structured as follows. In section 2, we review the

related event study research and identify the methodological challenges that extant

C

2014 John Wiley & Sons Ltd

UNDERSTAND COMPANY-SPECIFIC NEWS FLOW 793

event studies struggle with. In section 3, because our approach makes use of a relatively

new information platform (i.e., Twitter), we discuss the role of Twitter (versus the

business press and stock message boards) as an information intermediary in financial

markets from a theoretical perspective. We then derive the research questions that

guide our study. Section 4 describes our dataset and our methodological approach. In

section 5 we provide the results of our event study, showing that the market reaction

differs across multiple event types and that sentiment is a crucial factor in evaluating

news events. Finally, in section 6, we discuss the implications of our findings and

provide suggestions for further event study research.

2. LITERATURE REVIEW

(i) News as a Source to Identify Event Days

Within the event study literature, only a few studies use comprehensive datasets to

explore the market impact of stock-related news stories. For example, to explore

the effect of news on drift, Chan (2003) used news stories archived in the Dow

Jones Interactive Publication Library and focused on the differences between returns

after major news stories about a company on the one hand and returns after large

price movements in the absence of news on the other. The author found that stocks

associated with news, particularly bad news, exhibit momentum or drift of up to

12 months. However, event days were defined based on the presence or absence of

one or more news headlines in major publications, which provides no indication of

the intensity and salience of news coverage (Barber and Odean, 2008). Addressing

this limitation, Mitchell and Mulherin (1994) have compared the number of news

announcements (i.e., the news volume) reported by Dow Jones & Company to the

market activity and found a weak positive relationship between the news volume and

both the trading volume and the absolute value of firm-specific returns. Fang and

Peress (2009) support the notion that the breadth of information dissemination affects

stock returns but found empirical evidence in contrast to Mitchell and Mulherin

(1994), suggesting that stocks with no media coverage earn higher returns. Mitchell

and Mulherin (1994) refer to the news volume as a “measure of information” (p. 923).

We argue that this is a rather optimistic definition because the mere number of news

stories fails to capture many nuances of the information content, such as the sentiment

or importance of any particular news story.

In sum, event studies that comprehensively explore company-specific media cov-

erage are largely limited to the number of news items and do not take into account

news sentiment. However, as demonstrated in the extant literature, it is important to

consider the influence of news sentiment, as it has been shown to substantially affect

market reactions (e.g., Tetlock, 2007; Riordan et al., 2010; Storkenmaier et al., 2012).

(ii) Limitations of the Business Press as a Source of News

Most event studies (e.g., Mitchell and Mulherin, 1994; Chan, 2003; Antweiler and

Frank, 2006) use professionally edited news content such as the Dow Jones News Ser-

vice or the Wall Street Journal. Weargue that there are three methodological concerns

with respect to these data sources. First, professional news agencies do not necessarily

C

2014 John Wiley & Sons Ltd