K.7
First to “Read” the News: News Analytics and
Algorithmic Trading
von Beschwitz, Bastian, Donald B. Keim, and Massimo Massa
International Finance Discussion Papers
Board of Governors of the Federal Reserve System
Number 1233
July 2018
Please cite paper as:
von Beschwitz, Bastian, Donald B. Keim, and Massimo Massa
(2018). First to “Read” the News: News Analytics and
Algorithmic Trading. International Finance Discussion Papers
1233.
https://doi.org/10.17016/IFDP.2018.1233
Board of Governors of the Federal Reserve System
International Finance Discussion Papers
Number 1233
July 2018
First to “Read” the News: New Analytics and Algorithmic Trading
Bastian von Beschwitz, Donald B. Keim, and Massimo Massa
NOTE: International Finance Discussion Papers are preliminary materials circulated to stimulate
discussion and critical comment. References to International Finance Discussion Papers (other
than an acknowledgment that the writer has had access to unpublished material) should be
cleared with the author or authors. Recent IFDPs are available on the Web at
www.federalreserve.gov/pubs/ifdp/. This paper can be downloaded without charge from the
Social Science Research Network electronic library at www.ssrn.com.
1
First to “Read” the News:
News Analytics and Algorithmic Trading
Bastian von Beschwitz*
Federal Reserve Board
Donald B. Keim**
Wharton School
Massimo Massa***
INSEAD
May 16, 2018
Abstract
Exploiting a unique identification strategy based on inaccurate news analytics, we document a
causal effect of news analytics on the market irrespective of the informational content of the news.
We show that news analytics speed up the stock price and trading volume response to articles, but
reduce liquidity. Inaccurate news analytics lead to small price distortions that are corrected
quickly. The market impact of news analytics is greatest for press releases, which are timelier and
easier to interpret algorithmically. Furthermore, we provide evidence that high frequency traders
rely on the information from news analytics for directional trading on company-specific news.
JEL classification: G10, G12, G14
Keywords: Stock Price Reaction, News Analytics, High Frequency Trading, Press Releases.
* Bastian von Beschwitz, Federal Reserve Board, International Finance Division, 20th Street and Constitution Avenue N.W.,
Washington, D.C. 20551, tel. +1 202 475 6330, e-mail: bastian.vonbeschwitz@frb.gov (corresponding author).
** Donald B. Keim, Wharton School, University of Pennsylvania, Philadelphia, PA 19104; [email protected]
*** Massimo Massa, INSEAD, Finance Department, Bd de Constance, 77305 Fontainebleau Cedex, France, tel. + 33-(0)160-724-
481, email: [email protected]
An earlier version of this paper was titled "Media-Driven High Frequency Trading : Evidence from News Analytics". We are
grateful to RavenPack for providing their data, and Malcolm Bain in particular for his expertise on different RavenPack releases.
Thanks also to the technical personnel at WRDS, especially Mark Keintz, for making the construction of the intraday-market
indexes possible. We thank Joseph Engelberg, Nicholas Hirschey, Todd Gormley, Markus Leippold, Joel Peress, Ryan Riordan,
Paul Tetlock, Sarah Zhang and conference participants at the NBER Microstructure Meeting, European Winter Finance Summit,
FIRS, and DGF for valuable comments. We acknowledge the financial support of the Wharton-INSEAD Center for Global
Research and Education. All remaining errors are our responsibility. The views in this paper are solely the responsibility of the
authors and should not be interpreted as reflecting the views of the Board of Governors of the Federal Reserve System or of any
other person associated with the Federal Reserve System.
2
Introduction
A major purpose of financial markets is the assimilation of information into prices. Since the
advent of securities trading, informationally-relevant news has been read and processed by
humans, first directly from newspapers, then from news wires such as Dow Jones, Reuters, and
Bloomberg. However, in the last two decades, computer algorithms have increasingly been used
to read and interpret financial news. Given the importance of news for financial markets, it is
crucial to understand how the algorithmic processing of news releases by computers (“news
analytics”) affects financial markets. In particular, in what ways do news analytics affect stock
returns and trading volume? Who are the users of news analytics? And for which type of articles
are news analytics most important?
We address these questions using news analytics provided by RavenPack, the leading provider
of news analytics in the market. RavenPack uses computer algorithms to determine for each article
in the Dow Jones Newswire its relevance to each company mentioned in it, and whether the news
is positive or negative. This processed content is then electronically delivered to RavenPack’s
subscribers within a third of a second, allowing them to react to the news faster than humans
possibly could.
We address three broad questions in this paper. The first is whether news analytics have a
causal effect on the stock market, making prices and trading volumes react faster to news wire
articles and thereby increasing market efficiency. The second asks for which articles is the causal
effect of news analytics most important. In particular, we compare press releases that are directly
released by companies to other articles that are written by the journalists of Dow Jones. The third
asks whether news analytics are used only for directional trading or also to avoid adverse selection.
3
We study this question by focussing on high frequency traders (HFTs), which we argue are the
type of trader most likely to use news analytics to avoid adverse selection.
These questions are difficult to address in practice because the response to news analytics
normally cannot be distinguished from the reaction to the news itself. We are able to address this
distinction by exploiting a unique identification strategy based on inaccuracies in news analytics
that are revealed by comparing older and newer versions of RavenPack. We use the back-filled
analytics of increasingly more sophisticated versions of RavenPack to identify inaccuracies in the
old version that was released to the market. Finding evidence that markets react to such
inaccuracies would suggest a causal impact of RavenPack on the stock market.
To identify inaccuracies in news analytics, we focus on differences in RavenPack’s “relevance
score”, which measures the importance of an article for a certain company. The relevance score is
very important: highly relevant articles that are positive (negative) are on average followed by
positive (negative) stock returns, while there is almost no reaction to articles with a low relevance
score. Differences in relevance scores between the old and new RavenPack versions are due to
improvements in the algorithm when identifying companies in the article and determining the
article’s relevance to the company.
We use these differences in relevance scores to define three categories of articles: High-
relevance articles Released as High-relevance articles (HRH) are articles that were correctly
released to the market; Low-relevance articles Released as High-relevance articles (LRH) are false
positives, i.e. articles that are wrongly attributed to a company; and High-relevance articles
Released as Low-relevance articles (HRL) are false negatives, i.e. articles that the old version of
RavenPack failed to attribute to the correct company.
4
To assess the causal effect of Ravenpack, we start by focussing on LRH articles. We find that
the market indeed reacts to such false positives, but the effect does not persist. The market initially
overreacts to the incorrect information, realizes the inaccuracy, and quickly corrects after 30
seconds. This finding confirms the causal effect of RavenPack on stock prices but also suggests
that the market is quite resilient against disturbances from inaccurate news analytics.
To reinforce our finding of a causal effect, we examine the difference in the market’s reaction
to HRH and HRL articles. These two article types are of similar relevance according to the most
recent version of RavenPack, but only HRH articles were released to the market as highly relevant.
Because HRL articles were incorrectly released as not relevant, they should not trigger a causal
effect on stock prices. Thus, comparing the difference in market responses between HRH and HRL
articles allows us to assess the causal effect of RavenPack.
We find that the share of stock price reaction concentrated in the first 5 seconds after an article,
compared to the total reaction over 120 seconds, is significantly greater for HRH articles than for
HRL articles. This speed of the stock price response is 1.3 percentage points higher for HRH
articles, or 10% relative to the mean. The market not only reacts faster to HRH articles, but it also
reacts in the sentiment direction indicated by RavenPack. The RavenPack sentiment direction of
an article predicts the stock price reaction to HRH articles better than to HRL articles. This implies
that traders use RavenPack to trade in the direction of the sentiment indicator provided by the news
analytics.
In addition to the faster stock price response, we also document an increase in the share of
trade volume concentrated in the first 5 seconds compared to the two minutes after an article. This
increase in the speed of trade volume response is consistent with the theoretical prediction that
investors with a speed advantage trade aggressively on signals that they can exploit before other
5
traders (e.g., Foucault, Hombert, and Rosu (2016)). Taken together, these findings confirm that
RavenPack has a causal effect on the stock market, resulting in both prices and trading volume
reacting more quickly to the information delivered by news analytics and, thereby, improving
market efficiency.
Having established the baseline finding that news analytics affect the stock market, we ask for
which article types news analytics have the largest causal effect. We distinguish between press
releases that are directly released by companies and articles written by Dow Jones’ journalists, and
find that RavenPack has a statistically significantly larger effect for press releases. The speed of
stock price response increases 2.6% for press releases that are HRH, while it increases only 0.8%
for other HRH articles. This difference is even starker for trading volume: Being correctly covered
in RavenPack increases the speed of trade volume response by 1.4% for press releases but only by
an insignificant 0.2% for other articles. Taken together, these results confirm that the effect of
RavenPack is mainly concentrated in press releases.
Why do news analytics have a larger effect for press releases than for other articles? We show
that press releases are timelier: they are 8% more likely to be the first article of the day for the
company and 17% more likely to be a new news story rather than a reprint of an earlier story. In
addition, RavenPack sentiment is more accurate for press releases, correctly predicting the
direction of the stock price reaction in the 2 minutes around the article more often.
These findings are consistent with the notion that traders view RavenPack as being more
reliable for press releases. We extend this idea to the time series, by asking whether users of
RavenPack learn dynamically about its signal quality. We find that they do: the causal effect of
RavenPack on the 5-second return is stronger if RavenPack has been more informative in the past
6 months, measured by whether sentiment scores accurately predicted 2-minute returns following
6
the article. This finding suggests that algorithmic traders learn dynamically about the precision of
RavenPack, and that they rely more heavily on RavenPack’s sentiment scores if these scores have
been more informative in the past. Such learning could be programmed into their algorithms
(machine learning) or can come from manually updating their algorithms over time.
Next, we focus on the two ways in which traders can use news analytics. They can either use
them to get an informational edge to conduct directional trades, or they can use them to learn when
to get out of the market to avoid adverse selection or elevated order execution costs. The causal
effects of RavenPack on returns and trading volume clearly suggest that RavenPack is used for
directional trading. But is it also used to avoid adverse selection?
To examine this question, we focus on high frequency traders (HFTs). HFTs are a subset of
algorithmic traders that have invested heavily to gain a speed advantage, for example through co-
location at an exchange or hyper-fast connections between different exchanges (such as microwave
towers). Common trading strategies associated with HFTs include market making and cross-venue
arbitrage (Boehmer, Li, and Saar (2017), Zhang (2017)). While executing these strategies, HFTs
submit limit orders that are at risk of being picked off when new fundamental information reaches
the market. Therefore, we believe that HFTs are the class of trader most likely to use RavenPack
to avoid adverse selection. They would hold their usual algorithms and cancel their outstanding
orders whenever a new article about the firm is released.
To study this question, we use the NASDAQ high frequency trading data first used in
Brogaard, Hendershott and Riordan (2014) which identifies the traders that NASDAQ knows are
HFTs. Because this sample is limited to 120 stocks and just two years of data (2008-2009), a
comparison between HRL and HRH articles is not feasible. Instead we conduct a simple time-
series comparison on how the release of RavenPack to the market affects HFT trading for all
7
relevant articles. We focus on the fraction of HFT trading in the 5 seconds after an article,
standardized by the fraction of HFT trading in the 120 seconds after the article. If HFTs use
RavenPack to avoid adverse selection, we would expect this measure to decrease. Instead, it
increases by 1.8% after the release of RavenPack, indicating that HFTs make up a larger fraction
of trading in the 5 seconds after an article once RavenPack is live. We also find, in line with our
previous results, that this effect is much stronger for press releases. Moreover, we find that HFTs
mainly increase their liquidity demanding trades after an article and that these trades are
predominantly in the direction of the article sentiment. Taken together, these results suggest that
HFTs do not use RavenPack to avoid adverse selection but rather to place directional bets.
Given that even HFTs mainly engage in liquidity demanding trades after the release of an
article, we ask whether RavenPack causes a faster decline in liquidity following articles. The idea
is that the directional trades triggered by RavenPack hit existing quotes and cause liquidity to
decline following an article. We find that this is indeed the case. Both effective spreads and
Amihud illiquidity increase in the five seconds following an HRH article (compared to HRL
articles).
A series of robustness checks confirm our results. One potential concern is that HRH articles
may be systematically different from HRL articles. We address this concern in two ways. First, we
show that HRH and HRL articles are similar in terms of long-run stock price reactions and several
other characteristics. Second, we use the fact that RavenPack has back-filled the data of all versions
to February 2004 and conduct placebo tests during the time before RavenPack went live. If our
results are driven by actual differences between the two article types, rather than a causal impact
of RavenPack, then we should find significant differences in price reactions before RavenPack
went live. However, for all tests before RavenPack went live we find insignificant differences
8
(between HRH and HRL articles). Moreover, the stock price reactions to HRH and HRL articles
start to diverge precisely when RavenPack went live, and the resulting increase in the difference
between HRH and HRL articles is significant. All of this suggests that our results are robust.
In this paper we show that many algorithmic traders, including HFTs, use RavenPack for
directional trading. This results in RavenPack having a significant impact on the market in terms
of returns, trading volume, and liquidity. This effect goes beyond the underlying influence of the
news itself. While our study can only detect the effect of RavenPack, there are other providers of
news analytics, and traders may conduct algorithmic news processing in house. Thus, the total
effect of algorithmic news processing is likely much larger than the effect of RavenPack measured
in this paper. Also, given that RavenPack is the leading provider of news analytics, we expect the
results to carry through to wide-subscription news analytics services more generally.
Our results contribute to four major strands of literature. First, we contribute to the literature
on the causal effect of media on the stock market.
1
Methods to address the endogeneity of media
coverage include exogenous scheduling of journalists (Dougal, Engelberg, Garcia, and Parsons,
2011), local media coverage and its delay due to extreme weather (Engelberg and Parsons, 2011)
and newspaper strikes (Peress, 2014). We add to this literature in three ways. First, we study news
analytics, rather than news articles themselves. News analytics are special in that they are a
derivative of news articles that contain less information than the article itself. Their only advantage
is that they are easier to process algorithmically. Our results show that in the age of algorithmic
trading, processability is just as important as informational content. Second, we study the effect of
1
There is a wider literature on media and stock markets including for example Chan (2003), Tetlock (2007, 2011),
Fang and Peress (2009), Griffin, Hirschey, and Kelly (2011), Boudoukh et al. (2016), Loughran and McDonald (2013),
Garcia (2013), Ferguson et al. (2015), Hu, Pan, and Wang (2017)). For a review on textual analysis in finance see
Kearney and Liu (2014).
9
such news analytics on algorithmic traders rather than private investors. This focus increases the
policy relevance of our findings in a regulatory environment that is increasingly focused on news
analytics. Third, we show that the impact of news analytics on prices are particularly important for
press releases, which are not the subject of the prior studies.
Second, we contribute to the literature on news analytics. Prior papers in this literature study
the correlation between the market and news analytics without passing judgment on whether there
is a causal impact of news analytics on the market (e.g. Dzielinski and Hasseltoft (2017), Riordan,
Storkenmaier, Wagener, and Zhang (2013), Gross-Klugmann and Hautsch (2011), Sinha (2016),
Heston and Sinha (2016)). In contrast, our paper is the first to show the causal impact of news
analytics on stock markets.
Third, we contribute to the growing empirical literature on algorithmic and high frequency
trading.
2
Several papers show that high frequency traders use information from order flow (e.g.
Hirschey (2018) or information from related asset prices (e.g., Chaboud et al. (2014), Boehmer,
Li, and Saar (2017), Zhang (2017)). In contrast to these studies, we show that HFTs do not only
trade on market information, but also enter directional bets based on news analytics, which contain
new, company-specific information that is not yet reflected in any market prices.
Fourth, our results are consistent with recent models of high frequency trading in which some
traders have an informational advantage. For example, Foucault, Hombert, and Rosu (2016) model
a situation in which a speculator receives information one period ahead of the market maker in a
2
Examples of this literature include Brogaard, Hendershott and Riordan (2014), Boehmer, Fong and Wu (2015),
Hendershott and Riordan (2013), Hendershott, Jones, and Menkveld (2011), Baron, Brogaard, Hagströmer, and
Kirilenko (2017), Menkveld (2013), Jovanovic and Menkveld (2010), Riordan and Storkenmaier (2012), Boehmer,
Fong, and Wu (2015), Hasbrouck and Saar (2013), Benos and Sagade (2016), Clark-Joseph (2013), Hirschey (2018),
Brogaard et al. (2014), Chordia, Green, and Kottimukkalur (2017). A survey of this literature is provided by Jones
(2013).
10
set-up similar to Kyle (1985); in Martinez and Rosu (2013) some agents have a short-lived
informational advantage; and in Dugast and Foucault (2017), speculators face a trade-off between
processing a signal faster or more accurately. Faster traders in these models make markets more
informationally efficient, but also more unstable. We find support for both effects.
2. Test design, identification strategy, and data sources
In this section we first describe the RavenPack news analytics data and how it is used in our
identification strategy and tests. After briefly describing our stock market data, we then present
summary statistics for the variables used in our tests. Variable definitions are in Appendix 1.
2.1 RavenPack
RavenPack provides real-time news analytics based on the Dow Jones (DJ) Newswire. This service
analyzes all the articles on the DJ Newswire with a computer algorithm and delivers article-level
relevance and sentiment metrics to its users. It determines which companies are mentioned in the
article, how relevant the article is to the company and reports different sentiment indicators about
whether the article is good or bad news for the company. The latency – i.e. the time from the
release of the DJ Newswire to the release of the RavenPack metrics – is approximately 300
milliseconds. RavenPack claims it has the “timeliest company sentiment indicators in the
marketplace.”
3
As such, RavenPack is ideally suited for the use of traders engaging in algorithmic
news trading.
2.1.1 Ravenpack – definition of variables
We extract from RavenPack the following variables. Article Category is a variable determining
the topic of the article and the role played by the company in the article. For example, Article
3
“RavenPack Enables Trading Programs with Sentiment on 10,000 Global Equities,” RavenPack press release from
May 28, 2009.
11
Category might be “acquisition – completed – acquirer” for a company announcing the completion
of an acquisition of another company or “rating – change – negative – rater” for a rating company
that just downgraded another company. The identification of the news topic is based on a purely
algorithmic approach, and a large percentage of articles cannot be classified in this way. Article
Category Identified is a dummy variable equal to 1 if Article Category is identified by RavenPack,
and zero otherwise.
There are two major sentiment scores in RavenPack. The Composite Sentiment Score (CSS)
is based on several individual RavenPack sentiment measures. It takes a value ranging from 100
(positive) to 0 (negative), where 50 is a neutral article. It is available for each article. The Event
Sentiment Score (ESS) is coded in the same way as CSS, but available only if the category of the
article can be identified. We aggregate these two scores into a single sentiment variable called
Sentiment Direction, which is primarily based on ESS and uses CSS only if ESS is either missing
or equal to 50 (neutral).
Relevance is an index provided by RavenPack that indicates the relevance of an article to the
company. Relevance takes values ranging from 0 (least relevant) to 100 (most relevant). If the type
of the article can be identified and the company plays an important role in the main context of the
story – e.g. is an acquirer or announces a buyback – then Relevance is 100. If the company is
mentioned in the title, but the type of article cannot be identified, then Relevance ranges between
90 and 100. If the company is mentioned, but plays an unimportant role, then it gets a low
Relevance score – e.g., a bank advising an acquisition might get a score of 20. We would not expect
such articles to affect the bank’s stock price very much.
In line with this, RavenPack recommends “filtering for Relevance greater than or equal to 90
as this helps reduce noise in the signal”. To examine this claim, Figure 1 plots the market reaction
12
to news as a function of Relevance. We plot the cumulative returns relative to news events from
April 1, 2009 to September 10, 2012. We multiply returns by the article’s sentiment direction. The
articles with Relevance greater than 90 do indeed have an important effect on stock prices, but
there is no reaction to articles with Relevance below 90. Thus, we will refer to articles with
Relevance below 90 as low relevance. This analysis suggests that RavenPack is good at identifying
both relevance and sentiment of an article.
That the reaction to high relevance articles starts about 60 seconds before the article suggests
that some of the news events are covered in other news sources before they are covered in the DJ
Newswire (used by RavenPack). Cases where the DJ Newswire is not the first to report an event
should only work against us by making it more difficult to find a causal impact of RavenPack. We
have no reason to believe that this issue should bias the results because it should be unrelated to
whether RavenPack makes a mistake interpreting the article. While some trades in the 5 seconds
after a RavenPack article may be due to human traders reacting to earlier coverage of the news
elsewhere, this trading should affect both HRL and HRH articles. Thus the additional trading
following HRH articles (relative to HRL articles) should only be due to algorithmic traders
reacting to the coverage in RavenPack itself.
2.1.2 Ravenpack – test design using different product versions
RavenPack released its first version (v. 1.0) to the market on April 1, 2009,
4
5
and a revised version
of the service (v. 2.0) with additional features on June 6, 2011. The most recent version we use (v.
4
Even though the official release date of the RavenPack service was May 2009, some customers had access to the
service as early as from April 1, 2009. Thus, we refer to April 1, 2009 as the introduction of RavenPack. Before April
2009 RavenPack had a pre-existing service that also released sentiment information on the Dow Jones News Wire.
However, this service was meant for longer term news analysis, such as charting sentiment over several days. The
prior service was not provided timely enough to be used at high frequency.
5
RavenPack 1.0 was actually released on Sept 6, 2010. A predecessor to v.1.0, that was similar to v.1.0, is the version
that was released on April 1, 2009. This predecessor version was not made available to us, but RavenPack confirmed
that it was very similar to RavenPack 1.0.
13
3.0) was released on September 10, 2012. RavenPack has provided us with data from each of the
release-specific algorithms, each having been back-filled to February 2004. RavenPack does not
continuously update its algorithm, so as not to distort its customers’ trading strategies which might
be based on specific variable definitions. Rather, RavenPack rolls out any changes to its algorithm
when releasing a new version, meaning that stock-specific metrics from the three releases can
sometimes differ.
6
These differences are often related to the way companies are identified in an
article and how the relevance of an article to a company is determined.
7
Thus, there are articles
that might be associated with a particular company in one RavenPack release, but not in another.
Such differences in the relevance of articles to companies in different versions provide the basis
for our tests. Assuming the most recent version of RavenPack (v. 3.0, hereafter New RavenPack)
is the most accurate, we can identify inaccuracies in RavenPack 1.0 and RavenPack 2.0 (hereafter
Old RavenPack) that were released to the market. If the market reacts to these inaccuracies, it is
an indication of a causal effect of RavenPack on the stock market.
Our analysis can be thought of as assuming two types of traders: Algorithmic traders that
subscribe to RavenPack and human traders that manually read the article to determine its content.
Further, we assume that human traders can more precisely derive the relevant signal from the
article, while algorithmic traders have an advantage in terms of speed (a setting modelled by
Dugast and Foucault (2017)). This means that RavenPack allows its subscribers to trade faster on
a possibly less precise signal. In the short run, when only algorithmic traders can react to news,
6
Because the algorithm is proprietary, we do not know exactly what changes RavenPack implemented but some
examples of articles where the two versions disagree are provided in the Internet Appendix.
7
In addition, the number of companies covered by RavenPack has also increased between releases. There are 156
companies (3%), which are only covered in New RavenPack. We ensure by using company fixed effects that this
difference in coverage is not driving our results.
14
RavenPack will have the largest impact; while in the long run human traders determine the price
reaction because their signal is more precise.
In the empirical implementation we choose specific time intervals to constitute the short and
long run. We define the short run to be 5 seconds, because this is long enough to capture the full
reaction of algorithmic traders (and accommodates slower algorithmic traders that are not co-
located and not trading within milliseconds), but is too short for a human trader to read an article,
process it and make a trading decision based on it. We choose two minutes as the long run because
this permits enough time to read an article and trade on it, whereas longer time windows will be
more affected by noise. In the Internet Appendix, we provide robustness checks in which we use
both 1 and 10 seconds for the short run and 5 minutes for the long run.
We define the following article types that we also list in Panel A of Table 1. High relevance
article Released as High relevance article (HRH) is defined as an article classified as relevant in
both Old and New RavenPack. We predict that such a correctly released article creates a fast and
persistent market reaction. High relevance article Released as Low relevance article (HRL) is
defined as an article with high relevance in New RavenPack, but incorrectly assigned low
relevance in Old RavenPack. Low relevance means either the article was not assigned to the
company or the relevance score was below 90. We expect an HRL article to have a similar long-
run market reaction as an HRH article because they are of similar relevance according to New
RavenPack. However, we would expect a slower market reaction to an HRL article as it was not
released originally as relevant. Low relevance article Released as High relevance article (LRH) is
an article that was incorrectly released to investors as having high relevance but has low relevance
in New RavenPack. For these articles we expect an initial overreaction of algorithmic traders,
which might later be reversed by human traders. Examples of all three article groups are provided
15
in the Internet Appendix. A fourth article category is Low relevance articles Released as Low
relevance articles (LRL); these articles have a relevance score below 90 in both versions.
8
We do
not expect much market reaction to LRL articles.
These predictions allow for two possible empirical set-ups using the two types of mistakes.
First, we could study overreaction to false positives by examining LRH articles or we could study
underreaction to false negatives by examining HRL articles. In both cases, it would be desirable
to have a control group of articles where RavenPack determined the relevance accurately. Having
a control group would enable us to include time and firm fixed effects and a host of control
variables. The ideal control group would consist of articles that are similar in terms of importance
and content but have been released to the market with a different relevance. The obvious candidates
are articles that have the same relevance (according to New RavenPack) but were released to the
market as having a different relevance (in Old RavenPack). Thus, the candidate control group for
HRL articles are HRH articles and the candidate control group for LRH articles are LRL articles.
For these control groups to be appropriate, we would require HRL and HRH articles to be of
similar relevance and LRH and LRL articles to be of similar relevance. This is the case if New
RavenPack determines the relevance of an article accurately and if Old RavenPack contains no
information on the relevance of the article over and above the information contained in New
RavenPack. This is a fairly strong assumption, but it is testable. Because we have data before
RavenPack went “live” in 2009, we can compare the market impact to the different types of articles
during the time period when RavenPack could not have had any causal market impact, because it
was not yet “live”. To examine the market impact of the different article types, we regress absolute
8
LRL articles also include articles that have a relevance score below 90 in either Old RavenPack or New RavenPack
and are not assigned to the company in the other version.
16
return and turnover in the two minutes after the article’s release on dummy variables equal to 1 for
HRH, HRL and LRH (with LRL being the omitted category). To control for firm- and time-specific
effects, we include firm, date and hour-of-the-day fixed effects.
The results are presented in Panel B of Table 1. When we test the difference between LRH
and LRL articles, we find a large and significant difference for both absolute returns and turnover,
suggesting LRH articles are significantly more important than LRL articles. Thus, a test of
overreaction comparing LRH and LRL articles is not possible because the two article types are
fundamentally different. Thus, instead of comparing LRH and LRL articles, we rely on graphical
evidence and also compare the reaction to LRH articles before and after RavenPack went “live”.
When we test the significance of the difference between the coefficients on HRH and HRL,
however, we find the difference is small and insignificant for both turnover and absolute returns.
This finding suggests that HRH and HRL articles have very similar relevance. Thus, any difference
in their market reaction can be attributed to the causal effect of RavenPack. Therefore, most of our
tests are based on the comparison between these two article types. To ensure further that
differences in importance between HRH and HRL are not driving our results, we control for article
characteristics and study the speed of market reaction, i.e. the size of the short run reaction relative
to the long-run reaction. We also conduct placebo checks for our tests showing that our results are
not driven by differences between HRH and HRL articles, but by the causal effect of RavenPack.
2.2 Stock market data
We use intraday quotes and trade data from TAQ.
9
We use the TAQ National Best Bid and Offer
(NBBO) file provided by WRDS for quotes. As a first step, we aggregate trading volume at the
9
We use the usual filters of excluding all trades with zero size, negative prices, correction code different from 0 and
bid ask quotes where the bid is above the asked.
17
frequency of one second, and compute second-by-second returns based on end-of-second bid-ask
midpoints. We use bid-ask midpoints rather than trading prices to avoid bid-ask bounce effects.
Even after this aggregation, the data for all stocks in our 8-year sample is far too large to be used
in a second-by-second panel set-up. But we are interested only in the market reaction around
specific company news events, so we limit our analysis to a few minutes around these events. This
simplification allows us to study all US common stocks over the full 8-year sample period.
To control for the overall market movements taking place during this period, we compute a
second-by-second intraday market index from the total TAQ universe. We compute second-by-
second returns, turnover, and value-weighted volatility for the market index. We also compute
returns for industry-specific indices for the 12 Fama French industries. The details of the index
construction are explained in Internet Appendix 1. To control for stock-specific information, we
use the CRSP daily stock file and compute the prior month’s return, volatility, turnover, Amihud
(2002) illiquidity measure, and market capitalization.
We employ the following filters: To be included in our sample, a stock must be covered in
CRSP and TAQ, must have SHRCD 10 or 11, must have a beginning of the day stock price of at
least $1 and must have a beginning of the day percentage bid-ask spread of less than 10%. We
exclude articles that occur outside trading hours or in the first or last 20 minutes of trading in the
day. To avoid distortions from overlapping windows around articles, we exclude articles for which
the company had an article in the prior 15 minutes. We also exclude four companies that appear
in articles mainly as information providers: McGraw-Hill, NASDAQ, CME and Moody’s.
Because we need an initial bid-ask midpoint to compute a first return and because we want to avoid
a stock’s turnover influencing the stock price we measure, we use seconds t−480 to t−1 as a burn-
18
in period. Only articles for which the stock has a quote in those 8 minutes before the article are
included in our analysis.
2.3 Summary statistics and comparison between HRH and HRL
The final sample consists of 321,912 article-firm combinations, starting with the release of
RavenPack 1.0, over the period April 1, 2009 to September 10, 2012. In Panel A of Table 2, we
report descriptive statistics for all our variables for the combined sample of articles classified as
HRH and articles classified as HRL.
As alluded to previously, a concern is that the information content of HRH and HRL articles
might be different. Therefore, we compare their difference in terms of observable variables in
Panel B. For this purpose, we regress each article characteristic on a dummy variable, D(HRH)
equal to 1 if the article is HRH, and relevance, category, hour and date fixed effects. We report the
coefficient of D(HRH) as well as a t-statistic two-way clustered at the firm and date level. There
is no statistical significant difference between the HRH and HRL articles in terms of firm size,
sentiment scores, time since the last article, turnover, and illiquidity. Most importantly, we find no
evidence that HRH are more important than HRL articles: the absolute returns both over the 2
minutes following an article and on the full trading day of the article are actually (insignificantly)
lower for HRH articles. The only significant difference is that companies in HRH articles have a
slightly lower return (0.03%) in the prior month than those associated with HRL articles, and that
HRH articles cover fewer firms per article. However, these differences are small in economic terms
(0.05 and 0.22 standard deviations) and we account for these differences with control variables in
all our subsequent tests. In addition, we show in the Internet Appendix that our results are robust
to including fixed effects for the number of firms mentioned in the article. That the characteristics
of HRH and HRL articles are similar alleviates worries that our results are driven by differences
19
in the article types. In addition, we run placebo tests to confirm that unobservable differences are
not driving our results.
3. News analytics and impact on the market
3.1 News analytics and temporary price distortions
Here we examine whether inaccuracies in news analytics lead to temporary price distortions, i.e.
to an overreaction in stock price that is afterwards reversed. As explained in Section 2.1.2, we
expect the market to overreact to LRH articles, i.e. articles that New RavenPack identifies as
having low relevance, but that were incorrectly released as having high relevance in Old
RavenPack.
We first consider graphical evidence. In Figure 2, we compare the market reaction of articles
HRH and LRH articles. We focus on the cumulative returns from t−60 to t+120 seconds around
the article (measured relative to t=0). We focus on signed returns—i.e. returns multiplied with the
sentiment direction of the article—to be able to combine positive and negative news in one
analysis. We exclude articles with neutral sentiment. Figure 2 shows that the market overreacts to
LRH articles. In the short-run these articles have a price reaction that is very similar to HRH
articles. However, after approximately 30 seconds – a reasonable time for a fast human trader to
verify that no important news for the stock has arrived – the stock price reaction to LRH articles
starts to retreat. After approximately 2 minutes, a large part of the short-run reaction to these
articles has reversed.
10
In contrast, articles classified as HRH have a longer-term effect on price,
lasting more than two minutes. This finding is consistent with a causal effect of RavenPack that
leads algorithmic traders to trigger an initial overreaction to the article that is then corrected by
10
We would never expect full mean reversion given that LRH articles contain at least some information (see Table 1
Panel B).
20
other traders. While the magnitude of this effect is only about 1 basis point on average, this is due
to the fact that the average price reaction to company-specific news is fairly small. The average
absolute return reaction in the first 2 minutes after an article is only 11.4 basis points and the
average signed return is only 1.9 basis points (see Table 2).
Next, we provide a multivariate analysis. The problem in studying LRH articles in a regression
set-up is that we do not have an appropriate control group for these articles as they are more
relevant than LRL articles, but less relevant than HRH articles (see Table 1 Panel B). Therefore,
we use LRH articles from the period before RavenPack went “live” as the control group.
11
During
this period, the fact that these articles would have been marked as relevant by RavenPack should
not affect stock prices. However, after the introduction of RavenPack, LRH articles can have a
causal effect on the market, which should lead to a short-run overreaction to the article. Thus, we
study whether LRH articles have a stronger short run stock price impact and a larger reversal after
RavenPack goes “live”, as compared to before. We estimate the following article-level regression
including only LRH articles:
where Sentiment Direction indicates whether the article is positive or negative news and
RavenPack Release is a dummy variable equal to 1 after RavenPack went “live” in April 2009.
The coefficient of interest is the interaction between Sentiment Direction and RavenPack Release.
In addition, we include various combinations of control variables and fixed effects. To control for
stock-specific information, we use its market capitalization, return, volatility, and turnover
11
We confirm in IA-Table 1 in the Internet Appendix that the relevance of LRH articles does not change after the
introduction of RavenPack.
21
measured over the prior month, and our illiquidity measure based on Amihud (2002). For brevity,
the coefficients on these control variables are reported only in Internet Appendix 3. To control for
characteristics of the news announcement, we add fixed effects for the article category (e.g.
mergers and acquisitions), the relevance score, and the hour during the day in which the article
was released. In regressions 3 and 6, we add as additional controls the absolute return, turnover,
and volatility for each industry and the market over the horizon from t−1 to t+5 seconds around
the article. All standard errors are two-way clustered at the firm and date level. Appendix 1
contains a detailed description of all variables.
We present the results in Table 3. In regressions 1-3, the dependent variable is the short-run
stock return from 1 second before to 5 seconds after the article. We find that the short-run return
associated with LRH articles is significantly more positively correlated with the sentiment of the
article after RavenPack went “live” in 2009. Because the relevance of LRH articles was the same
both before and after the introduction of RavenPack, this finding indicates an overreaction to these
articles. Indeed, it seems plausible that algorithms would trade in the direction of the article’s
sentiment, because RavenPack (incorrectly) labelled such articles as highly relevant.
Next, we ask whether this overreaction is subsequently reversed. In regressions 4-6, we use
the stock price reaction from 6 to 120 seconds after the LRH article as the dependent variable. We
find that it is more negatively correlated with the article sentiment after RavenPack went “live”,
consistent with a reversal. While this result is not statistically significant, the negative magnitude
of this coefficient is about the same as the positive magnitude of the coefficient in regressions 1-
3, implying that almost all of the short run overreaction is reversed in the two minutes after the
article. This result is not significant because two-minute returns are much noisier than the five-
second returns studied in regressions 1 to 3.
22
Taken together, our graphical and regression analyses of LRH articles results provide
evidence that inaccuracies in news analytics can cause short term overreaction that is afterwards
reversed.
3.2 News analytics and speed of stock price response
In this section, we study whether news analytics improve market efficiency by increasing the speed
with which stock prices and traders react to news. For this purpose, we focus on false negatives,
i.e. HRL articles that are highly relevant according to New RavenPack, but were released as having
low relevance in Old RavenPack. For these articles we have a good control group in the form of
articles that have been reported as having high relevance in both versions (HRH). Comparing the
market reaction to those two article groups allows us to see whether the market underreacts to
relevant news when RavenPack does not classify it as relevant.
3.2.1 Regression analysis – speed of stock price response
We consider two alternative analyses for market reaction. First, we examine whether stock
prices respond faster to HRH articles irrespective of the direction of the reaction. Then we study
whether the sentiment of HRH articles predicts the directional stock price response better than the
sentiment of HRL articles. For the first analysis, we define Speed of Stock Price Response as:
over the 120 seconds around the news event.
12
This variable
measures the amount of the two-minute price change that takes place in the first five seconds after
the news release. It is in the spirit of DellaVigna and Pollet (2008). It captures the degree of under-
reaction by decomposing the market reaction into its short- and long-term components. The higher
12
We use
rather than
in the
denominator to constrain the variable between 0 and 1 rather than to allow it to approach infinity in cases where
is close to zero.
23
the value of Speed of Stock Price Response, the more the reaction to the news event concentrates
in the first few seconds after the event – i.e., the less under-reaction. We run the following article-
level regression including only HRH and HRL articles:
where D(HRH) is a dummy variable that equals one if the article was released to the market as
highly relevant (HRH) and zero if it was (incorrectly) released as having low relevance (HRL).
The regression is estimated at the article level, thus allowing for both HRH and HRL articles that
were released for the same firm or on the same day. This allows us to control in all regressions for
unobserved heterogeneity with firm fixed effects and daily fixed effects (
and
). In addition, we
include the same article-level fixed effects and control variables as before.
We report the results in Table 4. In regressions 1 to 3, we estimate our main specification
during the time in which RavenPack was live (Apr 1, 2009 – Sept 10, 2012). We observe a positive
and significant relation between Speed of Stock Price Response and D(HRH), indicating that the
stock price response is much quicker for an HRH article than for an HRL article. This result holds
across all the different specifications and samples. It is not only statistically significant, but also
economically relevant. HRH articles exhibit a Speed of Stock Price Response that is 1.3 percentage
points higher (10% relative to the mean). We find similar results if we compute Speed of Stock
Price Response using market-adjusted and industry-adjusted returns (reported in Internet
Appendix 3). This finding suggests that news analytics increase market efficiency by increasing
the speed of reaction to news.
Although we showed in section 2.3 that HRH and HRL articles are similar along most
dimensions, one potential concern in this set-up is that the results are driven by the two article
categories having different informational content. To address this issue, we use the fact that
24
RavenPack has back-filled the data to February 2004. If our results are driven by general
differences in the two categories, then there should be a difference in stock price reaction before
RavenPack went live. In regressions 4 to 6, we report a placebo test for the time period where
RavenPack was not yet released to investors (February 1, 2004 ‒ March 31, 2009). This placebo
test does not show a statistically significant relation between D(HRH) and the Speed of Stock Price
Response, thereby confirming that our main test is appropriate.
Another potential concern is that there might be a trend in the difference of informational
content between HRH and HRL articles, which is driving our results. To address this concern, we
examine the relation between Speed of Stock Price Response and D(HRH) for different years
before and after RavenPack went live. We follow Gormley and Matsa (2011) and plot in Figure 3
the point estimates of a modified version of regression 3 in Table 4, in which we interact D(HRH)
with yearly fixed effects
13
. We plot the coefficients of the interaction of D(HRH) with one-year
dummy variables in Panel A. In Panel B, we report the same regression but interacting D(HRH)
with two-year dummy variables. We report 95% confidence intervals for the coefficients in both
panels. In Panel C, we report the simple difference between Speed of Stock Price Response for
HRH and HRL articles without any controls over different years.
The results confirm that RavenPack has a causal effect on the Speed of Stock Price Response.
Before the introduction of RavenPack, the difference between HRH and HRL hovers around zero
and there is no obvious time trend. After the introduction of RavenPack, the difference is much
larger. This suggests the delivery of news analytics by RavenPack has an impact on the market
13
Because RavenPack went “live” in the second quarter of 2009, we assign the first quarter of every year to the prior
year. This way, years 2004 to 2008 were entirely before RavenPack went live, while years 2009 to 2011 were
completely after RavenPack went live.
25
that is distinct from the underlying informational content of the news and that our results are not
driven by a spurious trend.
3.2.2 Regression analysis – directional stock price response
We now ask whether there is a relation between the stock price response and the sentiment
direction of the news. That is, does the magnitude of the RavenPack-related stock price response
(via correctly-labelled HRH articles) depend on whether the news is positive, negative, or neutral?
For this purpose, we ask whether the sentiment indicator in RavenPack better predicts the short
run stock price reaction if an article is correctly classified as relevant (HRH) in RavenPack. We
use the following article-level regression specification:
We use the same fixed effects as before, but exclude any sentiment-related control variables as the
effect of sentiment will be captured by Sentiment Direction.
We report the results in Table 5. We observe a significant positive relation between returns
and the interaction of D(HRH) and Sentiment Direction at the time where RavenPack was live
(regressions 1 to 3). That is, the RavenPack-induced stock price reaction is significantly more
positive for positive news stories than for negative news stories. This result holds across all the
different specifications. Similar results for market- and industry-adjusted returns are reported in
Internet Appendix 3. As before, the placebo test in Regressions 4 to 6 shows no effect on returns
before RavenPack was live. These results confirm that news analytics have a directional impact
on stock prices over and above the one of the underlying news.
While RavenPack improves market efficiency, the information contained in RavenPack does
not fully get incorporated into prices instantaneously. In the Internet Appendix in IA-Table 2, we
26
show that RavenPack sentiment predicts returns from t+5 to t+120 seconds after the article by
about 1.35 basis points. Comparing this result with the causal effect of RavenPack on t-1 to t+5
second returns of 0.45 basis points in Table 5, we estimate that about 25% of the information of
RavenPack gets incorporated into prices in the first five seconds after the article.
14
3.3 News analytics and trade volume response
We have showed above that news analytics increase the speed of price adjustment after news is
publicly released via the DJ Newswire. While the DJ Newswire constitutes a public signal,
RavenPack enables faster reaction to such signals for its subscribers. Such a reaction speed
advantage is modelled by Foucault, Hombert, and Rosu (2016) who predict that investors trade
very aggressively when they receive a signal earlier than other market participants.
Therefore, we investigate whether the faster stock price response to an HRH article is
accompanied by a faster trade volume response as well. We define Speed of Trade Volume
Response as:
. The variable is defined using the same intervals as Speed of Stock
Price Response. It captures the amount of trade volume that is concentrated in the first 5 seconds
after the news event relative to the trading volume in the two minutes following the news event.
We regress Speed of Trade Volume Response on D(HRH) using the same fixed effects and control
variables defined above.
We report the results in Panel A of Table 6. We find a strong positive and significant relation
between Speed of Trade Volume Response and D(HRH). This result holds across all specifications.
14
This calculation assumes that the information of the article is fully incorporated into prices within 2 minutes. If the
market was fully efficient, RavenPack sentiment information would not predict returns from t+5 to t+120 seconds
after the article (because all RavenPack sentiment information would be instantaneously incorporated into prices).
Instead, we observe that only 25% (0.45/(0.45+1.35)) of this information is incorporated into prices. This finding
also shows that it remains profitable to trade on RavenPack sentiment information.
27
Speed of Trade Volume Response is 0.5 percentage points larger for HRH articles than for HRL
articles, or 9% relative to the mean. The placebo test shows no significant difference in the speed
of trade volume response between HRH and HRL articles before RavenPack went live.
We expect traders using RavenPack to trade in the direction of the article sentiment, so we
now examine directional trading volume. Using the methodology of Lee and Ready (1991), we
first determine whether a trade is initiated in the direction of the article, i.e. buyer initiated for
positive articles and seller initiated for negative articles.
15
We then define Speed of Directional
Trade Volume Response as
, and use it as the dependent
variable using the same regression specification (Panel B, Table 6). Speed of Directional Trade
Volume Response is 0.4% larger for HRH than for HRL articles. Comparing it to the 0.5 percentage
point increase in Speed of Trade Volume Response suggests that close to 80 percent of the increase
in trading volume in the 5 seconds after the article is due to trading in the direction of the article.
This finding suggests that RavenPack triggers fast and informed trading.
Taken together, the results in this section show that stock prices react faster and traders trade
more aggressively after the release of articles covered in RavenPack. Combined, these results
confirm that news analytics have a measurable impact on stock prices in addition to the information
content of the news itself and improve price efficiency.
15
Different from Lee and Ready (1991), we use the quote at the end of the previous second as the prevailing quote
rather than the quote 5 seconds ago to account for the faster trading processes today.
28
4. News analytics and article type
4.1 Types of News: Press releases vs other media-reported news
Our analyses above, suggest that RavenPack has a causal effect on stock prices and turnover and
thereby makes markets more efficient. In this section, we dig a little deeper and examine for which
types of articles the effect of RavenPack is most relevant. The main distinction between different
articles on the Dow Jones News Wire is whether they are press releases that are directly released
by companies or articles that are written by Dow Jones’ journalists. Press releases make up 23.7%
of articles in our sample.
Ex ante, it is not obvious whether we would expect RavenPack to have a stronger or weaker
effect for press releases. It might have a stronger effect if press releases are timelier or include
more important information. On the other hand, it might have less of an effect if press releases
contain less relevant information or the positive spin given by the company makes it difficult for
RavenPack to determine the true sentiment of the article.
We examine this question by re-estimating specifications from Tables 4 to 6 but interacting
D(HRH) with D(Press Release), which is a dummy variable equal to one if the article is a press
release (and zero otherwise). We report the results in Table 7. In regressions 1 and 2, the dependent
variable is Speed of Stock Price Response. The coefficient of the interaction between D(HRH) and
D(Press Release) is economically large at about 1.8% and statistically significant at the 5%
significance level. This suggests that RavenPack has a significantly stronger effect on the Speed
of Stock Price Response for press releases than for other articles. In fact, D(HRH) is only
marginally significant, suggesting that our findings are almost exclusively due to press releases
and are insignificant for other articles.
29
In regressions 3 and 4, we find similar results using Speed of Trade Volume Response as the
dependent variable. Here, the difference between press releases and non-press releases is even
larger: Being correctly covered in RavenPack increases the Speed of Trade Volume Response by
about 1.4% for press releases but only by an insignificant 0.2% for other articles.
Finally, we study the directional effect of RavenPack in regressions 5 and 6 of Table 7.
Studying the directional effect is interesting because it requires RavenPack to be able to identify
the sentiment of the article correctly, which may be more difficult for press releases that usually
entail a positive spin. The dependent variable is the signed return from t-1 to t+5 seconds around
the article. The coefficient of the interaction between D(HRH) and D(Press Release) is once again
positive and statistically significant, suggesting that also the directional effect of RavenPack is
mainly driven by press releases. Taken together, our results suggest that RavenPack mainly affects
the market reaction to press releases while having limited effects on other articles.
In Panel B of Table 7, we repeat the same analysis using the time period before RavenPack
was live. In this robustness check, we do not find significant results and the coefficient of the
interaction between D(HRH) and D(Press Release) is generally small and insignificant. This
finding suggests that our results are not driven by a mechanical difference between press releases
and other articles.
4.2 Why are press releases special?
In this section we ask why RavenPack has a larger effect for press releases than for other articles.
We look at several characteristics by comparing their mean for press releases and non-press
releases in Table 8. We start by comparing whether there are more incorrectly classified articles
for press releases than for other articles. We find that this is not the case: for both press releases
and other articles 97% of articles are correctly released as relevant in both versions of
30
RavenPack, i.e. they are HRH articles. This suggests that the results are not driven by a different
fraction of HRH articles among press releases.
Next, we look at a variety of market reaction measures. There is no clear sign that there is a
stronger market reaction to press releases. While the return reaction in the 2 minutes after the
article is 2 basis points higher, the turnover reaction over the same time period is actually lower.
Similarly, the return reaction over the whole trading day is also lower for press releases. This
suggests that the results are not driven by the fact that press releases are generally more
important.
We confirm that press releases have a more positive sentiment on average, which makes
sense considering that companies give stories a positive spin and might not release negative news
during trading hours.
Finally, we show that press releases are less stale: They are 8% more likely to be the first
article for the company on that day and 17% more likely to be a new news story rather than a
reprint as identified by RavenPack. This may be one of the explanations why RavenPack has a
larger effect for press releases. The other is that RavenPack is more accurate for press releases.
For press releases, the RavenPack sentiment predicts the market reaction in the 2 minutes around
the article correctly 56% of the time, while it is only 53% for other articles. This difference is
statistically significant. Taken together, our results suggest that RavenPack has a larger market
impact for press releases because they are timelier and RavenPack is better in evaluating their
sentiment.
4.3 Learning about precision in news analytics
We find that RavenPack has a stronger effect for press releases in part because it is more accurate
for press releases. This suggests that algorithmic traders have learned that RavenPack is more
31
accurate for press releases and thus use RavenPack more for trading on press releases. This finding
raises the more general question whether RavenPack users also learn dynamically about the
changing signal precision of RavenPack over time. Such learning could be programmed into their
algorithms (machine learning) or come from manual updates to the algorithms. If algorithmic
traders learn about the precision of RavenPack, we expect them to rely more on RavenPack’s
sentiment indicators if these indicators were more informative in the past. In that case, there should
be a stronger price reaction to news analytics at times where news analytics have been more
informative in the past.
This raises two issues: how to best measure the informativeness of RavenPack and whether
this informativeness is persistent and thus predictable. We choose to measure informativeness by
studying how well the Sentiment Direction of RavenPack predicts stock returns in the two minutes
following an article.
16
In IA-Table 3 in the Internet Appendix, we show that informativeness is
persistent. In particular, Sentiment Direction predicts two-minute post-article returns better if it
was better in predicting these returns in the prior 3-6 months. Thus, we define Past Informativeness
as the average signed two-minute post-article return for all articles during the previous six months.
We study whether RavenPack subscribers trade more on articles with higher Past
Informativeness by estimating the following article-level regression specification:
The explanatory variable of interest is the interaction between D(HRH) and Past Informativeness.
Its coefficient tests whether RavenPack’s causal effect on 5-second signed returns following the
article is stronger if RavenPack has been more informative in the past.
16
It may seem more intuitive to use the number of “mistakes” RavenPack makes in assigning the relevance score.
However this information is not available to the market because it uses the New RavenPack dataset, which was only
released years later. Therefore, traders could not have conditioned their trading on the number of “mistakes”.
32
The results are presented in Table 9. We find a significant increase in the causal effect of
RavenPack on 5 second signed stock returns if Past Informativeness is high. In regressions 4 to 6,
we show in a placebo test that this effect does not occur before Ravenpack went “live”. In IA-
Table 4 in the Internet Appendix, we report robustness checks using different definitions of Past
Informativeness. In particular, we use 3 months instead of 6 months and we use Stock-Level Past
Informativeness, where we only use articles for that specific stock in the past 6 month when
measuring past informativeness. Furthermore, we use informativeness based on whether the
direction of the news and the two-minute stock price reaction agree. This last robustness check
shows that our results are not driven by any persistence in volatility. For all three alternative
measures of informativeness, we obtain similar results as those presented in Table 9.
In total, these results suggest that algorithmic traders learn dynamically about the precision of
RavenPack and base their trades more on RavenPack’s sentiment scores if these scores have been
informative in the past. This finding has interesting implications for how the market reacts to
inaccurate news analytics. Because inaccurate news analytics are uninformative, subscribers will
base their future trades less on RavenPack’s sentiment score. Thus, inaccurate news analytics can
potentially reduce the market’s responsiveness to news analytics for several months following the
inaccurate news analytic.
5. High frequency traders (HFTs) and news analytics
Our results above suggest that algorithmic traders use news analytics to trade on news releases. So
far, we have identified algorithmic traders mainly through the fact that they are able to react to the
article within 5 seconds, which is too fast for a human trader. In this section, we focus on a subset
of algorithmic traders, namely high frequency traders (HFTs). HFTs are traders that use special
technology such as co-location, microwave towers between exchanges, and specialized algorithms
33
to trade extremely fast. Common trading strategies associated with HFTs include market making
and cross-venue arbitrage (Boehmer, Li, and Saar (2017), Zhang (2017)). These activities require
submitting limit orders which are at risk of being picked off when corporate news hits the market.
Therefore, they are amongst the traders most likely to use news analytics to avoid adverse
selection. Whether they use news analytics to avoid adverse selection or to make directional bets
or whether they do not use them at all remains an empirical question that we address in this part
of the paper.
5.1 HFT response to news analytics – total trading
To analyze HFT trading, we use NASDAQ high frequency trading data first used in Brogaard,
Hendershott and Riordan (2014). These data contain all trades on NASDAQ for a sample of 120
stocks for the years 2008 and 2009. For each trade, the data identifies which side of the trade
demanded liquidity and whether that trader is an HFT or a non-HFT (which can include
algorithmic traders that are not HFTs). HFTs are identified by NASDAQ based on the firms’
trading styles and the description on their websites.
We merge the NASDAQ data with our RavenPack data. Unfortunately, due to the fact that
the NASDAQ data contain only 2 years and only 120 stocks, our sample is significantly reduced.
For example, the merged data include only 112 HRL articles in the period after RavenPack was
“live”. Therefore, we are not able to use the regression set-up estimated in the previous sections.
Instead, we use a simpler set up in which we compare the reaction to all articles with high relevance
(HRL and HRH) before and after RavenPack was released on April 1, 2009. By using both HRL
and HRH articles, our sample size is significantly increased. We ask how HFTs changed their
trading behavior following company specific articles after the release of RavenPack.
34
We focus on HFT Trading Fraction, which we define as shares traded by HFTs (double
counting a trade if HFTs are on both sides of the trade) divided by total shares traded on NASDAQ.
In line with our previous analysis, we focus on the HFT Trading Fraction Response, which we
define as
. Thus it measures the HFT Trading
Fraction in the 5 seconds after an article standardized by the HFT Trading Fraction in the 120
seconds after the article. Standardizing in this way controls for the overall level of HFT trading in
that stock during that time of the day and focuses on how HFT trading might be different
immediately after the article.
We regress HFT Trading Fraction Response on D(RavenPack Release), which is a dummy
variable equal to 1 after RavenPack went “live” on April 1, 2009 and 0 before that. We also include
our usual article and firm specific controls and two-way cluster standard errors at the firm and date
level. We report the results in Table 10. The coefficient of D(RavenPack Release) is about 1.8 and
statistically significant at the 1% level. This suggests that the HFT Trading Fraction Response
increases by 1.8 percentage points after RavenPack is released to the market, which is 3.6% relative
to the unconditional median. This suggests that HFTs make up a larger fraction of trading in the 5
seconds after an article once RavenPack is live. This finding is consistent with HFTs using
RavenPack to trade on company specific news releases rather than to reduce their trading to avoid
adverse selection.
Next, in regressions 3 to 6, we split the sample into press releases and other articles. Similar
to our prior results, we find a much stronger reaction in press releases, where HFT Trading
Fraction Response increases by 3.8 percentage points (compared to a 1.3 percentage point increase
for other articles).
35
5.2 HFT response to news analytics – Liquidity demanding and directional trades
If HFTs conducted directional trades based on RavenPack information, we would expect them
to conduct more liquidity-demanding trades after an article. In contrast, if their limit orders were
being picked off by other traders using RavenPack, we would expect them to have more liquidity-
supplying trades. We address this question in Panel A of Table 11. We compute two dependent
variables: Active HFT Trading Fraction Response and Passive HFT Trading Fraction Response.
Active (Passive) HFT Trading Fraction Response is constructed in the same way as HFT Trading
Fraction Response but uses Active (Passive) HFT Trading Fraction, which is defined as shares
traded using liquidity-demanding (liquidity-supplying) trades by HFTs divided by all shares traded
on NASDAQ. Using the same specification as above, we regress these two variables on
D(RavenPack Release). After the introduction of RavenPack, we observe an increase in Active
HFT Trading Fraction Response of 2.3% that is significant at the 5% level. In contrast, the increase
in Passive HFT Trading Fraction Response is only 0.9% and not statistically significant. This
finding suggests that HFTs mainly increase their liquidity-demanding trades following news
articles after the RavenPack release, which is consistent with them exploiting the information in
RavenPack to conduct directional trades rather than to be picked off by other traders using
RavenPack.
Next, we study whether HFTs indeed trade in the direction of the RavenPack sentiment.
Analogously to the variables above, we compute With News HFT Trading Fraction Response and
Against News HFT Trading Fraction Response by counting only those trades that HFTs do in the
direction of the article sentiment as identified by RavenPack (we exclude articles with neutral
sentiment). The results are presented in Panel B of Table 11. After the introduction of RavenPack,
we observe an increase in With News HFT Trading Fraction Response of 1.9% that is significant
36
at the 10% level. In contrast, the increase in Against News HFT Trading Fraction Response is only
0.6% and not statistically significant. This finding is consistent with HFTs trading more in the
direction of the RavenPack sentiment.
To conclude, we do not find any evidence that HFTs use RavenPack to avoid being picked off
following an article. Rather, at least some NASDAQ HFT traders seem to use RavenPack for
speculative bets on company-specific news releases: After RavenPack goes “live” there is more
HFT trading in the 5 seconds after an article, and this trading consists mainly of liquidity
demanding trades and trades in the direction of the article sentiment. As a caveat, given that the
results are based on a purely time-series comparison, they might be driven by other factors that
change in the news trading environment at the same time RavenPack gets released.
5.3 News analytics and market liquidity
The results above suggest that RavenPack causes HFTs to conduct more liquidity demanding
trades after news releases. In this section, we ask whether such trades lead to a worsening of
liquidity after an article is released. Such a drop in liquidity may be caused by users of RavenPack
picking off liquidity-supplying orders in the order book. In addition, liquidity providers (that are
not using RavenPack) may notice such trading and reduce liquidity provision to avoid being picked
off. Both of these effects would lead to a decrease in liquidity after the release of the article. We
study whether RavenPack contributes to such a drop in liquidity.
Because we do not have to rely on the HFT data for this question, we can employ our more
robust regression set-up using the comparison between HRH and HRL articles. In particular, we
regress changes in liquidity on our D(HRH) dummy and a set of control variables defined as in the
previous specifications. We use two standard proxies for liquidity – the effective spread is the most
widely-used and reliable measure of the bid-ask spread when using transaction-level data like the
37
TAQ data used here, and the Amihud (2002) illiquidity measure is widely-considered the most
reliable measure of price impact (see Goyenko, Holden and Trzcinka (2009)). The Amihud
illiquidity measure is defined as:
,
where r
it
is the return for stock i during second t; dolvol
it
is the dollar volume for stock i during
second t; and N
ij
is the number of seconds in which stock i traded during interval j. Effective spread
is defined as:
,
where buys
it
(sells
it
) is the number of stocks bought (sold) for stock i during second t;
is
the last execution price for stock i during second t;
is the last bid-ask midpoint for
stock i during second t and N
ij
is the number of seconds in which stock i traded during interval j.
Because these liquidity measures are positively autocorrelated, we standardize them with
respect to their average, computed in the t-5 to t-2 minutes before the article is released.
Specifically, we compute:
; and
.
The regression set-up is the same as in Tables 4 and 6 and we report the results in Table 12.
During the time period where RavenPack was live (Panel A), we observe an increase in both
Amihud illiquidity and effective spread if an article is correctly released as relevant (HRH), while
there is no significant effect in the placebo sample (Panel B). These results show that illiquidity
38
increases (liquidity decreases) more after a news release delivered via RavenPack. This confirms
that the directional trading triggered by RavenPack reduces liquidity after an article is released.
6. Robustness checks
6.1 Difference in difference specification
Until now we have mainly focused on the significant effect of RavenPack on the stock market
during the period when it was live and showed in placebo tests that there is no effect for the pre-
RavenPack period. However, it is possible that the placebo tests might not find significant results
because of weak power. Therefore, we estimate a difference-in-difference specification for our
entire sample period (February 1, 2004 ‒ September 10, 2012) to study whether the difference
between the pre- and post-RavenPack periods is statistically significant.
We report the results in Table 13. In Regressions 1-2 and 3-4, the dependent variables are
Speed of Stock Price Response and Speed of Trade Volume Response, respectively. In Regressions
1 and 4, we estimate the following difference-in-difference specification:
where
and
are firm and date fixed effects; RavenPack Release is a dummy variable equal to
1 after the release of RavenPack on April 1, 2009, and zero otherwise; and D(HRH) is a dummy
equal to 1 for HRH articles and 0 for HRL articles. Endogenous control variables should not be
included in difference-in-difference specifications (see for example Gormley and Matsa (2014)).
To account for the fact that our control variables may be endogenous, we report specifications
without any control variables in regressions 1, 3 and 5. The usual specification with firm controls
is reported in Internet Appendix 3. The difference between the two specifications is minimal.
39
For specifications 1-4, the explanatory variable of interest is the interaction between D(HRH)
and RavenPack Release. Its coefficient is significantly positive, implying the effect of an HRH
article on the speed of reaction increased significantly after the introduction of RavenPack. This is
in line with our previous findings.
In regression 5 and 6, the dependent variable is the return from 1 second before to 5 seconds
after the article. Because our baseline analysis used an interaction, the difference-in-difference
specification uses a triple interaction:
The explanatory variable of interest is this triple interaction between D(HRH), RavenPack Release
and Sentiment Direction. Its coefficient is statistically significantly positive, suggesting an increase
in the effect of Sentiment Direction on returns for articles classified as HRH after RavenPack went
live and confirming our previous finding.
6.2 RavenPack overfitting its algorithm
One concern is that our results may be related to RavenPack overfitting the algorithm used to
identify the relevance score by training it on market reactions to articles. We think this issue is
unlikely to be relevant for two reasons. First, RavenPack’s relevance score, which we use for
identification, is solely based on where and how often the company name appears in an article. It
is not based or trained on market reactions to specific articles. Second, if RavenPack were to train
its relevance scoring algorithm on past returns, we would expect a change in the 2-minute market
reaction to LRH and HRL articles after RavenPack went live. The intuition is that Old RavenPack
could not have been trained on articles released after RavenPack went “live” while New
40
RavenPack could have been. Therefore, HRL articles that New RavenPack classifies as more
relevant, should have a larger 2-minute stock market reaction (leading New RavenPack to upgrade
them to highly relevant), while LRH articles should have a smaller 2-minute stock market reaction
(leading New RavenPack to downgrade them to low relevance). However, as we show in the
Internet Appendix in IA-Table 1, neither HRL nor LRH articles experience a change in 2-minute
stock price or trading volume response after RavenPack went live. This finding suggests that
RavenPack did not overfit its relevance-score algorithm to market responses.
6.3 Alternative placebo tests
Our base sample for the placebo test is Feb 2004 − Apr 2009. This time period includes the
financial crisis and the introduction of Regulation National Market System (Reg NMS) which
brought several changes to market structure that increased high frequency trading (Hasbrouck and
Saar (2013)) and fragmentation of U.S. markets (O’Hara and Ye (2011)). To address the possible
effect of these events on our analysis, we conduct two additional placebo tests: one that covers the
period Feb 1, 2004 to Dec 31, 2007, thereby excluding the financial crisis; and one that covers the
period July 9, 2007 to April 1, 2009, thereby excluding the time before the introduction of Reg
NMS. We report the results in Panel A and B of IA-Table 5 in the Internet Appendix. For both
alternative placebo tests and for all specifications, we find no significant effects and the
coefficients of interest are small. This suggests that the absence of significant results in our placebo
sample is not driven by the inclusion of the financial crisis or the pre-Regulation NMS time period.
6.4 “Old RavenPack” definition: RavenPack 1.0 versus RavenPack 2.0
In our main specification, Old RavenPack includes both RavenPack 1.0 and RavenPack 2.0. A
concern is that the difference in reaction before and after the release of New RavenPack might be
related to the transition from v.1.0 to v.2.0 in July 2011. Therefore, our next robustness check
41
focuses only on RavenPack 1.0. We re-estimate the same specifications as before, but include only
the period when RavenPack 1.0 was live, i.e. April 1, 2009 to July 6, 2011. We report results in
Panel C of IA-Table 5 in the Internet Appendix. All specifications confirm previous results and
are similar in terms of economic magnitude.
6.5 Adding fixed effects for number of firms mentioned in the article
As we have shown in Section 2.3, HRL articles mention somewhat more firms than HRH articles.
Thus it is a concern that this difference in the number of firms drives the difference in market
reaction to these two article types. In our main tests, we address this issue by including a control
variable for the number of firms and running placebo tests. In this robustness check, we control
even more carefully for this issue using fixed effects. In particular, we exclude articles that mention
more than 3 firms and use fixed effects for whether the article mentions 1, 2, or 3 firms. The results
are presented in Panel D of IA-Table 5 (Internet Appendix). All specifications confirm previous
results and are similar in terms of economic and statistic magnitude.
6.6 Controlling for the information environment
Another concern is that HRL and HRH articles have different transmission rates among non-
algorithmic traders. This is unlikely given that we do not find any difference between HRL and
HRH in our placebo tests before RavenPack went live, but, nonetheless, we conduct an additional
robustness check to address this issue. Following Manela (2014), we use media coverage as a
measure for the rate of transmission among non-algorithmic traders. Accordingly, we control for
different transmission rates by including the following controls: three dummy variables equal to
one if there was an article for the same firm in (1) the last hour, (2) earlier in the trading day or (3)
in the past trading day. In addition, we add log(1+number of articles in last week) to control for
the number of articles in the prior week. The results are presented in Panel E of IA-Table 5 in the
42
internet appendix. The additional controls do not meaningfully change our results, suggesting that
they are not driven by differences in the transmission speeds of non-algorithmic traders.
6.7 Alternative length of event window
In our analyses we compare the stock price reaction in the short run, during which only algorithmic
traders can react to an article, to the stock price reaction in the long run during which human traders
will have read and traded on the article. In all of our prior analyses, we used 5 seconds as the short-
run window and 120 seconds as the long-run window. Here we consider other window lengths.
In IA-Table 6 in the Internet Appendix, we show robustness to choosing longer windows lengths.
In particular, we use 10 seconds as the alternative short window and 300 seconds as the alternative
long window. The results vary slightly in magnitude, but remain statistically significant.
7. Conclusion
News analytics have revolutionized the way in which financial markets process news. In this paper,
we study how news analytics affect market efficiency, for which articles news analytics are most
relevant, and which traders use news analytics.
To establish a causal effect of news analytics on the market, we exploit an identification
strategy based on inaccuracies in news analytics that were released to the market by RavenPack, a
major provider of news analytics for algorithmic traders. Comparing the market reaction to similar
news items depending on whether the news has been correctly released to customers or not, we
are able to determine the causal effect of news analytics on stock prices, irrespective of the
informational content of the news.
We show that the market temporarily responds to errors in news analytics but the reaction is
small and reverts after 30 seconds, suggesting that the market is quite resilient. More generally,
43
news analytics increase market efficiency: The speed of adjustment of both stock prices and trade
volume in response to a highly-relevant article is faster if the article was originally released by
RavenPack as being relevant than if it was incorrectly released as not relevant. Interestingly, this
finding is almost completely concentrated in press releases, likely because they are timelier and
news analytics tend to be more accurate for press releases.
This finding suggests that users of news analytics understand for which articles news analytics
are most accurate. Furthermore, we show that they understand when news analytics are most
accurate: RavenPack has a stronger market impact at times after it had been more informative in
the past. This finding suggests that algorithmic traders learn about the informativeness of news
analytics dynamically. Thus, inaccuracies in news analytics can reduce the market’s sensitivity to
news analytics for a prolonged period of time.
Finally, we show that high frequency traders trade more following company-specific articles
after the release of RavenPack. In particular, they engage in more liquidity demanding trades in
the direction of the article sentiment. This finding suggests that HFTs, which are usually seen as
trading exclusively on price signals, might also trade on corporate-specific news using news
analytics. We show that such trading leads to a decrease in liquidity following an article.
Our findings have normative implications in terms of the recent regulatory debate on high-
speed information and the effects of algorithmic and high-frequency trading. We show that news
analytics improve market efficiency by speeding up the market reaction to news, in particular for
press releases. In contrast, temporary price distortions due to erroneous news analytics are
corrected quickly. News analytics provide an additional source of information advantage to high
frequency traders, who use it for directional speculative bets. However, these directional bets
absorb liquidity and contribute to a decrease in liquidity following the articles.
44
References
Amihud, Y., 2002. Illiquidity and Stock Returns: Cross-Section and Time-Series Effects. Journal
of Financial Markets 5, 31–56.
Baron, Matthew, Jonathan Brogaard, Björn Hagströmer, and Andrei Kirilenko, 2017, The Risk
and Return in High Frequency Trading, Journal of Financial and Quantitative Analysis,
forthcoming.
Benos, Evangelos and Satchit Sagade, 2016, Price discovery and the cross-section of high-
frequency trading, Journal of Financial Markets 30, 54-77.
Biais, Bruno, Thierry Foucault and Sophie Moinas, 2015, Equilibrium Fast Trading, Journal of
Financial Economics.
Boehmer, Ekkehart, Kingsley Fong, and Julie Wu, 2015, International evidence on algorithmic
trading, Working Paper.
Boehmer, Ekkehart, Dan Li, and Gideon Saar, 2018, The Competitive Landscape of High-
Frequency Trading Firms, Review of Financial Studies, forthcoming.
Boudoukh, Jacob, Ronen Feldman, Shimon Kogan and Matthew Richardson, 2016, Information,
Trading, and Volatility: Evidence from Firm-Specific News, Working Paper.
Brogaard, Jonathan, Terrence Hendershott, and Ryan Riordan, 2014, High frequency trading and
price Discovery, Review of Financial Studies 27, 2267-2306. .
Brogaard, Jonathan, Al Carrion, Thibaut Moyaert, Ryan Riordan, Andriy Shkilko, Konstantin
Sokolov, 2015, High-Frequency Trading and Extreme Price Movements, Working Paper.
Chaboud, Alain, Ben Chiquoine, Erik Hjalmarsson, and Clara Vega, 2014, Rise of the Machines:
Algorithmic Trading in the Foreign Exchange Market, Journal of Finance 69, 2045-2084.
Chan, W. S., 2003, Stock price reaction to news and no-news Drift and reversal after headlines,
Journal of Financial Economics 70, 223-260.
Chordia, Tarun, T. Clifton Green, and Badrinath Kottimukkalur, 2017, Rent Seeking by Low
Latency Traders: Evidence from Trading on Macroeconomic Announcements, Working Paper.
Clark-Joseph, Adam D., 2013, Exploratory Trading, Working Paper.
Das, Sanjiv R., and Mike Y. Chen, Yahoo! for Amazon: Sentiment Extraction from Small Talk on
the Web, Management Science 53, 1375-1388.
DellaVigna, S., and Pollet, J., 2009, Investor Inattention, Firm Reaction, and Friday Earnings
Announcements, Journal of Finance 64, 709-749.
Dougal, Casey, Joseph Engelberg, Diego Garcia and Christopher Parsons, 2012, Journalists and
the Stock Market, Review of Financial Studies 25, 639-679.
Dugast, Jerome and Thierry Foucault, 2017, Data Abundance and Asset Price Informativeness,
Working Paper.
Dzielinski, Michal, and Henrik Hasseltoft, 2017, News Tone Dispersion and Investor
Disagreement, Working Paper.
45
Engelberg, Joseph and Parsons, Christopher A., 2011, The Causal Impact of Media in Financial
Markets, Journal of Finance 66, 67-97.
Fang, Lily H. and Peress, Joel, 2009, Media coverage and the cross-section of stock returns,
Journal of Finance 64, 2023-2052.
Ferguson, Nicky J., Dennis Philip, Herbert Y.T. Lam and Jie Michael Guo, 2015, Media Content
and Stock Returns: The Predictive Power of Press, Multinational Finance Journal 19, 1-31.
Foucault, Thierry, Johan Hombert and Ioanid Rosu, 2016, News Trading and Speed, Journal of
Finance 71, 335-382.
Gai, Jiading, Chen Yao and Mao Ye, 2013, The Externalities of High Frenquency Trading,
Working Paper.
Garcia, Diego, Sentiment during Recessions, Journal of Finance 68, 1267–1300.
Gerig, Ausin, 2015, High-Frequency Trading Synchronizes Prices in Financial Markets, Working
Paper.
Golub, Anton, John Keane and Ser-Huang Poon, 2012, High Frequency Trading and Mini Flash
Crashes, Working Paper.
Gormley, Todd A., and David A. Matsa, 2011, Growing Out of Trouble? Corporate Responses to
Liability Risk, Review of Financial Studies 24, 2781-2821.
Griffin, John M., Nicholas H. Hirschey, and Patrick J. Kelly, How Important Is the Financial Media
in Global Markets?, Review of Financial Studies 24, 3941-3992.
Groß-Klußmann, Axel and Nikolaus Hautsch, 2011, When machines read the news: Using
automated text analytics to quantify high frequency news-implied market reactions, Journal of
Empirical Finance, 18, 321-340.
Hasbrouck, Joel and Gideon Saar, 2013, Low-Latency Trading, Journal of Financial Markets 16,
646–679.
Hendershott, Terrence and Ryan Riordan, 2013, Algorithmic Trading and the Market for Liquidity,
Journal of Financial and Quantitative Analysis 48, 1001-1024.
Hendershott, Terrence, Charles M. Jones and Albert J. Menkveld, 2011, Does Algorithmic Trading
Improve Liquidity?, Journal of Finance, 66, 1-33.
Heston, Steven L. and Nitish R. Sinha, 2016, News versus Sentiment: Predicting Stock Returns
from News Stories, Working Paper.
Hirschey, Nicholas, 2018, Do High-Frequency Traders Anticipate Buying and Selling Pressure?,
Working Paper.
Hu, Grace X., Jun Pan and Jiang Wang, 2017, Early Peek Advantage? Efficient price discovery
with tiered information disclosure, Journal of Financial Economics 126, 399-421.
Jegadeesh, Narasimhan, and Di Wu, 2013, Word power: A new approach for content analysis,
Journal of Financial Economics 110, 712-729.
Jones, Charles, 2013, What do we know about high-frequency trading?, Working Paper.
46
Jovanovic, Boyan and Albert J. Menkveld, 2011, Middlemen in Limit-Order Markets, Working
Paper.
Kyle, Albert, 1985, Continuous auctions and insider trading, Econometrica 53, 1315-1336.
Lee, Charles M. and Mark J. Ready, Inferring Trade Direction from Intraday Data, Journal of
Finance 46, 733-746.
Loughran, Tim and Bill McDonald, 2011. When is a liability not a liability? Textual analysis,
dictionaries, and 10-Ks, Journal of Finance 66, 35-65.
Manela, Asaf, 2014, The value of diffusing information, Journal of Financial Economics 111,
181-199.
Martinez, Victor H., and Ioanid Rosu 2013, High Frequency Traders, News and Volatility,
Working Paper.
Menkveld, Albert, 2013, High frequency trading and the new market makers, Journal of Financial
Markets, 16, 712-740.
O’Hara, Maureen and Mao Ye, 2011, Is market fragmentation harming market quality?, Journal
of Financial Economics, 100, 459–474.
Peress, Joel, 2014, The Media and the Diffusion of Information in Financial Markets: Evidence
from Newspaper Strikes, Journal of Finance, 69, 2007-2043.
Riordan, Ryan, Andreas Storkenmaier, Martin Wagener and S. Sarah Zhang, 2013, Public
information arrival: Price discovery and liquidity in electronic limit order markets, Journal of
Banking and Finance, 37, 1148-1159.
Riordan, Ryan and Andreas Storkenmaier, 2012, Latency, liquidity and price discovery, Journal
of Financial Markets, 15, 416-437.
Sinha, Nitish Rajan, 2016, Underreaction to News in the US Stock Market, Quarterly Journal of
Finance 6.
Tetlock, Paul, 2007, Giving content to investor sentiment The role of media in the stock market,
Journal of Finance, 62, 1139–1168.
Tetlock, Paul, 2011, All the News That's Fit to Reprint: Do Investors React to Stale Information?,
Review of Financial Studies 24, 1481-1512.
Weller, Brian, 2015, Efficient Prices at Any Cost: Does Algorithmic Trading Deter Information
Acquisition?, Working Paper.
Zhang, Sarah S., 2017, Need for Speed: Hard information processing in a high-frequency world,
The Journal of Futures Markets 38, 3-21.
47
Figure 1: Market Reaction by Relevance Score
This figure displays cumulative signed return (relative to the time of the article) from 60 seconds before to 120 seconds after the article for news
events from April 1, 2009 to September 10, 2012 (the time when RavenPack was live). Signed returns are returns are multiplied with the sentiment
direction of the article. We exclude articles with neutral sentiment. Low Relevance refers to articles with a Relevance Score below 90 in both
RavenPack versions, while High Relevance refers to articles that have a Relevance Score greater or equal than 90 in both RavenPack versions.
Cumulative signed return
-1.5
-1
-0.5
0
0.5
1
1.5
2
2.5
-60 -30 0 30 60 90 120
Cumulative Signed Return (Basis Points)
Low Relevance
High Relevance
48
Figure 2: Difference in Stock Price Response between HRH and LRH Articles
This figure displays cumulative signed return (relative to the time of the article) from 60 seconds before to 120 seconds after the article for news
events from April 1, 2009 to September 10, 2012 (the time when RavenPack was live). Returns are multiplied with the sentiment direction of the
article. We exclude articles with neutral sentiment. HRH refers to articles that have a relevance scores greater or equal 90 in both RavenPack
versions, while LRH refers to articles that had a relevance score greater or equal 90 in the old RavenPack version while having Relevance below
90 in the new RavenPack version.
-1.5
-1
-0.5
0
0.5
1
1.5
2
2.5
-60 -30 0 30 60 90 120
Cumulative Signed Return (Basis Points)
LRH
HRH
49
Figure 3: Difference in Speed of Stock Price Response Over Time
The figure in Panel A reports the point estimates from an OLS regression of Speed of Stock Price Response (
)
expressed in percent on D(HRH) interacted with yearly dummy variables from 2004 to 10 September 2012. We assign the first quarter of a year to
the prior year, i.e. the 2009 dummy covers a time period from 1 April 2009 to 1 April 2010. Controls and fixed effects are the same as in table 3
regression 3. The vertical line indicates the introduction of RavenPack on 1 April 2009. In Panel B, we report the same regression but interacting
the HRH dummy variable with two-year dummy variables (with the first quarter shifted backwards). In Panel C, we report the difference between
Speed of Stock Price Response for HRH and HRL articles over different years (with the first quarter shifted backwards).
Panel A: Estimate of coefficient on D(HRH) interacted with yearly dummies
Panel B: Estimate of coefficient on D(HRH) interacted with two-year dummies
Panel C: Comparing the difference in mean
-3
-2
-1
0
1
2
3
4
2004 2005 2006 2007 2008 2009 2010 2011 2012
Speed of Stock Price
Response
Coefficients
95% CI
-2
-1
0
1
2
3
4
04 05-06 07-08 09-10 11-12
Speed of Stock Price
Response
Coefficients
95% CI
-1
0
1
2
3
2003 2004 2005 2006 2007 2008 2009 2010 2011
Share of Stock Price
Reaction
Difference
50
Table 1: Overview of Four Article Types
In Panel A, we present our predictions for the market reaction to different articles. In Panel B, we present the results of article-level regressions that
examine the market reaction to different types of articles in the time period where RavenPack was not yet sold to investors. The dependent variables
are the absolute returns and turnover in the two minutes after the article. Returns are based on mid-quotes and expressed in basis points. The
explanatory variable of interest are D(HRH), D(HRL) and D(LRH), which are dummy variables for these article categories (LRL is the omitted
category). At the bottom of the table we display the t-statistic for the difference between HRH and HRL articles and for the difference between
RLH and LRL articles. All standard errors are two-way clustered by stock and date. T-statistics are below the parameter estimates in parenthesis;
***, **, * indicate significance at the 1%, 5%, and 10% level, respectively.
Panel A: Predictions for the market reaction of the different article types
New RavenPack
High Relevance Article
Low Relevance Article
Old RavenPack
High Relevance Article
HRH: Fast and persistent market reaction
LRH: Fast market reaction that
mean-reverts (overreaction).
Low Relevance Article
HRL: Slow market reaction (underreaction)
LRL: No stock price reaction
Panel B: Stock price reaction to the different article types BEFORE RavenPack went “live”
Dependent Variable:
Absolute Return t−1,t+120 (in bp)
Turnover t−1,t+120 (in bp)
D (HRH)
3.235
***
0.335
***
(33.02)
(30.30)
D (HRL)
3.118
***
0.325
***
(11.66)
(8.90)
D (LRH)
2.494
***
0.311
***
(9.07)
(9.27)
Number of Observations
2214677
2214677
R
2
0.151
0.184
Hour Fixed Effects
Yes
Yes
Date and Firm Fixed Effects
Yes
Yes
Difference between HRH and HRL
0.117
0.011
(t-stat)
(0.43)
(0.29)
Difference between LRH and LRL
2.494
***
0.311
***
(t-stat)
(9.07)
(9.27)
51
Table 2: Summary Statistics – Relevant articles, Apr 2009 to Sept 2012
This table displays summary statistics for the 321,912 article-company combinations after RavenPack went “live” (April 1, 2009 to September 10,
2012). These article-company observations are classified as relevant in the new RavenPack (i.e. they are HRH or HRL). Market capitalization is
the number of shares outstanding multiplied by the prior day closing price. Average volatility prior month is the mean of daily squared return in the
20 trading days before the article. Average turnover prior month is the mean of daily trading volume divided by shares outstanding in the 20 trading
days before the article. Absolute return t−1, t+5 is the absolute stock return from 1 second before to 5 seconds after the article. Speed of Stock Price
Response is defined as
. Turnover t−1, t+5 is trading volume divided by shares outstanding from 1 second
before to 5 seconds after the article. Speed of Trade Volume Response is defined as
. Return on trading day is the stock return over
the entire trading day that the article was released. Absolute return on trading day is its absolute value. Time since last company article is the time
since the company was last mentioned in an article. Number of firms in article defines the number of companies mentioned in the article. Composite
Sentiment Score is a sentiment score that is provided by RavenPack and takes a value from 100 (positive) to 0 (negative). Absolute Composite
Sentiment Score is defined as Abs (Composite Sentiment Score – 50). Neutral Composite Sentiment Score is a dummy variable equal to 1 if the
Composite Sentiment Score equals 50. Article Category Identified is a dummy variable equal to 1 if the article category (e.g. merger and acquisitions)
is identified by RavenPack. Event Sentiment Score is a sentiment score that is provided by RavenPack and takes a value from 100 (positive) to 0
(negative); this is available only for articles for which the category is identified. Absolute Event Sentiment Score is defined as Abs (Event Sentiment
Score – 50). Neutral Event Sentiment Score is a dummy variable equal to 1 if the Event Sentiment Score equals 50. In Panel A, we report descriptive
statistics. In Panel B we report the difference between articles that were consistently released as relevant in both RavenPack versions (HRH) and
those that were released as having low relevance (HRL). The difference is defined as the regression coefficient of D(HRH) in a regression of the
respective variable on D(HRH) and Relevance, Category, Hour and Date Fixed Effects. D(HRH) is a dummy equal to 1 if the article is HRH. We
also report t-statistics for the coefficient two-way clustered by stock and date. ***, **, * indicate significance at the 1%, 5%, and 10% level,
respectively.
Panel A: Descriptive Statistics
Mean
25
th
Percentile
Median
75
th
Percentile
Standard
Deviation
Market capitalization ($ million)
13185.0
157.4
1782.9
30027.4
37016.1
Average return prior month (%)
0.12
-0.57
0.10
0.79
0.65
Average volatility prior month (%)
9.69
1.19
4.79
20.4
17.7
Average turnover prior month (%)
1.17
0.27
0.83
2.29
1.23
Absolute Return t−1,t+5 (basis points)
1.95
0
0
4.43
9.46
Absolute Return t−1,t+120 (basis points)
11.4
0
5.00
27.4
21.7
Speed of Stock Price Response (%)
13.2
0
0
50
24.7
Signed Return t−1,t+5 (basis points)
0.60
-1.38
0
1.97
10.2
Signed Return t−1,t+120 (basis points)
1.89
-15.1
0
18.5
25.6
Turnover t−1,t+5 (basis points)
0.041
0
0
0.084
0.14
Turnover t−1,t+120 (basis points)
0.86
0
0.24
1.81
2.22
Speed of Trade Volume Response (%)
5.80
0
0
16.8
13.9
Return on trading day (%)
0.23
-3.29
0.056
3.92
4.02
Absolute return on trading day (%)
2.48
0.22
1.45
5.65
3.18
Time since last company article (hours)
32.2
0.49
6.42
103.1
57.6
Number of companies in article
2.14
1
1
3
4.30
Composite Sentiment Score
50.0
47
50
52
4.19
Absolute Composite Sentiment Score
2.07
0
2
5
3.65
Neutral Composite Sentiment Score
0.47
0
0
1
0.50
Article category identified
0.35
0
0
1
0.48
Event Sentiment Score
51.8
37
50
67
12.9
Absolute Event Sentiment Score
3.83
0
0
13
6.71
Neutral Event Sentiment Score
0.69
0
1
1
0.46
Past Informativeness 6 month 12FF
1.67
0.58
1.39
3.46
1.15
Past Informativeness 3 month 12FF
1.58
0.41
1.22
3.29
1.23
Past Informativeness 6 month 30FF
1.69
0.39
1.34
3.66
1.46
Number of Observations
321,912
52
Panel B: Comparison between Accurately Classified as Relevant (HRH) vs. Misclassified (HRL)
Standard Deviation
Difference between
HRH and HRL after
fixed effects
T- Statistic
Difference in terms
of Standard
Deviations
Market capitalization ($ million)
37016.1
-921.79
-0.25
-0.0249
Average return prior month (%)
0.65
-0.0319
***
-3.01
-0.04908
Average volatility prior month (%)
17.7
-1.525
*
-1.70
-0.08621
Average turnover prior month (%)
1.23
-0.0773
-0.93
-0.06285
Average illiquidity prior month (percentile)
26.4
-2.0886
-0.84
-0.07907
Absolute Return t−1,t+120 (basis points)
21.7
-0.3433
-0.77
-0.01582
Turnover t−1,t+120 (basis points)
2.22
0.0731
1.17
0.032928
Return on trading day (%)
4.02
-0.0953
-1.54
-0.02371
Absolute return on trading day (%)
3.18
-0.1242
-0.92
-0.03906
Time since last company article (hours)
57.6
3.32
1.24
0.057639
Number of companies in article
4.30
-0.95
***
-3.13
-0.22093
Composite Sentiment Score
4.19
-0.0679
-0.63
-0.01621
Absolute Composite Sentiment Score
3.65
0.0105
0.06
0.002877
Neutral Composite Sentiment Score
0.50
-0.0353
-0.67
-0.0706
Event Sentiment Score
12.9
-0.7440
-1.10
-0.05767
Absolute Event Sentiment Score
6.71
-0.0986
*
-1.75
-0.01469
Neutral Event Sentiment Score
0.46
-0.0020
-0.98
-0.00435
53
Table 3: Overreaction to News Analytics (LRH articles)
This table contains the results of article-level regressions that examine how well the sentiment direction of LRH articles predicts stock returns
before and after the release of RavenPack. In regressions 1 to 3, the dependent variable is the return from 1 second before to 5 seconds after the
article (measured in basis points). In regressions 4 to 6, we study the return from 6 to 120 seconds after the article to determine a potential reversal
of the short run reaction. Returns are based on mid-quotes. The explanatory variable of interest is an interaction between RavenPack Release and
Sentiment Direction. RavenPack Release is a dummy variable equal to 1 during the time in which RavenPack was “live” (April 1, 2009 – September
10, 2012) and equal to 0 before RavenPack was “live” (February 1, 2004 – March 31, 2009). Sentiment Direction is a variable indicating the
sentiment of the article derived from RavenPack sentiment indices; it takes the value +1 for positive sentiment, 0 for neutral sentiment and −1 for
negative sentiment. In all regressions we include the following firm specific control variables: Company size, Return prior month, Volatility prior
month, Turnover prior month, Illiquidity prior month. In regressions 2, 3, 5 and 6 we add fixed effects for the article category (e.g. mergers and
acquisitions), the relevance score (from 90 to 100) and the hour during the day in which the article was released. In regressions 3 and 6, we include
absolute return, turnover, and volatility each for industry and market from t−1 to t+5 seconds around the article. All variables are defined in
Appendix 1. All standard errors are two-way clustered by stock and date. T-statistics are below the parameter estimates in parenthesis; ***, **, *
indicate significance at the 1%, 5%, and 10% level, respectively.
Dependent Variable:
Return t−1, t+5 (in bp)
Return t+6, t+120 (in bp)
(1)
(2)
(3)
(4)
(5)
(6)
RavenPack Release * Sentiment Direction
0.465
**
0.533
**
0.563
***
-0.602
-0.666
-0.660
(2.29)
(2.46)
(2.60)
(-1.00)
(-1.09)
(-1.10)
RavenPack Release
-0.252
-0.258
-0.309
-1.739
***
-1.659
***
-1.643
***
(-1.48)
(-1.43)
(-1.60)
(-3.88)
(-3.67)
(-3.36)
Sentiment Direction
0.258
**
0.098
0.099
1.532
***
1.517
***
1.538
***
(2.31)
(0.82)
(0.83)
(4.96)
(4.41)
(4.49)
Article category identified
-1.177
***
-1.217
***
0.043
-0.316
(-3.14)
(-3.05)
(0.04)
(-0.26)
Time since last article
0.112
**
0.107
**
0.373
***
0.357
***
(2.41)
(2.31)
(2.97)
(2.83)
Number of firms in article
-0.179
**
-0.171
**
-0.397
*
-0.415
*
(-2.15)
(-2.08)
(-1.68)
(-1.78)
Number of Observations
20588
20586
20586
20588
20586
20586
R
2
0.003
0.008
0.014
0.007
0.013
0.018
Relevance, Category and Hour Fixed Effects
No
Yes
Yes
No
Yes
Yes
Market control variables
No
No
Yes
No
No
Yes
Firm specific control variables
Yes
Yes
Yes
Yes
Yes
Yes
54
Table 4: Speed of Stock Price Response to News Articles
This table contains the results of article-level regressions that examine the effect of an article covered in RavenPack on stock price, measured by
absolute returns. The dependent variable is Speed of Stock Price Response (in percent) defined as
and
measured in seconds around an article. Returns are based on mid-quotes. The explanatory variable of interest is D(HRH), a dummy variable equal
to 1 if an article was consistently released as highly relevant in both RavenPack versions and 0 if it was originally released as having low relevance
(HRL). In regressions 1 to 3, we estimate the various specification during the time in which RavenPack was “live” (April 1, 2009 – September 10,
2012). In regressions 4 to 6, we run a placebo test for the time period where RavenPack was not yet sold to investors (February 1, 2004 – March
31, 2009). In all regressions we include firm and date fixed effects and the following firm specific control variables: Company size, Return prior
month, Volatility prior month, Turnover prior month, Illiquidity prior month. In regressions 2, 3, 5 and 6 we add fixed effects for the article category
(e.g. mergers and acquisitions), the relevance score (from 90 to 100) and the hour during the day in which the article was released. In regressions 3
and 6, we include additional controls: the absolute return, turnover, and volatility each for industry and market and for the two horizons from t−1
to t+5 and t−1 to t+120 seconds around the article. All variables are defined in Appendix 1. All standard errors are two-way clustered by stock and
date. T-statistics are below the parameter estimates in parenthesis; ***, **, * indicate significance at the 1%, 5%, and 10% level, respectively.
Dependent Variable:
Speed of Stock Price Response (in %)
Main Test - RavenPack is “live”
Placebo Test - Before RavenPack is “live”
(1)
(2)
(3)
(4)
(5)
(6)
D(HRH)
1.469
***
1.333
***
1.321
***
-0.048
-0.058
-0.012
(3.39)
(3.06)
(3.17)
(-0.17)
(-0.20)
(-0.04)
Absolute Composite Sentiment Score
-0.004
-0.002
-0.032
**
-0.032
**
(-0.23)
(-0.12)
(-2.35)
(-2.38)
Neutral Composite Sentiment Score
-0.107
-0.115
-0.338
***
-0.357
***
(-0.91)
(-1.02)
(-3.44)
(-3.67)
Article category identified
0.522
1.394
-3.575
-4.142
(0.10)
(0.28)
(-0.57)
(-0.71)
Absolute Event Sentiment Score
0.098
***
0.089
***
0.021
*
0.022
*
(5.20)
(4.86)
(1.73)
(1.88)
Neutral Event Sentiment Score
-0.818
-0.958
*
-1.202
***
-1.150
***
(-1.35)
(-1.65)
(-2.99)
(-2.89)
Time since last article
0.099
***
0.086
**
0.069
***
0.062
**
(2.78)
(2.46)
(2.67)
(2.40)
Number of firms in article
-0.061
-0.058
-0.149
***
-0.170
***
(-0.67)
(-0.67)
(-2.59)
(-2.99)
Number of Observations
248849
248849
248849
400158
400158
400158
R
2
0.034
0.038
0.084
0.032
0.033
0.048
Relevance, Category and Hour Fixed Effects
No
Yes
Yes
No
Yes
Yes
Date and Firm Fixed Effects
Yes
Yes
Yes
Yes
Yes
Yes
Market control variables
No
No
Yes
No
No
Yes
Firm specific control variables
Yes
Yes
Yes
Yes
Yes
Yes
55
Table 5: Directional Stock Price Response to Article Sentiment
This table contains the results of article-level regressions that examine how well the sentiment direction of an article predicts the 5-second return
reaction to an article depending on whether the article is covered in RavenPack. The dependent variable is the return from 1 second before to 5
seconds after the article (measured in basis points). Returns are based on mid-quotes. The explanatory variable of interest is an interaction between
D(HRH) and Sentiment Direction. D(HRH) is a dummy variable equal to 1 if an article was consistently released as highly relevant in both
RavenPack versions and 0 if it was originally released as having low relevance (HRL). Sentiment Direction is a variable indicating the sentiment
of the article derived from RavenPack sentiment indices; it takes the value +1 for positive sentiment, 0 for neutral sentiment and −1 for negative
sentiment. In regressions 1 to 3, we estimate the various specification during the time in which RavenPack was “live” (April 1, 2009 – September
10, 2012). In regressions 4 to 6, we run a placebo test for the time period where RavenPack was not yet sold to investors (February 1, 2004 – March
31, 2009). In all regressions we include firm and date fixed effects and the following firm specific control variables: Company size, Return prior
month, Volatility prior month, Turnover prior month, Illiquidity prior month. In regressions 2, 3, 5 and 6 we add fixed effects for the article category
(e.g. mergers and acquisitions), the relevance score (from 90 to 100) and the hour during the day in which the article was released. In regressions 3
and 6, we include absolute return, turnover, and volatility each for industry and market from t−1 to t+5 seconds around the article. All variables are
defined in Appendix 1. All standard errors are two-way clustered by stock and date. T-statistics are below the parameter estimates in parenthesis;
***, **, * indicate significance at the 1%, 5%, and 10% level, respectively.
Dependent Variable:
Return t-1, t+5 (in bp)
Main Test - RavenPack is “live”
Placebo Test - Before RavenPack is “live”
(1)
(2)
(3)
(4)
(5)
(6)
D(HRH) * Sentiment Direction
0.407
***
0.452
***
0.452
***
0.081
0.116
0.114
(3.06)
(3.40)
(3.41)
(0.79)
(1.10)
(1.09)
D(HRH)
0.187
*
0.125
0.125
0.137
0.103
0.102
(1.86)
(1.19)
(1.19)
(1.29)
(0.95)
(0.94)
Sentiment Direction
0.118
-0.010
-0.009
0.421
***
0.184
*
0.187
*
(0.91)
(-0.08)
(-0.07)
(4.25)
(1.82)
(1.85)
Article category identified
1.152
**
0.933
*
0.304
0.370
(2.06)
(1.96)
(0.81)
(0.97)
Time since last article
0.043
***
0.043
***
0.237
***
0.239
***
(3.44)
(3.47)
(13.19)
(13.26)
Number of firms in article
-0.150
***
-0.145
***
-0.221
***
-0.220
***
(-4.96)
(-4.80)
(-9.24)
(-9.18)
Number of Observations
321762
321762
321762
481852
481852
481852
R
2
0.060
0.063
0.067
0.056
0.062
0.063
Relevance, Category and Hour Fixed Effects
No
Yes
Yes
No
Yes
Yes
Date and Firm Fixed Effects
Yes
Yes
Yes
Yes
Yes
Yes
Market control variables
No
No
Yes
No
No
Yes
Firm specific control variables
Yes
Yes
Yes
Yes
Yes
Yes
56
Table 6: Speed of Trade Volume Response to News Articles
This table contains the results of article-level regressions that examine the effect of an article covered in RavenPack on the market for a stock,
measured by turnover. In Panel A, the dependent variable is Speed of Trade Volume Response (in percent), defined as the turnover from 1 second
before the article to 5 second after the article divided by the turnover from 1 second before the article to 120 seconds after the article. The explanatory
variable of interest is D(HRH), a dummy variable equal to 1 if an article was consistently released as highly relevant in both RavenPack versions
and 0 if it was originally released as having low relevance (HRL). In regressions 1 to 3, we estimate the various specification during the time in
which RavenPack was “live” (April 1, 2009 – September 10, 2012). In regressions 4 to 6, we run a placebo test for the time period where RavenPack
was not yet sold to investors (February 1, 2004 – March 31, 2009). In all regressions we include firm and date fixed effects and the following firm
specific control variables: Company size, Return prior month, Volatility prior month, Turnover prior month, Illiquidity prior month. In regressions
2, 3, 5 and 6 we add fixed effects for the article category (e.g. mergers and acquisitions), the relevance score (from 90 to 100) and the hour during
the day in which the article was released. In regressions 3 and 6, we include additional controls: the absolute return, turnover, and volatility each
for industry and market and for the two horizons from t−1 to t+5 and t−1 to t+120 seconds around the article. In panel B, we split the trading volume
by who initiated the trade using the algorithm of Lee and Ready (1991). In Panel B, the dependent variable is Speed of Directional Trade Volume
Response, defined as the turnover initiated in direction of the article from 1 second before the article to 5 second after the article divided by the
turnover from 1 second before the article to 120 seconds after the article. The control variables in Panel B are the same as in Panel A, but not
reported for brevity. All variables are defined in Appendix 1. All standard errors are two-way clustered by stock and date. T-statistics are below the
parameter estimates in parenthesis; ***, **, * indicate significance at the 1%, 5%, and 10% level, respectively.
Panel A: Trade Volume
Dependent Variable:
Speed of Trade Volume Response (in %)
Main Test - RavenPack is “live”
Placebo Test - Before RavenPack is “live”
(1)
(2)
(3)
(4)
(5)
(6)
D(HRH)
0.656
***
0.516
**
0.533
**
0.033
-0.002
0.022
(3.06)
(2.36)
(2.49)
(0.24)
(-0.02)
(0.16)
Absolute Composite Sentiment Score
-0.008
-0.006
-0.018
***
-0.017
**
(-0.97)
(-0.77)
(-2.66)
(-2.54)
Neutral Composite Sentiment Score
-0.171
**
-0.164
**
-0.085
-0.089
*
(-2.57)
(-2.52)
(-1.62)
(-1.69)
Article category identified
-4.034
***
-3.973
***
-0.716
-0.691
(-3.02)
(-3.47)
(-0.47)
(-0.45)
Absolute Event Sentiment Score
0.059
***
0.056
***
0.023
***
0.024
***
(5.82)
(5.50)
(3.55)
(3.68)
Neutral Event Sentiment Score
-0.973
***
-1.018
***
-0.394
*
-0.368
*
(-2.82)
(-3.02)
(-1.86)
(-1.74)
Time since last article
0.109
***
0.101
***
0.091
***
0.087
***
(5.58)
(5.21)
(5.98)
(5.82)
Number of firms in article
-0.107
**
-0.112
***
-0.168
***
-0.173
***
(-2.55)
(-2.70)
(-6.04)
(-6.25)
Number of Observations
272019
272019
272019
418029
418029
418029
R
2
0.028
0.031
0.058
0.025
0.026
0.037
Relevance, Category and Hour Fixed Effects
No
Yes
Yes
No
Yes
Yes
Date and Firm Fixed Effects
Yes
Yes
Yes
Yes
Yes
Yes
Market control variables
No
No
Yes
No
No
Yes
Firm specific control variables
Yes
Yes
Yes
Yes
Yes
Yes
Panel B: Trade volume in direction of article
Dependent Variable:
Speed of Directional Trade Volume Response (in %)
Main Test - RavenPack is “live”
Placebo Test - Before RavenPack is “live”
(1)
(2)
(3)
(4)
(5)
(6)
D(HRH)
0.472
***
0.376
**
0.385
**
0.080
0.085
0.082
(2.67)
(2.12)
(2.20)
(0.60)
(0.64)
(0.61)
Number of Observations
168278
168278
168278
272477
272477
272477
R
2
0.046
0.051
0.063
0.039
0.040
0.045
Relevance, Category and Hour Fixed Effects
No
Yes
Yes
No
Yes
Yes
Date and Firm Fixed Effects
Yes
Yes
Yes
Yes
Yes
Yes
Market control variables
No
No
Yes
No
No
Yes
Firm specific control variables
Yes
Yes
Yes
Yes
Yes
Yes
57
Table 7: Effect of News Analytics: Press Releases vs. Non-Press Releases
This table contains the results of article-level regressions examining if our results are stronger for press releases. The dependent variables are Speed
of Stock Price Response (in percent), Speed of Trade Volume Response (in percent), and the signed return from 1 second before to 5 seconds after
the article (measured in basis points). The explanatory variable of interest is the interaction between D(HRH) and D(Press Release). D(Press
Release) is a dummy variable taking the value of 1 for press releases. In all regressions we include firm and date fixed effects. In regressions 2, 4,
and 6, we add the following firm specific control variables: Company size, Return prior month, Volatility prior month, Turnover prior month,
Illiquidity prior month. In these regressions, we also add fixed effects for the article category (e.g. mergers and acquisitions), the relevance score
(from 90 to 100) and the hour during the day in which the article was released as well as additional controls: the absolute return, turnover, and
volatility each for industry and market from t−1 to t+5 seconds around the article. In regression 2 and 4, we also include those values for t−1 to
t+120 seconds around the article. All variables are defined in Appendix 1. In Panel A, we run the various specifications during the time in which
RavenPack was “live” (April 1, 2009 – September 10, 2012). In Panel B, we run a placebo test for the time period where RavenPack was not yet
sold to investors (February 1, 2004 – March 31, 2009). In Panel B, we use the same control variables as in Panel A. All standard errors are two-
way clustered by stock and date. T-statistics are below the parameter estimates in parenthesis; ***, **, * indicate significance at the 1%, 5%, and
10% level, respectively.
Panel A: Main Test
Dependent Variable:
Speed of Stock Price
Response (in %)
Speed of Trade Volume
Response (in %)
Signed Return t-1, t+5
(in bp)
(1)
(2)
(3)
(4)
(5)
(6)
D(Press Release) * D(HRH)
1.686
**
1.813
**
1.044
***
1.195
***
1.038
***
0.992
***
(2.05)
(2.25)
(2.85)
(3.32)
(3.15)
(2.97)
D(HRH)
1.003
*
0.791
0.348
0.181
0.184
0.053
(1.96)
(1.63)
(1.50)
(0.78)
(1.07)
(0.30)
D(Press Release)
-0.885
-1.487
*
-0.760
**
-1.019
***
-0.160
-0.484
(-1.09)
(-1.87)
(-2.11)
(-2.90)
(-0.49)
(-1.41)
Sentiment Direction
0.276
***
0.231
***
(8.93)
(7.10)
Absolute Composite Sentiment Score
-0.001
-0.006
(-0.06)
(-0.72)
Neutral Composite Sentiment Score
-0.116
-0.165
**
(-1.03)
(-2.54)
Absolute Event Sentiment Score
0.090
***
0.056
***
(4.90)
(5.53)
Neutral Event Sentiment Score
-1.108
*
-1.095
***
(-1.90)
(-3.21)
Article category identified
1.167
-4.090
***
0.932
(0.23)
(-3.55)
(1.61)
Time since last article
0.079
**
0.097
***
0.041
**
(2.25)
(4.95)
(2.51)
Number of firms in article
-0.061
-0.113
***
-0.222
***
(-0.70)
(-2.73)
(-5.29)
Number of Observations
248888
248849
272058
272019
198263
198225
R
2
0.034
0.084
0.028
0.058
0.086
0.095
Relevance, Category and Hour Fixed Effects
No
Yes
No
Yes
No
Yes
Date and Firm Fixed Effects
Yes
Yes
Yes
Yes
Yes
Yes
Market control variables
No
Yes
No
Yes
No
Yes
Firm specific control variables
Yes
Yes
Yes
Yes
Yes
Yes
Panel B: Placebo Test
Dependent Variable:
Speed of Stock Price
Response (in %)
Speed of Trade Volume
Response in (%)
Signed Return t-1, t+5
(in bp)
(1)
(2)
(3)
(4)
(5)
(6)
D(Press Release) * D(HRH)
0.265
0.344
0.038
0.059
0.396
0.363
(0.48)
(0.63)
(0.13)
(0.20)
(1.07)
(0.98)
D(HRH)
-0.140
-0.095
0.034
0.008
0.177
0.107
(-0.43)
(-0.30)
(0.22)
(0.06)
(1.06)
(0.65)
D(Press Release)
0.112
0.076
0.222
0.089
1.283
***
0.938
***
(0.20)
(0.14)
(0.76)
(0.30)
(3.59)
(2.68)
Number of Observations
400304
400158
418187
418029
316945
316783
R
2
0.031
0.048
0.025
0.037
0.076
0.081
Relevance, Category and Hour Fixed Effects
No
Yes
No
Yes
No
Yes
Date and Firm Fixed Effects
Yes
Yes
Yes
Yes
Yes
Yes
Market control variables
No
Yes
No
Yes
No
Yes
Firm specific control variables
Yes
Yes
Yes
Yes
Yes
Yes
58
Table 8: Comparison Press Releases vs. Non-Press Releases
This table compares the characteristics of press releases and non-press releases. We report means for both groups and their difference as well as
significance levels based on two-way clustered standard errors by stock and date. D(HRH) is a dummy variable equal to 1 if an article was
consistently released as highly relevant in both RavenPack versions and 0 if it was originally released as having low relevance (HRL). Absolute
return t-1, t+120 and Turnover t-1,t+120 are absolute return and turnover from 1 second before to 120 seconds after the article. Absolute return on
trading day is the absolute stock return over the entire trading day on which the article was released. Sentiment Direction is a variable indicating
the sentiment of the article derived from RavenPack sentiment indices; it takes the value +1 for positive sentiment, 0 for neutral sentiment and −1
for negative sentiment. D(First company article of the day) is a dummy variable equal to one if it is the first article for that company on that day.
D(New story) is a dummy variable equal to 1 if it is a new news story (RavenPack ENS=100) and 0 if it is a reprint. This variable is missing if the
category of the article cannot be identified. D(Sentiment predicts return) is a dummy variable equal to 1 if the sign of the return from 1 second
before to 120 seconds after the article matches the sign of the article sentiment (it is set to missing for articles with neutral sentiment). We use the
full sample from February 1, 2004 to September 10, 2012. ***, **, * indicate significance at the 1%, 5%, and 10% level, respectively.
Press Releases
(Mean)
Non-Press Releases
(Mean)
Difference
D(HRH)
0.97
0.97
-0.00
Absolute Return t−1,t+120 (basis points)
14.6
12.5
2.17
***
Turnover t−1,t+120 (basis points)
0.92
1.05
-0.12
***
Absolute return on trading day (%)
2.09
3.18
-1.01
***
Sentiment Direction
0.41
0.07
0.34
***
D(First company article of the day)
0.85
0.77
0.08
***
D(New story)
0.95
0.78
0.17
***
D(Sentiment predicts return)
0.56
0.53
0.03
***
59
Table 9: Directional Stock Price Response conditional on Past Informativeness of
RavenPack
This table contains the results of article-level regressions that examine how well the past performance of RavenPack affects the stock price impact
of RavenPack. The dependent variable is the signed return from 1 second before to 5 seconds after the article (measured in basis points), i.e. returns
multiplied with the sentiment direction of the article. Returns are based on mid-quotes. The explanatory variable of interest is the interaction between
D(HRH) and Past Informativeness. Past Informativeness is the average signed return (in basis points) from t-1 to t+120 seconds around articles
over the previous 6 month. D(HRH) is a dummy variable equal to 1 if an article was consistently released as highly relevant in both RavenPack
versions and 0 if it was originally released as having low relevance (HRL). In regressions 1 to 3, we estimate the various specification during the
time in which RavenPack was “live” (April 1, 2009 – September 10, 2012). In regressions 4 to 6, we run a placebo test for the time period where
RavenPack was not yet sold to investors (February 1, 2004 – March 31, 2009). In IA-Table 4 in the Internet Appendix, we report a robustness check
using Stock-Level Past Informativeness measured over the previous three months (instead of a six), using Stock-Level Past Informativeness and
using Direction-Based Past Informativeness. In all regressions we include firm and date fixed effects and the following firm specific control
variables: Company size, Return prior month, Volatility prior month, Turnover prior month, Illiquidity prior month. In regressions 2, 3, 5 and 6,
we add fixed effects for the article category (e.g. mergers and acquisitions), the relevance score (from 90 to 100) and the hour during the day in
which the article was released. In regressions 3 and 6, we add additional controls: the absolute return, turnover, and volatility each for industry and
market and for the two horizons from t−1 to t+5 seconds around the article. Control variables are defined in Appendix 1. All standard errors are
two-way clustered by stock and date. T-statistics are below the parameter estimates in parenthesis; ***, **, * indicate significance at the 1%, 5%,
and 10% level, respectively.
Dependent Variable:
Signed Return t−1, t+5 (in bp)
Main Test - RavenPack is “live”
Placebo Test - Before RavenPack is “live”
(1)
(2)
(3)
(4)
(5)
(6)
Past Informativeness * D(HRH)
0.342
***
0.276
**
0.295
**
-0.191
-0.189
-0.181
(3.03)
(2.45)
(2.58)
(-0.63)
(-0.63)
(-0.60)
D(HRH)
-0.253
-0.173
-0.208
0.551
0.635
0.608
(-1.18)
(-0.80)
(-0.95)
(0.80)
(0.92)
(0.89)
Article category identified
1.040
*
0.732
-0.037
0.034
(1.93)
(1.31)
(-0.09)
(0.08)
Time since last article
0.056
***
0.055
***
0.251
***
0.252
***
(3.43)
(3.41)
(12.40)
(12.46)
Number of firms in article
-0.207
***
-0.197
***
-0.170
***
-0.167
***
(-4.98)
(-4.75)
(-5.24)
(-5.13)
Number of Observations
198225
198225
198225
310657
310657
310657
R
2
0.085
0.090
0.094
0.073
0.079
0.080
Relevance, Category and Hour Fixed Effects
No
Yes
Yes
No
Yes
Yes
Date and Firm Fixed Effects
Yes
Yes
Yes
Yes
Yes
Yes
Market control variables
No
No
Yes
No
No
Yes
Firm specific control variables
Yes
Yes
Yes
Yes
Yes
Yes
60
Table 10: HFT Trading Response to News Articles
This table contains the results of article-level regressions that examine the effect of an article covered in RavenPack on the fraction of HFT trading
on NASDAQ. The dependent variable is HFT Trading Fraction Response (in percent) defined as
and measured in seconds around an article. HFT Trading Fraction is defined as
The explanatory variable of interest
is RavenPack Release, a dummy variable equal to 1 during the time in which RavenPack was “live” (April 1, 2009 – December 31, 2009) and equal
to 0 during the time where RavenPack was not yet sold to investors (January 1, 2008 – March 31, 2009). In all regressions we include the following
firm specific control variables: Company size, Return prior month, Volatility prior month, Turnover prior month, Illiquidity prior month. In
regressions 2, 4, and 6 we add fixed effects for the article category (e.g. mergers and acquisitions), the relevance score (from 90 to 100) and the
hour during the day in which the article was released. All variables are defined in Appendix 1. All standard errors are two-way clustered by stock
and date. T-statistics are below the parameter estimates in parenthesis; ***, **, * indicate significance at the 1%, 5%, and 10% level, respectively.
Dependent Variable:
HFT Trading Fraction Response (in %)
All Articles
Press Releases
Non-Press Releases
(1)
(2)
(3)
(4)
(5)
(6)
RavenPack Release
1.735
***
1.784
***
3.538
**
3.880
**
1.258
*
1.343
*
(3.10)
(3.20)
(2.54)
(2.50)
(1.91)
(1.92)
Absolute Composite Sentiment Score
-0.031
-0.248
0.007
(-0.39)
(-1.40)
(0.09)
Neutral Composite Sentiment Score
-0.070
-2.758
*
0.437
(-0.12)
(-1.88)
(0.65)
Article category identified
-8.47e+05
-8.84e+04
0.000
(-0.00)
(-0.00)
(0.00)
Absolute Event Sentiment Score
0.148
**
0.397
**
0.099
(2.06)
(2.40)
(1.12)
Neutral Event Sentiment Score
1.064
1.755
-0.431
(0.35)
(0.27)
(-0.10)
Time since last article
-0.126
0.687
-0.422
(-0.42)
(0.97)
(-1.26)
Number of firms in article
-0.681
*
1.323
-0.706
*
(-1.78)
(1.13)
(-1.82)
Number of Observations
5163
5162
1014
1012
4149
4147
R
2
0.010
0.019
0.016
0.044
0.009
0.022
Relevance, Category and Hour Fixed Effects
No
Yes
No
Yes
No
Yes
Firm specific control variables
Yes
Yes
Yes
Yes
Yes
Yes
61
Table 11: HFT Trading Response split by active/passive and direction
This table contains the results of article-level regressions that examine the effect of an article covered in RavenPack on HFT trading on NASDAQ.
In Panel A, the dependent variable Active HFT Trading Fraction Response (in percent) is defined as
measured in seconds around an article. Active HFT Trading Fraction is defined as
. The dependent variable Passive HFT Trading Fraction Response is defined
analogously using the stocks traded in liquidity supplying trades. The explanatory variable of interest is RavenPack Release, a dummy variable
equal to 1 during the time in which RavenPack was “live” (April 1, 2009 – December 31, 2009) and equal to 0 during the time where RavenPack
was not yet sold to investors (January 1, 2008 – March 31, 2009). In all regressions we include the following firm specific control variables:
Company size, Return prior month, Volatility prior month, Turnover prior month, Illiquidity prior month. In regressions 2 and 4 we add fixed
effects for the article category (e.g. mergers and acquisitions), the relevance score (from 90 to 100) and the hour during the day in which the article
was released. In Panel B, we use the same set-up but the dependent variables are With News HFT Trading Fraction Response and Against News
HFT Trading Fraction Response, which are defined as above but using the fraction of trading that HFTs do in direction of the news sentiment or
against the news sentiment (articles with neutral sentiment are excluded). All variables are defined in Appendix 1. All standard errors are two-way
clustered by stock and date. T-statistics are below the parameter estimates in parenthesis; ***, **, * indicate significance at the 1%, 5%, and 10%
level, respectively.
Panel A: Split by liquidity demanding (active) and liquidity supplying (passive) trades
Dependent Variable:
Active HFT Trading Fraction Response
(in %)
Passive HFT Trading Fraction Response
(in %)
(1)
(2)
(3)
(4)
RavenPack Release
2.394
**
2.345
**
1.054
0.935
(2.54)
(2.60)
(1.21)
(1.06)
Absolute Composite Sentiment Score
0.136
-0.070
(1.53)
(-0.67)
Neutral Composite Sentiment Score
1.090
0.245
(1.34)
(0.28)
Article category identified
-4.15e+06
25834.284
(-0.00)
(0.00)
Absolute Event Sentiment Score
0.297
***
0.094
(2.78)
(0.94)
Neutral Event Sentiment Score
4.500
0.694
(0.96)
(0.18)
Time since last article
0.085
-0.597
*
(0.22)
(-1.93)
Number of firms in article
-0.361
-0.721
(-0.66)
(-1.35)
Number of Observations
5116
5115
5088
5087
R
2
0.009
0.023
0.015
0.025
Relevance, Category and Hour Fixed Effects
No
Yes
No
Yes
Firm specific control variables
Yes
Yes
Yes
Yes
Panel B: Split by whether HFTs trade with the news sentiment or against the news sentiment
Dependent Variable:
With News HFT Trading Fraction Response
(in %)
Against News HFT Trading Fraction
Response (in %)
(1)
(2)
(3)
(4)
RavenPack Release
2.084
*
1.881
*
0.496
0.601
(1.96)
(1.84)
(0.44)
(0.56)
Number of Observations
5116
5115
5088
5087
R
2
0.009
0.023
0.015
0.025
Relevance, Category and Hour Fixed Effects
No
Yes
No
Yes
Firm specific control variables
Yes
Yes
Yes
Yes
62
Table 12: Liquidity Response to News Articles
This table contains the results of article-level regressions that examine the effect of an article being covered in RavenPack on illiquidity. In
regressions 1 to 3, the dependent variable is Change in Amihud Illiquidity (in percent) defined as
in
seconds around the article. In regressions 4 to 6, the dependent variable is Change in Effective Spread (in percent) defined as
in seconds around the article. The explanatory variable of interest is D(HRH), a dummy variable
equal to 1 if an article was consistently released as highly relevant in both RavenPack versions and 0 if it was originally released as having low
relevance (HRL). In Panel A, we estimate our main specification during the time in which RavenPack was live (April 1, 2009 – September 10,
2012). In Panel B, we run a placebo test in the time period where RavenPack was not yet being sold to investors (February 1, 2004 – March 31,
2009). In all regressions we include firm and date fixed effects and the following firm specific control variables: Company size, Return prior month,
Volatility prior month, Turnover prior month, Illiquidity prior month. In regressions 2, 3, 5 and 6, we add fixed effects for the article category (e.g.
mergers and acquisitions), the relevance score (from 90 to 100) and the hour during the day in which the article was released. In regressions 3 and
6, we include additional controls: the absolute return, turnover, and volatility each for industry and market from t−1 to t+5 seconds around the
article. All variables are defined in Appendix 1. All standard errors are two-way clustered by stock and date. T-statistics are below the parameter
estimates in parenthesis; ***, **, * indicate significance at the 1%, 5%, and 10% level, respectively.
Panel A: Main Specification – RavenPack is “live”
Dependent Variable:
Change in Effective Spread (in %)
Change in Amihud Illiqudity (in %)
(1)
(2)
(3)
(4)
(5)
(6)
D(HRH)
0.883
*
1.268
***
1.281
***
1.633
**
1.408
**
1.357
**
(1.96)
(2.84)
(2.85)
(2.35)
(2.03)
(1.99)
Absolute Composite Sentiment Score
0.042
**
0.046
**
-0.025
-0.011
(2.16)
(2.35)
(-0.88)
(-0.38)
Neutral Composite Sentiment Score
-0.398
***
-0.403
***
0.052
0.080
(-2.76)
(-2.80)
(0.22)
(0.36)
Article category identified
-12.021
-12.719
-4.158
-3.938
(-1.47)
(-1.50)
(-0.50)
(-0.45)
Absolute Event Sentiment Score
0.073
***
0.072
***
0.067
**
0.065
**
(3.36)
(3.32)
(2.31)
(2.27)
Neutral Event Sentiment Score
-1.658
**
-1.768
**
-2.197
**
-2.341
**
(-2.21)
(-2.36)
(-2.05)
(-2.20)
Time since last article
-0.436
***
-0.459
***
0.350
***
0.253
***
(-10.70)
(-11.35)
(5.15)
(3.75)
Number of firms in article
1.427
***
1.433
***
-0.637
***
-0.542
***
(14.32)
(14.74)
(-4.65)
(-4.02)
Number of Observations
252077
252077
252077
115630
115630
115630
R
2
0.162
0.165
0.170
0.091
0.092
0.122
Relevance, Category and Hour Fixed Effects
No
Yes
Yes
No
Yes
Yes
Date and Firm Fixed Effects
Yes
Yes
Yes
Yes
Yes
Yes
Market control variables
No
No
Yes
No
No
Yes
Firm specific control variables
Yes
Yes
Yes
Yes
Yes
Yes
Panel B: Placebo Test - Before RavenPack is “live”
Dependent Variable:
Change in Effective Spread (in %)
Change in Amihud Illiqudity (in %)
(1)
(2)
(3)
(4)
(5)
(6)
-0.032
0.480
0.445
0.861
0.687
0.631
-0.032
(-0.11)
(1.60)
(1.49)
(1.43)
(1.14)
(1.05)
(-0.11)
Number of Observations
411734
411734
411734
178102
178102
178102
R
2
0.156
0.160
0.167
0.067
0.069
0.077
Relevance, Category and Hour Fixed Effects
No
Yes
Yes
No
Yes
Yes
Date and Firm Fixed Effects
Yes
Yes
Yes
Yes
Yes
Yes
Market control variables
No
No
Yes
No
No
Yes
Firm specific control variables
Yes
Yes
Yes
Yes
Yes
Yes
63
Table 13: Difference in Difference Analysis
This table contains the results of article-level regressions implementing a difference in difference set-up for our whole sample from February 1,
2004 to September 10, 2012 as a robustness check to tables 4 to 6. In regressions 1 and 2, the dependent variable is Speed of Stock Price Response
(in percent), defined as
and measured in seconds around an article. In regressions 3 and 4, the dependent
variable is Speed of Trade Volume Response (in percent), defined as the turnover from 1 second before the article to 5 second after the article
divided by the turnover from 1 second before the article to 120 seconds after the article. In regressions 1 to 4, the explanatory variable of interest is
the interaction between D(HRH) and RavenPack Release. D(HRH) is a dummy variable equal to 1 if an article was consistently released as highly
relevant in both RavenPack versions and 0 if it was originally released as having low relevance (HRL). RavenPack Release is a dummy variable
taking the value of 1 for articles after RavenPack went “live” on April 1, 2009, and zero otherwise. In regressions 5 and 6, the dependent variable
is the return (in percent) measured from 1 second before to 5 seconds after the article. The explanatory variable of interest is a triple interaction
between HRH, RavenPack Release and Sentiment Direction, where Sentiment Direction is a variable indicating the sentiment of the article derived
from RavenPack sentiment indices. It takes the value +1 for positive sentiment, 0 for neutral sentiment and −1 for negative sentiment. In all
regressions we include firm and date fixed effects. In regressions 2, 4, and 6, we add the following firm specific control variables: Company size,
Return prior month, Volatility prior month, Turnover prior month, Illiquidity prior month. In these regressions, we also add fixed effects for the
article category (e.g. mergers and acquisitions), the relevance score (from 90 to 100) and the hour during the day in which the article was released
as well as additional controls: the absolute return, turnover, and volatility each for industry and market from t−1 to t+5 seconds around the article.
In regression 2 and 4, we also include those values for t−1 to t+120 seconds around the article. All variables are defined in Appendix 1. All standard
errors are two-way clustered by stock and date. T-statistics are below the parameter estimates in parenthesis; ***, **, * indicate significance at the
1%, 5%, and 10% level, respectively.
Dependent Variable:
Speed of Stock Price
Response (in %)
Speed of Trade Volume
Response (in %)
Return t-1, t+5 (in bp)
(1)
(2)
(3)
(4)
(5)
(6)
RavenPack Release * D(HRH) * Sentiment
Direction
0.334
**
0.330
**
(2.10)
(2.08)
RavenPack Release * D(HRH)
1.744
***
1.709
***
0.506
*
0.488
*
0.121
0.109
(3.63)
(3.88)
(1.79)
(1.71)
(0.90)
(0.77)
RavenPack Release * Sentiment Direction
-0.307
**
-0.294
*
(-2.00)
(-1.91)
D(HRH) * Sentiment
0.079
0.114
(0.77)
(1.09)
D(HRH)
-0.081
-0.057
0.084
0.039
0.087
0.025
(-0.28)
(-0.21)
(0.59)
(0.27)
(0.89)
(0.24)
Sentiment Direction
0.431
***
0.222
**
(4.35)
(2.21)
Absolute Composite Sentiment Score
-0.027
***
-0.013
***
(-2.62)
(-2.66)
Neutral Composite Sentiment Score
-0.280
***
-0.111
***
(-3.72)
(-2.70)
Absolute Event Sentiment Score
0.044
***
0.035
***
(4.46)
(6.28)
Neutral Event Sentiment Score
-1.012
***
-0.529
***
(-3.10)
(-3.01)
Article category identified
-1.672
-2.300
**
0.580
**
(-0.39)
(-2.12)
(2.38)
Time since last article
0.054
***
0.085
***
0.165
***
(2.63)
(7.15)
(13.08)
Number of firms in article
-0.109
**
-0.132
***
-0.201
***
(-2.24)
(-5.63)
(-10.77)
Number of Observations
649435
649247
690462
690268
804002
803725
R
2
0.025
0.054
0.018
0.037
0.046
0.052
Relevance, Category and Hour Fixed Effects
No
Yes
No
Yes
No
Yes
Date and Firm Fixed Effects
Yes
Yes
Yes
Yes
Yes
Yes
Market control variables
No
Yes
No
Yes
No
Yes
Firm specific control variables
No
Yes
No
Yes
No
Yes
64
Appendix 1: Variable Definitions
This table displays the variable definitions for all variables used in the regressions. Article variables (sentiment scores, relevant scores, etc.) are
based on RavenPack 3. When we winsorize, we set outliers to the allowed extreme value; e.g., “smaller 10” means that any value below 10 is set
to 10. For all variables, winsorizing affects less than 1% of observations on either side.
Variable Name
Definition
Winsorizing
HRH
High relevance article Released as High relevance article. Article with relevance score of 90 or
higher in both RavenPack versions.
None
HRL
High relevance article Released as Low relevance article. Article with a relevance score of 90
or higher in the new RavenPack version, but was not assigned to the company or had a
relevance score below 90 in the old RavenPack version.
None
LRH
Low relevance article Released as High relevance article. Article with a relevance score below
90 or not assigned to the company in the new RavenPack version, but had a Relevance Score
greater or equal than 90 in the old RavenPack version.
None
RavenPack Release
Dummy variable equal to 1 after RavenPack is “live” (April 1, 2009) and equal to 0 before.
None
Company size
Log(prior day closing price * shares outstanding)
Smaller 10
Volatility prior month
Average of daily squared returns of the stock in the prior 20 trading days
Larger 2%
Turnover prior month
Average of daily volume divided by shares outstanding in the prior 20 trading days
Larger 10%
Return prior month
Average return in the prior 20 trading days
Larger 3% &
Smaller -3%
Illiquidity prior month
Percentile rank of all article-firm combinations of a day according to
. The most illiquid firms are assigned 100.
None
Relevance
Score provided by RavenPack that indicates the relevance of an article to a company and takes
values from 0 (least relevant) to 100 (most relevant).
Event Sentiment Score
Sentiment score that is provided by RavenPack; takes a value from 100 (positive) to 0 (negative).
It is available only for articles for which the category is identified.
None
Absolute Event Sentiment
Score
Abs (Event Sentiment Score – 50)
None
Neutral Event Sentiment
Score
Dummy variable equal to 1 if Event Sentiment Score equals 50 or if it is missing.
None
Composite Sentiment
Score
Sentiment score that is provided by RavenPack; takes a value from 100 (positive) to 0 (negative).
It is available for each article.
None
Absolute Composite
Sentiment Score
Abs (Composite Sentiment Score – 50)
None
Neutral Composite
Sentiment Score
Dummy variable equal to 1 if Composite Sentiment Score equals 50.
None
Sentiment Direction
Takes the values 1 (positive sentiment), 0 (neutral sentiment) and −1 (negative sentiment). It is
first based on Event Sentiment Score (ESS). If ESS is larger 50, this variable is 1, if ESS is smaller
than 50, it is −1. If ESS is missing or 50, we consult Composite Sentiment Score (CSS). If CSS
is greater than 50 we set this variable to 1, if CSS is smaller than 50 we set it to −1, if CSS equals
50 we set it to zero.
None
Article category
identified
Dummy variable equal to 1 if the category (e.g. “merger”) of the article is identified
None
Number of firms in article
Log ( Number of firms in article)
None
Time since last article
Log (Time since last article in seconds)
None
Return t-1, t+5
Stock return from 1 second before to 5 seconds after the article. Returns are computed from
mid-quotes.
Larger 2%,
smaller -2%
Return t+6, t+120
Stock return from 6 seconds after to 120 seconds after the article. Returns are computed from
mid-quotes.
Larger 2%,
smaller -2%
Speed of Stock Price
Response
None
Speed of Stock Price
Response – Market Adj.
Set to missing if:
= 0.
None
Speed of Stock Price
Response – Industry
Adjusted
Set to missing if:
= 0.
None
65
Speed of Trade Volume
Response
None
Speed of Directional
Trade Volume Response
, where Turnover in Direction of the Article is buyer
initiated turnover for articles with positive Sentiment Direction and seller initiated turnover for
articles with negative Sentiment Direction. The direction of a trade is determined using the Lee
and Ready (1991) methodology, but using the quote at the end of the previous second as the
prevailing quote rather than the quote 5 seconds ago.
None
D(Press Release)
Dummy variable equal to 1 if the article is a press release.
None
D(First company article
of the day)
Dummy variable equal to 1 if the article is the first article of the day within our sample period.
None
D(New Story)
Dummy variable equal to 1 if the article has Event Novelty Score (ENS) of 100, which means
it is a new news story.
None
D(Sentiment predicts
return)
Dummy variable equal to 1 if the sign of the return from 1 second before to 120 seconds after
the article matches the sign of the article sentiment (it is set to missing for articles with neutral
sentiment)
Signed Return t-1, t+120
This variable is set to missing if Sentiment Direction is equal to zero.
Larger 2%,
smaller -2%
Past Informativeness
, the mean is taken over the prior six calendar months.
None
HFT Trading Fraction
Response
where
None
HFT Active Trading
Fraction Response
where
None
Passive HFT Trading
Fraction
Defined analogous to HFT Active Trading Fraction Response counting only trades using
liquidity supplying trades.
With News Trading
Fraction
Defined analogous to HFT Active Trading Fraction Response counting only trades in direction
of news sentiment (missing for neutral sentiment).
Against News Trading
Fraction
Defined analogous to HFT Active Trading Fraction Response counting only trades against
direction of news sentiment (missing for neutral sentiment).
, where r
it
is the return for stock i during second t; dolvol
it
is the dollar volume for
stock i during second t; and N
ij
is the number of seconds in which stock i traded during interval
j.
Larger 2
, where buys
it
(sells
it
) is the number of
shares traded and initiated by the buyer (seller) for stock i during second t;
is the last
execution price for stock i during second t;
is the last bid-ask midpoint for stock i
during second t and Nij is the number of seconds in which stock i traded during interval j. The
direction of a trade is determined using the Lee and Ready (1991) methodology, but using the
quote at the end of the previous second as the prevailing quote rather than the quote 5 seconds
ago.
Larger 3
Change in Amihud
Illiquidity
.
None
Change in Effective
Spread
.
None
Market return t-1, t+5
Value-weighted return of all common stocks in TAQ (which are also in CRSP) from 1 second
before to 5 seconds after the article. Returns are computed from mid-quotes.
None
Market turnover t-1, t+5
Total dollar trading volume of all common stocks in TAQ (which are also in CRSP) from 1
second before to 5 seconds after the article divided by total market capitalization at t-2.
None
Market volatility t-1, t+5
Value weighted average squared second return of all common stocks in TAQ (which are also in
CRSP) averaged from 1 second before to 5 seconds after the article.
Larger 20 bp
Market adjusted return t-
1, t+5
Return (t-1, t+5) − Market Return (t-1, t+5)
Larger 2%,
smaller -2%
Industry variables
Follow the same definition as market variables but include only stocks within the firm’s 12
Fama French industry
Same as
market
variables
