Front Page News: The Effect of News Positioning on Financial Markets

Front Page News:

The Eﬀect of News Positioning on Financial Markets

Anastassia Fedyk

∗

March 12, 2018

Click Here for the Latest Version

Abstract

This paper estimates the eﬀect of presentation of information on ﬁnancial mar-

kets, using a natural experiment in prominent “front page” positioning of news on the

Bloomberg terminal. The front page and non-front page articles are indistinguishable

by either algorithmic analysis or by the target audience of active ﬁnance professionals.

Front page positioning induces 280% higher trading volumes and 180% larger price

changes within the ﬁrst ten minutes after news publication, followed by a strong drift

for 30-45 minutes. Subsequently, non-front page news begins to catch up, but the

incorporation of this information is substantially more gradual, and the initial eﬀects

of positioning persist for days after publication. The short-term eﬀects induced by

positioning are even stronger than diﬀerences between articles of varying editorial im-

portance.

Keywords: information diﬀusion, news positioning, asset pricing, trading volume

∗

Harvard University, Department of Economics and Harvard Business School. Mail: Baker Library 244C,

25 Harvard Way, Boston, MA 02163. Email: [email protected]. I am particularly indebted to John Camp-

bell, Lauren Cohen, David Laibson, Chris Malloy, Andrei Shleifer, and Jeremy Stein for their invaluable

guidance and advice. I am also grateful to Daniel Andrei, Malcolm Baker, John Beshears, Joshua Coval,

Zhi Da (discussant), Kent Daniel, Stefano DellaVigna, Tatiana Fedyk, Xavier Gabaix, Tom Glocer, Robin

Greenwood, Sam Hanson, James Hodson, Marc Kaufmann, Patrick Kelly (discussant), Julien Penasse (dis-

cussant), Matthew Rabin, Stas Sokolinski, Adi Sunderam, Carmen Wang, Louis Yang (discussant), Vladimir

Zdorovtsov, Florian Zimmermann, and seminar participants at the Federal Reserve Board, Emory Univer-

sity, Georgetown University, Harvard University, Institut Jozef Stefan, London Business School, London

School of Economics, Stanford University, State Street Global Advisers, UC Berkeley, UCLA, University of

Virginia, University of Zurich, the American Finance Association, the European Finance Association, the

Trans-Atlantic Doctoral Conference, and the Western Finance Association for insightful comments. This

paper received the 2017 Best Ph.D. Paper Award from the European Finance Association and was a ﬁnalist

for the Hillcrest Behavioral Finance Award. I thank Lauren Cohen and Chris Malloy for sharing the Quant-

Quote data and Judes Echauz and Katelyn Haruki for providing outstanding research assistance. Funding

for this research was generously provided by the Harvard Business School, Hillcrest Asset Management, and

the Pershing Square Venture Fund for Research on the Foundations of Human Behavior.

1 Introduction

How does information get incorporated into asset prices? A number of theoretical models

propose potential frictions that may prevent even publicly available information from being

instantaneously reﬂected in prices.

Multiple empirical studies lend suggestive evidence to

this view.

However, tracing out incorporation of information in real time remains diﬃcult,

and requires a detailed understanding of the variation across individual pieces of information.

In this paper, I capture the causal eﬀect of presentation of news on the way the underlying

information is incorporated into asset pricing, using a natural experiment in the way news

articles are pinned to the top of the Bloomberg terminal news screen, in the “front page”

positions. I show that when the news is prominently positioned, the price response occurs

within an hour of publication. By contrast, for comparable non-front page news, the price

formation process takes 10-15 days to achieve the same response. In particular, within

ten minutes of publication, news articles that get pinned to the front page induce 280%

higher trading volumes and 180% larger price changes. Beyond that, front page articles

are accompanied by a strong price drift for approximately 30-45 minutes after publication,

consistent with the fact that these articles remain prominently positioned for approximately

half an hour. After that, the information in front page news appears to be fully incorporated,

and the reactions to non-front page articles begin to gradually catch up. However, the

incorporation of non-front page information is much slower: although the price paths after

front page and non-front page news eventually converge, this process takes multiple days.

Interestingly, diﬀerences in news positioning have an even stronger eﬀect on short-term

market dynamics than diﬀerences between news articles marked with distinct importance

labels by the editorial staﬀ.

My empirical design exploits a natural experiment based around a category of Bloomberg

news articles whose placement depends on the contemporaneous volume of other articles,

See, for example, Peng and Xiong (2006), DellaVigna and Pollet (2009), and Andrei and Hasler (2014)

on limited attention to publicly available information; and Harris and Raviv (1993), Kandel and Pear-

son (1995), Cao and Ou-Yang (2009), and Banerjee and Kremer (2010) on diﬀerential interpretations of

public information. A large related literature including Kyle (1985), Holden and Subrahmanyam (1992),

Wang (1994), Hirshleifer, Subrahmanyam, and Titman (1994), Cao, Coval, and Hirshleifer (2002), Foucault,

Hombert, and Ro¸su (2016), and Andrei and Cujean (2017), among others, considers incorporation of private

information.

See, for example, Foster, Olsen, and Shevlin (1984), Bernard and Thomas (1989), and Peress (2008) on

the post-earnings announcement drift; Bali, Peng, Shen, and Tang (2014) on inattention and underreaction

to liquidity shocks; Loh (2010), Da, Engelberg, and Gao (2011, 2015), Drake, Roulstone, and Thornock

(2012, 2015, 2016), Schmidt (2013), Curtis, Richardson, and Schmardebeck (2014), and Ben-Rephael, Da,

and Israelsen (2017) on predictability of market dynamics from proxies of attention; Huberman and Regev

(2001), Tetlock (2011), Gilbert, Kogan, Lochstoer, and Ozyildirim (2012), and Fedyk and Hodson (2015)

on reactions to stale news; and Carvalho, Klagge, and Moench (2011) and Marshall, Visaltanachoti, and

Cooper (2014) on reactions to false news.

rather than on their own content. I focus on news articles about individual U.S. equity

securities, and hand-collect a sample of news between March 2014 and December 2015. The

news articles in my hand-collected sample fall into three categories: “primary important,”

“secondary important,” and “all other” news. News articles marked as “primary important”

are always pinned to the prominent front page positions, displacing the previous front page

news and remaining on the front page for, on average, twenty to forty miutes. News articles

marked as “all other” are never placed into front page positions. News articles marked

as “secondary important” constitute the category of interesting variation. Any particular

news article in this category is given a front page slot if and only if, at the precise moment

when that article is released, there is at least one such slot remaining from the “primary

important” news. As a result, “secondary important” news articles that make it to the front

page and those that do not are marked as equally signiﬁcant. Their positions vary due to

contemporaneous numbers of “primary important” articles, rather than their own underlying

content.

I structure the empirical analysis of the market dynamics following front page versus non-

front page “secondary important” news articles using a theoretical framework that reﬂects

standard models of limited attention and gradual information diﬀusion.

The three-period

model considers a news signal published by the main news source of interest and also reported

by alternative news sources. The framework incorporates two standard features from models

of gradual information diﬀusion: (1) only a fraction of investors are attentive to the news

signal from each source in each period; and (2) investors update their beliefs in a na¨ıve

Bayesian manner, incorporating their own information but not rational expectations of the

information available to other investors. Front page positioning is represented by more

prominent and longer-lasting reporting by the main news source. A larger incidence of

investors are attentive to the news signal from the main source when it is published on the

front page, and this persists into the second period.

This framework generates several predictions. First, the front page news articles are

accompanied by larger immediate trading volumes and absolute price changes. Second, the

initial returns accompanying front page news articles are more likely to continue in the

short-term. Third, the front page news articles induce lower longer-term price continuation.

My empirical results conﬁrm these predictions. There are signiﬁcant diﬀerences in market

dynamics following “secondary important” news articles that are pinned to the front page

and those that are not. Consistent with the ﬁrst prediction, front page news articles are

accompanied by substantially higher trading volumes and absolute price changes for the

For models of limited attention and gradual information diﬀusion, see, for example, Hong and Stein

(1999), DellaVigna and Pollet (2009), and Andrei and Hasler (2014).

tagged securities immediately after publication. For example, these articles are, on average,

accompanied by 280% larger trading volumes and 180% larger absolute price changes during

the ten minutes following article publication.

Since pinning a news article to a front page position makes it visible for a longer period

of time, the front page positioning also induces more persistent short-term market reactions,

conﬁrming the second prediction of the gradual information diﬀusion framework. Front page

news articles are accompanied by signiﬁcantly higher serial correlation in price changes over

a variety of short-term horizons. For example, these articles are, on average, followed by

17% larger serial correlations in price changes across consecutive ﬁve-minute intervals.

I also ﬁnd empirical support for the third prediction of the model: after the initial period

of about forty-ﬁve minutes, the price drift is signiﬁcantly stronger for non-front page news

articles. Most front page articles are displaced from their prominent positions after twenty to

forty minutes, and the incorporation of front page information is virtually complete within an

hour of news publication. The incorporation of non-front page information takes much longer.

We begin to see some convergence during the ﬁrst few hours after the news. For example,

the initial returns from the ﬁrst 30-45 minutes after publication of non-front page news

are accompanied by a drift of 14-29% over the subsequent hour. However, the convergence

is quite gradual, and diﬀerences in price eﬀects persist even days after news publication.

Securities mentioned in front page news articles see 34 basis points larger absolute cumulative

returns measured from the moment of publication to two days later, relative to securities

mentioned in non-front page news; this diﬀerence is statistically signiﬁcant at the 10% level.

Five days out, the diﬀerence declines to a statistically insigniﬁcant 25 basis points, and

ﬁfteen days after the news, the diﬀerence is a statistically indiscernible 8 basis points.

To contextualize the economic importance of these ﬁndings, I compare the market eﬀects

of news positioning against the eﬀects of news importance. In particular, I estimate market

dynamics following two sets of news articles that receive equally prominent positions but

that diﬀer in their importance, as marked by the editorial staﬀ. These are: (1) “secondary

important” articles that make it to the front page; and (2) “primary important” articles,

all of which make it to the front page by default. Articles in both of these categories are

prominently positioned, but the articles in the second category are marked by the editorial

staﬀ, ex ante, to be more important than those in the ﬁrst category.

I ﬁnd that news importance is not as signiﬁcant in driving short-term market activity

as positioning. Trading volumes induced by news publication are not statistically diﬀerent

for securities mentioned in more (“primary important”) versus less important (“secondary

important”) news articles pinned to the front page. Absolute price changes are 80% (66%)

larger during the ﬁrst ﬁve (ten) minutes following the more important news articles, but

the relative diﬀerence is smaller than that induced by the news positioning. The short term

price drift is statistically indistinguishable for more and less important news articles, holding

front page position constant. Overall, the results indicate that news positioning plays an

even larger role for short-term market dynamics than editorial markings of the importance

of the underlying news.

I perform a number of additional analyses to conﬁrm that the results are not driven

by systematic diﬀerences between “secondary important” articles that receive a front page

slot and those that do not. First, I consider the possibility that, due to market participants’

distraction during periods with high volumes of news, articles published during quieter times

garner larger reactions. To address this possibility, I hold position constant and compare non-

front page “secondary important” articles released during times with diﬀerent amounts of

contemporaneous news activity. I document that the non-front page articles released during

quiet times are, if anything, accompanied by less substantial reactions than the non-front

page articles released during busy times.

Second, using techniques from machine learning and a representative corpus of ﬁnancial

news from Reuters, I learn the mixtures of topics generally discussed in ﬁnancial news, such

as earnings announcements, technology, and litigation. I then use the trained model to com-

pare the distributions of identiﬁed topics appearing in the text of the individual Bloomberg

news articles in my hand-collected samples. I ﬁnd no systematic diﬀerences between the dis-

tributions of topics discussed in the front page versus non-front page “secondary important”

news articles. The distribution of topics covered by the “primary important” news articles,

by comparison, does diﬀer slightly from the distribution of topics appearing in “secondary

important” news.

Third, a survey of 150 active ﬁnance professionals indicates that absent salient posi-

tioning, market participants ﬁnd front page “secondary important” headlines to be indis-

tinguishable from the non-front page ones. The survey participants consist of key decision

makers at a broad range of ﬁnancial institutions, including broker dealers such as Bank of

America and Goldman Sachs, investment management ﬁrms such as BlackRock and PIMCO,

hedge funds such a Bridgewater and AQR, and private equity ﬁrms such as Blackstone and

Warburg Pincus. The ﬁnance professionals conﬁrm Bloomberg editorial staﬀ’s judgment of

news importance. They consistently identify the “primary important” news articles as, on

average, more impactful than “secondary important” news articles (“primary important”

headlines are chosen as more impactful 61% of the time, signiﬁcantly diﬀerent from 50%).

By contrast, these ﬁnance professionals identify the front page “secondary important” news

articles as more impactful than their non-front page counterparts only 48% of the time, not

signiﬁcantly diﬀerent from 50%. I also repeat this analysis using a smaller survey of 27 MBA

students from top business schools and ﬁnd qualitatively similar results.

My ﬁndings build on the growing literature evaluating the impact of media on ﬁnancial

markets.

Prior empirical strategies for estimating the causal impact of media use exoge-

nous variation in news arrival through weather-related disruptions (see Engelberg and Par-

sons (2011)), newspaper strikes (see Peress (2014)), disruptions to boat routes (see Koudijs

(2016)), and staggered implementation of robo-journalism (see Blankespoor, deHaan, and

Zhu (2017)), as well as variation in security relevance tags (see von Beschwitz, Keim, and

Massa (2017)) and headline complexity and degree of quantiﬁcation (see Umar (2017) and

Huang, Nekrasov, and Teoh (2017)). Klibanoﬀ, Lamont, and Wizman (1998) ﬁnd that for

closed-end country funds, the incidence of news on the front page of the New York Times is

correlated with a higher elasticity of price with respect to asset value. Huberman and Regev

(2001) further highlight the importance of prominent news positioning by analyzing a case of

(mostly) stale information, initially reported in Nature in November 1997, getting reprinted

on the front page of the New York Times in May 1998. Furthermore, Lawrence, Ryans, Sun,

and Laptev (2017) present compelling evidence that promotion of earnings announcement

news on Yahoo! Finance to a subset of website visitors increases the abnormal return on the

announcement date.

The present paper contributes to the literature by providing a clear counterfactual. The

natural experiment in positioning of news on the Bloomberg terminal oﬀers clean variation

in institutional investor attention in an important setting that represents the main source of

information for a large set of ﬁnance professionals. This allows me to document two things.

On the one hand, when information is especially saliently highlighted, the market response

is quite eﬃcient: prices respond within an hour (and largely within the ﬁrst minutes) of

news publication. These highlighted news events, for which limited attention and cognitive

processing limitations play a minor role, illustrate a best-case scenario. On the other hand,

the price formation process is this eﬃcient only for especially highlighted news. In other cases

– even with public, easily accessible news consumed by sophisticated institutional investors

– attention is more gradual and the price formation process takes substantially longer, on

the order of days or even weeks.

See Busse and Green (2002), Barber and Loeﬄer (1993), Chan (2003), Fehle, Tsyplakov, and Zdorovtsov

(2005), Antweiler and Frank (2006), Barber and Odean (2008), Fang and Peress (2009), Engelberg, Sasseville,

and Williams (2012), Solomon (2012), Dougal, Engelberg, Garc´ıa, and Parsons (2012), Rogers, Skinner, and

Zechman (2013), Hillert, Jacobs, and M¨uller (2014), Ahern and Sosyura (2014), Liu, Sherman, and Zhang

(2014), Yuan (2015), and Boulland, Degeorge, and Ginglinger (2017).

The importance of prominent positioning and alphabetical ordering has also been documented in other

contexts; see, for example, Ho and Imai (2008), Jacobs and Hillert (2015), and Feenberg, Ganguli, Gaule,

and Gruber (2017). In ﬁnancial markets, presentation has been shown to drive mutual fund ﬂows (Kaniel

and Parham (2017)) and attract attention to securities independent of information ﬂows (Wang (2017)).

These ﬁndings provide systematic evidence that it is not enough to make ﬁnancially-

relevant information easily accessible: how saliently the information is presented plays an

important role in determining whether the information is immediately reﬂected in asset

prices. The price impact induced by front page positioning occurs quickly, but the compa-

rable non-front page information takes surprisingly long to converge, given that these are

all easily accessible news articles available on the Bloomberg terminal. For more obscure or

private information, similar mechanisms are likely to apply at longer horizons, generating

phenomena such as months-level momentum.

The remainder of the paper proceeds as follows. Section 2 describes the data and the

natural experiment in news positioning. Section 3 outlines the conceptual framework of

market dynamics following more and less prominently positioned news. Section 4 presents

the key empirical ﬁndings on the diﬀerential market dynamics following front page and non-

front page news articles. Section 5 explores the eﬀect of news importance, holding position

constant, by comparing “secondary important” news articles that are positioned on the front

page against “primary important” front page news articles. Section 6 presents additional

analyses of news content, conﬁrming that the front page “secondary important” articles in

the sample are indistinguishable from their non-front page counterparts by both algorithmic

analysis and the target audience of market participants. Section 7 concludes.

2 Data Sources and Empirical Strategy

In order to capture the casual eﬀect of news presentation on trading volumes and returns,

I use quasi-random variation in positioning of news articles on the Bloomberg terminal.

Two key features of these data make them especially well-suited to the current analysis.

First, Bloomberg is one of the largest ﬁnancial news providers and a main source of news

for ﬁnance professionals, making it an ideal setting to estimate the eﬀect of attention to

news on ﬁnancial markets. Second, the data include a natural experiment of quasi-random

positioning for a subset of news articles. The news data are merged with market data to

relate news presentation to trading volumes and price formation.

2.1 Natural Experiment in News Positioning

In this subsection, I describe the quasi-random variation in news positioning that I use

in my research design. In particular, I concentrate on a subset of news articles that are

sometimes prominently positioned, and sometimes not, depending on the volume of other

articles released around the same time and not on the characteristics of the news articles

themselves.

The full sample of news passing through the Bloomberg terminal is aggregated from a

variety of sources in real-time. The sources of news include key national and international

news wires from a comprehensive set of news organizations, company ﬁlings, press releases,

and content from web sources, including blogs and social media. The news articles are

disseminated electronically to over 300,000 ﬁnance professionals through the subscription-

based terminal. Overall, there are millions of articles tagged with U.S. equity securities

during the sample period of March 22, 2014 - December 31, 2015.

There are diﬀerences in how Bloomberg presents individual news articles on the terminal.

Generally, the news screen features a scrolling list of news articles, where newly published

articles replace the older ones at the top of the screen. However, some of the news articles

written directly by Bloomberg News get pinned to the top of the screen. At any given point

in time, there are at most three such pinned articles. Figure 1 shows a screenshot of a

default Bloomberg news screen covering all company-relevant news. The top three articles

are pinned and remain at the top, while the articles below continually move down as new

publications arrive. It is these positions, highlighted in yellow font at the top of the default

company news screen in Figure 1, that I term “front page” throughout this paper.

Eﬀectively, there are three broad categories of news articles passing through the Bloomberg

terminal: the primary “primary important” (PI) articles; the “secondary important” (SI)

articles, and “all other” articles. The assignment of individual articles to these categories

reﬂects the journalistic and editorial opinion regarding the importance of a given piece of

news. Each of the two important categories, PI and SI, comprises roughly 0.1-0.5% of all

news, so both of these categories of articles capture news of fairly rare perceived importance.

I exclude market wrap articles, in order to focus on new information relevant to individual

securities, and hand-collect all articles that are tagged with at least one publicly traded U.S.

equity security, that are published between 8AM and 5PM EST during the sample period,

and that are either in the PI category (1,419 unique PI articles) or the SI category (4,887

unique SI articles).

For the most part, PI articles represent signiﬁcant company news, such as earnings reports

and M&A decisions. A few representative examples of the sample of PI news are provided

in Panel 1 of Table 1.

SI articles likewise include signiﬁcant events, such as changes in regulation and drug

approvals. However, this set of news also features articles that are likely to capture the

readers’ curiosity, but that are less immediately relevant to ﬁnancial markets, such as moves

of top well-known traders and perks in ﬁnancial ﬁrms. A few representative examples of SI

articles are presented in Panel 2 of Table 1.

The classiﬁcation of articles into categories of relative importance plays a role in how

prominently the articles are positioned. When an article from the PI category is released, it

is immediately placed in a prominent front page position, displacing whichever news article

was in that position previously. Once on the front page, a news article remains there until

the earlier of two things occurs: either a new PI article comes out and displaces the old

article, or a predeﬁned amount of time (on the order of hours) elapses. Occasionally, there

are not enough PI articles at a given point in time to ﬁll all of the front page slots. In this

case, the next SI article to be published, upon its release, takes the available front page

position. The process of article positioning is depicted in Figure 2.

As a result, there are two categories of news articles deemed equally important but having

diﬀerent positions: the SI articles that come out at a time when there are available front

page slots and the SI articles that come out at a time when front page slots are unavailable.

I hand-collect the positions of the SI articles in my sample. This subset of the news sample

– SI articles in various positions – forms the basis for my causal analysis.

Screening of the articles conﬁrms that there are no systematic diﬀerences in content

between SI news articles that are placed on the front page and those that are not. Both

include signiﬁcant events, such as:

• “T-Mobile Said to Plan to Turn Down Iliad’s $15 Billion Oﬀer” (not front page)

• “Chipotle Probed for New Outbreak of Diﬀerent E. Coli Strain” (front page)

But both front page and non-front page SI news articles also feature news events that

carry less immediately relevant impact for ﬁnancial markets. For example:

• “Morgan Stanley Gets 90,000 Applications for Summer Program” (not front page)

• “Pimco Said to Have Discussed Firing Gross Before Exit to Janus” (front page)

In Section 6.2, I compare the texts of the front page and non-front page SI news articles

formally using machine learning techniques. I ﬁnd no systematic diﬀerences between the

two categories of news. Similarly, a survey of active ﬁnance professionals and MBA students

from top business school programs indicates that human ﬁnancial experts do not perceive the

front page SI news articles to be any more signiﬁcant than their non-front page counterparts.

Table 2 presents the distribution over time of PI and SI news articles published between

the hours of 8AM and 5PM. All numbers are cited in ticker-articles, so that articles tagged

with more than one U.S. equity security ticker are included one time for each tagged U.S.

security. Overall, there are 2,362 PI article-tickers in the sample and 8,233 SI article-tickers,

of which 1,274 are given a front page position. The articles are roughly evenly distributed

across the months of the year, with a lower volume of articles in January and February (and,

to a lesser extent, March), since the sample begins on March 22, 2014 and hence does not

cover these months in 2014. Over hours of the day, PI news articles peak at the start and end

of the business day, during 8-10AM and especially during 4-5PM, while the SI news articles

are more evenly distributed during the day. Consistent with the SI articles’ positioning being

determined by the concurrent volume of PI news, a lower percentage of SI articles makes it

to the front page during the hours that see a higher volume of PI articles. The correlation

between the hourly numbers of PI articles and the hourly likelihoods of SI articles receiving

front page positions is -81%.

Examining the timing of news releases in the sample, I ﬁnd no evidence of strategic

release timing of the SI news articles. Of all the front page SI news articles, only 1.4% have

a non-front page article released up to one minute before or after the front page article’s

publication. Similarly, only 0.7% of the front page SI articles are accompanied by non-front

page SI articles within 30 seconds before or after. A mere 0.2% of the articles in the front

page SI sample have a non-front page article released within 10 seconds of their publication.

This low volume of SI news articles leaves little scope for inﬂuencing article position by

strategically timing the exact seconds of when the articles are released. As a result, the

process is unlikely to be contaminated by editorial staﬀ being faced with multiple SI articles

to be released at the same time and strategically releasing the more important ones ﬁrst.

I also ﬁnd that the article volume does not appear to be driven by editorial targets. I

observe the distribution of articles across days and ﬁnd that the volumes of PI and SI news

articles vary dramatically from day to day. The number of PI news articles ranges from 0 to

40 per day, while SI articles can number anywhere between 0 and 67 per day. There is also

little relationship between the numbers of PI and SI articles on any given day. The daily

numbers of the two types of articles display a low correlation of 25%. As shown in Figure 3,

any given day can see a large number of PI articles accompanied by few SI articles, and vice

versa. Overall, the distribution of PI and SI articles across days indicates that the editorial

staﬀ is not targeting particular numbers of high-importance articles. Instead, the patterns

are more consistent with the evaluation of each article’s importance being based on its own

merit, independently of the volume of other news.

2.2 Market Data

I use the security ticker tags to merge the news position data with market data from

several sources. Industry classiﬁcation, market capitalization, and shares outstanding come

from Compustat. High frequency price and trading data come from QuantQuote. The

second-resolution QuantQuote data include all tickers listed on NYSE and NASDAQ ex-

changes, and provide prices and numbers of shares traded for each second during the market

open. The data are adjusted for splits, dividends, and symbol changes.

The high frequency tests are run using news articles tagged with all ﬁrms for which

there are pricing data in QuantQuote, and shares outstanding and NAICS industry codes in

Compustat. The merged sample includes 948 front page SI article-ticker pairs, 4,930 non-

front page SI article-ticker pairs, and 1,650 PI article-ticker pairs. All of these article-ticker

pairs have at least one price data point in QuantQuote on the day of publication, but not

necessarily within shorter windows. Recall that PI news articles are more likely to come out

during the hours of 8-9AM EST and especially 4-5PM EST. As a result, the empirical tests,

which require market data within short windows of publication, reduce the PI news sample

more substantially than the two SI news samples.

3 Conceptual Framework

In this section, I present a conceptual framework formalizing the intuition regarding the

diﬀerences between front page and non-front page news articles. I outline two key aspects in

the way investors are likely to pay diﬀerential attention to news articles in diﬀerent positions,

and then trace out the implications of these aspects for the process of incorporation of

information into asset prices.

The conceptual framework follows the setups in Hirshleifer and Teoh (2003) and Della-

Vigna and Pollet (2009). There is a risk-free asset with a zero rate of return and a single

risky security with a stochastic payoﬀ R normally distributed with mean R and variance σ

realized in an unmodeled ﬁnal period T . In the relatively short-term empirical settings that

I consider, the realized value R can be taken to denote, for example, the price on which an

asset settles in the days following an earnings announcement or the price of the combined

enterprise following an acquisition. The risky asset is in ﬁxed supply X. For expositional

simplicity, I ﬁx X = 0, so that the asset is in zero net supply; this simpliﬁes the notation

without aﬀecting the results.

There is a continuum of investors with total mass equal to 1, who maximize mean-variance

utility. In particular, let W

(i)

denote investor i’s ﬁnal wealth at the end of the game at time

T . Then at any point in time t, investor i maximizes expected utility of the form

i,t

(i)

} −

(i)

V ar

i,t

(i)

} (1)

with respect to his current holdings. For expositional simplicity, I take the risk-aversion

coeﬃcient to be identical across investors and normalize it to one: ∀i, A

(i)

= 1. Each

investor i is initially endowed with wealth W

(i)

. There are no liquidity constraints.

Information in this framework is modeled as a signal arriving at a particular point in

time and gradually diﬀusing across the population of investors. In particular, there are four

periods in the model. In period 0, investors form prior expectations regarding the distribution

of R. In period 1, a noisy signal (news) is released, and investors update their expectations

accordingly. In periods 2 and 3, investors continue to update their beliefs following the news

signal. At the end of the game, in the unmodeled period T , the true value of R is realized

and the investors consume their ﬁnal wealth. I assume the following form for the news signal:

N = R + , where  is a normally distributed noise term, independent of R, with mean 0

and variance σ



The news signal is not immediately observed by all investors. Instead, the main news

source, S, reports the news signal N for some number of periods. Mass γ > 0 of investors

are attentive to the main source S in each period t. Thus, in each period t that S reports

the news signal N, a fraction γ of investors who had not observed the news signal prior to t

now become aware of N.

I model the diﬀerence between front page and non-front page news with two key features.

First, front page news articles induce more attention overall, so that the fraction of investors

attentive to the news signal is higher: γ = γ in the case of front page news and γ = γ < γ

in the case of non-front page news. Second, front page new corresponds to the signal being

reported by S for longer. Thus, for non-front page news, investors can observe the signal N

from the main source S only in period 1. For front page news, by contrast, investors can

also observe the signal from the main source S in period 2.

Investors may also learn the news from alternative sources, albeit at a lower rate. In

particular, in any period when the news is not being reported by S, a fraction ξ > 0 of

uninformed investors still observe the news signal. This additional information channel can

be interpreted as investors ﬁnding the news through ﬁlters or active searches once it scrolls

oﬀ the top of the Bloomberg terminal screen, or reading the news from other providers. This

channel is a minor one in the model, and I assume that most investors who receive the news

do so from the main source S. In particular, I assume that:

ξ <

1 − γ

γ (2)

This condition ensures that once the main news source stops actively reporting the news

(i.e., when the news is not on the front page), the fraction of informed investors does not

increase faster than when the source continues to report (front page news). Consistent with

the information disseminating relatively slowly over the short horizons considered in my

empirical analysis, I also assume that both γ and ξ are small: γ, ξ << 1/2.

The model timeline is depicted in Figure 4. In each period t, let I

denote the set of

informed investors, who observe the news signal either during or prior to t, and let F

= |I

be the share of informed investors. I denote the remaining uninformed investors by U

. Let

F P

and F

NF P

denote the values of F

in the cases of front page and non-front page news,

respectively. Figure 4 illustrates the arrival of information and the evolution of the share of

informed investors for both front page and non-front page news.

The key frictions in the model are that (1) some investors are inattentive; and (2) investors

update their beliefs in a na¨ıve Bayesian manner. Namely, some of the investors do not

observe the public signal, and all investors update their beliefs with respect to only their

own information, without taking into account the information sets and actions of others. In

particular, while all investors observe equilibrium prices in all periods, they do not use the

information contained in the price history to update their beliefs. These assumptions are

standard modeling devices in models of gradual information diﬀusion (see Hong and Stein

(1999), Hirshleifer and Teoh (2003), or Peng and Xiong (2006)).

I characterize the price path and trading volume following a news signal as a function

of the fraction of attentive investors F

(Section 4.1). The empirical predictions for the

diﬀerences in market dynamics following front page and non-front page news are then derived

in Section 4.2.

3.1 Evolution of Prices and Trading Volumes

I begin by characterizing the price levels and trading volumes in terms of the fraction

of attentive investors F

, without distinguishing whether the news signal is reported on the

front page or not.

Price levels. First, note that the uninformed investors hold the prior beliefs that the

return R is normally distributed with mean R and variance σ

. The informed investors

attend to the signal and update their beliefs in a na¨ıve Bayesian manner. Hence, their

beliefs are given by:

∀t ∈ {0, 1, 2, 3}, i ∈ I

: E

(i)

{R} =



R + σ

+ σ



; V ar

(i)

{R} =



+ σ



(3)

Next, note that optimization of the mean-variance preferences given by (1) with the above

beliefs results in the following demand functions by the two groups of investors during any

period t:

∀t ∈ {0, 1, 2, 3}, i ∈ I

: x

(i)



(R − P

) + σ

(N − P

)



(4)

∀t ∈ {0, 1, 2, 3}, i ∈ U

: x

(i)

R − P

(5)

where P

denotes the price of the risky asset in period t.

The market clearing condition each period is that the total demand from the informed

and uninformed investors must equal the zero net supply. Hence, in each period t, the price

of the asset P

must satisfy:

∀t ∈ {0, 1, 2, 3} : F



(R − P

) + σ

(N − P

)



+ (1 − F

)

R − P

= 0 (6)

Solving this equation gives the following expression for the price of the asset during each

period t:

∀t ∈ {0, 1, 2, 3} : P



+ F

R +



+ F

N (7)

Absolute price changes. Taking the ﬁrst diﬀerences yields the absolute price change

between any two consecutive periods:

∀t ∈ {1, 2, 3} : |∆P

| = |P

− P

t−1

| =

− F

t−1

)σ



|N − R|

(σ



+ F

)(σ



+ F

t−1

)

(8)

Price continuation. In order to calculate the continuation in the price path, recall that

the news signal has the form N = R + , where R and  are independent normal variables

with R ∼ N (R, σ

) and  ∼ N (0, σ



). Hence, price continuation, measured as the slope in

a regression predicting the price change in period t + 1 from the price change in period t, is

given by:

∀t ∈ {1, 2} : Cont(t, t + 1) =

Cov(∆P

, ∆P

t+1

)

V ar(∆P

)



t+1

− F

t−1





+ F

t−1



+ F

t+1



(9)

Note that this expression is deﬁned for any setting where a non-trivial set of investors

learns the news during the earlier period t. This holds for both the front page and the

non-front page news in my setting, since even in absence of reporting by the main source S,

news diﬀuses at the low but nonzero hazard rate ξ.

Trading volumes. Trading volume in each period t consists of all holdings that ex-

change hands between periods t − 1 and t. In each period, the newly informed investors, i.e.

investors i ∈ I

∩ U

t−1

change their demand following receipt of the news signal, inducing a

change in the equilibrium price and the other investors’ equilibrium holdings. Let x

(I)

de-

note the equilibrium holdings, in period t, of an investor i ∈ I

; similarly, let x

(U)

denote the

equilibrium holdings of an investor u ∈ U

. Trading volume in each period can be expressed

as a function of the newly informed investors’ holdings as follows:

∀t ∈ {1, 2, 3} : T V

= (F

− F

t−1

)|x

(I)

− x

(U)

t−1

| (10)

Taking the holdings from (4)-(5) and the equilibrium price levels from (7) then gives the

following expression for each period’s trading volume:

∀t ∈ {1, 2, 3} : T V

= (F

− F

t−1

)

(1 − F

)σ



+ F

t−1

(σ



+ σ

)

(σ



+ F

)(σ



+ F

t−1

)

|N − R| (11)

3.2 Empirical Predictions

I now compare the expressions for price changes, trading volumes, and price continuation

for front page and non-front page news, and derive empirical predictions for diﬀerential

market dynamics following diﬀerent article positions.

Before proceeding, I note the evolution of the share of informed investors, F

, in the cases

of front page and non-front page news. In the ﬁrst period, F

NF P

= F

F P

= 0. After that,

the share of informed investors following non-front page news evolves as follows:

NF P











γ for t = 1

γ + (1 − γ)ξ for t = 2

ξ + (1 − ξ)(γ + (1 − γ)ξ) for t = 3

(12)

Following front page news, meanwhile, the share of informed investors evolves as follows:

F P











γ for t = 1

γ + (1 − γ)γ for t = 2

ξ + (1 − ξ)(γ + (1 − γ)γ) for t = 3

(13)

Combining the shares of informed investors in (12)-(13) with the price changes in (8)

gives the immediate absolute price changes after non-front page and front page news:

|∆P

NF P

| =

γσ



+ γσ

|N − R|; |∆P

F P

| =

γσ



+ γσ

|N − R| (14)

Given that γ > γ, the ﬁrst-period absolute price change is larger following front page

news than following non-front page news.

Similarly, trading volumes at the news release in the ﬁrst period are given by:

T V

NF P

γ(1 − γ)

(σ



+ γσ

)

|N − R|; T V

F P

γ(1 − γ)

(σ



+ γσ

)

|N − R| (15)

The relationship between immediate trading volume around the news signal and the

percentage of immediately informed investors is non-monotonic. Trading volume is low if

either all or none of the investors see the news immediately, and trading volume is maximized

when the split between immediately attentive and inattentive investors is roughly even.

Recall that γ,  << 1/2, reﬂecting the empirical setting I consider, where the proportion

of the population who see any news article immediately (within the ﬁrst few minutes of

publication) is relatively low even for front page news. As a result, the split of attentive

versus inattentive investors is more equal and the immediate trading volume is higher when

the news is pinned to the front page.

Together, the price and volume expressions give the ﬁrst empirical prediction regarding

the immediate market response to front page and non-front page news.

Prediction 1 (Immediate Market Response) Front page news articles are followed by

larger trading volumes and absolute price moves immediately (within minutes) after the news.

How does the price response play out outside of the immediate window? To see this, I

turn to the continuation in the price path. I begin with the short-term continuation:

Cont

NF P

(∆P

, ∆P

) =

(1 − γ)ξ



+ [γ + (1 − γ)ξ]σ

(16)

Cont

F P

(∆P

, ∆P

) = (1 − γ) ×



+ [γ + (1 − γ)γ]σ

(17)

Note that from condition (2), the ﬁrst term of Cont

F P

(∆P

, ∆P

) is larger than the

ﬁrst term of Cont

NF P

(∆P

, ∆P

). The second term is larger in Cont

NF P

(∆P

, ∆P

), since

γ > γ > ξ. However, for suﬃciently low levels of immediate attention γ and γ, the former

eﬀect dominates. This results in the following empirical prediction.

Prediction 2 (Immediate Return Continuation) Front page news articles are accom-

panied by higher continuation in the short-term price changes.

While front page news articles are followed by a larger immediate reaction that continues

in the short-term, the longer term dynamics are quite diﬀerent. To see this, note that the

continuation in returns from the second to the third period for front page and non-front page

news is given by:

Cont

NF P

(∆P

, ∆P

) = (1 − ξ) ×



+ γσ



+ [ξ + (1 − ξ)(γ + (1 − γ)ξ)]σ

(18)

Cont

F P

(∆P

, ∆P

) =

(1 − γ)ξ



+ γσ



+ [ξ + (1 − ξ)(γ + (1 − γ)γ)]σ

(19)

Note that with ξ < γ < γ << 1/2, expressions (18)-(19) imply that the continuation from

the second period to the third is actually lower for front page news compared to non-front

page news. This yields the third empirical prediction of the gradual information diﬀusion

framework.

Prediction 3 (Delayed Return Continuation) Front page news articles induce lower

continuation in the long-term price changes.

In the next section, I test Predictions 1, 2, and 3 by observing the market dynamics

following front page and non-front page Bloomberg news articles in my hand-collected sam-

ple. For the immediate news release window, t = 1, I look at the 5-10 minutes following

publication of each individual news article. As the short-term subsequent window, t = 2, I

consider 30-45 minutes following the news, as the front page news articles tend to remain

prominently positioned for approximately half an hour to an hour. For the longer horizon,

t = 3, I consider windows of 60, 90, and 120 minutes following the news release.

4 News Positioning and Market Dynamics

Using the natural experiment in news positioning, in this section, I empirically estimate

the causal eﬀect of front page news positioning on ﬁnancial markets.

4.1 News Positioning and Short-Term Market Dynamics

I begin the analysis of diﬀerential activity following comparable front page and non-front

page news articles by observing the short-term trading volume surges and price dynamics

following the two types of SI news. Placing a piece of news on the front page is associ-

ated with substantially larger trading volumes and absolute price changes within minutes of

publication, as well as with higher continuation in the short-term price paths.

Consistent with Prediction 1, the more saliently positioned front page news articles induce

signiﬁcantly higher trading volumes. The median 15-second trading volume, computed as

the percentage of shares turned over during the ten minutes before and after SI news articles,

is displayed in Panel 1 of Figure 5. The median non-front page SI news article is accompanied

by virtually no increase in trading volume (plotted in light blue in the ﬁgure) relative to the

pre-news baseline. There is, however, a pronounced increase in the trading volumes following

SI news articles that appear on the front page (displayed in dark blue). The diﬀerence in

averages is even starker. Over the ten minutes after a news release, the average non-front

page SI news article is accompanied by a total of 0.05% turnover. The average ten-minute

trading volume after front page news is almost four-fold larger, at 0.19%. The diﬀerence is

statistically signiﬁcant at the 1% level, with a t-statistic of 4.52, as reported in Panel 1 of

Table 3. The estimated diﬀerence remains identical when controlling for month and hour

ﬁxed eﬀects, log market capitalization, and industry ﬁxed eﬀects.

Does the increased market activity reﬂected in trading volume correspond to larger price

changes? Panel 2 of Figure 5 presents the average absolute percentage price changes following

front page and non-front page SI news articles. The absolute price changes are calculated

separately for each ﬁrm over every ﬁve-second interval. The graph averages the price changes

in event time over the cross-section of ﬁrms. As a reference, the graph also plots, in dashed

lines, the baseline price changes computed over the same time period for the same securities

24 hour prior to the publication of the news articles. Conﬁrming the comparability of the two

sets of articles, the pre-news baselines are statistically indistinguishable for the two samples

of news articles. After publication, both front page news articles and non-front page news

articles are accompanied by larger absolute price changes than their respective baselines.

Two patterns emerge from a visual inspection of the absolute price changes. First, the

overall price change from the time of news publication to ten minutes later is much larger for

SI news articles that are positioned on the front page than for those that are not. Second,

corresponding to the more persistent attention garnered by the front page news articles being

saliently positioned for longer, price changes after these news articles are more persistent. I

consider these two eﬀects in greater detail below.

I begin the statistical analysis of price eﬀects by looking at the diﬀerential immediate

price reactions to front page and non-front page SI news articles. Lending further support to

Prediction 1, the average absolute price change within the ﬁrst ten minutes after front page

SI news articles is 60 basis points, compared to 21 basis points for non-front page SI news.

The diﬀerence of 39 basis points is statistically signiﬁcant at the 1% level, with a t-statistic

of 5.91, as can be seen from Panel 2 of Table 3. The result is robust to the inclusion of

controls: the estimated diﬀerence is 40 basis points when accounting for month and hour

ﬁxed eﬀects, and 36 basis points when also controlling for log market capitalization and

industry ﬁxed eﬀects. The results are similar at a shorter horizon of ﬁve minutes following

the news, with an average absolute price change of 42 basis points accompanying front page

news articles, compared to 16 basis points for non-front page news articles (t-statistic on

the diﬀerence is 6.19). The contrast is less stark, but still signiﬁcant when the window is

extended to one hour following the news. The average absolute price change over the hour

following front page SI news articles is 0.98%, whereas the average absolute price change

over the hour following non-front page SI news articles is 0.51% (t-statistic on the diﬀerence

is 5.01).

Having established empirical support for the ﬁrst prediction of my conceptual framework,

I now turn to Prediction 2. The theoretical prediction states that price paths following the

front page SI news articles should display more short-term continuation, reﬂecting the more

persistent attention garnered by news articles that stay at the top of the terminal screen for

longer. I test the extent to which front page positioning induces higher short-term return

continuation formally by estimating the following speciﬁcation:

Ret

s,i,[t+t

,t+t

]

= α + β

Ret

s,i,[t,t+t

]

+ β

F P

+ β

Ret

s,i,[t,t+t

]

× F P

(20)

+γX

i,t

+ 

s,i,[t+t

,t+t

]

where Ret

s,i,[t,t+t

]

denotes the return on security i during the immediate period [t, t + t

]

after publication of news article s, and Ret

s,i,[t+t

,t+t

]

is the return during the delayed period

[t+t

, t+ t

]. F P

is an indicator variable equal to one for SI news articles that are pinned to

the front page and zero for SI news articles not on the front page. The controls X

i,t

include

month and hour of day ﬁxed eﬀects, as well as log ﬁrm size and industry ﬁxed eﬀects. The

tests are run over the following time windows: (t

, t

) ∈ {(3 min, 5 min), (5 min, 10 min),

(5 min, 15 min), (5 min, 20 min), (10 min, 20 min), (10 min, 30 min)}.

Conﬁrming Prediction 2, front page news articles are followed by higher serial correlation

in returns at all considered short-term horizons, except for the shortest horizon of (t

, t

) =

(3 min, 5 min). The coeﬃcient of interest, β

, is positive and statistically signiﬁcant across

the other time speciﬁcations, as displayed in Table 4. For example, relative to non-front page

SI news articles, front page SI news articles induce 17% more continuation in returns from

the ﬁrst ﬁve minutes after publication to the next ﬁve minutes. This result is economically

sizable. For every 1% price move within the ﬁrst ﬁve minutes after a front page SI news

articles, there is an additional 17 basis points move in the same direction during the following

5 minutes, compared to non-front page SI news articles. The eﬀect is also precisely estimated,

with a t-statistic of 5.70 without controls, 5.67 with month and hour ﬁxed eﬀects, and 5.57

with the full set of controls including log ﬁrm size and industry ﬁxed eﬀects. Results over

other windows are qualitatively similar, with the coeﬃcient β

falling between 0.17 and 0.32,

depending on the considered time windows.

Interestingly, the non-front page SI news articles are actually followed by short-term re-

turn reversal from the ﬁrst ﬁve minutes to the next ﬁve to ten minutes, consistent with

the literature on short-term price reversals.

The coeﬃcient on Ret

s,i,[t,t+t

]

not inter-

acted with the front page indicator is negative and statistically signiﬁcant for (t

, t

) ∈

{(5 min, 10 min), (5 min, 15 min), (5 min, 10 min)}. Eﬀectively, these news articles, which

are prominently positioned at the top of the terminal screen only for short periods of time,

see the initial ﬁve-minute price reactions partially reverse within the following minutes. On

the other hand, front page SI news articles, which are prominently positioned for longer, are

followed by a strong price drift over the short term.

4.2 News Positioning and Longer-Term Price Dynamics

Placing a piece of news on the front page induces sizable short-term price eﬀects; do we

see the non-front page information eventually catch up? As the front page news articles get

removed from their prominent positions, the diﬀerences in diﬀusion of information contained

in these articles and the non-front page articles gradually diminish. The conceptual frame-

work predicts that at longer horizons, front page news articles should see less continuation

in returns. I ﬁnd evidence in support of this prediction: over longer horizons of one to two

hours after the news, non-front page information induces substantially more price drift than

front page news. The incorporation of non-front page information is much slower, however,

and full convergence does not occur for days after the news.

I begin evaluating longer-term price continuation by estimating speciﬁcation (20) over

the following windows: t

∈ {5 min, 10 min} and t

∈ {45 min, 60 min, 90 min}. The results

are reported in Panel 1 of Table 5.

The results reveal an interesting pattern of dynamics: the immediate returns over the

ﬁrst ﬁve minutes after news publication are more positively predictive of subsequent returns

following front page news than following non-front page news, up to approximately forty-

ﬁve minutes. But over longer horizons of sixty or ninety minutes, the eﬀect is no longer

present. Continuation in returns from the ﬁrst ﬁve minutes to the remainder of the ﬁrst

hour is statistically indistinguishable for front page versus non-front page SI news articles.

From the ﬁrst ﬁve minutes to the remainder of the ﬁrst hour and a half, there is slightly less

continuation following front page news (signiﬁcant at the 10% level).

Similarly, the initial ten-minute returns induced by front page news are followed by a

stronger drift for about forty-ﬁve minutes. During the ﬁrst forty-ﬁve minutes, front page

news articles induce an additional drift of approximately 31% of the initial ten-minute return.

See, for example, Atkins and Dyl (1990), Ederington and Lee (1995), Fung, Mok, and Lam (2000),

Chordia, Roll, and Subrahmanyam (2002), Zawadowski, Andor, and Kert´esz (2006), and Heston, Korajczyk,

and Sadka (2010).

Expanding the window to sixty minutes, the continuation becomes statistically indistinguish-

able between front page and non-front page news, and at ninety minutes there is weakly more

continuation for article that are not pinned to the front page.

As I shift the window even further, the results lend empirical support to Prediction 3.

Panel 2 of Table 5 reports estimates of speciﬁcation (20) over the following windows: t

∈

{30min, 45min} and t

∈ {90min, 120min}. The non-front page news articles are followed,

on average, by 25-27% continuation in returns from the ﬁrst half-hour to the remainder of

the 90-120 minutes. Front page news articles, however, see 14-19% less continuation. The

diﬀerences are highly statistically signiﬁcant. Similarly, the returns from the ﬁrst forty-ﬁve

minutes are substantially less likely to continue if the news article is pinned to the front

page. Non-front page news articles see a continuation of 19-22% from the ﬁrst 45 minutes

to the remainder of the ﬁrst 90-120 minutes. By contrast, front page news articles actually

see no return continuation over the same time windows.

Coupled with the results in Table 4, the longer-term price dynamics highlight the diﬀer-

ences in the speed of incorporation of front page and non-front page information. Pinning a

piece of news on the front page induces a stronger drift in returns up to forty-ﬁve minutes,

and the reactions to non-front page articles begin to catch up over the remainder of the

ﬁrst couple of hours after news publication. Theoretically, these patterns are fully consis-

tent with the gradual information diﬀusion framework outlined in Section 3. Practically,

the results indicate that for news articles consumed by sophisticated ﬁnance professionals

through a subscription-based platform such as Bloomberg, the market dynamics track the

discretionary positioning in real time.

While the price impact of front page information occurs quickly, it takes substantially

longer for non-front page information to be fully reﬂected in asset prices. Table 6 presents

the average diﬀerences in trading volumes and absolute price changes one, two, ﬁve, ten,

and ﬁfteen days after front page and non-front page SI news articles. The diﬀerences are

estimated controlling for month and hour ﬁxed eﬀects, log market capitalization, and indus-

try ﬁxed eﬀects. For the trading volume tests, I look at the total trading volume over a

10-minute window d days after news publication, where d ∈ {1, 2, 5, 10, 15}. Similarly, the

absolute price changes are calculated as the absolute percentage diﬀerence in price from the

time of news publication to exactly d days later.

The results indicate that some of the non-front page information is reﬂected in prices

within the ﬁrst few days, but a portion of the gap in market reactions induced by positioning

remains even days after the news. The elevated level of trading volume accompanying front

page news lasts for at most two days (the diﬀerence is statistically signiﬁcant at the 1% level

one day out, and at the 10% level two days out). The diﬀerence in absolute price changes one

day after the news is highly statistically signiﬁcant, but milder than the diﬀerence from just

the ﬁrst hour, tabulated in Table 3. This diﬀerence declines slightly but remains economically

similar two days after the news, signiﬁcant at the 5% level. Five days after news publication,

the price impact of front page positioning declines substantially and is no longer statistically

signiﬁcantly diﬀerent from the price impact of non-front page news, although the diﬀerence

remains economically visible at around 25 basis points. The gap is milder still, at 18 basis

points, ten days out after the news, and converges to a statistically indiscernible 8 basis

points ﬁfteen days out after the news.

These patterns are reinforced in a graphical evaluation of the directional price paths

following front page and non-front page news, displayed in Figure 6. I group the SI news

articles in my sample along two dimensions: (1) their position (front page versus non-front

page) and (2) the direction of the initial ﬁve-minute price move (positive versus negative). I

take the average cumulative price paths across news articles in each category, in event time

from the time of publication to various time windows. These price paths are plotted in solid

lines for front page news and in dashed lines for non-front page news. The price paths for

articles accompanied by positive initial ﬁve-minute price changes are shown in blue, while

the price paths for articles accompanied by negative initial price changes are shown in red.

In each case, the price change accompanying a given piece of news is computed relative to

the market return over the same time period, in order to screen out the directional equity

premia at longer horizons. Standard error bars are shaded in gray.

The ﬁgure shows a variety of time windows ranging from minutes to days after the news.

In the immediate term (0-10 minutes, displayed in the ﬁrst quarter of the ﬁgure), front page

news articles are accompanied by larger price changes, in both the positive and the negative

domains, consistent with the absolute price change results reported in Panel 2 of Table 3.

This gap widens for about 45 minutes, and then begins to narrow, as can be seen from the

price paths over the ﬁrst hour after news publication. The narrowing of the gap continues

for hours after the news, as front page articles see no additional price moves, while non-front

page information continues being incorporated into prices. In the last quarter of the ﬁgure,

I show the price responses from publication to one, two, ﬁve, ten, and ﬁfteen days out after

the news. Although the standard errors become very wide at these horizons, the economic

magnitudes show no diﬀerence in the long-term reactions to front page versus non-front

page news, due to the non-front page news articles gradually catching up to their front page

counterparts.

Overall, the eﬀect of diﬀerential news positioning is stark and quick, and takes a while to

converge. The gradual catching up of the reactions to non-front page information begins as

early as an hour after publication, but the diﬀusion of information in non-front page news

is quite slow. As a result, the eﬀect of news positioning can be statistically noticeable and

economically meaningful even several days after the news.

5 News Positioning versus News Importance

In this section, I compare the estimated eﬀects of news positioning against the eﬀects

of news importance, as marked by the editorial staﬀ. I estimate the relationship between

news importance and market dynamics by concentrating on news articles that are all equally

prominently positioned but that vary in importance – i.e., by comparing front page news

articles from the PI and SI categories. The diﬀerence in market reactions following these two

types of news is qualitatively diﬀerent from and quantitatively weaker than the diﬀerence

induced by front page positioning.

I limit my attention only to news articles that are pinned to the front page, so that there is

no variation in the prominence of the article positions. I include all front page news articles,

regardless of their importance markings, and estimate the diﬀerence in market reactions

following the more (“primary important”) and less important (“secondary important”) news

articles.

First, I note that the trading volumes immediately following front page PI news articles

are not statistically diﬀerent from the trading volumes following front page SI news articles.

As displayed in Panel 1 of Table 7, during the ﬁrst ﬁve minutes after a front page news article,

on average, an additional 0.09% of shares turn over when the article is from the PI category,

but this diﬀerence is not statistically signiﬁcant. Similarly, during the ﬁrst ten minutes, front

page PI news articles are followed by an additional 0.10% in trading volume compared to

front page SI news articles, signiﬁcant only at the 10% level. The pattern remains similar

over longer horizons, with an average of 0.18% additional shares turned over during the hour

following front page PI news articles, with the diﬀerence remaining statistically insigniﬁcant.

Second, while PI articles are accompanied by larger price impact than SI articles, the

eﬀect is less signiﬁcant and less persistent than the diﬀerence in absolute price changes

induced by front page positioning. As can be seen from Panel 2 of Table 7, in the ﬁrst

ﬁve minutes, front page PI news articles are followed by an additional 0.35% absolute price

change, an increase of 80% over the front page SI articles; the diﬀerence is signiﬁcant but

statistically weaker than the diﬀerence between front page and non-front page SI articles.

The diﬀerence in absolute price changes following front page PI news articles versus front

page SI news articles remains similar in magnitude and declines in statistical signiﬁcance as

the window is extended to ten and then sixty minutes. Overall, PI news articles are followed

by larger price reactions immediately in the ﬁrst ﬁve to ten minutes, but do not see further

diﬀerences from the front page SI news articles over longer horizons. This contrasts with the

diﬀerence between front page and non-front page SI news articles documented in Table 3,

which continues to grow over the hour following the news.

This result is corroborated by a comparison of the continuation in the price paths follow-

ing PI and front page SI news articles, which I estimate using the following speciﬁcation:

Ret

s,i,[t+t

,t+t

]

= α + β

Ret

s,i,[t,t+t

]

+ β

P I

+ β

Ret

s,i,[t,t+t

]

× P I

(21)

+γX

i,t

+ 

s,i,[t+t

,t+t

]

where Ret

s,i,[t,t+t

]

denotes the return on security i during the immediate period [t, t + t

]

after publication of news article s, and Ret

s,i,[t+t

,t+t

]

is the return during the delayed period

[t + t

, t + t

]. P I

is an indicator variable equal to one if the front page article comes from

the “primary important” category and zero if the article is from the “secondary important”

category. The controls X

i,t

include month and hour ﬁxed eﬀects, log ﬁrm size, and industry

ﬁxed eﬀects. The considered time windows are (t

, t

) ∈ {(5 min, 10 min), (5 min, 15 min)}.

Table 8 presents the results.

The estimated coeﬃcient on Ret

s,i,[t,t+t

]

× P I

indicates that front page PI news articles

are not accompanied by any more short-term price drift compared to front page SI news

articles. The diﬀerence is neither economically notable, nor statistically signiﬁcant. Over

the same time horizons, the diﬀerence in price drift following front page and non-front page

SI news articles is 17% and highly statistically signiﬁcant (see Table 4).

Recall that this analysis considers only news articles positioned on the front page, but of

both categories: “primary important” and “secondary important.” Whereas the results in

Section 4.2 keep article importance constant (only SI articles) and vary front page position-

ing, the analyses in this section keep the positioning constant but vary article importance.

As can be seen from a comparison of Tables 7-8 against Tables 3-4, diﬀerences in article

importance correspond to milder diﬀerences in market dynamics than diﬀerences in article

positioning. These ﬁndings suggest that article positioning is even more instrumental in

driving market reactions than diﬀerences in article importance, as marked by Bloomberg’s

journalistic and editorial staﬀ and conﬁrmed by the target audience of ﬁnance professionals

(see Section 6.3 below).

5.1 Discussion: Attention versus Inference

The comparison of diﬀerential reactions to news position and news importance helps

highlight the channel behind the market response to front page positioning. The drift patters

associated with prominent positioning, which are not observed for diﬀerential importance,

indicate that the positioning eﬀect is driven by attention patterns rather than inference

regarding the importance of the underlying news.

Eﬀectively, there are two mechanisms that could induce heightened market activity fol-

lowing front page news articles relative to non-front page articles. First is the attention

channel highlighted by the conceptual framework in Section 3: front page news articles re-

ceive more immediate attention, corresponding to higher trading volumes and absolute price

changes. The second mechanism is inference regarding the importance of the underlying

news: investors perceive the superior position to signal greater importance of the front page

articles.

While both channels produce increases in immediate trading volumes and price changes,

only the attention channel predicts the type of subsequent dynamics observed in the data.

As captured by Predictions 2 and 3, the attention channel predicts that front page articles

should be accompanied by more short-term drift and less continuation at longer horizons.

If instead the initial reactions are driven by inference regarding the articles’ importance,

there is no reason to observe a pattern of higher short-term drift and subsequent gradual

convergence.

The results on the diﬀerences between PI and SI front page news articles further support

the gradual information diﬀusion interpretation. The diﬀerences between reactions to articles

of actual varying importance are immediate, inducing no diﬀerential drifts. If the eﬀect of

positioning were driven primarily by the inference channel, then the timing of the positioning

eﬀect should be comparable to the eﬀect of importance. Instead, diﬀerential positioning

induces diﬀerences in incorporation of information that creates predictability in returns at a

variety of horizons. This corresponds more closely to the timing of the conceptual model of

gradual information diﬀusion.

Altogether, the eﬀects of news positioning are not only more substantial than the dif-

ferences between articles of varying editorial importance in the immediate term, but also

induce diﬀerences in return predictability further out. These results support the importance

of gradual information diﬀusion and highlight news consumption as playing a signiﬁcant role

in causally driving market dynamics around information releases.

6 Additional Analyses

I present additional analyses conﬁrming the exclusion restriction of my natural experi-

ment design: that the SI news articles that are pinned to the front page do not systematically

diﬀer from those that are not. First, I show that, holding position constant, the news arti-

cles published during quiet times (when more front page slots are available) do not generally

induce stronger market reactions than the articles published during busy times. Second, I

use machine learning techniques to show that the distributions of topics discussed in the

texts of the front page and non-front page SI news articles do not systematically diﬀer.

Lastly, I conduct a survey of active ﬁnance professionals and MBA students at top business

school programs to highlight that, in absence of the diﬀerential positioning, the target audi-

ence ﬁnds the two sets of news to be indistinguishable in terms of importance and expected

market impact.

6.1 Quiet Times vs. Busy Times

In this subsection, I address the potential concern that the diﬀerential reactions to front

page and non-front page SI news articles are driven by the fact that the former are released

during generally quieter times (when there are fewer PI news articles), rather than by dif-

ferent amounts of attention to the two types of articles. A few points are worth noting

here.

First, to the extent that increased market activity during quiet times reﬂects increased

attention dedicated to the security due to few other contemporaneous events, the results

would still capture the attention channel. In fact, various indicators of “quiet times” have

been used as indirect proxies for attention in prior work

The analysis in Section 4 captures

the variation in attention more precisely through the salience of news positioning.

Second, in my sample, without the diﬀerential positioning, SI news articles published

during quiet times, when little goes on in the markets, are likely to be accompanied, if

anything, by less market activity than the SI news articles published during the busier

times. This would push in the direction of ﬁnding less market activity after the front page

SI news articles (i.e., SI news articles released during quieter times), dampening my results.

I document this ﬁnding by comparing the non-front page SI news articles released during

relatively quiet times with non-front page SI news articles released during relatively busy

times. Thus, I hold news position (non-front page) and news importance (SI) constant, and

vary only the numbers of contemporaneous news releases.

To diﬀerentiate busy times from quiet times, I consider the contemporaneous volumes of

articles within three time intervals: (1) on the same day as a given non-front page SI article;

See DellaVigna and Pollet (2009) on earnings announcements released on Fridays, and Hirshleifer,

Lim, and Teoh (2009) on earnings announcements released contemporaneously with other announcements.

Accordingly, deHaan, Shevlin, and Thornock (2015) and Niessner (2015) provide evidence that ﬁrms strate-

gically respond to investors’ limited attention by timing their releases. In other contexts, distraction has

been shown to aﬀect liquidity provision (Corwin and Coughenour (2008)) and corporate actions (Kempf,

Manconi, and Spalt (2016))

(2) within ﬁve hours of a given non-front page SI news article; and (3) within two hours of

a given article. Non-front page SI news articles for which the contemporaneous volumes of

other news fall below the median form the “quiet times” sample. Non-front page SI news

articles with at or above-median contemporaneous volumes of other news form the “busy

times” sample.

As displayed in Table 9, holding editorial importance markings and position constant,

SI news articles that come out during quieter times are not accompanied by larger trading

volumes and absolute price changes than the SI news articles published during busier times.

If anything, price changes and trading volumes are smaller following non-front page SI news

articles published during quiet times. These patterns are qualitatively consistent across

deﬁnitions of quiet and busy times using the one day, ﬁve hour, and two hour windows.

Statistically, the diﬀerences in absolute price changes and trading volumes after non-front

page articles that come out during quiet and busy times are only discernible when the

volume of contemporaneous news is measured on a daily level. Economically, the diﬀerences

are small across the board, within a range of 1-3 basis points.

These results conﬁrm that the diﬀerential market reactions following front page and non-

front page SI news articles are not driven by the SI news articles that come out during quiet

times (and are therefore more likely to take an available front page position) carrying more

important content than the articles that come out during busy times.

6.2 Distributions of Topics

To rule out systematic diﬀerences in the content of front page and non-front page arti-

cles, I directly analyze the text of the news articles across diﬀerent positions and levels of

importance. The distribution of topics discussed in front page SI news articles is statistically

indistinguishable from the distribution of topics covered by non-front page SI articles. By

contrast, the distribution of topics discussed in PI news articles does diﬀer somewhat from

the SI news articles, with a larger focus on company operations and the healthcare industry,

and lower coverage of regulations and the ﬁnancial services industry.

Topic analysis provides an intuitive way to compare the content value of diﬀerent news

articles. The existing literature on the eﬀect of news on ﬁnancial markets considers textual

characteristics such as sentiment,

grammatical structure,

and complexity.

The method-

ology in this section contributes to the literature by proposing an intuitive approach to

See Tetlock (2007), Das and Chen (2007), Tetlock, Saar-Tsechansky, and Macskassy (2008), Loughran

and McDonald (2011), Bollen, Mao, and Zeng (2011), Garc´ıa (2013), and Uhl (2014).

See Engelberg (2008).

See Li (2008), You and Zhang (2008), Miller (2010), Lehavy, Li, and Merkley (2011), Loughran and

McDonald (2014), and Umar (2017).

identifying common topics in ﬁnancial news and representing the news articles in terms of

these prototypical areas of focus.

The topic analysis proceeds in two steps. First, I use a large corpus of news articles

from Reuters to analyze textual patterns in ﬁnancial news in general by representing the

articles in the space of meaningful features and identifying a set of broadly applicable topics.

Second, I apply the trained topic model to the news articles in the PI, front page SI, and

non-front page SI samples of hand-collected Bloomberg news articles.

For the ﬁrst step of the process, I use the Latent Dirichlet Allocation algorithm proposed

by Blei, Ng, and Jordan (2003), following similar methods employed in genetics (see, for

example, Pritchard, Stephens, and Donnelly (2000)). The Latent Dirichlet Allocation ap-

proach is particularly well suited to the problem at hand, because it represents all documents

as being generated from an underlying set of topics by a latent process. This admits mod-

eling out-of-sample documents as mixtures over the topics identiﬁed from the training data

– i.e., modeling the news articles from the various Bloomberg categories in terms of topics

identiﬁed from the larger sample of Reuters news. For a description of the Latent Dirichlet

Allocation methodology, please refer to Appendix A.1.

In order to train the topic model on a dataset that is similar yet distinct from the

Bloomberg news articles that I ultimately classify and evaluate, I use the Thomson Reuters

Text Research Collection 2 (TRC2), part of the Thomson Reuters Research Collection de-

scribed in Lewis, Yang, Rose, and Li (2004). This training corpus includes approximately

1.8M news articles spanning the full spectrum of ﬁnancial news reported by Reuters during

the period of 2008-2009, and is available from the National Institute of Standards and Tech-

nology. Appendix A.2 describes the pre-processing of the news articles in order to represent

them in terms of meaningful textual features ready to be inputted into the Latent Dirichlet

Allocation algorithm.

The output of the Latent Dirichlet Allocation model provides an intuitive conceptualiza-

tion of the identiﬁed topics in terms of the most frequently occurring words conditional on

each topic. I estimate the model for k = {10, 15, 20, 25, 30, 35, 40} topics, and observe that

the speciﬁcation with 15 topics performs best in terms of model log likelihood (see Appendix

A.3 for details on the topic model estimation). The topics identiﬁed in this speciﬁcation are

presented in Table 10. For each topic, the table displays the ﬁfteen terms in the vocabulary

that are most likely to appear conditional on that topic. For each of the topics, the set

of common terms forms a single coherent theme; for clarity of reference, each topic is la-

beled with a concise name capturing its theme. For example, the topic whose most common

terms are “court,” “case,” “judge,” “federal,” etc. is labeled “Litigation;” while the topic

whose most common terms are “deal,” “oﬀer,” “price,” “bid,” etc. is labeled “Mergers &

Acquisitions.”

The topics in Table 10 are listed in order of their estimated frequencies, which are pre-

sented in the last column. The most common topic to appear in the training corpus of

Reuters ﬁnancial news relates to technology, followed by ﬁnancial reports such as earnings,

and then news regarding ﬁnancial institutions such as hedge funds and banks. Other common

topics include automobile and air transport industries, litigation, and management. Over-

all, the identiﬁed topics are generally applicable and representative of concepts discussed in

ﬁnancial news.

For the second part of the process, I take advantage of the Latent Dirichlet Allocation’s

ability to represent out-of-sample documents as mixtures over the identiﬁed topics. I apply

this to characterize the distribution of topics in news articles from three categories: (1) PI

news articles; (2) SI news articles that appear on the front page; and (3) non-front page SI

news articles.

The results suggest that there are some distinct topic patterns for the select set of news

articles marked as especially important by the editorial staﬀ (PI news), but no diﬀerences

between the content of SI news articles that make it to the front page and those that do

not. The distribution of topics for each category of news is displayed in Figure 7. All three

categories of news overweight content regarding the ﬁnancial services industry, regulations,

the retail industry, and company employees. Coverage of technology, earnings reports, and

the healthcare industry is also common, although technology is far less ubiquitous than in

the training corpus. PI news articles are more likely to cover news related to the healthcare

industry and company operations; they have a lower focus on the ﬁnancial services industry

and regulations. Front page and non-front page SI news articles are very similar in terms of

the distribution of topics, with only minor diﬀerences (non-front page SI news articles are

more likely to feature news about M&A deals and company employees, whereas front page

articles contain more discussions of litigation and the ﬁnancial services industry).

For a formal comparison of the distributions of topics across the diﬀerent categories of

news, I perform a Pearson χ-square test of independence pairwise between any two categories

(see Rao and Scott (1981)). The results are tabulated in Table 11. In the main speciﬁcation

with 15 topics, the distribution of topics in PI news articles is weakly statistically signiﬁcantly

diﬀerent, at the 10% level, from the distribution of topics covered by the front page SI news

articles (see Panel 2 of Table 11). This is robust to varying the number of topics, with the

diﬀerence becoming signiﬁcant at the 5% level in the speciﬁcation with 25 topics but falling

short of the 10% statistical signiﬁcance threshold when the number of topics is reduced

to 10. More importantly, the front page and non-front page SI articles are statistically

indistinguishable in terms of their textual content, with a p-value of 87.76% in the primary

speciﬁcation with 15 topics. The similarity in the two distributions is robust to varying the

topic model speciﬁcation, with all p-values above 75%.

Overall, the results point to some distinction in the content of “primary important”

news articles from the content of “secondary important” articles. But the distributions of

topics are statistically indistinguishable across front page and non-front page SI news articles.

This supports the identifying assumption of independence of the prominence of the SI news

articles’ positions from their underlying content.

6.3 Market Participants’ Perceptions of News

In order to directly assess the market’s perceptions of the underlying news articles in my

hand-collected sample, I survey the target audience of the news: active ﬁnance professionals

and current MBA students at top business schools. Without the diﬀerential positioning,

these individuals do not perceive front page SI headlines to be any more impactful than

non-front page ones. They do, however, perceive the PI news articles to be more impactful,

supporting Bloomberg editorial staﬀ’s decisions to mark these articles as more important.

For this part of the analysis, I survey 150 active professionals from a number of ﬁnancial

institutions, as well as 25 current students at top MBA programs. The breakdown of these

individuals across aﬃliations is presented in Table 12. The majority of the sample (78.6%)

covers active professionals from a representative landscape of ﬁnancial institutions. The re-

mainder consists of current MBA students at Harvard Business School, the Wharton School,

Columbia Graduate School of Business, the University of Chicago Booth School of Business,

UVA Darden School of Business, and the McDonough School of Business at Georgetown

University.

The sample of active ﬁnance professionals is representative of the full landscape of the

ﬁnancial services industry. The bulk (81%) of the active professionals come from large

banks and broker dealers such as JP Morgan and Morgan Stanley, investment management

ﬁrms such as BlackRock and State Street, hedge funds such as Bridgewater Associates and

AQR Capital Management, and private equity ﬁrms such as the Blackstone Group and

Warburg Pincus. The remainder of the sample spans consulting ﬁrms such as the Boston

Consulting Group, government agencies such as the Federal Reserve Board, ﬁnancial oﬃces of

corporations such as Nike and Walt Disney, pension funds such as North Carolina Retirement

System, insurance companies such as Liberty Mutual, and other areas of the ﬁnancial services

industry.

The respondents largely constitute key decision makers within their respective ﬁrms.

Many of the professionals from larger corporations such as banks, broker dealers, and large

investment management ﬁrms are at the principal or managing director levels within their

organizations, including heads of regional oﬃces. The sample also includes chairmen, part-

ners, and C-level executives. This sample is broadly reﬂective of the client base consuming

Bloomberg news through the terminal. Approximately 87% of the professionals in my sam-

ple report having used a Bloomberg terminal at some point, with 63% actively using the

terminal on an ongoing basis.

In the survey, each respondent is asked to answer a series of twenty-ﬁve questions about

news headlines. The respondent is told that the headlines come from a news provider who

chooses how prominently the headlines are displayed based in part on the importance and

market impact of the underlying news. Each question presents two headlines, and asks the

respondent to specify which headline the respondent thinks had larger market impact and

deserves more prominence. A screenshot with an example question is displayed in Figure 8.

The survey questions span two sets of comparisons: (1) between front page SI news

articles and PI news articles; and (2) between front page SI news articles and non-front page

SI news articles. In particular, in each question, one of the two headlines (in random position

– either on the left or on the right of the screen) is randomly selected from the front page

SI news category. The other headline is randomly selected, with equal likelihoods, from the

categories of PI news (approximately 37.5% of the questions) and non-front page SI news

(approximately 62.5% of the questions).

The respondents are incentivized to identify the relative news importance as accurately

as they can. Each respondent receives a $10 gift (an Amazon.com gift card or a lunch

voucher to a venue of the respondent’s choice) for completing the survey. In addition, the

ﬁve respondents whose answers most closely match actual diﬀerences in positioning by the

news provider receive additional prizes of $90 each.

The results indicate that the ﬁnancial experts in the sample do not distinguish between

front page and non-front page SI news articles. Panel 1 of Table 13 presents the incidence

of front page SI news articles being chosen as more important than non-front page SI news

articles, with standard errors clustered by participant. The sample of 150 ﬁnance profession-

als identiﬁes the front page articles as more impactful 48.24% of the time, not statistically

diﬀerent from 50%. Similarly, the smaller sample of 26 MBA students choose the front page

news 45.05% of the time, which is, if anything, lower than 50% (marginally statistically

signiﬁcant at the 10% level). Pooling across both samples, front page articles are chosen as

more impactful 47.78% of the time. The results are very similar when I exclude attritors

(participants who do not answer all 25 questions). Thus, absent diﬀerential positioning, the

target audience of ﬁnance professionals does not perceive the front page SI news articles as

being any more important than their non-front page counterparts.

The market participants do, however, identify the “primary important” articles as more

impactful, validating Bloomberg’s importance markings. Active ﬁnance professionals choose

PI news articles over front page SI news articles 61.16% of the time, signiﬁcantly higher than

50% at the 1% level. MBA students are somewhat weaker at identifying “primary important”

news, choosing them 57.54% of the time, signiﬁcant at the 5% level. Pooling all responses, PI

stories are chosen as more impactful 60.58% of the time, signiﬁcantly higher than 50% at the

1% level. Overall, the results point to the Bloomberg editorial staﬀ correctly identifying, on

average, the news most relevant for the target demographic: the higher importance ranking

assigned to the PI news articles is corroborated by the surveyed market participants.

Similar patterns hold at the individual level. For each respondent, I calculate the per-

centage of times that the respondent chooses a front page SI news article over a non-front

page one, and the percentage of times that the respondent chooses a PI news article over

an SI one. A histogram of these individual-level percentages is displayed in Figure 9. The

incidence of choosing front page articles over non-front page ones is presented in blue; the

distribution is centered around 50%, is symmetric, and resembles a normal distribution.

Overall, this distribution is consistent with there being no distinction between the two sets

of articles, and the diﬀerences between individuals’ choices coming from noise and the vari-

ation in the randomly selected questions faced by diﬀerent individuals. The incidence of

choosing PI news articles over SI ones, presented in gray, paints a diﬀerent picture. Very

few respondents choose PI news articles less than 40% of the time, and the distribution is

centered around 60%, with a number of respondents choosing the PI news articles as often

as 90-100% of the time.

Overall, the target audience of the news perceives no systematic diﬀerences between the

SI news articles that get placed on the front page and those that do not. This is consistent

with the quasi-random positioning of these news articles. There is a stark juxtaposition

between the signiﬁcantly diﬀerent market dynamics following these two sets of news and

the market participants’ lack of distinction between them in the survey. This juxtaposition

highlights the extent to which salient news positioning can induce diﬀerent reactions to

otherwise identically important content.

7 Conclusion

This paper takes advantage of a natural experiment in news positioning to directly es-

timate the eﬀect of news consumption on ﬁnancial markets. For two news articles of equal

importance, pinning one to a prominent position induces 280% higher trading volume during

the ten-minute window after the news, 180% larger absolute price change, and substantially

higher short-term return continuation. Interestingly, diﬀerences in news positioning play an

even larger role for market dynamics than diﬀerences in the editorial markings of importance

of the underlying news articles’ content.

The results in this paper highlight the importance of how information is presented for the

way in which the information is incorporated into asset prices. In the modern informational

environment, where investors face millions of news articles per day, the distinction between

public and private information becomes somewhat blurred, and even public information may

not be immediately and eﬃciently priced.

My analysis traces out incorporation of infor-

mation in real time using a natural experiment on a highly relevant platform, the Bloomberg

terminal. My results capture momentum in price responses to information, and show that

the speed of incorporation depends on the method of dissemination. For more obscure or

private information, similar mechanisms are likely to apply at longer horizons, generating

phenomena such as month-level momentum.

References

[1] Ahern, K. R. and D. Sosyura (2014). “Who writes the news? Corporate press releases

during merger negotiations.” The Journal of Finance, 69.1: 241-291.

[2] Andrei, D., and J. Cujean (2017). “Information percolation, momentum and reversal.”

Journal of Financial Economics, 123.3: 617-645.

[3] Andrei, D. and M. Hasler (2014). “Investor attention and stock market volatility.” Review

of Financial Studies, 28.1: 33-72.

[4] Antweiler, W. and M. Z. Frank (2006). “Do US stock markets typically overreact to

corporate news stories?” Working paper, University of British Columbia.

[5] Atkins, A. B. and E. A. Dyl (1990). “Price reversals, bid-ask spreads, and market eﬃ-

ciency.” Journal of Financial and Quantitative Analysis, 25.4: 535-547.

[6] Bali, T.G., L. Peng, Y. Shen, and Y. Tang (2014). “Liquidity Shocks and Stock Market

Reactions.” Review of Financial Studies, 27.5: 1434-1485.

[7] Banerjee, S. and I. Kremer (2010). “Disagreement and learning: Dynamic patterns of

trade.” The Journal of Finance, 65.4: 1269-1302.

Dugast and Foucault (2017) propose a mechanism whereby increased data availability does not nec-

essarily lead to more eﬃcient prices in the long run. Farboodi, Matray, and Veldkamp (2017) document

declining price informativeness for ﬁrms outside of the S&P 500.

[8] Barber, B. M. and D. Loeﬄer (1993). “The ‘Dartboard’ column: Second-hand informa-

tion and price pressure.” Journal of Financial and Quantitative Analysis, 28.2: 273-284.

[9] Barber, B. M. and T. Odean (2008). “All that glitters: The eﬀect of attention and news

on the buying behavior of individual and institutional investors.” The Review of Financial

Studies, 21.2: 785-818.

[10] Ben-Rephael, A., Z. Da, and R. D. Israelsen (2017). “It Depends on where you search:

A comparison of institutional and retail attention.” Review of Financial Studies, forth-

coming.

[11] Bernard, V. L. and J. K. Thomas (1989). “Post-earnings-announcement drift: Delayed

price response or risk premium?” Journal of Accounting Research, 27.Supplement: 1-36.

[12] von Beschwitz, B., D.B. Keim, and K. Massa (2015). “First to ‘read’ the news: News

analytics and high frequency trading.” Working Paper.

[13] Blankespoor, E., E. deHaan, and C. Zhu (2017). “Capital market eﬀects of media syn-

thesis and dissemination: Evidence from robo-journalism.” Review of Accounting Studies,

forthcoming.

[14] Blei, D. M., A. Y. Ng, and M. I. Jordan (2003). “Latent dirichlet allocation.” Journal

of Machine Learning Research, 3.1: 993-1022.

[15] Bollen, J., H. Mao, and X. Zeng (2011). “Twitter mood predicts the stock market.”

Journal of Computational Science, 2.1: 1-8.

[16] Boulland, R., F. Degeorge, and E. Ginglinger (2017). “News dissemination and investor

attention.” Review of Finance, 21.2: 1-31.

[17] Busse, J. A. and T. C. Green (2012). “Market eﬃciency in real time.” Journal of Fi-

nancial Economics, 65.3: 415-437.

[18] Cao, H. H., J. D. Coval, and D. Hirshleifer (2002). “Sidelined investors, trading gener-

ated news, and security returns.” Review of Financial Studies, 15.2: 615-648.

[19] Cao, H. H. and H. Ou-Yang (2009). “Diﬀerences of opinion of public information and

speculative trading in stocks and options.” Review of Financial Studies, 22.1: 299-335.

[20] Carvalho, C., N. Klagge, and E. Moench (2011). “The persistent eﬀects of a false news

shock.” Journal of Empirical Finance, 18.4: 597-615.

[21] Chan, W. (2003). “Stock price reaction to news and no-news: drift and reversal after

head- lines.” Journal of Financial Economics, 70.2: 223-260.

[22] Chordia, T., R. Roll, and A. Subrahmanyam (2002). “Order imbalance, liquidity, and

market returns.” Journal of Financial Economics, 65.1: 111-130.

[23] Corwin, S. A. and J. F. Coughenour (2008). “Limited attention and the allocation of

eﬀort in securities trading.” The Journal of Finance, 63.6: 3031-3067.

[24] Curtis, A., V.J. Richardson, and R. Schmardebeck (2014). “Investor attention and the

pricing of earnings news.” Working paper, University of Arkansas.

[25] Da, Z., J. Engelberg, and P. Gao (2011). “In search of attention.” The Journal of

Finance, 66.5: 1461-1499.

[26] Da, Z., J. Engelberg, and P. Gao (2015). “The sum of all fears investor sentiment and

asset prices.” Review of Financial Studies, 28.1: 1-32.

[27] Das, S. R. and M. Y. Chen (2007). “Yahoo! for Amazon: Sentiment extraction from

small talk on the web.” Management Science, 53.9: 1375-1388.

[28] deHaan, E., T. Shevlin, and J. Thornock. (2015). “Market (in)attention and the strate-

gic scheduling and timing of earnings announcements.” Journal of Accounting and Eco-

nomics, 60.1: 36-55.

[29] DellaVigna, S. and J. M. Pollet (2009). “Investor inattention and Friday earnings an-

nouncements.” The Journal of Finance, 64.2: 709-749.

[30] Dougal, C., J. Engelberg, D. Garc´ıa, and C. A. Parsons (2012). “Journalists and the

stock market.” Review of Financial Studies, 25.3: 639-679.

[31] Drake, M. S., D. T. Roulstone, and J. R. Thornock (2012). “Investor information de-

mand: Evidence from google searches around earnings announcements investor informa-

tion demand.” Journal of Accounting Research, 50.4, 1001-1040.

[32] Drake, M. S., D. T. Roulstone, and J. R. Thornock (2015). “The determinants and con-

sequences of information acquisition via EDGAR.” Contemporary Accounting Research,

32.3: 1128-1161.

[33] Drake, M. S., D. T. Roulstone, and J. R. Thornock (2016). “The usefulness of historical

accounting reports.” Journal of Accounting and Economics, 61.2: 448-464.

[34] Dugast, J. and T. Foucault (2017). “Data abundance and asset price informativeness.”

Working Paper, HEC Paris.

[35] Ederington, L. H. and J. H. Lee (1995). “The short-run dynamics of the price adjustment

to new information.” Journal of Financial and Quantitative Analysis, 30.1: 117-134.

[36] Engelberg, J. E. (2008). “Costly information processing: Evidence from earnings an-

nouncements.” Working Paper, University of California San Diego.

[37] Engelberg, J. E. and C. A. Parsons (2011). “The causal impact of media in ﬁnancial

markets.” The Journal of Finance, 66.1: 67-97.

[38] Engelberg, J., C. Sasseville, and J. Williams (2012). “Market madness? The case of

mad money.” Management Science, 58.2: 351-364.

[39] Fang, L. and J. Peress (2009). “Media coverage and the cross-section of stock returns.”

The Journal of Finance, 64.5: 2023-2052.

[40] Farboodi, M., A. Matray, and L. Veldkamp (2017). “Where has all the big data gone?”

Working Paper, Princeton University.

[41] Fedyk, A. and J. Hodson (2015). “When can the market identify stale news?” Working

Paper, Harvard Business School.

[42] Feenberg, D., I. Ganguli, P. Gaule, and J. Gruber (2017). “It’s good to be ﬁrst: Order

bias in reading and citing NBER working papers.” Review of Economics and Statistics,

99.1: 32-39.

[43] Fehle, F., S. Tsyplakov, and V. Zdorovtsov (2005). “Can companies inﬂuence investor

behaviour through advertising? Super bowl commercials and stock returns.” European

Financial Management, 11.5: 625-647.

[44] Foster, G., C. Olsen, and T. Shevlin (1984). “Earnings releases, anomalies, and the

behavior of security returns.” Accounting Review, 59.4: 574-603.

[45] Foucault, T., J. Hombert, and I. Ro¸su (2016). “News trading and speed.” The Journal

of Finance, 71.1: 335-382.

[46] Fung, A. K. W., D. M. Y. Mok, and K. Lam (2000). “Intraday price reversals for index

futures in the US and Hong Kong.” Journal of Banking & Finance, 24.7: 1179-1201.

[47] Garc´ıa, D. (2013). “Sentiment during recessions.” The Journal of Finance, 68.3: 1267-

1300.

[48] Gilbert, T., S. Kogan, L. Lochstoer, and A. Ozyildirim (2012). “Investor inattention

and the market impact of summary statistics.” Management Science, 58.2: 336-350.

[49] Griﬃths, T. L. and M. Steyvers (2004). “Finding scientiﬁc topics.” Proceedings of the

National Academy of Sciences, 101.suppl1: 5228–5235.

[50] Harris, M. and A. Raviv (1993). “Diﬀerences of opinion make a horse race.” Review of

Financial Studies, 6.3: 473-506.

[51] Heston, S. L., R. A. Korajczyk, and R. Sadka (2010). “Intraday patterns in the cross-

section of stock returns.” The Journal of Finance, 65.4: 1369-1407.

[52] Hillert, A., H. Jacobs, and S. M¨uller (2014). “Media makes momentum.” Review of

Financial Studies, 27.12: 3467-3501.

[53] Hirshleifer, D., S. S. Lim, and S. H. Teoh (2009). “Driven to distraction: Extraneous

events and underreaction to earnings news.” The Journal of Finance, 64.5: 2289-2325.

[54] Hirshleifer, D., A. Subrahmanyam, and S. Titman (1994). “Security analysis and trading

patterns when some investors receive information before others.” Journal of Finance,

49.5: 1665-1698.

[55] Hirshleifer, D. and S. H. Teoh (2003). “Limited attention, information disclosure, and

ﬁnancial reporting.” Journal of Accounting and Economics, 36.1: 337-386.

[56] Ho, D. E., and K. Imai (2008). “Estimating causal eﬀects of ballot order from a ran-

domized natural experiment.” Public Opinion Quarterly, 72: 216?240.

[57] Hofmann, T. (1999). “Probabilistic latent semantic indexing.” Proceedings of the 22nd

Annual International ACM SIGIR Conference on Research and Development in Infor-

mation Retrieval: 50–57.

[58] Holden, C. W. and A. Subrahmanyam (1992). “Long-lived private information and

imperfect competition.” The Journal of Finance, 47.1: 247-270.

[59] Hong, H. and J. C. Stein (1999). “A uniﬁed theory of underreaction, momentum trading,

and overreaction in asset markets.” The Journal of Finance, 54.6: 2143-2184.

[60] Huang, X., A. Nekrasov, and S. H. Teoh (2017). “Headline salience and over-and un-

derreactions to earnings.” Working paper.

[61] Huberman, G. and T. Regev (2001). “Contagious speculation and a cure for cancer: A

nonevent that made stock prices soar.” The Journal of Finance, 56.1: 387-396.

[62] Jacobs, H., and A. Hillert (2015). “Alphabetic bias, investor recognition, and trading

behavior.” Review of Finance, 20.2: 693-723.

[63] Kandel, E. and N. D. Pearson (1995). “Diﬀerential interpretation of public signals and

trade in speculative markets.” Journal of Political Economy, 103.4: 831-872.

[64] Kaniel, R. and R. Parham (2017). “WSJ Category Kings?The impact of media attention

on consumer and mutual fund investment decisions.” Journal of Financial Economics,

123.2: 337-356.

[65] Kempf, E., A. Manconi, and O. Spalt (2016). “Distracted shareholders and corporate

actions.” The Review of Financial Studies, 30.5: 1660-1695.

[66] Klibanoﬀ, P., O. Lamont, and T. A. Wizman (1998). “Investor reaction to salient news

in closed-end country funds.” The Journal of Finance, 53.2: 673-699.

[67] Koudijs, P. (2016). “The boats that did not sail: Asset price volatility in a natural

experiment.” The Journal of Finance, 71.3: 1185-1226.

[68] Kyle, A. S. (1985). “Continuous auctions and insider trading.” Econometrica, 53.6:

1315-1335.

[69] Lawrence, A., J. Ryans, E. Sun, and N. Laptev (2017). “Earnings Announcement Pro-

motions: A Yahoo Finance Field Experiment.” Working paper, London Business School.

[70] Lehavy, R., F. Li, and K. Merkley (2011). “The eﬀect of annual report readability on

analyst following and the properties of their earnings forecasts.” The Accounting Review,

86.3: 1087-1115.

[71] Lewis, D. D., Y. Yang, T. G. Rose, and F. Li (2004). “RCV1: A new benchmark

collection for text categorization research.” Journal of Machine Learning Research, 5.4:

361-397.

[72] Li, F. (2008). “Annual report readability, current earnings, and earnings persistence.”

Journal of Accounting and economics, 45.2: 221-247.

[73] Liu, L., A. Sherman, and Y. Zhang (2014). “The long-run role of the media: Evidence

from initial public oﬀerings.” Management Science, 60.8: 1945-1964.

[74] Loh, R. K (2010). “Investor inattention and the underreaction to stock recommenda-

tions.” Financial Management, 39.3: 1223-1252.

[75] Loughran, T. and B. McDonald (2011). “When is a liability not a liability? Textual

analysis, dictionaries, and 10-Ks.” The Journal of Finance, 66.1: 35-65.

[76] Loughran, T. and B. McDonald (2014). “Measuring readability in ﬁnancial disclosures.”

The Journal of Finance, 69.4: 1643-1671.

[77] Marshall, B. R., N. Visaltanachoti, and G. Cooper (2014). “Sell the rumour, buy the

fact?” Accounting & Finance, 54.1: 237-249.

[78] Miller, B. P. (2010). “The eﬀects of reporting complexity on small and large investor

trading.” The Accounting Review, 85.6: 2107-2143.

[79] Niessner, M. (2015). “Strategic disclosure timing and insider trading.” Working paper,

Yale University.

[80] Peng, L. and W. Xiong (2006). “Investor attention, overconﬁdence and category learn-

ing.” Journal of Financial Economics 80.3: 563-602.

[81] Peress, J. (2008). “Media coverage and investors’ attention to earnings announcements.”

Working Paper, INSEAD.

[82] Peress, J. (2014). “The media and the diﬀusion of information in ﬁnancial markets:

Evidence from newspaper strikes.” The Journal of Finance, 69.5: 2007-2043.

[83] Pritchard, J. K., M. Stephens, and P. Donnelly (2000). “Inference of population struc-

ture using multilocus genotype data.” Genetics, 155.2: 945–959.

[84] Rao, J. and A. J. Scott (1981). “The analysis of categorical data from complex sam-

ple surveys: chi-squared tests for goodness of ﬁt and independence in two-way tables.”

Journal of the American Statistical Association, 76.374: 221-230.

[85] Rogers, J. L., D. J. Skinner, and S. L. C. Zechman (2013). “The role of the media in

disseminating insider-trading news.” Review of Accounting Studies, 21.3: 711-739.

[86] Schmidt, D. (2013). “Investors’ attention and stock covariation.” Working Paper, HEC

Paris.

[87] Solomon, D. H. (2012). “Selective publicity and stock prices.” The Journal of Finance,

67.2: 599-638.

[88] Tetlock, P. C. (2007). “Giving content to investor sentiment: The role of media in the

stock market.” The Journal of Finance, 62.3: 1139-1168.

[89] Tetlock, P. C. (2011). “All the news that’s ﬁt to reprint: Do investors react to stale

information?” Review of Financial Studies, 24.5: 1481-1512.

[90] Tetlock, P. C., M. Saar-Tsechansky, and S. Macskassy (2008). “More than words: Quan-

tifying language to measure ﬁrms’ fundamentals.” The Journal of Finance, 63.3: 1437-

1467.

[91] Uhl, M. W. (2014). “Reuters sentiment and stock returns.” Journal of Behavioral Fi-

nance, 15.4: 287-298.

[92] Umar, T. (2017). “Attention grabbers when seeking alpha.” Working Paper, University

of Chicago Booth.

[93] Wang, B. (2017). “Ranking and Salience.” Working Paper, Fordham University.

[94] Wang, J. (1994). “A model of competitive stock trading volume.” Journal of Political

Economy, 102.1: 127-168.

[95] Yuan, Y. (2015). “Market-wide attention, trading, and stock returns.” Journal of Fi-

nancial Economics, 116.3: 548-564.

[96] You, H. and X. Zhang (2009). “Financial reporting complexity and investor underreac-

tion to 10-K information.” Review of Accounting Studies, 14.4: 559-586.

[97] Zawadowski, A. G., G. Andor, and J. Kert´esz (2006). “Short-term market reaction after

extreme price changes of liquid stocks.” Quantitative Finance, 6.4: 283-295.

Figure 1: Bloomberg terminal screen displaying company news as of 3:01PM EST on December

15, 2016. The ﬁrst three lines, highlighted in yellow font, are the articles pinned to the “front

page.” All other articles scroll oﬀ the screen with the arrival of more recent news.

Journalist writes

News article

Journalist labels

importance

Editor decides on

ﬁnal importance

Secondary

important

Primary

important

All other

news articles

FRONT PAGE NEWS NON-FRONT PAGE NEWS

Is there currently

space on the front

page?

Yes

Figure 2: Process illustrating how Bloomberg news articles are pinned to the prominent front page

positions at the top of the news screen.

0 5 10 15 20 25 30 35 40

Number of "Primary Important" Articles

Number of "Secondary Important" Articles

Daily Numbers of "Primary Important" and "Secondary Important" Articles

Figure 3: Distribution of daily volumes of “primary important” (PI) and “secondary important”

(SI) news articles across the days in the sample.

Period 0 Period 1 Period 2 Period 3 Period T

Form priors

True prior

distribution:

R ∼ N (R, σ

)

News arrives

Signal

N = R + ,

 ∼ N (0, σ



)

Some investors observe N

and update beliefs

Payoﬀ R

realized

Investors

consume wealth

Non-front page news:

= 0

Hazard rate γ

of observing N

= γ

Hazard rate ξ

of observing N

= γ + (1 − γ)ξ

Hazard rate ξ

of observing N

= ξ + (1 − ξ)

×[γ + (1 − γ)ξ]

Front page news:

= 0

Hazard rate γ

of observing N

= γ

Hazard rate γ

of observing N

= γ + (1 − γ)γ

Hazard rate ξ

of observing N

= ξ + (1 − ξ)

×[γ + (1 − γ)γ]

Figure 4: Model timeline, illustrating the diﬀusion of information for front page and non-

front page news articles, as well as corresponding shares of informed investors F

in each

period t.

-10 min -8 min -6 min -4 min -2 min 0 min 2 min 4 min 6 min 8 min 10 min

Time since news publication

0.5

1.5

Trading volume (share turnover)

-6

Median Trading Volume

Non Front Page News

Front Page News

0 min 1 min 2 min 3 min 4 min 5 min 6 min 7 min 8 min 9 min 10 min

Time since news publication

0.25%

0.50%

Cumulative absolute return

Cumulative Absolute Price Change

Non Front Page News

Front Page News

Pre-News Baseline (FP)

Pre-News Baseline (NFP)

Figure 5: Market dynamics following front page and non-front page SI news articles. Panel 1

displays the median trading volume by 15-second intervals during the ten minutes before and

the ten minutes after news publication. The dark blue line plots trading volume around SI news

articles that are prominently placed on the front page; trading volume around non-front page SI

news articles is displayed in light blue. Panel 2 presents the average absolute percentage price

changes from publication to up to ten minutes later, for front page SI news articles (in dark blue)

and non-front page SI news articles (in light blue), as well as baseline price changes in absence of

news (in gray).

0.2%

0.4%

0.6%

-0.2%

-0.4%

-0.6%

2m 4m 6m 8m 10m 20m 30m 40m 50m

1h 2h 3h 4h 5h 6h 1d 2d 5d 10d 15d

Cumulative Signed Price Change

Time since news publication

Front page news, positive price move within the ﬁrst 5 min

Non-front page news, positive price move within the ﬁrst 5 min

Front page news, negative price move within the ﬁrst 5 min

Non-front page news, negative price move within the ﬁrst 5 min

Figure 6: Price paths after front page and non-front page SI news, sliced by the direction

of the initial 5-minute moves. The solid lines display price paths following front page news,

while the dashed lines indicate reactions to non-front page news. Blue lines correspond to

article-ticker pairs where the price moves in the positive direction within the ﬁrst ﬁve minutes

after news. Red lines represent article-ticker pairs with negative price moves within the ﬁrst

ﬁve minutes after news. Standard errors bars are shaded in gray.

Distributions of Topics in Different Categories of News

Technology

Earnings

Financials

Automobile

Air Transport

Litigation

Management

Healthcare

Operations

Strategy

M&A

Advertising

Regulations

Retail

Employees

10%

12%

14%

16%

Topic Incidence

Primary Important

Front Page SI

Non-Front Page SI

Figure 7: Distributions of topics across news articles from diﬀerent categories. The ﬁgure presents

the frequency of the identiﬁed topics in three types of news articles: PI news, SI news pinned to

the front page, and non-front page SI news.

Figure 8: Example question from the survey administered to active ﬁnance professionals and MBA

students. One of the presented headlines comes from the front page SI sample, while the other is

either a non-front page SI headline or a PI headline.

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

% Times Chose First Type of Article (Front Page over not; PI over SI)

12%

18%

24%

30%

Share of Respondents

Individual-Level Response Proportions

Choice: Front Page SI over non-Front Page SI

Choice: PI over Front Page SI

Figure 9: Individual-level responses from the survey of ﬁnancial experts. For each respondent, I

calculate the percentage of times that respondent chooses a front page SI headline over a non-front

page one (plotted in blue), and the percentage of times the respondent chooses a PI headline over

a front page SI one (in gray). The ﬁgure tabulates the distribution of these percentages across

respondents.

Table 1: Examples of articles in the “primary important” and “secondary important” categories.

Panel 1 presents randomly selected headlines in the “primary important” sample, all pinned to

the front page. Panel 2 lists randomly selected headlines in the “secondary important” sample,

indicating whether each of the example articles was pinned to the front page.

Panel 1: “Primary important” news articles

Date Headline

3/25/2014 Walgreen 2Q Adj. EPS Misses Est.

4/7/2014 Tekmira Says FDA Modiﬁes TKM-Ebola Drug to Partial Hold

4/25/2014 United Technologies Reports SEC Formal Investigation, Subpoena

8/14/2014 Icahn Reports 6.63% Stake in Gannett, Urges Splitting Co.

10/28/2014 Amgen Restarts Buyback, Boosting Dividend; 2015 View Tops Ests.

12/23/2014 Stryker Said to Plan Smith & Nephew Takeover Bid Within Weeks

1/27/2015 Amgen 4Q Adj. EPS, Rev. Top Ests.; Ivabradine, T-vec Delayed

5/13/2015 Nissan Forecasts 6% Gain in Proﬁt on U.S. Demand, Weak Yen

5/19/2015 Computer Sciences Corp. to Split Into Two Companies

7/30/2015 Sanoﬁ Proﬁt Beats Estimates as Multiple Sclerosis Drugs Gain

9/10/2015 Morrison Earnings Miss Analysts’ Estimate as Grocer Cut Prices

9/14/2015 Standard Chartered Said to Plan Cutting 250 Managing Directors

11/24/2015 Fed Says It’s Overhauling Standards for Large-Bank Examiners

1/15/2016 Wal-Mart to Close 269 Stores in U.S., Globally

2/25/2016 Apple Says U.S. Can’t Force It to Unlock Terrorist’s IPhone

Panel 2: “Secondary important” news articles

Date Headline FP

4/7/2014 Honeywell CEO Makes Biggest Executive Shift Naming Vice Chairmen N

5/19/2014 AstraZeneca Chairman “Surprised” Pﬁzer Took Last Oﬀer Public N

6/3/2014 Robertson’s Stock Picker Singh Said to Become Newest Tiger Cub Y

6/24/2014 Morgan Stanley Gets 90,000 Applications for Summer Program N

7/10/2014 TRW Said to Receive Takeover Approach From ZF Friedrichshafen Y

8/1/2014 Judge Grants Preliminary Approval to Apple e-Book Settlement N

9/26/2014 Pimco Said to Have Discussed Firing Gross Before Exit to Janus Y

12/5/2014 CNN’s Candy Crowley to Leave Cable News Network After 27 Years N

1/20/2015 FXCM Plunges as Bailout Lets Leucadia Force Sale of Brokerage Y

3/12/2015 Viacom Says Chairman Redstone Will Miss Monday’s Annual Meeting N

4/28/2015 McDonald’s Axes Seven Sandwiches in Push to Get Its Menu Right Y

6/3/2015 Pandora Internet Radio Wins U.S. Nod to Buy South Dakota Station N

6/11/2015 Biotech Led by 29-Year-Old CEO Now Worth Billions With No Sales Y

7/29/2015 High-Density Drone Flights Possible Within Decade, Google Says N

9/21/2015 Clinton’s Tweet on High Drug Prices Sends Biotech Stocks Down Y

10/22/2015 Amazon Sales Top Estimates on Prime Day Event, Cloud Computing Y

12/21/2015 Chipotle Probed for New Outbreak of Diﬀerent E. Coli Strain Y

1/14/2015 Apple, Ericsson Sue Each Other Over Phone Patent Royalties N

2/27/2016 Lenovo to Purge Adware From New PCs After Superﬁsh Controversy N

Table 2: Summary statistics of the hand-collected news sample. Panel 1 presents the breakdown

by month of publication of PI news articles, SI news articles, and SI news articles that are positioned

on the front page. Panel 2 presents the breakdown by hour of publication, and includes the

percentage of SI articles that are positioned on the front page. The sample is restricted to the

articles published between 8AM and 5PM EST and tagged with at least one U.S. equity security.

Panel 1: News Articles By Month

Hour of Day PI articles SI articles FP SI articles

January 106 346 78

February 125 305 63

March 208 461 85

April 284 891 123

May 222 830 104

June 232 776 90

July 245 1,009 132

August 152 640 97

September 238 495 134

October 239 757 125

November 157 854 104

December 154 869 139

Total 2,362 8,233 1,274

Panel 2: News Articles By Hour

Hour of Day PI articles SI articles FP SI articles % SI articles on FP

8AM - 9AM 370 745 134 18%

9AM - 10AM 285 1,054 135 13%

10AM - 11AM 189 1,090 174 16%

11AM - 12PM 173 942 155 16%

12PM - 1PM 147 935 142 15%

1PM - 2PM 171 896 147 16%

2PM - 3PM 213 819 158 19%

3PM - 4PM 147 808 134 17%

4PM - 5PM 667 944 95 10%

Total 2,362 8,233 1,274 15%

Table 3: Comparison of trading volumes and absolute price changes immediately following SI

news articles that are pinned to the front page and those that are not. Panel 1 looks at trading

volumes within ﬁve, ten, and sixty minutes of publication, while Panel 2 considers the absolute

price changes over the same time periods.

Panel 1: Trading Volume

Front Page SI News Non-Front Page SI News Diﬀerence (FP − NFP)

First 5 min 0.10% 0.02% 0.07%**

Standard Error (0.012%) (0.001%) (0.013%)

# Observations 847 4,095 –

First 10 min 0.19% 0.05% 0.14%**

Standard Error (0.030%) (0.002%) (0.031%)

# Observations 858 4,233 –

First 1 hour 0.58% 0.26% 0.32%**

Standard Error (0.143%) (0.012%) (0.143%)

# Observations 897 4,459 –

Panel 2: Absolute Price Changes

Front Page SI News Non-Front Page SI News Diﬀerence (FP − NFP)

First 5 min 0.42% 0.16% 0.26%**

Standard Error (0.041%) (0.006%) (0.042%)

# Observations 847 4,095 –

First 10 min 0.60% 0.21% 0.39%**

Standard Error (0.065%) (0.006%) (0.066%)

# Observations 858 4,233 –

First 1 hour 0.98% 0.51% 0.47%**

Standard Error (0.091%) (0.020%) (0.094%)

# Observations 897 4,459 –

** denotes signiﬁcance at the 1% level.

Table 4: Short-term continuation in returns after front page and non-front page SI news articles.

Each column runs the following speciﬁcation:

Ret

s,i,[t+t

,t+t

]

= α + β

Ret

s,i,[t,t+t

]

+ β

F P

+ β

Ret

s,i,[t,t+t

]

× F P

+ γX

i,t

+ 

s,i,[t+t

,t+t

]

where F P

is a dummy variable equal to 1 if the news article s is pinned to the front

page; Ret

s,i,[t,t+t

]

denotes the return on security i during the immediate period [t, t + t

] af-

ter publication of news article s, and Ret

s,i,[t+t

,t+t

]

is the return during the delayed period

[t + t

, t + t

]. The main coeﬃcient of interest is β

on the interaction term Ret

s,i,[t,t+t

]

F P

(highlighted in blue). The tests are run over the following time windows: (t

, t

) ∈

{(5min, 10min), (3min, 5min), (5min, 15min), (5min, 20min), (10min, 30min), (10min, 30min)}.

Columns marked with (1) do not include any controls. Columns marked with (2) include hour and

month ﬁxed eﬀects. Columns marked with (3) also control for log ﬁrm size and industry ﬁxed

eﬀects.

= 5 min, t

= 10 min t

= 3 min, t

= 5 min

(1) (2) (3) (1) (2) (3)

Ret

s,i,[t,t+5 min]

-0.076** -0.077** -0.079** 0.040* 0.035† 0.034†

Standard error (0.025) (0.028) (0.028) (0.018) (0.018) (0.018)

Ret

s,i,[t+5 min]

× F P

0.171** 0.170** 0.167** 0.033 0.041† 0.038†

Standard error (0.030) (0.030) (0.030) (0.022) (0.022) (0.022)

c c c c c c

# FP SI articles 859 859 858 848 848 847

# Non-FP SI articles 4,235 4,235 4,233 4,097 4,097 4,095

= 5 min, t

= 15 min t

= 5 min, t

= 20 min

(1) (2) (3) (1) (2) (3)

Ret

s,i,[t,t+5 min]

-0.120** -0.121** -0.117** -0.170** -0.168** -0.168**

Standard error (0.025) (0.025) (0.025) (0.031) (0.031) (0.031)

Ret

s,i,[t+5 min]

× F P

0.261** 0.258** 0.254** 0.313** 0.313** 0.317**

Standard error (0.031) (0.031) (0.031) (0.037) (0.038) (0.038)

c c c c c c

# FP SI articles 864 864 863 871 871 869

# Non-FP SI articles 4,267 4,267 4,265 4,273 4,273 4,271

= 10 min, t

= 20 min t

= 10 min, t

= 30 min

(1) (2) (3) (1) (2) (3)

Ret

s,i,[t,t+5 min]

0.043† 0.044† 0.046† -0.011 -0.010 -0.014

Standard error (0.026) (0.026) (0.027) (0.034) (0.034) (0.034)

Ret

s,i,[t+5 min]

× F P

0.173** 0.174** 0.170** 0.208** 0.208** 0.212**

Standard error (0.028) (0.028) (0.029) (0.037) (0.038) (0.038)

c c c c c c

# FP SI articles 871 871 869 892 892 890

# Non-FP SI articles 4,273 4,273 4,271 4,310 4,310 4,308

**, *, and † denote signiﬁcance at the 1%, 5%, and 10% levels, respectively.

Table 5: Continuation in returns over longer horizons following front page and non-front page SI

news articles. Each column runs the following speciﬁcation:

Ret

s,i,[t+t

,t+t

]

= α + β

Ret

s,i,[t,t+t

]

+ β

F P

+ β

Ret

s,i,[t,t+t

]

× F P

+ γX

i,t

+ 

s,i,[t+t

,t+t

]

where F P

is a dummy variable equal to 1 if the news article s is positioned on the front page;

Ret

s,i,[t,t+t

]

denotes the return on security i during the immediate period [t, t+t

] after publication

of news article s, and Ret

s,i,[t+t

,t+t

]

is the return during the delayed period [t + t

, t + t

]. The

vector of controls X

i,t

is empty in columns marked with (1), consists of month and hour ﬁxed eﬀects

in columns marked with (2), and also includes log ﬁrm size and industry ﬁxed eﬀects in columns

marked with (3). The main coeﬃcient of interest is β

on the interaction term Ret

s,i,[t,t+t

]

× F P

(highlighted in blue).

Panel 1 looks at the following time windows: t

∈ {5 min, 10 min}, t

∈ {45 min, 60 min, 90 min}.

Panel 2 considers the following time windows: t

∈ {30 min, 45 min}, t

∈ {90 min, 120 min}.

Panel 1: Return continuation from the ﬁrst 5-10 minutes up to 45-90 minutes

= 5 min, t

= 45 min t

= 10 min, t

= 45 min

(1) (2) (3) (1) (2) (3)

Ret

s,i,[t,t+5 min]

0.084** 0.083* 0.080* 0.121** 0.120** 0.123**

Standard error (0.033) (0.033) (0.033) (0.040) (0.040) (0.040)

Ret

s,i,[t+5 min]

× F P

0.338** 0.342** 0.344** 0.315** 0.314** 0.317**

Standard error (0.058) (0.058) (0.059) (0.048) (0.048) (0.048)

c c c c c c

# FP SI articles 894 894 892 894 894 892

# Non-FP SI articles 4,421 4,421 4,418 4,421 4,421 4,418

= 5 min, t

= 60 min t

= 10 min, t

= 60 min

(1) (2) (3) (1) (2) (3)

Ret

s,i,[t,t+5 min]

0.284** 0.283** 0.281** 0.154** 0.155** 0.157**

Standard error (0.048) (0.048) (0.049) (0.050) (0.050) (0.051)

Ret

s,i,[t+5 min]

× F P

-0.123 -0.122 -0.126 0.081 0.083 0.080

Standard error (0.079) (0.081) (0.081) (0.059) (0.060) (0.060)

c c c c c c

# FP SI articles 899 899 897 899 899 897

# Non-FP SI articles 4,462 4,462 4,459 4,462 4,462 4,459

= 5 min, t

= 90 min t

= 10 min, t

= 90 min

(1) (2) (3) (1) (2) (3)

Ret

s,i,[t,t+5 min]

0.225** 0.228** 0.229** 0.124* 0.120* 0.122*

Standard error (0.048) (0.049) (0.050) (0.053) (0.053) (0.054)

Ret

s,i,[t+5 min]

× F P

-0.115† -0.113* -0.109† -0.103† -0.106† -0.107†

Standard error (0.071) (0.071) (0.071) (0.060) (0.060) (0.061)

c c c c c c

# FP SI articles 901 901 899 901 901 899

# Non-FP SI articles 4,475 4,475 4,472 4,475 4,475 4,472

**, *, and † denote signiﬁcance at the 1%, 5%, and 10% levels, respectively.

Panel 2: Return continuation from the ﬁrst 30-45 minutes up to 90-120 minutes

= 30 min, t

= 90 min t

= 45 min, t

= 90 min

(1) (2) (3) (1) (2) (3)

Ret

s,i,[t,t+5 min]

0.254** 0.248** 0.249** 0.191** 0.187** 0.186**

Standard error (0.029) (0.029) (0.030) (0.016) (0.016) (0.016)

Ret

s,i,[t+5 min]

× F P

-0.143** -0.142** -0.145** -0.215** -0.215** -0.214**

Standard error (0.032) (0.032) (0.032) (0.020) (0.020) (0.020)

c c c c c c

# FP SI articles 901 901 899 901 901 899

# Non-FP SI articles 4,475 4,475 4,472 4,475 4,475 4,472

= 30 min, t

= 120 min t

= 45 min, t

= 120 min

(1) (2) (3) (1) (2) (3)

Ret

s,i,[t,t+5 min]

0.266** 0.267** 0.255** 0.226** 0.224** 0.221**

Standard error (0.035) (0.035) (0.036) (0.017) (0.017) (0.017)

Ret

s,i,[t+5 min]

× F P

-0.185** -0.183** -0.188** -0.264** -0.268** -0.273**

Standard error (0.032) (0.032) (0.033) (0.021) (0.021) (0.021)

c c c c c c

# FP SI articles 903 903 901 903 903 901

# Non-FP SI articles 4,491 4,491 4,488 4,491 4,491 4,488

**, *, and † denote signiﬁcance at the 1%, 5%, and 10% levels, respectively.

Table 6: Diﬀerences in trading volumes and absolute price changes following front page and non-

front page SI news articles over longer horizons. The trading volumes are measured over 10-minute

intervals d days after the news. The absolutely price changes are computed from news publication

to d days after the news.

Diﬀerence in:

Number of Days after News Trading Volume Absolute Price Change

d = 1 0.02%** 0.38%**

Standard Error (0.01%) (0.14%)

# Obs – SI FP 892 892

# Obs – SI NFP 4,432 4,432

d = 2 0.03%† 0.34%†

Standard Error (0.02%) (0.19%)

# Obs – SI FP 888 888

# Obs – SI NFP 4,415 4,415

d = 5 -0.01% 0.25%

Standard Error (0.02%) (0.20%)

# Obs – SI FP 890 890

# Obs – SI NFP 4,422 4,422

d = 10 -0.01% 0.18%

Standard Error (0.03%) (0.22%)

# Obs – SI FP 878 878

# Obs – SI NFP 4,403 4,403

d = 15 0.01% 0.08%

Standard Error (0.03%) (0.25%)

# Obs – SI FP 885 855

# Obs – SI NFP 4,411 4,411

** and * denote signiﬁcance at the 1% and 5% levels, respectively.

Table 7: Trading volumes and absolute price changes following PI and front page SI news articles.

Panel 1 compares trading volumes over ﬁve-, ten-, and sixty-minute windows following PI news

articles against trading volumes following front page SI news articles.

Panel 2 compares absolute price changes over ﬁve-, ten-, and sixty-minute windows following PI

news articles against absolute price changes following front page SI news articles.

Panel 1: Trading Volume

Front Page SI News PI News Diﬀerence (PI−SI)

First 5 min 0.10% 0.18% 0.09%

Standard Error (0.02%) (0.04%) (0.04%)

# Observations 847 1,291 –

First 10 min 0.19% 0.29% 0.10%†

Standard Error (0.02%) (0.04%) (0.05%)

# Observations 858 1,306 –

First 60 min 0.57% 0.74% 0.18%

Standard Error (0.08%) (0.10%) (0.13%)

# Observations 897 1,349 –

Panel 2: Absolute Price Changes

Front Page SI News PI News Diﬀerence (PI−SI)

First 5 min 0.44% 0.79% 0.35%**

Standard Error (0.04%) (0.05%) (0.05%)

# Observations 847 1,291 –

First 10 min 0.61% 1.01% 0.40%**

Standard Error (0.07%) (0.06%) (0.09%)

# Observations 858 1,306 –

First 60 min 0.99% 1.40% 0.41%**

Standard Error (0.09%) (0.07%) (0.11%)

# Observations 897 1,349 –

** and * denote signiﬁcance at the 1% and 5% levels, respectively.

Table 8: Serial correlation in returns following PI and SI front page news articles. Each column

estimates the following speciﬁcation:

Ret

s,i,[t+t

,t+t

]

= α + β

Ret

s,i,[t,t+t

]

+ β

P I

+ β

Ret

s,i,[t,t+t

]

× P I

+ γX

i,t

+ 

s,i,[t+t

,t+t

]

where P I

is a dummy variable equal to 1 if the news article s is marked as “primary important”;

Ret

s,i,[t,t+t

]

denotes the return on security i during the immediate period [t, t+t

] after publication

of news article s, and Ret

s,i,[t+t

,t+t

]

is the return during the delayed period [t+t

, t+t

]. The main

coeﬃcient of interest is β

on the interaction term Ret

s,i,[t,t+t

]

× P I

(highlighted in blue). The

vector of controls is empty in columns marked with (1), consists of month and hour ﬁxed eﬀects

in columns marked with (2), and also includes log ﬁrm size and industry ﬁxed eﬀects in columns

marked with (3). The considered time intervals are (t

, t

) ∈ {(5 min, 10 min), (5 min, 15 min)}.

= 5 min, t

= 10 min t

= 5 min, t

= 15 min

(1) (2) (3) (1) (2) (3)

Ret

s,i,[t,t+5 min]

0.093* 0.091* 0.089* 0.137** 0.134** 0.136**

Standard error (0.037) (0.037) (0.037) (0.039) (0.039) (0.039)

Ret

s,i,[t+5 min]

× P I

-0.016 -0.020 -0.027 -0.025 -0.020 -0.026

Standard error (0.043) (0.044) (0.044) (0.045) (0.045) (0.045)

c c c c c c

# PI articles 1,294 1,294 1,291 1,310 1,310 1,306

# FP SI articles

859 859 858 864 864 863

**, *, and † denote signiﬁcance at the 1%, 5%, and 10% levels, respectively.

Table 9: Comparison of trading volumes and absolute price changes within ten minutes of non-front

page SI articles published during quiet versus busy times. Quiet times are deﬁned as below-median

volume of news measured over the following windows around a given article: one day, ﬁve hours,

and two hours. Panel 1 considers trading volumes, while Panel 2 looks at absolute price changes.

Panel 1: 10-Minute Trading Volume

News in Quiet Times News in Busy Times Diﬀerence

Window: 1 day 0.04% 0.06% -0.02%*

Standard Error (0.003%) (0.006%) (0.007%)

# Observations 3,383 3,576 –

Window: 5 hours 0.05% 0.06% -0.01%

Standard Error (0.002%) (0.006%) (0.008%)

# Observations 2,011 4,948 –

Window: 2 hours 0.05% 0.06% -0.01%

Standard Error (0.003%) (0.006%) (0.009%)

# Observations 3,224 3,735 –

Panel 2: 10-Minute Absolute Price Changes

News in Quiet Times News in Busy Times Diﬀerence

Window: 1 day 0.20% 0.24% -0.03%*

Standard Error (0.01%) (0.01%) (0.02%)

# Observations 3,383 3,576 –

Window: 5 hours 0.21% 0.22% -0.01%

Standard Error (0.01%) (0.01%) (0.01%)

# Observations 2,011 4,948 –

Window: day 0.21% 0.21% -0.01%

Standard Error (0.01%) (0.01%) (0.01%)

# Observations 3,224 3,735 –

* denotes signiﬁcance at the 5% levels.

Table 10: This table lists and labels the topics identiﬁed from a large and representative dataset

of ﬁnancial news from Reuters. The topics are listed in order from most to least common in the

corpus. For each topic code, I provide the ﬁfteen terms in the vocabulary that are most likely to

appear conditional on that topic, as well as the frequency of that topic in the corpus.

Topic Label Most Common Terms Frequency

#1 Technology data, technology, companies, security, information, 23.83%

comment, including, according, government, card,

software, credit, declined, store, did

#2 Earnings & percent, year, sales, quarter, million, analysts, share, rose, 9.84%

Performance revenue, estimates, proﬁt, earnings, fell, cents, average

#3 Financial million, year, bank, today, oﬃcer, ﬁnancial, chief, 8.29%

Services according, statement, executive, ﬁrm, largest, new york,

investment, unit

#4 Automobile vehicles, cars, downturn, sales, automaker, deliveries, 6.48%

turnover, air, current, safety, in-house, auto, backlog,

switches, parts

#5 Air transport internet, service, search, aircraft, today, ﬂight, plane, 6.22%

contract, engine, carrier, air, airline, satellite, web, traﬃc

#6 Litigation court, case, judge, federal, workers, law, claims, million, 5.96%

ﬁled, trial, state, lawsuit, ruling, lawyers, attorney

#7 Management ceo, president, job, board, women, chairman, director, vice, 5.70%

named, executive, world, role, according, chief, leave

#8 Healthcare drug, patients, care, percent, ﬂu, health, treatment, disease 4.92%

immunize, today, study, research, treatments, medical,

medicines

#9 Operations according, years, got, little, long, later, industry, great, 4.69%

left, good, costs, international, commercial, saying, end

#10 Business & year, percent, executive, chief, market, oﬃcer, brand, today, 4.63%

Strategy products, global, world, plans, month, second, sales

#11 Mergers & deal, oﬀer, price, people, bid, shares, comment, buy, 4.40%

Acquisitions companies, analyst, takeover, shareholders, matter,

investors, call

#12 Advertising tv, like, food, video, according, subscribers, products, 4.15%

review, pay, media, content, digital, cable, website,

advertising

#13 Regulations oﬀer, regulator, today, agency, government, information, 3.66%

review, adjudicate, statement, public, rules, letter, asked,

questions, mailed

#14 Retail stores, chain, retailer, sales, retail, years, home, online, 3.62%

customers, holiday, shoppers, foods, black, season, target

#15 Employees companies, time, people, make, week, including, work, 3.61%

interview, want, just, need, way, making, does, spokesman

Table 11: This table presents the results from pairwise comparisons between sets of news articles

from diﬀerent positions and diﬀerent levels of importance. Each cell gives the p-value of a χ-square

test of independence between the topic distributions in the two speciﬁed sets of news articles. Topics

are estimated on the large training dataset from Reuters, and the results display robustness to the

number of topics varying from 10 to 25 (baseline with 15 topics is highlighted in blue).

Panel 1 compares front page SI news articles against non-front page ones.

Panel 2 compares PI news articles against front page SI news articles.

Panel 1: Front Page SI versus Non-Front Page SI

# Topics in Model p-value

10 topics 0.8670

15 topics 0.8776

20 topics 0.8731

25 topics 0.7801

Panel 2: PI versus Front Page SI

# Topics in Model p-value

10 topics 0.1236

15 topics 0.0836†

20 topics 0.0526†

25 topics 0.0417*

* and † denote signiﬁcance at the 5% and 10% levels, respectively.

Table 12: Summary statistics of the ﬁnancial experts surveyed regarding the news. The table

presents the breakdown of the survey respondents across MBA students and active ﬁnance profes-

sionals, as well as the breakdown of precise aﬃliations in each group across speciﬁc schools and

institutions.

Aﬃliation Type Institution Percentage

MBA Students 21.4% of Total

Breakdown:

Harvard Business School 42.3%

The Wharton School 42.3%

Columbia Graduate School of Business 3.9%

University of Chicago Booth School of Business 3.9%

UVA Darden School of Business 3.9%

McDonough School of Business, Georgetown 3.9%

Professionals 78.6% of total

Breakdown:

Hedge Funds 8.0%

Bridgewater Associates, AQR Capital Management,

Tudor Investment Corp, BlueMountain Capital

Management, Blue Ridge Capital, QTrade Capital,

Bluegrass Capital, One East Partners

Investment Managers 20.7%

BlackRock, The Vanguard Group, State Street,

Fidelity, Paciﬁc Investment Management Company,

Wellington Management Company, Northern Trust

Company, T. Rowe Price, Dodge & Cox Funds,

Acadian Asset Management, Eachwin Capital, Crane

Asset Management, Wafra Investment Advisory

Group, Cambridge Associates, Broadﬁn Capital

Pension Funds 0.7%

North Carolina Retirement System

Private Investors 0.7%

Banks and Broker Dealers 40.7%

JP Morgan, Morgan Stanley, Goldman Sachs,

Bank of America Merrill Lynch, BNP Paribas,

Credit Suisse, Deutsche Bank, Wells Fargo,

Royal Bank of Canada, UBS, Standard Chartered

Bank, Citizens Bank, The NEX Group, HSBC,

Edelweiss, Royal Bank of Scotland, SunTrust Bank,

Berliner Volksbank, First Republic Bank

Table 12 (Continued): Summary statistics of the ﬁnancial experts surveyed regarding the news.

The table presents the breakdown of the survey respondents across MBA students and active ﬁ-

nance professionals, as well as the breakdown of precise aﬃliations across speciﬁc schools and

institutions.

Aﬃliation Type Institution Percentage

Professionals

Breakdown:

Investment Banks 2.7%

Barclays Capital, Lazard

Insurance 3.3%

Massachusetts Mutual Life Insurance, Voya Financial,

Nippon Life Insurance, Liberty Mutual

Government Agencies & Sovereign Wealth Funds 2.0%

The Federal Reserve Board, Abu Dhabi Investment

Authority, World Bank Group

Corporations 4.0%

Nike, Inc., Shaw’s Supermarkets, Tiﬀany & Co.,

The Walt Disney Company, Philips, ReBio LLC

Private Equity & Venture Capital 11.3%

Blackstone Group, Warburg Pincus, Motive Partners,

Garrison Investment Group, ATL Partners, Tamarisc,

Pomona Capital, Clearview Capital, Cerberus

Capital Management, Madrona Partners

Consulting 0.7%

Boston Consulting Group

Non-Proﬁt 0.7%

Ford Foundation

Financial Advisory, Taxes, and Real Estate 1.3%

Princeton Tax Services, DK Partners, Condor Partners

Media 0.7%

Other ﬁnance professionals 2.0%

Table 13: This table presents the aggregated responses of ﬁnancial experts to the news survey.

The results are displayed for the full sample and separately for the sub-samples of active ﬁnance

professionals and MBA students. For all samples, the results are presented with and without the

respondents who do not complete the survey in full. Standard errors are clustered by participant.

Panel 1 reports the frequency with which ﬁnance professionals and MBA students identify front

page SI news articles as more impactful than their non-front page counterparts.

Panel 2 reports the incidence of ﬁnancial experts choosing PI news articles as more impactful

than front page SI news articles.

Panel 1: Front Page SI versus Non-Front Page SI

Respondent Type Choosing Front Page Standard Error # Respondents

Finance Professionals 48.24% (1.21%) 150

MBA students 45.05%† (2.65%) 26

All Respondents 47.78%* (1.11%) 176

Finance Professionals (excl. attritors) 48.16% (1.24%) 136

MBA Students (excl. attritors) 44.83%† (2.69%) 25

All Respondents (excl. attritors) 47.67%* (1.14%) 161

Panel 2: PI versus Front Page SI

Respondent Type Choosing PI Standard Error # Respondents

Finance Professionals 61.16%** (2.13%) 150

MBA Students 57.54%* (3.55%) 26

All Respondents 60.58%** (1.87%) 176

Finance Professionals (excl. attritors) 61.59%** (2.20%) 136

MBA Students (excl. attritors) 57.66%* (3.61%) 25

All Respondents (excl. attritors) 60.95%** (1.93%) 161

** denotes a proportion diﬀering from 50% with signiﬁcance at the 1% level.

Appendix A Technical Details

A.1 Latent Dirichlet Allocation

I brieﬂy present the Latent Dirichlet Allocation methodology for identifying representa-

tive topics covered by the ﬁnancial news in the training corpus. For additional details on

the methodology, please refer to Blei et al (2003).

Let D denote the set of ﬁnancial news documents in the training corpus, with d ∈ D

representing an individual document. Each document d is a sequence of N words: d =

, ..., w

), where w

is the nth term to appear in the document d. All terms come from

the vocabulary W , which is constructed as described in Appendix A.2.

The latent set of topics is denoted by T , where each element t ∈ T is a unit vector

in k-dimensional space. The parameter k is the desired number of topics, speciﬁed by the

researcher.

The Latent Dirichlet Allocation algorithm conceptualizes each document d as a sequence

of words drawn from a latent distribution D

over topics. The distribution D

is itself ran-

domly determined for each document: in particular, for each document d, D

is a multinomial

distribution whose parameters are a random variable drawn from a pre-speciﬁed Dirichlet

prior.

Speciﬁcally, the generative process assumed by the Latent Dirichlet Allocation algorithm

is as follows.

• Pre-specify model parameters: ξ, α, β.

• To construct each new document d:

1. Choose the document length N

∼ P oisson(ξ).

2. Choose a distribution over topics θ

∈ Dir(α).

3. Fill the N words in the document d by sequentially choosing each word w

follows:

(a) Choose a topic t

∼ Multinomial(θ

(b) Choose a word w

from P{w

, β}, the conditional probability distribution

over words in the vocabulary given the chosen topic t

The model relies on three parameters: ξ, α, and β. The parameter ξ is chosen to best

match the set of document lengths in the corpus, assuming that the lengths are drawn from a

Poisson distribution. This parameter is independent of the rest of the process, and therefore

I forego it in the remainder of the discussion.

The key model parameters of interest are α and β: α is a k-dimensional vector that

governs the relative frequencies of the k topics, and β is a k-by-2, 000 matrix that speciﬁes

the likelihood of each word in the vocabulary conditional on each of the k topics. Thus, the

element in the ith row and jth column of β is β

i,j

= P{w

= 1|t

= 1}.

In theory, the parameters α and β are estimated to maximize the likelihood of observing

the actual corpus of documents D. The conditional probability of observing a document d

given the model parameters α and β is given by:

P{d|α, β} =

Γ(Σ

i=1

)

i=1

Γ(α

)



i=1

−1





n=1

i=1

j=1

(θ

i,j

)



dθ, (A.1)

where w

denotes the jth component of the nth word vector w

, and Γ(·) is the Gamma

function.

Note that unlike other methods such as the probabilistic Latent Semantic Indexing ap-

proach (see Hofmann (1999)), the Latent Dirichlet Allocation method does not require the

parameters to be estimated individually for each document; there is a single set of parame-

ters ξ, α, β for the entire model. This oﬀers two advantages highlighted by Blei et al (2003).

First, by reducing the number of estimated parameters, the Latent Dirichlet Allocation ap-

proach reduces the computational complexity of the estimation problem. Second, and more

importantly, the Latent Dirichlet Allocation method allows for the generation of any arbi-

trary document and facilitates the evaluation of the likelihood of out-of-sample documents.

This ability to represent out-of-sample documents in terms of the identiﬁed topics is essential

to the application in this paper.

A.2 Text Preprocessing

The Latent Dirichlet Allocation algorithm takes as its input a set of documents, each

represented by a sequence of terms from a pre-speciﬁed vocabulary. Before applying the topic

modeling methodology, I need to identify a relevant vocabulary to represent the ﬁnancial

news documents. I proceed in three steps.

First, in order to focus on the set of relevant terms, I begin by stripping out all “stop

words.” To identify “stop words,” I use the list provided by the University of Glasgow

Information Retrieval Group.

Second, I construct the vocabulary using not only single words appearing in the TRC2

news corpus, but also common pairs of words. The Latent Dirichlet Allocation method is

a bag-of-words method, meaning that the algorithm ignores the ordering of terms within a

The full list of stop words can be accessed at http://ir.dcs.gla.ac.uk/resources/linguistic utils/stop words.

document and treats each term as an independently drawn random variable. In theory, this

may be a problematic assumption, particularly for ﬁnancial news, where some concepts are

captured by phrases, for example “traditional enterprise” or “stock exchange.” In order to

account for this feature of the data, I augment the vocabulary of unigrams (single words)

appearing in the corpus with bigrams (pairs of words).

Lastly, I limit my attention to the most common and representative terms. In particular,

I focus on the terms that appear in at least two distinct documents and that appear in

no more than 70% of the documents in the training corpus. Furthermore, the terms are

ranked according to frequency in order to capture relative importance; the ﬁnal vocabulary

is comprised of the top 2, 000 terms.

A.3 Topic Model Estimation

I estimate the model varying the number of iterations and the number of identiﬁed topics.

The model’s ﬁt ﬂattens out at around ﬁfteen topics.

In practice, the expression in (A.1) is intractable, and hence parameter estimation re-

lies on approximate inference methods. Following Griﬃths and Steyvers (2004), I estimate

the parameters from the training corpus of documents using a collapsed Gibbs sampling

algorithm.

I vary the number of iterations of the sampling algorithm from 30 to 1,000, and ﬁnd that

the marginal improvement in the model’s ﬁt is largest up to approximately 250 iterations,

and mostly ﬂattens out after 500 iterations (see Panel 1 of Figure A.1, which plots the log

likelihood as a function of the number of iterations for a model with k = 15 topics). The

results in the paper come from the estimation algorithm with 500 iterations for all considered

speciﬁcations.

The results from estimating the parameters of the Latent Dirichlet Allocation topic model

for k ∈ {10, ..., 40} indicate that the model’s ﬁt is best at around k = 15 topics. Panel 2

of Figure A.1 plots the model ﬁt for k ∈ {10, ..., 40}. For each number of topics, the model

is estimated using the collapsed Gibbs sampler with 500 iterations. The ﬁgure shows the

ﬁnal log likelihood for each speciﬁcation. The model’s ﬁt improves somewhat as the number

of topics increases from ten to ﬁfteen, with an increase in log likelihood from −6.01 × 10

to −6.00 × 10

. Increasing the number of topics to 20, 25, or 30 does not oﬀer marginal

improvements over the k = 15 speciﬁcation. Increasing the number of topics further to 35

or 40 markedly decreases the estimated log likelihood. Overall, the k = 15 speciﬁcation

achieves the best ﬁt after 500 iterations.

100 200 300 400 500 600 700 800 900 1000

Number of Iterations

-6.4

-6.3

-6.2

-6.1

-6

-5.9

Log Likelihood, Model with 15 Topics

Panel 1: LDA Estimated with Collapsed Gibbs Sampler, 30-1000 Iterations

10 15 20 25 30 35 40

Number of Topics

-6.05

-6.04

-6.03

-6.02

-6.01

-6

Model Log Likelihood with 500 Iterations

Panel 2: LDA Model Log-Likelihood for 10-40 Topics

Figure A.1: Log likelihood for diﬀerent estimations of the Latent Dirichlet Allocation model.

Panel 1 plots model ﬁt as a function of the number of iterations, using collapsed Gibbs sampling

estimation and ﬁxing the number of topics at k = 15. Panel 2 displays the model log likelihood

for the number of topics varying from k = 10 to k = 40, estimating each speciﬁcation with 500

iterations using the collapsed Gibbs sampler.