When is a MAX not the MAX?

How news
resolves information uncertainty
Article

Accepted Version

Creative Commons: Attribution-Noncommercial-No Derivative Works 4.0

Tao, R., Brooks, C. and Bell, A. R. (2020) When is a MAX not

the MAX? How news resolves information uncertainty. Journal
of Empirical Finance, 57. pp. 33-51. ISSN 0927-5398 doi:
https://doi.org/10.1016/j.jempfin.2020.03.002 Available at
https://centaur.reading.ac.uk/89795/

It is advisable to refer to the publisher’s version if you intend to cite from the
work. See Guidance on citing .

To link to this article DOI: http://dx.doi.org/10.1016/j.jempfin.2020.03.002

Publisher: Elsevier

All outputs in CentAUR are protected by Intellectual Property Rights law,

including copyright law. Copyright and IPR is retained by the creators or other
copyright holders. Terms and conditions for use of this material are defined in
the End User Agreement .

www.reading.ac.uk/centaur

CentAUR
Central Archive at the University of Reading
Reading’s research outputs online
 Tomorrow’s Fish and Chip Paper? Slowly incorporated
News and the Cross-section of Stock Returns
∗ † ‡
Ran Tao Chris Brooks Adrian Bell
March 8, 2020

Abstract
A large literature debates the link between news and investor decision making. Relying
on unique U.S. firm-level news data between 1979 and 2016, we document a cross-
sectional difference in the speed of diffusion of the information contained in news. We
distinguish news articles as being either slowly or quickly incorporated into contem-
poraneous stock prices. The return spread between these two types of news yields a
statistically significant profitability (139 basis points per month) and this effect can-
not be explained by other well-known risk factors. By employing novel attention data
(Google Search Volume Index and Bloomberg News Readerships Index), we find that
this news-induced anomaly can be attributed to limited-attention theory where firm-
specific news is not read by investors. Our research refines the role of news regarding
information dissemination in the financial markets.

JEL classifications: G12; G14

Keywords: News sentiment; information dissemination; stock return predictability; investor
attention; anomalies

We are grateful to Chris Adcock (the Editor), an Associate Editor, and two anonymous referees for
constructive comments and suggestions. We thank Tian Han for advice about textual data mining and the
Google Cloud Platform (GCP) providing research credits for our sentiment analysis. We also thank C.S
Agnes Cheng, Steven Young, Simone Varotto, Tony Moore, the 2019 BAFA doctoral conference and the
FMA Europe Doctoral Student Consortium for useful comments. All errors are our own.
∗
ICMA Centre, Henley Business School, University of Reading. E-mail: [email protected]
†
ICMA Centre, Henley Business School, University of Reading. E-mail: [email protected]
‡
ICMA Centre, Henley Business School, University of Reading. E-mail: [email protected]
1. Introduction
News articles usually have limited-time effects due to quick information transmission.
This intuition can be summed up by the well-known idiom that “today’s news is tomorrow’s
fish and chip paper”.1 Although today’s financial markets build on a continuous flow of news,
the analogy does not apply if we follow the existing body of news predictability studies. So
far, the evidence has revealed that returns following news arrivals are somewhat predictable.
While the literature attributes such a predictable pattern to either investor underreaction
or overreaction based on ex-post measures,2 a close look at the ex ante interaction between
the information embedded in news articles and contemporaneous stock reactions seems to
be ignored and might be worthwhile.
Arguably, the information conveyed by news articles will start to be incorporated into
prices immediately on its arrival, and so it is possible that different news articles might incor-
porate information into stock markets at different speeds. For example, one negative news
event could impact on the stock return gradually and persistently whereas the influence of
another might be quick and prominent. Therefore, in this paper we take a unique perspective
by studying ex ante instances when news tone does not match with contemporaneous stock
reactions. The essential question we ask is: what is the impact of future return performance
when contemporaneous stock reactions mismatch with the tone of news flows?
There are two important building blocks to derive our central hypothesis: First, the
monthly aggregated news sentiment scores can serve as a proxy for a firm’s economic fun-
damentals. Second, investors tend to underreact to the news momentum phenomenon (see
Wang, Zhang, and Zhu (2018)). To measure how quickly information is incorporated in
an ex ante instance, we use a sequential double-sorted approach based on monthly stock
returns and news sentiment scores. The idea is that there should be a monotonic relation-
ship between stock returns and aggregate news sentiment scores at a monthly horizon (e.g.,
Wang et al. (2018)). If there is a mismatch between the two, it is likely that investors have
underreacted to news fundamentals and a delayed reaction will be observed.
By allowing for a heterogeneous relationship between contemporaneous stock returns and
news tone, we are able to classify a news item in terms of whether it conveys information
1
In the UK, fish and chips (a takeaway treat) was traditionally wrapped in newspaper in order to ab-
sorb grease. This demonstrated that a newspaper was only valuable for the news it carried on the day of
publication.
2
For example, Chan (2003) conclude that investors underreact to public information and overreact to
private information by measuring long-run stock return performance; Tetlock (2007) documents an investor
overreaction pattern by constructing a VAR (Vector Autoregression) model. Other related studies include
Tetlock, Saar-Tsechansky, and Macskassy (2008), Garcia (2013), Ahmad, Han, Hutson, Kearney, and Liu
(2016), Jiang, Li, and Wang (2017), and Kräussl and Mirgorodskaya (2017).

1
quickly or slowly to the markets. For instance, a stock which has positive stock-related news
stories accompanied by modest or perhaps negative returns might suggest an inability to
incorporate information effectively. In other words, this fundamental information is not yet
incorporated into stock prices. As such, a central hypothesis underlying this research design
is that the predictable future returns will be stronger among these slowly incorporated (SI)
news stocks than those where news is quickly incorporated (QI).
We perform the following analysis: we first assign a sentiment score to each news article
by employing the Loughran and McDonald (2011) dictionary method. This dictionary is
based on the ‘bag-of-words’ model which pre-labels each word as either positive, negative,
model-weak, litigious, etc. This technique is advantageous as it can gauge business tone.
We then define an SI news stock and a QI news stock in two ways. Under the assumption
that stock returns are a function of stock-related news, we evaluate the investor reaction
to news based on the current monthly stock return. Specifically, we construct monthly
double-sorted portfolios by sorting stocks into terciles based on their current returns and
further sorting each return group into another tercile based on their aggregate news sentiment
scores. The good (bad) news accompanying negative (positive) stock returns is defined as
slowly incorporated news, whereas news with matching stock returns we refer to as quickly
incorporated news.
Next, we examine the post-formation stock returns of these two types of news-influenced
stocks. To do so, we compute equally-weighted average portfolio returns for the two groups
in the following month. A long/short portfolio consisting of buying the slowly incorporated
good news (LRHS) stocks and selling the slowly incorporated bad news (HRLS) stocks
generates abnormal future returns even after controlling for well-known risk factors. In
sharp contrast, the other strategy of buying quickly incorporated good news (HRHS) stocks
and selling quickly incorporated bad news (LRLS) stocks earns negative future returns.
The profitability figures for both sets of portfolios are robust to various specifications. By
including quintile portfolios, splitting into different sample periods, using a weekly frequency
of analysis and employing a different measure of news sentiment scores, the results are
quantitatively similar, suggesting that these SI news effects are unlikely to arise from data
mining or biased measures.
To further understand what drives SI news effects, we propose two hypotheses. One
explanation follows limited attention theory, suggesting that investors tend to narrow down
their focus to a few stocks instead of dispersing their selection evenly throughout the entire
stock universe, see (Hirshleifer, Lim, and Teoh, 2009). To be more specific, these limited-
attention investors, especially retail investors, tend to concentrate on certain stocks such
as big size, high media coverage, high trading volume stocks and firms with a high level of

2
analyst coverage. Undoubtedly, these firms have a better information environment where
investors face less information asymmetry and therefore it is more efficient for them to react
to news. In contrast, small and less-information rich firms are less likely to be under the
spotlight, causing delayed reactions to news by investors.
Second, the speed of incorporation could result from the different nature of the news
item, namely complexity and informativeness. Prior literature has studied the relationship
between textual complexity and corresponding market reactions (Loughran and McDonald,
2014; Lawrence, 2013). For example, Loughran and McDonald (2014) document that the
readability of financial disclosures has a significant impact on post-filing date returns, analyst
dispersion, and standardized unexpected earnings (SUE). Umar (2008) finds that a long news
headlines lead to a 40-basis point return under-reaction on the Seeking Alpha forum. Thus, it
is likely that slowly incorporated news conveys information that is challenging to interpret for
less sophisticated investors. Moreover, the literature also argues that different news articles
exhibit different informativeness features. You, Zhang, and Zhang (2017) show that market-
oriented media tend to be more comprehensive in reporting corporate events compared to
state-controlled media. Presumably, investors might react quickly to these more informative
news articles but under-react to those less informative articles which are labelled as slowly
incorporated news in our setting.
To test the two hypotheses, we examine whether slowly incorporated news is concentrated
among firms with less attention, proxied by small size, low media coverage, and low trading
volume. We find consistent evidence that slowly incorporated news has stronger effects
among low-attention firms. Furthermore, we investigate limited attention theory by applying
two direct attention proxies: the Google Search Volume Index (SVI) and Bloomberg News
Reading Activity Index (AIA, Abnormal Institutional Attention). As conventional attention
proxies cannot guarantee that investors are actually reading firm-specific news (albeit they
have high coverage), the literature utilises these alternative measures in order to better
capture investor attention (Engelberg and Gao, 2011). Our results suggest that SI news
with low investor attention, particularly without retail investor attention, predicts stronger
stock returns. In contrast, we find mixed evidence for cross-sectional differences between the
effects of news as measured by textual complexity and by informativeness features. While
SI news is not exclusively earnings-irrelevant and too hard to read (i.e. low readability), we
do find that these news articles tend to be less accurate (proxied by business uncertainty
tone) and less comprehensive (measured by article length). Collectively, these results do not
systemically support the notion that SI news is more complex and less informative.
Our paper contributes to the literature in several aspects: First, we propose a novel
ex ante approach to categorise news articles as slowly incorporated and quickly incorpo-

3
rated into prices. The predictable return patterns after the news will only emerge when
ex ante news sentiment scores and contemporaneous returns are mismatched. Second, our
results support limited-attention theory in which the investor attention radar keeps certain
stocks below the horizon and therefore they react to stock-related news slowly (e.g., Da,
Gurun, and Warachka (2014) and Fang and Peress (2009)). Finally, our findings also have
some implications for practitioners. The refined predictive model based on both news and
contemporaneous returns provides a promising avenue for a potentially profitable trading
strategy.
The rest of the paper is organised as follows: Related literature and hypothesis devel-
opment will be discussed in Section 2. Section 3 presents the news data collection process
and how we define SI and QI news. Section 4 reports baseline empirical results, additional
robustness tests and empirical designs for two potential explanations of our findings. In the
final section, we present our conclusions and outline further directions for research.

2. Related Literature and Hypothesis Development

Our paper is related to several existing strands of the literature. First, our research is
relevant to the topic of investor attention. The literature argues that it is either limited-
attention theory or prominence theory that determines the impact of news on stock prices.
Limited-attention theory, as suggested by Fang and Peress (2009), explains that a firm with
no media coverage earns higher returns than others, indicating that no attention can lead
to stock price rises. In contrast, prominence theory believes in the value of the visibility. In
other words, more media coverage leads to better asset returns – e.g., Hillert, Jacobs, and
Müller (2014) finds that the media makes momentum profitability. They attribute this to an
overconfidence-driven overreaction where investors tend to be attracted by a news story with
high coverage and therefore are overly optimistic. Recently, a new source of data has added
to the investor attention literature, namely Google SVI, Bloomberg AIA (Engelberg and
Gao, 2011; Ben-Rephael, Da, and Israelsen, 2017). Following the majority of the previous
literature, we examine various measures of attention including firm-specific media coverage,
firm-level Google SVI and Bloomberg AIA. The results support limited-attention theory
where a firm within an inferior information environment and with less attention tends to
have news that is slowly reacted to and thus achieves better stock returns.
Second, the readability literature is also relevant to our study. Acccounting research
documents that the complexity of corporate releases has certain impacts on post-filing
stock returns (Lawrence, 2013; Dougal, Engelberg, Garcia, and Parsons, 2012; Chakrabarty,
Seetharaman, Swanson, and Wang, 2018). A major measure of document complexity is the

4
‘Fog Index’ – which is defined as a linear combination of average sentence length and the
percentage of complex words. Loughran and McDonald (2014) argues that the complexity
of 10-Ks certainly affects investor reaction as simple and concise materials take investors
and analysts much less time to digest and to determine the valuation-relevant information.
Moreover, news materials can exhibit similar effects (Umar, 2008). Inspired by this strand of
the literature, we investigate whether our identified slow information incorporation is largely
attributed to document complexity. However, we do not find any strong evidence to support
this argument.
In addition, several studies of media characteristics are related to this paper. Fedyk and
Hodson (2017) find that ‘Front page’ news in Bloomberg terminals always induces higher
trading volumes and larger price drifts. In contrast, the reaction to ‘non-front page’ news
stories is much smaller. Boulland, Degeorge, and Ginglinger (2017) document that the
market quick reaction to firm-specific events is after which these news items are printed
in English and circulated in electronic format. Engelberg and Parsons (2011) argue that
media outlets seem to have different broadcasting radii, where local mass media significantly
drives local trading behaviour. Newspaper articles published by state-owned newspapers
and market-oriented newspapers exhibit different effects upon stock performance (You et al.,
2017).
Our research is also inspired by the link between textual analysis and investor decision
making (Tetlock, 2007; Tetlock et al., 2008; Garcia, 2013; Hadzic, Weinbaum, and Yehuda,
2015; Ahmad et al., 2016; Jiang et al., 2017; Caporale, Spagnolo, and Spagnolo, 2018).
Tetlock (2007) documents that the tone of the Abreast Of The Market column can predict
DJIA index price movements on the next trading day, and is followed by a subsequent
reversal. Garcia (2013) further confirms this effect by using different newspaper columns
and finds that the market response to news tone is stronger during periods of recession. In
addition, firm-level studies are explored in terms of media tone – e.g., Tetlock et al. (2008)
document that firm-level media tone can predict earnings surprises and daily stock returns.
Hadzic et al. (2015) shows that investors sometimes misunderstand public information and
will reverse the stock return in the following months. Ahmad et al. (2016) select 20 big
firms with consecutive daily news and categorises this news into being either informative
or noise. They show that informative news can predict persistent future returns whereas
‘noise news’ forecasts a subsequent reversal. Moving to the intra-day frequency, the effects
of media tone becomes fairly simple (i.e., good news predicts positive asset returns while
bad news predicts negative returns) (Jiang et al., 2017). Our study adds to this literature
as slowly incorporated news can be shown to forecast asset returns.
Finally, this paper also touches upon momentum and reversal literature. Return contin-

5
uations and reversals are normally measured on a weekly or monthly basis. In particular,
momentum strategies are operationalised by constructing a portfolio with a formation pe-
riod from 3- to 12-months whereas the short-term reversal is measured on a monthly basis
(Jegadeesh and Titman, 1993, 1995, 2001). Wang et al. (2018) study news momentum on a
monthly basis and find that the media tone of firm-level news flows is more likely consistent.
Our research adds to this literature by highlighting media-induced anomalies. That is, slowly
incorporated news tends to lead to momentum profits, whereas quickly incorporated news
generates return reversals.
Based on these studies, we develop our central hypothesis. It has been widely accepted
that both investor underreaction and overreaction can cause predictable return patterns
following measurable news. Chan (2003) documents that investors underreact to public
information (i.e., measurable DJNS news) and overreact to private information while Tetlock
et al. (2008) find that stock prices will have positive drifts after good DJNS news and
become negative when the news is bad. Wang et al. (2018) document a news momentum
phenomenon: stocks in the highest news sentiment score portfolio will typically continue to
have high news sentiment scores in the next month, and vice versa. Their further results
show that the future return will be high (low) for those stocks from the highest (lowest)
news sentiment score portfolio. On the other hand, Tetlock (2007) documents an investor
overreaction pattern: the tone of the Abreast Of The Market column positively predicts
stock index price movements on the next trading day and negatively predicts it in the short
term future, suggesting a return reversal. Borrowing these findings and applying them to
our research setting, it seems reasonable to infer that investors will underreact to these
measurable news stories when contemporaneous stock reactions do not quickly match with
the news tone. As a result, one can expect a delayed reaction in the following months. On
the other hand, investors might overreact to these news items when both news sentiment
scores and stock returns are in the top (bottom) bins.
Collectively, this discussion motivates our central hypothesis. The future return pattern
regarding SI and QI news is as follows:

Hypothesis: SI news tends to lead to high future returns and QI news tends to produce
low future returns.

6
3. Data and Methodology
3.1. News Collection and Sentiment Analysis
Our primary data comprises news taken from Dow Jones Newswire Archive.3 It contains
all of the news from the Dow Jones newswire and all Wall Street Journal newspapers. To
examine slowly incorporated and quickly incorporated news and their cross-sectional return
patterns, we construct a sample of firm-level news released by the common U.S. stocks listed
on the NYSE, AMEX and NASDAQ between 1979 and 2016.
The firm-level news articles are collected with a few filtering conditions. To be included
in our sample, firms should have at least one news story disseminated by the Dow Jones
Newswire. Second, we require stocks to have prices greater than one dollar and non-missing
values of market capitalisation and book-to-market ratio. Third, we only select firms listed
on the NYSE, AMEX and NASDAQ stock exchanges with share codes 10 or 11 as these two
share types allow us to retrieve all U.S. common stocks.
To obtain all potential news items from the Dow Jones Newswire Archive, we downloaded
the historical name change file from the Center for Research in Security Prices (CRSP). All
firm identifiers are changed to their historical names/tickers according to the period in which
they were officially registered. Eventually, via this method, we identified 21,242 stocks from
the database.
We assign news items to the particular firm permco (CRSP’s permanent company iden-
tifier) if the ticker tagged in the news matches the stock’s within the valid period. For
example, the ticker symbol for Apple Inc is AAPL, which would also be tagged in the news
if the story is relevant to the Apple company. But arguably, this method has pitfalls where
news will be collected regardless of how relevant it is for the firm. Presumably, an article
analysing the cell phone industry certainly differs from a news article addressing iPhone
products specifically in terms of the likely strength of any stock price reaction. Thus, in
the Appendix, we perform robustness checks by collecting firm-level news with more strict
conditions to increase its relevance.
Through this method, our initial sample has 14,079 Dow Jones news firms out of 21,242
targets from the CRSP universe between July 1979 and December 2016, 13,406 of which
have at least ten news items in the whole lifespan. One concern is that since the Dow Jones
Newswire database is underdeveloped and somehow incomplete during the early stage, it
might be the case that a number of firms have news reports in the Newswire but these
3
Dow Jones Newswire is a global leading real-time news product and is commonly used in many research
papers – e.g., (Tetlock, 2010, 2011; Engelberg, Reed, and Ringgenberg, 2012; Engelberg, McLean, and Pontiff,
2018).

7
articles do not have tickers attached. Therefore, our procedure does not capture these firm-
specific news items. On the other hand, these missing firms might not be covered by the
Newswire at all. In an unreported test, we compare two statistics: the total number of
stocks from the CRSP universe and the number of matched Dow Jones News stocks. During
the early period (1979 to 1995), the percentage of stocks covered by measurable Dow Jones
news reports is around 20% to 30%. However, this statistic increases rapidly to over 95%
after 1995. As a result, there is strong evidence to believe that most missing firms are
clustered in the early years. To further allay readers’ potential concerns that our baseline
results might be significantly influenced by these unmatched firms in the early years, we later
perform a sub-sample analysis and the results remain positive in all four sub-periods. Overall,
concerning the total number of firm-level news items, our sample stocks have 10,751,978 news
observations in total. On average, there are 763 news articles for each firm.
In addition, we also plot the histogram of news observations. The distribution of our
news data is highly skewed. As can be seen in the Figure 1, the percentage of firms with
larger news items gradually decreases. To be more specific, there are 4492 firms with 1 to
100 news items, accounting for 32% of total sample. This statistic declines to 15% when it
comes to firms with 100-200 news items, and only 8% for those with 200-300 news items.
Collectively, it can be concluded that most news items belong to a small number of firms.
This pattern is consistent with Tetlock et al. (2008), who find that large firms tend to have
news coverage every day whereas small firms only have sparse observations. One implication
regarding this property is that our baseline results might be driven by the small size effect.
However, our further robustness checks have eliminated this concern as we remove stocks
whose prices are below $5 per share and our results remain positive. We also incorporate
the size factor into the Fama-French factor models which should also soak up this effect.

———————————— Insert Figure 1 here ————————————

To quantify the sentiment from our firm-specific news, we employed the Loughran and
McDonald dictionary method as this approach builds on the work of Henry (2008).4 This
sentiment analysis method is a reliable and popular technique as it is tailored specifically to
finance applications (e.g., see Loughran and McDonald (2011) and Loughran and McDonald
(2015)). For instance, the word “abandoned” is pre-assigned as a negative sentiment word
4
Although Henry (2008) is the first work that contributes to sentiment analysis, we do not employ the
word list developed by therein for two reasons: First, Henry’s list only has a very limited number of sentiment
words (e.g., 85 negative words) whereas that of Loughran and McDonald (2011) includes 2329 such words.
More importantly, the most frequent LM negative words based on 10-K annual reports such as loss, losses,
claims, impairment, against, adverse, restated, adversely, restructuring, and litigation, do not appear in
Henry’s list. This suggests that Henry’s dictionary may not sufficiently capture all potential negative tone
from the text.

8
in this dictionary. Any document which contains the word “abandoned” will be counted and
increases the negative percentage. Although our core analysis builds on the sentiment score
measured by this method, we later consider employing other sentiment analysis to justify
that our results are not sensitive to different tools.
The very first step is to construct an article-level sentiment score. For each news story, we
adapt the method proposed by Garcia (2013), constructing an article-level news sentiment
score using both positive and negative sentiment words as indicators. Specifically, the article-
level news sentiment score is computed by taking the number of positive words minus negative
words divided by the total number of words.
In the second step, we standardise these news sentiment scores at a stock-level. For
each stock, the news sentiment score is standardised by subtracting the rolling mean of the
previous 12 months and dividing by the standard deviation of the same period for each firm.
In the following paragraphs, N SS is used to denote the standardised news sentiment score.
The final step is aggregation. To fit the N SS into our empirical setting (i.e., monthly
portfolio analysis), we aggregate these N SS every month. N SSi,t represents the aggregated
news sentiment score standardised in a time-series manner for each firm i in each month t.

N o. of pos − N o. of neg
nss = (1)
total number of words

nss − µnss
N SS = (2)
σnss
where µnss is the rolling mean of the previous 12 months news sentiment score (denoted by
N SS) and σnss is the rolling standard deviation of N SS during the same period for each
firm. In other words, N SS is standardised in a time-series manner, separately for each firm.

3.2. Slowly and quickly incorporated News

In this subsection, we derive two building blocks of this paper: (1) the monthly aggregated
news sentiment scores reflect a firm’s economic fundamentals. (2) investors underreact to
news momentum. Specifically, Wang et al. (2018) document a news momentum phenomenon:
By aggregating news articles’ sentiment scores on a monthly basis for each firm, they find
that stocks in the highest (lowest) news sentiment score portfolio continue to have high (low)
news sentiment scores in the future. Such a pattern is driven by the firm’s fundamentals.
When linking this news momentum pattern with stock prices, they find that those stocks
with the highest (lowest) news sentiment scores have high (low) future returns, which is
caused by investors’ underreaction.

9
 Our empirical design of SI and QI news portfolio construction builds on these two blocks.
To measure how quickly information is incorporated, we use a sequential double-sorted ap-
proach based on monthly stock returns and news sentiment scores. The idea is that there
should be a monotonic relationship between stock returns and aggregated news sentiment
scores at a monthly horizon (e.g., Wang et al. (2018)). If there is a mismatch between the two
patterns, it is likely that investors have underreacted to news fundamentals and a delayed
reaction will be observed.
To ensure that the stock return we utilise is firm-specific, we calculate the DGTW-
adjusted returns following Daniel, Grinblatt, Titman, and Wermers (1997) to remove the
expected return components of common risk factors including SMB (Small-Minus-Big), HML
(High-Minus-Low), UMD (Up-Minus-Down). Specifically, we independently sort the entire
stock universe into quintile portfolios based on firm size, industry-adjusted book-to-market
ratio, and industry-adjusted momentum. Next, we compute these 125 (i.e. 5 × 5 × 5) value-
weighted benchmark returns. The “cleaned” return is each stock’s raw return minus the
corresponding portfolio benchmark return, which can be called an abnormal return. The
double-sorted portfolios are constructed as follows: in each month, all stocks are partitioned
into tercile portfolios based on contemporaneous DGTW-adjusted returns. Low return,
medium return and high return subsets correspond to LR, M R and HR groups respectively.
We then form another three subsets in each return group individually based on stock-level
news sentiment scores (i.e. LS, M S and HS refer to low, medium and high news sentiment
subsets). By doing so, we end up with nine portfolios having different stock return and news
sentiment characteristics. Next, we label each portfolio given its news return characteristics.
For example, the HRLS portfolio comprises the stocks with high contemporaneous stock
returns and low news sentiment scores. Similarly, we define the LRHS and HRLS portfolios
as the slowly incorporated news groups and the HRHS and LRLS portfolios as quickly
incorporated news groups. The rationale is that a one-month window is enough for a stock
to adjust prices given arriving news. If returns cannot match news sentiment, it can be
accepted that stock prices did not catch the news signal promptly.
In Table 1, we report univariate analysis for the four types of news. The LRHS, HRLS,
LRLS and HRHS portfolios are presented in each column. Specifically, both news sentiment
scores and stock returns statistics exhibit considerable differences across the groups. For
example, the LRHS portfolio has average stock return of −12% per month and an average
news sentiment score of 0.94, which suggests that the stock in this portfolio tends to have
negative returns and positive news. In other words, LRHS stocks are slowly incorporating
good news. Consistent with this notion, we then find that the HRLS portfolio exhibits
negative news (the news sentiment score is -0.96) and positive return performance (average

10
15% per month), -1.17 news sentiment and -15% monthly return for the LRLS portfolio,
and those of HRHS are 1.02 and 15% respectively.
Moving to news volume, LRHS stocks release relatively fewer news articles compared
to the other three counterparts (an average of 8.68 for LRHS stocks, 11.75 for HRLS
stocks, 10.87 for HRHS stocks, and 11.71 for LRLS stocks). The increasing number of
news items in the following month can be observed for all four types of stock. This can be
interpreted as suggesting that news stocks are likely to have a continuous information flow,
especially SI news stocks. Readers may be concerned that these next month news items
could more heavily drive their contemporaneous stock returns than those news released in
previous periods. However, later in the robustness checks, we show that this is not the case
by dropping all news observations in the portfolio holding periods.
Overall, it seems that our sequential double-sorted approach separates stocks into differ-
ent types of news portfolios effectively.

———————————— Insert Table 1 here ————————————

4. Empirical Results
4.1. Baseline Model
We use a sequential double-sorted calendar-time portfolio approach to examine SI and
QI news return predictability. In each month from July 1979 to December 2016, we partition
all stocks into tercile portfolios based on their contemporaneous monthly DGTW-adjusted
stock returns. The LR, M R and HR groups contain stocks with low returns, medium returns
and high returns respectively. For each return portfolio, we further rank stocks into another
three portfolios based on news sentiment scores. LS, M S and HS therefore collect stocks
with low, medium and high new sentiment scores. Subsequently, we track the performance
of each subset over the following month by computing their equally-weighted stock returns.
By rolling this monthly window through the entire sample period, we obtain a time-series
return performance for all nine portfolios. In Panel A of Table 2, it can be observed that
news predictability monotonically decreases across different return groups. For example,
the LR and HS portfolios predict a 1.45% return per month in the post-formation period
whereas the LR and LS conjunction only achieves 1.05% every month. Similarly, it is also
evident that the predicting power of return performance monotonically decreases for each
news group from LR to HR.
We further study SI and QI news effects in Panel B. SI is a long/short portfolio obtained
by buying slowly incorporated good news (LRHS) stocks and selling slowly incorporated bad

11
news (HRLS) stocks. QI is a long/short portfolio obtained by buying quickly incorporated
good news (HRHS) stocks and selling quickly incorporated bad news (LRLS) stocks. If
investors do indeed react slowly to arriving news, we should observe that slowly incorporated
news predicts stronger post-formation stock returns than quickly incorporated news.
Panel B of Table 2 supports this hypothesis: the SI news portfolio (i.e., the spread
between LRHS and HRLS) earns 101 basis points per month (t-statistic: 5.85) whereas
the QI news portfolio (i.e., the spread between HRHS and LRLS) has negative return
predictability. The difference between the two, known as SMQ (Slow Minus Quick), gains
139 bps per month, which is significant at the 1% level.
Further, we perform a Fama-French regression analysis of these results and we report the
alpha for the two news groups. Specifically, the risk-adjusted returns of the SI news portfolios
remain consistently significant across all models, even when the short-term return reversal
factor is included. This result suggests that the SI news effect cannot be simply interpreted
as a short-term return reversal. In contrast, the QI news portfolio becomes insignificant after
controlling for the Fama-French three factors and five factors. The alpha of the SMQ portfolio
is significant across risk-adjusted models, suggesting that SI news indeed has stronger return
predictability compared to QI news. Since most of the long-short portfolio returns come
from SI news rather than QI news, we next focus on SI news predictability instead of SMQ
in the regression analysis.

———————————— Insert Table 2 here ————————————

4.2. Fama-MacBeth Regressions

The asset pricing literature identifies a number of stock characteristics as future return
predictors. Therefore, we are able to estimate several Fama and MacBeth (1973) regression
models to study the predictability of our SI news and QI news on the following months stock
returns after controlling for additional stock returns predictors. Specifically, we perform this
regression model cross-sectionally in each month:

EXReti,t+1 = α + β1 ∗ N SSi,t + β2 ∗ (N SS ∗ SIN ws)i,t + β3 ∗ (N SS ∗ QIN ws)i,t

+β4 ∗ SIGdN wsi,t + β5 ∗ SIBdN wsi,t + β6 ∗ QIGdN wsi,t (3)
+β7 ∗ QIBdN wsi,t + Xi,t + i,t

where EXReti,t+1 is the excess return of stock i at time t+1. N SS is the news sentiment
scores of stock i at time t. SIGdN ws, SIBdN ws, QIGdN ws and QIBdN ws are dummy

12
variables if a stock has a news article identified as slowly incorporated good news, slowly
incorporated bad news, quickly incorporated good news or quickly incorporated bad news.
N SS ∗ SIN ws and N SS ∗ QIN ws are the interaction terms where news sentiment scores
multiply with the SIN ws or QIN ws dummy variables. The literature and previous sections
in this paper imply a positive β2 and a negative β3 coefficient on the interaction variables
N SS∗SIN ws and N SS∗SIN ws respectively. In the X vector, a number of control variables
are added to capture various stock characteristics. Specifically, we first include SIZE and
BT M to control for the Fama and French (1993) predictors. LRET is the lagged one-month
stock return to control for short-term reversals. Furthermore, M OM is added in order
to capture momentum effects, which is computed by taking past twelve-month cumulative
return with at least eight-month valid observations. BET A is included and computed by
following Scholes and Williams (1977); Dimson (1979). We include IV OL (Idiosyncratic
volatility) in the regression motivated by Fu (2009), who argues that idiosyncratic volatility
represents how fast the firm-level information is incorporated into stock prices. We also add
ILLIQ, a firm-level illiquidity proxy by Amihud (2002), which is computed by the average
value of daily absolute stock returns divided by the dollar trading volume over the previous
year.
Table 3 reports all primary variable coefficients from the three regression specifications.
First, the N SS positively predicts future returns, indicating that our news articles are indeed
informative. Second, our SIBdN ws and N SS ∗ SIN ws variables are statistically significant
under all regression specifications. In particular, the coefficient of N SS ∗ SIN ws is 0.2004
(t-statistic=2.21) whereas the N SS ∗ QIN ws term is insignificant. The positive SIBdN ws
coefficient suggests that the return continuation mainly comes from slowly incorporated
bad news rather than good news occurring during the formation period. Comparing the
magnitudes of N SS and N SS ∗ SIN ws, the N SS ∗ SIN ws term is more than double that
of N SS (0.2004 for N SS ∗ SIN ws vs 0.0959 for N SS). All three specifications include a
battery of control variables. Collectively, these tests confirm that our primary news variable,
N SS ∗ SIN ws, has strong stock return predictability and is robust to the incorporation of
a number of control variables.

4.3. Robustness Checks

4.3.1. Alternative Information Incorporation Measure

Another measure of news incorporation would be to examine beta coefficients between

the contemporaneous stock return and news sentiment scores. Inspired by the calculation of
market beta in the CAPM model as in (Frazzini and Pedersen, 2014), we regress the daily

13
news sentiment score on the corresponding stock return in a monthly rolling OLS regression
model to obtain the coefficient. As such, it should improve measurement accuracy when the
data frequency is on a daily basis. The regression equation is presented as follows:

Reti,t = α + β ∗ N SSi,t + i,t (4)

where Reti,t is the daily return of stock i on date t. N SSi,t is the daily pessimism value
of stock-level news for stock i on date t. This regression is then performed using a rolling
monthly window. As a result, we obtain time-varying beta parameters between the stock
return and news sentiment for each stock and each calendar month. To address the concern
of overnight news or weekend news, which may have lagged return impacts and therefore
might not be captured by our regression model, we allow a lagged three-trading-day window
to capture market lagged reactions. For example, Monday night news will not have return
impacts until Tuesday morning. Friday night news will also not influence stock returns until
the next Monday when the market opens. It is then believed that our time-varying beta
should capture all these delayed news reactions. Moreover, we only include a stock in the
sample when it has at least five different news stories on five different trading days in the
case of our regression having sparse observations. As a result, a steep sloping coefficient β
indicates news that is quickly incorporated, whereas a gentle slope implies slow information
incorporation.
To trace the performance of different coefficients between the daily news sentiment score
and daily contemporaneous stock returns, we employ a single-sorted calendar-time portfolio
approach with a one-month rolling window. Specifically, in each month, we partition stocks
into three terciles based on the β coefficients (i.e., LoBET A, M eBET A and HiBET A
portfolios). Naturally, the higher is the coefficient, the quicker the stock return responses.
The top portfolio HiBET A then contains high beta coefficients, suggesting that news is
incorporated quickly into stock prices. The bottom group, LoBET A, comprises those stocks
with low beta coefficients, which indicates slow information incorporation. Finally, equally-
weighted portfolio returns are computed during the post-formation period for the three
portfolios separately.
The result of this alternative definition is reported in the Panel A of Table 4. The
coefficient of the top portfolio (SINws group) is 1.28, which is significant at the 1% level,
whereas the QINws portfolio achieves 80 bps per month (t-statistic: 2.27). The Slow-Minus-
Quick (SMQ) portfolio reports an average 49 bps in each month, which is significant at
1%. Compared to the SMQ portfolio constructed in the first way (94 bps per month), the
overall profitability of SMQ created by news beta is halved. Presumably, our sample size is
significantly reduced with a number of small size stocks excluded due to the fact that the

14
stock included in calculation must have at least five news stories published on five different
days in a month. It is unlikely that small firms have more than five different news releases
in any month. Although this implies that the SI news effect is partially contributed by the
small size effect, our SI news theory, on the other hand, seems to be more convincing under
these two different designs.

4.3.2. Further Robustness Checks

We examine the robustness of our primary SI and QI news definitions. The sequential
double-sorted of SI and QI news leads us to a potential issue. Since news sentiment scores
and stock returns are positively correlated, the second sort on returns may create further
variation in the next month’s stock performance. As such, we perform an independent
double-sorted on news sentiment and stock returns in the first robustness test. The SI and
QI news defined by independent double-sorted displays the same pattern as those in Table
2, with return predictability achieving 100 bps (t-statistic=5.80) for SI news and -40 bps for
QI news.
A potential concern regarding our findings is that the effect is largely driven by penny
stocks. To avoid the bid-ask bounce and illiquidity, we exclude stocks whose prices are below
$ 5 per share at the end of each particular formation month. As panel B of Table 4 shows,
the SI news effect still holds, albeit with reduced magnitudes.
One could argue that the predictability of SI news is due to a number of news articles
arriving in the subsequent months. This might be true given the fact that news volume
increases over time, see Table 1. The SI news effect could be driven by news released in the
holding period rather than caused by the news in the previous formation period. To address
this concern, we eliminate all news observations occurring within the holding periods. As
can be seen in Panel C, the SI news effects remain statistically significant.
In addition, we attempt to quantify news sentiment by different sentiment indicators,
or even different techniques. To mitigate the concern that some news items contain high
volumes of both positive and negative sentiment words and lead to low levels of pessimism
(i.e. N eg - P os), we also report the results measured by only employing either N eg or
P os indicators by the Loughran and McDonald (2011) dictionary method. Furthermore,
we attempt to apply the different sentiment analysis tool. Earlier methodologies have been
criticised in three main ways: First, the underlying bag-of-words model only detects single
words as the leading judgment of sentiment analysis. It fails to take into account that some
sentences contain “navigating words” such as doesn’t, don’t, can’t, etc., which can change
the sentiment of the entire sentence. Second, this model fails to consider “modifier words”
such as really, too much, etc. as these words or phrases sometimes enhance the positive

15
and negative tone. Third, without analyzing grammatical structures, the Loughran and
McDonald (2011) dictionary method finds it difficult to deal with part-of-speech tagging.5
For example, the firm-specific news of company “Best buy” will be more positive than others
given its firm name contains a positive sentiment word “Best”. To address these pitfalls and
the potential bias caused by them, we employ the Google Natural Language API as an
alternative sentiment analysis.6 Overall, the results remain quantitatively similar.
As for any definition, our measure of SI and QI news is somewhat subjective. For instance,
we could examine the slowly incorporated news effect with a one-week time interval instead
of a one-month window with or without matching returns. In our sample, SI news predicts a
53 bps return on a weekly basis (equivalent to 2.21% per month), compared to 1.01% using
monthly windows. QI news exhibits similar patterns (-0.79% per month). The rationale
behind this finding is that the relevance of news is time-limited and its effects will fade away
over longer horizons.
Collectively, the above robustness checks of the SI news effect should eliminate concerns
of data mining or biased measures. In the next subsection, we study the time-variation of
SI news effects within our sample.

———————————— Insert Table 4 here ————————————

4.3.3. Subsample Analysis

Undoubtedly, technical advances have facilitated information dissemination, therefore

reducing the gap between information released and information received by investors. If SI
news is caused by information that travels slowly, we should observe that the predictive power
of SI news has gradually diminished over time. To test this hypothesis, we evenly split our
sample into four different periods, the first two of which might be termed pre-Internet and
the later two are post-Internet periods. Within each period, we perform the same analysis
as previously discussed above. In Table 5, SI news significantly predicts the next month’s
stock returns during all sub-sample periods (93 bps during 1979 to 1987, 108 bps between
1988 and 1997, 132 bps from 1998 to 2007 and 62 bps between 2008 and 2016), which are
all significant. Interestingly, the largest magnitude of SI news emerges during the 1998-2007
period where the number of news items dramatically increases. Until the crisis period (i.e.,
2008 - 2016), we observe a reduced SI news effect.
5
Part-of-speech tagging is a computational linguistic method to label the category of words as noun,
adverb, adjective, etc.
6
The Google Natural Language API is a newly-built natural language processing tool by Google Cloud,
details of which can be found at https://cloud.google.com/natural-language/.

16
 Given the relatively stable performance of SI news across each sub-sample, it is implau-
sible to conclude that SI news is caused by the slow information dissemination hypothesis,
at least not for the post-Internet period.

———————————— Insert Table 5 here ————————————

4.3.4. Long-run Performance

It is natural to ask whether SI news is informative or noisy. If our news source contains
genuine information about a firm’s fundamentals, the return predictability of SI news should
not subsequently reverse in the long-run. On the other hand, QI news has incorporated this
information into the market price and therefore we should not observe any further return
movements.
To examine long-run performance, we conduct a calendar-time portfolio approach over
four different holding periods as follows: 3 months (i.e., month 2 to month 4), 6 months
(i.e., month 2 to month 7), 9 months (i.e., month 2 to month 10) and 12 months (i.e.,
month 2 to month 13). These post-formation periods all skip the first month. However,
the current month (i.e., month 1 to month 2) is also included as a comparison. We again
construct a SI news portfolio that buys stocks with low returns but high news sentiment
(LRHS) and sells stocks with high returns but low news sentiment (HRLS), and a QI news
portfolio comprising of buying HRHS stocks and selling LRLS stocks. These portfolios are
then rebalanced monthly and their performance monitored over different holding periods.
The results, reported in Table 6, show that all periods observe insignificant coefficients for
SI news except the first post-formation period. It is evident that the return predictability
of SI news does not reverse in the long run. Similarly, we can also observe that the slow-
minus-quick factor gains significantly during the post-formation period and these portfolio
returns do not reverse in the following months.
On the other hand, the results for QI news are surprising as this effect starts obtaining
profits even after just 6 months. Presumably, this is because QI news stocks tend to be
prominent and more visible for investors and therefore lead to further momentum effects
after which the informative role of news dies away. Here, readers should distinguish the
effect between prominence channels and information channels where the former only provides
visibility but the latter gives new information, as documented by Kaniel and Parham (2017).
Unfortunately, it is difficult to find an appropriate technique to apply to distinguish the
effects of these two channels in this setting.
Overall, the long-run performance test confirms that SI news does contain new informa-
tion and indeed this is incorporated into the market at a later stage.

17
 ———————————— Insert Table 6 here ————————————

4.3.5. Trading Strategy

Our findings could potentially have useful implications for practitioners. In this sub-
section, we evaluate the impact of considering reasonable transaction costs on our trading
strategy’s profitability. The strategy we focus on is the slow-minus-quick portfolio that has
both long and short positions. Since it is difficult to obtain precise transaction cost estimates
for each trade, we therefore make a rough assumption that each trade incurs a total cost
between 10 and 100 bps. This estimated cost will include all necessary expenses such as
commissions, exchange fees, bid/ask spreads, market impact costs, and occasionally taxes.
To recalculate the trading strategy returns, we trace the position of each stock included in
our portfolio per month. A transaction cost will be incurred if a stock’s position is opened
or closed during the formation month. A double transaction cost will be incurred if a stock’s
position changes from buy to sell or from sell to buy. No transaction cost will be recorded
if a stock stays in the portfolio and the position remains unchanged. Table 7 presents the
abnormal and raw monthly trading returns under different cost assumptions.
As we can see from Table 7, the profitability of the trading strategy gradually declines and
becomes negative after accounting for 100 bps per transaction. It is likely that sophisticated
investors can keep transactions costs to a low level or manage a refined trading strategy using
different weighting schemes that might achieve better profits, and for such investors, the
trading strategy would still be profitable net of costs. At a minimum, our paper sheds some
light on one possible avenue for employing a news-based trading strategy for practitioners.

———————————— Insert Table 7 here ————————————

4.4. Limited Attention Theory?

4.4.1. Conventional attention proxies

Since these news portfolios exhibit cross-sectional differences in future return perfor-
mance, we should ask what drives this cross-sectional difference of stock returns. One expla-
nation regarding market delayed reactions is limited-attention theory. The growing literature
of this strand includes Engelberg and Gao (2011) on the volume of Google searches for stock
names, Da et al. (2014) on continuous information and momentum effects, and Fang and
Peress (2009) on non-news stock returns. The theory predicts that investors under-react
regarding stocks with low volumes of attention and thus the following return predictability
is the result of this delayed reaction. Intuitively, stocks with small size, low media coverage,

18
low turnover and low analyst coverage are generally associated with a poor information envi-
ronment. In other words, it is information asymmetry that slows down investors’ perceptions
of arriving news. If so, we would expect that our SI news effects are concentrated in the low
attention-grabbing stocks whereas QI news stocks would receive a relatively high volume of
attention.
In this section, we employ four different investor attention proxies to study SI news ef-
fects following the previous literature (in particular (Da et al., 2014; Huang, 2018)), which
includes size, media coverage, trading volume and analyst coverage. We first examine the
cross-sectional difference between small and large firms. To do so, we perform a double-sorted
and construct calendar-time portfolios. Specifically, we rank the firm’s market capitalisa-
tion values and partition these firms into either small size groups or big size groups before
constructing SI and QI news portfolios. The market capitalization (SIZE) is computed by
taking the natural logarithm of stock market values re-balanced at the end of June. We then
form our SI and QI portfolios with small size and big size groupings. The next-month stock
return performance is then traced for each month and the tabulated results are reported in
the panel A of Table 8.
As Table 8 shows, it is possible to observe that small firms predict higher returns in the
following month compared to large firms. Statistically, SI news earns returns more than three
times higher among small firms than large firms – 161 basis points versus 50 basis points
respectively. The difference between the two the subsets is significant at the 1% level (with
a t-statistic of 4.54). This suggests that SI news effects are concentrated in small sized firms,
and this finding is consistent with limited-attention theory. However, there is no evidence
for cross-sectional differences between small and big firms with QI news. Finally, our SMQ
factor also positively confirms the SI news effect.
Following the literature on the impact of media attention (e.g., (Fang and Peress, 2009;
Hillert et al., 2014)), changes of media coverage are often applied as the representation of
investors’ attention. Stocks with higher ∆ media coverage can naturally diffuse information
in a way which is much quicker than lower ∆ media covering stocks. Thus, SI news is more
likely to be observed among low-attention stocks and QI news concentrated in high-attention
stocks. Empirically, Panel A supports this limited-attention theory. With the monthly ∆
media coverage shifting from low to high, SI news predicts next-month returns from 144
basis points per month to only 9 bps. The difference between the two coverage groups is
significant at the 1% level. The SMQ factor also suggests that information diffusion is likely
to be slower within the low media volume portfolio (2.21% with a t-statistic of 7.56 for the
cross-sectional difference between low and high media coverage groups), which is consistent
with limited attention theory.

19
 Continued Panel A also tests the hypothesis regarding whether stocks with lower trading
volume predict higher SI news profitability. The literature documents that trading volume
can be a proxy for investor attention (Barber and Odean, 2007). To perform this test,
we again construct a double-sorted calendar-time portfolio by employing stock-level trading
volume. The variable is calculated as the natural log of the average share volume divided
by the number of shares outstanding over each month using daily data. We then trace
the portfolio performance by a rolling one-month window to observe SI news effects under
either high or low trading volume groups. In the Table 8, we report all coefficients with t-
statistics for each subset of stocks. Specifically, SI news stocks again earn 1.36% per month
(t-statistic=8.94). The difference in profitability between low and high turnover SI news
is statistically significant at the 1% level. This is consistent with our theory that stocks
without active trading activities are more likely to be under an investor’s attention radar.
Overall, comparing low and high turnover profits, we can conclude that low trading volume
stocks are associated with higher SI news profitability.
Finally, we examine how analyst coverage is associated with cross-sectional difference in
SI news predictability. The literature uses analyst coverage as a proxy for investor attention
(Hirshleifer and Teoh, 2003). If investor inattention slows down information incorporation
from news articles, the predictability should be concentrated among firms with low analyst
coverage. The analyst coverage data is retrieved from Datastream for one-year forward
earnings per share forecast for each firm. The sample is then divided into two groups based
on the median number of analysts covering the stock, before which SI and QI news stock
portfolios have been constructed in each month. In the continued Panel A, the performance
of each portfolio has been reported. Stocks with low analyst coverage have higher levels of SI
news compared to those with high coverage – 1.03% and 0.57% per month respectively. The
t-statistic of low-minus-high spread is 1.19, which is significant at the 5% level. In contrast,
there is no evidence of QI news effects in either low or high sub-samples.
To summarise, four different investor attention proxies consistently support limited-
attention theory, which indicates that SI news predictability is positively associated with
low investor attention. In the next section, we will examine this theory with further tests.

———————————— Insert Table 8 here ————————————

4.4.2. Google SVI and Bloomberg AIA attention proxies

Although our empirical results positively confirm limited attention theory, we have had
to assume that investors would have paid attention to the news if a company’s name is
mentioned. To address this issue, we further study limited-attention theory by utilising two

20
direct attention proxies: Google Search Volume Index and Bloomberg news reading activity
(AIA, i.e., abnormal Institutional investor Attention).
The direct attention proxies (Google SVI and Bloomberg AIA) have been proposed and
studied in the literature, and have provided compelling evidence involving certain financial
variables (Engelberg and Gao, 2011; Ben-Rephael et al., 2017). The Google SVI directly
measures investor attention by using aggregated web users searching frequency based on
Google search engines. That is to say, Google will count the number of visits for a particular
key word during a time period. For instance, users might type “Apple Inc” and look up
news related to the firm, which then is recorded and presented by the Google Search Volume
Index (SVI). As shown in Figure 2 below, the SVI index of the key word Apple Technology
company rises to a the peak of 100 on September 12, 2018 when the main launch event was
held by Apple. As a consequence, this SVI index brings a direct linkage between the interest
of the general public and firm-level news events.

———————————— Insert Figure 2 here ————————————

On the other hand, each Bloomberg terminal provides a news reader function where
the news reading activity of any firm is monitored on daily basis. Due to the financial
cost and required expertise of using Bloomberg, most terminal users are from the financial
services sector, working as portfolio managers, stock analysts, or traders, etc. In other words,
Bloomberg news reading activity is highly likely to represent institutional investor attention
compared to the Google search engine which tends to represent retail investor attention.
To download Google SVI data, we use a Python web crawler to automatically send
keywords (firm legal names) and retrieve data for each firm following certain procedures.
Due to the mechanism of this data request, Google censors often return empty values for
certain firm names. We then conduct several techniques to address this issue, the details of
which can be found in the Appendix. Although given the fact that misrepresentation and
bias of our data can exist, we believe that such issues have been minimised in this case.
Finally, for a total number of 7,864 U.S. firms, we have evidence for 4,545 firms. As for
the Bloomberg AIA data, we follow the procedure by Ben-Rephael et al. (2017) in which
they download the Russel 3000 stocks from the year 2010 forwards.7 We then move to data
pre-processing. Specifically, Bloomberg news reading activity data is a time-series measure
of the rolling prior 30-days terminal users reading activity on an hourly basis. The value is 0,
1, 2, 3 or 4 if the rolling average is below 80%, between 80% and 90%, between 90% and 94%,
between 94% and 96%, and above 96% respectively. Following the method in Ben-Rephael
et al. (2017), the AIA measure is a dummy variable which takes the value of zero when the
7
Bloomberg AIA data is missing for the periods 12/6/2010 - 1/7/2011 and 8/17/2011 - 11/2/2011.

21
score is 0, 1, or 2 and a value of one otherwise. To be consistent with Bloomberg AIA, we
also assign the score for the Google SVI data based on the same method. The Google ASVI
(Abnormal SVI) value will be one if the corresponding score is 3 or 4 and zero otherwise.
These procedures allow us to capture the right tail of the distribution.
With these two attention proxies, we then ask whether low attention could explain
stronger news predictability, particularly for slowly incorporated news. To examine this,
we conduct a calendar-time portfolio approach with one trading day as a formation period
and use the following ten trading days as a holding period for either the high or the low
attention sub-sample. We change the research setting in this case for two reasons: first,
both Google ASVI and Bloomberg AIA are reported on a daily basis and have been scaled
in a proprietary manner prior to downloading. We are cautious to keep the natural data
frequency in order to capture investor attention precisely. Second, due to the sample range,
our period examined approaches recent years and both the SI and QI effects become weaker,
as discussed in Section 4.3.3. The weaker effects might be partially attributed to advanced
newswire transmissions such as Twitter, which speed up information dissemination. Overall,
daily-basis settings can better capture the SI news effects.
Moreover, to address the concern that part of the SI news is released around the market
closing bell and investors do not have enough time to fully digest it, we therefore categorise
the news on the date t at least 30 minutes prior to the market closing and label them on the
date t + 1 otherwise. To trace both SI and QI news portfolio performances, we again employ
a double-sorted calendar-time portfolio approach. Specifically, we first sort stocks into either
high or low attention groups based on stock-level attention values. We then form SI and QI
news portfolios based on contemporaneous stock returns and news sentiment scores. The SI
news portfolio is constructed by buying stocks with bad returns but good news and selling
stocks with good returns but bad news; the QI news portfolio is formed by buying good
stocks for both returns and news (HRHS) and selling bad stocks (LRLS). These portfolios
are rebalanced each trading day and the performances are traced over the next ten trading
days.
In the Table 8, it is possible to observe that Google ASVI predicts a stronger SI news
effect. The difference between the low and high Google ASVI is even significant at the
1% (with t-statistic 4.44). The coefficients for the QI news subgroup are nearly zero (all
statistically insignificant). Unfortunately, we do not observe any significant evidence for
the Bloomberg AIA (our proxy for institutional investor attention). The results, therefore,
suggests that it is retail investors rather than institutional investors for whom genuine news
slips under their radar.
A further test is conducted to examine the relationship between SI news and retail at-

22
tention. Figure 3 nicely visualises the interaction between SI news and retail attention. To
perform a CAR analysis, we compute the characteristics-adjusted return based on (Daniel
et al., 1997) and donated as DGTW. The slowly incorporated good news and bad news
without retail attention achieves the highest and lowest abnormal returns respectively, while
the return of news under retail attention lies in between. As a consequence, direct attention
proxies allow us to take a close look at how limited investor attention is associated with SI
news effects.

———————————— Insert Figure 3 here ————————————

4.5. Complexity and Informativeness?

SI news effects can also potentially be explained by news characteristics. The literature
identifies that variation in news content does affect market reaction (Dougal et al., 2012;
Umar, 2008). For instance, (especially retail) investors might find a complex news article
challenging, in the sense that they might be uncertain about how to interpret the news
content. Likewise, an ambiguous news tone could also fail to give investors a clear trad-
ing signal, which therefore discourages them form quickly responding. On the other hand,
investors might under-react to a short news article because it means a less comprehensive
report about a corporate event. They are also more likely to be interested in earnings-related
news since these news items contain more valuable information (Tetlock et al., 2008; Chen,
De, Hu, and Hwang, 2014). If SI and QI news is driven by the nature of the news in this
setting, we would expect that easily-understandable and more informative news predicts
lower future stock returns. In other words, investors will react quickly to news stories with
fewer ambiguous words, fewer complex sentences, more comprehensive content, and more
earnings-related topics.
To capture the textual complexity of our news documents, we apply the fog index and
weak modal words measures, respectively. The Fog index is one of the most commonly
applied readability measures for firm’s financial disclosures (Dougal et al., 2012; Lawrence,
2013; Loughran and McDonald, 2014). It captures document readability by computing the
average sentence length and the percentage of complex vocabularies. Thus, the higher the
Fog index value, the more complex the material is. Weak modal words were first introduced
by Loughran and McDonald (2011) in their financial dictionary to gauge uncertain business
tone. Words such as ”maybe”, ”appears”, ”might” frequently shown in the news suggest
that these stories are less likely to attract investors. For example, Ahern and Sosyura (2015)
uses this word category to gauge merger and acquisition rumours. As a result, we argue
these two proxies are commonly used in the literature and should be effective to measure

23
complexity in this paper.
To measure the news informativeness features, we first utilize article length as a proxy for
comprehensiveness, as suggested by You et al. (2017). A long news article tends to convey
more useful information and helps (especially retail) investors better understand underlying
events. Next, we assess news informativeness by categorizing the news topic as earnings-
related or non-earnings-related. Specifically, we detect keywords from each news article using
the word stem “earn”. We identify the topic of a news article related to earnings if at least
one word stem “earn” is found. The rationale behind this is that an article mentioning the
word stem “earn” contains more information about a firm’s fundamentals and is more likely
to be value-relevant (Tetlock et al., 2008). Overall, it can be argued that a news story with
long length and more “earn” word stems represents greater informativeness.
To test our hypothesis, we examine the cross-sectional difference of news characteristics
individually. Specifically, we first calculate the median level of weak tone at each time point,
rank the stocks with higher weak tone into an Ambiguous group and lower weak tone into
an Accurate group. Similarly, we rank more readable stocks into a Concise group and less
readable stocks into a Complex group based on the median Fog index in each month. For
article length, we cross-sectionally rank all stocks into Short and Long groups based on a
median level of word counts in each month. Lastly, we split our sample into two groups:
EarningsEx, where stocks without any news articles containing earnings topics are collected,
and EarningsIn, where stocks with only earnings-related news articles are included. To ex-
amine the statistical difference between each group, we construct Ambiguous − Accurate,
Complex − Concise, Short − Long, and EarningsEx − EarningsIn, respectively. If com-
plexity and informativeness do affect investor trading behaviour, we expect that these four
cross-sectional differences would be positively significant for the SI news portfolio. The
equally-weighted portfolio return is then computed for each subset sample to trace its post-
formation performance.
Panel A of Table 9 reports the cross-sectional differences by textual complexity. The
evidence is mixed: although the cross-sectional difference between ambiguous accurate news
stocks is positively and statistically significant (t-statistic=2.29 for SI news), there is no
clear evidence showing a significant difference between complex and concise news. This
indicates that SI news predictability is not likely to come from hard-to-read news articles.
Moving to Panel B, it is evident that the cross-sectional difference between short news (less
comprehensive) and long news (more comprehensive) is significant (t-statistic of 2.70) for SI
news stocks. Interestingly, when we split our sample into earnings-related and non-earnings-
related topics, the cross-sectional difference between these two is insignificantly different from
zero. This suggests that SI news is not exclusively earnings irrelevant and at least contains

24
some valuable information.
Collectively, it can be concluded that SI news tends to use ambiguous tone and reports
less comprehensively but such stories are not too complex to read and are not completely
earnings-irrelevant. The empirical results do not systemically support complexity and infor-
mativeness theory.

5. Conclusions
In this paper, we first distinguish a firm-level news type as slowly incorporated, whereby
stock returns do not promptly respond to corresponding news content, or quickly incorpo-
rated, where stock return performance matches the news sentiment score. The feature of SI
news, specifically, is having a good (bad) sentiment score but with bad (good) stock returns.
A long/short portfolio is then constructed to capture the SI news effect as buying a stock
with low returns but a high news sentiment score and selling a stock with high returns but a
low news sentiment score. As a result, we find that this news-induced anomaly can achieve
139 basis points per month (equivalent to 16.68% per year).
The SI news effect can be explained by limited-attention theory. According to this theory,
investors tend to focus on a few stocks instead of evenly dispersing their attention throughout
the entire stock universe. This then leads to the phenomenon where a group of firm-related
news items drop under an investor’s radar. The stock returns thus react very slowly. In
empirical tests, we proxy less investor attention by small size, low media coverage, low
trading volume, low analyst coverage, low Google SVI and low Bloomberg AIA. The results,
strikingly, show stronger stock return predictability for less attention-grabbing stocks.
The empirical results, however, do not support complexity and informativeness theory
in which the nature of news might cause investors’ slow reaction. Generally, the textual
complexity of news content should influence investors’ trading behaviour, particularly for
retail investors, whereby they might have no idea how to interpret complex news items or
how to respond to ambiguous news articles. In addition, a short news article tends to be
less comprehensive in reporting underlying events and therefore discourages investors from
responding quickly. An article that does not contain value-relevant information could also
fail to make investors take interest. Yet the evidence does not align with this theory.
To conclude, our paper refines the role of news coverage regarding information dissemi-
nation as it relates to the large gap between “news supply” and “news demand”. If investors
keep their focus on high attention-grabbing stocks, even genuine news for low-attention
stocks cannot guarantee readership. Perhaps, from the news providers’ point of view, news
a month ago is only fish and chip paper. However, it might not be the case for investors if

25
a firm-related news article does not receive attention.
Our findings are of relevance to both academics and practitioners: For academics, we add
to the literature that the predictable return pattern after news emerges will only exist when
prior news sentiment scores and contemporaneous returns are mismatched. For practitioners,
our refined predictive model based on both news and contemporaneous returns provides a
promising avenue for a potentially profitable trading strategy.

26
References
Ahern, Kenneth R, and Denis Sosyura, 2015, “Rumor has it: Sensationalism in financial
media,”Review of Financial Studies 28, 2050–2093.

Ahmad, Khurshid, JingGuang Han, Elaine Hutson, Colm Kearney, and Sha Liu, 2016,
“Media-expressed negative tone and firm-level stock returns,”Journal of Corporate Fi-
nance 37, 152–172.

Amihud, Yakov, 2002, “Illiquidity and stock returns: cross-section and time-series ef-
fects,”Journal of Financial Markets 5, 31–56.

Barber, Brad M, and Terrance Odean, 2007, “All that glitters: The effect of attention and
news on the buying behavior of individual and institutional investors,”Review of Financial
Studies 21, 785–818.

Ben-Rephael, Azi, Zhi Da, and Ryan D Israelsen, 2017, “It depends on where you search:
institutional investor attention and underreaction to news,”Review of Financial Studies
30, 3009–3047.

Boulland, Romain, François Degeorge, and Edith Ginglinger, 2017, “News dissemination and
investor attention,”Review of Finance 21, 761–791.

Caporale, Guglielmo Maria, Fabio Spagnolo, and Nicola Spagnolo, 2018, “Macro news and
bond yield spreads in the euro area,”European Journal of Finance 24, 114–134.

Chakrabarty, Bidisha, Ananth Seetharaman, Zane Swanson, and Xu Wang, 2018, “Manage-
ment risk incentives and the readability of corporate disclosures,”Financial Management
47, 583–616.

Chan, Wesley S, 2003, “Stock price reaction to news and no-news: drift and reversal after
headlines,”Journal of Financial Economics 70, 223–260.

Chen, Hailiang, Prabuddha De, Yu Jeffrey Hu, and Byoung-Hyoun Hwang, 2014, “Wisdom of
crowds: The value of stock opinions transmitted through social media,”Review of Financial
Studies 27, 1367–1403.

Da, Zhi, Umit G Gurun, and Mitch Warachka, 2014, “Frog in the pan: Continuous informa-
tion and momentum,”Review of Financial Studies 27, 2171–2218.

27
Daniel, Kent, Mark Grinblatt, Sheridan Titman, and Russ Wermers, 1997, “Measuring mu-
tual fund performance with characteristic-based benchmarks,”Journal of Finance 52,
1035–1058.

Dimson, Elroy, 1979, “Risk measurement when shares are subject to infrequent trad-
ing,”Journal of Financial Economics 7, 197–226.

Dougal, Casey, Joseph Engelberg, Diego Garcia, and Christopher A Parsons, 2012, “Jour-
nalists and the stock market,”Review of Financial Studies 25, 639–679.

Engelberg, Joseph, and Pengjie Gao, 2011, “In search of attention,”Journal of Finance 66,
1461–1499.

Engelberg, Joseph, R David McLean, and Jeffrey Pontiff, 2018, “Anomalies and
news,”Journal of Finance 73, 1971–2001.

Engelberg, Joseph E, and Christopher A Parsons, 2011, “The causal impact of media in
financial markets,”Journal of Finance 66, 67–97.

Engelberg, Joseph E, Adam V Reed, and Matthew C Ringgenberg, 2012, “How are shorts in-
formed?: Short sellers, news, and information processing,”Journal of Financial Economics
105, 260–278.

Fama, Eugene F, and Kenneth R French, 1993, “Common risk factors in the returns on
stocks and bonds,”Journal of Financial Economics 33, 3–56.

Fama, Eugene F, and Kenneth R French, 2017, “International tests of a five-factor asset
pricing model,”Journal of Financial Economics 123, 441–463.

Fama, Eugene F, and James D MacBeth, 1973, “Risk, return, and equilibrium: Empirical
tests,”Journal of Political Economy 81, 607–636.

Fang, Lily, and Joel Peress, 2009, “Media coverage and the cross-section of stock re-
turns,”Journal of Finance 64, 2023–2052.

Fedyk, Anastassia, and James Hodson, 2017, “When can the market identify stale news?,”
Harvard University Working Paper.

Frazzini, Andrea, and Lasse Heje Pedersen, 2014, “Betting against beta,”Journal of Finan-
cial Economics 111, 1–25.

Fu, Fangjian, 2009, “Idiosyncratic risk and the cross-section of expected stock re-
turns,”Journal of Financial Economics 91, 24–37.

28
Garcia, Diego, 2013, “Sentiment during recessions,”Journal of Finance 68, 1267–1300.

Hadzic, Muris, David Weinbaum, and Nir Yehuda, 2015, “News content, investor misreac-
tion, and stock return predictability,” Syracuse University Working Paper.

Henry, Elaine, 2008, “Are investors influenced by how earnings press releases are writ-
ten?,”Journal of Business Communication 45, 363–407.

Hillert, Alexander, Heiko Jacobs, and Sebastian Müller, 2014, “Media makes momen-
tum,”Review of Financial Studies 27, 3467–3501.

Hirshleifer, David, Sonya Seongyeon Lim, and Siew Hong Teoh, 2009, “Driven to distraction:
Extraneous events and underreaction to earnings news,”Journal of Finance 64, 2289–2325.

Hirshleifer, David, and Siew Hong Teoh, 2003, “Limited attention, information disclosure,
and financial reporting,”Journal of Accounting and Economics 36, 337–386.

Huang, Jiekun, 2018, “The customer knows best: The investment value of consumer opin-
ions,”Journal of Financial Economics 128, 164–182.

Jegadeesh, Narasimhan, and Sheridan Titman, 1993, “Returns to buying winners and selling
losers: Implications for stock market efficiency,”Journal of Finance 48, 65–91.

Jegadeesh, Narasimhan, and Sheridan Titman, 1995, “Overreaction, delayed reaction, and
contrarian profits,”Review of Financial Studies 8, 973–993.

Jegadeesh, Narasimhan, and Sheridan Titman, 2001, “Profitability of momentum strategies:

An evaluation of alternative explanations,”Journal of Finance 56, 699–720.

Jiang, Hao, Sophia Zhengzi Li, and Hao Wang, 2017, “News momentum,” Michigan State
University Working Paper.

Kaniel, Ron, and Robert Parham, 2017, “Wsj category kings–the impact of media attention
on consumer and mutual fund investment decisions,”Journal of Financial Economics 123,
337–356.

Kräussl, Roman, and Elizaveta Mirgorodskaya, 2017, “Media, sentiment and market perfor-
mance in the long run,”European Journal of Finance 23, 1059–1082.

Lawrence, Alastair, 2013, “Individual investors and financial disclosure,”Journal of Account-

ing and Economics 56, 130–147.

29
Loughran, Tim, and Bill McDonald, 2011, “When is a liability not a liability? textual
analysis, dictionaries, and 10-ks,”Journal of Finance 66, 35–65.

Loughran, Tim, and Bill McDonald, 2014, “Measuring readability in financial disclo-
sures,”Journal of Finance 69, 1643–1671.

Loughran, Tim, and Bill McDonald, 2015, “The use of word lists in textual analysis,”Journal
of Behavioral Finance 16, 1–11.

Scholes, Myron, and Joseph Williams, 1977, “Estimating betas from nonsynchronous
data,”Journal of Financial Economics 5, 309–327.

Tetlock, Paul C, 2007, “Giving content to investor sentiment: The role of media in the stock
market,”Journal of Finance 62, 1139–1168.

Tetlock, Paul C, 2010, “Does public financial news resolve asymmetric information?,”Review
of Financial Studies 23, 3520–3557.

Tetlock, Paul C, 2011, “All the news that’s fit to reprint: Do investors react to stale infor-
mation?,”Review of Financial Studies 24, 1481–1512.

Tetlock, Paul C, Maytal Saar-Tsechansky, and Sofus Macskassy, 2008, “More than words:
Quantifying language to measure firms’ fundamentals,”Journal of Finance 63, 1437–1467.

Umar, Tarik, 2008, “Complexity aversion when seeking alpha,” University of Chicago Work-
ing Paper.

Wang, Ying, Bohui Zhang, and Xiaoneng Zhu, 2018, “The momentum of news,” Working
Paper, Central University of Finance and Economics.

You, Jiaxing, Bohui Zhang, and Le Zhang, 2017, “Who captures the power of the
pen?,”Review of Financial Studies 31, 43–96.

30
6. Appendix
6.1. Name-ticker News collection
In this section, we address the issue that news might be less relevant to the firm by only
matching tickers. We impose additional matching criteria roughly following Tetlock et al.
(2008). Briefly, the procedure proceeds as follows:
1. Download a list of company names and tickers from CRSP (all common stocks traded
on the NYSE, AMEX and NASDAQ between 1979 and 2016).
2. The firm name string must appear in the first 25 words of a news article including the
headline.
3. The firm name must be detected at least twice in the main body of news stories.
4. News reports with fewer than 50 words are excluded.
We tweaked the firms’ names depending on the searching quality we had. This is because
CRSP provides a very unique and different name strings in comparison of the common names
applied in the Dow Jones Newswire Database. The tweaking details are in the following:
1. CRSP puts space between the abbreviated letters. e.g., F D X CORP. We delete the
space in this case.
2. The firm name ends with Inc, Ltd, Corp, Co. Then the suffix will be removed. Note
that we apply this rule to all sample firms including Apple Inc. We insist on doing so because
the keyword Apple will not match many noise news reports related to apple the fruit after
conditioning tickers.
3. We replaced the abbreviated words INTL, MFG, CHEM with International, Manu-
facturing and Chemical respectively.
4. The name string which ends with NEW or OLD to specify the company’s status was
tweaked to keep the only name before these words.
Overall, our sample of news items halves to only 4,530,243 non-repeated firm-level news
stories and the number of sample stocks decreases to 12,072. The result remains consistent
as beforehand. Unfortunately, it is tricky to define the relevance between firm-level news
and the firm. Relevance scores designed by databases such as LexisNexis and RavenPack
seem to be unrealistic at this stage.

31
6.2. Google Search Volume Index collection
Unlike Engelberg and Gao (2011) using a sample of Russell 3000 stocks, our sample
extends to all U.S. common stocks and we also acknowledge the fact that not all stocks
have a non-zero SVI. We therefore carefully deal with the potential bias and errors when
retrieving the Google SVI from the Google Trends webpage. The procedure is as follows:
1. Instead of using firm legal names from the CRSP historical name file as Google
censor keywords, we utilise these firm names as search inputs but look into the Google auto-
suggestion menu. The optimal keyword would be chosen if the category of a keyword is
“company” or “corporation”. The rationale behind this is to convert firm legal names to the
names commonly searched by users or mentioned in the financial press.
2. The SVI data has been scaled by Google based on the sample range. In this paper,
we set up a 90-day period for each data request and Google censors therefore return daily
data.
3. To get rid of “noise” attention, the underlying search source is restricted to business
news only. Given that investors could come from the outside of the U.S., geographical
location is worldwide when downloading.
4. We improve the data quality by manually validating the firm name sought with the
original one. If there is a significant difference that can be visualised, the data will be
excluded from our sample.
In total, we finally obtain a 4,545-firm sample out of 7,864 targets.

32
 Fig. 1. Distribution of Firm-level News Observations

Figure 1 reports the frequency of news observations. X-axis is the number of news items for each firm.
Y-axis is the percentage of firms among total sample. Sample period is between 1979 and 2016.

33
 Fig. 2. Google Searching Volume Index for Apple Technology Company

Figure 2 plots the search volume index of Apple Technology company by the Google search engine between
4 July and 30 September, 2018. The raw search volumes have been scaled and display as a time-series value
between 0 and 100.

34
 Fig. 3. Retail Attention and slowly incorporated News

Figure 3 plots the effect of slowly incorporated news under different retail attention which is proxied by Google
ASVI data: (1) slowly incorporated good news (LRHS) under zero of retail attention (SIGdN ASVI0). (2)
slowly incorporated bad news (HRLS) under zero of retail attention (SIBdN ASVI0). (3) slowly incorporated
good news (LRHS) under one of retail attention (SIGdN ASVI1). (4) slowly incorporated bad news (HRLS)
under one of retail attention (SIBdN ASVI1). The LRHS and HRLS news portfolios are defined as follows:
On each trading day, firms are sorted into three groups based on their past abnormal returns, and then
within each group, stocks are further ranked into three groups based on their news sentiment scores. We
label the news on date t at least 30 minutes prior to the market closing bell, otherwise it is classified on
date t + 1. SIGdN is slowly incorporated good news, donated as LRHS and SIBdN is slowly incorporated
bad news donated as HRLS. The cumulative abnormal return is the DGTW-adjusted return computed by
following (Daniel et al., 1997). The sample period ranges from 2004 to 2016 on a daily basis.

35
 Table 1: Summary Statistics

Table 1 reports the summary statistics across all four types of news portfolios. The monthly sequential double-sorted approach is used. In each month,
firms are sorted into three groups based on their past abnormal returns, and then within each group, stocks are further ranked into three groups
based on their news sentiment scores. Abnormal return is the DGTW-adjusted return computed by following (Daniel et al., 1997). SIGoodN ews
is defined as those with low current stock returns but good positive news; SIBadN ews are those with high current stock returns but bad negative
news; QIGoodN ews refers to news portfolios with high current stock returns and good positive news. QIBadN ews are low stock return and bad
news stocks. N ews V olumet is the number of news articles on current month.

LRHS HRLS HRHS LRLS

SI Good News SI Bad News QI Good News QI Bad News
News Sentiment Scoret 0.94 -0.96 1.02 -1.17
Abnormal Returnt -0.12 0.15 0.15 -0.14
News Volumet 8.68 11.75 10.87 11.71
36

News Volumet+1 10.70 12.84 13.21 12.37

 Table 2: Slowly incorporated and quickly incorporated News

Table 2 Panel A reports the return and news predictability across terciles. The formation period and
estimation period are both one month. In each month, firms are sorted into three groups based on their
past abnormal returns, and then within each group, stocks are further ranked into three groups based on
their news sentiment scores. Abnormal return is the DGTW-adjusted return computed by following (Daniel
et al., 1997). All returns are reported in monthly average donated in percentages respectively. The sample
period ranges from 1979 to 2016. Panel B reports time-series portfolio return with the risk-adjusted alphas
including the Fama and French (1993) three-factor model, the Fama-French-Carhart four-factor model, the
Fama and French (2017) five-factor model, a liquidity-augmented Fama-French-Carhart four-factor model,
and short-term augmented Fama-French-Carhart four-factor model. The alpha estimates are obtained by
regressing monthly portfolio excess returns on the monthly returns from the risk factors. The definition of
SI and QI presents as follows:

SI = LRHS − HRLS
QI = LRLS − HRHS

where SI news is a long/short portfolio of buying slowly incorporated good news (LRHS) and selling slowly
incorporated bad news (HRLS). QI news is a long/short portfolio of buying quickly incorporated good news
(HRHS) and selling quickly incorporated bad news (LRLS). t-statistics are reported in parentheses and *,
**, *** refers to the 10%, 5%, and 1% significance levels respectively.

Panel A HS Mid LS
LR 1.45% 1.22% 1.05%
Mid 0.94% 0.77% 0.84%
HR 0.67% 0.53% 0.45%

Panel B TS Ret FF3F FF4F FF5Falpha FF4F Liq FF4F Rev

SI News 1.01%*** 0.92%*** 1.09%*** 0.85%*** 1.09%*** 0.82%***
t-stat (5.85) (3.66) (3.68) (3.00) (4.00) (4.11)
QI News -0.39%** -0.27% -0.64%*** -0.29% -0.61%*** -0.38%**
t-stat (-2.09) (-1.14) (-2.70) (-0.94) (-2.55) (-2.42)
Slow-Minus-Quick 1.39%*** 1.19%*** 1.73%*** 1.14%** 1.69%*** 1.20%***
t-stat (4.26) (2.59) (3.40) (2.01) (3.70) (3.75)

37
Table 3: Predicting Stock Returns by slowly incorporated News and quickly incorporated
News

Table 3 reports the stock return predictability of slowly incorporated news and quickly incorporated news
based on various regression specifications. SIN ws and QIN ws are dummy variables if a news is slowly
incorporated and quickly incorporated respectively. N SS ∗ SIN ws and N SS ∗ QIN ws are the interaction
terms where N SS is a measure of news sentiment score. CON T ROLS is a battery of control variables
including LRET , SIZE, BT M , BET A, IV OL, M OM , ILLIQ. SIZE is computed by taking the natural
logarithm of stock market values in each previous month. LRET is the lagged one-month stock return.
BT M is computed by taking the natural logarithm of stock market values divided by firm book values
adjusted at each end of June. BET A is calculated following Scholes and Williams (1977); Dimson (1979).
M OM is the stock’s most recent 12-month cumulative returns. IV OL is the idiosyncratic risk computed by
the standard deviation of residuals from the Fama-French-Corhart four-factor model over the month using
daily returns. ILLIQ is a proxy for stock liquidity based on (Amihud, 2002). The sample period ranges
from 1979 to 2016. t-statistics are reported in parentheses and *, **, *** refers to the 10%, 5%, and 1%
significance levels respectively.

Dep = EXRet
1 2 3
CONST 2.0814*** 2.0905*** 2.0721***
(6.22) (6.24) (6.22)
NSS 0.0868*** 0.1249*** 0.0959***
(3.09) (4.44) (3.11)
NSS*SINws 0.2080** 0.2004**
(2.30) (2.21)
NSS*QINws -0.1651 -0.1351
(-1.63) (-1.30)
SIGdNws -0.0555 -0.0488
(-0.52) (-0.45)
SIBdNws 0.2653** 0.2768**
(2.36) (2.44)
QIGdNws 0.1339 0.1581
(1.10) (1.28)
QIBdNws -0.1609 -0.1262
(-1.13) (-0.87)
CONTROLS YES YES YES

Nobs 794317 794317 794317

Adj R2 6.12% 6.14% 6.19%

38
 Table 4: Robustness Checks

Table 4 reports robustness checks of calendar-time portfolio tests. Panel A reports an alternative method of
portfolio construction for returns and news sentiment scores. Panel B excludes stocks with prices below than
$ 5 per share. Panel C eliminates stocks with news observations on the holding period. Panel D measures
news tone by different sentiment tools including Loughran and McDonald (2011) negative words and positive
words, and the Google Natural Language Sentiment scores. Panel E reports portfolio performance on a weekly
basis. t-statistics are reported in parentheses and *, **, *** refers to the 10%, 5%, and 1% significance levels
respectively.

Panel A SI News t-stat QI News t-stat SMQ t-stat

News Beta 1.28%*** (5.16) 0.80%** (2.27) 0.49%** (2.36)
Independent double-sorted 1.00%*** (5.80) -0.40%** (-2.22) 1.41%*** (4.33)

Panel B SI News t-stat QI News t-stat SMQ t-stat

Price>= $5 0.71%*** (4.81) -0.16% (-1.08) 0.87%*** (3.22)

Panel C SI News t-stat QI News t-stat SMQ t-stat

Excluded subsequent news 1.45%*** (4.59) -0.98%*** (-3.07) 2.43%*** (5.07)

Panel D SI News t-stat QI News t-stat SMQ t-stat

LM negative score 0.99%*** (5.47) -0.49%*** (-2.68) 1.47%*** (4.43)
LM positive score 0.74%*** (4.22) -0.69%*** (-3.65) 1.43%*** (4.23)
Google NL score 0.78%*** (4.53) -0.63%*** (-3.32) 1.41%*** (4.30)

Panel E SI News t-stat QI News t-stat SMQ t-stat

Weekly frequency 2.09%*** (12.32) -0.79%*** (-4.47) 2.88%*** (9.71)

39
 Table 5: slowly incorporated and quickly incorporated News over Time

Table 5 reports the subsample analysis. The sample is split into four different periods: 1979 - 1987, 1988 -
1997, 1998 - 2007, 2008 - 2016 respectively. All returns are reported and converted into monthly averages in
percentages. Panel A reports the SI and QI news portfolio returns on monthly. The definition of SI and QI
presents as follows:

SI = LRHS − HRLS
QI = LRLS − HRHS

where SI news is a long/short portfolio of buying slowly incorporated good news (LRHS) and selling slowly
incorporated bad news (HRLS). QI news is a long/short portfolio of buying quickly incorporated good news
(HRHS) and selling quickly incorporated bad news (LRLS). t-statistics are reported in parentheses and *,
**, *** refers to the 10%, 5%, and 1% significance levels respectively.

1979 - 1987 1988 - 1997 1998 - 2007 2008 - 2016

SI News 0.93%*** 1.08%*** 1.32%*** 0.62%**
t-stat (3.06) (4.75) (2.77) (2.20)
QI News 0.14% -0.53%** -0.51% -0.58%*
t-stat (0.47) (-1.97) (-1.03) (-1.65)
Slow-Minus-Quick 0.79% 1.61%*** 1.83%** 1.20%**
t-stat (1.61) (3.96) (1.97) (2.06)

40
 Table 6: The Long-run Performance of SI and QI News over Different Horizons

Table 6 reports the performance of slowly incorporated news and quickly incorporated news over different
holding periods. The portfolio is constructed by sorting stocks into tercile portfolios over five different holding
periods: one month, three months, six months, nine months and twelve months. In each month, firms are
sorted into three groups based on their past abnormal returns, and then within each group, stocks are further
ranked into three groups based on their news sentiment scores. Abnormal return is the DGTW-adjusted
return computed by following (Daniel et al., 1997). The definition of SI and QI presents as follows:

SI = LRHS − HRLS
QI = LRLS − HRHS

where SI news is a long/short portfolio of buying slowly incorporated good news (LRHS) and selling slowly
incorporated bad news (HRLS). QI news is a long/short portfolio of buying quickly incorporated good news
(HRHS) and selling quickly incorporated bad news (LRLS). The sample period ranges from 1979 to 2016.
t-statistics are reported in parentheses and *, **, *** refers to the 10%, 5%, and 1% significance levels
respectively.

Months SI News t-stat QI News t-stat Slow-Minus-Quick t-stat

[1,2] 1.01%*** (5.85) -0.39%** (-2.09) 1.39%** (2.43)
[2,4] -0.03% (-0.24) 0.20% (1.54) -0.22% (-1.01)
[2,7] -0.07% (-0.83) 0.20%* (1.93) -0.28% (-1.64)
[2,10] -0.06% (-0.83) 0.17%* (1.79) -0.23% (-1.53)
[2,13] -0.09% (-1.30) 0.14%* (1.76) -0.22%* (-1.90)

41
Table 7: The Sensitivity of Slow-Minus-Quick Trading Strategy Returns to Trading Cost
Assumptions

Table 7 shows estimates of the trading strategy’s profitability based on the Slow-minus-Quick portfolio after
considering transaction costs. We recalculate the average monthly profitability for 10 alternative assumptions
about an investor’s round-trip transaction costs, namely: 10, 20, 30 ... , and 100 basis points (bps) per trade.
The abnormal return is based on the full sample raw return of the SMQ trading portfolio after regressing
on the Fama-French five factors.

Trading Costs (bps) Abnormal Returns (%) Raw Returns (%)

0 1.14% 1.39%
10 0.99% 1.24%
20 0.83% 1.08%
30 0.68% 0.93%
40 0.53% 0.78%
50 0.38% 0.63%
60 0.22% 0.47%
70 0.07% 0.32%
80 -0.08% 0.17%
90 -0.23% 0.02%
100 -0.39% -0.14%

42
 Table 8: SI and QI News under Different Information Environment

Table 8 reports the performance of slowly incorporated news and quickly incorporated news under different information environments. In the panel A,
we independently sort all stocks into two portfolios based on their most recent market capitalization (Size), current monthly change of media coverage,
turnover ratio and analyst coverage. We report the ”Small-Large” SIZE, ”Low-High” ∆ MEDIA, ”Low-High”TURN, ”Low-High”AstCvg spread
profitability in the post-formation period. In Panel B, we independently sort all stocks into two portfolios based on their current investor attention
index. We report the ”Low-High”Google ASVI ATTN, ”Low-High”Bloomberg AIA ATTN spread profitability in the following ten trading days. The
Google ASVI is the Google Search Volume Index, which is a proxy for retail investor attention and Bloomberg AIA measures the Bloomberg users
news reading activity, which represents institutional investor attention. The panel B coefficients are ten-trading-day cumulative returns in percentages.
t-statistics are reported in parentheses and *, **, *** refers to the 10%, 5%, and 1% significance levels respectively.

Panel A Size Subsamples MEDIA Subsamples

SmallSIZE BigSIZE Small-Big LowMEDIA HighMEDIA Low-High
SI News 1.61%*** 0.50%*** 1.11%*** 1.44%*** 0.09% 1.35%***
(6.58) (3.04) (4.54) (7.52) (0.54) (7.10)
QI News -0.61%* -0.18% -0.43% -0.64%*** 0.22% -0.86%***
(-1.92) (-1.07) (-1.50) (-3.18) (1.41) (-4.55)
Slow-Minus-Quick 2.22%*** 0.67%** 1.55%*** 2.08%*** -0.13%*** 2.21%***
(4.67) (2.28) (3.65) (6.00) (-0.48) (7.56)
43

Panel A - Continued TURN Subsamples AstCvg Subsamples

LowTURN HighTURN Low-High LowAstCvg HighAstCvg Low-High
SI News 1.36%*** 0.56%** 0.80%*** 1.03%*** 0.57%*** 0.46%**
(8.94) (2.03) (2.95) (4.53) (3.06) (1.99)
QI News -0.34%** -0.29% -0.04% -0.08% -0.20% 0.12%
(-2.12) (-1.01) (-0.16) (-0.33) (-1.04) (0.50)
Slow-Minus-Quick 1.70%*** 0.85%* 0.85%* 1.11%*** 0.77%** 0.34%
(6.79) (1.71) (1.87) (2.76) (2.33) (0.94)

Panel B Google ASVI ATTN Subsamples Bloomberg AIA ATTN Subsamples

LowATTN HighATTN Low - High LowATTN HighATTN Low - High
SINws 0.05%*** 0.01%** 0.04%*** 0.02%** 0.00% 0.01%
(6.23) (2.22) (4.44) (2.24) (0.28) (0.85)
QINws 0.00% 0.01% -0.01% 0.00% -0.01% 0.00%
(-0.15) (1.22) (-1.12) (-0.17) (-0.83) (0.35)
Slow-Minus-Quick 0.05%*** 0.00% 0.05%*** 0.02%* 0.01% 0.01%
(4.22) (0.42) (3.92) (1.72) (0.74) (0.40)
 Table 9: SI and QI News in Different Nature of News

Table 9 reports the performance of slowly incorporated news and quickly incorporated news for different news characteristics. In Panel A, we
independently sort all stocks into two portfolios based on textual complexity. In Panel B, we again sort all stocks into two portfolios based on news
informativeness. We report the “Ambiguous-Accurate” Tone, “Complex-Concise” Readability, “Short-Long” Length, and “EarningsEx-EarningsIn”
Topic spread profitability in the post-formation period. t-statistics are reported in parentheses and *, **, *** refers to the 10%, 5%, and 1% significance
levels respectively.

Panel A SI and QI News across Accuracy Subsamples SI and QI News across Readability Subsamples
Ambiguous Accurate Ambiguous - Accurate Complex Concise Complex - Concise
SI News 1.05%*** 0.42% 0.63%** 1.10%*** 0.81%*** 0.30%
(3.95) (1.61) (2.29) (5.34) (3.97) (1.30)
44

QI News -0.51%** 0.02% -0.53%* -0.67%*** -0.32% -0.34%*

(-2.08) (0.07) (-1.91) (-2.93) (-1.54) (-1.68)
Slow-Minus-Quick 1.55%*** 0.40% 1.15%*** 1.77%*** 1.13%*** 0.64%**
(3.37) (0.83) (2.55) (4.77) (3.29) (2.11)

Panel B SI and QI News across Length Subsamples SI and QI News across Earnings Subsamples
Short Long Short - Long EarningsEx EarningsIn EarningsEx - EarningsIn
SI News 1.29%*** 0.70%*** 0.58%*** 0.92%*** 0.77%*** 0.16%
(5.78) (3.62) (2.70) (5.33) (3.61) (1.04)
QI News -0.79%*** -0.19% -0.60%*** -0.45%** -0.27% -0.18%
(-3.53) (-0.81) (-2.60) (-2.35) (-1.36) (-1.09)
Slow-Minus-Quick 2.08%*** 0.90%** 1.18%*** 1.38%*** 1.04%*** 0.34%*
(5.40) (2.45) (3.74) (4.10) (2.90) (1.65)