=Paper=
{{Paper
|id=Vol-2104/paper_210
|storemode=property
|title=The Predictive Power of the Sentiment of Financial Reports
|pdfUrl=https://ceur-ws.org/Vol-2104/paper_210.pdf
|volume=Vol-2104
|authors=Jan-Hendrik Meier,Walid Esmatyar,Rasmus Frost
|dblpUrl=https://dblp.org/rec/conf/icteri/MeierEF18
}}
==The Predictive Power of the Sentiment of Financial Reports==
https://ceur-ws.org/Vol-2104/paper_210.pdf
The predictive power of the sentiment of financial reports
Jan-Hendrik Meier1[0000-0002-3080-2210], Walid Esmatyar1, and Rasmus Frost1
1
Kiel University of Applied Sciences, Sokratesplatz 2, 24149 Kiel, Germany
jan-hendrik.meier@fh-kiel.de
Abstract. The present study examines the predictive power of the tone or sen-
timent of 10-K annual and 10-Q quarterly financial statements for future corpo-
rate development. The sentiment indicator was calculated using word lists de-
veloped for financial texts by Loughran and McDonalds [23] and Henry [14]
and applying a conventional and a tf-idf weighted word count. The results show
that the sentiment indicator is of significant incremental prognostic quality both
for the next quarter and the quarter following it. Unlike suggested by previous
literature, neither the scope and content of the word lists nor the weighting
method applied had a significant influence on forecasting quality.
Keywords: Sentiment Analysis, Sentiment, Text Mining, Text Analysis, Pre-
diction, Tone, tf-idf.
1 Introduction
Both professional financial analysts and interested private investors have been able to
develop their skills in analyzing financial statements in recent years. Sophisticated
computer-aided analysis tools help them to process increasing amounts of data effi-
ciently and free from bias to create decision-relevant information. Of interest in this
context is that the technical solutions almost exclusively focus on quantitative finan-
cial data and figures, while qualitative financial data, mainly texts, are hardly consid-
ered in the analyses. The reason is that figures are considered to be more reliable and
less prone to manipulation than textual data because of their better verifiability; in
contrast, texts offer more room for interpretation.
Numerous studies have shown that the texts of annual financial statements contain
important information that cannot be obtained directly from key figures. While com-
plex verbal statements – despite all advances in the field of natural language pro-
cessing – cannot yet be adequately interpreted and evaluated by applying a software,
implicit information in the tone of printed texts can be detected and evaluated by
means of sentiment analysis. The present analysis deals with the question of whether
a positive or negative tone in annual and quarterly financial statements can be used to
forecast future earnings figures, thus allowing insight into a company’s future devel-
opment. In order answer this question, a sentiment analysis based on a total of 19,390
annual and quarterly financial statements of the companies listed in the Standard &
Poor's 500 Index (S&P 500) for the years 2005 to 2015 was conducted. The results
show that the sentiment indicators for annual and quarterly financial statements pro-
vide a highly significant incremental contribution to predicting the return on assets
(ROA) for the next two quarters. The present study provides empirical evidence on
the basis of a very broad random sample and confirms its findings by using two estab-
lished text-mining methods and comparing their results. The study also assesses the
extent to which the scope of the word lists used and the weighting method influence
forecasting quality.
Section 2 of this study presents a literature review and provides an overview of the
current state of research. Section 3 describes the methodology used and describes the
sample. Section 4 presents and discusses the results. This study concludes with a
summary.
2 Literature Review and Current State of Research
Two different methods dominate the field of sentiment analysis in empirical account-
ing and capital market research. Early research used test subjects or the researchers’
own subjective assessment to categorize the tone of company publications as positive
or negative [11, 16, 19]. Core [5] points out that the work load necessary for this
manual approach is too high to evaluate sufficiently large samples, and notes that a
computer-assisted evaluation of the texts can help to reduce the workload and to in-
crease the analyses’ accuracy and objectivity. Regarding computer-assisted evalua-
tion, two approaches can be identified in the literature, a fully-automated and a semi-
automated approach [20]. The statistical approach uses machine learning and is ap-
plied among others by Li [21], who uses information from management discussion
and analysis sections (MD&A) in 10-K annual and 10-Q quarterly reports. The author
evaluates 30,000 randomly selected sentences and subjectively determines if their
tone is positive, negative, neutral or uncertain. The words of this training data set are
then used to train a naive Bayesian algorithm to determine the largest statistical corre-
lation of the remaining texts with one of the categories. However, research using text
analysis prefers a second method, which employs predefined word lists which deter-
mine for individual words whether they imply a positive or negative tone. In this
method, the number of positive and / or negative words in each text is counted, and it
is assumed that the order of the words is irrelevant for the tone [22, 24].
Four different word lists have been established in accounting and capital market re-
search by the following authors: Osgood [30], Hart [13], Henry [14], and Loughran
and McDonald [24]. While word lists by Loughran/McDonald and Henry were devel-
oped specifically for financial texts, the word lists by Osgood and Hart were original-
ly designed for research in psychology and, respectively, in political communication.
Empirical accounting and capital market research primarily analyzes the tone of
10-K annual financial statements, 10-Q quarterly statements, company press releases,
articles in print media, articles on Internet platforms, and other company publications
[18]. The sentiment of these publications is used to predict the development of com-
panies, that is, either the price development of their securities or the development of
assets, earnings, and financial ratios.
Tetlock [33] based his analysis on the word list by Osgood and showed that a pes-
simistic tone in press releases results in a negative investor reaction and thus a de-
crease in stock returns. The press releases were taken from the Wall Street Journal
column "Abreast of the Market". A number of other studies have also confirmed Tet-
lock's results with respect to short-term market responses to tone [1, 3, 8, 9, 26]. In
addition, both Solomon [32] and Huang, Teoh and Zhang [17] conclude that this ef-
fect is reversed over time: an initial reaction normalizes over time and is therefore to
be considered as a market overreaction. Garcia [12] points out that a negative tone in
newspaper articles results in a much stronger stock-market reaction when the econo-
my is in recession.
The two profit figures regularly forecasted in empirical studies are the return on as-
sets and the unexpected earnings. The latter figure is calculated as the difference be-
tween earnings per share and the average of earnings per share expected by analysts.
Tetlock, Saar-Tsechansky and Macskassy [34] come to the conclusion that a negative
tone in press releases causes negative earnings in the following quarters. Other studies
arrive at similar results based on the tone in 10-K annual financial statements or of
social media posts [3, 9]. Loughran and McDonald [22] on the other hand find that a
negative tone in 10-K financial statements follows a positive development of unex-
pected earnings. A distortion of the results due to the applied word list can be ruled
out since both results were obtained using the Osgood as well as the Loughran and
McDonald word list. Thus, the studies differ mainly due to the analyzed text types.
Thus, Loughran and McDonald [22] conclude with reference to the results of Tetlock,
Saar-Tsechansky and Macskassy [34] that due to the independency of journalists the
tone of newspaper articles is a better indicator of future profits, while managers of
companies can use 10-K annual financial statements to pursue impression or expecta-
tion management.
Davis, Piger and Sedor [6] show based on Hart's word list that the tone of corporate
earnings announcements in the press is a good predictor of the company's return on
assets for the next four quarters. Furthermore, they present evidence that the stock
market reacts directly to the sentiment of these communications. The authors argue
that managers can report their own earnings expectations more subtly by tone choice
than by simply presenting figures. Using tone, managers provide the capital market
with important signals.
Davis and Tama-Sweet [7] investigate 10-K and 10-Q quarterly statements in addi-
tion to press releases and confirm Davis, Piger and Sedor's results based on their ex-
panded sample. In addition, the authors find that managers’ tone choice is with signif-
icant frequency more optimistic and less pessimistic in earnings announcements in the
press than in the MD&A sections of financial statements. The authors conclude that
managers use the freedom afforded to them by press releases to strategically influence
the stock market.
In their analysis, Ferris, Hao and Liao [10] evaluate securities prospectuses of
companies in the technology sector that are about to go public. They show a forecast
of the return on assets for a period of three years into the future to be possible with
this information. Huang, Teoh and Zhang present opposite findings for the predictive
capacity of press sentiment for the return on assets. According to their results, the
return on assets shows a negative sign for three years if the press sentiment was pre-
viously positive. Davis, Piger and Sedor [6] and Davis and Tama-Sweet [7] use the
Hart word list in their investigations, whereas Huang, Teoh and Zhang [17] apply the
other three word lists.
Against the background of the inconsistency in the results of previous research, the
present analysis aims at investigating the predictive power of the sentiment in 10-K
and 10-Q financial statements for the return on assets on the basis of a representative
sample. For this purpose, two different word lists were used to assure that the results
are not influenced by word-list choice. The lists by Loughran/McDonald and Henry
were applied since they were designed for the analysis of financial texts.
3 Sample and Method
In order to analyze the predictivity of the sentiment in financial reports, a sentiment
analysis based on English-language 10-K annually and 10-Q quarterly US financial
statements was conducted and the predictivity of the sentiment on the company's per-
formance in terms of return on assets (ROA) was examined. The analysis was based
on the 10-K and 10-Q filings of all companies listed in the Standard & Poor's 500
Index (S&P 500) as of March 31, 2016, i.e., the 500 largest listed companies in the
United States. For these companies, data from 2005 to 2015 were retrieved. The data
were obtained from the Securities and Exchange Commission's (SEC) Electronic File
Gathering, Analysis and Retrieval System (EDGAR). The initial sample contained a
total of 20,435 annually and quarterly financial statements. The size of the sample
was reduced by 1,045 to 19,390 documents as not all relevant financial data pertain-
ing to an annual report were available, which was partly due to incomplete financial
years or quarters. Associated financial data were retrieved from the Thomson Reuters
Eikon database and randomly tested using annual report data.
The preparation of the texts for processing was based on the so-called parsing pro-
cedure by Loughran and McDonald [27, 28]. Since the documents were partly HTML
documents, HTML tags were removed [24]. Likewise, tables and figures were re-
moved with the exception of text tables.1 Furthermore, the texts were cleared of fre-
quently recurring words and words not related to the content (stopwords) [27]. In a
final step, numbers, special characters, and single letters were removed.
As mentioned above, this study used the word list by Loughran and McDonald and
– to countercheck the results – the much shorter list by Henry. The Loughran and
McDonald word list contains a total of six word categories for positive, negative,
uncertain, litigious, strongly modal, and weakly modal expressions [22]. This study
uses only the 355 positive and 2,355 negative words of this list. The word list by Hen-
ry contains a total of 105 positive and 85 negative words [14]. Loughran and McDon-
1
As in Loughran and McDonald [27, 28], tables embedded in HTML Code are analyzed to
ascertain if the proportion of numerals contained in them is greater than 15% of the total of
all numerals and letters. If this is the case, the entire table, including its content, is deleted
since it is assumed that it only presents financial-reporting figures and, therefore, is irrele-
vant for textual analysis.
ald [24] argue that managers have far more ways of implying positive or negative
sentiment than can be captured by Henry's word list. In particular, Henry's list lacks
entries such as loss, losses, adverse and impairment, all of which are considered nega-
tive in the corporate context.
The relevance of a positive or negative word for a certain text can be determined
unweighted, that is, by ordinary counting of the words, or by so-called term frequency
– inverse document frequency (tf-idf) weighting. The latter attributes more influence
to a positive or negative word if it occurs frequently in the document (term frequency,
tf), but lessens its influence if the word occurs in many documents and, as a result,
appears "ordinary" (inverse document frequency, idf). The tf-idf weighting logic used
in this study follows Chrisholm and Kolda [4]. The word frequency tfw,d is calculated
as follows [4, 22, 24]:
,
, = ( )
(1)
Here, tf´w,d reflects the frequency of word w in document d, while ad represents the
average word count in document d. The inverse document frequency idfw is calculated
based on the following equation:
= (2)
represents the total number of documents and dfw the number of documents in
which the word w occurs at least once. The tf-idf weight results from the multiplica-
tion of both terms, tfw,d∙idfw. The resulting weight measure is then used as the count
for the respective word, while in the unweighted approach each occurrence of a word
is counted equally.
The tf-idf weighting has been criticized by Henry and Leone [15] because its re-
sults are contingent upon the totality of documents used in a sample, and, thus, the
result for a particular document can vary significantly depending on the other docu-
ments included in the sample. The authors showed that the weighting does not lead to
improved results, thus refuting the findings of Loughran and McDonald [22], who
observed a significant improvement in the results if tf-idf weighting was applied. Due
to contradictory results, this study applied both methods, the unweighted count and
the tf-idf weighting, in parallel and tests them against each other.
Since for every company i only one 10-K annual or 10-quarter financial statement
is available in each quarter t, the document index d can clearly be replaced by the
index i,t. To determine the sentiment of each quarter or annual document, the senti-
ment indicator SENTi,t or, respectively, SENTd can be calculated according to the
following equation:
, , ,, ,,
, = = (3)
, , ,, ,,
tf´pos,d denotes the number of weighted or, respectively, unweighted positive words
in the document d, tf´neg,d denotes the number of weighted or, respectively, un-
weighted negative words [15]. Accordingly, the sentiment indicator can have a value
between 1 for a perfectly positive sentiment and -1 for a perfectly negative sentiment.
This study analyzed the predictivity of the sentiment for the return on assets; there-
fore, the return on assets ROAi,t+n with a lag of n quarters was considered as a de-
pendent variable in the econometrics. The return on assets ROAi,t+n was calculated
based on earnings before interest and taxes (EBIT) and the average assets of the ob-
served quarter.
The control variables used in this study were based on the relevant literature [6, 7,
17]. The natural logarithm of the market capitalization at the time t was used as an
indicator for the size of the company SIZEi,t. The correction for company size is ne-
cessitated due to the small firm effect according to Banz [2]. In order to control for
expected growth as well as other key balance sheet figures and company news, the
market to book ratio MBRi,t was used as an indicator that incorporates a variety of
publicly known information and news. To take the leverage effect and the findings of
Modigliani and Miller [29] into account, the leverage ratio LEVi,t is also used as a
control variable in the model. Since the return on assets ROAi,t is a good indicator of
the return on assets in the following quarters ROAi,t+n, it is also used a control varia-
ble. Finally, the Boolean variables DLOSSi,t for loss quarters and D10Ki,t for annual
reports are integrated into the model. They compensate for the effects of differing
company responses in loss quarters as well as the potential differences in the predic-
tivity of annual financial statements compared to quarterly financial statements. To
avoid outlier-induced results, all control variables were winsorized at the first and
99th percentile. For an overview of the variables used please refer to Table 1.
Table 1. Variables and data sources
Variable Expression Thompson Reuters Codes
ROAi,t , TR.EBIT
Explained
, , TR.TotalAssetsReported
Variable
SEN- ,, ,,
Ti,t
Explanatory
,, ,,
Variable
SIZEi,t ln( , ) TR.CompanyMarketCap
Control Va-
riables
MBRi,t , TR.CompanyMarketCap
, TR.TotalEquity
LEVi,t , TR.TotalLiabilities
, TR.TotalAssetsReported
DLOS 1, if Net Incomei,t < 0 TR.NetIncome
Si,t 0, else
D10Ki 1, if 10 K-Filing
,t 0, if 10 Q-Filing
In the analysis, the word lists by Loughran and McDonald and Henry were each
used both with and without tf-idf weighting to predict the return on assets for one and
two quarters. This results in four explainatory variables. While the list of Loughran
and McDonald returns negative sentiment indicators in most cases, the list of Henry
returns almost positive sentiments. This effect can be attributed to the strong deviation
of the two lists. All other variables show inconspicuous behavior.
As in the analysis, the word lists by Loughran and McDonald and Henry were each
used both with and without tf-idf weighting to predict the return on assets for one and
two quarters, this results in a total of eight variants to be calculated. The correspond-
ing pooled OLS, fixed-effects, and random-effects panel data models were calculated
for all variants. However, the F- and Hausmann specification tests used to compare
the models show the superiority of the fixed-effects model with fixed-effects for each
company and every quarter ; the model is applied according to the following
equation:
,+ = + + 1
∙ , + 2
∙ , + 3
∙ , + 4
∙ , +
5
∙ , + 6
∙ , + 7
∙ 10 , + , (3)
In addition to the panel data models, all variants were calculated by means of a ro-
bust least absolute deviations (LAD) regression in order to rule out outlier-induced
results. However, the panel data models and LAD regression were not significantly
different from each other; thus, the validity of the panel data models can be assumed.
Multicollinearity problems were excluded by the analysis of the variance inflation
factors that all showed numbers below three. All eight variants were tested for hetero-
scedasticity using the Breusch-Pagan test. All variants show heteroskedasticity prob-
lems. Therefore, the results shown in Table 1 include the heteroskedasticity-robust
standard errors according to White (HC 0) and the significance values calculated on
their basis. Endogeneity problems were precluded due to the time lag between the
explained variable and explanatory variables and, therefore, no separate consideration
is required.
4 Results and Discussion
The descriptive statistics and the correlation matrix are presented in Table 2 and Ta-
ble 3 respectively. The results of the analysis are summarized in Table 4 for one quar-
ter (Lag 1) and Table 5 for two quarters (Lag 2). For the purpose of clarity, individual
fix effects, which do not show any abnormalities, have been omitted. The same ap-
plies to the fixed quarterly effects, which, albeit particularly significant in financial
crisis quarters, do not show abnormalities.
All models show significant model statistics, which implies a high prediction char-
acter. The models with Lag 1 have an explained variance R2 of – depending on the
model – around 20%, which can be considered as high in this type of models. The
Lag 2 models are naturally weaker in terms of the explained variance as forecasts for
the distant future are generally associated with weaker prediction qualities. Accord-
ingly, experiments with higher lags only sporadically show significant models, which
is why they are not presented here.
The sentiment indicator has a significant positive impact in all tested models.
Therefore, it can be used as a predictor for future corporate development forecasts. It
is important to underline that in the present analysis the sentiment indicator was cor-
rected for the influence of several significant indicators. In particular, the market-to-
book ratio already contains the aggregated information of a large number of balance
sheet ratios and company news, which are incorporated in the price by the market.
The sentiment indicator thus contains incremental information beyond the information
imparted by balance sheet figures and thereby provides additional forecasting power.
This implies that companies use texts to impart information about positive or negative
developments which cannot yet be expressed in figures.
For sentiment analysis, it does not seem to matter whether the extensive word list
by Loughran and McDonald or the much shorter list by Henry is used. The quality
measures of both word lists seem to differ only marginally. It also appears to be irrel-
evant if the sentiment indicator is determined by mere counting or by applying the tf-
idf weighting. The advantages of tf-idf weighting in other areas of text analysis, such
as the creation of word clouds for faster collection of content, can hardly be trans-
ferred to sentiment analysis. Although the method is capable of identifying relevant
words in documents, this advantage is mainly limited to nouns that are frequent in
individual documents but rare in other documents. Adjectives and adverbs, which
often carry a positive or negative connotation in the corporate publications in ques-
tion, tend to be underweighted by the method because they are commonplace in eve-
ryday language.
Table 2. Descriptive Statistics
Explanatory Variables
Loughran/McDonald (tf-idf) Loughran/McDonald (unweighted)
positive negative SENTIi,t positive negative SENTIi,t
Min. - - - 0.864 - - - 0.888
1st Qu. 169.400 371.900 - 0.511 80.000 172.000 - 0.518
Median 236.700 602.800 - 0.426 133.000 341.000 - 0.405
Mean 269.200 688.900 - 0.405 188.100 487.700 - 0.390
3rd Qu. 350.300 945.200 - 0.319 248.000 647.000 - 0.278
Max. 880.200 2,506.300 0.567 2,620.000 5,347.000 0.619
Henry (tf-idf) Henry (unweighted)
positive negative SENTIi,t positive negative SENTIi,t
Min. - - - 0.745 - - 0.862
1st Qu. 243.300 178.500 0.085 163.000 119.000 0.025
Median 305.100 229.100 0.154 251.000 192.000 0.136
Mean 326.000 237.300 0.154 336.700 266.600 0.133
3rd Qu. 402.500 293.400 0.224 433.000 331.500 0.245
Max. 745.400 522.900 0.688 3,216.000 2,666.000 0.843
Control-Variables
SIZEi,t MBRi,t LEVi,t ROAi,t DLOSSi,t D10Ki,t
Min. 18.920 0.501 0.129 - 0.019 - -
1st Qu. 22.540 1.606 0.471 0.013 - -
Median 23.190 2.595 0.605 0.024 - -
Mean 23.300 3.649 0.600 0.028 0.077 0.238
3rd Qu. 23.930 4.215 0.741 0.038 - -
Max. 27.310 25.220 0.959 0.100 1.000 1.000
Table 3. Correlation Matrix
Explaine d Variables Explanatory Variables (SENT i,t) Control Variables
ROAi,t+1 ROA i,t+2 LMD (tf.idf) LMD Henry (tf-idf) Henry SIZEi,t MBRi,t LEVi,t ROA i,t DLOSSi,t D10K i,t
Explained ROAi,t+1 1.000
Variables ROAi,t+2 0.706 1.000
LMD (tf.idf) 0.133 0.126 1.000
Explanatory
LMD 0.127 0.122 0.909 1.000
Variables
He nry (tf-idf) 0.164 0.170 0.523 0.466 1.000
(SENTi,t)
He nry 0.260 0.252 0.408 0.377 0.771 1.000
SIZEi,t 0.108 0.098 - 0.061 - 0.058 - 0.001 0.066 1.000
MBRi,t 0.368 0.368 0.115 0.106 0.116 0.203 0.094 1.000
Control LEVi,t - 0.337 - 0.338 - 0.076 - 0.061 - 0.132 - 0.210 0.034 0.104 1.000
Variables ROAi,t 0.732 0.654 0.160 0.158 0.192 0.297 0.117 0.402 - 0.399 1.000
DLOSSi,t - 0.241 - 0.183 - 0.095 - 0.117 - 0.101 - 0.135 - 0.145 - 0.010 0.094 - 0.323 1.000
D10K i,t - 0.038 0.016 0.002 0.060 0.230 0.045 - 0.006 - 0.006 - 0.002 0.006 0.058 1.000
Table 4. Results of fixed-effects panel data models with heteroskedasticity-robust standard
errors (White's HC 0) for lag 1
Loughran/McDonald (tf-idf) Loughran/McDonald (unwe ighte d)
Coe ff. Std.Error t-Value p-Value Coe ff. Std.Error t-Value p-Value
SENTi,t 0.005 0.002 2.743 0.006 ** 0.006 0.001 3.840 0.000 ***
SIZEi,t 0.005 0.001 4.374 0.000 *** 0.005 0.001 4.339 0.000 ***
MBRi,t 0.001 0.000 5.936 0.000 *** 0.001 0.000 5.924 0.000 ***
LEVi,t - 0.006 0.005 - 1.202 0.229 - 0.006 0.005 - 1.162 0.245
ROAi,t 0.403 0.049 8.272 0.000 *** 0.402 0.049 8.240 0.000 ***
DLOSSi,t - 0.002 0.001 - 2.647 0.008 ** - 0.002 0.001 - 2.533 0.011 *
D10Ki,t - 0.004 0.002 - 1.895 0.058 . - 0.004 0.002 - 1.947 0.052 .
Obse rvations 19,390 19,390
2
R 0.198 0.199
2
Adjus te d R 0.175 0.176
F Statistic 91.2 0.000 *** 0.000 ***
(df = 51; 18844)
F-Te st OLS vs. FE (F) 8.5 0.000 *** 8.5 0.000 ***
(df = 494; 18844)
Hausmann FE vs . RE 2,655.9 0.000 *** 2,735.6 0.000 ***
(df = 51)
Breusch-Pagan (BP) 19,563.0 0.000 *** 19,490.0 0.000 ***
(df = 51)
Henry (tf-idf) Henry (unweighted)
Coe ff. Std.Error t-Value p-Value Coe ff. Std.Error t-Value p-Value
SENTi,t 0.011 0.002 4.631 0.000 *** 0.006 0.002 3.729 0.000 ***
SIZEi,t 0.004 0.001 4.137 0.000 *** 0.004 0.001 4.138 0.000 ***
MBRi,t 0.001 0.000 6.048 0.000 *** 0.001 0.000 5.981 0.000 ***
LEVi,t - 0.007 0.005 - 1.345 0.179 - 0.007 0.005 - 1.348 0.178
ROAi,t 0.399 0.048 8.246 0.000 *** 0.399 0.049 8.219 0.000 ***
DLOSSi,t - 0.002 0.001 - 2.683 0.007 ** - 0.002 0.001 - 2.699 0.007 **
D10Ki,t - 0.005 0.002 - 2.220 0.026 * - 0.004 0.002 - 1.937 0.053 .
Obse rvations 19,390 19,390
2
R 0.200 0.199
2
Adjus te d R 0.177 0.176
F Statistic 92.2 0.000 *** 91.7 0.000 ***
(df = 51; 18844)
F-Te st OLS vs. FE (F) 8.5 0.000 *** 8.4 0.000 ***
(df = 494; 18844)
Hausmann FE vs . RE 2,706.5 0.000 *** 2,694.8 0.000 ***
(df = 51)
Breusch-Pagan (BP) 19,559.0 0.000 *** 19,449.0 0.000 ***
(df = 51)
Table 5. Results of fixed-effects panel data models with heteroskedasticity-robust standard
errors (White's HC 0) for lag 2
Loughran/McDonald (tf-idf) Loughran/McDonald (unwe ighte d)
Coe ff. Std.Error t-Value p-Value Coe ff. Std.Error t-Value p-Value
SENTi,t 0.005 0.002 2.630 0.009 ** 0.006 0.002 3.821 0.000 ***
SIZEi,t 0.005 0.001 3.893 0.000 *** 0.005 0.001 3.857 0.000 ***
MBRi,t 0.002 0.000 5.971 0.000 *** 0.002 0.000 5.957 0.000 ***
LEVi,t - 0.020 0.006 - 3.451 0.001 *** - 0.020 0.006 - 3.408 0.001 ***
ROAi,t 0.166 0.081 2.054 0.040 * 0.165 0.081 2.031 0.042 *
DLOSSi,t - 0.000 0.001 - 0.063 0.950 0.000 0.001 0.018 0.985
D10Ki,t 0.001 0.001 1.444 0.149 0.001 0.001 1.311 0.190
Obse rvations 19,390 19,390
2
R 0.107 0.107
2
Adjus te d R 0.081 0.082
F Statistic (F) 44.1 0.000 *** 44.5 0.000 ***
(df = 51; 18844)
F-Te st OLS vs. FE (F) 11.6 0.000 *** 11.6 0.000 ***
(df = 494; 18844)
Hausmann FE vs . RE 5,650.7 0.000 *** 5,705.0 0.000 ***
(df = 51)
Breusch-Pagan (BP) 21,713.0 0.000 *** 21,677.0 0.000 ***
(df = 51)
Henry (tf-idf) Henry (unweighted)
Coe ff. Std.Error t-Value p-Value Coe ff. Std.Error t-Value p-Value
SENTi,t 0.015 0.003 5.264 0.000 *** 0.009 0.002 4.802 0.000 ***
SIZEi,t 0.004 0.001 3.583 0.000 *** 0.004 0.001 3.538 0.000 ***
MBRi,t 0.002 0.000 6.066 0.000 *** 0.002 0.000 6.017 0.000 ***
LEVi,t - 0.021 0.006 - 3.602 0.000 *** - 0.021 0.006 - 3.620 0.000 ***
ROAi,t 0.161 0.081 1.985 0.047 * 0.159 0.081 1.966 0.049 *
DLOSSi,t - 0.000 0.001 - 0.025 0.980 - 0.000 0.001 - 0.036 0.971
D10Ki,t 0.000 0.001 0.401 0.688 0.001 0.001 1.269 0.204
Obse rvations 19,390 19,390
2
R 0.110 0.110
2
Adjus te d R 0.084 0.084
F Statistic (F) 45.7 0.000 *** 0.000 ***
(df = 51; 18844)
F-Te st OLS vs. FE (F) 11.7 0.000 *** 11.6 0.000 ***
(df = 494; 18844)
Hausmann FE vs . RE 5,689.3 0.000 *** 5,683.1 0.000 ***
(df = 51)
Breusch-Pagan (BP) 21,573.0 0.000 *** 21,503.0 0.000 ***
(df = 51)
5 Conclusions
The present study demonstrated the predictive power of the sentiment of 10-K annual
and 10-Q quarterly statements for future corporate development. The sentiment indi-
cator has significant incremental prognostic quality both for the following quarter and
the subsequent quarter. If the forecast horizon was three quarters or more, only isolat-
ed cases of significant predictiveness were detected. As a result, analysts and inves-
tors should include the sentiment of corporate publications into their analyses to gath-
er latent information from company and detect subtle signals from management that is
of value in their decision-making. Corporate publications contain valuable "between
the lines" information which may be relevant to the assessment of potential opportuni-
ties and risks. Accordingly, a professionalized evaluation of textual data can provide
an information advantage.
The present study drew on two very different word lists, the Loughran and
McDonald and the Henry word lists, which are both geared toward financial texts.
Both were used in conventional word-counting as well as tf-idf-weighting approaches.
However, the quality of the word lists used and the weighting approach taken influ-
enced forecasting quality to a much lesser extent than suggested by the existing litera-
ture. The calculated models were almost identical in terms of quality.
References
1. Allee, K. D., DeAngelis, M. D.: The Structure of Voluntary Disclosure Narratives: Evi-
dence from Tone Dispersion. Journal of Accounting Research, 53(2), 241-274 (2015).
2. Banz, R.: The relationship between return and market value of common stocks. Journal of
Financial Economics, 9 (1981), 3-18.
3. Chen, H., De, P., Hu, Y., Hwang, B.: Wisdom of Crowds: The Value of Stock Opinions
Transmitted Through Social Media. Review of Financial Studies, 27(5), 1367-1403
(2014).
4. Chrisholm, E., Kolda, T. G.: New Term Weighting Formulas for the Vector Space Model
in Information Retrieval. 1999. Oak Ridge, Tennessee.
5. Core, J. E.: A review of the empirical literature discussion. Journal of Accounting and
Economics, 31, 441-456 (2001).
6. Davis, A. K., Piger, J. M., Sedor, L. M.: Beyond the Numbers: Measuring the Information
Content of Earnings Press Release Language. Contemporary Accounting Research, 29(3),
845-868 (2012).
7. Davis, A. K., Tama-Sweet, I.: Managers’ Use of Language Across Alternative Disclosure
Outlets: Earnings Press Releases versus MD&A. Contemporary Accounting Research,
29(3), 804–837 (2012).
8. Dougal, C., Engelberg, J., Garcia, D., Parsons, C. A.: Journalists and the stock market. Re-
view of Financial Studies, 25(3), 639-679 (2012).
9. Feldman, R., Govindaraj, S., Livnat, J., Segal, B.: Management’s tone change, post earn-
ings announcement drift and accruals. Review of Accounting Studies, 15(9), 915-953
(2010).
10. Ferris, S. P., Hao, Q., Liao, M.: The Effect of Issuer Conservatism on IPO Pricing and Per-
formance. Review of Finance, 17, 993-1027 (2010).
11. Francis, J., Philbrick, D., Schipper, K.: Shareholder Litigation and Corporate Disclosures.
Journal of Accounting Research, 32(2), 137-164 (1994).
12. Garcia, D.: Sentiment during Recessions. The Journal of Finance, 68(3), 1267-1300
(2012).
13. Hart, R.P.: Redeveloping DICTION: Theoretical Considerations. In: West, M. D. (Eds.):
Theory, method, and practice in computer content analysis, New York, 43-60.
14. Henry, E.: Are Investors Influenced by How Earnings Press Releases are Written? Journal
of Business Communication, 45(4), 363-407 (2008).
15. Henry, E., Leone, A. J.: Measuring Qualitative Information in Capital Markets Research:
Comparison of Alternative Methodologies to Measure Disclosure Tone. The Accounting
Review, 91(1), 153-178 (2016).
16. Hoskin, R. E., Hughes, J. S., Ricks, W. E.: Evidence on the Incremental Information Con-
tent of Additional Firm Disclosures Made Concurrently with Earnings. Journal of Ac-
counting Research, 24, 1-32 (1986).
17. Huang, X., Teoh, S. H., Zhang, Y.: Tone Management. The Accounting Review, 89(3),
1083-1113 (2014).
18. Kearney, C., Liu, S.: Textual sentiment in finance: A survey of methods and models. In-
ternational Review of Financial Analysis, 33, 171-185 (2014).
19. Lang, M. H., Lundholm, R. J.: Voluntary Disclosure and Equity Offerings: Reducing In-
for-mation Asymmetry or Hyping the Stock?. Contemporary Accounting Research, 17(4),
623-662 (2000).
20. Li, F.: Textual Analysis of Corporate Disclosures: A Survey of the Literature. Journal of
Accounting Literature, 29, 143-165 (2010).
21. Li, F.: The Information Content of Forward-Looking Statements in Corporate Filings – A
Naïve Bayesian Machine Learning Approach. Journal of Accounting Research, 48(5),
1049-1102 (2010).
22. Loughran, T., McDonald, B.: When Is a Liability Not a Liability? Textual Analysis, Dic-
tion-aries, and 10-Ks. The Journal of Finance, 66(1), 35-65 (2011).
23. Loughran, T., McDonald, B.: 2014 Master Dictionary. 03.2015. As of: 01.11.2017.
URL: https://www3.nd.edu/~mcdonald/Word_Lists.html
24. Loughran, T., McDonald, B.: Textual Analysis in Accounting and Finance: A Survey.
Journal of Accounting Research, 54(4), 1187-1230 (2016).
25. Mayew, W. J., Sethuraman, M., Venkatachalam, M.: MD&A Disclosure and the Firm’s
Ability to Continue as a Going Concern. The Accounting Review, 90(4), 1621-1651
(2015).
26. Mayew, W. J., Venkatachalam, M.: The Power of Voice: Managerial Affective States and
Future Firm Performance. The Journal of Finance, 67(1), 1-43 (2012).
27. McDonald, B.: Documentation for Stage One Parse. Last updated: 03.09.2013. As of:
13.08.2017, URL: https://www3.nd.edu/~mcdonald/Word_Lists.html.
28. McDonald, B.: Documentation of Master Dictionary and Document Dictionaries. Last up-
dated: 18.03.2014. As of: 13.08.2017, URL:
https://www3.nd.edu/~mcdonald/Word_Lists.html.
29. Modigliani, F., Miller, M.: Corporate Income Taxes and the Cost of Capital: A Correction.
The American Economic Review, 53(1), 433-443 (1963).
30. N.N.:Descriptions of Inquirer Categories and Use of Inquirer Dictionaries. As of
01.11.2017. URL: http://www.wjh.harvard.edu/~inquirer/homecat.htm.
31. Price, S. M., Doran, J. S., Peterson, D. R., Bliss, B. A.: Earnings conference calls and stock
returns: The incremental informativeness of textual tone. Journal of Banking & Finance,
36, 992-1011 (2012).
32. Solomon, D. H.: Selective Publicity and Stock Prices. The Journal of Finance, 67(2), 599-
637 (2012).
33. Tetlock, P. C.: Giving Content to Investor Sentiment: The Role of Media in the Stock Mar-
ket. The Journal of Finance, 62(3), 1139-1168 (2007).
34. Tetlock, P. C., Saar-Tsechansky, M., Macskassy, S.: More Than Words: Quantifying Lan-
guage to Measure Firms’ Fundamentals. The Journal of Finance, 63(3), 1437-1467 (2008).
35. Twedt, B., Rees, L: Reading between the lines: An empirical examination of qualitative at-
tributes of financial analysts’ reports. Journal of Accounting and Public Policy, 31, 1-21
(2012).