Buy on bad news, sell on good news: How insider trading analysis can benefit from textual analysis of corporate disclosures
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
Buy on bad news, sell on good news: How insider trading analysis can
benefit from textual analysis of corporate disclosures
Michael Hagenau1, Adriana Korczak2, Dirk Neumann1
1
Universität Freiburg, Platz der Alten Synagoge,
79085 Freiburg, Germany
{michael.hagenau, dirk.neumann}@is.uni-freiburg.de
2
School of Economics, Finance and Management, University of Bristol, 8 Woodland Road,
Bristol BS8 1TN, United Kingdom
Adriana.Korczak@bristol.ac.uk
Abstract. We demonstrate how insider trading analysis may benefit from textual
analysis. We analyze reported insider trading behavior and explain the association
between corporate as well as 3rd-party news announcements on directors’
dealings activity. Previous approaches are extended by adding the sentiment of
news to the research setting. We find strong evidence that insiders follow the
stock market adage “Buy on bad news, sell on good news”: They tend to buy
(sell) securities in those years where their respective companies issue negative
(positive) news. Likewise, insiders tend to buy (sell) stocks in years when 3rd-
party news coverage is pessimistic (optimistic). The impact of corporate news on
insider trading is higher than for 3rd-party news, as corporate news are subject to
direct influence by the insiders. We also find that insiders buy when next year’s
news improves compared to the current year. Looking more concretely into the
language, we also demonstrate that insiders buy when expressing insecurity and
uncertainty. Overall, the findings reveal additional insights for insider trading
analysis and demonstrate how finance may benefit from textual analysis.
Keywords: text mining, sentiment, insider trading, tone of news, corporate news1 Introduction
Literature is providing evidence that the way corporate news announcements are phrased
plays an important role in the capital markets. Indisputably, news carries valuable qualitative
information that may influence investor’s expectations about firm’s future performance
(Tetlock et. al, 2008).
Prior financial literature on textual analysis often investigates the relationship between the
tone of news – in this context often dubbed ‘sentiment’ – and its relation to company
performance and capital market reactions such as stock price returns, trading volumes and
volatility (e.g. Tetlock et. al. 2008, Loughran & MacDonald 2011). Tetlock et al. find that
negative words from a psychosocial dictionary in corporate news can predict company’s
earnings and stock returns. Loughran & MacDonald extend these findings by tailoring their
word lists to financial news and also linking the news measure to trading volume, return
volatility and fraud.
These news announcements are formulated within the company by executives or outside the
company by third parties such as analysts. By law, executives are banned from trading based
on unpublished superior material information about the company that may significantly
impact the stock prices. However, there are neither clear definitions of material information
nor clear regulations of how long in advance before a piece of news is revealed for which
directors are not permitted to trade in their shares1. Regardless of whether the news about a
company comes from within or from outside a company, there are no strict or binding rules
and guidance on the linguistic context with which news need to be written. Each piece of
corporate news reveals an important fact about a company and the mix of words used
determines the strength and the tone of news. Thus, reader’s perception and view of the
company’s current position and future prospects are influenced. In companies where
executives are shareholders, executives may strategically use their discretion in formulating
news in order to influence the stock prices and subsequently exploit these stock price
movements by own trading activities. When news about a company is announced by a third
party, it impacts the perception of all investors including executives about the company’s
current position and future prospects. Accordingly, directors lose the insider advantage as
news is novel to all investors equally.
1 There are explicit 60 (30) days blackout periods before preliminary annual or semi-annual (quarterly) earnings
announcement defined by law in the UK, but not before any other news types (see Korczak et al. 2010). In the US, some
companies chose to introduce blackout periods and restrict directors trading before earnings announcementFrom the past literature, it is known that executives trade around different corporate news announcements such as takeover bids (Seyhun, 1990), seasoned equity offerings (Karpoff & Lee, 1991), bankruptcy filings (Seyhun & Bradley, 1997) or earnings announcements (Ke, Huddart & Petron, 2003). In this way, insiders manage to outperform the market by earning abnormal returns from their trades. Ke et al. (2003), Piotroski & Roulstone (2005) and Huddart & Ke (2007) show that executives trade based on their superior knowledge about company’s prospects long before the news is announced. Ke et al. (2003) document that executives trade based on their superior knowledge about companies’ future accounting performance up to two years ahead of public announcement of that news. Likewise, Piotroski & Roulstone (2005) show that insider trading patterns can be explained by executives’ foreknowledge about next year’s results. Korczak, Korczak & Lasfer (2010) also show that insiders time their trades around different types of news to benefit from their superior knowledge about the company. At the same time, however, insiders are cautious to avoid potential litigation problems and a loss of reputation. Other studies find evidence that actively trading executives not only benefit from their superior knowledge, but also actively influence corporate disclosures. Aboody & Kasznik, (2000) find that executives tend to release bad news before stock option grants to fix lower strike prices. Vice versa, executives release very positive news before stock option exercises to obtain higher sales prices (Brockman, Martin & Puckett 2010). Similarly, executives who intend to buy shares for their own account tend to release bad news in the period just before their insider purchases (Cheng & Lo 2006). This insider trading pattern could appear to be another example of contrarian trading, but it also could represent active manipulation of stock market prices. Our paper extends literature in four ways: First, while ‘good’ and ‘bad’ news have been used as explanatory variable for insider trading, ‘good’ or ‘bad’ was always measured by the sign of abnormal returns around the news. However, the actual textual and qualitative content of the news has never been part of the analysis. Since the capabilities of textual analysis and news sentiment are not limited to stock price prediction, we use it to measure whether news is ‘good’ or ‘bad’. That is, we use news sentiment as an explanatory variable to analyze how the tone of news shapes insider trading behavior. This way of measuring whether news is ‘good’ or ‘bad’ also provides a decisive advantage: While the company’s situation is a given fact, executives do have flexibility in the scoping and wording of disclosures (Cheng & Lo 2006). Thereby, the tone of news is a more direct measure of executives’ intention than the abnormal stock price reaction around the news. The stock price reaction is also subject to the influence
of other factors such as market’s expectations on the company and the context of the news. Second, we do not only consider news published shortly before or after insider trades, but also examine intertemporal relationships to news in the previous and following year. Consequently, we establish an alternative measurement for management’s expectation about future firm performance which is currently expressed by future earnings and future stock returns. Third, we do not only consider management forecasts (as e.g. Cheng & Lo 2006, Brockman et al. 2008), but also look at a diverse set of US news announcements archived by Reuters over 8 years. We also distinguish between corporate news issued by the company and 3rd- party news published by independent brokers and investment analysts. Fourth, we further drill down into drivers of ‘good’ and ‘bad’ news and explain how confidence and uncertainty expressed in language impacts insider trading behavior. We find that directors sell in years when their companies issue positive news and buy when their companies issue negative news. This behavior follows the stock market adage “Buy on bad news, sell on good news” and is consistent with Cheng & Lo (2006). Analogously, for 3rd-party news, we find that insiders sell in years when news coverage is pleasant and buy when unpleasant. The impact of corporate news – being influenced by insiders – on insider trading is higher than for 3rd-party news. We also find that insiders buy when next year’s news tone improves compared to the current year. Similarly, insiders also buy their companies’ stock when news tone is worse than in the previous year. Thereby, insider trading is not only affected by ‘good’ or ‘bad’ news as a whole, but also by insecurity and uncertainty expressed in the language of the news. The findings demonstrate how textual analysis can add significant value for applications in finance when textual data is available. Our findings also indicate that apparent contrarian trading strategies could as well be a consequence of executives’ active manipulation of firm- specific information. The remainder of the paper is organized as follows: Section 2 discusses approaches for news sentiment analysis and our research design and methodology while Section 3 describes the data set. Section 4 presents the main empirical findings and its implications. Section 5 summarizes and concludes the paper.
2 Data and Research Design
In this section, we introduce the data set and our research design that we use to explain
insider trading behavior. We build on the research design proposed in Piotroski & Roulstone
(2005) by introducing a sentiment variable to account for the tone of news issued by
companies and 3rd-party providers. We start this section by describing the data set and
variables for the regression analysis. Thereby, we first describe our response variable
measuring insider trading. Having defined the essence of sentiment analysis, we elaborate on
how different sentiment measures are included in our analysis. Lastly, we describe the control
variables for our main regression and list descriptive statistics for our key variables.
2.1 Data set
Data for our analysis is collected from Reuters World Archive, Thomson Reuters Insider
Filings, CRSP, Compustat and ExecuComp. Our news sample is based on the Reuters World
Archive. We select all news relating to US companies and split them between corporate news
and 3rd-party news. Corporate news are issued by companies and consist of financial results
and forecasts, changes in management, regulatory issues, dividend announcements, merger
and acquisitions and other stock exchange activities. 3rd-party news is not issued by
companies, but by 3rd-party providers and mainly consists of broker research, press digests
and independent investment analysis. We only select those news messages which are focused
on a single company to avoid confusions with other companies. After filtering, we have
252,700 US corporate news and 48,600 3rd-party news from the years 2003 to 2010. Table 1
lists the content categories included in the news set. For news with multiple labels (e.g.
dividend announcements may be part of financial results reports as illustrated by the headline
“Staples' quarterly profit rises, initiates dividend”), only the main category is listed.
Corporate news mainly (66%) comprises financial results while the majority of 3rd-party
news (71%) stems from broker research.
News relates to 5,032 different US firms which are listed on the NYSE, NASDAQ and
AMEX stock exchanges for which we have information on directors’ trades from Thomson
Reuters Insider Filings. All stock prices and market indices needed to calculate returns are
collected from CRSP database. All accounting data based on financial statements used to
calculate return on assets (ROA) and book-to-market (BM) ratios are sourced fromCompustat. Data related to executive compensation variables (i.e. GRANTS,
OPTION_EXERCISE) are gathered from ExecuComp.
Table 1. Overview of news sample
No. of Corporate news 2003 2004 2005 2006 2007 2008 2009 2010 Total
Financial Results 13,514 16,043 16,076 19,391 22,118 25,673 28,469 25,713 166,997
Mergers & Acquisitions 1,063 1,356 1,947 3,337 4,692 3,669 3,426 3,329 22,819
New issues of shares/bonds 2,784 3,033 2,695 2,853 2,450 1,753 2,257 1,965 19,790
Regulatory issues 1,259 1,663 1,608 2,617 2,308 2,265 3,690 3,602 19,012
Management issues 204 1,120 1,389 2,668 2,579 2,515 2,606 1,980 15,061
Dividend 279 399 616 763 701 1,533 39 95 4,425
Company Events 13 197 - - - 227 880 450 1,767
Exchange activities 26 132 124 130 172 178 215 283 1,260
Bankruptcy 86 119 100 53 52 203 445 175 1,233
IPO 30 96 54 45 28 16 26 23 318
Total 19,258 24,158 24,609 31,857 35,100 38,032 42,053 37,615 252,682
No. of 3rd-party news 2003 2004 2005 2006 2007 2008 2009 2010 Total
Broker research 4,519 2,483 5,663 7,633 2,663 1,550 1,957 6,362 34,335
Investment analysis 15 10 19 45 1,410 4,194 3,198 4,799 14,246
Total 4,534 2,493 5,682 7,678 4,073 5,744 5,155 11,161 48,581
2.2 Measurement of Insider-Trading Behavior and Related Characteristics
Following Piotroski & Roulstone (2005), we measure insider-trading behavior using the
firm’s purchase ratio, defined as
PRi,t = BUYi,t / (BUYi,t + SELLi,t)
where BUYi,t (SELLi,t) equals the number of shares bought (sold) by directors of firm i
during fiscal year t. To be comparable with Piotroski & Roulstone (2005) and Rozeff &
Zaman (1998), we only include firm-years in which insiders engaged in open-market
transactions during the fiscal year.
2.3 News Sentiment Analysis
In news sentiment analysis, written texts are associated with a particular numbered value or
category in order to answer related research questions. Most sentiment approaches measure
how market participants perceive and react upon important corporate news. Accordingly, thetone of the news and its relationship to stock price reactions are analyzed (Liebmann et al.
2012; Li 2010a; Tetlock et al. 2008). Based on the stock price reactions following the
announcement, news is classified into either positive or negative.
As written text is a form of unstructured information as compared with e.g. financial
statements, it first has to be translated into a machine readable representation of the text2.
This step includes the identification of those words, which are most relevant and should thus
represent the text. Second, a numerical metric needs to be calculated for each announcement
as an aggregate based on the occurrences of the previously identified words. These two steps
have to be performed for any kind of sentiment analysis. Li (2010a) distinguishes sentiment
analyses into dictionary and statistical approaches.
The dictionary has a fixed set of words and often originates from a certain specific field,
such as the Harvard-IV-4 psychosocial dictionary which is used by Tetlock et al. (2008). For
aggregation into a metric, Tetlock et al. (2008) simply count the negative words and add up
all occurrences of negative words from the dictionary for each message. Other dictionaries
may focus on a combination of psycho-social and economic words. While this approach is
striking for its simplicity, it suffers from many drawbacks (Li 2010a). Few dictionaries exist
that are built for the setting of corporate financial statements and thus may prove
inappropriate for such settings (Loughran & McDonald, 2011). Furthermore, the simple
dictionary-based approach explicitly ignores the context of the sentences. For instance, if a
sentence is about cost, then the word ''increase'' qualifies it as being negative: However, it is
likely to be a positive word if the sentence is about "revenue” (Li 2010a). Lastly, some words
fully change their sentiment depending on the specific domain they are used in. While
“cancer” is a serious disease and labeled negative in the Harvard-IV-4 dictionary, it may be
part of a positive message when a new drug for cancer treatment has been developed.
To mitigate these drawbacks, statistical approaches have been developed to account for
correlations between specific keywords and the text type to classify the documents. This task
is mostly performed by machine learning approaches (Li 2010b; Hagenau et al. 2012).
However, some approaches translate their results into quantifiable metrics which can be used
for competitive comparison to dictionary approaches (Jegadeesh & Wu 2010, Liebmann et al.
2012).
2 In text mining, the step includes three sub-steps: Feature extraction (e.g. retrieving words from a text), Feature Selection
(e.g. determining which of the retrieved words are relevant), and Feature Representation (e.g. selecting the data format
how occurrences of words in a text shall be represented). We follow this approach, but simplify the description. In a
dictionary approach, there is no feature selection step as the predetermined dictionary itself defines which features to use.For the statistical approach, all words in the document corpus have to be extracted first.
Some of these words will be occurring in almost all text messages (i.e. stopwords such as
“the”, “it”, “like”, “or”). Thus, they are of low informative value and need to be excluded
from the analysis. Additionally, assuming that words with the same word stem convey the
same or similar meaning, we employ the Porter Stemmer, as in Porter (1980), which reduces
inflected words to their stem. Subsequently, the most informative words for sentiment
analysis out of the remaining words have to be selected3. For our explanation of insider
trading activity, we focus on two dictionary based approaches and one statistical approach.
Most informative refers in particular to those words which help to discriminate between
positive and negative news messages. If, for example, a word frequently occurs in both,
positive and negative news messages, it is considered to contain less information. However, if
a word occurs less frequently in the overall data set, but concentrates in either positive or
negative news messages, it is considered to be informative. To determine whether a news
message is positive or negative in an objective manner, abnormal stock price returns on the
day the news was disclosed can be used to label the messages. Actual market returns are used
to identify the most informative words. The market feedback however reflects the average
interpretation of all acting market participants which significantly differ from using a
predetermined dictionary as all words are selected that play a discriminating role in the given
text corpus and thus do not miss out on relevant terms that the creator of a predetermined
dictionary has not thought of.
In our empirical analysis, we use three different proxies for news sentiment, Tetlock-neg
(Tetlock et al. (2008)), Net-optimism (Loughran and MacDonald (2011)) and Tonality
Liebmann et al. (2012):
1. Tetlock-neg (H4): Tetlock et al. (2008) define a sentiment measure based on the
negative wordlist of the Harvard-IV dictionary. The authors take the “standardized
fraction of negative words” in each composite news story. Standardization is performed
by subtracting the prior year’s mean and dividing by the prior year’s standard deviation
of the fraction of negative words. The larger the sentiment value, the more negative the
interpretation might be. Consequently, there is a reciprocal relation between Tetlock-
3
This step is generally referred to as ‘Feature Selection’neg and stock price returns. As we expect more insider purchases in times of negative
news, we expect a positive correlation between Tetlock-neg and insider purchases.
2. Net-optimism: Loughran and MacDonald (2011) designed their wordlists for a
particular finance application based on the analysis of a comprehensive set of 10K
reports. They calculate a measure based on the difference between positive and
negative words and divide by “the total words in the text article”. Consequently, results
range between -1 and 1. This sentiment factor differs from Tetlock-neg as it also takes
positive words into account. Thus, it is named Net-optimism. The larger the Net-
optimism the more positive the news message is supposed to be. Hence, there is a
positive relation between Net-optimism and stock price returns and we expect a
negative relation to insider purchases.
3. Tonality: Liebmann et al. (2012) designed their own sentiment factor as a proxy for
announcement interpretation. Tonality differs significantly from the two approaches
above as it is specifically designed for the given data set and does not rely on any
general predetermined dictionary wordlist. Based on the Tonality word list, each
informative word in an announcement receives a weighting factor. The factor has a
positive sign for positive words and a negative sign for negative. The sum of weighting
factors over a whole announcement is standardized by the total number of words. The
larger the Tonality, the more positive the news message is supposed to be. Hence, there
is a positive relation between Tonality and stock price. As Tonality is a statistical
approach, tonalities for each word need to be calibrated on a different, but very similar
news sample with corresponding stock returns. It is important to note that Tonality is
always designed and calibrated out-of-sample. We make use of an out-of-sample
calibration on corporate disclosures of similar news content. For a detailed description
of the calculation of the Tonality, we refer to Liebmann et al. (2012) where motivation
and calculation of the Tonality is described in more detail.
Sentiment metrics as described in the previous subsections calculate a sentiment value for
each news message. As many different news messages are issued throughout the year,
sentiment values have to be aggregated to form an annual sentiment value. We choose an
annual timeframe because of the availability of control variables. Some of our control
variables which are based on a company’s financial reporting are only available on aquarterly basis, for some companies only on an annual basis. Additionally, choosing a quarterly timeframe would not only reduce the number of companies in the sample, but also decrease the number of available news per data point to one fourth. For creating an aggregated annual sentiment, we average sentiment values of news messages for each firm and fiscal year. Each of the news in our sample meets certain requirements that we impose to eliminate irrelevant stories. We require that each story contains at least 20 words in total and at least 3 words from the respective wordlist of the employed sentiment approach (c.f. Tetlock et al. (2008)). We impose these word count filters to eliminate news that contain only tables or lists with quantitative information, and to limit the influence of outliers on our sentiment measures. For creating a meaningful average, we also require a minimum of four news messages per fiscal year and firm. A higher number of news reduces noise and variability of the sentiment measure and thus generates results of higher confidence. This fact is further investigated in section 3.2 where we demonstrate the impact of different minimum values on the significance of results. 2.4 Measurement of Future Firm Performance and Contrarian Behaviour For each fiscal year where insiders were trading, we measure the firm’s future performance in the following fiscal year (i.e. year t+1). Following Piotroski & Roulstone (2005), we use two measures for next year’s performance: the firm’s future market-adjusted stock return and annual earnings innovation. The firm’s future market-adjusted stock return (MARETt+1) is measured over the 12-moth period in fiscal year t+1: This metric measures the director’s potential profit from trading in the firm’s stock compared to the buy-and-hold return of the market portfolio. If directors have superior knowledge about information influencing future stock returns, director purchases will be positively related to return performance in t+1. As neither insiders ex ante nor researchers ex post can perfectly predict or explain how market expectations and thus future returns will evolve (Piotroski & Roulstone, 2005), alternative measures of insiders’ information is needed. One measure of future firm performance is the earnings of the following year. As changes in a firm’s earnings performance influence stock prices, insiders have an incentive to trade based on superior information about future earnings. Accordingly, Piotroski & Roulstone (2005) expect insider
trading to reflect the sign of next year’s earnings change. Their primary measure of future earnings news is year t+l’s annual earnings innovation, defined as ∆ROAt+1 = ROAt+1 - ROAt where ROAt equals net income before extraordinary items standardized by average total assets which is assumed to represent market expectations about future earnings performance. Consequently, any changes in annual earnings represent private information of directors. In a corresponding manner, we also use this year’s annual earnings innovation (∆ROAt) as control variable. Following Rozeff & Zaman (1998), we test whether insiders follow contrarian strategies using the firm’s 12-month market-adjusted buy–and-hold stock return (MARETt) and the firm’s book-to-market ratio (BMt) in year t. All annual variables are calculated over the fiscal year. Companies are ranked based on MARETt and BMt and classified into return treciles (i.e. HRET_t and MRET_t) and book-to-market quintiles respectively. By choosing the same controls, we facilitate comparability with the results of Piotroski & Roulstone (2005) and Rozeff & Zaman (1998). 2.5 Controlling for other factors affecting insider trading Insiders’ trading behavior is influenced by changes in their holdings due to the granting of stocks and options and the exercising of stock options (Ofek & Yermack 2000; Piotroski & Roulstone 2005). We use two variables measuring compensation-related changes in insider holdings: number of shares of restricted stock and stock options granted (GRANTSt) and number of stock options exercised (OPTN_EXRCt) in year t. Following Piotroski & Roulstone (2005), due to the skewness of the data, we measure each variable as the log of one plus the ratio of number of shares granted or options exercised during the fiscal year, respectively, scaled by total shares outstanding at year-end. We also control for firm size (i.e. logarithm of total assets TAt) for each firm-year. 2.6 Descriptive Statistics of Key Variables In this subsection, we provide descriptive statistics on the sample we used in this study. Table 2 presents descriptive statistics for both our sentiment metrics as well as other control variables.
Given the current executive compensation schemes in which stock options account for a large part of variable compensation, the majority of our insider trading data consists of sell transactions. 70.3% of firm-years exhibit net sell transactions while 36.0% represent sell-only transactions. This distribution is also consistent with Piotroski & Roulstone (2005). As a minimum number of news is required per firm-year observations to calculate a sentiment value, the availability of news determines the number of observations in our empirical analysis. As there is less 3rd-party news in the sample, we have fewer observations available for sentiment metrics based on 3rd-party news. For corporate news based metrics, we obtain 13,881 firm-year observations. For 3rd-party news based metrics, we obtain 2,613 observations. Despite the different sample sizes, the distribution of these two subsets of news is very similar. The Tonality sentiment measure based on corporate news shows more positive than negative years with a median Tonality of 0.484. This can be attributed to companies reporting more positive than negative news. Tonality based on 3rd party news also measures more positive than negative news. However, with a median at 0.374 3rd party news messages are assigned a less positive meaning than corporate announcements. At a first glance, Tetlock- neg’s shows a different behavior with the median being closer to zero. However, Tetlock-neg only measures the negative content in news and is standardized on annual basis (Tetlock et al. 2008). Similar to Tetlock-neg, Net-optimism is not skewed to the positive side. This is not astonishing as the stemmed wordlist behind Net-optimism contains 85% negative words and is thus much more sensitive to negative content (Loughran & MacDonald 2011). Comparing corporate news and 3rd party news, all three sentiment measures attribute more positive meanings to corporate news than to 3rd-party news. Looking at control variables, future market-adjusted returns yield a different view than future earnings innovations. While 60.2% of firm-years are associated with a negative future return performance, only 47.2% exhibit a negative earnings innovation. The median of earnings innovations (∆ROAt) is only slightly above zero demonstrating that about an equal number of firm-years have positive and negative earnings changes. The wide spread in firm sizes measured by total assets (ln(TAt)) expresses the large variety of firms in the sample. Data availability for option exercises and grants is lower than for other control variables. Specifically, only 2,606 data points have values for Tonality (Corporate news) and grants. For 3rd-party news, the number decreases to 985 data points. Thus, we do not show option grants and exercises in our regression analyses, but use them for robustness checks only. All continuous variables are winsorized at the 1% level.
Table 2 - Descriptive statistics
This table presents descriptive statistics for sentiment metrics as well as other control variables. The number of
observations ‘N’ reflects the maximum number of observations in the data sample with a valid entry for the specified
variable. For each of the variables, only a fraction of the sample has available and valid entries. The number of
observations in the regressions will be lower as only those data points can be used for regression which have a valid entry
in all employed variables, i.e. for the univariate regression at least ‘PR’ and the independent variable is needed.
N Mean Stdev 5th Pctl 25th Pctl Median 75th Pctl 95th Pctl
PR 20,062 0.297 0.407 0.000 0.000 0.020 0.678 1.000
Tonality
13,881 0.463 0.587 -0.533 0.074 0.484 0.884 1.400
(Corp. news)
Tonality
2,613 0.363 0.434 -0.358 0.092 0.374 0.643 1.065
(3rd-party news)
Tetlock-neg
13,854 -0.103 0.524 -0.927 -0.453 -0.128 0.211 0.798
(Corp. news)
Tetlock-neg
2,642 0.019 0.650 -0.976 -0.412 -0.006 0.452 1.052
(3rd-party news)
Net-optimism
11,158 -0.001 0.014 -0.023 -0.009 0.000 0.008 0.021
(Corp. news)
Net-optimism
2,642 -0.006 0.014 -0.030 -0.014 -0.005 0.003 0.016
(3rd-party news)
MARETt+1 23,989 0.124 0.841 -0.560 -0.229 -0.004 0.273 1.104
∆ROAt+1 21,228 -0.002 0.278 -0.194 -0.021 0.000 0.022 0.196
∆ROAt 23,987 0.005 0.267 -0.197 -0.022 0.001 0.023 0.210
ln(TAt) 25,003 6.365 2.161 2.904 4.857 6.332 7.748 10.096
BMt 24,999 0.699 0.778 0.120 0.318 0.530 0.829 1.762
OPTN_EXRCt 11,268 0.004 0.145 0.000 0.000 0.001 0.003 0.012
GRANTSt 4,800 0.006 0.009 0.000 0.001 0.004 0.008 0.020
3. Results
In this section, we demonstrate how news sentiment of both corporate news and 3rd-party
news contributes to the explanation of insider trading behavior. After showing the results for
contemporaneous sentiment (i.e. this year’s sentiment), we examine drivers for significance
and also look into associations of intertemporal sentiment changes (i.e. changes between
previous, current and next year’s sentiment) and insider purchasing behavior. Lastly, we
present the impact of confidence and uncertainty in language on insider trading activity.3.1 Contemporary results
To analyze the association between contemporary sentiment and insider purchasing behavior,
we use the three sentiment measures described in the previous section. Table 3 presents basic
univariate relations between news sentiment measures and insider trading (i.e. purchase ratio
PR).
Table 3 – Univariate Regression: Prediction of Purchase Ratio (PR) of different
sentiment approaches
Factors Cons Coeff t-stat p-value R² N
Tonality
0.2862 -0.0878 -14.77 0.000 1.93% 11,080
(Corp. news)
Tonality
0.1673 -0.0749 -6.72 0.000 1.48% 3,001
(3rd-party news)
Tetlock-neg
0. 2510 0. 0388 5.65 0.000 0.09% 11,054
(Corp. news)
Tetlock-neg
0.1306 -0.0175 -1.84 0.071 0.16% 2,061
(3rd-party news)
Net-optimism
0.2175 -0.9895 -3.58 0.000 0.14% 9,060
(Corp. news)
Net-optimism
0.1170 -2.010 -4.53 0.000 0.99% 2,061
(3rd-party news)
Univariate regressions document that indeed sentiment measures are correlated to insider
trading activity. For corporate news, coefficients are negative for Tonality and Net-optimism:
This indicates that insiders sell in years when their companies issue positive news and
insiders buy when their companies issue negative news. Analogously, for 3rd-party news, the
negative coefficient indicates that insiders sell in years when news coverage is pleasant and
buy when unpleasant. The correlation coefficient is significantly higher for corporate news.
This indicates that the impact of corporate news (which may be influenced by insiders
themselves) on insider trading is higher than for 3rd-party news. Also Tetlock-neg for
corporate news confirms the indications above. As Tetlock-neg expresses the negativity of
news, the sign of the coefficient needs to be positive to confirm results. For 3rd-party news,
the sign of Tetlock-neg’s coefficient is not as expected. However, the correlation is not very
strong. Comparing sentiment approaches, the regression also shows that Tonality based on
both news types has the highest explanatory power.
The number of observations in the regression varies with the news type. As a minimum
number of news is required per year and firm for calculating a sentiment value, the
availability of news limits the number of observations. There is less 3rd-party news in thesample, thus, fewer observations are available for sentiment metrics based on 3rd-party news.
The intersection of both news types will lead to the lowest number of observations.
These univariate tests, however, fail to separate out the effects of trading against
misvaluation and trading with superior future information (Piotroski & Roulstone 2005). To
test whether the relations between insider purchase ratios and tone of the news are
incremental to variables capturing contrarian trading and future firm performance, we utilize
the methodology in Piotroski & Roulstone (2005) as our benchmark. In particular, we
estimate coefficients from the following cross-sectional model:
PRi;t = α + β1 * sentiment_metrics + β2 * MARETi;t + β3 * ∆ROAi;t+1 + β4 * ∆ROAi;t
+ β5 * HRETi;t + β6 * MRETi;t + β7 * BM1i;t+ β8 * BM2i;t + β9 * BM3i;t + β10 *BM4i;t
+ β11 * ln(TAt)
where sentiment_metrics use different sentiment measures (Tonality, Tetlock-neg, Net-
optimism) interchangeably in our regression equation. HRETt and MRETt are indicator
variables equal to one if the firm’s 12-month-market-adjusted return (MARETt) is in the top
and middle third of all sample firms that year, and zero otherwise. The indicator variables
BMlt, BM2t, BM3t, and BM4t are equal to one if the firm’s BM ratio ranks in the bottom,
second, third and forth quintiles, respectively, of annual BM ratios, and zero otherwise
(Piotroski & Roulstone 2005).
Furthermore, we tested for multicolinearity to ensure that no relevant abnormality
confounds the experiments’ results and could not find any biasing influences. We perform
nine separate regression models: Six regressions - one model for each sentiment approach
and each news type, one empty control model as benchmark, one model for Tonality
sentiment for corporate news and 3rd-party news, and one full combined model to compare
the interactions between sentiment approaches and news types. For each of the estimations,
we perform a set of ordinary least square (OLS) regressions to compare the sentiment
approaches against one another. Table 4 presents the results of all nine regression models for
the prediction of purchase ratio PRt during fiscal year t.
We estimate all models with firm fixed effects allowing us to control for unobservable firm
characteristics that may drive differences in trading behavior between firms (e.g. firm-
specific insider trading policies or restrictions). The standard errors of the coefficients in all
regressions are adjusted for clustering at the firm level to control for non-independence of
observations within a firm.Table 4. Multivariate Regression: Predictor Models for Purchase Ratio (PR)
Table 4
This table presents estimated coefficients of the OLS regressions with firm fixed effects to explain insider purchasing activity.
t-statistics are listed in parentheses. Sentiment metrics are based on all corporate news, 3rd-party news and on a combination of
both, respectively. Standard errors of the coefficients are adjusted for clustering at the firm level. All variables are as defined
in section 2. Significance is denoted at the 0.01 (***), 0.05 (**) and 0.1 (*) level, respectively.
(1) (2) (3) (4) (5) (6) (7)
(8) (9)
All Corp & All Corp &
News Focus - All Corp. All Corp. All Corp. 3rd-party 3rd-party 3rd-party
3rd-party 3rd-party
Tonality -0.061*** -0.096*** -0.094***
(Corp. news) (-7.41) (-4.56) (-3.68)
Tetlock-neg 0.028*** 0.069***
(Corp. news) (3.16) (3.01)
Net-optimism 0.243 2.229**
(Corp. news) (0.71) (2.46)
Tonality -0.076*** -0.053*** -0.044*
(3rd-party news) (-4.51) (-3.00) (-1.83)
Tetlock-neg -0.008 -0.022**
(3rd-party news) (-0.81) (-1.99)
Net-optimism -1.785*** -0.783
(3rd-party news) (-2.85) (-0.90)
MARETt+1 0.0735** 0.075*** 0.074*** 0.073*** 0.054** 0.061** 0.062** 0.055** 0.058**
(7.47) (7.66) (7.50) (7.45) (2.19) (2.23) (2.28) (2.12) (2.21)
0.224*** 0.201*** 0.211*** 0.226*** 0.211** 0.246** 0.231** 0.188* 0.172*
∆ROAt+1
(5.97) (5.32) (5.60) (5.98) (2.30) (2.45) (2.34) (1.89) (1.76)
0.011 0.006 0.005 0.012 -0.039 -0.072 -0.08 -0.086 -0.106
∆ROAt
(0.30) (0.16) (0.12) (0.33) (-0.44) (-0.83) (-0.92) (-0.90) (-1.11)
-0.042** -0.035*** -0.042*** -0.044*** -0.041** -0.058*** -0.043** -0.036* -0.036*
HRETt
(-4.40) (-3.62) (-4.25) (-4.43) (-1.98) (-2.73) (-2.05) (-1.67) (-1.68)
-0.032*** -0.026*** -0.031*** -0.033*** -0.041** -0.051** -0.044** -0.04** -0.038*
MRETt
(-3.40) (-2.72) (-3.2) (-3.41) (-2.09) (-2.56) (-2.21) (-2.03) (-1.89)
-0.165*** -0.141*** -0.16*** -0.166*** -0.214*** -0.246*** -0.236*** -0.193** -0.188**
BM1t
(-6.05) (-5.16) (-5.87) (-6.09) (-2.92) (-3.09) (-2.98) (-2.43) (-2.39)
-0.141*** -0.121*** -0.137*** -0.143*** -0.199*** -0.235*** -0.227*** -0.19*** -0.187***
BM2t
(-6.12) (-5.23) (-5.95) (-6.19) (-3.04) (-3.30) (-3.21) (-2.67) (-2.65)
-0.097*** -0.084*** -0.094*** -0.098*** -0.156*** -0.18*** -0.173*** -0.147** -0.149**
BM3t
(-4.80) (-4.12) (-4.62) (-4.84) (-2.64) (-2.80) (-2.71) (-2.29) (-2.31)
-0.047** -0.037** -0.044** -0.047*** -0.057 -0.068 -0.064 -0.053 -0.055
BM4t
(-2.56) (-2.02) (-2.40) (-2.57) (-1.11) (-1.22) (-1.14) (-0.94) (-0.99)
0.000 0.008 -0.002 -0.001 0.031 0.017 0.032 0.023 0.019
ln(TAt)
(0.01) (0.59) (-0.17) (-0.08) (1.17) (0.60) (1.09) (0.79) (0.62)
0.358*** 0.311*** 0.375*** 0.368*** 0.076 0.199 0.043 0.151 0.189***
cons (3.61) (3.17) (3.77) (3.69) (0.32) (0.77) (0.16) (0.56) (0.68)
R² 6.5% 7.4% 6.7% 6.5% 10.4% 8.7% 9.5% 12.3% 13.5%
N 10,012 10,012 9,986 9,986 2,137 1,956 1,956 1,931 1,915
Estimation (1) is our empty control model, which serves as benchmark confirming the
previously documented relations. Consistent with Piotroski & Roulstone (2005), we illustratethat executives are more likely to purchase securities during periods of falling stock prices. Table 4 documents the significant and negative coefficient of HRETt and MRETt. Insider purchases are positively related to the firm’s book-to-market ranking (with BM1 being the bottom quintile). We also confirm that insider purchase decisions are significantly positively associated with future firm performance, measured by both future earnings innovation (∆ROAt+1) and future market returns (MARETt+1). The control for firm sizes is insignificant for all of our estimations. Priotroski & Roulstone (2005) also show an inverse relation, albeit substantially weaker, between current earnings performance (∆ROA_t) and the firm’s purchase ratio. This relation would indicate that insiders sell their securities during periods of stronger firm performance. However, we cannot confirm this finding. Estimations (2), (3), and (4) examine the association between sentiment of corporate news and insider purchasing behavior. Consistent with the unconditional results in Table 3, there is a strong relation between a negative sentiment and a high insider purchasing activity and vice versa. This highlights again that insiders tend to sell in years when their company issues positive news and buy when the company issues negative news. The findings confirm previous research showing that executives publish bad news before purchasing shares (Cheng & Lo 2006). While Tonality and Tetlock-neg exhibit strong significance, Net-optimism is not significant at all. Similarly, estimations (5), (6), and (7) examine the association between sentiment of 3rd- party news and insider purchasing behavior. Consistent with the results in Table 3 and the estimation models based on corporate news sentiment, there is a relation between negative sentiment of 3rd-party news coverage and insider purchasing activity and vice versa. This demonstrates our hypothesis that insiders are also influenced by 3rd-party news in a way that they sell in years when news coverage is advantageous and buy when unpleasant. If sentiment metrics based on 3rd-party news are included into the regression, the current firm performance (∆ROA_t) exhibits its expected behavior, i.e. the coefficient turns negative. Comparing different sentiment metrics, the Tonality metric is most significant. As for the univariate regression Tetlock-neg based on 3rd-party news is not significant and carries a different sign than expected. Comparing the different news types, corporate news lead to more significant sentiment metrics. In contrast to corporate news in estimation (4), Net- optimism is significantly related to purchase ratio. The respective word list seems to be better tailored to 3rd-party news than to the language used in corporate news.
To even better compare the different news types, estimation (8) examines the association between Tonality sentiment on both corporate news and 3rd-party news and insiders purchasing behavior. Essentially, Tonality based on corporate news is more significant than Tonality based on 3rd-party news indicating that corporate news exhibit a higher influence on insider trading behavior than 3rd-party news. As corporate news is subject to direct influence by the insiders, this also suggests that corporate news may be used to actively manipulate stock prices. Having both sentiment metrics significant demonstrates that the combination of both news types adds additional explanatory power. Thus, both news types each have an influence on insider trading behavior. To corroborate the preceding results, control for omitted performance variables and compare the interactions between sentiment approaches and news types, we estimate a full model (estimation (9)) that includes all sentiment metrics for both news types. Tonality again proves to be the most significant sentiment metric with corporate news being more significant than 3rd-party news. The full model carries the highest explanatory power at R² = 9.9%. Our analysis builds on the assumption that insiders trade after the news is published. Although the research design does not control for the exact timing of insider transactions, trades before publication of major news are less likely due to regulatory reasons. Black-out periods before the announcement of scheduled news (such as quarterly or annual financial reports) prevent insiders from trading (see Footnote1 and Korczak et al. 2010) while insiders trading on major unpublished news are subject to prosecution. Furthermore, previous research controlled for the timing of disclosure and came to similar results (Cheng & Lo 2006). Our results are robust to the inclusion of stock option grants and stock option exercises, i.e. sign of coefficients stay the same and key tonality variables remain on similar significance levels. Still, they are not included as they drastically reduce the number of data points in our regressions due to lower data availability. The sentiment metrics in this section were produced requiring a minimum of 4 news messages per firm year (i.e. medium confidence as denoted in section 3.2). The following section will analyze robustness and significance for different minima. 3.2 Increasing significance of results In this section, we analyze two drivers for significance: First, we vary the number of news which we require as minimum per firm and year to calculate a Tonality average. If there is
less news in a firm-year, no Tonality average is calculated and the data point is omitted from
the regression. We denote the different thresholds as different confidence levels (i.e. low,
medium, and high). If more news messages contribute to the Tonality average per firm and
year, the average should reduce variability and the impact of outliers. Second, we analyze the
effect of different news types within the set of corporate news. In addition to metrics based
on all corporate news, we also base our sentiment metrics on financial results only, which
make up two-thirds of our corporate news data set. Other news categories may not be
analyzed separately as number of observations would become too low.
We perform nine separate regression models: One regression for each of the three
confidence levels and each of the three news categories (i.e. all corporate news, corporate
financial results, 3rd-party news). Table 5 presents the results of all nine regression models for
the prediction of purchase ratio PRt during fiscal year t.
Table 5 – Predictor Models for Purchase Ratio (PR) for different confidence levels and news categories
This table presents estimated coefficients of the OLS regressions with firm fixed effects to explain insider purchasing activity.
t-statistics are listed in parentheses. Sentiment metrics are based on all corporate news, corporate financial results (a subset of
the previous one), and 3rd-party news, respectively. Standard errors of the coefficients are adjusted for clustering at the firm
level. Tonality is calculated for different confidence levels, i.e. low confidence requires at least one news message per firm-year,
medium confidence requires a minimum of 4 messages while high requires 7 messages. For 3rd-party news, high confidence only
requires 5 news messages due to lower data availability. All other variables are as defined in section 2. Significance is denoted at
the 0.01 (***), 0.05 (**) and 0.1 (*) level, respectively.
(1) (2) (3) (4) (5) (6) (7) (8) (9)
Fin. Fin. Fin.
News Focus All Corp. All Corp. All Corp. Results Results Results 3rd-party 3rd-party 3rd-party
Tonality -0.019*** -0.029*** -0.031***
(low confidence) (-3.51) (-5.46) (-4.26)
Tonality -0.061*** -0.068*** -0.076***
(med. confidence) (-7.41) (-8.32) (-4.51)
Tonality -0.089*** -0.101*** -0.067***
(high confidence) (-8.42) (-9.83) (-2.89)
0.07*** 0.075*** 0.085*** 0.070*** 0.073*** 0.088*** 0.053*** 0.054** 0.049
MARETt+1
(9.02) (7.66) (6.75) (8.72) (7.05) (6.25) (4.03) (2.19) (1.55)
0.17*** 0.201*** 0.208*** 0.194*** 0.200*** 0.123** 0.184*** 0.211** 0.168
∆ROAt+1
(5.89) (5.32) (4.05) (6.26) (4.66) (2.15) (3.49) (2.30) (1.51)
0.018 0.006 0.026 0.020 -0.022 -0.003 -0.115** -0.039 -0.075
∆ROAt
(0.61) (0.16) (0.57) (0.64) (-0.57) (-0.06) (-2.17) (-0.44) (-0.67)
-0.04*** -0.035*** -0.035*** -0.038*** -0.033*** -0.029** -0.042*** -0.041** -0.024
HRETt
(-4.90) (-3.62) (-3.09) (-4.55) (-3.23) (-2.47) (-3.59) (-1.98) (-0.97)
-0.028*** -0.026*** -0.028*** -0.027*** -0.024** -0.023** -0.04*** -0.041** -0.026
MRETt
(-3.49) (-2.72) (-2.56) (-3.37) (-2.39) (-2.03) (-3.38) (-2.09) (-1.1)
-0.181*** -0.141*** -0.135*** -0.172*** -0.141*** -0.112*** -0.219*** -0.214*** -0.232***
BM1t
(-8.40) (-5.16) (-4.11) (-7.64) (-4.83) (-3.19) (-6.52) (-2.92) (-2.59)
-0.15*** -0.121*** -0.116*** -0.144*** -0.113*** -0.101*** -0.177*** -0.199*** -0.214***
BM2t
(-8.43) (-5.23) (-4.08) (-7.69) (-4.53) (-3.25) (-5.96) (-3.04) (-2.69)
-0.1*** -0.084*** -0.081*** -0.092*** -0.079*** -0.08*** -0.127*** -0.156*** -0.159**
BM3t
(-6.53) (-4.12) (-3.17) (-5.74) (-3.61) (-2.91) (-4.78) (-2.64) (-2.19)-0.057*** -0.037** -0.041* -0.049*** -0.026 -0.035 -0.082*** -0.057 -0.048
BM4t
(-4.19) (-2.02) (-1.77) (-3.44) (-1.33) (-1.37) (-3.41) (-1.11) (-0.76)
0.004 0.008 0.015 -0.003 0.000 0.01 0.009 0.031 -0.001
ln(TAt)
(0.39) (0.59) (0.96) (-0.26) (0.03) (0.59) (0.60) (1.17) (-0.02)
0.365*** 0.311*** 0.216* 0.404*** 0.353*** 0.224 0.273** 0.076 0.345
cons (4.92) (3.17) (1.71) (5.02) (3.36) (1.59) (2.22) (0.32) (1.02)
R² 6.0% 7.4% 9.0% 6.6% 8.2% 9.8% 7.6% 9.7% 8.0%
N 14,517 10,012 6,715 13,451 9,009 5,594 6,198 2,137 1,555
For all corporate news, significance of associations between purchasing activity and Tonality
increases with the confidence level (estimations (1), (2), and (3)). This suggests that
increasing confidence reduces noise and limits the impact of outliers. Still, lowest confidence
(i.e. only one news message per firm-year) already exhibits significant relation (estimation
(1)) showing the robustness of results. Significance of relations between purchasing activity
and sentiment further increases when Tonality is based on financial results only instead of the
full set of corporate news (estimations (4), (5), and (6)). Financial results seem to transport
more relevant content explaining insider trading than the average news. This is consistent
with Tetlock et al. (2008) who find that stories about fundamentals predict earnings and
returns more effectively than other stories. However, another possible explanation might also
be that it is just easier to influence and adjust the tone in financial results than in other news.
It is important to note that the number of observations significantly increases with lower
required confidence – we witness twice as many data points for low confidence compared to
high confidence.
Tonality based on 3rd-party news also benefits from medium confidence compared to low
confidence (estimations (7) and (8)). However, Tonality based on 3rd-party news does not
benefit from high confidence (estimation (9)) which could be attributed to the lower number
of news in the 3rd-party sample.
3.3 Intertemporal results
While the previous sections only considered contemporaneous Tonality, this section looks at
tone changes between the previous, current, and following year. In the same manner as future
firm performance is measured by MARETt+1 and ∆ROAt+1, the future tone of news is also of
importance. As MARET (market adjusted return) and ∆ROA (delta return on assets) are
relative measures for the change between two subsequent years, Tonality is also calculated asthe delta between two subsequent years. ∆Tonalityt+1 is calculated as the difference between
the next and the current year while ∆Tonalityt is the difference between the current and the
previous year. For this analysis, we focus on the Tonality metric only as this sentiment metric
showed the strongest relation between tone of news and insider purchasing activity in
previous regressions of this paper.
We perform nine separate regression models: Three for each news categories (i.e. all
corporate news, corporate financial results, 3rd-party news). Within each of the three news
categories, we run the current delta-Tonality, the future delta-Tonality and the combination of
both. Table 6 presents the results of all nine regression models for the prediction of purchase
ratio PRt during fiscal year t.
Table 6 – Predictor Models for Purchase Ratio (PR) for intertemporal sentiment measures
This table presents estimated coefficients of the OLS regressions with firm fixed effects to explain insider purchasing activity.
t-statistics are listed in parentheses. Sentiment metrics are based on all corporate news, corporate financial results (a subset of
the previous one), and 3rd-party news, respectively. Standard errors of the coefficients are adjusted for clustering at the firm
level. Intertemporal changes in Tonality are calculated as follows:
∆Tonality_t = Tonality_t – Tonality_t-1
∆Tonality_t+1 = Tonality_t+1 – Tonality_t
All other variables are as defined in section 2. Significance is denoted at the 0.01 (***), 0.05 (**) and 0.1 (*) level,
respectively.
(1) (2) (3) (4) (5) (6) (7) (8) (9)
Fin. Fin. Fin.
News Focus All Corp. All Corp. All Corp. Results Results Results 3rd-party 3rd-party 3rd-party
-0.03*** -0.028*** -0.031*** -0.029*** -0.008 0.035
∆Tonalityt (-4.32) (-3.42) (-4.55) (-3.64) (-0.55) (1.53)
0.023*** 0.008 0.021*** 0.01 0.022 0.074**
∆Tonalityt+1 (3.58) (0.91) (3.34) (1.17) (1.44) (2.34)
0.086*** 0.072*** 0.088*** 0.093*** 0.071*** 0.095*** 0.036 0.044 0.004
MARETt+1
(6.44) (6.76) (6.53) (6.57) (6.32) (6.48) (0.98) (1.18) (0.08)
0.231*** 0.226*** 0.235*** 0.237*** 0.253*** 0.226*** 0.208 0.309* 0.396*
∆ROAt+1
(5.06) (5.45) (4.94) (4.48) (5.60) (4.09) (1.53) (1.88) (1.76)
-0.029 -0.001 -0.006 -0.016 0.004 -0.021 -0.025 -0.076 -0.038
∆ROAt
(-0.61) (-0.03) (-0.13) (-0.29) (0.09) (-0.37) (-0.21) (-0.53) (-0.24)
-0.022* -0.043*** -0.025** -0.026** -0.044*** -0.027** -0.04 -0.025 -0.003
HRETt
(-1.83) (-4.31) (-2.00) (-2.05) (-4.12) (-2.08) (-1.03) (-0.78) (-0.06)
-0.017 -0.036*** -0.023* -0.024* -0.035*** -0.025* -0.043 -0.022 0.004
MRETt
(-1.42) (-3.60) (-1.79) (-1.87) (-3.24) (-1.87) (-1.37) (-0.78) (0.09)
-0.144*** -0.159*** -0.133*** -0.129*** -0.16*** -0.114*** -0.384*** -0.4*** -0.514***
BM1t
(-4.32) (-5.45) (-3.95) (-3.55) (-5.23) (-3.11) (-2.71) (-3.10) (-2.90)
-0.124*** -0.137*** -0.118*** -0.114*** -0.135*** -0.102*** -0.343*** -0.336*** -0.444***
BM2t
(-4.27) (-5.55) (-4.03) (-3.58) (-5.17) (-3.16) (-2.63) (-2.99) (-2.66)
-0.077*** -0.091*** -0.072*** -0.074*** -0.096*** -0.067** -0.207* -0.246** -0.259*
BM3t
(-2.94) (-4.20) (-2.71) (-2.61) (-4.17) (-2.31) (-1.89) (-2.47) (-1.95)
-0.054** -0.04** -0.053** -0.048* -0.039* -0.047* -0.106 -0.091 -0.121
BM4t
(-2.29) (-2.04) (-2.20) (-1.84) (-1.86) (-1.75) (-1.09) (-1.24) (-1.13)
0.008 0.007 0.006 -0.016 -0.003 -0.021 -0.015 0.01 -0.007
ln(TAt)
(0.44) (0.53) (0.30) (-0.79) (-0.21) (-0.93) (-0.33) (0.26) (-0.14)0.246* 0.292*** 0.257 0.418*** 0.362*** 0.445** 0.556 0.335 0.56 cons (1.66) (2.67) (1.61) (2.57) (3.05) (2.47) (1.23) (0.88) (1.01) R² 7.1% 6.9% 7.4% 7.7% 7.3% 7.9% 10.8% 13.3% 17.4% N 6,810 9,024 6,410 5,893 8,031 5,492 899 1,033 522 Estimations (1), (2), and (3) examine the association between Tonality changes based on all corporate news and insider purchasing behavior. ∆Tonalityt is negatively related while ∆Tonalityt+1 is positively associated with purchasing activity. This demonstrates that insiders buy on both worsening news between last and current year and improving news between the current and next year. The observation reconfirms that insiders trade on superior knowledge about future earnings innovations and future stock returns (Piotroski & Roulstone 2005). The relation of purchasing activity to ∆Tonalityt is stronger than to ∆Tonalityt+1 (estimation (3)) showing that impact of existing news is stronger than expected future news. While insiders profit from their superior knowledge about future earnings and returns, existing news – with direct impact on stock prices – still seem to be the stronger driver for insider purchasing behavior than future news. Estimations (4), (5), and (6) examine association between Tonality changes based on financial results only and insider purchasing behavior. They confirm the results based on all corporate results, but exhibit slightly stronger relationships. Again, our sentiment metrics based on financial results seem to be more relevant. The analysis is repeated for 3rd-party news (estimations (7), (8), and (9)). Thereby, ∆Tonalityt is still negatively related (positively related for ∆Tonalityt+1), but not significant any more. In particular, the combination of both time horizons in estimation (9) shows that ∆Tonalityt+1 is more strongly associated to insider purchasing activity than for ∆Tonalityt. This indicates that insiders also buy on improving 3rd-party news coverage. In contrast to corporate news emitted by the company, the improving news coverage is more important than the change to the previous year. A possible explanation may be that 3rd-party news cannot be directly influenced by executives and mirror actual development and condition of the company in a reflected manner. 3.4. Relating insider trading to confidence and uncertainty in news Literature does not only provide dictionaries for the tone of news, but also for other forms of expression (Loughran & MacDonald 2011). By means of these dictionaries, not only low or
You can also read
