Some People Say Linguistic Imprecision in Corporate Disclosures Table A.1: Variable Deﬁnitions

Some People Say

Linguistic Imprecision in Corporate Disclosures

October 19, 2018

Abstract

Imprecise language in corporate disclosures can convey valuable information during uncertain times. To evaluate this possibility, we develop a measure of linguistic imprecision using sentences marked with the “weasel tag” on Wikipedia, which is used to identify vague and un- verifiable information. We find that managers use imprecise language in their 10-K disclosures when their firms face greater product market threats or greater financial constraints in equity markets. Consistent with linguistic imprecision capturing managerial information on upcoming positive but yet uncertain prospects, firms that employ high linguistic imprecision exhibit significantly higher subsequent abnormal returns. In addition, firms with high linguistic imprecision subsequently experience greater upside return volatility relative to downside return volatility. After a disclosure with high linguistic imprecision, firms have higher investment opportunities, particularly in intangible assets such as R&D. 1 Introduction

A well-functioning ﬁnancial system depends critically on disclosure of ﬁnancial information, yet encouraging informative and accurate disclosure is challenging in practice (La Porta et al., 2006).

The vast majority of prior work on the accuracy of disclosure tends to analyze numerical measures that are straightforward to quantify (e.g., earnings manipulation following Dechow et al., 1996), but textual information has become increasingly more important to ﬁnancial markets over time as

financial texts become easier to access and process (e.g., scraping EDGAR filings). It has thus become more important to understand the content of financial disclosures, particularly with respect to the interpretation of the textual information. At the same time, qualitative information in financial text is susceptible to contain vague information, and regulators and textual analysts have few reliable tools at their disposal to define imprecise versus precise text and to understand managerial incentives for linguistic imprecision in financial disclosures (Hwang and Kim, 2017; Hoberg and Lewis, 2017).

This is a critical gap that needs to be addressed, especially in light of recent evidence that the qualitative aspects of ﬁnancial text matter beyond quantitative disclosures (Huang et al., 2014).

In this context, we introduce to the finance and accounting literatures a novel measure of linguistic imprecision, developed from “weasel words” extracted from Wikipedia. Wikipedia defines a weasel word as “an informal term for words and phrases aimed at creating an impression that a specific or meaningful statement has been made, when instead a vague or ambiguous claim has actually been communicated.” Wikipedia states that “it can be used in advertising and in political statements, where it can be advantageous to cause the audience to develop a misleading impression.”

Moreover, the idea of weasel words is not merely a construct within Wikipedia. Miriam-Webster dictionary deﬁnes the term weasel word to mean, “a word used in order to evade or retreat from direct or forthright statement or position.” From these deﬁnitions, weasel words provide a useful marker of linguistic imprecision in corporate disclosures.1

Our core idea is to quantify these hallmarks of linguistic imprecision, and apply these insights to better understand both the choice to use imprecise language in corporate disclosure and the ﬁnancial

1In the context of legally-required corporate disclosures, we expect the incentives to use “weasel words” to be distinct from other source texts such as political statements and informal conference calls. As we discuss later at length, this distinction is important for how to interpret the use of linguistic imprecision in our context.

1 implications of it. Nevertheless, it is challenging to assemble a reliable dictionary of imprecise words and phrases to use as the basis for our measure. The primary challenge is in constructing a list that is free of researcher subjectivity. Language that appears imprecise to one person may not appear so to another. A related issue is that there is no universally accepted list of imprecise words and phrases. To address both of these issues, we analyze the text of Wikipedia articles, and more importantly, the tags embedded into these articles to compile a list of keywords and phrases for linguistic imprecision (henceforth “imprecision keywords”) for use in our textual analysis of corporate disclosures. Most usefully for our analysis, Wikipedia users are advised to use the “weasel tag” when they encounter sentences or phrases in the text of Wikipedia articles that have vague phrasing that accompanies unveriﬁable information. Wikipedia uses this tagging strategy to identify and crowdsource solutions to correct imprecise language, which helps them provide a more precise online encyclopedia. We use the language in these weasel-tagged sentences that are outside of our textual corpus for main analysis to build a list of imprecision keywords that is free of researcher subjectivity.

Using our dictionary of imprecision keywords, we generate a measure of linguistic imprecision at the firm-year level (and at the paragraph-year level in some tests) by computing the fraction of words in each firm’s annual 10-K filing that are imprecision keywords. Consistent with our measure capturing linguistic imprecision, we find that 10-Ks that have greater linguistic imprecision tend to exhibit greater uncertainty, and also contain more words that convey differing shades of meaning

(weak modal words) and positive sentiment. Yet, we note that the information contained in our linguistic imprecision measure has distinctive and unique aspects compared to these other textual measures. For example, even after controlling for uncertainty and (strong and weak) modality textual measures, other textual measures from the Loughran and McDonald (2011) dictionary, and also a readability measure (the Fog index as in Li, 2008 for example), we find that disclosures with high linguistic imprecision are more likely to be issued by young and small firms with low profitability and low tangibility. These findings are consistent with the types of firms that would gain most from or would face less risk from using imprecise language in disclosures: firms when experiencing bad times (i.e., low or worsening profitability) and firms that are difficult to quantify

2 (i.e., young and intangible).

Next, we turn to exploring the product market conditions that lead to the use of imprecise language in disclosures. We find that firms that face heightened product market threats – measured by the product market fluidity of Hoberg et al. (2014) – increase the use of imprecise language in their subsequent 10-K filing. We also find that firms use more imprecise language when they face greater financial constraints (measured by the textual financial constraints of Hoberg and Maksi- movic, 2015, and also two non-textual measures introduced by Whited and Wu, 2006 and Hadlock and Pierce, 2010). These findings suggest that disclosures with greater linguistic imprecision are more likely in the presence of adverse market conditions for the firm.

Digging deeper, we present evidence that the use of imprecise language in disclosures tends to reflect valuable information, which is a sharp contrast to managerial obfuscation. Specifically, our linguistic imprecision measure captures what appears to be additional information on upcoming positive but yet uncertain prospects. Bloomfield (2002) and Bushee et al. (2018) offer the similar view that complexity could be necessary to convey information about firms’ future performance.

Consistent with this interpretation, we find that firms whose 10-K disclosures contain greater linguistic imprecision subsequently have higher buy and hold abnormal returns (BHAR) and market valuations (measured by Tobin’s Q), and invest more in both R&D and capital expenditures. More- over, the prospects of firms with 10-K disclosures containing greater linguistic imprecision are more uncertain, but the greater uncertainty is due to greater upside opportunities. Specifically, we find that disclosure with greater linguistic imprecision is followed by higher return volatility, which is significantly more pronounced on the upside than the downside (Barndorff-Nielsen et al., 2010; Se- gal et al., 2015). Taken together, these findings suggest that the imprecise language in corporate disclosures reflects valuable, but yet fundamentally uncertain earnings opportunities.

Finally, we perform several tests that complement these main ﬁndings. First, we split our tests of ﬁnancial constraints into equity constraints and debt constraints as in Hoberg and Maksimovic

(2015). Interestingly, we find that equity constraints are associated with greater linguistic imprecision, but debt constraints are not. Second, we find that firms in which management has a greater ownership stake in the firm’s stock use more linguistic imprecision in their 10-K disclosures, partic-

3 ularly when facing severe ﬁnancial constraints. Third, we ﬁnd that disclosure with greater linguistic imprecision appears to lead to the overall positive market reaction and dampen the stock market reaction to information shocks (measured by earnings surprises, SUE, as in Livnat and Mendenhall,

2006) with the dampening effect only for negative earnings surprises. We further note that this positive stock return effect is not immediate and also does not lead to a subsequent reversal. In ﬁnding a slow but eventually positive market reactions to imprecise language in 10-Ks, this result com- plements our interpretation that imprecise language in corporate disclosures reﬂects fundamentally valuable but immature information.

Our analysis makes several contributions, which should be of general interest. First, our evidence on the use of imprecise language in disclosure relates to the work on discretionary disclosure and persuasion through information revelation (e.g., Bloomﬁeld, 2002). Discretionary disclosure leads to full disclosure in a perfect information environment, but not in the presence of proprietary costs or other market frictions (Ross, 1979; Verrecchia, 1983; Kamenica and Gentzkow, 2011; Ely,

2017). Following this line of research, recent empirical applications have focused on how the disclosure of bad news can signal quality (Gormley et al., 2012; Gao et al., 2017). Our results that show the informational value of imprecise language provide new perspective on this question. The recent work by Bushee et al. (2018) decomposes underlying latent components of linguistic complexity in the conference call setting into information and obfuscation and shows that information

(obfuscation) is negatively (positively) associated with information asymmetry. As we show in our evidence later, our explicit measure is more related to information than obfuscation (information asymmetry), and our results suggest that managers act in their decisions to provide more voluntary information.

Second, our identification and analysis of linguistic imprecision provide a unique perspective on the SEC regulatory mandate (the Plain Writing Act) to use plain English in firm disclosures, studied in Hwang and Kim (2017). Linguistic imprecision is not especially discouraged in this SEC mandate that regulates the readability of firm disclosure documents for the general public because imprecision words and phrases accord with plain English, yet imprecise language reflects the informational content of disclosures and thus affects firm value. In finding that linguistic imprecision

4 in corporate disclosures affects how investors react to the revelation of information, our analysis suggests that more attention to the use of plain English may be needed.

Third, our evidence on the role of product market threats (e.g., Hoberg et al., 2014; Cookson,

2017) and financial constraints (e.g., Hoberg and Maksimovic, 2015; Buehlmaier and Whited, 2017) in amplifying the managerial incentives to disclose more positive information in the immature state provides a new perspective on the financial market consequences of heightened competition and frictions in raising capital. Notably, these results suggest that tighter constraints and tougher competition have effects on the precision of information disclosures, which is a novel finding beyond kindred effects on the complexity of products (as in Celerier and Vallee, 2017 and Carlin et al.,

2012).

Finally, our work is part of a growing literature within finance and accounting that makes use of text descriptions to study important aspects of corporate behavior. Recent text-based analyses in corporate finance have examined linkages between firms and industries, corporate risk management, the value of corporate culture, and innovation (e.g., Popadak, 2013; Agarwal et al., 2016; Hoberg and Moon, 2017; Bellstam et al., 2017; Bushee et al., 2018). Within the broader literature on text analysis in finance, our work is most closely related to applying textual analysis tools to analyze the tone of financial information (Hanley and Hoberg, 2010; Dougal et al., 2012; Loughran and

McDonald, 2013; Garcia, 2013; Jegadeesh and Wu, 2017; Hoberg and Lewis, 2017). As we show in our regression evidence later, our measure is sensibly related to, but distinct from the existing lexicon of measures – many of which are available at the Loughran and McDonald (2011) master dictionary. Relative to these other textual measures, our linguistic imprecision measure provides a useful description of imprecise language in ﬁnancial disclosures, which is distinctive unto itself. In this respect, we anticipate fruitful applications of the linguistic imprecision measure to understand better market participants’ incentives to inject imprecision into the information environment.

The remainder of the paper proceeds as follows. Section 2 provides a description of Wikipedia’s weasel tags, the construction of the imprecision keyword list, and the development of our textual measure. Section 3 describes the main results relating linguistic imprecision to other textual measures, firm characteristics, and main findings on competitive threats and financial constraints. Sec-

5 tion 4 examines mechanisms that can explain our ﬁndings, particularly the role of equity constraints, market reactions to linguistic imprecision, and subsequent real outcomes. Section 5 concludes with future directions for research.

2 Data and Variable Construction

2.1 Wikipedia and Weasel Words

To construct our measure of linguistic imprecision, we take the entire Wikipedia articles as our text corpus to detect sentences of imprecise language and compile a list of imprecision keywords.

A study by Wikipedia (Ganter and Strube, 2009) suggests three categories of weasel words: 1) numerically vague expressions (e.g., "many"), 2) the passive voice (e.g., "it is said"), and 3) adverbs that weaken (e.g., "probably"). Examples of these weasel words directly given by Wikipedia as style guidelines include “People are saying...”, “There is evidence that...”, and “It has been mentioned that.”2 Wikipedia users are then advised to avoid using weasel words and at the same time to detect and mark excessive uses of such words by others using a special weasel tag, {{Weasel- inline|{{subst:DATE}}}} for improvement. The examples below illustrate how the weasel tag is used in a sentence of each article:

“The Tic Tok Men” • Many{{weasel inline|date=March 2009}} consider this album to be the quintessential Tic Tok

sound.

“Manu Parrotlet” • It has been said{{weasel inline|date=January 2014}} that the Manu parrotlet can be seen

along the Man on top of trees across from the Altamira beach about 25 minutes from the

Manu Resort.

“Nathaniel Mather” • 2See Wikipedia’s own article about weasel words for more details at https://en.wikipedia.org/wiki/Weasel_word.

6 He ﬁnished his studies in England probably{{weasel inline|date=January 2014}} returning

with his brother [[Samuel Mather (Independent minister)|Samuel]] in 1650.

We process a recent Wikipedia dump completed on April 20, 2017 comprised of 17,483,910 articles and extract sentences that contain weasel tags.3 To do so, we follow the methodology in Ganter and

Strube (2009) with the following modiﬁcations. While Ganter and Strube (2009) examine the ﬁve words occurring right before each weasel tag, we consider all words in sentences that contain weasel tags. We also further consider the frequencies of all those words and their bigrams and trigrams as well to better identify potential weasel words and phrases. The bigrams and trigrams are particularly useful to capture weasel phrases that use passive voice and appeals to anonymous authority.

Because weasel tags are removed after the language is edited and improved, the tags are not frequently observed at any given snapshot of Wikipedia. Therefore, sentences containing weasel tags are not abundant, despite the large number of articles we process. We identify 433 sentences with weasel tags in 367 articles after removing corrupt or redundant sentences. Our number of weasel tags is slightly more than 328 weasel tags of Ganter and Strube (2009) who processed two

Wikipedia dumps with different completion dates.

The numbers of unique and total words in the extracted sentences containing weasel tags are approximately 6,000 and 16,000, respectively. We sort the roughly 6,000 unique words and their bigrams and trigrams by frequencies and assesses whether each word or phrase correctly qualiﬁes for a weasel word. In this raw word frequency sort, commonly used words tend to show up as most frequent, despite not being weasel words themselves (e.g., words like “the”, “and”, and “that”). This is a larger issue with the unigrams than it is with the bigrams or trigrams. For example, Panel (a) of

Table 1 displays the three separate lists of the top 10 most frequently mentioned unigrams, bigrams, and trigrams in our weasel-tagged sentences.

[Insert Table 1 Here]

To ensure we do not merely pick up common language in our keyword lists, we extract a control sample of sentences that occur three sentences later in the text of the same articles. Upon manually

3Wikipedia dumps are available for downloading at https://dumps.wikimedia.org/.

7 inspecting these sentences, these control sentences are free of weasel language, and have the virtue that they are on the same set of topics as the weasel text. Using these control sentences together with the weasel-tagged sentences, we compute the saliency of the words in the weasel-tagged sentences relative to control sentences based on Goldsmith-Pinkham et al. (2016). The saliency measure captures the degree to which the words are overused relative to common language, and is thus, appropriate for screening our list of common language. Panel (b) of Table 1 shows how effective the saliency screen is in ﬁltering out common language from the list of words.4

After ﬁltering out common language using the saliency screen on unigrams, we compile our

ﬁnal list of imprecision keywords (unigrams, bigrams, trigrams). Further, we expand the list of imprecision keywords using variations on these words such as the singular and plural forms for nouns and the past, present, and future tenses for verbs. We also manually eliminate redundancy in bigrams and trigrams (especially) in cases where including both would count the same language twice.5

The dictionary of imprecision keywords is distinct from notable alternatives. Speciﬁcally, Panel

(c) of Table 1 presents the top 10 most frequently used keywords in the 10-Ks using our dictionary of imprecision keywords, and for comparison, the same list for uncertainty words and weak and strong modal words taken from the Loughran and McDonald (2011) master dictionary. The most frequently used words in each of these dictionaries have minimal overlap with one another, indicating that our linguistic imprecision measure using the imprecision keywords provides unique information distinct from these related measures. For example, numerically vague expressions such as “other”, “number of”, and “various” are uniquely included in the top 10 most frequently used imprecision keywords.6 Also, a number of passive expressions such as “said”, “considered”, and

4We also consider a re-weighted version of the measure using text frequency, inverse document frequency weights (tf-idf) that mirrors the intuition of the Goldsmith-Pinkham et al. (2016) saliency filter. Results using the saliency filter and tf-idf weighting are nearly identical. Despite this robustness to a more sophisticated methodology, we prefer to use the main measure of imprecision because it is more transparent, and involves fewer researcher choices. 5In addition, Wikipedia has published guidelines for weasel words, giving specific examples to help users identify weasel language. Our methodology captures the vast majority of the example phrases offered by Wikipedia, but several example phrases in the guidelines are not in the Wikipedia dump we analyze. To maintain the most comprehensive list of imprecision keywords, we also include these guideline weasel words in our final list. The complete list of imprecision keywords can be obtained by contacting the authors. 6The most frequently used unigram, “other”, can be simply mentioned in 10-Ks to refer to an accounting item that contains “other”, for example as in “Other Comprehensive Income”, “Assets - Other”, “Liabilities - Other - Total”. We

8 “found” are frequently used imprecision keywords in 10-Ks, although those are not included in the top 10 list. In robustness exercises, we construct our linguistic imprecision measure purged of uncertainty and weak modal words, and show that all the main conclusions of our analysis go through.

2.2 10-K Disclosure and Firm Linguistic Imprecision Measure

The final step in our text processing procedure is to download all 10-K filings with report dates from 1997 to 2015 and extract the raw counts of how many times a given firm mentions each of the imprecision keywords in a given year. This generates a full panel of imprecision keyword vectors with 219,491 firm-year observations. Our final sample is reduced to 44,207 firm-year observations after merging with Compustat data, CRSP data, and the product market threats and financial constraints data from Hoberg et al. (2014) and Hoberg and Maksimovic (2015), respectively. We create our main linguistic imprecision measure, Imprecision, based on the vectors of our imprecision keywords. Imprecision is how many times the imprecision keywords are mentioned (i.e., the sum of all elements in the imprecision keyword vector) in a given firm’s 10-K filing in a given year scaled by the total word count in the filing in the percentage term. Throughout the paper, we focus on

Imprecision as our main variable of interest.

To provide a contextual understanding of our imprecision measure, we brieﬂy discuss several

firm and industry examples of high versus low linguistic imprecision. We find that the industry that most uses imprecision keywords is Security brokers, dealers and flotation industry at the 3-digit

SIC code level. Imprecision is also high in Membership organizations, Telegraph and other mes- sage communications, Local and suburban transit and interurban passenger transportation, Legal services, and Drugs industries. The industry that least uses imprecision keywords is Cigarettes industry. Forestry, Miscellaneous furniture and ﬁxtures, Electronic and other electrical equipment and components, except computer equipment, Steam and air-conditioning supply, and Land subdividers and developers industries follow.

We also examine which firms most or least use imprecision keywords in their filings. xG Tech- note that our findings discussed later are robust to excluding “other” from our dictionary of imprecision keywords.

9 nology, Inc., a company that sells communications equipment, has the highest Imprecision. Essen- dant Inc, a wholesale company of paper and paper products, has the lowest Imprecision. We present below short business descriptions of the two ﬁrms.

xG Technology, Inc. (the most linguistically imprecise ﬁrm) The overarching strategy of xG Technology, Inc. (“xG Technology", “xG", the “Com-

pany", “we", “our", “us") is to design, develop and deliver advanced wireless commu-

nications solutions that provide customers in our target markets with enhanced levels of

reliability, mobility, performance and efﬁciency in their business operations and mis-

sions. xG’s business lines include the brands of Integrated Microwave Technologies

LLC (“IMT"), Vislink Communication Systems (“Vislink"), and xMax. There is con-

siderable brand interaction, owing to complementary market focus, compatible product

and technology development roadmaps, and solution integration opportunities. In ad-

dition to these brands, xG has a dedicated Federal Sector Group focused on providing

next-generation spectrum sharing solutions to national defense, scientiﬁc research and

other federal organizations.

Essendant Inc. (the least linguistically imprecise ﬁrm) Essendant Inc. (formerly known as United Stationers, Inc.) is a leading national whole-

sale distributor of workplace items including janitorial, foodservice and breakroom sup-

plies (JanSan), technology products, traditional ofﬁce products, industrial supplies, cut

sheet paper products, automotive products and ofﬁce furniture.

Overall, the picture that emerges from these examples is that ﬁrms and industries that use more imprecise language are more likely to involve business operations that require expressing shades of possibility. In contrast, the ﬁrms and industries that use less imprecise language are more likely to have business operations that are certain and unambiguous.

Although the examples in this section are anecdotes drawn from the extremes of our data, they are consistent with a more systematic analysis of the data. As evidence on this point, Table 2 presents sample splits and two-sample t-tests for various characteristics available in our data. Con-

10 sistent with these examples, disclosures that have greater linguistic imprecision tend to come from small and young ﬁrms that are more difﬁcult to quantify. In Section 3, we show that the relationships and characteristics that are important in these sample splits are also robust to industry and year

ﬁxed effects, as well as controls for other important factors.

[Insert Table 2 Here]

2.3 10-K Disclosures and Empirical Strategy

Before describing our empirical tests, it is important to comment on the meaning of imprecision within the context of 10-K disclosures relative to other potential source texts. There are two notable features to discuss: within-ﬁrm persistence and the care that goes into 10-K disclosures.

First, regarding within-ﬁrm persistence of the measure, it is well known that the 10-K disclosures are highly persistent over time. In light of the persistence of 10-K disclosures, we expect the extent of imprecision in the 10-K disclosures contains more information about cross-ﬁrm differ- ences in characteristics than time-series changes. At the same time, recent work by Cohen et al.

(2016) has shown that there is information content in the minor changes from year to year that eventually is capitalized into asset prices. In light of this nature of underlying variation in the 10-K language, our main specifications that relate imprecision to economic considerations that firms face use industry fixed effects (to focus on relevant cross-firm variation in imprecision, which comprises the majority of the variation in our measure). In addition, we also estimate specifications with firm

fixed effects, recognizing that these specifications throw away most of the variation in language from the 10-Ks. These specifications rely on changes in language from year to year, and thus, map into similar variation studied by Cohen et al. (2016).

Second, we expect that the imprecision measure based on 10-K disclosures – which are required by Regulation S-K to include any information with material effects on the firm’s financial condition or results of operations, carefully curated by the firm’s legal team, and also should be audited– is likely different than other source texts that do not have the same degrees of difficulty of censoring and ex ante scrutiny (e.g., the question and answer portion of the earnings call). Because of this high degree of care in preparing the 10-Ks, imprecise language in the 10-Ks is more deliberate than

11 other source texts. With this background in mind, we expect our imprecision measure based on

10-K disclosures to contain genuine information that – because of market conditions or timing – is not possible to make precise at the time of the disclosure. This information is distinctively useful from the standpoint of investors in evaluating the likely consequences of adverse conditions that the

ﬁrm faces. It is important to keep this interpretation in mind when interpreting our tests on how linguistic imprecision relates to product market threats and ﬁnancial constraints.

3 Main Results

3.1 Relation to Other Textual Measures

Correlations with existing textual measures in the literature help validate our imprecision measure.

Although imprecise language is distinct from uncertainty and weak modal language, it should be positively related to uncertainty and weak modal language. Beyond the partial overlap in the word dictionary, we expect ﬁrms to use imprecise language in disclosure at times and in situations in which there is greater uncertainty. For this reason, we anticipate imprecision to positively asso- ciate with uncertainty and weak modal words, both of which indicate environments with greater uncertainty.

We validate this intuition of imprecision using data on uncertainty words, and modal words from the Loughran and McDonald (2011) master dictionary. Portraying a series of univariate comparisons to the use of imprecision words, Figure 1 presents side-by-side box plots of the amount of linguistic imprecision in 10-Ks by whether uncertainty, weak modality, and strong modality are above versus below the median. These plots show that linguistic imprecision is more commonly used in high uncertainty, high weak modality, and low strong modality 10-Ks. Although these relationships are strong, there is also useful residual variation in linguistic imprecision, holding constant other textual measures. The side-by-side box plots in Figure 1 also show this substantial overlap in the distributions of imprecision for high and low uncertainty, weak modality, and strong modality.

[Insert Figure 1 Here]

12 To examine these associations more systematically, we estimate the following regression speci-

ﬁcation:

Imprecisionit = a + b1Pct Uncertainit + b2Pct Weak Modalit + b3Pct Strong Modalit + ds + gt + hXit + eit ,

where Imprecisionit is the percentage of imprecision keywords (out of total words) used in ﬁrm i’s

10-K disclosure in year t, Pct Uncertainit , Pct Weak Modalit , and Pct Strong Modalit are the per- centages of uncertain words, weak modal words, and strong modal words taken from the Loughran and McDonald (2011) master dictionary, ds are SIC3 industry fixed effects, gt are year fixed effects, and Xit are control variables including textual measures for sentiment (the difference between percent positive words and percent negative words), interesting words, superfluous words, litigious words, constraining words, and fog words, and in some specifications, controls for lagged firm characteristics taken from Compustat. To account for serial correlation over time, the specifications cluster standard errors by firm.7

The multiple regression evidence in Table 3 shows that weak modality and uncertainty are each positively associated with the use of linguistic imprecision, even in a regression controlling for the other textual measures. In addition, strong modality is negatively associated with imprecision.

Interestingly, sentiment, measured by the percentage of positive words minus the percentage of negative words, is positively associated with imprecision, being consistent with the view that firms use less precise expressions when they discuss positive prospects that have not realized in negative situations. As the specifications in columns (2) through (4) show, these associations are robust to accounting for other available textual measures (the fog measure, interesting words, superfluous words, litigious words, and constraining words) and firm characteristics (ROA, Tobin’s Q, Sales growth, R&D/Sales, CAPX/Sales, and Leverage). Furthermore, the specification in column (5) shows that these associations are robust (in magnitude and significance) to including firm fixed effects.8 7Variable definitions in detail are given in Appendix Table A.1. 8We have conducted two additional tests for robustness for this specification. First, we have also conducted the

13 [Insert Table 3 Here]

Taken together, these ﬁndings validate that the content of our linguistic imprecision measure applied to the 10-K disclosures captures the underlying idea of imprecision. The evidence here suggests that – consistent with the motivating idea of weasel words from Wikipedia – imprecision language in the 10-Ks captures relatively positive tone with high uncertainty and high modality.

Because these aspects of the text are – to a large degree – part of the content of linguistic imprecision, we do not control for these measures in our main tests. We do, however, show robustness to controlling for these previously understood aspects of language.

3.2 Relation to Firm Characteristics

Ex ante, imprecise language ought to be more frequently used by firms that have more intangible assets or business models, and those that are otherwise difficult to quantify. We examine this intuition by relating lagged firm characteristics to our linguistic imprecision measure.

Figure 2 presents 95% confidence intervals for several notable firm characteristics by each quartile of the distribution of linguistic imprecision used in 10-K disclosures. Consistent with the notion that intangible firms are more likely to use imprecise language, younger and smaller firms with more growth opportunities (measured by higher Tobin’s Q and higher sales growth) tend to use less precise language. In a similar vein, firms that make more intangible investments, measured by

R&D/Sales, also tend to use linguistic imprecision with greater frequency. In addition, ﬁrms with lower proﬁtability tend to use more imprecision.

[Insert Figure 2 Here]

To examine these associations with firm characteristics more systematically, we estimate the following regression specification: analysis at the paragraph level, reaching the same conclusions about how linguistic imprecision relates to uncertainty and modal language. In the paragraph-level analyses, we are able to control for firm-year (i.e., report level) fixed effects, identifying only on the variation within 10-K disclosure. Second, beyond normalizing by calculating the percentage of imprecision keywords in our imprecision measure, we have run all of the specifications controlling for the log of the total number of words in the report related to readability in financial disclosure (e.g., Loughran and McDonald, 2014), and the findings are robust.

14 Imprecisionit = a + b1ROAit 1 + b2Firm Sizeit 1 + b3Tobin0sQit 1 + ds + gt + hXit 1 + eit ,

where Imprecisionit is the percentage of imprecision keywords (out of total words) used in ﬁrm i’s

10-K disclosure in year t, ROAit 1 is the return on assets for firm i in year t 1, Firm Sizeit 1 is the market valuation of firm i in year t 1, and Tobin0sQit 1 is the market to book ratio of firm i in year t 1 (included as a measure of growth opportunities anticipated by market investors), d s are SIC3 industry fixed effects, gt are year fixed effects, and Xit 1 are other controls for lagged firm characteristics taken from Compustat, as well as textual measures taken from the Loughran and

McDonald (2011) master dictionary. To account for serial correlation over time, the speciﬁcations cluster standard errors by ﬁrm.

The multiple regression evidence in Table 4 shows broadly that the associations between profitability, intangibility, growth opportunities and the use of linguistic imprecision indicated in Figure

2 are also present in the regression specifications that control for other firm characteristics, industry and year fixed effects, and the other textual measures present in the literature. Specifically, the positive and significant slope coefficient estimate of on ROA implies that a decrease in operating profitability is associated with an increase in the use of linguistic imprecision. The association with a proxy for growth opportunities measured by in Tobin’s Q is similar, exhibiting a slope coefficient estimate with a two times larger magnitude than that of profitability.

[Insert Table 4 Here]

In addition, column (3) presents a specification with firm fixed effects, which rely on within-firm variation in the linguistic imprecision measure (throwing out the majority of the variation, which tends to be focused across firms). Using this within-firm variation, these specifications paint a similar picture of the types of firms, though the magnitudes and statistical significance are weaker, as expected, especially for ROA. Although these regressions do not pin down causation precisely, they provide robust evidence on the types of firms that use linguistic imprecision and conditions under which firms use more imprecise language in disclosure. Two major themes emerge from these re-

15 sults on firm characteristics and linguistic imprecision: (1) firms with low profitability use more imprecise language in disclosure, and this finding is not explained by firm size, investment opportunities, and firm lifecycle, (2) proxies for intangibility (small, young firms with R&D investments and high market -to-book ratios) are positively associated with imprecise language in disclosure. We consider analogous tests using the MD&A sections of 10-K filings only and present results in the

Appendix Table A.2. Results using entire 10-K texts and the MD&A section texts are qualitatively similar.

3.3 Relation to Product Market Threats and Financial Constraints

Moving beyond the descriptive analysis, we now turn to a more systematic analysis of the incentives to use imprecise language in financial disclosures by relating our linguistic imprecision measure to proxies for product market threats and financial constraints. The analysis in this section is informative for an important strand of the corporate finance literature that has paid particular attention to how corporate policies relate to product markets and financial constraints. For example, Phillips

(1995) shows- that industries with greater leverage exhibit less aggressive behavior, and Khanna and Tice (2000) present evidence that ﬁrms with greater leverage retrench upon facing a threat (in the context of Walmart’s nationwide expansion). More recently, Cookson (2017) shows that leverage constrains the investment opportunities of casino ﬁrms that are facing entry threats, and on the

ﬁnancial side, Hoberg et al. (2014) show important changes in payouts in response to an increase in product market threats.

In this context, we examine how ﬁrms’ use of imprecise language in their disclosures changes upon facing greater product market threats (via product market ﬂuidity in Hoberg et al., 2014).

Relevant to this point, Panel (a) of Figure 3 presents a 95% conﬁdence interval of product market

ﬂuidity for each quartile of linguistic imprecision in 10-K disclosures. As the use of imprecise language increases, product market ﬂuidity increases as well, indicating a strong, positive relation.

[Insert Figure 3 Here]

Related to this notion that product market threats place pressure on firms to increase the use of imprecise language, firms that use more linguistic imprecision also tend to face greater financial

16 constraints. Panel (b) of Figure 3 shows this positive relation via plots of 95% confidence intervals of the financial constraints measure of Hoberg and Maksimovic (2015) by quartiles of linguistic imprecision. As with product market fluidity, there is a strong positive relation between financial constraints and the choice to use imprecise language in 10-K disclosures. These univariate comparisons on product market threats and financial constraints are consistent with the view that greater imprecision in 10-Ks can be a consequence of tighter profit margins.

To examine this rationale systematically, we estimate the following regression speciﬁcation:

Imprecisionit = a + b1Product Market Fluidityit 1 + b2Financial Constraintsit 1 + ds + gt + hXit 1 + eit ,

where Imprecisionit is the percentage of imprecision keywords (out of total words) used in ﬁrm i’s

10-K disclosure in year t, Product Market Fluidityit 1 is the measure of product market fluidity from Hoberg et al. (2014), Financial Constraintsit 1 is Hoberg and Maksimovic (2015)’s text- based measure of financial constraints (including separate measures for equity constraints and debt constraints), ds are SIC3 industry fixed effects, gt are year fixed effects, and Xit 1 are controls for lagged firm characteristics taken from Compustat. To account for serial correlation over time, the specifications cluster standard errors by firm.

The multiple regression evidence in Table 5 indicates that the simple associations between product market fluidity, financial constraints, and linguistic imprecision indicated in Figure 3 are robust to a more systematic approach that controls for industry and year fixed effects and firm characteristics. According to the specification in column (1) of Panel (a) of Table 5, a standard deviation increase in product market fluidity is associated with an increase of approximately one fifth of a standard deviation in linguistic imprecision (=0.44%). This estimated relation is statistically significant at the one percent level, and is robust to the inclusion of industry and year fixed effects. The magnitude and significance of this estimated coefficient is also stable and robust to the inclusion of

firm characteristics in columns (3) and (4) and firm fixed effects in columns (2) and (4).

[Insert Table 5 Here]

17 Similar to product market fluidity, financial constraints are robustly related to the use of imprecise language. From the specifications in Panel (a) of Table 5, a standard deviation increase in financial constraints is associated with approximately one tenth of a standard deviation more linguistic imprecision (=0.44%). The specifications with firm fixed effects in columns (2) and (4) show that these relations are robust in statistical significance – though slightly smaller in magnitude

– when focusing only on the relatively sparse within-ﬁrm variation.

Although controlling for other textual measures – as we discussed earlier – is partly over- controlling in the sense that it holds constant factors that partly deﬁne what linguistic imprecision means, it is useful to show that from the standpoint of the literature that our linguistic imprecision measure has content unto itself that is not contained in a recombination of previously understood textual measures. To evaluate this robustness to using alternative textual measures, we control for the full complement of other textual measures used in the literature (not just sentiment, uncertainty, weak modality, and strong modality) in Appendix Table A.3. Regardless of the set of controls, we

find that product market fluidity and financial constraints each exhibits a positive and statistically significant relation with linguistic imprecision after controlling for the full suite of other textual measures. These findings indicate that our linguistic imprecision contributes useful information beyond existing textual measures of tone.9

We also consider analogous tests using the MD&A sections of 10-K filings only and present results in the Appendix Table A.4. Results using the entire 10-K texts and the MD&A section texts are qualitatively similar for these tests. As another robustness check, in the Appendix Table A.5, we obtain similar results using the linguistic imprecision measure purged of uncertainty and weak modal words.10 9The findings here show that the additional information from linguistic imprecision is economically meaningful, and thus, motivates its use beyond our setting. Further supporting this interpretation, we find in column (4) of Appendix Table A.3 that the relation between imprecision and adverse market conditions (product market threats and financial constraints) is present even after accounting for other textual measure, time-varying firm characteristics, and firm fixed effects. 10One potential concern with our specification is that we employ a text-based proxy for financial constraints as an explanatory variable for linguistic imprecision, which is also a text-based measure. As these proxies are built from the same 10-Ks, it is a natural concern that the measures have a common source of error, which biases the regression estimates. We address this concern by relating our measure of imprecision to non-textual measures of financial constraints (specifically, we examine the Whited and Wu, 2006 financial constraints index and the SA Index of Hadlock and Pierce, 2010). As the results in the Appendix Table A.6 show, we obtain similar insights from this analysis, which validates our use of the text-based financial constraints measure in the main analysis.

18 Digging deeper, we follow Hoberg and Maksimovic (2015) and examine how equity constraints and debt constraints differ in their relation to linguistic imprecision. Interestingly, as the results in

Panel (b) of Table 5 show, the positive relation between ﬁnancial constraints and the use of imprecise language in disclosure comes from equity constraints, not debt constraints. As in Panel (a), these

findings are robust in statistical significance to using firm fixed effects in specifications (columns

2 and 4) that focus on within-firm changes to linguistic imprecision. The differential response of linguistic imprecision to equity versus debt financial constraints suggests that the incentive to use imprecise language in disclosure comes from its potential effects on equity valuations. This is a natural finding in light of the importance of information disclosure in the 10-Ks for equity markets relative to debt markets.

Further, as a contextual validation of these ﬁndings, we obtain several examples of the content of sentences that contain both imprecise language and economic constraints. For example, when faced by poor economic conditions in 1996-1997, Matson, Inc. disclosed the following in their

10-K in 1997:

Sales again will be challenged by adverse economic conditions, but it is reasonable to

anticipate some potential upside, based upon unplanned sales opportunities that may

present themselves, just as several did during 1996.

Another example comes from Claire’s Stores, Inc., which disclosed the following in 1997:

Although the Company faces competition from a number of small specialty store chains

and others selling fashion accessories, in addition to one chain of approximately 800

stores, the Company believes that its Fashion Accessory Stores comprise the largest

and most successful chain of specialty retail stores in the World devoted to the sale of

popular-priced women’s fashion accessories.

These examples are consistent with our bottom-line interpretation that product market threats and

financial constraints bring about deteriorating business conditions, while at the same time present- ing the firm with an opportunity to defer responsibility to a third party. Importantly, given that the declining performance of the firm needs to be disclosed, our evidence underscores the natu-

19 ral incentive to disclose potentially positive information imprecisely (weakening tone, non-speciﬁc attribution, and overall imprecision) when these conditions present themselves.

4 Mechanisms

4.1 Equity Market Incentives

In this section, we investigate more in depth the role of equity market constraints in driving the use of imprecise language in disclosure by examining how linguistic imprecision in 10-Ks depends on the degree to which management compensation is aligned with equity market performance of the

firm’s stock. More specifically, we consider these management incentives to use imprecise language in disclosure by estimating variants of the following regression specification:

Imprecisionit = a + b1High Incentivesit 1 + b2High Incentivesit 1 Financial Constraintsit 1 ⇥

+ b3Financial Constraintsit 1 + b4Product_Market_Fluidityit 1 + ds + gt + hXit 1 + eit ,

where Imprecisionit is the percentage of imprecision keywords (out of total words) used in ﬁrm i’s

10-K disclosure in year t, High Incentivesit 1 is an indicator variable that equals one if the fraction 11 of the manager’s compensation that is paid in stock is in the top quartile, Product Market Fluidityit 1 is the measure of product market ﬂuidity from Hoberg et al. (2014), Financial Constraintsit 1 is the

Hoberg and Maksimovic (2015) text-based measure of ﬁnancial constraints, ds are SIC3 industry

fixed effects, gt are year fixed effects, and Xit 1 are controls for lagged firm characteristics taken from Compustat. To account for serial correlation over time, the specifications cluster standard errors by firm.

In these specifications, the main effect b1 and the interaction b2 are informative of management’s equity market incentives. In specifications without the interaction terms, b1 measures the degree to which firm managers with stronger sensitivity to equity markets are more or less prone to using

11We do not employ firm fixed effects in these specifications because the variation in management incentives is persistent (even more so than textual disclosures in the 10-Ks), leaving very little useful variation within firm.

20 imprecise language in disclosure. The interaction term informs the degree to which these equity market incentives interact with the financing constraints measure from the previous section. If the equity market incentives are important in explaining managers’ use of imprecise language in disclosure, we should expect b1 to be positive in non-interacted specifications and b2 to be also positive in interactive specifications. Consistent with this prediction, the univariate evidence in

Figure 4 shows there is generally positive association between imprecision and the fraction of the

ﬁrm owned by its manager. Managers with stronger equity market incentives tend to use much more imprecise language in their disclosures.

[Insert Figure 4 Here]

The multiple regression evidence in Table 6 shows that there is a significant positive association between management equity market incentives and imprecise language, even after accounting for industry and year fixed effects, product market fluidity, and financial constraints. The estimated coefficient on the high incentives dummy in column(1) of Table 6 is 0.02, which is comparable but slightly smaller in magnitude to change in the use of imprecise language implied by a standard deviation increase in product market fluidity or financial constraints, and this estimate is statistically significant at the five percent level. Moreover, the interaction term with financial constraints is also positive and statistically significant. For a standard deviation above the mean of financial constraints, the interaction magnitude is nearly identical with the main effect of manager incentives.

Beyond the main effects of management incentives and financial constraints, this interaction – which captures the additional response to high financial constraints by managers with high equity market incentives – suggests that managers increase their usage of imprecise language to partially mitigate the costs of high financial constraints.

[Insert Table 6 Here]

4.2 Market Reactions to Imprecise Language in Disclosure

Thus far, our evidence indicates that ﬁrms that are performing worse (lower ROA), facing greater competitive threats in the product markets, and facing greater ﬁnancial constraints tend to employ

21 more imprecise language in their 10-Ks. In addition, managers with stronger equity market incentives are more likely to use imprecise language in firm disclosures, especially when financial constraints are high. These findings collectively suggest that imprecision can help mitigate the adverse consequences to stock market valuation when a firm faces greater competitive threats and

ﬁnancial constraints. Thus, we now examine how stock returns react to linguistic imprecision in

10-K disclosures.

Speciﬁcally, for each 10-K release, we compute the BHAR in weekly windows (counting calendar days) and relate these abnormal returns to our linguistic imprecision measure in the following regression speciﬁcation:

BHARitn =b1Imprecisionit + hXit + eitn,

where BHARitn is the buy and hold abnormal return in the window of the nth week after the 10- K release date in year t (following existing studies, the 1st week window starts from the four days after the 10-K release date), Imprecisionit is the percentage of imprecision keywords (out of the total words) used in firm i’s 10-K disclosure in year t, and Xit is a vector of controls employed in prior work (e.g., Loughran and McDonald, 2011), including the percentage difference between positive words and negative words (sentiment), log of book-to-market, percentage of institutional holdings, pre-filing date Fama-French alpha, log of share turnover, log of market capitalization on the day prior to the 10-K filing date, and most recent standardized unexpected earnings (SUE). We estimate this BHAR regression model for each week separately from the 1st through 10th weeks after the 10-

K release date. We use clustered standard errors by month to account for cross-sectional correlation of returns.

The coefficient of interest in this specification is b1, which captures the extent to which the market reacts to imprecise language in disclosure. We expect b1 > 0, and if the positive market reaction is based on fundamental information, we predict that this effect is not immediate and does not revert afterwards. The evidence from estimating the BHAR specification in Panel (a) of Table

7 confirms this intuition. We find a positive and significant coefficient estimate for Imprecisionit

22 only from the third through seventh weeks. The coefficient estimates for Imprecisionit in weeks eight and nine are positive but statistically insignificant, which indicates that the positive market reaction to imprecise language in disclosure does not revert. Panel (b) of Table 7 reports analogous test results over multiple-week BHAR windows to capture cumulative market reactions. The positive return effect of imprecise language in disclosure cumulatively emerges from the 3rd week and monotonically increases afterward. The economic interpretation of the positive return coefficient estimate, for example as in column (7), is that a one standard deviation shift in the use of imprecise language in disclosure (=0.48%) is associated with a nearly 0.9% higher cumulative BHAR through the 7th week after the release of disclosure. In Appendix Table A.7, we also find similar market reaction results by additionally controlling for R&D and a financial constraints measure. Including these additional control variables in the market reaction analysis shows that our results cannot be fully explained by the findings in Li (2011) and Whited and Wu (2006), respectively.

[Insert Table 7 Here]

As an alternative presentation of these results, Figure 5 plots cumulative BHARs in the following weeks after 10-K releases by high-imprecision versus low-imprecision disclosures. The solid black line shows that the cumulative returns increase approximately 5% over the 10-week period following highly imprecise 10-K releases, relative to a 3% increase for low-imprecision 10-Ks. The difference in cumulative returns between high and low imprecision disclosures is 1.8% over 10 weeks after the

10-K release dates. We also confirm in this graphical illustration that the positive return reaction does not lead to a return reversal. The positive and persistent market reaction to imprecise language in disclosure suggests that the imprecision in 10-Ks provides fundamentally valuable information that can help firms under adverse economic conditions in equity valuations, and the findings deepen our understanding of the greater use of linguistic imprecision in reaction to high product market threats and greater financial constraints.12

12To ensure that these market reactions are not driven by future information releases through possible earning announcements in the next 3rd to 7th weeks, we repeat an analogous test with a more refined sample that excludes all 10-K filings that have new earnings announcements during the 3rd to 7th week period after the 10-K release dates. Results in Appendix Table A.8 confirm that the positive market reaction in Table 7 is not due to future earning announcements during the corresponding period. We also consider an event study of the market reactions, especially to negative shocks and examine whether these market reactions are dampened by the use of imprecise language in 10-Ks. In Appendix Table

23 [Insert Figure 5 Here]

4.3 Subsequent Upside and Downside Return Volatility

The fact that the market eventually responds positively to the use of imprecise language in disclosure, and the market returns do not revert, suggests that there is value-relevant information contained in managers’ using linguistically imprecise language in corporate disclosure. To understand this relation on a deeper level, we examine the difference in the good and bad volatility components of subsequent stock returns. If the mechanism for the positive market reactions is that ﬁrms are releas- ing good but immature information during bad times, we expect there to be greater upside return volatility, capturing ﬁrm’s growth opportunities, relative to downside return volatility, capturing

firm’s risk. We also expect stronger market valuation proxied by Tobin’s Q and more investment in the future as in Segal et al. (2015). We first examine the relation between linguistic imprecision and the difference in the good and bad components of the subsequent return volatility using the following regression specification:

RetVolDi f fitm = b1Imprecisionit + gtm + hXit + eitm,

where RetVolDi f fitm is the net upside return volatility measure of interest (which is deﬁned as the difference between upside and downside return volatilities following Barndorff-Nielsen et al.,

2010; Feunou et al., 2011; Segal et al., 2015) for ﬁrm i over the window of future m months after the 10-K release date in year t(the window starts from the four days after the 10-K release date).

The right-hand side variable of interest is Imprecisionit , the percentage of imprecision keywords

(out of total words) used in ﬁrm i’s 10-K disclosure in year t, gtm is month ﬁxed effect, and Xit is a vector of controls, including the percentage difference between positive words and negative words

A.9, we compute the BHAR in a window from 0 to 180 days and relate these abnormal returns to our measure of imprecise language in disclosure and the recent SUEs. We find a negative and significant slope coefficient on the interaction term between the linguistic imprecision measure and SUE, which opposes the estimated positive and significant main effect of SUE in sign. Examining the split sample results for negative versus positive SUEs, the estimated interaction is only statistically significant in the negative SUE subsample. These results reinforce our findings that linguistic imprecision in disclosures can help insulate equity valuations from negative earnings information.

24 (sentiment), pre-filing date return volatility over the past one month before the 10-K release date, pre-filing date Amihud illiquidity measure over the past one month before the 10-K release date, log of book-to-market, and log of market capitalization on the day prior to the 10-K filing date. To account for serial correlation over time, the specifications cluster standard errors by firm.

The regression evidence in Table 8 is consistent with our hypothesis that firms disclose good but immature information in their disclosures using linguistic imprecision. Namely, we find a robust, significant, and positive relation between the use of imprecise language in current disclosure and subsequent net upside return volatility. Specifically, the economic impact of a standard deviation increase in the use of imprecise language (=0.48%) on net upside return volatility translates into approximately one fifth of the impact of a standard deviation increase in pre-filing date return volatility

(=1.87%). The ﬁndings are robust to controlling for pre-ﬁling date return volatility and illiquidity which are well-known predictors of future return volatilities in the literature.

[Insert Table 8 Here]

4.4 Subsequent Market Valuation and Real Investment

Appealing to the theoretical framework of Segal et al. (2015) on the roles of upside versus downside volatility, we expect linguistic imprecision to be associated with greater subsequent output and investment if imprecision tends to reflect future upside volatility. In this case, we also expect linguistic imprecision to be positively related to subsequent valuation ratios at the individual firm level, which contrasts with the prediction drawn if linguistic imprecision reflects downside volatility. Although both upside and downside volatilities positively contribute to asset returns, the eventual economic growth effects in a firm differ between the upside and downside return volatilities. Hence, as a complement to the evidence on the net upside return volatility, in Table 9 we now examine the relation between the use of imprecise language in disclosure and subsequent valuation as measured by

Tobin’s Q and investment opportunities captured by both R&D intensity and capital expenditures.

Speciﬁcally, we estimate variants of the following regression speciﬁcation:

25 Yit+1 = a + b1Imprecisionit + ds + gt + hXit + eit+1,

where Yit+1 is the future corporate outcome of interest (Tobin’s Q, R&D, or capital expenditures) for firm i in year t + 1, Imprecisionit is the percentage of imprecision keywords (out of total words) used in firm i’s 10-K disclosure in year t, ds are SIC3 industry fixed effects, gt are year fixed effects, and Xit are controls for firm characteristics taken from Compustat. To account for serial correlation over time, the specifications cluster standard errors by firm.

The regression evidence in Panel (a) of Table 9 is consistent with the event study evidence in

Tables 7 and 8in that it shows that firms with greater linguistic imprecision in disclosure today have higher long term valuations as measured by Tobin’s Q. Specifically, a standard deviation increase in imprecision is associated with 0.036 higher Tobin’s Q in column (1). In addition, firms with the greater use of imprecise language tend to also invest more in R&D and capital expenditures in subsequent years. These estimates control for firm characteristics, as well as industry and year fixed effects. The findings are robust to the use of firm fixed effects for Tobin’s Q and R&D in columns

(2) and (4), but the relation to capital expenditures becomes weaker and statistically insignificant with firm fixed effects in column (6).13

[Insert Table 9 Here]

Panel (b) of Table 9 further provides evidence that the positive effects of linguistic imprecision on predicting subsequent market valuations and real investments are indeed from conveying future information. To show this, we additionally introduce to the same regression speciﬁcations in Panel

(a) of Table 9 a measure of forward-looking disclosure, which is the number of forward-looking disclosure keywords from Muslu et al. (2015) scaled by the total word count in the ﬁling in the

13Moreover, related to potential downside information, we find in unreported results that our measure of linguistic imprecision in disclosure is not associated with the increased likelihood of receiving SEC comments letters or various kinds of corporate litigations. Using the comment letter data from the SEC EDGAR website for the sample period from 2006 to 2014 and the corporate litigation data from the Audit Analytics database for the sample period from 1997 to 2014, we find that linguistic imprecision in 10-Ks rather predicts less incidences of SEC comment letters and is not significantly related to the incidences of corporate lawsuits.

26 percentage term, and its interaction variable with our imprecision measure. If the positive effects of linguistic imprecision on subsequent financial and real outcomes are from managers’ conveying information about upcoming positive prospects of their firms, we should expect both coefficient estimates for the measure of forward-looking disclosure and its interaction term with the imprecision measure to be positive. The regression results in Panel (b) of Table 9 are consistent with this prediction. In general, the measure of forward-looking disclosure is positively related to Tobin’s

Q, R&D intensity, and capital expenditures, and the interaction term is also positively and more strongly related to those measures of growth and investment opportunities. The speciﬁcations with

firm fixed effects in columns (2), (4), and (6) of Panel (b) show that these relations in the interaction term are robust (both in magnitudes of the effects and statistical significance) to focusing only on the within-firm variation. The stronger effect of the interaction term effectively supports our intuition that firms are more likely to introduce linguistic imprecision when conveying forward-looking information that is by nature less specific and precise.

Taken together, the positive relation of linguistic imprecision in disclosure to market valuations and real investment – together with the evidence on market reactions, the net upside return volatility, and the combination of forward-looking information measure – suggests that there are useful real and ﬁnancial implications of the use of imprecise language in disclosure and managers will consider those effects when deciding appropriate disclosure responses to given business situations and economic conditions.

5 Conclusions

In this paper, we introduce a novel textual measure to the ﬁnance and accounting literature, which captures the degree of linguistic imprecision in ﬁrm disclosures. The measure is distinct from uncertainty, sentiment, and other textual measures that are previously developed in the literature.

On its own, our measure of linguistic imprecision has signiﬁcant explanatory power for identifying additional qualitative information in ﬁrm disclosures beyond quantitative information.

27 We find that firms are more likely to use imprecise language when faced with product market threats and financial constraints, and that the long-term market reaction to imprecise language in disclosure is positive. These findings suggest that firms use imprecise language to convey valuable information that is nonetheless too immature to make precise. Paradoxically, linguistic imprecision allows firms to make valuable disclosures instead of obfuscating. In contrast to obfuscation, we find that imprecise language in corporate disclosure is valued by investors as it implies managers’ dis- closing potential upside opportunities of the firms. Collectively, our findings and approach suggest that there is much to learn from the qualitative content of firm disclosures.

28 References

Agarwal, S., S. Gupta, and R. D. Israelsen (2016). Public and private information: Firm disclosure,

sec letters, and the JOBS Act. Working Paper SSRN 2891089.

Barndorff-Nielsen, O. E., S. Kinnebrock, and N. Shephard (2010). Measuring downside risk-

realised semivariance. Working Paper.

Bellstam, G., S. Bhagat, and J. A. Cookson (2017). A Text-Based Analysis of Corporate Innovation.

Working Paper.

Bloomﬁeld, R. J. (2002). The "incomplete revelation hypothesis" and ﬁnancial reporting. Account-

ing Horizons 16(3), 233–243.

Buehlmaier, M. M. M. and T. M. Whited (2017). Are ﬁnancial constraints priced? Evidence from

Textual Analysis. Working Paper.

Bushee, B. J., I. D. Gow, and D. J. Taylor (2018). Linguistic complexity in ﬁrm disclosures: Obfus-

cation or information? Journal of Accounting Research 56(1), 85–121.

Carlin, B. I., S. W. Davies, and A. Iannaccone (2012). Competition, comparative performance and

market transparency. American Economic Journal: Microeconomics 4, 202–237.

Celerier, C. and B. Vallee (2017, August). Catering to investors through security design: Headline

rate and complexity. Quarterly Journal of Economics 132(3), 1469–1508.

Cohen, L., C. Malloy, and Q. H. Nguyen (2016). Lazy prices. Working Paper SSRN 1658471.

Cookson, J. A. (2017, February). Leverage and Strategic Preemption: Lessons from Entry Plans

and Incumbent Investments. Journal of Financial Economics 123(2), 292–312.

Dechow, P. M., R. G. Sloan, and A. P. Sweeney (1996). Causes and consequences of earnings

manipulation: An analysis of ﬁrms subject to enforcement actions by the SEC. Contemporary

Accounting Research 13(1).

29 Dougal, C., J. Engelberg, D. Garcia, and C. A. Parsons (2012, March). Journalists and the stock

market. Review of Financial Studies 25(3), 639–679.

Ely, J. C. (2017). Beeps. American Economic Review 107(1), 31–53.

Feunou, B., M. R. Jahan-Parvar, and R. Tédongap (2011). Modeling market downside volatility.

Review of Finance 17(1), 443–481.

Ganter, V. and M. Strube (2009). Finding Hedges by Chasing Weasels: Hedge Detection Using

Wikipedia Tags and Shallow Linguistic Features. In Proceedings of the ACL-IJCNLP 2009 Con-

ference Short Papers, Number 2044, pp. 173–176. Association for Computational Linguistics.

Gao, J., C. Liang, K. Merkley, and J. Pacelli (2017). Do lenders promote the revelation of bad news?

Evidence from Lender-side loan defaults. Working Paper.

Garcia, D. (2013). Sentiment during recessions. Journal of Finance 68(3), 1267–1300.

Goldsmith-Pinkham, P., B. Hirtle, and D. Lucca (2016). Parsing the content of bank supervision.

Working Paper FRBNY Staff Report No. 770.

Gormley, T. A., B. H. Kim, and X. Martin (2012, March). Do ﬁrms adjust their timely loss recog-

nition in response to changes in the banking industry? Journal of Accounting Research 50(1),

159–196.

Hadlock, C. J. and J. R. Pierce (2010). New Evidence on Measuring Financial Constraints: Moving

Beyond the KZ Index. Review of Financial Studies.

Hanley, K. W. and G. Hoberg (2010). The information content of IPO prospectuses. Review of

Financial Studies 23(7), 2821–2864.

Hoberg, G. and C. Lewis (2017, April). Do fraudulent ﬁrms produce abnormal disclosure? Journal

of Corporate Finance, Forthcoming.

Hoberg, G. and V. Maksimovic (2015, May). Redeﬁning ﬁnancial constraints: A text-based analy-

sis. Review of Financial Studies 28(5), 1312–1352.

30 Hoberg, G. and S. K. Moon (2017, August). Offshore Activities and Financial vs Operational

Hedging. Journal of Financial Economics 125(2), 217–244.

Hoberg, G., G. Phillips, and N. Prabhala (2014). Product market threats, payouts, and ﬁnancial

ﬂexibility. Journal of Finance 69(1), 293–324.

Huang, A., A. Zang, and R. Zheng (2014, November). Evidence on the information content of text

in analyst reports. The Accounting Review 89(6), 2151–2180.

Hwang, B.-H. and H. H. Kim (2017). It pays to write well. Journal of Financial Economics,

Forthcoming.

Jegadeesh, N. and D. Wu (2017). Deciphering Fedspeak: The information content of FOMC meet-

ings. Working Paper SSRN 2939937.

Kamenica, E. and M. Gentzkow (2011). Bayesian Persuasion. American Economic Review 101,

2590–2615.

Khanna, N. and S. Tice (2000). Strategic Responses of Incumbents to New Entry: The Effect of

Ownership Structure, Capital Structure, and Focus. The Review of Financial Studies 13, 749–779.

La Porta, R., F. L. de Silanes, and A. Shleifer (2006). What works in securities laws? Journal of

Finance 61(1), 1–32.

Li, D. (2011). Financial constraints, r&d investment, and stock returns. The Review of Financial

Studies 24(9), 2974–3007.

Li, F. (2008). Annual report readability, current earnings, and earnings persistence. Journal of

Accounting and Economics 45(2-3), 221–247.

Livnat, J. and R. R. Mendenhall (2006). Comparing the Post-Earnings Announcement Drift for Sur-

prises Calculated from Analyst and Time Series Forecasts. Journal of Accounting Research 44(1),

177–205.

31 Loughran, T. and B. McDonald (2011). When is a liability not a liability? Textual analysis, dictio-

naries, and 10-Ks. Journal of Finance 66(1), 35–65.

Loughran, T. and B. McDonald (2013, August). IPO ﬁrst-day returns, offer price revisions, volatil-

ity, and form S-1 language. Journal of Financial Economics 109(2), 307–326.

Loughran, T. and B. McDonald (2014). Measuring readability in ﬁnancial disclosures. The Journal

of Finance 69(4), 1643–1671.

Muslu, V., S. Radhakrishnan, K. Subramanyam, and D. Lim (2015). Forward-looking md&a dis-

closures and the information environment. Management Science 61(5), 931–948.

Phillips, G. M. (1995). Increased Debt and Industry Product Markets: An Empirical Analysis.

Journal of Financial Economics 37, 189–238.

Popadak, J. A. (2013). A corporate culture channel: How increased shareholder governance reduces

ﬁrm value. Working Paper SSRN 2345384.

Ross, S. A. (1979). Disclosure Regulation in Financial Markets: Implications of Modern Finance

Theory and Signaling Theory. Issues in Financial Regulation. McGraw-Hill New York.

Segal, G., I. Shaliastovich, and A. Yaron (2015). Good and bad uncertainty: Macroeconomic and

ﬁnancial market implications. Journal of Financial Economics 117(2), 369–397.

Verrecchia, R. E. (1983). Discretionary disclosure. Journal of Accounting and Economics 5, 179–

194.

Whited, T. M. and G. Wu (2006). Financial Constraints Risk. Review of Financial Studies 19(2),

531–559.

32 Figure 1: Linguistic Imprecision: Uncertainty and Modality

Note: This figure presents plots of notable textual measures from the Loughran and McDonald (2011) master dictionary in comparison to the propensity of firms to use imprecise language in disclosure. Each panel presents side-by-side box plots of the distribution of linguistic imprecision using imprecision keywords by above and below the median of each textual measure from Loughran and McDonald (2011). The difference in means in each panel is statistically significant at the 1% level. To help describe the content of the imprecision measure, the textual measures we consider are (a) uncertainty, (b) weak modal words, and (c) strong modal words.

(a) Uncertainty Words (Loughran and McDonald 2011) (b) Weak Modal Words (Loughran and McDonald 2011)

33 Figure 2: Firm Characteristics versus Propensity to Use Imprecise Language

Note: This figure presents plots of notable firm characteristics taken from Compustat as they individually relate to the propensity of firms to use imprecise language in disclosure. Each panel presents a 95% confidence interval for the firm characteristic for firms in the first, second, third, and fourth quartile of the distribution of linguistic imprecision, respectively. In this figure, the firm characteristics we consider are Firm Age in Years, Firm Size, Tobin’s Q, Sales Growth, R&D/Sales, and ROA.

(a) Firm Age (b) Firm Size

(e) R&D/Sales (f) ROA

34 Figure 3: Product Market Fluidity and Financial Constraints versus Propensity to Use Imprecise Language

Note: This figure presents plots of notable characteristics related to product market threats and financial constraints as they individual relate to the propensity of firms to use imprecise language in disclosure. Each panel presents a 95% confidence interval for the firm characteristic for firms in the first, second, third, and fourth quartile of the distribution of linguistic imprecision respectively. Product market threats and financial constraints that we consider are from Hoberg et al., 2014 and Hoberg and Maksimovic, 2015, respectively.

(a) Product Market Fluidity (Hoberg-Phillips-Prabhala)

(b) Financial Constraints (Hoberg-Maksimovic)

35 Figure 4: Management Incentives versus Propensity to Use Imprecise Language

Note: This figure presents a plot of management stock market incentives as it relates to the propensity of firms to use imprecise language in disclosure. The plotted ranges are 95% confidence intervals for stock ownership in the first, second, third, and fourth quartile of the linguistic imprecision distribution, respectively.

36 Figure 5: Linguistic Imprecision and Buy and Hold Abnormal Returns

Note: This figure presents plots of cumulative buy and hold abnormal returns (BHARs) over weekly windows after 10-K filing dates for high vs low imprecision firms. The cumulative BHARs are measured from the fourth days after the 10-K release date until the end of each week. High Imprecision and Low Imprecision are disclosures with above and below median linguistic imprecision, respectively.

37 Table 1: Frequently Used and Salient Words in Sentences with Weasel Tags

Note: Panel (a) of the table lists the top 30 most frequently mentioned unigrams, bigrams, and trigrams in the 433 sentences that have weasel tags ({{Weasel-inline|{{subst:DATE}}}}) from an Wikipedia dump completed on April 20, 2017. The Wikipedia dump comprises of 17,483,910 articles and is available at https://dumps.wikimedia.org/. To illustrate the inﬂuence of our saliency screen, Panel (b) of the table presents the top 10 unigrams and the bottom 10 unigrams sorted on the saliency measure of Goldsmith-Pinkham et al. (2016). Panel (c) lists the top 10 most frequently mentioned imprecision keywords, uncertain words, and weak and strong modal words. The uncertainty, weak modality, and strong modality word lists are from the Loughran and McDonald (2011) master dictionary.

(a) Top 10 Unigrams, Bigrams and Trigrams in Sentences with Weasel Tags Rank Unigrams Bigrams Trigrams 1 the of the one of the 2 and in the it has been 3 some it is considered by many 4 that to be is considered by 5 was has been of the most 6 many to the is one of 7 for for the it can be 8 with one of may have been 9 has and the according to some 10 have that the be one of

(b) Top 10 Unigrams and Bottom 10 Unigrams, Sorted on Saliency Rank Top 10 Unigrams Bottom 10 Unigrams 1 some the 2 many and 3 although for 4 considered was 5 may from 6 said their 7 have new 8 argued united 9 believed also 10 often ﬁrst

(c) Top 10 Imprecision, Uncertain, and Modal Words Rank Imprecision Words Uncertain Words Weak Modal Words Strong Modal Words 1 other hidden may will 2 may may could must 3 clear could possible best 4 could approximately might highest 5 would risk depend never 6 number of intangible uncertain lowest 7 can believe depending always 8 well assumptions depends clearly 9 however risks appears strongly 10 various believes appearing undisputed

38 Table 2: Summary and Sample Splits by High versus Low Imprecision

Note: This table presents averages of several notable variables in our analysis, separately by disclosures with below median linguistic imprecision (“Low Imprecision”) and disclosures with above median linguistic imprecision (“High Imprecision”). All firm characteristics are standardized and winsorized at the 1% level. The t-test column reports the two-sample t-test, using standard errors that are clustered by firm. * denotes significance at the 10% level, ** at the 5% level, and *** at the 1% level.

Low Imprecision High Imprecision t-stat Product Markets & Financial Constraints Product Market Fluidity 5.8696 7.5492 32.743*** Financial Constraints -0.0260 0.0031 20.098*** Equity Constraints -0.0323 0.0024 26.132*** Debt Constraints 0.0104 -0.0034 -17.53***

Notable Firm Characteristics ROA 0.1449 -0.2274 -25.483*** Tobin’s Q -0.1041 0.1619 19.152*** Log Market Value 0.0443 -0.2791 -20.587*** Log Age 0.3590 -0.0779 -32.141***

Other Textual Measures Sentiment (Pct Positive - Pct Negative) -0.7489 -0.9405 -29.958*** Pct Uncertain 0.8058 1.1599 80.593*** Pct Weak Modal 0.3410 0.5462 88.210*** Pct Strong Modal 0.2400 0.3185 44.713*** Pct Fog 29.8136 29.6796 -2.495**

39 Table 3: Relation of Linguistic Imprecision to Other Measures of Textual Information

Note: This table presents OLS regressions of imprecise language in the 10-Ks on other textual measures. Modal words, uncertainty words, positive words, constraining words, superfluous words, interesting words and litigious words are from the Loughran and McDonald (2011) master dictionary. The fog words are the complex words based on the fog measure initially proposed by Robert Gunning in 1952, and used extensively in quantifying the lack of plain English (see Hwang and Kim, 2017). Variable definitions are given in Appendix Table A.1. (Z) indicates that the variable has been standardized to have mean 0 and standard deviation 1 for ease of interpretation. Standard errors that are clustered by firm are reported in parentheses. * denotes significance at the 10% level, ** at the 5% level, and *** at the 1% level.

(1) (2) (3) (4) (5) Pct Uncertain (Z) 0.173*** 0.186*** 0.158*** 0.157*** 0.163*** (0.004) (0.004) (0.005) (0.005) (0.006) Pct Weak Modal (Z) 0.132*** 0.122*** 0.128*** 0.123*** 0.117*** (0.003) (0.003) (0.003) (0.003) (0.004) Pct Strong Modal (Z) -0.004** -0.004** -0.006*** -0.008*** -0.012*** (0.002) (0.002) (0.002) (0.002) (0.002) Sentiment (Z) -0.009*** 0.033*** 0.032*** 0.031*** 0.030*** (0.002) (0.002) (0.002) (0.002) (0.002) Log(Total Words in 10-K) 0.007*** -0.012*** -0.022*** -0.014*** -0.024*** (0.003) (0.003) (0.003) (0.003) (0.004) Additional Controls Pct Constraining, Pct Litigious no yes yes yes yes Pct Superﬂuous, Pct Interesting, Pct Fog no no yes yes yes Lagged Firm Characteristics no no no yes yes Observations 44,207 44,207 44,207 44,207 44,207 Adjusted R2 0.745 0.779 0.786 0.790 0.853 Fixed Effects SIC3, year SIC3, year SIC3, year SIC3, year ﬁrm, year

40 Table 4: Relation of Linguistic Imprecision to Firm Characteristics

Note: This table presents OLS regressions of imprecise language in disclosure in the 10-Ks on firm characteristics. Textual measures are from the Loughran and McDonald (2011) master dictionary. Firm characteristics are taken from Compustat and winsorized at the 1% level. Variable definitions are given in Appendix Table A.1. (Z) indicates that the variable has been standardized to have mean 0 and standard deviation 1 for ease of interpretation. Standard errors that are clustered by firm are reported in parentheses. * denotes significance at the 10% level, ** at the 5% level, and *** at the 1% level.

(1) (2) (3)

ROAt 1 (Z) -0.015*** -0.008*** -0.001 (0.003) (0.002) (0.003) Tobin’s Qt 1 (Z) 0.029*** 0.013*** 0.012*** (0.003) (0.003) (0.003) Firm Sizet 1 (Z) -0.019*** -0.004 -0.029*** (0.004) (0.004) (0.010) Firm Aget 1 (Z) -0.066*** -0.063*** (0.003) (0.007) Sales Growtht 1 (Z) 0.006*** 0.002 (0.002) (0.002) R&D/Salest 1 (Z) 0.015*** 0.001 (0.003) (0.004) CAPX/Salest 1 (Z) -0.000 -0.003 (0.003) (0.003) Leveraget 1 (Z) -0.024*** -0.011*** (0.003) (0.004) Log(Total Words in 10-K) -0.019*** -0.024*** -0.052*** (0.004) (0.004) (0.005) Observations 44,207 44,207 44,207 Adjusted R2 0.494 0.514 0.648 Fixed Effects SIC3, year SIC3, year ﬁrm, year

41 Table 5: Imprecise Language in Disclosure, Competitive Threats, and Financial Constraints

Note: This table presents OLS regressions of imprecise language in disclosure in the 10-Ks on measures of competitive threats, financial constraints and one year lagged firm characteristics. The product market fluidity measure is taken from Hoberg et al. (2014). Financial constraints is the text-based financial constraints measure developed by Hoberg and Maksimovic (2015). Specifications in columns 3 and 4 control for standard firm characteristics taken from Compustat lagged one year (firm size, firm age, ROA, Tobin’s Q, sales growth, R&D/Sales, CAPX/Sales, and leverage). Variable definitions are given in Appendix Table A.1. (Z) indicates that the variable has been standardized to have mean 0 and standard deviation 1 for ease of interpretation. Firm characteristics are all lagged one year and winsorized at the 99% level. Standard errors that are clustered by firm are reported in parentheses. * denotes significance at the 10% level, ** at the 5% level, and *** at the 1% level.

(a) Product Market Fluidity and Financial Constraints (1) (2) (3) (4)

Product Market Fluidityt 1 (Z) 0.088*** 0.030*** 0.074*** 0.028*** (0.003) (0.005) (0.003) (0.005) Financial Constraintst 1 (Z) 0.041*** 0.020*** 0.034*** 0.018*** (0.002) (0.003) (0.002) (0.003) Log(Total Words in 10-K) -0.045*** -0.053*** -0.033*** -0.052*** (0.004) (0.005) (0.004) (0.005) Observations 44,207 44,207 44,207 44,207 Adjusted R2 0.521 0.647 0.535 0.649 Lagged Firm Characteristics no no yes yes Fixed effects SIC3, year ﬁrm, year SIC3, year ﬁrm, year

(b) Equity and Debt Constraints Separately (1) (2) (3) (4)

Product Market Fluidityt 1 (Z) 0.085*** 0.030*** 0.074*** 0.028*** (0.003) (0.005) (0.003) (0.005) Equity Constraintst 1 (Z) 0.042*** 0.020*** 0.034*** 0.018*** (0.002) (0.003) (0.002) (0.003) Debt Constraintst 1 (Z) -0.006*** -0.001 -0.001 -0.000 (0.002) (0.002) (0.002) (0.002) Log(Total Words in 10-K) -0.044*** -0.053*** -0.034*** -0.053*** (0.004) (0.005) (0.004) (0.005) Observations 44,207 44,207 44,207 44,207 Adjusted R2 0.521 0.647 0.534 0.649 Lagged Firm Characteristics no no yes yes Fixed effects SIC3, year ﬁrm, year SIC3, year ﬁrm, year

42 Table 6: Imprecise Language in Disclosure and Equity Market Ownership

Note: This table presents OLS regressions imprecise language in disclosure in the 10-Ks on measures of product market threats, financial market constraints, and an indicator for whether the CEO is in the top quartile of fraction of ownership (“High Incentives”). Firm characteristics are taken from Compustat and winsorized at the 1% level. Variable definitions are given in Appendix Table A.1. (Z) indicates that the variable has been standardized to have mean 0 and standard deviation 1 for ease of interpretation. Standard errors that are clustered by firm are reported in parentheses. * denotes significance at the 10% level, ** at the 5% level, and *** at the 1% level.

(1) (2) (3) High Incentives 0.020** 0.021** 0.001 (0.008) (0.008) (0.008) High Incentives Financial Constraintst 1 (Z) 0.015** 0.014* ⇥ (0.008) (0.008) Product Market Fluidityt 1 (Z) 0.070*** 0.070*** 0.059*** (0.005) (0.005) (0.005) Financial Constraintst 1 (Z) 0.036*** 0.032*** 0.028*** (0.004) (0.005) (0.004) Log(Total Words in 10-K) -0.126*** -0.126*** -0.110*** (0.010) (0.010) (0.011) Lagged Firm Characteristics no no yes Observations 17,413 17,413 17,413 Adjusted R2 0.574 0.574 0.583 Fixed Effects SIC3, year SIC3, year SIC3, year

43 Table 7: Imprecise Language in Disclosure and Buy and Hold Abnormal Returns (BHARs)

Note: This table presents OLS regressions that regress buy and hold abnormal returns (BHARs) over various weekly windows after the 10-K filing date on the percentage of imprecision keywords in disclosure. The specifications control for the percentage difference between positive and negative words (sentiment), log of market capitalization one day prior to the 10-K filing date, log of book-to-market, log of share turnover, pre-filing-date Fama-French alpha, percentage of institutional holdings, and the most recent SUE. The definitions of these variables are given in Appendix Table A.1. Stan- dard errors that are clustered by month to account for cross-sectional correlation of BHARs are reported in parentheses. * denotes significance at the 10% level, ** at the 5% level, and *** at the 1% level.

(a) Weekly BHAR Windows Week1 Week2 Week3 Week4 Week5 Week6 Week7 Week8 Week9 Week10 (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Imprecision (Z) -0.000 0.001 0.001*** 0.002*** 0.001** 0.002** 0.002*** 0.001 0.000 -0.000 (0.001) (0.001) (0.000) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.000)

Sentiment (Z) -0.000 0.001* 0.001 0.001 -0.001 0.001 0.001* -0.000 -0.000 -0.001 (0.001) (0.000) (0.001) (0.001) (0.001) (0.001) (0.000) (0.000) (0.000) (0.000) Observations 41,414 41,382 41,346 41,282 41,235 41,207 41,151 41,084 41,053 41,015 Adjusted R2 0.001 0.001 0.003 0.001 0.001 0.001 0.001 0.001 0.002 0.002 Other Controls yes yes yes yes yes yes yes yes yes yes

(b) Multiple-week BHAR Windows Week1 Week1-2 Week1-3 Week1-4 Week1-5 Week1-6 Week1-7 Week1-8 Week1-9 Week1-10 (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Imprecision (Z) -0.000 0.001 0.002* 0.003*** 0.005*** 0.006*** 0.009*** 0.010*** 0.010*** 0.010*** (0.001) (0.001) (0.001) (0.001) (0.002) (0.002) (0.002) (0.003) (0.003) (0.003) Sentiment (Z) -0.000 0.000 0.001* 0.002* 0.001 0.002 0.003 0.002 0.002 0.001 (0.001) (0.001) (0.001) (0.001) (0.001) (0.002) (0.002) (0.002) (0.002) (0.002) Observations 41,414 41,451 41,465 41,466 41,468 41,469 41,469 41,470 41,470 41,470 Adjusted R2 0.001 0.001 0.002 0.003 0.003 0.004 0.005 0.005 0.006 0.005 Other Controls yes yes yes yes yes yes yes yes yes yes

44 Table 8: Imprecise Language in Disclosure and Subsequent Return Volatilities

Note: This table presents OLS regressions that regress net upside return volatilities using daily return data over one or two months after the 10-K filing date on the percentage of imprecision keywords in disclosure. The net upside return volatility is the difference between upside return volatility and downside volatility following Barndorff-Nielsen et al. (2010); Feunou et al. (2011); Segal et al. (2015). The specifications control for the percentage difference between positive and negative words (sentiment), pre-filing-date return volatility, pre-filing-date Amihud illiquidity, log of market capitalization one day prior to the 10-K filing date, and log of book-to-market. The definitions of these variables are given in Appendix Table A.1. Standard errors that are clustered by firm to account for serial correlation of return volatilities over time are reported in parentheses. * denotes significance at the 10% level, ** at the 5% level, and *** at the 1% level.

1 Month 2 Months (1) (2) Imprecision (Z) 0.021*** 0.021*** (0.008) (0.007) Sentiment (Z) 0.002 -0.002 (0.007) (0.006) Pre-ﬁle Return Volatility (Z) 0.112*** 0.111*** (0.011) (0.009) Pre-ﬁle Amihud Illiquidity (Z) -0.004 0.019* (0.015) (0.011) Observations 46,762 46,767 Adjusted R2 0.053 0.068 Other Controls yes yes Fixed effect Month Month

45 Table 9: Imprecise Language in Disclosure and Subsequent Valuation and Investment

Note: This table presents OLS regressions that relate subsequent firm valuations (Tobin’s Q) and corporate outcomes to current imprecise language in disclosure. Variable definitions are given in Appendix Table A.1. Standard errors that are clustered by firm are reported in parentheses. * denotes significance at the 10% level, ** at the 5% level, and *** at the 1% level. (a) Imprecision and Subsequent Outcomes

Tobin’s Qt+1 R&D/Salest+1 CAPX/Salest+1 (1) (2) (3) (4) (5) (6) Imprecision (Z) 0.036*** 0.027*** 0.022*** 0.010* 0.024*** 0.010 (0.006) (0.008) (0.006) (0.005) (0.006) (0.006) Observations 35,072 35,072 35,072 35,072 35,072 35,072 Adjusted R2 0.406 0.569 0.600 0.722 0.427 0.600 Date t Firm Characteristics yes yes yes yes yes yes Fixed Effects SIC3, year firm, year SIC3, year firm, year SIC3, year firm, year

(b) Imprecision, Forward-looking Context, and Subsequent Outcomes

Tobin’s Qt+1 R&D/Salest+1 CAPX/Salest+1 (1) (2) (3) (4) (5) (6) Imprecision (Z) 0.044*** 0.042*** 0.028*** 0.024*** 0.023*** 0.019** (0.009) (0.010) (0.008) (0.009) (0.009) (0.009) Forward-Looking (Z) 0.013* 0.015* 0.013** -0.008 0.035*** 0.014 (0.007) (0.009) (0.006) (0.006) (0.009) (0.009) Imprecision (Z) Forward-Looking (Z) 0.024*** 0.028*** 0.017*** 0.019*** 0.023*** 0.024*** ⇥ (0.006) (0.007) (0.006) (0.007) (0.007) (0.007) Observations 35,072 35,072 35,072 35,072 35,072 35,072 Adjusted R2 0.405 0.588 0.598 0.733 0.425 0.617 Date t Firm Characteristics yes yes yes yes yes yes Fixed Effects SIC3, year firm, year SIC3, year firm, year SIC3, year firm, year

46 Appendix to:

Some People Say Linguistic Imprecision in Corporate Disclosures Table A.1: Variable Deﬁnitions

Imprecision is the number of imprecision keywords scaled by the total word count in the filing in the percentage term. Pct Positive is the number of positive words from the master dictionary by Loughran and McDonald (2011) scaled by the total word count in the filings in the percentage term. Pct Negative is the number of negative words from the master dictionary by Loughran and McDonald (2011) scaled by the total word count in the filings in the percentage term. Pct Uncertain is the number of uncertain words from the master dictionary by Loughran and McDonald (2011) scaled by the total word count in the filings in the percentage term. Pct Modal is the number of modal words from the master dictionary by Loughran and McDonald (2011) scaled by the total word count in the filings in the percentage term. Pct Constraining is the number of constraining words from the master dictionary by Loughran and McDonald (2011) scaled by the total word count in the filings in the percentage term. Pct Litigious is the number of litigious words from the master dictionary by Loughran and McDonald (2011) scaled by the total word count in the filings in the percentage term. Pct Superfluous is the number of superfluous words from the master dictionary by Loughran and McDonald (2011) scaled by the total word count in the filings in the percentage term. Pct Interesting is the number of interesting words from the master dictionary by Loughran and McDonald (2011) scaled by the total word count in the filings in the percentage term. Pct Fog is the number of words of three or more syllables that are not hyphenated words or two- syllable verbs made into three with -es and -ed endings, scaled by the total word count in the filing in the percentage term. Sentiment is the percentage of positive words minus the percentage of negative words. Product Market Fluidity is a measure of the competitive threats faced by a firm in its product market that captures changes in rival firms’ products relative to the firm, from Hoberg, Phillips and Prabhala (2014). Financial Constraints is a financial constraints measure from Hoberg and Maksimovic (2015) with higher values indicating that firms are more at risk of delaying their investments due to issues with liquidity. Equity Constraints is a financial constraint measure from Hoberg and Maksimovic (2015) with higher values indicating that the firms are more at risk of delaying their investments due to issues with liquidity and have plans to issue equity in their financial statements. Debt Constraints is a financial constraint measure from Hoberg and Maksimovic (2015) with higher values indicating that the firms are more at risk of delaying their investments due to issues with liquidity and have plans to issue debt in their financial statements. High Incentives is an indicator for whether the CEO is in the top quartile of fraction of equity ownership. Equity ownership is the number of shares owned excluding options by the CEO (’shrown excl opts’ in the Execucomp data) divided by total number of shares outstanding. Firm Size is the log of market value of total assets (market value of common equity plus book value of preferred stock, long-term and short-term debt, and minority interest). Firm Age is the log of one plus firm age based on first appearance in Compustat. Tobin’s Q is market value of assets divided by book value of assets. ROA is net operating income divided by total assets in the prior year. Leverage is the ratio of total debt to the market value of assets. R&D/Sales is research and development expenditures divided by sales. CAPX/Sales is capital expenditures divided by sales. Sales Growth is the percentage growth in sales in a given year. Share Turnover is the volume of shares traded over one month before 10-K filing date divided by the number of shares outstanding on 10-K filing date.

ii Pre-file Fama-French Alpha is the intercept estimated by regressing daily excess returns on Fama-French’s three factors over one year before 10-K filing date. For each stock, at least 60 observations of daily returns are required to be included in the sample. Institutional Ownership is the percentage of institutional ownership available from the CDA/Spectrum database for the most recent quarter before 10-K filing date. The variable is treated as missing for negative values and winsorized to 100% on the positive side. Book-to-market is the book-to-market ratio using the book value from firm’s annual report known as of the end of the previous fiscal year and the market value known as of December of the year before the year of analysis. Following Fama and French (1992), stocks with negative book values are excluded. Pre-file Return Volatility is the standard deviation of daily returns over one month before 10-K filing date. Pre-file Illiquidity is the average of daily Amihud illiquidity measure (the ratio of absolute value of daily return to dollar trading volume) over one month before 10-K filing date. Forward-looking is the number of forward-looking disclosure keywords scaled by the total word count in the filing in the percentage term. The forward-looking disclosure keywords are from Muslu et al. (2015). SUE1 is standardized (by price) earnings surprise based on a rolling seasonal random walk model in Livnat and Mendenhall (2006). SUE2 is standardized (by price) earnings surprise based on a rolling seasonal random walk model in Livnat and Mendenhall (2006) accounting for exclusion of special items.

iii Table A.2: Relation of Imprecise Language in Disclosure to Firm Characteristics – MD&A Section Only

Note: This table presents OLS regressions of imprecise language in disclosure in the MD&A section of 10-K filings on firm characteristics. Textual measures are from the Loughran and McDonald (2011) master dictionary. Firm characteristics are taken from Compustat and winsorized at the 1% level. Variable definitions are given in Appendix Table A.1. (Z) indicates that the variable has been standardized to have mean 0 and standard deviation 1 for ease of interpretation. Standard errors that are clustered by firm are reported in parentheses. * denotes significance at the 10% level, ** at the 5% level, and *** at the 1% level.

(1) (2) (3) (4)

ROAt 1 (Z) -0.041*** -0.033*** -0.013* (0.006) (0.007) (0.008) Firm Sizet 1 (Z) -0.040*** -0.026*** -0.015 (0.010) (0.010) (0.023) Tobin’s Qt 1 (Z) 0.038*** 0.021*** 0.016** (0.006) (0.007) (0.008) Firm Aget 1 (Z) -0.060*** -0.096*** (0.008) (0.024) Sales Growtht 1 (Z) 0.005 0.003 (0.004) (0.004) R&D/Salest 1 (Z) 0.023*** 0.014 (0.008) (0.009) CAPX/Salest 1 (Z) 0.003 -0.003 (0.006) (0.007) Leveraget 1 (Z) -0.030*** -0.020** (0.007) (0.010) Log(Total Words in 10-K) 0.033*** 0.029*** -0.003 (0.009) (0.009) (0.008) Observations 40,259 40,259 40,259 Adjusted R2 0.069 0.077 0.510 Fixed Effects SIC3, year SIC3, year ﬁrm, year

iv Table A.3: Determinants of Imprecise Language in Disclosure – Accounting for Other Textual Measures

Note: This table presents OLS regressions of imprecise language in disclosure on product market fluidity and financial constraints. The product market fluidity measure is taken from Hoberg et al. (2014). Financial constraints is the text-based financial constraints measure developed by Hoberg and Maksimovic (2015). In contrast to the main specifications in Table 5, these specifications control for other textual measures. Modal words, uncertainty words, positive words, constraining words, superfluous words, interesting words and litigious words are from the Loughran and McDonald (2011) master dictionary. The fog words are the complex words based the fog measure initially proposed by Robert Gunning in 1952, and used extensively in quantifying the lack of plain English (see Hwang and Kim, 2017). Variable definitions are given in Appendix Table A.1. (Z) indicates that the variable has been standardized to have mean 0 and standard deviation 1 for ease of interpretation. Standard errors that are clustered by firm are reported in parentheses. * denotes significance at the 10% level, ** at the 5% level, and *** at the 1% level.

(1) (2) (3) (4)

Product Market Fluidityt 1 (Z) 0.022*** 0.006** 0.025*** 0.009*** (0.002) (0.003) (0.002) (0.003) Financial Constraintst 1 (Z) 0.011*** 0.009*** 0.011*** 0.009*** (0.002) (0.002) (0.002) (0.002)

Notable Textual Controls Pct Uncertainty (Z) 0.164*** 0.168*** 0.160*** 0.168*** (0.005) (0.006) (0.005) (0.005) Pct Weak Modal (Z) 0.120*** 0.116*** 0.117*** 0.114*** (0.003) (0.004) (0.003) (0.004) Pct Strong Modal (Z) -0.002 -0.006*** -0.006*** -0.007*** (0.002) (0.002) (0.002) (0.002) Sentiment (Z) 0.034*** 0.034*** 0.031*** 0.032*** (0.002) (0.002) (0.002) (0.002) Controls for Other Textual Measures yes yes yes yes Lagged firm characteristics no no yes yes Observations 44,207 44,207 44,207 44,207 Adjusted R2 0.787 0.852 0.7923 0.853 Fixed Effects SIC3, year firm, year SIC3, year firm, year

v Table A.4: Imprecise Language in Disclosure, Competitive Threats, and Financial Constraints – MD&A Section Only

Note: This table presents OLS regressions of imprecise language in disclosure in the MD&A section of the firm’s 10-K filings on measures of competitive threats, financial constraints and one year lagged firm characteristics. Aside from the change in the dependent variable, the specifications are the same as in Table 5. The product market fluidity measure is taken from Hoberg et al. (2014). Financial constraints is the text-based financial constraints measure developed by Hoberg and Maksimovic (2015). Specifications in columns 3 and 4 control for standard firm characteristics taken from Compustat lagged one year (firm size, firm age, ROA, Tobin’s Q, sales growth, R&D/Sales, CAPX/Sales, and leverage). Variable definitions are given in Appendix Table A.1. (Z) indicates that the variable has been standardized to have mean 0 and standard deviation 1 for ease of interpretation. Firm characteristics are all lagged one year and winsorized at the 99% level. Standard errors that are clustered by firm are reported in parentheses. * denotes significance at the 10% level, ** at the 5% level, and *** at the 1% level.

(1) (2) (3) (4)

Product Market Fluidityt 1 (Z) 0.043*** 0.022*** 0.036*** 0.020** (0.009) (0.009) (0.009) (0.009) Financial Constraintst 1 (Z) 0.078*** 0.065*** (0.006) (0.006) Equity Constraintst 1 (Z) 0.090*** 0.073*** (0.007) (0.007) Debt Constraintst 1 (Z) 0.004 0.013** (0.006) (0.006) Log(Total Words in 10-K) -0.004 0.022*** -0.004 0.021** (0.009) (0.008) (0.009) (0.008) Observations 40,259 40,259 40,259 40,259 Adjusted R2 0.074 0.085 0.077 0.086 Lagged Firm Characteristics no yes no yes Fixed effects SIC3, year SIC3, year SIC3, year SIC3, year

vi Table A.5: Imprecise Language in Disclosure, Competitive Threats, and Financial Constraints – Purged Measure

Note: This table presents OLS regressions of imprecise language in disclosure on measures of competitive threats, financial constraints and one year lagged firm characteristics. The measure of imprecision in this table only contains imprecision keywords that are distinctive from uncertain and weak modal words from the Loughran and McDonald (2011) master dictionary. Aside from the change in the dependent variable, the specifications are the same as in Table 5. The product market fluidity measure is taken from Hoberg et al. (2014). Financial constraints is the text-based financial constraints measure developed by Hoberg and Maksimovic (2015). Specifications in columns 3 and 4 control for standard firm characteristics taken from Compustat lagged one year (firm size, firm age, ROA, Tobin’s Q, sales growth, R&D/Sales, CAPX/Sales, and leverage). Variable definitions are given in Appendix Table A.1. (Z) indicates that the variable has been standardized to have mean 0 and standard deviation 1 for ease of interpretation. Firm characteristics are all lagged one year and winsorized at the 99% level. Standard errors that are clustered by firm are reported in parentheses. * denotes significance at the 10% level, ** at the 5% level, and *** at the 1% level.

(1) (2) (3) (4)

Product Market Fluidityt 1 (Z) 0.017*** 0.009*** 0.013*** 0.007*** (0.002) (0.002) (0.002) (0.002) Financial Constraintst 1 (Z) 0.014*** 0.007*** (0.002) (0.002) Equity Constraintst 1 (Z) 0.024*** 0.014*** (0.002) (0.002) Debt Constraintst 1 (Z) -0.011*** -0.006*** (0.001) (0.001) Log(Total Words in 10-K) -0.026*** -0.012*** -0.026*** -0.012*** (0.002) (0.003) (0.002) (0.003) Observations 44,207 44,207 44,207 44,207 Adjusted R2 0.087 0.003 0.098 0.130 Lagged Firm Characteristics no yes no yes Fixed effects SIC3, year SIC3, year SIC3, year SIC3, year

vii Table A.6: Imprecise Language in Disclosure, Competitive Threats, and Financial Constraints – Size-Age Index and Whited Wu Index

Note: This table presents OLS regressions of imprecise language in disclosure on measures of competitive threats, financial constraints and one year lagged firm characteristics. In contrast to the main specifications, this table presents results using two financial constraints indexes that are not textual, found in the literature: the Whited and Wu (2006) financial constraints index (WW index) and the size-age index (SA Index) of Hadlock and Pierce (2010). Specifications in columns 3 and 4 control for standard firm characteristics taken from Compustat lagged one year (firm size, firm age, ROA, Tobin’s Q, sales growth, R&D/Sales, CAPX/Sales, and leverage). Variable definitions are given in Appendix Table A.1. (Z) indicates that the variable has been standardized to have mean 0 and standard deviation 1 for ease of interpretation. Firm characteristics are all lagged one year and winsorized at the 99% level. Standard errors that are clustered by firm are reported in parentheses. * denotes significance at the 10% level, ** at the 5% level, and *** at the 1% level.

(1) (2) (3) (4)

Product Market Fluidityt 1 (Z) 0.090*** 0.093*** 0.086*** 0.087*** (0.003) (0.003) (0.003) (0.003) WW Index (Z) 0.031*** 0.069*** (0.003) (0.009) SA Index (Z) 0.039*** 0.039*** (0.003) (0.005) Log(Total Words in 10-K) -0.026*** -0.025** -0.024** -0.025*** (0.004) (0.004) (0.004) (0.004) Observations 44,207 44,207 44,207 44,207 Adjusted R2 0.510 0.512 0.515 0.516 Lagged Firm Characteristics no no yes yes Fixed effects SIC3, year SIC3, year SIC3, year SIC3, year

viii Table A.7: Imprecise Language in Disclosure and Buy and Hold Abnormal Returns (BHARs) - R&D and Financial Constraints Controlled

Note: This table presents OLS regressions that regress buy and hold abnormal returns (BHARs) over various weekly windows after the 10-K filing date on the percentage of imprecision keywords in disclosure when controlling for the effects from R&D and financial constraints. R&D expenses divided by sales and the Whited and Wu (2006) financial constraints index (WW index) are used as controls for the effects from R&D and financial constraints, respectively. The specifications also control for the percentage difference between positive and negative words (sentiment), log of market capitalization one day prior to the 10-K filing date, log of book-to-market, log of share turnover, pre-filing-date Fama- French alpha, percentage of institutional holdings, and the most recent SUE. The definitions of these variables are given in Appendix Table A.1. Standard errors that are clustered by month to account for cross-sectional correlation of BHARs are reported in parentheses. * denotes significance at the 10% level, ** at the 5% level, and *** at the 1% level.

(a) Weekly BHAR Windows Week1 Week2 Week3 Week4 Week5 Week6 Week7 Week8 Week9 Week10 (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Imprecision (Z) -0.000 0.001 0.002*** 0.002*** 0.001** 0.002*** 0.003*** 0.001 0.000 -0.001 (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.000) Sentiment (Z) -0.000 0.001** 0.001* 0.001* -0.000 0.001 0.001** -0.000 -0.000 -0.001 (0.001) (0.000) (0.001) (0.001) (0.001) (0.001) (0.000) (0.000) (0.000) (0.000) Observations 40,984 40,951 40,916 40,854 40,808 40,783 40,726 40,659 40,630 40,592 Adjusted R2 0.002 0.003 0.003 0.002 0.001 0.001 0.002 0.002 0.003 0.002 Other Controls yes yes yes yes yes yes yes yes yes yes

(b) Multiple-week BHAR Windows Week1 Week1-2 Week1-3 Week1-4 Week1-5 Week1-6 Week1-7 Week1-8 Week1-9 Week1-10 (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Imprecision (Z) -0.000 0.001 0.003** 0.004*** 0.005*** 0.007*** 0.010*** 0.011*** 0.012*** 0.011*** (0.001) (0.001) (0.001) (0.002) (0.002) (0.002) (0.003) (0.003) (0.003) (0.003) Sentiment (Z) -0.000 0.001 0.002** 0.002** 0.002 0.002 0.003* 0.003* 0.003 0.002 (0.001) (0.001) (0.001) (0.001) (0.001) (0.002) (0.002) (0.002) (0.002) (0.002) Observations 40,984 41,020 41,034 41,035 41,037 41,038 41,038 41,039 41,039 41,039 Adjusted R2 0.002 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.009 0.007 Other Controls yes yes yes yes yes yes yes yes yes yes

ix Table A.8: Imprecise Language in Disclosure and Buy and Hold Abnormal Returns (BHARs) - Excluding Future Earnings Announcements

Note: This table presents OLS regressions that regress buy and hold abnormal returns (BHARs) over various weekly windows after the 10-K filing date on the percentage of imprecision keywords in disclosure with a refined sample that excludes all 10-K filings that have new earnings announcements during the next 3rd to 7th week period. The specifications control for the percentage difference between positive and negative words (sentiment), log of market capitalization one day prior to the 10-K filing date, log of book-to-market, log of share turnover, pre-filing-date Fama-French alpha, percentage of institutional holdings, and the most recent SUE. The definitions of these variables are given in Appendix Table A.1. Standard errors that are clustered by month to account for cross-sectional correlation of BHARs are reported in parentheses. * denotes significance at the 10% level, ** at the 5% level, and *** at the 1% level.

(a) Weekly BHAR Windows Week1 Week2 Week3 Week4 Week5 Week6 Week7 Week8 Week9 Week10 (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Imprecision (Z) -0.000 0.000 0.001 0.002*** 0.001* 0.002** 0.002*** 0.001 0.000 -0.000 (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) Sentiment (Z) -0.000 0.000 0.000 0.000 -0.001 0.001 0.001 -0.001 -0.001 -0.001 (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) Observations 24,868 24,841 24,820 24,756 24,713 24,695 24,650 24,591 24,559 24,526 Adjusted R2 0.001 0.001 0.001 0.003 0.001 0.001 0.001 0.002 0.002 0.001 Other Controls yes yes yes yes yes yes yes yes yes yes

(b) Multiple-week BHAR Windows Week1 Week1-2 Week1-3 Week1-4 Week1-5 Week1-6 Week1-7 Week1-8 Week1-9 Week1-10 (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Imprecision (Z) -0.000 0.000 0.001 0.003* 0.004** 0.005*** 0.008*** 0.009*** 0.009*** 0.009*** (0.001) (0.001) (0.001) (0.001) (0.002) (0.002) (0.002) (0.002) (0.003) (0.003) Sentiment (Z) -0.000 0.000 0.000 0.001 -0.000 0.000 0.001 0.000 -0.000 -0.001 (0.001) (0.001) (0.001) (0.001) (0.001) (0.002) (0.002) (0.002) (0.002) (0.003) Observations 24,868 24,897 24,906 24,906 24,908 24,908 24,908 24,908 24,908 24,908 Adjusted R2 0.001 0.001 0.001 0.003 0.003 0.004 0.005 0.005 0.006 0.005 Other Controls yes yes yes yes yes yes yes yes yes yes

x Table A.9: Imprecise Language in Disclosure Mitigates Negative Market Information: Buy and Hold Abnormal Returns

Note: This table presents Fama-Macbeth regressions (using 74 quarters of data from 1997Q4 to 2016Q1) that regress buy and hold abnormal returns from 0 to 180 calendar days after the 10-K filing date on our linguistic imprecision measure, the most recent SUE, and their interaction term. The specifications in odd columns construct the SUE as in Livnat and Mendenhall (2006) using a rolling seasonal random walk model (SUE1). The specifications in even columns also account for exclusion of special items (SUE2). The specifications also control for log of market capitalization one day prior to the 10-K filing date, log of book-to-market, log of share turnover, pre-filing date Fama-French alpha, and percentage of institutional holdings. Variable definitions are given in Appendix Table A.1. Newey-West corrected standard errors that account for serial correlations in estimated FM slope coefficients are reported in parentheses. * denotes significance at the 10% level, ** at the 5% level, and *** at the 1% level.

All 10-Ks Negative SUE Positive SUE (1) (2) (3) (4) (5) (6) Imprecision SUE -0.533** -0.893** -0.700** -1.014** -0.432 -0.8288 ⇥ (0.219) (0.364) (0.297) (0.502) (0.357) (0.5324) SUE 0.800** 1.455** 0.799* 1.256 0.779 1.689* (0.368) (0.641) (0.471) (0.832) (0.578) (0.987) Imprecision 0.026 0.026 0.029 0.024 0.024 0.031 (0.022) (0.022) (0.030) (0.032) (0.020) (0.022) Observations 39,739 39,739 15,556 15,334 24,233 24,455 # of Cross-sections (Quarters) 74 74 74 74 74 74 Event-level Controls yes yes yes yes yes yes