Product Dynamics, Characteristics and Dividends

Transcription

1 Product Dynamics, Characteristics and Dividends Gerard Hoberg*, Gordon Phillips** and Nagpurnanand Prabhala April 3, 2011 University of Maryland, **University of Maryland and NBER. Hoberg can be reached at Phillips can be reached at and Prabhala can be reached at We thank Harry DeAngelo, Gustavo Grullon and participants at the UBC Winter Finance Conference and seminar audiences at HKUST, National University of Singapore, Michigan State University, Northwestern University, Notre Dame University, Temple University, and University of Southern California for helpful comments. Any remaining errors are ours alone. Copyright c 2010 by Gerard Hoberg, Gordon Phillips and Nagpurnanand Prabhala. All rights reserved.

2 Product Dynamics, Characteristics and Dividends ABSTRACT We show that a firm s product characteristics are significantly related to its dividend and share repurchase policies. Using text-based analysis of product descriptions in firm 10-Ks, we examine how the propensity to pay dividends and repurchase shares are affected by a firm s product fluidity, the local competition it faces, and the nature of the clientèle for the firm s products. Dividend payers are more likely to be stable firms with less product fluidity. Firms also are more likely to use dividends but less likely to repurchase when their products are more likely to be sold to a business clientèle, consistent with firms using dividends to indicate stability to customers who value stability. The results are consistent with new hypotheses that firms payout policies are significantly shaped by product stability and the dynamics of their product market space.

3 1 Introduction Our paper examines how a firm s product markets and its product characteristics shape its payout policy. We use computational linguistics to analyze over 42,000 individual firm business descriptions from firm 10-Ks to construct new measures of the structure and evolution of the product space occupied by firms. These measures include product fluidity or stability, which reflect changes in a firm s products relative to those of its rivals. We use the text-based measures to test hypotheses about how product fluidity, product-market competition, and product customer type affect payout policy. Our focus on product fluidity and competition can be motivated by the evidence in Brav, Graham, Harvey, and Michaely (2005). Figure 1 in their survey indicates that 70% of managers perceive the... stability and sustainability of future earnings as being central to their choice of payout policy. This stable environment is less likely when product markets are fluid, particularly when the fluidity is due to competitive moves by rival firms. We hypothesize that managers view product market fluidity and measured cash flow risk differently, and that both can explain payout policy. For instance, firm s managers facing competitive threats may change their products frequently and maintain ex post stable profits, and hence cash flow risk may not reflect ex-ante competitive threats. Consistent with this view, product market fluidity and cash flow risk are only moderately correlated, and both are negatively associated with the likelihood of paying dividends. Apple Computer, for example, has changed its product mix frequently over the last decade. It has high profits and low measured cash flow risk - yet pays no dividends. Our central finding is that product market fluidity has an economically significant impact on dividend policy, both in the cross section and in explaining firm decisions to change their payout policy (initiations, omissions, increases and decreases). Firms with fluid products whose description is changing over time relative to rival firms 1

4 are less likely to pay dividends or repurchase shares. These firms also pay lower dividends, as measured by the dividend yield or dividends scaled by assets. These results are most pronounced for dividends, which entail a long-term commitment to pay out cash flows that is not easily reversed. This role for product fluidity is consistent with the industrial organization literature on competitive threats relating to product life cycle, as in Abernathy and Utterback (1978) or Klepper (1996). Firms should be more willing to make a dividend commitment when their product has reached a stable point and when they foresee fewer competitive threats. In contrast, retaining liquid assets can provide flexibility to firms in less stable markets, allowing them to react more aggressively to competitive threats when they do materialize. Alternatively, DeAngelo, DeAngelo, and Skinner (2009) discuss a life cycle theory with managerial agency, in which cash payouts prevent a wasteful managerial mindset arising from resource abundance. Payouts may be less necessary when firms are disciplined by ongoing competitive threats from the product market. We find that product market fluidity, which is forward looking and relates to future competitive threats, is distinct from measures of current product market competition and differentiation introduced by Hoberg and Phillips (2010a). We also show that the product fluidity result is especially strong in concentrated industries. Firms in these industries are especially sensitive to product market fluidity, consistent with their having more to lose should competitive threats materialize. The results are also consistent with the results of Gaspar and Massa (2006), Hou and Robinson (2006), and or Peress (2010) that firms in concentrated industries are less risky and thus more likely to make payouts. In secondary tests, we examine the role of product customer type in explaining dividend and repurchase policy. We hypothesize that firms with business clientèle are more likely to pay dividends. The relation arises because only firms who are stable in the long-term can afford to make dividend commitments, and this stability 2

5 is valued by business partners. Interestingly, the dividend literature suggests that this relation should not extend to repurchases. Lintner (1956) and Brav, Graham, Harvey, and Michaely (2005) point out that while managers are very conservative in setting dividends, 1 repurchase policies are not as conservative (see, e.g., their Section 4.1). Thus, while both dividends and repurchases pay out cash from firms, there is an additional long-term commitment implicit to paying dividends. The stability indicated by this commitment could be valued by business clientèle. The testable proposition is that firms with business clientèle are more likely to pay dividends but not repurchase shares. We test this hypothesis using new proxies for business customers based on product text merged with the input output vertical relatedness tables. We find evidence of asymmetry in which business clientèle is positively related to dividends but negatively related to repurchases. This hypothesis and evidence are unique as existing clientèle theories assume that the intended targets are investors. We show that the targets may be business partners. 2 Consistent with Brav, Graham, Harvey, and Michaely (2005), the two forms of payout are not necessarily a... fluid, one-for-one substitution. We conduct extensive robustness tests. We explicitly consider whether our results can be explained by growth options. We include multiple measures of potential growth options including firm market-to-book, asset growth, firm age, R&D expenditure and patenting intensity. We also consider whether our product fluidity measures are distinct from cash flow risk and thus include measures of historical risk (Hoberg and Prabhala (2009)), firm maturity (DeAngelo, DeAngelo, and Stulz (2006)), and earnings losses (DeAngelo, DeAngelo, and Skinner (1992)). 1 See also Section 5.2 in DeAngelo, DeAngelo, and Skinner (2009) or the fifth of the six major facts about dividend policy outlined in Allen and Michaely (2003). As DeAngelo et al. emphasize in their Section 1.1, the roots of dividend conservatism go further back to Alfred Sloan, the well-known head of General Motors, who discusses dividend policy in a 1935 article in the New York Times. Brav et al. write of executives tell stories of selling assets, layoffs, and so on before... slaying the sacred cow by cutting dividends. 2 As DeAngelo, DeAngelo, and Skinner (2009) conclude, investor clientèle driven demand are unlikely to explain dividends. Related points are made by Grinstein and Michaely (2005), Graham and Kumar (2006) or Jain (2007). 3

6 We also find that our results are also robust in a subsample that excludes the largest dividend payers, which have been shown to pay a large fraction of dividends by DeAngelo, DeAngelo, and Skinner (2004). They also remain significant in explaining changes in dividend policy including initiations, omissions, and dividend changes by current payers. Our results suggest that product descriptions are likely to capture changing opportunities in the product market before their effects filter through to firm financials. Our study is the first to illustrate the extent to which some product markets are highly fluid over time while others are not. Our new text-based measures build upon Hoberg and Phillips (2010a,b), who introduce text based industry classifications, and illustrate how they can be used to test important theories in corporate finance and industrial organization. The current paper develops this research further, and constructs novel new measures of the dynamics of the product market space over time, and shows how these dynamics relate to payout policy. 3 The advantages of using product text descriptions over current methods of classifying firms include the following. First, product descriptions must meet regulatory standards. Item 101 in Regulation S-K specifies that the descriptions must be representative and significant. Thus, product descriptions contain timely and forward looking information about the challenges that firms face in its product markets, as seen through the eyes of senior management. Second, they improve on SIC codes in that they offer extensive detail regarding the full distribution of rivals surrounding each firm, including continuous measures of pairwise similarity. Third, changes in product descriptions empower the measurement of how products change over time. SIC codes are static regarding product location, and firm assignments change only rarely, as shown by Hoberg and Phillips (2010a). 3 Other studies using text to analyze finance theories include Antweiler and Frank (2004), Boukus and Rosenberg (2006), Li (2006), Tetlock (2007), Tetlock, Saar-Tsechanksy, and Macskassy (2008), Hanley and Hoberg (2010), Loughran and McDonald (2010), Hoberg and Phillips (2010b), and Rauh and Sufi (2010). 4

7 Our results add to the existing dividend and repurchase literature on several dimensions. Our work is related to Grullon, Michaely, and Swaminathan (2002) and DeAngelo, DeAngelo, and Stulz (2006), who argue that firm maturity is a key determinant of dividends. 4 Our results suggest that maturity has two components. One is a firm level component that reflects the gradual maturation of a firm over its life cycle. Maturity also has a distinct product market component, which reflects new threats and opportunities faced by a firm at its boundaries. Our evidence suggests that both aspects of maturity matter in setting payout policy. We also contribute to the literature on dividend clientèle, 5 by stressing the role of business rather than investor clientèle. The business clientèle result offers fresh perspective on the dividend-repurchase substitution issue (see Jagannathan, Stephens, and Weisbach (2000), Fama and French (2001), and Grullon and Michaely (2002)). The remainder of the paper is organized as follows. In Section II, we discuss the relation of our work to the existing literature on dividends. We develop our three central hypotheses in this section. The data and text-based measure of fluidity and similarity are in section III. Section IV presents the results of our analysis and Section V concludes. 2 Product Characteristics and Dividends In this section we describe and develop new hypotheses on how dividend and share repurchase policy may be affected by product market characteristics and how these change over time. We also discuss how these hypotheses relate to the existing literature. We have two hypotheses relating payout policy to product fluidity and product customer type. 4 Related work includes Venkatesh (1989), Fink, Fink, Grullon, and Weston (forthcoming), and Chay and Suh (forthcoming). 5 For an extensive survey of the literature, see chapters 8 and 10 of DeAngelo, DeAngelo, and Skinner (2009) or section 6 in Allen and Michaely (2003). 5

8 2.1 Product Fluidity We postulate that product fluidity or instability are important given that firms pay dividends based on their expectations of the future market for their products. When product markets are in flux, the future is less certain. It is unlikely that such firms would have expectations of sustainable, stable earnings that 70% of the executives surveyed in Brav, Graham, Harvey, and Michaely (2005) cite as the most critical determinant in making payouts. While stability and risk matter for both repurchases and dividends, they are perhaps more critical for setting dividends whose relative rigidity and inflexibility make managers especially conservative with regard to dividend policy. The importance of product fluidity is also shown in the literature on product life cycles and the competitive threats that firms face. Prominent work in this area includes Abernathy and Utterback (1978) and Klepper (1996), who suggest that products and their associated industries develop over time. These product changes give rise to a natural life cycle. Relevant to forming our hypotheses, firms selling newer products or products that have more technological risk face more fluid product markets and thus more future competitive threats. In turn, such firms might refrain from making payouts until their product markets mature, as payout conservatism can strengthen the firm s competitive positioning should a threat materialize. As firms products begin to stabilize, they may build some excess cash and repurchase shares. Because some product instability is still present, though, they might only be willing to repurchase shares and not pay dividends (as repurchases entail less of a longer-term commitment). If and when a mature product dominant design emerges and competition moderates, then firms are in a position to make a dividend commitment. Our main measure of fluidity measures the extent to which rival firms change their products relative to a given firm s existing products. This approach incorporates competitive pressure potentially related to technological change in related markets. 6

9 While we develop and describe the fluidity measure in greater detail in Section 3, we emphasize that it is a forward looking variable that is distinct from measures of growth options, other accounting numbers disclosed by firms - including firm R&D, historical cash flow risk and, or the negative earnings indicator suggested by DeAngelo, DeAngelo, and Skinner (1992). Because the product descriptions in a firm s 10-Ks are legally required to be accurate and current, they are likely to reflect updated assessments of top management about the opportunities and threats that a firm faces in its product markets. Thus, product descriptions are likely to contain information beyond that in accounting numbers that tend to reflect a firm s history. Product fluidity may, but need not be, related to firm maturity. As Grullon, Michaely, and Swaminathan (2002) point out, maturation is the cumulative product of a firm s history played out over a long period of time. DeAngelo, DeAngelo, and Stulz (2006) make a similar point, arguing that maturity is reflected in a build up of capital over several years. Thus, firm maturity is a largely stable cross-sectional characteristic based on a firm s past history. Product fluidity provides a complementary perspective of maturity based on more specific mechanisms. For instance, older firms and firms that have retained a large portion of their past earnings might be viewed as mature firms. However, some older firms may face renewed technological change or product innovation. Our fluidity measure is well suited to identify such changes. Our measure is also distinct from firm maturity because we consider effects from rivals in related product markets. Whether our fluidity variable adds to conventional measures of firm maturity or not is ultimately an empirical question. We include other controls including firm age and stock-price risk, in addition to R&D expenditures that may be high when firms spend more money on changing products. We thus consider the following hypothesis: H1: Firms whose products change more over time - both relative to product changes by rival firms or relative to their own past products - thus having more competitive threats, will have a lower propensity to pay dividends. 7

10 Our main measure of fluidity is local. It is based only on words that tend to persistently appear with a localized group of related words (ie, they appear in word cliques as identified by the local clustering coefficient). 6 Intuitively, product market words have this property: the words cola, beverage, and drink should appear together in a 10-K if they appear at all. Because words in the local dictionary are represented by words unique to local product markets, we use local product fluidity as our primary variable of interest. This measure identifies product instabilities and competitive pressure arising from peer firms, not instability in a firm s own product vocabulary. A high score means the firm s product words experience high turnover by other firms in related product spaces. Table 1 presents initial summary statistics as well as transition matrices examining how fluidity changes over time. We define fluidity formally and describe how it is computed in the next section. We present some preliminary summary statistics in this section to show that the effects we are describing are large, and that they fit in well with a product life cycle/competitive threat interpretation. [Insert Table 1 Here] Panel A presents summary payout statistics for the firms in each quintile of local product market fluidity. The results show that 48.6 percent of firms with the highest product stability pay dividends in contrast to only 9.1 percent of firms with the most fluid product descriptions. Dividend yields are also close to four times larger for firms with the highest product stability vs. firms with the most fluid product descriptions. Repurchases also have a strong negative relationship with fluidity, but results for dividend variables are stronger than those for repurchases. Panels B to D present transition matrices examining ex post local product market 6 To address potential endogeneity concerns, where rivals make changes in response to the firm s choices, we also examine a broad measure of product fluidity. In particular, using a broad measure of product fluidity assesses the change in a firm s product description with respect to the broad set of words that all firms use. This set of all firms is most akin to the overall market index, which is exogenous from a single firm s perspective, as it is unlikely that the market is responding to one particular firm s choices. 8

11 fluidity outcomes as a function of initial local product market fluidity. The rows indicate the initial fluidity quintile of the firm and the columns indicate which fluidity quintile the firm is in t years later. The panels are thus transition matrices. For example, in Panel B, row 1, column 1, 77.9 percent of firms that have the most stable product description (lowest quintile of fluidity) remain in this lowest quintile of product fluidity 1 year later. In the transition matrices that begin in Panel B, we first limit our sample to firms having a track record of at least 3 years as of a given year t (although this condition does not materially affect results). Panel B displays one year transition probabilities while Panels B and C examine three year and six year transition probabilities, respectively. For the three year test, we divide our twelve year sample into four non-overlapping three year intervals, and we examine how product market fluidity changes for firms from one interval to the next. For the six year test, we divide our twelve year sample into two six year non-overlapping intervals and repeat the same test. In each case, we assign firms to quintiles in each ex-ante interval based on the level of ex-ante product market fluidity. Holding breakpoints fixed for each interval, we group observations into quintiles based on their ex-post product market fluidity. The result is a 5x5 grid containing the empirical distribution of transitions. The results in Panels B to D show that there is strong persistence in local product market fluidity. Firms with stable products ex ante are very likely to have stable products ex-post. This result holds for one, three and six year horizons. Also, firms with less stable products ex-ante generally remain less stable in the subsequent period - but persistence is weaker relative to firms having stable products. For example a firm in the most fluid quintile remains in this category 71.1% of the time, while firms in the most stable quintile retain this designation 77.9% of the time. The table also shows that firms are more likely to graduate to a category with more stability than they are to move to a category with less stability over time. For example, in panel B, firms with the most stable products move into category 2 with only a 16.8% 9

12 likelihood. In contrast, firms move from the highest fluidity category to the more stable 4th category with a 23.6% likelihood. A similar and stronger pattern emerges over the three and six year windows. This evidence is consistent with a product life cycle interpretation in which product fluidity decreases over time. 2.2 Product Customer Type We also conduct an analysis on whether the type of customers who buy the firm s products matters to firm dividend policy. We examine whether a firm s type of customer affects a firm s choice to pay dividends or repurchase shares. We hypothesize that firms with business customers (not retail customers) should favor dividends over repurchases despite potential tax disadvantages for dividends. This hypothesis rests on the premise that managers are extremely conservative towards dividends (but not repurchases). This view dates back to at least Lintner (1956) and is affirmed recently in the Brav, Graham, Harvey, and Michaely (2005) survey of 384 financial executives working at 256 public firms. A firm that can pay dividends will thus be seen as more stable in the long-term, and may attract more business customers because they value the ability to minimize future supply chain interruptions. Repurchase decisions are not as conservative (Guay and Harford (2000), Jagannathan, Stephens, and Weisbach (2000)). As Brav et al. (p. 500) write, managers see little penalty in case their repurchase decisions are later reversed. The prediction is thus that firms with business customers (not retail customers) should favor dividends over repurchases. This theory differs from existing dividend clientèle theories, in which the targets are investors. The investor-driven view of payout policy is not favored by the executives surveyed in Brav, Graham, Harvey, and Michaely (2005), who do not indicate that they believe that institutions as a class prefer dividends over repurchases. Furthermore, Grinstein and Michaely (2005) show that while institutions may prefer dividend payers due to the 1978 ERISA (Brav and Heaton (1998)), institutional holding has little relation to dividend payouts. In 10

13 short, there is not a clear cut relation between institutional clientèle and dividend policy in prior work. Our work has a different focus. We start with the hypothesis that the benefits of dividends are product market (real-side) driven and intended for potential business partners, not investors. Operationally, we use the BEA input/output tables to get a list of product words that have over 90% of their output going to other businesses, and thus less than 10% to retail. We then compute each firm s textual similarity to this list of businessclientèle products. Firms scoring highly should be dividend payers. This gives rise to our second product-based hypothesis: H2: Firms who sell more products to business customers should pay higher dividends and should have a higher propensity to pay dividends than to repurchase shares. 3 Data, Methodology and Key Variables In this section we describe the data and the methodology we use to extract the 10- K text, and how we form the key text-based variables used in our analysis. These variables come purely from 10-K text extracted from business descriptions. The first step is to use web crawling and text parsing algorithms to construct a database of business descriptions from 10-K annual filings on the SEC Edgar website from 1997 to These descriptions are found in a separate section of each 10-K filed by each firm. Business descriptions are legally required to be accurate, as Item 101 of Regulation S-K legally requires that firms describe the significant products they offer to the market, and these descriptions must also be updated and representative of the current fiscal year of the 10-K. This recency requirement is important, as our goal is to accurately measure how product characteristics and industry structure change from year to year. 11

14 3.1 The CRSP and COMPUSTAT Sample We construct our COMPUSTAT-CRSP sample following Fama and French (2001) and Hoberg and Prabhala (2009). We start with 61,136 firm-years from 1997 to 2008 that have adequate COMPUSTAT and CRSP data. The years are chosen based on the availability of our text-based data. We then apply the same screens as Hoberg and Prabhala (2009). After discarding regulated utilities (SIC codes between 4900 and 4949) and financials (SIC codes between 6000 and 6999), we have 48,159 observations. We then screen out observations in which firms have book values of less than $250,000 or assets of less than $500,000. This leaves us with 45,631 observations. Following Fama and French (2001), we identify companies that are dividend payers if their dividends per share (Compustat annual data item 26) is greater than zero. We identify companies that are share repurchasers in a given year using the method suggested in Grullon and Michaely (2002). In the COMPUSTAT universe, we define stock repurchases as annual data item 115 (purchase of common and preferred stock) less the reduction in the value of any preferred stock outstanding (annual data item 56). We label a firm as a repurchaser of shares if this difference is greater than zero. We also separately analyze large repurchasing firms, defined as those whose repurchases exceed 1% of total assets. 3.2 The Sample of 10-Ks Our sample of 10-K s comes from Hoberg and Phillips (2010a), who search the Edgar database for filings that appear as 10-K, 10-K405, 10-KSB, or 10-KSB40. The business descriptions appear as Item 1 or Item 1A in most 10-Ks and are scanned using PERL web crawling scripts, APL programming, and human intervention when documents are non-standard. The document is processed for text information and a company identifier, CIK. A very small number of firms (roughly 100) did not contain a valid description or had a business description of less than 1,000 characters. Such firms are excluded from our analysis. 12

15 Our primary sample includes filings associated with firm fiscal years ending in calendar years 1997 to We also use 1996 data to compute text-based variables requiring lagged data, but do not use 1996 data otherwise. We merge each firm s text product description to the CRSP/COMPUSTAT database using the central index key (CIK), which is the primary key used by the SEC to identify the issuer. 7 Of the 45,631 observations available in CRSP and COMPUSTAT database noted above, we are left with 43,904 after requiring that same-year text data is available. The final sample is then 42,878, covering years 1997 to 2008, after requiring that lagged textbased data is also available. Lagged text based data is needed to compute variables including product market fluidity, which are based on how product markets change from one year to the next. As discussed in Hoberg and Phillips (2010a), this text based database generally has uniform coverage of the CRSP and COMPUSTAT sample during these years. 3.3 Word Vectors and Cosine Similarity We employ methods used in Hoberg and Phillips (2010a) and Hoberg and Phillips (2010b) to construct word vectors and measure similarity. The first step is to parse the product descriptions from the firm 10-Ks and form word vectors for each firm. These vectors can be used to compute cosine similarities that can measure how similar each firm s vector is to other vectors, which include key word lists, word lists of other firms, or vectors that measure product market flux. To construct each firm s word vector, we first omit common words that are used by more than 5% of all firms. We then consider the universe of all product words in the 10-K universe in each year. Let M t denote the number of such words. For a firm i in year t, we define its word vector W i,t as a binary M t -vector, having the value one for a given element when firm i uses the given word in its year t 10-K business 7 We thank the Wharton Research Data Service (WRDS) for providing us with an expanded historical mapping of SEC CIK to COMPUSTAT gvkey, as the base CIK variable in COMPUSTAT only contains the most recent link. 13

16 description. We then normalize each firm s word vector to unit length, resulting in the normalized word vector N i,t. The cosine similarity for any two word vectors W i,t and W j,t is their dot product W i,t W k,t. Cosine similarities are bounded in the interval [0,+1] when both vectors are normalized to have unit length, and when they do not have negative elements, as will be the case for the quantities we consider here. If two firms have similar products, their dot product will tend towards 1.0 while dissimilarity moves the cosine similarity toward zero. We use the cosine similarity method because it is widely used in studies of information processing (see Kwon and Lee (2003)). It measures the angle between two word vectors on a unit sphere. 3.4 Product Market Fluidity A key new variable we use in our study is a measure of how fluid or stable a firm s products are. Our fluidity variable is based on observing how 10-K business descriptions change year to year. Our first measure computes how much a firm s current products change relative to its own past-year products. We label this variable self product fluidity. This variable is simply one minus the cosine similarity of the text in the current year s business description and its previous year s business description. We consider a second measure of product fluidity based not on a firm s own product description changes, but on how actively rivals in the broader product space are changing products. We label this variable product market fluidity. To construct this variable, we first define the broad change in usage for each word from year t-1 to year t (word vectors W i,t are defined above): D t 1,t Σ j (W j,t W j,t 1 ) (1) The variable D t 1,t is a vector having the same number of elements as each W j,t, and this variable describes the extent to which each word is changing in usage across all product descriptions. This vector does not have an i subscript, and is not firm specific. We define each firm s product market fluidity as cosine similarity between 14

17 each firm s word usage vector and D t 1,t. Product Market Fluidity i N i,t D t 1,t D t 1,t (2) This measure identifies product instabilities and competitive pressure arising from peer firms, not instability in a firm s own product vocabulary. A high score means the firm s product words experience high turnover by other firms in related product spaces. A further refinement is suggested by the work of Hoberg and Phillips (2010b) (see their Appendix B for details). We compute fluidity variables using three dictionaries: (A) all words in all business descriptions (the overall dictionary ), (B) limit words to those with a local clustering coefficient in the top two terciles (the local dictionary ), and (C) those with a local clustering coefficient in the lowest tercile (the broad dictionary ). While the formula for similarities remains Eq. (2), the calculation is based on different dictionaries in each case. Local product market fluidity is thus based only on words that tend to persistently appear with a localized group of related words (ie, they appear in word cliques as identified by the local clustering coefficient). Intuitively, product market words have this property: the words cola, beverage, and drink should appear together in a 10-K if they appear at all. Because words in the local dictionary are represented by words unique to local product markets, we use local product fluidity as our primary variable of interest. 8 It is worth stressing that the words in the broad dictionary are not common words such as the or and because words in more than 5% of 10-K s are screened out, as discussed earlier. Rather, they are non-common words that do not generally appear in cliques (local clustering coefficient in the lowest tercile). Although these words are generally not specific to a local product market, they are often indicative of a firm having assets that can be easily redeployed to other product markets. Examples of words in the broad dictionary that are consistent with this interpretation include 8 However we also report in Table 9 that our results are similar if we use overall product market fluidity or broad fluidity. 15

18 container, painting, pallet, frames, forecasting, learning, fabric, and salespeople. In the literature studying the interaction of industrial organization and corporate finance, a potential endogeneity issue is that the agents, e.g., senior management, who set the financial policies also choose the product market strategies. This issue is somewhat mitigated in our study. While a firm s own top management certainly sets its payout policies, fluidity reflects moves by rival firms competing in a firm s product space. We also use both local and broad measures of product fluidity and examine if our results are robust across measures. In particular using a broad measure of product fluidity focuses on the changes in the broad set of words that all other firms in the Compustat/CRSP universe use. This broad set is like the familiar market index return in analyzing returns. We believe it is likely that changes in the broad vector of all words is exogenous from any one firm s perspective, just as the market return is considered exogenous to any single firm s return. 3.5 Business Clientèle Similarity We next compute the degree to which a firm likely sells its products to business clientèle. We consider the input/output tables from the Bureau of Economic Analysis, which report who buys the output produced for a comprehensive array of product categories. These tables separately report output sold as personal consumption, as well as output sold to other industries (business clientèle). Important from our perspective, the tables are organized based on verbal product lists, and these verbal lists are what we extract and use for textual analysis. We first extract the product words for products that have over 90% of their output sold to business (non-retail) customers. We then compute the cosine similarity of each firm s own product words (based on the vector N i,t ) to these business clientèle words. The result is a measure of the degree to which firm i likely sells its products to business clientèle in year t. This variable is recomputed for each firm in each year to capture potential changes to a firm s product mix over time. 16

19 Table 2 reports more detailed information about our text-based variables. We present these results to provide more insight regarding the specific words that drive our results. The results suggest that our text-based variables generate plausible classifications. [Insert Table 2 Here] In Panel A, we list the product market words used to score firm product descriptions based on the degree of business clientèle they have. Some products in this list, such as minerals, metal, and wood are pure natural resources. Other products are manufactured intermediate inputs into various manufacturing functions, including chemicals, plastics, coatings, and adhesives. Other products are uniformly needed by an array of businesses, such as paperboard, transportation, and refrigeration. What is common to virtually all products in this list is that they are sold to businesses more than they are sold to retail consumers. These products also are related to numerous different industries, indicating that our results likely are not driven by a small group of industries. In Panels B to E, we report firms that score in the lowest 25 or highest 25 based on their local product market fluidity in 1997 or 2008 (the first and last years of our sample). In these panels, firms are sorted from most extreme to least extreme. In Panels B and D, we report firms that have the lowest local product fluidity (firms with stable products) based on their fiscal year 1997 and 2008 SEC filings. Many of these firms are commodity firms as well as department stores. In Panels C and E, we report firms that have the highest local product fluidity (firms with unstable products) based on their fiscal year 1997 and 2008 SEC filing respectively. A considerable fraction of these firms are in the biotechnology sector, communications, and gaming industries. Biotechnology, at this time, is likely in a high state of flux due to impending regulatory changes, as well as the ongoing shocks related to innovation and government approval. Communications and media experienced changes at this time with ongoing battles between cable, television, and online media. 17

20 3.6 Control Variables We include controls suggested by the dividend literature in addition to others that could be correlated with our text measures. Following Fama and French (2001), we control for firm size using NYP (NYSE size percentile), which is the fraction of NYSE firms having equal or smaller capitalization than firm i in year t. Earnings/Assets (profitability), which is earnings before extraordinary items plus interest expense plus income statement deferred taxes divided by assets. Following Hoberg and Prabhala (2009), we control for a firm s risk by including the standard deviation of its daily stock returns from CRSP in the given calendar year. We also control for measures of growth options including market a firm s marketto-book ratio, asset and sales growth, R&D, and a firm s patenting intensity. These variables are constructed as follows. Book Equity (BE) is Stockholder s Equity minus Preferred Stock plus Balance Sheet Deferred Taxes and Investment Tax Credit minus Post Retirement Asset. If stockholder s equity is not available, it is replaced by either Common Equity plus Preferred Stock Par Value, or Assets - Liabilities. Preferred Stock is Preferred Stock Liquidating Value [or Preferred Stock Redemption Value, or Preferred Stock Par Value. Market Equity is the fiscal year closing price times shares outstanding. We control for R&D in addition because firms spending more money on R&D may invest more in creating new products. We also consider whether our measure provides any information not already contained in R&D and thus include lagged R&D divided by sales as an additional control. Asset growth, is the percent growth in assets from year t-1 to year t. We measure patenting activity through a combination of two data sources: (A) the NBER patent citations file as extended by Bronwyn Hall and (B) usage of the words {patent, patents} in each firm s product description. The Text + Applied Patents variable is one if the given firm mentions patents in its product description, or if it applied for a patent in the most recent three year window. 18

21 We include additional variables that could be correlated with our product text variables. We control for firm age by including the natural log of one + firm age. We compute age for a given firm in a given year as the current year minus its founding date. For the 15.2% of firms in our sample missing age data, we use the CRSP listing vintage as a substitute for the founding date. 9 We also include a negative earnings dummy in all specifications as DeAngelo, DeAngelo, and Skinner (1992) have shown this has a large impact on dividend payout. This variable equals one when a firm reports negative earnings in a given year and is zero otherwise. Lastly, we account for the current product market competition faced by firms. We consider the each firm s HHI (Herfindahl-Hirschman Index) based on the Variable Industry Classification (VIC) industries formed using firm-by-firm similarity measures as in Hoberg and Phillips (2010a). These measures are updated each year. Firm i s industry cluster comprises firms j whose product descriptions have similarity to i s products exceeding a threshold, as discussed in Section V.A. of Hoberg and Phillips (2010a). Following these procedures, we compute HHI concentration measures based on sales for all firms in each firm s VIC-7.06 industry. 10 This measure also excludes firm pairs in industries that are more than 1% vertically related based on BEA input-output tables. 3.7 Summary Statistics Table 3 presents summary statistics for the variables we use in our study covering 1997 to Statistics for payout policy variables are reported in Panel A. The table shows that 24.5% of the firms are dividend payers while a larger set of firms, 41.2%, are repurchases. There is some separation of firms among these payout variables as only 15.7% of firms both repurchase shares and pay dividends in the same 9 Our results are similar if we use listing vintage for all observations instead of the founding date. We thank Gustavo Grullon and James Weston for generously providing us with the data on firm founding dates used to construct this control variable. 10 The similarity cutoff of 7.06 is chosen so that industry classifications are as coarse as three digit SIC code industries. 19

22 year. Although many firms increase dividends (8.9%), less than 1% of firms decrease dividends in any year. [Insert Table 3 Here] Statistics for our key text-based variables are presented in Panel B, and key control variables are reported in Panel C. Self Fluidity is simply one minus the cosine similarity between firm i s year t product description and its year t-1 product description. Higher values indicate that the firm is changing its own product composition. This variable is non-negative by construction, and the statistics suggest that some firms experience very little change from year to year, whereas others experience much more substantial changes. Our product market fluidity variable measures the extent to which product market words used by firm i are being adopted and dropped by other firms at a high rate. This measure is multiplied by 100 for convenience, and is also non-negative by construction. The variables in Panel C include other variables known to explain payer status as noted in Hoberg and Prabhala (2009). Table 4 reports the Pearson correlations between our variables. The first five columns display results for our key text-based product measures. Perhaps not surprisingly, product market fluidity is positively correlated with the VIC-7.06 total similarity measure at the industry level (65.7%). Firms that are more similar to their local industries have rivals that are more likely to travel in the product space, and reside in markets with greater flux. The other moderate correlation is between the VIC based HHI and the VIC similarity measure (-33.7%). This correlation reflects the intuition that the more similar firms are, the lower is concentration as market shares are more evenly dispersed when firms are similar. Later we confirm that our results are not affected by these modest correlations. [Insert Table 4 Here] Product fluidity is -34.4% correlated with log firm age, and self fluidity is -13.5% correlated with log firm age. This confirms the basic intuition that firms generally 20

23 achieve higher product stability over time. However, the correlations are far less than 100% confirming that product life cycles are also distinct. For example, older firms might experience technological shocks, forcing product markets back to a more fluid state. 4 Results 4.1 Do Text Based Variables Add Explanatory Power? Before presenting more extensive regression evidence, we first analyze whether our text-based variables of product stability and competition capture additional explanatory power over other accounting variables such as those in Fama and French (2001). We estimate the propensity to pay dividends for our panel from 1997 to 2008 including the base Fama and French (2001) and risk variables excluding our text variables. Holding the propensity to pay constant, we then examine whether the text variables are different for payers and non-payers. We estimate a base specification without the text variables, split firms into propensity to pay quartiles and then compare the mean and median of the text variables within each propensity quartile. Table 5 reports the results. Explanatory variables for the logit include non-text variables: total risk log firm age, firm size, M/B, asset growth, the negative earnings loss dummy of DeAngelo, DeAngelo, and Skinner (1992), R&D/Sales and year fixed effects. Within each quartile of the predicted propensity to pay, we report the mean and median of the five text-based product characteristics for payers and non-payers. [Insert Table 5 Here] The results in Table 5 show that there are significant differences in our text variables for each propensity quartile. Within all quartiles of predicted dividend propensity, firms that actually pay dividends have significantly lower values of local fluidity. They are also more likely to sell to business clientèle. The text variables appear to contain additional explanatory power over and above variables that have 21

24 been shown to affect payout propensity in past studies. DeAngelo, DeAngelo, and Skinner (2004) and DeAngelo, DeAngelo, and Skinner (2009) argue that there are several large companies with long dividend paying histories that continue to pay dividends. Time series variation in dividend paying propensity are thus driven by small to mid-size firms. In unreported tables, we also exclude the beginning of period payers in the top size quartile and find similar results, suggesting that our results are not driven by the very large firms with little variation in their payouts. 4.2 Multivariate Evidence The Propensity to Pay Dividends In this section we examine the propensity to make payouts using multivariate logistic and tobit regressions. We start with the propensity to pay dividends, and test the key implications of hypotheses H1 and H2. We then examine changes in payout policy including dividend initiations, omissions, increases and decreases. In all the logit models estimated here, we standardize the independent variables so they have unit standard deviation. This scaling does not effect significance levels or the economic impact of our variables. Panel A of Table 6 presents the results of estimating a logit model where the dependent variable is one if the firm pays dividends. With the exception of row (7), which uses a linear probability model, Panel A displays the results of logit regressions using a panel specification with time fixed effects and standard errors adjusted for clustering by firm. Panel B estimates panel tobit regressions where the dependent variable in row (8) is a firm s dividend yield - where yield is expressed as dividends divided by fiscal year-end price and is dividends divided by assets in row (9). [Insert Table 6 Here] Both panels in Table 6 show similar results. Firms in fluid product markets 22

25 are less likely to pay dividends and pay lower dividend yields. 11 The multivariate specifications for the dividend payer dummy and dividend yields in Rows (6) to (9) show that this effect is attributed mainly to local product market fluidity (the tendency of rival firms in a localized product market to change their products). This result confirms our first hypothesis that firms with stable products are more likely to pay dividends. These results are consistent with competitive risk (Hypothesis H1) and the product life cycle theories of Abernathy and Utterback (1978) and Klepper (1996). These findings are also robust to including other text-based product market variables and other explanatory variables including firm risk, R&D and firm maturity. Table 6 also shows that our localized measure of competition (VIC-7.06 HHI) is positively related to dividend propensity. However, its economic effects are not large. This result is different from Grullon and Michaely (2007), and hence we investigate this relationship further. Our samples are different, as Herfindahl measures based on the Census data are only available for manufacturing firms, and also our sample begins in 1997 due to the required availability of machine readable 10-Ks. Thus while our sample is broader as we include non-manufacturing industries, it is also shorter as we do not have a long panel of data. We reestimate our model including the Census Herfindahl based on SIC codes and find that this measure of concentration is insignificant in our sample, and our fluidity and business clientèle variables remain significant. The business clientèle text variable, which captures the extent to which a firm s products are likely to be sold to business clientèle, is also significant in all specifications. Consistent with Hypothesis H2, this suggests that firms selling to business clientèles are more likely to pay dividends. In later tests, we compare this result to our results for share repurchases, as existing literature suggests that repurchases are not as strongly linked to long-term stability relative to dividends. 11 As discussed earlier, we base our main results on local product market fluidity, which restricts 10-K words to those that appear in localized product market cliques. Our results are robust if we instead use all words. 23