Persistence of the Complementary Relation between Earnings and Private Information

Transcription

1 Persistence of the Complementary Relation between Earnings and Private Information Ian D. Gow Harvard Business School Daniel J. Taylor The Wharton School University of Pennsylvania Robert E. Verrecchia The Wharton School University of Pennsylvania May 30, 2012 We thank Mary Barth, Judson Caskey, John Core, Ilia Dichev, John McInnis, and seminar participants at Carnegie-Mellon, MIT, University of North Carolina, and the 2012 FARS Midyear Meeting for helpful comments. We gratefully acknowledge the financial support of the Harvard Business School, and the Wharton School. We thank John Core for sharing his estimates of accrual quality and Rodrigo Verdi for sharing his estimates of institutional investor concentration. 1

2 Persistence of the Complementary Relation between Earnings and Private Information Abstract This paper examines whether, for some firms, the complementary relation between earnings information and private information is persistent. First, consistent with earnings information being a complement to the private information held by some investors, we show that more precise earnings are associated with a larger increase in information asymmetry at the time of the earnings announcement. Second, consistent with the complementary nature of earnings leading to persistent differences in information asymmetry for some firms, we find more precise earnings are associated with higher levels of information asymmetry over the subsequent year. As a result, for these firms, we show that when equity markets are imperfect, earnings information reduces expected returns by reducing market risk, but increases expected returns incremental to market risk by increasing information asymmetry. Keywords: complementary information, information precision, information asymmetry, market competition, disclosure, expected returns, cost of capital, risk, earnings volatility JEL Classification: D41, G12, G14, M41 2

3 1. Introduction The purpose of this paper is to examine whether the complementary relation between earnings and private information previously documented in short windows around earnings announcements is more persistent than commonly thought. Conventional wisdom suggests that the public disclosure of more precise information reduces information asymmetry and results in lower expected returns. This wisdom, however, relies on the notion that public information substitutes for the private information held by some investors. When public information complements the private information held by some investors, the disclosure of more precise information may actually exacerbate information asymmetry and result in higher expected returns. The well-known increase in information asymmetry at the time of the earnings announcement is perhaps the most prominent example of the existence of a complementary relation between public disclosure and investors private information (e.g., Lee, Mucklow, and Ready, 1993; Krinsky and Lee, 1996). 1 The theoretical literature offers little guidance as to whether public and private information are substitutes or complements, and, in the case of the latter, the persistence of the complementary relation. For example, some theoretical analyses posit that public information substitutes for the private information held by investors, and thus more precise public information reduces information asymmetry (e.g., Diamond and Verrecchia, 1991). Other theoretical analyses posit that heterogeneity in the cost of processing public information, heterogeneity in information processing biases, and/or the existence of private information useful in conjunction with public information can result in more precise public information 1 Additional evidence of the complementary relation between earnings and private information comes from studies that suggest the private information imbedded in analyst forecasts increases around earnings announcements (e.g., Barron, Byard, and Kim, 2002). 3

4 disproportionately benefiting more sophisticated investors at the expense of less sophisticated investors (e.g., Indjejkian, 1991; Kim and Verrecchia, 1994, 1997; Fischer and Verrecchia, 1999). In other words, the theoretical literature leaves room for public information to act as either a substitute for, or complement to, private information gathering activities. In addition, in most theoretical analyses there is no notion of time : most analyses are couched in terms of a single period. Thus, irrespective of whether public information substitutes for or complements private information, the theoretical literature offers little guidance as to the persistence of this relation. While prior empirical studies suggest that earnings complement investors private information around earnings announcements (e.g., Lee et al., 1993; Krinsky and Lee, 1996; Barron et al., 2002), these papers leave unexamined whether, for some firms, this complementary relation results in persistent differences in information asymmetry. The goal of this paper is to examine the persistence of the complementary relation between earnings and private information. In theory, the extent to which complementary information increases information asymmetry depends on the precision of the information, with more precise information leading to a larger, and potentially persistent, increase in information asymmetry. Accordingly, we examine whether more precise earnings are associated with higher levels of information asymmetry, and whether this association persists over long windows. If, for some firms, the complementary nature of earnings results in persistently higher levels of information asymmetry, it raises the possibility that the complementary nature of earnings will manifest in expected returns. In perfectly competitive markets, information only affects market risk through the average level of the precision of investors information, and there is no separate effect from information asymmetry (see Lambert, Leuz, and Verrecchia, 2012). 4

5 Thus, in perfectly competitive markets, more precise earnings reduce expected returns by increasing investors average precision (and thereby reducing market risk) regardless of whether earnings complements or substitutes for private information (e.g., Hughes, Liu, and Liu, 2007; Lambert, Leuz, and Verrecchia, 2007). However, in markets that depart from perfect competition, information asymmetry has a separate effect on expected returns that affects expected returns incremental to market risk (e.g., Armstrong, Core, Taylor, and Verrecchia, 2011; Akins, Ng, and Verdi, 2012). Thus, if the complementary relation between earnings and private information persists over long windows, more precise earnings will reduce expected returns by reducing market risk, but will increase expected returns incremental to market risk by increasing information asymmetry. Importantly, the persistence of a complementary relation between earnings and private information over long windows can manifest in a positive relation between earnings precision and expected returns. This potentially explains the puzzling and counter-intuitive result in prior literature that firms with less volatile earnings have higher expected returns. 2 While in principle there are a variety of empirical proxies for earnings precision, precision itself is a well-defined statistical construct that measures the reciprocal of the variance of an information signal, in this case earnings (e.g., Kim and Verrecchia, 1994, 1997). Accordingly, we closely follow the literature and use the variance of earnings as the basis for our measures of earnings precision. We discuss several measurement issues relating to empirical measures of earnings precision, and distinguish between the variance of earnings as it relates to 2 Hints of a positive relation between earnings precision and expected returns are found in the literature, but have been either dismissed as anomalous, or attributed to an optimistic bias on the part of investors. For example, Gebhardt, Lee, and Swaminathan (2001) and Francis, Lafond, Olsson, and Schipper (2004) are surprised to find less volatile (more precise) earnings are associated with a higher implied cost of capital but only mention this finding in passing. Huberts and Fuller (1995) and Myers, Myers, and Skinner (2007) also find that firms with less volatile earnings have higher expected returns, but attribute this to an optimistic bias on the part of investors. 5

6 fundamental volatility and measurement error. We report results using four different proxies for earnings precision. We find results are robust to controlling for fundamental volatility, robust to using different deflators (e.g., market value, assets, and the absolute value of average earnings), and robust to alternative measures of earnings precision including two measures that control for fundamental volatility (i.e., the variance of the component of earnings orthogonal to fundamental value, and the value relevance of earnings). We also validate each of our measures of earnings precision by demonstrating that firms with more precise earnings have higher earnings response coefficients. 3 Our results are as follows. First, we confirm the complementary relation between earnings and private information. We find that firms with more precise earnings tend to have greater increases in information asymmetry at the time of the earnings announcement. Second, consistent with the complementary nature of earnings information persisting over long windows, we find that firms with more precise earnings have higher levels of information asymmetry over the subsequent year. Third, we find that earnings precision is associated negatively with the firm s market risk (beta), but associated positively with expected returns incremental to market risk (alpha). Consistent with our predictions, we find the negative relation between earnings precision and beta does not vary with the level of market competition, but the positive relation between earnings precision and alpha is strongest when market competition is low. In additional analyses, we use industry membership to identify variation in earnings precision outside the control of the firm. Consistent with our firm-level results, we find firms in 3 Our predictions apply to complementary information in general. However, earnings information where earnings information communicates firm value with error is closest to the spirit and intent of the theoretical literature on information complementarities (e.g., Kim and Verrecchia, 1994, 1997). For example, the theoretical literature commonly examines whether and how the precision of earnings information attenuates or exacerbates information asymmetry among investors. 6

7 industries with more precise earnings tend to have (i) higher information asymmetry, (ii) lower market risk (beta), and (iii) higher expected returns incremental to market risk (alpha) when market competition is low. The results suggest that the complementary relation between earnings and private information can have long-lasting effects on information asymmetry and expected returns. Counter to conventional wisdom, our results suggest that when public and private information are complements, more precise public information can actually increase information asymmetry and expected returns. Importantly, the results also suggest a rational, information-based explanation for previously documented anomalous patterns in expected returns. The remainder of the paper proceeds as follows. Section 2 reviews the relevant prior literature. Section 3 describes how we measure key variables and the research design of our empirical tests. Section 4 describes our sample. Section 5 presents our findings and robustness tests, and Section 6 concludes. 2. Summary of hypotheses and review of related literature 2.1 Complementary information and information asymmetry Most studies that examine the economic consequences of providing investors with more information explicitly or implicitly assume that public information decreases the information advantage of sophisticated investors, and leads to leveling the playing field among investors (see Botosan, 2006, Leuz and Wysocki, 2009, Bamber, Barron, and Stevens, 2012 for reviews). In this characterization of public information, public and private information are said to be substitutes. An alternative characterization, however, is one in which public information complements private information held by sophisticated investors (e.g., Lundholm, 1988). For example, prior research shows that public information can disproportionately enhance the 7

8 knowledge of sophisticated investors if there is heterogeneity in processing costs (Indjejikian, 1991), if sophisticated investors have private information useful in conjunction with the disclosure (Kim and Verrecchia, 1994, 1997), or if unsophisticated investors do not process public information efficiently (e.g., Fischer and Verrecchia, 1999). 4 Thus, if public and private information are complements (substitutes), more precise public information will increase (decrease) information asymmetry. Despite the conventional wisdom that the public disclosure of information reduces the information advantage of sophisticated investors, prior research finds that earnings announcements actually stimulate private information gathering (e.g., Barron et al., 2002). Consistent with the notion that earnings information complements the private information of sophisticated investors, Lee et al. (1993) and Krinsky and Lee (1996) find pronounced increases in intra-day spreads around the time of the earning announcement. However, these papers do not examine cross-sectional variation in the increase in information asymmetry, or whether the complementary nature of earnings can lead to persistent differences in information asymmetry. In theory, the effect of complementary information depends on its precision, with more precise information leading to a larger increase in information asymmetry. Accordingly, we examine the relation between earnings precision and the increase in information asymmetry at the time of the earnings announcement. Importantly, because standard theoretical analyses that posit a positive relation between earnings precision and information asymmetry are based on a 4 When public information complements private information, more precise public information will increase the information of both unsophisticated and sophisticated investors, but will increase the information of sophisticated investors by a greater extent. Throughout the paper we use the term "sophisticated investors" to describe investors that have lower information processing costs, possess private information useful in conjunction with public information, and/or are more efficient processors of information. 8

9 single period (e.g., Indjejikian, 1991; Kim and Verrecchia 1994, 1997), they provide no guidance on whether the relation is transitory or persistent. One possibility is that the impact of earnings precision on information asymmetry is temporary because subsequent information events quickly eliminate the information-processing advantage or private information of sophisticated investors. However, it is also possible that the information-processing advantage enjoyed by sophisticated investors is persistent (e.g., differences in processing costs), and/or the private information of sophisticated investors is not revealed in a timely manner following the earnings announcement. Ultimately, it is an empirical question whether the complementary nature of earnings information persists beyond earnings announcements. If the effect of earnings on the information advantage of sophisticated investors persists, we predict firms with more precise earnings have persistently higher levels of information asymmetry. Our focus on the precision of earnings differs from prior work that examines the relation between information asymmetry and disagreement among security analysts (e.g., Affleck- Graves, Callahan, and Chipalkatti, 2002). Disagreement among agents about the interpretation of an information signal is a different concept than the precision of information (e.g., Aumann, 1976: Bamber et al., 2012). 5 Disagreement relates to how agents differentially update their beliefs conditional on observing the same information, as opposed to some underlying characteristic of the information signal itself. Importantly, in our empirical results we find a positive association between information asymmetry and the dispersion of analyst forecasts, and an incremental positive association between information asymmetry and earnings precision. This 5 Affleck-Graves et al. (2002) use the dispersion of analyst forecasts to measure earnings predictability, but most subsequent work uses the dispersion of analyst forecasts to measure disagreement (e.g., Diether, Malloy, and Scherbina, 2002) and properties of earnings to measure earnings predictability (e.g., Dichev and Tang, 2009). 9

10 is consistent with the relation between information asymmetry and forecast dispersion reported in Affleck-Graves et al. (2002), and is consistent with forecast dispersion and earnings precision measuring different constructs. The focus on the precision of earnings itself also distinguishes this paper from prior work that examines the relation between information asymmetry and the volatility of discretionary accruals or volatility in how accruals map into earnings (i.e., accrual quality see Jayaraman, 2008; Bhattacharya, Desai, and Venkataraman, 2012). Rather than examine the volatility of earnings itself, these papers conjecture that when accruals are more volatile than what would otherwise be expected, information is garbled. Garbling, in turn, results in increased information asymmetry. These papers find poor quality accruals are associated with higher information asymmetry. Importantly, earnings precision and accruals quality are two distinct concepts (e.g., Liu and Wysocki, 2006; Wysocki, 2009; Dechow, Ge, and Schrand, 2010). Holding constant the quality of accruals, if the information in earnings complements private information, the theory makes clear that more precise earnings should increase information asymmetry. Consistent with this, in our empirical results, we find a negative association between information asymmetry and accruals quality and an incremental positive association between information asymmetry and earnings precision. This is consistent with the relation between information asymmetry and accruals quality reported in prior work and suggests accruals quality and earnings volatility measure different constructs. 2.2 Complementary information and expected returns An extensive literature examines the relation between the precision of information, information asymmetry, and expected returns. One approach used in this literature is to study the relation between market-based measures of information asymmetry, such as the bid-ask spread 10

11 (Amihud and Mendelson, 1986) or the probability of informed trade (PIN) (Easley, Hvidkjaer, and O Hara, 2002), and expected returns. The evidence, however, has been mixed. Recent theory-based work highlights the importance of distinguishing two potential channels through which more precise information can affect expected returns. The first channel, which we call the average precision effect, is the effect that more precise information has on the average precision of investors information. The second channel, which we call the information asymmetry effect, is the effect that more precise information has on the extent to which investors are dissimilarly informed (i.e., the degree of information asymmetry among investors). Figure 1 summarizes these effects and their predicted relations with expected returns. Prior literature shows that when markets are perfectly competitive, more precise information affects expected returns only through its affect on the average level of the precision of investors information (i.e., the average precision effect). In other words, in perfectly competitive markets, more precise information reduces expected returns by reducing market risk, and there is no incremental effect from information asymmetry (e.g., Hughes et al., 2007; Lambert et al., 2007). 6 When markets are imperfectly competitive, however, Lambert et al. (2012) show that information asymmetry can affect expected returns incremental to market risk (i.e., the information asymmetry effect). Thus, in imperfect markets, these papers suggest that more precise information affects expected returns in two ways: through the average precision effect (i.e., through market risk), and through the information asymmetry effect (i.e., incremental to market risk). 6 Beaver, Kettler, and Scholes (1970) is perhaps the first paper to find evidence of a negative relation between variance of earnings and market risk. However, they view their evidence as descriptive and do not offer an explanation for the relation. Related, Ng (2011) finds a negative relation between earnings precision and liquidity risk. 11

12 To the extent that earnings complements private information and this complementarity is persistent, these two effects will be countervailing. On the one hand, more precise earnings will reduce expected returns by increasing average precision and reducing market risk. On the other hand, more precise earnings will increase expected returns incremental to market risk by increasing information asymmetry. While the opposing direction of the effects facilitates their empirical detection, it is unclear which of the information asymmetry or average precision effects will dominate. If the average precision effect dominates, then, true to the conventional wisdom, more precise earnings will be negatively related to expected returns. If, however, the information asymmetry effect dominates, then counter to the conventional wisdom, more precise earnings will be positively related to expected returns. In this regard, our findings shed light on a number of puzzles in the literature. While prior literature has pointed to the possibility of a counter-intuitive positive relation between earnings precision and expected returns, researchers have attributed this relation largely to behavioral explanations or dismissed it as anomalous. The earliest evidence appears in Huberts and Fuller (1995), who find that a hedge portfolio based taking a long (short) position in firms with low (high) earnings volatility earns annual abnormal returns in excess of 4% per year. Huberts and Fuller (1995) attribute this to prices of high volatility stocks reflecting excessive optimism on the part of investors. In later work, Gebhardt et al. (2001, p. 165) find more volatile earnings are associated with lower implied cost of capital but dismiss the result as anomalous. Similarly, Francis et al. (2004, p.993) find a reliably negative relation between the variance of unexplained earnings and the implied cost of capital despite predicting a positive one. Myers et al. (2007) also find that firms with less volatile earnings have higher future returns, and similar to Huberts and Fuller (1995), attribute this to an optimistic bias on the part of 12

13 investors. In contrast to these studies, Frankel and Litov (2009) and McInnis (2010) find no evidence of a relation between earnings volatility and future returns. 7 Collectively, the evidence of a negative relation appears not to have been explored in detail, perhaps because of mixed evidence, perhaps because the evidence has been perceived as anomalous given the conventional wisdom that less volatile (more precise) earnings should result in lower expected returns. By building on the complementary nature of earnings announcements, this study suggests a rational, information-based explanation for these seemingly counter-intuitive results. 3. Research design In this section, we describe the measurement of key variables and our empirical tests. First, we describe our measures of earnings precision, information asymmetry, and market competition. Second, we describe the set of tests we use to examine the relation between earnings precision and information asymmetry. Third, we describe the set of tests we use to examine the relation between earnings precision and the expected returns Measurement of key variables Measurement of earnings precision Earning precision is a well-defined statistical concept that measures the ability of earnings information to communicate firm value (e.g., Kim and Verrecchia, 1994, 1997). More precisely, the theoretical models on which we base our empirical predictions express a firm s earnings, E t, as a noisy communication of the unknown or uncertain value of the firm, V t : E t = V t + ε t (1) 7 Our results suggest the omission of market competition as a conditioning variable from prior work may explain the mixed findings across these studies. 13

14 where V t has standard deviation σ V and ε t has standard deviation σ ε. The variable ε t measures the noise in the communication of firm value through earnings. The reciprocal of σ 2 ε, (σ 2 ε ) -1, measures the precision of earnings information: higher (σ 2 ε ) -1 implies more precise earnings information, and lower (σ 2 ε ) -1 implies less precise earnings information. To measure ε t and thereby infer the precision of earnings information, ideally one needs to control for investors beliefs about V t, as ε t represents the value of E t having accounted for (or conditioned over) V t. Failure to control for V t in our empirical tests will result in a measure of earnings precision that reflects both the variance of ε t ( measurement error ) and the variance in V t ( fundamental volatility ). This will confound inferences regarding the effect of earnings precision, and will bias against finding a positive association between earnings precision and information asymmetry. For example, prior theoretical research suggests that fundamental volatility, σ 2 V, is positively related to information asymmetry regardless of whether earnings information complements or substitutes for private information (e.g., Kim and Verrecchia, 1994, 1997). 8 In contrast, the relation between measurement error, σ 2 ε, and information asymmetry depends on whether earnings and private information are substitutes or complements. If earnings and private information are substitutes, then measurement error will be positively related to information asymmetry. In this case, measurement error and fundamental volatility have similar effects on information asymmetry, and thus distinguishing between the two empirically will be difficult. However, if earnings and private information are complements, then measurement error will be negatively related to information asymmetry. In this case, measurement error and 8 The intuition for this result is that fundamental volatility measures ex ante uncertainty about value, and greater ex ante uncertainty increases information asymmetry. This is consistent with extant empirical work that documents a positive relation between firm risk and information asymmetry. Liu and Wysocki (2006) suggest measures of accruals quality primarily measure fundamental volatility, which is an alternative explanation for why firms with worse quality accruals (more volatile accruals) have higher information asymmetry. 14

15 fundamental volatility have countervailing effects on information asymmetry. Importantly, this allows us to distinguish empirically these effects. We use two approaches to control for V t in our empirical tests. First, we include controls for fundamental volatility (e.g., return volatility) in our regressions. Second, in addition to controlling for fundamental volatility in our regressions, we also estimate two measures of earnings precision designed to exclude the effect of V t. Regardless, because measurement error and fundamental volatility have countervailing effects when earnings and private information are complements, a measure of earnings precision that commingles measurement error and fundamental volatility will bias against finding that less measurement error (more precise earnings) is associated with higher information asymmetry. Consistent with prior literature, our first proxy for earnings precision, Precision, is calculated as the time-series standard deviation of earnings (Compustat Item IB) scaled by average market value and multiplied by negative one. We scale by market value to control for differences due to scale (i.e., nominal size of the firm), and multiply by negative one for expositional reasons (i.e., higher values of Precision correspond to more precise earnings). To ensure our results are not sensitive to scaling by market value, our second measure of earnings precision, Precision2, is constructed similarly but is scaled by total assets rather than by market value. 9 While similar to prior research (e.g., Dichev and Tang, 2009), these measures conflate measurement error and fundamental volatility and will thus bias against finding empirical support for our predictions. In addition to controlling for fundamental volatility in our regressions, we also estimate two alternative measures of earnings precision designed to exclude the effect of V t. The first of 9 We also find inferences are robust to scaling by the absolute value of earnings over the ten year period (e.g., Minton, Schrand, and Walther, 2002), but for parsimony do not tabulate results for this measure. 15

16 these two alternative measures (Precision3) is the variance of the component of earnings orthogonal to value (V t ). In particular, for each firm we estimate the regression E t = δ 0 + δ 1 V t + ε t (2) where E t is earnings in year t and V t is the value of the firm at the end of year t. Because the value of the firm is not directly observable, we use the market value of equity as of the fiscal year-end to measure value. To the extent that market prices are an unbiased estimate of true value, ε t will represent the component of earnings orthogonal to value (more precisely, investors beliefs about value). 10 Similar to our earlier measures, Precision3 is constructed as the timeseries standard deviation of the component of earnings orthogonal to value, multiplied by negative one and scaled by average market value. In this regard Precision3 is a direct measure of the variance of ε t The second alternative measure of earnings precision (Precision4) does not attempt to measure earnings precision directly, but rather measures the returns-earnings relation, and relies on the notion that a stronger returns-earnings relation is the result of more precise earnings. In particular, Precision4 is calculated as the adjusted-r 2 from a regression of buy-and-hold returns over the twelve months period ended three months after the fiscal year-end on the level and change in earnings, both scaled by market value (e.g., Barth, Konchitchki, and Landsman, 2011). For each firm-year, we estimate all four measures using a rolling ten-year window that ends in the current year (i.e., t = 9,, 0). Using a rolling ten-year window has the advantage of minimizing the extent to which earnings in any given year affect the measurement of precision. 11 In untabulated analyses, we validate our measures of earnings precision by finding that firms 10 To the extent that market prices are a biased estimate of value, the estimate of δ 1 will capture this bias, and ε t will represent the component of earnings orthogonal to both fundamental value and the bias in market prices. 11 This is consistent with measuring a commitment to a given level of earnings precision. 16

17 with more precise earnings have higher earnings response coefficients (e.g., Barth, Landsman, and Lang, 2008). We tabulate results for our primary tests using all four measures of earnings precision, but present and discuss results for the last three measures at the end of the study Measurement of information asymmetry Our research design requires a measure of information asymmetry around earnings announcements as well a measure of information asymmetry over long windows. Drawing on prior research, we measure information asymmetry using four market-based measures of information asymmetry: Illiquidity, Spread, λ GH, and λ MRR. Our first measure of information asymmetry, Illiquidity, is the Amihud (2002) illiquidity measure defined as R t Illiquidit yt = (3) DVolumet where R t is the daily return and DVolume t is the daily dollar volume. Importantly, this measure of information asymmetry is calculated using data on CRSP. The remaining three measures of information asymmetry are calculated using intraday data on the Trades and Automated Quotation (TAQ) database. Our second measure of information asymmetry, Spread, is the average effective bid-ask spread during the day scaled by trade price and is defined as N Spread t = 1 Ask n Bid n, (4) N price n n=1 where Ask n (Bid n ) is the quoted ask (bid) for trade n on day t, and price n is the trade price for trade n on day t and N is the number of trades on day t. Because information asymmetry is not the only determinant of the bid-ask spread, we also consider two measures of the information asymmetry component of the bid-ask spread, or λ. Our first measure is the Glosten and Harris (1988) measure of the information asymmetry component of the bid-ask spread, λ GH (e.g., Akins 17

18 et al., 2012). Our second measure is the Madhavan, Richardson, and Roomans (1997) measure of the information asymmetry component of the bid-ask spread, λ MRR, (e.g., Armstrong et al. 2011). 12 When examining the change in information asymmetry around earnings announcements, we measure each of the four variables at the daily level and calculate the change in information asymmetry around the earnings announcement as the average value on the day of and day following the earnings announcement (t = 0, +1) less the average value over the fifty days prior to the announcement (t = 1,, 50), requiring data for at least ten days during the preannouncement period. When measuring information asymmetry over long windows, we measure each of the four variables at the monthly level and then take the average monthly value over the twelve month period beginning three months after the fiscal year end. All measures are winsorized at the 1st and 99th percentiles prior to calculating the average values over the respective measurement window Measurement of market competition We follow prior literature and use the number of shareholders on record as of the fiscal year-end (Compustat Item CSHR) as our primary measure of market competition, MktComp (e.g., Armstrong et al. 2011). This annual measure is report in firms annual 10-K filings and available for a large number of firms beginning in Our rationale for this choice is that prior research observes a wide range in the number of investors in U.S. firms, with some in the hundreds of thousands, thereby seemingly consistent with assumptions about perfect competition and price-taking behavior, but others in the hundreds, thereby less plausibly associated with perfect competition and price taking. Among the limitations of this measure are that it is 12 Details on the estimation procedure appear in Appendix B. 18

19 available only once per year and may be noisy because the SEC requires only that firms give the approximate number of shareholders of record, a figure that may not include individual shareholders when shares are held in street name (Dyl and Elliott, 2006). In untabulated analyses, we find that inferences are robust to using two alternative measures of market competition: the number of institutional investors and institutional investor concentration (e.g., Akins et al., 2012). 3.2 Empirical tests Earnings precision and information asymmetry We use two sets of tests to examine the relation between earnings precision and information asymmetry. Our first set of tests builds on literature that examines changes in information asymmetry around earnings announcements. In particular, we examine the relation between earnings precision and the change in information asymmetry at the time of the earnings announcement. In our second set of tests, we examine whether the relation between earnings precision and information asymmetry persists over longer windows. In particular, we examine the relation between earnings precision and the level of information asymmetry over the following year. We examine whether the change in information asymmetry at the time of the earnings announcement varies with earnings precision by estimating a series of regressions that take the form: ΔInfoAsym t+1 = φ 0 + φ 1 Precision t + θ Controls t + ε t+1, (5) where t denotes year, and ΔInfoAsym is a measure of the increase in information asymmetry at each of the quarterly earnings announcements for fiscal year t+1 (i.e., ΔIlliquidity, ΔSpread, Δλ GH, or Δλ MRR ), Precision is the measure of earnings precision, and Controls is a vector of 19

20 control variables. All independent variables are measured at the end of year t and, for ease of interpretation, are ranked into quintiles and scaled to range from 0 to We include the following variables as controls. RetVol is the standard deviation of monthly returns over the fiscal year. Beta is the estimated slope coefficient from a regression of monthly stock returns over the fiscal year on the market returns. MktComp is the number of shareholders of record as of the fiscal year-end (Compustat Item CSHR), Turnover is the average monthly share turnover (trading volume scaled by shares outstanding) during the year. Coverage is the number of analysts providing one-year-ahead earnings forecasts as of the fiscal year-end. ΔEarn is the change in earnings (Compustat Item IB) scaled by market value of equity at the end of the fiscal year. Loss is an indicator variable for whether the firm reported a loss for the year. Size is the market value of equity at the end of the fiscal year. BM is the book value of equity (Compustat Item CEQ) scaled by market value of equity at the end of the fiscal year. 1/Price is the inverse of share price (adjusted for stock splits) at the end of the fiscal year. NYSE (NASDAQ) is an indicator variable for whether the firm is listed on the New York Stock Exchange (NASDAQ). We also tabulate a specification that includes accruals quality (AccrQual) and analyst forecast dispersion (σanalyst) as additional controls. AccQual, is calculated following Francis et al. (2004) and Core, Guay, and Verdi (2008). 14 We multiply AccQual by negative one so that 13 Inferences in all specifications are similar if we use the raw values of the variables. Our independent variables are measured with a one-year lag so as not to induce any mechanical correlation between our measure of earnings precision and the current period's earnings surprise. Because our predictions relate to the notion of fixed or permanent earnings precision, the lag value of earnings precision is a valid instrument for current period earnings precision. Nonetheless, inferences across all specification are stronger if we repeat our tests measuring earnings precision contemporaneously with information asymmetry. 14 Specifically, AccQual is the standard deviation of residuals from a regression of total current accruals on lagged, current, and future cash flows, change in revenue, and property, plant and equipment all scaled by average total assets. Total current accruals are calculated as the change in current assets, less the change in current liabilities, less 20

21 larger values correspond to higher quality accruals. σanalyst is calculated following Gebhardt et al. (2001) and Diether et al. (2002) as the standard deviation of one-year ahead analyst earnings forecasts scaled by the absolute value of the average forecast, both measured as of the end of the fiscal year. Importantly, requiring analyst coverage and sufficient data to calculate accrual quality (e.g.,. ten years of accruals) reduces the sample by more than 50%. For this reason, we present results both with and without including these additional controls. In either specification, if the complementary relation between earnings information and private information persists over long windows, we expect a positive relation between earnings precision and the increase in information asymmetry around earnings announcements (i.e. φ 1 > 0). We examine whether the relation between earnings precision and the level of information asymmetry persists over the following year by estimating a series of regressions that take the form: InfoAysm t+1 = φ 0 + φ 1 Precision t + θ Controls t + ε t+1, (6) where t denotes year, and InfoAysm is a measure of information asymmetry, Precision is the respective measure of earnings precision, and Controls is a vector of control variables. As before, we measure the level of information asymmetry using four measures (i.e., Illiquidity, Spread, λ GH, or λ MRR ) and tabulate a specification that includes accruals quality and analyst forecast dispersion as additional controls. However, because we are interested now in longwindow measures of information asymmetry, we take the average monthly value of each of the information asymmetry variables over the twelve-month period beginning three months after the fiscal year-end. If the complementary relation between earnings information and private the change in cash, less depreciation, plus the change in long-term debt in current liabilities. Similar to our measures of earnings precision, the regression is estimated for each industry-year using a ten-year rolling window. 21

22 information persists over long windows, we expect a positive relation between earnings precision and information asymmetry over the subsequent year (i.e., φ 1 > 0) Earnings precision and expected returns To test our predictions regarding the relation between earnings precision and the expected returns, each year we form twenty-five (5 5) portfolios by independently sorting firms into five quintiles based on a proxy for earnings precision and five quintiles based on a proxy for market competition. We calculate expected returns for each portfolio and decompose expected returns into a market risk component and the component incremental to market risk (i.e., abnormal returns). 15 Following a large literature in asset pricing, we test our hypotheses about the relation between earnings precision and the expected returns using market beta as a measure of the market risk. To mitigate concerns that market betas are estimated with error, we follow the standard practice in the asset-pricing literature and group firms into portfolios, and conduct our analysis at the portfolio level. In our tests, we form portfolios by sorting based on the variables of interest (either earnings precision and/or market competition), and estimate market risk (β) and abnormal returns (α) for each portfolio. This calendar-time portfolio approach is used extensively in the finance literature to test asset-pricing models (e.g., Fama and French, 1993; Core et al., 2008). The primary advantages of this approach are that it does not assume that returns are linear in the variable of interest (i.e., the sort variable) and that it collapses the cross- 15 The benefit of independent sorts is that they ensure there is similar variation in earnings precision across each market competition portfolio. To the extent that the two sorting variables are correlated, the cost is a reduction in power (because of sparsely populated cells). However, our sample size should mitigate this concern: each portfolio contains at least 50 stocks in any given month. 22

23 section of returns (on a given date) into a single time-series observation, thereby alleviating concerns about cross-sectional dependence and outliers. 16 Following Fama and French (1993), portfolios are formed in June of each year t based on quintiles of earnings precision (and/or market competition) calculated for the last fiscal year ending in year t 1. For each portfolio, we calculate equal-weighted returns over the subsequent twelve months and rebalance every June. 17 A six-month lag between measuring the sort variable and portfolio returns ensures that any subsequent patterns in returns are the result of persistent risk-related characteristics of expected returns (e.g., Fama and French, 2008). We then estimate the following equation using the time-series of monthly portfolio returns: (R p R f ) = α + β MKTRF + s SMB + h HML + ε, (7) where p {Q1,, Q5}, R p is the monthly portfolio return, R f is the risk-free rate, and (R mkt R f ), SMB, and HML are the three Fama-French factors. The coefficients of interest are the estimated intercepts (α) and the estimated market beta (β). Given that our tests are joint tests of our predictions and the Fama-French asset-pricing model, we attempt to rule out the possibility that our results are explained by the omission of a correlated risk factor. In particular, we estimate additional specifications of equation (7) that augment the standard three-factor model with the momentum factor (UMD) and the Pastor and Stambaugh (2003) liquidity factor (LIQ) Inferences are robust to conducting our analyses at the firm level rather than the portfolio level (i.e., estimating market risk and abnormal returns for each firm and replacing factors with the respective firm characteristics). 17 If a firm delists in a given month during the sample period, we follow Beaver, McNichols, and Price (2007) and compute the return by compounding the monthly return and the delisting return. At the end of the month any delisting proceeds are reinvested in the portfolio. 18 There is an ongoing debate in the literature on the relative merits of using realized returns over implied cost of capital estimates (e.g., Core et al., 2008). Many implied cost of capital measures are a function of predictably biased analyst forecasts (e.g., McInnis, 2010). For example, Dichev and Tang (2009) find that the bias in analyst forecast errors is decreasing in earnings volatility. This suggests the bias in implied cost of capital measures varies mechanically with our variable of interest, and that such measures are inappropriate in our setting (see also, Hughes, Liu, and Liu, 2009). As an additional concern, a chief interest of our study is the relation between earnings precision 23

24 Given that the average precision effect is not conditional on the level of market competition, within each quintile of market competition we predict that β decreases as we move from quintile 1 to quintile 5 of earnings precision. The information asymmetry effect, however, is conditional on the level of market competition. As the market becomes increasingly competitive, we expect to find no evidence of an information asymmetry effect. Accordingly, we predict a positive association between earnings precision and abnormal returns (α) in the lower quintiles of market competition, and no association between earnings precision and abnormal returns in the upper quintiles of market competition. 4. Sample and Descriptive Statistics 4.1 Sample construction We construct our sample using data from Compustat, CRSP, and TAQ. To be included in the sample, a firm must have common stock traded on the NYSE, NASDAQ, or AMEX (CRSP share codes 10 or 11, and CRSP exchanges codes 1, 2, or 3). The firm must also have earnings before extraordinary items, market value, and total assets for years t 9 through t, and number of shareholders at the end of year t. We begin forming portfolios in 1976, when our measure of market competition (the number of shareholders) becomes available on Compustat, and conclude in This timeframe allows for calculation of portfolio returns until June of The first four columns of Table 1 show the number of observations each year for the various measure of earnings precision. The next four columns of Table 1 show the number of observations available for the various measures of information asymmetry, and the final column shows the number of and the cost of capital in firms with low market competition. Because low-competition firms tend to have little to no analyst following, we cannot calculate implied cost of capital estimates for these firms. 24

25 observations available for our measure of market competition. In addition, because some of our analyses require additional control variables (e.g., analysts forecasts and accruals quality), the number of observations reported in Table 1 reflects the maximum number of observations used in subsequent analysis. In order to maximize the sample size for any given test, we only require availability of the variables for that test. For example, when we test our hypotheses about earnings precision and information asymmetry using Precision and λ MRR, we use all available firm-years for which we have estimates of both Precision and λ MRR (i.e., from 1993 to 2007), but when we test the same hypothesis using Precision and Illiq, we extend the sample period to 1976 to 2007 (i.e., we use all available firm-years for which we have estimates of both Precision and Illiq). 4.2 Descriptive statistics Table 2 presents descriptive statistics for our sample and correlations between variables used in our analysis. Panel A presents the distribution of variables used in our analysis. Panel B presents the distribution of variables used in our analysis by quintile of our primary measure of earnings precision, Precision. From Panel B, it is evident that the univariate associations between each of the four measures of earnings precision and each of the six measures of information asymmetry in the subsequent year point to a positive relation between earnings precision in year t and information asymmetry in year t+1. Illiquidity, Spread, λ GH and λ MRR are all monotonically increasing in earnings precision. Across all four measures, the portfolio of firms with the highest earnings precision (i.e., quintile 5) has significant higher levels of information asymmetry than the portfolio of firms with the lowest earnings precision (i.e., quintile 1). Panel C presents the correlation among the variables used in our analysis, with Spearman (Pearson) correlations above (below) the diagonal. Because of our interest in cross-sectional 25