The Information Content of Earnings Announcements: New Insights from Intertemporal and Cross-Sectional Behavior

Transcription

1 The Information Content of Earnings Announcements: New Insights from Intertemporal and Cross-Sectional Behavior William H. Beaver Joan E. Horngren Professor (Emeritus) Graduate School of Business, Stanford University, Stanford, CA, 94305, USA Maureen F. McNichols Marriner S. Eccles Professor Graduate School of Business, Stanford University, Stanford, CA, 94305, USA Zach Z. Wang Assistant Professor University of Illinois at Urbana-Champaign, Champaign, IL, 61821, USA March 14, 2015 We thank Paul Reist and Nora Richardson for their assistance in data collection. We assume responsibility for any remaining errors.

2 The Information Content of Earnings Announcements: New Insights on Intertemporal and Cross-Sectional Behavior Abstract This study examines the information content of quarterly earnings announcements. We first use a nonparametric approach to investigate whether quarterly earnings announcements are informative between 1971 and 2011 and find unequivocal evidence that earnings announcements convey significantly more information relative to non-announcement periods. We also find that the information content increases over time with a dramatic increase from 2001 onward, a period that includes the implementation of Sarbanes Oxley reforms and the worst economic downturn since the Great Depression. We then investigate cross-sectional variation in information content. We find that the information content of earnings announcements is positively associated with profitability, firm size and analyst coverage. Keywords: capital markets; earnings announcements; information content; return volatility

3 The Information Content of Earnings Announcements: New Insights on Intertemporal and Cross-Sectional Behavior 1 I. INTRODUCTION This paper examines three questions about the information content of earnings announcements. First, how robust are the earlier findings that revisions to security prices are significant when earnings are released? Second, has the magnitude of the price response to earnings announcements increased over time? Third, does the magnitude of the price response to earnings announcements vary for firms with potentially different information environments? Specifically, does the magnitude of the price response vary with the profitability of the firm, its size, or the extent of its analyst coverage? Beginning with Ball and Brown (1968) and Beaver (1968), a large literature has examined the information content of earnings. While the early findings in this literature document that earnings announcements have significant information content, more recent literature has explored whether and how the information content of earnings varies over time and across firms. Although the question of whether earnings announcements have information content might be viewed by some as having long been settled, a number of papers question this, including Bamber et al. (2000), Ball and Shivakumar (2008), and Ball (2013). Furthermore, there are many reasons to hypothesize changes to the information content of earnings, including changes in financial reporting standards and regulation over time, changes in firms business models that are potentially less well captured by the current accounting model, and changes to the institutions and incentives for private production and dissemination of information. The latter 1 We refer to the information content of earnings announcements rather than earnings as we study the reaction of prices on earnings announcement days, which will reflect the reaction to all information released on those days. This potentially includes information about the balance sheet and statement of cash flows, as well as management guidance or other disclosures. 1

4 includes dramatic changes in information technology that have led to significant reduction in costs of providing and disseminating information, and an increase in the availability of information throughout the year on a continuous basis. These factors potentially would serve to reduce the incremental information content of earnings announcements. The passing of time allows us to revisit this literature with a longer and potentially more varied sample period, to examine the robustness of the earlier findings and to assess whether and how the information content of earnings announcements has changed. Three measures of the information content of earnings announcements are predominantly employed in the literature, the earnings response coefficient (ERC), for which Ball and Brown (1968) is a precursor, a measure of abnormal volume, and a measure of abnormal return volatility based on the U-statistic introduced by Beaver (1968). This study examines the information content of earnings announcements based on the abnormal return volatility measure, and therefore focuses on the information conveyed to investors at earnings announcements. We develop a U-statistic motivated by Beaver (1968), Patell (1976), and Landsman and Maydew (2002), which we refer to as TCU, for three-day cumulative U-statistic. First, we examine whether the information released at earnings announcement days is greater than the information released at other times during our sample period. Using a distribution-free test, we find unequivocal evidence that earnings announcements convey more information to the markets than the information conveyed in non-earnings announcement periods. Specifically, we find that the mean U-statistic observed at earnings announcement periods is substantially above the 99 th percentile of the distribution of U-statistics drawn from non-earnings announcement periods for each of the 41 years in our 1971 to 2011 sample period. Our statistic measuring relative price revision activity in earnings announcement periods from 2

5 1971 to 2011 is 2.54, compared to a mean of 1.18 in non-earnings announcement periods and 1.67 for the sample in Beaver (1968). To incorporate the skewed nature of the U-statistic and make inferences about the breadth of the effect across reporting firms, we focus on the median as well as the mean. Similar to the mean tests, the median tests indicate that the median firm exhibits substantially greater return variability at earnings dates than in non-report periods. These findings indicate the mean results are not driven by a small subsample of firms with extreme reactions. Second, using a sample that extends ten years beyond the time period of prior research, we examine whether the relative variability of returns at earnings announcements has continued to increase over time. We find the variability of returns at earnings announcements has continued to rise, and that the rate of the increase is substantially greater than in earlier years. In particular, we document a dramatic increase from 2001 onward, a period that includes the implementation of SOX reforms and the worst recession since the Great Depression. Moreover, the relative price revisions at earnings announcements have increased substantially over the past decade, with the mean TCU reaching a peak of 4.28 in 2006 and 4.15 in our final sample year Thus the rate of growth is substantially greater than documented in earlier studies. Furthermore, we document that the increase in information content is not monotonic over time, with increases and decreases in the 2000 s, a pattern that warrants future research. We next examine several cross-sectional determinants of relative return variability at earnings announcements. First, we examine the association between the information content of earnings announcements and profitability. Prior research by Basu (1997) finds that the conservatism inherent in accounting causes more timely recognition of losses than gains, which could result in greater revision of prices at earnings announcements by loss firms. Alternatively, 3

6 the transitory nature of losses, as documented by Hayn (1995), could result in less revision of prices. Further, the discretionary component of losses may be larger (e.g. due to taking big baths) or may be preempted by disclosure occurring prior to the narrow 3-day earnings announcement window (Kasznik and Lev, 1995). We find that the variability of returns at earnings announcement dates of firms reporting profits is greater than that for firms reporting losses, suggesting that the effect of greater persistence of profits offsets the potentially greater timeliness of losses. Second, we examine the association between information content and size, or market capitalization. Early studies on this relation, such as Atiase (1985), documented that smaller firms had more pronounced price reactions to earnings. However, there have been many changes in the financial reporting environment since the time periods examined in his and other early studies and ex ante it is not clear how these changes might affect the relative informativeness of earnings announcements. We find that the average market reaction to earnings was negatively related to size through the late 1980 s, consistent with Atiase (1985). However, this relation reversed in the 1990 s and 2000 s, with more pronounced reactions to earnings for each quartile of market capitalization. Third, we find a positive association between relative return variability at earnings announcements and analyst following. Currently, there is considerable ambiguity on the relation between analyst coverage and information content of earnings announcements, due in part to competing predictions of various theories and models. We conjecture that there are several potential forces associated with analyst following. On one hand, greater analyst following could increase the total market reaction to earnings announcements to the extent that analysts serve as information intermediaries, and are an important part of the process through which information is 4

7 incorporated into price. Also, analysts may have incentives to cover firms where earnings are more informative because, for example, it gives them the opportunity to discuss and interpret price reaction to earnings. In addition, managers may have incentives to produce more informative earnings disclosures where there is greater analyst following. If one or more of these forces is in effect, we would expect a positive association between information content and analyst coverage. On the other hand, competition among analysts could motivate analysts to seek more timely information that would at least partially preempt the earnings announcement. Selection by analysts might favor their covering stocks with weaker information environments, in the manner hypothesized by Barth, Kasznik and McNichols (2001), or providing more informative reports as documented by Chen, DeFond and Park (2002), and DeFond and Hung (2003). This second force could lead to a negative association between information content and analyst coverage. Our evidence suggests that the first set of forces dominates the second set. Furthermore, our evidence based on multivariate tests indicates that the association between relative return variability and size is often insignificant or changes sign when analyst coverage is included in the regression, whereas analyst coverage remains highly significantly positive. In addition to our findings on the information content of earnings announcements, our paper contributes to the methodology of testing for information content. First, we develop a distribution-free test of the hypothesis that the observed U-statistic at earnings announcements is greater than the U-statistic estimated for randomly chosen intervals in the non-report period. This approach allows us to test whether the mean or median, or any other characteristic of the U- statistic, is significantly higher in announcement windows than in randomly chosen nonannouncement windows without making assumptions about the underlying return distributions. Second, we develop a U-statistic (denoted as the TCU-statistic) that allows for serial correlation 5

8 of returns between adjacent trading days in both announcement and non-announcement windows and non-normality of returns. Third, we present extensive descriptive and visual evidence that permits a richer picture of how information content of earnings has varied over time and across firms. The remainder of the paper proceeds as follows. Section II discusses the prior literature and our hypotheses. Section III discusses the research design. Section IV presents sample properties. Section V discusses the results, and Section VI provides concluding remarks. II. PRIOR LITERATURE AND HYPOTHESIS DEVELOPMENT By showing both trading volume and return volatility increasing at the time of annual earnings announcements, Beaver (1968) provides evidence that earnings reports provide information to investors, as measured by the U-statistic and abnormal volume. 2 Beaver (1968) defines the U-statistic to be the announcement week squared residual returns divided by the mean squared residual returns in the non-announcement period. Patell (1976) derived the distribution of the U-statistic under the assumption that security returns are normally distributed and serially independent. With these assumptions, he showed the U-statistic has an F-distribution with an expected value of slightly greater than 1 under the null hypothesis that earnings announcements do not convey more information than in non-announcement periods. Furthermore, the F distribution is in general skewed to the right so the median U-statistic is less than one under the null. 2 The U-statistic and the abnormal volume measures are based upon the notion that a necessary condition for a signal to have information content is that beliefs are altered and that the change in beliefs must be sufficient to alter actions across signals. 6

9 In moving from weekly returns to daily returns, it has been common practice to measure the mean U-statistic over a three-day interval surrounding the earnings announcement, due to uncertainty over the exact timing of the earnings release (e.g., Landsman and Maydew, 2002). We refer to the Landsman and Maydew calculation of the U-statistic as LMU. As we describe in greater detail later, we develop a measure that compounds returns over the three-day interval surrounding the earnings announcement to calculate a U-statistic (hereafter, TCU). Because the distribution of TCU is not expected to be symmetric, we examine the median of the distribution as well as the mean. By focusing on the median as well as the mean, we can examine the extent to which relative return variability differs during earnings report periods across the cross-section of firms. While the mean is expected to be one under the null hypothesis, the median is expected to be less than one. Our hypotheses concern whether earnings announcements have information content, and whether the information content varies with time, profitability, size, and analyst coverage. The numerator of the U-statistic captures the information released during earnings announcement windows. The denominator of the U-statistic captures the information released during non-report windows. Any factor that increases the information released during earnings announcement windows or decreases the information released during non-report windows will increase the information content of earnings announcements. Hence, to examine the explanatory factors of the U-statistic, we explore factors that might affect how much information is released within and outside earnings announcement windows. Our first hypothesis is a replication of the tests conducted in prior studies for a more comprehensive sample. Specifically, our sample includes a longer time period, , and allows us to examine the past decade, which includes the Great Recession. We interpret the information content of earnings announcements to mean the 7

10 ability to change the market s belief in a systematic manner such that equilibrium price revisions are greater at the time of earnings announcements than at other times in the year (i.e., the nonreport period). The non-report period is not a no information benchmark because there is other information occurring during the non-report period. It contains a mixture of days with information and days with no new information. While some may view the information content of earnings announcements as having long been established, a number of recent papers question this, including Bamber et al. (2000), Ball and Shivakumar (2008), and Ball (2013). For instance, Bamber et al. (2000) conclude that Beaver s original findings are not generalizable to larger firms and are driven by a small set of sample firm-years. For Hypothesis 1, we include all firms with earnings reports dates from Compustat. Our hypothesis is stated in null form and our alternative is two-sided: Hypothesis 1: Earnings announcements do not convey more information to investors than is conveyed on randomly chosen days in non-announcement periods. A number of studies examine whether the information content of earnings announcements has increased or decreased over the 1970 s through 1990 s and explore potential explanations for this trend. Recent contributions to the literature include those of Lo and Lys (2001); Landsman and Maydew (2002); Francis, Schipper and Vincent (2002a); and Collins, Li and Xie (2009). Using a sample of firms from 1972 to 2000, Lo and Lys (2001) find that the information content of earnings announcements does not increase over time but the value relevance of earnings decreases over time. They attribute the difference in the findings for information content and value relevance to unrecognized disclosure at the same time as earnings. Using a sample of firms from 1972 to 1998, Landsman and Maydew (2002) instead find that the information content of earnings announcements increases over time after controlling for factors 8

11 such as firm size, intangible intensity and presence of a loss. Francis et al. (2002a) limit their sample to firms that are present in Compustat throughout the entire period from 1980 to 1999 and find that an increasing trend in market reactions exists in spite of declining absolute unexpected earnings and ERC s. Using earnings press releases, they find that over-time increases in the amount of concurrent disclosures and over-time increases in investor reactions to concurrent disclosures explain the increase in market reactions. Using a sample of firms drawn from IBES from 1985 to 2000, Collins et al. (2009) find an increasing trend in the informativeness of earnings announcements, which they attribute to an increasing reaction to Street earnings. Because the sample periods in these studies ended by 2000, they do not provide evidence on capital market response to earnings in more recent years. The post-2000 period is a fascinating period in our economic history that includes the Internet boom and bust, Reg FD, SOX reforms, and several quarters of significant economic crisis as well as several quarters of recovery on the information content of earnings. The significant uncertainty created during the Internet bust and the financial crisis could potentially put a premium on more timely information. For example, industry or macroeconomic news could be a dominating factor during these time periods and could be preemptive in nature. If the preemptive nature of other information has increased, the information content of earnings announcements may have declined. On the other hand, during unfavorable economic times there may be a premium on relatively higher quality signals such as the earnings announcement, which are subject to audit, and audit scrutiny likely increased during this time. In addition, the information disclosed at earnings announcements may have increased. 9

12 While the question of the time trend of the information content of earnings announcements is an intuitively appealing initial question, it can be potentially limiting in at least two respects. First, the variable time is the ultimate sponge because it is a proxy for some underlying factor that is varying through time. A complete understanding of the factors that drive information content would drive the time variable to insignificance. Second, it tends to relegate the other variables into a secondary status as control variables. For example, as discussed, Landsman and Maydew (2002) describe an expanded model with additional variables but do not report the results beyond indicating that the time variable is still significant. Collins et al. (2009) include an essentially different set of control variables based on prior research on volume determinants. Francis et al. (2002a) contain neither set of the control variables used in these two studies to focus instead on other information disclosed in earnings press releases. By contrast, our study explicitly considers hypotheses related to the additional variables of size, analyst following and profitability, which have cross-sectional as well as time-series variation. This approach elevates explaining the cross-sectional variation to equal status in terms of understanding the determinants of the information content of earnings announcements. In this sense, it complements earlier approaches addressing why the information content of earnings might change in different periods. The preceding discussion suggests two questions of interest. First, is the information content of earnings announcements associated with time? This leads to H2a, stated in null form: Hypothesis 2a: The information content of earnings announcements is not associated with time. Our alternative hypothesis is two-sided. Second, has the association with time changed in the post-2000 period? We select 2000 as a cut-off year because the sample period of the above-mentioned studies concludes in 2000, 10

13 and because of the significant events in the post-2000 period. This cutoff year also results in a similar number of firm-year observations in each subsample. This leads to H2b, stated in null form: Hypothesis 2b: The rate of change in the information content of earnings announcements post is equal to the rate of change in the information content in the pre-2000 period. Our alternative hypothesis is two-sided. Hayn (1995), Basu (1997) and Givoly and Hayn (2000) suggest that accounting standards, such as the change in the impairment standards introduced by SFAS 144, requires more timely recognition of losses over time. If investors treat losses as timely information about poor operational conditions, we should observe significant market reactions to losses. However, these studies also document that losses are less persistent than gains. For example, losses might reflect managers decision to take a big bath (Healy 1985). If investors do not extrapolate losses in firm valuation, the lower degree of persistence could lower the information content of earnings announcements. For example, Collins et al. (1999), Barth, et al. (1998) and Francis et al. (2002a) find that the earnings coefficients in regressions of market value or stock returns on earnings are significantly lower for loss firm-years than for profit firm-years. The profitability of a firm may also affect firms voluntary disclosure decisions, as Miller (2002) documents. Firms experiencing increasing profitability are more likely to disclose more information voluntarily. The effect on the information content at earnings announcements depends on whether increased voluntary disclosure occurs with the earnings announcement or prior to it. The competing effects of timeliness and persistence of losses and the effect of profitability on voluntary disclosure motivate our third null hypothesis, for which our alternative is two-tailed. 11

14 Hypothesis 3: The information content of earnings announcements does not differ between firms reporting profits and firms reporting losses. Prior literature hypothesizes a relation between the information content of earnings and firm size through the effects of variables correlated with size on firms information environments. Atiase (1985) hypothesizes that alternative sources of pre-disclosure information increase as a function of firm size. If most of the information contained in earnings announcements is preempted by other information sources, we should observe smaller market reactions to earnings of large firms. Using a sample of firms with market capitalization either greater than $400 million or smaller than $20 million from 1971 to 1972, he finds that the information content of earnings announcements is negatively associated with size. Shores (1990) examines the cross-sectional determinants of the information content of earnings announcements of OTC firms from 1983 to For this specific group of firms, she finds a negative association between the information content of earnings announcements and proxies for the information environment such as market capitalization and analyst coverage. An alternative factor is that earnings for large firms may be more informative due to their business models, to higher quality audits or greater litigation risk. Large firms are also more likely to provide investors with more concurrent disclosure through conference calls, such as management guidance. Furthermore, the post-earnings announcement literature (Bernard and Thomas, 1989) finds that investors react in a more timely manner to earnings announcements of large firms than small firms. Because we view the relative magnitude of these opposing forces as unknown, the test of our fourth hypothesis is two-tailed. Our fourth null hypothesis is: Hypothesis 4: The information content of earnings announcements is not associated with size. 12

15 Analyst following has the potential to influence the information content of earnings announcements in competing ways. To the extent that analysts serve as information intermediaries and are part of the process through which information is incorporated into price prior to the release of earnings, an increase in analyst following could increase the total market reaction to earnings announcements. For example, analysts may ask relevant questions in conference calls and make timely forecast and recommendation revisions in earnings announcement windows. The association between information content and analyst coverage may also be due to the factors that analysts consider in their coverage decisions. Analysts may have incentives to cover firms that have more informative earnings announcements, for example, because it provides them an opportunity to discuss and interpret price reactions to earnings. These forces would lead to a positive association between information content and analyst following. Using a sample of firms from 1986 to 1995, Francis et al. (2002b) find that the information content of earnings announcements is positively associated with the information content of analyst reports and conclude that the informativeness of earnings announcements is not eroded by competing information in the form of analyst reports. On the other hand, competition among analysts could motivate analysts to seek and release more timely information that would at least partially preempt the earnings announcement. For example, Chen et al. (2002) and DeFond and Hung (2003) find that analysts respond to market incentives and release more information about firms when earnings information is less informative. Relatedly, Barth et al. (2001) find analysts choose to cover firms for which the financial reporting model works less well so there is greater uncertainty about valuation. These forces could lead to a negative association between information content and analyst following. Given the strength of the arguments for increasing and decreasing relations, our tests are two- 13

16 tailed, and our alternative hypothesis is that the information content of earnings announcements differs for firms with levels of analyst coverage. Our null hypothesis is: Hypothesis 5: The information content of earnings announcements is not associated with analyst following. III. EMPIRICAL METHODOLOGY Measures of the Information Content of Earnings Announcements Our measure of the information content of earnings announcements, the Three-day Cumulative U-Statistic ( TCU ) captures the magnitude of the average squared residual return during the announcement period, hereafter the testing period TP, to averaged squared residual returns during the non-announcement period, hereafter the estimation period EP. 3 For each quarterly earnings announcement (day 0), we require that return data are available for each of the three trading days in the test period TP (days -1, 0, +1). We choose three-day announcement windows following prior research. 4 An estimation period, EP, is defined as the period from 130 to 10 days prior to the earnings announcements and days 10 to 130 days after the announcement. In a procedure to be explained shortly, we also make the same requirements for each as if non-report earnings announcement randomly selected from the non-report period. Following the Landsman and Maydew (2002) convention, we aggregate quarterly earnings announcements into calendar years based on the dates of earnings release. We 3 The difference between our measure TCU and the Landsman and Maydew measure LMU is that LMU implicitly assumes that daily returns are not correlated in both the estimation period and the testing period, whereas TCU does not make such assumptions. Our untabulated analysis indicates that the first order serial correlation of daily returns is significantly negative in both the estimation and testing periods. 4 We focus on three-day rather than one-day returns for two reasons. First, prior studies use a three-day period and we wish to compare our results with theirs. Second, an examination of one-day returns would require the earnings announcements to be time-stamp in order to identify announcements that were made after the close of trading. The time stamp data is not available in computer-readable form until

17 then calculate the means and medians of TCU for each calendar year. Refer to Appendix 1 for further details of TCU. Development of the Nonparametric Distribution As discussed earlier, the distribution of TCU is not expected to be symmetric. If the underlying return variables are independently, normally distributed, the numerator and denominator are each expected to be distributed as a chi-square distribution, and the ratio would be distributed as an F distribution, which is skewed to the right. Hence, the median is expected to be below 1. Of course, daily returns (or residual returns) may be neither normally nor independently distributed. If so, testing the significance of these differences would require the derivation of the underlying sampling distribution in a distribution-free manner that does not assume normality or serial independence of returns. To generate a sampling distribution under the null hypothesis, for each quarterly earnings announcement, we randomly select a day during its non-report or estimation period. That day is treated as if it were an announcement day. We then calculate TCU in the same manner as the actual earnings announcement, and obtain the mean and median of TCU for that sample of randomly-chosen announcement dates. Repeating the procedure 1,000 times produces a sampling distribution, which is then compared with the values obtained for the actual earnings announcements. 5 The procedure is described in greater detail in Appendix 2. The nonparametric procedure differs significantly from that of Beaver (1968), where each mean U-statistic reported for the non-report period is chronologically aligned (and thus subject to cross-sectional dependence) and is based on a smaller number of observations. The sampling of 5 Drawing the randomly chosen announcement date from the estimation period for each earnings announcements ensures that the randomly chosen date is from approximately the same calendar period as the actual earnings announcement date. Given that our sample is drawn over 41 calendar years, this procedure seems preferable to allowing the randomly chosen announcement date to occur at any time in the firm s history. 15

18 the non-report periods used here repairs both of these potential deficiencies. The randomsampling procedure mitigates potential cross-sectional dependence of observations aligned chronologically (even after taking out the market-wide factor), and we obtain a sampling distribution of the mean U-statistic for the non-report periods based on 1,000 trials. With the advent of technology that permits data-intensive techniques, we are able to conduct the 7*10 8 (700,000 earnings time 1,000 iterations) simulations that underlie the non-report distribution. Definitions of Other Variables We examine the association of TCU with several variables, including time, loss, size and analyst following. T is a trend variable that takes values from 0 to 27 for calendar years from 1984 to POST2000 is an indicator variable that is equal to 1 for firm-quarters in calendar years from 2001 to 2011 and 0 otherwise. TPOST2000 is T times the POST2000 indicator, to capture the relation between TCU and time in the post-2000 period. LOSS is an indicator variable that is equal to one when PTEBS is negative and zero otherwise, where PTEBS is the sum of Net Income before Extraordinary Items, Tax Expenses and Minority Interest before Special Items. CV is the market capitalization value of the firm s common stock at fiscal quarter end from Compustat. Following Atiase (1985), we take the natural log of CV, LCV, as a proxy for size. The variable for analyst following, NUMESTQ, or the number of analysts making forecast of the upcoming quarterly earnings at fiscal quarter end, is drawn from the IBES analysts forecast database. For firm-quarters that are not in the IBES database, we assume zero analyst coverage. NUMn is an indicator variable that equals one if n analysts cover a specific firm quarter and zero otherwise, and measures analyst following. We control for participation in financial services, reporting lags, fiscal-year end, and unexpected earnings, variables we expect to be associated with the information content of earnings releases. FIN (NONFIN) is an indicator 16

19 variable that is equal to 1(0) when the four-digit SIC code is between 6000 and 6999 and 0 (1) otherwise. LAG is the number of days after the end of the fiscal quarter that earnings are announced. NONDEC31 is an indicator variable that is equal to 1 for firm-quarters with Non- Dec31 fiscal year end and 0 otherwise. Following Collins et al. (2009), ABSFE_STREET is the absolute value of the difference between IBES realized earnings per share and IBES median consensus earnings per share at fiscal quarter end scaled by price per share at fiscal quarter end. Initially, the hypotheses are treated in an unconditional, bivariate manner (i.e., one variable at a time). In the subsequent section, we also conduct multivariate regression analysis to determine if the results of these hypotheses tests are preserved conditional on all of the variables included in the regression. IV. SAMPLE PROPERTIES We start with the universe of firms listed on the NYSE, AMEX and NASDAQ markets for which quarterly reporting dates are available from Compustat and return data are available from CRSP. For each quarterly earnings announcement included, we require that return data are available for each of the three trading days in the event period (days -1,0,+1),TP, and the number of trading days with nonzero return data in estimation period EP is at least 40. We also make the same requirements for each simulated non-report earnings announcement. For our first hypothesis, our analysis covers the quarterly 41-year period from 1971 through 2011 and the final sample size is 700,000 firm-quarters. For the additional hypotheses, we impose additional restrictions on the sample to ensure each firm-quarter observation has data for the required variables. Because Compustat does not provide sufficient data to calculate the profitability variable, PTEBS, before 1976, our analysis 17

20 on LOSS starts from 1976, and the sample size is 554,796. Because most firms do not have IBES coverage before 1984, our analysis of analyst following begins in 1984, leaving us with a sample size of 586, For our analysis on unsigned Street earnings, IBES coverage is required and therefore also begins in This analysis has a sample size of 335,092. For our multivariate regression models, we require included firm-quarters to have all required variables except for unexpected earnings. We winsorize all variables at the 1st and 99th percentiles. The number of observations included in each analysis is reported in the respective tables and figures. V. RESULTS Are Earnings Announcements Informative and Has Information Content Changed over Time? Figure 1 serves two purposes. It compares the mean (Panel A) and median (Panel B) of TCU over each of the years from 1971 through In each quarter, the figure also compares these values with the respective sampling distribution constructed from the non-announcment period. For each calendar year, we plot the value of TCU_mean (TCU_median) for our event sample against the values of 1000 TCU_means (TCU_medians) from the null distribution. 7 The mean and median values are above the 99th percentile in each of the 41 calendar years. The evidence provides striking evidence that earnings announcements convey significantly more information than in non-announcement periods. Given the prior research that has also rejected the null hypothesis, the finding with respect to the mean U-statistic is not completely surprising. However, prior research has also 6 Our convention of equating no coverage by IBES with no analysts coverage would contain substantial error if we included years prior to 1984, given the much less comprehensive sample available for earlier years. For firms that are not covered by IBES, we assume zero analyst coverage. 7 The pattern of the mean and medians of LMU is similar, and in the interests of brevity are untabulated. 18

21 raised the issue of whether the skewed nature of the underlying distribution affects the robustness of results based on means, which may be dominated by a few extreme observations. The median results provide new evidence that the finding is robust with respect to the use of medians, which are less influenced by extreme observations. In fact, the median value is not only above the 99 th percentile in each of the years, it is higher than every one of the 1,000 non-announcement values in every year from 1971 to Table 1 Panels A and B report the mean and median values of TCU by year for each event period, along with the distribution of the respective values for the non-announcement period. For instance, in 2011, the mean value of TCU in Panel A is 4.15 and the 99th percentile of the non-report distribution is Similarly, the median value of TCU for 2011 in Panel B is 0.98 and the 99th percentile is Although the values of the event period statistics in 2011 are among the highest across the sample period, the results are similar for the other years where the earnings announcement period values lie consistently above the 99th percentile of the non-report distribution. Over the 41 years, the mean value of TCU in event windows is 2.54, which is 2.2 times 1.18, the mean value of TCU in simulated non-report windows. To test our Hypothesis 1 more formally, we also conduct a rank test, and the details are reported in Table 2. Using Bonferroni s Correction Method, we reject the null of no information content at the 1% significance level for every year. As expected, the distribution of TCU is not symmetric. The mean TCU in the non-report period is 1.18 over the years and above one in every year, which is greater than would be 8 The retesting of the null hypothesis is also nontrivial in the sense that another 11 years has been added to the analysis, reflecting turbulent economic times and several regulatory changes. It is possible, although unlikely, that the information content may have declined sufficiently to be unable to reject the null. The retesting also shows that the TCU transformation of the U-statistic produces essentially the same results as prior research using common years. 19

22 expected under the assumptions of independence and normality. A comparison of the means of the TCU median distribution in Panel B to those of the TCU mean distribution in Panel A shows the skewed nature of the underlying distribution: in each year the mean of the TCU median distribution is considerably lower than that of the TCU mean distribution in Panel A. These findings confirm the value of developing a nonparametric test, and indicate that the inferences based on the mean TCU are not due to a subset of extreme observations. With respect to the time trend in TCU, an increase is evident and is in fact highest in the last 10 years. While prior research has imposed a linear time trend on the U-statistic, it is clear from Figure 1 that the time trend is decidedly nonlinear. In fact, the slope dramatically increases in the last ten years. By comparison, the time trend of the prior time period upon which prior research is based appears to be small by comparison. In the multivariate analysis, we will test directly whether the increasing time trend effect is stronger after 2000 than before Overall, the null hypothesis of no information content is rejected for each and all of the time periods. Moreover, there appears to be a positive time trend that is more pronounced in the later years, which includes the most significant economic recession since the 1930s. Firm-Quarter Specific Variables In this section, we examine the ability of profitability, size, number of analysts and unexpected earnings to explain the differences in TCU. We present a series of figures that illustrate the differences over time. This form of presentation permits us to examine cross sectional differences, holding time constant, and time-series differences holding the cross sectional variable constant. The subsequent section reports the findings of multivariate analysis that combines all of the variables examined. Profits versus Loss 20

23 Figure 2 presents a description of the mean and median values of TCU for both profitable and loss firms. We assign firm-quarters with positive PTEBS to the profitable group and firmquarters with negative PTEBS to the loss group. The figures indicate that firm-quarters with losses are associated with lower information content than firm-quarters with profits. The mean and median values for the profitable group are higher than those for the loss group in each of the 41 years. Our results suggest that the low persistence of losses dominates the timely information role of losses. Furthermore, the difference in the market reaction to earnings announcements of profit vs. loss firms is more pronounced in the later years. Figure 3 presents a figure with the percentage of loss firm-quarters over time. The frequency of loss increases from 10% in 1976 to 40% in 2000, which is consistent with the findings in Hayn (1995) and Givoly and Hayn (2000). After 2000, the frequency of loss decreases gradually to 25% in 2006 and then increases to about 40% in the Great Recession. Consistent with Beaver et al. (2012), these results suggest that the frequency of losses is the joint effect of accounting standards and underlying economic conditions. Size Figure 4 presents a description of the mean and median values of TCU in each LCV, or log of capitalized value, percentile. We rank firm-quarters by the log of market capitalization in the pooled sample and assign LCV percentiles accordingly. We assign firm-quarters with the lowest LCV to percentile 1 and firm-quarters with the highest LCV to percentile 100. The mean values are essentially flat until the 50 th percentile, but there is a positive association between TCU and LCV percentiles from the 50th percentile to the 100th percentile in the mean graph. The mean TCU for the 50 th LCV percentile is 1.92, and the mean TCU for the 100 th LCV percentile is The median graph shows that there is a monotonic positive association 21

24 between TCU and LCV percentiles. Overall, we see a positive association between size and the information content of earnings announcements. We defer an explicit test for significance of size until the discussion of the multivariate analysis. Figure 5 documents the relation between size and information content over time. It presents a time-series plot of the mean and median values of TCU in each of the LCV quartiles by calendar year. For each calendar year, we assign firm-quarters with the lowest LCV to quartile 1 and firm-quarters with the highest LCV to quartile 4. Before 1988, the U-statistic for the smallest quartile observations is higher than for the largest, which is consistent with the preemption effect documented in Atiase (1985). From 1988 through 1994, the largest and smallest size quartiles have approximately the same U-statistic. However, from 1995 to 2011, the TCU is larger for the largest firms relative to the smallest. This is consistent with the view that LCV is a proxy for other characteristics of the information environment such as higher analyst coverage, and these characteristics dominate the preemption effect documented in Atiase (1985). In addition, the increase in information content after 2000 is more pronounced for firms in the top two LCV quartiles than for firms in the bottom two LCV quartiles. One characteristic of size as a variable is its sponge-like ability to soak up the effect of other correlated variables. In that spirit, we examine analyst coverage, which is more directly related to information environment. Analyst Coverage Figure 6 presents a description of the mean and median values of TCU in each of the analyst coverage categories. The mean graph shows that there is a positive unconditional association between TCU and analyst coverage. The mean TCU for the zero analyst following category is 1.92 and the mean TCU for firm-quarters with more than 26 analysts is The 22

25 median graph shows a similar pattern. Compared to size, analyst coverage can be argued to be a more refined measure of the information environment. Figure 7 presents the mean TCU for the quartiles of the distribution of analyst coverage and for firms with no coverage. Consistent with Atiase s preemption hypothesis, mean TCU is highest for firms with less coverage in the earliest years. Consistent with our findings for size, this relation reverses to a positive association between TCU and analyst coverage. For this later period, our results suggest that the information intermediary role of analysts, the analysts tendency to cover more earnings-sensitive firms, and/or managers' incentives to provide more information when analyst coverage is greater dominates the preemption effect of greater analyst coverage. We cannot distinguish between these explanations but consider the positive association between analyst following and information content of earnings announcements an intriguing phenomenon for future research. Multivariate Regression Results While Figures 1 through 7 are helpful in analyzing the bivariate relation between the individual explanatory variables and TCU, ultimately we are also interested in the role these variables play conditional on the remaining explanatory variables. The purpose of the multivariate analysis is to assess the incremental association of the variables jointly estimated. As a prelude, Table 3 reports the descriptive statistics, and Pearson and Spearman correlations for the variables in the regressions. With respect to the descriptive statistics in Panel A, the mean TCU of 2.34 is well above 1.0 and above the mean based on the randomly selected event windows. Note that these means are slightly different than those reported earlier, because they are based on a shorter time period (1984 to 2011). The median TCU of 0.59 is also well above the median based on the randomly 23

26 selected event windows. The percentage of loss firm-quarters is 28 percent. Notably, 44% of the firm-quarter observations are from 2001 to 2011, which suggests that the number of observations before 2000 is comparable to the number of observations after 2000 in our sample. With respect to the correlations in Panel B, TCU is positively associated with T (Pearson 0.097, Spearman 0.093), POST2000 (Pearson 0.099, Spearman 0.090), TPOST2000 (Pearson 0.107, Spearman 0.103), LCV (Pearson 0.078, Spearman 0.119), NUMESTQ (Pearson 0.108, Spearman 0.138), NONDEC31 (Pearson 0.014, Spearman 0.010) but negatively associated with LOSS (Pearson , Spearman ), FIN (Pearson , Spearman ) and LAG (Pearson , Spearman ). ABSFE_STREET is not highly correlated with TCU (Pearson , Spearman ). These results are in general consistent with what we observe in the figures. LCV is positively associated with NUMESTQ (Pearson 0.640, Spearman 0.663). The high correlation between LCV and NUMESTQ is consistent with size in part proxying for the number of analysts. Tables 4 through 6 report the main results from the multivariate regressions. In Table 4, analyst following is a discrete variable NUMESTQ; in Table 5, we include a series of indicator variables to allow for nonlinearity between analyst following and TCU. In Table 6, unsigned unexpected earnings are included in the regressions. With respect to Table 4, the relationships observed in the bivariate analyses reported earlier are largely preserved with the exception of the size effect when both LCV and NUMESTQ are included. With respect to the control variables, for each of models 1-7, the coefficients on FIN and LAG are significantly negative and the coefficient on NONDEC31 is significantly positive, as expected. For the specification in Model 1, we see a positive association between T and TCU, which is consistent with the general trends in the figures (coefficient 0.069, 24