Evidence of a recent increase in the usefulness of quarterly earnings announcements

Transcription

1 Evidence of a recent increase in the usefulness of quarterly earnings announcements Michael J. Smith 1 12 September Associate Professor, Boston University, 595 Commonwealth Avenue, Boston, MA Phone: (617) msmith22@bu.edu. I acknowledge the financial support of the Boston University School of Management and thank the workshop participants at Boston University for their comments.

2 Abstract I document an increase in the usefulness of quarterly earnings announcements subsequent to the implementation of Regulation FD, Sarbanes Oxley and the SEC mandated acceleration of filing deadlines. I measure the usefulness of earnings announcements as the percentage of total annual excess returns that occurs on or around quarterly earnings announcements. I interpret this annual measure as the share of total uncertainty about firm value resolved by earnings announcements. In my main sample, approximately 18.4% of uncertainty is resolved by quarterly earnings announcements from compared to 14.5% from , an increase of 27%. I include firm-specific controls for variations in an individual firms underlying characteristics, information environment, voluntary disclosure practices, and a proxy for the amount of other information in earnings announcements. I create a quarterly variation of my annual measure, allowing me to track changes in usefulness on a month-by-month basis. Informal analysis suggests that the increase in usefulness begins in June 2002, the first month for which the filing date of the 10-Ks and 10-Qs associated with the earnings announcements is after the implementation of, and therefore subject to, the Sarbanes Oxley Act.

3 1 Introduction There have been three important changes in the financial reporting environment in recent years. In October 2000, the SEC implemented Regulation Fair Disclosure (Reg FD), requiring firms to disclose information to all investors simultaneously, rather than selectively. In July 2002, Congress enacted the Public Company Accounting Reform and Investor Protection Act of 2002 (Sarbanes-Oxley, or SOX). The legislation has many provisions, the most important for the purposes of this study that CEOs and CFOs must certify financial statements prior to their release. Finally, the SEC adopted an accelerated deadline for filing 10-K and 10-Q reports, beginning in the 2002 fiscal year with a three-year phasein period. All three changes could potentially change the amount of new information (usefulness) in earnings announcements. Reg FD might induce firms to defer until the earnings announcement information previously disclosed selectively prior to the earnings announcement, leading to an increase in new information on announcement dates. It is also possible that Reg FD might induce firms to make public voluntary disclosure of information previously disclosed selectively. This would lead to less new information in earnings announcements, because all investors would already know the relevant information instead of only a privileged subset of investors. SOX, by mandating stronger internal controls and making senior management more accountable for the accuracy of financial statements, could also produce more accurate and more reliable financial statements. Finally, the accelerated filing deadline could also lead to more timely earnings announcements. In this study, I examine the usefulness of earnings announcements from the period to determine if there is any change coincident with the recent changes in the financial reporting environment. I measure the usefulness of earnings announcements as the percentage of total annual excess returns occurring on or around earnings announcement dates. I motivate this measure by showing analytically that the expected sum of the variance of price changes during the life of an investment depends only 1

4 on the variance of the underlying cash flows and is, in particular, independent of the quality of the interim information signals. One can think of the variance of the cash flows as creating a fixed pool of uncertainty about firm value. As information arrives, uncertainty is resolved, leading to price changes. The relative qualities of the information signals determine the pattern of uncertainty resolution, in other words what proportion of the total price variance will occur at the realizations of the respective signals. An increase in the noise of the first information signal, for example, defers the resolution of uncertainty, thereby increasing the price reactions to the realizations of the subsequent signals, but has no effect on the total variance of price changes. Holthausen and Verrecchia (1988) derive this result, but do not show that the sum of the variances of the price reactions is fixed with respect to the quality of the signals. The last result is important because it provides the justification for deflating the earnings announcement excess returns by total annual excess returns. To implement the measure, I divide each firm-year into 36 intervals of (approximately) 7 trading days. Four of the 36 intervals straddle earnings announcement dates. For each interval, I compute the squared cumulative excess return. I then sum the squared cumulative excess returns both for the four earnings announcement intervals and for all 36 intervals. The ratio of these two forms the main empirical measure in the paper, AvgExcess. If earnings announcement intervals provided no more useful information than non-earnings intervals, then the AvgExcess measure would be 4 36 =11.1%. The AvgExcess measure is qualitatively similar to returns-based measures in which researchers scale earnings announcement excess returns by the variance of excess returns, but has an intuitive appeal and theoretical basis lacking in the others. The cross-section annual mean of AvgExcess fluctuates between 14% and 18% in the period, and increases starting in 2002 to a peak of 22% in The annual mean of AvgExcess always exceeds 11.1%, implying that investors revise their expectations more during announcement periods than non-announcement periods. Another interpretation of the time series of the AvgExcess measure 2

5 is that uncertainty resolution (information flow) was about 44% higher during announcement periods than during non-announcement periods from , but more than 103% higher from In the first empirical test, I regress AvgExcess on a time variable and an indicator variable for fiscal years 2002 and later. Consistent with other research, I find an increasing trend in usefulness. The coefficient on the indicator variable is also significant, representing a 20% increase in usefulness. Next, I include controls for firm-specific characteristics such as industry, whether or not it is a loss firm, the variance of operating cash flows, size, analyst coverage and book-to-market ratio. I also include controls for voluntary disclosure using the First Call company guidance database, and for the timeliness of the earnings announcement. Unlike Francis, Schipper and Vincent (2002b), who find that increases in qualitative information (such as guidance on earnings) included in earnings announcements explain the positive trend in the usefulness of earnings announcements between 1980 and 1999, I continue to obtain a significant coefficient on the indicator variable after controlling for the incidence of management forecasts on earnings announcement dates, my proxy for the amount of qualitative information in earnings announcements. The size of the coefficient implies an increase in usefulness of 20% over the period. I also find that increased timeliness of earnings announcements does not explain the increase in usefulness. The coefficient on the days lag between the end of the quarter and the earnings announcement is negative, as expected. I find, however, that thought the lag between the earnings announcement and the filing date has decreased sharply in recent years, the lag between the end of the quarter and the earnings announcements has actually increased slightly. To guarantee that the results are not a function of changes in the composition of the sample over time, I estimate the same regressions on a smaller data set including only firms with data for all 30 years. The results are qualitatively similar with an increase in usefulness of approximately 25% in the period. The post-2001 indicator variable and time variable are strongly significant. The results suggest a discrete shift in the usefulness of earnings announcements in the

6 period. I cannot use the AvgExcess measure to identify more precisely the most likely date because it is an annual measure aligned by fiscal year. To address this concern, I create a variation of the measure, QExcess, which is the percentage of total excess returns occurring at a specific quarterly earnings announcement. The four firm-specific QExcess measures for a given fiscal year add up to AvgExcess. I align the QExcess measures by calendar month. Then, for each month in the data set starting with the second and ending with the penultimate, I regress the pooled, time-series QExcess on the explanatory variables and an indicator variable equal to 1 if the period is after the month in question. In all, I estimate the equation 358 times. The purpose of this procedure is to determine which monthly regression has the highest explanatory power. I argue informally that this procedure identifies the most likely month of the discrete change in usefulness. The results indicate that the June 2002 regression has the highest explanatory power. The coefficient on the indicator variable also attains its highest value and significance for this regression. I argue that this is at least circumstantial evidence that the recent increase in usefulness is associated with the implementation of SOX rather than of Reg FD or of the accelerated filing deadline. SOX pertains to 10-K and 10-Q filings, not earnings announcements. The average lag between earnings announcement and filing dates in 2002 is 44. June 2002, therefore, is the first month in which earnings announcements correspond to filings that will take place after July , the effective date for SOX. The study contributes to two different streams of literature. First, there is an ongoing debate about the appropriateness of the financial reporting model and the usefulness of financial statements and earnings for investment decisions. While researchers studying the association between capital market variables and financial variables (for example, Francis and Schipper [1999] and Lev and Zarowin [1999]) typically find that the association has declined or stayed steady over time, researchers studying returnsbased measures of the usefulness of earnings announcements (for example, Landsman and Maydew [2002], Francis, Schipper, and Vincent [2002], and Kross and Kim [2000]), typically find that the short- 4

7 window excess returns on announcement dates have increased. I contribute to this stream of research both by introducing a different returns-based measure of usefulness, by examining more recent data, and by showing that there has apparently been a substantial increase in usefulness in recent years. Second, the paper contributes to the literature assessing the effects of the Sarbanes Oxley Act. Though the business press has focused on the implementation costs of Sarbanes Oxley (see Solomon and Bryan-Low [2004]), especially Section 404, the academic literature has also attempted to identify the economic benefits, especially in terms of cost of capital. The evidence is mixed. While Ashbaugh- Skaife, Collins, Kinney and Lafond (2006) find that stronger internal controls are associated with a lower cost of capital, Ogneva, Raghunandan and Subramanyam (2006) find no such association. While my study does not directly address the benefits of the recent changes in the reporting environment, either to investors or to firms, it does suggest that SOX has altered the flow and timing of information relevant to investors. The empirical results do not necessarily imply that financial disclosure on earnings announcement dates has improved. It is also possible that information flow during non-announcement periods has decreased. Regulation FD would be a natural explanation for this phenomenon. My study, however, fails to find a link between Reg FD and the recent increase, consistent with other studies failing to find statistically significant implications of Reg FD. In Section 2, I discuss the model. I define the variables and discuss the hypothesis in Section 3. I perform the main tests of the hypotheses in Section 4, and the secondary tests in Section 5. Section 6 concludes the paper. 2 Information model underlying empirical measure I motivate the empirical approach with the following disclosure model. The model lasts for one period, which is subdivided into three subperiods. At the initial date 0, the firm acquires an asset whose terminal cash flows is Ṽ = v +ũ, whereũ v N(0,s v ). There are two signals. Signal X arrives at date 5

8 1, the conclusion of the first sub-period. I assume that X = v +ũ x, where ũ x N(0,xs v ). Signal Ỹ arrives at date 2, the conclusion of the second sub-period, with Ỹ = v +ũ y, where ũ y N(0,ys v ). The cash flow is realized at the terminal date 3. 1 I interpret X as an information signal about firm value unrelated to the release of the financial statements. For example, X could be an analyst report. Signal Ỹ is an accounting signal that may only partially reveal the terminal cash flow. The risk-neutral investors price the firm according to the expected value of the cash flows conditional on the realized signals. The precision, or quality, of both signals is inversely related to x and y respectively. Accounting quality, therefore, is decreasing in y. I seek to understand the relation between the quality of the signals and the variance of price changes occurring at their respective realizations. Before deriving the expressions for variance, I must first derive the prices. Let P i be the price at date i, and let the ˆ notation denote a realization of a random variable. By the properties of normal random variables, P 0 = v, P 1 = 1 1+x ( ˆX v), P 2 = y y+x(1+y) ( ˆX x v)+ y+x(1+y) (Ŷ v), and P 3 =ˆv. Also,Var(P 1 P 0 )= sv 1+x,Var(P s 2 P 1 )= vx 2 (1+x)(x+y+xy), and Var(P 3 P 2 )= svxy x+y+xy. The first observation pertains to the sum of the variances. Observation 1 The sum of the variances of the price changes is s v, independent of the quality of the signals. The variance of the underlying the cash flows defines the total amount of uncertainty to be resolved by the subsequent information. If X is perfectly informative (x = 0), all of the price movement occurs at date 1. The second signal and the cash flow realization are redundant. Similarly, if X is perfectly uninformative (x = ) andỹ is perfectly informative (y = 0), then all of the uncertainty resolution 1 The results in the Observations are identical for the more general model in which ũ x N(0,s X)ũ y N(0,s y). The advantage of my specification is that the relative quality of the signals is independent of s v. 6

9 and price movement occur at date 2. Finally, if both signals are uninformative, then all uncertainty resolution and price movement are deferred until the realization of cash flows. The logic in the extreme cases is straightforward. The observation shows that the conservation of variance also holds for the non-extreme cases in which both signals are partially informative (0 <x,y< ). In the next observation, I examine the properties of the date 1 and date 2 variances deflated by the total variance, s v, i.e., the percentage of total variance that occurs on those dates. In particular, I derive the comparative statics of the measures with respect to the respective signal qualities. Because the total variance is independent of the quality of the information signals, the deflation is appropriate. Observation 2 The percentage of uncertainty explained by the first signal, X, is 1 1+x,increasingin the quality of X. The percentage of uncertainty explained by the second signal, Ỹ,is x 2 (1+x)(x+y+xy), decreasing in the quality of X and increasing in the quality of Ỹ. The result is straightforward. The uncertainty resolved by each signal is increasing in its own quality. Also, increasing the precision of X reduces the amount of residual uncertainty to be explained by Ỹ. Note that comparing the percentage of uncertainty explained by each signal is not a valid way to compare the quality of the signals. If the signals are equally precise (x = y), signal X will resolve more uncertainty because it arrives first: 1 1+x for X and 1 x 1+x 2+x for Ỹ. The result that increasing the quality of the first signal decreases the variance of the second price change is similar to the result derived in Holthausen and Verrecchia (1988), though the authors do not derive the conservation of variance result. Shin (2006), in his model rationalizing post-earnings announcement drift, exploits the notion that there is only so much fundamental uncertainty associated with the projects, and early 7

10 resolution of uncertainty implies that there is less uncertainty to be resolved later on (p. 345). The conservation of variance result justifies deflating by total variance in the empirical measure I will derive. Some researchers (for example, Lang and Lundholm [1993] and Leuz and Verrecchia [2000]) posit that higher quality accounting produces lower variance returns, the intuition being that noise in earnings produces noise in price. This does not happen in my model; noise in earnings causes Bayesian investors to attach a lower weight to earnings announcements, dampening the price reaction, a dampening that is completely offset by a stronger price reaction to subsequent information. The basic conservation of variance result holds in a more general, multiple-period setting in which the steadystate firm replenishes uncertainty by replacing expiring investments with new investments. I conjecture that it would also hold in a model with noise traders. It would not hold, however, if investors are uncertain about the variance of the underlying distribution of firm value, if investors have incorrect beliefs about the precision of the information signals or act in a non-bayesian manner, or if investors have a time-varying risk premium. The properties of the percentage of uncertainty explained by Ỹ motivate the empirical methodology in the following sections. I will construct a measure capturing the percentage of total uncertainty resolved by quarterly earnings announcements, and examine changes in this measure over time. An increase in the percentage uncertainty resolved by earnings announcements does not necessarily imply improved earnings quality. It might instead be the case that the quality of non-accounting information has declined, leaving more uncertainty resolution to the earnings announcements. The measure, therefore, can unambiguously capture only changes in the information environment. The econometric measure will not map perfectly into the theoretical measure. In particular, to ensure comparability to the prior literature, I compute the measure using returns instead of price changes. Also, some of the elements compromising the conservation of variance result may be present in real trading environments. The purpose of the theory model is to provide the intuition for the 8

11 relation between uncertainty resolution and return variance, and to emphasize that the variance of the underlying cash flows determines the amount of uncertainty to be resolved, whereas signal quality only allocates it to signal realization dates. 3 Hypotheses and variable definitions To implement the theoretical measure of the usefulness of earnings announcements literally, I would need to observe the variance of the entire distribution of price changes at each date. In practice, I can observe only the realization on each date, a single point from the underlying distribution. To create the measure, then, I must use a proxy for the variance of the distribution of excess returns on each date. As proxy, I use the square of the excess return. I compute the excess return manually as Excess i,j,t =ret i,j,t (ˆα i,t + ˆβ i,t MktRet j,t ), where the subscript i designates the firm, the subscript j designates the day/month, and the subscript t designates the year. Unlike other studies examining the excess return, which require Excess i,j,t only on the event date, I require Excess i,j,t on all trading days. This is an important data limitation, forcing me to compute the excess returns manually rather than rely on the CRSP beta-excess returns, which are available for fewer firm-days. 2 There are approximately 252 trading days in each year. I divide each year into 36 seven-day intervals, four of which are centered on the four quarterly earnings announcement dates. 3 In each interval, I square the sum of the seven raw excess returns. For example, the squared excess return for 2 For computational ease, I compute the firm-specific β only once per year. On a small sample of data, I compared the Excess i,j,t variable with the CRSP beta-excess variable and found a correlation of An alternative design is to divide the year into 84 three-day intervals, or simply to focus the attention on the announcement date excess return. In an unreported robustness check, I compute the scaled excess return variable common in the literature and find similar time-series properties. I conjecture that all of the related methods of measuring excess return would yield similar empirical results. 9

12 the 1 st interval in year t for firm i is RawExcess i,t,1 = ( Π 7 n=1excess i,n,t ) 2, where n designates the trading day in the year. Evenly spaced quarterly earnings announcements would occur in the 5 th,14 th,23 rd,and32 nd intervals. Quarterly announcements may not be evenly spaced. When I implement the measure, I use the actual earnings announcement dates. The 2000 year, for example, begins 28 trading days after the release of fiscal 1999 fourth-quarter earnings, or, equivalently, 28 trading days before the release of fiscal 2000 first-quarter earnings. The total excess return occurring at earnings announcement dates for firm i in period t, then, is: EAExcess i,t = RawExcess i,t,5 + RawExcess i,t,14 + RawExcess i,t,23 + RawExcess i,t,32. I now derive the metric for total annual uncertainty. The total annual excess return, defined for firm i in year t as : 36 TotalExcess i,t = RawExcess i,t,k. k=1 Finally, I compute the percentage of total excess return occurring at the earnings announcement dates, defined for firm i in year t as: AvgExcess i,t = EAExcess i,t TotalExcess i,t. This is the primary measure of earnings announcement usefulness in the study. If the rates of investment in new assets and information arrival were continuous and uniform, the expected value of AvgExcess i,t would be 4/36 =.111, or 11.1% of the uncertainty about firm value would be resolved during earnings announcement intervals. The principal hypothesis in the study follows: 10

13 Hypothesis 1 There has been no change in the usefulness of quarterly earnings announcements. 4 Evidence 4.1 Sample and descriptive data The Main Sample includes all firm-years from 1976 to 2005 that have complete sets of daily returns from CRSP, all four quarterly earnings announcement dates on Compustat, information on analyst forecasts (I/B/E/S), and enough information to calculate the 10-year firm-specific variance of operating cash flows. I set 1976 as the earliest date in my study because analyst forecast data is available only for a small number of firms prior to that date. Table 1 summarizes the composition of the main sample by year. Column 2 shows the total number of firms on Compustat with sufficient data to compute the market value of equity, book-to-market ratio and price-to-earnings ratio. In the early years, most of the data points lost are due to the unavailability of analyst data. In later years, most are lost due to the lack of 10 years of cash flow from operations information. The percentage of total firms in the Main Sample has increased steadily over time, largely due to increased analyst coverage. Table 2 summarizes the characteristics of the Main Sample, providing the median values of assets, market capitalization, book-to-market ratio, and the price-to-earnings ratio. The numbers in brackets are the median for the entire Compustat sample. On average, the firms in the Main Sample are larger, have a higher capitalization, similar book-to-market, and higher price-earnings ratio than the full Compustat sample. I obtained the information about voluntary disclosures from the First Call Company Issued Guidance database. There is little coverage before 1994 in this database. If there is no entry in the database for a firm year, I assign a value of 0. Panel A of Figure 1 shows the mean and median values of AvgExcess for the Main Sample. The mean AvgExcess for the Main Sample is 16.8% with a maximum of 22.0% in 2004 and a minimum of 11

14 13.7% in AvgExcess is higher than 11.1% in all years, suggesting that more uncertainty resolution occurs in announcement periods than non-announcement periods. There is a slight increase from 1976 to 2001 and a sharp increase starting in Figure 1a illustrates the principal finding in the paper. The subsequent empirical tests establish that the increase in usefulness is statistically significant. It is possible that the observed trend is caused by changes in the sample over time. In particular, the Main Sample comprises a much higher percentage of total Compustat firms in the period than it did previously. As a robustness check, I perform all of the econometric tests on the 91 firms that are in the sample for all 30 years. The All-Year-Sample comprises 2,730 of the 40,535 data points in the Main Sample. In general, these firms are larger, with larger analyst following, and lower variance of operating cash flows. Panel B of Figure 1 shows that a similar trend holds for the All-Year-Sample. The average AvgExcess over the period is 16.9%, with a maximum of 24.5% in 2004 a minimum of 11.7% in Tests of hypothesis In the first test of the hypothesis, I estimate the following regression: AvgExcess i,t = γ 0 + γ 1 Year i,t + γ 2 post2001 i,t + e i,t. (1) Year i,t is the calendar year of the observation (1976 = 1, 1977 = 2, etc). Consistent with the findings of other researchers, I predict a positive coefficient. The main variable of interest is post2001, an indicator variable that takes a value of 1 for fiscal years 2002 and later, and 0 otherwise. I choose the 2002 fiscal year as the cutoff because it is the first fiscal year for which it is plausible that all three changes in the reporting environment could have had an effect. Though Regulation FD might have had an effect prior to this date, the other two had not been implemented yet. I report the results in the columns labeled MAIN SAMPLE in Table 5. The coefficient on Year is positive and significant, consistent with Francis, Schipper, and Vincent (2002b) and Landsman and 12

15 Maydew (2002), who also find an increase in usefulness over time using earlier data. The coefficient on post2001 is also positive and significant, refuting the null hypothesis that there has been no change in the usefulness of earnings announcements. The coefficients imply that on average the amount of uncertainty resolved by earnings announcements increases from 14.03% in 1976 to 16.53% in 2001 and then increased to 20.07% in That is, the recent increase is much larger than the increase in usefulness already documented in the literature using earlier data. AvgExcess is a measure of information flow. An average of 15.3% of total uncertainty was resolved during the quarterly earnings intervals, or 3.82% per announcement, in the period. The remaining 84.7% of uncertainty, therefore, was resolved during the 32 non-announcement intervals, or 2.65% per interval. Thus, in the period, information flow was 44% higher during announcement periods than non-announcement periods. Information flow, however, was 103% higher during announcement periods than non-announcement periods in the interval (5.06% compared to 2.49%), suggesting a significant change in the financial reporting environment. [INSERT TABLE 3 HERE] It is possible that the positive coefficients on Year and post2001 can be explained by firm-specific factors. Thus, I estimate the following regression: AvgExcess i,t = γ 0 + γ 1 post2001 i,t + γ 2 Year i,t + γ 3 Lossfirm i,t + γ 4 VarCF i,t + γ 5 Analyst i,t + γ 6 Assets i,t + γ 7 BTM i,t + γ 8 HiTech i,t + γ 9 LoTech i,t +ForeOn i,t + γ 10 ForeOff i,t + γ 11 Lag i,t + e i,t.(2) Lossfirm is an indicator variable with a value of 1 if the firm has experienced a losing quarter during the fiscal year, 0 otherwise. The Lossfirm classification can change from year to year for an individual firm. Loss firms are more likely to preempt earnings announcements by early disclosure of bad news. Therefore, I predict this coefficient will be negative. VarCF is the 10 year time-series variance of operating cash flows. Because my sample begins before the introduction of the statement of 13

16 cash flows, I use the balance sheet method to compute operating cash flows. To the extent that highly volatile cash flows are related to unreliable accrual signals, I predict a negative coefficient. Analyst is the number of analyst forecasts during the fiscal year. The coefficient could take either sign. If analyst information complements (substitutes for) earnings information, then the sign will be positive (negative). Francis, Schipper, and Vincent (2002a) find that analyst forecasts and earnings announcements are complementary information. Assets is the size of the firm. Because more information is available for larger firms, I predict a negative coefficient. BTM is the book-to-market ratio. HiTech is an indicator variable with a value of 1 if the three-digit SIC code is 283, 357, 360-8, 381, 737, or 873, and LoT ech is an indicator variable equal to 1 if the three-digit SIC code is 020, 160, 170, 202, 220, 240, 245, 260, 300, 307, 324, 331, 356, 371, 399, 401, 421, 440, 451, and I expect the sign to be positive for Hitech because investors must rely more heavily on financial statements for more information about their relatively more complex operations. The opposite logic holds for Lotech. ForeOn is an indicator variable taking a value of 1 if the firm has forecasted future earnings in at least one of its quarterly earnings announcements during the year, 0 otherwise. ForeOff is an indicator variable taking a value of 1 if the firm has forecasted future earnings at least once on a non-earnings announcement date during the year, 0 otherwise. I predict a positive coefficient for ForeOn because investors respond to the news in the earnings forecast as well as the news in earnings. I predict a negative coefficient on ForeOff because the non-earnings announcement date forecasts may preempt the information in earnings announcements. Finally, Lag is the number of days between the last day of the fiscal year end month and the fourth quarter earnings announcement. 5 I predict a negative sign on Lag because the information in less timely earnings announcements is more likely to have been preempted by other sources. All the non-indicator variables are winsorized at the 1% level. The correlation matrix for the 4 I follow the Francis, Schipper and Vincent [2002a] classification. 5 The correlation between the fourth quarter lag and the other lags is approximately.70. The results are unchanged when I define Lag as the average of the four quarterly lags. Panel B of Figure 2 shows the average lag for the Main Sample over time. 14

17 explanatory variables is in Table 3, and their descriptive statistics are reported in Table 4. I present the results for the multivariate regression in the columns of Table 5 labeled Main Sample. The coefficients on post2001 and Yearare significant and positive, and their relative sizes are similar to the univariate regression. The coefficient on ForeOn also has the predicted positive sign. The coefficients on Lossfirm, LoTech, ForeOff and Lag are negative as predicted, and significant. The coefficient on VarCF is significantly negative at the 5% level. Finally, the positive coefficient on Analyst suggests that analyst forecasts and management forecasts are complements rather than substitutes. This corroborates the results in Francis, Schipper, and Vincent (2002a). The significance of the post2001 coefficient despite the inclusion of Lag implies that the recent increase is usefulness has not been caused by a reduction in the time between the end of the quarter and the earnings announcement, ruling out the accelerated 10-K filing deadline as the causal effect. In fact, Figure 2, Panel A shows that the lag between the end of the quarter and the earnings announcement has increased slightly in recent years. Hand collection (untabulated) of a small sample of 10-K filing dates indicates that the lag between the end of the quarter and the filing deadline has fallen, consistent with the new regulation, but the reduction occurs in the time between the earnings announcement and the filing, not between the end of the quarter and the earnings announcement. The inclusion of the ForeOn explanatory variable is important in distinguishing my paper from Francis, Schipper, and Vincent (2002b). These authors find that the increase in the usefulness of earnings announcements in the period can be explained by the simultaneous increase in the amount of other information disclosed by firms in earnings announcements. Replication of their research design, which requires hand-collection of data from the texts of earnings announcements, is beyond the scope of my study because of the larger sample size (their hand-collected dataset is 2,190 firm-years whereas I have 40,535 firm-years in my main sample). As Figure 2, Panel B shows, the percentage of firms issuing earnings guidance in earnings announcements increased from 0.2% in 1994, 15

18 the first year of coverage, to a peak of 44.3% in The ForeOn variable, thus, serves as a proxy for the amount of other disclosure in earnings announcement press releases. Including ForeOn and ForeOff reduces the coefficient on post2001 from 3.70 (untabulated) to That is, controlling for the incidence of non-earnings disclosures attenuates the strength of the finding, but the relevant coefficient remains highly significant and large relative to the size of the overall time trend Robustness checks The results in the previous section lead to the rejection of the null hypothesis that there has been no change in the usefulness of earnings announcements. The composition of the firms has changed over time, however. In particular, Table 1 reports that the Main Sample has increased from 19-20% of all Compustat firms in the pre-2002 period, to 30-35% of the firms from To verify that the previous results are not simply a function of the changing sample, I estimate equations (1) and (2) using only the 91 firms that have been in the sample all 30 years. I report the results in Table 5 in the columns labeled ALL-YEAR SAMPLE. The main finding is that the coefficients on post2001 and Year are still significant and of similar sizes to the coefficient in the Main Sample regression. Thus, the basic finding in the previous section that there has been a slow increase in usefulness that has magnified in recent years still holds. Only LoT ech and ForeOff of the remaining explanatory variables are still significant. The coefficient on Analyst is negative, unlike for the Main Sample, but insignificant. Lo and Lys (2000) and Buchheit and Kohlbeck (2002) examine whether the increase in usefulness documented by other researchers and find that it applies only to large firms. I address this issue by estimating equations (1) and (2) with the sample of firms excluded from the Main Sample because of 6 As a further control for other disclosure, I hand-collected data on the number of words in quarterly earnings announcements for the 91 firms in the All-Year Sample for the period. The number of words increases steadily throughout the period. The coefficient on post2001 remains significant and of similar magnitude after the inclusion of the words as an explanatory variable. 16

19 lack of analyst coverage or insufficient time series data to compute the variance of operating cash flows. 7 These firms are on average much smaller than the firms in the Main or All-Year samples (statistics in Table 4). Panel C of Figure 1 shows that the pattern of AvgExcess from for the Small Firm sample is similar to the other two samples. I report the regression results in the last two columns of Table 5. The post2001 coefficient is positive and significant, though only 70% as large as the coefficients in the Main and All-Year regressions. That is, the same effect holds, though it is of smaller magnitude. The overall time trend is also smaller for the Small Firm sample, though also positive. The coefficients on the voluntary disclosure and Lag variables are significant with the predicted sign. This test does not fully address the concerns of Lo and Lys (2000) and Buchheit and Kohlbeck (2002). Because the minimum data requirements of my study are stringent (a full set of returns and earnings announcement date), very small firms may not be included. 5 Identifying date of increase in usefulness The preceding sections present evidence consistent with a discrete change in the usefulness of earnings announcements occurring sometime after the implementation of Regulation FD, Sarbanes Oxley, and the accelerated filing deadline for Form 10-K. In this section, I conduct informal analysis attempting to identify more precisely the date of the shift. In order to examine the time-series of usefulness measures in more detail, I create a quarterly measure, QExcess, defined for firm i in year t and quarter j as QExcess i,t,j. = RawExcess i,t,j TotExcess i,t. 7 In untabulated results, I also estimated the regression for the smallest quarter of firms, regardless of analyst coverage or time series data. The mean (median) size for this sample was 31 (29). The empirical results are similar to the Small Firms results in Table 5. 17

20 In other words, this is the percentage of total annual excess return that occurs at the quarterly announcement dates. 8 I begin with the premise that there has been a discrete change in the usefulness of accounting earnings, and then estimate the following regression to determine the most likely date for the change: QExcess i,t = γ 0 + γ 1 Cut i,t + γ 2 Year i,t + γ 3 Analyst i,t + γ 4 Assets i,t + γ 5 Lossfirm i,t + γ 6 ForeOff i,t + e i,t The variable Cut is an indicator variable taking a value of 1 if the data point occurs after the test month and 0 otherwise. I estimate the regression for each test month, starting with February 1976 as the first test month and ending with November 2005 as the last test month, 358 separate regressions in all. The definitions of the other variables are the same as in the earlier sections. I argue informally that the test-month regression yielding the highest explanatory power is the most likely month in which the asserted discrete change in usefulness occurred. In Panel A of Figure 3, I plot the R 2 yielded by the regressions. The Figure shows that the R 2 curve reaches its maximum in June The results suggest that the change is most closely related to Sarbanes Oxley. Because SOX was enacted on 31 July 2002, the first 10-K filings to which it applies occur in August Given the lag between the earnings announcement and the filing of the 10-K, the filings in August 2002 are associated with earnings announcements in June In Panel B of Figure 3, I plot γ 1, the coefficient on Cut. It also attains its maximum in the June 2002 test-month regression. I note, however, that this test is informal and that the evidence of a link between SOX and the increase is circumstantial. 6 Discussion and Conclusion In light of recent changes in the financial reporting environment (Regulation Fair Disclosure, Sarbanes- Oxley, and accelerated filing dates), I examine the usefulness of quarterly earnings announcements 8 The distributions of the four quarterly returns are similar so I make no adjustments. 18

21 between 1976 and I measure usefulness as the percentage of total annual excess returns that occurs on or around earnings announcement dates (AvgExcess). I motivate the deflation by total annual excess returns by demonstrating analytically that the total variance of price changes depends only on the variability of underlying cash flows, not on the quality of the information signals. That is, a conservation of variance holds in which high (low) quality signals accelerate (defer) the resolution of uncertainty but have no effect on the total amount of uncertainty to be resolved. I compute the measure by dividing each fiscal year into (approximately) 36 7-day periods, 4 of which straddle earnings announcements, 32 of which do not. I provide empirical evidence that the usefulness of earnings announcements increased significantly beginning in the 2002 fiscal year. The increase is approximately 20-25%, much larger than the overall upward trend documented by other researchers using earlier data, and statistically significant. The empirical finding holds for the sample of 91 firms that meet the data requirements for all 30 years as well as for a sample of smaller firms. Francis, Schipper, and Vincent (2002b) document an increase in the usefulness of earnings announcements from and show that it can be explained by increases in other disclosures included in the earnings announcement over that time period. Though controlling for the incidence of management forecasts on earnings announcement dates reduces the magnitude of the increase in my study, it is still significant and large relative to the underlying time trend. In addition, I control for the timeliness of earnings announcements, which also cannot explain the increase. The results suggest the possibility that it is a property of earnings that is driving the increase in the usefulness, not a general increase in firm disclosure on earnings announcement dates. I leave for future research an exploration of those properties (earnings quality, for example). The evidence that there has been a recent change in the way in which information flows in the market seems robust. Relatively more information has been arriving in earnings announcements in recent years. This does not necessarily imply an improvement in the quality of financial disclosure 19

22 during earnings announcement periods. The cause could instead be worse disclosure at non-earnings announcement periods. This suggests a link to Regulation Fair Disclosure, which explicitly prohibits firms from selectively disclosing information. I derive a quarterly measure of usefulness to attempt to isolate the specific date of change in the usefulness time series. I make the presumption that there has been a one-time change in the time series. I find that the regression in which the change date indicator variable is June 2002 has the highest explanatory power. This informal analysis is consistent with the Sarbanes Oxley Act, implemented on July 30, 2002, being the trigger for the change, not Regulation FD, which was implemented in November The lack of a link to Regulation FD is not surprising given the earlier researchers who have failed to find any consequences (see Bailey, Mao and Zhong [2003], Eleswarapu, Thompson and Venkataraman [2004], Francis, Nanda, and Wang [2004], Gadarowski and Sinha [2002]). I leave for future research empirical tests linking the change in usefulness more rigorously to the Sarbanes Oxley Act. 20

23 7 References 1. Ashbaugh-Skaife, H., Collins, D., Kinney, W. and R. Lafond The effect of SOX internal control deficiencies and their remediation on accrual quality. Working paper, MIT. 2. Bailey, W., Li, H., Mao, C. and R. Zhong Regulation Fair Disclosure and earnings information: market, analyst, and corporate responses. Journal of Finance 58 (6): Beaver. W The information content of annual earnings announcements. Journal of Accounting Research 6 (Supplement): Buchheit, S. and M. Kohlbeck Have earnings announcements lost information content? Journal of Accounting, Auditing and Finance 17 (Spring): Dechow, P. and I. Dichev The quality of accruals and errors: the role of accrual estimation errors. The Accounting Review 77 (Supplement): Eleswarapu, V., Thompson, R. and K. Venkataraman Measuring the fairness of Regulation Fair Disclosure through its impact on trading costs and information asymmetry. Journal of Financial and Quantitative Analysis 39 (June): Francis, J., Nanda, D. and X. Wang Re-examining the effects of Regulation Fair Disclosure using foreign-listed stocks to control for concurrent shocks. Journal of Accounting and Economics 33 (August) Francis, J. and K. Schipper Have financial statements lost their relevance? Journal of Accounting Research 37 (2): Francis, J., Schipper, K. and L. Vincent. 2002a. Earnings announcements and competing information. Journal of Accounting and Economics 33 (3):

24 10. Francis, J., Schipper, K. and L. Vincent. 2002b. Expanded disclosure and the increased usefulness of earnings disclosures. The Accounting Review 77 (July): Gadarowski, C. and P. Sinha On the efficacy of Regulation Fair Disclosure: theory and evidence. Working paper, Cornell University. 12. Gompers, P., Ishii, J., and A. Metrick Corporate governance and equity prices. Quarterly Journal of Economics 18 (February): Heflin, F., Subramanyam, K. and Y. Zhang Regulation FD and the financial information environment: early evidence. Accounting Review 78 (January): Holthausen, and R. Verrecchia The effects of sequential information releases on the variance of price changes in an intertemporal multi-asset market. Journal of Accounting Research 26 (1): Kross, W. and M. Kim Differences between market responses to earnings announcements in the 1990s versus the 1960s. Working paper, Purdue University. 16. Landsman, W. and E. Maydew has the information content of quarterly earnings announcements declined in the past three decades? Journal of Accounting Research 40 (3): Lang, M. and R. Lundholm Cross-sectional determinants of analyst ratings of corporate disclosures. Journal of Accounting Research 31 (Autumn): Leuz, C. and R. Verrecchia The economic consequences of increase disclosure. Journal of Accounting Research 38 (Supplement): Lev, B. and P. Zarowin The boundaries of financial information and how to extend them. Journal of Accounting Research 37(2):

25 20. Lo, K. and T. Lys Bridging the gap between value relevance and information content. Working paper, Northwestern University. 21. Ogneva, M., Raghunandan, K., and K.R. Subramanyam Internal control weakness and cost of equity: evidence from SOX Section 404 disclosures. Working paper, University of Southern California. 22. Shane, P., Soderstrom, N. and S. Yoon Earnings and price discovery in the post Regulation FD information environment: a preliminary analysis. Working paper, University of Colorado, Boulder. 23. Shin, H Disclosure risk and price drift. Journal of Accounting Research 44 (2):

26 Table 1 Breakdown of Number of Firms in Main Sample by Year (1) (2) (3) (4) (5) (6) Year Compustat Announcement Announcement MAIN % of Compustat CRSP CRSP/Analyst SAMPLE in MAIN ,005 1, ,977 1, ,171 1, ,476 1, ,709 1, ,313 1, ,314 1,667 1, ,665 2,229 1,312 1, ,744 2,416 1,459 1, ,749 2,780 1,607 1, ,995 2,720 1,577 1, ,210 2,821 1,658 1, ,003 2,927 1, ,848 3,039 1,752 1, ,783 3,097 1,822 1, ,867 3,230 1,892 1, ,140 3,318 1,924 1, ,145 3,666 2,145 1, ,563 4,089 2,463 1, ,308 4,656 2,818 1, ,942 4,850 3,030 1, ,996 5,119 3,306 1, ,892 5,185 3,470 1, ,122 5,219 3,518 1, ,000 5,472 3,587 1, ,484 5,519 3,567 1, ,044 5,247 3,533 2, ,683 4,930 3,524 2, ,458 4,842 3,608 2, ,147 4,676 3,643 2, TOTAL 201, ,114 65,116 40, Column (1) is the fiscal year. Column (2) is the total number of firms in Compustat with asset data. Column (3) is the total number of firms with a complete set of daily returns. Column (4) is the total number of firms that have a complete set of daily returns and quarterly earnings announcement dates. Column (5) is the total number of firms with a complete set of daily returns, quarterly earnings announcements, and analyst following data. This is the Main Sample in the study. Column (6) is % of total Compustat firms represented by the Main Sample. 24

27 Table 2 Descriptive Statistics (Medians) on Main Sample [All Compustat Firms] (1) (2) (3) (4) (5) Year Assets ($ mil) Market Value of Market-to-Book Price-Earnings Equity ($ mil) Ratio [70.1] [26.6] 0.83 [1.18] 9.21 [6.88] [76.3] [30.3] 0.93 [1.16] 7.69 [6.86] [75.4] [31.7] 1.02 [1.12] 6.95 [6.45] [70.3] [31.2] 1.00 [1.04] 6.76 [6.15] [65.2] [35.7] 0.93 [0.89] 8.00 [7.19] [49.9] [26.3] 0.97 [0.90] 7.80 [6.72] [48.7] [28.3] 0.90 [0.84] 9.72 [7.45] [49.1] [40.8] 0.71 [0.62] [9.26] [47.6] [32.9] 0.78 [0.71] [8.18] [50.3] [39.4] 0.69 [0.61] [9.36] [51.8] [40.4] 0.64 [0.57] [10.15] [54.3] [36.1] 0.71 [0.65] [8.46] [60.9] [38.9] 0.67 [0.64] [8.35] [66.2] [43.2] 0.63 [0.60] [8.75] [69.9] [34.9] 0.76 [0.75] [7.46] [74.2] [54.9] 0.63 [0.59] [10.16] [78.3] [67.5] 0.57 [0.54] [11.14] [110.1] [79.4] 0.52 [0.52] [11.87] [119.6] [75.5] 0.57 [0.58] [11.16] [108.5] [85.0] 0.50 [0.51] [11.78] [111.2] [96.1] 0.48 [0.47] [12.16] [122.9] [114.2] 0.42 [0.43] [13.63] [140.5] [95.0] 0.49 [0.52] [9.99] [147.5] [102.6] 0.53 [0.48] 12,71 [7.91] [160.6] [79.3] 0.51 [0.57] [15.09] [159.7] [84.4] 0.49 [0.54] [0.24] ,053.2 [177.9] [79.6] 0.58 [0.62] [3.31] ,108.2 [209.5] 854,8 [158.6] 0.46 [0.43] [10.88] ,062.3 [250.8] [209.5] 0.43 [0.41] [13.22] ,090.4 [311.2] [237.8] 0.43 [0.41] [13.24] 25