The Conservatism Principle and the Asymmetric Timeliness of Earnings: An Event-Based Approach

Pervin K. Shroff * University of Minnesota Ramgopal Venkataraman Southern Methodist University Suning Zhang George Mason University September 2007

We thank Sudipta Basu, Richard Dietrich, Peter Easton, Chandra Kanodia, Wayne Landsman, Karl Muller, Doron Nissim, John Phillips, Douglas Schroeder, Rajdeep Singh, and workshop participants at the University of Minnesota, the Ohio State University, CUNY-Baruch College, George Washington University, Columbia University, University of Cincinnati and participants of the 2004 Financial Accounting and Reporting Section Conference and the 2004 American Accounting Association Conference for valuable comments and suggestions.
*

University of Minnesota, Carlson School of Management, 321, 19th Ave South, Minneapolis, MN 55455. Tel.: (612) 626-1570. email: shrof003@umn.edu , ramgopal@smu.edu, szhang4@gmu.edu

The Conservatism Principle and the Asymmetric Timeliness of Earnings: An Event-Based Approach

Abstract In this paper, we test the conservatism and asymmetric timeliness hypothesis by using information in extreme events as a measure of good/bad news. Our focus on extreme events is motivated by two arguments. First, the accounting concept of materiality in conjunction with litigation risk influences managers and auditors to make more conservative choices with respect to material events. Second, focusing on large events minimizes the probability that accounting slack may obscure the effect of asymmetric timeliness (Beaver and Ryan, 2005). We identify individual events using short-window returns, since long-window returns may aggregate multiple events and thus limit our ability to detect conservatism. Taken together, these features of our research design provide more powerful tests of asymmetric timeliness. Consistent with prior studies, we document that the correlation between bad news and concurrent earnings is significantly higher than that between good news and concurrent earnings. Our event-based research design obtains results even in a quarterly analysis that are easy to interpret as evidence of asymmetric timeliness. Unlike some prior studies, we find no evidence of asymmetric timeliness in cash flows using our approach. Moreover, our analysis of extreme events enables us to document higher correlation of good news with earnings three or more quarters ahead. This is in contrast to prior studies that did not focus on extreme events and were unable to document asymmetry in the relation between returns and subsequent earnings in the opposite direction to that between returns and current earnings.

1. Introduction Accounting conservatism is often described as a reporting objective that calls for anticipation of all losses and expenses but defers recognition of gains or profits until they are realized (Weil, Stickney, and Davidson, 1990, p. 22). Ceteris paribus, this implies that bad news is recognized earlier than good news in reported earnings. In an influential paper, Basu (1997) tested the asymmetric timeliness hypothesis by regressing annual earnings on contemporaneous returns using positive (negative) annual stock returns as a proxy for good (bad) news. He found that the contemporaneous sensitivity of earnings to negative returns is significantly higher than that of earnings to positive returns. 1 In this paper, we rely on the accounting concept of materiality in conjunction with litigation risk and focus on extreme events to develop powerful tests of the asymmetric timeliness hypothesis. Our approach has several desirable properties that add to the weight of the evidence in favor of conservatism. We test the asymmetric timeliness hypothesis using an event-based approach that modifies Basus proxy for good/bad news. Asymmetric timeliness relates to when the information in an individual economic event is recorded in accounting earningsearlier if the event conveys bad news and later if it conveys good news. Basu (1997) uses the aggregate positive or negative effect of all events occurring during a year as the proxy for good or bad news. If the incremental information in individual economic events is recorded with asymmetric timing, then Givoly, Hayn and Natarajan (2007) show that aggregation of all good news and bad news events that occur during a period may

Other papers that have since used this method include (among others) Pope and Walker (1999), Ball, Kothari, and Robin (2000), Giner and Rees (2001), and Ball, Robin, and Wu (2003), which study international differences in conservatism; Basu, Hwang, and Jan (2002), which studies conservatism and audit quality; Holthausen and Watts (2001), Ryan and Zarowin (2003) and Sivakumar and Waymire (2003), which study the importance of conservatism over time.

obscure this effect and thus impair our ability to detect conservatism. 2 In this paper, we develop a proxy for good or bad news reflected in individual economic events which will be more powerful in detecting asymmetric timeliness than an aggregate proxy. Because the economic impact of an individual event is not directly observable, we rely on short-window returns as the measure of value-relevant information in an event. We assume that a good news (bad news) event has occurred if we observe an unusually high (low) 3-day marketadjusted stock return for the firm during a given fiscal quarter. We argue that, assuming (semistrong) market efficiency, economic events are incorporated in prices as soon as they are publicly revealed. Thus, in our setting, an economic event occurs and is revealed, the market reacts to it, and the 3-day market-adjusted return captures the value-relevant information in that event. We test whether this information is incorporated in accounting earnings earlier if it reflects bad news than if it reflects good news. We define good/bad news by identifying extreme stock return behavior based on (i) a uniform cut-off for all firms, and (ii) a firm-specific cut-off. Specifically, good (bad) news is defined as: (i) 3-day market-adjusted return greater (less) than or equal to 10% (-10%) during a given quarter for a given firm, or (ii) 3-day market-adjusted return greater (less) than or equal to the mean plus (minus) 2.58 times the standard deviation of 3-day adjusted returns for a given firm. When identifying extreme returns, we exclude the returns of the earnings announcement window. There are some conceptual reasons for testing the effect of asymmetric timeliness of events that have a material economic impact. The accounting concept of materiality is a key factor governing the decisions of managers, auditors, litigators, and regulators. Our focus on extreme
Givoly et al. (2007) discuss this issue and, using a simulation procedure, show that aggregation of multiple random shocks in an interval impairs the power of the differential coefficient on negative returns to detect conservatism.
2

returns is motivated by the materiality concept in conjunction with shareholder litigation. Beaver (1993) and Watts (1993) note that the threat of litigation under the Securities Acts leads to conservative reporting, because the likelihood of litigation is higher when earnings and net assets are overstated rather than understated. Since omissions or misstatements that are material in magnitude are more likely to trigger litigation (as indicated by the emphasis on material in various statements of the Securities Act of 1933), and/or enforcement actions by the Securities and Exchange Commission (SEC), we expect managers and auditors to especially apply conservative accounting to events that are material in magnitude. Beaver and Ryan (2005) also provide support for the use of events of large magnitude in tests of asymmetric timeliness. In their model of conditional and unconditional conservatism, they show that the asymmetric relation between earnings and returns depends on the amount of accounting slack for recorded assets and the uncertainty of the write-down trigger. Their model and simulation results show that using negative events that are too small to use up the available accounting slack reduces the power of asymmetric timeliness tests. Consistent with their model, our focus on extreme events enables more powerful tests of asymmetry in the relation between returns and both concurrent and subsequent earnings. Our proxy for good/bad news enhances the power of our tests in three ways: (i) by focusing on short return windows to identify events; (ii) by identifying firm-specific events with material economic impact; and (iii) by excluding the effect of market-wide events. To identify events with material economic impact, we focus on positive or negative returns that are significant in

magnitude. 3 We examine rolling 3-day return windows and identify the window where we find significant returns indicating that a material economic event has occurred. Since we infer that an economic event has occurred from the return performance, the short 3-day window improves our ability to isolate the economic effect of the event, thereby improving the power of our test. We use market-adjusted returns in order to exclude market-wide events that may not have any impact on an individual firms earnings. Our objective in testing only extreme firm-specific events is to maximize the power of our tests which may come at the cost of generalizability of results to a broader set of events. In support of the above arguments, we also demonstrate how using longer return windows to capture value-relevant information in events and focusing our tests on events with lower economic impact impairs our ability to detect asymmetric timeliness in the data. Our results show that the R2 from a regression of earnings changes on extreme returns is significantly higher for the sample of firms with negative extreme returns relative to the sample with positive extreme returns. Also, the estimated differential coefficient on extreme negative returns is positive and significant. Moreover, as one would expect, we obtain positive and significant coefficient estimates on extreme positive as well as extreme negative returns, which allows us to easily interpret the positive differential coefficient estimate on extreme negative returns as evidence of asymmetric timeliness. 4 Additionally, we provide evidence on the dual effects of asymmetric timeliness. If accounting is indeed conservative (i) bad news should have a higher correlation with earnings of the
3

As another example, Bernard and Thomas (1990) explain that they focus on the extreme deciles of standardized unexpected earnings to provide a more powerful test that enhances their ability to document the post-earningsannouncement drift.

While prior studies using the Basu method find positive and significant coefficient estimates in annual regressions, tests using quarterly data yield significant negative coefficients on positive returns, which are inconsistent with the welldocumented positive returns-earnings association. These negative coefficients make it difficult to interpret the differential

concurrent quarter relative to good news; and (ii) good news should have a higher correlation with subsequent earnings. We find that, although good news continues to have a lower positive correlation with earnings changes cumulated over the subsequent two quarters, it is more positively correlated with earnings changes cumulated over the subsequent three quarters or longer. This confirms the argument in Beaver and Ryan (2005) that focusing on events of large magnitude allows us to clearly document that the asymmetric relation between returns and subsequent earnings is directionally opposite to that between returns and current earnings. Conservatism is an accrual accounting property, and hence should have minimal effect on cash flows. Consistent with this notion, Givoly and Hayn (2000) find that while earnings are negatively skewed in the cross-section, no such bias is apparent in operating cash flows. Similarly, we do not find any evidence of asymmetric timeliness in operating as well as free cash flows using our event-based approach. In summary, our evidence supports both predictions of the asymmetric timeliness hypothesisthat is, bad news is contemporaneously reflected in accounting earnings and returns, while the accounting recognition of good news is delayed to later periods. We contribute to the growing literature on conservatism by modifying the Basu methodology to enhance the power of the test of asymmetric timeliness. Our methodology can be applied to situations where tests have failed to identify asymmetric timeliness, for example, in the case of code law countries as documented by Ball, Kothari, and Robin (2000). It may be premature to label the accounting systems of these countries as less conservative without the benefit of a more powerful test of the hypothesis. The rest of the paper is organized as follows. Section 2 describes our hypotheses. Section 3

coefficient on negative returns since the differential is relative to a base estimate that is not consistent with expectations.

discusses the data and research design. Empirical results are reported in Section 4, followed by concluding remarks in Section 5.

2. Hypotheses Development Traditionally, conservatism in accounting means when in doubt choose the solution that will be least likely to overstate assets and income (Kieso, Weygandt, and Warfield, 2007, p. 46). Para. 95 of the Statement of Financial Accounting Concepts (SFAC) No. 2 states: ....if two estimates of amounts to be received or paid are about equally likely, conservatism dictates using the less optimistic estimate. In emphasizing the conservatism principle, accounting practice prefers skepticism in recognizing gains when some uncertainty attaches to the successful completion of the transaction and requires a higher standard of verification before recognizing gains. On the other hand, less stringent requirements are imposed in recognizing expenses/losses from incomplete transactions, as mandated, for example, by standards on pensions, environmental liabilities, and asset impairment. 5 As a consequence of conservative accounting, bad news is more likely to be incorporated in accounting earnings on a timely basis compared to good news. Prior studies that document accounting conservatism have used various measures of conservatism, for example, choice among alternative accounting methods (Bowen et al., 1995), the level and rate of accumulation of negative non-operating accruals, skewness of earnings relative to skewness of operating cash flows (Givoly and Hayn, 2000), unrecorded or hidden reserves

Watts and Zimmerman (1986, pp. 205-206) and Watts (2003) discuss alternative explanations for conservative reporting: contracting (conservatism as a means of efficient contracting with debt-holders and managers); shareholder litigation (conservatism reduces expected litigation costs); taxation (conservative reporting reduces the present value of taxes); and accounting regulation (conservatism reduces the political costs imposed on standard-setters and regulators).

(Penman and Zhang, 2002), and the market-to-book ratio (Stober, 1996). Basu (1997) and related papers focus on the timing dimension of conservatism. They rely on the conservative accounting rule, anticipate no profits, but anticipate all losses, which leads to more timely recognition of bad news relative to good news. 6 These studies use contemporaneous positive/negative annual stock returns as the measure of publicly available good/bad news. They estimate a regression of (scaled) accounting earnings on contemporaneous stock returns with differential coefficients for negative versus positive returns. They find that the contemporaneous sensitivity of earnings to negative returns is significantly higher than that of earnings to positive returns. Recent research has shown that this methodology for testing asymmetric timeliness yields results that are difficult to interpret when using quarterly data. For example, Basu, Hwang, and Jan (2002) report a positive coefficient estimate on negative returns as expected, but a negative coefficient estimate on positive returns which is inconsistent with the well-documented positive returns-earnings association. 7 The positive differential coefficient estimate on negative returns is not readily interpretable as evidence of asymmetric timeliness since it is relative to a base coefficient that is inconsistent with our expectations. If asymmetric timeliness relates to the recording of the incremental information in individual economic events, the use of long-interval returns over the entire period as an aggregate proxy for good/bad news may reduce the power of the test (see Givoly et al., 2007). Besides the concern about the power of the test, negative returns over long intervals are
6

This type of accounting reflects the notion of conservatism that is labeled by Beaver and Ryan (2005) as conditional (or news dependent), meaning that book values are written down under sufficiently adverse circumstances but not written up under favorable circumstances. Similar distinctions in the type of conservatism are made by Ball and Shivakumar (2005), Basu (1997, 2001), and Ball, Kothari, and Robin (2000). We similarly obtain negative coefficient estimates on positive returns from the Basu-regressions using recent annual data, both for the pooled regression and for 8 out of 14 annual regressions over the 1992-2005 time-period.

bounded from below (-100%) while positive returns are unbounded. This results in negative returns having lower variance than positive returns. In fact, consistent with a paper by Dietrich, Muller, and Riedl (2007), in our sample, we find that the variance of positive annual returns is about seven times larger than the variance of negative annual returns. This difference in variances of negative and positive returns leads to a difference in the ratio of earnings variance to return variance of the two groups. The interpretation of the differential coefficient on negative returns as evidence of asymmetric timeliness becomes difficult, to the extent this coefficient is driven by the lower bound on annual returns. Our tests of asymmetric timeliness using extreme 3-day event returns are not subject to the above concerns regarding differential variances. In our sample, we find that the variance of extreme market-adjusted event returns for positive news firms (0.0284) is not substantially different from that for negative news firms (0.0212). Moreover, we obtain similar results when we standardize the model, so that the returns of both positive and negative news firms have equal (unit) variance.

2.1 Extreme 3-day returns as a measure of good/bad news In an efficient market, prices incorporate the effect of economic events as soon as they are revealed. We assume that unusually high/low returns over a 3-day interval capture information in an economic event that has just been revealed. Thus, we measure good (bad) news as unusually high (low) 3-day market-adjusted returns for a firm during a given quarter. Note that an unusual 3-day return is our measure of value-relevant information contained in a publicly disclosed news event (although we make no attempt to find out the specifics of this news event). Thus, we take an event perspective rather than study the association between accounting earnings and contemporaneous

returns (either fiscal period or announcement date to announcement date). As discussed earlier (and demonstrated later), our use of extreme short event window returns (as opposed to annual/quarterly interval returns) increases the power of our tests, thus increasing our ability to detect asymmetric effects. Because we infer that an economic event has occurred from the return performance, the short 3-day window enhances our ability to isolate the economic impact of the event. Our focus on material economic events is motivated by conceptual as well as statistical reasons. Shareholder litigation is considered by many as a potential source of conservatism (see Watts, 2003). Litigation produces asymmetric payoffs: overstating earnings and net assets is more likely to generate litigation costs than understating earnings and net assets. Conservatism, by understating earnings and net assets, reduces the firms expected litigation costs. Consistent with this view, Skinner (1994) finds that managers voluntarily disclose bad news to avoid costs of potential litigation. While in general managers and auditors have the incentive to recognize losses on a timely basis relative to gains to reduce litigation risk, the incentive is more pronounced when losses are material in magnitude. Materiality plays a significant role in determining when litigation is triggered and hence is an important decision factor for managers, auditors, litigators, and regulators. The Financial Accounting Standards Board (FASB) defines materiality in SFAC No. 2.8 Several accounting pronouncements also emphasize materiality, for example, APB Opinion 30: Reporting the results of operations. Auditing standards too emphasize the concept of materiality (see Statement of Auditing Standards No. 47). The auditors consideration of materiality is a matter of professional judgment and is affected by what the auditor perceives as the view of a reasonable person who is relying on the financial statements (Boynton, Johnson, and Kell, 2001, p. 171). The
8

SFAC No. 2 states: Materiality is the magnitude of omission or misstatement of accounting information that, in the light of surrounding circumstances, makes it probable that the judgment of a reasonable person relying on the information

auditor is required to provide reasonable assurance that all material misstatements are detected. While no exact rules for determining materiality are prescribed, a common rule of thumb applied by auditors is that total (aggregated) misstatements of more than about 3% to 5% of net income before taxes would cause financial statements to be materially misstated. 9 Suits against auditors under the Securities Act of 1933 are usually based on material misstatements or omissions.10 Moreover, under the civil provisions of the 1933 Act, monetary damages recoverable equal the magnitude of loss suffered by the plaintiff based on the prevailing price of the security. Thus, managers and auditors are likely to be more conservative in recognizing negative events that are material in magnitude, since the likelihood of litigation and its potential cost is higher for material misstatements or omissions. As discussed earlier, we define a firm as a good (bad) news firm for a quarter, if the cumulative market-adjusted return in any 3-day window during the quarter is greater (less) than or equal to 10% (-10%), or is greater (less) than or equal to the mean plus (minus) 2.58 times the standard deviation of 3-day adjusted returns for that firm (calculated over the previous five years). As a further restriction, 3-day adjusted returns that are less than 5% in absolute value are not included in the good/bad news samples. To identify these extreme returns, we apply our criteria to rolling 3-day adjusted returns in a quarter for each firm. If a firm-quarter has consecutive rolling 3day adjusted returns that fall outside our cut-off, we cumulate returns over a period longer than 3 days in order to include the total reaction to the news event. For example, if we find two
would have been changed or influenced by the omission or misstatement. 9 Consistent with the importance of materiality, Wright and Wright (1997) find, using data from actual audit engagements, that the auditors decision to waive an audit adjustment (which reflects the auditors discovery of a potential breach in the clients accounting system) is correlated with the materiality of the proposed adjustment.
10

For example, see Section 11, Civil Liabilities on Account of False Registration Statement.

10

consecutive rolling 3-day adjusted returns outside our cut-off, we consider the 4-day window to obtain the total reaction to the news. Further, since our purpose is to obtain information in an economic event other than an earnings announcement, we exclude returns that fall during a 3-day interval surrounding an earnings announcement. When multiple extreme 3-day adjusted returns occur during the quarter, we cumulate the 3-day adjusted returns during the quarter if they are all of the same sign. If there are multiple 3-day intervals in the same quarter with a mix of extreme positive and extreme negative adjusted returns, we exclude that firm-quarter observation. 11 Thus, our final sample includes firm-quarters when a significant event(s) that conveys clearly positive or clearly negative news occurred as evidenced by the unusual price reaction. The market adjustment of returns is based on the value-weighted market index. We use market-adjusted returns in order to exclude market-wide events that may not have any impact on an individual firms earnings. By focusing on firm-specific returns, it is possible that we exclude some market-wide events that may have an impact on an individual firms earnings. Our objective in testing only firm-specific events is to maximize the power of our tests. Another advantage of excluding market-wide events is mitigation of calendar-time event clustering and the resultant crosscorrelations in error terms. At the same time, market-adjustment may lead to calendar-time clustering in the case of thinly traded stocks. Elimination of penny stocks from our sample and consistent results for large-sized firms help mitigate concerns with respect to this potential problem. We estimate a regression of change in earnings per share (EPS) for quarter t, deflated by beginning-of-quarter price, on the unusual 3-day adjusted return (or cumulated return) observed during quarter t, separately for the good news and the bad news samples. Quarterly earnings
11

This is consistent with Givoly et al. (2007) who use a simulation procedure to show that aggregating multiple events with different signs reduces the power of the test of asymmetric timeliness.

11

changes are defined as EPS of the current quarter minus EPS of the same quarter of the previous year. We use EPS after extraordinary items and discontinued operations because extreme news may be incorporated in accounting earnings as part of these items. 12 We hypothesize that, on average, bad news in quarter t is incorporated in accounting earnings of quarter t, while the accounting recognition of quarter-t good news may be delayed to subsequent periods. Thus, we expect that the R2 and the slope coefficient for the bad news sample will be significantly higher than that for the good news sample. However, this test addresses only one aspect of asymmetric timeliness. To provide complete evidence of asymmetric timeliness, we also need to test whether good news is in fact incorporated in earnings with a delay. Finding a lower coefficient on positive relative to negative extreme returns could also mean that the market incorporated good news with a lag and in fact the good news was incorporated in earnings of the previous period. To ensure that this is not occurring, we test whether good news is in fact incorporated in subsequent periods earnings. We expect the R2 and the slope coefficient from a regression of earnings changes of subsequent quarters on the unusual 3-day adjusted return observed in quarter t to be higher for the good news sample than for the bad news sample.

2.2 Subsequent-period analyses by prior research A few previous papers examine the effect of current good/bad news on earnings of subsequent periods. For example, in examining cross-country differences in the degree of

We repeat our analyses with earnings before extraordinary items and discontinued operations with substantially similar results.

12

12

accounting conservatism, Ball, Kothari, and Robin (2000) use accumulated net income over two years (t and t+1) as the dependent variable and regress it on returns of year t allowing for a differential coefficient on negative returns (reported in their Table 7). They find that, while the coefficient estimates generally increase (compared to those estimated with net income of year t as the dependent variable), contrary to expectation the estimated coefficient on positive returns (for the U.S. sample) is negative and the differential coefficient on negative returns in fact increases. Similarly, Basu (1997) estimates a regression of aggregate earnings over a four-year period on returns over the same period and finds that the differential coefficient on negative returns continues to be positive and significant although it is smaller compared to the one for the annual regression (reported in his Table 5). It appears that the purpose of these tests was to document a reduction in the effect of conservatism on earnings over longer intervals and not to test whether positive returns have a higher correlation with subsequent earnings than negative returns. Pope and Walker (1999) include lagged returns as additional independent variables in the regression of earnings on contemporaneous returns and find that, for the U.S. sample, the differential coefficient on negative lagged returns continues to be positive although it declines with longer lags. From their result (reported in their Table 6), positive returns do not have a higher correlation than negative returns with earnings as far as three years ahead. Ryan and Zarowin (2003) document similar results over different sub-periods from 1966 to 2000. Beaver and Ryan (2005) provide an explanation for the asymmetry being in the same direction for concurrent as well as subsequent earnings. They argue that negative returns that are too small to use up the available accounting slack for recorded assets may exhibit a higher correlation with subsequent earnings than positive returns. We believe that, consistent with their finding, our focus on extreme events will enable us to

13

document asymmetry with respect to current earnings to be in the opposite direction to that of subsequent earnings.

3. Data and Research Design Our initial sample includes all firms with data available on CRSP and Quarterly Compustat tapes covering the period 1982-2000. We use the preceding 20 quarters (five years) to estimate the mean and the standard deviation of 3-day market-adjusted returns for each firm. 13 Hence, our final sample includes data from 1987 to 2000. Of a total of 353,642 firm-quarter observations with data on CRSP, 67,080 (19%) observations are identified as those with extreme good/bad news. The final sample includes 61,867 (17.5%) firm-quarter observations identified as extreme good/bad news with required accounting data on the Quarterly Compustat tape. 14 As expected, in the extreme news sample, the percentage of observations with good news (66.8%) is significantly higher than the percentage with bad news (33.2%). 15 To test whether bad news is incorporated in accounting earnings earlier than good news, we estimate the following regression separately for the good news and bad news samples: Eit = 0 + 1ERit + it where Eit = EPS of quarter t of year minus EPS of quarter t of year -1 for firm i, divided by the security price at the beginning of quarter t.
We require at least three prior years of daily returns data for calculating the mean and standard deviation of 3-day adjusted returns for each firm. We exclude observations where the beginning-of-quarter price is less than 10 cents. Further, to ensure that our results are not affected by outliers, we exclude firm-quarters with price-deflated EPS in the upper or lower 1% of observations.
14 13

(1)

14

ERit = Extreme 3-day market-adjusted return of firm i occurring during the quarter t (as defined in the previous section). 0 is the intercept, 1 is the slope coefficient measuring the impact on accounting earnings of valuerelevant information conveyed by extreme news, and it is the error term. We expect the R2 from regression (1) to be higher for the bad news sample relative to that for the good news sample, indicating that bad news is reflected in accounting earnings on a more timely basis than good news. We use seasonal earnings changes as the dependent variable instead of earnings levels because we examine the incremental earnings effect of the economic content of individual events captured by extreme returns. 16 We also estimate the regression after pooling the good and bad news samples and expect the estimate of the differential coefficient on negative extreme returns, 3, to be positive and significant if bad news is recognized in earnings earlier than good news: Eit = 0 + 1D + 2ERit + 3ERit*D + it where D = 1 if ERit < 0, and zero otherwise. Since events other than the extreme one occurring during the quarter may also impact accounting earnings, the R2 from regression (1) may include the effect of the common variation explained by the extreme event and other events (if the two variables are correlated). Thus, we modify regression (1) and include an additional variable that captures the effect of other events that may occur during the quarter--the market-adjusted returns for the rest of the quarter (that is, excluding the extreme event return). 17 For the good news and the bad news samples separately, we
In comparison, using annual positive versus negative returns, 41% of observations are classified as bad news for the sample period 1963-90 in Basu (1997), and 37% for the sample period 1976-96 in Pope and Walker (1999). 16 Using a similar specification, Givoly, Hayn and Natarajan (2007) simulate the relation between individual events and incremental earnings response versus aggregate events and aggregate earnings response.
17 15

(2)

Naturally, accounting earnings will also reflect other information that is only revealed to the public when earnings are announced. This will not affect our analysis as long as the publicly available news in the fiscal quarter is uncorrelated

15

estimate the regression: Eit = 0 + 1ERit + 2ORit + uit (3)

where ORit = Other returns, or the market-adjusted returns of firm i for quarter t, excluding the extreme event returns, ERit. To estimate the incremental effect of the extreme event returns (or the unique variation explained by the extreme event), we estimate the regression (separately for the good news and the bad news samples): Eit = 0 + 2ORit + vit (4)

and compare the incremental R2 of regression (3) relative to regression (4) for the good news versus the bad news samples. We expect the impact on accounting earnings of extreme bad news as indicated by the incremental R2 to be significantly higher than the impact of extreme good news. To test whether the accounting recognition of good news is delayed until subsequent quarters, we estimate the regression: Eit+1,t+j = 0 + 1ERit + it+1,t+j (1a)

where Eit+1,t+j is change in earnings per share for the period comprising quarters t+1 to t+j (j=1,..,4), deflated by price at the beginning of quarter t. We expect the R2 to be higher for the good news sample relative to the bad news sample. We also estimate the regression after pooling the good and bad news samples and expect the estimate of the differential coefficient on negative extreme returns, 3, to be negative and significant, if more good news relative to bad news gets incorporated in accounting earnings of subsequent quarters: Eit+1,t+j = 0 + 1D + 2 ERit + 3 ERit*D + it+1,t+j Similar to the concurrent-period regressions, we also estimate the regressions (2a)

with other information conveyed through the earnings announcement.

16

Eit+1,t+j = '0 + '1ERit + '2Rit+1,t+j + uit+1,t+j where Rit+1,t+j = Market-adjusted return of fiscal quarters t+1 to t+j for firm i, and Eit+1,t+j = '0 + '2Rit+1,t+j + vit+1,t+j

(3a)

(4a)

The incremental R2 of regression (3a) relative to regression (4a) is expected to be higher for the good news sample compared to that for the bad news sample.

4. Empirical Results 4.1 Descriptive statistics Table 1 reports descriptive statistics for the positive news and negative news samples. The mean extreme 3-day adjusted return is -18.6% for the negative news sample relative to 21.7% for the positive news sample. Correspondingly, the mean change in EPS as a percentage of beginning-ofquarter price is -0.35% for the negative news sample relative to 0.32% for the positive news sample. Extreme returns as well as (price-deflated) change in EPS of the two samples are significantly different using the Fama-MacBeth t-test and Wilcoxon signed rank test over 56 quarters of the sample period. Firm size, measured by market value (price times common shares outstanding) as well as sales, is not significantly different for the two samples.

4.2 Extreme event returns and concurrent earnings changes Table 2 reports the results of the cross-sectional time-series regression (1) for the samples of positive news and negative news firms. The regression results are consistent with the hypothesis that bad news is incorporated in accounting earnings earlier than good news. Comparing the

17

positive and negative news samples, we find that the R2 from the regression of scaled earnings changes of quarter t (Et) on extreme returns of quarter t (ERt) is higher for the negative news sample (1.06%) relative to the positive news sample (0.16%). However, simply comparing the magnitudes of R2s of the two samples does not indicate whether they are (statistically) significantly different, given that the variances of returns and earnings for the positive versus the negative news samples may be asymmetric (see Dietrich, Muller, and Riedl, 2007).18 We resolve this issue by first estimating regression (1) using standardized variables (mean=0 and variance=1). We then conduct an F-test of the ratio of the variance of Y-hat of the negative news sample to the variance of Y-hat of the positive news sample, where Y-hat is the predicted value of scaled earnings changes estimated from the standardized regression (1). Note that the variance of standardized scaled earnings changes (Y) equals one for both samples and hence the ratio of variances of Y-hat is equivalent to the ratio of R2s and has an F distribution. The F-test (in the last row of panel A) indicates that the difference in R2 is significant at the 1% level. 19 The estimate of the slope coefficient, 1, is also higher for the negative news sample (0.036) compared to that of the positive news sample (0.012). The regression of earnings changes on extreme returns estimated by pooling the positive news and negative news samples (regression 2) yields a differential slope coefficient, 3, of 0.024 for the negative news firms, which is significant at the 1% level. Thus, overall, the results based on extreme returns as measures of good/bad news are consistent with the asymmetric timeliness hypothesis.

The discussion in Basu (1999) also points out the difficulty of interpreting ratios of coefficients and R2s in Pope and Walker (1999) absent statistical tests.
19

18

The R2s and significance levels of coefficient estimates of the standardized model are substantially the same as that of the non-standardized model reported in Table 2.

18

4.2.1 Effect of (i) aggregation over shorter windows and (ii) increasing magnitude of extreme news We focus on extreme positive/negative returns over 3-day event windows to enhance the power of our tests to detect asymmetric timeliness as discussed in section 2. Givoly, Hayn and Natarajan (2007) analyze the properties of the differential timeliness coefficient and the ratio of the coefficient on negative versus positive news. They first show that the differential coefficient on negative returns estimated from a firm-specific Basu regression is positive and significant for roughly twice the number of firms using quarterly time-series data relative to the annual time-series. Second, using a simulation procedure, these authors show that aggregation of economic events or shocks (from 1 to 20) monotonically reduces the differential timeliness measure. Third, the differential coefficient is higher in magnitude when either positive or negative news dominates the cumulative news conveyed by multiple events relative to when neither positive nor negative news dominates. Our approach combines the advantages indicated by their findings, in that we focus on short 3-day windows to isolate the economic effect of individual events and we only retain observations which indicate clearly positive or clearly negative news in a given quarter and exclude mixed news observations. Based on our data and analyses, we further demonstrate the effect of shortening the event window over which returns are cumulated, and the effect of identifying more extreme events, on the measure of asymmetric timeliness. In Table 3, we report the following results from estimating regression (2): the coefficient estimate on positive extreme returns (2), the differential coefficient estimate on negative extreme returns (3), and the ratio of the coefficient estimates of negative and positive extreme returns, (2 + 3)/ 2. In panel A, we compare results for the same set of firms

19

when the independent variable is measured as the 3-day market-adjusted return (as reported in Table 2) versus (i) the market-adjusted return cumulated over 15 days (i.e. we extend the return window by 6 days on either side of the 3-day window), and (ii) the market-adjusted return cumulated over the full fiscal quarter. We find that the differential timeliness ratio, (2 + 3)/ 2, decreases

monotonically from 3.0 to 2.4 to 1.6 as the return window is lengthened from 3 days to 15 days to the full quarter. In panel B, we identify different magnitudes of extreme 3-day event returns using cut-offs of 5%, 8%, and 10% (we do not apply the firm-specific cut-off for this test). We find that the differential timeliness ratio, (2 + 3)/2, decreases monotonically from 2.6 to 2.1 to 1.8 as the cutoff for identifying the magnitude of extreme news decreases from 10% to 8% to 5%. Overall, these results show that our focus on extreme returns over short event windows enhances the power of our tests to detect the asymmetric timeliness of negative versus positive news.

4.2.2 Incremental R2 results From Table 4, the incremental R2 of regression (3) relative to regression (4) is higher for the sample of negative news firms (1.65%) relative to the sample of positive news firms (0.49%).20 Thus, even after excluding the effect of the common variation explained by the extreme event and other events occurring during the quarter, the R2 is significantly higher for the negative news sample relative to the positive news sample. To test whether the difference in incremental R2 is statistically

The low explanatory power of quarterly returns-earnings regressions is consistent with previous evidence. For example, Warfield and Wild (1992) report an R2 of 0.4% from a regression of quarterly returns on (deflated) quarterly earnings estimated over the period 1983-86; Freeman and Tse (1992) also report an R2 of 0.4% from a regression of unexpected quarterly returns on (deflated) quarterly unexpected earnings over the period 1984-87. This low explanatory power of quarterly earnings is consistent with the argument in Easton, Harris, and Ohlson (1992) and Dechow (1994) that the timing difference in accounting recognition of value-relevant events is more pronounced in short return intervals.

20

20

significant, we invoke the Frisch-Waugh-Lovell (FWL) theorem that proves the equivalence between the OLS estimator for a sub-vector of parameters and that obtained from OLS on the model after projecting the regressand and the regressors of interest orthogonally from other regressors (see Davidson and Mackinnon, 1993). That is, for the positive news and negative news samples separately, we estimate in the first stage (i) the regression of Eit on ORit and (ii) the regression of ERit on ORit, and then in the second stage we estimate the regression of standardized errors from (i) on standardized errors from (ii). We conduct an F-test of the ratio of the variance of Y-hat of the negative news sample to the variance of Y-hat of the positive news sample, where Y-hat is the predicted value estimated from the second stage regression using standardized variables. Based on the F-test, the difference in incremental R2 of the positive versus the negative news samples is significant at the 1% level.

4.2.3 Robustness tests Table 5 examines the robustness of our results with respect to the time-period examined. We find that in 13 of the 14 sample years, the R2 is higher for the negative news sample relative to the positive news sample. The incremental R2 is likewise higher for the negative news sample in 13 of the 14 sample years. The frequency of years with higher R2 (and incremental R2) for the negative news sample is significant at the 1% level, using a normal approximation of the binomial probabilities and assuming that under the null hypothesis the probability of higher R2 for either sample is 0.5. From the first column of Table 5, the estimate of the differential coefficient from regression 2 (3) is positive and significant in 11 of the 14 sample years (mean=0.025). Further, the Fama-MacBeth t-statistic for 3 estimated from 14 yearly regressions is positive and significant at

21

the 1% level. 21 Overall, the year-by-year analysis suggests that the finding of bad news having a higher impact on concurrent accounting earnings than good news is not driven by certain subperiods. Our definition of extreme event returns includes 3-day adjusted returns relating to single as well as multiple events during a quarter. We analyze these sub-samples separately to examine whether the asymmetric timeliness of negative versus positive news is evident for single as well as for multiple events. Our results show that the negative news sub-sample has a higher R2 and slope coefficient estimate than the positive news sub-sample for both multiple and single event samples (not tabulated). We also examine whether the asymmetric timeliness of negative versus positive news is more pronounced in the fourth quarter relative to interim quarters. In view of the legal liability exposure of auditors, fourth quarter earnings are expected to be more conservatively reported than interim earnings (which are generally unaudited). We find that negative news has a higher correlation with earnings changes than positive news in both interim and fourth quarters (not tabulated). The difference in R2s of the negative news versus the positive news samples is marginally higher in the fourth quarter relative to interim quarters (fourth quarter: 0.82% versus interim quarters: 0.70%). Consistent with Basu (1997) and Basu, Hwang, and Jan (2003), our yearly results indicate that conservatism is higher in periods in which auditors face a high legal liability exposure relative to periods in which their legal liability exposure is relatively low. For the portion of our sample period that overlaps with that of Basu, Hwang, and Jan (2003), we designate the periods 1987, and 1992-98 as low liability periods and 1988-91 as a high liability period. Further, we designate 1999
The coefficient on positive extreme returns (2) is positive and significant for 13 of the 14 sample years (mean = 0.014) and the Fama-MacBeth t-statistic for 2 is also positive and significant at the 1% level.
21

22

as a low liability period (due to the continuing effects of the Securities Litigation Uniform Standards Act passed in 1998), and 2000 as a high liability period due to the effects of the economic downturn and the stock market decline. From Table 5, the mean difference in R2 of the negative versus the positive news samples is -0.47% for 1987 (low liability), 1.2% for 1988-91 (high liability), 0.77% for 1992-99 (low liability), and 2.77% for 2000 (high liability). Similarly, the estimate of the differential coefficient on negative extreme returns is -0.007 for 1987 (low liability), 0.039 for 198891 (high liability), 0.019 for 1992-99 (low liability), and 0.051 for 2000 (high liability). 22

4.2.4 Extreme event returns and changes in cash flows Conservatism is an accrual accounting property and hence should have minimal effect on cash flows. While cash flow from operations are unaffected by operating accruals, anecdotal evidence suggests that managers may have some degree of discretion in the reporting of operating cash flow via the classification of items as operating or investing. 23 Hence, it may be possible to observe asymmetric timing of bad news with respect to operating cash flows similar to accrual accounting earnings albeit to a lesser degree, as documented in Basu (1997). We replicate our Table 2 with change in operating cash flow (CFO) as the dependent variable instead of change in earnings. From Table 6, the R2s from the regression of changes in operating cash flows of quarter t (scaled by Pt-1) on extreme event returns of quarter t are not substantially different for the samples of negative versus positive extreme returns. Also, the differential

Regressions pooled over each of the four sub-periods obtain results that are substantially similar to those indicated by the means of annual regressions over these sub-periods. For example, a firm may treat capital expenditure as a routine expense, which reduces cash flow from operations instead of cash flow from investing activities and leads to early recognition of bad news even in operating cash flows. The reverse occurred in the case of WorldCom where line maintenance expenses were capitalized instead of being
23

22

23

coefficient estimate on negative extreme returns estimated from the pooled regression is insignificant, indicating that the asymmetric timing of recognition of bad news versus good news is not evident in operating cash flows as it is in accrual accounting earnings. Since free cash flows are unaffected by either operating or investing accruals, we should in general not observe any asymmetry in the timing of bad news versus good news in free cash flows (see Ijiri and Nakano, 1989). 24 Results with changes in free cash flow as the dependent variable indicate no asymmetry in the recognition of bad news versus good news in free cash flows (reported in Table 6). Overall, our finding of no asymmetric timeliness with respect to operating as well as free cash flows is consistent with the previous finding that, unlike earnings, cash flows are not negatively skewed in the cross-section (Givoly and Hayn, 2000).

4.3 Extreme event returns and subsequent earnings changes Table 7 presents the results of regression (1a) for the samples of positive and negative news firms and the pooled regression (2a). If negative news is incorporated in accounting earnings earlier than positive news, we expect a higher correlation between subsequent quarters earnings changes and extreme returns of quarter t for positive news firms relative to negative news firms. From Panel A, we find that extreme negative returns of quarter t continue to have higher explanatory power for earnings changes of the first subsequent quarter (t+1) and earnings changes cumulated over two subsequent quarters (t+1 to t+2). On the other hand, from Panel B, the R2s with respect to the period comprising the subsequent three (t+1 to t+3) and subsequent four (t+1 to t+4) quarters are

expensed. 24 These authors implicitly argue that financing accruals are not as conservative as operating and investing accruals because liabilities are not accounted for under a higher of cost or market value rule to parallel the lower of cost or market value rule for assets. We thank an anonymous reviewer for pointing this out.

24

significantly higher for the positive news sample. Similarly, while the estimates of the differential coefficient on negative extreme returns are positive but insignificant in regressions of the subsequent quarter and the subsequent two quarters (Panel A), the estimate is negative when earnings changes are cumulated over the subsequent three quarters and negative and significant when cumulated over the subsequent four quarters (Panel B). Further, when earnings are cumulated over the subsequent three (four) quarters, the R2 is higher for the sample of positive news firms in 7 out of 14 years (10 out of 13 years) and the differential coefficient is negative and significant in 10 out of 14 years (11 out of 13 years). While the number of years in which the sample of positive news firms has a higher R2 is significant when four subsequent quarters are cumulated, the differences in R2s for other accumulation periods are insignificant (using a normal approximation of the binomial probabilities and assuming that under the null hypothesis the probability of higher R2 for either sample is 0.5). Similarly, the FamaMacBeth t-statistic is negative and significant only when earnings changes are cumulated over the subsequent four quarters. Overall, our results in Table 7 are robust with respect to time-periods examined. The results of incremental explanatory power of extreme returns for subsequent earnings changes are reported in Table 8. From Panels A and B, we find that the incremental R2s for extreme negative returns continue to be significantly higher than that for extreme positive returns when earnings changes are cumulated over the subsequent one and two quarters. On the other hand, consistent with Table 7, when earnings changes are cumulated over the subsequent three and four quarters (Panels C and D), the incremental R2s are significantly higher for the sample of positive news firms. Thus, even after excluding the effect of common variation explained by extreme returns

25

and the returns of subsequent quarters, it appears that the impact of good news on accounting earnings is delayed by at least three quarters relative to the impact of bad news. Overall, our results in Tables 7 and 8 document an asymmetric relation between returns and subsequent earnings in the opposite direction to that between returns and concurrent earnings. Thus, we provide consistent evidence in relation to both aspects of the asymmetric timeliness hypothesis.

5. Concluding Remarks In this paper, we test whether bad news is incorporated in accounting earnings on a more timely basis than good news. We use extreme returns to identify good/bad news and find that negative extreme returns have a significantly higher explanatory power for current earnings changes than positive extreme returns. On the other hand, positive extreme returns have a higher correlation with earnings changes cumulated over the subsequent three and four quarters. This is in contrast to prior studies that find that, contrary to expectation, negative returns continue to have a higher correlation with future earnings up to three years ahead (see Pope and Walker, 2000 and Ryan and Zarowin, 2003). Our focus on extreme events is consistent with the argument in Beaver and Ryan (2005), that tests of asymmetric timeliness have more power in samples of large magnitude returns, and enables us to document this aspect of conservatism. If managers strategically disclose information to the market, it is possible that the pattern we observe is merely a result of their news dissemination strategy. Specifically, the pattern we observe may result from managers delaying the announcement of bad news to the market until its disclosure in earnings is imminent (Gigler and Hemmer, 2001), but disclosing good news some periods prior to its recording. This seems a less likely scenario than our stated hypothesis. First, it is not clear why

26

managers would pre-announce good news and then not have their current earnings deliver. In fact, prior empirical evidence (Skinner,1994) suggests that managers announce bad news more promptly to avoid potential litigation. Second, not all news events that arrive in the market are announced by managers and moreover managers may not have the discretion to strategically disclose all news events (for example, acquisitions). Third, we identify events occurring during a fiscal quarter and not over the period from the previous quarters earnings announcement to the current quarters earnings announcement. Hence, the effect of pre-announcements by managers that usually occur after the quarter-end but before the earnings announcement will not be included in our analysis. As discussed earlier, research using contemporaneous positive/negative returns as proxies for good/bad news has encountered difficulties in interpreting results of quarterly analysis. We find strong evidence of asymmetric timeliness in the reporting of quarterly earnings. Further, we find no evidence of asymmetric timeliness in operating as well as free cash flows, consistent with the prior observation that, unlike earnings, operating cash flows do not exhibit negative skewness. Given that our event-based approach has the advantage of providing powerful tests of asymmetric timeliness for both statistical and conceptual reasons, it can be effectively used to address other issues related to accounting conservatism.

27

References Ball, R., Kothari, S.P., Robin, A. 2000. The effect of international institutional factors on properties of accounting earnings. Journal of Accounting and Economics 29, 1-52. Ball, R., Robin, A., and Wu, J. 2003. Incentives versus standards: properties of accounting income in four East Asian countries. Journal of Accounting and Economics 36, 235-270. Ball, R., Shivakumar, L. 2005. Earnings quality in U.K. private firms: Comparative loss recognition timeliness. Journal of Accounting and Economics 39(1), 83-128. Basu, S. 1997. The conservatism principle and the asymmetric timeliness of earnings. Journal of Accounting and Economics 24, 3-37. Basu, S. 1999. Discussion of international differences in the timeliness, conservatism, and classification of earnings. Journal of Accounting Research 37, 89-99. Basu, S., Hwang, L., Jan, C. 2002. Differences in conservatism between big eight and non-big eight auditors. Working paper, Emory University. Basu, S., Hwang, L., Jan, C. 2003. Auditor conservatism and quarterly earnings. Working paper, Emory University. Beaver, W.H. 1993. Conservatism. Working paper, Stanford University (presented at the American Accounting Association national meeting, San Francisco, CA). Beaver, W.H., Ryan, S. 2005. Conditional and unconditional conservatism: concepts and modeling. Review of Accounting Studies 10, 269-309. Bernard, V., Thomas, J. 1990. Evidence that stock prices do not fully reflect the implications of current earnings for future earnings. Journal of Accounting and Economics 13, 305-340. Bowen, R., DuCharme, L., Shores, D. 1995. Stakeholders implicit claims and accounting method choice. Journal of Accounting and Economics 20, 255-295. Boynton, W., Johnson, R., and Kell, W. 2001. Modern Auditing. Seventh Edition. John Wiley & Sons, New York. Davidson, R. and Mackinnon, J. 1993. Estimation and inference in econometrics. Oxford

28

University, New York. Dechow, P. 1994. Accounting earnings and cash flows as measures of firm performance: The role of accounting accruals. Journal of Accounting and Economics 18, 3-42. Dietrich, R., Muller, K.A., Riedl, E.J. 2007. Using stock returns to determine bad versus good news to examine the conservatism of accounting earnings. Review of Accounting Studies12(1): 95-124. Easton, P., Harris, T., Ohlson, J. 1992. Aggregate accounting earnings can explain most of security returns: The case of long return intervals. Journal of Accounting and Economics 15, 119-42. Freeman, R.N., Tse, S.Y. 1992. A nonlinear model of security price responses to unexpected earnings. Journal of Accounting Research 30 (2), 185-209. Gigler, F., Hemmer, T. 2001. Conservatism, optimal disclosure policy, and the timeliness of financial reports. The Accounting Review 76 (4), 471-493. Giner, B., W. Rees. 2001. On the asymmetric recognition of good and bad news in France, Germany and the United Kingdom. Journal of Business Finance and Accounting 28, 1285-1331. Givoly, D., Hayn, C. 2000. The changing time-series properties of earnings, cash flows and accruals: Has financial reporting become more conservative? Journal of Accounting and Economics 29, 287-320. Givoly, D., Hayn, C., Natarajan, A. 2007. Measuring reporting conservatism. The Accounting Review 82 (1): 65-106. Holthausen, R., Watts, R. 2001. The relevance of the value-relevance literature for financial accounting standard setting. Journal of Accounting and Economics 31, 3-75. Ijiri, Y., Nakano, I. 1989. Generalizations of cost-or-market valuation. Accounting Horizons 3(3): 111. Kieso, D.E., Weygandt, J.J., Warfield, T.D. 2007. Intermediate Accounting. Twelfth edition. John Wiley & Sons, New York. Penman, S., Zhang, X. 2002. Accounting conservatism, quality of earnings, and stock returns. The

29

Accounting Review 77, 237-264. Pope, P., Walker, M. 1999. International differences in the timeliness, conservatism, and classification of earnings. Journal of Accounting Research 37, 53-87. Ryan, S., Zarowin, P. 2003. Why has the contemporaneous linear returns-earnings relation declined? The Accounting Review 78 (2): 523-553. Sivakumar, K., Waymire, G. 2003. Enforceable accounting rules and income measurement by early twentieth century railroads. Journal of Accounting Research 41(2): 397-432. Skinner, D.J. 1994. Why firms voluntarily disclose bad news. Journal of Accounting Research 32, 38-60. Stober, T. 1996. Do prices behave as if accounting book values are conservative? Cross-sectional tests of the Feltham-Ohlson [1995] valuation model. Working paper, University of Notre Dame. Warfield, T., Wild, J. 1992. Accounting recognition and the relevance of earnings as an explanatory variable for returns. The Accounting Review 67 (4), 821-842. Watts, R. 1993. A proposal for research on conservatism. Working paper, University of Rochester (presented at the American Accounting Association national meeting, San Francisco, CA) Watts, R. 2003. Conservatism in accounting Part I: explanations and Implications. Accounting Horizons 17, 207-221. Watts, R., Zimmerman, J. 1986. Positive Accounting Theory. Prentice Hall, Upper Saddle River, NJ. Wright, S., and Wright, A. 1997. The effect of industry experience on hypothesis generation and audit-planning decisions. Behavioral Research in Accounting 9, 273-294.

30

Table 1 Descriptive statistics of the samples of positive news and negative news firms over all quarters during the period 19872000.
Positive News (N=41,344) Mean EPS E ER MV ($ millions) Sales ($ millions) 0.012 0.32% 21.74% 1,489 301 Median 0.015 0.25% 15.21% 165 43 SD 0.045 0.049 0.169 1,284 8,844 Mean 0.004 -0.35% -18.64% 1,619 341 Negative News (N=20,523) Median 0.012 0.07% -12.85% 155 50 SD 0.048 0.050 0.146 1,289 9,550 t-test
*

Variables

p-value Wilcoxon** (<0.0001) (<0.0001) (<0.0001) (0.4203) (0.1531)

(<0.0001) (<0.0001) (<0.0001) (0.9640) (0.3139)

Variable definitions: EPS equals quarterly earnings per share after extraordinary items and discontinued operations divided by the beginning-of-quarter price. E equals the change in EPS calculated relative to the same quarter of the previous year and expressed as a percentage of the beginning-of-quarter price. ER is extreme 3-day market-adjusted return occurring during a quarter. Extreme returns are defined (i) by a 10% cutoff applied to all firms, or (ii) by a firmspecific cut-off based on the mean 2.58 times the standard deviation of 3-day adjusted returns for each firm. Positive/negative news is based on the sign of ER. MV is market value of the firm at the end of the quarter. Sales ($millions) refers to quarterly sales in millions. SD denotes standard deviation. *Fama-MacBeth t-test of difference in means of the positive news and the negative news samples over 56 quarters. **Wilcoxon signed rank test of difference in distributions of the positive news and the negative news samples over 56 quarters.

31

Table 2 Effect of negative versus positive news on concurrent-quarter change in earnings: Results of regressions of earnings changes on extreme negative/positive event returns Eit = 0 + 1ERit + it Eit = 0 + 1*D + 2ERit + 3ERit*D + it
Positive News Negative News

(1) (2)
Pooled

N Intercept

41,344 0.001
(1.74)

20,523 0.003
(5.48)**

61,867 0.001
(1.73)

0.002
(3.54)**

ERit

0.012
(14.88)**

0.036
(18.62)**

0.012
(8.16)**

ERit*D

0.024
(8.63)**

Adj-R2 Ratio of R s (p-value)

2 a

0.16%

1.06% (0.0001)

0.88%

t-statistics in parentheses. **Significant at the 1% level. Variable definitions: Eit is change in earnings per share for firm i quarter t (relative to the same quarter of the previous year), deflated by price at the beginning of quarter t. ERit is extreme 3-day market-adjusted return of firm i occurring during quarter t. Extreme returns are defined (i) by a 10% cutoff applied to all firms, or (ii) by a firm-specific cut-off based on the mean 2.58 times the standard deviation of 3-day adjusted returns. Positive /negative news is based on the sign of ERit. D is a dummy variable that takes on a value of one if ERit <0, and zero otherwise.
a

F-test of the ratio of the variance of Y-hat of the negative news sample to the variance of Y-hat of the positive news sample, where Y-hat is the predicted value estimated from standardized models.

32

Table 3
Effect of (i) aggregation over shorter windows and (ii) increasing magnitude of extreme returns on the differential timeliness ratio. Results of pooled regressions of earnings changes on extreme negative/positive event returns: Eit = 0 + 1*D + 2ERit + 3ERit*D + it (2)

Panel A: Effect of varying the return window to capture value-relevant information in extreme events
Quarter (1) Coefficient on Positive News (2) Differential Coefficient (3) Differential Timeliness Ratio [(2+3)/2] 0.0259
(19.98)**

15-day (2) 0.0187

(10.35)**

3-day (3) 0.0120

(8.16)**

0.0165
(7.20)**

0.0261
(7.98)**

0.0240
(8.63)**

1.64

2.39

3.00

Panel B: Effect of varying the magnitude of extreme returns as the measure of good/bad news
|ER| > 5% (1) Coefficient on Positive News (2) Differential Coefficient (3) Differential Timeliness Ratio [(2+3)/2] 0.0147
(4.83)**

|ER| > 8% (2) 0.0111

(5.88)**

|ER| > 10% (3) 0.0096

(4.98)**

0.0118
(3.10)**

0.0128
(4.51)**

0.0152
(4.69)**

1.80

2.15

2.58

In Panel A, reported coefficients for all return windows are estimated from the pooled regression (2) for the samples of positive and negative news firms that are identified on the basis of the 3-day event return window. While negative and positive news firms are identified based on the initial 3-day return criteria (as in Table 2), extreme returns are measured for the identified samples as market-adjusted returns cumulated over the entire fiscal quarter in column (1), and, in column (2), over a 15-day window by adding 6 days on either side of the 3-day event window. Column (3) presents the same results as reported in Table 2. In Panel B, positive and negative news firms are identified on the basis of varying extreme return cut-offs (5%, 8%, and 10%) applied to all firms (we do not apply the firm-specific cut-off for this test). Other variables are defined in Table 2. t-statistics are reported in parentheses. ** denotes significance at the 1% level.

33

Table 4 Incremental explanatory power of extreme event returns for earnings changes Eit = 0 + 1ORit + 2 ERit + uit Eit = 0 + 1 ORit + vit
Positive News (N=41,344)

(3) (4)
Negative News (N=20,523)

Intercept

0.005
(19.05)**

0.001
(3.11)**

-0.004
(-12.20)**

0.004
(6.86)**

ERit

0.022
(14.26)**

0.046
(18.62)**

ORit

0.023
(15.08)**

0.031
(19.08)**

0.028
(12.10)**

0.039
(16.48)**

Adj -R2 Incremental R2 F-test (p-value)a Ratio of Incremental R2s (p-value)b

0.54%

1.03% 0.49% (0.0001)

0.70%

2.35% 1.65% (0.0001) (0.0001)

t-statistics in parentheses. **Significant at the 1% level. ORit is market-adjusted returns of firm i for quarter t, excluding the extreme event return, ERit. Other variables are defined in Table 2. Incremental R2 refers to the difference in R2 of regression (3) and regression (4) for each sample. Test of significance of incremental R2 of unrestricted regression (3) relative to restricted regression (4). Test of the ratio of incremental R2s of the negative news and the positive news samples. As explained in the text, we apply the Frisch-Waugh-Lovell (FWL) theorem (Davidson and Mackinnon, 1993) to obtain residuals used in the secondstage standardized regression and conduct an F-test of the ratio of the variance of Y-hat of the negative news sample to the variance of Y-hat of the positive news sample.
a b

34

Table 5 Effect of negative versus positive news on concurrent-quarter earnings changes for the period 1987-2000: Year by year results Eit = 0 + 1ERit + it (1) Eit = 0 + 1*D + 2ERit + 3ERit *D + it (2) Eit = 0 + 1ORit + 2 ERit + uit (3) Eit = 0 + 1 ORit + vit (4)
Regression (1) R2 Incremental R2 Differenceb [(3)-(2)] -0.47% 1.16%** 1.54%** 1.06%** 1.04%** 0.91%** 0.05%** 0.80%** 0.14%** 1.25%** 0.82%** 1.77%** 0.38%** 2.77%** Positive Newsc (5) 1.22%** 0.58%** 1.04%** 0.60%** 1.17%** 1.50%** 0.43%** 0.17%* 0.27%** 0.61%** 0.75%** 0.12%* 0.34% 0.00% Negative Newsc (6) 1.04%** 2.07%** 2.65%** 2.15%** 2.44%** 2.15%** 0.49%* 1.30%** 0.42%* 2.23%** 1.91%** 2.55%** 0.86%** 3.35%** Differenced [(6)-(5)] -0.18% 1.49%** 1.61%** 1.55%** 1.27%** 0.65%** 0.06%** 1.13%** 0.15%** 1.62%** 1.16%** 2.43%** 0.52%** 3.35%**

Year

Differential Coefficient

Positive Newsa (2) 0.63%* 0.13%* 0.41%** 0.23%* 0.36%** 0.59%** 0.09% 0.04% 0.02% 0.10%* 0.30%** -0.02% 0.11%* -0.03%

Negative Newsa (3) 0.16% 1.29%** 1.95%** 1.29%** 1.40%** 1.50%** 0.14% 0.84%** 0.16% 1.35%** 1.12%** 1.75%** 0.49%** 2.74%**

(1)
1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 -0.007 0.041* 0.048** 0.030** 0.035** 0.027** 0.007 0.028* 0.011 0.025** 0.014* 0.029** 0.014* 0.051**

**, * Significant at the 1% and 5% levels respectively. Column (1) presents the differential coefficient estimate, 3, estimated year-wise from regression (2). Columns (2) and (3) present the R2s of regression (1) estimated year-wise for the positive and negative news samples, respectively. Columns (5) and (6) present incremental R2s (i.e. the difference in R2s of regression 3 and regression 4) estimated yearwise for the positive and negative news samples, respectively. Variables are defined in Tables 2 and 4.
a

Significance based on F-test of regression (1) for each sample. Significance based on F-test of the ratio of the variance of Y-hat of the negative news sample to the variance of Y-hat of the positive news sample, where Y-hat is the predicted value estimated from standardized models. c Significance based on F-test of the incremental R2 of unrestricted regression (3) relative to restricted regression (4). d Significance based on test of the ratio of incremental R2s of the negative news and the positive news samples.
b

35

Table 6 Effect of negative versus positive news on change in cash flow from operations (CFO) and free cash flow (FCF): Results of regressions of changes in the cash flow variable on extreme negative/positive event returns CFit = 0 + 1ERit + it CFit = 0 + 1*D + 2ERit + 3ERit*D + it

Dependent variable = CFOit Positive News (1) N Intercept 29,530 0.003

(2.83)**

Dependent variable = FCFit Positive News (4) 22,221 -0.009

(-1.17)

Negative News (2) 15,550 0.003

(1.85)

Pooled (3) 45,030 0.003

(2.87)**

Negative News (5) 11,790 0.031

(2.13)*

Pooled (6) 34,011 -0.009

(-1.02)

0.0001
(0.23)

0.040
(2.68)**

ERit

0.014
(3.71)**

0.011
(1.87)

0.014
(3.76)**

0.080
(2.86)**

0.058
(0.93)

0.080
(2.48)*

ERit*D

-0.003
(-0.44)

-0.021
(-0.35)

Adj-R2

0.04%

0.02%

0.09%

0.03%

0.001%

0.01%

**, * Significant at the 1% and 5% levels respectively. Variable definitions: CFit equals CFOit in columns 1-3 and FCFit in columns 4-6. CFOit equals the cash flow from operations for firm i quarter t. FCFit is the free cash flow for firm i quarter t measured as CFO plus cash flow from investing activities plus after-tax interest expense (where the tax rate is assumed to be 38%). CFO (FCF) equals the change in CFO (FCF) calculated relative to the same quarter of the previous year and expressed as a percentage of the beginning-of-quarter price. Other variables are defined in Table 2.

36

Table 7 Effect of positive news versus negative news on earnings changes of subsequent quarters Eit+1,t+j = '0 + '1ERit + it+1,t+j Eit+1,t+j = 0 + 1D + 2 ERit + 3 ERit*D + it+1,t+j (j=1,.., 4) (j=1,.., 4) (1a) (2a)

Panel A: Dependent variable is change in earnings of the subsequent (Eit+1) and subsequent 2 quarters (Eit+1,t+2)
Dependent variable: Positive News 39,646 0.0007
(1.71)

Eit+1 Negative News 20,131 -0.0007

(-1.22)

Eit+1,t+2 Pooled 59,777 0.0007

(1.72) -0.0014 (-1.98)*

N Intercept

Positive News 37,793 0.0008

(1.16)

Negative News 19,390 -0.0003

(-0.30)

Pooled 57,183 0.0008

(1.17) -0.0012 (-0.92)

ERit

0.0098
(6.72)**

0.0126
(5.36)**

0.0098
(6.74)**

0.0259
(9.82)**

0.0306
(7.34)**

0.0259
(9.91)**

ERit*D

0.0028
(1.01)

0.0047
(0.95)

Adj-R2 Ratio of R2 (p-value)a

0.11%

0.14% (0.0001)

0.44%

0.25%

0.27% (0.0001)

0.73%

Panel B: Dependent variable is change in earnings of subsequent 3 quarters (Eit+1,t+3) and 4 quarters (Eit+1,t+4)
Dependent variable: Positive News 36,228 0.0002
(0.19)

Eit+1,t+3 Negative News 18,678 -0.0016

(-1.21)

Eit+1,t+4 Pooled 54,906 0.0002

(0.19) -0.0018 (-1.07)

N Intercept

Positive News 34,714 0.0015

(1.16)

Negative News 17,789 -0.0016

(-0.91)

Pooled
52,503 0.0015
(0.19) -0.0031 (-1.40)

ERit

0.0385
(10.41)**

0.0275
(4.80)**

0.0385
(10.55)** -0.0110 (-1.58)

0.0607
(12.41)**

0.0383
(5.03)**

0.0607
(12.65)** -0.0224 (-2.42)*

ERit*D

Adj-R2 Ratio of R2 (p-value)a

0.30%

0.12% (0.0001)

0.62%

0.44%

0.14% (0.0001)

0.89%

37

Table 7 continued Variable definitions: Eit+1,t+j is the change in earnings per share for quarters t+1 to t+ j (j=1,..,4), deflated by price at the beginning of quarter t. Change in earnings per share is calculated relative to the same quarter of previous year. ERit is extreme 3-day market-adjusted return of firm i occurring during quarter t. Extreme returns are defined (i) by an arbitrary 10% cutoff applied to all firms, or (ii) by a firm-specific cut-off based on the mean 2.58 times the standard deviation of 3-day adjusted returns. Positive/negative news is based on the sign of ERit. D is a dummy variable that takes on a value of one if ERit <0, and zero otherwise. t-statistics are reported in parentheses. **, * denote significance at the 1% and 5% levels respectively. a F-test of the ratio of the variance of Y-hat of the negative news sample to the variance of Y-hat of the positive news sample, where Y-hat is the predicted value estimated from standardized models.

38

Table 8 Incremental explanatory power of extreme event returns for earnings changes of subsequent quarters: Positive versus negative news Eit+1,t+j = '0 + '1ERit + '2Rit+1,t+j + uit+1,t+j Eit+1,t+j = '0 + '2Rit+1,t+j + vit+1,t+j (j=1,..,4) (j=1,..,4) (3a) (4a)

Panel A: Dependent variable is change in earnings of subsequent quarter (Eit+1)

Positive News (4a) (3a) 39,646 39,646 0.0028
(11.67)**

Regression N intercept

Negative News (4a) (3a) 20,131 20,131 -0.0028

(-8.26)**

0.0009
(2.42)**

-0.0005
(-0.91)

ERit

0.0087
(5.98)**

0.0123
(5.25)**

Rit+1

0.0235
(23.25)**

0.0233
(23.04)**

0.0231
(16.59)**

0.0231
(16.55)**

Adj-R2 Incremental R2 F-test (p-value)a Ratio of Incremental R2 (p-value)b

1.34%

1.43% 0.09% (0.0001)

1.34%

1.47% 0.13% (0.0001) (0.0001)

Panel B: Dependent variable is change in earnings of subsequent 2 quarters (Eit+1,t+2)

Positive News (4a) (3a) 37,793 37,793 0.0067
(15.63)**

Regression N Intercept

Negative News (4a) (3a) 19,390 19,390 -0.0048

(-8.32)**

0.0018
(2.65)**

0.0003
(0.33)

ERit

0.0227
(8.72)**

0.0281
(6.84)**

Rit+1,t+2

0.0380
(33.56)**

0.0376
(33.25)**

0.0402
(24.81)**

0.0399
(24.66)**

Adj-R2 Incremental R2 F-test (p-value)a Ratio of Incremental R2 (p-value)b

2.89%

3.08% 0.19% (0.0001)

3.07%

3.30% 0.23% (0.0001) (0.0001)

39

Table 8 continued Panel C: Dependent variable is change in earnings of subsequent 3 quarters (Eit+1,t+3)
Positive News (4a) (3a) 36,228 36,228 0.0090
(15.09)**

Regression N Intercept

Negative News (4a) (3a) 18,678 18,678 -0.0048

(-6.06)**

0.0020
(-2.07)*

0.0000
(-0.03)

ERit

0.0328
(9.05)**

0.0263
(4.66)**

Rit+1,t+3

0.0466
(39.26)**

0.0462
(38.90)**

0.0399
(25.45)**

0.0398
(25.43)**

Adj-R2 Incremental R2 F-test (p-value)a Ratio of Incremental R2 (p-value)b

4.08%

4.29% 0.21% (0.0001)

3.35%

3.45% 0.10% (0.0001) (0.0001)

Panel D: Dependent variable is change in earnings of subsequent 4 quarters (Eit+1,t+4)

Positive News (4a) (3a) 34,714 34,714 0.0154
(19.85)**

Regression N Intercept

Negative News (4a) (3a) 17,789 17,789 -0.0050

(-4.91)**

0.0041
(3.21)**

0.0019
(1.15)

ERit

0.0538
(11.30)**

0.0385
(5.19)**

Rit+1,t+4

0.0567
(44.64)**

0.0562
(44.30)**

0.0547
(31.94)**

0.0547
(31.96)**

Adj-R2 Incremental R2 F-test (p-value)a Ratio of Incremental R2 (p-value)b

5.43%

5.77% 0.34% (0.0001)

5.42%

5.56% 0.14% (0.0001) (0.0001)

40

Table 8 continued Variable definitions: Rit+1,t+j = Market-adjusted return of fiscal quarters t+1 to t+j for firm i. Other variables as defined in Table 7. Incremental R2 refers to the difference in R2 of regression (3a) and regression (4a) for each sample. t-statistics are reported in parentheses. **, * denote significance at the 1% and 5% levels, respectively. Significance based on F-test of incremental R2 of unrestricted regression (3a) relative to restricted regression (4a). Significance based on test of difference in incremental R2 of the positive news versus the negative news samples. We apply the Frisch-Waugh-Lovell (FWL) theorem to obtain residuals used in the second-stage standardized regression and conduct an F-test of the ratio of the variance of Y-hat of the negative news sample to the variance of Y-hat of the positive news sample.
a b

41