Capital Markets Hypotheses and Samples description

949 proxy for differences in opinion amongst investors. In a similar vein, Bryan and Tiras 2007 provide evidence that analyst forecasts are a proxy for other information and that higher level of dispersion indicates high information asymmetry. Zhang 2006 also uses analyst forecast dispersion as a measure of information uncertainty. Gebhardt, Lee and Swaminathan 2001 on the other hand, uses variance of analysts forecast as risk. These characterizations are not without problems and possible contradictions. For example, Diether, Malloy and Scherbina 2002 show that higher analysts forecast variance generate a higher level of stock return, thus contradicting the risk interpretation of the measure. Evidence consistent with the interpretation of informational uncertainty has been provided by many studies such as Diether et al., 2002 and Han and Manry 2000 who document that firms with higher variance of analysts‘ earnings forecast have relatively lower future stock returns and ROE. In another important paper, Easterwood and Nutt 1999 document that either extreme good news or extreme bad news is associated with increased uncertainty about earnings. Consistent with Easterwood and Nutt 1999, Gu and Xue 2007 document that variances of analysts forecasts are indeed higher for extreme good news and bad news, thus confirming the characterization of informational uncertainty. However, no direct evidence of informational uncertainty has been documented so far. Theoretical models strive for rational explanations Trueman 1990; Verrecchia 1983; Dye 1985 and empirical studies provide documentation Chambers and Penman 1984; Givoly and Palmon 1982; Kothari et al, 2007 that good news about corporate performance comes out earlier than the bad news. In the context of annual earnings, the implications of later arrival of bad 950 news could be that of higher informational uncertainty, increased private information acquisition by individual investors and financial analysts, and increased variance in analysts‘ forecasts of earnings for bad news. In this study, we investigate and offer evi dence that the variances of analysts‘ forecasts are systematically different in Good News and Bad News environment and how the extent of good and bad news may affect such varinces 21 Following the measures developed by Barron et. al. 1998 we also offers evidence that such differential in variance can be explained by the underlying information environment of public vs. private information of analysts. We also examine if the variance of analysts‘ forecasts differ systematically during the fiscal year for good-news and bad-news environments and for the different amount of good or bad news. Finally we also examine the economic significance of such differential variances in connection to earnings forecasts and future stock returns. More specifically, this study examines the level and change in the variance of analysts‘ forecasts over the fiscal year under good-news and bad- news environments preceding earnings announcement. It uses four definitions of good-news and bad-news events – two ex ante definitions and two ex post definitions. For each definition, the variance of analysts‘ forecasts of earnings is compared at various points of time prior to the annual earnings announcement to answer the following four research questions: First, is the variance of an alysts‘ earnings forecasts significantly different for forecasts associated with good and bad earnings news, and if so, how does it differ 21 Payne and Robb 2000 have used a reverse argument that lower variance of analyst forecast would be a motivating factor for managers to expend extra effort to meet market expectations. Hence a ―good news‖ by way or meet or beat earnings forecast is more likely to follow a lower forecast variance. 951 across different levels of good and bad news? Second, does the differential in variances can be explained by the amount of private v. public information with analysts? Third, does the variance of analysts‘ forecasts changes differentially and predictably over the fiscal year in good-news and bad-news environments? Fourth, , does the variance of financial analysts‘ earnings forecasts explain cross-sectional variation in the level of earnings based on expected news good-news or bad-news in earnings? Finally, do arbitrage portfolios based on the analysts‘ forecast variance and the nature of expected underlying news generate positive abnormal stock returns? In this final test we confirm the results of Diether, Malloy and Scherbina 2002 and refine their results to show that the interaction of the nature of news and the forecast variance generate an even higher level of stock return. Our empirical findings are that after controlling for firm, year, and pre- earnings announcement period quarter, the variance of analysts‘ earnings forecasts is smaller when there is good-news about earnings. We also confirm the Gu and Xue 2007 results that variances of analysts forecast is greater when the underlying news is extreme. We show that after controlling for the level of good news and bad news, analysts forecast variance is higher for the bad news variance. Further, using the measure developed by Barron et. al. 1998 we show that a the amount of private information of analysts are always higher in a bad-news environment and b the amount of private information is higher when the news is extreme. Thus, higher variance is always associated with the higher private information of the analysts. In addition, the variance in analysts‘ annual earnings forecasts decreases over time, in part, because of interim disclosure of mandatory information. The 952 evidence further suggests that the reduction in variance is earlier in the good- news environment, consistent with the view that good news disclosures take place earlier and relatively less information is systematically produced for bad- news announcements. Consistent with theory, we show that the private information of analysts also decreases over time. At an aggregate level, reported earnings are negatively associated with the variance of the analysts‘ earnings forecasts, and firms with higher variance have lower earnings and bad news. Thus, if a higher variance is caused by either the nature of the news good v. bad or the size of the news extreme v. small our result indicate that on the average, the effect of the nature of the bad news, dominate that of the size of the good news. , As a result, the market consensus forecast is smaller than the mean of analysts‘ earnings forecasts when analysts‘ forecasts variance is higher. We suggest the use of different adjustment factors in different quarters for forecasts made in good and bad- news environments. Finally, our study documents that arbitrage portfolios constructed at the end of each quarter on the basis of the variance of analysts‘ forecasts and the nature of the underlying news good-news or bad- news generate year-end abnormal returns in the range of 6-10. The issue of information dissemination is important for capital market researchers. The predictive power of the analysts‘ earnings forecast variance, if any, of the future earnings news at announcements is important for investors and other market participants alike. If information in the variance of analysts‘ earnings forecasts improves the prediction of earnings, then the market‘s expectation of earnings should assign a weight to this measure. To the extent the market fails to do so, it provides an opportunity to earn 953 abnormal returns around future earnings announcements and to develop better earnings forecasts. The finding of this study, that the relative disclosure timing of good-news and bad-news Kothari et al., 2007 has differential implication on the variance of analysts‘ forecasts, make several important contributions to the literature. No prior study has examined and documented that both a the variance of analysts‘ earnings forecasts, and b the change in t he variance of analysts‘ earnings forecasts over the fiscal year, vary systematically for good and bad-news firms and explain the variation through the extent of private v. public information possessed by the analysts. The finding that the variance of ana lysts‘ earnings forecasts contains information that can be used to improve the prediction beyond the consensus analysts‘ forecast is also not present in the prior literature and has direct implications for all capital market studies using consensus analyst forecast as market earnings expectation. Finally, though it has been documented that the variance of analysts‘ earnings forecasts can be used to generate subsequent abnormal monthly returns Diether et al., 2002, our results indicate that these abnormal returns are driven by firms with good and not bad-news. Furthermore, by forming portfolios at different points of time, this study documents that the variance of analysts‘ forecast has information that can be used as early as the first quarter of the fiscal year to generate abnormal returns based on expected earnings news. II. HYPOTHESES DEVELOPMENT Information Arrival in Good-News and Bad-News Environments Theoretical arguments support managerial incentives to delay the announcement of bad news. In a model that makes information manipulation 954 costly, Trueman 1990 shows that a strategic manager may delay releasing bad news to receive additional information that could improve his or her ability to judge the costs and benefits of information manipulation. Furthermore, if disclosure leads to proprietary costs, information gets released only if the benefits exceed the costs Dye 1985; Verrecchia 1983. Non-disclosure can represent an extreme form of disclosure delay. Managers also have incentives to build their personal or corporate reputations to create a ―resolution preference‖ that encourages them to hasten the delivery of good news and postpone bad news Hirschleifer 1993. Finally, the manipulation of information is time consuming and bad news is always manipulated to the greatest extent possible before its release, which may lead to the delay. A simple explanation for this intentional delay posits that the passage of time is of value to a manager because interim actions and unforeseen events might ameliorate the consequences of the bad-news event. Overwhelming empirical evidence provides corroborating evidence that good news about corporate earnings comes out early, and bad news comes out later Chambers and Penman 1984; Givoly and Palmon 1982; Kross and Schroeder 1984; McNichols 1988; Mendenhall and Nichols 1988; Patell and Wolfson 1982; Penman 1984; Begley and Fischer 1998; Kothari et al. 2007. Kothari et al. 2007 consider this as a career concern of the manager because such disclosure strategy has the potential to alter management compensation. They rightly observe that in cases of recent scandals managers explicitly withheld bad news from outside investors and such 955 behavior reinforces the belief that managers‘ private incentives significantly influence the characteristics of corporate disclosures. 22 The literature points to other motives for timing disclosures. For instance, managers may delay the release of bad news to benefit from the opportunity to exercise stock options. Several studies Baginski et al., 1994, Yermack, 1997, and Aboody, Barth and Kasznik, 2004 have shown that managers accelerate bad news andor withhold good news in the period immediately preceding the option grant dates to lower the exercise price of the options thus increasing the value of their option grants. Other studies Kasznik and Lev 1995, and Skinner 1994, 1997 show that managers have a propensity to forewarn analysts of losses bad news to prevent increased exposure to litigation risk. This advance notice, however, refers only to a warning prior to the statutory release of the information and does not suggest communication at the inception of the bad-news event or that it would appear earlier if the firm had good news to report. 23 Other explanations for delay include audit complexities associated with bad news, delays by smaller firms that have more bad news on average, and industry practices that cause more bad news during the test period. None of these explanations receive strong empirical support over time e.g., Givoly and Palmon, 1982. To summarize, earlier as well as more recent studies Begley and Fischer, 1998, Kothari, 2007 provide evidence that after considering various 22 A more appropriate but extreme example of the value of the passage of time is the following story retold by Ro Verrecchia Verrecchia 1983. When sentenced to death by the King, a knave offers to teach the King‘s horse how to talk in a year‘s time. He explains later to his friends, ―Within a year anything can happen, the King may die, the horse may die, or the horse may even learn to talk.‖ 23 It seems that under an experimental setting, analysts seem to believe that disclosures are downwards biased and managers have a definite pre-announcement strategy Tan et al. 2002. 956 incentives to manage news releases, managers release good news earlier than bad news. Analysts Incentive Interactions in Good News and Bad News Environments Empirical evidence indicates that financial analyst incentives differ in analyzing firms with good news versus bad news. εcNichols and O‘Brien 1997 show that financial analysts are generally reluctant to be the first bearers of bad news and may choose to delay their information release. In extreme cases of bad news, financial analysts may drop the firm rather than repeatedly release unfavorable news. Womack 1996 and Michaely and Womack 1999 provide similar evidence when analysts face conflicts of interest arising from underwriting relationships. Because most analysts issue buy-side recommendations, they do not want to antagonize their clients by possibly premat urely ―crying wolf‖. Furthermore, firms experiencing poor performance may have an inherently higher level of uncertainty, whose resolution takes longer, which would imply a later release of information. The notion that the nature of the news affects the communication, interaction, and information flows among managers and analysts is well documented in empirical research. Burgstahler and Eames 2006 and Matsumoto 1998 provide evidence that firms manage both earnings upwards and forecasts downwards in their effort to create good news by meeting or beating analysts‘ expectations. They further argue that firms manage their forecasts by deliberately biasing their communications with analysts to lower forecasts. Such forecast management can take many forms, 957 including but not limited to calls to analysts, and public ―pre-announcements‖ of bad news Kasznik and Lev, 1995, Skinner, 1994 1997, Bamber and Cheon, 1998 at least in qualitative terms. Firms releasing good news could be more precise than for bad news. For example, a good-news release could refer to improvements in earnings per share, whereas a bad-news release might offer only general terms Skinner 1994. Brown 2001 contends that managers‘ incentives to manage earnings and analysts differ dramatically when they report losses versus profits. When reporting bad news, managers do not forewarn analysts. Degeorge et al. 1999 demonstrate the decreased increased communication among firms and analysts when firms report losses profits, perhaps to reduce increase the frequency of bad news good news. 24 Taken together, these characteristics suggest significant differences in good and bad news forecast environment, which can influence the properties of analysts‘ forecasts. Impact of Inf ormational Characteristics on the Variance of Analysts‟ Forecasts Empirical studies document that financial analysts revise and update their forecasts in response to quarterly reports Abdel-khalik and Espejo, 1978, management forecasts Jennings, 1987, other analysts‘ forecasts Stickel, 1990, and changes in stock prices Brown et al., 1985. Changes in the frequency and nature of interactions with managers can influence the properties of analyst forecasts. Until recently, analysts routinely conducted informal discussions with managers on a regular basis. However, with the 24 Lev and Penman 1990 find that earnings forecasts, as a rule, differentiate firms with ―good‖ annual earnings from other firms and that, on average, forecast news is ―bad‖. 958 passage of SEC Regulation Fair Disclosure, this communication is not possible anymore. Researchers have used the variance in analysts‘ forecasts to measure earnings predictability and the degree of consensus in the market. Imhoff and Lobo 1992 show that the ex ante uncertainty of earnings predictability, as captured by the variance in analysts forecasts, correlates with the earnings response coefficient. Daley et al. 1988 indicate that the variance in analysts forecasts correlates with the magnitude of unexpected earnings and, by implication, the ex-post variance of realized earnings. They also show that the variance in analysts forecasts captures the ex ante variance of stock prices, similar to Patell and Wolfson 1979. Ziebart 1990 uses the variance measure as a surrogate for market belief consensus, which correlates positively with trading volume. Several studies argue that variance of analysts‘ forecasts capture the nature of information environment that determines forecast accuracy. Brown, Richardson and Schwager 1987 construct their measure of forecast accuracy of analysts by comparing the variances of tome series forecasts and analysts forecasts. Other studies document that the accuracy of analyst forecasts is correlated with the variance coefficient of variationdispersion of the analysts forecasts Elton, Gruber and Gultekin 1981,. Gu and Xue 2007 show that the variance of analysts forecasts is much larger when the news is either extremely good or extremely bad. They show a U-shaped relationship between news and variance, but fail to identify that the U shape is asymmetric with larger steepness on the left bad news than the right good news. A comparison of the relative magnitude of variance under good news and bad 959 news provides additional insights on the role of news on analysts‘ earnings forecast variance that our study seeks to exploit. Definitions of Good News and Bad News Environments Degeorge et al. 1999 consider three earnings thresholds that induce income manipulation to generate good bad news: meet or beat not meet or beat prior performance, report profits losses, and meet or beat not meet or beat analysts‘ expectations. Because our focus is on the properties of analysts‘ forecasts, we could only use the first two measures. We contend that consensus mean forecasts the threshold for meeting and beating expectation and variance of analysts‘ forecasts may not be independent. We define ex post good and bad news as follows: A. A reported increase in earnings from last year is good news, and a reported decrease is bad news. B. A reported profit is good-news, and a reported loss is bad news. We also consider ex ante definitions of good and bad news. Unlike the ex post grouping, which is based on realized earnings, the ex ante grouping depends on whether the expected news about the firm is good or bad. Following our earlier arguments and our definition of good-news and bad- news scenar ios, we expect the variance of analysts‘ forecasts to be smaller in the case of good-news announcements over all, and also to be so after controlling for the level of the good bad news. H1: The variance of analysts‘ forecasts of annual earnings is smaller in good-news environments. Our interpretation of the variance in analysts forecasts, though consistent with the literature, also posits that good and bad news information 960 environments lead to two distinct distributions of analyst‘s forecasts characterized by differential variances. We conclude that bad-news leads to a decrease in the amount of information available to analysts at the time of their forecasts. Thus, the forecast generated by financial analysts in bad-news environments is conditional on relatively less information that is likely to be noisier. The lack of news may also affect the degree of private information available to the analysts and the weights they assign to this information. Following Barron et al. 1998 and Barron et al. 2002, such lack of precision and lack of consensus would lead to a higher variance in analysts‘ forecasts of earnings in bad-news environments. Thus our second hypothesis follows: H2: A smaller larger variance of analysts‘ forecasts of annual earnings is associated with a large proportion of public private information with the analysts. If consensus building takes place over time, both good and bad-news firms should exhibit a steady reduction of forecast variance over time. This could be partly due to the interim release of the mandatory quarterly earnings reports. Blackwell and Dubins 1962 show that opinions about an unknown event tend to converge as the amount of available information increases. If good-news arrives faster than bad-news, the forecast variance under good news should decrease faster over time than for bad news, at least in the early quarters. Kasznik and Lev 1995 point out that later fourth quarter of each year has by far the largest number of analysts‘ forecasts which suggests an increased rate of information arrival. Therefore, information acquisition both private and public is likely to be significantly more in later periods that are closer to the annual earnings announcement. We treat the fourth later 961 quarter information environment differently also because it coincides with the release of annual earnings, which are audited and more closely scrutinized. Accordingly, our hypothesis is: H3: i The variance of analysts‘ forecasts of annual earnings declines over the year; ii the rate of variance reduction is higher in earlier periods for firms with good-news. To further exploit the role of analysts earnings forecast variance, we examine its ability to predict future earnings. Given Hypothesis 1, the analyst earnings forecasts variance should convey information beyond the mean forecast information. Therefore, analysts earnings forecast variance should play a role in predicting earnings. However, with the passage of time and because good news arrives early, the role of variance should diminish for good-news announcements. For bad-news announcements, analysts earnings forecast variance may continue to play a role until such time when the underlying uncertainty is resolved. Ceteris paribus, if higher variances are associated with bad news, then the effect of increased variance on earnings forecasts should be more pronounced for firms with bad-news. H4: For a given level of annual earnings forecast, i actual earnings are negatively associated with t he variance of analysts‘ forecasts, and ii this association is more pronounced in the bad-news environment. Finally, we investigate the possible economic impact of differential analysts earnings forecast variance under the two news environments on stock returns. Though Diether et al. 2002 have documented that portfolios formed on the variance of analysts‘ forecasts can be used to earn positive abnormal returns, we further partition the variance levels by the nature of the expected news, and state Hypothesis 4 as follows: H5: i For any news environment, the buy-and-hold stock returns will be higher for stocks with smaller forecast variance, and ii the largest 962 arbitrage return is between the good news-low variance and the bad news-high variance portfolios.

III. DATA Sample and Variables

We draw our data from three sources: IBES, Compustat, and CRSP, from 1996-2002. We use the IBES detail database for individual analysts forecasts of annual earnings per share EPS. This database provides a the cusip identifier, fiscal year and the reported EPS of the firm, and b analyst‘s identification code and dated forecast. We obtain quarterly and annual earnings announcement dates from the 2006 Quarterly Compustat files. The Compustat and IBES databases are merged to create an initial sample of firms with availability of both i analysts‘ forecasts, and ii quarterly announcement dates. In this initial sample, annual earnings forecasts are made at various times, from the beginning of the fiscal year 25 to the annual earnings announcement date. To ensure that a subset of forecasts are conditional on the same earnings information, we classify all forecasts into one of four pre-earnings announcement periods PEAPs that roughly correspond to the four fiscal quarters. The first PEAP includes only those forecasts made after annual earnings announcement and prior to the announcement of first quarter earnings. When the first quarterly earnings announcement is delayed relative to the prior year‘s corresponding announcement, we delete the forecast to ensure that all forecasts use the previous year‘s annual earnings alone, not the inference associated with any delay in the quarterly earnings 25 Jan 1 st for firms with December ending fiscal year. 963 announcement of the current quarter. In the second third PEAP, we include only those forecasts made after the announcement of first second quarter earnings and prior to either 1 the date of announcement of second third quarter earnings or 2 the expected earnings announcement date determined by the previo us year‘s corresponding quarterly announcement date, whichever is earlier. Finally, the fourth PEAP includes only those forecasts made after the announcement of third quarter earnings and prior to the announcement of annual earnings or the expected earnings announcement date determined by the previous years annual announcement date, whichever is earlier. Thus, the four PEAPs represent four disjointed time periods, with potential discontinuities, between the two annual earnings announcement dates. The sample includes all forecasts for a firm in a given PEAP-year. To avoid dependency induced by an analyst issuing multiple forecasts in a PEAP-year, we select only the last forecast of an analyst in each PEAP-year for a firm. Furthermore, we require at least four distinct analysts‘ forecasts in each PEAP-year. To ensure that there are enough years of data for each firm, we retain only those firms that have at least full 7 years of data over the period examined. Finally, we delete firms in the financial services and utilities services industries because these industries are regulated. These sample selection constraints result in a final sample with 12,368 PEAP-years for 334 firms. We compute the mean, M itp , and variance, V itp, of analysts‘ forecasts for each firm PEAP-year. 964

IV. RESEARCH DESIGN AND EMPIRICAL RESULTS

In this section, we explain and present the multivariate tests of our hypotheses. To ensure that firm-specific factors, such as relative beta risk, size, and industry-specific characteristics, do not drive differences in variances, we include a separate dummy variable for each firm in our specification. By providing this explicit control, we capture firm-specific differences through the dummy variable coefficients. Because we cover a 10- year period, variations in macroeconomic factors could confound our findings. We control for these time-specific effects by including a dummy variable for each year. In addition, a dummy for earnings quarter was included for tests involving differences across PEAPs. Unless stated otherwise, all of the specifications tested share these common control variables. Good News and Bad News Measures We construct indicator variables corresponding to the two ex post and the two ex ante definitions of good-news and bad-news. Using definitions A and B, outlined in Section 2, the ex post indicator variable for good-news, GN, is constructed as follows: A. GN it = 1 if E it ≥ E it-1 If current year earnings otherwise exceed previous year earnings B. GN it = 1 if E it ≥0 If profits are reported in the otherwise current year BN it , is the complementary indicator variable for bad-news announcements. In contrast to ex post definitions, the ex ante definitions proxy for the market‘s expectation of good-news or bad- 965 news during each PEAP. We compute the mean of all analysts‘ forecasts in each PEAP as an estimate of the markets‘ expectation. If the mean of firm i in year t and PEAP p, M itp , equals or exceeds is less than that firms‘ previous years‘ earnings, E it-1 , we infer that the market expects current years‘ earnings to be higher lower than the previous year‘s earnings. If ε itp is equal or greater less than 0, we infer the market expectation to be that the firm will report profits loses. These indicator variables, denoted as C D, provide ex ante version of good news and bad news defined as A and B. C. GN itp = 1 if M itp ≥ E it-1 If current year earnings are otherwise expected to exceed previous year earnings D. GN itp = 1 if M itp ≥0 If the firm expects to report otherwise profits in the current year Under definitions CD, the news expectation can change from PEAP to PEAP. This allows for the same firm to be classified as expecting good-news in one PEAP and bad news in another. These definitions contrast with definitions AB, where a firm is classified into the same news category for all four PEAPs of the year. Descriptive Statistics In Table 1, we present the industry distribution and descriptive statistics for the sample. Panel A provides evidence that our data is well represented and all of the major industries are included in our sample. As per Table 1, Panel B, total assets Compustat item 44 vary from 30,016 million to 527,715 million, consistent with representation of mostly large firms in the 966 sample. Average firm ROA, income before extraordinary items Compustat data item 8 to average total assets, varies from 3.30 to 24.10 percent. The mean median return on equity ROE, computed as the ratio of income before extraordinary items to total stockholders‘ equity Compustat data item 60, is 13.31 13.01. The mean of earnings per share Compustat data item 11 is 1.46 and the maximum EPS is 35.00. Market value of equity is calculated by multiplying number of shares outstanding Compustat data item 15 by the closing price on the 3 rd month of the quarter Compustat data item 14. The mean median market capitalization is 15,731 4,473 million. Table 2 presents the variance of analyst forecasts for the two ex post and the two ex ante definitions of good-news, by year and PEAP. Also presented are the results of the univariate tests of difference in the mean analysts‘ forecast variances under the good-news and the bad-news environments. Under all four definitions A-D, our tests indicate that in all 11 yearly comparisons 100, the variance is smaller for the good-news announcements, with 10 9 out 11 differences statistically significant at 0.01 level or better for definitions A-C D. To ensure that these results are not driven by any particular PEAP, we also test for difference in variances in good-news and bad-news environments across PEAPs. For all four definitions, the variance in good-news environments is significantly smaller than the variance in bad-news environments for each PEAP 1-4. Based on this univariate analysis, the variance of analysts‘ forecasts is significantly smaller for firms expecting good news than for firms expecting bad news for all eleven yearly comparisons and all four PEAP comparisons. 967 Test of Hypothesis 1 Association of nature of news and the forecast variance As per Hypothesis 1, the variance of analyst forecasts should be lower when the current year‘s earnings convey good news. We test this hypothesis by regressing analyst forecast variance on dummy variables corresponding to good and bad news, after controlling for the firm, year and PEAP effects using the dummy variables and suppressing the intercept term as follows:                1 1 1 1 1 1 2 1 itp ~ ~ I i T t P p itp it it p p t t i i BN GN P Y F V       1 where F i , Y t , and P p denote dummy variables corresponding to firms, years, and PEAPs, respectively, and  i ,  t , and  p are the corresponding parameters. I, T and P denote the number of firms, years and PEAPs.  1 and  2 represent average variance for good news and bad news firms. To make comparisons of analysts‘ forecasts variance under good news and bad news, a test of the null hypothesis of the equality of  1 and  2 is conducted against the directional alternate that  1  2 . Estimation results for 1 appear in Table-3. Due to a large number of parameters in the model, we only present the parameters of interest  1 and  2 , but compute F-statistics on the test of equality of the subset of coefficients corresponding to firm, year and PEAP to evaluate the effectiveness of these controls. In each of the estimations of 1 reported in the table, the F-statistics corresponding to the firm, year and PEAP effects were significant, justifying these controls. The results are first presented for the pooled sample, and then separately for each PEAP. In the estimation involving data pooled over all PEAPS for firms reporting increasing profits definition A, the coefficient for the good-news 968 dummy is 0.02 while the coefficient for the bad-news dummy is 0.06. Even though the individual coefficients are not, the difference in these coefficients is significantly different from zero at conventional levels of significance 0.05 or better. Similar findings are obtained when analysis is performed over PEAPS. For PEAP 1 2, 3, 4 the coefficients for the good news and bad-news dummies are 0.03 0.02, 0.007, 0.01 and 0.06 0.05, 0.07, 0.04, respectively. For each PEAP, the magnitude of the parameter estimate of good-news dummy is significantly smaller than that of the bad-news dummy. These results provide supporting evidence for Hypothesis-1. For firms reporting profits, we get similar results and draw the same conclusions. At the pooled level, the coefficient on the good-news dummy is 0.03 which is significantly smaller than the coefficient for the bad-news dummy, 0.14. For PEAP 1 2, 3, 4 the coefficient for the good news dummy 0.03 0.03, 0.02, 0.02 is significantly smaller than the corresponding coefficient for bad-news 0.15 0.16, 0.16, 0.08. For firms expecting higher profits in the current year compared to previous years‘ profits ex ante definition C, the estimated coefficient for the good-news dummy is 0.02 and that for the bad-news dummy is 0.05 for the pooled data. Our test rejects the null hypothesis of equality of these two coefficients at conventional levels of significance 0.05 or better. For PEAP 1 2, 3, 4 the coefficient for the good news dummy, 0.04, 0.02, 0.01, 0.02 is significantly smaller than the corresponding coefficient for the bad-news dummy 0.06 0.05, 0.06, 0.04 p 0.01. Similar results are obtained when we consider firms expected to report profits definition D. For the pooled sample, the coefficient on the good-news dummy is 0.02, significantly smaller 969 than the coefficient for the bad-news dummy that equals 0.14, and we reject the test of equality at conventional levels of significance 0.05 or better. For PEAP 1 2, 3, 4 the coefficient for the good news dummy 0.02 0.03, 0.02, 0.02 is significantly smaller than for the bad-news dummy 0.170.20, 0.18, 0.09 at conventional levels of significance 0.05 or better. Collectively, our results provide evidence that the variance of analysts‘ forecasts is smaller under good news environments. In order to verify whether the size of the news affect our result, we partition our data into four quartiles for each PEAP and run our tests. Though the quartiles with extreme news have a higher variance than those in the middle, we show that for each comparable quartile, the good news variances are lower than those of bad news. discuss the table - - to be written Test of Hypothesis 2 More private public information with larger variances to be written Tests of Hypothesis 3 Variance reduction over time Hypothesis-3 suggests that early consensus building occurs when firms expect good news and the consensus building is delayed for bad news. Given faster resolution of uncertainties for good news firms, we expect the rate of decline in forecast variance for good-news firms to be higher than for bad-news firms. We test this by estimating the following equation: I 1 T 1 itp i i t t g itp 1 it b itp 1 it itp i 1 t 1 V F Y V GN V BN                  2 Estimation of equation 2 involves regression of V itp on V itp-1 for good- and bad-news firms after controlling for firm and year fixed effects. Because the 970 lagged variance is a right hand side variable, equation 1 can be estimated using the data for PEAPS 2-4 only. The model is estimated separately for each PEAP because we are interested in the reduction of variance over-time and our predictions vary over PEAPS. In this specification, we expect both  g and  b , to be less than one, consistent with reduction of variance of forecasts over time. A lower value of the coefficient of prior period variance  lies a faster reduction of variance over time. We expect  b to be larger than  g in the earlier PEAPS and  b to be smaller than  g, in the later PEAPs 26 . Results from the estimation of equation 2 for all four definitions are presented in Table 4. The e stimated coefficient of ρ g varies from 0.37 PEAP 4, definition C to 0.74 PEAP 1, definitions AC, and that of ρ b varies from 0.23 PEAP 4, definition B to 1.04 PEAP 3, definitions AC. All the coefficients are significantly smaller than one at the conventional levels of significance 0.01 or better, except for PEAP 3 under definitions AC. However, this estimate is not significantly larger than one consistent with the view that the variance is not increasing. The conclusion from these results is that the variance either decreases or remains same for each PEAP, for all news types. To test for the relative rates of variance reduction under good news and bad news scenarios, we compare the magnitudes of ρ g and ρ b in each PEAP. For PEAPs two three and four, the respective estimates of ρ g ρ b are 0.74 0.67, 0.41 1.04 and 0.44 0.27 under definition A, 0.61 0.77, 0.59 0.71 and 0.48 0.23 under definition B, 0.74 0.78, 0.43 1.04, and 0.37 0.26 under definition C, and 0.59 0.77, 0.48 0.75 and 0.45 0.24 under 26 This is because the total reduction of variance is expected to be same for all firms, on average. 971 definition D. A pair-wise comparison shows that for all four definitions, the rate of variance reduction is significantly higher for good news firms ρ g ρ b in PEAPs 2 and 3, but the rate of variance reduction is faster for bad news firms ρ b ρ g in PEAP 4. The only exception is PEAP 2 under definition A in which the variance reduction is faster for firms with bad news ρ b ρ g . Collectively, the finding is that variance reduces at a faster rate for firms with good news in the earlier PEAPS. However, in PEAP 4, the variance reduces at a faster rate for firms with bad news because consensus is built amongst analysts about the upcoming bad news about the firm either through the private information acquisition and its dissemination in prices or because delay of news is accepted by all analysts as a precursor to the arrival of bad news, or because of voluntary release by the management. These results provide support for our hypothesis. Tests of Hypothesis 4 Earnings Predictions Hypothesis-4 suggests that actual earnings should be lower when the variance of analysts‘ forecasts is higher. Furthermore, the association between news and variance is stronger for firms with bad news after controlling for analysts‘ forecasts, nature of the expected news and the firm and year effects. We test this hypothesis by regressing earnings on the mean and variance of analysts‘ forecasts of earnings, with a separate parameter for the good-news and bad-news firms as follows: I 1 T 1 it i i t t gp itp it bp itp it gp itp it bp itp it it i 1 t 1 E F Y M GN M BN V GN V BN                   3 Table 5 presents results based on two ex ante definitions CD. The F- statistics for the goodness of fit of the models are significant for all PEAPs. 972 The predictive ability of the models also increases over time as adjusted R 2 increases from PEAP one through PEAP four. This finding is consistent with the notion that arrival of news over time explains larger portion of cross- sectional variation in reported earnings. In t his specification, ρ gp ρ bp are the estimated parameters for the mean forecasts of the firm in PEAP p. An estimate of ρ less greater than one is consistent with optimism pessimism in the forecasts, and ρ=1 is reflects no bias in the forecasts. 27 For PEAPs one, two, three and four, the respective estimates of ρ gp ρ bp are 0.93 0.89, 1.02 0.97, 1.03 1.04, and 0.97 0.96 under definition C, and 1.02 0.76, 1.02 1.03, 1.00 1.03 and 1.01 0.91 under definition D. In each of these pairs, the estimate of ρ gp is closer to one than the estimate of ρ bp , consistent with smaller forecast errors for good news firms in each PEAP for both the definitions. Though no clear predictions are made about changes in ρ bp over PEAPs, pessimism will get embedded in forecasts for firms expecting bad news because of late arrival of news. Under both the definitions, there is optimism for firms expecting bad news in PEAP 1. However, with the passage of time, the optimism changes to pessimism. The coefficient of ρ bp gradually increases from 0.89 0.76 to 0.97 1.03 to 1.04 1.03 when moving from PEAP one to two to three under definition C D. For PEAPs one, two, three and four, the respective estimates of τ gp τ bp , the coefficients for the variance of analysts ‘ forecast variable for firms expecting good bad news, are -0.05 -0.09, -0.01 -0.11, -0.04 -0.14, and 0.005 -0.17 under definition C, and -0.08 -0.17, -0.07 -0.13, -0.07 -0.38, and -0.001 -0.12 under definition D. For the bad good news firms, all 7 27 We also estimate a more conventional specification of forecast error model by subtracting the mean forecast from the actual earnings, and regress it on the remaining variables in equation 3. The results not reported provide similar interpretation. 973 out of 8 estimates are negative, with seven four out of eight significantly less than zero. These finding are indicative of smaller earnings when the variance is higher, especially for firms with bad news. A temporal analysis of the role of variance in bad news scenarios indicates that the role of variance keeps getting more prominent as evidenced by the increase in the absolute value of τ bp from -0.09 to -0.11 to -0.14 to -0.17 under definition C with three out of four estimates significantly smaller than zero. Under definition D, all four estimates -0.17, -0.13, -0.38 and -0.12 are negative and significant, consistent with variance playing a role in all time periods. A pair- wise comparison shows that the absolute value of τ gp is always smaller than the absolute value of τ bp with four out eight differences significantly smaller than zero. Collectively, these results point to smaller earnings when the variance is high, especially for firms expecting bad news. These results provide support for our hypothesis that high variance is associated with lower earnings and that variance plays a stronger role for firms with bad news. In sum, we conclude that variance is an important factor in predicting earnings and that there are significant differences in the role of forecast variance in good-news and bad-news scenarios. Test of Hypothesis 5 Portfolio Returns To examine the economic significance of the improvement in earnings forecasts by incorporating forecast variances, arbitrage portfolios are constructed on the basis of the nature of the news and the variance of analysts‘ forecasts. The abnormal returns from these arbitrage portfolios become another basis for evaluating the usefulness of the information in analysts‘ forecast variance. Since our evidence suggests systematic 974 difference of forecast variance as early as the first PEAP, we begin construction of our arbitrage portfolio based on the variance of forecasts at the end of first PEAP and then liquidate our positions at the year end. We perform the same exercise at the end of PEAP 2, 3, and 4. At the end of each PEAP, we form four portfolios based on the quartiles of the distribution of the variances. This strategy captures returns for three two and one quarters because positions created after PEAP 1 2 and 3 have only three 2 and 1 quarters remaining before the year end. For the portfolios constructed at the end of PEAP 4, we hold the portfolio till the end of the following year for four quarters and liquidate our position at the end of the following year. Since the objective is to examine the economic significance of forecast improvement based on the variances, only the ex ante measures are reported. First, we classify firms into good news and bad news based on our ex ante definitions CD and rank them by variances from high to low within each group. This way we create four portfolios: good news high variance, good news low variance, bad news high variance and bad news low variance. After the formation of portfolios, we implement three trading strategies. For the first trading strategy, we buy good news low variance stocks and hold them till the year-end and calculate the average raw returns. We also calculate the average raw returns for a good news high variance portfolio. The difference in raw returns between these two portfolios represents our measure of arbitrage return for this trading strategy. For the second trading strategy, we buy and hold the firms with bad news low variance, calculate the raw return at the end of the year, and subtract from this the returns on the portfolio with bad news high variance. Finally, the third trading strategy 975 creates the arbitrage return measure by subtracting the buy and hold returns of the bad news high variance portfolio from those of the good news low variance portfolio. We use raw returns because the market return factor that would be common for all portfolios would cancel out in our arbitrage returns measure. Table 6 presents the results of the three trading strategies for the two different ex ante definitions of good news and bad news for the four portfolio holding periods. For each trading strategy, there are positive abnormal returns based on the variance as well as the nature of the news. The difference in the low and high variance portfolios with good news is 0.0583 0.0272, 0.0063 and 0.0517 for definition C and 0.055 0.025, 0.0133, and 0.0686 for definition D for the four holding periods. Similarly the difference in the low and high variance portfolios with bad news is -0.0068 -0.005, 0.0534 and 0.14 for definition C and 0.139 -0.0437, 0.0594, and 0.13 for definition D for the four holding periods. Seven one positive differences out of the total eight are significant at conventional levels of significance 0.05 or better for the good bad news firms. None of the negative differences are significant. Thus, variance seems to play a more significant role for firms with good news than those with bad news. We conclude from these results that, in general, high variance portfolios have lower returns than low variance portfolios after controlling for the nature of the expected news. As per Hypothesis 4, the third trading strategy good news low variance versus bad news high variance should generate highest returns. Under definition C, this strategy provides highest returns for three out of the 4 holding periods. The only exception is the full one year return, where other 976 confounding factors could have diluted the effect. However, under definition D, we do not find support for the argument that good news low variance and bad news high variance should generate highest returns. The highest returns are for the good news firms with low and high variance. Overall, our results are consistent with the hypothesis that abnormal returns can be obtained when the variance of analysts‘ forecasts are conditional upon the good-news and bad-news scenarios. These results also provide insights into the findings of Diether et al. 2002 that abnormal returns based on variance are driven mostly by the firms expecting good news.

5. DISCUSSION AND CONCLUDING REMARKS

Consistent with the notion that significantly increased informational uncertainties mark a bad news environment compared with a good news environment, we find a greater variance in analysts‘ forecasts in all bad news environments. Although this finding is consistent with the notion of managers‘ differential incentives to select information release timing, alternative explanations do not enable us to arrive at such a conclusion. Instead, our evidence indicates that such strategic behavior, if any, affects the informational uncertainty of the forecast environment, as well as the forecasts issued during at least the first three quarters of the fiscal year. To the extent our results are driven by management‘s tendency to withhold bad news, our results indicate that even if there were alternative information sources, these alternative sources do not seem to act as credible substitutes for the firm management during this period. We also provide some evidence that differential distributional properties of the forecasts, specifically the conditional 977 variance, under good-news and bad-news environments have incremental predictive power for predicting earnings levels. Our evidence is also consistent with the idea that the distributions of analysts earnings forecasts are significantly different in good news or a bad news environment. This finding has some implication for any study that assumes analysts‘ forecast variances represent random variables from a single population. Our evidence indicates a need to consider forecast variances conditioned on good news and bad news. The existence of arbitrage return for portfolios constructed with analyst forecast variance is consistent with the predictive power of forecast variances of annual earnings. We refrain from making any conclusions regarding market efficiency because these returns do not consider factors such as transaction costs. Future research is needed in these areas.