Blundering Herd - Mark Moore by hedongchenchen

VIEWS: 4 PAGES: 44

									                             The Blundering Herd.


                                   Robert B. Durand*
                                      Mark E. Moore
                                      Adam Sullivan

                          JEL classification: _Behavioral Finance___

               Keywords: _Analyst Forecasts, Overconfidence, Herding____



                                            Version Two




Robert Durand and Adam Sullivan are from the Department of Accounting and Finance, University of
Western Australia, 35 Stirling Highway, Crawley, Western Australia, 6009. Mark Moore is Department of
Finance at the Texas Tech University at Lubbock, Texas. Despite all this help, any errors, omissions and
flawed deliberations remaining in the paper are due to the authors.
*       Corresponding author (US): Mark Moore; Email: mmoore@ba.ttu.edu
        Phone: (806) 742-3298
        Corresponding author (Australia): Robert Durand
          Phone + 61 8 6488 3764. Fax + 61 8 6488 1047.
        E-mail: Robert.Durand@uwa.edu.au
        Web: http://www.ecom.uwa.edu.au/information_about/staff/durand_robert
                            This draft: Wednesday, August 15, 2012




Robert Durand and Adam Sullivan are from the Department of Accounting and Finance, University of
Western Australia, 35 Stirling Highway, Crawley, Western Australia, 6009. Mark Moore is Department of
Finance at the Texas Tech University at Lubbock, Texas. Despite all this help, any errors, omissions and
flawed deliberations remaining in the paper are due to the authors.
*       Corresponding author (US): Mark Moore; Email: mmoore@ba.ttu.edu
        Phone: (806) 742-3298
        Corresponding author (Australia): Robert Durand
          Phone + 61 8 6488 3764. Fax + 61 8 6488 1047.
        E-mail: Robert.Durand@uwa.edu.au
        Web: http://www.ecom.uwa.edu.au/information_about/staff/durand_robert
       The Blundering Herd.




ABSTRACT



[ROBERT TO INSERT ABSTRACT HERE].




                   1
               No home, no rest, no sleep, no content, no life worth the living.

               He must be a lone wolf or he must herd among men obnoxious to

               him.

                                                 Zane Grey, The Lone Star Ranger



Analysts’ zeal to agree, their propensity to “herd”, is puzzling and perhaps, to users of

their services, disquieting. Investors pay for analysts’ research. Many of our most

talented graduates seek careers as analysts. It is reasonable to assume that financial

analysts may engage in behavior that adds to the wealth of the consumers of their

analysis (Womack, 1996).1 Yet analyst’s forecasts are systematically influenced by the

prevailing consensus of other analysts (Welch, 2000). Analysts play follow-the-leader.

It is difficult to believe that analysts are engaging in economically useful activity if they

are simply echoing their competitors’ views.



While there are potentially rational stories that may explain herding, 2 recent analyses

have provided evidence that herding is motivated by quasi-rational influence. Welch

(2000) found that analysts forecasts are positively related to existing forecasts but that, as

this is unaffected by subsequent stock returns,3 herding is not the result of rational and

efficient information aggregation (consistent with Shafstein and Stein, 1990). Cooper,

Day and Lewis (2001) found that more timely analysts have a greater impact on share

prices than those lagging behind them. This is finding is relevant to herding. Leader


1
       Barber, Lehavy, McNichols and Trueman, 2001, provide a counter argument that trading rules
       based on analysts’ recommendations incur punitive trading costs.
2
       For example, Bakchandani and Sharma (2000) point to the potential role of preserving reputation.
       Welch (2000,) provides a summary of pertinent literature.
3
       This would suggest that analysts’ forecasts were, in some way, related to economically valuable
       information.
                                                 2
analysts potentially have no consensus to follow in contrast the analysis laggards.

Cooper et al. also found that, when ranked on timeliness, leaders were more skilled than

laggards. Graham (1999) found that newsletters with a high reputation but low ability (as

measured by forecast accuracy) tend to follow Value Line’s forecasts. Such a findings

suggest that analysts for these newsletters are not fully confident in their skills.



We study the behavior of analysts whose behavior, we believe, characterizes the extremes

of the herd. We examine the forecasting behavior of each “leader” and “laggard” in

every group of seven or more analysts providing quarterly earnings forecasts for a firm

between the first quarter of 1993 to the third quarter of 2002. We imagine an analyst

conga. At the beginning of each quarter, the herd’s leader provides an initial forecast of

the quarter-end earnings4 and each analyst, in turn, follows until the laggard provides her5

initial forecast for the quarter. By focusing on the behavior of the two analysts at the

extremes of timeliness, we believe that we can provide insights into the behavior of the

whole herd.



We find that analyst forecasting behavior is conditioned by behavioral biases moderated

by the difficulty of the task (which, in turn, is effected by the information environment).

Thus, our analysis provides further support for a quasi-rational basis to herding. We also

find that Fair Disclosure (FD) regulations, which aimed to prevent favoritism in

information provision, increased the supply of information to analysts. Given that the



4
       And, potentially, other quarters, or year-end forecasts. We do, not, however, consider those in our
       present study.
5
       Noting the current practice of alternating the gender of the third-person pronoun, we have adopted
       the feminine for the laggard (and the masculine for the leader). Unfortunately, we were not able to
                                                   3
information has been processed by quasi-rational analysts, however, the effects of FD

may not have had their intended effect.



The structure of this paper is as follows. In Section 1 we introduce and motivate the

variables we utilize in our analysis. In Section 2 we present summary statistics for our

data and make some initial comments and in Section 3, we analyze the determinants of

forecast timeliness. We provide further evidence on the interpretation of our proxies in

Section 4 when we discussion the influence of optimism in analysts’ forecasts. We

discuss the determinants of the number of forecasts mad in Section 5 of the paper. In

Section 6 we discuss the determinants of herding [THIS IS THE SECOND

STATISTICAL ANALYSIS WHICH NEEDS TO BE DONE – and will be

completed in the next draft]. Section 7 concludes the paper.



1.     Data and variables



This study employs the IBES Earnings Forecast database6, in particular, the details file,

the database of earnings forecasts and the changes, and the actuals files – a list of actual

reported earnings - using data on analysts’ ‘one quarter ahead’ forecasts from Q1 1993 to

Q3 2002.7 The total sample of per firm quarterly analyses available to us is reported in

Table 1. We study herds of seven or more analysts, approximately 21% of the firm

quarters available to us. While the definition of “herd” is perhaps arbitrary, we believe

that fewer than seven analysts does not facilitate a sufficient number to generate the


       incorporate a gender variable in our analysis. In our case, the choice may well be fortuitous
       [ROBERT TO REFERENCE BARBER AND ODEAN’S BOYS WILL BE BOYS].
6
       Daily data, as required by our study design, is only available from 1993.
                                                4
leader/laggard contrast, and its interplay with group dynamics, we hope to capture and

study.



                               [INSERT TABLE 1 ABOUT HERE]



1.1      “Leaders”, “laggards” and “ghosts”.



From the initial forecast of each analyst, we rank forecasts by timeliness8 and define the

“leader” as the first analyst for firm i in quarter t for whom an end-of-quarter-t forecast is

recorded. The “laggard” is simply the analyst who makes the last recorded forecast. We

often find that leaders are tied, i.e., there are two or more analysts making the first

forecast for a company, and, in these cases we choose the “alpha” leader9 by examining

the previous quarter’s forecasts made by these analysts for the firm in question. The first

of these analysts making a forecast was chosen as the leader for quarter t. If there was

found to be another tie in quarter t-1, or if neither of the “alpha” leader contenders

forecast in quarter t-1, we applied the decision rule to quarter t-2, and, if the same

situation applied, we applied the rule to quarter t-3. If the “alpha” leader could not be

determined by examining quarter t-3 data, the observation was dropped from the sample.

By choosing an alpha leader we aim to capture the analyst who might be thought of as the

natural leader and thus best reflect the propensities we are studying.10


7
         IBES included Canadian dollar forecasts and these are removed leaving a pure U.S. sample.
8
         Defined in Section 1.2.i below. [DOUBLE CHECK REFERENCE BEFORE FINALISING]
9
         At this stage, we mix our metaphors from those of the herd to those of the pack.
10
         Each group of forecasts for firm i in quarter t were considered distinctly. Our study does not
         preclude, nor consider, that leaders for one firm may be laggards for another (and the whole range
         of other permutations of the ordering). We are currently engaged in a longitudinal study of the
         analysts included in our current study to, in part, tease out such effects..
                                                    5
Determining the alpha laggard in the presence of ties involved a different selection rule.

If all candidates for “alpha” laggard forecasted in the previous quarter, the last analyst to

forecast in quarter t-2, was chosen as the laggard for quarter t-1. We found that many

candidate alpha laggards did not forecast in quarter t-1. Perhaps such analysts are not

representative of the “true herd”. Such “ghosts”, as we denoted them, might fall into one

of three categories. Firstly, she may a rookie for the firm: that is, she is forecasting for

that firm (that is, the firm about which the herd is forecasting) for the first time.

Secondly, she might be a “seasonal forecaster”: she may only make quarterly forecasts

for the firm at certain times of the year (for example, she might make a final quarter

forecast in conjunction with an annual estimate). Finally, she might be an opportunist.

Such an opportunist may only make occasional forecasts when it suits her. Where a

candidate laggard is found to be a ghost, she is removed from the analysis and the alpha

laggard is chosen using the same decision-rule as for the alpha leader. We leave the

question of the determinants of such apparitions for further study but we use a

dichotomous variable taking the value of 1 for observations from those firm quarters

where a ghost has been sighted. Use of the ghost variable in our analyses will control for

the possibility that there is something different, or special, about ghostly apparitions.



1.2     Other variables.



1.2.i   Timeliness.

Cooper et al. (2001) found that timely analysts are less likely to herd. They also found

in keeping with Welch (2000), that, when ranked by timeliness, lead analysts have greater
                                              6
impact of prices and volume. The bottom line of Cooper et al. timeliness is positively

associated with skill. We expect timeliness, therefore, to be positively associated with

variables such as experience which is associated with skill as well as variables capturing

the quality of the information set. In addition, we also expect that timeliness will be have

a positive association with confidence

Timeliness, the dependent variable analyzed in Tables 2 to 6 is computed as the number

of calendar days since the date the earnings result from the last period is released, to the

date the leader or laggard’s initial forecast. While firm quarters may begin before this

date, we contend that analysts forecasting quarter t before the results of quarter t-1 results

are known might be thought more to be revising their t-1 forecasts rather than providing

“fresh” quarter t forecasts utilizing a current information set.



1.2.ii Forecast errors, consensus and task complexity



If leaders are more skillful than laggards, it is not unreasonable to assume that their

forecast errors for firm i in quarter t will be more accurate than the corresponding

laggards’ forecasts and, in addition, their forecasts for firm i in quarter t-1 will also be

more accurate. In addition, leaders, if they are relatively more skilled, should be more

accurate than the consensus forecast. Yet skill involves more than success at a given

task. Skill in a task is positively associated with the ability to reflect on the outcomes and

associated process and an ability to purposefully incorporate these reflections towards




                                              7
improving skill at a task (the process of meta-cognition).11 Therefore, if timeliness

equates to skill, we might see evidence of leaders learning from their errors.



Overconfidence, however, involves reducing negative dissonance and increasing positive

assonance. Signals confirming misperceived skills will be psychically “amplified” while

disconfirming signals will be under-weighted. The effects of overconfidence can be

dramatic: for example, Bitmead, Durand and Ng (2004) argue that overconfidence

propelled the Internet stock bubble. Hilary and Menzly (2001) find that accurate analysts

tend subsequently to be poorer than the consensus subsequently suggesting that such

analysts suffer from an attribution bias and become overconfident in either their

information set or their abilities to process it. If leaders are overconfident, we might

expect that they are less affected by errors, and perhaps unaffected, by their or others’

errors in the previous quarters. The contrast in the effects of error between leaders and

laggards should be informative.



Laggards may, in contrast to leaders, be underconfident, Stickel (1990) found that

changes in the consensus forecast is a good predictor of analysts’ earnings estimates

suggesting that, in following the herd, such herding analysts not being confident in their

own abilities.

We utilize each leader and laggard’s forecast error firm i in quarter t and firm i in quarter

t-1 defined as the forecasted earnings minus actual earnings12 as a percentage of actual


11
       The American Idol auditions, where we see so many aspirants’ dreams dashed by Simon’s
       comments, provide a clear example of the association of task and meta-cognitive skills. What is
       truly surprising is that failing aspirant idols have not realized the hopelessness of their quest long
       before their auditions; their subsequent denial what is often so obvious provides an excellent
       example of the strong impetus to reduce cognitive dissonance.
                                                    8
earnings.13 We also incorporate the consensus forecast error (the mean forecast minus

actual earnings).



Confidence, however, is influenced by task complexity. Overconfidence is greater when

tasks and inputs are familiar. It may be that biases may be more apparent for “easier”

rather than “difficult” firms. Additionally, as we discuss below (1.2.vi), there may be a

relationship of forecast optimism and the difficulty of forecasting for a firm (Das, Levine

and Sivaramakrishnan, 1998). Recognizing this, we split our sample into three groups:

high forecast error (FE), middle FE and low FE (in effect, hard, medium and easy) where

each group represents , respectively, the largest 30%, middle 40% and lowest 30% of

firms ranked by average forecast errors14 for each firm i in the quarters in which our

criterion of being followed by seven or more analysts is met.



1.2.iii FD (Fair Disclosure) Dummy.

PostFD is a dichotomous variable taking the value of one if the observed forecast is

recorded after the date Regulation Fair Disclosure was passed (October 23rd, 2000) and

zero otherwise. As we have noted, FD aimed for simultaneous intentional disclosure and

prompt and wide distribution of information if such information has been disclosed

unintentionally with the aim of facilitating information flow and, presumably, its

incorporation into analysis. Mohanaram and Sunder (2002) found that, while aggregate



12
       In many cases the type of earnings (primary or diluted) reported by the company in the actuals file
       differed from that being forecast due to the effect of FASB’s SFAS 128 in 1998 addressing
       changes in the reporting of EPS. In these cases, forecasted earnings were adjusted by the dilution
       factor reported in IBES.
13
       This resulted in a great range of values. Our maximum likelihood estimate of the Tobit
       regressions assumes that the data conform to an extreme value distribution.
                                                   9
forecast error and forecast dispersion have increased post-FD, “All-Star” analysts have

been able to distinguish themselves post-FD. Heflin, Subramanyam and Zhang (2003)

argue that Mohanaram and Sunder’s findings do not hold post-FD, that is, that there is no

change in aggregate error and dispersion as the latter do not control for variations in the

economic climate in which the forecasts are made. That FD has not “improved” forecasts

is noteworthy in itself. Our working hypotheses are that post-FD:

     a.       a more level playing field may reduce any confidence gulf found between

              leaders and laggards;

     b.       improved information flow will increase the timeliness of leaders and

              laggards;

     c.       the interval between leaders’ and laggards’ forecasts will fall;

     d.       forecast frequency will increase;

     e.       any effect of forecast optimism will diminish.



1.2.iv Experience.



Hong, Kubik and Solomon (2000) argue that older analysts are less likely to herd than

younger analysts. While it may be comforting to assume that skill increases, and the

need to herd correspondingly falls, with age, there is conflicting evidence suggesting that

this might not be the case with investment. As investors age, however, they seem to

suffer less from overconfidence and, as such, age may increase herding propensities. Our




14
          Where the average forecast error is the average of the leader and laggard’s forecast for each firm
          in each quarter.
                                                     10
data precludes the inclusion of age as a variable. In any case, age need not correspond

with experience and skill.



Our analysis incorporate experience, the number of days since the leader or laggard

began reporting on the IBES database until the date of the current forecast. With data

only from 1993, this variable is truncated and some observations will be biased

downwards. Our large sample, however, suggests that the impact any such error will be

negligible.



If analysts become more confident with experience, we might expect to find that

behavioral biases associated with overconfidence will increase. Overconfident analysts

might not be affected by the consensus and may be quicker to forecast (and hence be

leaders).     Such analysts would not be affected by the consensus.          If experience is

associated with skill, we should expect this variable to be associated with lower forecast

error. Conversely, inexperienced analysts may suffer from poor meta-cognition: they

may be unskilled and unaware of it. In such a case we might see experience associated

with high forecast error. If so, we should also see overconfident analysts relatively less

affected by their previous forecast error (as overconfident analysts will be slower to

revise their beliefs in the face of disconfirming evidence). Of course, if experienced

analysts are overconfident we might expect them to be unaffected by their previous error;

we assume, however, that skill will ameliorate this affect to some extent.



1.2.v Optimism.



                                            11
Studies have suggested that, on average, analysts tend to make optimistic forecasts (Fried

and Givoly, 1985; Abarbanell, 1991). Das, Levine and Sivaramakrishnan (1998) find

that this is especially the case for harder to predict firm. They argue that companies

reward (or “rewarded” given the implementation of FD) optimistic analysts through and

argue that this is a means by which managers facilitate information transfer (something

FD should have abolished).



We model optimism using a dichotomous variable taking the value 1 if the forecast error

for a quarter was positive and zero otherwise.



2.     Summary Statistics



Summary statistics for leaders and laggards are present in Table 2. Summary statistics

for the differences between leaders and laggards for each firm i in quarter t are presented

in Table 3. In both of these tables, Panel A reports results for the entire sample and

Panels B to D report results for low, middle and high (easy, moderate and hard) FE firms

respectively. While the number of firms in each difficulty cohort fits the 30/40/30 split,

the number of firm quarters in each group do not. For the hardest to forecast firms, fewer

observations are available: there are fewer quarters for firms in this group where the

criterion of being followed by seven or more analysts was met.



                          [INSERT TABLE 3 ABOUT HERE]



                          [INSERT TABLE 3 ABOUT HERE]
                                            12
The summary statistics provide some telling results. Panel A of Table 2 shows that, on

average, leaders overestimate earnings for quarter t after making underestimating

earnings in quarter t-1 although medians suggest that the distribution is skewed. Panels B

to D show that this result is driven by the results for easy and moderate firms while, for

difficult firms, overestimate (positive errors) increase. Laggards, by way of contrast, tend

on average to underestimate earnings in both quarters t and t-1 save for the hardest to

forecast firms where the pattern of their errors is similar to those of the leaders (although

the quarter t errors are much lower).15 The changed sign for leaders’ forecast errors

would be consistent with patterns of overreaction to forecast error. The pattern observed

for the laggards is consistent with conservatism and anchoring. The summary statistics

for the difference in forecast errors presented in Table 3 are consistent with this story:

leaders underestimate earnings in quarter t-1 and follow these with positive errors that are

t-1.



By definition, we find that leaders are more timely than laggards in quarter t. Focusing

on Table 3, leaders forecast almost 54 days before the laggard. Both Tables 2 and 3 show

that the leader also tend to beat the laggard in the previous quarter suggesting we are

capturing some sort of systematic difference. In both quarters t and t-1, forecasts seem to

be provided earlier for harder to forecast firms than easier to forecast firms. This is not

what might be expected should the difficulty of a task make analysis longer. Perhaps the

competition to be first is greater when analysis is more difficult.


15
         The consensus, in keeping with Fried and Givoly (1985 and Abarbanell (1991) mark optimistic
forecasts. It is the extremes who provide the exceptions to the rule.
                                                13
What is most surprising is that leaders seem to be less experienced than laggards: Table 2

shows that leaders have an average of 1,360 days in the database while laggards have

1,383. Table 3 shows that the difference, although less than two months, is significant

(the p-value of the median difference of –46 is less than .0001). The difference is greater,

though not much more so, for those firms which are hardest to forecast (approximately 49

days). If experience equates to skill, as has been proposed by Cooper et al. (2001) and

also in our paper, these initial findings, taken with the pattern of forecast errors discussed

above, suggest that lesser skilled analysts are more prone to overreaction and more

skilled experienced analysts are more prone to anchoring and conservatism. On one

hand, the magnitude of the difference we find in analyst experience would, however,

suggest that gap between leaders and laggards is closed within the space of a couple of

months.    As efficient market theorists propose, the failure in competitive financial

markets is costly and provides painful stimulus to get it right or get out. On the other

hand, a median difference of 46 days might suggest that it may be misleading to strongly

equate this variable with skill (at least as it is commonly understood).



Tables 2 and 3 also shows that leaders make more forecasts than their partner laggards

and Panels B to D suggest that this difference becomes greater as forecasting becomes

more difficult. Perhaps leaders simply have more available time in a quarter to make

more forecasts. If leaders were overconfident, given our discussion in Section 4, we

would expect them to make fewer forecasts, ceteris paribus.



3.     Timeliness
                                             14
The summary statistics presented in the preceding section suggest that leaders and

laggards may be prone to different behavioral biases (overreaction versus conservatism

and anchoring) although the difference in their experience (our proxy for skill) is not as

large as might be expected. Contrary to our expectations that leaders might be more

overconfident, and hence, less likely to revise their forecast, we found evidence that

leaders make more forecasts than laggards although, once again, the difference is not as

great as we might expect.      In Table 4 we present results of Tobit analyses of the

timeliness of first forecasts denoting laggards’ forecasts with a dummy variable taking

the value of one. We focus on Table 4 in our discussion and use Table 5, in which we

present the results of regressions on the paired differences in leader and laggard

timeliness, to highlight important points. In both Table 4 and 5, the first three columns of

results present estimated parameters, summary statistics and p-values for each variable in

the regression for the entire sample. The analysis is repeated for the lowest FE, highest

FE and middle FE cohorts in the remaining columns.



                          [INSERT TABLE 4 ABOUT HERE]



                          [INSERT TABLE 5 ABOUT HERE]



On average, the negative coefficient for experience for the complete sample in Table 3

confirms the picture presented by our summary statistics, the greater an analyst’s

experience, the more timely are his forecasts. The positive coefficients for experience in

Table 4 confirm this story: the greater the difference in experience between leader and
                                            15
laggard, the greater the difference in timeliness is. The findings are consistent with

Cooper et al. (2001) who equate skill with experience and timeliness. The parameter

estimate for experience in Table 4, however, is driven by the low and middle FE (the

easy) firms. For high FE firms, the coefficient is negative. The interaction of experience

and laggard is also instructive, for harder firms, laggards are less timely. While such

findings are prima facie consistent with leader overconfidence for harder to forecast

firms, we perhaps should expect overconfidence to be greatest where tasks are easier

rather than those which are difficult. Conversely, in the case of the extremes of the herd,

overconfidence is manifested only in the analyses which are most difficult; that is,

overconfidence is greatest where meta-cognitive skills are lowest.



We find no relationship between either the analysts own error or the consensus forecast

error in quarter t-1 and timeliness in quarter t.             In Table 4, significantly negative

coefficients in on these variables (and their higher order effects) might suggest a

propensity to pause for reflection and, perhaps, learning while significantly positive

coefficients might suggest recklessness. The heedlessness we find may be consistent

with overconfidence (that is, the failure to take disconfirming signals into account) if the

process of reflection and learning is manifested in temporal delay.16



In Table 4, the estimated parameter for timeliness in the previous quarter is, for the whole

sample, negative (-0.0589) and significant although the result is driven by high FE firms

(the results for middle FE is positive and significant).              Consistent with the picture


16
       Which, in turn, would seem to imply a process of conscious deliberation which is not necessarily a
       function of the cognitive biases we discuss. The biases operate subliminally.
                                                  16
presented by the summary statistics, it appears that tardiness in the previous quarter is

reversed in the following quarter.      The interaction of timeliness(t-1) and laggard,

however, shows that laggards, on average, become significantly less timely in the

following quarter. The parameters for timeliness(t-1) reported in Table 5 indicate that,

overall and for each difficulty cohort, the a positive relationship between the difference in

timeliness between leader and laggard in quarter t. The significance and sign of the

interaction term is difficult to reconcile with stories linking timeliness to skills (as,

indeed, is the 46 day difference between leaders and laggards discussed in Section 2).

Leaders becoming quicker in the tasks at which they are less adept suggests

overconfidence rather than skill.



Interactions of variables with the Post FD Dummy are instructive. Regulation FD seems

to have resulted in more timely forecasts and the effect is concentrated in high FE firms.

Our findings initially suggest that analysts covering harder to analyze firms are the ones

to have benefited from the potentially richer post-FD information environment. The

positive and statistically significant coefficient on the interaction of Post FD and

experience suggests, however, that increased information is making inexperience (and,

given the indications in our preceding analysis, more overconfident) analysts more

timely. The further interaction of the laggard and post-FD dummies suggests, however,

that laggards have also benefited for easy and middle FE firms while they have become

more tardy for High FE firms.



Our findings suggest that timeliness, the feature by which we differentiate leaders and

laggards, the extremes of our herd, is less a function of skill and more a function of
                                             17
confidence. Such confidence is greatest where meta-cognitive ability is lowest (high FE

firms).



By way of a digression, it is also worth considering the impact of the Ghost Dummy.

Table 4 indicates that ghost apparitions correspond to slower leader forecasts but more

timely laggard forecasts. Table 5 confirms that the difference between leaders and

laggards is narrowed when ghosts are active. Post FD, the gap between leaders and

laggards is still narrower. Thus, if it is accepted that the leaders are more timely than

laggards due to their confidence and the information environment, the findings suggest

that apparitions occur when either the factors determining differing levels of confidence

are lowest, information differences are lowest or if both occur. We leave this to further

research to explore.



4.        Optimism



Table 4 also reveals negative and statistically significant coefficients for optimistic

forecasts in the previous quarter and timeliness. The effect is greatest for high FE firms.

This might be expected if Das et al. (1998) are correct. If analysts are rewarded for

optimism in quarter t-1 with information in time t, we might expect more timely forecasts

by analysts having that information.     In contrast, when optimism is interacted with

laggard, it appears that laggards become commensurately later for high FE firms

(suggesting that it is only leaders who may be rewarded in the way Das et al. argue).

Interactions with the Post FD dummy suggest that FD has exacerbated the first-order

effect optimism for low and middle FE firm but, for high FE firms, FD has reversed this
                                            18
effect. This finding suggests that FD has had its desired effect only for those firms that

are most difficult to forecast. The behavior captured by Das et al. still pertains for the

other two cohorts.       Table 6 provides evidence on the determinants of optimism (a

dichotomous variable taking the value of 1 if the forecast earnings is greater than actual)

by reporting the results of logit analyzes for all of the companies in the sample and then

low, high and middle FE cohorts.



                             [INSERT TABLE 6 ABOUT HERE]



For leaders, the analysis in Table 6 is consistent with the reversal in optimism suggested

by the summary statistics in Table 2, for all firms, and each of the cohorts, the coefficient

of optimism(t-1) are negative and statistically significant.               The magnitude of the

estimated parameters indicates that optimism is most likely followed by pessimism and

vice versa. The size of the forecast error in the previous period, however, is always

found to be insignificantly different from zero. The findings in section 6 are consistent

with leader overreaction but are inconsistent with the arguments in Das et al. which, if

taken with our findings, would imply that managers rewarding optimism, and presumably

happy with such forecasts, in quarter t-1 systematically provide information that

encourages pessimist forecasts in the subsequent quarter.17 The interaction of optimism(t-

1) with PostFD indicates that the negative effect is somewhat reduced post FD. The

effect, however, is driven by low and middle FE firms. If our overreaction story holds,

the potentially richer information set post-FD reduces leaders’ propensity to overreact but


17
       Das et al.’s findings are also at odds with literature in earnings management suggesting managers
       will favor behavior leading to positive earnings surprises.
                                                 19
only for the easy and moderate firms. For those firms for which the analysis is hardest,

there is no evidence that FD has reduced analysts disposition to overreact. This findings

supports our overreacting leaders hypothesis – we expect the bias of overreaction to be

strongest in matters where difficult, and consequent uncertainty, are greatest.



The estimated parameters for the laggard dummy are positive and statistically significant.

Their magnitude suggest that laggards have a greater propensity for optimism than

leaders but the impact of optimism(t-1) is not significantly different from them. Such a

finding is not consistent with the arguments we made in Section ??? when discussing

indications in the summary statistics that laggards were conservative and tended to

anchor. Our findings indicate that laggards overreaction to their errors are no different to

leaders. Rather, laggards propensity to the optimistic is greater to begin with.



In the previous section, we suggested that ghost apparitions were associated with a

leveling of the playing field. The analyses in Table 6 indicate that ghost apparitions are

associated with a greater likelihood of optimistic forecasts suggesting that the conditions

that level the playing field are perhaps also those that induce bullish sentiments in the

analysts we study.



5.     Number of forecasts.



If overconfidence is driving leaders, we might expect leaders to make fewer revisions to

their forecasts. Confidence breeds certainty and certainty breeds obstinacy. The analysis

in the previous section, however, suggests that analysts, especially leaders, respond to the
                                             20
direction of their previous error and tend to pessimism following optimism and vice

versa. This is not the behavior that would be expected of overconfident analysts reducing

the cognitive dissonance of previous errors.        Our Tobit analyzes of the number of

forecasts made by leaders and laggards examines these issues further. As with the

analyzes presented in Tables 4 and 5, the first three columns of results present estimated

parameters, summary statistics and p-values for each variable in the regression for the

entire sample and then the analysis is repeated for the lowest FE, highest FE and middle

FE cohorts in the remaining columns.



                           [INSERT TABLE 7 ABOUT HERE]



In Section 2, our summary statistics pointed to leaders making fewer forecasts than

laggards and we suggested that this may be a function of timeliness: the later an analyst

makes a forecast, the less time he has to revise it. Our analysis in Table 7 controls for

this and we find that, for the entire sample, the coefficient of timeliness is negative and

statistically significant but that, for the difficulty cohort, this result is found for the low

FE firms only. When we examine the interaction of timeliness(t) and laggard we find

that the estimated parameter is negative and statistically significant. Thus, there appears

to be a basis for our “time to revise” explanation for the difference between the number

of forecasts made by leaders and laggards. Even when time constraints have been taken

into account, the negative and statistically significant parameter for laggard suggests a

greater propensity for laggards to forecast less often than leaders.




                                              21
In Section 3 we argued that timeliness was associated with confidence. If this is the case,

the negative and statistically significant parameter estimate for timeliness(t) is interesting.

The more timely, and corresponding more confident, the leader is, the more likely it is

that he will issue fewer forecasts. Examining the difficulty cohorts shows that the effect

is concentrated is the low and middle FE firms. A lower propensity to revise, as we have

argued in Section ???, is associated with overconfidence. The positive and significant

estimate for experience(t) suggests that greater experience is associated with increased

forecasting activity and, thus, that overconfidence is ameliorated through experience (as

expected). The significant interaction between experience(t) and laggard suggests that

this effect is confined to leaders.



There are significant effects of the analysts error in the previous period (error(t-1)) and

error compared with the consensus (error(ConRel t-1)); the effect is positive for the

former and negative for the latter. As with experience(t) the magnitude of the interaction

of these variables and laggard suggest that the effects are confined to leaders. The

effects of the previous errors provide mixed evidence to the role that (dis)confirming

stimuli play in determining forecasting frequency and do not provide clear evidence on

confidence, and overreaction, stories.



By was of contrast to the relative error measurements, the optimism dummy, the variable

indicating the optimism or pessimism in the previous quarter’s forecast is positive and

statistically significant for the entire sample as well as each difficulty cohort. We have

seen that analysts overreact to the direction of their previous error and the finding here



                                              22
confirms this story. Overreaction is associated with higher arousal18 and higher arousal is

associated with greater levels of activity



Therefore, after controlling for the time available for revisions, we see that the number of

forecasts made is influenced by behavioral biases stemming from confidence and

overreaction.    The findings in this section also confirm our interpretation that the

explanatory variables are proxying well for underlying behavioral phenomena.



6.     Herding

To be completed




7.     Conclusion




                                             23
References



Abarbanell, J., 1991. Do analysts’ earnings forecasts incorporate information in prior

sstock price changes? Journal of Accounting and Economics 14, 147-166.



Barber, B., Lehavy, R., McNichols, M., Trueman, B., 2001. Can Investors Profit from

the Prophets?    Security Analyst Recommendations and Stock Returns, Journal of

Finance 56, 531-563.



Barber, B.M., Odean, T., 2001. Boys Will Be Boys:          Gender, Overconfidence and

Common Stock Investment, Quarterly Journal of Economics ????, 261-292.



Bickchandani, S., Hirshleifer, D., Welch, I., 1992. A theory of fads, fashion, custom, and

cultural change as information cascades, Journal of Political Economy 100, 992-1026.



Bickchandani, S., Sharma, S., 2000. Herd Behavior in Finance Markets: A Review, IMF

Working Paper.



Bitmead, Aaron, Durand, Robert B. and Ng, Hock Guan, 2004, Bubblelepsy.               The

Behavioral Wellspring of the Internet Stock Phenomenon, Journal of Behavioral

Finance, forthcoming.

Cooper, R., Day, T., Lewis, C., 2001. Following the leader: A study of individual

analysts’ earnings forecasts, Journal of Financial Economics 61, 383-416.



                                           24
Das, S., Levine, C., Sivaramakrishnan, K., 1998. Earnings predictability and bias in

analysts’ earnings forecasts, Accounting Review 73, 277-295.



Fried, D., Givoly, D., 1985.      Financial analysts’ forecasts of earnings, Journal of

Accounting and Economics 4, 85-106.



Graham, J., 19999.     Herding among investment newsletters:        Theory and evidence,

Journal of Finance 54, 237-268.



Heflin, F., Subramanyam, K., Zhang, Y., 2003.             Regulation FD and the financial

information environment: Early evidence, Accounting Review 78, 1-38.



Hilary, G., Menzly, L., 2001. Does past success lead analysts to become overconfident?,

working paper.



Hong, H., Kubik, J., Solomon, A., 2000. Security analysts’ career concerns and herding

of earnings forecasts, the Rand Journal of Economics 31, 121-145.



Kruger, J., Dunning, D., 1999.       Unskilled and unaware of it: How difficulties in

recognizing one’s own incompetence lead to inflated self assessments, Journal of

Personality and Social Psychology 77, 1121-1134.



Mohanrm, P., Sunder, S., 2002.        Has regualation fair disclosure affected financial

analysts’ ability to forecast earnings?, working paper.
                                            25
Scharfsetin, D.S., Stein, J.C., 1990. Herd behavior and investment, American Economic

Review 80, 456-479.



Stickel, S., 1990. Predicting individual analyst earnings forecasts, Journal of Accounting

Reseawrch 28, 409-508.



Welch, I., 2000. Herding among security analysts, Journal of Financial Economics 58,

369-396.



Womack, K.L., 1996.      Do Brokerage Analysts’ Recommendations Have Investment
     Value?, Journal of Finance 51, 137-167.




                                           26
Table 1

                                     Analysts Per Firm Quarter

          Total No.       No. of Firm   % of All          Total No.     No. of Firm    % of All

          of Analysts     Quarters      Firms Qtrs        of Analysts   Quarters       Firms Qtrs

                      1       37368          26.84                22           303           0.22

                      2       25468          18.29                23           244           0.18

                      3       17891          12.85                24           205           0.15

                      4       12593           9.04                25           134           0.10

                      5         9114          6.55                26           135           0.10

                      6         7056          5.07                27            95           0.07

                      7         5591          4.02                28            68           0.05

                      8         4413          3.17                29            64           0.05

                      9         3491          2.51                30            41           0.03

                  10            2896          2.08                31            41           0.03

                  11            2297          1.65                32            26           0.02

                  12            1894          1.36                33            26           0.02

                  13            1631          1.17                34            10           0.01

                  14            1273          0.91                35            15           0.01

                  15            1121          0.81                36               7         0.01

                  16             940          0.68                37               6         0.00

                  17             803          0.58                38               3         0.00

                  18             630          0.45                39               3         0.00

                  19             543          0.39                42               2         0.00

                  20             422          0.30                43               1         0.00

                  21             367          0.26                47               1         0.00




                                                     27
Table 2

                                           Summary Statistics for Leaders and Laggards
The sample is 23,420 pairs of ‘timely’ leaders (the first analyst to make an initial forecast in period t) and laggards (the last analyst to
make an initial forecast in period t) taken from the IBES Earnings Forecast database for quarters during the years 1994-2002. Below are
the summary statistics for forecast error measured as a percentage of actual earnings for quarter t (initial error) and quarter t-1 (final
error), timeliness measured in days from the beginning of quarter t and quarter t-1, experience measured in days since the analysts first
appearance on the database, the number of forecasts made in quarter t-1 and the consensus error in quarter t-1 (reported in laggard
column). Statistics are reported for the group as a whole and by difficulty of earnings predictability (measured by average forecast error
of the firm).
                               Forecast error                Timeliness           Experience Number of Consensus
                             (t)            (t-1)        (t)           (t-1)          (t)       Forecasts (t) Error(t-1)


                                                          Panel A. Complete Sample
                                                                Firms = 2697
                                                            Firm Quarters=23,420


Leaders:
 Mean                          6.345        -1.936        1.529          12.896       1360.778        1.512
 Median                       -1.188        -2.381        1.000          1.000        1211.000        1.000
 Standard Deviation            212.119       117.899      4.232          23.336       822.584         0.748
 Sign Test Statistic
 p-value                      -1451.5       -3432.5       9074           10519         11704          11710
                              <.0001         <.0001       <.0001         <.0001        <.0001         <.0001
Laggards:
 Mean                         -0.582        -1.808        56.018         31.759       1383.257        1.337         1.219
 Median                       -2.326        -2.273        62.000         23.000       1248.000        1.000        -1.784
 Standard Deviation            126.419       119.280      27.867         29.942        818.857        0.600         139.950
 Sign Test Statistic          -3292.5        -3392        11691          11386.5       11710          11710         -2599
 p-value                       <.0001         <.0001      <.0001         <.0001        <.0001         <.0001        <.0001




                                                                    28
                                            Panel B. Low FE Firms
                                                  Firms=809
                                             Firm Quarters=9215


Leaders:
 Mean                  1.722     -2.124    1.388          12.998    1357.809   1.450
 Median                 0.000    -1.719    1.000          1.000     1197.000   1.000
 Standard Deviation    46.494     45.184   3.037          23.665    824.836    0.697
 Sign Test Statistic    -585     -1433     3484           4103       4605.5    4607.5
 p-value                <.0001    <.0001   <.0001         <.0001     <.0001    <.0001

Laggards:
 Mean                  -1.783    -2.018    58.311         33.513    1379.714   1.286    -1.063
 Median                -1.538    -1.613    64.000         27.000    1246.500   1.000    -1.216
 Standard Deviation    30.117     47.967   27.112         30.480    803.503    0.550     34.830
 Sign Test Statistic   -1369     -1392     4597           4477       4602      4602     -1136
 p-value               <.0001     <.0001   <.0001         <.0001     <.0001    <.0001    <.0001




                                                     29
                                              Panel C. Mid FE Firms
                                                   Firms=1079
                                               Firm Quarters=9175


Leaders:
 Mean                   3.250     -3.405    1.5440          12.897    1385.006   1.535
 Median                -1.597     -2.985    1.000           1.000     1247.000   1.000
 Standard Deviation     98.229    66.545    3.692           23.319     822.688   0.761
 Sign Test Statistic    -527      -1340     3548            4112.5     4585.5    4587.5
 p-value                <.0001     <.0001   <.0001          <.0001     <.0001    <.0001

Laggards:
 Mean                  -2.672    -3.209     55.457          30.901    1395.071   1.341    -1.245
 Median                -3.125    -3.125     61.000          21.000    1265.500   1.000    -2.381
 Standard Deviation     62.095    71.613    27.816          29.850     827.786   0.610    96.238
 Sign Test Statistic   -1319     -1362.5    4575.5          4449.5     4586      4586     -985
 p-value                <.0001     <.0001   <.0001          <.0001      <.0001   <.0001   <.0001




                                                       30
                                              Panel D. High FE Firms
                                                    Firms=809
                                               Firm Quarters=5030



Leaders:
 Mean                   20.460    1.090      1.760          12.708     1322.022   1.585
 Median                -3.643     -3.845     1.000          1.000      1176.000   1.000
 Standard Deviation    433.258     229.995   6.445          22.760     816.832    0.802
 Sign Test Statistic   -339.5     -659.5     2042           2303.5     2513       140.164
 p-value               <.0001     <.0001     <.0001         <.0001     <.0001     <.0001

Laggards:
 Mean                   5.409     1.121      52.855         30.119     1368.242   1.362     9.862
 Median                -4.037     -3.774     58.000         21.000     1212.000   1.000     -3.333
 Standard Deviation     255.936    229.203   28.941         28.946     830.066    0.631      267.960
 Sign Test Statistic   -604.5     -637.5     2518.5         2460       2522       2522      -478
 p-value                <.0001     <.0001    <.0001         <.0001     <.0001     <.0001     <.0001




                                                       31
Table 3

                            Summary Statistics for the Differences between Leaders and Laggards
The sample is 23,420 pairs of ‘timely’ leaders (the first analyst to make an initial forecast in period t) and laggards (the last analyst to
make an initial forecast in period t) taken from the IBES Earnings Forecast database for quarters during the years 1994-2002. Below are
the summary statistics for the differences (leader minus laggard) in forecast error measured as a percentage of actual earnings for quarter t
and quarter t-1, timeliness measured in days from the beginning of quarter t and quarter t-1, experience measured in days since the
analysts first appearance on the database, and the number of forecasts made in quarter t-1. Statistics are reported for the group as a whole
and by difficulty of earnings predictability (measured by average forecast error for each firm).



                                  Forecast error                      Timeliness              Experience         Number of
                            (t)                    (t-1)       ( t)                (t-1)         (t)             Forecasts(t)


                                                           Panel A. Complete Sample
                                                                 Firms = 2697
                                                             Firm Quarters=23,420

 Mean                     7.035                -0.213       -53.850            -18.592        -25.612              0.189
 Median                    0.000                0.000       -61.000             -9.000        -46.000              0.000
 Standard Deviation       172.160               95.550      29.662              36.167        900.543              0.838
 Sign Test Statistic      1041                  21.5         -11336            -5723.5         -2009               1618
 p-value                 <.0001                 0.756       <.0001              <.0001        <.0001              <.0001




                                                                        32
                                              Panel B. Low FE Firms
                                                    Firms=809
                                               Firm Quarters=9215

Mean                     3.472      -0.128   -56.340          -20.300    -23.625     0.163
Median                   0.000       0.000   -63.000          -12.000    -46.000     0.000
Standard Deviation       46.526     52.943    29.024           37.076    -48.000     0.802
Sign Test Statistic      498          -34     -4481             -2329    -880.5       562
p-value                 <.0001       0.422    <.0001           <.0001    <.0001     <.0001



                                              Panel C. Mid FE Firms
                                                   Firms=1079
                                               Firm Quarters=9175

 Mean                   5.860     -0.141     -53.346        -17.734     -14.754     0.197
 Median                0.000       0.000     -53.346         -8.000     -43.000     0.000
 Standard Deviation    94.216     67.609      29.357         35.793     902.987     0.840
 Sign Test Statistic   439.5       59.5      -4446.5         -2181       -716.5     657
 p-value               <.0001      0.167     <.0001          <.0001     <.0001     <.0001



                                             Panel D. High FE Firms
                                                   Firms=809
                                              Firm Quarters=5030

 Mean                  15.675     -0.500     -50.220        -17.035     -48.980     0.220
 Median                  0.000     0.000     -56.000        -8.0000     -48.000     0.000
 Standard Deviation    342.704    170.204     30.926         35.027      944.383    0.894
 Sign Test Statistic     103.5      -4       -2408.5         -1213.5      -412       399
 p-value                0.0019     0.913     <.0001          <.0001      <.0001    <.0001




                                                       33
Table 4

                                                                              Timeliness(t)
The sample is 23,420 pairs of ‘timely’ leaders (the first analyst to make an initial forecast in period t) and laggards (the last analyst to
make an initial forecast in period t) taken from the IBES Earnings Forecast database for quarters during the years 1994-2002. Below are
the results for the determinants of timeliness for quarter t for the combined group.

                                Complete Sample                           Lowest FE                        Highest FE                             Middle FE
                                     Firms=2697                           Firms=809                        Firms=809                              Firms=1079
                             Firm Quarters=23,420                 Firm Quarters=9215                Firm Quarters=5030                     Firm Quarters=9175
                         Parameter    Chi-Sq     Pr > ChiSq   Parameter Chi-Sq Pr > ChiSq      Parameter    Chi-Sq       Pr > ChiSq   Parameter    Chi-Sq     Pr > ChiSq
            Intercept     20.3157      1160.44      <.0001      1.5395       11.29    0.0008      27.782      333.49        <.0001       0.6222        1.63        0.201
     Timeliness (t-1)      -0.0589        33.5      <.0001      -0.0088       1.28    0.2588     -0.0677        6.48        0.0109       0.0515       38.88      <.0001
           Error (t-1)       0.007         2.9      0.0888      0.0044        0.16    0.6876      0.0074        1.67        0.1958       0.0017        0.15      0.6941
   Error (ConRel t-1)       -0.005        6.67      0.0098      -0.0012       0.05    0.8176     -0.0022        0.76        0.3827       -0.001        0.18      0.6694
       Experience (t)      -0.0106     1194.95      <.0001      0.0006        5.95    0.0147     -0.0144      344.19        <.0001        0.001       14.96      0.0001


Post FD Dummy             -22.4593      221.99      <.0001      2.0366        2.55    0.1101    -32.4952       90.85        <.0001        2.673        4.22      0.0398
     *Timeliness(t-1)       0.8859      986.07      <.0001       0.017        0.52    0.4725      1.1614      289.11        <.0001       0.1924       60.44      <.0001
          *Error(t-1)      -0.0096        1.49      0.2217      0.0325        2.14    0.1435     -0.0106        1.04        0.3078      -0.0055        0.31      0.5796
  *Error (ConRel t-1)       0.0088         4.9      0.0268      -0.0326       3.31    0.0687      0.0034         0.5        0.4789       0.0054        1.33      0.2492
      *Experience (t)       0.0094      229.24      <.0001      -0.0006       1.45    0.2284      0.0131       77.53        <.0001       0.0009        3.05      0.0809


Laggard Dummy             55.2461      3162.86      <.0001     79.6129      10149     <.0001    46.1136       336.28        <.0001     78.2878      9026.99      <.0001
     *Timeliness(t-1)       0.1131       62.67      <.0001      0.0441       15.84    <.0001      0.1244       11.03        0.0009      -0.0215        3.22      0.0726
          *Error(t-1)      -0.0046        0.75      0.3863      0.0384        3.44    0.0637      -0.004        0.33        0.5666      -0.0035        0.21        0.644
  *Error (ConRel t-1)       0.0034        1.37      0.2424      -0.0088       1.03    0.3097     -0.0001             0      0.9718       0.0025        0.48      0.4868
      *Experience (t)       0.0085      246.95      <.0001      -0.0027       35.7    <.0001      0.0118       74.23        <.0001      -0.0021       21.63      <.0001




                                                                                      34
Laggard*PostFD          13.6833     30.34   <.0001    -5.1829     6.07   0.0138   20.7223     12.87   0.0003    -9.0809    18.32   <.0001
     *Timeliness(t-1)    -0.8738   526.71   <.0001    -0.0329     1.05    0.305    -1.1355   154.45   <.0001    -0.1732    27.59   <.0001
          *Error(t-1)    -0.0012     0.01   0.9182    -0.0861     4.68   0.0305    -0.0002       0    0.9873    -0.0038     0.06   0.8022
  *Error (ConRel t-1)    -0.0016     0.07    0.791     0.043      1.25   0.2633    0.0043      0.29   0.5875    0.0006      0.01   0.9217
      *Experience (t)    -0.0067    41.05   <.0001    0.0012      1.67    0.196    -0.0089    12.88   0.0003    0.0004      0.25   0.6182


Optimism Dummy           -8.8363   332.13   <.0001    -0.7936     4.21   0.0401   -13.3414   117.22   <.0001    -0.9774      5.9   0.0151
          *Error(t-1)    -0.0029     0.47   0.4933    -0.0034     0.11   0.7432    -0.0057     0.98   0.3219    -0.0043     0.34   0.5618


Optimism*PostFD         14.2917    139.14   <.0001    -3.1641     9.94   0.0016   21.9321     56.73   <.0001     -1.452     2.18   0.1402
          *Error(t-1)    -0.0043     0.25   0.6184    0.1285     13.95   0.0002    0.0026      0.05   0.8256    0.0014      0.01    0.917


Laggard*Optimism         7.5129     80.13   <.0001    0.1031      0.02   0.8818   11.4524     29.75   <.0001    -0.9352     1.77   0.1838
          *Error(t-1)    -0.0008     0.01   0.9181    -0.0299      2.1   0.1473    0.0001        0    0.9892    0.0075      0.41   0.5218


Laggard*PostFD*Opt -15.7946          54.7   <.0001    2.8532       2.2   0.1379   -23.9368    24.02   <.0001    1.5718      0.77   0.3789
          *Error(t-1)    0.0054      0.05   0.8229    -0.0413      0.3   0.5834    0.0034      0.01   0.9138     -0.025     0.81   0.3695


Ghost Dummy             12.6891    669.31   <.0001    1.5744     18.01   <.0001   17.8374    199.12   <.0001    1.1729      8.78    0.003
Ghost*Laggard           -18.8828   233.52   <.0001   -19.4604   355.65   <.0001   -14.5881    21.62   <.0001   -11.3226   119.55   <.0001
     *Timeliness(t-1)    0.0668     17.16   <.0001    0.0432     12.61   0.0004    0.0834      3.64   0.0565    0.0001        0    0.9923
          *Error(t-1)    -0.0014     0.05   0.8242    -0.0301     1.79   0.1813    -0.0074     0.52   0.4694     0.004      0.24   0.6237
  *Error (ConRel t-1)    -0.0049     1.33   0.2481    -0.0002       0    0.9913     0.001      0.04    0.849    -0.0142     9.17   0.0025
      *Experience (t)    -0.0079   157.84   <.0001    0.0004      0.42   0.5146    -0.0132    69.42   <.0001    -0.0033    37.85   <.0001




                                                                         35
Ghost*PostFD             -7.4281   40.01   <.0001    5.3374    29.92   <.0001   -11.3652    16.4   <.0001   -6.4174   43.18   <.0001
Ghost*Lag*PostFD        -15.5595   26.81   <.0001    -2.3526    0.73   0.3939   -31.7279   20.32   <.0001   1.1077     0.18   0.6682
     *Timeliness(t-1)    0.3167    69.16   <.0001    -0.0692    4.67   0.0306     0.567     36.7   <.0001   0.1435    20.78   <.0001
          *Error(t-1)    -0.0006      0    0.9635    0.1084     3.33    0.068    0.0052     0.08   0.7761   0.0271     1.31   0.2524
  *Error (ConRel t-1)    0.0051     0.29   0.5912    -0.0077    0.03   0.8645    -0.0016    0.02   0.8972   -0.0239    4.84   0.0279
      *Experience (t)    0.0107     80.1   <.0001    0.0037    10.66   0.0011    0.0181    41.52   <.0001   0.0003     0.11   0.7383


Ghost*Lag*Opt            -5.1082    25.2   <.0001    0.8978     1.12   0.2905    -9.4672   14.33   0.0002   -2.5193    8.35   0.0039
          *Error(t-1)     0.008     0.81   0.3685    -0.0367    0.73   0.3917    0.0124     0.91   0.3402    -0.004     0.1    0.749


Ghost*Lag*PostFD         -2.8266    1.29   0.2553   -14.8211   42.88   <.0001    5.8971     1.07      0.3   -0.5674    0.07   0.7854
          *Error(t-1)    -0.0085    0.07   0.7886    -0.1181    1.62   0.2031    -0.0243    0.25   0.6159   0.0425     1.47    0.225




                                                                       36
Table 5

                                                            Timeliness Difference(t)
The sample is 23,420 pairs of ‘timely’ leaders (the first analyst to make an initial forecast in period t) and laggards (the last analyst to
make an initial forecast in period t) taken from the IBES Earnings Forecast database for quarters during the years 1994-2002. Below are
the results for the determinants of the difference in the timeliness between the pairs (leaders minus laggards) in quarter t. Each
independent variable is the leader value minus the laggard value for each firm quarter
                        Complete Sample                    Lowest FE                            Highest FE                         Middle FE
                          Firms=2697                       Firms=809                            Firms=809                         Firms=1079
                      Firm Quarters=23,420             Firm Quarters=9215                Firm Quarters=5030                   Firm Quarters=9175
                   Parameter Chi-Sq Pr > ChiSq     Parameter Chi-Sq Pr > ChiSq      Parameter    Chi-Sq      Pr > ChiSq   Parameter Chi-Sq Pr > ChiSq
       Intercept -65.88879 -225.77        <.0001   -68.07269 -148.48    <.0001      -63.35887       -96.44      <.0001     -65.13831 -142.22   <.0001
      Error(t-1)    -0.00445   -1.74      0.0827     0.00362    0.53    0.5993       -0.00591        -1.89      0.0582       -0.0031   -0.46   0.6438
 Timeliness(t-1)     0.04299    6.31      <.0001     0.03936    3.82    0.0001        0.03187         1.99      0.0464      0.04842     4.44   <.0001
  Experience (t)     0.00144    4.37      <.0001     0.00138    2.63    0.0086        0.00145         2.01      0.0447      0.00146     2.84   0.0046


Post FD Dummy        8.18958   11.03      <.0001     7.13106    5.82    <.0001        8.83815         5.72      <.0001        8.3794    7.16   <.0001
      Error(t-1)     0.00848    1.72      0.0854    -0.00421   -0.13    0.8929        0.00938         1.66      0.0976      0.00554     0.42   0.6759
 Timeliness(t-1)     0.00117    0.06       0.951    -0.00168   -0.05    0.9567        0.02489         0.64       0.522      -0.00904   -0.29   0.7704
  Experience (t)    -0.00123   -2.26      0.0238    -0.00179   -2.02    0.0434      0.0006053         0.53      0.5992      -0.00172   -1.99   0.0462


Ghost Dummy         25.46614   60.18      <.0001    25.36836   38.36    <.0001       25.86476       27.29       <.0001     25.44004    37.99   <.0001
      Error(t-1)    -0.00365   -0.82      0.4096    -0.00768   -0.72    0.4733       -0.01048        -1.67      0.0945      0.00442     0.49   0.6276
 Timeliness(t-1)     0.06485    6.27      <.0001     0.07202    4.65    <.0001        0.05103         2.11      0.0352      0.05928     3.54   0.0004
  Experience (t) -0.0001698    -0.35      0.7248   0.0006754    0.89    0.3747       -0.00102        -0.97      0.3301    -0.0005356   -0.69   0.4874


Ghost*PostFD         5.68891    5.54      <.0001     6.65684    3.95    <.0001         8.4223         3.89      0.0001      3.38121     2.09   0.0363
      Error(t-1)    -0.00999   -1.17      0.2429      0.0565    1.46        0.143    -0.01337        -1.25      0.2132      -0.00867   -0.44   0.6588
 Timeliness(t-1)     0.06852    2.51      0.0121     0.05743    1.32    0.1881        0.16314         2.79      0.0053      0.04295     0.98   0.3264
  Experience (t)    -0.00309   -3.93      <.0001     -0.0032    -2.5    0.0125       -0.00401        -2.44      0.0146       -0.0022   -1.76       0.079




                                                                             37
Table 6

                                                                    Optimistic Forecast (t)
The sample is 23,420 pairs of ‘timely’ leaders (the first analyst to make an initial forecast in period t) and laggards (the last analyst to
make an initial forecast in period t) taken from the IBES Earnings Forecast database for quarters during the years 1994-2002. Below are
the results for the determinants of an optimistic forecast (whether forecasted eps > actual eps) for quarter t (Note: logit regression).
                                Complete Sample                          Lowest FE                        Highest FE                        Mid FE
                                      N=2697                              N=809                            N=809                            N=1079
                              Firm Quarters=23,420                Firm Quarters=9215               Firm Quarters=5030               Firm Quarters=9175
                                                   Pr >                                Pr >                             Pr >                           Pr >
                          Parameter     Chi-Sq    ChiSq      Parameter       Chi-Sq   ChiSq   Parameter       Chi-Sq   ChiSq    Parameter      Chi-Sq ChiSq
              Intercept     0.6839    818.0953    <.0001       0.7734       410.571 <.0001      0.5634      115.7765   <.0001     0.6593     292.4083 <.0001
      Timeliness (t-1) -0.00236        14.4929    0.0001      -0.00283       8.4511 0.0036     -0.00235       2.9485    0.086    -0.00194      3.7967 0.0514
            Error (t-1) -0.00003        0.0564    0.8123      -0.00043       0.5248 0.4688     -0.00016        1.031   0.3099   0.000547       1.8411 0.1748
 Optimistic Fcast (t-1)      -0.824   610.3973    <.0001       -0.8067     220.7065 <.0001      -0.6063      71.5265   <.0001     -0.9594    299.5869 <.0001


Post FD Dummy               -0.1861      9.917    0.0016       -0.3418      12.2976 0.0005      -0.0579       0.2259   0.6346     -0.0928      0.9482 0.3302
      *Timeliness(t-1) -0.00032         0.0329       0.856   0.000675         0.053 0.8179     0.00101        0.0766   0.7819    -0.00184      0.4594 0.4979
           *Error(t-1) 0.000046          0.026    0.8718      0.00109         0.441 0.5067    0.000146        0.1914   0.6617    -0.00026      0.1193 0.7298
*Optimistic Fcast (t-1)     0.2817     11.8135    0.0006         0.469      10.5031 0.0012      0.0111        0.0044   0.9468     0.2548       3.7877 0.0516


Laggard Dummy               0.6447    208.5884    <.0001       0.7384       99.6217 <.0001      0.5717       35.3983   <.0001     0.6237      78.6501 <.0001
      *Timeliness(t-1)      -0.0008      0.712    0.3988      -0.00162       1.1619 0.2811     0.00126        0.3591    0.549    -0.00139      0.8571 0.3545
           *Error(t-1) 0.000158         0.5141    0.4734      0.00179        2.2908 0.1301    0.000181        0.5831   0.4451   0.000093       0.0203 0.8867
*Optimistic Fcast (t-1)     0.0724      1.5807    0.2087       0.0316        0.1086 0.7418      -0.0592       0.2339   0.6286     0.1581        2.822    0.093




                                                                                      38
Laggard*PostFD             0.0574    0.2877   0.5917     0.3089   2.713 0.0995     -0.1771    0.6808   0.4093   -0.00168   0.0001 0.9921
      *Timeliness(t-1)    0.00277    1.2044   0.2724   0.000172   0.0016 0.9685   -0.00049    0.0092   0.9235   0.00651    2.6604 0.1029
           *Error(t-1) 0.000787      1.5195   0.2177   -0.00147   0.2044 0.6512   0.000845    1.4468    0.229   0.00129    0.6246 0.4293
*Optimistic Fcast (t-1)   -0.4001    7.3284   0.0068    -0.1648   0.3715 0.5422     -0.419    2.0529   0.1519    -0.4652   3.8003 0.0512


  Ghost Dummy              0.1243   17.0724   <.0001     0.1174   5.9898 0.0144     0.1022     2.479   0.1154    0.1393    8.3294 0.0039
Ghost*Laggard             -0.2906   21.9889   <.0001    -0.3128   9.4652 0.0021    -0.4237   10.4577   0.0012    -0.1983   4.0069 0.0453
      *Timeliness(t-1)    0.00294    7.1292   0.0076    0.00337   3.7378 0.0532    0.00107    0.2002   0.6546   0.00318    3.1995 0.0737
           *Error(t-1) 8.97E-06       0.001   0.9751   -0.00086   0.4346 0.5097   -0.00008    0.0611   0.8048    -0.0003   0.1699 0.6802
*Optimistic Fcast (t-1)     -0.15    4.5264   0.0334    -0.0889   0.5892 0.4427     0.0255    0.0291   0.8645    -0.2722    5.577 0.0182


Ghost*PostFD              -0.0104    0.0201   0.8872    -0.0352   0.0829 0.7734     0.2279    2.2819   0.1309     -0.123   1.1299 0.2878
Ghost*Lag*PostFD          -0.0988    0.4825   0.4873     -0.154   0.3888 0.5329    -0.0705    0.0608   0.8053     -0.133   0.3483 0.5551
      *Timeliness(t-1) -0.00351      1.7762   0.1826   -0.00141   0.0946 0.7584    0.00108    0.0412   0.8391   -0.00703   2.8879 0.0892
           *Error(t-1) -0.00158      2.8133   0.0935   -0.00048   0.0146 0.9038   -0.00145    2.1441   0.1431   -0.00289   1.3978 0.2371
*Optimistic Fcast (t-1)     0.512    8.5892   0.0034     0.1035   0.1038 0.7473     0.5687    2.7568   0.0968     0.791    7.7634 0.0053




                                                                         39
Table 7

                                                                  Number of Forecasts(t)
The sample is 23,420 pairs of ‘timely’ leaders (the first analyst to make an initial forecast in period t) and laggards (the last analyst to
make an initial forecast in period t) taken from the IBES Earnings Forecast database for quarters during the years 1994-2002. Below are
the results for the determinants of the number of forecasts in quarter t for the combined group
                                Complete Sample                         Lowest FE                             Highest FE                             Middle FE
                                  Firms=2697                            Firms=809                             Firms=809                             Firms=1079
                              Firm Quarters=23,420                Firm Quarters=9215                     Firm Quarters=5030                   Firm Quarters=9175
                         Parameter   Chi-Sq    Pr > ChiSq   Parameter    Chi-Sq     Pr > ChiSq    Parameter    Chi-Sq      Pr > ChiSq   Parameter    Chi-Sq      Pr > ChiSq
            Intercept       1.8827   14969.7      <.0001       1.6612   5358.24          <.0001      1.9816     3045.35       <.0001       2.0871    6963.31        <.0001
        Timeliness(t)      -0.0055     26.46      <.0001       -0.022     71.82          <.0001     -0.0013        0.29       0.5879      -0.0114      23.04        <.0001
           Error (t-1)     -0.0002      3.77      0.0521      -0.0012         7          0.0082     -0.0001        0.52       0.4705            0       0.02        0.8886
   Error (ConRel t-1)       0.0003     44.85      <.0001        0.001     24.87          <.0001      0.0001        4.42       0.0355       0.0004      32.41        <.0001
       Experience (t)       0.0001      63.1      <.0001       0.0001     94.68          <.0001      0.0001       20.22       <.0001            0          0        0.9679

Post FD Dummy               0.0852      4.38      0.0364       0.2461      11.5          0.0007      0.0053           0       0.9467      -0.1003       2.42        0.1199
       *Timeliness(t)      -0.0093     34.28      <.0001       0.0062      1.88          0.1702     -0.0123       16.51       <.0001      -0.0137      11.69        0.0006
           *Error(t-1)     -0.0003      1.77       0.183      -0.0007      0.36          0.5462     -0.0001        0.23       0.6292      -0.0008       1.48        0.2243
  *Error (ConRel t-1)       0.0007     84.74      <.0001        0.003      10.1          0.0015      0.0009       50.29       <.0001      -0.0002       0.73        0.3933
      *Experience (t)       0.0001     32.96      <.0001      -0.0001      5.49          0.0191      0.0002       23.36       <.0001       0.0002      52.16        <.0001

Laggard Dummy              -0.3715     96.64      <.0001      -0.2323        15          0.0001     -0.3915       23.23       <.0001      -0.5517      86.61        <.0001
       *Timeliness(t)      -0.0004      0.15      0.6991        0.017      40.6          <.0001     -0.0047        3.52       0.0607       0.0048       3.87         0.049
           *Error(t-1)      0.0003      3.21       0.073       0.0008      0.76          0.3831      0.0002        0.91          0.34      0.0001       0.15         0.697
  *Error (ConRel t-1)      -0.0003     15.54      <.0001      -0.0008      3.86          0.0493     -0.0002         3.5       0.0614      -0.0003       3.89        0.0487
      *Experience (t)      -0.0001     15.86      <.0001      -0.0001     27.24          <.0001     -0.0001        13.3       0.0003            0        2.4        0.1217




                                                                                    40
Laggard*PostFD            0.0155      0.03   0.8598   -0.3275      4.67        0.0306    0.1327    0.59    0.443    0.2344     2.94   0.0862
       *Timeliness(t)     0.0082     20.37   <.0001   -0.0059      1.54        0.2145    0.0105    8.99   0.0027    0.0126     8.88   0.0029
           *Error(t-1)   -0.0001      0.14    0.711    0.0017      1.03        0.3093         0       0   0.9877    -0.001      1.4   0.2375
  *Error (ConRel t-1)    -0.0006     19.94   <.0001    -0.005     10.24        0.0014   -0.0007   14.67   0.0001    0.0003     0.64   0.4237
      *Experience (t)    -0.0001     11.42   0.0007    0.0001      3.38        0.0658   -0.0001     4.8   0.0285   -0.0002    24.38   <.0001

Optimism Dummy             0.427   1016.28   <.0001    0.6922   1113.03        <.0001    0.2583   69.46   <.0001    0.3161   226.15   <.0001
         *Error(t-1)     -0.0002      2.98   0.0844   -0.0001      0.02        0.8773   -0.0001    0.58   0.4468   -0.0012     7.25   0.0071

Optimism*PostFD           0.1457     18.83   <.0001    0.1563      8.89        0.0029    0.2692   12.64   0.0004   -0.0014        0   0.9784
         *Error(t-1)     -0.0006      5.18   0.0229   -0.0074     21.17        <.0001   -0.0011    10.3   0.0013    0.0023     8.84   0.0029

Laggard*Optimism         -0.3618    232.38   <.0001   -0.6179     284.3        <.0001   -0.1864   11.89   0.0006   -0.264     51.31   <.0001
         *Error(t-1)      0.0002      0.95   0.3298    0.0002      0.05        0.8319    0.0001    0.19   0.6598   0.0011      3.02   0.0822

Laggard*PostFD*Opt       -0.1422      5.49   0.0191   -0.0971      0.94        0.3329   -0.4214   10.98   0.0009    0.0757     0.64    0.424
         *Error(t-1)      0.0008      1.26   0.2616    0.0065      2.43        0.1192    0.0009    1.23   0.2675   -0.0002     0.02   0.8974

Ghost Dummy              -0.0742     32.07   <.0001   -0.0944     23.47        <.0001    0.0174    0.33   0.5664   -0.1177    32.18   <.0001
Ghost*Laggard             0.1171       6.5   0.0108    0.1466      4.08        0.0435   -0.0806    0.66   0.4169    0.2412       11   0.0009
       *Timeliness(t)     0.0002      0.17   0.6785   -0.0006      0.59        0.4435    0.0016    2.27   0.1319   -0.0004     0.18    0.669
          *Error(t-1)    -0.0001      0.19   0.6633   -0.0003      0.15        0.6989         0    0.02   0.8923         0        0    0.984
 *Error (ConRel t-1)      0.0002      2.89   0.0893    0.0007      3.03        0.0819    0.0002    1.79   0.1813   -0.0002     0.49   0.4834
      *Experience (t)          0      0.67   0.4124         0       0.1        0.7484    0.0001    2.14   0.1431   -0.0001     6.36   0.0117




                                                                          41
Ghost*PostFD             0.1002    9.35   0.0022     0.018   0.12        0.7247    0.1271   3.18   0.0746    0.1352    7.02   0.0081
Ghost*Lag*PostFD        -0.0045       0    0.966   -0.0671   0.14        0.7041    0.1526   0.54   0.4609   -0.1714    1.07    0.301
       *Timeliness(t)   -0.0022    3.58   0.0583    0.0001      0        0.9717   -0.0038   2.64   0.1043    -0.001    0.25    0.618
          *Error(t-1)    0.0005     1.6   0.2055    0.0008   0.11        0.7365   -0.0001   0.02   0.9015    0.0044   17.48   <.0001
 *Error (ConRel t-1)    -0.0002    0.62   0.4296    0.0009   0.22        0.6417         0      0   0.9962   -0.0008    2.16   0.1419
      *Experience (t)         0    2.12   0.1457         0   0.02        0.8927         0   0.04   0.8487    0.0001    6.85   0.0088

Ghost*Lag*Opt            0.1459   25.06   <.0001   0.1289     8.4        0.0038    0.1616   6.15   0.0131    0.1419    9.61   0.0019
         *Error(t-1)    -0.0004    1.75   0.1853   0.0015    1.21        0.2717   -0.0004   1.26   0.2612   -0.0005    0.67   0.4139

Ghost*Lag*PostFD        0.0662     0.87   0.3523    0.1715   2.08        0.1488   0.1054    0.51   0.4751    -0.044    0.16   0.6893
         *Error(t-1)    0.0011     1.29   0.2569   -0.0027   0.39        0.5322   0.0016     1.8   0.1798   -0.0024    1.11   0.2924




                                                                    42

								
To top