Journal of accounting, auditing finance, tập 26, số 03, 2011 7

Ó The Authors 2011 Reprints and permission: sagepub.com/journalsPermissions.nav DOI: 10.1177/0148558X11401556http://jaaf.sagepub.com Analysts’ Recommendation Revisions and Subsequent Ear

Ó The Author(s) 2011 Reprints and permission: sagepub.com/journalsPermissions.nav DOI: 10.1177/0148558X11401556

http://jaaf.sagepub.com

Analysts’ Recommendation

Revisions and Subsequent

Earnings Surprises: Pre- and

identify-Keyword

Regulation FD, analyst recommendations, earnings surprises, portfolio analysis

Introduction

Earnings-related selective disclosure was one of the most publicized cases of unfair

Exchange Commission (SEC) received several thousand comment letters expressing tion on the basis of the belief that corporations were giving earnings-related informationonly to a select group of financial analysts and institutional investors Relying on this infor-mation, analysts then made recommendations to their clients prior to earnings announce-ments, thus giving them an unfair advantage over other investors Inevitably, suchdisclosure policies helped certain selected investors earn profits or avoid losses at theexpense of other investors In response, the SEC passed Regulation FD and listed earnings-related disclosure on top of the list of potential material information that needs to be dis-closed simultaneously to all market participants

This study examines the extent to which analysts’ recommendations helped their clients

ana-lysts, as alleged, were privately communicating earnings-related information with managersand providing consistent investment advice to their clients, then analysts’ recommendationsshould have exhibited predictive power of upcoming earnings surprises in the pre-Regulation FD period Furthermore, to the extent that Regulation FD was effective in cur-tailing selective disclosure, the predictive value of recommendations should have declinedafter Regulation FD took effect

We estimate recommendations’ predictive value of upcoming earnings surprises usingthe association between recommendation revisions and unexpected earnings calculatedbased on (a) time-series earnings expectations, (b) analysts’ earnings expectations, and(c) earnings announcement returns In addition, we examine the association between earn-ings surprises and recommendation revisions using regression analysis controlling forpostearnings announcement drift, return momentum, accruals anomaly, and institutionaltrading Finally, we construct a trading strategy designed to capture recommendation revi-sions’ predictive power of earnings surprises and compare the abnormal returns accrued bythis portfolio during the pre- and post-Regulation FD periods

We find that prior to Regulation FD’s acceptance, upgraded firms exhibited 3-daymarket-adjusted earnings announcement returns that were on average 0.93% higher thandowngraded firms After Regulation FD took effect, the return differential betweenupgraded and downgraded firms declined 55% to 0.43% The regression analysis providessimilar results and supports the inference that the association between recommendationrevisions and subsequent earnings surprises declined after Regulation FD took effect.Finally, the trading strategy analysis reveals an approximately 70% decline in the portfo-lio performance after Regulation FD took effect Overall, our results are consistent withRegulation FD having been effective in limiting selective disclosure and reducing analystrecommendations’ predictive power of upcoming earnings surprises

Regulation FD was preceded with intense objection that the rule would harm the level

of corporate disclosure Consistently, prior studies on Regulation FD focused primarily onwhether the rule damaged corporate disclosure level, increased earnings volatility, andreduced forecast accuracy The literature suggests that Regulation FD did not have signifi-cant adverse effects on corporate disclosure This article contributes to the extant literature

by comparing the extent to which recommendations were valuable to analysts’ clients inidentifying earnings surprises before Regulation FD and how much of this predictive valuewas eliminated after Regulation FD took effect We also examine the impact of Regulation

FD on the abnormal performance of investors who followed analysts’ recommendationrevisions with the intent of benefiting from analysts’ earnings-related private information.Hence, our analysis provides important insights relating to the fundamental concern of cer-tain investors earning abnormal profits at the expense of other investors based on selectivedisclosure

Literature Review and Hypotheses Development

In response to growing concerns of select individuals obtaining access to inside tion, the SEC passed Regulation FD, which was concerned with the fair disclosure of non-public material information Regulation FD required managers to disseminate any materialinformation simultaneously to all market participants and prohibited selective disclosure.Many securities markets professionals and institutional investors argued that bringing

informa-additional restrictions on corporate disclosure would reduce the quantity and quality ofinformation available to capital markets.

After Regulation FD took effect, there was increased concern by practitioners thatRegulation FD increased return volatility, adversely affected analysts’ earnings forecasts,reduced corporate disclosure, and increased information asymmetry However, academicresearch investigating Regulation FD’s impact provided mixed inferences Table 1 presentsstudies that investigated Regulation FD’s impact and summarizes their findings

On the return volatility aspect, while Heflin, Subramanyam, and Zhang (2003),Eleswarapu, Thompson, and Venkataraman (2004), and Sinha and Gadarowski (2010)found a significant decline after Regulation FD, Bailey, Li, Mao, and Zhong (2003) andFrancis, Nanda, and Wang (2006) found no change in volatility after Regulation FD tookeffect Analogous to studies on return volatility, no consensus was reached on whetheranalysts’ forecasts suffered or improved after Regulation FD Heflin et al., Bailey et al.,and Francis et al found no significant impact of Regulation FD, whereas Agrawal,Chadha, and Chen (2006) and Mohanram and Sunder (2006) documented deterioration.Similarly, Heflin et al and Francis et al reported no change in forecast dispersion afterRegulation FD, whereas Bailey et al., Irani and Karamanou (2003), and Agrawal et al.found forecast dispersion to have increased Results on corporate disclosure were alsomixed Heflin et al and Bailey et al documented an increase in conference call fre-quency, and Irani (2004) found that conference calls became more useful in helping ana-lysts increase forecast accuracy In contrast, Bushee, Matsumoto, and Miller (2004) foundthat corporate conference call policy did not change after Regulation FD

Studies examining Regulation FD’s impact on information asymmetry reached ences ranging from an increase to a decrease in information asymmetry after Regulation

infer-FD took effect On one hand, Eleswarapu et al (2004), Chiyachantana, Jiang,Taechapiroontong, and Wood (2004), and Ahmed and Schneible (2007) documented evi-dence consistent with an improvement after Regulation FD On the other hand, Sidhu,Smith, Whaley, and Willis (2008) and Gomes, Gorton, and Madureira (2007) reported evi-dence suggesting an increase in information asymmetry Finally, Charoenrook and Lewis(2009) and Collver (2007) found no change in information asymmetry

Another strand of literature examined the informativeness of analyst reports before andafter Regulation FD took effect Gintschel and Markov (2004) studied the value of ana-lysts’ information outputs using the return volatility surrounding analysts’ announcements.They found that the absolute price impact of financial analysts’ forecasts and recommenda-tions declined by 28% after Regulation FD consistent with Regulation FD having curtailedselective disclosure Cornett, Tehranian, and Yalcin (2007) provided further evidence byevaluating the impact of Regulation FD on affiliated versus unaffiliated analysts Theirresults suggested that the market reaction to affiliated analysts’ recommendation changesdecreased significantly after the passage of Regulation FD Francis et al (2006) providedsupporting evidence to Gintschel and Markov’s (2004) results using American DepositaryReceipt (ADR) firms to control for confounding events

The prior literature provides extensive evidence on the effect of Regulation FD on thepreearnings announcement informational efficiency, analysts’ forecast accuracy and disper-sion, the informativeness of analysts’ reports, and corporate disclosure However, little isknown about Regulation FD’s impact on the usefulness of analysts’ advice in identifyingearnings surprises The unfair disclosure of information causing some investors to takepositions prior to earnings announcements with advance knowledge of the outcome wasone of the central issues in the debate surrounding Regulation FD We contribute to the

literature by examining the extent to which analysts’ recommendation revisions were useful

in predicting earnings surprises in the pre- and post-Regulation FD periods and estimatingthe level of abnormal returns that investors could have earned by following analysts’recommendations preceding earnings announcements

Analysts disclose forecasts and recommendations in their reports to their clients.Forecasts represent analysts’ predictions of various financial statement line items (e.g.earnings, sales) and do not necessarily convey information about analysts’ assessments ofcompanies’ intrinsic values relative to their stock prices (e.g., overvalued or underva-lued) Conversely, recommendation ratings represent direct indication of analysts’ assess-ment of the valuation of the company Consistently, investors react more strongly to

who intend to give early warnings to their clients about earnings announcements aremore likely to communicate this through their recommendation ratings rather than theirearnings forecasts

Furthermore, analysts can more effectively communicate information they receivedthrough selective disclosure via recommendation ratings rather than earnings forecastsbecause they may not have received a precise forecast from management Many analystsargued against Regulation FD on the basis that they only receive ‘‘soft’’ information frommanagers, which could not be communicated to the public in a form other than selectivedisclosure Managers are likely to be reluctant to give precise figures about upcomingearnings to analysts via selective disclosure because they themselves may not have preciseknowledge of upcoming earnings at the time Anecdotal evidence suggests that managersoften limited their communications to general directional guidance such as ‘‘earningsare likely to be better than expected’’ or ‘‘the current earnings expectations are unrealis-

are likely to limit their earnings forecast revisions In contrast, through recommendations,analysts can signal upcoming negative or positive earnings news without giving precise infor-mation about the degree of the earnings surprise

If analysts, as alleged, were privately communicating earnings-related information withmanagers and providing consistent investment advice to their clients, then analysts’ recom-mendations should have possessed predictive power of upcoming earnings surprises duringthe pre-Regulation FD period Furthermore, to the extent that Regulation FD was successful

in reducing selective disclosure, the association between recommendation revisions andearnings surprises should have weakened in the post-Regulation FD period

Hypothesis 1 (H1): The association between changes in analysts’ recommendationsand subsequent earnings surprises declined after Regulation FD took effect

Supporters of Regulation FD argued that selective disclosure helped select analysts andtheir clients reap economic benefits at the expense of other investors If, as alleged, analystsguided their clients to earn profits based on the private earnings guidance they receivedfrom management, we should observe significant abnormal returns associated with imple-menting a trading strategy that follows analysts’ recommendations and liquidates after earn-ings announcements Furthermore, to the extent that Regulation FD limited earnings-relatedselective disclosure, we should observe a reduction in the abnormal performance of thetrading strategy after Regulation FD took effect

Hypothesis 2 (H2): The profitability of a trading strategy intended to capture the vate earnings-related information conveyed by recommendation revisions declinedafter Regulation FD took effect.

pri-Research Design

The period before an earnings announcement corresponds to a time in which firms arelikely to have prepared financial statements and managers have the greatest knowledge ofthe current quarter’s earnings If firm executives selectively disclose earnings-related infor-mation to analysts, then this information is most likely to be privately communicated toanalysts during the period before earnings announcements

To isolate recommendations that may be associated with analysts’ communications withmanagers about upcoming earnings, we limit our sample to analysts’ recommendation revi-

We then examine Regulation FD’s impact on these recommendations’ predictive value ofupcoming earnings surprises by estimating and comparing the association between recom-mendation revisions and subsequent earnings surprises during the pre- and post-Regulation

the timeline and the main tests of this study

Univariate Analysis

The univariate analysis examines the mean earnings surprise that follows upgrades anddowngrades and tests whether a significant change is evident after Regulation FD tookeffect For robustness, earnings surprise is computed using four alternative methods:(a) standardized unexpected earnings based on time-series expectations, (b) standardizedunexpected earnings based on analyst expectations, (c) 3-day earnings announcementabnormal returns, and (d) 2-day earnings announcement abnormal returns

Standardized unexpected earning based on time-series expectation is computed asfollows:

st;t8

deciles are constructed and transformed to range between 20.5 and 10.5 The construction

of SUE deciles controls for common marketwide effects and reduces the influence ofextreme values on the results

Standardized unexpected earning based on analysts’ expectations is computed asfollows:

ASUEit ¼et ^et

pt

;

defined as the median of all analysts’ latest earnings forecasts made after the previous

ASUE deciles are constructed and transformed to range between 20.5 and 10.5

Unexpected Earnings Difference (Time-Series)

Unexpected Earnings Difference (Analyst)

Earnings Surprise

Return Difference CAR (0,1) CAR (– 1,1)

Unexpected Earnings Difference (Time-Series)

Unexpected Earnings Difference (Analyst)

Firm generates revenues

and incurs expenses

Firm generates revenues

and incurs expensesFirm generates revenues and incurs expenses

Firm generates revenues

and incurs expenses

Fiscal Quarter Ends

Fiscal Quarter Ends

Management obtains a rough idea of upcoming earnings.

Potential period for related selective disclosure.

earnings-Management obtains a rough idea of upcoming earnings.

Potential period for related selective disclosure.

earnings-Management obtains a rough idea of upcoming earnings.

Potential period for related selective disclosure.

earnings-Management obtains a rough idea of upcoming earnings.

Potential period for earnings- related selective disclosure Earnings

are reported

Earnings are reported

Analysts receiving related selective disclosure revise recommendations.

earnings-(-22,-2)

Upgrades & Downgrades

Analysts receiving earnings- related selective disclosure revise recommendations.

(–22,–2)

Upgrades and Downgrades

Earnings Surprise

1 CAR

2 SUE

3 UE

Earnings Surprise

1 CAR

2 SUE

3 UE

Figure 1 Research design

Panel A: Timeline of the empirical analysis

Panel B: Association test for the pre- and post-Regulation periods

Note: This figure provides a visual overview of the research design Panel A illustrates which analystrecommendation revisions are included in the analysis in relation to the firm’s quarterly cycle Panel Bshows the main test of the article and lists which earnings surprise measures are used

Earnings announcement abnormal returns are computed as follows:

m¼1

for Research in Security Prices (CRSP) New York Stock Exchange (NYSE)/American

For each quarter, we compute the mean earnings surprise for upgraded and downgradedfirms As recommendation revisions can be driven by marketwide information that affectsnumerous firms, earnings surprises that follow revisions may be correlated across firms.Therefore, we follow the Fama and Macbeth (1973) procedure and first compute averagequarterly earnings surprises and then calculate time-series averages and t-statistics of quar-terly average surprises

Regression Analysis

The regression analysis allows us to examine the change in recommendations’ predictivevalue while controlling for confounding factors To measure the change in the associationbetween recommendation revisions and earnings surprises, we estimate the followingregression models:

ð1Þ

ð2Þ

ð3Þ

ð4Þ

In the regression analyses above, we regress alternative earnings surprise measures onthe recommendation revisions made during the 3-week period before earnings announce-ments (REV), a Regulation FD indicator variable that takes a value of 1 for calendar quar-ters during the post-Regulation FD period (FD) and the interaction of REV and FDvariables (REV 3 FD) The coefficient of the REV variable tests whether recommendation

revisions have predictive power of subsequent earnings surprises The interaction of theREV and FD variables, labeled REV 3 FD, tests whether the association between analysts’revisions and subsequent earnings surprises declined after Regulation FD.

In addition, we examine whether the Sarbanes-Oxley Act had a significant impact onthe association between recommendation revisions and earnings surprises by including

a dummy variable labeled SOX that takes a value of 1 for calendar quarters ending afterDecember 31, 2003, and the interaction of this variable with the recommendation revisionvariable (REV 3 SOX) The coefficient of REV 3 SOX tests whether the predictive value

of recommendation revisions changed after the Sarbanes-Oxley Act took effect

There are several potential confounding factors that may be correlated with analysts’revisions Bernard and Thomas (1989) found that stock prices underreact to earningsannouncements and this leads to a postearnings-announcement drift that is concentratedaround subsequent earnings announcements Therefore, it is possible that analysts revisetheir recommendation ratings in reaction to the previous quarter’s earnings results ratherthan to their own information acquisition efforts or selective disclosure To control for thepotential effect of postearnings announcement drift, we include the previous quarter’s earn-ings announcement market-adjusted returns (LANCRET) in the empirical model To theextent that there is an underreaction to the previous quarter’s earnings announcement,LANCRET will be positively correlated with the current quarter’s earnings surprise

Furthermore, Jegadeesh and Titman (1993) found that past winners outperform pastlosers Analysts aware of the positive association between past and future returns mayrevise their recommendations accordingly, and this can result in an association betweenrevisions and subsequent earnings surprises To control for the momentum effect, we mea-sure the 3-month buy-and-hold return ending in the 2nd month of the firm quarter (LRET)

In addition, Sloan (1996) found evidence suggesting that investors fixate on bottomline earnings and ignore the accruals component of earnings He finds that accruals arenegatively associated with the subsequent year’s abnormal returns and demonstrates thatthe mispricing is mainly corrected during subsequent quarterly earnings announcements

As previously announced accruals are negatively associated with subsequent earningsannouncement returns, analysts may be revising their recommendation ratings in response

to past accruals rather than to their private communication with management Therefore,

we also control for the accrual component in the regression analysis We compute

The results are similar when we alternatively use discretionary accruals estimated fromthe modified Jones model (Dechow, Sloan, & Sweeney, 1995), or from the performance-augmented modified Jones model (Kothari, Leone, & Wasley, 2005), or using statement

of cash flows data

Finally, Ali, Durtschi, Lev, and Trombley (2004) document that institutional investorshave superior knowledge of upcoming earnings results and that institutional trades are posi-tively correlated with future earnings announcement returns Again, analysts’ revisions may

be correlated with subsequent earnings surprises because analysts simply respond to temporaneous institutional trading To control for this possibility, we measure the change

con-in con-institutional ownership durcon-ing the most recent calendar quarter (CHNG_IO) and con-include

it in the regression model Table 2 provides detailed descriptions of the variables used inthe regression analysis

Trading Strategy Analysis

The trading strategy analysis involves constructing a portfolio that aims to capture mal returns earned by investors who followed recommendations with the intent of exploit-ing analysts’ earning-related private information The trading strategy focuses onrecommendation revisions made after the fiscal quarter-end date and before the earningsannouncement date The hedge portfolio purchases (sells) shares of upgraded (downgraded)firms 1 day after the revision date and holds these shares until 1 day after the earnings areannounced In this strategy, all dates are known in event time; hence, no hindsight bias isintroduced to the analysis

abnor-To estimate abnormal returns, we first compute the daily raw returns that accrue toupgrade and downgrade portfolios Each firm that is upgraded after the fiscal quarter-end

Table 2 Variable Definitions

SSUE Quarterly decile of unexpected earnings defined as (et2 et–4) scaled by the

standard deviation of unexpected earnings (st,t–8) e is basic earnings per shareexcluding extraordinary items (epsfxq), adjusted for stock splits and stockdividends The variable is transformed to range between 20.5 and 10.5SASUE Quarterly decile of unexpected earnings defined as actual earnings reported by IBES

minus median earnings estimate of analysts scaled by the price at the end of theprevious fiscal quarter The deciles are transformed to range between 20.5 and10.5

CAR (21, 11) Cumulative market-adjusted returns during the 3-day period centered on the

earnings announcement date The CRSP NYSE/AMEX/NASDAQ value-weightedindex return is used as the market return Returns are adjusted for delistingCAR (0, 11) Cumulative market-adjusted returns during the 2-day period beginning on the

earnings announcement date The CRSP NYSE/AMEX/NASDAQ value-weightedindex return is used as the market return Returns are adjusted for delistingREV Recommendation revisions during the 3-week period ending 2 days before the

earnings announcement date

FD An indicator variable that takes a value of 1 for fiscal quarters ending after

October 23, 2000, and 0 for prior fiscal quarterSOX An indicator variable that takes a value of 1 for fiscal quarters ending after

December 31, 2003, and 0 for other quartersLANCRET Previous fiscal quarter’s earnings announcement return, which is defined as the

cumulative market-adjusted return during the 3-day period centered on theearnings announcement date The CRSP NYSE/AMEX/NASDAQ value-weightedindex return is used as the market return Returns are adjusted for delistingLRET The 3-month buy-and-hold return ending at the end of the fiscal quarter’s 2nd

monthACCR DCA – DCL – DEP scaled by average total assets where DCA is the change in

current assets (act) minus the change in cash and short-term investments (che),DCL is the change in current liabilities (lct) minus the sum of changes in debt incurrent liabilities (dlc) and income taxes payable (txp), and DEP is depreciationand amortization (dp)

CHNG_IO The change in institutional ownership percentage compiled from the

Thomson-Reuters Institutional Holdings (13F) databaseNote: This table lists and defines the variables used in the study The first column reports the variable name and the second column provides the definition.

enters the upgrade portfolio 1 day after the revision date and remains in the portfolio until

1 day after the firm announces its quarterly earnings results Similarly, each firm that isdowngraded after the fiscal quarter-end enters the downgrade portfolio 1 day after the revi-sion date and remains in the portfolio until 1 day after the firm announces its quarterlyearnings results We calculate value-weighted daily returns for each portfolio as follows:

Rpt ¼nXp;t1

j¼1

xjt1Rj;t;

market capitalization of all the firms in the portfolio The daily portfolio returns are thencompounded to monthly returns:

the portfolio on day t

We construct a hedge portfolio that goes long on the upgrade portfolio and short on thedowngrade portfolio The hedge portfolio’s return is equal to the difference between thereturns of the upgrade and downgrade portfolios Finally, we subtract the risk-free ratefrom the upgrade and downgrade portfolios to compute the excess returns of the twoportfolios.10

We estimate the average monthly abnormal return associated with each portfolio by mating the four-factor model represented by the equation below The intercept of this equa-tion (Jensen’s a) serves as an estimate of the average monthly abnormal return:

where Mkt is the market risk premium that is equal to the market return minus the risk-freerate, SMB is the average return on three small-market capitalization portfolios minus theaverage return on three large-market capitalization portfolios on day t, HML is the averagereturn on two high book-to-market equity portfolios minus the average return on two lowbook-to-market equity portfolios for day t, and UMD is the average of the returns on two(big-sized and small-sized) high prior return portfolios minus the average of the returns ontwo low prior return portfolios, where a big-sized company is identified as being largerthan the median NYSE market cap

Sample

The initial sample consists of all firms traded in the NYSE and AMEX and NASDAQ thathave data available in both CRSP and Compustat files The sample spans the fiscal quartersbetween 1995Q1 and 2009Q2 The sample begins in 1995Q1 because the InstitutionalBrokers’ Estimate System (IBES) recommendation file is sparse for the period before 1994,and we require 1 year of prior recommendation data to compute analysts’ previous recom-mendations The sample ends in 2009Q2 because that is the latest fiscal quarter for which

we have data on Compustat files We exclude closed-end funds, investment trusts, units,

and foreign companies We also exclude firms with share prices less than US$1 at the end

of the previous quarter to avoid outliers from biasing the results

Analyst recommendations ratings are obtained from IBES (ibes.recddet), and

revision as the action of a particular analyst to change his or her prior recommendationrating If the recommendation is revised to a more favorable one, we identify it as anupgrade If the recommendation is revised to a less favorable one, we identify it as

Analysts’ quarterly earnings forecasts are obtained from the IBES-unadjusted detail file

adjustment factors when necessary To compute analyst expectations, we retain the lastquarterly earnings forecast made by each analyst before the earnings announcement date

Finally, we obtain institutional ownership data from the Thomson-Reuters InstitutionalHoldings (13f) database We use the Wharton Research Data Services (WRDS) re-createdshares outstanding to compute institutional ownership and exclude observations where thefiling and reporting dates are not equal to avoid erroneous observations from entering thesample

Table 3 reports the descriptive statistics of the sample employed in this study The finalsample consists of 41,833 firm quarters Due to numerous data requirements (CRSP,

Table 3 Descriptive Statistics

observa-at fiscal quarter-end, PRC is share price observa-at fiscal quarter- end, and BETA is the market model beta computed using

60 months of prior return data All other variables are defined in Table 2.

Compustat, IBES, and TFN) imposed by the research design, the final sample is composed

of relatively large firms The average firm in the sample has a market capitalization ofUS$8.3 billion with a mean share price of US$30 and an average book-to-market ratio of0.481

Table 4 reports the correlation matrix of the variables used in the multiple regressionanalysis The four earnings surprise measures, SSUE, SASUE, CAR (-1, 1), and CAR (0, 1),are positively correlated because they all intend to capture the earnings surprise component.The REV variable, which represents the recommendation revisions during the 3-weekperiod before the earnings announcement period, is positively correlated with earnings sur-prise measures This is consistent with recommendation revisions having predictive power

of subsequent earnings surprises The interaction of REV and FD is also positively lated with earnings surprise measures but at a weaker level, which suggests that the associ-ation between recommendations and earnings surprises declined after Regulation FD tookeffect The LANCRET variable, which is the previous quarter’s 3-day earnings announce-ment return, is positively correlated with the four earnings surprise measures consistentwith the postearnings announcement drift documented in Bernard and Thomas (1989).Furthermore, as suggested by the momentum effect (Jegadeesh & Titman, 1993), LRET ispositively correlated with the earnings surprise measures The accrual component (ACCR),

corre-as shown in Sloan (1996), is negatively correlated with subsequent earnings surprises TheCHNG_IO variable is positively correlated with subsequent earnings surprises with theexception of CAR (0, 1) In short, the control variables employed in the multiple regressionanalysis are correlated with the earnings surprise measures consistent with the results docu-mented in the prior literature Finally, we do not find a strong correlation among the inde-

Empirical Results

Univariate Analysis

The univariate analysis results suggest that recommendation revisions were useful in dicting upcoming earnings surprises during the pre-Regulation FD period On average,upgraded firms reported earnings above expectations and downgraded firms reported earn-ings below expectations Table 5 Panel A reports that before Regulation FD took effect, theearnings surprise differential between upgraded and downgraded firms was 0.0531 based

pre-on time-series expectatipre-ons (2nd column) and 0.0855 based pre-on analyst expectatipre-ons (3rdcolumn) The average earnings announcement return differential between upgraded anddowngraded firms was 0.93% based on 3-day returns (4th column) and 0.67% based on2-day returns (5th column) These results suggest that investors who followed analysts’advice in the pre-Regulation FD period were able to benefit from subsequent earnings sur-prises Overall, the results suggest that some form of information acquisition or interpreta-tion either through selective disclosure or effort was taking place These results areconsistent with pre-Regulation FD concerns that analysts were receiving early peeks atearnings results

In the post-Regulation FD period, a weaker association between recommendation sions and earnings surprises is evident The mean earnings surprise differentials betweenupgraded and downgraded firms based on time-series and analyst expectations decline to0.0279 and 0.0399, respectively Similarly, the mean 3- and 2-day earnings announcementreturn differentials between upgraded and downgraded firms decline to 0.42% and 0.24%,

respectively The decline in the strength of the relationship between recommendation sions and earnings surprises is statistically significant The final row of Panel A reports thatchanges in time-series and analyst-based earnings surprise differentials are 20.0252 and20.0456, respectively Both changes are statistically and economically significant Finally,results based on earnings announcement returns also indicate a significant decline afterRegulation FD Table 5 Panel A reports that the mean 3-day market-adjusted earningsannouncement returns declined 55% from 0.93% to 0.42% and the mean 2-day market-adjusted earnings announcement returns declined 64% from 0.67% to 0.24%.

revi-Table 5 Recommendation Revisions and Subsequent Earnings Surprises

Period

Time-seriesexpectations

Analystexpectations CAAR (21, 11) CAAR (0, 11)Panel A: Upgrades–Downgrades

***, **, and * denote significance at the 1%, 5%, and 10% significance levels, respectively.

Overall, the univariate results in Table 5 Panel A reveal that the association betweenrevisions and earnings surprises weakened substantially after the Regulation FD period.These results are consistent with Regulation FD having significantly reduced analysts’power in predicting earnings surprises In the post-Regulation FD period, analyst recom-mendations appear to be less useful in guiding investors to firms that later experience earn-ings surprises.

Analysis of revisions separately by upgrades and downgrades yields similar results.Table 5 Panels B and C report that the mean time-series-based earnings surprise declined45% from 0.0339 to 0.0185 for upgraded firms and increased 51% from 20.0192 to20.0094 for downgraded firms Although neither change is statistically significant in itsown, the combination of the two changes is significant Results based on analyst expecta-tions and market-adjusted returns are similar and support the conclusion that recommenda-tion revisions’ usefulness in identifying earnings surprises declined after Regulation FD

Regression Analysis

The regression analyses examining the association between recommendation revisions andearnings surprises controlling for confounding factors also reveal a significant decline in thepredictive value of recommendation revisions after Regulation FD took effect Table 6reports the estimation results of Equation 1, which investigates the impact of Regulation FD

on the association between recommendation revisions and time-series-based unexpected ings The coefficient of REV is positive and significant in Model I, suggesting recommenda-tion revisions to have predictive value of upcoming earnings surprises Consistent with thedownturn in the economy, the FD indicator variable, which takes a value of 1 for fiscal quar-ters after Regulation FD took effect, is negative The hypothesis variable, REV 3 FD, is20.013 and significantly negative, implying that the association between recommendationrevisions and earnings surprises declined significantly after Regulation FD took effect TheSOX indicator variable is significantly positive, indicating an improvement in earnings sur-prises as markets recovered from the tech bubble burst starting in 2004 However, differentfrom the results based on the REV 3 FD variable, REV 3 SOX is estimated to be insignifi-cant, suggesting that the information dynamics relating to analysts’ recommendation revisionsdid not change significantly after the Sarbanes-Oxley Act took effect

earn-In Model II, we include the LANCRET and LRET variables to control for announcement drift and momentum effects The LANCRET variable is the earningsannouncement return in the previous quarter, and it is estimated to have a positive coeffi-cient This is consistent with the prior literature and suggests that firms that experiencedearnings surprises in the prior quarter continue to do so in the next quarter The positivecoefficient on the LRET variable that controls for the past 3-month buy-and-hold returnsuggests that firms that experienced superior market performance during the recent monthsexhibited positive earnings surprises in the subsequent quarter In Model II, REV, FD, andREV 3 FD coefficients are estimated to be 0.022, 20.008, and 20.013, respectively REVand REV 3 FD are both statistically significant The negative coefficient on REV 3 FDindicates that the extent to which recommendation revisions predicted earnings surprisesdeclined after Regulation FD took effect The REV 3 SOX variable is insignificant anddoes not point to a significant impact on the association between recommendation revisionsand earnings surprises that can be attributed to the Sarbanes-Oxley Act

postearnings-In Model III, we control for accruals with the ACCR variable The ACCR variable is tistically significant, and its coefficient is estimated to be 20.089 This is consistent with the

sta-prior literature and suggests that firms that had income increasing accruals during the ous fiscal year reported disappointing earnings results during the subsequent fiscal quarter.Results relating to the impact of Regulation FD on the predictive value of recommendationrevisions are unchanged by the inclusion of the ACCR variable The REV, FD, and REV 3

previ-FD variables are all statistically significant, and their coefficients are estimated to be 0.023,20.015, and 20.012, respectively These results indicate a significant decline in the predic-tive value of recommendation revisions after the enactment of Regulation FD

Finally, in Model IV, we incorporate the change in institutional ownership to our cal model to control for the positive association between institutional trading activity andearnings surprises While the CHNG_IO variable is estimated to have a positive coefficient,

empiri-Table 6 Regression Analysis of Subsequent Earnings Surprises (Based on Time-Series Expectations)

Note: This table reports the estimation results of the empirical model:

where SSUE is the standardized unexpected earnings decile based on a time-series earnings expectation model REV is the recommendation revision during the preearnings announcement period, FD is a post-Regulation FD indi- cator variable, and REV 3 FD is the interaction of REV and FD variables SOX is a Sarbanes-Oxley indicator variable and REV 3 SOX is the interaction of REV and SOX variables LANCRET is the previous quarter’s 3-day earnings announcement return (market adjusted), and LRET is the past 3-month buy-and-hold return ending a month before the fiscal quarter-end ACCR is total accruals scaled by average total assets as in Sloan (1996), and CHNG_IO is the change in percentage ownership during the most recent calendar quarter Four specifications of the above model are estimated t-statistics based on firm clustered standard errors are reported in parentheses.

***,**, and * denote significance at the 1%, 5%, and 10% significance levels, respectively.

it is not statistically significant The REV, FD, and REV 3 FD variables confirm the ences based on Models I to III The predictive value of recommendation revisions appears

infer-to be significantly lower after Regulation FD infer-took effect

We repeat the regression analysis using unexpected earnings based on analyst tions and report the estimation results in Table 7 The REV variable is estimated to have

expecta-a coefficient of 0.042 expecta-and suggests expecta-a significexpecta-antly positive expecta-associexpecta-ation between dation revisions and earnings surprises The FD variable is estimated to be 0.003, which isinsignificant, suggesting no change in the average level of earnings surprises afterRegulation FD The interaction variable REV 3 FD is estimated as 20.017, which is

recommen-Table 7 Regression Analysis of Subsequent Earnings Surprises (Based on Analyst Expectations)

Note: This table reports the estimation results of the empirical model:

***,**, and * denote significance at the 1%, 5%, and 10% significance levels, respectively.

statistically significant, implying a significant reduction in the association between mendation revisions and earnings surprises The SOX and REV 3 SOX variables reveal

recom-a further decline in the predictive vrecom-alue of recommendrecom-ation revisions recom-after the Srecom-arbrecom-anes-Oxley Act

Sarbanes-In Model II, we control for postearnings-announcement drift and momentum effects andfind largely similar results Both factors, LANCRET and LRET, have the expected signs.The interaction of REV and FD is significantly negative, suggesting a decline in the associ-ation in the predictive value of recommendation revisions Different from Model I, inModel II we do not observe evidence suggesting a significant reduction in the predictivevalue of recommendation revisions after the Sarbanes-Oxley Act took effect

In Model III, we control for the accruals anomaly by incorporating the ACCR variableinto the empirical model The ACCR variable is estimated to be significantly negative TheREV 3 FD is negative, implying a significant decline in the predictive value of recom-mendation revisions The REV 3 SOX variable is insignificant, suggesting no substantialchange in the predictive value of recommendation revisions after the Sarbanes-Oxley Act

In Model IV, we control for changes in institutional ownership The estimated cient on the CHNG_IO is positive, as expected Firms with increased institutional owner-ship appear to experience positive earnings surprise This is consistent with institutionalinvestors anticipating earnings surprises The REV 3 FD coefficient indicates a significantdecline in the predictive value of recommendations after Regulation FD We do not identifyany change in the association between recommendation revisions and earnings surprisesafter the Sarbanes-Oxley Act

coeffi-Table 8 reports the estimation results of Equation (3), where 3-day market-adjusted ings announcement returns are regressed on recommendation revisions and confoundingfactors The results echo the previous findings from the time-series and analyst-based earn-ings surprise measures The REV 3 FD interaction variable is significantly negative, sug-gesting a substantial decline in the association between recommendation revisions andsubsequent earnings surprises Finally, Table 9 reports the regression analysis of 2-daymarket-adjusted earnings announcement returns The results are consistent with prior resultsand confirm the existence of a significant decline in the association between recommenda-tion revisions and earnings surprises in the post-Regulation FD period

earn-Overall, the regression analysis results suggest a significant reduction in tion revisions’ ability to identify earnings surprises after Regulation FD took effect Theresults show that the level and change in the predictive value of recommendation revisionscannot be explained by confounding factors that were also shown to exhibit predictivepower of earnings surprises In conclusion, during the post-Regulation FD period, analystsappear to be less successful in guiding their clients to firms that experience positive earn-ings surprises and warning their clients away from firms that experience negative earningssurprises These results are consistent with a reduction in selective disclosure, which was

recommenda-a mrecommenda-ajor source of informrecommenda-ation for recommenda-anrecommenda-alysts during the pre-Regulrecommenda-ation FD period

Trading Strategy Analysis

During the pre-Regulation FD period, the trading strategy that followed analysts’ advice tocapture earnings surprises appears to have earned significant abnormal returns controllingfor market risk, size, book-to-market, and momentum effects Table 10 reports the perfor-mance of the hedge, upgrade, and downgrade portfolios The results suggest that investorswho followed analysts’ recommendation revisions with the intent of capturing earnings

surprises earned an average monthly abnormal return of 4.6% before transaction costsduring the pre-Regulation FD period The average abnormal returns associated withupgrade and downgrade portfolios are 1.9 and 22.7%, respectively Both abnormal returnestimates are statistically significant and consistent with investor complaints relating to cer-tain select investors receiving privileged access to management via financial analysts.During the post-Regulation FD, a considerable reduction in the performance of the port-folios is evident The performance of the hedge portfolio declines 70% from 4.6% to 1.4%.Although the hedge portfolio’s abnormal performance is statistically significant in the post-Regulation FD period, it is substantially lower Similarly, we see a 21% reduction in the

Table 8 Regression Analysis of Subsequent Earnings Surprise (Based on 3-Day Market-AdjustedReturns)

Note: This table reports the estimation results of the empirical model:

where CAR(21,1) is the 3-day market-adjusted earnings announcement return REV is the recommendation sion during the preearnings announcement period, FD is a post-Regulation FD-indicator variable, and REV 3 FD is the interaction of REV and FD variables SOX is a Sarbanes-Oxley indicator variable, and REV 3 SOX is the interac- tion of REV and SOX variables LANCRET is the previous quarter’s 3-day earnings announcement return (market adjusted), and LRET is the past 3-month buy-and-hold return ending a month before the fiscal quarter-end ACCR is total accruals scaled by average total assets as in Sloan (1996), and CHNG_IO is the change in percentage owner- ship during the most recent calendar quarter Four specifications of the above model are estimated t-statistics based on firm clustered standard errors are reported in parentheses.

revi-***, **, and * denote significance at the 1%, 5%, and 10% significance levels, respectively.

performance of the upgrade portfolio, decreasing from 1.9% to 1.5% Finally, the most matic reduction is evident in the performance of the downgrade portfolio During the post-Regulation FD period, abnormal returns associated with downgrades vanish.

dra-Overall, our results are consistent with analysts’ recommendation revisions conveyingless information about upcoming earnings surprises in the post-Regulation FD Thereduction in the private information transmitted via recommendation revisions appears tomanifest itself as a reduction in the association between recommendation revisions and

a reduction in the performance of the portfolios constructed surrounding recommendationrevisions and earnings announcements The results support the conclusion that analysts’

Table 9 Regression Analysis of Subsequent Earnings Surprise (Based on 2-Day Market-AdjustedReturns)

Note: This table reports the estimation results of the empirical model:

where CAR(0, 11) is the 2-day market-adjusted earnings announcement return beginning on the earnings announcement date REV is the recommendation revision during the preearnings announcement period, FD is

a post-Regulation FD-indicator variable, and REV 3 FD is the interaction of REV and FD variables SOX is

a Sarbanes-Oxley indicator variable, and REV 3 SOX is the interaction of REV and SOX variables LANCRET is the previous quarter’s 3-day earnings announcement return (market adjusted) LRET is the past 3-month buy-and-hold return ending a month before the fiscal quarter-end ACCR is total accruals scaled by average total assets as in Sloan (1996), and CHNG_IO is the change in percentage ownership during the most recent calendar quarter Four specifications of the above model are estimated t-statistics based on firm-clustered standard errors are reported

in parentheses.

***,**, and * denote significance at the 1%, 5%, and 10% significance levels, respectively.

recommendation revisions are less useful in helping analysts’ clients earn abnormalprofits.

Sensitivity Analysis

In addition to confounding factors, analysts may also be revising their recommendations

in response to key corporate developments To the extent that the impact of corporatedevelopments is less during the post-Regulation FD period, there may be a change in theassociation between revisions and earnings surprises that is due to the impact of corporatedevelopments, rather than due to the information that analysts convey through their rec-ommendation revisions To control for this possibility, we use a hand-collected publicinformation arrival database and exclude revisions that took place within a 3-trading-daywindow surrounding dates of Wall Street Journal news articles during the period 1995-

2006 and find that the results are largely unaffected when we add this additionalcontrol.15

Conclusion

This article examines the extent to which analyst recommendations are useful in identifyingfirms that experience earnings surprises during the pre- and post-Regulation FD periods.The empirical analysis suggests a significant decline in the predictive value of recommen-dation revisions after Regulation FD took effect Prior to Regulation FD, upgraded firms

Table 10 Portfolio Performance

a time-series regression of each portfolio’s excess monthly return on the four-factor returns: excess market return (b—4th column), a zero-investment size portfolio (SMB—5th column), a zero-investment book-to-market portfo- lio (HML—6th column), and a zero-investment momentum portfolio (UMD—7th column) The intercept of this

in parentheses.

***, **, and * denote significance at the 1%, 5%, and 10% significance levels, respectively.

outperformed downgraded firms by 0.93% on earnings announcements After Regulation

FD, the earnings announcement return differential between upgraded and downgradedfirms declined by 55% to 0.42% In addition, we find that the abnormal returns associatedwith following analysts’ recommendation revisions made shortly in advance of earningsannouncements declined considerably following the enactment of Regulation FD The port-folio analysis reveals that the trading strategy yielded an average monthly abnormal return

of 4.6% during the pre-Regulation FD period compared with 1.4% during the Regulation FD period

post-The passage of Regulation FD was preceded with intense objection and frustration byinvestors that certain analysts and institutional investors were enjoying privileged access tomaterial nonpublic information Disclosure of earnings-related information was one of themost scrutinized practices of selective disclosure Investors complained that managers weregiving private information to analysts about upcoming earnings results and analystswere then advising their clients accordingly The empirical results provided in this studysuggest a considerable decline in analysts’ ability to guide their clients to and away fromfirms that are likely to experience earnings surprises These results are consistent with

a decline in selective disclosure and suggest that Regulation FD was effective in achievingits objective

Authors’ Note

We are grateful to the editor Kashi R Balachandran, Jean Bedard, Yoel Beniluz, Mahendra R.Gujarathi, Rani Hoitash, James Hunton, Bikki Jaggi, Jay Thibodeau, an anonymous reviewer, andseminar participants at Bentley University and Rutgers University for their valuable comments andsuggestions We thank Tesfalidet Tukue for excellent research assistance We also thank theWhitcomb Center for Research in Financial Services for providing research support through use ofthe WRDS system All errors are the authors’ responsibility

Declaration of Conflicting Interests

The author(s) declared that they had no conflicts of interests with respect to their authorship or thepublication of this article

Funding

The author(s) received no financial support for the research and/or authorship of this article

Notes

1 Selective Disclosure and Insider Trading, Release No 33-7881, August 15, 2000

2 Forecasts also correspond to an important aspect of analysts’ reports; however, they do not vide investors direct advice regarding buy or sell decisions In contrast, recommendation revi-sions are a direct way in which analysts advise their clients to buy or sell shares

pro-3 Francis and Soffer (1997) find that stock recommendation revisions contain information mental to the information in earnings forecast revisions

incre-4 Our interviews with analysts and cases of Regulation FD violations (e.g., http://www.sec.gov/litigation/admin/34-48461.htm and http://www.sec.gov/litigation/admin/34-46897.htm) confirmthis conclusion

5 The results remain qualitatively similar when we use a 2-week or 1-month period

6 In early 2002, the National Association of Securities Dealers (NASD) proposed Rule 2711 andthe New York Stock Exchange (NYSE) proposed a modification to its Rule 472,

Communications with the Public The Securities and Exchange Commission (SEC) approvedthese proposals on May 8, 2002 Barber, Lehavy, McNichols, and Trueman’s (2006) results indi-cate that the approved proposals had a significant change in the distribution of analysts’ buyrecommendations during the first two calendar quarters of the year 2002 As the changes in rat-ings during this period are likely to be due to changes in regulations, they may not necessarilyreflect changes in analysts’ private information For robustness, we replicate all analyses exclud-ing the first two calendar quarters of the year 2002 and find that the results remain qualitativelysimilar when we exclude these quarters.

7 Brown and Warner (1985), using simulations, demonstrate that the choice of benchmark modelleads to minor differences in abnormal returns when calculating short-term event window returnsusing daily returns As the computation of market-adjusted returns does not require preevent esti-mation period data and imposes the least data requirements, we present the empirical resultsbased on market-adjusted returns However, the results are similar when we alternatively com-pute abnormal returns using market, three-factor, or four-factor models as the normal return gen-erating models

8 We compute 3-month returns ending at the end of the 2nd month of the fiscal quarter to avoid

a potential correlation between analysts’ revisions during the preearnings announcement periodand returns Nevertheless, the results are similar when we compute the buy-and-hold returnduring the fiscal quarter

9 We assume that annual financial statements are filed within 3 months

10 We do not subtract the risk-free rate from the hedge portfolio as this portfolio is a self-financingportfolio

11 IBES does not provide a readily available revision variable Therefore, to compute tion revisions, analysts’ prior and current recommendation ratings are necessary We identify ananalyst’s prior recommendation rating by using IBES’s analyst code and finding the analyst’sprevious recommendation rating As IBES assigns a code of 000000 for all anonymous analysts,

recommenda-it is not possible to compute recommendation revisions for anonymous analysts Hence, we inate them from the sample In untabulated results, we assume that only one analyst in eachbroker covers the same firm and compute revisions based on the broker ID and obtain similarresults

elim-12 We use the unadjusted detail file as opposed to the summary or the adjusted detail files becausePayne and Thomas (2003) show that stock-split adjusted files do not have enough precision tounadjust the data without experiencing severe rounding errors

13 Only earnings estimates made after the previous quarter’s earnings announcement enter thesample

14 In addition to the correlation matrix, the variance inflation factors are computed to ascertainthere is no multicolinearity issue in the regression analysis

15 The Wall Street Journal news in our database relate to mergers and acquisitions, divestitures,appointments, executive changes, earnings, litigation, suspensions, dismissals, research anddevelopment investments, resignations, antitrust actions, drugs, bankruptcies, takeovers, frauds,FDA, inventions, patents, initial public offerings, products, strikes, insider trading, dividend,stock splits, equity issues, debt issues, and contracts

References

Agrawal, A., Chadha, S., & Chen, M A (2006) Who is afraid of Reg FD? The behavior and mance of sell-side analysts following the SEC’s fair disclosure rules Journal of Business, 79,2811–2834

perfor-Ahmed, A S., & Schneible, R A (2007) The impact of regulation fair disclosure on investors’ priorinformation quality—Evidence from an analysis of changes in trading volume and stock pricereactions to earnings announcements Journal of Corporate Finance, 13, 282–299

Ali, A., Durtschi, C., Lev, B., & Trombley, M (2004) Changes in institutional ownership and quent earnings announcement abnormal returns Journal of Accounting, Auditing & Finance, 19,221–248.

subse-Bailey, W., Li, H., Mao, C X., & Zhong, R (2003) Regulation fair disclosure and earnings tion: Market, analyst, and corporate responses Journal of Finance, 58, 2487–2514

informa-Barber, B., Lehavy, R., McNichols, M., & Trueman, B (2006) Buys, holds, and sells: The tion of investment banks’ stock ratings and the implications for the profitability of analysts’recommendations Journal of Accounting & Economics, 41, 87–117

distribu-Bernard, V L., & Thomas, J K (1989) Post-earnings-announcement drift: Delayed price response orrisk premium? Journal of Accounting Research, 27(Suppl.), 1–36

Brown, S J., & Warner, J B (1985) Using daily stock returns: The case of event studies Journal ofFinancial Economics, 14, 3–31

Bushee, B J., Matsumoto, D A., & Miller, G S (2004) Managerial and investor response to sure regulation: The case of Reg FD and conference calls Accounting Review, 79, 617–643.Charoenrook, A., & Lewis, C M (2009) Information, selective disclosure, and analyst behavior.Financial Management, 38, 39–57

disclo-Chiyachantana, C N., Jiang, C X., Taechapiroontong, N., & Wood, R A (2004) The impact of ulation fair disclosure on information asymmetry and trading: An intraday analysis FinancialReview, 39, 549–577

reg-Collver, C D (2007) Is there less informed trading after regulation fair disclosure? Journal ofCorporate Finance, 13, 270–281

Cornett, M M., Tehranian, H., & Yalcin, A (2007) Regulation fair disclosure and the market’s tion to analyst investment recommendation changes Journal of Banking & Finance, 31, 567–588.Dechow, P M., Sloan, R G., & Sweeney, A P (1995) Detecting earnings management AccountingReview, 70, 193–225

reac-Eleswarapu, V R., Thompson, R., & Venkataraman, K (2004) The impact of regulation fair sure: Trading costs and information asymmetry Journal of Financial and Quantitative Analysis,

Payne, J L., & Thomas, W B (2003) The implications of using stock-split adjusted I/B/E/S data inempirical research Accounting Review, 78, 1049–1067.

Sidhu, B., Smith, T., Whaley, R E., & Willis, R H (2008) Regulation fair disclosure and the cost ofadverse selection Journal of Accounting Research, 46, 697–728

Sinha, P., & Gadarowski, C (2010) The efficacy of regulation fair disclosure Financial Review, 45,331–354

Sloan, R G (1996) Do stock prices fully reflect information in accruals and cash flows about futureearnings? Accounting Review, 71, 289–315

Ó The Author(s) 2011 Reprints and permission: sagepub.com/journalsPermissions.nav DOI: 10.1177/0148558X11401555

http://jaaf.sagepub.com

Analyst Quality, Optimistic

Bias, and Reactions to Major

a firm As a result, the authors expect the asymmetric response to be reduced for superioranalysts Using the stock return/recommendation changes relationship, they find that theasymmetric reaction is less for analysts with characteristics that are indicative of higherquality Furthermore, the reduction is more pronounced for analysts in the top decile and

is only present when analysts have negative private information This article therefore tributes to the research on differing analyst characteristics and report quality, and providesadditional insights on analyst bias

asymmet-to large negative price shocks and effectively do not respond asymmet-to large positive price shocks.Other research examining analysts’ earnings forecast revisions provide evidence of analystoptimism as suggested by analysts’ tendency to underreact to prior bad news (see, forexample, Easterwood & Nutt, 1999; Mikhail, Walther, & Willis, 2003) However,what causes the optimistic bias in analysts’ reactions to major news has yet to be systemati-cally studied

An underlying presumption is that brokerage firms’ investment banking affiliationscreate incentives for analyst optimism (e.g., see Michaely & Womack, 1999; O’Brien,McNichols, & Lin, 2005) Along this line of reasoning, Conrad et al (2006) conjecture andtest conflicts of interest as a possible cause of the asymmetry but find little evidence to sup-port their conjecture However, Ljungqvist, Marston, and William (2006) find no evidencethat aggressive analyst behavior leads to investment banking business Furthermore, morereputable analysts/banks tend to be less aggressive suggesting that analysts and brokers’characteristics may be an important omitted factor that discourages optimism Therefore,our primary objective is to better understand the incentives and reporting quality factorsthat differentiate analysts and subsequently explain the asymmetric pattern in analysts’response.

Prior research has identified characteristics of high-status analysts (e.g., Brown, 2001;Clement, 1999; Clement & Tse, 2003; Jacob, Lys, & Neale, 1999; Mikhail, Walther, &Willis, 1997) and documented that high-status analysts issue more profitable stock recom-mendations than low-status analysts (e.g., Mikhail, Walther, & Willis, 2004; Stickel, 1995).Drawing on evidence from prior studies, we examine whether similar analyst characteristicsare associated with analysts’ effectiveness in processing public information and avoidingoptimistic bias, as opposed to analysts’ economic incentives

We predict that the asymmetric response is less for superior analysts First, superior lysts have a reputation advantage to attract new banking clients As a result, other analystsare more likely to avoid negative views to attract new business A sufficiently large nega-tive shock permits these other analysts to downgrade while still maintaining an optimisticbias Second, we expect that superior analysts have incentives to reveal their negative pri-vate information to the market Compared to analysts identified as providing no upgrade,superior analysts can provide a relative upgrade but still be relatively less optimistic inreaction to a negative price shock Combined, we expect less asymmetry for analysts exhi-biting superior attributes

ana-Using a sample of large price changes, we corroborate Conrad et al (2006) by findingasymmetry in the analyst response to large positive and large negative information shocks.More importantly, we find that our proxy for analyst quality is inversely associated withthe probability of recommendation downgrades following large negative price shocks, indi-cating a reduction in asymmetry as analyst quality improves Furthermore, the reduction inasymmetry is driven by analysts in the top group Analysts in the upper decile of analystquality do not appear to respond to either positive or negative news events, whereas themajority analysts in lower deciles of analyst quality exhibit varying degrees of asymmetricresponse

Further analysis indicates that the reduction in the asymmetry is only present whenanalysts have negative private information based on their own earnings forecasts.However, we find no evidence that banking affiliation is associated with the observedasymmetry Together, our findings are consistent with the asymmetry being associatedwith a general information processing bias among lower quality analysts Importantly,such a bias affects superior analysts less due, at least in part, to their effectiveness intranslating earnings forecasts into recommendations These findings are consistent withthe stream of research that differentiates stock picking ability between superior and infe-rior sell-side analysts (Li, 2005; Mikhail et al., 2004) We complement our findings byanalyzing future market price adjustments following large price shocks Although there islimited evidence that the downgrades following large, negative price shocks are

informative to the market, we find that the subsequent price discovery process is fasterfor superior analysts.

Examining the properties of analyst recommendations is important because priorresearch suggests that analyst research is particularly informative in bad news scenarios.For example, Hong, Lim, and Stein (2000) argue that analysts play a more significant role

in the dissemination of bad news because management has stronger incentives to highlightgood news than bad news Frankel, Kothari, and Weber (2006) suggest that bad news cre-ates uncertainty in firms’ information environment and thus presents traders with an oppor-tunity to gain from information acquisition that would mitigate the uncertainty It seemsthat the arrival of bad news increases the market demand for analysts’ services, providinganalysts additional incentives to deliver high-quality research Our evidence that superioranalysts are less likely to defer adverse views about companies is relevant to investors whofocus on downside risk and trade on security analysts’ stock recommendations

We add to Conrad et al (2006) by providing evidence on the extent to which analystcharacteristics are associated with the asymmetry in the analyst response Our articleextends the literature that examines the quality variation in analyst recommendations(Barber, Lehavy, McNichols, & Trueman, 2006; Li, 2005; Mikhail et al., 2004; Stickel,1995) by exploring the factors affecting the effectiveness of the analysts in processingpublic information Also, our results are related to the analyst forecast literature on behav-ioral biases (Herrmann & Thomas, 2005; Mikhail et al., 2003) Like underreaction orrounding in analyst forecasts, we show that a similar bias exists in analyst recommenda-tions Finally, by linking the asymmetry to the optimistic bias in recommendation revisions,our research provides additional insights into evaluation of potential biases in analysts’recommendations and improvement of research analyst integrity addressed in the Securitiesand Exchange Commission’s (SEC) Regulation Analyst Certification (SEC, 2003)

The article proceeds as follows The section ‘‘Literature Review and HypothesisDevelopment’’ discusses prior literature and forms our hypotheses Section ‘‘ResearchDesign’’ describes the formation of our sample and research design Descriptive statisticsand our empirical results are reported in the sections ‘‘Sample Descriptive Statistics’’ and

‘‘Empirical Results.’’ Section ‘‘Additional Tests’’ provides additional analyses and we clude in the section titled ‘‘Concluding Remarks.’’

con-Literature Review and Hypothesis Development

Analysts revise recommendations when they anticipate changes in company fundamentals

or when they react to earnings and other corporate news releases Analysts also respond toprice changes because analysts have private valuations about securities and use value-to-price comparisons when setting recommendations (Womack, 1996) and because stock pricemovements are potentially driven by the revelation of public information (Ryan & Taffler,2004) In the following paragraphs, we first discuss analysts’ information response Then

we discuss analyst characteristics leading to our hypothesis

Analysts are assumed to have private information relative to the market If an analyst’sobjective is to generate the most profitable stock recommendations, her or his prior stockrecommendation should fully incorporate her or his private information As a consequence,when a portion of an analyst’s private information is revealed through stock price move-ments, she or he may revise the recommendation to reflect the new value-to-price compari-son Analysts are therefore expected to downgrade (upgrade) following extreme priceincreases (decreases)

However, management tends to highlight good news more than bad news (Hong, Lim,

recommendation downgrades due to brokerage firms’ investment banking business (Conrad

et al., 2006) This adds difficulty to an analyst’s information-gathering process and reducesthe possibility of anticipation of bad news in the analyst’s reports

If analysts’ recommendations are upwardly biased (e.g., Stickel, 1995), then the arrival

of negative information may precipitate a greater response from analysts than good news.The result is asymmetry between good and bad news recommendation revisions (Conrad

et al., 2006) However, the existence of such asymmetry does not necessarily suggest thatanalysts are homogeneously biased or inefficient in anticipating bad news in their recom-mendations In contrast, Mikhail et al (2003) and Herrmann and Thomas (2005) find evi-dence that two known behavioral biases in analyst forecasts—underreaction androunding—are reduced with individual analysts’ ability and knowledge It follows that theasymmetry in the analyst response between good and bad news may be inversely associatedwith analysts’ quality

The literature identifies a number of analyst-specific factors that have the potential toreduce the bias and inefficiencies in analysts’ recommendation generation Womack (1996)states that generating stock recommendations are prediction tasks requiring industryand firm-specific knowledge and environmental cues Expert analysts possess the ability

of knowing which environmental cues need to be monitored and which can beignored (Dawes & Corrigan, 1974) Analysts with more experience are likely to developsuperior private information about a company’s economics the longer they follow it (Jacob

et al., 1999)

Stickel (1992) suggests that reputable analysts possess more timing ability He finds thatInstitutional Investor All-Stars analysts supply forecasts more often than other analysts.More frequent forecasts could be more advantageous for recommendation generationbecause they can incorporate the latest earnings-relevant information (e.g., recent interimearnings announcements and industry reports) Ke and Yu (2007) also find that analystsimprove their effectiveness in translating earnings forecasts into recommendations as theirexperience increases

If expertise, reputation, and experience proxy analysts’ innate ability and learning, thenthese characteristics could be associated with analysts’ efficiency in incorporating privateand public information into their stock recommendations Furthermore, relying on moreaccurate earnings forecasts (Bonner, Walther, & Young, 2003; Brown, 2001; Clement,1999; Clement & Tse, 2003, 2005; Jacob et al., 1999; Mikhail et al., 1997; O’Brien, 1988;Stickel, 1992), analysts with superior characteristics may incorporate their private valua-tions more fully because the resulting recommendations will have greater economic conse-

of their private information in recommendations are both direct and indirect

Analysts have incentives to build, and maintain, reputation for objectivity throughouttheir tenure, and such incentives are greater for superior analysts (e.g., Ljungqvist et al.,2006) Consistent with this view, prior research finds that analysts who have been identified

as providing superior performance are more likely to experience favorable career outcomes,such as moving up to a high-status brokerage house (e.g., Hong & Kubik, 2003; Hong,Kubik, & Solomon, 2000; Jacob et al., 1999) Reputational advantage or career concernsshould moderate the conflict of interest between investment banking and research, andincrease the ability of analysts to express their true opinions in recommendations Being

a top analyst also attracts business decreasing the pressure to bias recommendations

Combined, we expect the asymmetric response to be less prominent in analysts withsuperior characteristics for three primary reasons First, superior analysts are better able toforesee bad news and more willing to capture negative signals in a firm’s information envi-ronment and incorporate timely bad news in their stock recommendations Second, superioranalysts are more efficient in using earnings forecasts in stock recommendations, and thusare more likely to change their recommendations based on their private information Third,superior analysts have a reputation advantage and are less likely to hide their negativeviews to attract new underwriting business (or obtain access to management in the pre-regulation fair disclosure [FD] period).

Our hypothesis stated in the alternative is as follows:

Hypothesis 1 (H1): The asymmetry in bad versus good news stock recommendationrevisions is less for analysts with superior characteristics

Our article differs in several important ways from Conrad et al.’s (2006) We believe thatfactors other than conflicts of interest can affect analysts’ incentives A number of the analysts

do not have investment banking conflicts (Cowen, Groysberg, & Healy, 2006) In addition,high-status analysts and brokers rely at least partially on their reputation to attract new clientsand business, which may relieve the firm pressure More importantly, to maintain their superiorstatus, such analysts have greater incentives to reveal their negative private information to themarket through their reports (e.g., Clarke, Khorana, Patel, & Rau, 2007; Ljungqvist et al.,2006) We examine whether the asymmetry in the analyst response is caused by low-qualityanalysts after controlling for any influence of investment banking relationships Moreover, wetest our conjecture that the differential asymmetric response between high-status analysts andlow-status analysts is only present when analysts possess negative private information

We follow Conrad et al (2006) and use extreme price movements (returns) to identifynews events Price movements are determined as the 3-day, market-adjusted return(ADJRET) as follows:

ðRi;t Rm;tÞ=si;tðRi;t Rm;tÞ ð1Þ

net-of-market returns, calculated by using a sample of nonoverlapping 3-day net-of-net-of-market returnsduring days 23 to 2249 relative to day t

An extreme price movement (return) is defined as one in the top or bottom 1% of thedistribution of 3-day returns across all firms followed by I/B/E/S We identify extreme 1%return events based on the ADJRET cutoff points calculated from all firms included in

That is, once a 3-day period is identified as an extreme return event, we do not select thenext two (overlapping) 3-day return events as extreme return events This procedure results

in 168,249 large return events

We require the recommendations (upgrade, downgrade, or reiteration) to be issuedwithin 1/220 days of the event period, where day 0 is the day following the end of theextreme 3-day return event we identified above We consider recommendations issuedwithin this 41-day period to add cross-sectional variation in returns without which the

retain 46,264 large return events for which the firm receives at least one recommendationfrom its analysts within 1/220 days of the event period The final sample consists of82,386 analyst-firm observations (representing 7,980 unique firms and 6,574 unique ana-lysts from 481 brokerage houses), an average of almost two recommendation observationsper large return event

Analyst-Specific Characteristics

Our main focus is whether characteristics that are consistent with superior analysts areassociated with a reduction in the asymmetric response of stock recommendations Wedraw on prior research in the following paragraphs to identify analyst and brokerage-firmcharacteristics that are indicative of analysts’ superior ability in generating stockrecommendations

Analysts’ reputation is connected to the quality of research they produce (e.g., Cowen

et al., 2006; Ljungqvist et al., 2006; Stickel, 1992, 1995) The first two reputation measuresare based on inclusion on the All-Star list of the most recent Institutional Investor(II_STAR) or inclusion of the analyst’s brokerage firm on II All-American Research Teamfor the most recent team announcement prior to the recommendation change (II_TEAM).More reputable banks also devote more resources toward securities research We considertwo measures for brokerage resources: (a) the percentile rankings of the total number ofanalysts employed by the brokerage firm relative to other brokerage firms (B_SIZE) and(b) the percentage of the analyst’s brokerage house analysts that follow the company’sindustry during the current year (B_IND)

Analysts’ research efforts are suggestive of analyst expertise (see, for example, Ertimur

et al., 2007; Jacob et al., 1999) We include the number of companies followed by the lyst (COMP), an indicator variable that captures the timeliness of issuing recommendationrelative to earnings announcements (TIME), number of recommendations issued by theanalyst on this company during the current year (FREQ), and the analyst’s industry special-ization measured as the percentage of companies followed by the analyst with the sametwo-digit SIC code as the company (SPEC)

ana-Analysts’ ability is expected to be increasing in experience (e.g., Bowen, Chen, &Cheng, 2008; Ertimur et al., 2007) We use two variables that capture analyst experience.First, firm experience is measured as the natural log of the number of quarters the analysthas issued recommendations for the company prior to day t (FIRMEXP) Second, generalexperience is the natural logarithm of the number of quarters the analyst has recommenda-tions in IBES prior to day t (GENEXP).

Finally, career mobility is expected to be associated with recommendation ability as lysts who are able to move suggests the individual analyst may be in greater demand (e.g.,Hong & Kubik, 2003; Hong, Kubik, et al., 2000; Jacob et al., 1999) Three variables areused First, we measure the proportion of analysts who left the analyst’s brokerage houserelative to the total number of analysts who worked for that brokerage during the currentyear (P_OUT) Second, we capture the other side with the proportion of analysts whojoined the analyst’s brokerage house relative to the total number of analysts who workedfor that brokerage during the current year (P_IN) Finally, MOVEUP is an indicator vari-able that equals 1 if the analyst moved to a larger brokerage house during the current year,and 0 otherwise

ana-As each proxy measures an analyst’s overall ability in recommendation generation witherror and we are interested in an analyst’s general ability in recommending stocks ratherthan in any one particular source of that ability, we perform principal components analysis

to aggregate these 13 financial analyst characteristics This helps reduce the measurementerror inherent in each individual characteristic and allows us to incorporate multiple corre-lated measures into a set of uncorrelated variables that capture analyst quality

Panel A of Table 1 reports descriptive statistics of the sample observations for the 13analyst characteristics we consider Only 30.6% of the analysts are included on the all-American team and 16.4% on the all-star list suggesting that these lists are discriminating.Firm experience suggests that most analysts have issued recommendations on the companyfor less than four quarters Although there appears to be significant turnover at the broker-age houses, only 7.2% of the analysts in our study moved to a larger brokerage houseduring the year Finally, there is also substantial variability in the number of recommenda-tions the analysts issued on the sample company during the year and the number of compa-nies followed by the analysts On average, each analyst follows a mean of 10.6 samplefirms The average number of recommendations issued by the analyst for the firm is 2.2,consistent with recommendation changes clustering around extreme return events

Next, we report the results of our principal components analysis (Panel B of Table 1).Factor patterns are generated by a maximum likelihood estimation procedure with a varimaxrotation of the factors Based on an analysis of the Eigenvalues for each factor, we retain

factors relate to (a) REPUTATION, based on B_SIZE, II_TEAM, II_STAR, and B_IND; (b)EXPERIENCE, based on FIRMEXP and GENEXP; (c) CAREER, based on P_IN, P_OUT,and MOVEUP; and (d) EXPERTISE, based on FREQ and COMP As predicted,each factor has positive loadings on the underlying variables (except for B_IND) Weuse the average of the four factors as a combined measure (FACTOR) as our analyst

Asymmetric Response Model

Conrad et al (2006) estimate an ordered probit regression model for separate positive andnegative return observations and find that analysts systematically respond only to large

Table 1 Principal Components Analysis of Financial Analyst Characteristics

Panel A: Descriptive Statistics

bro-to the recommendation issue date, and 0 otherwise; B_IND is the percentage of the analyst’s brokerage house analysts that follow the company’s industry during the current year; FIRMEXP is the natural logarithm of the number of quarters the analyst has issued recommendations for the company prior to day t; GENEXP is the natural logarithm of the number of quarters the analyst has recommendations in I/B/E/S prior to day t; P_IN is the proportion of analysts who joined the analyst’s brokerage house relative to the total number of analysts who worked for that brokerage during the current year; P_OUT is the proportion of analysts who left the analyst’s brokerage house relative to the total number

of analysts who worked for that brokerage during the current year; MOVEUP is the indicator variable that is 1 if the analyst moved to a larger brokerage house during the current year, and 0 otherwise; FREQ is the number of recommen- dations issued by the analyst for the company during the current year; COMP is the number of companies followed by the analyst during the current year; TIME is the indicator variable equal to 1 if the recommendation is issued during the 3-day quarterly or annual earnings announcement window, and 0 otherwise; and SPEC is the percentage of companies followed by the analyst with the same two-digit SIC code as the company during the current year.

b

Factor patterns are generated by a maximum likelihood estimation procedure with a varimax rotation of the tors TIME and SPEC have the lowest individual Kaiser-Meyer-Olkin (KMO) statistics These two variables are therefore dropped from the principal components analysis to reduce multicollinearity and ensure sampling ade-

negative price shocks To test our hypothesis, we expand the Conrad et al model to estimatethe following models We start with a baseline model, Equation (2a), that includes ouroverall analyst quality proxy We replace this proxy with indicator variables for analystquality rankings to examine differences between superior and inferior analysts, inEquation (2b).

RECCHi;j;t ¼ b0þ b1ADJRETi;tþ b2Di;tþ b3ADJRETi;t3Di;tþ b4FACTORi;j;t

3ADJRETi;t3Di;tþ b5FACTORi;j;tþ b6FACTORi;j;t3Di;t

þ b7FACTORi;j;t3ADJRETi;tþ b8AFILi;j;t3ADJRETi;t3Di;tþ b9AFILi;j;t

þ b18LPRECi;tþ b19SMALLi;tþ b20SMALLi;t3ADJRETi;tþ ei;j;t

ð2aÞ

3ADJRETi;t3Di;tþ b6;7RANKi;j;tþ b8;9RANKi;j;t3Di;t

þ b10;11RANKi;j;t3ADJRETi;tþ b12AFILi;j;t3ADJRETi;t3Di;t

þ b13AFILi;j;tþ b14AFILi;j;t3Di;tþ b15AFILi;j;t3ADJRETi;t

ð2bÞ

where for firm i, analyst j, and day t, the variables are defined as follows: RECCH is therecommendation change for firm i by analyst j on day t, t is the 220, ,120; the change

is reordered such that the recommendation change variables ranges from 24 (downgrade)

pre-ceding the recommendation change; D is the indicator variable equaling 1 if the dation change occurs within the 1/2 20 day window of a large negative return event and

vari-ables for top and bottom decile analysts, TOP and BOTTOM; AFIL is the indicator variableequaling 1 if analysts’ firm has an investment banking relationship with firm i, and 0 other-wise; LREC is the previous recommendation level; MVE is the beginning-of-year marketvalue of equity (in thousands); AGE is the approximate age of the firm measured as thenumber of years between current year and year the firm first appears on CRSP; NUMREC

is the number of analysts following the firm; PRICE is the per share stock price for firm i

at day t; LMNREC is the mean recommendation change for the 10 days preceding analyst’srecommendation change date; LPREC is the percentage of analysts following firm i duringyear t that change their recommendation during the 10 days preceding analyst’s recommen-dation change; and SMALL is the indicator variable equaling 1 if the firm is in the smallestequity market value decile for all sample observations during the year

changes relationship If analysts have private information that is revealed with stock pricemovements, then the estimated coefficient would assume a negative sign A negative coef-ficient indicates that analysts are likely to downgrade (upgrade) following positive (nega-tive) returns However, a negative coefficient may not be observed for ADJRET if some

analysts still upgrade (downgrade) after positive (negative) returns due to divergence ofopinions on whether the prices will continue to rise or fall An insignificant coefficient onADJRET is then expected if the upgrades and downgrades are about equal.

However, if analyst recommendations anticipate future good news more fully thanfuture bad news (optimistic bias), analysts likely react more to the arrival of bad news thangood news and are more likely to downgrade following large negative returns (i.e., theasymmetric response) We therefore predict that the incremental main effect for bad news

We include analyst quality variables interacted with ADJRET and ADJRET 3 D to

Equation (2a), FACTOR is used to capture overall analyst quality Consistent with ourhypothesis, we predict a negative coefficient on FACTOR 3 ADJRET 3 D suggesting thatthe likelihood of downgrades (and thus the asymmetric response) is reduced for analystswith superior characteristics The expected sign of the coefficient on FACTOR 3 ADJRET

is ambiguous because analysts are free to express their opinions when there is less concernfor conflicts of interest in absence of bad news (i.e., past positive returns) For complete-ness, we also include FACTOR and its interaction with D without predictions

Past research differentiates stock picking ability between superior and inferior analystsand suggests that the increasing difference between top-performing and underperforminganalysts is primarily attributable to the performance of the few analysts in the top group(Li, 2005; Mikhail et al., 2004) We therefore replace FACTOR (and its interaction terms)

in Equation (2b) with two indicator variables (and their interactions) for analysts in the topand bottom deciles, thereby allowing the response coefficients to vary across analysts Weexpect analysts in the top decile of analyst quality to have the least tendency to downgradefollowing large negative returns and thus the most pronounced reduction in asymmetry.The remaining variables in Equations (2a) and (2b) are control variables drawn fromprior research We include an indicator variable for the presence of an investment bankingrelationship (AFIL) as a main effect as well as interacted with the asymmetric response

presence of the investment banking relationship is expected to have a positive sign ifinvestment banking influences analysts’ impartiality (e.g., Michaely & Womack, 1999;O’Brien et al., 2005) To investigate the investment banking explanation in Conrad et al.(2006), we further augment the model with a three-way interaction term AFIL 3 ADJRET

3 D (and the interactions of AFIL with ADJRET and D) A positive coefficient on thethree-way interaction provides evidence that analysts who are subject to firm pressures aremore likely to defer negative recommendations until after the arrival of bad news

Prior recommendation level (LREC) ranging from 1 (a strong buy) to 5 (a strong sell) isincluded as the existing recommendation level limits the potential for any change The pre-dicted sign is positive because there is little or no room to upgrade (downgrade) when theinitial recommendation level is low (high)

We include market value of equity (MVE), firm age (AGE), number of analysts ing the firm (NUMREC), and liquidity (PRICE) to control for the amount of publicly avail-able information as well as analysts’ incentives to perform a private information searchabout the company Although the relationship between the magnitude of a recommendationrevision and firm information environment has not been examined, Hong, Lim, et al.(2000) suggest that analysts are more likely to incorporate bad news in firms with lowerinformation asymmetry This may imply that the revised recommendations after the arrival

follow-of the news would generally be less negative in firms with better information flow (e.g.,larger firms, newly-listed stocks, and stocks with greater liquidity) Accordingly, we predictpositive signs for MVE, NUMREC, and PRICE and a negative sign for AGE.

In addition to the event day return information, analysts may incorporate actions taken

by their peers that follow the same stock (i.e., herding) We therefore include the age of analysts following the firm that change their recommendation (LPREC) and themean recommendation change (LMNREC), both during the 10 days preceding the analyst’srecommendation change Both variables are expected to be positive, based on studies ofanalyst herding (e.g., Welch, 2000)

percent-Finally, we include an indicator variable for firms in the decile of smallest firms (based

on market value of equity) in our sample (SMALL) and its interaction with market-adjustedreturn because analysts may respond differently to large returns for the smallest set offirms (e.g., Hong, Lim, et al., 2000) If bad news gets out more slowly (relative to goodnews) in smaller firms, then smaller firms would be more likely to experience downgrades(often with a greater magnitude) than larger firms Although opinions may differ on theinterpretation of this variable (see Hong, Lim, et al., 2000), we predict that SMALL (indi-cating the sign of the response) will be negative and its interacting variable with ADJRET(indicating the magnitude) will be positive

Sample Descriptive Statistics

Panel A of Table 2 presents a transitional matrix of the distribution of observations based

on the previous and current recommendation Overall, 51% of the observations areassociated with recommendation upgrades, whereas only 36% are associated with recom-mendation downgrades (the balance are reiterations—no change—of the previous recom-mendation) The positive bias in the level of recommendations is consistent with thosereported elsewhere The majority of changes are one classification moves—right below orabove the diagonal

In Panel B, we partition the sample of recommendation changes based on the sign of theextreme 3-day return event and report the return descriptive statistics by classification ofreturn sign and recommendation change Downgrades are usually issued around the mostnegative and most positive returns

Finally, Panel C further breaks down the data by number of grades We find that ple downgrades are approximately three times that of upgrades following negative returns

multi-In contrast, multiple downgrades are as common as they are among upgrades following

sig-nificantly different (p value \ 001) The distribution is consistent with the asymmetry in

Table 3 reports descriptive statistics for the variables included in the probit analysis.Consistent with the distributional data, the mean and median recommended change is

a slight downgrade and the market-adjusted return is negative However, the mean andmedian of the previous recommendation is 2.15 or a buy We also find that the recommen-dation level prior to bad news events for higher status analysts (those with above-medianFACTOR scores) is less optimistic than that for lower status analysts (untabulated meanrecommendation is 2.21 vs 2.13, p value \ 01), consistent with the arguments leading toour hypothesis Only 5% of the observations are associated with investment banking rela-tionships The distributional properties among many of the control variables suggest signifi-cant variation among the sample observations that we control for in Equation (2)

Table 4 presents the Pearson and Spearman correlations among the explanatory variablesfor recommendation changes No unusual correlations are noted that would affect ourestimation.

Table 2 Recommendation Changes Within 1/220 Days of Large Return Events

Panel A: Transitional Matrix of Recommendations (N = 82,386)

Panel C: Recommendation Change Categories by Event Return Sign

Gradeb Negative return sign Positive return sign

stan-dard deviation of 3-day market adjusted returns over day 2249 to 23.

b

Both upgrades and downgrades have 1 to 4 recommendation grades: recommendation upgrades to either strong buy, buy, hold, or sell; and recommendation downgrades to either buy, hold, sell, or strong sell Upgrades are coded as positive changes (e.g., a move from hold to strong buy is coded as Upgrade 12) and downgrades as nega- tive changes (e.g., a move from strong buy to hold is coded as Downgrade 22).