when are analyst recommendation changes inﬂuential? · 2017. 12. 22. · when are analyst...

When Are Analyst RecommendationChanges Influential?

Roger K. LohLee Kong Chian School of Business, Singapore Management University

Rene M. StulzFisher College of Business, Ohio State University, NBER, and ECGI

The existing literature measures the contribution of analyst recommendation changes us-ing average stock-price reactions. With such an approach, recommendation changes canhave a significant impact even if no recommendation has a visible stock-price impact. In-stead, we call a recommendation change influential only if it affects the stock price ofthe affected firm visibly. We show that only 12% of recommendation changes are influ-ential. Recommendation changes are more likely to be influential if they are from leader,star, previously influential analysts, issued away from consensus, accompanied by earn-ings forecasts, and issued on growth, small, high institutional ownership, or high forecastdispersion firms. (JELG14, G20, G24)

Market observers at times attribute large stock-price changes to analyst rec-ommendation changes. For instance, according toThe Wall Street Journal,Kenneth Bruce from Merrill Lynch issued a recommendation downgrade onCountrywide Financial on August 15, 2007, questioning the giant mortgagelender’s ability to cope with a worsening credit crunch. The report sparked asell-off in Countrywide’s shares, which fell 13% on that day. In another exam-ple, when Meredith Whitney (CIBC World Markets) downgraded Citigroup onNovember 1, 2007, the stock price dropped 6.9%, the CEO quit two days later,and she apparently received death threats.1 Though the finance literature findsthat significant average abnormal returns are associated with recommendation

We thank an anonymous referee, Oya Altinkilic, Brad Barber, Francois Degeorge, Harrison Hong, AlexanderLjungqvist (the editor), Siew Hong Teoh, Rong Wang, and Jay Ritter for many helpful comments and sug-gestions. We are also grateful to Gilbert Ho for research assistance, Thomson Financial for providing theI/B/E/S broker translation file, and the Office of Research [Grant number C207/MSS8B010] at SingaporeManagement University for financial support. Send correspondence to Rene M Stulz, Department of Finance,806 Fisher Hall, 2100 Neil Avenue, Columbus, OH 43210, USA; telephone: (614) 292-1970. E-mails: Loh:[email protected]; Stulz: [email protected].

1 The above examples are from the following articles: “Countrywide’s woes multiply,” by James R. Hagerty andRuth Simon,The Wall Street Journal, August 17, 2007, and “CIBC analyst got death threats on Citigroup,” byJonathan Stempel, Reuters, November 4, 2007.

c© The Author 2010. Published by Oxford University Press on behalf of The Society for Financial Studies.All rights reserved. For Permissions, please e-mail: [email protected]:10.1093/rfs/hhq094 Advance Access publication October 15, 2010

The Review of Financial Studies / v 24 n 2 2011

changes, the typical estimate associated with a recommendation change is toosmall to be considered a significant abnormal return for the stock of a firm.Consequently, with the typical recommendation change, investors following afirm cannot distinguish the impact of the recommendation change from noise.However, at times, a recommendation change, such as the Bruce call on Coun-trywide, is viewed by observers as having a large identifiable impact on thestock price. In this article, we investigate how frequently recommendationchanges visibly impact stock prices, which we assess to be the case when arecommendation change has a significant stock-price impact, and we try to un-derstand better when and why analyst recommendation changes have such animpact.

The existing literature that assesses the impact of recommendation changesdoes not make it possible to answer the questions we are interested in. Thisliterature focuses on average effects in large samples and generally investi-gates whether some type of analyst recommendation change has a significantaverage abnormal return. By averaging across a large number of announce-ments, the researcher hopes to eliminate the influence of confounding effectson the study and therefore to obtain an estimate of the “pure” recommendationchange effect. At the same time, however, such an approach is of little use toevaluate claims about the ability of analysts to visibly impact stock prices ofindividual firms. To wit, in our sample, the median abnormal return associatedwith a downgrade is roughly−1%. For the typical firm, a−1% abnormal re-turn is noise. However, an abnormal return of the magnitude associated withthe recommendation change of Bruce for Countrywide is a highly significantabnormal return for the typical firm. The existing literature that focuses onaverage abnormal returns does not make it possible to understand whether an-alysts can visibly move prices or how often they do so. Such an understandingis critical to assessing the role of analysts in generating information about firmsand in influencing investors and management. In particular, it would be hard toargue that analysts influence investors and management systematically if theydo not visibly move prices.

Our contribution is to identify recommendation changes that are impactfulbased on stock-level abnormal returns. We define a recommendation change asinfluential in returns if its associated abnormal return is in the same directionas the recommendation change and is statistically significant. An analyst mightnot affect the stock price, but she might lead investors to trade in response toher analysis. Therefore, we also use an alternative definition of an influentialrecommendation change based on turnover. With this definition, a recommen-dation change is influential if it leads to a statistically significant increase inturnover at the firm-level. These approaches ensure that the recommendationchanges that we eventually label as influential are indeed those that are noticedby investors following the firm.

An important component of our contribution is that we conduct our maintests with recommendation changes that occur on days without firm-specific

594

When Are Analyst Recommendation Changes Influential?

news. Analysts often write reports on days of firm-specific news, and recom-mendation changes on such days are more likely to be favorable if the firmhas positive news. Though the traditional event study method reduces or eveneliminates the impact of confounding news on the average abnormal return,it does so only when news and the probability of occurrence of the event areuncorrelated. In the case of analysts, there is no reason to believe that this con-dition holds. It is therefore important to construct a sample of recommendationchanges where the impact of confounding firm-specific news is minimized. Notsurprisingly, eliminating firm-specific news days reduces the stock-price reac-tion to analyst recommendation changes, but the average stock-price reactionremains statistically significant.

We find that about 12% of recommendation changes in our sample (thatminimizes the impact of firm-specific news) are influential in returns and about13% are influential in turnover. However, about one out of four analysts neverhad any influential recommendation change. Conditional on an analyst havingan influential recommendation change, one in five of the analyst’s recommen-dation changes are influential. This finding illustrates that influential recom-mendation changes come only from a subset of skilled analysts and that theseinfluential recommendation changes are infrequent even for analysts withinthis subset.

Meredith Whitney’s Citigroup downgrade on November 1, 2007, was asso-ciated with a drop in Citigroup’s stock price of 6.9%. Yet, as aWall Street Jour-nal article recently reported, other analysts in the weeks before downgradedthe stock with reports that had similar content.2 Consequently, a recommen-dation change is not influential simply because of its content—other factorsmust affect whether the recommendation change is influential. We use a probitmodel to investigate the factors that make it more likely that a recommendationchange will be influential. We consider a battery of analyst, recommendation,and firm variables. We find that recommendations away from the consensusand recommendations accompanied by any sort of earnings forecasts are morelikely to be influential. Influential recommendations are also more likely to befrom Institutional Investor–ranked analysts and analysts who have a history ofbeing ahead of the herd in issuing recommendations. Analysts have hot handsin influential recommendations: An analyst who has had an influential recom-mendation in the past is more likely to have one in the future. It is harder foran analyst to have an influential recommendation when more analysts follow afirm and when the firm is larger. However, greater diversity of opinion about afirm makes it more likely that a recommendation change will be influential.

When analyst recommendation changes are influential, they should leadto more analyst and investor activity in the stock as investors adjust their

2 “When Meredith Whitney calls, should you listen?” by David Weidner,The Wall Street Journal, April 9, 2009.

595


holdings to the new information produced by analysts. We find this to be thecase. Analyst activity increases after an influential analyst recommendationchange compared to before. Forecast revisions by analysts following such achange are much larger than forecast revisions before such a change. Stockvolatility and turnover are much larger in the three months following an in-fluential analyst change than in the three months before. Finally, the firm’sindustry is also more likely to have a large return coinciding with the recom-mendation event—consistent with the analyst research containing an industryelement affecting similar firms.

We are not the first to examine the differential impact of stock recommenda-tion changes. For instance,Stickel (1995) finds that recommendation changesof star analysts have more impact, andFang and Yasuda(2008) show thatthey are more profitable.Irvine (2004) provides evidence that the market re-acts more strongly to initiations than to other recommendations.Ivkovic andJegadeesh(2004) demonstrate that the timing of recommendation changes inrelation to earnings announcements affects their impact.Asquith, Mikhail, andAu (2005) provide evidence that the impact of recommendation changes is af-fected by the content of analyst reports.Chen, Francis, and Schipper(2005)find that the average analyst recommendation or earnings forecast producesa price impact that is no different from the average stock-price movementon non-recommendation days.Frankel, Kothari, and Weber(2006) examinewhether firm characteristics affect the impact of earnings forecast revisions,but they do not consider analyst characteristics or stock recommendations. Akey distinguishing feature of our approach from this literature is that we donot focus on average effects. Most authors find a significant effect of analystrecommendations on average for certain samples, and some authors find no sig-nificant effect. Our study is not about average effects, but rather about whetherindividual recommendations are influential. We could find evidence that somerecommendations are influential even if the average recommendation in a sam-ple has an insignificant stock-price reaction; alternatively, we could find thatno recommendations are influential even if the stock-price reaction to analystrecommendations is significant on average.

A related paper,Altinkilic and Hansen(2009), reports evidence that the av-erage recommendation revision does not produce an economically meaning-ful reaction after removing recommendations that piggyback on firm news,such as earnings announcements. They go on to conclude that analyst recom-mendations are therefore uninformative.Chen, Francis, and Schipper(2005)find that the average analyst recommendation or earnings forecast producesa price impact that does not differ from the average stock-price movementon non-recommendation days. Because these papers focus on average effects,they do not discuss or identify subsets of recommendations that are influential.Although our findings agree that the majority of recommendations are unin-formative, we argue that analysts add value to financial markets by virtue ofthe fact that they can produce influential recommendations (e.g., as anecdotally

596


illustrated in the Citigroup and Countrywide cases).3 Since our threshold for aninfluential recommendation (ignoring the requirement of correct sign) corre-sponds to the 5% probability level, we expect 5% of recommendation changesto be significant by chance alone. We find that the percentage of influentialrecommendation changes is more than twice the percentage we would ex-pect by chance alone. At the same time, our evidence shows that producingan influential recommendation change requires a combination of skills andcircumstances that makes such recommendation changes infrequent.

Analysts also produce earnings forecast revisions. Prior work on the impactof earnings forecast revisions has focused on differentiating reaction magni-tudes according to firm and analyst characteristics, for example, inClementand Tse(2003) andGleason and Lee(2003).4 Therefore, we estimate the frac-tion of earnings forecast revisions that are influential. We find that roughly 5%of earnings forecast revisions are influential. Earnings forecast revisions ac-companied by recommendations are twice as likely to be influential. Further, arecommendation change is more likely to be influential if it is accompanied byan earnings forecast. We conjecture that analyst research is more likely to beinfluential (according to our definition) when conveyed through a recommen-dation change, since it is a clear call to buy or sell a stock that can receive agreat deal of attention in the press.

The rest of the study is organized as follows. Section1 details the dataand sample. Section2 describes the average recommendation event abnormalreturn. Section3 identifies which recommendations are influential and theircharacteristics and consequences. Section4 investigates predictive variablesfor influential recommendations. Section5 considers robustness tests, andSection 6 concludes.

1. Data and Sample

1.1 Recommendations dataThe stock recommendations sample is from Thomson Financial’s InstitutionalBrokers Estimate (I/B/E/S) U.S. Detail File, augmented with dates from theFirst Call Database. We build our sample starting from I/B/E/S ratings is-sued by individual analysts from 1993 to 2006, with ratings ranging from 1(strong buy) to 5 (sell). Ratings are reversed (e.g., strong buy now denotedby 5) so that higher ratings correspond to more favorable recommendations.We focus on recommendation changes issued from 1994 onward since 1993

3 Our research design also has more power to identify informational effects of analysts since we investigate eachrecommendation change at the individual firm level. Average stock-price reactions disproportionately reflectrecommendations at firms with greater analyst coverage. For instance, a firm with 30 analysts will typically haveten times more observations in a sample than a firm with three analysts. Yet the firm covered by 30 analystswould be one for which an individual recommendation is less likely to be informative.

4 These papers are different from our approach in that they do not examine recommendations, rely on our methodof identifying influential events, or examine volume reactions to the events.

597


observations are sparse (1993 data is used for prior ratings when available).Ljungqvist, Malloy, and Marston(2009) report that matched records in theI/B/E/S recommendations data were altered between downloads from 2000 to2007. They also document that Thomson Financial, in response to their pa-per, fixed the alterations in the recommendation history file as of February 12,2007. The dataset we use is dated March 15, 2007, and hence reflects theserecent corrections by Thomson.

We focus on recommendation revisions and not levels, since prior researchconfirms that recommendation changes are more informative than mere levels(e.g.,Boni and Womack 2006; andJegadeesh and Kim 2010). The recommen-dation change (recchg) is computed as the current rating minus the prior ratingby the same analyst. By construction,recchgranges between−4 and+4. Arating is assumed to be outstanding according to the definition inLjungqvist,Malloy, and Marston(2009). Specifically, a rating is outstanding if it has beenconfirmed by the analyst (in the I/B/E/S review date field) in the last twelvemonths and has not been stopped by the broker (in the I/B/E/S Stopped File).We exclude observations where there is no outstanding prior rating from thesame analyst (i.e., analyst initiations or re-initiations are excluded). We removeanalysts coded as anonymous by I/B/E/S since it is not possible to track theirrecommendation revisions. To ensure that our sample focuses on firms that areof sufficient interest to investors, we also remove observations for which fewerthan three analysts have valid outstanding ratings.

We also deal with overall rating distribution changes due to the NationalAssociation of Securities Dealers (NASD) Rule 2711 in 2002. Many brokersreissued stock recommendations in response to the rule, with many of themchanging to a three-point (buy, hold, sell) scale instead of a five-point (strongbuy, buy, hold, underperform, sell) scale (Kadan, Madureira, Wang, and Zach2009). As a result, 2002 contains the largest number of recommendations inI/B/E/S compared to any other sample year (Barber, Lehavy, McNichols, andTrueman 2006). We account for this structural break by using the I/B/E/SStopped File to locate these rating distribution changes and adopt a three-pointrating scale for the affected brokers.5 These adjustments code 40% of I/B/E/Sobservations after September 2002 on three-point rating scales so that therecchgfor these affected brokers would range between−2 and+2.

To ensure that our recommendation dates are reliable, we augment the I/B/E/S sample with real-time recommendation dates from First Call. A wrongdate may result in us not capturing the true event date of the recommendationchange and understating the influence of analysts. To insert First Call dates

5 For 2002, we check for cases where a broker stopped all the recommendations in its coverage universe andresumed coverage in the subsequent days using only a three-point rating. We check one year post-resumptionfor the new distribution of ratings. When the new distribution contains only three distinct ratings [∈(1,3,5)or (2,3,4)], we assume this broker uses three-point ratings beginning with the resumption date. In the probitestimations, the rating change explanatory variable is based on three-point scales for the affected brokers. Weverified that removing all three-point scale-based observations does not affect our results in Table4.

598


into matched observations in I/B/E/S, we do the following. The broker names(bro name) on First Call are matched by hand to the I/B/E/S translation filebroker name (baname).6 We then look seven days on either side of the I/B/E/Srecommendation date to find a First Call observation that is matched on bro-ker, firm, and recommendation level. When there are duplicate matches, theclosest date observation is chosen (earlier date for ties). We found matches for52% of the I/B/E/S observations (Ljungqvist, Malloy, and Marston 2009alsoreport a similar match rate of 46.8%). About 77% of these had recommenda-tion dates unchanged, 21% had dates brought back by one day, and 2% haddates brought forward by one day. We use this First Call augmented samplefrom now on, although our results hold even when we use the I/B/E/S samplealone.

We adopt a two-day event window to incorporate the daily return reflectingthe recommendation change.7 To compute the two-day cumulative buy-and-hold abnormal return (CAR) for a recommendation changei , we define

CARi =1∏

t=0

(1+ Rit )−1∏

t=0

(1+ RDGTW

it

). (1)

Rit is the raw return of the stock on dayt , and RDGT Wit is the return on a

benchmark portfolio with the same size, book-to-market (B/M), and momen-tum characteristics as the stock (Daniel, Grinblatt, Titman, and Wermers 1997,thereafter DGTW).8 Day 0 is either the First Call augmented recommendationdate or the next trading day (for recommendations on non-trading days or rec-ommendations between 4:30PM and 11:59PM on a trading day). We removeobservations where the lagged price is less than one dollar on day 0 to preventour results from being driven by low-priced stocks.

6 First Call has the practice of sometimes recycling broker codes and backfilling the new broker name onto oldrecommendations. To mitigate this problem, we also rely on a file containing historical linkages between FirstCall broker codes and broker names in matching the broker names between First Call and I/B/E/S. This file isalso used inLjungqvist, Marston, and Wilhelm(2009), and we thank Alexander Ljungqvist for providing thedata.

7 We find similar results with a three-day window from day−1 to+1. We also examine the average abnormalreturn around the event and find that days 0 and+1 account for almost all of the cumulative abnormal return inthe−5 to+5 period. Hence, we believe that our recommendation dates are accurately aligned with contempo-raneous stock-price reactions.

8 The results are similar when we use the sum of abnormal returns rather than buy-and-hold abnormal returns.The DGTW portfolios are computed as follows. Every July, firms are first sorted into quintiles based on theirsize (market cap on June 30 of each year) using break-points determined from NYSE stocks. Second, firms arethen sorted within each size quintile into quintiles based on their B/M ratios. B/M ratios are computed as inFama and French(2006). Third, firms within each size-B/M group are sorted into momentum quintiles everymonth based on the buy-and-hold return over the prior 12 months skipping the most recent month. Therefore,the size and B/M rankings are updated every 12 months while the momentum rankings are updated monthly.Finally, the stocks within each characteristic portfolio are equally weighted at the beginning of each month andthe buy-and-hold average daily returns are computed.

599


1.2 Importance of removing recommendations made in responseto firm news

If a stock recommendation has an immediate impact on a firm’s stock price, itdoes so because it reveals information about the firm. In determining whetherthe analyst produced any material information, one should be careful to re-move recommendations that merely repeat the information contained in firm-specific news releases. As already discussed,Altinkilic and Hansen(2009) goso far as to argue that once the impact of other corporate news is removed,analyst recommendation changes do not have a material impact.Malmendierand Shanthikumar(2007) andLoh (2010) report that 12%–13% of stock rec-ommendations occur in the three days around quarterly earnings announce-ments. Since there are 252 trading days in a year, one would expect only 4.8%of all recommendations to be issued around earnings announcements if thelikelihood of a recommendation is uniformly distributed throughout the year.Therefore, not removing recommendations associated with earnings announce-ments falsely gives credit to the analyst recommendation for producing theearnings announcement price impact (see alsoFrankel, Kothari, and Weber2006). To apply this screen, we obtain quarterly earnings announcement datesfrom Compustat.

Another type of firm-specific news release is earnings guidance issued byfirms. Chen, Francis, and Schipper(2005) suggest that such days should alsobe taken out when determining the price impact of stock recommendations.We obtain earnings guidance dates from the First Call Guidelines database.Finally, Bradley, Jordan, and Ritter(2008) contend that clustering in recom-mendation changes usually occurs because of firm-specific news. Therefore,we also identify days on which the I/B/E/S universe records multiple analystsissuing recommendations for the firm as potential firm-specific news events.

2. The Average CAR of Recommendation Changes

In this section, we estimate the average CAR of recommendation changesto provide a benchmark for our later analysis and to show how minimizingthe impact of firm-specific news affects the estimate of the average CAR ofrecommendation changes.

2.1 Descriptive statistics of recommendation changesOur main sample contains 154,134 recommendation changes. Panel A ofTable1 shows the transition probabilities of recommendation changes. We seethat recommendation levels are predominantly optimistic, with sell and under-perform ratings making up only a small percentage of all recommendations.Figure1 plots the transition probabilities in Panel A of Table1. Looking at thebars in Figure1, we see that there is a tendency for recommendations that arenot holds to get revised into holds, while hold ratings themselves tend to getupgraded to buys.

600


Table 1Descriptive statistics of recommendation changes

Panel A: Transition probabilities of recommendationchanges

CurrentRec

1 2 3 4 5Prior Rec Sell Underperform Hold Buy Strong Buy Total

1 (Sell) 90 130 2,008 253 223 2,7043.3% 4.8% 74.3% 9.4% 8.2% 100%

2 (Underperform) 125 1,079 5,191 801 160 7,3561.7% 14.7% 70.6% 10.9% 2.2% 100%

3 (Hold) 2,139 5,554 7,661 24,098 12,195 51,6474.1% 10.8% 14.8% 46.7% 23.6% 100%

4 (Buy) 333 1,065 30,406 8,093 15,079 54,9760.6% 1.9% 55.3% 14.7% 27.4% 100%

5 (Strong Buy) 371 338 16,740 15,594 4,408 37,4511.0% 0.9% 44.7% 41.6% 11.8% 100%

Total 3,058 8,166 62,006 48,839 32,065 154,134

Panel B: Recommendation change categories

Rec Change Frequency Percentage

−4 371 0.2%−3 671 0.4%−2 19,944 12.9%−1 51,679 33.5%

0 21,331 13.8%+1 44,498 28.9%+2 15,004 9.7%+3 413 0.3%+4 223 0.1%

Total 154,134 100%

The sample of recommendation (rec) changes are from I/B/E/S Detail U.S. File 1994 to 2006. Each rec change(or rating change) is an analyst’s current rating minus his prior rating (prior ratings may be from 1993, but currentratings are from 1994 onward). Analyst initiations or ratings with no prior outstanding ratings are excluded. Arating is outstanding if it has been confirmed by the analyst (in the I/B/E/S review date field) in the last twelvemonths and has not been stopped by the broker (in the I/B/E/S Stopped File). Ratings are coded as 1 (sell) tostrong buy (5), and rating changes lie between−4 and 4. Anonymous analysts and observations with less thanthree analysts with outstanding ratings are excluded. Panel A reports the transition probabilities of rec changes.For example, in column 1, when the prior rec is a sell, it has a 4.8% probability of transiting to an underperformrating. Panel B reports the frequencies of each rec change category.

Next, Panel B of Table1 summarizes the number of recommendations ac-cording to the sign and magnitude of the rating change. The two rating-changegroups that have the largest number of recommendations are one-point up-grades (+1) and one-point downgrades (−1). The+1 group contains 44,498recommendations (28.9% of the sample), and the−1 group contains 51,679recommendations (33.5%). Reiterations (rating change of 0) make up 13.8%of all recommendation changes.

2.2 Histogram of recommendation CARFigure2 plots the histogram of two-day CARs of recommendation changes forone-point magnitude rating changes since these categories contain the largest

601


Figure 1Transition probabilities of recommendation changesThe sample of recommendation (rec) changes is from I/B/E/S Detail U.S. File 1994 to 2006. Each rec change (orrating change) is an analyst’s current rating minus his prior rating (prior ratings may be from 1993, but currentratings are from 1994 onward). Analyst initiations or ratings with no prior outstanding ratings are excluded. Arating is outstanding if it has been confirmed by the analyst (in the I/B/E/S review date field) in the last twelvemonths and has not been stopped by the broker (in the I/B/E/S Stopped File). Ratings are coded as 1 (sell) tostrong buy (5), and rating changes lie between−4 and 4. Firms with less than three analysts making up theconsensus are excluded. The chart plots the transition probabilities of rec changes—the probability that a priorrec transits to any of the five rating categories.

number of observations. The first chart shows the distribution of event CARsfor one-point downgrades with the percentage of CARs that fall within 100basis point bins. The histogram reveals two prominent trends. First, the zerobin (representing CARs of−0.5% to 0.5% and shaded black) accounts formore than 10% of all one-point downgrades. This forms initial evidence that asizable number of recommendation changes may have little significant impacton stock prices. The distribution also does not appear to resemble a normaldistribution, given that there are more left-tail observations than there are right-tail observations, implying negative skewness in the distribution. The secondchart in Figure2 shows the distribution of CARs for one-point upgrades. Thechart here tells a similar story in that the zero bin contains a sizable number ofobservations and that tail observations may have a large influence so that thetypical upgrade CAR may be very different from the mean upgrade CAR.

2.3 Impact of firm news events and influential observations on mean CARTable2 shows the distribution statistics of recommendation change subsam-ples sequentially from−4 to +4. These descriptive statistics illustrate the

602


Figure 2Histogram of CARs for one-point upgrades and downgradesThe sample of recommendation (rec) changes is from I/B/E/S Detail U.S. File 1994 to 2006. Each one-pointrec change (or rating change) is an analyst’s current rating minus his prior rating. Analyst initiations or ratingswith no prior outstanding ratings are excluded. The above shows the histogram of two-day [0,1] event CARsof one-point downgrades and one-point upgrades, respectively. CAR is the two-day buy-and-hold return aroundthe rec less the corresponding return on a size-B/M-momentum matched DGTW characteristic portfolio. Eachbin in the histogram is a CAR interval of 100 basis points (1%). The bin centered on a CAR of zero is shadedin black.

key point that the CAR distributions are not normal and that firm-specificnews-contaminated recommendations and outliers have a strong impact onthe means. For example, we illustrate with the fourth panel the rating change

603


of −1. Sample 1 is the full set of−1 downgrades. The average CAR is astatistically significant−3.551%, based on standard errors clustered by cal-endar day. However, the median CAR is only−1.716%, and the modal CARis just−0.5% (mid-point of 50 bps modal group). The Kolmogorov-SmirnovD statistic rejects the normality of the CAR distribution consistent with theobserved skewness and kurtosis. Another interesting statistic (third column) isthe percentage of positive-signed CARs. This shows that 30.2% of−1 ratingchanges actually had stock-price reactions of the wrong sign. Similar findingsare in the other panels of Table2.

We examine the impact of removing observations that are contaminatedby contemporaneous firm-specific news releases. First, we exclude observa-tions that fall in the three-day window around quarterly earnings announce-ment dates. This reduces the average CAR to−2.976% (see sample 2 inTable 2). Next, we also remove recommendations that fall in the three-daywindow around management earnings guidance days, and the average CARdrops dramatically to−1.913%. Finally, we remove days with multiple rec-ommendations, and the average CAR now becomes−1.562%. Although thisaverage is still statistically significant, we see that moving from sample 1 tosample 4 shaves the economic magnitude of the average CAR by more thanhalf, from−3.551% to−1.562%. The median CAR also falls from−1.716%to −1.148%. This halving of the mean CAR is also evident in some otherpanels of the table and highlights that a large fraction of the average recom-mendation CAR could be attributed to contemporaneous firm news releasesrather than to the recommendation itself, consistent withChen, Francis, andSchipper(2005) andAltinkilic and Hansen(2009).

Finally, we consider the impact of removing outlier observations from thissample that is uncontaminated by firm news using two different approaches.The first approach trims 5% from both tails of the sample distribution, and wefind that the average CAR reduces to−1.422% (see sample 5 row in Table2).The second approach uses the least trimmed squares (LTS) method (e.g.,Knezand Ready 1997) to identify outliers. Specifically, we estimate a regressionusing LTS with the CAR against a constant. We then compute the mean CARby excluding the LTS-identified outliers. The average CAR is now−1.264%.These results show that the removal of outliers from both tails further tempersthe magnitude of the typical recommendation absolute value average CAR.

3. Influential versus Non-influential Recommendation Changes

3.1 Methods for classifying recommendation changesIn this section, we identify recommendation changes that are influential andcompare them with non-influential recommendation changes. We report resultsfor two definitions of influential. The first method classifies a recommendationchange as influential if the CAR is in the correct directionand is statistically

604


Tabl

e2

The

impa

ctof

vario

usfil

ters

onre

com

men

datio

nev

entp

erce

ntag

eC

AR

Per

cent

iles

Filt

ered

Sam

ples

Mea

nM

ode

%C

AR+

Ske

wK

urt

KS

test

99%

75%

Med

ian

25%

1%#

Obs

Rec

omm

enda

tion

Cha

nge =−

4

1)F

ulls

ampl

e−

3.67

9 ***

−1.

50.

348

−2.

2910

.56

0.26

1***

30.4

80.

94−

1.25

8 ***

−4.

01−

70.9

837

12)

No

earn

ings

annc

days

−3.

135*

**−

1.5

0.34

8−

2.12

11.0

40.

256*

**30

.48

0.89−

1.22

3***

−3.

92−

55.1

131

93)

No

earn

ings

annc

orm

gtfo

reca

sts

days

−2.

114 *

**−

1.5

0.36

7−

2.30

15.3

10.

255*

**34

.67

1.08−

1.06

3 ***

−3.

44−

70.9

829

74)

No

earn

ings

annc

,mgt

fore

cast

sor

mul

tiple

rec

days−

1.27

8 **

−1.

50.

360

−0.

8925

.71

0.19

3***

14.8

40.

94−

1.00

1 ***

−3.

08−

36.1

024

25)

Rem

ove

5%fr

ombo

thta

ilsof

(4)

−1.

025 *

**−

1.5

0.34

4−

0.07

0.08

0.05

25.

310.

77−

1.00

1 ***

−2.

82−

7.59

218

6)R

emov

eLT

S-id

entifi

edou

tlier

sfr

om(4

)−

1.02

9 ***

−1.

50.

348

−0.

080.

690.

057*

6.78

0.78−

1.00

1 ***

−2.

86−

9.33

224

Rec

omm

enda

tion

Cha

nge=−

3

1)F

ulls

ampl

e−

2.91

6***

−0.

50.

392

−1.

437.

680.

203*

**27

.65

1.37−

0.84

0***

−4.

55−

43.9

267

12)

No

earn

ings

annc

days

−2.

234 *

**0

0.40

6−

1.56

10.2

60.

209*

**27

.86

1.36−

0.68

0 ***

−3.

75−

46.4

158

43)

No

earn

ings

annc

orm

gtfo

reca

sts

days

−1.

158*

**−

0.5

0.42

2−

1.46

16.0

50.

189*

**27

.86

1.53−

0.53

4***

−2.

91−

41.7

853

84)

No

earn

ings

annc

,mgt

fore

cast

sor

mul

tiple

rec

days−

0.31

3−

0.5

0.42

70.

1610

.38

0.16

3***

27.6

51.

64−

0.40

1 ***

−2.

46−

21.7

647

85)

Rem

ove

5%fr

ombo

thta

ilsof

(4)

−0.

475 *

**−

0.5

0.41

90.

090.

720.

061*

**8.

351.

30−

0.40

1 ***

−2.

16−

8.01

432

6)R

emov

eLT

S-id

entifi

edou

tlier

sfr

om(4

)−

0.56

9 ***

−0.

50.

414

−0.

120.

500.

063*

**6.

811.

27−

0.43

3 ***

−2.

17−

8.95

430

Rec

omm

enda

tion

Cha

nge=−

2

1)F

ulls

ampl

e−

4.00

7 ***

−0.

50.

315

−1.

8310

.45

0.19

5***

16.4

00.

61−

1.57

2 ***

−5.

48−

46.4

619

944

2)N

oea

rnin

gsan

ncda

ys−

3.37

2***

−0.

50.

331

−1.

9112

.59

0.20

5***

16.6

90.

71−

1.31

2***

−4.

58−

45.9

416

325

3)N

oea

rnin

gsan

ncor

mgt

fore

cast

sda

ys−

1.94

1***

−0.

50.

353

−1.

5324

.23

0.17

6***

16.9

30.

86−

1.02

6***

−3.

54−

33.8

114

569

4)N

oea

rnin

gsan

nc,m

gtfo

reca

sts

orm

ultip

lere

cda

ys−1.

432 *

**−

0.5

0.35

60.

0030

.74

0.13

9***

13.2

10.

83−

0.93

4 ***

−3.

20−

20.3

912

599

5)R

emov

e5%

from

both

tails

of(4

)−

1.27

7***

−0.

50.

341

−0.

520.

190.

059*

**4.

490.

61−

0.93

4***

−2.

90−

9.00

1134

16)

Rem

ove

LTS

-iden

tified

outli

ers

from

(4)

−1.

076 *

**−

0.5

0.35

9−

0.18

0.50

0.05

0***

6.89

0.79−

0.85

1 ***

−2.

85−

9.31

1173

3

Rec

omm

enda

tion

Cha

nge =−

1

1)F

ulls

ampl

e−

3.55

1 ***

−0.

50.

302

−1.

6313

.86

0.17

7***

14.3

00.

51−

1.71

6 ***

−5.

26−

39.3

751

679

2)N

oea

rnin

gsan

ncda

ys−

2.97

6 ***

−0.

50.

315

−1.

7616

.83

0.18

4***

13.7

30.

59−

1.48

0 ***

−4.

45−

37.6

741

490

3)N

oea

rnin

gsan

ncor

mgt

fore

cast

sda

ys−

1.91

3 ***

−0.

50.

332

−1.

0831

.43

0.15

3***

13.5

50.

72−

1.22

6 ***

−3.

69−

26.8

437

232

4)N

oea

rnin

gsan

nc,m

gtfo

reca

sts

orm

ultip

lere

cda

ys −1.

562*

**−

0.5

0.33

70.

3841

.12

0.12

1***

11.9

60.

73−

1.14

8***

−3.

44−

18.9

832

564

5)R

emov

e5%

from

both

tails

of(4

)−

1.42

2 ***

−0.

50.

318

−0.

400.

010.

045*

**4.

490.

50−

1.14

8 ***

−3.

13−

8.94

2930

86)

Rem

ove

LTS

-iden

tified

outli

ers

from

(4)

−1.

264*

**−

0.5

0.33

7−

0.14

0.47

0.04

2***

7.03

0.68−

1.07

4***

−3.

14−

9.77

3064

6

(co

ntin

ue

d)

605


Tabl

e2

Con

tinue

d

Per

cent

iles

Filt

ered

Sam

ples

Mea

nM

ode

%C

AR +

Ske

wK

urt

KS

test

99%

75%

Med

ian

25%

1%#

Obs

Rec

omm

enda

tion

Cha

nge=

01)

Ful

lsam

ple

−0.

089 *

*0

0.49

00.

2442

.71

0.13

4***

15.4

41.

94−

0.06

5 ***

−1.

97−

18.1

721

331

2)N

oea

rnin

gsan

ncda

ys−

0.10

2**

−0.

50.

486

0.38

59.9

60.

131*

**13

.84

1.81−

0.08

9***

−1.

89−

16.7

818

530

3)N

oea

rnin

gsan

ncor

mgt

fore

cast

sda

ys0.

027−

0.5

0.48

81.

6478

.47

0.11

7***

13.6

01.

79−

0.07

4 ***

−1.

83−

12.9

317

877

4)N

oea

rnin

gsan

nc,m

gtfo

reca

sts

orm

ultip

lere

cda

ys0.

070−

0.5

0.48

92.

8695

.29

0.11

1***

12.9

91.

78−

0.06

8***

−1.

79−

12.1

616

456

5)R

emov

e5%

from

both

tails

of(4

)0.

002−

0.5

0.48

80.

13−

0.11

0.02

4***

5.94

1.52

−0.

068 *

**−

1.57

−5.

4514

812

6)R

emov

eLT

S-id

entifi

edou

tlier

sfr

om(4

)−

0.04

6−

0.5

0.48

40.

060.

470.

037*

**7.

401.

61−

0.09

3 ***

−1.

70−

7.24

1560

4

Rec

omm

enda

tion

Cha

nge =+

1

1)F

ulls

ampl

e2.

503*

**0.

50.

676

2.06

17.9

40.

132*

**25

.68

4.47

1.49

8***−

0.67

−11.3

044

498

2)N

oea

rnin

gsan

ncda

ys2.

074*

**0.

50.

663

2.46

27.3

80.

128*

**22

.00

3.90

1.29

8***−

0.73

−10.0

835

436

3)N

oea

rnin

gsan

ncor

mgt

fore

cast

sda

ys2.

050*

**0.

50.

663

2.91

30.4

80.

128*

**21

.06

3.82

1.27

4***−0.

72−

9.23

3412

54)

No

earn

ings

annc

,mgt

fore

cast

sor

mul

tiple

rec

days

1.91

4***

0.5

0.66

12.

3123

.63

0.11

7***

19.5

03.

701.

228*

** −0.

72−

9.01

3077

25)

Rem

ove

5%fr

ombo

thta

ilsof

(4)

1.61

7***

0.5

0.67

90.

57−

0.01

0.05

4***

9.57

3.39

1.22

8***

−0.

51−

3.88

2769

66)

Rem

ove

LTS

-iden

tified

outli

ers

from

(4)

1.35

2***

0.5

0.65

20.

210.

360.

041*

**10

.06

3.32

1.09

5*** −

0.76

−6.

9729

144

Rec

omm

enda

tion

Cha

nge=+

2

1)F

ulls

ampl

e2.

303*

**0.

50.

658

2.32

20.4

30.

137*

**25

.62

4.14

1.23

1***−

0.72

−10.0

115

004

2)N

oea

rnin

gsan

ncda

ys1.

919*

**0.

50.

647

2.64

29.4

10.

133*

**22

.27

3.63

1.06

8***−

0.77

−9.

1612

187

3)N

oea

rnin

gsan

ncor

mgt

fore

cast

sda

ys1.

885*

**0.

50.

646

3.00

32.4

70.

133*

**20

.56

3.56

1.04

2***−0.

77−

8.53

1177

34)

No

earn

ings

annc

,mgt

fore

cast

sor

mul

tiple

rec

days

1.74

5***

0.5

0.64

52.

6321

.26

0.12

0***

18.7

83.

410.

991*

**−0.

76−

8.12

1068

55)

Rem

ove

5%fr

ombo

thta

ilsof

(4)

1.42

7***

0.5

0.66

10.

650.

100.

070*

**9.

183.

060.

991*

**−

0.57

−3.

6396

176)

Rem

ove

LTS

-iden

tified

outli

ers

from

(4)

1.14

9***

0.5

0.63

30.

250.

410.

052*

**9.

542.

970.

844*

**−

0.80

−6.

7810

121

Rec

omm

enda

tion

Cha

nge =+

3

1)F

ulls

ampl

e1.

111*

**0

0.55

41.

7013

.67

0.12

4***

22.3

63.

500.

530*

*−

2.00

−20.0

041

32)

No

earn

ings

annc

days

1.05

0***

00.

539

1.92

15.5

00.

131*

**27

.81

3.50

0.45

0−

1.95

−20.0

036

23)

No

earn

ings

annc

orm

gtfo

reca

sts

days

1.21

6***

00.

544

2.24

16.8

60.

136*

**27

.81

3.59

0.46

1−1.

86−

13.8

835

54)

No

earn

ings

annc

,mgt

fore

cast

sor

mul

tiple

rec

days

1.11

0***

00.

543

0.98

6.85

0.13

1***

22.3

63.

410.

458−

1.76

−13.8

832

25)

Rem

ove

5%fr

ombo

thta

ilsof

(4)

0.87

1***

00.

548

0.59

0.20

0.08

2***

10.7

22.

660.

458−

1.46

−5.

9729

06)

Rem

ove

LTS

-iden

tified

outli

ers

from

(4)

0.75

2***

00.

537

0.36

0.50

0.08

2***

11.5

12.

730.

439−

1.73

−8.

6530

7

(co

ntin

ue

d)

606


Tabl

e2

Con

tinue

d

Per

cent

iles

Filt

ered

Sam

ples

Mea

nM

ode

%C

AR +

Ske

wK

urt

KS

test

99%

75%

Med

ian

25%

1%#

Obs

Rec

omm

enda

tion

Cha

nge=+

4

1)F

ulls

ampl

e1.

359*

**−

0.5

0.55

21.

082.

410.

128*

**17

.18

2.93

0.67

3−

1.20

−7.

8822

32)

No

earn

ings

annc

days

1.24

7***−

0.5

0.56

51.

173.

160.

138*

**17

.30

2.64

0.73

5*−

1.17

−9.

0619

13)

No

earn

ings

annc

orm

gtfo

reca

sts

days

1.16

8***−

0.5

0.56

51.

143.

390.

136*

**17

.30

2.50

0.72

3*−

1.10

−9.

0618

64)

No

earn

ings

annc

,mgt

fore

cast

sor

mul

tiple

rec

days

1.08

8***−

0.5

0.55

11.

233.

850.

143*

**17

.30

2.40

0.53

1−

1.20

−9.

0616

75)

Rem

ove

5%fr

ombo

thta

ilsof

(4)

0.84

8***−

0.5

0.55

60.

750.

570.

069*

8.71

2.31

0.53

1−

1.08

−3.

7115

16)

Rem

ove

LTS

-iden

tified

outli

ers

from

(4)

0.54

3**−

0.5

0.53

20.

270.

610.

060

8.42

2.23

0.24

8−

1.26

−6.

8515

6

The

desc

riptiv

est

atis

tics

ofth

eab

norm

alre

turn

sto

reco

mm

enda

tion

(rec

)ch

ange

sar

ere

port

ed.R

ecom

men

datio

nsar

efr

omI/B

/E/S

Det

ailU

.S.F

ile19

93–2

006.

Rec

date

sar

efr

omF

irst

Cal

lfor

obse

rvat

ions

com

mon

betw

een

I/B/E

/San

dF

irstC

all.

Eac

hre

cch

ange

isan

anal

yst’s

curr

entr

atin

gm

inus

the

anal

yst’s

prio

rra

ting.

Ana

lyst

initi

atio

nsar

eex

clud

ed.R

atin

gsar

eco

ded

as1

(sel

l)to

stro

ngbu

y(5

),an

dra

ting

chan

ges

liebe

twee

n−

4an

d4.

Obs

erva

tions

with

few

erth

an3

anal

ysts

havi

ngou

tsta

ndin

gra

tings

are

excl

uded

.Eac

hpa

nelr

epor

tssu

mm

ary

stat

istic

sfo

rth

etw

o-da

y(0

,1)

buy-

and-

hold

even

tC

AR

(inpe

rcen

t)of

are

cch

ange

grou

p.D

aily

abno

rmal

retu

rnis

the

raw

retu

rnle

ssth

ere

turn

ona

size

-B/M

-mom

entu

mm

atch

edpo

rtfo

liow

ithC

AR

obse

rvat

ions

whe

reth

ela

gged

pric

eon

day

0is

<$1

excl

uded

.T

hem

ode

colu

mn

repo

rts

the

mid

poin

tof

the

50bp

sin

terv

alm

odal

grou

p.P

-val

ueof

the

med

ian

isco

mpu

ted

from

asi

gned

test

.Kur

tis

exce

ssku

rtos

isso

that

ano

rmal

dist

ribut

ion

wou

ldha

veK

urt

=0.

KS

test

isth

eK

olm

ogor

ov-S

mirn

ovDst

atis

ticte

stin

gfo

rthe

norm

ality

ofth

esa

mpl

edi

strib

utio

nw

here

aste

risks

repr

esen

trej

ectio

nof

the

null

ofno

rmal

ity.T

hesa

mpl

esar

efil

tere

dac

cord

ing

toth

em

entio

ned

crite

ria.N

oea

rnin

gsan

ncda

ysre

fer

toa

filte

rsa

mpl

eex

clud

ing

rec

chan

ges

that

occu

rin

the

thre

e-da

yw

indo

war

ound

the

firm

’sC

ompu

stat

quar

terly

earn

ings

anno

unce

men

tdat

e.N

om

gtfo

reca

stda

ysm

eans

we

excl

ude

rec

chan

ges

that

occu

rin

the

thre

e-da

yw

indo

war

ound

the

firm

’sm

anag

emen

tea

rnin

gsgu

idan

ceda

tes

prov

ided

byF

irst

Cal

lGui

delin

es.

No

mul

tiple

rec

days

refe

rsto

excl

udin

gda

ysw

here

mor

eth

anon

ean

alys

tis

sues

recs

onth

efir

m.

Rem

ove

5%fr

ombo

thta

ilsof

refe

rsto

the

rem

oval

ofth

eex

trem

e5%

ofou

tlier

sfr

omth

efil

tere

dsa

mpl

e.Le

ast

trim

med

squa

res

outli

ers

(LT

S)

are

iden

tified

byes

timat

ing

aLT

Sre

gres

sion

ofC

AR

agai

nsta

cons

tant

.*,*

*,an

d**

*re

pres

entt

hatt

henu

llof

zero

(nor

mal

ity)

isre

ject

edat

the

10%

,5%

,and

1%le

velr

espe

ctiv

ely,

for

the

mea

nan

dth

em

edia

n(K

Ste

st).

Sta

tistic

alsi

gnifi

canc

eof

the

mea

nis

base

don

stan

dard

erro

rscl

uste

red

byca

lend

arda

y.

607


significant using the market model. Specifically, we check if the CAR is in thesame direction as the recommendation change and the absolute value CAR ex-ceeds 1.96×

√2× σε. We multiply by

√2 since the CAR is a two-day CAR.

σε, the idiosyncratic volatility, is the standard deviation of residuals from adaily time-series regression of past three-month (trading days−69 to−6) firmreturns against market returns and the Fama-French factors SMB and HML.This measure roughly captures recommendation changes that observers wouldjudge to be influential, namely those that are associated with noticeable abnor-mal returns that can be attributed to the recommendation changes.

The second approach classifies a recommendation change as influential whenthe increase in abnormal turnover (abturn) is statistically significant. With thismeasure, a recommendation change is influential because it leads investorsto trade. FollowingLlorente, Michaely, Saar, and Wang(2002), abturn =log turnover− log turnover, where logturnover is the average of daily logturnover over the past three months, and logturnover= log (turnover+0.00000255).9 Specifically, we check if the cumulativeabturn is > 1.96×√

2× σabturn, whereσabturn is the standard deviation of the stock’sabturn inthe past three months (days−69 to−6 from the recommendation date).

The first row of Table3 reports the number of recommendation changes(reiterations, i.e.,recchg= 0, are excluded here) that are categorized into eachdimension of success. We see that 11.7% of all recommendation changes aredefined as influential in returns and 12.8% are defined as influential in turnover.While the typical recommendation change is not influential, more than onerecommendation out of ten is influential; 4.8% of recommendation changesare influential in both returns and turnover, 6.9% are influential in returns butnot turnover, and 8.0% are influential in turnover but not returns.

3.2 Analyst characteristics of influential recommendation changesWe characterize influential recommendation changes by examining severalanalyst-, firm-, and recommendation-level characteristics. We start with ana-lyst characteristics. We examine the relation between these variables and thelikelihood of an influential recommendation in both a univariate and a probitsetting.

1) Forecast accuracy: Loh and Mian(2006) show that analysts who pos-sess more accurate earnings forecasts issue more profitable contempo-raneous stock recommendations. It is possible that such analysts havemore impact. We compute theForecast accuracyquintile of an ana-lyst by sorting analysts within a firm-year into quintiles using the last

9 Daily turnover is from CRSP and defined as number of shares traded divided by the number of sharesoutstanding. Firms from NASDAQ have their shares traded divided by two to adjust for inter-dealer doublecounting.

608


Tabl

e3

Com

parin

gan

alys

tand

firm

char

acte

ristic

sof

influ

entia

lver

sus

non-

influ

entia

lrec

omm

enda

tion

chan

ges

Influ

entia

lbas

edon

firm

’sab

norm

alre

turn

sIn

fluen

tialb

ased

onfir

m’s

abno

rmal

turn

over

Diff

eren

ceD

if fer

ence

Cha

ract

eris

tics

Not

Influ

Influ

entia

lIn

flu–

Not

t−st

atN

otIn

fluIn

fluen

tial

Influ

–N

ott−

stat

Pan

elA

:Ana

lyst

and

reco

mm

enda

tion

char

acte

ristic

sN

umbe

rof

rec

chan

ges

77,5

3710

,292

76,5

9511

,234

11.7

%12

.8%

For

ecas

tacc

urac

yqu

intil

e2.

810

2.77

1−

0.03

9***

(−

2.75

)2.

810

2.77

6−

0.03

4**

( −2.

45)

Aw

ayfr

omco

nsen

sus

0.35

80.

416

0.05

8***

(9.8

4)0.

355

0.43

40.

079*

**(1

4.15

)S

tar

anal

yst

0.15

90.

205

0.04

6***

(9.5

3)0.

160

0.19

80.

038*

**(8

.19)

Abs

olut

ean

alys

texp

erie

nce

(#Q

trs)

28.4

529

.64

1.19

2***

(4.6

3)28

.44

29.6

11.

169*

**(4

.62)

Rel

ativ

ean

alys

texp

erie

nce

2.65

23.

354

0.70

2***

(5.0

3)2.

671

3.16

20.

491*

**(3

.52)

Con

curr

ente

arni

ngs

fore

cast

0.44

60.

510

0.06

4***

(10.

15)

0.44

80.

492

0.04

4***

(7.1

6)In

fluen

tialb

efor

e(a

nyst

ock)

0.56

20.

664

0.10

2***

(15.

57)

0.56

20.

656

0.09

4***

(14.

45)

Influ

entia

lbef

ore

(sam

est

ock)

0.10

00.

137

0.03

7***

(9.9

9)0.

100

0.13

20.

032*

**(8

.96)

Pan

elB

:Firm

char

acte

ristic

spr

ior

tore

com

men

datio

n

B/M

ratio

0.48

40.

498

0.01

4***

(2.7

8)0.

485

0.48

5−

0.00

1(−

0.12

)S

ize

($m

)89

71.7

7451

.0−

1520.6

***

(−

5.49

)88

10.9

8674

.9−

136.

03( −

0.47

)In

stitu

tiona

low

ners

hip

0.61

0.63

0.02

7***

(8.4

4)0.

610.

640.

029*

**(9

.69)

Dis

pers

ion ×

100

14.1

6715

.119

0.95

2(1

.31)

14.2

6714

.368

0.10

1(0

.15)

Idio

sync

ratic

vola

tility×

100

2.50

42.

195

−0.

310 *

**(−

13.6

4)2.

501

2.24

0−

0.26

1 ***

(−

11.2

8)To

talv

olat

ility×

100

2.87

62.

556

−0.

319 *

**(−

11.7

9)2.

872

2.60

7−

0.26

5 ***

(−

9.56

)D

aily

turn

over×

100

0.65

50.

603

−0.

052*

**(−

7.42

)0.

657

0.59

2−

0.06

5***

(−

9.59

)A

naly

stac

tivity

(#E

PS

fore

cast

s)86

.669

72.4

22−

14.2

47**

*(−

15.5

6)86

.156

77.1

12−

9.04

4 ***

(−

9.93

)

(co

ntin

ue

d)

609


Tabl

e3

Con

tinue

d

Influ

entia

lbas

edon

firm

’sab

norm

alre

turn

sIn

fluen

tialb

ased

onfir

m’s

abno

rmal

turn

over

Diff

eren

ceD

if fer

ence

Cha

ract

eris

tics

Not

Influ

Influ

entia

lIn

flu–

Not

t -st

atN

otIn

fluIn

fluen

tial

Influ

–N

ott -

stat

Pan

elC

:Cha

nge

infir

men

viro

nmen

taro

und

reco

mm

enda

tion

Lead

er-F

ollo

wer

Rat

io(L

FR

)of

rec

2.03

23.

176

1.14

4***

(12.

21)

1.95

53.

619

1.66

4***

(16.

57)

1To

talv

olat

ility

ofda

ilyre

t×10

0−

0.08

10.

350

0.43

1***

(21.

41)

−0.

067

0.21

80.

285*

**(1

3.60

)1

Idio

sync

ratic

vola

tility×

100

−0.

082

0.33

30.

414*

**(2

4.90

)−

0.06

80.

204

0.27

1***

(16.

09)

1D

ispe

rsio

n×10

00.

876

0.37

5−

0.50

1(−

0.55

)0.

834

0.69

1−

0.14

3(−

0.17

)1

Dai

lytu

rnov

er×

100

0.00

40.

096

0.09

2***

(18.

95)

−0.

012

0.20

20.

215*

**(3

3.71

)1

Ana

lyst

activ

ity(#

fore

cast

s)−

0.35

94.

978

5.33

7***

(9.6

5)−

0.15

13.

111

3.26

3***

(5.6

1)1

inF

Y1|F

orec

astR

evis

ion |×

100

−0.

010

0.01

10.

021

(0.8

9)−

0.00

90.

005

0.01

4(0

.62)

1in

FY

2| F

orec

astR

evis

ion|×

100

0.03

40.

127

0.09

3***

(3.9

3)0.

032

0.13

90.

107*

**(4

.38)

1in

LTG|F

orec

astR

evis

ion |×

100

0.04

00.

121

0.08

1***

(3.1

6)0.

035

0.15

60.

122*

**(4

.20)

Fra

ctio

nw

ithla

rge

indu

stry

vwre

turn

0.04

20.

123

0.08

0***

(18.

45)

0.04

80.

079

0.03

1***

(9.1

7)F

ract

ion

with

larg

ein

dust

ryew

retu

rn0.

067

0.12

30.

056*

**(1

2.12

)0.

071

0.09

30.

022*

**(5

.62)

Influ

entia

l(In

flu)

reco

mm

enda

tions

(rec

s)ar

eco

mpa

red

with

non-

influ

entia

lrec

s.In

fluen

tiald

efini

tions

are

base

dei

ther

onpr

ior

retu

rns

oron

prio

rab

norm

altu

rnov

er.

For

exam

ple,

for

retu

rns,

are

cch

ange

isin

fluen

tiali

sits

corr

ect-

sign

ed[0

,1]

CA

Ris

1.96

stan

dard

devi

atio

nsgr

eate

rth

anex

pect

edba

sed

onth

efir

m’s

prio

rth

ree-

mon

thId

iosy

ncr

atic

vola

tility

ofda

ilyre

turn

s.T

hesa

mpl

eis

from

I/B/E

/S(1

994–

2006

)w

ithea

rnin

gsan

nc.

days

,m

gt.

fore

cast

days

,an

dm

ultip

le-r

ecda

ysre

mov

ed(s

ampl

e4

inTa

ble

2).

Rei

tera

tions

are

excl

uded

.P

anel

Are

port

sth

eav

erag

ean

alys

tor

rec

char

acte

ristic

.F

orec

ast

accu

racy

quin

tile

isth

equ

intil

era

nk(lo

wer

rank

=gr

eate

rac

cura

cy)

ofth

ean

alys

tba

sed

onhi

sla

stun

revi

sed

FY

1ea

rnin

gsfo

reca

st.A

rec

mov

esAw

ay

fro

mco

nse

nsu

s(dum

my)

whe

nth

eab

solu

tede

viat

ion

ofth

ene

wre

cfr

omth

eco

nsen

sus

isla

rger

than

the

abso

lute

devi

atio

nof

the

prio

rre

cfr

omth

eco

nsen

sus.

Sta

ra

na

lyst

s(dum

my)

are

anal

ysts

who

are

rank

edinInst

itutio

na

lInv

est

or.A

na

lyst

exp

erie

ncei

sth

eno

.ofq

uart

ers

sinc

eth

ean

alys

tiss

ued

the

first

estim

ate

orre

con

I/B/E

/S.

Co

ncu

rre

nt

ea

rnin

gs

fore

cast=

1w

hen

the

anal

ysti

ssue

dan

yea

rnin

gsfo

reca

stin

the

thre

e-da

yw

indo

war

ound

the

rec.

Influ

en

tialb

efo

rein

dica

tes

that

the

anal

ystw

asin

fluen

tiali

nth

epa

st.P

anel

Bco

mpa

res

firm

char

acte

ristic

sw

ithva

riabl

ede

finiti

ons

foun

din

Sec

tion

3.P

anel

Cco

mpa

res

the

chan

ges

inth

efir

men

viro

nmen

tar

ound

the

rec.

Le

ad

er-

Follo

we

rR

atio

(LF

R)

isth

ega

psu

mof

the

prio

rtw

ore

csdi

vide

dby

the

gap

sum

ofth

ene

xttw

ore

cs.A

ratio

>1

indi

cate

sa

lead

eran

alys

twho

sere

csar

equ

ickl

yfo

llow

edby

othe

rbr

oker

s’re

cs.W

eal

soco

mpu

teth

e1

inTo

talv

ola

tility

ofda

ilyre

turn

sfr

omth

epr

ior

thre

em

onth

sto

the

thre

em

onth

saf

ter

the

rec

(exc

ludi

ng±

5da

ysar

ound

the

even

t).

We

also

com

pute

inth

esa

me

man

ner

othe

rch

ange

varia

bles

.The

repo

rted

valu

esfo

rTo

tal,

Idio

syn

cra

ticvo

latil

ity,D

isp

ers

ion,

and

Tu

rnov

era

rem

ultip

lied

by10

0(e

.g.,

1.0

repr

esen

ts1.

0%).

1| F

ore

cast

Rev

isio

n| is

com

pute

dfo

rth

eF

Y1

and

FY

2an

dlo

ng-t

erm

grow

th(L

TG

)co

nsen

sus

anal

ystf

orec

astr

evis

ions

from

thre

em

onth

sbe

fore

toth

ree

mon

ths

afte

rth

ere

c.F

Y1

and

FY

2co

nsen

sus

fore

cast

revi

sion

sar

esc

aled

bypr

ice

(out

lier

revi

sion

sm

ore

than

100%

ofpr

ice

are

rem

oved

),an

dLT

Gfo

reca

sts

are

inpe

rcen

tage

s.La

rge

Fam

a-F

renc

h49

indu

stry

retu

rns

=1

whe

nre

com

men

datio

nsco

inci

dew

itha

larg

e( >

1.96

stde

vth

anex

pect

ed)

abno

rmal

retu

rnm

ovem

enti

nth

efir

m’s

indu

stry

valu

e-w

eigh

ted

(vw

)or

equa

l-wei

ghte

d(e

w)

retu

rns

over

days

[0,1

].A

ster

isks

inth

edi

ffere

nce

colu

mns

indi

cate

stat

istic

alsi

gnifi

canc

eus

ing

stan

dard

erro

rscl

uste

red

byca

lend

arda

yw

here

*,**

,and

***

repr

esen

tsig

nific

ance

leve

lsof

10%

,5%

,and

1%,r

espe

ctiv

ely.

610


unrevised FY1 forecast of the analyst on the I/B/E/S Detail U.S. File.Only firms with at least five analysts are included. TheForecast accu-racy rank (1 being the most accurate) is assigned to the analyst for therecommendations that the analyst issues during the 12-month windowthat overlaps three months into the next fiscal year, followingLoh andMian (2006). Overlapping the 12-month period into the next fiscal yearallows the accuracy rank to be applied during the months when the fis-cal year’s actual earnings are announced. Note that this rank is a perfectforesight rank and is not known at the time of the recommendation sinceactual earnings have not yet been announced.

2) Away from consensus: Jegadeesh and Kim(2010) formulate a test forherding and contend that if analysts herd, recommendations that go to-ward the consensus would have a smaller price impact than those thatgo away from the consensus. Following their paper, we define recom-mendations that go away from consensus as those where the absolutedeviation of the new recommendation from the consensus is larger thanthe absolute deviation of the prior recommendation from the consensus.The consensus recommendation is defined as the mean recommenda-tion level that includes the most recent non-stale recommendation is-sued by all analysts covering the firm (see Section1.1for the definitionof stale). This variable is defined based on three-point ratings to accountfor rating distribution differences between brokers.

3) Star analyst: This is an indicator variable that equals one if the analystis ranked as an All-American (first, second, third, or runner-up teams) inthe annual polls in theInstitutional Investormagazine. Analyst namesin I/B/E/S are matched toInstitutional Investorpolls (published in theOctober issue), and an analyst maintains the star status for 12 monthsbeginning the November after the polls. TheStar analystindicator vari-able proxies for the reputation of the analyst and the market’s attentive-ness to the recommendation (the market could pay more attention tocalls from star analysts).

4) Analyst experience: Mikhail, Walther, and Willis(1997) show that ana-lysts improve their earnings forecast accuracy with experience. Hence,it is possible that experience could be related to the impact of stockrecommendation changes.Analyst experienceis measured as the numberof quarters since the analyst issued the first earnings forecasts or stockrecommendation on I/B/E/S. Two measures of experience are computed.The first isAbsolute analyst experience, which is the number of quartersthat he appeared on I/B/E/S. The second is theRelative analyst experi-ence, which is the number of quarters the analyst has covered that specificfirm minus the average experience for all analysts covering the firm.

5) Concurrent earnings forecast: Kecskes, Michaely, and Womack(2009)report that stock recommendations accompanied by earnings forecast

611


revisions are more profitable and have larger price reactions. Therefore,we include aConcurrent earnings forecastindicator variable indicatingwhether the same analyst issued any type of earnings forecast in thethree-day window around the recommendation change.

6) Influential before: If an influential recommendation is related to analystskill that is persistent, being influential in the past could be related tothe current likelihood of being influential.

We compute the average of these analyst-specific variables for the differ-ent rating-change groups. Table3 reports the averages for observations wherethese variables can be computed. The average analystForecast accuracyquin-tile of influential recommendation changes is 2.771, versus 2.810 for non-influential recommendations. The difference is statistically significant, but itseconomic importance is small; 41.6% of influential recommendation changesmove away from the consensus, while only 35.8% of non-influential recom-mendation changes move away from the consensus—the difference is signif-icant and sizable. Also, star analysts are responsible for 20.5% of influentialrecommendation changes and 15.9% of non-influential recommendations. In-fluential recommendations are also associated with higherAbsoluteandRel-ative analyst experience. A larger proportion of influential recommendationchanges have concurrent earnings forecasts issued together with the recom-mendation change. Finally, being influential in the past for the same stock, aswell as for any stock, appears to be positively related to the current recommen-dation becoming influential. Using the second definition of influential (basedon increases in abnormal turnover) yields similar patterns of differences. Ofthe many variables we report, those associated with economically larger dif-ferences areAway from consensus, Star analyst, Concurrent earnings forecast,andInfluential beforevariables.

Many of the variables we consider are correlated. To assess more preciselythe relation between these variables and the probability that an analyst makesan influential recommendation, we estimate a probit model. This model allowsus to estimate not only the incremental contribution of each variable, but alsoits economic significance. We cluster the standard errors in the probits by an-alyst as well as by firm (two-way clustering suggested byThompson 2010).Although some of these variables have been examined in the literature assess-ing the determinants of the stock-price reaction to analyst recommendationchanges, they have not been considered in a unified fashion, nor have they beenexamined in the context of identifying an influential recommendation changein the manner we defined.

The probit estimates in Table4 show that a recommendation change is sig-nificantly more likely to be influential if it is by an analyst who has made aninfluential recommendation before. The marginal effect ofInfluential before(any stock)is 2.88%. Such an effect is large, since the unconditional probabil-ity of a recommendation change being significant is 11.70%. This means that

612


when an analyst made at least one influential recommendation change in thepast for any stock, the probability that the analyst’s current recommendationchange will be influential increases by 2.88%. In addition, if the analyst wereinfluential before for the same stock, the probability of being influential goesup by another 1.21%. Other analyst variables that also have large economiceffects are the variableRec away from consensus, which has a marginal effectof 2.74%,Star analyst(3.84%), andConcurrent earnings forecast(2.22%).

Table 4Predicting when a recommendation change will be influential

Influential based on Influential based onfirm’s abnormal returns firm’s abnormal turnover

Explantory Variable Coefficient Marg. Eff Coefficient Marg. Eff

Influential before (any stock) 0.154*** 2.88% 0.130*** 2.58%(8.35) (7.17)

Influential before (same stock) 0.065*** 1.21% 0.052** 1.03%(2.93) (2.54)

Rec level 0.045*** 0.80% 0.008 0.16%(4.01) (0.74)

Absolute value ofrecchg −0.017 −0.16% 0.001 0.01%(−1.00) (0.03)

Upgrade Dummy 0.080*** 1.50% −0.032 −0.64%(4.01) (−1.63)

Reg FD Dummy 0.206*** 3.85% 0.219*** 4.37%(8.82) (9.24)

Settlement Dummy 0.093*** 1.73% 0.171*** 3.42%(3.84) (6.81)

Past forecast accuracy quintile −0.011* −0.24% −0.009 −0.21%(−1.90) (−1.59)

Away from consensus 0.147*** 2.74% 0.148*** 2.96%(9.87) (10.18)

Star analyst 0.207*** 3.87% 0.156*** 3.11%(9.36) (7.22)

Absolute analyst experience −0.001* −0.41% −0.001 −0.31%(−1.96) (−1.61)

Relative analyst experience 0.001 0.14% 0.000 0.07%(0.84) (0.36)

Concurrent earnings forecast 0.119*** 2.22% 0.110*** 2.19%(7.87) (7.71)

Past Leader-Follower Ratio (LFR) 0.006*** 0.36% 0.004** 0.28%(2.74) (1.98)

Log(B/M) −0.100*** −1.51% −0.061*** −0.99%(−9.66) (−6.15)

Log(Size) −0.082*** −2.49% 0.017** 0.53%(−10.67) (2.46)

Price momentum 0.031** 0.34% 0.042*** 0.49%(2.38) (3.40)

Log(Institutional ownership) 0.049** 0.38% 0.138*** 1.15%(2.15) (5.30)

Log(Turnover) 0.042*** 0.62% −0.084*** −1.32%(2.82) (−5.36)

Log(Idiosyncratic volatility) −0.351*** −3.45% −0.090*** −0.95%(−15.10) (−4.04)

Dispersion 0.032*** 0.38% 0.033*** 0.42%(3.13) (2.92)

(continued)

613


Table 4Continued

Influential based on Influential based onfirm’s abnormal returns firm’s abnormal turnover

Explantory Variable Coefficient Marg. Eff Coefficient Marg. Eff

Log(Analyst activity) −0.138*** −2.17% −0.148*** −2.48%(−11.49) (−12.86)

Pseudo R-sq 0.04745 0.03654# Observations 58384 58384Chi-Sq test 1485.68*** 1323.57***

The binary dependent variable is whether a recommendation (rec) is influential and the sample is sample 4 fromTable2 with reiterations excluded and firm and analyst characteristics required. Marginal effects are reported be-low the coefficient estimates. The marginal effect for continuous (dummy) explanatory variables represents thechange in the predicted probability when the independent variable changes by one standard deviation (changesfrom 0 to 1). There are two definitions of influential. First, influential recs are those when a correct-signed CARis 1.96 standard deviations greater than expected based on the firm’s prior three-monthIdiosyncratic volatilityofdaily returns. The second uses an abnormal turnover greater than 1.96 standard deviations of that expected fromthe abnormal daily turnover in the stock’s prior three-month history. TheRec Levelis the rating level after the recchange (recchg) (1= sell to 5= strong buy). The absolute value of therecchg, Upgradedummy,Reg FDdummy(=1 after Aug 2000), andSettlementdummy (=1 in 2003 and after) are also included. PastForecast accuracyquintile is the average quintile rank of the analyst (smaller ranks denote greater accuracy).Away from consensus= 1 when the absolute deviation of the new rec from the consensus is larger than the absolute deviation of theprior rec from the consensus.Star analysts= 1 for ranked analysts in the most recentInstitutional Investorpolls.Absolute analyst experienceis measured as the # of quarters in I/B/E/S.Relative analyst experiencesubtractsthe average experience of other covering analysts.Concurrent earnings forecast= 1 when the same analystissued any earnings forecast in the three-day window around the rec.Leader-Follower Ratioslarger than onedenote leader analysts.Turnover, Idiosyncratic volatility, andDispersionare based on prior three-month aver-ages.Analyst activityis total number forecasts issued by all analysts in the prior three months. *,**, and ***denote significance levels of 10%, 5%, and 1%, respectively, using standard errors clustered in two dimensions(by analyst and firm) withz statistics in parentheses.

PastForecast accuracyandAnalyst experiencedo not seem to provide muchincremental predictability in identifying influential recommendations.

We also include the analyst’s prior yearLeader-Follower Ratio(LFR) as apredictive variable.Cooper, Day, and Lewis(2001) use this ratio to gauge theextent to which a forecast event leads other analysts to revise their estimates.A ratio larger than one denotes a leader analyst.10 The coefficient on pastLFRis statistically significant, although the marginal effect on predicting influen-tial changes is modest at 0.36% (marginal effects of continuous variables arethe change in the probability of making an influential recommendation whenthe explanatory variables change by one standard deviation). Altogether, wesee that several analyst-specific variables are important in predicting influ-ential recommendation changes. Our second definition of influential providessupportive evidence on the role of the variables discussed in this section. Be-cause skill-related variables are strongly related to the influential likelihood,

10 To compute this, the gaps between the current recommendation and the previous two recommendations fromother brokers are computed and summed. The same is done for the next two recommendations. The leader-follower ratio is the gap sum of the prior two recommendations divided by the gap sum of the next two rec-ommendations. A ratio larger than one shows that other brokers issue new ratings quickly in response to theanalyst’s current recommendation.

614


we believe that it is analyst ability rather than chance that allows for thegeneration of influential recommendation changes.

3.3 Firm characteristics of influential recommendationsWe consider how firm characteristics differ between the influential and non-influential subsamples of Panel B of Table3. Certain firm characteristics couldcreate conditions that make it more likely for analysts to make significant rec-ommendation changes. For example, analysts may have more influence whenthe value of the firm depends more on growth options that are harder to valuethan assets-in-place. We see that influential recommendation changes tend tobe issued on firms that have a lower B/M ratio and are therefore more likelyto be growth firms. Also, the influential subsample is associated with smallersize, higherInstitutional ownership, lower Total volatility and Idiosyncraticvolatility, lowerTurnover, and lower level ofAnalyst activityas proxied by thenumber of earnings forecasts during the prior three months. The results sug-gest that analysts can more easily affect investors’ beliefs about a firm whenthey are speaking in a smaller crowd. However, institutional investors are themain consumers of analyst reports, so it is not surprising that analysts are morelikely to have a significant impact for highInstitutional ownershipfirms. Thesecond definition of influential recommendation changes yields similar results.

The probit estimation in Table4 provides supplementary results. In the inf-luential in returns column, the firms that are more likely to receive impactfulrecommendation changes are growth firms, small firms, highInstitutionalownershipfirms, high priorTurnoverfirms, low Idiosyncratic volatilityfirms,higherearnings forecastDispersionfirms,and lowpriorAnalystactivity(numberof forecasts issued) firms. Some firm variables have large marginal effects onthe probability of a recommendation change being influential. A one-standard-deviation change in size has a−2.49% impact on being influential. The effectfor B/M is also high, at−1.51%. The impact ofAnalyst activityis −2.17%.These results are generally consistent with the idea that an analyst is able tocontribute the most when the information environment surrounding the firm isuncertain (e.g., small size, low B/M, and highDispersion). Similar results areobtained in the probit estimation using the influential in turnover definition.

3.4 Changes to firm characteristics in response to influential reco-mmendations

Since an influential recommendation change by definition causes a significantstock-price or volume reaction, it must impact the way that investors view orvalue the firm. TheLFR provides some insight from the perspective of otheranalysts (see Panel C of Table3). If other analysts begin to issue recommen-dations quickly in response to the influential event, theLFR computed basedon the influential recommendation would be large. Indeed, this is the case.The LFR of influential recommendations is 3.176, about 57% larger than the

615


averageLFR of 2.032 in the non-influential subsample. The difference is evenstarker in the influential in turnover definition, where the averageLFR of in-fluential recommendation changes is about 85% larger than the averageLFRof the non-influential subsample. Analyst activity also increases by 4.978 fore-casts three months after the event. This increase is about 7% of the originallevel of analyst activity (Panel B of Table3 shows there are 72.422 forecastsin the three months prior to recommendation). Finally, average dailyTurnovergoes up by 0.096 from a benchmark of 0.603% prior to the recommendation—about a 15% increase. We also compare the change in average absolute magni-tudes of monthly consensus earnings forecast revisions (scaled by price, withoutlier observations larger than 100% of price removed) three months afterthe event to three months before. This change is higher for the influential sub-samples compared to the non-influential subsample (e.g., for FY2 revisions,0.127% is more than three times 0.034%).

Volatility also increases after an influential recommendation change. Wecompare volatility three months before and three months after the event, skip-ping the ten days around the event. The change in dailyIdiosyncratic volatil-ity of returns is 0.333% for influential recommendations, while it is slightlynegative or close to zero for non-influential changes. Based on the averageId-iosyncratic volatilityof 2.195% (in Panel B of Table3), a 0.333% increaserepresents a sizable increase of about 15% from the prior level.

Finally, we investigate whether influential recommendations have an impacton other firms in the same industry as the firm on which the recommendationis made. If the analyst’s influential recommendation contains industry infor-mation, similar firms could also be impacted by the revision. For example,banking stocks were also negatively affected by Meredith Whitney’s down-grade of Citigroup. Evidence of such positive spillover is also consistent withthe theoretical work ofVeldkamp(2006) on information markets that predictthat agents have incentives to produce information with implications for a sub-set of assets. To allow a reasonable number of firms in the industry, we usetheFama and French(1997) 49 industry groups and show both value-weightedand equal-weighted industry returns. The industry’s return is considered largewhen its two-day (0,1) market-adjusted return is in the same direction as therecommendation change and is greater than 1.96×

√2× σi , whereσi is the

standard deviation of residuals from a time-series regression of the industry’sdaily returns against market returns for the past three months. We show thatinfluential recommendations are associated with a larger fraction of large-impact industry returns. In Table3, Panel C, about 12.3% of the influentialrecommendation changes are influential for the industry as well.

3.5 Post regulation probability of being influentialOur sample period straddles the pre- and post-regulation periods governing se-curity analysis. Increased regulation can either stifle analysts’ ability to produce

616


impactful research (e.g., by limiting useful information channels) or increasethe likelihood of influential recommendations (e.g., by mitigating conflicts ofinterest). Our probit estimations include indicator variables for Regulation FairDisclosure (Reg FD= 1 from Sep 2002 onward) and the Global Analyst Set-tlement (Settlement= 1 from year 2003 onward). AfterReg FDwas passed inAugust 2000, analysts were no longer allowed to get access to private infor-mation from firm executives. If such private information is one of the mainsources of influential recommendation changes, we would expect influentialrecommendation changes to abate after the passage of the law. For exam-ple,Gintschel and Markov(2004) show evidence that selective disclosure wascurtailed afterReg FDand the absolute price impact of analyst output wasreduced. We find, however, that the coefficient forReg FDis significantly pos-itive, implying that influential recommendation changes are even more likelyafterReg FD(the marginal effect is a large 3.58%).

For the period after theSettlement, Kadan, Madureira, Wang, and Zach(2009) find that the overall informativeness of recommendations was reduced(their sample goes up to 2004 only).Boni (2006) also finds similar evidence.Although the overall informativeness of recommendations decreased,Kadan,Madureira, Wang, and Zach(2009) also find that the falling number of op-timistic recommendations could have caused optimistic recommendations tobecome more informative. Our probit estimations find some evidence that arecommendation change is more likely to be influential after theSettlement(marginal effect is 1.73%). When we add an interaction variableupgrade×Settlementto the estimations, we find that this interaction term is statisti-cally insignificant for both definitions of influential. Altogether, our resultssuggest that, although the overall impact ofReg FDand theSettlementcouldhave reduced the mean price impact of recommendations, the probability thata recommendation change is influential actually increased.11 It is thereforepossible that regulation scrutiny improved the overall quality of analystrecommendations.

3.6 Do influential recommendation changes come from only a subsetof analysts?

Finally, we investigate whether influential recommendation changes are pro-duced by only a subset of analysts. Our evidence so far points to the fact thatinfluential recommendations are associated with certain analyst skills and firmenvironments that promote the utilization of such skills. If this were true, influ-ential recommendation changes should emanate from only a subset of analysts

11 One alternative explanation for why recommendations are more influential in the post-regulation period is thatrecommendation dating became more accurate later in the sample period and hence aligns better with contempo-raneous stock price reactions. To test this, we drop the first three years (1994 to 1996) of the recommendationsin the probits. Although the coefficients are attenuated, we still find marginal effects of about 1.14% to 2.75%for theReg FDandSettlementvariables.

617


Figure 3Histogram of proportion of an analyst’s influential recommendation changesWe plot the histogram of the proportion of an analyst’s recommendation (rec) changes that are influential. Wefocus on analysts who made at least five recs in the 1994 to 2006 period. The first chart uses the firm’s pastIdiosyncratic volatilityof daily returns to determine if a rec change is influential (1.96 standard deviationsaway). The second uses the history of prior abnormal turnover to determine if a rec change leads to an increasein abnormal turnover (1.96 standard deviations more).

that possess such skills. Figure3 plots the histogram of the proportion of ananalyst’s recommendation changes that are influential. We limit this analy-sis to analysts who made at least five recommendation changes in the sample

618


period. If all analysts were equally capable of making influential recommen-dation changes, we would expect the distribution to peak around the averageproportion of influential recommendation changes in the entire sample. Thefigure shows otherwise. The first chart uses the returns definition of influential.Although the unconditional proportion of influential recommendation changesis 11.7% in the sample, 24.9% of all analysts never issue a single influentialrecommendation change in their recommendation histories (in our sample).For the turnover definition of influential, about 20.8% of analysts have neverissued an influential recommendation change. This skewed distribution indi-cates that only some analysts are influential and that there is a sizable pro-portion of analysts whose recommendation changes never have a noticeablestock-price impact, consistent with the skill story. These distribution statisticstell a consistent story with our analyst characteristics analysis.

Table5 shows that, if an analyst has ever been influential, about 22.1% ofthe analyst’s recommendations are influential. This number is 23.5% for theturnover definition of influential. Table5 also compares the characteristics ofanalysts who issue at least one influential recommendation change (Ever influ-ential) in the sample versus the other (Never influential) analysts. One wouldexpect theEver influentialgroup to dominate theNever influentialgroup interms of skill, experience, star status, etc. Indeed, this is the case.Ever in-fluentialanalysts have better average analyst earnings forecast accuracy ranks.They are also more likely to have once been aStar analystand have greater ab-solute and relativeAnalyst experiencecompared toNever influentialanalysts.The difference in theStar analystproportion is especially large, with 25.3%stars for the Ever influential group versus 9.2% for the Never group. It is pos-sible, however, that an analyst is more likely to be selected as aStar analystbecause he or she has had influential recommendations. There is only mixedevidence that analysts in theEver influentialgroup issue more recommenda-tionsAway from consensusand issueConcurrent earnings forecastswith theirrecommendations.

4. Robustness Tests and Different Samples

4.1 Other definitions of influentialWe also use other methods of classifying recommendation changes as influ-ential and use different samples with generally similar results (results not re-ported here are in the online appendix of this article). We summarize thosemethods here.

First, we are sensitive to potential issues arising from the usage of an an-alyst’s own prior recommendation in the computation ofrecchg. If the priorrecommendation is stale,recchgcould contain stale information and we couldunderstate the fraction of influential recommendations. However, we are carefulto use the review date in I/B/E/S and the stopped date so that we have confi-dence that the prior rating is still outstanding whenever we compute arecchgvalue.

619


Tabl

e5

Ana

lyst

sw

hoha

dat

leas

tone

influ

entia

lrec

omm

enda

tion

chan

ge

Influ

entia

lbas

edon

firm

’sab

norm

alre

turn

sIn

fluen

tialb

ased

onfir

m’s

abno

rmal

turn

over

Diff

eren

ceD

if fer

ence

Cha

ract

eris

tics

Nev

erE

ver

Influ

entia

lE

ver–

Nev

ert-

stat

Nev

erE

ver

Influ

entia

lE

ver–

Nev

ert-

stat

Num

ber

ofA

naly

sts

935

2,82

178

12,

975

75.1

%79

.2%

%in

fluen

tialr

ecs

for

typi

cala

naly

st0.

0%22

.1%

0.0%

23.5

%F

orec

asta

ccur

acy

qunt

ile2.

894

2.81

9−

0.07

6 ***

(−

3.20

)2.

907

2.81

9−

0.08

8 ***

(−

3.38

)A

way

from

cons

ensu

s0.

357

0.36

70.

010

(1.5

2)0.

348

0.36

90.

021*

**(2

.88)

Was

once

aS

tar

anal

yst

0.09

20.

253

0.16

1***

(12.

91)

0.10

10.

243

0.14

2***

(10.

60)

Abs

olut

ean

alys

texp

erie

nce

(#Q

trs)

18.0

924

.42

6.32

9***

(10.

11)

18.7

123

.93

5.21

8***

(7.5

7)R

elat

ive

anal

yste

xper

ienc

e−

1.26

1.36

2.62

4***

(9.6

1)−

0.71

1.08

1.78

8***

(5.9

4)C

oncu

rren

tear

ning

sfo

reca

st0.

477

0.46

2−

0.01

5*( −

1.66

)0.

461

0.46

70.

005

(0.5

5)

Thi

sta

ble

repo

rts

the

aver

age

anal

ysto

rre

com

men

datio

n(r

ec)

char

acte

ristic

byan

alys

tsw

how

ere

ever

influ

entia

lver

sus

thos

ew

how

ere

neve

rin

fluen

tial.

The

perc

enta

geof

anal

ystr

ecs

that

are

influ

entia

lis

the

aver

age

prop

ortio

nof

anin

divi

dual

anal

yst’s

recs

that

are

influ

entia

lcon

ditio

nalo

nth

ean

alys

teve

rbe

ing

influ

entia

lin

the

who

lesa

mpl

e.O

nly

anal

ysts

who

had

atle

ast

5re

com

men

datio

ns(r

ecs)

inth

e19

94–2

006

perio

dar

ein

clud

ed.

The

rear

etw

ode

finiti

ons

ofin

fluen

tial.

Firs

t,in

fluen

tialr

ecs

are

thos

ew

hen

aco

rrec

t-si

gned

[0,1

]C

AR

is1.

96st

anda

rdde

viat

ions

grea

ter

than

expe

cted

base

don

the

firm

’spr

ior

thre

e-m

onth

Idio

syn

cra

ticvo

latil

ityof

daily

retu

rns.

The

seco

ndus

esan

abno

rmal

turn

over

grea

ter

than

1.96

stan

dard

devi

atio

nsof

that

expe

cted

from

the

abno

rmal

daily

turn

over

inth

est

ock’

spr

ior

thre

e-m

onth

hist

ory.

The

sam

ple

isfr

omI/B

/E/S

with

earn

ings

anno

unce

men

tday

s,m

anag

emen

tfor

ecas

tda

ys,

and

mul

tiple

-rec

days

rem

oved

(sam

ple

4in

Tabl

e2)

.I/B

/E/S

rec

date

sar

ere

plac

edby

thos

efr

omF

irst

Cal

lwhe

neve

rav

aila

ble.

Rei

tera

tions

(rec

chan

ge=

0)ar

eal

soex

clud

ed.

For

ecas

tacc

urac

yqu

intil

eis

the

quin

tile

rank

ofth

ean

alys

tbas

edon

his

last

unre

vise

dF

Y1

earn

ings

fore

cast

for

that

fisca

lyea

rac

cord

ing

toLo

han

dM

ian

(200

6).A

rec

mov

esA

wa

yfr

om

con

sen

sus(d

umm

yva

riabl

e)w

hen

the

abso

lute

devi

atio

nof

the

new

rec

from

the

cons

ensu

sis

larg

erth

anth

eab

solu

tede

viat

ion

ofth

epr

ior

rec

from

the

cons

ensu

s(a

sin

Jega

dees

han

dK

im20

10).

Sta

ra

na

lyst

s(dum

my

varia

ble)

are

anal

ysts

who

are

rank

edas

All-

Am

eric

ans

inth

em

ostr

ecen

tann

ual

Inst

itutio

na

lInv

est

orp

olls

.Ab

solu

tea

na

lyst

exp

erie

nceis

mea

sure

das

the

#of

quar

ters

inI/B

/E/S

.Re

lativ

ea

na

lyst

exp

erie

nces

ubtr

acts

the

aver

age

expe

rienc

eof

othe

rco

verin

gan

alys

ts.

Co

ncu

rre

nte

arn

ing

sfo

reca

stisa

dum

my

indi

catin

gw

heth

erth

ean

alys

tis

sued

any

type

ofea

rnin

gsfo

reca

stin

the

thre

e-da

yw

indo

war

ound

the

rec.

Ast

eris

ksin

the

diffe

renc

eco

lum

nsin

dica

test

atis

tical

sign

ifica

nce

usin

gst

anda

rder

rors

clus

tere

dby

cale

ndar

day,

whe

re*,

**,a

nd**

*re

pres

ents

igni

fican

cele

vels

of10

%,5

%,a

nd1%

,res

pect

ivel

y.

620


Table 6Different definitions of influential

Method Using alternative methods to define influential Total Non-Influential InfluentialPercentage

0 Recchg Original 87,829 77,537 10,292 11.7%1 Recchg Last 50,024 44,932 5,092 10.2%2 Recchg Consensus 96,282 86,942 9,340 9.7%3 Influential based on raw returns 87,829 78,379 9,450 10.8%4 Using rec-free days to compute prior volatilties 87,829 77,203 10,626 12.1%5 Removing obs with large pre-event absolute return 76,638 68,100 8,53811.1%

The fraction of influential recommendation changes (recchg) for different subsamples (methods) is reported.The definition of influential here is based on abnormal returns. The originalrecchg(method 0) variable is theanalyst’s current rec minus the analyst’s own prior rec. A prior rec is outstanding only if it has a less-than-one-year-old review date and has not been stopped by the broker. Arecchgis considered influential when the [0,1]CAR is the same sign as therecchgand the CAR is 1.96 standard deviations greater than expected based onthe firm’s prior three-monthIdiosyncratic volatilityof daily returns. Reiterations are excluded since therecchgis basically zero. Methods 1 and 2 amend therecchgdefinition: 1 is based on the current rec minus the last recfrom any analysts (recchg last), and 2 is based on the current rec minus the most recent consensus (recchg con).Method 3 is method 0 except that arecchgis influential if the cumulative raw (not abnormal) return is influential.Method 4 amends method 0 by using only rec-free days to compute the prior return volatility of the firm. Thisremoves days with recommendations from the prior three-month [−69,−6] period used to computeIdiosyncraticvolatility. Method 5 uses method 0 except that it removes observations where the pre-event [−2,−1] containsan influential abnormal return in any direction.

Nevertheless, two alternativerecchgvariables are considered. The first isthe recommendation change value computed as the current recommendationminus the last recommendation by any analyst (we denote this asrecchg last).This captures a revision in the time series of recommendations. The second isthe current recommendation minus the most recent consensus recommenda-tion (recchg con). We estimate a multiple regression of CAR against the orig-inal recchg, recchg last, andrecchg convariables and show that the two newrecommendation-change variables do have incremental statistical power forthe event CAR, although the originalrecchgvariable has the largest economicsignificance (coefficients are 140, 8, and 19 basis points, respectively). Theserecommendation-change variables are based on three-point ratings (mappedfrom five-point rating scales), since it is not possible to form a consensus thatmixes different rating scales. We then redo our analysis with the two alter-native recommendation-change measures. We show that the influential (in re-turns) fraction forrecchg lastandrecchg conis 10.2% and 9.7%, respectively.Table6 reports these fractions. The influential fraction remains similar. In fact,the originalrecchgvariable is the best avenue to locate influential recommen-dation changes. Re-estimating Tables2 and4 with these two new measures ofrecchgdoes not materially affect any of our other results.

We investigate whether coding a recommendation change as influential if itis influential in raw returns as opposed to abnormal returns makes a differenceto our conclusions. The motivation for this analysis is that analyst recom-mendations could sometimes elicit market-wide responses (e.g., the Citigroupdowngrade was accompanied by a drop in the S&P 500) and hence a methodof identifying influential recommendations that nets out market returns mayunderstate the influence of analysts. Using raw returns to define influential

621


eliminates this problem. We show that the raw return-based measure of in-fluential obtains a 10.8% fraction of influential observations, not higher butlower than our original 11.7%. Therefore, we are confident that our baselineapproach to identify influential recommendations based on abnormal returns isnot biased.

Fourth, one might argue that our approach is too conservative because thesignificance of the stock-price impact of recommendation changes does notdepend on a benchmark, which itself might include influential recommen-dation changes. To evaluate this issue, we repeat our analysis by removingnon-corporate event recommendation days from the computation of priorId-iosyncratic volatilityof stock returns. This slightly increases the fraction ofrecommendation changes that are influential. The fraction is now 12.1%. Thiscould be due to the fact that only a small fraction of recommendations areinfluential. Therefore, the typical estimate ofIdiosyncratic volatilityusing allobservations is not an overly high hurdle under our main influential definition.

Fifth, one potential concern is that influential recommendations have a largereaction because they piggyback on some firm-related event not captured byour screens. Because it is difficult to mount an exhaustive news wire search forcorporate events in our large sample, we proxy for a contaminating event usingthe pre-recommendation stock returns. We impose this additional screen of re-moving recommendations if the absolute value of the day [−2,−1] pre-eventreturn is more than 1.96×

√2 standard deviations of the firm’s priorIdiosyn-

cratic volatility of returns. About 13% of observations are removed using thisscreen. However, none of our results are changed with this new sample. Withthis reduced sample, 11.1% of the recommendation changes are still classifiedas influential in returns.12

In sum, our results are robust to alternative definitions of recommendationchanges, influential definitions, and different samples.13

12 An additional concern is that there could be more important company-specific news events on or immediatelybefore days that influential recommendations are made than on other days, so our identification of analyst rec-ommendation changes as influential might be spurious. To investigate whether news events are more likely ondays with influential recommendations, we select 100 influential and 100 non-influential observations randomlyand use Factiva to search for corporate news in Dow Jones and Reuters newswires from day−2 to the recom-mendation date. We limit our search to corporate news relating to mergers and acquisitions, lawsuits, securityissuance, dividend changes, debt rating news, and earnings guidance or announcements. We find 19 observa-tions with such events in the influential sample and 25 in the non-influential sample. Hence, there is no evidencethat the influential subsample contains more corporate news-motivated recommendation changes. Of the 44 ob-servations, most do not elicit any large stock-price reaction. Our pre-event screen is hence especially useful forremoving large reaction observations. The pre-event price screen removes 10 observations. The average absolutevalue of pre-event returns for these 10 is 17.9%. For the remaining 34 observations, the average absolute value ofpre-event returns is only 1.9%. This provides evidence that most of the corporate events our screens miss do notexert a large impact on the stock price. We believe that using returns to identify pre-events is a better approachthan searching for news articles because news articles may not always elicit discernible stock-price reactions.

13 One other test we do relates to addressing the concern that I/B/E/S provides only an incomplete record of theuniverse of stock recommendations (e.g.,Ljungqvist, Malloy, and Marston 2009). We include broker-matchedrecommendations from First Call not found on I/B/E/S. We let these observations inherit an I/B/E/S analystidentifier if the closest prior and future (two-year window centered on the First Call observation) I/B/E/Srecommendations have the same analyst identifier. With this combined sample, the percentage of influentialrecommendation changes remains similar, at 11.5%. Probit estimations also yield similar results.

622


4.2 Comparing influential fraction in earnings forecasts versus reco-mmendation samples

Our article focuses on stock recommendations. However, analyst reports alsocontain another important output—revisions of earnings forecasts. Our prioris that stock recommendations are more likely to be influential. A recommen-dation is an explicit statement on whether an investor should buy or sell thestock, while an earnings forecast revision is not. Recommendations can re-ceive a great deal of attention in the press, as evidenced by the examples atthe start of this article; the same attention is rarely given to earnings forecastchanges. Finally, a recommendation can encompass the joint impact of cashflow and discount rate news on the stock price, while standalone earnings fore-casts contain only cash flow news. However, we now investigate whether theevidence is consistent with our prior.

We consider three forecast horizons: one-quarter-ahead, one-year-ahead, andlong-term-growth (LTG) forecasts from I/B/E/S. We conduct a similar exercisein hand-matching broker names between I/B/E/S and First Call (FC) and uti-lize First Call dates whenever possible. For each forecast horizon, we classifyearnings forecasts into upward and downward revisions based on three meth-ods:revision own, revision last, andrevision con, in a manner akin to our threerecommendation-change measures. We then remove firm news-contaminatedobservations using the sample 2 to 4 screens as in Table2. The replicationof Table2 using earnings forecasts shows that revisions elicit statistically sig-nificant stock-price reactions, although their reaction magnitudes are smallerthan those of recommendation changes (the online appendix contains theseresults).

We then report the influential fraction in Table7. Interestingly, the fractionhovers around 5%. This is close to the percentage of influential observationsthat we would have expected by chance alone. However, whenever the earningsforecast revision is accompanied by a stock recommendation in the three-daywindow around the revision, the influential fraction goes up to about 10%,consistent with the results from our recommendation change sample. We alsoshow evidence from the flip side—if the current recommendation change isaccompanied by an earnings forecast in the three-day window, the influentialfraction increases from 11.7% to 13.2%. This is in line with the findings ofKecskes, Michaely, and Womack(2009).

5. Conclusion

Recommendation changes are sometimes associated with extremely large ab-normal returns, and these changes are typically the ones that the press focuseson. Such changes are associated with stock-price reactions that are quite differ-ent from the stock-price reaction of the typical recommendation change. Theexisting literature on analyst recommendation changes focuses on the averagestock-price reaction. We show that when proper care is taken to account for

623


Table 7Influential fraction of earnings forecasts with and without stock recommendations

Panel A: Forecast revisionssample

All forecast revisions Revisions withrecommendations

Not Influ Influential Percent Not Influ Influential PercentForecast Influential basedrevision sample onabnormal:

Annual Returns 286,813 13,402 4.5% 18,094 2,023 10.1%Turnover 283,672 16,543 5.5% 17,676 2,441 12.1%

Quarterly Returns 105,570 5,346 4.8% 6,268 799 11.3%Turnover 104,125 6,791 6.1% 6,081 986 14.0%

LTG forecasts Returns 42,258 1,750 4.0% 3,119 310 9.0%Turnover 41,055 2,953 6.7% 2,930 499 14.6%

Panel B: Recommendation changessample

All recommendation changes Rec changes with earningsforecasts

Not Influ Influential Percent Not Influ Influential PercentInfluential based

Sample onabnormal:

Rec Change Returns 77,537 10,292 11.7% 34,608 5,253 13.2%Turnover 76,595 11,234 12.8% 34,331 5,530 13.9%

We examine what fraction of earnings forecasts revisions are influential (Influ) according to our definition. Earn-ings forecasts revisions based on yearly (FY1), quarterly (Q1), and long-term-growth (LTG) horizons are fromthe I/B/E/S Detail file from 1994 to 2006 where a revision is coded as an upward or downward revision basedthe current forecast minus the prior outstanding forecast by the same analyst. Dates from First Call are used asrevision dates whenever available through hand-matching of broker name. The samples are then screened forcontaminating events based on the sequence shown in Table2. The final sample considered here is based onforecast revisions without contemporaneous earnings announcements, company guidance, and multiple same-horizon forecasts days. A revision is considered influential if its two-day event [0,1] return reaction is in thecorrect direction and 1.96 standard deviations more than expected based on the prior three-month dailyId-iosyncratic volatilityof stock returns. Abnormal turnover is also used as a benchmark for being influential. Anearnings forecast is treated as accompanied by a recommendation when there is a recommendation issued bythe same analyst in the three-day window around the earnings forecast revision date. Panel B, we list influentialfractions based on the original recommendation-change sample in Table4.

confounding news, the typical stock-price reaction is small enough that foran individual stock it would not be identifiable with a firm-level event study.When analyst recommendation changes have an identifiable impact at the firmlevel, we call them influential recommendation changes.

We show that some analyst recommendation changes lead to substantialchanges in how a firm is assessed and valued by investors, leading to largereturns and turnover relative to the history of the firm. We investigate the fre-quency of such recommendation changes, when a recommendation change islikely to be influential, and how the firm’s information environment changesaround influential recommendation changes. Using our criterion for influentialrecommendation changes ensures that the observations identified as influentialcan actually be noticed by investors following the firm. We find that about 12%of the recommendation changes are influential after eliminating recommenda-tion changes associated with confounding firm news. Strikingly, a quarter ofthe analysts in our sample have no influential recommendation change in theirrecommendation histories. We find that influential recommendations are morelikely to be from analysts with larger leader-follower ratios and more accurateearnings forecasts. Recommendations that go away from the consensus andare issued contemporaneously with earnings forecasts are also more likely tobe influential. Having had an influential recommendation change before also

624


improves the likelihood of having an impactful recommendation change, giv-ing credit to the view that analyst skill is an important determinant of impactfulresearch. Further, growth firms, small firms, high institutional ownership firms,and low analyst activity firms are more likely to be associated with influentialrecommendations.

Why is it that an analyst at times can make recommendations that are asso-ciated with a significant firm-level abnormal return or turnover? We conjecturethat at times analysts can change how a corporation is viewed and that such“paradigm shifts” are responsible for the large impact of some of the recom-mendation changes. This perspective is related toHong, Stein, and Yu(2007),who study the implications of learning in an environment in which the truemodel of the world is a multivariate one, but agents update only over the classof simple univariate models. When sufficient evidence accumulates against theincumbent simple model, agents switch to another simple model, and pricesin the underlying stock move to reflect this paradigm shift. While our analy-sis does not amount to a test of their model, our evidence is consistent withan influential analyst recommendation change being able to precipitate such aparadigm shift. It is not surprising, therefore, that firms experiencing influentialrecommendation changes see their stock turnover increase, their volatility in-crease, and analysts make more and bigger earnings forecast changes. Further,industry returns are also impacted around the recommendation event.

Our evidence shows that focusing on the average stock-price reaction tochanges in analyst recommendations leads to an incomplete assessment ofthe value of the information produced by analysts. The size of the averagestock-price reaction to recommendation changes is small enough that investorswould not notice how a recommendation affects a stock, if it does at all. Thisis because many analysts do not make influential recommendations and not allrecommendations are influential. However, our evidence suggests that someanalysts do make recommendation changes that change how a firm is assessedby investors.

Supplementary DataSupplementary data are available online athttp://www.sfsrfs.org/addenda/.

ReferencesAltinkilic, O., and R. S. Hansen. 2009. On the Information Role of Stock Recommendation Revisions.Journalof Accounting and Economics48:17–36.

Asquith, P., M. B. Mikhail, and A. Au. 2005. Information Content of Equity Analyst Reports.Journal ofFinancial Economics75:245–82.

Barber, B. M., R. Lehavy, M. McNichols, and B. Trueman. 2006. Buys, Holds, and Sells: The Distributionof Investment Banks’ Stock Ratings and the Implications for the Profitability of Analysts’ Recommendations.Journal of Accounting and Economics41:87–117.

Boni, L. 2006. Analyzing the Analysts after the Global Settlement. In Y. Fuchita and R. E. Litan (eds.),Finan-cial Gatekeepers: Can They Protect Investors?Brookings Institution Press and the Nomura Institute of CapitalMarkets Research.

625

http://www.sfsrfs.org/addenda/


Boni, L., and K. L. Womack. 2006. Analysts, Industries, and Price Momentum.Journal of Financial & Quanti-tative Analysis41:85–109.

Bradley, D. J., B. D. Jordan, and J. R. Ritter. 2008. Analyst Behavior Following IPOs: The “Bubble Period”Evidence.Review of Financial Studies21:101–33.

Chen, Q., J. Francis, and K. Schipper. 2005. The Applicability of the Fraud on the Market Presumption toAnalysts’ Forecasts. Working Paper, Duke University.

Clement, M. B., and S. Y. Tse. 2003. Do Investors Respond to Analysts’ Forecast Revisions as If ForecastAccuracy Is All That Matters? Accounting Review 78:227–49.

Cooper, R. A., T. E. Day, and C. M. Lewis. 2001. Following the Leader: A Study of Individual Analysts’ EarningsForecasts.Journal of Financial Economics61:383–416.

Daniel, K., M. Grinblatt, S. Titman, and R. Wermers. 1997. Measuring Mutual Fund Performance withCharacteristic-based Benchmarks.Journal of Finance52:1035–58.

Fama, E. F., and K. R. French. 1997. Industry Costs of Equity.Journal of Financial Economics43:153–93.

. 2006. The Value Premium and the CAPM.Journal of Finance61:2163–85.

Fang, L., and A. Yasuda. 2008. Are Stars’ Opinions Worth More? The Relation between Analyst Reputation andRecommendation Values. Working Paper, University of Pennsylvania.

Frankel, R., S. P. Kothari, and J. Weber. 2006. Determinants of the Informativeness of Analyst Research.Journalof Accounting and Economics41:29–54.

Gintschel, A., and S. Markov. 2004. The Effectiveness of Regulation FD.Journal of Accounting and Economics37:293–314.

Gleason, C. A., and C. M. C. Lee. 2003. Analyst Forecast Revisions and Market Price Discovery.AccountingReview78:193–225.

Hong, H., J. C. Stein, and J. Yu. 2007. Simple Forecasts and Paradigm Shifts.Journal of Finance62:1207–42.

Irvine, P. J. 2004. Analysts’ Forecasts and Brokerage-firm Trading.Accounting Review79:125–49.

Ivkovic, Z., and N. Jegadeesh. 2004. The Timing and Value of Forecast and Recommendation Revisions.Journalof Financial Economics73:433–63.

Jegadeesh, N., and W. Kim. 2010. Do Analysts Herd? An Analysis of Recommendations and Market Reactions.Review of Financial Studies23:901–37.

Kadan, O., L. Madureira, R. Wang, and T. Zach. 2009. Conflicts of Interest and Stock Recommendations: TheEffects of the Global Settlement and Related Regulations.Review of Financial Studies22:4189–217.

Kecskes, A., R. Michaely, and K. L. Womack. 2009. What Drives the Value of Analysts’ Recommendations:Earnings Estimates or Discount Rate Estimates? Working Paper, Darthmouth College.

Knez, P. J., and M. J. Ready. 1997. On the Robustness of Size and Book-to-market in Cross-sectional Regres-sions.Journal of Finance52:1355–82.

Ljungqvist, A., C. J. Malloy, and F. C. Marston. 2009. Rewriting History.Journal of Finance64:1935–60.

Ljungqvist, A., F. Marston, and W. J. Wilhelm. 2009. Scaling the Hierarchy: How and Why Investment BanksCompete for Syndicate Co-management Appointments.Review of Financial Studies22:3977–4007.

Llorente, G., R. Michaely, G. Saar, and J. Wang. 2002. Dynamic Volume-return Relation of Individual Stocks.Review of Financial Studies15:1005–47.

Loh, R. K. 2010. Investor Inattention and the Underreaction to Stock Recommendations.Financial Management39:1223–51.

Loh, R. K., and G. M. Mian. 2006. Do Accurate Earnings Forecasts Facilitate Superior Investment Recommen-dations?Journal of Financial Economics80:455–83.

626


Malmendier, U., and D. Shanthikumar. 2007. Are Investors Naıve About Incentives?Journal of FinancialEconomics85:457–89.

Mikhail, M. B., B. R. Walther, and R. H. Willis. 1997. Do Security Analysts Improve Their Performance withExperience?Journal of Accounting Research35:131–57.

Stickel, S. E. 1995. The Anatomy of Buy and Sell Recommendations.Financial Analysts Journal51:25–39.

Thompson, S. 2010. Simple Formulas for Standard Errors That Cluster by Both Firm and Time.Journal ofFinancial Economics, forthcoming.

Veldkamp, L. L. 2006. Information Markets and the Comovement of Asset Prices.Review of Economic Studies73:823–45.

627

when are analyst recommendation changes inﬂuential? · 2017. 12. 22. · when are analyst...

Documents