Determinants of the informativeness of analyst … Kothari and Weber...informative. However, if market participants are assumed to ﬁxate on or under- or over-react to analyst reports,

ARTICLE IN PRESS

Journal of Accounting and Economics 41 (2006) 29–54

0165-4101/$ -

doi:10.1016/j

$We are

Zhayoang G

Rochester, a�CorrespoE-mail ad

www.elsevier.com/locate/jae

Determinants of the informativeness ofanalyst research$

Richard Frankela, S.P. Kotharib,�, Joseph Weberb

aOlin School of Business, Washington University in St. Louis, St. Louis, MO 63130-6431, USAbSloan School of Management, Massachusetts Institute of Technology, Cambridge, MA 02142, USA

Abstract

We examine cross-sectional determinants of the informativeness of analyst research, i.e., their

effect on security prices, controlling for endogeneity among the factors affecting informativeness.

Analyst reports are more informative when the potential brokerage profits are higher (e.g., high

trading volume, high volatility, and high institutional ownership) and lower when information

processing costs (e.g., more business segments) are high. We also find that the informativeness of

analyst research and informativeness of financial statements are complements.

r 2006 Elsevier B.V. All rights reserved.

JEL classification: G12; G14; M4

Keywords: Analyst’s forecasts; Capital markets; Informativeness

1. Introduction

We study the determinants of the magnitude (i.e., absolute value) of stock price reactionto analyst reports. We define the informativeness of an analyst report (referred to in thepaper as AI, analyst informativeness) as the average magnitude of a firm’s stock price

see front matter r 2006 Elsevier B.V. All rights reserved.

.jacceco.2005.10.004

grateful to Mike Barclay, Jennifer Francis (the referee), Dave Larcker, Thomas Lys (the editor),

u, Wayne Guay, Ane Tamayo, Jerry Zimmerman, and workshop participants at Pittsburgh,

nd Wharton, for helpful comments.

nding author. Tel.: +1617 253 0994; fax: +1 617 253 0603.

dress: [email protected] (S.P. Kothari).

www.elsevier.com/locate/jae

ARTICLE IN PRESSR. Frankel et al. / Journal of Accounting and Economics 41 (2006) 29–5430

reaction to analysts’ reports in a given year. Our measure of informativeness is related tothe variance measure of informativeness of earnings reports pioneered by Beaver (1968)and used in subsequent research (see Landsman and Maydew (2002), and Francis et al.,(2002) for recent applications).Much of previous research studies the informativeness of consensus forecast, whereas

our primary focus is on examining the average price impact of individual analyst reportsand determinants of the informativeness of a report.1 Such research on the informativenessof analyst reports is important for several reasons. First, recently voiced concern about theintegrity and objectivity of analysts in the academic literature and in the financial pressraises skepticism about the informativeness of analyst research. Our analysis of theinformativeness of analyst reports is helpful in ascertaining whether and how much price-sensitive information exists in a typical analyst report.2 Second, analysts as informationproviders in capital markets decide on which firms to follow and how many reports to issueon each firm. By investigating the determinants of analyst informativeness (AI) we examinewhether the incentive to provide informative reports influences analysts’ decisions tofollow a firm. Finally, previous research suggests that there is also demand for analysts tosimply repackage and re-transmit and also to interpret corporate disclosures to generateinvestment banking and brokerage business, potentially making analyst reports lessinformative. The net effect of the aforementioned three reasons on the averageinformativeness of analyst reports is an empirical issue. Moreover, cross-sectionalvariation in the informativeness of the reports is determined by multiple factors, withsome of those being endogenous, the focus of our study.Our examination of the determinants of cross-sectional variation in AI is similar to

examining why some firms’ returns are more volatile than others. Clearly, event-specificfactors cause price movements and therefore return volatility, but in explaining differencesin volatility one seeks to understand why news comes in larger doses for some firms and insmaller doses for others. Our study provides evidence on why for some firms the averagemagnitude of news in analyst reports is greater than that for others.Previous research has typically examined the relation between analyst following and firm

characteristics under the assumption that analyst-following proxies for the resourcesdevoted to information collection and thus the informativeness of analyst reports (seeBhushan, 1989a, O’Brien and Bhushan, 1990; Lang and Lundholm, 1996; Hong et al.,2000). Unlike prior research, we do not assume that the richness of the information setattributable to analysts is automatically an increasing function of the number of analystsfollowing for a firm. We examine analyst following as one factor among many that canpositively or negatively influence the informativeness of analyst reports. In addition, our

1Lys and Sohn (1990) is an exception, but they also do not directly focus on estimating analyst informativeness.2Note that in an efficient market, analyst reports that lack integrity and objectivity are likely to be less

informative. However, if market participants are assumed to fixate on or under- or over-react to analyst reports,

then analyst research would appear informative (i.e., generate a stock-price reaction), but it would also cause

returns to be predictable—contrary to market efficiency. In recent years, behavioral finance theory has gained

acceptance among academics and practitioners, and thus hypothesis assuming market inefficiency should also be

considered. Return-predictability tests that are designed to discriminate between analyst informativeness in

efficient markets and under/overreaction are beyond the scope of this study. Preliminary analysis conducted in an

earlier version of the paper suggest lack of return predictability, which is consistent with analyst informativeness

as if market participants filter out irrelevant and misleading information in analyst report.

ARTICLE IN PRESSR. Frankel et al. / Journal of Accounting and Economics 41 (2006) 29–54 31

analysis of the endogenous relation between AI and various firm characteristics extendsprevious research by O’Brien and Bhushan (1990), Alford and Berger (1999), and others.

1.1. Summary of results

We analyze analyst forecasts, stock returns, and firm characteristics for almost 24,000firm-year observations from 1995 to 2002. We measure informativeness as the absoluteabnormal stock price reaction on the dates that analysts release forecast revisions for afirm. AI is calculated for each firm in each year by averaging the absolute stock pricereaction to all the analyst forecast revisions for the firm in a given year.3 We find analystresearch, on average, is significantly informative and the informativeness exhibitssubstantial cross-sectional variation in our sample.

A two-stage-least-squares regression analysis reveals that, as predicted, AI increases inreturn volatility and trading volume. This result suggests that analysts provide moreinformation when profit opportunities for informed traders increase. Under thesecircumstances, we expect investors to seek more information from analysts. Our result isthus consistent with analysts responding to increased investor demand for private information.

We also find that AI decreases when there is a high correlation between a firm’s returnand the market’s return, and that AI decreases when a firm has relatively more lines ofbusiness. The finding of lower informativeness for firms with multiple segments suggeststhat investors benefit less from analysts’ research because analysts specialize in individualindustries and reports for multi-segment firms are likely to be less accurate.

The empirical analysis also shows that the marginal effect of analyst following on theinformativeness of analysts’ research reports is not statistically distinguishable from zero.It is reasonable to expect that competition among analysts would result in a negativemarginal effect of analyst following on informativeness. However, our results suggest that,while the supply of analysts is likely to rise with opportunities to provide informativeresearch, it does not go beyond the point when the marginal effect of analyst following oninformativeness becomes negative.

We also find that AI and the timeliness of financial information (measured as thecontemporaneous association between security prices and financial information) arecomplements. Thus, instead of preempting the information content of analyst reports,more timely financial information is associated with more informative reports. Our findingis at odds with the predictions of models where investors’ reaction to the analyst reportdecreases in the quality of other information (e.g., accounting information) available to themarket participants (see Holthausen and Verrecchia (1988); Demski and Feltham (1994),and Subramanyam (1996) for examples). Our findings, however, corroborate Francis etal.’s (2002, p. 317) conclusion that their results ‘‘do not in general support the predictedsubstitution relation between earnings announcements and competing informationsourcesy .’’

3An alternative measure of the analyst report informativeness used in the literature (e.g., Lys and Sohn, 1990;

Mikhail et al., 1997) is the correlation between the stock price reaction to the report and the surprise in an analyst

forecast (i.e., analyst forecast minus the market expected forecast at the time). This measure is appealing because it

incorporates the magnitude of the earnings surprise. However, the correlation between returns and the surprise in

analysts’ forecasts also requires the researcher to develop a measure of the market’s earnings expectation. In our

research setting, measurement error in determining the market’s expectation of earnings is likely to be correlated

with some of our determinants of analyst informativeness, making this measure less appealing.


Finally, we find analyst research is more informative when analysts issue a negativeforecast revision than a positive revision. This result implies the market has moreforeknowledge of the content of analysts’ positive forecast revisions than negative forecastrevisions or that investors are more suspicious of positive analyst reports. The resultsupports Hong et al. (2000) hypothesis that bad news propagates more slowly than goodnews.Summarizing, we find that, on average, analysts’ reports are informative. We also find

that analysts are more informative when potential brokerage profits are higher and lessinformative when information processing costs are higher. These results are consistent withand complement those reported in Lys and Sohn (1990). Using an alternative measure ofinformativeness, discussed below, Lys and Sohn also find that analysts’ reports are, onaverage, informative, even when they are preceded by other analyst forecasts or corporatedisclosures. Our paper provides evidence that the demand for and cost of supplyinginformation affect the extent to which analyst reports convey information.

1.2. Outline

Section 2 surveys previous research on analysts’ incentives to provide informativereports. Section 3 develops hypotheses as to the factors likely to affect the informativenessof an analyst’s report. Section 4 describes the data, regression model, and results fromestimating the relation between AI and its determinants. We offer concluding remarks insection 5.

2. Background and hypothesis development

In this section we discuss competing hypotheses about analysts’ role in capital markets,the effect of analysts’ incentives on the contents of their reports, and the likely impact ofanalyst research on security prices. This discussion provides a foundation for theeconometric analysis of the informativeness of analysts’ reports presented in Section 4.

2.1. Analysts’ role as information providers

Analysts are prominent information intermediaries in capital markets. They engage inprivate information search, perform prospective analysis aimed at forecasting a firm’sfuture earnings and cash flows, and conduct retrospective analysis that interprets pastevents (Beaver, 1998, p. 10). Regulators and other market participants view analysts’activities and the competition among analysts as enhancing the informational efficiency ofsecurity prices.4 The importance of analysts’ role in capital markets has spurred researchshowing that analysts influence the informational efficiency of capital markets. Specifically,the speed with which prices reflect public information increases with analyst following(e.g., Hong et al., 2000; Elgers et al., 2001). Prior research (e.g., Givoly and Lakonishok,

4Justice William O. Douglas is an early believer of analysts’ activities making markets efficient. Douglas (1933)

observes, ‘‘Even though an investor has neither the time, money, or intelligence to assimilate the mass of

information in the registration statement, there will be those who can and who will do so, whenever there is a

broad market. The judgment of those experts will be reflected in the market price.’’


1979, Lys and Sohn, 1990, Francis and Soffer, 1997) also shows analyst reports, onaverage, convey information to the capital market.

2.2. Factors reducing the informativeness of analyst research

Analyst research may not provide value relevant information to the market.5 First, withthe objective of generating investment banking and brokerage business for their firms,analysts provide information and other services to market participants. One of theseservices is to simply repackage and re-transmit corporate disclosures, i.e., providefundamental analysis to individual investors and money managers, which serve as an inputinto investment decisions. Thus, some analysts might follow companies or produceadditional research reports even when they are unable to provide information beyond thatimpounded in prices at the time of their reports. Lang and Lundholm (1996) label this asanalysts’ information intermediary role. While such reports are useful to institutions and/or individuals in their investment decisions, and thus help analysts generate commissionsby routing institutional and individual investors’ trades through the brokerage housesemploying the analysts (see Irvine, 2000; Lin and McNichols, 1998; Cowen et al., 2005),they will not necessarily enrich the information set investors use in setting prices.

Second, other analysts and timely (voluntary or mandated) corporate disclosures arelikely to serve as partial substitutes for the informativeness of an individual analyst’sresearch by preempting their analysis (e.g., Bhushan, 1989b; Beaver, 1998; Francis et al.,2002; Frankel and Li, 2004). This seems inconsistent with the evidence that corporatedisclosure quality and the amount of analyst research are complements (see Lang andLundholm, 1996). However, whereas analyst activity is likely to be related to the(perceived) quality and quantity of corporate disclosures, it does not necessarily translateinto greater informativeness of analyst research.

Finally, academic research suggests that analysts have incentive to compromise theirobjectivity and optimistically bias their forecasts, stock recommendations, and analysis(e.g., Lin and McNichols, 1998; Michaely and Womack, 1999; Dechow et al., 2000;Bradshaw et al., 2003; Lin et al., 2003). Analyst incentives to misinform, combined withmounting evidence of market inefficiency with respect to analyst reports (i.e., market’sfixation or under- or over-reaction to analyst reports), implies analyst research cannot beunambiguously interpreted as serving to enhance informational efficiency of the capitalmarkets. A less harmful outcome compared to the market’s fixation on uninformative ormisleading analyst reports is that market participants would ignore analysts’ reportsbecause they are perceived to be tainted by analysts’ conflict of interest. Consistent with thenotion that market participants ignore analyst reports, Barth and Hutton (2004, p. 92)conclude that while ‘‘yanalyst forecast revision signals identify firms with low persistenceaccruals and, thus, earnings, but investors do not incorporate this information into shareprices.’’

In sum, multiple forces influence the informativeness of analyst research. Researchsuggests that some of these forces are offsetting and has not identified their net effect

5Other researchers also doubt whether quantity of disclosure adequately proxies for the amount of information

to the market. For example, Barth (2002) wonders ‘‘y whether more disclosure results in more information to the

market’’ in the context of the informativeness of voluntary management forecasts. The concern is that quantity is

likely to affect quality and/or result in substitution. Also see Schipper (1991) and Beaver (1998).


on AI. Therefore, we study the extent to which analyst research conveys information tocapital markets. In particular, we examine the average magnitude of a firm’s stock pricereaction to analysts’ reports to provide evidence on whether, on average, analyst’s reportsare informative. We then investigate the cross-sectional determinants of AI. In the nextsection we motivate our selection of cross-sectional determinants by relating them to thesupply and demand for analyst services.

3. Determinants of AI

Following Bhushan (1989b), who studies the determinants of analyst following, wemodel the informativeness of analyst reports as a function of factors proxying for thedemand for and supply of analyst services. Naturally, we expect informativeness toincrease in firm and institutional characteristics that proxy for the demand for analysts’services and the informativeness to wane in the analysts’ cost of supplying new informationabout a firm. However, AI, the demand for informative reports, and the costs of supplyinginformative reports are all likely to affect each other. Below we develop empiricalpredictions for the various factors we expect to affect AI and in Section 4 we present theempirical model we use to address the resulting endogeneity.6

3.1. Return variance

High return volatility is indicative of high uncertainty about a security’s cash flow andthus presents traders with an opportunity to gain from information acquisition, whichwould mitigate the uncertainty. Return volatility can also result from informationasymmetry between management and outside investors. However, voluntary disclosure bymanagement as well as analysts’ private information gathering activities can reduceinformation asymmetry. Therefore, return volatility spurs demand for informative analystresearch and should be positively correlated with AI.

3.2. Trading volume

AI and trading volume should be positively related for at least three reasons. First,brokerage houses offer analyst research as a service for their clients to generate increasedtrading business. By increasing their trading business, brokerage houses can increase theircommissions. A brokerage firm’s success in generating trading volume is likely to bepositively influenced by the informativeness of their analysts’ research.Second, (uninformed) liquidity traders contribute to the overall trading volume in a

security. Holding the supply of shares constant, more liquidity trading makes the priceprocess noisier, i.e., increase volatility (see Verrecchia, 1982; Bhushan, 1989a). This noisecreates a profit opportunity for informed traders, which increases (i) trading volume, and

6We acknowledge an additional determinant of informativeness, namely ownership concentration and insiders’

holdings. External investor demand for corporate financial information and analysis is low in firms with high

levels of owner–manager holding or insider holdings. In such firms, insiders have access to the information, and

with limited outside ownership, outsiders’ demand for financial information and analysis is low. We are unable to

examine the role of insider ownership on analyst informativeness for lack of machine-readable data. However,

high insider ownership is not the norm for our sample of relatively large publicly traded firms.


(ii) the demand for informative analyst research that traders use to become informed andcapitalize on the profit opportunity.

Finally, trading volume can arise because of heterogeneous beliefs prior to informa-tion arrival (e.g., Beaver, 1968; Karpoff, 1986; Kim and Verrecchia, 1991a, b) ordifferential interpretation of the information (e.g., Karpoff, 1986; Harris andRaviv, 1993, Kandel and Pearson, 1995; Kim and Verrecchia, 1997). Regardless ofits source and timing, divergence in investor beliefs increases the demand for ana-lyst research that can be used to resolve the disagreement and restore consensus. Thus,the demand for AI is positively related to trading volume via the divergence of beliefsamong investors.

3.3. Institutional ownership percentage

Analyst research is an important input in an investor’s decision to invest in a stock(e.g., Merton, 1987; Brennan and Hughes, 1991). For information purposes as well asfor fiduciary responsibility reasons, institutional investors seek analyst reports (seeO’Brien and Bhushan, 1990). Thus, institutional ownership in a stock would increasethe demand for informative analyst research. If the demand arose primarily to satisfymoney managers’ fiduciary responsibility to make prudent investment decisions, theninstitutional ownership would create a demand for analyst reports, but not necessarily forinformative research.

3.4. Number of shareholders

The relation between number of shareholders and AI is ambiguous. The number ofshareholders should increase the demand for analyst research directly as well as through itseffect on trading volume. The larger the number of shareholders, the greater theopportunities for brokerage firms to sell their research to investors and generatecommissions from the resulting trading volume. This implies a positive influence of thenumber of shareholders on AI because the brokerage business would be more successful astheir analysts’ research becomes more informative.

However, an offsetting negative effect of the number of shareholders on informativenessarises because increases in the demand for analyst’s research can lead to competitionamong analysts, reducing the informativeness of the average report. In addition, ceterisparibus, each shareholder’s ownership declines as the number of shareholders increases. Asa result, each shareholder’s incentive to gather information declines and the demand forfinancial intermediation decreases.7

7Another factor influencing the relation between number of shareholders and informativeness is that an

investor’s interest in a security is likely to be a function of analyst informativeness. While informative research

attracts some investors to a security, its net effect on shareholder interest in a security is ambiguous. Informed

analysts are likely to make uninformed investors reluctant to invest in the security for fear that they are at a

disadvantage with respect to analysts’ preferred clients. In fact, as we mention above, the informativeness of the

typical analyst diminishes as more analysts follow the firm because of competition among analysts. Thus,

investors are likely to find the security attractive for investment because they perceive little information

disadvantage. So, a (net) negative relation between the number of shareholders and the informativeness of analyst

reports is also plausible.


3.5. Growth opportunities

A firm’s growth and/or investment opportunities are reflected in the market-to-bookratio. The effect of growth opportunities on AI is unclear. Growth firms have moreunrecorded, intangible assets, whose valuation depends heavily on future profitability.Forecasting future prospects requires expertise and the collection of data beyond thefinancial statements, implying higher information supply costs for analysts. However,investors who lack expertise and access to data helpful in valuing growth firms will demandincreased analyst guidance for these firms. The equilibrium quantity of informationsupplied, i.e., AI, is uncertain given these countervailing influences on demand and supply.

3.6. Firm size

Previous research suggests the demand for analyst research and information producedabout a security increase in firm size (e.g., Bhushan, 1989b; Collins et al., 1987; Lang andLundholm, 1996). Bhushan (1989b) argues that investors in general and informed tradersin particular are likely to value information about a large firm more because large stocksare more liquid. Thus, holding the price impact constant, informed traders could executelarger trades in large firms. While firm size spurs the demand for analyst research, it alsoinfluences the supply of analyst research through the effect of firm size on the cost ofproducing analyst research. Because many analysts are likely to follow a large firm, thecost of producing an informative report is likely to be high. On the other hand, because ofthe large number of sources of price relevant information for a large firm, individualanalysts are in a position to produce informative research notwithstanding the large supplyof analysts following a large firm. This conclusion, however, must be tempered by the factthat a large firm’s cash flows are likely to be more diversified than that of a small firm, sothe marginal impact of a piece of cash flow information about a large firm on its stockprice is likely to be smaller. Thus, the net effect of firm size on the informativeness ofanalyst reports is ambiguous.

3.7. Number of firms in an industry

The number of firms in an industry is likely to reduce AI because information about onefirm also provides information about other firms in an industry. Evidence of an industryfactor in returns and earnings dates back at least to King (1966) and Brown and Ball(1967). Research also documents evidence of intra-industry information transfers (e.g.,Foster, 1981). Thus, informativeness of an analyst report preempts the informativeness ofresearch reports on other firms in the industry and thus raises the cost of informativeresearch. Therefore, we predict a negative impact of the number of firms in an industryon AI.

3.8. Correlation between the firm and market return

The higher the correlation between a firm’s stock return and the market, the larger themacroeconomic component of the firm’s return and smaller the impact of firm-specificinformation on the firm’s returns (Bhushan, 1989b). Thus, analysts’ informationacquisition costs would decrease in the correlation of a firm’s returns with the market


and analyst following would rise. The resulting increased competition among analysts andthe fact that macroeconomic information accounts for a relatively larger fraction of afirm’s return variability, combine to diminish the informativeness of analyst reports. Thelogic here is similar to the effect of the number of firms in an industry on AI.

3.9. Number of lines of business

Bhushan (1989b) and others argue that analysts’ information gathering costs increase ina firm’s number of lines of business. Increased costs reduce analyst following. The highcost of information gathering and lower analyst following will decrease the amount ofinformation in each analyst report and thus lower AIs. Alternatively, fewer analyst reportsreduce competition and lessen preemption of information, making each research reportmore informative. Therefore, the net effect on AI is ambiguous.

3.10. Price-earnings association

The price-earnings association (i.e., the strength of contemporaneous associationbetween security prices and accounting earnings as an indicator of the informativeness ofaccounting numbers) affects AI for at least two reasons.8 First, a high price-earningsassociation means a firm’s earnings are relatively timely in reflecting the impact ofeconomic events affecting a firm’s stock price (see, for example, Kothari and Zimmerman,1995). Timeliness makes analysts’ task of uncovering price-relevant information moredifficult and therefore lowers AI. However, the greater information gathering difficultyraises analysts’ costs and thus reduces the supply of analysts, which can positively influenceinformativeness. These offsetting effects suggest that the price-earnings association’s neteffect on AI is indeterminate.

3.11. Number of analysts

The relation between AI and the number of analysts following the firm is ambiguous.The number of analysts following a firm is viewed in the research as a proxy for theinformativeness of their research as reflected in the firm’s prices, i.e., aggregateinformativeness. Consistent with this characterization, Brennan and Subramanyam(1995) argue that the number of investment analysts researching a firm is a proxy forthe number of individuals producing information about the value of the firm and thatincreases in analyst following will reduce information asymmetry as proxied for by theadverse selection cost of trading in a stock.9

The relation between the number of analysts and the average informativeness of analystreports could be positive or negative. On one hand, greater opportunities to be informativeinduce analysts to follow a firm, so one should expect a positive association. On the otherhand, as discussed in the introduction, (i) in addition to supplying information to the

8Price-earnings association is also referred to as the return-earnings association, where the former is on the basis

of a price-level regression, while the latter implies a specification in changes. Economically both specifications are

equivalent, and choosing between the two is guided by econometric considerations (see Kothari and Zimmerman,

1995).9Also see Hong et al. (2000), Frankel and Li (2004), and Barth and Hutton (2004) who assume that the

reduction in information asymmetry is proportional to analyst following.


market, analysts also simply repackage available information in their research reports andthus perform an information intermediary role (see Lang and Lundholm, 1996); or (ii) theyhave incentives to misinform the market, presumably for private gains; or (iii) analysts arepartial substitutes for each other. For these reasons, informativeness of analyst reports canbe negatively related to the number of analysts following a firm.

3.12. Asymmetric informativeness of good and bad news in research reports

Competing arguments predict an asymmetric market reaction to good and bad news inanalysts’ reports. Hong et al., (2000) argue that management has stronger incentives tohighlight good news than bad news, and therefore absent analysts, bad news is expected topropagate through prices more slowly. Thus, analysts are likely to play a more significantrole in the dissemination of bad news. This argument predicts greater AI when analysts reviseearnings forecasts downward, i.e., disseminate bad news. On the other hand, contractingconsiderations and litigation create an asymmetric demand for accounting conservatism, i.e.,recognition of bad news is more timely than recognition of good news (see Skinner, 1994;Basu, 1997; Ball et al., (2000). Therefore, management is more likely to preempt bad newsfrom analyst reports and the price-effect of bad news analyst reports will be reduced. In sum,the relation between informativeness and the sign of news conveyed is ambiguous.

4. Regression model, data, and results

This section describes the econometric procedures and results. Section 4.1 describes ourmeasure of AI and Section 4.2 presents the econometric model of the determinants ofinformativeness. Section 4.3 presents data sources and sample. We discuss the descriptivestatistics and cross-correlations in the data in Section 4.4. Section 4.5 contains our mainresults and Section 4.6 presents the results of our sensitivity tests.

4.1. Measuring AI

We begin our analysis by developing a measure of informativeness, and providingevidence on whether analysts are, on average, informative. We define AI, as follows. We firstsum the absolute size-adjusted returns on all the forecast revision dates for a firm in a givencalendar year.10 We then divide it by the sum of absolute size-adjusted returns for all tradingdays for the firm for the calendar year. Finally, we divide this ratio by the number of forecastrevision dates in a given calendar year. Because the divisor is number of forecast revisiondates, AI is a measure of the average informativeness of an analyst report date.11 Specifically,

AI ¼

Pt¼1toNREVSjRt;s �RSIZEt;sjP

t¼1to250jRt;s �RSIZEt;sj�

1

NREVS,

10The event window for each analyst report date is a single day. Expanding the event window is not preferable

as more than 25% of the sample firms have more than 40 analyst report dates per year. For these firms moving to

a larger window will contaminate our informativeness variable with noise.11If multiple forecast revisions are issued on the same day for a firm, our informativeness measure treats the

collection of reports as one report. In Section 4.6 we investigate how our measure of informativeness is affected by

multiple reports being issued on the same day. We find that controlling for multiple reports issued on the same day

does not significantly impact our results.


where Rt,s is a firm’s stock return on day t and the firm belongs to the NYSE size decileportfolio s, t ¼ 1 to NREVS are analyst forecast revision dates for the firm in a given yearwith 250 trading days, NREVS is the number of unique days on which at least one analystforecast appears on the I/B/E/S detail tape, where the analyst forecast can be a revisionor a reiteration, and RSIZEt,s is the return on the NYSE decile portfolio s on day t.

If analysts supply no information then the absolute return on a forecast revision dateshould be equal on average to the absolute return on any other trading date. In a givencalendar year with 250 trading days, each trading day would yield, on average, 1/250th ofthe sum of the absolute returns and thus AI would equal 0.004.

Our AI measure is similar to the variance measure of information content first used inBeaver (1968). The difference between our measure and a variance measure is that we usethe average absolute stock price reaction, not the square of the reaction. Average absolutevalue, however, is highly positively correlated with the variance, especially when thedistribution is symmetric. In the sensitivity analysis section of the paper (Section 4.6), wepresent results when the AI measure is computed using squared returns, i.e., analogous tothe variance measure.

A second alternative to the AI measure developed in this paper is to measure theinformation content of an analyst report on the basis of the correlation between theunexpected component of an analyst’s forecast and the stock return for the day ofthe forecast (e.g., Lys and Sohn, 1990). The correlation measure of informativenesscaptures the magnitude of the news content, i.e., some analyst forecasts are a large surpriseto the market, generating a correspondingly large stock price reaction. Fortunately, thisproperty is not lost when using our AI measure. Large forecast surprises generate largeabsolute stock price reaction and thus the AI measure is positively affected by the contentof the analyst forecast. In Section 4.6 we present evidence consistent with this intuition.

4.2. Modeling the determinants of analyst informativeness

Following O’Brien and Bhushan (1990) and Alford and Berger (1999), we base our mainanalysis of cross-sectional the variation in AI, on a two-stage-least-squares estimation of asystem of simultaneous equations because some of the determinants of AI are likely to alsobe affected by whether analyst reports are informative. In the presence of endogeneity,ordinary least squares estimation of the determinants of AI yields biased and inconsistentcoefficient estimates.

Use of two-stage least squares requires exogenous determinants of AI. However, manyof our hypothesized determinants of AI described in Section 3 are not ‘‘purely exogenous,’’but the extent of the endogeneity, in our assessment, for some of the variables is less thanfor others. To identify the system, we assume that some of our variables are exogenous.Specifically, we model two of our hypothesized determinants of informativeness, returnvariance and trading volume, as endogenous and treat the rest of our hypothesizeddeterminants as exogenous (i.e., Institutional ownership, price-earnings association,number of analysts, etc.).

We believe return variance and trading volume are endogenous in that our measure ofinformativeness directly affects returns, and thus is highly likely to affect volume andvariance. We also recognize that institutional ownership and the informativeness offinancial reports (i.e., price-earnings association) are also likely to be endogenous, in thatinstitutional investors’ decision to invest depends in part on the informativeness of reports,


and the informativeness of a firm’s financial reports is affected by the informativeness ofanalyst reports. However, in both of these cases, we believe that institutional ownershipand the price-earnings association are primarily affected by other factors and relativelylittle by the informativeness of reports. For example, diversification, tracking a benchmarkindex like the S&P 500, and industry prospects influence an institutional investor’s decisionto invest in a security more so than the average informativeness of analyst reports for thatsecurity. Similarly, informative research will affect the timing of the news received by themarket, but will only have a small impact on the association between annual changes inprices and earnings (i.e., price-earnings association). Therefore, we treat institutionalownership and price-earnings association as exogenous.Finally, a reverse causality can also exist in the case of number of analysts and AI. If

analysts are concerned about the informativeness of their reports and the number ofanalysts following a firm affects the informativeness of each analyst report, then thenumber of analysts following a firm will adjust in response to AI. For example, if analystsact as substitutes for each other and the average informativeness of an analyst report fallsbelow the level acceptable to analysts currently following the firm, then in the newequilibrium, fewer analysts will follow the firm. From an econometric perspective, theimportance of the bias that results from treating number of analysts as exogenous dependson the source of this ‘‘fall’’ in AI. If its source is the other included independent variables,then this bias will be limited, because the coefficient on analyst following represents theeffect on analyst following on informativeness conditional on the other variables. If itssource is a change in analyst following, then the coefficient on analyst following will bebiased, but we would argue that this feedback effect is of second-order importance. Basedon these arguments, AI is modeled as the following system of equations:

AI ¼ aþ b1Fittedððs2RÞÞ þ b1FittedðLnVOLÞ þ b3INSTþ b4LnOwners

þ b5MBþ b6LnMVþ b7NSICþ b8MMRsqþ b9NSEGSþ b10AccRsq

þ b11LnAnalystþ b12GNEWSþ e1; ð1Þ

s2ðRÞ ¼ aþ b1INSTþ b2LnOwnersþ b3MBþ b4LnMVþ b5NSIC

þ b6MMRsqþ b7NSEGSþ b8AccRsqþ b9r_LnVOLþ e2, ð2Þ

LnVOL ¼ aþ b1INSTþ b2LnOwnersþ b3MBþ b4LnMVþ b5NSIC

þ b6MMRsqþ b7NSEGSþ b8AccRsqþ b9r_s2ðRÞe3; ð3Þ

where s2(R) is the daily return variance for firm i in year t computed from the CRSP dailyreturn file, LnVOL is the natural log of total trading volume for firm i in year t (Compustat#28), INST is the percentage of shares held by institutions for firm i in year t computed asinstitutional shareholding at year-end as reported on CDA Spectrum divided by sharesoutstanding at the end of year t (Compustat #25), LnOwners is the natural log of thenumber of thousands of shareholders of firm i in year t (Compustat #100), MB is themarket-to-book ratio for firm i in year t ([Compustat #24�Compustat #25]/Compustat#60), and we exclude firm-year observations when the book value of equity is negative,LnMV is the natural log of the market value of firm i at the end of year t (CRSP), NSIC isthe number of firms in firm i’s four digit industry classification in year t (CRSP four digitSIC code) divided by the total number of firms on CRSP in year t, MMRsq isthe correlation between the firm and market return calculated as the R2 from firm i’s


market-model regression in year t (CRSP), NSEGS is the number of industry segmentson the Compustat business information file in year t, AccRsq is a measure of the priceearnings association which is derived from the fitted residuals from a pooled cross-sectional regression of prices on the book values of shareholders’ equity and earnings.Each year the pooled cross-sectional regression is estimated using data from the prior5-year period, and the 5-year window rolls forward by 1 year each year. Each firm’sannual residual from the pooled regression is scaled by price and squared. Each year wethen calculate an average squared residual for each firm using the five time-seriesobservations for the firm. We subtract this average from the corresponding average forthe entire population of firms to create AccRsq, a relative measure of the firm’saccounting-based pricing errors. The larger the value of AccRsq, the better the ability ofnet income and book value of equity in explaining prices.12 GNEWS is an indicatorvariable equal to one if the number of positive revision dates exceeds the number ofnegative revision dates in firm-year t, zero otherwise. A revision date is considered to bepositive if more analysts revise up than down on that date, LnAnalyst is the log of thenumber of analysts following the firm in year t, r_ indicates that the variable is assigned anumber 0, 1, or 2 on the basis of ranking each year all firm-year observations for thevariable into three equal portfolios, and Fitted( � ) indicates predicted values produced byEqs (2) and (3).

4.2.1. Endogeneity and 2SLS

We expect return variance to influence AI and vice versa. However, most of thedifference in the variance between high volatility stocks (e.g., young, growing and hightechnology stocks) and low volatility stocks (e.g., low growth, mature, regulated industrystocks) is unlikely to be attributable to variation in AI. Instead, endogeneity likely pertainsto the variation within the high volatility stocks that is likely to be related to the variationin AI. Similar arguments apply to trading volume.

2SLS requires the identification of exogenous instruments. Ideal instruments correlatehighly with the level of the endogenous variables but not with the (relatively small)variation around those levels, which is likely due to the endogenous nature of therelationship among the variables (Greene, 2000, pp. 370–375).13 However, all three

12Our measure of AccRsq differs from the typical measure based on a firm-specific time-series regression of

prices on financial variables. Our preference for a pseudo-explanatory power measure of the price-earnings

relation derived from a pooled regression is for practical reasons. Use of a pooled regression instead of firm-

specific time-series regressions avoids burdensome data-availability requirements. Firm-specific time-series

regressions using historical data (i.e., time series ending before the start of the sample period), would require data

availability of at least 15–20 years, which would result in a loss of 70–80% of the sample.

Another alternative to using a price-level-based measure of accounting informativeness would be to use a

return-earnings regression to assess accounting informativeness. The economics underlying the levels and return

specifications is identical and therefore the choice ought to be guided only on econometric considerations (see

Kothari and Zimmerman, 1995). Since the residuals from a price-level regression are an estimate of the extent to

which prices contain forward-looking information that is missing from contemporaneous accounting numbers,

the explanatory power of the price-level regression is an estimate of the informativeness of contemporaneous

accounting numbers. Recent examples of previous research assessing accounting informativeness using price-level

regressions include Barth et al. (1996); Collins et al., (1997) and Francis and Schipper (1999). The latter study also

uses return-based models to assess the relevance of financial statement numbers.13We reiterate that because all variables are fundamentally related to the firm’s information environment, they

are all related to each other and thus endogenous. Indeed, given this fundamental relation, arguing that a given

variable can be used as an instrument to identify the system and therefore be included in some equations and


equations contain large subsets of the variables such that the system is not identified inspite of the large number of exogenous variables.To identify this system of equations, we follow Hentschel and Kothari (2001), and

construct an instrumental variable for return volatility as follows. We rank annually thesample firms according to return volatility and assign firms to three portfolios, with thelowest (highest) volatility portfolio rank of 0 (2).14 These portfolio ranks are used asinstruments to proxy for the level of volatility, but the instruments are not correlated withthe endogenous variability around those levels.15 That is, we assume that whether a stock isin the low, medium, or high return volatility portfolio is unlikely to be due to theendogenous relation between AI and return volatility. However, the variation in returnvolatility among the stocks within the portfolio of high volatility stocks is likely to beendogenously determined by the AI of the stocks in the high volatility portfolio of firms.The volatility rank therefore meets the requirements of a good instrument. The same logicapplies to trading volume. We use portfolio rank values instead of actual volatility ortrading volume in Eqs. (2) and (3) above. The fitted values from these two equations arethen substituted in Eq. (1) to explain cross-sectional variation in AI. Under theassumptions about the properties of the instruments described above, the use of the fittedvalues circumvents the endogeneity problems of estimation.

4.3. Data

The empirical analysis is based on data gathered from four sources: I/B/E/S, CRSP,Compustat, and CDA Spectrum. We begin with all individual analyst forecast revisions onthe I/B/E/S detailed tape from 1995 to 2002. The forecast revisions represent the release ofnew reports by individual analysts with revisions in quarterly, annual, and/or long-termgrowth forecasts. Because our goal is to assess the informativeness of analyst reports, wedo not restrict our analysis to examining any one type of forecast revision (e.g., only 1-year-ahead earnings forecasts). The sample period begins in 1995 because conversationswith I/B/E/S representatives indicate that the ‘‘Estimate date’’ in the I/B/E/S detail filedoes not accurately represent the release dates of analyst forecasts prior to this time. Forall firms on I/B/E/S, we compile individual analyst forecast revision (or reiteration) datesfor each company-year. Each calendar day with at least one analyst forecast revisionreported on the I/B/E/S tape is treated as a forecast revision date.Stock return data are from the CRSP tapes. We compute size-adjusted returns on each

forecast revision date by subtracting the size-matched-decile return from the firm’s rawreturn. We obtain trading volume, return variability, and firm size, i.e., market capital-ization, from the CRSP tapes. Financial data and the number of shareholders data comefrom Compustat. We obtain data on institutional holdings at calendar-year ends from

(footnote continued)

excluded from others can be problematic. As in any econometric formulation, we exercise judgment and hope to

circumvent the problem by treating all the variables except analyst informativeness, volatility, and trading volume

as exogenous.14We evaluate the sensitivity of our results using decile ranks (as opposed to 3 ranks). Even though we expect

decile ranking to make the instrument less effective, we still find that our results remain qualitatively invariant to

this research design choice.15See Lys and Sabino (1992) who also recommend the use of three portfolios to maximize efficiency of the tests

in the presence of measurement error.


CDA Spectrum. We lose approximately one-half of the firm-year observations availablefrom the I/B/E/S detail database, with corresponding CRSP data, due to missingCompustat segment data and missing CDA Spectrum data on institutional ownership. Wedelete firms with negative shareholders’ equity to permit meaningful market-to-bookratios. The final sample comprises 23,798 firm-year observations for 6121 firms.

4.4. Descriptive statistics and cross-correlations

Table 1 reports descriptive statistics for the pooled sample of 23,798 firm-year observationsfrom 1995 to 2002. AI averaged across all firms is 0.0046. As noted earlier, since we measureAI as a ratio of price reaction to analyst reports normalized by the price movement duringthe entire year. Under the null of zero incremental informativeness of an analyst report, AI,would be 0.0040 ( ¼ 1/250 trading days in a year). Both mean and median AI estimates inTable 1 exceed 0.0040, which indicates analyst reports are informative on average and isconsistent with the findings in the research on the properties of analyst forecasts.

Recall that AI is calculated as a ratio of sums of absolute returns. Because of Jensen’sinequality, the mean of the distribution of firm-specific ratios can be biased upward relativeto the ratio of means, which would lead to a spurious inference of AI. Therefore, weexamine whether the reported mean AI is biased. We construct a bootstrap distribution forAI using randomly selected event dates for the sample firms. For each firm year in oursample, the number of random event dates is set equal to the actual number analystrevision dates for that firm year. This process produces a sample of 23,798 AI observationswith random analyst forecast dates. The sample mean (median) is 0.0040 (0.0039). With250 trading days in a year, this is exactly the average value expected when event dates areno more informative than non-event dates.

We next examine whether our estimate of AI is biased upward because of the possibilitythat analyst report dates are clustered around earnings announcement dates. Since theearnings announcement period has long been shown to be one of greater return volatility(e.g., Beaver, 1968), our measure of AI using absolute returns around analyst report datesis potentially contaminated by earnings announcement period volatility and thusmechanically indicates information in analyst reports. We therefore recalculate AI deletingall analyst report dates within a 3-day window centered on quarterly and annual earningsannouncement dates. Approximately 8% of the forecast revision dates are lost, but themean average AI measure remains at 0.0044 (p-valueo0.01 that the mean equals 0.004)while the median declines to 0.0042 (p-valueo0.01 that the median equals 0.004).

Descriptive statistics for the remaining variables in Table 1 indicate that firms in oursample are large with several analysts following them. The average market-to-book ratio is3.16, which is due in part to the bull market of the 1990s.16 The number of firms in anindustry, NSIC, is deflated by the number of firms on CRSP in each year and therefore theaverage is only 0.0087. It translates into an average of 103 firms in each four-digit SIC codeindustry. NSIC refers to the number of all the firms in a four-digit SIC code industry onCRSP, not just the number of firms for an industry included in our sample.

More than half of the sample firms report financial information for more than onesegment. The 25th percentile for NSEGS is 1. There are about 755,000 unique firm-analystforecast revision dates in our sample over from 1995 to 2002. The last row in Table 1

16We winsorize the extreme 1% values of the market-to-book ratio variable. Without winsorization, a few

observations assume a market-to-book ratio in excess of 1,000.

ARTICLE IN PRESS

Table 1

Descriptive Statistics

Mean Std. dev. Q1 Median Q3 Min Max

AI 0.0046 0.0016 0.0038 0.0044 0.0051 0.0000 0.0346

s2(R) 0.0018 0.0022 0.0005 0.0011 0.0023 0.0000 0.0064

LnVOL 3.08 1.57 2.01 2.98 4.10 �4.5190 9.5380

INST 0.46 0.25 0.25 0.46 0.66 0.00 1.00

LnOwners 1.34 1.21 0.40 1.01 1.92 0.00 11.49

MB 3.16 3.61 1.27 2.03 3.52 0.32 24.00

LnMV 12.78 1.81 11.44 12.65 13.89 6.92 20.22

NSIC 0.0073 0.0107 0.0008 0.0023 0.0092 0.0001 0.0765

MMRsq 0.0882 0.1028 0.0148 0.0504 0.1242 0.0000 0.7750

NSEGS 2.00 1.67 1.00 1.00 3.00 1.00 23.00

AccRsq 2.249 1.593 �2.079 2.342 3.105 �4.431 4.478

LnAnalyst 1.67 0.90 1.10 1.61 2.30 0.00 3.93

GNEWS 0.42 0.49 0.00 0.00 1.00 0.00 1.00

LnNREVS 3.01 1.05 2.40 3.14 3.76 0.00 5.25

This table presents univariate statistics of the variables used in the subsequent tests. The sample is derived from all

analyst earnings forecast revisions on the I/B/E/S detail database between January 1, 1995 and December 31,

2002. Firm-years with I/B/E/S data are merged with CRSP, Compustat, and CDA Spectrum for data on stock

returns, financial information, and institutional ownership. Firm years with missing data are discarded. Market-

to-book ratio and AccRsq are winsorized because of extreme values. Values of institutional ownership (INST)

greater than 100 percent are coded as 100 percent. The statistics are based on a final sample of 23,798 firm-year

observations.

Variable definitions (for firm i for use in a regression in year t)

AI is the sum of absolute size-adjusted returns on analyst earnings forecast revision dates for a firm in

year t divided by the sum of absolute size-adjusted returns for the firm for all the trading days in

year t, all divided by the number of analyst revision dates for the firm in year t,

s2(R) is the daily return variance for year t computed from the CRSP daily return file,

LnVOL is the natural log of total trading volume in year t�1 (Compustat #28),

INST is the percentage of shares held by institutions for firm i in year t computed as institutional

shareholding at year end as reported on CDA Spectrum divided by shares outstanding at the end

of year t (Compustat 25),

LnOwners is the natural log of the number of shareholders of firm i in year t (Compustat #100),

MB is the market-to-book ratio for firm i in year t ([Compustat #24 � Compustat #25]/Compustat

#60),

LnMV is the natural log of the market value of firm i at the end of year t (CRSP),

NSIC is the number of firms in firm i’s four digit industry classification in year t (CRSP four digit SIC

code) divided by the total number of firms on CRSP in year t,

MMRsq is the R2 from firm i’s market-model regression in year t using CRSP daily returns,

NSEGS is the number of industry segments on the Compustat business information file in year t,

AccRsq is derived from the fitted residuals from a pooled cross-sectional regression of prices on the book

values of shareholders’ equity and earnings using data from the prior 5 years. Each year this

window is roll forward by 1 year. Each firm’s five annual residuals from the pooled regression are

scaled by lagged price and squared. We then each year calculate an average squared residual for

each firm using the five time-series observations for the firm. We subtract this average from the

average for the entire population of firms in a pooled regression to create AccRsq, a relative

measure of the firm’s accounting-based pricing errors,

LnAnalyst is the log of the maximum number of analysts issuing 1-year-ahead earnings forecasts in firm-year

t,

GNEWS is an indicator variable equal to one if the number of positive revision dates exceeds the number of

negative revision dates in firm-year t. A revision date is considered to be positive if more analysts

revise up than down on that date, and

NREVS is the number of unique analyst earnings forecast revision dates from I/B/E/S for firm i in year t.

R. Frankel et al. / Journal of Accounting and Economics 41 (2006) 29–5444


reports statistics for the log of the number of analyst forecast revision dates, LnNREVS,for our sample. Since the median number of analysts following a firm is about 5, (e1.61), amedian number of revisions (NREVS) of about 23 (e3.13), implies the median analyst issuesalmost 5 earnings forecasts a year.

We report cross-correlations among the variables in Table 2. Correlations are estimatedusing data pooled across the 8-year sample period. AI exhibits significant positive product-moment correlation with volatility and trading volume. The rank correlation reinforcesthis result. Endogenous relations among these variables preclude us from drawing stronginferences from univariate correlations. Firm size, volatility, trading volume, and analystfollowing are highly cross-correlated variables. A negative relation between firm size andreturn volatility and a positive relation between firm size and trading volume and analystfollowing are well known from previous research. Our measure of the price-earningsassociation, AccRsq, is positively correlated with firm size, with a product-momentcorrelation of 0.38. This is consistent with the evidence in Hayn (1995) and Collins et al.(1997) that small firms have weaker price-earnings relation than large firms, perhapsbecause of the higher frequency of losses among small firms.

4.5. Results

Table 3 reports results of estimating the two-stage-least-squares (2SLS) regression model(i.e., the system of Eqs. (1)–(3)) of the determinants of AI. We estimate the model each yearand tabulate the time-series means of the 2SLS regression coefficients for the first equation.(We omit tabulating the results for the second and third equations for parsimony.) The t-statistics are for the sample mean coefficients, calculated using the standard deviation ofthe respective time series of estimated coefficients (see Fama and MacBeth, 1973).17 Wereport results with and without trimming extreme observations from the data, whereextreme observations are those in the lowest and highest percentiles of the annualdistribution of AI. Since the two sets of results are quite similar, we focus our discussion onthe results using the outlier-trimmed sample.

Overall, the results are largely consistent with the predicted relations between AI and itsdeterminants. Return volatility and trading volume positively influence the informative-ness of analyst research. Number of shareholders reduces AI, which suggests that theincremental demand for analyst services from shareholders increases analysts’ supply tothe point that the marginal effect of shareholders on informativeness is negative. Lowinformativeness is attractive to shareholders because they are unlikely to (perceivethemselves to) be informationally disadvantaged.

We also find that AI significantly declines in both market model R2 (MMRsg) andnumber of segments (NSEGS). The result of reduced informativeness for firms withmultiple segments suggests investors seeking guidance from analyst reports in cases whereinformation analysis is likely to be costly are helped little by analysts’ research. Sinceanalysts specialize in individual industries, reports for multi-segment firms are less accurate

17While we throughout discuss results of the 2SLS estimation to address endogeneity, we find that OLS

estimation of Eq. (1) provides qualitatively similar results as those from the 2SLS estimation. This suggests that

either endogeneity is not a severe factor in the relation between analyst informativeness and its determinants and/

or our 2SLS approach is not effective in modeling endogeneity such that it does not differ much from OLS

estimation. We cannot differentiate between these competing explanations, but are inclined to favor the 2SLS

analysis because of the economic intuition underlying the model.

ARTICLE IN PRESS

Table

2

Cross-C

orrelationsamongvariables

AI

s2(R

)LnVOL

INST

LnOwners

MB

LnMV

NSIC

MMRsq

NSEGS

AccRsq

LnNREVS

LnAnalyst

GNEWS

AI

0.06

0.04

0.02

�0.05

0.01

�0.03

0.05

0.00

�0.01

0.06

�0.02

�0.02

�0.01

s2(R

)0.12

0.05

�0.30

�0.22

0.06

�0.37

0.29

�0.06

�0.11

�0.15

�0.26

�0.22

�0.05

LnVOL

0.11

�0.08

0.41

0.43

0.18

0.67

0.12

0.55

0.17

0.27

0.62

0.69

0.02

INST

0.08

�0.29

0.43

0.18

0.06

0.54

�0.11

0.38

0.14

0.34

0.52

0.51

0.15

LnOwners�0.04

�0.35

0.34

0.19

0.05

0.57

�0.13

0.30

0.26

0.14

0.38

0.45

0.01

MB

0.03

�0.02

0.23

0.17

0.07

0.26

0.17

0.13

�0.06

0.03

0.10

0.13

0.16

LnMV

0.03

�0.47

0.63

0.58

0.47

0.40

�0.08

0.57

0.27

0.38

0.70

0.77

0.17

NSIC

0.02

0.29

0.12

�0.13

�0.13

0.12

�0.07

0.02

�0.14

0.01

�0.04

�0.02

0.04

MMRsq

0.08

�0.09

0.57

0.44

0.25

0.26

0.61

�0.10

0.18

0.31

0.44

0.44

0.11

NSEGS

�0.05

�0.17

0.13

0.14

0.24

0.09

0.24

�0.16

0.20

0.14

0.12

0.16

�0.02

AccRsq

0.11

�0.11

0.30

0.32

0.13

0.04

0.37

0.00

0.32

0.20

0.34

0.36

0.02

LnNREVS

0.05

�0.28

0.65

0.53

0.36

0.23

0.73

�0.02

0.51

0.11

0.30

0.89

0.06

LnAnalyst

0.04

�0.29

0.68

0.52

0.39

0.24

0.77

0.00

0.49

0.14

0.30

0.92

0.07

GNEWS�0.01

�0.06

0.01

0.15

0.00

0.24

0.18

0.03

0.14

�0.02

0.03

0.06

0.07

Thistablecontainsvalues

ofthecorrelationsbetweenthevariablesusedin

thesubsequenttests.Spearm

an(Pearson)correlationsare

below(above)

thediagonal.The

sampleisderived

from

allanalystearningsforecastrevisionsontheI/B/E/S

detaildatabase

betweenJanuary

1,1995andDecem

ber

31,2002.Firm

years

withI/B/E/S

data

are

merged

withCRSP,Compustat,andCDA

Spectrum

fordata

onstock

returns,financialinform

ation,andinstitutionalownership.Firm-years

withmissing

data

are

discarded.Thestatisticsare

basedonafinalsample

of23,798firm

-yearobservations.Correlationssignificantatthe0.05level,tw

o-tailed,appearin

bold.

Variable

definitionsappearin

Table

1.

R. Frankel et al. / Journal of Accounting and Economics 41 (2006) 29–5446

ARTICLE IN PRESS

Table 3

Annual two-stage-least-squares estimation

Independent variables Predicted sign Eq. (1) Eq. (1) AI outliers trimmed

Coefficient Coefficientt-stat t-stat

Fitted (σ2(R )) + 0.10037 3.44 0.11609 4.28Fitted (LnVOL) + 0.00008 3.90 0.00004 2.19INST + 0.00035 3.52 0.00039 4.26LnOwners ? −0.00005 −4.15 −0.00005 −4.42MB ? 0.00000 −0.41 0.00000 −0.92 LnMV ? 0.00002 0.59 0.00002 0.61NSIC −

− 0.00027 0.22 −0.00075 −0.54

MMRsq −4.31 −0.00068 −4.28NSEGS ? −0.00002

−0.00082−2.49 − 0.00002 −2.19

AccRsq ? 0.00007 4.09 0.00006 4.98LnAnalyst ? −0.00012 −2.43 −0.00003 −0.65GNEWS ? −0.00004 −1.61 −0.00006 −2.25Adjusted R2 1.83% 2.28%(F-Stat) (6.06) (8.26)

This table presents the time-series means of coefficients, t-statistics for the averages of the coefficients, and average

adjusted R2s from annual cross-sectional two-stage-least-squares regression model below. The sample is derived

from all analyst earnings forecast revisions on the I/B/E/S detail database between January 1, 1995 and December

31, 2002. Firm years with I/B/E/S data are merged with CRSP, Compustat, and CDA Spectrum for data on stock

returns, financial information, and institutional ownership. Firm years with missing data are discarded. The final

sample consists of 23,798 firm-year observations.

AI ¼ aþ b1Fittedðs2ðRÞÞ þ b2FittedðLnVOLÞ þ b3INSTþ b4LnOWNERS

þ b5MBþ b6LnMVþ b7NSICþ b8MMRsqþ b9NSEGSþ b10AccRsq

þ b11LnNREVSþ b12LnANALYSTþ b13GNEWSþ e1, ð1Þ

s2ðRÞ ¼ aþ b1INSTþ b2LnOWNERSþ b3MBþ b4LnMVþ b5NSIC

þ b6MMRsqþ b7NSEGSþ b8AccRsqþ b9r_LnVOLþ e2, ð2Þ

LnVOL ¼ aþ b1INSTþ b2LnOWNERSþ b3MBþ b4LnMVþ b5NSIC

þ b6MMRsqþ b7NSEGSþ b8AccRsqþ b9r_s2ðRÞ þ e3: ð3Þ

R_ indicates that the variable is assigned a number of 0, 1, or 2 on the basis of ranking each year all

firm-year observations for the variable into three equal portfolios, and

Fitted(.) indicates predicted values produced by Eqs. (2) and (3).

Variable definitions appear in Table 1.

R. Frankel et al. / Journal of Accounting and Economics 41 (2006) 29–54 47

(see Gilson et al., 2001). An alternative interpretation of the negative impact of MMRsg,NSEGS, and the number of shareholders on AI is that all of these variables reduce AIthrough their high correlation with firm size (see Table 2). This inference is weakened,however, because firm size (LnMV) is included in the estimation.

We find that AI increases in the strength of the price-earnings association, AccRsq.18

Like Francis et al. (2002), our result runs counter to the models of Holthausen and

18The association between accounting information and security prices may be clustered by industry. To control

for this potential problem we recompute AccRsq including industry fixed effects in the regression. The statistical


Verrecchia (1988), Demski and Feltham (1994), and Subramanyam (1996).19 In thesemodels Bayesian investors are expected to place lower weight on analyst reports in settingprices when corporate accounting information is timely, i.e., contemporaneously highlyassociated with security price movements. The complementarity we observe contradictsthis notion, but is consistent with the findings in previous empirical research. Specifically,Francis et al. (2002) conclude that the information in earnings announcements does notdecrease in competing information sources like the information in analyst reports.Similarly, Lang and Lundholm (1996) and others find that timely (often interpreted as highquality) accounting disclosures, i.e., high AccRsq, and analyst following are complements(e.g., Lang and Lundholm, 1996). We show that accounting reports’ informativeness andinformation in analyst reports are complements. The result is consistent with high AccRsqreducing the demand for AI while, at the same time, leading to a reduction in analysts’ costof supplying informativeness. The resulting outward shift in supply of informativenessmore than offsets the effect of reduced demand on equilibrium informativeness.The results in Table 3 provide mixed evidence on whether the ability to provide

incremental information is critical to analysts when deciding to follow a firm. We find asignificantly negative relation between number of analysts following the firm and theinformativeness of an analyst report, AI, for the entire sample (t-statistic ¼ �2.43, p-value ¼ 0.01), which implies an additional analyst will follow a firm even if theinformativeness declines as a result. Since the analyst derives benefits not only from beinginformative, but also by catering to the demand for repackaging and retransmittingfinancial information, it is plausible that the marginal analyst follows a firmnotwithstanding a decline in the price informativeness of the reports. However, this resultis attributable to outliers, as the relation becomes insignificant when outliers of AI aredeleted from the distribution. A finding of no relation between AI and the number ofanalysts following would be expected if (a) analyst following and report frequency rise withanalysts’ ability to discover valuable information and (b) the supply of reports does not gobeyond the point where competition among analysts reduces the price reaction to eachreport. However, lack of power and misspecification can also lead to the lack of afinding.20


significance of the other coefficients in the analysis remains unchanged, and the industry-adjusted AccRsq

continues to be positive and statistically significant with a p-value less than 0.01.19We assume the price-earnings association to be exogenous with respect to analyst informativeness. However,

we recognize that the informativeness of analyst reports can affect the price-earnings association, which means the

latter is endogenous. As we discuss above, we believe that the price-earnings association estimated using annual

data is unlikely to be significantly affected by the informativeness of analyst reports because analyst

informativeness likely influences the timing when information is reflected in stock prices, but that over a 1-year

horizon the total information set incorporated into prices is relatively unaffected. This supports our treatment of

price-earnings association as exogenous. To the extent our assumption is violated and the relation is endogenous,

it is possible that the sign and magnitude of the coefficient on AccRsq are incorrect.20Instead of estimating the incremental informativeness of a report with respect to an analyst, we also estimate

it with respect to an analyst report. That is, we use LnNREVS instead of LnAnalysts as the independent variable

in the regression model to explain AI, the informativeness of a report. However, because the number of analysts

reports (LnNREVS) and the number of analysts (LnAnalysts) exhibit a correlation of about 0.90 (see Table 2), the

results using LnNREVS are similar to those in Table 3 using LnAnalysts. Additionally, when both LnAnalysts

and LnNREVS are included in the model, due to multi-collinearity, we find both to load up highly significantly,

but with opposite signs, such that their net joint effect is indistinguishable from zero. Thus, the implication


Finally, the good news dummy, GNEWS, proxying for whether average earningsforecast news for a firm is good or bad during a year, is significant and negative. This resultimplies bad news analyst forecasts have greater price impact than upward forecastrevisions. The result suggests market have greater foreknowledge of the information inanalysts’ positive forecast revisions, perhaps because management has an incentive todisseminate good news to the market through financial reports and/or voluntarydisclosures. The evidence is also consistent with the following two explanations: (i) theHong et al. (2000) hypothesize that bad news travels more slowly than good news, so themarket is surprised by analysts’ negative forecast revisions, and (ii) analysts are reluctantto revise earnings forecasts downward, so when they do revise a forecast downward, itindicates severe bad news.

As we mention above, we do not tabulate the results of regressions 2 and 3. The coefficientsin these models are, for the most part, statistically significant in the hypothesized direction.

4.6. Sensitivity analysis

Because this paper is the first to measure AI using a scaled measure of the absolute valueof the market’s reaction to analyst reports, we perform additional analysis to add to theintuitive appeal of our measure and to demonstrate the robustness of the results. As notedearlier, previous research shows a positive relation between the surprise in analyst forecastrevisions and stock returns (see Lys and Sohn, 1990). Therefore, if our measure of AI, i.e.,the absolute stock price reaction to an analyst’s forecast revision, is well specified, then weexpect it to be increasing in the magnitude of the surprise in the analyst’s forecast revisionof annual earnings. To investigate this conjecture, for each 1-year-ahead forecast-revisiondate in our sample we measure the unexpected component of the analyst’s forecast ofannual earnings (UAF) as follows:

UAFitm ¼AFitm � ConForm�1j j

Pm�1,

where AFitm is the Analyst i’s 1-year-ahead earnings forecast on day t in month m,ConForm�1 the consensus forecast reported on IBES in month m�1, and Pm�1 the shareprice for the previous month.

We rank the absolute values of the surprise in analyst forecasts and place them into threeportfolios.21 We recalculate analyst informativeness (defined earlier) by firm year within eachUAF portfolio and examine the association between UAF and analyst informativeness.

Table 4 reports average absolute surprise and average AI for the three portfolios rankedby absolute forecast surprise. As expected, AI increases in the magnitude of surprise inanalyst forecasts. The mean (median) AI for the lowest absolute forecast surprise portfolio


remains unchanged, i.e., the marginal effect of analyst following or the number of analyst reports on the

informativeness of the reports is not statistically discernible.21We delete observations with share price below $2 to avoid the potential impact of a small denominator.

Furthermore, if more than one analyst has issued a forecast revision on any given date, we calculate the surprise as

the average absolute value of the surprises in analysts’ forecasts, UAF, on that date. Finally, not all of the forecast

revisions for a firm in a given year are assigned to the same portfolio. For example, a firm year containing five

forecast revision dates might have three assigned to the portfolio of smallest forecast surprises and two to the

portfolios consisting of largest absolute forecast surprises.


is 0.0043 (0.0040) compared to 0.0051 (0.0044) for the portfolio comprising the largestabsolute forecast surprises.22 When we test whether the mean and median AI are differentacross portfolios, we find that the mean and median AI of portfolio 2 are statistically largerthan portfolio 1, and those of portfolio 3 are larger than portfolio 1. Thus, the results inTable 4 are consistent with the market’s reaction (as captured in AI) to the surprise inanalysts’ forecast revisions being proportional to the magnitude of the surprise.

4.6.1. AI and Squared returns

We also investigate the sensitivity of our results to the use of absolute vales. FollowingBeaver (1968), we recalculate AI using squared returns. We find the correlation between AIcomputed using absolute returns and AI computed from squared returns is 0.91. Weestimate our system of equations using the squared return AI measure as the dependentvariable. Overall, the results, not tabulated, are remarkably similar to those shown inTable 3. In particular, we find that AI is increasing in return variance, volume, institutionalholdings and accounting R2, while it is decreasing in number of owners, market-model R2,number of segments, and when analysts report good news.

4.6.2. Model estimates with random event dates

To examine whether firms-specific factors such as risk lead to a mechanical relationbetween our measure of AI and our explanatory variables, we re-compute AI for eachfirm-year using random event dates instead of analyst forecast release dates. We find nosignificant coefficients at the 10% level. While we cannot rule out risk-based explanationfor the coefficients estimates presented in Table 3, this results suggests that returns onanalysts forecast-release dates must be incorporated in such an explanation. In our viewthis reduces the likelihood of such an explanation.

4.6.3. Multiple reports per report date

AI measures the average informativeness of an analyst report date. This is notnecessarily the same as the informativeness of an average analyst report, because multipleanalyst reports can be released on a given date. To provide a sense of the importance ofthis distinction, we create a variable to measure the average number of analyst reportsreleased per report date in a given firm year. We do this by using the I/B/E/S detail tape tocompute the average number of different analysts issuing revisions on a given day providedthat that day has at least one report. We then average this daily measure by firm-year toproduce a variable consistent with the other variables in our 2SLS model. For our sample-firm years, the mean (median) number of analysts issuing a report on a report date is 1.32(1.18). The standard deviation is 0.39. Thus for most firm-years in our sample, analystreports occur at a rate slightly greater than one per day, though relatively few firms in anyyear average more than two analyst report on a report day. We re-estimate the 2SLS modelin Table 3 adding this yearly average of reports-per-report-day variable as an independentvariable. The results are similar to those shown in Table 3, in that all variables that are

22Informativeness, AI, averaged across the three portfolios in Table 4 indicates the average informativeness of

analyst reports is greater in Table 4 than reported earlier. This arises because of differences in the samples. For

Table 4, we impose fewer data-availability requirements (e.g., price-earnings association or institutional

ownership data are not required). As a result, we obtain a bigger sample that is skewed toward including smaller

stocks compared to the samples used in Tables 1–3. Since informativeness is likely to be greater for small firms, we

observe larger values of AI in Table 4 than reported earlier.

ARTICLE IN PRESS

Table 4

Forecast Revision and Analyst Informativeness

Unexpected analyst

forecast (UAF) Rank

AI mean median

(obs.)

UAF mean

median (obs.)

T-stat (Z-stat) for test

of differences in means

(medians) across AI

portfolios

0.0044 0.0005

1 0.0041 0.0006 —

(23,334) (23,334)

0.0048 0.0031 (Port2�Port1)

2 0.0042 0.0031 11.74***

(26,536) (26,536) (9.26)***

0.0052 0.0374 (Port3�Port2)

3 0.0044 0.0146 13.30***

(24,388) (24,388) (11.70)***

0.0048 0.0136

All 0.0042 0.0031

(74,258) (74,258)

*** Statistically significant at the 0.01 level.

This table gives means and medians of AI and UAF (Unexpected Analyst Forecast) across observations in each

UAF portfolio. UAF is computed for each 1-year-ahead forecast date on the I/B/E/S detail tape from 1995 and

2003 as follows:

UAFitm ¼AFitm � ConForm�1j j

Pm�1,

where AFitm is analyst i’s 1-year-ahead earnings forecast on date t in month m, ConForm�1 is the consensus

forecast reported on IBES in month m�1, and Pm�1 is the share price for the previous month. The consensus

forecast taken from the I/B/E/S summary tape. We delete observations with share price below $2. If more than

one analyst issues a year-ahead forecast revision on a given date, we calculate UAF as the average over all analyst

forecasts on that date. We rank UAF observations and place them into three portfolios. We then compute AI for

each firm-year portfolio. AI is described in the notes to Table 1.

R. Frankel et al. / Journal of Accounting and Economics 41 (2006) 29–54 51

significant in Table 3 remain significant with the addition of the reports-per-report-dayvariable and no variables that were insignificant in Table 3 become significant.Interestingly, the coefficient on the reports-per-report-day variable is positive andsignificant with at t-statistic of 2.11 (2.03 with outliers deleted). Thus, when more revisionsoccur on a given day, AI increases. This result is not surprising in light of the other resultsin the paper because, like our other results, it suggests analysts issue a report when they canprovide information. In sum, this analysis suggests that the interpretation of our resultsapplies to both report dates and reports.

5. Summary and conclusions

In this paper we provide further evidence that, on average, analysts’ reports areinformative. That is, the market’s reaction on dates analysts issue reports is greater than(in absolute value) on other days, on average. We then focus on the cross-sectionaldeterminants of AI, controlling for endogeneity among the factors affecting informative-ness. Our main findings are: (i) AI increases in uncertain environments that are likely to


lead to high brokerage profits (i.e., high return volatility, high trading volume and highinstitutional ownership); (ii) AI is reduced in circumstances of increased informationprocessing costs (e.g., more business segments); (iii) analysts are more informative whenfinancial statements are more highly related to prices, which is consistent with analystscomplementing financial disclosure; and (iv) the marginal impact of an (additional) analystor an analyst report on the informativeness of analyst reports is indistinguishable fromzero.Our finding that the estimated informativeness of an analyst report does not decline with

the number of analysts (or analyst reports) for a firm has two implications. First, the resultsupports the reliability of the analyst following as a proxy for information productionabout a firm. This proxy is widely used in the literature (e.g., Bhushan, 1989b; O’Brien andBhushan, 1990; Brennan and Subramanyam, 1995; Lang and Lundholm, 1996). Second,the result suggests that the ability to supply information is an important impetus for ananalyst when deciding to follow a company.Overall, these results suggest that analyst activities are sensitive to factors associated

with the demand for and the supply of informative research. Thus, they represent achallenge to those who allege that analyst activities are primarily a marketing device forbrokerage firms that provides little information to investors. Instead analysts become moreinformative when investors can derive greater benefits from private information andbecome less informative when the costs of supplying private information are high. Theseresults suggest that desire to supply private information is a significant impetus for analystactivities.

References

Alford, A., Berger, P., 1999. A simultaneous equations analysis of forecast accuracy, analyst following, and

trading volume. Journal of Accounting, Auditing, and Finance 14, 219–240.

Ball, R., Kothari, S., Robin, A., 2000. The effect of institutional factors on properties of accounting earnings.

Journal of Accounting & Economics 29, 1–51.

Barth, M., 2002. Discussion of compensation policy and discretionary disclosure. Journal of Accounting &

Economics 34, 311–318.

Barth, M., Hutton, A., 2004. Analyst earnings forecast revisions and the pricing of accruals. Review of

Accounting Studies 9, 59–96.

Barth, M., Beaver, W., Landsman, W., 1996. Relative pricing multiples on equity book value and net income: an

examination of the abandonment option hypothesis. Working paper, Stanford University.

Basu, S., 1997. The conservatism principle and the asymmetric timeliness of earnings. Journal of Accounting &


Beaver, W., 1968. The information content of annual earnings announcements. Journal of Accounting Research

Supplement 6, 67–92.

Beaver, W., 1998. Financial Reporting: An Accounting Revolution, third ed. Prentice-Hall, Upper Saddle River,

NJ.

Bhushan, R., 1989a. Collection of information about publicly traded firms: Theory and evidence. Journal of

Accounting & Economics 11, 183–206.

Bhushan, R., 1989b. Firm characteristics and analyst following. Journal of Accounting & Economics 11, 255–274.

Bradshaw, M., Richardson, S., Sloan, R., 2003. Pump and dump: an empirical analysis of the relation between

corporate financing activities and sell-side analyst research. Working paper, University of Michigan Business

School.

Brennan, M., Hughes, P., 1991. Stock price and the supply of information. Journal of Finance 46, 1665–1691.

Brennan, M., Subramanyam, A., 1995. Investment analysis and price formation in securities markets. Journal of

Financial Economics 38, 361–381.


Brown, P., Ball, R., 1967. Some preliminary findings on the association between the earnings of a firm, its industry

and the economy. Journal of Accounting Research Supplement 5, 55–77.

Collins, D., Kothari, S., Rayburn, J., 1987. Firm size and information content of prices with respect to earnings.

Journal of Accounting & Economics 9, 111–138.

Collins, D., Maydew, E., Weiss, I., 1997. Changes in the value-relevance of earnings and book values over the past

forty years. Journal of Accounting & Economics 24, 39–67.

Cowen, A, Groysberg, B., Healy, P., 2005. What types of analyst firms make more optimistic forecasts? Journal of

Accounting & Economics, forthcoming.

Dechow, P., Hutton, A., Sloan, R., 2000. The relation between analysts’ long-term earnings forecasts and stock

price performance following equity offerings. Contemporary Accounting Research 17, 1–32.

Demski, J., Feltham, G., 1994. Market response to financial reports. Journal of Accounting and Economics 17,

3–40.

Douglas, W., 1933. Protecting the investor. Yale Review 21, 523–524.

Elgers, P., Lo, M., Pfeiffer, R., 2001. Delayed security price adjustments to financial analysts’ forecasts of annual

earnings. Working paper, University of Massachusetts, Amherst.

Fama, E., MacBeth, J., 1973. Risk, return, and equilibrium: empirical tests. Journal of Political Economy 81,

607–636.

Foster, G., 1981. Intra-industry information transfers associated with earnings releases. Journal of Accounting &


Francis, J., Schipper, K., 1999. Have financial statements lost their relevance? Journal of Accounting Research 37,

319–352.

Francis, J., Soffer, L., 1997. The relative informativeness of analysts’ stock recommendations and earnings

forecast revisions. Journal of Accounting Research 35, 193–211.

Francis, J., Schipper, K., Vincent, L., 2002. Earnings announcements and competing information. Journal of


Frankel, R., Li, X., 2004. Characteristics of a firm’s information environment and the information asymmetry

between insiders and outsiders. Journal of Accounting & Economics 37, 229–259.

Gilson, S., Healy, P., Noe, C., Palepu, K., 2001. Analyst specialization and conglomerate stock breakups. Journal

of Accounting Research 39, 565–582.

Givoly, D., Lakonishok, J., 1979. The information content of financial analysts’ forecasts of earnings. Journal of


Greene, W., 2000. Econometric Analysis, fourth ed. Prentice-Hall, Upper Saddle River, NJ.

Harris, M., Raviv, A., 1993. Differences of opinion make a horse race. Review of Financial Studies 6, 473–506.

Hayn, C., 1995. The information content of losses. Journal of Accounting & Economics 20, 125–153.

Hentschel, L., Kothari, S., 2001. Are corporations reducing or taking risks with derivatives? Journal of Financial

and Quantitative Analysis 36, 93–118.

Holthausen, R., Verrecchia, R., 1988. The effect of sequential information releases on the variance of price

changes in an intertemporal multi-asset market. Journal of Accounting Research 26, 82–106.

Hong, H., Lim, T., Stein, J., 2000. Bad news travels slowly: Size, analyst coverage and the profitability of

momentum strategies. Journal of Finance 55, 265–295.

Irvine, P., 2000. Do analysts generate trades for their firms? Evidence from the Toronto Stock Exchange. Journal

of Accounting & Economics 30, 209–226.

Kandel, E., Pearson, N., 1995. Differential interpretation of public signals and trade in speculative markets.

Journal of Political Economy 103, 831–853.

Karpoff, J., 1986. A theory of trading volume. Journal of Finance 41, 1069–1087.

Kim, O., Verrecchia, R., 1991a. Trading volume and price reactions to public announcements. Journal of

Accounting Research 29, 302–321.

Kim, O., Verrecchia, R., 1991b. Market reactions to anticipated announcements. Journal of Financial Economics

30, 273–309.

Kim, O., Verrecchia, R., 1997. Pre-announcement and event-period private information. Journal of Accounting &


King, B., 1966. Market and industry factors in stock price behavior. Journal of Business 39, 139–190.

Kothari, S., Zimmerman, J., 1995. Price and return models. Journal of Accounting & Economics 20, 155–192.

Landsman, W., Maydew, E., 2002. Has the information content of quarterly earnings announcements declined in

the past three decades? Journal of Accounting Research 40, 797–808.


Lang, M., Lundholm, R., 1996. Corporate disclosure policy and analyst behavior. The Accounting Review 71,

467–492.

Lin, H., McNichols, M., 1998. Underwriting relationships, analysts’ earnings forecasts and investment

recommendations. Journal of Accounting & Economics 25, 101–127.

Lin, H., McNichols, M., O’Brien, P., 2003. Analyst impartiality and investment banking relationships.Working

paper, Stanford University.

Lys, T., Sabino, J., 1992. Research design issues in grouping-based tests. Journal of Financial Economics 32,

355–387.

Lys, T., Sohn, S., 1990. The association between revisions of financial analysts’ earnings forecasts and security

price changes. Journal of Accounting & Economics 13, 341–363.

Merton, R., 1987. A simple model of capital market equilibrium with incomplete information. Journal of Finance

42, 483–510.

Michaely, R., Womack, K., 1999. Conflict of interest and the credibility of underwriter analyst recommendations.

Review of Financial Studies 12, 573–608.

Mikhail, M., Walther, B., Willis, R., 1997. Do security analysts improve their performance with experience?

Journal of Accounting Research Supplement 35, 131–157.

O’Brien, P., Bhushan, R., 1990. Analyst following and institutional ownership. Journal of Accounting Research

28, 55–76.

Schipper, K., 1991. Analysts’ forecasts. Accounting Horizons 5, 105–121.

Skinner, D., 1994. Why firms voluntarily disclose bad news. Journal of Accounting Research 32, 38–60.

Subramanyam, K., 1996. Uncertain precision and price reactions to information. The Accounting Review 71,

207–220.

Verrecchia, R., 1982. Information acquisition in a noisy rational expectations economy. Econometrica 50,

1415–1430.

Documents

Determinants of the informativeness of analyst … Kothari and Weber...informative. However, if market participants are assumed to ﬁxate on or under- or over-react to analyst reports,