ABSTRACT PUBLIC INFORMATION: EVIDENCE FROM RECOMMENDATION REVISIONS. Zheng Wang, Doctor of Philosophy, Accounting and Information Assurance

Transcription

1 ABSTRACT Ttle of Document: ANALYSTS SUPERIORITY IN PROCESSING PUBLIC INFORMATION: EVIDENCE FROM RECOMMENDATION REVISIONS Zheng Wang, Doctor of Phlosophy, 2006 Drected by: Professor Olver Km Accountng and Informaton Assurance In ths paper, I study analysts superorty over the market n processng publcly dsclosed earnngs nformaton by examnng a sample of recommendaton revsons ssued subsequent to annual earnngs announcements wthn a short perod of thrty tradng days. The man fndngs of ths study are as follows: Frst, I provde strong evdence that these recommendaton revsons convey valuable nformaton to the market for clarfyng the long term mplcatons of recently released earnngs. These revsons sgnfcantly alter the market's belef about the value mplcatons of announced earnngs, suggestng that analysts do have superorty over the market n processng pubc

2 nformaton. Also, the extent of ths superorty s postvely related to analysts performance n pckng stocks and forecastng earnngs. Recommendaton revsons ssued by analysts wth superor performance can make the market revse ts assessment about the value mplcatons of prevous earnngs to a much greater extent than those ssued by analysts wth moderate performance. Moreover, the extent of ths superorty ncreases wth the level of nformaton complexty of earnngs sgnals. Analysts nformaton s even more valuable to the market for reevaluatng prevous earnngs when the earnngs nformaton s more dffcult to analyze. Lastly, on average, the extent of ths superorty declnes after Regulaton Far Dsclosure, but stll remans sgnfcant, suggestng that analysts do not solely rely on nsde nformaton from the management to nterpret publc nformaton. Actually, the declne n the extent of superorty s more lkely due to a great ncrease n the number of revsons ssued by analysts whose expertse s not n processng publc nformaton. Pror studes document that nvestors also use subsequent earnngs announcements to adjust ther estmate of the value mplcatons of prevous earnngs. Ths study fnds ntal evdence that when analysts nformaton and subsequent earnngs announcements provde consstent predctons on how prevous earnngs s msnterpreted, subsequent earnngs announcements become less useful to nvestors for updatng ther belefs regardng the mplcatons of prevously released earnngs. Ths paper also compares the extent of analysts superorty n processng publcly released earnngs nformaton across ndustres and fnd that analysts exhbt a greater degree of superorty for frms n the manufacturng and retal ndustry.

3 ANALYSTS SUPERIORITY IN PROCESSING PUBLIC INFORMATION: EVIDENCE FROM RECOMMENDATION REVISIONS By Zheng Wang Dssertaton submtted to the Faculty of the Graduate School of the Unversty of Maryland, College Park n partal fulfllment of the requrements for the degree of Doctor of Phlosophy 2006 Advsory Commttee: Professor Olver Km, Char Professor Martn Loeb Professor Taewoo Park Professor Jm Bedngfeld Professor Sarah Eno

4 Copy Rght by Zheng Wang 2006

5 ACKNOWLEDGMENTS Frst, I would lke to acknowledge my advsor Olver Km for hs support, patence and encouragement throughout my graduate studes. He has taught me many lessons on the workngs of ths dssertaton and also academc research n general. It has been my true pleasure to work wth and learn from such an excellent researcher. Second, thanks are due to Professors Taweoo Park, Martn Loeb and Bedngfeld for agreeng to serve on my dssertaton commttee and for ther valuable comments on varous versons of ths dssertaton. I also want to thank Professor Sarah Eno for servng on my commttee as the Dean s representatve. Thrd, I would also lke to extend my personal thanks to Professor Mark Chen n the Fnance department and Professor Rchard Mendenhall at Unversty of Notre Dame for ther expert comments on ths dssertaton. Fnally, I owe my deepest thanks to my famly my mother, Yuan Zhhong, my father, Wang Xanju and my husband, Wang Le for ther uncondtonal love and support. I would also lke to express my grattude to my frends and colleagues n the PhD program for ther help and frendshp that I wll chersh forever.

6 Table of Contents Chapter 1: Introducton... 1 Chapter 2: Lterature Revew... 9 Chapter 3: Methodology and Hypothess Development Methodology Development Hypothess Development Chapter 4: Sample Selecton and Desgn of Emprcal Tests Sample Selecton Desgn of Emprcal Tests Chapter 5: Results and Dscusson Descrptve Statstcs Regresson Results Analysts' superorty n processng publc nformaton The Relatonshp between Analysts Performance and Ther Ablty to Analyze Publc Informaton Informaton Complexty The Impact of Reg FD Chapter 6: Subsequent Earnngs Announcements Chapter 7: Industry Analyss Chapter 8: Concluson References... 69

7 Lst of Fgures Fgure 1: Groups of Observatons wth Increased or Decreased ERC Fgure 2: Daly Average Cumulatve Abnormal Returns v

8 Lst of Tables Table 1: Sample Descrptve Statstcs Table 2: Estmated Results of Belef Revson of Value Implcaton Table 3: Estmated Results of Stock Pckng Ablty and Forecastng Ablty Table 4: Estmated Results of the Effect of Informaton Complexty Table 5: Estmated Results of the Impact of Reg FD Table 6: New Analysts v.s. Experenced Analysts Table 7: Estmated Results of the Impact of Reg FD on Good News v.s. Bad News Table 8: Subsequent Earnngs Announcements Table 9: Industry Analyss v

9 Chapter 1: Introducton Pror studes consstently document that the market reacts to analysts' forecast and recommendaton revsons. It s mplctly assumed n those studes that the value of analysts' nformaton s attrbuted to ther ablty to collect and create nformaton that s totally new to the market. For example, analysts have access to the management s nsde nformaton before t becomes publc (Francs et al., 1997; Bowen et al., 2002). However, t could be that another crucal aspect of analysts expertse, analysts' superor ablty to process publc nformaton, whch has been neglected by pror lterature, also makes a sgnfcant contrbuton to the value of analysts nformaton. In fact, ths aspect of analysts' expertse may have become the domnant factor of ther ablty to make recommendatons and forecasts valued by the market after the adopton of Regulaton Far Dsclosure (Reg FD). Ths regulaton prohbts selectve dsclosure of materal nformaton to fnancal professonals ncludng analysts and requres broad, non-exclusonary dsclosure of such nformaton. 1 Elmnatng selectve dsclosure could create more ncentves for analysts to research on publc nformaton release such as earnngs announcements. Even before the mplementaton of Reg FD, earnngs announcements already drve a large fracton of forecast and recommendaton revsons, mplyng that analysts rely on materal publc nformaton to ssue forecasts and recommendatons (Stckel, 1989; Ivkovc and Jegadeesh, 2004). 1 Recent studes have shown evdence that Reg FD s takng effect (e.g., Sunder, 2002; Ztzewtz, 2002; Eleswarapu et al., 2004). 1

10 Ths paper studes analysts' superorty n processng publc nformaton by examnng a sample of ther recommendaton revsons ssued wthn thrty tradng days after annual earnngs announcements. Whle analysts could acqure prvate nformaton from the management any tme durng the year, revsons ssued followng the release of materal publc nformaton are more lkely to come from ther analyss of announced pubc sgnals. If analysts were more sklled at processng publc nformaton, these revsons would be useful to the market for reevaluatng prevously released publc nformaton. Durng earnngs announcements, the management publcly dsclose sgnfcant nformaton about frms' performance. Prce senstvtes to earnngs announcements are determned by the value mplcatons of currently announced earnngs for future earnngs such as ts persstence (Kormend and Lpe, 1987; Easton and Zmjewsk, 1989). However, the value mplcaton of earnngs s uncertan to both analysts and the market, and can only be estmated by them based on ther respectve knowledge. 2 If analysts were superor to the market n nterpretng publc nformaton, they would be able to make a more accurate estmaton than the market. 3 Ths paper argues that by revsng ther recommendatons followng earnngs announcements, analysts reveal ther belef about the value mplcatons of recently released earnngs. After observng those recommendaton revsons, the market would realze that t may have msreacted to pror earnngs news and would revse ts own belef accordngly. Thus, the extent of the market's belef revson represents the extent of analysts' superorty over the market n processng publc nformaton. Ivkovc and 2 There are many studes nvestgatng nvestors' learnng process under parameter uncertanty (e.g., Lewellen and Shanken, 2002; Chen, Francs and Jang, 2005) 3 Km and Verreccha (1994) suggest that earnngs announcements provde nformaton that allows certan traders to make judgments about a frm's performance that are superor to the judgments of other traders. 2

11 Jegadeesh (2004) menton that the concentraton of recommendaton revsons followng earnngs announcements suggests that analysts dentfy nstances of msprcng durng ths perod. Otherwse, there s no reason to expect that recommendaton revsons are trggered by publc earnngs announcements. A smple model developed n ths paper explans how analysts recommendaton revsons help the market correct securty msprcng attrbuted to ts pror msassessment about the value mplcatons of announced earnngs. Upward (downward) recommendaton revsons followng good (bad) earnngs news reveal that as compared to analysts, the market has underestmated the value mplcaton of prevously released earnngs. Such revsons would cause the market to revse ts belef about value mplcaton upward. In contrast, downward (upward) recommendaton revsons ssued subsequent to good (bad) earnngs news suggest that the market has overestmated the value mplcaton of earnngs durng the announcement. The market would revse ts belef about value mplcaton downward. Therefore, after earnngs announcements, a statstcal relaton should be found between return responses and pror earnngs nnovatons surroundng analysts recommendaton revsons f the market reassesses the value mplcatons of pror earnngs nnovatons accordng to analysts nformaton released through those revsons. The drecton of the relaton depends on whether analysts' assessment about value mplcaton s hgher or lower than the market's pror assessment. The strength of the relaton, whch represents the extent of the market's belef revson, could be used to measure the value of analysts' expertse n processng publc nformaton. The above settng provdes a methodology to examne analysts' superorty over the market n processng publc nformaton. Usng ths methodology, the followng 3

12 questons have been studed: 1. Do analysts ndeed have superorty over the market n processng publc nformaton? If so, what s the extent of ths superorty? 2. Does the extent of ths superorty dffer among analysts? 3. When nformaton complexty ncreases, does ths superorty become more (less) sgnfcant? 4. Does ths superorty truly come from analysts ntellectual skll n processng publc nformaton or gudance from the management? The man fndngs regardng the above four questons are as follows: Frst, I fnd a sgnfcant statstcal relaton between returns to recommendaton revsons and pror earnngs nnovatons n the predcted drectons after earnngs announcements. Prce responses to recommendaton revsons are postvely (negatvely) related to pror earnngs nnovatons for the group of observatons, where analysts' assessment about the value mplcatons of announced earnngs s hgher (lower) than the market's pror assessment. Ths s consstent wth the conjecture that recommendaton revsons are nformatve about the value mplcatons of prevously released earnngs, whch causes the market to revse ts pror belef. Results reveal that the extent of the market's belef revson can be as hgh as about 14%, suggestng that analysts can process publc nformaton better than the market to a great extent. Second, analysts possess dfferng abltes to nterpret publc nformaton. Analysts who have performed better n forecastng earnngs and pckng stocks also have exhbted greater expertse n analyzng the value relevance of publc nformaton. Followng pror studes, I use average excess returns earned by analysts' past recommendaton revsons to proxy for ther stock pckng ablty and ther past forecast accuracy to proxy for ther forecastng sklls. Usng both proxes, I fnd consstent evdence that recommendaton 4

13 revsons ssued by sklled analysts can alter the market's assessment about the value mplcatons of prevous earnngs at least 110% greater than those ssued by moderately sklled analysts. It seems that ths superorty s an ntegral part of fnancal analysts ablty as nformaton ntermedares, an mportant part of whose jobs s to forecast earnngs and pck proftable stocks. Thrd, nvestors need to use more of analysts nformaton to clarfy the mplcatons of prevously released earnngs news when the nformaton complexty (uncertanty) of earnngs news s hgh. The extent of analysts superorty ncreases wth the complexty of earnngs sgnals. When earnngs nformaton s dffcult to analyze, the market may need more gudance from the analysts to help them correctly nfer ts longterm mplcatons for frms future performance. Followng pror studes, I use the measure of three-day wndow abnormal tradng volumes surroundng the earnngs announcement to represent ts nformaton complexty. The results show that the extent of the market's belef revson caused by analysts recommendaton revsons goes up by at least 43% wth the ncrease of nformaton uncertanty. Fnally, by examnng the mpact of Reg FD on analysts superorty n processng publc nformaton, ths study reveals some evdence on the sources of ths superorty. Snce the adopton of Reg FD, analysts have lost ther access to the management s nsde nformaton. However, f they contnue to process publc nformaton better than the market, then t means that ntellectually, they are ndeed more skllful than the market as to analyze the value relevance of publc nformaton. The estmated results on the mpact of Reg FD show that prce reactons surroundng recommendaton revsons are stll sgnfcantly assocated wth prevous earnngs nnovatons even n the post Reg FD 5

14 perod, although the extent of the assocaton dd declne. Therefore, t seems that at least analysts do not solely rely on the gudance from the management to help themselves nterpret publcly dsclosed earnngs nformaton. Actually, the declne of the extent of ths superorty s more lkely due to the fact that analysts who are not so sklled at processng publc nformaton started ssung recommendaton revsons followng earnngs announcements after Reg FD. The methodology adopted n ths paper s smlar to that used n Freeman and Tse (1989), Mendenhall (1991), and Koch and Sun (2004). These studes fnd that the market learns about the value mplcatons of prevously announced earnngs n lght of postannouncement nformaton such as current-quarter earnngs announcements, analyst forecasts, or dvdend announcements. However, ths study dffers from thers n several crtcal aspects. Ths paper s the frst one to apply ths methodology to closely nvestgate analysts' superorty over the market n processng publc nformaton. Also, I explctly propose the extent of the market's belef revson about the value mplcatons of earnngs as a measure of the value of analysts' nformatonal advantage as publc nformaton processors. In addton, n contrast wth Mendenhall's study, I use recommendaton revsons rather than forecast revsons to nvestgate the role of analysts nformaton n sgnalng the value mplcatons of prevous earnngs. Compared to forecast revsons, analysts recommendaton revsons drectly reflect analysts' evaluaton on securty value relatve to ts prce. Therefore, recommendaton revsons trggered by earnngs announcements are more ndcatve as to whether the securty s msprced because of the market's msreacton to prevously released earnngs nformaton. 6

15 An nterestng by-product of ths paper s the fndng that analysts nformaton can substtute for subsequent earnngs announcements n terms of helpng nvestors reevaluate pror earnngs announcements. Freeman and Tse (1989) fnd that nvestors use subsequent earnngs announcements to adjust ther estmate of the persstence of prevous perods earnngs. I document that when analysts nformaton and subsequent earnngs announcements provde consstent predctons on how the value mplcatons of pror earnngs news are msestmated, subsequent earnngs announcements become less useful to nvestors for clarfyng the mplcatons of prevous earnngs news. Ths fndng also confrms that analysts nformaton s valuable to nvestors for helpng correct ther msreactons to publcly dsclosed earnngs news. In ths study, I also compare the role of analysts and the extent of ther superorty across ndustres. Accordng to the SIC ndustry classfcaton code, I dvde the whole sample nto seven ndustres: Manufacturng, Agrculture, Mnng, Wholesale, Retal, Servce, Constructon and Transportaton and Publc Utlty. I fnd that analysts exhbt a greater degree of superorty n processng publcly released earnngs nformaton for frms n the manufacturng and retal ndustry. A possble explanaton for ths result s that manufacturng and retal frms usually report fnancal nformaton that s more dffcult to analyze than frms n the other ndustres due to ther complex operaton processes. Therefore, gudance from analysts who follow these two ndustres s even more mportant to nvestors. The rest of the paper s organzed as follows: Chapter 2 hghlghts related pror research; Chapter 3 develops the man hypothess; Chapter 4 descrbes sample selecton and desgn of emprcal tests; Chapter 5 dscusses the emprcal results; Chapter 6 7

16 examnes the nformaton content of subsequent earnngs announcements wth the exstence of analysts nformaton. Chapter 7 presents the ndustry analyss. Chapter 8 concludes. 8

17 Chapter 2: Lterature Revew Numerous studes have shown that analysts nformaton s valuable to the market by examnng market s reactons to revsons of analysts forecast or recommendatons. Early studes such as Lloyd-Daves and Canes (1978) fnd on average, an event day abnormal return of 0.93% (-2.37%) for new favorable (unfavorable) recommendatons, whch appear n the Wall Street Journal. Lys and Sohn (1990) provde evdence that ndvdual analysts earnngs forecasts are nformatve, even when they are preceded by earnngs forecast, whch s made by other analysts or by corporate accountng dsclosure. More recent studes (Stckel, 1992; Womack, 1996; Mkhal et al., 2005) also provde consstent evdence that forecast and recommendaton revsons have nformaton content to the market. Asquth, Mkhal and Au (2005) nvestgate the contents of analyst reports n ther entrety and fnd that the other elements of ther reports are also sgnfcantly and postvely assocated wth the market's reacton at the tme a report s released. Some studes record that the nformaton content of analysts forecasts or recommendatons s related to factors such as analyst reputaton, frm sze, brokerage profts, and brokerage sze, etc. (Stckel, 1992; Gleason and Lee, 2003; Frankel et al., 2003). Partally, the value of analysts' forecasts and recommendatons stems from ther ablty to collect non-publc nformaton from the management, whch has caught the attenton of the SEC. The SEC s concerned that the ssuers' selectve dsclosure of materal non-publc nformaton to securty analysts has hurt unnformed nvestors. 9

18 Despte the controversal evdence as to whether prvate communcatons between analysts and management make the market better or worse off (Bushman, 1991; 4 Das et al., 1998; 5 Francs et al., ), Reg FD became effectve on October 23, Motvated by Reg FD, Ivkovc and Jegadeesh (2004) evaluate the nformaton content of analysts' one-quarter ahead earnngs forecast and recommendaton revsons at varous ponts n tme relatve to earnngs announcement dates. They fnd that prce reactons and the senstvty of prce reactons to revsons are weaker for revsons n the week after earnngs announcements than n the week before earnngs announcements and conclude that analysts' nformatonal advantage as prvate nformaton collector s more mportant than as publc nformaton processor. Asquth, Mkhal and Au (2005) suggest that analysts may play a role n nterpretng nformaton prevously released by examnng market reactons to analyst reports whch occurred smultaneously wth other nformaton releases. Park and Pncus (2000) look nto a smlar ssue by examnng whether analysts recommendaton revsons have ncremental nformaton content beyond current earnngs surprses durng earnngs announcements. I fnd evdence of analysts superorty n processng publc nformaton n the sense that the mpact of analysts nformaton s related to pror earnngs surprses. Therefore, my study drectly shows how analysts nterpretaton of publcly dsclosed earnngs nformaton s valued by the market. 4 Bushman (1991) draws on the concluson that f frms are gven the power to alter the structure of the prvate nformaton market through selectve dsclosures of prvate nformaton before any publc release, traders may be made better off. 5 Das et. al (1998) provde evdence that analysts ssued optmstc forecasts to facltate ther access to management's non-publc nformaton. 6 Francs et al (1997) fnd that both frms and analysts beneft from corporate presentatons to securty analysts. 10

19 The evdence on the market's msreacton to earnngs announcements s mxed. There are several studes tryng to explan post-earnngs announcement drfts as an ntal underreacton to earnngs and a subsequent delayed prce response n the same drecton (Bernard and Thomas, 1990; Wggns, 1991; Mendenhall, 1991; Abarbanell and Bernard, 1992; Shane and Brous, 2001; Lang, 2003;). There are also studes documentng pror overreactons to news events (DeBondt and Thaler, 1985, 1987, 1990; Klen, 1990; Chopra and Rtter, 1992;). Abarbanell and Bernard also fnd that value lne analysts appear to underreact to recent earnngs nformaton. However, stock prces appear to underreact to an even greater degree. Mendenhall's evdence also ndcates that stock prces reflect less nformaton than analysts' forecasts. The methodology aspect of ths paper s n the sprt of Freeman and Tse (1989), Mendenhall (1991), and Koch and Sun (2004). Freeman and Tse (1989) report that the market learns about the value mplcatons of prevously announced earnngs from current earnngs announcements. 7 Koch and Sun (2004) show that nvestors reassess the persstence of recently announced earnngs usng nformaton released durng dvdend announcements. Ther studes also suggest that the market may overreact or underreact to earnngs news. Fnally, ths study jons a growng lterature examnng the mpact of Reg FD on analysts' nformaton envronment. Baley et al. (2003) show that analysts' forecast 7 Mendenhall (1991) fnds that nvestors correct ther prevous underweghts to earnngs when analyst forecast revsons have the same sgn wth pror earnngs nnovatons. In contrast to recommendaton revsons, analysts naturally update ther forecasts followng earnngs announcements. Even though forecast revsons have the same sgn as pror earnngs nnovatons, the extent of the revson would be larger or smaller than the market's pror expectaton change on future earnngs durng earnngs announcements, whch means that the market could underweght or overweght pror earnngs news. Therefore, n Mendenhall's paper, the R 2 of the regresson results s very smaller. 11

20 dsperson and other measures of dsagreement ncreases after the adopton of Reg FD. Ztzewtz (2002) fnds that mult-forecast days that typcally follow publc announcements or events now account for over 70 percent of the new nformaton about earnngs, up from 35% before Reg FD. Agrawal, Chadha and Chen (2006) document that earnngs forecasts become less accurate and forecast dsperson across ndvdual analysts followng a frm ncreases post-reg FD. 8 My paper augments ths lterature by drectly examnng whether analysts' nformatonal advantage n processng publc nformaton has changed n the post-reg FD perods. 8 Some of the other papers that examne the mpact of Reg FD on analysts' nformaton envronment are Mohanran and Sunder (2001), Hefln, Subramanyam and Zhang (2003), Iran and Karamanou (2003). 12

21 Chapter 3: Methodology and Hypothess Development 3.1 Methodology Development Pror studes suggest that the Earnngs Response Coeffcent (ERC) s an ncreasng functon of current earnngs' value mplcatons for future earnngs,.e., persstence and relablty, under the assumpton of dscounted earnngs valuaton. However, durng an earnngs announcement, the real value mplcaton s unknown or uncertan to the market. For example, nvestors may not be able to dscern how much of current earnngs nnovaton wll persst n the future perods or how relable the reported earnngs number s. Therefore, the market has to react to the earnngs news based on ts assessment about ts value mplcaton. After observng the market s reacton to the earnngs announcement, an analyst revses her recommendaton for the frm's stock f she beleves that the market msprces pror earnngs news accordng to her own assessment about earnngs value mplcaton. Assume, durng earnngs announcement, the ntal prce reacton to the announcement of E t (perod t's reported earnngs) s determned as: R = ERC ( E F ), (1) t where ERC s the value multple attached to the announced earnngs by the market whch s determned by varous factors related to the value mplcaton of earnngs such as persstence, relablty and others based on the market's ntal assessment of these factors. t t 13

22 F t t denotes the consensus analyst forecast for E t before t s announced. After analyzng the earnngs news released durng the announcement, the analyst beleves that the prce should react to the earnngs nnovaton - E wth an ERC equal to ERC A based on t t F t her own assessment about the value mplcaton of earnngs nnovaton. Therefore, to the analyst, the prce reacton to the announced earnngs should be equal to: A A R = ERC ( E F ). (2) t t t Any dfference between ERC and ERC A means that the analyst would thnk the market msprces pror earnngs news and would revse her recommendaton based on the sgn of ( R A R ), whch s equal to: A A t R R = ( ERC ERC) ( E F ). (3) When ERC A > ERC and the earnngs nnovaton s postve (negatve), the analyst would revse her recommendaton upward (downward) because the securty s underprced (overprced). When ERC A < ERC and the earnngs nnovaton s postve (negatve), the analyst would revse her recommendaton downward (upward) because the securty s overprced (underprced). After observng the recommendaton revson, the market would know whether t underreacted or overreacted to pror earnngs news durng the announcement and correct ts pror msreacton by the amount of ( R A R ). Therefore, after earnngs announcement, prce response surroundng the analyst s recommendaton revson would stll be related to pror earnngs nnovaton as equaton (3). The sgn of the t t 14

23 coeffcent on ( E F ) can be dentfed by the sgns of pror earnngs nnovatons and t t t analysts recommendaton revsons as llustrated n Fgure 1. 9 [INSERT FIGURE 1 HERE] For Group 1, the analyst revses her recommendaton upward (downward) and pror earnngs nnovaton s postve (negatve). Ths means that the frm's stock s underprced (overprced) due to the market's underreacton to pror good (bad) news. For Group 2, the analyst revses her recommendaton downward (upward) and pror earnngs nnovaton s postve (negatve). Ths means that the frm's stock s overprced (underprced) due to the market's overreacton to pror good (bad) news. 10 From equaton (3), the magntude of the coeffcent on ( E F ) t t t s equal to ERC A ERC, whch represents the dfference between the analyst s opnon on the value mplcaton of the earnngs nnovaton and that of the market. If the market follows the analyst s opnon, then ERC A ERC can also be understood as the extent to whch the market s belef about the value mplcaton of prevous earnngs s altered by the analyst s nformaton. Therefore, the magntude of ths coeffcent measures the value of 9 I assgn a numercal value for each recommendaton: 5 strong buy; 4 buy; 3 hold; 2 underperform; 1 sell. Therefore, postve revson means that the recommendaton s revsed upward, whle negatve revson means that the recommendaton s revsed downward. 10 I use the recommendaton revson rather than the recommendaton tself (Buy or Sell) to group the observatons because of two reasons: frst, recommendaton revsons reflect the change of analysts' opnons on securty prces caused by the market's reacton to earnngs announcements. Second, analysts are reluctant to ssue ''sell'' recommendatons. In a sample of 17,093 recommendatons examned by ths paper, there are only 1,412 ''underperform'' or ''sell'' recommendatons and 15,680 ''buy'' or ''strong buy'' recommendatons. 15

24 analysts nformaton advantage over the market n processng publcly dsclosed earnngs nformaton. 3.2 Hypothess Development From the prevous dscussons, I hypothesze that f analysts can nterpret publc nformaton better than the market, ther recommendaton revsons ssued after earnngs announcements convey valuable nformaton to the market for reassessng the value mplcaton of recently released earnngs nformaton. If so, as ndcated by equaton (3), there would exst a statstcal relaton between returns surroundng analysts recommendaton revsons and pror earnngs nnovatons post earnngs announcements. The sgn of ths relaton s postve (negatve) f the observaton falls nto group 1 (2) as descrbed n Fgure 1. The strength of ths relaton - the absolute magntude of the coeffcent on t Et F t, represents the extent of analysts superorty over the market n processng publcly released earnngs news. Specfcally, I predct: H1: For Group 1 (2), returns to recommendaton revsons ssued after earnngs announcements are postvely (negatvely) related to pror earnngs nnovatons. H1 nvestgates whether generally analysts are superor to the market n processng publc nformaton, specfcally publcly dsclosed earnngs nformaton. However, f ths superorty n processng publc nformaton s an ntegral part of analysts abltes as nformaton ntermedares, shouldn t the extent of ths superorty also dffer among analysts themselves? Shouldn t analysts wth superor performance exhbt greater extent of ths superorty? Otherwse, the expertse of analyzng publc nformaton would not be regarded as useful sklls to analysts. 16

25 Pror studes evaluate analysts superorty (performance) based on ther stock pckng and earnngs forecastng records. Mkhal et al. (2004) fnd that securty analysts exhbt persstent dfferences n ther stock pckng ablty. Analysts whose recommendaton revsons earned more (less) excess returns n the past contnue to outperform (underperform) n the future. Stckel (1992) and Mkhal et al. (1999) document that analysts' promotons and job termnatons are related to ther forecastng ablty gauged by ther past forecast accuracy. The next set of tests focuses on examnng whether analysts ablty to process publc nformaton contrbutes to ther abltes to pck stocks and forecast earnngs. If so, analysts performance would be postvely related to the extent of ther superorty n processng publc nformaton. In ths study, the extent of ths superorty s measured by how much the market s assessment about the value mplcatons of prevous earnngs can be altered by ther recommendaton revsons. Therefore, I expect that the revsons ssued by superor analysts would be able to cause a greater assessment revson than those ssued by moderate analysts, whch means that the strength of the assocaton between returns to those revsons ssued by superor analysts and pror earnngs nnovatons should be stronger. Followng pror studes, I use hgh excess returns earned by the analyst s past recommendaton revsons to represent superor stock pckng ablty and her accurate forecast record as a proxy for superor forecastng ablty. Hypothess 2 s as follows: H2: The strength of the statstcal relaton between returns to recommendaton revsons ssued after earnngs announcements and pror earnngs nnovatons ncreases 17

26 wth the average excess returns earned by the analyst's past recommendaton revsons and the accuracy of her past forecast record. If I fnd that analyst nformaton can help the market solve ts uncertanty about the value mplcatons of earnngs, then consequently, another research queston arses: can analysts help resolve more uncertanty for the market as the level of nformaton uncertanty (complexty) ncreases? When earnngs nformaton s really dffcult to analyze, would analysts nformaton become more or less valuable to nvestors n terms of clarfyng the value mplcatons of prevous earnngs? When the level of nformaton uncertanty s hgher, t s expected that prce msreactons to earnngs announcements would be more severe due to nvestors less accurate estmate of the mplcatons of earnngs news. One may thnk that analysts gudance would become more mportant and valuable to nvestors for reevaluatng prevous earnngs news and thus under such stuatons, analysts would exhbt a greater extent of superorty over the market. However, as earnngs nformaton becomes more dffcult to analyze, analysts would also have more dffcultes n correctly assessng ts value mplcaton themselves. If analysts do not have enough expertse to deal wth the dffcultes, t s not certan that ther nformaton would be more useful to nvestors and could cause nvestors to adjust ther estmate about the value mplcatons of prevous earnngs as much as when earnngs nformaton s less complex. As elte nformaton processors, analysts are expected to provde valuable gudance to nvestors for nterpretng publcly released nformaton especally when the nformaton s dffcult to analyze. Therefore, the next hypothess s developed to study whether the extent of analysts superorty over the market becomes greater or smaller 18

27 when the level of nformaton complexty of earnngs announcements ncreases. Pror theoretcal studes suggest that tradng volume s related to the market's dfferent opnons of a sgnal (Holthausen and Verreccha, 1990, Km and Verreccha, 1991, 1994). Hgh abnormal tradng volume reflects a greater degree of heterogenety of nvestors' opnons on the same nformaton sgnal when they are uncertan about the sgnal. Snce t s not clear how nformaton uncertanty affects the extent of analysts' superorty over the market, I predct: H3: The strength of the statstcal relaton between returns to recommendaton revsons drven by earnngs announcements and pror earnngs nnovatons s not related to the nformaton uncertanty of earnngs announcements. Untl now, an mportant queston that has not been addressed yet s the source of analysts superorty n processng publc nformaton f t does exst. Does ths superorty ndeed stem from analysts ntellectual sklls? Or actually analysts rely on nsde nformaton from the management to help themselves nterpret publc nformaton. Reg FD provdes an deal envronment to nvestgate whether analysts channel to the management s non-publc nformaton s the sole or major source of ths superorty. Snce Reg FD was adopted on October 23, 2000, the management have been requred to dssemnate any materal nformaton smultaneously to all market partcpants. Therefore, after Reg FD, analysts channel to the management s nsde nformaton has been blocked. Ztzewtz (2002) reports evdence that Reg FD has had ts desred effect of reducng selectve dsclosure of nformaton about frms' future performance to ndvdual analysts. If analysts access to the management s nsde nformaton s the sole source of ther superorty, after Reg FD, they would no longer exhbt sgnfcant advantage over the 19

28 market n processng publc nformaton. At the same tme, analysts would revse ther recommendatons less frequently after Reg FD because they can not get as much the management s prvate nformaton as before Reg FD. However, f analysts are ndeed more skllful than nvestors n processng publc nformaton, after ther access to the management s nsde nformaton has been blocked, they would stll be able to process publc nformaton better than nvestors. Meanwhle, snce prvate communcatons of materal nformaton between the management and analysts have been prohbted after Reg FD, analysts whose expertse s n seekng nsde nformaton from the management would be forced to follow materal publc nformaton release such as earnngs announcements to make recommendatons. Therefore, n the post-reg FD perod, there would be an ncrease n the number of recommendaton revsons ssued subsequent to earnngs announcements by analysts who may lack the expertse to process publc nformaton. Hence, on average, analysts recommendaton revson would be less nformatve to the market n the sense that they would not be able to cause the market to revse ts belef about the value mplcatons of announced earnngs as much as before Reg FD. As dscussed above, after Reg FD, t s expected that the extent of analysts superorty over the market may have declned, but t should reman sgnfcant f analysts access to the management s nsde nformaton s not the sole source of ths superorty. If not, then analysts superorty n processng publc nformaton would have been gone after Reg FD. H4 s developed to nvestgate the mpact of Reg FD on the extent of ths superorty. 20

29 H4: The extent of the statstcal relaton between returns to recommendaton revsons drven by earnngs announcements and pror earnngs nnovatons has changed after the adopton of Reg FD. 21

30 Chapter 4: Sample Selecton and Desgn of Emprcal Tests 4.1 Sample Selecton The above hypotheses are tested on a sample of recommendaton revsons made after November 1993 ncluded n the 2004 I/B/E/S database. The annual earnngs announcement dates and reported values of earnngs also come from I/B/E/S. The stock returns and prce data come from the 2004 Center for Research n Securty Prces (CRSP) fles. A sample frm must meet the followng data requrements to be ncluded n the tests: (1) at least one analyst revses her recommendaton followng ts annual earnngs announcement wthn 30 tradng days wth the revson date recorded n I/B/E/S ; (2) annual earnngs nformaton wth actual reported value and actual reported date recorded n I/B/E/S; (3) analysts forecasts used to calculate earnngs nnovatons for announced earnngs; (4) daly stock returns and begnnng of perod prce for the 3-day wndow centered on the recommendaton revson date n the CRSP daly fles. I elmnate observatons wth recommendaton revsons equal to zero n order to dentfy nstances of msreactons. Observatons wth earnngs nnovatons equal to zero are also excluded because t s not meanngful to study the value mplcatons of earnngs nnovatons that are equal to zero. The varables of earnngs nnovatons and three-day cumulatve abnormal returns are wnsorzed at the 1% and 99% level n my sample. The 22

31 above selecton crtera yeld 17,093 observatons of recommendaton revsons wth 4,553 frms n the sample for H Desgn of Emprcal Tests H1 predcts that recommendaton revsons ssued followng earnngs announcements could alter the market's belef about the value mplcatons of prevous earnngs and cause prces to react as equaton (3). In order to test ths hypothess, I estmate the followng regresson equaton usng the above sample: CAR = α + β PER FERROR + β NPER FERROR t,, j 1 t,, j t, 2 t,, j t, + β3rev t,, j + ε t,, j, (4) where CAR t,, j s the cumulatve 3-day sze adjusted abnormal returns surroundng analyst j s recommendaton revson for frm. FERROR t, s frm s earnngs nnovaton (forecast error) whch s equal to the dfference between announced earnngs and ts consensus forecast (mean forecast) deflated by the frm's begnnng prce of the 3-day wndow centered on the recommendaton revson date. PER t,, j ( NPER t,, j ) s a dummy varable wth the value equal to one f the recommendaton revson falls nto Group 1 (2) and zero otherwse. Group 1 (2) conssts of all observatons where pror earnngs nnovatons and analyst recommendatons have the same (opposte) sgn and the market's belef about earnngs value mplcatons s revsed upward (downward) as n Fgure 1. Whle there are two dummy varables for two categores, each dummy varable s nteracted wth FERROR t,. Therefore, the above equaton does not suffer from perfect 23

32 collnearty. REV t,, j s the dfference between the numercal values of analyst s j s recommendaton ssued after the earnngs announcement and her pror recommendaton ssued before the earnngs announcement. REV t,, j s ncluded n the regresson to control for the analyst's non-earnngs nformaton. The predcted sgn for the coeffcent on REV t,, j s postve. However, snce REV t,, j already contans nformaton related to analysts' nterpretaton of pror earnngs news, a certan degree of multcollneary wll be ncurred after ncludng REV t,, j n the regresson. Therefore, I wll manly use the estmated results of equaton (4) wthout REV t,, j to test the man hypothess. 11 β 1 ( β 2 ) measures the extent of the market's belef revson caused by analysts nformaton for Group 1 (2). Snce for Group 1 (2), the market revses ts belef about value mplcaton upward (downward), β 1 ( β 2 ) s predcted to be postve (negatve). Therefore, the predcted sgns of the coeffcents n equaton (4) should be: β > 0, β < 0, β > Snce REV t,, j s expected to be hghly correlated wth CAR t,, j, to check the robustness of the estmated results of equaton (4), I also test H1 by regressng the followng equaton: 11 Followng Freeman and Tse (1989) and Mendenhall (1991), I do not nclude an nteracton varable of the ntercept and dummy varables n equaton (4). To check the robustness of the results, I estmated the followng equaton usng the same sample: CARt j = α 1PERt j + α 2NPERt j + β1pert j FERRORt + β 2NPERt j FERRORt + β3revt j + ε t j Both α 1 and α 2 are smlar to α and the other coeffcents are also smlar to those estmated usng equaton (4).,,,,,,,,,,,,,,,,, 24

33 CAR = a + b PER1 FERROR + b PER2 FERROR t,, j 1 t,, j t, 2 t,, j t, + b NPER1 FERROR + b NPER2 FERROR 3 t,, j t, 4 t,, j t, + b 5 REV t,, j + ε t,, j, (5) where PER1 t,, j ( PER2 t,, j ) s equal to one f earnngs nnovaton and recommendaton revson both have postve (negatve) sgns as n the upper left (rght) dagonal of Fgure 1 and otherwse zero. NPER1 t,, j ( NPER2 tj,, ) s equal to one f earnngs nnovaton s postve (negatve), but recommendaton revson s negatve (postve) as n the lower left (rght) dagonal of Fgure 1 and otherwse zero. Therefore, the expected sgns of b 1 and b 2 are postve and the expected sgns of b 3 and b 4 are negatve. 12 It s conjectured n H2 that the magntude of the coeffcents on pror earnngs nnovatons ncreases wth analysts' stock pckng skll and earnngs forecastng skll measured usng the excess returns earned by ther past recommendaton revsons and the accuracy of ther past forecast record respectvely. In order to test ths predcton, I run the regresson usng the followng equaton: CAR = α + β' PER FERROR + β' NPER FERROR + β' SUPER t,, j 1 t,, j t, 2 t,, j t, 3 t, j PER FERROR + ' SUPER NPER FERROR t,, j t, β 4 t, j t,, j t, + β' 5 REV,, + β' 6 SUPER, REV,, + ξ, (6) t j t j t j t,, j 12 I also estmate equaton (4) by groupng the observatons as: Group 1: (Innovaton +, Buy); (Innovaton -, Sell) Group 2: (Innovaton +, Sell); (Innovaton -, Buy). The mplcatons of the results reman unchanged, but adjusted R 2 of equaton (4) becomes very small because there are only a very small fracton of ''underperform'' or ''sell'' recommendatons n the sample as dscussed n footnote 6. 25

34 where SUPER t, j s a dummy varable representng analyst j s ablty of pckng stocks or forecastng earnngs. For the analyst's ablty of pckng stocks, for each year I calculate the average cumulatve sze-adjusted 3-day (-1,0,1) wndow returns by takng long (short) postons n her upward (downward) recommendaton revsons ssued n the past year wth 0 representng the revson date. 13 Because of addtonal data requrements, the sample sze s reduced to 12,332 observatons. Then, for each year, analysts are ranked nto two groups based on the average excess returns earned by ther recommendaton revsons n the past year. SUPER t, j takes the value of one f analyst j has a hgh rank and zero otherwse. All the other varables are as defned n equaton (4). For analyst j s skll n forecastng earnngs, I estmate her past mean absolute forecast errors usng rollng three-year wndows. 14 For ths test, the sample sze s reduced to 16,232 observatons because of the lack of past analyst forecast nformaton for some frm years. Then for each year, f the analyst s past mean absolute forecast error s lower than the mean of the sample, the dummy varable of SUPER t, j takes the value of one and zero otherwse. 15 Therefore, for analysts wth moderate sklls n pckng stocks or forecastng earnngs, the coeffcent on FERROR t, s equal to β' 1 ( β' 2 ) for observatons n Group 1 (2). For analysts wth superor sklls, the coeffcents on FERROR t, are equal to β' + β' Followng Mkhal et. al. (2004), reterated recommendaton revsons are elmnated from estmatng the excess returns earned by the analysts' past recommendaton revsons because reterated revsons have lttle nformaton content as recorded n pror studes. 14 Gu and Wu (2003) argue that analysts seek to mnmze mean absolute forecast errors. 15 The results are robust usng contnuous varable of excess returns earned by the analyst's past recommendaton revsons and the accuracy of the analyst's past forecast record. The estmated results usng the dummy varables of SUPER t, j are presented n ths paper n convenence of explanaton. 26

35 and β' + β' for observatons n Group 1 and 2 respectvely. Snce the absolute magntude 2 4 of the coeffcents on FERROR t, s predcted to ncrease wth analysts' performance as nformaton processors, β 3 should be postve, and β 4 should be negatve. β 6 s predcted to be postve because recommendaton revsons ssued by analysts wth superor sklls are expected to cause larger prce reactons. Therefore, the predcted sgns for all coeffcents should be: β' > 0, β' < 0, β' > 0, β' < 0, β' > 0, β' > H3 s developed to nvestgate whether analysts nformaton can provde more gudance to nvestors for reevaluatng recently announced earnngs when they are more uncertan about ts value mplcatons. In the followng equaton, a dummy varable whch represents nformaton complexty s nteracted wth PER FERROR, t,, j t and NPER FERROR, t,, j t : CAR = α + β'' PER FERROR + β'' NPER FERROR t,, j 1 t,, j t, 2 t,, j t, + β'' AVOL PER FERROR + β'' AVOL 3 t, t,, j t, 4 t, NPER FERROR + '' REV t,, j t, β 5 t,, j + β'' 6 AVOL, REV,, + ς,,, t t j t j (7) where AVOL t, measures abnormal tradng volume around the three-day announcement wndow. Lke Landsman and Maydew (2002), the varable AVOL t, s estmated as: 1 t, tn,, t, t, n= 1 AVOL = ( V V ) / σ, 27

36 where V tn,, s the tradng volume of frm durng day n at year t deflated by shares outstandng of the frm on day n durng the 3-day wndow (-1,0,1) at year t, wth day 0 beng the announcement date. For each frm year, I estmate V t, and σ t, - the mean and standard devaton of frm s daly tradng volume dvded by daly shares outstandng of the frm durng the estmaton perod (-242, -20). Then for each year, the observatons are ranked nto two groups based on AVOL t,. Then, the value of AVOL t, s replaced by one f the observaton has a hgh rank and zero otherwse. 16 For earnngs announcements of hgh nformaton uncertanty, the coeffcents on FERROR t, are equal to β'' 1+ β'' 3 and β'' + β'' for Group 1 and 2 respectvely. Snce t s unclear whether the magntude of 2 4 the coeffcents ncreases or decreases wth the level of nformaton uncertanty as dscussed n Secton 4, there are no predcted sgns for β'' 3 and β'' 4. For recommendaton revsons ssued around earnngs announcement dates, AVOL t, would be hghly correlated wth CAR t,, j. Therefore, I exclude 5,913 observatons of recommendaton revsons ssued wthn three tradng days from the announcement dates. 17 Because addtonal data are requred to estmate nformaton uncertanty, the sample of ths test s further reduced to 10,919 observatons. The last set of tests s to study the mpact of Reg FD on analysts' superorty as publc nformaton processors n order to examne the sources of ths superorty. I nvestgate whether the extent of analysts' superorty has changed after the adopton of Reg FD by estmatng the followng regresson. 16 I also run the test usng contnuous varable of AVOL t,. The results are robust. The results estmated usng dummy varable are presented n ths paper n convenence of nterpretaton. 17 Equaton (7) s also estmated usng a sample of observatons ncludng recommendaton revsons ssused n (0, 2). The results are smlar. 28

37 CAR = α + β''' PER FERROR + β''' NPER FERROR tj,, 1 tj,, t, 2 tj,, t, + β''' POST PER FERROR + β''' POST 3 t, t,, j t, 4 t, NPER FERROR + ''' REV t,, j t, β 5 t,, j + β6 ''' POST, REV,, + ζ, (8) t t j t,, j where POST t, s the dummy varable representng post-reg FD perods. POST t, takes the value of one f the earnngs announcement date s after the adopton of Reg FD (10/23/2000) and zero otherwse. If β '' + ''' and β '' + ''' are sgnfcantly dfferent 1' β 3 2' β 4 from zero, t suggests that analysts can stll process publc nformaton better than the market after Reg FD. 29

38 Chapter 5: Results and Dscusson 5.1 Descrptve Statstcs Panel A of Table 1 shows the breakdown of the sample nto Group 1 and 2 as descrbed n Fgure 1. Group 1 contans 9,102 observatons, and n Group 2, there are a total number of 7,973 observatons. Thus, my sample contans 1,129 more observatons of market underreacton than overreacton to earnngs news. [INSERT TABLE 1 HERE] Table 1 also reports sample descrptve statstcs. Panel B shows that the cumulatve three-day market adjusted abnormal returns surroundng recommendaton revsons range from to wth mean equal to and medan equal to As more analysts revse recommendatons downward than upward n ths sample as shown n Panel A, the mean and medan of recommendaton revsons are both negatve. The mean of FERROR t, s negatve, and the medan s postve suggestng that there are more frms reportng postve earnngs nnovatons even though the magntude of postve earnngs nnovatons s smaller than that of negatve earnngs nnovatons. The correlaton table presented n Panel C shows that the cumulatve three-day returns surroundng recommendaton revsons are sgnfcantly postvely correlated wth the revsons, and the extent of the correlaton s strong ( ρ =0.3215). Returns are also 30