Access to Management and the Informativeness of Analyst Research

Transcription

1 Access to Management and the Informativeness of Analyst Research T. Clifton Green, Russell Jame, Stanimir Markov, and Musa Subasi * September 2012 Abstract We study the effects of broker-hosted investor conferences on the informativeness of analyst research. We find analysts stock recommendations have significantly larger price impacts when the broker has a conference-hosting relationship with the firm. The incremental effect is most pronounced in the quarter following the conference and remains significant for three quarters. The post-conference effect is stronger for small, volatile stocks and when the analyst has more experience covering the firm. Analysts at brokers with a conference-hosting relation also issue more accurate earnings forecasts than non-hosts in the postconference period. Our findings suggest access to management remains an important source of analysts informational advantage following the passage of Regulation Fair Disclosure. JEL: G14 * Green is from Goizueta Business School, Emory University. Jame is from School of Banking and Finance, University of New South Wales. Markov is from the School of Management, University of Texas at Dallas. Subasi is from Trulaske College of Business, University of Missouri. 1

2 1. Introduction A large literature establishes the important informational role that brokerage research analysts play in financial markets. Analysts earnings forecasts have been documented to be generally more accurate than statistical models (Brown and Rozeff, 1978; Bradshaw et al. 2012), and another line of research shows that analysts stock recommendations tend to be profitable (Womack 1996, Barber et al. 2001, Jegadeesh et al. 2004). Although analysts expertise could arise from skilled processing of public information, a common explanation for analysts' forecasting skill relies on superior access to management. Brokerage analysts place considerable emphasis on interacting with firm management through visits to company headquarters, investor office meetings, and broker-hosted investor conferences. Despite the widespread nature of these costly activities, relatively little is known about the extent to which access to management provides analysts with value-relevant information. The enactment of Regulation Fair Disclosure (Reg FD) in 2000 requires that management disclose material information to all investors at the same time, which would seem to diminish the value of private meetings with management. Indeed, several studies find evidence that Reg FD largely eliminates the benefits of access to management. For example, Cohen, Frazzini, and Malloy (2010) document that analysts with educational ties to managers issue more informative research than other analysts, but only in the pre-reg FD period, which suggests value-relevant information may no longer flow along social networks. Chen and Matsumoto (2006) find that analysts providing optimistic recommendations issue more accurate forecasts, but exclusively in the pre-regulation period, suggesting analysts providing favorable research may no longer be rewarded with value-relevant information (see also Gintschel and Markov, 2004). 2

3 On the other hand, existing work on management access relies on relatively noisy proxies based on geographic proximity (Malloy, 2004), the timing of earnings announcements (Ivkovic and Jegadeesh, 2004), analyst optimism (Chen and Matsumoto, 2006), or educational ties (Cohen, Frazzini, and Malloy 2010). In this article, we analyze a direct measure of management access using a large sample of broker-hosted investor conferences and examine whether analysts with access to management produce more informative stock recommendations and earnings forecasts. Broker-hosted investor conferences are organized to provide select investing clients with opportunities to interact with corporate managers, yet the analyst host also may reap informational benefits. The typical conference format includes formal company presentations followed by Q&A sessions, often led by the analyst-host, and series of one-on-one meetings between management and select clients, also often led by the analyst-host (see Bushee, Jung, and Miller, 2011; and Green et al for institutional evidence). With other analysts excluded from these private interactions, investor conferences present an ideal opportunity for measuring and evaluating the informational benefits of management access. We hypothesize that interaction with management at investor conferences provides analysts with an informational advantage that leads to more informative research. Research published immediately following a conference may be particularly informative, and examining contemporaneous research from non-host analysts provides a control for any public information releases. We measure the information content of analyst research as the buy-and-hold abnormal return following stock recommendation changes. 1 Our methodology involves regressing the market reaction to recommendation changes on indicator variables related to the timing and 1 Our emphasis is on recommendation changes since they generally produce larger market reactions than earnings forecast revisions, although we find similar evidence using both measures of analyst research. 3

4 source (host or non-host) of the recommendation, as well as various firm, analyst, and broker characteristics to control for factors influencing the informativeness of analyst research (Loh and Stulz, 2010). Our analysis of 2,749 investor conferences hosted by 107 brokerages reveals convincing evidence that investor conferences provide their analysts hosts with value-relevant information. We find analysts at brokerages with a hosting relation with a firm issue more informative recommendations than non-hosts. The results are robust to a variety of controls and hold with analyst and firm fixed effects. Brokerage analysts hosting firms at conferences issue especially informative research in the post-conference period. In particular, recommendation changes in the three months following conferences induce incremental abnormal returns of 30 to 50 basis points depending on the specification estimated. The post-conference effect is particularly strong for small, volatile stocks and when the analyst has more experience covering the firm. We find no evidence that recommendation changes for conference stocks by non-hosts induce incremental abnormal returns during this period, which suggests that only the host obtains value-relevant information during the conference. The informational benefits of access to management at investor conferences likely decrease over time, and we explore this conjecture by partitioning the post-conference period into six sub-periods. Relative to pre-conference recommendations, intuitively we find that recommendation revisions in the first three months following the conference induce the largest market reaction, yet recommendation changes produce significantly larger market responses for up to nine months following the conference. Moreover, after controlling for known determinants of research informativeness, we find brokers that host a firm at any point during the sample 4

5 period issue more informative research than non-hosts, which suggests that conference attendance may signal a more general connection to management. A potential alternative explanation for the market impact findings is that hosts do not obtain value-relevant information on conference days, but that market participants nevertheless perceive their post-conference research to be more informative. We address this concern by also studying the effects of management access on analysts earnings forecast accuracy. Consistent with the market impact results, we find evidence of increased forecast accuracy for conference hosts but not for other analysts in the post-conference period. Specifically, in the three months following the conference, the hosting analyst issues forecasts that are 8% to 13% more accurate than non-hosts. Together, the improved forecast accuracy and larger market response to recommendation changes in period following investor conferences provides compelling evidence that access to management is an important determinant of analysts information advantage. 2 Although analysts spend significant resources meeting with management, our findings provide some of the first direct evidence that such meetings lead to more informative research, particularly following the enactment of Regulation Fair Disclosure. Soltes (2012) examines the private interactions between sell-side analysts and the senior management of a single large-cap NYSE firm over a one year period and finds no evidence that private interactions leads to more informed research. In contrast to Soltes (2012), we examine analyst research for over 3000 different companies over a seven year period and find that management access leads to more informative recommendation changes and more accurate earnings forecasts. Our results also 2 We caution against concluding that analysts obtain material nonpublic information at investor conferences in violation of Regulation FD. Analysts may have the ability to produce value-relevant information by piecing together public information and nonmaterial information from management (i.e. the mosaic theory), and Regulation FD allows the transfer of nonmaterial information. While the issue of whether analysts specifically obtain material nonpublic information from management is important, this level of analysis is beyond the scope of the data. 5

6 compliment recent work by Bushee, Jung and Miller (2012), and Soltes and Solomon (2012), who find that institutional investors benefit from private interactions with firm management. Our findings suggest broker-hosted investor conferences provide a measure of access to management that is more effective than indirect measures based on the timing of earnings announcements (Ivkovic and Jegadeesh, 2004), analyst optimism (Chen and Matsumoto, 2006), or educational ties (Cohen, Frazzini, and Malloy, 2010). Moreover, the evidence that analysts at brokers with a hosting relation with a firm issue more informative research prior to conferences suggests conference attendance may proxy for other forms of management access such as company visits or investor office meetings The remainder of the paper is organized as follows. Section 2 describes the investor conference and analyst research data and presents descriptive statistics; Section 3 examines the effects of investor conferences on the informativeness of analyst research; and Section 4 concludes. 2. Data and Descriptive Statistics 2.1 Brokerage Research Reports Our sample consists of data on brokerage research reports and broker-hosted investor conferences. We obtain data on stock recommendations from the Institutional Broker Estimate (I/B/E/S) Recommendation History dataset. The recommendation history file contains the recommendations of individual analysts with ratings ranging from 1 (strong buy) to 5 (strong sell). We focus on recommendation revisions since prior research finds that recommendation changes are more informative than levels (see e.g. Jegadeesh et al., 2004). Recommendation changes are computed as the current rating minus the prior rating by the same analyst. We limit the sample to recommendation revisions made between 2004 and 2010 to match the sample of 6

7 investor conferences. We remove analysts coded as anonymous by I/B/ES since it is not possible to track their recommendation revisions. We also exclude reiterations of earlier recommendations. Our initial sample consists of 82,849 recommendation changes. Altinkilic and Hansen (2009) highlight that recommendation revisions often react to recent corporate news. In assessing the informativeness of analyst research, it is therefore important to remove revisions that are merely responding to firm-specific news releases. To control for firm-specific news, we follow Loh and Stulz (2010) and exclude revisions that fall in the three day window [-1,1] around quarterly earnings announcement dates (obtained from Compustat) or management earnings guidance days (as reported in Fall Call's Company Issued Guidelines Database). We also exclude recommendation revisions where multiple analysts issued a recommendation on the same day. After these filters, 46,903 recommendation changes remain. We next merge our recommendation revision sample with CRSP and Compustat. For each firm, we collect data on share price, stock returns and volume from CRSP and we obtain data on book value of equity from Compustat. We drop any firms that have missing return or volume date over the prior year, as well as any firms with missing or negative book values of equity. Our final sample includes 45,840 recommendation changes. Prior research finds that recommendation revisions have a greater impact on stock prices than revisions of earnings forecasts (see e.g. Loh and Stulz (2010)). As a result, our primary focus is on recommendation revisions. Nevertheless, as an additional test, we also examine earnings forecast revisions. We obtain data on individual analyst s earnings forecasts from the (I/B/E/S) Detail History dataset. Forecast revisions are computed as the current forecast for one- 7

8 year ahead earnings minus the prior forecast by the same analyst. 3 We impose all the same filters that were applied to our recommendation revisions sample. Our initial sample consists of 371,059 forecast revisions. This number is reduced to 171,402 after excluding firm-specific news days, and 168,285 after dropping firms with missing data in CRSP and Compustat. 2.2 Broker-Hosted Conferences We obtain data on broker-hosted investor conferences for the period January 2004 to December 2010 from the Bloomberg Corporate Events Database. The database includes information on the conference name, date, and hosting organization, as well as the presenting company name. We eliminate conferences that are not hosted by I/B/E/S-listed equity research providers which employ at least 5 analysts in a given year. We then match companies attending investor conferences by name or ticker with the CRSP and COMPUSTAT databases. Our final sample consists of 68,194 presentations, by 4,394 companies, at 2,749 conferences, hosted by 107 I/B/E/S-listed brokers. We merge our revision samples with our conference data by both broker and stock. For each revision we create four conference indicator variables: Host: An indicator variable equal to one if the recommendation revision is for a firm that attended an investor conference hosted by the analyst's brokerage house at any point over the sample period. 4 Non-Host: An indicator variable equal to one if the recommendation revision is for a firm that has never attended a conference hosted by the analyst's brokerage house at any point over the sample period. Host_Post-Conf: An indicator variable equal to one if the recommendation revision is issued in the 60 trading days following an investor conference, and the report is authored by the conference host. 5 3 We also examine forecasts of quarterly earnings and find very similar results. 4 We define Host at the broker level rather than the analyst level since the broker s resources are required to host the conference and therefore the hosting relation may not travel with analysts across brokers. We find similar results if we define Host at the analyst level. 5 Table 5 explores horizons beyond 60 days. 8

9 Non-Host_Post-Conf: An indicator variable equal to one if the recommendation revision is issued in the 60 trading days following an investor conference, and the report is authored by a Non-Host. We conjecture that firms attending broker-hosted investor conferences will have a closer ongoing relationship with the hosting analyst than non-hosts, resulting in more private interactions (e.g., more company visits and meetings with management), and a continual flow of value-relevant information throughout the sample period. We therefore hypothesize that analysts generally issue more informative research for firms that attend their conferences, that is, Host revisions are more informative than Non-Host revisions. 6 In addition to providing a signal of access to management, investor conferences also provide a specific opportunity for the transfer of value-relevant information. We therefore predict that analysts issue unusually informative research for firms that attend their conference in the post-conference period, i.e. Host_Post-Conf revisions are more informative than Host revisions. We note that Host_Post-Conf revisions are a subset of Host revisions. Bushee, Jung, and Miller (2011) find evidence that firms disclose more information publicly on conference days, which raises the concern that the increase in the informativeness of the host s research in the post-conference period is due to the analyst s ability to interpret publicly disclosed information. If this concern is valid, then Non-Host_Post-Conf revisions should also be as informative as Host_Post-Conf revisions, and including this dummy variable provides an effective control. Panel A of Table 1 describes the sample of recommendation changes. Of the 45,840 recommendation changes in our sample, 30,520 are classified as Non-Host recommendation changes, while the remaining 15,320 are Host recommendation changes. Our sample includes 6 We acknowledge the potential for endogeneity (e.g. firms may be more likely to attend conferences hosted by analysts who issue more informative research), and thus we primarily focus on the post-conference period. 9

10 1938 Host_Post-Conf recommendation changes, 916 (1022) of which are upgrades (downgrades). Panel B of Table 1 presents results for earnings forecast revisions. Our sample includes 111,396 Non-Host forecast revisions and 56,889 Host forecast revisions, of which 6,997 are Host_Post-Conf forecast revisions. 2.3 Other Variable Construction and Descriptive Statistics For each revision, we compute a number of characteristics about the revision, the analyst and brokerage firm making the revision, and the firm for which the revision is being made. In this section we discuss these characteristics and motivate their inclusion as controls. The details of the variable construction are presented in the Appendix. We first examine characteristics of the revision itself. We include an Upgrade dummy to control for the fact that upgrades and downgrades may have a differential effect on prices. We also include Abs(Rec Change) and Abs(Revision)/Price as measures of the magnitude of the recommendation and forecast revision, respectively. Kesckes, Michaely, and Womack (2011) find that stock recommendations accompanied by earnings forecast revisions lead to larger price reactions. Thus for recommendation revisions we include a Concurrent Earnings Forecast dummy, and for forecast revisions we include a Concurrent Recommendation dummy. Ivkovic and Jegadeesh (2004) show that revisions prior to (after) an earnings announcement lead to greater (weaker) price responses. We control for these effects by including a Pre_Earnings (Post_Earnings) dummy variables which equals one if the revision was made in the two weeks prior to (after) an earnings announcement. Earnings forecast and recommendation revisions that go away from the consensus lead to larger price impacts (see e.g. Gleason and Lee, 2003 and Jegadeesh and Kim, 2010). To capture this effect, we include an Away from Consensus dummy. 10

11 Lastly, we include an Affiliation dummy since the presence of an investment banking relationship with the firm may influence the informativeness of analyst research (Malloy, 2005). We next include analyst characteristics. Stickel (1995) finds that recommendation changes made by all-star analysts have greater price effects, so we create an All-Star dummy variable. We also include prior Forecast Accuracy since Loh and Mian (2006) show that analysts who possess more accurate earnings forecasts also issue more profitable recommendations. Mikhail, Walther, and Willis (1997) highlight the importance of analyst experience as a forecast accuracy determinant. We include two measures of experience: Total Experience which measures the number of years since the analyst issued research on any stock, and Firm Experience which measures the number of years the analyst has covered that specific firm minus the average experience for all other analysts covering the firm. Finally, we include Broker Size, which reflects resources available to the analyst (Clement, 1999), and several firm characteristics: Book-to-Market (BM), Size, Turnover, Volatility, Momentum, Analyst Coverage, and Conference Attendance. Panel A of Table 2 presents descriptive statistics for our sample of recommendation changes. Columns 1 and 2 reveal a number of substantial differences between Host and Non- Host recommendation changes. First, we observe that affiliated analysts account for 3% of Host revisions and 1% of Non-Host revisions. Since affiliated brokers tend have a closer relationship with firm management, this finding is consistent with the view that brokers are more likely to invite a firm to its conference if they have a close relationship with the firm's management. We also find that Host revisions are more likely to be made by All-Stars, analysts with greater firmspecific experience, and analysts who work at larger brokerage firms. In addition, Host revisions are more likely to be made for smaller firms and firms with low BM. This is consistent with 11

12 Green et al. (2012) who find that, relative to published research, investor conferences tend to overweight smaller firms with high intangible assets. We find similar, but generally amplified differences, when we compare Host_Post-Conf revisions to Non-Host_Post-Conf revisions. In particular, relative to Non-Host_Post-Conf revisions, Host_Post-Conf revisions are significantly more likely to be made by affiliated analysts, all-star analysts, analysts with greater firm-specific experience, and analysts working for larger broker houses. 7 They are less likely to be made immediately after an earnings announcement and also more likely to be bold recommendations (i.e. move away from the consensus), on smaller stocks with less analyst coverage. Panel B of Table 2 presents analogous results for our sample of earnings forecast revisions. Although our forecast revision sample consists of larger stocks with greater analyst coverage (relative to our recommendation revision sample), the patterns across the two samples are very similar. Overall, the findings from Table 2 suggest analysts hosting investor conferences have characteristics associated with more informative research. 3. Empirical Analyses 3.1 Informativeness of Analyst Revisions: Univarite Results We measure the informativeness of analyst revision (recommendation or forecast) as the stock-price reaction in the two-day event window [0,1], where day 0 is the announcement date of the revision. Following Loh and Stulz (2010), we compute the two-day cumulative buy-and-hold abnormal return (CAR) for revision i as: ( ) ( ) (1) 7 Statistical significance is estimated based on standard errors clustered by analyst and firm. 12

13 R it is the raw return of stock i on day t and is the return on day t of a benchmark portfolio with the same size, book-to-market, and momentum characteristics as the stock. 8 We winsorize CAR i at the 99 th and 1 st percentile for upgrades and downgrades separately. 9 We begin by looking at the two-day abnormal returns around recommendation upgrades and downgrades for our four conference variables. The results are presented in Panel A of Figure 1. Consistent with analysts obtaining value-relevant information at investor conferences, we find that Host_Post-Conf upgrades generate the largest two-day abnormal returns (333 bps), and Host_Post-Conf downgrades generate the most negative two-day abnormal returns (-319 bps). We also find that Host upgrades generate larger returns than Non-Host Upgrades (287 bps vs. 185 bps), and we find similar patterns for downgrades. This is consistent with the view that hosting brokers have closer relationship with the firms they invite to conferences and are thus able to issue more informative research. Lastly we see that Non-Host_Post-Conf revisions are less informative than Non-Host revisions, which is inconsistent with non-hosting analysts obtaining valuable information from conference presentations. Panel B of Figure 1 presents analogous results around forecast revisions. Consistent with prior literature, the price effects associated with forecast revisions are significantly smaller than recommendation changes. Nevertheless, a similar pattern emerges in relative informativeness across our four conference variables. In particular, Host_Post-Conf upgrades are associated with the largest two day returns (96 bps), followed by Host, Non-Host, and Non-Host_Post-Conf (50 bps). A nearly identical pattern emerges for downgrades. 3.2 Informativeness of Analyst Revisions: Regression Evidence 8 See Daniel, Grinblatt, Titman and Wermers (1997) for a more detailed discussion of the construction of the DGTW benchmark portfolio. 9 Winsorzing helps reduce the impact of extreme 2 day returns that are likely driven by firm-specific news but are not captured by our filters. Nevertheless, our results are qualitatively similar if we use non-winsorized returns. 13

14 The results from Figure 1 suggest access to management at investor conferences determines the informativeness of analyst research. In this section, we more formally investigate this idea using a regression framework. Given the largely symmetric patterns for upgrades and downgrades in Figure 1, we estimate regressions on the full sample of revision (regardless of the direction of the revision) and create a new dependent variable: CAR_IND i equal to CAR i multiplied by an indicator variable equal to 1 (-1) when the revision is an upgrade (downgrade). 10 We begin by estimating the panel regression: CAR _ IND Host Host _ Post-Conf Non-Host _ Post- Conf. (2) i 1 i 2 i 3 i Specification 1 of Panel A of Table 3 presents the results of the regression for our sample of recommendation revisions. The results confirm the findings of Figure 1 and also verify that the estimates are highly significant, where statistical significance is computed from standard errors clustered by analyst and firm. The intercept of 189 bps reflects the average two-day abnormal return for Non-Host recommendation revisions. Host revisions are associated with an incremental 80 bps return relative to Non-Host revisions. Host_Post-Conf revisions are associated with an additional 56 bps return relative to Host revisions. Thus, Host_Post-Conf revisions outperform Non-Host revisions by 136 bps and are substantially more informative. There is no evidence that nonhosting analysts issues more informative research around conference presentations. In fact, the coefficient on Non-Host_Post-Conf revisions is significantly negative (perhaps non-hosts are at an informational disadvantage following conferences). 10 Estimating the regression separately for upgrades and downgrades yields very similar coefficients, but statistical significance is reduced due to the smaller sample size. 14

15 In the bottom row, we report our estimate of A positive estimate indicates that the increase in the informativeness of host s research following conferences exceeds that of non-hosts research. This estimate is positive and statistically significant. 11 In specification 2 we add the recommendation, analyst, and broker characteristics discussed in section 2.3. To ease the interpretation of the continuous variables, we scale Total Experience, Firm Experience and Broker Size to have a standard deviation of 1. To reduce skewness, Broker Size is reported in natural logs. The coefficients on the controls variables are in line with prior literature. For example, revisions that are issued with concurrent earnings forecasts, revisions that move away from the consensus, and revisions that are made by analysts working at larger brokerage houses are more informative, while revisions made immediately after earnings are less informative. However, these controls have very little impact on our conference variables. Host and Host_Post-Conf remain positive and highly significant, while Non-Host_Post-Conf remains significantly negative. In specification 3 we add firm characteristics. All firm characteristics are scaled by their standard deviation, and all variables except the two momentum variables, are reported in natural logs. Not surprisingly, recommendations for smaller firms, firm with less analyst coverage, more volatile firms, and growth firms have larger price responses. After including firm characteristics, the coefficients on Host, Host_Post-Conf, and Post-Conf_Diff remain positive and highly significant. In specification 4 we augment specification 3 by including analyst fixed effects. This helps control for time-invariant analyst characteristics, such as persistent analyst skill (e.g. Mikhail, Walther, and Willis, 2004). Intuitively, the coefficient on Host tests whether a given 11 An alternative test is that the incremental benefits of a firm attending an analyst hosted conference is greater for the hosting analyst than the non-hosting analyst (i.e. β 2 > β 3 ). In unreported results we formally test β 2 > β 3 and find statistically significant differences across all specifications. 15

16 analyst issues more informative revisions for stocks that attend conferences hosted by the analyst s broker relative to firms that do not attend the broker s conferences. 12 Similarly, the coefficient on Host_Post-Conf now tests whether a given analyst issues more informative revisions for stocks that recently attended a conference hosted by the analyst s broker relative to firms that have attended (or will attend) conferences hosted by the broker, but not in the past 60 trading days. The coefficients on Host, Host_Post-Conf, and Post-Conf_Diff remain economically and statistically significant. In specification 5 we replace analyst fixed effects with firm fixed effects, with minimal impact on the coefficients of interest. Lastly, in specification 6 we include both analyst and firm fixed effects. The only source of variation in Host is the relatively small sample of analysts who switch brokerage houses. The coefficient on Host_Post-Conf now tests whether a given analyst issues more informative revisions for a specific stock in the 60 days after that stock attended a conference hosted by the analysts' broker. Despite the relatively low power of this test, the coefficient on Host and Host_Post-Conf remain positive and statistically significant. Moreover Host_Post-Conf revisions generate two-day abnormal returns that are 96 bps higher than Non- Host_Post-Conf revisions. To further explore the robustness of our Table 3 finings, we re-estimate specification 3 each year from Figure 2 reports the coefficients on Host, Host_Post-Conf, Non- Host_Post-Conf, and Post-Conf_Diff for each year. The figure indicates that our results are stable over time. Host is positive in all 7 years, Host_Post-Conf is positive in 6 of 7 years, and Post- Conf_Diff is positive in all seven years, ranging from a low of 49 bps in 2007 to a high of 127 bps in For a small subset of analysts who switch brokerage firms, the effects also pick up whether the analyst issues more informative research for a given firm when the analyst is a working at a broker that has invited the firm to one of its conferences. Results are nearly identical after removing analysts who switch brokerage houses. 16

17 Panel B of Table 3 repeats the analysis of Panel A, but substitutes earnings forecasts revisions for recommendation revisions. The coefficients on Host is positive in all six specifications, and is statistically significant in four specifications. The coefficient on Host_Post- Conf is also positive in all specifications and statistically significant in two specifications. Post- Conf_Diff ranges from 14 to 46 bps and is significant in all specifications except specification six, which suffers from relatively low power. In sum, the forecast revision results yields similar but weaker results than the recommendation changes. The weaker results for forecast revisions is not surprising in light of prior research finding relatively small price reactions to forecast revisions. In light of the generally weak market response to earnings forecast revisions, our remaining tests focus on recommendation revisions. 3.3 Event-Time Analysis The positive and significant coefficient on Host_Post-Conf in Table 3 supports our prediction that recent access to management allows analysts to issue more informative research. We proxy for recent access to management by identifying cases where analysts meet with management in the prior 60 days, but the benefits of access to management may persist for longer periods. For example, interactions with firms' management may allow analysts to better interpret information released by the company one or two years after the meeting. However, if much of the information is time-sensitive, then the hosting analyst s information will likely decline over time. We further explore the dynamics of the benefits of access to management by introducing additional indicator variables based on the timing of recommendation changes relative to investor conferences. As before, recommendation revisions are classified as Host_Post-Conf if the issuing analyst works for a broker that hosted the firm at a conference in the past quarter (60 17

18 trading days). For clarity in Table 4 we label this as Host_Post-Conf_Q1, and add indicator variables for Quarter 2 (days ), Quarter 3 (days ), Quarter 4 (days ), Year 2 (days ), and Beyond Year 2 (days >480). Note that by including indicators for the entire post-conference period, the coefficient on Host now captures the abnormal returns around recommendation changes issued by analysts on firms who have not yet attended a conference hosted by the analysts' brokerage firm (as of the beginning of our sample in 2004), but who will do so by the end of our sample period in 2010). Specification 1 of Table 4 presents the results of the regressions (3). Relative to Non-Host revisions, Host revisions outperform by 60 bps. The larger market response to host analyst s research prior to the conference is consistent with investor conferences signaling a more general relationship between the analyst and firm management. For example, conference attendance may proxy for other forms of access to management such as company visits or investor office meetings. 13 However, even after controlling for this general effect, revisions made by the hosting analyst in the first quarter after the firm attends its conference outperform by an additional 76 bps. This difference falls slightly to 56 and 63 bps for revisions made 2 and 3 quarters after the conference. The informational advantage falls further to 28 bps in the fourth quarter after the conference, 16 bps in the 2nd year after the conference, and 10 bps for revisions made on firms that attended the hosting brokers' conference over two years ago. 14 Specifications 2 through 6, which are analogous to the specifications reported in Table 3 except they now include the additional Post_Conf variables, yield largely similar results. The 13 Future conference attendance may also signal the existence of past conference attendance before our sample begins in Consistent with this view, the coefficient on conference Host is largest in 2004 (82 bps). However, the coefficient on Host exceeds 35 bps in each year of the sample, which suggests past conference attendance is at most a partial explanation. 14 In unreported results, we also split the first quarter into two six-week periods (days 1-30 and 31-60). The coefficient for recommendations released by the host analyst within six weeks of the conference is 82 bps, and the coefficient for the second six week period is 69 bps. 18

19 coefficient on Host_Post-Conf is the largest in the first quarter following the conference, and generally remains statistically significant for three quarters following the conference. Overall, the results are largely consistent with access to management providing analysts an immediate informational advantage that decays over time. 3.4 Cross-Sectional Determinants In this section we examine cross-sectional variation in the informational advantage of the hosting analyst in the period immediately following the conference. More specifically, we estimate the following panel regression: CAR _ IND Host Host _ Post-Conf Non-Host _ Post-Conf i 1 i 2 i 3 i β X β Host_Post-Conf * Z. 4 i 5 i i (3) X i is vector that contains all of the recommendation, analyst, broker, and firm characteristics included as controls in specification 3 of Table 3. Z i is also a vector of recommendation, analyst, broker, and firm characteristics and is a subset of X i. We include a number of variable in Z i that we believe may influence the magnitude of the Host_Post-Conf coefficient. First, we examine whether such effects are significantly larger (or smaller) for Upgrades. Managers may be more willing to disclose positive information, suggesting that upgrades may be more informative. On the other hand, the market may discount upward revisions if they believe analysts are simply rewarding management for attending the conference. Recommendations that deviate from the consensus or recommendations accompanied by earnings forecast revisions may also be particularly informative. In addition, analysts with more skill, more experience, and better resources may disproportionately benefit from meeting with management. We include All-Star and Prior Forecast Accuracy as measures of skill, Firm Experience and Total Experience as measures of experience, and Broker Size as a 19

20 measure of resources. Lastly, meetings with management may be particularly valuable for hardto-value firms such as growth firms, small firms, and more volatile firms. Table 5 reports the results of Equation (3). For brevity, we report only the coefficient on Host_Post-Conf and the interaction terms (i.e. β 5 ). Most of the coefficients are small and statistically insignificant. However, there is some evidence that Host_Post-Conf revisions are more informative when the hosting analyst also makes a concurrent earnings forecast. This is consistent with part of the hosting analyst's information advantage being driven by their ability to better forecast subsequent earnings, a possibility we more formally examine in Section 3.6. We also find that revisions are significantly more informative for smaller stocks and more volatile stocks. Intuitively, access to management is more valuable for harder-to-value firms. The sample of Host_Post-Conf recommendation changes is relatively small at 1,938. We also consider a longer post-conference period to increase the power of the test. Motivated by the long-horizon results in Table 4, we redefine Host_Post-Conf to be equal to 1 if the analyst issues a recommendation change for a firm that attended a conference hosted by the analyst s broker during the past 180 trading days (instead of 60). Lengthening the post-conference period increases the number of Host_Post-Conf recommendation changes to 5,173 and generally produces stronger interaction effects. The incremental informativeness of post-conference recommendation changes remain significantly stronger for smaller and more volatile stocks, and the effects become statistically stronger for revisions made around concurrent earnings forecasts. We also now find that the information advantage is larger for analysts with greater firm-specific experience. As firm-specific experience increases, analysts may obtain better access to management or may better interpret information revealed by management. 3.5 Longer-Horizon Returns 20

21 Our analysis relies on two-day returns around the release of revisions as a measure of the informativeness of the analysts' research. Such a measure is reasonable if markets are efficient. However, it is possible that the market incorrectly believes that the hosting analyst has more information and overreacts to revisions released by the hosting analyst. Alternatively, it is possible that market underreacts to the hosting analysts information, in which case two-day returns understate the true benefits of access to management. If the larger abnormal returns around Host and Host_Post-Conf revisions effects are driven by overreaction we would expect such revisions to exhibit reversals over longer horizons. To explore this possibility, we re-estimate the equation in Table 3, but now the dependent variable is the cumulative buy-and-hold abnormal returns over the subsequent month (CAR 21 _IND i ), subsequent six months (CAR 126 _IND i ), or subsequent year (CAR 252 _IND i ). 15 We continue to winsorize all returns at the 99th and 1st percentile for upgrades and downgrades, respectively, although non-winsorized results are similar. Figure 3 plots the coefficients on Host, Host_Post-Conf, Non-Host_Post-Conf, and Post- Conf_Diff for the three longer-horizon returns, as well as the original two-day return. The coefficients on Host is relatively flat across all time-horizons. The coefficient on Host_Post-Conf declines slightly over the first month (from 43 bps to 14 bps) and then increases to a statistically significant 203 bps (t=2.06) over the subsequent year. Similarly, Non-Host_Post-Conf increases from -9 bps over the two-day window to 90 bps over the subsequent year, however none of the estimates for Non-Host_Post-Conf are significantly different from zero. Overall, the return results are more consistent with underreaction than overreaction; however it is difficult to draw any definitive conclusions given the imprecision in estimating longer-horizon returns. In the next 15 The figure results use controls as in Specification 3 of Table 3. We find similar results using any of the other specifications in Table 3. 21

22 section, we present a more powerful test to preclude the possibility that the higher market reaction to recommendations by conference hosts is driven by overreaction. 3.6 Forecast Accuracy The results from Table 5 suggest that recommendation changes in the post conference period are particularly informative when the hosting analysts also issues an earnings forecast revision. This result is consistent with the view that access to management provides information that allows analysts to better estimate future earnings. We specifically test this conjecture by examining whether analysts issue more accurate earnings estimates for firms that recently attended a conference hosted by their brokers. Such a finding would also provide evidence that the larger market response to post-conference recommendation changes reflects information regarding fundamentals rather than overreaction by market participants. We estimate forecast accuracy using annual earnings forecasts, although quarterly earnings forecasts generate similar results. If an analyst issues more than one forecast for the firm-year, we use the most recent forecast prior to the earnings announcement date. 16 Following Clement (1999) we measure forecast accuracy as the proportional mean forecast error (PMAFE), which is calculated as: i, j, t j, t j, t AFE AFE AFE. (5) AFE i, j, t is the absolute forecast error for analyst i's forecast of firm j for year t earnings, and AFE jt, is the mean absolute forecast error for firm j in year t. We demean by firm and year to help control for differences in forecast difficulty that vary by firm-year. 16 We rely on the most recent estimate to make the results more comparable to prior literature on forecast accuracy (e.g. Mikhail, Walther, and Wallis, 1997; Jacob, Lys, and Neale, 1999; and Clement, 1999). Our findings are robust to using all forecast estimates. 22

23 We next estimate the following panel regression: PMAFE dhost dhost _ Post-Conf dnon-host _ Post- Conf. (6) ijt 1 i 2 i 3 i dhost is the firm-year demeaned value of Host as defined in section 3.2, but modified for earnings estimates. Specifically, Host now equals 1 if the analyst is issuing an earnings estimate for a firm that attended a conferences hosted by the analyst s broker at any point during the sample period. dhost_post-conf and dnon-host_post-conf are defined analogously. Since all dependent and independent variables are demeaned by firm-year averages, the regression becomes equivalent to a firm-year fixed effects regression. Consequently, the regression is estimated without an intercept. The results of regression (6) are presented in Table 6. We find evidence that Host and Host_Post-Conf earnings estimates are significantly more accurate. Specifically, Host_Post-Conf earnings estimates are 4.7% more accurate relative to Host estimates and 13% more accurate than Non-Host estimates. The findings indicate that access to management at broker conferences improves analysts forecast accuracy. The coefficient on Non-Host_Post-Conf is insignificantly different from zero, suggesting that the information advantage immediately after conferences accrues only to the hosting broker. In specification 2, we include the following control variables: Total Experience, Firm Experience, Broker Size, Forecast Age, Forecast Frequency, and Firms Followed. The construction of the control variables is presented in the Appendix. Again, all control variables are demeaned by firm-year averages. The coefficients on the controls are in line with prior literature (Clement, 1999). For example older forecasts are less accurate while forecasts made by analysts with greater firm-specific experience are more accurate. Adding the controls eliminates the 23

24 statistical significance of Host, but Host_Post-Conf and Post-Conf_Diff remain highly significant. In specification 3 we add analyst fixed effects to control for differences in innate ability. The coefficient on Host_Post-Conf remains highly significant. Similarly, the coefficient on Post- Conf_Diff is largely unchanged at 8.8%, an estimate that is both statistically and economically significant. Overall, the results suggest the access to management at investor conferences allows the hosting analyst to obtain better estimates of future earnings. This finding clarifies the nature of the value-relevant information transmitted, and precludes the alternative explanation that the market incorrectly perceives hosts research on conference stocks to be more informative. 4. Conclusion Broker-hosted investor conferences provide analysts with private opportunities for interactions with firm management. With other market participants excluded from these interactions, brokerage-hosted conferences provide an excellent opportunity for studying whether analysts obtain superior information in periods when they have greater access to management. Our analysis of 2,749 investor conferences hosted by 107 brokerages reveals convincing evidence that investor conferences provide their analysts hosts with value-relevant information. We find analysts at brokerages with a hosting relation with a firm issue more informative recommendations changes than non-hosts, and that this difference is the largest in the postconference period. In particular, recommendation revisions in the three months following conferences induce incremental abnormal returns of 30 to 50 basis points depending on the estimated specification. We consider an alternative explanation for our results, which is that market participants wrongly perceive hosts post-conference research to be more informative. We find no evidence 24

25 of return reversal in the months subsequent to the recommendation change. More importantly, we find evidence of increased forecast accuracy for conference hosts but not for other analysts in the post-conference period. These findings support our hypothesis that access to management at investor conferences generates informational advantages for the hosting analyst. While investor conferences appear to be an important mechanism through which analysts obtain management access, there are many other ways analysts interact with management. For example, analysts routinely take clients to meet management at company headquarters. Analysts also spend significant amount of time communicating with management over the phone and via . The importance of management access as a source of analysts information advantage is therefore likely to be greater than what our evidence suggests. Our finding that analysts with a hosting relationship with the firm generally issue more informative research than non-hosts throughout the sample period suggests that investor conferences may serve as a more general proxy for access to management. 25

26 Appendix: Description of Control Variables This appendix describes the construction of a number of variables describing the characteristics of the recommendation revision, earnings forecast revision, or earnings forecast. The characteristics are partitioned into three groups: Revision Characteristics, Analyst and Broker Characteristics, and Firm Characteristics. Revision Characteristics: Host_Post-Conf - a dummy variable equal to one if the revision is for a firm that attended a conference hosted by the analysts' brokerage house over the past 60 trading days. Host - a dummy variable equal to one if the revision is for a firm that attended a conference hosted by the analysts' brokerage house at any point over the sample period. Non-Host_Post-Conf - a dummy variable equal to one if the revision is for a firms that has never attended a conference hosted by the analysts' brokerage house, but who has attended a conference hosted by a different brokerage house over the past 60 trading days. Non-Host - a dummy variable equal to one if the revision is for a firm that has never attended a conference hosted by the analysts' brokerage house. Upgrade - a dummy variable equal to one if the revision is favorable (e.g. a recommendation change from hold to buy or an upward revised earnings forecast) Abs(Rec Change) the absolute value of the magnitude of the recommendation change. For example, going from a hold (=3) to a strong buy (=1), would have a value of 2. Abs(Revision)/Price the absolute value of the forecast revision change scaled by the price of the stock two days prior to the revision change. This value is winsorized at the 99%. Concurrent Forecast - a dummy variable equal to one if the recommending analyst issued an earnings forecast for the stock in the 3 days surrounding the recommendation [- 1,1] and the forecast was in the same direction as the revision. Concurrent Recommendation - a dummy variable equal to one if the analyst issuing a forecast revision also issued a recommendation change for the stock in the 3 days surrounding the forecast revision [-1,1] and the recommendation change was in the same direction as the revision. 26