What is the value of sell-side analysts? Evidence from coverage initiations and terminations

Journal of Accounting and Economics 60 (2015) 141–160

Contents lists available at ScienceDirect

Journal of Accounting and Economics journal homepage: www.elsevier.com/locate/jae

What is the value of sell-side analysts? Evidence from coverage initiations and terminations$ Kevin K. Li a, Haifeng You b,n a b

School of Business Administration, University of California Riverside, Anderson Hall, 900 University Avenue, Riverside, CA 92521, USA School of Business and Management, Hong Kong University of Science and Technology, Clear Water Bay, Kowloon, Hong Kong

a r t i c l e i n f o

abstract

Available online 11 September 2015

We investigate three potential channels of analyst value creation: improving fundamental performance through monitoring, reducing information asymmetry, and increasing investor recognition. We show that changes in investor recognition have consistent explanatory power for the market reaction to coverage initiations and terminations but find mixed evidence for changes in information asymmetry and no evidence for changes in fundamental performance as determinants of the market reaction. These results suggest that analysts create value for firms under their coverage by improving their investor recognition and not by monitoring or reducing information asymmetry. & 2015 Elsevier B.V. All rights reserved.

JEL classification: G14 G24 G29 G32 Keywords: Coverage initiation Exogenous coverage termination Firm value Investor recognition Information asymmetry

1. Introduction Prior studies document a positive market reaction to analyst coverage initiations (e.g., Branson et al., 1998; Demiroglu and Ryngaert, 2010; Irvine, 2003) and a negative market reaction to exogenous coverage terminations due to brokerage mergers or closures (Kelly and Ljungqvist, 2012). This evidence is consistent with analysts adding significant value to the firms they cover. However, it remains unclear how analysts add value. We attempt to answer this question by examining the potential channels of analyst value creation. We find robust evidence that analysts create value for the firms under their coverage by increasing the investor recognition of these companies. Equity valuation theory prescribes that the value of a firm equals the present value of its expected future cash flows. For analysts to increase firm value, their coverage should either help to improve future cash flows (i.e., the fundamental effect), reduce the cost of capital (i.e., the discount rate effect), or both (e.g., Campbell, 1991; Campbell and Shiller, 1988). Following this framework, we examine three potential channels of analyst value creation, namely, that analysts can (1) improve firms’ ☆ We greatly appreciate the comments and suggestions from John Core (the editor), Robert Hansen (the referee and the discussant), Mary Barth, Joy Begley, Mark Bradshaw, Sandra Chamberlain, Kevin Chen, Yiwei Dou, Pingyang Gao, Mingyi Hung, Amy Hutton, Ping-Sheng Koh, SP Kothari, Mark Lang, Christian Leuz, Russell Lundholm, Stan Markov, Steven Matsunaga, Miguel Minutti-Meza, Shiva Rajgopal, David Reeb, Haresh Sapra, Catherine Schrand, Charles Shi, Pervin Shroff, Derrald Stice, Jim Wahlen, Michael Welker, Guochang Zhang, and workshop participants at the National University of Singapore, Hong Kong University of Science and Technology, University of British Columbia, University of California Riverside, 2014 CAPANA Research Conference, the Eighth Annual Rotman Accounting Research Conference, and the 2014 Journal of Accounting and Economics Conference. We also thank Jake Thornock for kindly sharing the EDGAR search data with us. n Corresponding author. Tel.: þ 852 2358 7576; fax: þ852 2358 1693. E-mail addresses: [email protected] (K.K. Li), [email protected] (H. You).

http://dx.doi.org/10.1016/j.jacceco.2015.08.006 0165-4101/& 2015 Elsevier B.V. All rights reserved.

142

K.K. Li, H. You / Journal of Accounting and Economics 60 (2015) 141–160

fundamental performance by monitoring managers (e.g., Chen et al., 2014; Jung et al., 2012; Yu, 2008), (2) reduce firms’ cost of capital via decreasing information asymmetry (e.g., Bowen et al., 2008; Bradley et al., 2014; Kelly and Ljungqvist, 2012; Wu, 2013), or (3) reduce firms’ cost of capital through increasing investor recognition (Merton, 1987). We employ a multivariate regression to examine which of these channels drives the market reaction to analyst coverage initiations and exogenous coverage terminations. We use the difference between future actual earnings and analyst consensus forecasts before the coverage change to proxy for changes in fundamental performance, changes in the adverse selection component of the bid-ask spread to proxy for changes in information asymmetry, and changes in the breadth of institutional ownership to proxy for changes in investor recognition. Consistent with the prior literature, we find the market reacts significantly to both coverage initiations and exogenous coverage terminations. Our univariate tests show that compared to a sample of control firms, firms with coverage initiations (exogenous coverage terminations) exhibit a significant decrease (increase) in information asymmetry and a significant increase (decrease) in investor recognition in the following year. In contrast, we find that changes in fundamental performance are indistinguishable between our initiation and control samples and are higher for our exogenous termination sample than for the control sample, which is inconsistent with analysts creating value via improving fundamental performance. Although all three potential value drivers exhibit significant associations with the initiation/termination period returns in the underspecified univariate regressions, the coefficient of changes in fundamental performance becomes insignificant in the presence of proxies of the other two value creation channels. After controlling for the normal associations between the value drivers and stock returns, changes in information asymmetry are only marginally associated with the initiation/termination period returns. In contrast, the coefficient of changes in investor recognition remains highly significant. We conduct a battery of robustness analyses. Specifically, we expand the measurement window of the three value drivers to fully capture the effects of analyst coverage changes. We replace the ex post proxies for market expectations with ex ante proxies measured by the average changes in fundamental performance, information asymmetry, and investor recognition triggered by the analyst’s prior initiations. We also adopt various alternative proxies for the value drivers. For example, we use changes in the number of searches on the SEC’s EDGAR website for a firm’s filings to proxy for changes in investor recognition, changes in the probability of informed trading to proxy for changes in information asymmetry, and changes in return on assets to proxy for changes in fundamental performance. We find that the initiation/termination period returns are always significantly associated with the proxies for changes in investor recognition but are not correlated with any proxies for changes in fundamental performance. The evidence on the association between initiation/termination period returns and changes in information asymmetry is mixed. The association is insignificant in over 70% of the tests and is only significant at the 10% level in most of the remaining tests. These results suggest that changes in investor recognition are the most important determinant of the market reaction to analyst coverage changes. In contrast, the hypothesis that analysts create value via improving (reducing) fundamental performance (information asymmetry) is not supported in any (most) of the robustness tests. The significant association between changes in investor recognition and the initiation/termination period returns is consistent with the “value creation hypothesis”—investors react favorably (unfavorably) to coverage initiations (terminations) because they understand that analysts create value for firms by promoting the stocks to more investors. However, the results for the initiation sample are also consistent with an alternative explanation that analysts tend to initiate coverage on stocks that they expect to have higher investor recognition (the “anticipation hypothesis”).1 The anticipation hypothesis is unlikely to explain our results for several reasons. First, our early results that firms with exogenous coverage terminations experience a significant decrease in investor recognition suggest that changes in analyst coverage lead to changes in investor recognition. Second, we find that changes in investor recognition and initiation period returns tend to be higher for initiations by analysts who devote more time and effort to promoting the newly covered stocks, but they are unrelated to the analyst’s ability to predict future earnings. This suggests that the positive initiation period return reflects the increase in investor recognition generated by the analyst’s promotion of the stock, rather than the analyst’s favorable private information about the future change in investor recognition. Our paper makes several contributions. First, we synthesize the prior studies and examine the three potential channels of analyst value creation simultaneously in a multivariate setting. We show that changes in fundamental performance, information asymmetry, and investor recognition are highly correlated. Thus analyzing only one channel without considering the others may overstate its effect (e.g., Irvine, 2003; Kelly and Ljungqvist, 2012; Mola et al., 2013). Our multivariate design helps to alleviate this concern and allows us to investigate the relative importance of the channels. Using this research design, we document compelling evidence that analysts create value for the firms they cover by increasing their investor recognition. Second, we provide new evidence on analysts’ role in reducing firms’ information asymmetry. Consistent with prior studies (e.g., Bowen et al., 2008; Kelly and Ljungqvist, 2012; Wu, 2013), we document weak evidence that information asymmetry decreases (increases) after analysts initiate (terminate) coverage on a stock in univariate test. However, our multivariate regression analysis reveals that changes in information asymmetry do not exhibit a consistent association with the initiation and termination period returns. The results suggest that analysts’ role in reducing information asymmetry may be overstated without properly controlling for the contemporaneous changes in investor recognition. 1 The anticipation hypothesis does not apply to the termination sample because these terminations are due to exogenous reasons such as brokerage mergers or closures.

K.K. Li, H. You / Journal of Accounting and Economics 60 (2015) 141–160

143

Third, we contribute to the literature on analysts’ role in monitoring managers. Several recent studies show that analyst coverage helps to monitor managers (e.g., Chen et al., 2014; Jung et al., 2012; Yu, 2008). We find that the analysts’ monitoring fails to generate any discernable improvement in fundamental performance, and that the market reaction to analyst coverage changes is uncorrelated with changes in fundamental performance. Overall, our evidence suggests that the monitoring benefits of analyst coverage may be economically insignificant. Fourth, we contribute to the debate on whether analysts provide information to investors. Several recent studies argue that prior research may overestimate the market impact of analyst forecasts and recommendations (e.g., Loh and Stulz, 2011), and that analysts’ research contains little information after controlling for confounding events (e.g., Altınkılıç and Hansen, 2009; Altınkılıç et al., 2013; Kim and Song, forthcoming). We find that initiations are not associated with future fundamental performance and that they do not significantly reduce information asymmetry. These findings suggest that analysts’ initial reports provide little new fundamental information for the average investors. The rest of the paper is structured as follows. Section 2 reviews the literature. Section 3 presents the data and sample selection procedures. Section 4 presents the main empirical results. Section 5 discusses and analyzes the alternative explanations of the results, and Section 6 concludes.

2. Related literature and research question Despite the voluminous literature on financial analysts (see the survey papers by Bradshaw 2011 and Ramnath et al. 2008), evidence on how analysts create value for the firms they cover remains limited. Prior research documents a positive market reaction to analyst coverage initiations (e.g., Branson et al., 1998; Demiroglu and Ryngaert, 2010; Irvine, 2003) and a negative reaction to exogenous coverage terminations due to brokerage mergers or closures (Kelly and Ljungqvist, 2012). The combined evidence is consistent with the notion that analysts can add significant value to covered firms. Valuation theories prescribe that the value of a firm equals the present value of expected future cash flows. For an analyst to increase firm value, his coverage should lead to either an improvement in future cash flows, a reduction in the cost of capital, or both (e.g., Campbell, 1991; Campbell and Shiller, 1988). Evidence on whether analyst coverage improves firms' fundamental performance is mixed. On the one hand, analysts may monitor managers (e.g., Chen et al., 2014; Jung et al., 2012; Yu, 2008). If the monitoring effects are economically significant, analyst coverage can improve future operating performance. On the other hand, analyst following may put excessive pressure on managers and induce them to myopically focus on boosting short-term performance at the expense of long-term value (He and Tian, 2013). Francis and Philbrick (1993) show that analysts have incentives to please managers so that they can receive preferential disclosures of private information. Hence, a priori, it is unclear whether analysts can monitor managers effectively and improve firms’ performance. The second potential channel of analysts’ value creation is to reduce the cost of capital by decreasing information asymmetry. Some research suggests that analysts provide useful information to the market (e.g., Bradley et al., 2014; Fried and Givoly, 1982; Gleason and Lee, 2003), which helps reduce information asymmetry among investors and lower the cost of capital (Bowen et al., 2008; Kelly and Ljungqvist, 2012; Wu, 2013). However, other studies show that analysts’ forecasts or recommendations are rather inaccurate, as analysts often bias them for economic incentives, such as securing underwriting business and boosting trading volume (e.g., Bradshaw, 2004; Bradshaw et al., 2012, 2013; Cowen et al., 2006; Irvine, 2004; Jackson, 2005; Lin and McNichols, 1998; Michaely and Womack, 1999; Niehaus and Zhang, 2010). Recent research also argues that prior studies may overestimate the information content or the market impact of analyst forecasts and recommendations (Loh and Stulz, 2011), as they fail to control for the confounding earnings announcement or other news (Altınkılıç and Hansen, 2009; Altınkılıç et al., 2013; Kim and Song, forthcoming). Finally, Irvine et al. (2007) and Juergens and Lindsey (2009) argue that, even when analysts can produce new information, they may distribute their private information to a select group of investors, which could exacerbate the information asymmetry of the stocks (Chung et al., 1995). Hence, it is also unclear whether analysts create significant value by reducing information asymmetry for the firms they cover. Another important, but less studied, channel for analysts to reduce the cost of capital is by improving the investor recognition of stocks. Merton (1987) predicts that higher investor recognition leads to a lower cost of capital and a higher stock price. Under the assumption that investors can only hold stocks that they know, stocks with less investor recognition are held by fewer investors. Due to a lack of demand, these stocks trade at relatively lower prices in equilibrium in order for the market to clear. Building on Merton (1987), a large body of empirical studies documents that investor recognition is negatively associated with the cost of capital (e.g., Bodnaruk and Ostberg, 2009; Foerster and Karolyi, 1999; Kadlec and McConnell, 1994; Lehavy and Sloan, 2008; Richardson et al., 2012). Given Merton’s theory, an analyst should be able to increase firm values if his coverage increases the investor recognition of the stocks.2 One of the primary functions of sell-side analysts is to promote securities to investors. Analysts’ compensation and career success are closely tied to their ability to sell securities in brokerage and investment banking businesses (e.g., 2 It is also worth emphasizing that investor recognition and information asymmetry are related but distinct effects. More investors knowing about a stock does not necessarily reduce the information asymmetry of the stock. For example, if the new investors who get to know the stock are informed traders, the information asymmetry of the stock may actually increase after more informed traders enter the market.

144

K.K. Li, H. You / Journal of Accounting and Economics 60 (2015) 141–160

Groysberg et al., 2011; Hong and Kubik, 2003; Juergens and Lindsey, 2009; Niehaus and Zhang, 2010). To facilitate sales, analysts frequently distribute research reports to investors (Asquith et al., 2005), communicate with them directly, and arrange meetings between corporate executives and potential investors. By constantly directing investors’ attention to a firm, these activities may considerably improve investor recognition of the stock, thus leading to a lower cost of capital and a higher stock price.3 Consistent with this prediction, a recent study by Mola et al. (2013) shows that, after a complete loss of analyst coverage, the number of institutions holding the stock decreases.4

3. Sample selection and variable measurement We employ two samples of change in analyst coverage: coverage initiation and exogenous coverage termination due to brokerage mergers or closures. Both samples have merits and limitations. Because an analyst’s decision to initiate coverage is endogenous by nature (e.g., McNichols and O’Brien, 1997; O’Brien and Bhushan, 1990), the analyses using the initiation sample are subject to potential endogeneity bias. In contrast, analyst coverage terminations due to brokerage mergers or closures are largely exogenous shocks to the affected firms. However, the coverage termination sample also has some limitations. First, its size is much smaller than that of the initiation sample. Second, after the termination of coverage, analysts from other brokerage firms often initiate coverage on the stocks.5 Hence the effects of the exogenous decrease in analyst coverage in those stocks are likely short lived due to the regained coverage. This is an important limitation for our research, as our objective is to understand how analysts create value for firms, and it often takes time for analysts to create value through the three potential channels. Because both the initiation and exogenous termination samples have their own limitations, we use both samples in our main analysis. We collect the data of analyst coverage initiation from the I/B/E/S Recommendation Detail File. Based on prior research (e.g., Crawford et al., 2012; Ertimur et al., 2011; Irvine, 2003; Irvine et al., 2007), we define coverage initiation as the first time a brokerage issues a recommendation for a firm as well as the first time an analyst issues a recommendation for the firm. These two conditions ensure that neither a recommendation carried by an analyst from one brokerage to another, nor one transferred from one analyst to another within a brokerage is counted as an initiation. In addition, we require that the recommendation be issued after the first two years of the I/B/E/S recommendation data (i.e., starting in 1996) to exclude recommendations added due to I/B/E/S data addition, after the first 12 months of the firm’s appearance on CRSP to exclude the potentially mechanical initiations for IPO firms, and after the first six months of the brokerage’s or analyst’s appearance on I/B/E/S to exclude initiations due to new brokerages or analysts expanding coverage. Finally, to reduce the effects of confounding events, we require the recommendation be issued without concurrent (same-day) initiations on the same firm by other analysts and without an earnings announcement or management forecast/guidance issued in the five trading days centered on the initiation date (i.e., initiation day  2 to initiation day þ2, hereinafter “the initiation period”). We obtain earnings announcement dates from I/B/E/S and management forecast/guidance dates from the Company Issued Guidelines (CIG) of the First Call Historical Database.6 As summarized in Appendix A, there are 55,428 initiations satisfying all of the seven conditions over the sample period of 1996 to 2012, consisting of recommendations from 7,805 unique analysts for 8,825 unique firms.7 We denote the initiation quarter as quarter t. The exogenous termination sample includes 32 mergers and 22 closures of brokerages identified by Hong and Kacperczyk (2010) and Kelly and Ljungqvist (2012) from 1984 to 2008. We manually search the merger and closure dates on Factiva and cross check these dates with the I/B/E/S stop file. To identify the stocks for which analyst coverage was terminated due to brokerage mergers or closures, we follow the same procedures used by Kelly and Ljungqvist (2012). Specifically, for the 32 mergers, we require the following: the stocks must be covered by the target during the year before the 3 Barber and Odean (2008) provide a related, but slightly different, explanation of why investor recognition affects stock prices. They argue that individual investors tend to be net buyers of attention-grabbing stocks. The resulting buying pressure may temporarily increase stock prices. If analysts can attract investor attention to the stocks they cover, the resulting buying pressure may increase stock prices. Another related explanation is that some investors prefer to delegate their investment decisions to analysts, as they believe that analysts’ stock picks “certify fair pricing.” Consequently, they are more likely to buy stocks that receive coverage. This hypothesis also implies that the demand for a stock increases after coverage initiation. We thank the referee for pointing out this alternative explanation. 4 There are several important distinctions between Mola et al. (2013) and our paper. First, Mola et al. do not test whether a change in investor recognition is associated with firm value. Second, they do not distinguish whether the loss of analyst coverage leads to the decrease in investor recognition or analysts drop coverage in anticipation of the loss of investor interest. Finally, they focus on a sample of very small firms. The median market capitalization for their sample is less than $28 million. Hence it is unclear whether their findings are generalizable. 5 In an untabulated test, we find that the number of analysts following the termination firms drops significantly in the termination month but gradually increases over the next 12 months. At the end of that period, the termination firms have similar coverage as a group of control firms. 6 The CIG data are available from January 1, 1996, to June 30, 2011. Hence we cannot identify and exclude initiations with concurrent management forecasts/guidance after June 30, 2011. As a robustness check, we exclude all initiations after June 30, 2011 and obtain similar results. 7 Consistent with prior literature (e.g., Crawford et al., 2012; Ertimur et al., 2011; Irvine, 2003; Irvine et al., 2007), we define analyst coverage initiation using recommendations. To test the robustness of our results, we also examine initiations defined using forecast data. The two types of initiations overlap significantly. From 1996 to 2012, there are 53,803 forecast initiations satisfying the same seven conditions, consisting of forecasts from 8,110 unique analysts for 8,982 unique firms. Approximately 95% of the firms and 90% of the analysts exist in both initiation samples. We replicate all of the tests using forecast initiations, and the results are similar, both qualitatively and quantitatively. Furthermore, only 4% of the recommendation initiations have prior forecast initiations. When these observations are excluded, the results are essentially the same.

K.K. Li, H. You / Journal of Accounting and Economics 60 (2015) 141–160

145

merger, and the coverage must stop after the merger; the stocks must be covered by the acquirer during the year before the merger, and the coverage must continue after the merger; the stocks must not be terminated by the target before the merger; and the stocks must not be reinstated by the acquirer after the merger.8 For the 22 closures, we require the following: the stocks must be covered by the broker during the year before the closure; they must not be terminated by the broker before the closure; and the coverage must stop after the closure. As summarized in Appendix A, this sample selection process produces 6,549 exogenous terminations. To isolate the effect of coverage changes, we compare the initiation (termination) sample against a benchmark sample that has similar characteristics but without initiation (termination). We construct the control sample for initiations using propensity-score matching (Heckman et al., 1998; Rosenbaum and Rubin, 1983), which helps us control for differences in the observable determinants of initiations and therefore mitigates the selection bias issue.9 For each initiation in quarter t, we select a control firm from the same quarter that does not have an initiation in quarter t, does not announce earnings in the initiation period, and has a propensity score closest to the initiation firm. The matching is done without replacement. We impose the constraint that the control firm be within a distance (i.e., a caliper) of 0.01 of the initiation firm’s propensity score to guarantee similarity of the observable variables between the initiation and control samples. Appendix B provides more details about the propensity-score model and the matching procedure. During the sample period, the propensityscore matching generates 18,424 initiation-control pairs. Table B1 reports the pooled logistic regressions, before and after matching, with z-statistics adjusted for two-way (by firm and quarter) cluster-robust standard errors (e.g., Gow et al., 2010; Petersen, 2008). All of the determinants significantly predict the probability of initiation, except for one variable. After matching, none of the determinants are significant, suggesting that the matching effectively reduces the differences in these observable determinants of initiation between the initiation and control samples. Because the terminations are exogenous, they are hardly predictable. To ensure that the control firms have similar observable determinants of coverage changes as the termination firms, we use the coefficients reported in Table B1 to compute a pseudo-propensity score for the termination firms and the potential control firms. For each termination in quarter t, we then select a matching firm from the same quarter that does not have a termination in quarter t, does not announce earnings in the five trading days centered on the termination date (hereinafter “the termination period”), and has a propensity score closest to the termination firm. The matching for the termination sample is also done without replacement. We require that the control firm be within a caliper of 0.01 of the termination firm’s propensity score. The matching procedure generates a final sample of 4,029 termination-control pairs. The mean values of the determinants and propensity scores for the initiation and termination samples and their respective control samples are reported in Table B2. The differences between the initiation and control samples are statistically insignificant for all but one variable, while the differences between the termination and control samples are all statistically insignificant.10 We measure market reaction to initiation (termination) using the size-adjusted return over the five-trading-day initiation (termination) period (CAR).11 We use analyst forecast surprise (Unexearn) to measure changes in firms’ future fundamental performance. Unexearn is the difference between the first actual annual earnings per share (EPS) announced after the initiation period and the last corresponding analyst one-year-ahead (FY1) consensus EPS forecast issued in quarter t  1, scaled by the stock price at the time of the t  1 consensus. We use changes in the adverse selection component of the bid-ask spread (ΔAdvsel) to measure changes in information asymmetry.12 ΔAdvsel is the mean adverse selection component of the bid-ask spread (Advsel) of quarters tþ1 to tþ4 minus the mean Advsel of quarters t  4 to t 1. Following Hendershott et al. (2011), we measure Advsel using the five-minute price impact of a trade: qt(mt þ 5 min–mt)/mt, where qt is 8 Reinstatement means the acquirer terminated the coverage at the time of the merger and reinstated it later. As Kelly and Ljungqvist (2012) argue, both the termination and reinstatement may be endogenous. 9 For example, a particular endogeneity/selection bias issue is that analysts tend to initiate coverage on stocks with increased investor recognition or decreased information asymmetry in recent periods. By identifying control firms with past changes in investor recognition and information asymmetry similar to those of the initiation firms, the propensity-score matching mitigates the concern that greater future changes in investor recognition and information asymmetry are due to analysts initiating stocks with greater changes in investor recognition and information asymmetry in the recent past. 10 To examine the robustness of the results, we also adopt a simple matching procedure. We pair each initiation (termination) with a control in quarter t, which is a firm from the same industry, and with the number of analysts following closest to the initiation (termination) firm. This simple matching generates 35,004 initiation-control pairs (5,201 termination-control pairs). We obtain similar results using these alternative control samples. 11 To verify the accuracy of the announcement date of I/B/E/S recommendations, we randomly select 50 initiations each year between 1996 and 2012 (850 observations in total, representing 4.6% of our sample). We verify the initiation dates in I/B/E/S by cross checking them against Investext. We find that I/B/E/S date errors are unlikely to affect our results significantly. Specifically, 633 observations have corresponding initiation reports on Investext. Among them, 421 initiation dates match precisely; 177 initiation dates in Investext are within the five trading days centered on the I/B/E/S initiation dates; and 35 initiation dates in Investext fall outside of the five-trading-day window. Hence a five-day return window adequately captures the true initiation date for the majority (94.5%) of the observations. In addition, prior research on analyst initiations (e.g., Irvine, 2003; Irvine et al., 2007) also confirms that the errors in the I/B/E/S initiation dates are not likely to be a significant issue. 12 In the recent literature, the most widely used proxies for information asymmetry include the probability of informed trading (PIN), the bid-ask spread, Amihud’s (2002) illiquidity measure, and the adverse selection component of the bid-ask spread. However, the literature has reached no consensus regarding the best measure of information asymmetry. For example, Duarte and Young (2009) show that PIN is only priced to the extent that it proxies for illiquidity rather than information asymmetry. The bid-ask spread and Amihud’s illiquidity measure capture much more than information asymmetry (Clarke and Shastri, 2000). The adverse selection component of the bid-ask spread, albeit theoretically appealing, appears to be uncorrelated with measures of information uncertainty (Van Ness et al., 2001). Given the inconclusive evidence, we use the adverse selection component of the bid-ask spread constructed by Hendershott et al. (2011) as the primary proxy for information asymmetry and check the robustness of the results using other alternative proxies.

146

K.K. Li, H. You / Journal of Accounting and Economics 60 (2015) 141–160

the buy-sell indicator (þ 1 for buys, 1 for sells), mt is the midpoint prevailing at the time of the tth trade, and mt þ 5 min is the quote midpoint five minutes after the tth trade. The daily Advsel is the average of Advsels of all trades of a stock in a given day, and the quarterly Advsel is the average of the daily Advsels of that stock. We obtain the intraday trade and quote data from the Trade and Quote (TAQ) database. Finally, following Lehavy and Sloan (2008) and Richardson et al. (2012), we use changes in the breadth of institutional ownership (ΔBreadth) to measure changes in investor recognition. ΔBreadth is the mean institutional ownership breadth (Breadth) of quarters tþ1 to t þ4 minus the mean Breadth of quarters t  4 to t 1, where Breadth is the number of 13F filers holding a firm’s stock, divided by the total number of 13F filers. Detailed definitions of all variables are provided in Appendix A.

4. Empirical results 4.1. Changes in analyst coverage, fundamental performance, information asymmetry, and investor recognition We first test whether a change in analyst coverage (initiation or termination) is associated with changes in any of the three value drivers. Panel A of Table 1 reports the mean initiation/termination period return (CAR) and the changes in fundamental performance (Unexearn), information asymmetry (ΔAdvsel), and investor recognition (ΔBreadth) for the initiation and termination samples and their respective control samples. The t-statistics are computed from two-way (by firm and quarter) cluster-robust standard errors. Consistent with prior studies (e.g., Branson et al., 1998; Demiroglu and Ryngaert, 2010; Irvine, 2003), we find that stock markets react positively to analyst coverage initiations. On average, the firms in the initiation sample have a mean sizeadjusted return of 0.787% (t¼10.03) during the five-trading-day initiation period, compared to 0.024% (t¼0.52) for the control firms. The difference in CAR between the two groups is highly significant (t¼8.94). Compared to the control firms, the initiation firms are associated with a lower mean ΔAdvsel (0.079 vs. 0.368) and a higher mean ΔBreadth (0.165 vs. 0.051) during the year after initiations. The differences in ΔAdvsel and ΔBreadth between the two groups are statistically significant at the 5% level (t ¼  2.17) and the 1% level (t¼ 5.26), respectively. In contrast, the difference in Unexearn between the two groups is insignificant (t ¼1.52). The results suggest that in univariate tests, the initiation of analyst coverage is associated with a decrease in information asymmetry and an increase in investor recognition and is not associated with any significant change in fundamental performance. Table 1 Descriptive statistics of initiation/termination returns and changes in fundamental performance, information asymmetry, and investor recognition. Panel A: Summary statistics for the initiation, termination, and control samples N

CAR

Unexearn

ΔAdvsel

ΔBreadth

Initiation

18,424

Control

18,424

0.787nnn (10.03) 0.024 (0.52) 0.763nnn (8.94)

 0.529nnn (  7.03)  0.617nnn (  9.10) 0.088 (1.52)

0.079 (0.19) 0.368 (0.85)  0.289nn (  2.17)

0.165nnn (4.21) 0.051n (1.99) 0.114nnn (5.26)

 0.885nn (  2.45)  0.056 (  0.14)  0.829nnn (  3.27)

 0.461nnn (  6.06)  0.872nnn (  6.94) 0.411nnn (5.59)

1.499 (1.19) 0.837 (0.71) 0.662n (1.90)

 0.332nnn (  3.51)  0.185nnn (  3.63)  0.147nn (  2.70)

Difference

Termination

4,029

Control

4,029

Difference

Panel B: Pearson (above diagonal) and Spearman (below diagonal) correlations Initiation sample (N¼ 18,424) Variable CAR Unexearn ΔAdvsel ΔBreadth

CAR 0.058nnn  0.050nnn 0.118nnn

Unexearn 0.037nnn  0.070nnn 0.280nnn

Termination sample (N¼4,029) ΔAdvsel  0.063nnn  0.074nnn  0.167nnn

ΔBreadth 0.126nnn 0.170nnn  0.158nnn

CAR 0.164nnn  0.134nnn 0.171nnn

Unexearn 0.100nnn  0.107nnn 0.319nnn

ΔAdvsel  0.118nnn  0.120nnn  0.224nnn

ΔBreadth 0.146nnn 0.280nnn  0.194nnn

See Appendix A for variable definitions and sample selection, and Appendix B for the propensity-score matching procedure. Panel A reports the sample size and the pooled mean of the initiation/termination period returns (CAR) and the proxies for the three value creation channels (Unexearn, ΔAdvsel, and ΔBreadth) for the initiation and termination samples and their control samples. Panel B reports Pearson (above the diagonal) and Spearman (below the diagonal) correlations for the initiation and termination samples. The numbers in parentheses are t-statistics adjusted for two-way cluster-robust standard errors (clustered by firm and quarter). nnn, nn, and n denote two-tailed significance at the 0.01, 0.05, and 0.10 levels, respectively.

K.K. Li, H. You / Journal of Accounting and Economics 60 (2015) 141–160

147

The results for the termination sample are largely consistent with those of the initiation sample. Compared to the control firms, the termination firms have a lower mean CAR (0.885% vs.  0.056%), a higher mean ΔAdvsel (1.499 vs. 0.837), and a lower mean ΔBreadth (0.332 vs.  0.185), with the differences significant at the 1%, 10%, and 5% levels, respectively. Surprisingly, the termination firms also report a significantly higher mean Unexearn than the control firms ( 0.461 vs.  0.872). The significant decrease in institutional ownership breadth contrasts interestingly with the finding of Kelly and Ljungqvist (2012) that the percentage of institutional ownership increases after the exogenous coverage termination. The seemingly contradictory results are consistent with Merton’s investor recognition theory. Exogenous termination of analyst coverage reduces the recognition of the stock for both institutional and retail investors. Hence the number of both institutions and retail investors holding the stock declines after coverage termination. However, the effect is more significant for retail investors, as they arguably have fewer information sources about the firm than institutional investors. The remaining investors who still know about the stock—more likely the institutional investors—have to hold more of the shares, thus leading to an increase in the percentage of institutional ownership. In equilibrium, these investors would demand higher returns to hold a disproportionate fraction of the stock, thereby resulting in a higher expected rate of return and a lower stock price. Panel B of Table 1 reports Pearson (above diagonal) and Spearman (below diagonal) correlations between CAR and the three value drivers. In both the initiation and termination samples, CAR is positively associated with Unexearn and ΔBreadth and is negatively associated with ΔAdvsel. Furthermore, the three value drivers are also significantly correlated. For example, in the initiation sample, ΔAdvsel is negatively correlated with Unexearn (Pearson:  0.074; Spearman:  0.070) and ΔBreadth (Pearson:  0.158; Spearman:  0.167). The significant correlations among the value drivers highlight the importance of simultaneously analyzing the three potential value creation channels, as focusing on only one of them without properly controlling for the others may lead to false inferences. Next, we turn to regression analysis to provide more direct evidence on whether any of the three value creation channels drives the market reaction to coverage initiation/termination. In Table 2, Panel A, the first set of columns reports the univariate regressions of CAR on Unexearn, ΔAdvsel, and ΔBreadth, respectively, for the initiation firms. CAR is positively associated with Unexearn (t¼2.20) and ΔBreadth (t¼12.69) and is negatively associated with ΔAdvsel (t ¼  4.86). The Table 2 Explanatory power of the three value drivers on initiation/termination returns. Panel A: Regression of initiation/termination returns on the three value drivers Initiation sample (N¼ 18,424) Intercept Unexearn

0.823nnn (10.99) 0.087nn (2.20)

0.791nnn (10.21)

 0.053nnn (  4.86)

ΔAdvsel ΔBreadth

Termination sample (N¼ 4,029) 0.674nnn (9.08)

0.684nnn (12.69)

Recom

Adj. R2

0.1%

0.4%

3.471nnn (20.79) 0.025 (0.69)  0.034nnn (  3.50) 0.510nnn (9.13)  1.280nnn (  19.28)

 0.638n (  1.97) 0.536nn (2.76)

4.9%

0.7%

1.6%

 0.726nnn (  2.99)

 0.190nnn (  3.52)

0.9%

 0.516 (  1.50)

1.110nnn (6.66)

 0.336 (  1.62) 0.283 (1.52)  0.145nnn (  2.79) 0.892nnn (5.45)

2.1%

3.1%

Panel B: Comparison of the explanatory power of the three value drivers on initiation/termination returns between the initiation/termination and control samples

Intercept Unexearn ΔAdvsel ΔBreadth

N Adj. R2

Initiation

Control

Difference

Termination

Control

Difference

0.698nnn (10.14) 0.032 (0.89)  0.036nnn (  3.65) 0.634nnn (12.01)

0.020 (0.48) 0.022 (0.93)  0.016nnn (  2.85) 0.455nnn (7.90)

0.678nnn (8.61) 0.010 (0.30)  0.020n (  1.80) 0.179nnn (2.74)

 0.336 (  1.62) 0.283 (1.52)  0.145nnn (  2.79) 0.892nnn (5.45)

0.046 (0.17) 0.176 (1.46)  0.080nnn (  3.65) 0.255nn (2.40)

 0.382nn (  2.24) 0.107 (0.96)  0.065n (  1.91) 0.637nnn (4.09)

18,424 1.8%

18,424 0.7%

–

4,029 3.1%

4,029 1.1%

–

See Appendix A for variable definitions and sample selection, and Appendix B for the propensity-score matching procedure. Panel A reports the OLS regressions of the initiation/termination period returns (CAR) on the proxies for the three value creation channels (Unexearn, ΔAdvsel, and ΔBreadth) and initiation recommendations for the initiation sample, and the OLS regressions of CAR on Unexearn, ΔAdvsel, and ΔBreadth for the termination sample. Panel B reports the OLS regressions of CAR on Unexearn, ΔAdvsel, and ΔBreadth. The regressions are separately estimated for the initiation and termination samples and their control samples. The numbers in parentheses are t-statistics adjusted for two-way cluster-robust standard errors (clustered by firm and quarter). nnn, nn, and n denote two-tailed significance at the 0.01, 0.05, and 0.10 levels, respectively.

148

K.K. Li, H. You / Journal of Accounting and Economics 60 (2015) 141–160

adjusted R2 of the regression on ΔBreadth is 1.6%, which is significantly higher than the adjusted R2 of the regressions on Unexearn and ΔAdvsel (0.1% and 0.4%, respectively). The results indicate that ΔBreadth has much higher explanatory power over CAR than either ΔAdvsel or Unexearn. In addition, Panel A also reports the multivariate regression of CAR on Unexearn, ΔAdvsel, ΔBreadth, and Recom, the level of the initiation recommendations coded by I/B/E/S. In the multivariate regression, ΔAdvsel and ΔBreadth still exhibit significant associations with the initiation period return, while the coefficient of Unexearn becomes insignificant. The second set of columns in Panel A reports the results for the termination sample. In the multivariate regression, we do not control for the recommendation level before terminations because it does not contain new information and should not be associated with the market reaction to terminations. Overall, the results for the termination sample corroborate the results for the initiation sample. Changes in fundamental performance, information asymmetry, and investor recognition are value relevant. Such information may be anticipated by investors and gradually reflected in stock prices. Consequently, the stock returns of any random days may be associated with the changes in these value drivers. The results in Panel A of Table 2 may simply reflect this normal association instead of the market reacting to the effects of coverage changes. To isolate the effects of the changes in coverage, we examine the associations between CAR and Unexearn, ΔAdvsel, and ΔBreadth after controlling for their normal associations. We estimate the normal associations using the control sample by regressing the control firms’ CAR on their Unexearn, ΔAdvsel, and ΔBreadth. Because there is no change in coverage during the return window for the control sample, the associations between CAR and Unexearn, ΔAdvsel, and ΔBreadth estimated in this group should approximate the normal relation between short-window returns and the value drivers.13 The first set of columns in Panel B of Table 2 reports the results of the initiation sample. Among the control firms, CAR is also not associated with Unexearn (t¼0.93), is negatively associated with ΔAdvsel (t¼  2.85), and is positively associated with ΔBreadth (t ¼7.90). These results confirm the existence of a normal association between short-window stock returns and the value drivers. They also highlight the importance of controlling for such an association. Otherwise, one may falsely attribute the normal association to the effect of coverage changes. The third column compares the regression coefficients between the initiation sample and the control sample. The difference in the coefficients of ΔBreadth between the two samples is significant at the 1% level with a tstatistics of 2.74. In contrast, the difference in the coefficients of Unexearn is statistically insignificant (t¼ 0.30) and the difference in the coefficients of ΔAdvsel is only marginally significant at the 10% level (t¼ 1.80). The rest of Panel B reports the regression results of the termination sample. Similar to the results of the initiation sample, ΔBreadth is highly significantly associated with the termination period returns after controlling for its normal associations with CAR. In contrast, the coefficient of Unexearn is statistically insignificant, and the coefficient of ΔAdvsel is only marginally significant. In summary, the results in Table 2 show that, among the three potential value drivers, changes in investor recognition have the highest explanatory power on the market reaction to coverage changes, and this explanatory power remains significant after controlling for the normal association between changes in investor recognition and short-window returns.14 In contrast, the association between the market reaction and changes in information asymmetry is only marginally significant after controlling for the normal association between changes in information asymmetry and shortwindow returns. Finally, the results do not support the hypothesis that analysts create value through improving fundamental performance as the association between the market reaction and changes in fundamental performance is indistinguishable between the initiation (termination) and control samples. 4.2. Alternative model specifications The above results are based on Unexearn, ΔAdvsel, and ΔBreadth, which are noisy ex post proxies for the market expectation of future changes in fundamental performance, information asymmetry, and investor recognition. In this section, we examine the robustness of the results to alternative proxies. 4.2.1. Longer measurement windows In the above tests, we focus on the changes in fundamental performance, information asymmetry, and investor recognition over the first year after coverage changes, which may not fully capture the effects of coverage changes if these effects take a longer time to emerge. To examine this possibility, we expand the measurement window of Unexearn, ΔAdvsel, and ΔBreadth to two and three years after coverage changes. In the first column of Table 3, Panel A, we report the regression of CAR on Unexearn, ΔAdvsel, and ΔBreadth over two years after coverage changes (denoted as Unexearn2, ΔAdvsel2, and ΔBreadth2, respectively) for the initiation sample. Similar to the results in the first column of Table 2, Panel B, CAR is not 13 In addition, we also use stock returns over the five randomly selected non-event trading days within quarter t for each initiation (termination) as a self-control benchmark to estimate the normal associations. To select the random trading days, we first exclude all of the five-trading-day windows centered on the announcement dates of the analyst recommendations for the firm (from all analysts, hence including the initiation date), the firm’s earnings announcement date, the announcement dates of the analyst EPS forecasts for the firm (from all analysts), or the management forecast/guidance issuance dates. We then randomly select five trading days from the remaining days in quarter t. We obtain similar results using this alternative approach. 14 In untabulated tests, we find that the more positive market reaction to initiations by star analysts documented by Branson et al. (1998) is driven primarily by increases in investor recognition. The evidence suggests that the market expects star analysts to trigger a larger increase in investor recognition through initiations, given their prominent status and hence reacts more favorably to their initiations. Similarly, we find that exogenous terminations by star analysts are associated with a more negative market reaction and a larger decrease in investor recognition.

K.K. Li, H. You / Journal of Accounting and Economics 60 (2015) 141–160

149

Table 3 Initiation/termination return and changes in fundamental performance, information asymmetry, and investor recognition over longer horizons. Panel A: Two years after initiation/termination Initiation Intercept Unexearn2 ΔAdvsel2 ΔBreadth2

N Adj. R2

nnn

Control nn

Difference nnn

Termination nn

Control

Difference

0.750 (10.03) 0.025 (1.63)  0.020nn (  2.58) 0.305nnn (7.75)

0.138 (2.48) 0.013 (1.03)  0.010 (  1.57) 0.206nnn (5.69)

0.612 (6.91) 0.012 (0.67)  0.010 (  1.02) 0.099nn (2.13)

 0.644 (  2.14) 0.094nn (2.69)  0.057n (  1.74) 0.420nnn (4.53)

0.007 (0.02) 0.035 (0.96)  0.030 (  1.48) 0.004 (0.07)

 0.651nn (  2.53) 0.059 (1.57)  0.027 (  1.01) 0.416nnn (4.25)

14,591 1.1%

14,591 0.4%

–

3,344 2.1%

3,344 0.3%

–

Panel B: Three years after initiation/termination

Intercept Unexearn3 ΔAdvsel3 ΔBreadth3

N Adj. R2

Initiation

Control

Difference

Termination

Control

Difference

0.695nnn (7.27)  0.006 (  0.72)  0.012 (  1.47) 0.254nnn (7.22)

0.161nnn (2.68) 0.004 (0.52)  0.012 (  1.63) 0.143nnn (4.26)

0.534nnn (4.94)  0.010 (  1.08) 0.000 (0.03) 0.111nn (2.53)

 0.745nn (  2.24) 0.021 (1.31)  0.054 (  1.23) 0.321nnn (3.52)

 0.091 (  0.25) 0.028n (1.80) 0.015 (1.02) 0.028 (0.62)

 0.654nnn (  3.73)  0.007 (  0.34)  0.069 (  1.39) 0.293nnn (3.31)

12,028 0.8%

12,028 0.3%

–

3,196 1.4%

3,196 0.2%

–

See Appendix A for variable definitions and sample selection. This table reports the OLS regressions of initiation/termination period returns (CAR) on the three value creation proxies (Unexearn, ΔAdvsel, and ΔBreadth) over two and three years after initiation/termination. The regressions are separately estimated for the initiation and termination samples and their control samples. The numbers in parentheses are t-statistics adjusted for two-way clusterrobust standard errors (clustered by firm and quarter). nnn, nn, and n denote two-tailed significance at the 0.01, 0.05, and 0.10 levels, respectively.

associated with Unexearn2 (t¼ 1.63), is negatively associated with ΔAdvsel2 (t¼  2.58), and is positively associated with ΔBreadth2 (t¼7.75). In the second column, we observe that the CAR of the control firms is positively associated with ΔBreadth2 (t ¼5.69), while the coefficients of Unexearn2 and ΔAdvsel2 are insignificant (t¼1.03 and  1.57, respectively). After controlling for the normal associations between firms’ returns and the three long-term proxies, only ΔBreadth2 is significantly associated with the abnormal return over the initiation period, as shown in the third column. The results of the termination sample resemble those of the initiation sample, showing that only ΔBreadth2 is still significantly associated with termination period returns after controlling for the normal associations between CAR and the three value drivers. Panel B of Table 3 shows similar results when we extend our analysis to three years after coverage changes. Overall, the results in Table 3 support the hypothesis that analysts create value for firms via improving investor recognition and do not support the hypotheses that analysts create value via reducing information asymmetry or improving fundamental performance. 4.2.2. The ex ante market expectation proxies In the previous analyses, we use realized Unexearn, ΔAdvsel, and ΔBreadth as proxies for the market expectation of the changes in future fundamental performance, information asymmetry, and investor recognition generated by coverage changes. However, the realized values are not observable to the market at the time of changes in coverage. In this section, we investigate whether the market reaction to coverage changes is associated with the ex ante proxies for the market expectation, measured by the mean Unexearn, ΔAdvsel, and ΔBreadth triggered by an analyst’s prior initiations (denoted as PUnexearn, PΔAdvsel, and PΔBreadth, respectively). The first column in Panel A of Table 4 reports the regression results of CAR on PUnexearn, PΔAdvsel, and PΔBreadth for the initiation sample. The coefficient of PΔBreadth is significantly positive (t¼4.54), while the coefficients of PUnexearn and PΔAdvsel are both insignificant (t¼0.44 and 0.99, respectively). The second column reports the results for the termination sample. The coefficient of PΔBreadth is negative and significant at the 10% level (t¼ 1.83), while the coefficients of PUnexearn and PΔAdvsel are both insignificant (t¼0.74 and 0.35, respectively). The results show that investors react more positively (negatively) to coverage initiated (terminated) by analysts whose prior initiations triggered a larger increase in investor recognition, thereby corroborating early evidence that investors view a potential change in investor recognition as one of the most important effects of coverage changes. Once again, the evidence fails to support the hypotheses that analysts create value via reducing information asymmetry or improving fundamental performance.

150

K.K. Li, H. You / Journal of Accounting and Economics 60 (2015) 141–160

Table 4 Alternative proxies for the market expectation of changes in fundamental performance, information asymmetry, and investor recognition. Panel A: Ex ante market expectation proxies Initiation sample

Termination sample

nnn

Intercept PUnexearn PΔAdvsel PΔBreadth

N Adj. R2

0.669 (9.25) 0.011 (0.44) 0.013 (0.99) 0.425nnn (4.54)

 0.432 (  0.81) 0.700 (0.74) 0.052 (0.35)  1.073n (  1.83)

15,422 0.3%

3,333 0.1%

Panel B: Use changes in ROA as the proxy for changes in fundamental performance

Intercept ΔROA ΔAdvsel ΔBreadth

N Adj. R2

Initiation

Control

Difference

Termination

Control

Difference

0.679nnn (9.10) 0.011n (1.84)  0.036nnn (  3.44) 0.625nnn (11.86)

0.013 (0.28) 0.007 (1.55)  0.016nnn (  2.70) 0.404nnn (7.84)

0.666nnn (8.00) 0.004 (0.55)  0.020n (  1.69) 0.221nnn (3.37)

 0.445n (  2.01) 0.011 (0.64)  0.153nnn (  3.71) 0.940nnn (6.16)

0.140 (0.38) 0.001 (0.07)  0.081nnn (  3.77) 0.224 (1.61)

 0.585nn (  2.33) 0.010 (0.41)  0.072nn (  2.06) 0.716nnn (4.35)

18,151 1.8%

18,151 0.8%

–

4,029 2.9%

4,029 1.1%

–

Panel C: Use changes in PIN as the proxy for changes in information asymmetry

Intercept Unexearn ΔPIN ΔBreadth

N Adj. R2

Initiation

Control

Difference

Termination

Control

Difference

0.618nnn (6.64) 0.005 (0.15)  0.069nnn (  3.52) 0.628nnn (10.87)

 0.033 (  0.62) 0.016 (0.83)  0.016 (  1.59) 0.465nnn (7.83)

0.651nnn (6.03)  0.011 (  0.29)  0.053nn (  2.46) 0.163nn (2.16)

 0.413 (  1.25) 0.284 (1.49)  0.115n (  1.90) 0.946nnn (5.30)

 0.115 (  0.37) 0.187 (1.45)  0.033 (  0.92) 0.325nnn (3.16)

 0.298 (  1.51) 0.097 (0.34)  0.082 (  1.55) 0.621nnn (3.63)

13,493 1.9%

13,493 0.7%

–

3,825 2.4%

3,825 0.3%

–

Panel D: Use changes in EDGAR searches as the proxy for changes in investor recognition Initiation Intercept Unexearn ΔAdvsel ΔEDGAR

N Adj. R2

Control

Difference

0.636 (3.64) 0.028 (0.40)  0.060 (  1.70) 1.564nn (2.54)

0.074 (0.97)  0.005 (  0.18) 0.001 (0.15) 0.058 (0.71)

0.562nn (2.36) 0.033 (0.63)  0.061 (  1.71) 1.506nn (2.45)

2,690 0.8%

2,690 0.1%



nnn

See Appendix A for variable definitions and sample selection. This table reports the OLS regressions of initiation/termination period returns (CAR) on alternative proxies for the market expectation of changes in fundamental performance, information asymmetry, and investor recognition. The regressions are separately estimated for the initiation and termination samples and their control samples. The numbers in parentheses are t-statistics adjusted for twoway cluster-robust standard errors (clustered by firm and quarter). nnn, nn, and n denote two-tailed significance at the 0.01, 0.05, and 0.10 levels, respectively.

K.K. Li, H. You / Journal of Accounting and Economics 60 (2015) 141–160

151

4.2.3. Alternative proxies for changes in fundamental performance We investigate whether our results are sensitive to alternative ways of measuring changes in fundamental performance. In Panel B of Table 4, we report the results using changes in return on assets (ΔROA) as an alternative proxy for changes in fundamental performance. Similar to the results in Panel B of Table 2, the third and sixth columns show that ΔROA is not associated with the market reaction to coverage changes (t ¼0.55 and 0.41 for the initiation and termination samples, respectively). In contrast, ΔBreadth is significantly associated with the initiation and termination period returns at the 1% level with t-statistics of 3.37 and 4.35, respectively. Finally, the coefficients of ΔAdvsel are significant at the 10% (t¼  1.69) and 5% (t¼ 2.06) levels for the initiation and termination samples, respectively. We obtain similar results (untabulated) using standardized unexpected earnings (see Bernard and Thomas, 1989) or analyst forecast revisions (i.e., changes in consensus forecasts) as alternative proxies for changes in fundamental performance. 4.2.4. Alternative proxies for changes in information asymmetry We also investigate whether our results are sensitive to alternative ways of measuring changes in information asymmetry. In Panel C of Table 4, we report the results using changes in the probability of informed trading (ΔPIN) as an alternative proxy for changes in information asymmetry. Following Brown and Hillegeist (2007), PIN is computed using the Venter and De Jongh (2006) model. The results are generally consistent with those using ΔAdvsel (Table 2, Panel B). The difference in the coefficients of ΔPIN between the initiation and control samples is significant at the 5% level (t ¼  2.46), while the difference in the coefficients of ΔPIN between the termination and control samples is insignificant (t¼ 1.55). More importantly, ΔBreadth continues to be significantly associated with the market reaction to coverage changes, as shown in the third and sixth columns. We obtain similar results (untabulated) using changes in the number of days with zero or missing returns, changes in Amihud’s (2002) illiquidity measure, and changes in the bid-ask spread as alternative proxies for changes in information asymmetry. 4.2.5. Alternative proxy for changes in investor recognition Measuring investor recognition about a firm has been a challenge for academic research because the number of investors who know about a particular security is not directly observable. Following prior research (e.g., Lehavy and Sloan, 2008; Richardson et al., 2012), we use changes in the breadth of institutional ownership as the primary proxy for changes in investor recognition. To verify the robustness of our results, we adopt an alternative proxy—changes in the number of searches on the SEC’s EDGAR website for a firm’s filings (ΔEDGAR). Ceteris paribus, we expect that an increase in investors’ awareness about a stock, i.e., an increase in investor recognition, would lead to an increase in the search for information about the firm on EDGAR. Hence ΔEDGAR serves as a reasonable alternative proxy for changes in investor recognition. The SEC maintains a log file of all activities performed by users on EDGAR. Following Drake et al. (forthcoming), we exclude searches by automated computer programs, identified by a high frequency of search requests (more than five requests per minute or more than 1,000 requests per day from a unique IP address). The intersection of changes in EDGAR search data and our initiation sample consists of 2,690 initiation-control pairs in 2009 and 2010.15 Table 4, Panel D, reports the regressions of CAR on Unexearn, ΔAdvsel, and ΔEDGAR for the initiation and control samples and compares the coefficients between the two groups. The coefficient of ΔEDGAR is positive for the initiation sample (t¼2.54) but insignificant for the control sample (t¼0.71). The difference in the coefficients of ΔEDGAR between the two samples is significant at the 5% level, with a t-statistic of 2.45. These results are consistent with those using ΔBreadth as a proxy for changes in investor recognition. In addition, the association between ΔAdvsel and CAR is insignificant in this sample. To summarize, the association between changes in investor recognition and the market reaction to coverage changes is robust to all of the alternative proxies for the three value drivers and the different model specifications examined in Tables 3 and 4. In contrast, changes in fundamental performance do not exhibit any significant association with initiation or termination period returns. Finally, the association between changes in information asymmetry and the market reaction to coverage changes is not supported in eight out of the eleven analyses in Tables 3 and 4.

5. Alternative explanations 5.1. Value creation hypothesis vs. anticipation hypothesis The results so far suggest that changes in investor recognition are the most significant and robust determinant of the crosssectional variation in the market reaction to coverage changes. The evidence is consistent with the value creation hypothesis— investors react positively (negatively) to initiations (terminations) because they understand that analysts create value for firms by promoting the stocks to more investors. However, the results of the initiation sample are also consistent with an alternative hypothesis that analysts tend to initiate coverage on stocks they expect to have higher investor recognition in the future (the anticipation hypothesis) and initiations are merely a positive information event revealing these expectations. 15 Because the EDGAR search data start from 2008 and the exogenous termination sample ends in January 2008, we cannot use this alternative proxy for the termination sample.

152

K.K. Li, H. You / Journal of Accounting and Economics 60 (2015) 141–160

But the anticipation hypothesis is unlikely to explain the results of the exogenous termination sample. As Kelly and Ljungqvist (2012) show, these terminations are for exogenous reasons unrelated to analysts’ expectations about the future. Thus any market reaction and changes in investor recognition associated with these terminations should not be attributed to analysts’ expectations. Tables 1 and 2 show that investor recognition decreases significantly after exogenous coverage terminations, and changes in investor recognition have significant explanatory power over the termination period returns. These results suggest that changes in analyst coverage lead to changes in investor recognition, rather than merely anticipating future changes in investor recognition. To further distinguish the value creation vs. anticipation hypotheses for the initiation results, we develop three tests based on the initiating analyst's ability to predict the future and his time and effort to promote the stock. 5.1.1. Analysts’ ability to predict the future If the anticipation hypothesis is correct, we expect the magnitude of both the market reaction and changes in investor recognition to be larger for coverage changes by analysts who have a better ability to predict the future. Mikhail et al. (2003) show that analysts may have different abilities to forecast the future because they have different levels of experience. Lee et al. (2012) use the accuracy of the management forecast to proxy for a CEO's ability to anticipate future events. Accordingly, we use an analyst's EPS forecast accuracy (Accuracy) over the 90 days before his coverage change to proxy for his ability to anticipate future events. We sort the sample into quintiles based on Accuracy. The first set of columns in Panel A of Table 5 reports the results for the initiation sample. The mean CAR is 1.133% (t¼7.56) for firms in the bottom quintile of Accuracy and 0.801% (t¼7.45) for firms in the top quintile of Accuracy. The difference between the two extreme quintiles is insignificant (t¼  1.46). In addition, there is no clear pattern in CAR across the Accuracy quintiles, suggesting that CAR is unrelated to Accuracy. Similarly, we find no clear pattern in Unexearn, ΔAdvsel, or ΔBreadth across the Accuracy quintiles. The results show that the initiations by analysts who can better predict the future do not generate higher initiation period returns or larger increases in investor recognition. Thus the evidence fails to support the anticipation hypothesis. For completeness, we also perform the analysis for the termination sample. As expected, the rest of Panel A of Table 5 shows that neither the termination period returns nor the changes in the three value drivers are associated with analysts’ ability to forecast the future. 5.1.2. Analysts’ effort to promote the stock If the value creation hypothesis is correct, we expect the magnitude of both the market reaction and changes in investor recognition to be larger for coverage changes by analysts who exert more effort to promote the stocks. We use the average number of EPS forecasts that the analyst issues for each firm under his coverage over the 90 days before his coverage change as the proxy for his promotion effort (Effort). We sort the sample into quintiles based on Effort. As Panel B of Table 5 shows, the CAR of the initiation sample increases with Effort, ranging from 0.565% (t¼3.93) for firms in the bottom quintile of Effort to 1.017% (t¼8.21) for firms in the top quintile of Effort. The difference between the two extreme quintiles is significant at the 1% level (t¼2.88). In addition, the ΔBreadth of the initiation sample also increases with Effort, ranging from 0.119 (t¼2.60) for the bottom quintile of Effort to 0.210 (t¼4.97) for the top quintile of Effort. The difference (0.091) is statistically significant at the 1% level (t¼2.79). In contrast, there is no clear pattern in Unexearn or ΔAdvsel across the Effort quintiles. The results suggest that the initiations by analysts who exert more promotion effort tend to generate larger increases in investor recognition. Consequently, investors react more favorably to coverage initiated by these analysts. Using the termination sample, we find weaker but consistent results. Specifically, both the CAR and ΔBreadth of the termination sample decrease with Effort. The differences in the CAR and ΔBreadth between the two extreme Effort quintiles are both significant at the 10% level. The evidence indicates that the terminations by analysts who exert more promotion effort tend to generate larger decreases in investor recognition. As a result, investors react more negatively to these terminations. 5.1.3. Time devoted to the stock If the value creation hypothesis is correct, the magnitude of both the market reaction and changes in investor recognition should be larger for coverage changes by analysts who devote more time to promoting the stocks. Ceteris paribus, analysts who cover more stocks may have less time to promote each one. We use the inverse of the number of firms for which the analyst issues recommendations over the 90 days before his coverage change as the proxy for the time (Time) he uses to promote the stocks. We sort the sample into quintiles based on Time. As Panel C of Table 5 shows, the CAR of the initiation sample increases with Time, ranging from 0.180% (t¼1.27) for firms in the bottom quintile of Time to 0.976% (t¼9.10) for firms in the top quintile of Time. The difference between the two extreme quintiles is significant at the 1% level (t¼5.26). In addition, the ΔBreadth of the initiation sample also increases with Time, ranging from 0.063 (t¼1.54) for the bottom quintile of Time to 0.210 (t¼5.70) for the top quintile of Time. In contrast, there is no clear pattern in Unexearn or ΔAdvsel across the Time quintiles. The results suggest that the initiations by analysts who devote more time to promoting the stocks tend to generate larger increases in investor recognition, which in turn trigger more favorable market reactions. Once again, the results of the termination sample are consistent with those of the initiation sample. In sum, the above results show that both changes in investor recognition and market reactions to coverage changes are uncorrelated with analysts’ ability to forecast the future, but are significantly correlated with the time and effort analysts devote to promoting the stocks. The overall evidence supports the hypothesis that coverage changes affect firm value because they affect investor recognition.

K.K. Li, H. You / Journal of Accounting and Economics 60 (2015) 141–160

153

Table 5 The effects of forecast accuracy and the effort and time spent by the analyst to promote the stock. Panel A: Forecast accuracy Portfolio ranking on Accuracy

Initiation sample N

Bottom 2 3 4 Top Top-Bottom

CAR

3,100 1.133nnn (7.56) 3,136 0.756nnn (5.23) 3,143 0.938nnn (7.00) 3,133 0.817nnn (5.61) 3,110 0.801nnn (7.45)  0.332 (  1.46)

Termination sample

Unexearn

ΔAdvsel

ΔBreadth

N

CAR

 0.499nnn (  3.73)  0.562nnn (  5.06)  0.648nnn (  5.54)  0.527nnn (  4.76)  0.441nnn (  5.83) 0.058 (0.41)

0.157 (0.33) 0.133 (0.30)  0.064 (  0.14) 0.033 (0.08) 0.221 (0.54) 0.064 (0.29)

0.204nnn (5.22) 0.183nnn (4.66) 0.203nnn (5.10) 0.203nnn (4.89) 0.160nnn (3.42)  0.044 (  1.42)

731

 0.845 (  1.15)  0.404 (  0.50)  1.482nnn (  3.53)  0.558 (  1.45)  1.548nn (  2.23)  0.703 (  0.99)

716 753 723 711

Unexearn

ΔAdvsel

 0.456nnn 1.857 (  4.03) (1.49)  0.420nnn 0.878 (  3.43) (0.50)  0.549nnn 1.105 (  4.14) (0.94) nnn  0.514 2.066 (  4.13) (1.56)  0.412nnn 1.457 (  4.59) (1.12) 0.044  0.400 (0.32) (  0.43)

ΔBreadth  0.332nnn (  2.83)  0.293nnn (  3.64)  0.409nn (  2.73)  0.293n (  1.98)  0.352nn (  2.24)  0.020 (  0.18)

Panel B: The effort spent by the analyst to promote the stock Portfolio ranking on Effort

Initiation sample N

Bottom

3,561

2

3,819

3

3,687

4

3,692

Top

3,665

Top-Bottom

CAR

Unexearn nnn

0.565 (3.93) 0.796nnn (6.18) 0.617nnn (4.93) 0.932nnn (8.03) 1.017nnn (8.21) 0.452nnn (2.88)

nnn

 0.559 (  6.06)  0.452nnn (  4.51)  0.476nnn (  4.67)  0.620nnn (  5.17)  0.544nnn (  4.70) 0.015 (0.14)

ΔAdvsel 0.272 (0.76)  0.001 (  0.00) 0.096 (0.21) 0.087 (0.19)  0.049 (  0.11)  0.321 (  1.17)

Termination sample ΔBreadth nn

0.119 (2.60) 0.136nn (2.22) 0.169nnn (4.56) 0.191nnn (4.42) 0.210nnn (4.97) 0.091nnn (2.79)

N 801 796 816 815 801

CAR  0.376 (  0.61)  0.909 (  1.23)  0.902 (  1.40)  1.013nn (  2.60)  1.223nn (  2.32)  0.847n (  1.92)

Unexearn nnn

 0.467 (  4.00)  0.435nnn (  2.82)  0.454nnn (  5.41)  0.493nnn (  4.20)  0.453nnn (  5.11) 0.014 (0.12)

ΔAdvsel

ΔBreadth

1.545 (1.09) 1.462 (1.05) 1.547 (1.01) 1.483 (0.98) 1.456 (1.00)  0.089 (  0.23)

 0.193 (  1.30)  0.282nn (  2.62)  0.326nnn (  2.77)  0.437nnn (  4.39)  0.421nnn (  3.96)  0.228n (  2.02)

Panel C: The time spent by the analyst to promote the stock Portfolio ranking on Time

Initiation sample N

Bottom

3,599

2

3,692

3

3,558

4

3,762

Top

3,813

Top-Bottom

CAR 0.180 (1.27) 0.522nnn (4.65) 1.003nnn (8.31) 1.231nnn (7.73) 0.976nnn (9.10) 0.796nnn (5.26)

Unexearn nnn

 0.545 (  4.69)  0.591nnn (  5.35)  0.475nnn (  3.52)  0.509nnn (  5.17)  0.527nnn (  5.93) 0.018 (0.17)

ΔAdvsel 0.257 (0.66) 0.304 (0.72)  0.041 (  0.08)  0.135 (  0.30) 0.017 (0.04)  0.240 (  1.01)

Termination sample ΔBreadth 0.063 (1.54) 0.106n (1.72) 0.209nnn (4.94) 0.234nnn (5.06) 0.210nnn (5.70) 0.147nnn (3.93)

N 818 833 816 785 777

CAR  0.383 (  0.45)  0.696 (  1.15)  0.634nn (  2.56)  0.852 (  1.54)  1.913nn (  2.57)  1.530n (  1.99)

Unexearn nnn

 0.309 (  3.44)  0.609nnn (  5.62)  0.394nnn (  3.25)  0.500nnn (  4.91)  0.492nnn (  6.50)  0.183 (  1.54)

ΔAdvsel

ΔBreadth

1.979 (1.55) 1.522 (1.46) 1.456 (1.02) 1.654 (1.13) 0.858 (0.84)  1.121 (  1.13)

 0.244nn (  2.16)  0.305nnn (  2.88)  0.303nn (  2.46)  0.382nn (  2.70)  0.435nnn (  3.95)  0.191nn (  2.51)

See Appendix A for variable definitions and sample selection. This table reports the pooled mean of initiation/termination period returns (CAR) and the three value creation proxies (Unexearn, ΔAdvsel, and ΔBreadth) for the quintiles formed on the analyst’s forecast accuracy (Panel A), the effort used to promote the stocks (Panel B), and the time devoted to the stocks (Panel C). The numbers in parentheses are t-statistics adjusted for two-way cluster-robust standard errors (clustered by firm and quarter). nnn, nn, and n denote two-tailed significance at the 0.01, 0.05, and 0.10 levels, respectively.

5.1.4. Implications on the information asymmetry and fundamental performance channels In theory, the value creation vs. anticipation argument is also applicable to the information asymmetry channel or the fundamental performance channel. Specifically, the value creation hypothesis predicts that analyst coverage reduces information asymmetry or improves fundamental performance. The anticipation hypothesis says that analysts tend to initiate coverage on stocks for which they anticipate a decrease in information asymmetry or an increase in fundamental performance. Both hypotheses predict a negative (positive) association between initiation period returns and changes in

154

K.K. Li, H. You / Journal of Accounting and Economics 60 (2015) 141–160

Table 6 Control for changes in liquidity and trading volume. Initiation Intercept Unexearn ΔAdvsel ΔBreadth ΔIlliquidity ΔVolume

N Adj. R2

nnn

Control nnn

Difference nn

Termination

Control

Difference

1.297 (5.45) 0.011 (0.39)  0.033nnn (  3.41) 0.544nnn (10.03)  1.604nnn (  5.24) 0.365 (1.66)

0.612 (2.79) 0.011 (0.70)  0.014nn (  2.48) 0.402nnn (7.47)  1.006nnn (  3.92)  0.179 (  0.72)

0.685 (2.33) 0.000 (0.01)  0.019n (  1.80) 0.142nn (2.07)  0.598 (  1.67) 0.544 (1.59)

 1.435 (  1.54) 0.294 (1.70)  0.144nnn (  3.06) 0.894nnn (5.06) 1.461 (0.68) 0.744 (1.20)

0.964 (1.28) 0.185 (1.23)  0.081nnn (  3.38) 0.249nn (2.20)  1.017 (  1.27) 0.811 (1.04)

 2.399 (  1.66) 0.109 (0.99)  0.063n (  1.74) 0.645nnn (4.07) 2.478 (1.17)  0.067 (  0.17)

18,424 2.1%

18,424 0.8%

–

4,029 3.0%

4,029 1.2%

–

See Appendix A for variable definitions and sample selection. This table reports the OLS regressions of initiation/termination period returns (CAR) on the three value creation proxies (Unexearn, ΔAdvsel, ΔBreadth) and changes in illiquidity (ΔIlliquidity) and trading volume (ΔVolume). The regressions are separately estimated for the initiation and termination samples and their control samples. The numbers in parentheses are t-statistics adjusted for two-way cluster-robust standard errors (clustered by firm and quarter). nnn, nn, and n denote two-tailed significance at the 0.01, 0.05, and 0.10 levels, respectively.

information asymmetry (fundamental performance). However, the lack of a robust association between initiation period returns and changes in information asymmetry or fundamental performance suggests that neither hypothesis is supported. Hence we do not attempt to distinguish between the two competing hypotheses for the information asymmetry channel and the fundamental performance channel. 5.2. Changes in liquidity and trading volume as alternative explanations Irvine (2003) argues that analyst coverage initiation enhances competition between informed traders and reduces the asymmetric information component of the bid-ask spread, which in turn improves liquidity. Similar arguments are also made by Roulstone (2003). Irvine shows that the liquidity gain has significant explanatory power over the market reaction to initiations. Because a larger investor base is associated with higher liquidity, a potential concern about our results is that the positive association between changes in institutional ownership breadth and initiation period returns simply reflects the effects of the improved liquidity brought about by coverage initiation. Relatedly, recent studies in the trading literature (e.g., Jackson, 2005; Juergens and Lindsey, 2009) show that analysts use recommendations to boost brokerage trading income. Thus an additional concern may be that the investor recognition effect simply reflects changes in trading volume after initiations. Noting that, in our early analyses, we already include changes in information asymmetry, one of the key drivers of changes in liquidity, as an independent variable, which should alleviate these concerns. Nevertheless, to further address these concerns, we include changes in illiquidity (ΔIlliquidity) and changes in trading volume (ΔVolume) as additional control variables. Consistent with Irvine (2003), the first column in Table 6 shows that ΔIlliquidity is negatively associated with the CAR of the initiation sample (t¼  5.24). However, after controlling for the normal association between changes in illiquidity and stock returns, ΔIlliquidity is no longer associated with the abnormal initiation period returns, as shown by the t-statistic of  1.67 in the third column. Similarly, ΔVolume is not significantly associated with CAR (t ¼1.59). Finally, ΔBreadth is still significantly associated with CAR (t¼2.07) in the predicted directions after controlling for ΔIlliquidity and ΔVolume. We find qualitatively similar results for the termination sample. Overall, the evidence suggests that changes in liquidity and trading volume cannot explain our findings.

6. Conclusion This paper attempts to answer the important question of how financial analysts create value for the firms under their coverage. We tackle this question by examining what drives the market reaction to initiations and exogenous terminations of analyst coverage. We identify three potential channels of analysts’ value creation: improving fundamental performance, reducing information asymmetry, and increasing investor recognition of the stocks. Our analysis shows that among the three potential channels, only changes in investor recognition have significant and robust explanatory power over the crosssectional variation in initiation/termination period returns. We also provide direct evidence that changes in analyst coverage lead to—rather than merely anticipating—significant changes in investor recognition by investigating the dynamics of investor recognition following exogenous terminations of analyst coverage. Furthermore, we document that the magnitude of both changes in investor recognition and the market reaction to coverage changes is uncorrelated with analysts’ ability to forecast the future but tends to be larger when analysts

K.K. Li, H. You / Journal of Accounting and Economics 60 (2015) 141–160

155

devote more time and effort to promoting the stock. Collectively, these results provide compelling evidence that, at least from the perspective of the market, analysts create value for the firms they cover through improving investor recognition, rather than reducing information asymmetry or improving fundamental performance.

Appendix A. Sample selection and variable definition

Sample

Definition

Sample size at each step

Initiation and control

The initiation sample is from 1996 to 2012. We define initiations as: The first time a brokerage issues a recommendation for a firm, and it is also the first time an analyst issues a recommendation for the firm. In addition, we require that: a. the recommendation be issued after the first two years of the I/B/E/S recommendation data (i.e., starting in 1996) and before 2013; b. after the first 12 months of the firm’s appearance on CRSP; c. after the first six months of the broker’s or analyst’s appearance on I/B/E/S; d. without concurrent (same-day) initiation on the same firm by other analysts; e. and without an earnings announcement or management forecast/guidance issued in the five trading days centered on the initiation date (i.e., initiation day  2 to initiation day þ 2, or “the initiation period”). The quarter of the initiation is denoted as quarter t. We pair each initiation with a control firm in quarter t that does not have an initiation in quarter t, does not announce earnings in the initiation period, and has a propensity score closest to the initiation firm. See Appendix B for the propensity-score matching procedure. Final initiation sample with matched controls:

Termination and The termination sample includes 32 mergers and 22 closures of brokerages identified by Hong and Kacperczyk control (2010) and Kelly and Ljungqvist (2012). For the 32 mergers, we require the following: a. the firms must be covered by the target during the year before the merger and the coverage must stop after the merger; b. the firms must be covered by the acquirer during the year before the merger and the coverage must continue afterward; c. the firms must not be terminated by the target before the merger; d. the firms must not be reinstated by the acquirer after the merger. For the 22 closures, we require the following: a. the firms must be covered by the brokerage during the year before the closure and the coverage must stop afterward; b. the firms must not be terminated by the broker before the closure. The quarter of the termination is denoted as quarter t. We pair each termination with a control firm in quarter t that does not have a termination in quarter t, does not announce earnings in the termination period, and has a propensity score closest to the termination firm. See Appendix B for the propensity score matching procedure. Final termination sample with matched controls: Variable

150,415

111,935 88,382 62,481 62,023 55,428

18,424

12,324 4,603 3,520 3,485 4,253 3,064

4,029

Definition

Abnormal return

CAR

For initiation firms, CAR is the size-adjusted return (adjusted by the return of the CRSP sizematched decile portfolio) over the initiation period (i.e., initiation day  2 to initiation day þ 2). For termination firms, CAR is the size-adjusted return over the five trading days around the exogenous termination date. For control firms, CAR is the size-adjusted return over the same initiation or termination period. We measure CAR in percentages (i.e., return  100).

Change in breadth

ΔBreadth, ΔBreadth2, ΔBreadth3, PΔBreadth

ΔBreadth (ΔBreadth2, ΔBreadth3) is the mean institutional ownership breadth (Breadth) of quarters tþ 1 to tþ 4 (t þ5 to tþ 8, tþ 9 to tþ 12), subtracting the mean Breadth of quarters

Unexpected earnings

Unexearn, Unexearn2, Unexearn3, PUnexearn

t  4 to t  1. PΔBreadth is the average ΔBreadth on all of the initiating analyst’s prior initiations. Breadth is measured as the number of 13F filers holding a firm’s stock, divided by the total number of 13F filers. The resulting values are multiplied by 100. Similar to the measurement window of ΔBreadth, Unexearn is the difference between future annual earnings and the expected earnings measured at quarter t  1. Specifically, Unexearn (Unexearn2, Unexearn3) is the difference between the first (second, third) actual annual EPS announced after the initiation period and the last corresponding analyst consensus FY1 (FY2, FY3) EPS forecast issued in quarter t  1, scaled by the stock price at the time of the t  1 consensus. We replace the missing FY2 or FY3 forecast with the implied forecast computed using analyst long-term growth forecast (e.g., FY2¼ FY1  (1þ LTG/100)).

156

K.K. Li, H. You / Journal of Accounting and Economics 60 (2015) 141–160 PUnexearn is the average Unexearn on all of the initiating analyst’s prior initiations. The resulting values are multiplied by 100.

Change in the adverse selection component of the bid-ask spread

ΔAdvsel, ΔAdvsel2, ΔAdvsel3, PΔAdvsel

ΔAdvsel (ΔAdvsel2, ΔAdvsel3) is the mean adverse selection component of the bid-ask spread (Advsel) of quarters tþ 1 to tþ 4 (tþ 5 to tþ 8, tþ 9 to tþ 12), subtracting the mean Advsel of quarters t 4 to t  1. PΔAdvsel is the average ΔAdvsel on all of the initiating analyst’s prior initiations. Following Hendershott et al. (2011), we measure Advsel using the five-minute price impact of a trade: qt(mt þ 5 min–mt)/mt, where qt is the buy-sell indicator ( þ1 for buys,  1 for sells), mt is the midpoint prevailing at the time of the tth trade, and mt þ 5 min is the quote midpoint five minutes after the tth trade. The daily Advsel is the average of Advsels of all trades of a stock in a given day, and the quarterly Advsel is the average of the daily Advsels of that stock. The resulting values are multiplied by 10,000. The intraday trade and quote data are from the Trade and Quote (TAQ) database.

Initiation recommendation

Recom

The level of the initiation recommendation coded by I/B/E/S: 1 for Strong Buy, 2 for Buy, 3 for Hold, 4 for Underperform, and 5 for Sell.

Change in return on assets

ΔROA

ΔROA is the sum of the quarterly income before extraordinary items (IBQ) from quarters tþ 1 to tþ 4, scaled by the total assets (ATQ) of quarters t, subtracting the sum of IBQ from quarters t  4 to t  1, scaled by the ATQ of quarter t  5. The resulting values are multiplied by 100.

Change in the probability of informed trading

ΔPIN

The mean probability of informed trading (PIN) of quarters t þ1 to tþ 4 minus the mean PIN of quarters t  4 to t  1. The resulting values are multiplied by 100. Following Brown and Hillegeist (2007), PIN is computed using the Venter and De Jongh (2006) model. The PIN data are downloaded from Stephen Brown’s website (http://www.rhsmith.umd.edu/faculty/ sbrown/) and are available from 1993 to 2010.

Change in the number of searches on the EDGAR website

ΔEDGAR

The logarithm of the ratio of the number of unique searches on the SEC’s EDGAR website for a firm’s filings (EDGAR) from quarters t þ1 to tþ 4, divided by EDGAR from quarters t  4 to t  1. Following Drake et al. (forthcoming), we exclude searches by automated computer programs, identified by a high frequency of search requests (more than five requests per minute or more than 1,000 requests per day from a unique IP address). The EDGAR search data are available from 2008 to 2011.

Forecast accuracy

Accuracy

Accuracy is computed in the following steps: (1) for each firm in the analyst’s existing portfolio, compute the forecast error, which is the absolute value of the difference between analysts’ one-quarter-ahead forecast (FPI¼ 6) and the actual EPS, divided by the prevailing stock price, on the last review date (REVDATS) before a new initiation (termination), provided that REVDATS is within the 90 days before the initiation (termination) date and that the actual EPS is announced before the initiation (termination) date; (2) compute the average forecast error of all firms in the analyst’s existing portfolio before the initiation (termination); (3) Accuracy is the inverse of the average forecast error. Firms are sorted into quintiles every quarter based on Accuracy.

Analyst’s effort to market the stock

Effort

The average number of EPS forecasts that the initiating (terminating) analyst issues for each firm under his coverage over the 90 days before a new coverage initiation (termination). Firms are sorted into quintiles every quarter based on Effort.

Analyst’s time to market the stock

Time

The inverse of the size of the initiating (terminating) analyst’s portfolio, which is the number of firms for which the analyst issues recommendations over the 90 days before new coverage initiation (termination). Firms are sorted into quintiles every quarter based on Time.

Change in illiquidity

ΔIlliquidity

The mean illiquidity measure (Illiquidity) of quarters t þ1 to tþ 4 minus the mean Illiquidity of quarters t  4 to t  1. Following Amihud (2002), we measure Illiquidity as the quarterly average of the daily ratio of the absolute stock return to its dollar trading volume. We standardize ΔIlliquidity separately for stocks traded on the NYSE/AMEX vs. those traded on NASDAQ to account for the different market microstructures (Atkins and Dyl, 1997). We sort on

ΔIlliquidity within each quarter (separately for NYSE/AMEX and NASDAQ stocks) and

assign percentile ranks to each observation, ranging from 0 (low ΔIlliquidity) to 99 (high ΔIlliquidity). We then standardize the percentiles by dividing them by 99. Change in trading volume

ΔVolume

The mean turnover of quarters tþ 1 to t þ4 minus the mean turnover of quarters t  4 to t  1. Turnover is the quarterly average of the daily ratio of the trading volume, divided by the total shares outstanding. To account for the different market microstructures, we standardize

ΔVolume in the same way as ΔIlliquidity.

K.K. Li, H. You / Journal of Accounting and Economics 60 (2015) 141–160

157

Determinants of initiation in propensity score matching Stock return

Returnt  1

Cumulative stock return over the 12 months before quarter t.

Change in sales

ΔSalet  1

The sum of sales (SALEQ) from quarters t  1 to t  4 minus the sum of SALEQ from quarters t  5 to t  8, divided by the sum of SALEQ from quarters t  5 to t  8. The ratio is multiplied by 100.

Change in equity issuance

ΔEquityt  1

The sum of equity issuance from quarters t  1 to t  4 minus the sum of equity issuance from quarters t  5 to t 8, divided by the market value of equity in quarter t  4. The ratio is multiplied by 100. Quarterly equity issuance is calculated from SSTKY.

Change in institutional ownership

ΔIOt  1

The percent of institutional ownership (Psint) in quarter t  1 minus Psint in quarter t  5. Psint is total shares held by 13F filers, divided by the total shares outstanding.

Change in trading volume

ΔVolumet  1

The mean turnover of quarters t  1 to t  4 minus the mean turnover of quarters t  5 to t  8. Turnover is the quarterly average of the daily ratio of the trading volume, divided by the total shares outstanding. To account for the different market microstructures, we standardize

ΔVolume in the same way as ΔIlliquidity.

Change in absolute forecast error

ΔAfet  1

Absolute forecast error (Afe) in quarter t  1 minus Afe in quarter t 5. Afe is the absolute difference between a firm’s actual EPS and the latest consensus forecast, divided by the price at the time of the consensus.

Change in coverage

ΔCovt  1

The number of recommendations (Numrec) in quarter t 1 minus Numrec in quarter t  5, divided by Numrec in quarter t  5.

Unexpected earnings

Unexearnt  1

The difference between the last actual annual EPS announced before the initiation quarter t and the last corresponding analyst consensus FY1 EPS forecast issued in quarter t  5, scaled by the stock price at the time of the consensus. The resulting values are multiplied by 100.

Change in breadth

ΔBreadtht  1

The mean Breadth of quarters t  1 to t  4 minus the mean Breadth of quarters t  5 to t  8. The resulting values are multiplied by 100.

Change in the adverse selection component of the bid-ask spread

ΔAdvselt  1

The mean Advsel of quarters t  1 to t  4 minus the mean Advsel of quarters t  5 to t  8. The resulting values are multiplied by 10,000.

Appendix B. Control sample selected using propensity score matching We use propensity-score matching (Heckman et al., 1998; Rosenbaum and Rubin, 1983) to select the control sample. For each initiation in quarter t, we select a matching firm from the same quarter that does not have an initiation in quarter t, does not announce earnings in the initiation period, and has a propensity score closest to the initiation firm. The matching is done without replacement. We impose the constraint that the control firm be within a distance (i.e., a caliper) of 0.01 of the initiation firm’s propensity score to guarantee similarity of the observable variables between the initiation and control samples. Initiation represents an addition in analyst coverage. We develop the following model of initiation, based on the determinants of changes in analyst coverage identified by Anantharaman and Zhang (2011) (see Appendix A for variable definitions):16 Initiationt ¼ αþ β1 Returnt−1 þ β2 ΔSalet−1 þ β3 ΔEquityt−1 þβ4 ΔIOt−1 þ β5 ΔVolumet−1 þβ6 ΔAf et−1 þ β7 ΔCovt−1 þβ8 Unexearnt−1 þ β9 ΔBreadtht−1 þ β10 ΔAdvselt−1 þ Industry f ixed ef f ects þYear f ixed ef f ects þ εt

ðB:1Þ

We estimate the pooled logit model on all firms in the I/B/E/S universe with available data from 1996 to 2012. The industry fixed effects are based on I/B/E/S industry sector groupings. The propensity score matching procedure generates a final sample of 18,424 initiation-control pairs over the sample period from 1996 to 2012. Table B1 reports the pooled logistic regressions, before and after matching, with z-statistics adjusted for two-way (by firm and quarter) cluster-robust standard errors. All of the determinants significantly predict the probability of initiation, except for ΔEquity. After matching, none of the determinants are significant, suggesting that the matching effectively reduces the differences in these observable determinants of initiation between the initiation and control samples. 16 Our results are not sensitive to the determinants included in the model. As a robustness check, we include all of the control variables in Anantharaman and Zhang’s model. The results are quantitatively similar.

158

K.K. Li, H. You / Journal of Accounting and Economics 60 (2015) 141–160

Table B1 Logistic regression of initiation on the determinants. Before matching ***

Intercept Returnt  1 ΔSalet  1 ΔEquityt  1 ΔIOt  1 ΔVolumet  1 ΔAfet  1 ΔCovt  1 Unexearnt  1 ΔBreadtht  1 ΔAdvselt  1 Industry fixed effect Year fixed effect N Pseudo R2

After matching

 2.704 (  25.86) 0.168*** (5.50) 0.223*** (9.11) 0.001 (0.94) 0.618*** (5.05) 0.222*** (2.74)  0.646*** (  2.94) 0.111*** (6.19) 0.009*** (5.10) 0.299*** (12.80)  0.011*** (  4.84) Yes Yes 325,048

 0.108 (  1.09)  0.003 (  0.15) 0.147 (1.45) 0.001 (0.86)  0.054 (  0.38) 0.022 (0.30) 0.076 (0.28)  0.007 (  0.24) 0.006 (1.64)  0.015 (  1.74) 0.001 (0.50) Yes Yes 36,848

8.2%

0.2%

Table B2 Mean values of the determinants and propensity score for initiation, termination and control samples.

Propensity Score Returnt  1 ΔSalet  1 ΔEquityt  1 ΔIOt  1 ΔVolumet  1 ΔAfet  1 ΔCovt  1 Unexearnt  1 ΔBreadtht  1 ΔAdvselt  1 N

Initiation

Control

Difference

Termination

Control

Difference

0.102 0.185 0.166 0.290 0.026 0.522 0.001 0.165  0.479 0.173 0.153 18,424

0.102 0.186 0.151 0.166 0.026 0.520 0.000 0.165  0.516 0.172 0.091 18,424

0.000  0.001 0.015*** 0.124 0.000 0.002 0.001 0.000 0.037 0.001 0.062

0.549 0.047 0.215  0.754 0.027 0.495 0.003 0.107 0.086 0.077  0.448 4,029

0.549 0.057 0.224  0.332 0.030 0.488 0.003 0.095  0.049 0.070  0.453 4,029

0.000  0.010  0.009  0.422  0.003 0.007 0.000 0.012 0.135 0.007 0.005

See Appendix A for variable definitions. Table B1 reports the pooled logistic regressions of initiation on its determinants before and after propensity-score matching. The numbers in parentheses are z-statistics adjusted for two-way cluster-robust standard errors (by firm and quarter). Table B2 compares the mean values of the determinants and propensity scores between the initiation/termination and control samples. nnn, nn, and n denote two-tailed significance at the 0.01, 0.05, and 0.10 levels, respectively.

Because the terminations are exogenous (e.g., due to brokerage mergers or closures), they are hardly predictable. To ensure that the control firms have similar observable determinants of coverage changes as the termination firms, we use the coefficients reported in Table B1 to compute a pseudo-propensity score for the termination firms and the potential control firms. For each termination in quarter t, we then select a matching firm from the same quarter that does not have a termination in quarter t, does not announce earnings in the termination period, and has a propensity score closest to the termination firm. The matching for the termination sample is also done without replacement. We require that the control firm be within a caliper of 0.01 of the termination firm’s propensity score. The matching generates a final sample of 4,029 termination-control pairs. The mean values of the determinants and propensity scores for the initiation and termination samples and their respective control samples are given in Table B2. The differences between the initiation and control samples are statistically insignificant for all but one variable (ΔSale), while the differences between the termination and control samples are all statistically insignificant.

K.K. Li, H. You / Journal of Accounting and Economics 60 (2015) 141–160

159

References Altınkılıç, O., Hansen, R.S., 2009. On the information role of stock recommendation revisions. Journal of Accounting and Economics 48, 17–36. Altınkılıç, O., Balashov, V., Hansen, R.S., 2013. Are analysts’ forecasts informative to the general public? Management Science 59, 2550–2565. Amihud, Y., 2002. Illiquidity and stock returns: cross-section and time-series effects. Journal of Financial Markets 5, 31–56. Asquith, P., Mikhail, M., Au, A., 2005. Information content of equity analyst report. Journal of Financial Economics 75, 245–282. Atkins, A., Dyl, E., 1997. Transactions costs and holding periods for common stocks. Journal of Finance 52, 309–325. Anantharaman, D., Zhang, Y., 2011. Cover me: managers’ responses to changes in analyst coverage in the post-regulation FD period. The Accounting Review 86, 1851–1885. Barber, B., Odean, T., 2008. All that glitters: the effect of attention and news on the buying behavior of individual and institutional investors. Review of Financial Studies 21, 785–818. Bernard, V., Thomas, J., 1989. Post-earnings announcement drift: delayed price response or risk premium? Journal of Accounting Research 27, 1–36. Bodnaruk, A., Ostberg, P., 2009. Does investor recognition predict returns? Journal of Financial Economics 91, 208–226. Bowen, R.M., Chen, X., Cheng, Q., 2008. Analyst coverage and the cost of raising equity capital: evidence from underpricing of seasoned equity offerings. Contemporary Accounting Research 25, 657–699. Bradley, D., Clarke, J., Lee, S., Ornthanalai, C., 2014. Are analysts’ recommendations informative? Intraday evidence on the impact of time stamp delays. Journal of Finance 69, 645–673. Bradshaw, M., 2004. How do analysts use their earnings forecasts in generating stock recommendations? The Accounting Review 79, 25–50. Bradshaw, M., 2011. Analysts’ forecasts: what do we know after decades of work? Working paper, Boston College. Available via 〈http://ssrn.com/ abstract¼1880339〉. Bradshaw, M., Brown, L.D., Huang, K., 2013. Do sell-side analysts exhibit differential target price forecasting ability? Review of Accounting Studies 18, 930–955. Bradshaw, M., Drake, M., Myers, J., Myers, L., 2012. A re-examination of analysts’ superiority over time-series forecasts of annual earnings. Review of Accounting Studies 17, 944–968. Branson, B.C., Guffey, D.M., Pagach, D.P., 1998. Information conveyed in announcements of analyst coverage. Contemporary Accounting Research 15, 119–143. Brown, S., Hillegeist, S., 2007. How disclosure quality affects the level of information asymmetry. Review of Accounting Studies 12, 443–477. Campbell, J., 1991. A variance decomposition for stock returns. Economic Journal 101, 157–179. Campbell, J., Shiller, R., 1988. The dividend-price ratio and expectations of future dividends and discount factors. Review of Financial Studies 1, 195–228. Chen, T., Harford, J., Lin, C., 2014. Do analysts matter for governance? Evidence from natural experiments. Journal of Financial Economics 115, 383–410. Chung, K.H., McInish, T.H., Wood, R.A., Wyhowski, D.J., 1995. Production of information, information asymmetry, and the bid-ask spread: empirical evidence from analysts’ forecasts. Journal of Banking and Finance 19, 1025–1046. Clarke, J., Shastri, K., 2000. On information asymmetry metrics. Working paper, Georgia Institute of Technology. Available via 〈http://ssrn.com/ abstract¼251938〉. Cowen, A., Groysberg, B., Healy, P., 2006. Which types of analyst firms are more optimistic? Journal of Accounting and Economics 41, 119–146. Crawford, S., Roulstone, D., So, E., 2012. Analyst initiations of coverage and stock return synchronicity. The Accounting Review 87, 1527–1553. Demiroglu, C., Ryngaert, M., 2010. The first analyst coverage of neglected stocks. Financial Management 39, 555–584. Drake, M., Roulstone, D., Thornock, J., 2015. The determinants and consequences of information acquisition via EDGAR. Contemporary Accounting Research 32, 1128–1161. Duarte, J., Young, L., 2009. Why is PIN priced? Journal of Financial Economics 91, 119–138. Ertimur, Y., Muslu, V., Zhang, F., 2011. Why are recommendations optimistic? Evidence from analysts’ coverage initiation. Review of Accounting Studies 16, 679–718. Foerster, S.R., Karolyi, A.G., 1999. The effects of market segmentation and investor recognition on asset prices: evidence from foreign stocks listing in the United States. Journal of Finance 54, 981–1013. Francis, J., Philbrick, D., 1993. Analysts’ decisions as products of a multi-task environment. Journal of Accounting Research 31, 216–230. Fried, D., Givoly, D., 1982. Financial analysts’ forecasts of earnings: a better surrogate for market expectations. Journal of Accounting and Economics 4, 85–107. Gleason, C., Lee, C., 2003. Analyst forecast revisions and market price discovery. The. Accounting Review 78, 193–225. Gow, I., Ormazabal, G., Taylor, D., 2010. Correcting for cross-sectional and time-series dependence in accounting research. The Accounting Review 85, 483–512. Groysberg, B., Healy, P., Maber, D.A., 2011. What drives sell-side analyst compensation at high-status investment banks? Journal of Accounting Research 49, 969–1000. He, J., Tian, X., 2013. The dark side of analyst coverage: the case of innovation. Journal of Financial Economics 109, 856–878. Heckman, J., Ichimura, H., Todd, P., 1998. Matching as an econometric evaluation estimator. Review of Economic Studies 65, 261–294. Hendershott, T., Jones, C., Menkveld, A., 2011. Does algorithmic trading improve liquidity? Journal of Finance 66, 1–33. Hong, H., Kubik, J., 2003. Analyzing the analysts: career concerns and biased earnings forecasts. Journal of Finance 58, 313–351. Hong, H., Kacperczyk, M., 2010. Competition and bias. Quarterly Journal of Economics 125, 1683–1725. Irvine, P.J., 2003. The incremental impact of analyst initiation of coverage. Journal of Corporate Finance 9, 431–451. Irvine, P.J., 2004. Analysts’ forecasts and brokerage-firm trading. The Accounting Review 79, 125–149. Irvine, P.J., Lipson, M., Puckett, A., 2007. Tipping. Review of Financial Studies 20, 741–768. Jackson, A., 2005. Trade generation, reputation and sell-side analysts. Journal of Finance 60, 673–717. Juergens, J., Lindsey, L., 2009. Getting out early: an analysis of market making activity at the recommending analyst’s firm. Journal of Finance 64, 2327–2359. Jung, B., Sun, K., Yang, S., 2012. Do financial analysts add value by facilitating more effective monitoring of firms’ activities? Journal of Accounting, Auditing and Finance 27, 61–99. Kadlec, G., McConnell, J., 1994. The effect of market segmentation and illiquidity on asset prices: evidence from exchange listings. Journal of Finance 49, 611–636. Kelly, B., Ljungqvist, A., 2012. Testing asymmetric-information asset pricing models. Review of Financial Studies 25, 1366–1413. Kim, Y., Song, M., 2015. Management earnings forecasts and value of analyst forecast revisions. Management Science. (forthcoming). Lee, S., Matsunaga, S., Park, C.W., 2012. Management forecast accuracy and CEO turnover. The Accounting Review 87, 2095–2122. Lehavy, R., Sloan, R., 2008. Investor recognition and stock returns. Review of Accounting Studies 13, 327–361. Lin, H., McNichols, M., 1998. Underwriting relationships, analysts’ earnings forecasts and investment recommendations. Journal of Accounting and Economics 25, 101–127. Loh, R.K., Stulz, R.M., 2011. When are analyst recommendation changes influential? Review of Financial Studies 24, 593–627. McNichols, M., O’Brien, P., 1997. Self-selection and analyst coverage. Journal of Accounting Research 35, 167–199. Merton, R.C., 1987. A simple model of capital market equilibrium with incomplete information. Journal of Finance 42, 483–510. Michaely, R., Womack, K., 1999. Conflict of interest and the credibility of underwriter analyst recommendations. Review of Financial Studies 12, 653–686. Mikhail, M., Walther, B., Willis, R., 2003. The effect of experience on security analyst underreaction. Journal of Accounting and Economics 35, 101–116. Mola, S., Rau, P.R., Khorana, A., 2013. Is there life after the complete loss of analyst coverage? The Accounting Review 88, 667–705. Niehaus, G., Zhang, D., 2010. The impact of sell-side analyst research coverage on an affiliated broker’s market share of trading volume. Journal of Banking and Finance 34, 776–787.

160

K.K. Li, H. You / Journal of Accounting and Economics 60 (2015) 141–160

O’Brien, P., Bhushan, R., 1990. Analyst following and institutional ownership. Journal of Accounting Research 28, 55–76. Petersen, M., 2008. Estimating standard errors in finance panel data sets: comparing approaches. Review of Financial Studies 22, 435–480. Ramnath, S., Rock, S., Shane, P., 2008. The financial analyst forecasting literature: a taxonomy with suggestions for future research. Internal Journal of Forecasting 24, 34–75. Richardson, S., Sloan, R., You, H., 2012. What makes stock prices move? Fundamentals vs. investor recognition. Financial Analyst Journal 68, 30–50. Rosenbaum, P., Rubin, D., 1983. The central role of the propensity score in observational studies for causal effects. Biometrika 70, 41–55. Roulstone, D.T., 2003. Analyst following and market liquidity. Contemporary Accounting Research 20, 551–578. Van Ness, B., Van Ness, R., Warr, R., 2001. How well do adverse selection components measure adverse selection? Financial Management 30, 77–98. Venter, J.H., De Jongh, D., 2006. Extending the EKOP model to estimate the probability of informed trading. Studies in Economics and Econometrics 30, 25–39. Wu, W., 2013. Information asymmetry and insider trading: evidence from exogenous changes in analyst coverage. Working paper, University of Chicago. Available via 〈http://ssrn.com/ abstract¼ 2323537〉. Yu, F., 2008. Analyst coverage and earnings management. Journal of Financial Economics 88, 245–271.