Playing to the Gallery: Corporate Policies and Equity Research Analysts

Playing to the Gallery: Corporate Policies and Equity Research Analysts François Degeorge University of Lugano - Swiss Finance Institute François Derrien HEC Paris Ambrus Kecskés Virginia Tech Sébastien Michenaud Rice University Abstract It is well known that equity analysts choose to cover firms about which they have favorable views. We exploit this fact and use analysts coverage decisions to infer their preferences for corporate policies. Next, we show that firms cater to analysts preferences. Firms exogenously losing analyst coverage subsequently change their corporate policies away from the preferences of the analyst that drops coverage. Our findings suggest that analysts affect corporate policies. The effects are economically significant and are bigger for small firms and firms with low analyst coverage. Our evidence is consistent with firms altering their corporate policies to cater to the preferences whether based on beliefs or skills of equity research analysts. 3/15/2012 We thank Francesco Franzoni, Patrick Gagliardini, Johan Hombert and seminar participants at Imperial College and Rice University for useful comments and suggestions. All errors are our own. 0

Equity research analysts have strong views about the policies of the firms they cover. They express them in their research reports, in their conference calls with investors, in private meetings with investors and managers, and even in the media. This paper investigates, first, whether individual analysts have preferences for particular corporate policies, and, second, whether their preferences influence the corporate policies of the firms that they cover. In a nutshell, we find in the affirmative to both questions. There are several reasons to think that analysts influence corporate policies. First, analysts may be better informed than company management about important industry trends. They are specifically trained to analyze financial information, and they have access to many sources of information, including the management of competitors of the firms they follow. They may also reflect to a large extent the opinions of investors the biggest of which are in regularly contact with sell side analysts. Consistent with this information gathering and processing role of analysts, a large literature suggests that analyst recommendations move stock prices (e.g. Womack (1996) and Jegadeesh et al. (2004)), and consequently they carry weight with company management. Second, analysts do not simply produce information for investors about the firms that they cover; they also monitor these firms, at least indirectly. Yu (2008) finds that increases in analyst coverage are associated with a decrease in earnings management. Analysts may also facilitate monitoring by others. Investor pressure to undertake acquisitions, for example, may be harder to ignore once analysts publicly opine in support of inorganic firm growth over organic growth. Third, analysts typically choose to cover stocks that they like (McNichols and O Brien 1997). The flip side, of course, is that analysts can choose to drop coverage of stocks they dislike. Losing analyst coverage can be quite costly for firms. Kelly and Ljungqvist (2011) 1

find that the loss of an analyst results in a decrease in liquidity and an increase in the cost of capital, which, as Derrien and Kecskés (2011) find, causes a decrease in investment and financing. Therefore, at least for some firms, catering to an analyst s corporate policy preferences may be a reasonable price to pay if doing so ensures that the analyst continues to cover the firm. In undertaking this study, we face two empirical challenges. First, while analysts views are often observed anecdotally and qualitatively, we need to measure them systematically and quantitatively. We exploit the fact that analysts choose to cover firms about which they have positive views (McNichols and O Brien 1997). We interpret an analyst's decision to cover a firm as his approval of the firm s policies. For example, we posit that an analyst who tends to cover firms that undertake many acquisitions (such as Cisco during the 1990s) prefers such firms. These preferences can reflect analysts beliefs or skills: the analyst may believe that undertaking acquisitions is the optimal investment policy for the firm to pursue, or he may better understand firms that undertake acquisitions. As far as the firm is concerned, it does not matter why the analyst likes acquisitions: what matters is that he does like acquisitions and thus it is worthwhile for the firm to undertake them. To identify analyst preferences, we examine how an analyst s coverage decisions are related to eight firm policies: capital expenditures, R&D expenditures, acquisitions expenditures, cash holdings, leverage, sales growth, dividends and share repurchases. We use the word policy in a broad sense. Some policies we consider are decided by firm management (e.g. investment); others are the outcome of a broad array of company decisions and competitive interactions in the product marketplace (e.g. sales growth); still others are a mixture of both (e.g. leverage). We regress each corporate policy on analyst fixed 2

effects, firm characteristics and controls. The coefficients on the analyst fixed effects represent the preferences of each individual analyst for the corporate policy in question. For all eight policies, we find that analyst preferences are persistent over time, and do not appear to be the result of a random process. 1 Our second challenge is to find a counterfactual situation in which analyst views do not influence corporate policies. We posit that an analyst is more likely to influence a firm when he covers it than when he does not. We examine how firms change their policies when an analyst who was covering them disappears from the I/B/E/S database. (We recognize that such analyst disappearances are likely to be endogenous, and we address this issue below.) We find that after an analyst disappears, firms shift their corporate policies away from the preferences of the analyst in question. For instance, firms dial back (ramp up) their acquisitions when the analyst covering them that disappears has a high (low) acquisitions fixed effect. This effect is symmetric for positive and negative analyst fixed effects, is not capturing a broker fixed effect, and is robust to alternative definitions of analyst preferences and specifications. We interpret our finding as consistent with the notion that firms cater to analysts preferences. The impact of analyst fixed effects on firm policies should vary with the relative influence of the analyst. We assume that a single analyst is more important to firms that are covered by few analysts than to firms that are covered by many analysts. For example, if one analyst drops coverage of a firm covered by two analysts, the firm loses half of its coverage; by contrast, if a firm covered by twenty analysts loses one analyst, it loses only five percent 1 The framework we use to estimate analyst fixed effects is similar to Bertrand and Schoar (2003), who study the impact of CEOs styles on firm policies, and Cronqvist and Fahlenbrach (2009), who study the impact of blockholders on firm policies. However, our approach to investigating whether analyst preferences influence corporate policies is quite different econometrically from these authors' approach and is much closer that of Fee, Hadlock and Pierce (2010). 3

of its coverage. The same logic applies to firm size since smaller firms are typically covered by fewer analysts. Consistent with our hypothesis, we find that firms with less analyst coverage and smaller firms respond more strongly with their corporate policies to the disappearance of an analyst. We run several robustness tests of our main results. We restrict our broad sample of analyst disappearances to a subset of disappearances caused by exogenous broker closures and mergers (see Hong and Kacperczyk (2010), Kelly and Ljungqvist (2011), and Derrien and Kecskés (2011)). Our results are also robust to this refinement although some of them are less strong statistically because of the much smaller sample size. To ensure that our results are not driven by mean reversion in corporate policies, we run placebo tests in which we repeat the analysis one year before the actual analyst disappearance date. In these tests, analyst preferences generally do not affect corporate policy changes. In spite of decades of trying, neoclassical theories of corporate finance have had limited success in empirically explaining key corporate policies such as capital structure (e.g. Welch 2004). Several more recent studies provide evidence for non-traditional determinants of corporate policies. For example, Bertrand and Schoar (2003) find that managerial style explains a significant part of the variation in corporate policies; 2 Cronqvist and Fahlenbrach (2008) find that blockholders similarly matter. Our contribution is to identify another influence on firm policies from a hitherto unexplored source: equity research analysts. Another literature finds that analysts' research reports cater to the preferences of corporate managers especially through the issuance of optimistic investment recommendations (see Lin and McNichols (1998) and Michaely and Womack (1999)). Our study provides evidence 2 However, this finding is not uncontroversial (see Fee, Hadlock, and Pierce (2010)). 4

of catering going in the opposite direction: managers cater with their corporate policies to the preferences of the analysts covering their firm. Our study also contributes to the growing literature on catering (Baker and Wurgler (2011)). Examples of corporate policies guided by catering considerations include, among others: investment (Polk and Sapienza (2009)), mergers and acquisitions (Shleifer and Vishny (2003)), dividends (Baker and Wurgler (2004)), and growth vs. profits (Aghion and Stein (2008)). Relative to these articles our study differs in two ways. First, the audience to which firms cater sell-side analysts is firm-specific as opposed to market-wide. Second, in our context, firms cater to financial intermediaries rather than directly to investors. The article is organized as follows. Section I discusses the construction of the sample, the data and the variables used in our analysis. Section II discusses the estimation of analysts fixed effects. Section III studies the influence of analysts on corporate policies. Section IV discusses robustness checks and alternative explanations while section V concludes. I. Sample, Data, and Variables Definition We build our dataset by merging the Compustat Fundamentals Annual database (after excluding utility and financial services firms) with the I/B/E/S unadjusted detailed earnings estimates database for the years 1983 to 2009. A total number of 14,792 analysts have at least one forecast in I/B/E/S over our sample period. We then posit that analyst i covers firm j in year t if analyst i has at least one forecast for firm j in year t, and in such a case, we keep only the latest forecast for that analyst/firm/year. We further exclude analysts who have less than 5

50 firm-year observations, which restricts our sample to a total number of 3,113 analysts. 3 The reason for this restriction is threefold. First, we can estimate analyst fixed effects with higher precision if we observe more coverage decisions for a given analyst. Second, due to the incidental parameter problem the estimator of the fixed effects variables need not converge if the number of analyst fixed effects to estimate grows large as the number of firms grows large. Therefore restricting our sample to the most active analysts also makes sense from an econometric perspective. Finally, the estimation time is an exponential function of the number of fixed effects to estimate. For all those reasons we restrict our analysis to the subset of the most active analysts. 4 [Insert Table 1 about here.] We end up with a database of 625,285 analyst-firm-year observations from 1983 to 2009 to estimate the analyst fixed effects. This database corresponds to an unbalanced panel of 82,649 firm/year observations (3,061 firms per year on average). Table 1 presents summary statistics of this sample. When we examine the change in corporate policy after the drop in coverage by an analyst in Section III, we focus on firms that lose an analyst and ask whether future changes in the firms policies depend on the preferences of this analyst. We first consider firms that were covered by an analyst that exited the I/B/E/S database before 2008, and for each firm, we focus on the fiscal year that follows the disappearance of the analyst. We assume that 3 For example, an analyst that is active in I/B/E/S during five years and follows more than 10 firms per year during those five 5 years will be included in our sample while an analyst that is active during 5 years following strictly less than 10 firms per year on average will be excluded from our sample. The results hold when we change the threshold. 4 Our results are robust to the use of alternative thresholds. 6

when an analyst disappears from the I/B/E/S database he leaves the profession, because he retires or simply loses his job. We then compute the changes in corporate policy for all the firms followed by that analyst before his disappearance. 2,041 analysts working for 287 brokers exit the I/B/E/S database before 2008. These analysts covered a total of 14,568 firms (each of these analysts covered about seven firms on average). In our robustness tests, we repeat the analysis on a subsample of analyst disappearances, which follow brokerage house mergers as in Hong and Kacperczyk (2010) or due to brokerage house closures as in Kelly and Ljungqvist (2011). We follow the same methodology as Derrien and Kecskés (2011) to identify analysts that were fired after a brokerage house merger or closure. We first identify 52 broker closures and mergers, and we identify the analysts who work for these brokers using I/B/E/S. We assume that an analyst working for one of these brokers disappears if he does not issue any earnings forecast in the year following the disappearance of his broker, and we assume that a firm loses an analyst if the analyst who disappears covered the firm before his broker disappearance. In the case of broker mergers, we also require that both brokers covered the firm before they merged, so that the coverage drop is not a choice but is the consequence of redundancy in coverage. In this subsample of coverage drops, we can attribute with little ambiguity the analyst disappearance to a shock (the merger or closure of the broker the analyst is working for) that is independent from any anticipated changes in corporate policies. 5 Because of these strict requirements, this sample of coverage drops contains only 993 observations of firms losing an analyst. 5 See Hong and Kacperczyk (2010), Kelly and Ljungqvist (2011), and Derrien and Kecskés (2011) for a discussion on the exogeneity of this shock with respect to corporate policies. 7

We study eight different corporate policies: capital expenditures, investment in research and development, acquisitions, cash holdings, capital structure decisions in the form of leverage, dividend payments, share repurchases, and sales growth. Appendix 1 provides details on all of our variables including the control variables. II. Estimation of Analyst Preferences We identify analyst preferences by regressing the corporate policy for firm i in period t on analyst coverage dummy variables for all J analysts in period t, analyst activity dummy variables for all J analysts in period t, year-fixed effects, firm-fixed effects, and various controls. Formally we estimate: (1) Policy i,t is one of the eight corporate financial policy variables of interest, α i is the firm fixed effect, η t is the year fixed effect. The analyst coverage dummy variable is equal to 1 if the analyst covers the firm in period t and is equal to 0 if the analyst does not cover the firm in period t, where coverage is defined as the issuance of at least one EPS forecast for the fiscal year for firm i. We collect the analysts fixed effects δ j and interpret them as a preference measure for each analyst j and each policy variable. The analyst activity dummy variable is equal to 1 if the analyst is active during the fiscal year (i.e. the analyst issues at least one EPS forecast for the year for any firm in I/B/E/S) and is equal to 0 otherwise. By including a control for the analyst activity we condition our estimation of the analyst fixed effect on the analyst being active in I/B/E/S during the period. Therefore if the analyst is 8

not active for an extended period of time the coefficient on the coverage dummy variable will not take into account the inactivity period of the analyst to estimate the fixed effect δ j. Our approach to measuring analyst preferences for policies relies on the fact that analysts coverage decisions are endogenous and indicated a positive opinion of the firm on the part of the analyst. McNichols and O Brien (1997) find that analysts choose to cover firms about which they have favorable views. We posit that an analyst s choice of covering a firm represents an endorsement of the firm s policies. Because we estimate analyst preferences in a regression setting, we have an errors-invariable problem that will bias the estimates towards 0 when we study the influence of analyst preferences on corporate policies. This works against us finding any effect of analysts on policies in the data. We resort to a number of strategies to alleviate these concerns that we explain in the next section. Table 2, Panel A presents the distribution of analyst fixed effects for each corporate financial policy. We do not test whether all fixed effects are jointly different from 0 as F-tests typically over reject the null-hypothesis due to a number of econometric issues discussed in Fee, Hadlock and Pierce (2011) 6. [Insert Table 2 about here.] We compare these distributions of analyst fixed effects to the distributions we would obtain if analysts fixed effects were randomly generated. To do so, we randomize coverage decisions and re-estimate analyst fixed effects. We repeat this exercise 100 times, thus 6 Wooldridge (2002, p. 274) points out that using an F-test for testing the significance of a large number of fixed effects relies on the normality distribution of the fixed effects. To illustrate this issue, Fee, Hadlock, and Pierce (2011) use an F-test and reject the null hypothesis in a randomly generated fixed effect dataset that mimics the observed data properties of CEO fixed effects. 9

obtaining the distribution of analyst fixed effects one would obtain if coverage decisions were independent of corporate policies. More precisely, each year and for each analyst, we calculate the number of earnings forecasts the analyst issues in each SIC2, and we allocate firms to the analyst so that the analyst has the same number of earnings forecasts per industry as he has in reality. To minimize processing-time issues, in each of our 100 randomizations of coverage decisions, we focus only on 15% of the analysts we use in our main analysis. Then we compute analyst fixed effects for each of our eight corporate policies and each of our randomized experiments. We then compare the fixed effect distributions obtained when we randomize coverage decisions with the true distributions that appear in Table 2 using a Kolmogorov-Smirnov test. For each corporate policy, the tests reject the null hypothesis of equality of the distribution with very high statistical significance. [Insert Figure 1 about here.] To have a better sense of how simulated fixed effect distributions differ from true fixed effect distributions, we plot these distributions in Figure 1. In each of the eight graphs corresponding to the eight policies we are considering, the true distributions have higher variance than the simulated ones. This indicates that coverage decisions and corporate policies are correlated at the level of individual analysts, and that the sign and magnitude of the correlation vary across analysts more than they would if coverage decisions were independent from corporate policies. In other words, coverage decisions reflect analyst preferences. In Table 2, Panel B, we present the R squared with and without the fixed effects. The model fit increases substantially when we add analyst fixed effects in the regressions. This 10

suggests that analysts fixed effects may explain a large portion of the variation in the corporate financial policy variables that we study. Since our goal is to identify time-invariant analyst preferences, we test whether estimated analyst fixed effects are reasonably consistent over time. To do so, we split our sample period into two equal sub-periods: 1983-1995 and 1996-2009, and we calculate analyst fixed effects separately for each sub-period. We impose that analysts have at least 50 observations in each sub-periods, which reduces our sample size. In Table 2, Panel C, we present correlations between the 1983-1995 and the 1996-2009 fixed effects, as well as bootstrapped p-values for those correlations. The bootstrapped standard errors are computed as follows: we compute the correlation between pre-1996 fixed effects and the fixed effects of randomly selected analysts from the post-1995 period. We repeat this 1,000 times, and report the fraction of cases in which the simulated correlation is higher than the true one. For all eight corporate policies, analyst fixed effects are positively correlated across the two sub-periods. These correlations are statistically positive at the 10% level or better in five cases out of eight. For the three remaining variables (acquisitions, leverage, and share repurchases), correlations are small and not statistically significant. In another analysis of the distribution of the analysts fixed effects we present the correlation across the different types of corporate policies. The results are presented in Table 3. We find that analyst fixed effects on Capex, R& D, Acquisitions, Cash, Sales Growth are all significantly positively correlated with each other. Analyst fixed effects on Leverage is significantly negatively correlated with analyst fixed effects on Capex, Cash, and Share Repurchases and positively correlated with Acquisitions. Analyst fixed effects on Dividends are 11

positively correlated with analyst fixed effects on R&D and negatively correlated with fixed effects on Sales Growth. III. The Influence of Analyst Preferences on Corporate Policies A. Identification of the Influence of Analysts on Corporate Policies We study the influence of financial analysts by looking at the changes in corporate policies following exogenous drops in analyst coverage. We hypothesize that an analyst with a preference for a high value of a given corporate policy induces the firms he covers to increase its value. For example, a firm covered by an analyst with a preference for high Capex tends to increase its Capex. If the analyst drops coverage, the firm will stop catering to his preferences. To be able to establish causality unambiguously we need to be confident that the coverage drop is not driven by the analyst anticipating a change in corporate policies. To do so, we focus on situations in which the analyst disappears, which is at least somewhat exogenous to the corporate policies of the firms he covers. We consider two types of analyst disappearances. First, we consider all analyst disappearances from the I/B/E/S database. These occur when an analyst retires or leaves the profession. These analyst disappearances may also be driven by other factors, including the decision of a broker to stop reporting to I/B/E/S. Even though we do not expect a large number of analyst disappearances to be caused by anticipations of changes in the policies of the firm covered by these analysts, we consider another sample of analyst disappearances for which the exogeneity of the analyst 12

disappearance is unambiguous: analyst disappearances that follow broker mergers or closures. We present the results for this subsample of coverage drops in the robustness checks section of the paper. We estimate the change in corporate policy on the two subsamples described above. Formally, we test the following specification: (2) The change in policy is measured from year t to year t+1 if the analyst drops coverage in year t (i.e., if the analyst either last appears in I/B/E/S in year t or is fired in year t after a brokerage house merger or closure and never appears again in I/B/E/S afterwards). We include year fixed effects β t in the regressions and a number of controls such as the levels and first differences of market-to-book, price earnings ratio, returns on assets, cash holdings, and the levels of cash flows, of the logarithm of total assets and of momentum returns. We also include control variables to mitigate the effect of reversion toward the mean of the dependent variable. The analyst fixed effect may indeed capture analyst coverage during periods of abnormally high (low) levels of the corporate policy. By including the lagged corporate policy and the lagged average industry level of the corporate policy we control for this trend reversal. The variable of interest in the equation is the analyst fixed effect δ j. We either use the estimated fixed effect δ j directly in the specification or construct variants of the analyst fixed effects using discrete or continuous transformations of δ j. These variants use the sign and statistical significance of the estimated analyst fixed effects, or the relative analyst fixed effects. In robustness checks we also estimate model (2) one year before the actual disappearance of the analyst from I/B/E/S to ensure that the change in policy that we 13

identify is really driven by the drop in coverage and not by a trend that would be captured in the first stage estimation of analyst fixed effects. B. Main Results Table 4 presents the tests of specification (2). We find that after a drop in analyst coverage caused by an analysts disappearance from I/B/E/S, corporate policies move away from the preferences on the analyst dropping coverage. The analyst fixed effects for seven out of eight corporate policy variables are significantly negatively correlated at the 1% level with the change in the corporate policies following analysts disappearance from I/B/E/S. We interpret this result as consistent with the firms moving away from the analyst's preferences after the drop in coverage as it no longer needs to cater to the analyst s preferences. The economic magnitudes are relatively important for all eight corporate policies. For instance, a one standard deviation increase (decrease) in the estimated analyst preference towards R&D results in a decrease (increase) in R&D equal to 4% of the standard deviation of R&D. We follow the same approach as above to estimate the economic magnitude of a shock on analyst preferences on the remaining six significant corporate policies. We find that the economic magnitude is equal to 10% of the standard deviation of Acquisitions, 6% of standard deviation of Cash, 6% of the standard deviation of Leverage, 10% of the standard deviation of Sales Growth, 5% of the standard deviation of Dividends, and 10% of the standard deviation of Share repurchases. Analyst preferences for corporate policies are estimates from Equation (1). Using those estimates as regressors in Equation (2) may create an errors-in-variable problem that will bias 14

the estimates in Equation (2) towards 0. This will work against us finding any effect in the data. To alleviate this issue we create a discrete transformation for each one of the eight analyst fixed effects variables. We set the variable Discretized Analyst fixed effect equal to 1 if the Analyst Fixed Effect is positive and has a t-stat superior to +1.65, equal to -1 if the Analyst Fixed Effect is negative and has a t-stat inferior to -1.65, and equal to zero otherwise. We then estimate specification (2) with this new variable instead of Analyst Fixed Effect. Table 5, Panel A presents the results. All coefficients on the Discretized Analyst fixed effect variable for each of the eight corporate policies are negative and significant at the 5% level or better. In this new specification we estimate the economic magnitude of a discrete change from no significant preference - Discretized Analyst fixed effect is equal to 0 - to a positive or negative and significant preference - Discretized Analyst fixed effect is equal to 1 or -1. For Capital expenditures, losing an analyst who likes (dislikes) capex leads to an average decrease (increase) of Capital expenditures of 0.36% of total assets, or 6% of the standard deviation of Capital expenditures. Following the same methodology as above to estimate the economic magnitude of a shock on analyst preferences on the remaining 7 corporate policies, we find that the economic impact of an analyst loss is equal to 13% of the standard deviation of R&D, 19% of the standard deviation of Acquisitions, 11% of the standard deviation of Cash, 13% of the standard deviation of Leverage, 24% of the standard deviation of Sales Growth, 11% of the standard deviation of Dividends, and 16% of the standard deviation of Share repurchases. A possible concern with the results of Table 4 might be that analysts are more likely to disappear in bad economic times, which likely coincide with cuts in many of our policy variables (Capital expenditures, R&D, Acquisitions, Cash, Sales Growth, Dividends, and Share 15

repurchases) and an increase in Leverage. Under this scenario, we might be picking up a spurious correlation between analyst preferences and corporate policy: policies would change after an analyst coverage drop, but not because of it. Rather, both policies and the analyst coverage drops would be caused by the same factor bad economic times. To address this concern we run a new specification in Table 5, Panel B, in which we separate the positive and significant analyst fixed effects (Positive Analyst Fixed Effects) and the negative significant analyst fixed effects (Negative Analyst Fixed Effects). Positive Analyst Fixed Effects is equal to 1 if the Analyst Fixed Effect is positive and has a t-stat larger than +1.65, and is equal to zero otherwise. Negative Analyst Fixed Effects is equal to 1 if the Analyst Fixed Effect is negative and has a t-stat smaller than -1.65, and is equal to zero otherwise. One key prediction of our analysts catering hypothesis is that the influence of the analyst is conditional on the sign of the analyst preference: if the analyst dropping coverage scored high on, say, dividend preference, we would expect a drop in dividends post-coverage drop. If he scored low on dividends, we would expect an increase in dividends post-coverage drop. Our results should be symmetric when we separate the Positive Analyst Fixed Effects and the Negative Analyst Fixed Effects: we should observe negative coefficients on Positive Analyst Fixed Effects and positive coefficients on Negative Analyst Fixed Effects. By contrast, if the results in Table 4 were simply due to a spurious correlation between coverage drops and a cut in policy variables in bad economic times, we should observe the coefficients on Negative Analyst Fixed Effect for all eight variables to be positive and significant. The results of this specification are presented in Table 5, Panel B. They are generally inconsistent with the negative shock interpretation of our results. All coefficients on the Positive analyst fixed effect are negative while all coefficients on the Negative analyst fixed effect are 16

positive. The results seem consistent with our analysts catering hypothesis. However, the effect of analyst preferences does not seem to be perfectly symmetric, and not all variables are significant at conventional levels. The coefficient on the negative fixed effect for Capital Expenditures and Cash is positive but not significant at conventional levels while the coefficient on the positive fixed effect for R&D and Leverage is negative but not significant at conventional levels. One possible interpretation of our results is that catering to analyst preferences may be more costly depending on the direction of the analyst preference. For instance, reducing Capital Expenditures or Cash to comply with analysts that have a negative analyst fixed effect may be more costly than increasing Capital Expenditures or Cash to comply with analysts that have a negative analyst fixed effect with regards to the corporate policy considered. In Table 5, Panel C we run specification (2) using a relative analyst fixed effect. Relative Analyst Fixed Effect measured in year t is the difference between the Analyst Fixed Effect of the analyst that drops coverage of the firm in year t and the average analyst fixed effect of all the other analysts that cover the firm (and do not drop coverage) in year t. This variable measures the distance of the analyst s preference relative to all other analysts that maintain coverage of the firm, and it might better capture the effect of coverage drops on future policies. Indeed, a disappearing analyst may have a positive fixed effect for a given policy, but if all other analysts who cover the firm have an even higher positive fixed effect, the disappearance of the analyst should lead to an increase, not a decrease in the policy. Again, all eight relative preferences on the corporate policy are negatively correlated with the change in the policy, even though the economic effects are somewhat lower than those on absolute analyst preferences (Table 5, Panel A). 17

Another explanation for our results is that our analyst fixed effects are in fact capturing the preferences of the brokers analysts work for. It might be the case that brokers themselves have some preferences, which reflect those of their main clients. This view is consistent with the results of Cronqvist and Fahlenbrach (2008), who find a correlation between individual institutional investors and some specific firm policies. We consider this possible explanation by adding broker fixed effects to our tests. In Table 5, Panel D we report the results of a regression in which we include both a Pure Analyst Fixed Effect and a Broker Fixed Effect. We estimate Broker Fixed Effect by regressing the Analyst Fixed Effect obtained from equation (1) on a set of dummy variable for all brokerage houses identified in I/B/E/S. Formally, the model that we estimate is presented in equation (3) below where δ j,i,t is the analyst fixed effect estimated in (1) for analyst j that covers firm i at time t, and Brokerage House k,i,t is a dummy variable equal to 1 if analyst j covers firm i for brokerage house k at time t and is equal to 0 otherwise. All β k are then collected and set equal to the Broker Fixed Effect while the residuals ε j,i,t are set equal to the Pure Analyst Fixed Effect. (3) We report the results of specification (2) in Table 5, Panel D, where we regress the change of corporate policy variable on the Broker Fixed Effect and the Pure Analyst Fixed Effect. The coefficients on Pure Analyst Fixed Effect are negative and significant at the 1% level for seven out of the eight corporate policy variables. Likewise coefficients on Broker Fixed Effect are negative and significant for six out of the eight corporate policies that we study. Only Capital Expenditures and Dividends do not move away significantly from the brokerage house 18

preferences after the coverage drop. The economic significance of the results for the Pure Analyst Fixed Effect is generally two to five times larger than the economic effect of the Broker Fixed Effect. Taken together, these results suggest that the analyst fixed effects capture preferences of the analysts whether they are based on analysts' skills or their beliefs beyond those of the brokerage houses that employ them. These results are also robust to an alternative computation of the brokerage house fixed effects in which we estimate the Broker Fixed Effect along with the Analyst Fixed Effect in the first stage in specification (1) for all eight corporate policies. C. Cross-Sectional Results We now focus on firm characteristics that may lead firms to be more sensitive to the influence of sell-side analysts. Derrien and Kecskés (2011) find that the adverse effects of coverage drops on firm policies are more pronounced for small firms and firms that are covered by less analysts because the relative influence of analysts is larger for these firms. In our sample, we focus on small firms and firms with low analyst coverage to figure out whether these firms characteristics lead to an increased sensitivity to analyst fixed effects. We run model (2) in which we add an interaction term between a cross sectional variable and the Analyst Fixed Effect, the Analyst Fixed Effect itself as well as the variable of interaction. The results of this analysis are presented in Table 6. The variables that we interact with Analyst Fixed Effect are Small Firms, Smallest Firms, and Log(Analyst). Small Firms is a dummy variable equal to 1 if the firm year observation falls below the median of Total Assets and is equal to 0 otherwise. Smallest Firms is a dummy variable equal to 1 if the firm/year observation falls in the bottom quintile of Total Assets 19

and is equal to 0 otherwise. Log(Analyst) is the logarithm of the lagged number of analysts covering the firm. We find that small firms are more sensitive to the analyst fixed effects in Capital Expenditures, and Cash using Small Firms as the interaction variable. We find that small firms are more sensitive to the analyst fixed effects in R&D, Acquisitions, and Sales Growth using Smallest Firms as the interaction variable. Firms with lower analyst coverage are more sensitive to the analyst fixed effects in Capital Expenditures, R&D, Acquisitions, Cash, and Sales Growth. IV. Robustness Checks and Alternative Explanations We run a number of robustness checks to ensure that our results are not driven by the endogeneity of the analyst decisions to disappear from I/B/E/S with respect to firm corporate policies or by the way we calculate analysts fixed effects. To address the first issue, we test our main specification in a subsample of coverage drop events that we can attribute with certainty to analysts disappearing following brokerage house mergers (as in Hong and Kasperczyk, 2010) or brokerage houses closures (as in Kelly and Ljungqvist, 2011). Derrien and Kecskés (2011) study these events to analyze firms response to an exogenous drop in analyst coverage on corporate investment. We use Derrien and Kecskés (2011) sample. Note that there is no overlap between the period over which we estimate the analyst fixed effects and the period over which we measure the impact of the exogenous drop in coverage conditional on the analyst preference. 20

We present the results of this analysis in Table 7. Despite the much lower number of observations over which we test the main hypothesis the coefficient on the analyst fixed effect is negative and significant for six out of the eight corporate policies. Firms reduce Capital Expenditures, R&D, Acquisitions, Sales Growth, Dividends and Share repurchases after the exogenous drop in coverage of analysts with positive fixed effects for the policy. The economic magnitude of the influence of the analyst preferences is stronger than in the sample of analyst disappearance from I/B/E/S. A one standard deviation increase in the estimated analyst preference towards Capital expenditures results in a decrease in Capital expenditures equal to 9% of the standard deviation of Capital expenditures. We follow the same approach as above to estimate the economic magnitude of a shock on analyst preferences on the remaining seven corporate policies. We find that the economic magnitude is equal to 9% of the standard deviation of R&D, 14% of the standard deviation of Acquisitions, 13% of the standard deviation of Sales Growth, 15% of the standard deviation of Dividends, and 15% of the standard deviation of Share repurchases. A plausible concern is that our results might be driven by mean reversion in corporate policies. If a policy variable takes on high values right before an analyst drops coverage of the firm, it will tend to generate a high estimated analyst fixed effect. After the coverage drop, the policy will tend to revert to the mean but the coverage drop is not the cause of the change in policy. To evaluate the severity of this issue we follow the following strategy. We first run a falsification test in which we run Equation (2) one year before the actual analyst disappearance from I/B/E/S. If mean reversion is driving our results and the mean reversion process has already started before the analyst disappearance from I/B/E/S, then we would expect the coefficients on all eight corporate policies to be negative and significant 21

and to have the same magnitudes as in our main tests presented in Table 4. The results of this falsification test are presented in Table 8. Only the analyst fixed effects on Capital expenditures and Dividends are negative and significant. The coefficients on the analyst fixed effect on Capital expenditures and Dividends are smaller than in the exogenous shock specification in Table 7 (-18 vs -37.12 for Capital expenditures, -17.55 vs. -73.58 for Dividends). We interpret these results as consistent with mean reversion possibly driving some but not all of the negative correlation between the analyst fixed effects and the change in Capital expenditures and Dividends but not meaningfully so for the other six policy variables that we consider. V. Conclusion We exploit the well-known fact that equity research analysts choose to cover firms about which they have positive views, and we use coverage decisions by analysts to infer their preferences for corporate policies. We find that analysts exhibit preferences for policies that are persistent over time, and do not appear to be the result of a random process. Next, we show that firms cater to analysts preferences. Firms exogenously losing analyst coverage subsequently change their corporate policies away from the preferences of the analyst that drops coverage. The influence of analysts on corporate policies is economically significant, and is larger for small firms, and firms with low coverage. We conclude that firms play to the gallery of the analysts following them, perhaps in order to avoid the negative consequences of losing coverage. 22

References Aghion, P. and J. Stein. 2008. Growth versus Margins Destabilizing Consequences of Giving the Stock Market What It Wants. Journal of Finance, 63, (3), 1025-1058. Baker, M. and J. Wurgler. 2004. A Catering Theory of Dividends. Journal of Finance 59, 271-288. Baker, M. and J. Wurgler. 2011. Behavioral Corporate Finance: An Updated Survey. NYU Working Paper. Bertrand, M., and A. Schoar. 2003. Managing with Style: The Effect of Managers on Firm Policies. Quarterly Journal of Economics, 118, 1169-1208. Brennan, M, and A Subrahmanyam. 1995. Investment analysis and price formation in securities markets. Journal of Financial Economics, 38, 361 381. Chang, X., S. Dasgupta, and G. Hilary, 2006, Analyst Coverage and Financing Decisions. The Journal of Finance, 61, 3009 3048. Cronqvist, H. and R. Fahlenbrach. 2008. Large shareholders and corporate policies. Review of Financial Studies, Derrien, F and A Kecskés. 2011. The Real Effects of Financial Shocks: Evidence from Exogenous Changes in Analyst Coverage. HEC Paris Working Paper. Easley, D., M. O'Hara, and J. Paperman. 1998. Financial analysts and information-based trade. Journal of Financial Markets, 1, 175 201. Fee, CE, and CJ Hadlock. 2010. Managers Who Lack Style: Evidence from Exogenous CEO Changes. Michigan State Working Paper. Graham, J, C Harvey, and S Rajgopal. 2005. The Economic Implications of Corporate Financial Reporting. Journal of Accounting and Economics, 40, 3 73. Hong, H. and J. D. Kubik. 2003. Analyzing the Analysts: Career Concerns and Biased Earnings Forecasts. Journal of Finance, 58, 313-351. Jegadeesh, N., J. Kim, S. Krische, and C. Lee. 2004. Analyzing the Analysts: When Do Recommendations Add Value? Journal of Finance, 59, 1083-1124. Jensen, M. 2005. Agency Costs of Overvalued Equity. Financial Management, 34, 5-19. Kelly, B., and A. Ljungqvist. 2011. Testing asymmetric-information asset pricing models. forthcoming Review of Financial Studies. Lin, H., and M. F. McNichols. 1998. Underwriting relationships, analysts' earnings forecasts and investment recommendations. Journal of Accounting and Economics 25, 101-127. 23

McNichols, M. and P. O Brien. 1997. Self-Selection and Analyst Coverage. Journal of Accounting Research, 35, 167-199. Michaely, R., and K. Womack. 1999. Conflict of interest and the credibility of underwriter analyst recommendations. Review of Financial Studies 12, 653-686. Polk, C. and P. Sapienza. 2009. The Stock Market and Corporate Investment: A Test of Catering Theory. Review of Financial Studies, 22, 187-217. Shleifer, A. and R. Vishny. 2003. Stock Market Driven Acquisitions. Journal of Financial Economics, 70, 295-312. Stickel, S.. 1995. The Anatomy of the Performance of Buy and Sell Recommendations. Financial Analysts Journal, 51, (5), 25 39. Welch, I.. 2004. Capital Structure and Stock Returns. Journal of Political Economy, 112, 106-132. Womack, K.. 1996. Do Brokerage Analysts' Recommendations Have Investment Value? The Journal of Finance, 51, 137 167. Yu, F.. 2008. Analyst Coverage and Earnings Management. Journal of Financial Economics, 88, 245-271. 24

Appendix 1 Definition of Main Variables Acquisitions Analyst Fixed Effect Broker Fixed Effect CAPEX Cash Holdings Cash flow Changes in Total Assets Discretized Analyst Fixed Effect Dividends Leverage Market-to-Book ratio Negative Analyst Fixed Effect Past profitability Acquisitions (Compustat item aqc) scaled by start-of-year total assets (item at) Measure of analyst propensity to cover a firm given the level of a corporate policy. We compute the measure separately for each of the eight corporate policies considered (Capital Expenditures, R&D, Acquisitions, Cash, Leverage, Sales Growth, Dividends, Repurchases) by regressing the said policy on a number of control variables (lagged market-to-book, lagged price earnings ratio, lagged ROA, log of lagged Total Assets, momentum returns), a dummy variable equal to 1 if the analyst is active during the year and equal to 0 otherwise (the activity dummy variable), and a dummy variable for each analyst equal to 1 if the analyst covers the firm in the year and equal to 0 otherwise (the coverage dummy variable). The estimated coefficient on the coverage dummy variable obtained from each one of these 8 regression is the analyst fixed effect corresponding to the dependent variable. Difference between Analyst Fixed Effect and Pure Analyst Fixed Effect Capital expenditures (item capx) scaled by start-of-year total assets Cash (item che) scaled by start-of-year total assets (item at) Net income before extraordinary Items (item ib) + depreciation and amortization expenses (item dp) scaled by start-of-year total assets Total Assets (item at) divided by start-of-year Total Assets minus one Equal to -1 if Analyst Fixed Effect is negative with a t-stat inferior to -1.65, equal to +1 if Analyst Fixed Effect is positive and has a t-stat superior to +1.65, and equal to 0 otherwise. Common shares dividend (item dvc) and preferred shares dividend (item dvp) scaled by start-of-year total assets Long term debt (item dltt) plus short term debt (item dlc) scaled by debt and equity capital (item dltt + item dlc + item seq) Market value of equity (Item prcc_f multiplied by Item csho) plus book value of assets minus book value of equity minus deferred taxes (when available) (Item at - Item ceq - Item txdb), scaled by book value of total assets (Item at). Variable is lagged one year Dummy variable equal to 1 if Analyst Fixed Effect is negative with a t-stat inferior to -1.65 and equal to 0 otherwise. Ratio of operating income before depreciation and amortization (item oibdp) to start-of-year total assets. Variable is lagged one year 25

Positive Analyst Fixed Effect Price earnings ratio Pure Analyst Fixed Effect Dummy variable equal to 1 if Analyst Fixed Effect is positive with a t-stat superior to 1.65 and equal to 0 otherwise. Stock price (item prcc) times common shares outstanding (item csho) divided by earnings before interest and taxes (item ebit) Residual of a regression where Analyst Fixed Effect is regressed on dummy variables for each brokerage house in the database equal to 1 if the analyst works for the brokerage house during the year and equal to 0 otherwise. Relative Analyst Fixed Effect Repurchases Research and Development Sales growth Total assets Difference between Analyst Fixed Effects for the analyst j that drops coverage of the firm and the firm's average Analyst Fixed Effect in the year following the drop in coverage by j. Purchase of common and preferred stock (item prstkc) scaled by start-of-year total assets Research and Development Expenses (Item xrd) scaled by start-of-year total assets (Item at) Sales (item sale) divided by end of previous year sales minus one Start-of-year total assets (Item at) (in million USD) 26

Table 1 - Summary Statistics Data are collected from Compustat Industrial database and I/B/E/S. We exclude financial services firms (SIC code 6000-6999), regulated utilities (SIC code 4900) and firms that are not covered by analysts. The Sample: Estimation of Analysts Fixed Effects is the sample in which we estimate the Analyst Fixed Effect. The Sample: Last Year of Analyst Presence in I/B/E/S is the sample in which we estimate the influence of the analysts' fixed effects on corporate policies. It only includes analyst-firm-year observations corresponding to the last year of presence in I/B/E/S before the analyst stops covering firms in the database. All variables are described in Appendix 1. Panel A Sample: Estimation of Analyst Fixed Effects Sample N Mean Median Std. Dev. Total Assets 81,658 3,770 335 15,528 Number of Analysts 82,649 7.57 5.00 7.56 Capex 75,336 8.08 4.80 11.21 R&D 82,649 5.61 0.00 12.84 Acquisitions 72,789 3.49 0.00 10.77 Cash 81,440 25.19 8.90 54.06 Leverage 81,487 33.59 30.07 34.00 Sales growth 80,742 22.85 10.87 62.41 Dividends 81,253 1.10 0.00 2.93 Repurchases 69,668 1.70 0.00 4.32 Market to Book 81,806 2.02 1.39 2.64 ROA 80,613 8.74 12.10 32.58 Cash Flows 79,153 3.08 7.96 37.90