Changes in Analyst Coverage: Does the Stock Market Overreact?

Transcription

1 Changes in Analyst Coverage: Does the Stock Market Overreact? AMBRUS KECSKÉS and KENT L. WOMACK * Preliminary Version 1.0, October 19, 2006 ABSTRACT A sell-side analyst s decision to add or drop coverage of a firm typically reflects better or worse operating performance, respectively, both in the year of the decision and the next year. But, the stock market overreacts to analysts coverage decisions. When the number of analysts following a firm increases, future returns are lower, and, conversely, when the number of analysts decreases, future returns are higher. The decrease-increase return spread is 6.4 percentage points. The overreaction is more pronounced when changes in analyst following are confirmed by changes in analysts consensus recommendations or changes in institutional ownership, and the overreaction depends on valuation levels. * Kecskés is from the Joseph L. Rotman School of Management, University of Toronto. Womack is from the Amos Tuck School of Business, Dartmouth College. 0

2 1. Introduction There are numerous reasons for brokerage analysts, so-called sell-side analysts, to add or drop research coverage of stocks in the industries they follow. Analysts compensation and career prospects are closely tied to high standings in the annual polls conducted by Institutional Investor and The Wall Street Journal. Within their industry, analysts strive to identify up-and-coming firms as well as to cover the large capitalization firms that are of interest to most institutional investors. Analysts also tend to cover firms with solid stock market and operating performance. 1 Indeed, Jegadeesh, Kim, Krische, and Lee (2004) show that analysts tilt their recommendations towards glamour firms, i.e., firms with past market outperformance, high market valuations, high trading volume, etc. Another motivation for analysts coverage decisions is the generation of trading commissions stemming from the information content in analysts earnings estimates and investment recommendations. For example, Irvine (2003) finds that initiations of analyst coverage are followed by increases in liquidity. Analysts are also expected to assist their bank s investment banking division by covering firms with possible securities issuance needs and/or mergers and acquisitions prospects. Analysts are supposed to stimulate investor interest in these firms, which generates lucrative banking fees if it leads to the analyst s bank being chosen to intermediate these deals. As Krigman, Shaw, and Womack (2001) report from a survey of firms that switch underwriters, 88% of executives cite research coverage as one of the top three reasons for switching. Moreover, Michaely and Womack (1999) show that pressure from investment bankers may be sufficiently great as to bias analysts investment 1 See Bhushan (1989) and O Brien and Bhushan (1990). 1

3 recommendations in favor of firms that their bank has recently taken public, particularly when these firms have performed poorly. In sum, analysts coverage decisions seem to cater to the market s demand for information as well as to generating trading commissions and investment banking fees. Analysts decisions to drop coverage of certain stocks stem from related causes. Analysts have finite resources and so they generally cannot cover all of the firms in their industry group. For example, Boni and Womack (2006) show that the typical analyst covers 10 firms even though the typical industry has 177 firms. Therefore, an analyst s decision to add coverage of one firm tends to imply a decision to drop another firm. Firms for which analysts drop coverage have poorer past stock market and operating performance and demand fewer investment banking services. Generally speaking, analysts will add coverage of stocks about which they are bullish and drop coverage of stock about which they are bearish (e.g., McNichols and O Brien (1997)). This study examines how the market reacts to analysts decisions to add and drop coverage. Does firm value increase when analysts add coverage and vice versa? Does the market efficiently impound the information contained in the change in coverage or is there a subsequent return drift or reversal? We measure changes in analyst following and returns over calendar years and find that excess returns are higher in the year in which the number of analysts following a stock increases, and excess returns are lower in the year in which analyst following decreases. Surprisingly, however, in the year after a change in analyst following, excess returns are negative (-1.7 percentage points) when following has increased and positive (-4.7 percentage points) when following has decreased. In 2

4 other words, the market appears to overreact in the year of a change in analyst following and reverses itself the next year. This finding stands up to a battery of robustness tests. To put our findings in a rational theoretical framework, we consider the investor recognition model of Merton (1987). Suppose that the cost of capital for a firm is decreasing in the number of investors who include the firm in their portfolio choice problem. If investor recognition increases, the cost of capital falls, and so realized returns are higher than expected today but expected returns are lower going forward. Our results are consistent with this investor recognition model. Nevertheless, investor recognition cannot be the whole story. Changes in investor recognition impact the cost of capital but they should not systematically impact cash flows. However, we find that firms for which analysts add coverage have significantly better operating performance than firms for which analysts drop coverage. This performance difference persists beyond the year of the change in coverage. Moreover, when increases in analyst following are confirmed by changes in analysts consensus recommendations, operating performance is even better than when changes in analyst following are contradicted. The converse is also true for decreases in analyst following. In other words, it appears that analysts add (drop) coverage of firms that have better (worse) operating performance, both in the year of the change in coverage and the next year. The puzzle, then, is that excess returns reverse in the year after a change in analyst coverage. Whether for rational or behavioral reasons, the bottom line is that firm value increases (decrease) when analyst following increases (decreases). Further results suggest that the market gets carried away by the information value of changes in analyst 3

5 following, resulting in mispricing that is subsequently reversed. If the market does overreact, then, for a given change in analyst following (an increase, no change, or decrease), firms with higher valuations (e.g., as measured by book-to-market) should have lower excess returns next year relative to firms with lower valuations. Indeed, extreme glamour firms for which analyst following increases have the worst excess returns the next year. Extreme value firms for which analyst following increases have much better excess returns but their excess returns are only slightly worse than for glamour firms for which analyst following decreases. Extreme value firms for which analyst following decreases have the highest excess returns the next year. The evidence from the glamour-value distinction lends support to the market overreaction explanation of analysts decisions to add or drop coverage. The rest of this paper is organized as follows. Section 2 outlines the sample selection and data sources. Section 3 describes the sample. Section 4 presents the main results. Section 5 examines explanations of the main result. Section 6 presents robustness tests. Section 7 concludes. 2. Sample Selection and Data Sources This study primarily examines the time-series variation of the number of analysts following firms (providing earnings forecasts to IBES). Typically, an analyst will at least provide one-year-ahead earnings estimates on the firms he follows whether or not he provides anything else. Accordingly, we assume that an analyst follows a firm if he provides a one-year-ahead earnings estimate on that firm. We extract the number of analysts providing earnings estimates ( analyst following ) from the monthly I/B/E/S Summary History Summary Statistics file every December for every firm. Most of our 4

6 returns and operating performance statistics are for calendar years unless otherwise stated. We only retain a firm-year observation for which there is at least one analyst following the firm in two consecutive years. This restriction filters out potentially large changes in firms information environment resulting from changes in analyst following of some to zero analysts and from zero to some analysts. Since I/B/E/S earnings estimates coverage becomes comprehensive in 1983 and since our analysis requires one year of future returns data, we compute the change in analyst following for every year between 1984 and 2004 inclusive. We extract the mean recommendation for every firm from the monthly I/B/E/S Recommendations Summary Statistics file for the same time period. We use these data to study the impact of a change in analyst following that is confirmed or contradicted by analysts consensus recommendation changes. We merge our extracts from the estimates summary and recommendations summary files. We extract the return, closing price, shares outstanding, volume, exchange code, and share code from the CRSP monthly stock file for every firm, for every month. As is common practice, we retain only firms with share codes 10, 11, and 12 (operating companies). We match the remaining CRSP firms to our I/B/E/S firms, and we retain only firms that have both I/B/E/S earnings estimates data and CRSP data. Our final sample consists of a maximum of 66,627 firm-years and 10,619 firms with available earnings estimates and a maximum of 38,358 firm-years and 7,946 firms with both earnings estimates and recommendations. Additionally, we extract from CRSP the valueweighted index, which we use as our market index. We also extract the first date of listing for every firm as well as a list of S&P 500 constituents from CRSP. 5

7 We study how changes in analyst following relate to firm characteristics, market and operating performance, financing and investment activity, and valuations. Therefore, we extract operating performance data from the Compustat industrial annual and quarterly files. We extract annual data on total assets (item #6), sales (item #12), book value of equity (item #60), capital expenditures (item #128), and income before extraordinary items (item #237). We extract quarterly data on sales (item #2), depreciation and amortization (item #5), income before extraordinary items (item #8), and book value of equity (item #59). We obtain yearly NYSE capitalization decile breakpoints from Ken French s website. 2 We extract from Securities Data Company equity offering dates and M&A transaction dates. We extract from the CDA/Spectrum Institutional (13f) Holdings database, at the fourth quarter of every calendar year, for every firm, the number of institutional shareholders as well as the total number of institutions in the database. 3. Sample Description We begin by examining the characteristics of firms that analysts follow and why changes in analyst following occur. Therefore, we first examine analyst following conditional upon market capitalization and institutional ownership for a typical year, We sort all sample firms into deciles based on analyst following, based on NYSE market capitalization decile breakpoint, and based on the number of institutional shareholders of the stock. Panels A and B of Table I show separately for all firms as well as only for S&P 500 firms, respectively, the well-known strong positive relation between analyst following and market capitalization. Panel C shows that there is also a strong 2 6

8 positive relation between analyst following and institutional ownership and Panel D shows that there is a strong positive relation between market capitalization and institutional ownership. These relationships are a consequence of the fact that analysts follow firms for which they can generate commission and banking revenues. The results for other years are similar. In summary, Table I shows that both institutions and analysts follow larger firms. [Insert Table I about here] Next, we examine how the number of analysts following a firm changes from year to year. Figure 1-A shows the mean market capitalization of firms followed by analysts and of S&P 500 firms. The typical firm followed by analysts is about a quarter the size of the typical S&P 500 firm throughout the years. Figure 1-B shows the distribution of analyst following (for firms followed by at least one analyst). This distribution is stable over the years. Mean (median) following is about 6.5 (4) analysts and very few firms are followed by more than 20 analysts. [Insert Figure 1 about here] We now turn to the distribution of changes in analyst following from year to year. Figure 2-A shows the distribution of changes in analyst following and Figure 2-B shows the relative percentage of the changes in analyst following (increases, no changes, and decreases). The distribution of changes in analyst following is stable over the years, with the typical change being close to zero. However, this stability belies the substantial fluctuations in aggregate analyst following. The percentage of firms with no changes is fairly stable from year to year, between 24% and 32% of changes. By contrast, increases and decrease fluctuate substantially from year to year at each other s expense. 7

9 Oftentimes, increases (decreases) occur in years of high (low) market returns (not tabulated). Armed with a sketch of the dynamics of analyst following and the change in analyst following, we now turn to examining the determinants of the change in analyst following for a firm from year to year. [Insert Figure 2 about here] We rely primarily on the literature on analyst following to guide our choice of the determinants of changes in analyst following. Determinants include: Market capitalization decile change: Market capitalization is a well-known determinant of analyst following (e.g., Bhushan (1989)). Change in institutional breadth: We define change in institutional breadth for a given stock in a given year as the change in the number of 13f filers that hold that stock between last year and this year, all divided by the total number of 13f filers (for all stocks) last year. Thus change in institutional breadth is simply scaled changes in institutional holdings. Scaling is necessary because the number of institutions in our data source increases several fold during our sample period. O Brien and Bhushan (1990) find that the number of institutions is positively related to the number of analysts. We emphasize that throughout the paper, except in Panel A of Table IV, we take change in institutional breadth at face value as change in institutional ownership. Change in turnover percentile: Greater analyst following may be a profitable activity if it generates more trading and hence commissions for banks. Every year, for every exchange, we sort all firms in CRSP with a full year of monthly turnover data into turnover percentiles. We do so separately for the NYSE, 8

10 AMEX, and NASDAQ because volume measurement differs substantially between exchanges. Raw stock return and market return: O Brien and Bhushan (1990) find that an increase in analyst following is associated with higher excess of market returns, and at the yearly frequency aggregate analyst following is positively related to the level of the market. Accordingly, we use both the raw stock return this year and last year and the market return. Change in book-to-market: Book-to-market measures valuation. Higher valuation may capture greater growth opportunities and/or greater mispricing, but in both cases information production by analysts may be valuable. Equity issuance dummy and acquirer dummy: These measure whether the firm has issued equity and acquired another firm, respectively, this year. A firm may be more likely to choose as underwriter a particular bank if an influential analyst at that bank follows the firm. The equity issuance dummy variable equals one if a given firm in a given year issues equity according to SDC and zero otherwise. The acquirer dummy variable equals one if a given firm in a given year acquires another firm according to SDC and zero otherwise. Change in return on equity: As already noted, O Brien and Bhushan (1990) find that an increase in analyst following is associated with higher excess of market returns, so we consider the operating performance counterpart of stock return, return on equity. Return on equity is computed as income before extraordinary items this year scaled by the mean of book value of equity this year and last year. 9

11 Change in sales growth: O Brien and Bhushan (1990) find that the number of firms entering an industry is positively related to the number of analysts following firms in that industry. This finding inspires our somewhat simpler choice of sales growth, which is computed for every firm as the growth rate of sales this year versus last year. Change in capital expenditures: As a measure of real investment, capital expenditures may capture growth opportunities that are not captured by sales growth since sales growth is derived from assets in place. Both firms that invest heavily themselves and institutions interested in these firms may benefit from greater analyst following of these firms. Capital expenditures are scaled by the mean of total assets this year and last year. We regress changes in analyst following on contemporaneous determinants listed above. We also calculate the effect of a one standard deviation increase in each of our explanatory variables on change in analyst following. Since our sample consists of a cross-section of firms across time, we implement a firm fixed effects regression. The Appendix describes why we choose firm fixed effects and explains this methodology. Table II presents the results. All of the relations are in the predicted direction, except that book-to-market change, the acquirer dummy, return on equity change, and sales growth change are not statistically significant. The results overall are fairly intuitive. Firms for which size, institutional ownership, turnover, returns, and investment increase also experience an increase in analyst following as is the case for firms that issue equity. Following also increases for firms for which size, valuation, and trading activity increase. Additionally, analyst following increases when the broader market performs better and 10

12 when the firm is involved in financings and acquisitions. The most important determinant of change in analyst following is equity issuance, followed by raw stock return last year, change in institutional breadth, and the market return. Equity issuance is associated with an increase of 0.49 analysts. A one standard deviation increase in the raw stock return last year, change in institutional breadth, and the market return are associated with increases of 0.42, 0.34, and 0.26 analysts, respectively. Changes in analyst following do reflect changes in the market and operating performance of firms. Are changes in analyst following related to future returns? [Insert Table II about here] 4. Main Result We now examine the relation between changes in analyst following and returns. We are mindful of the possibility that an additional analyst may have less of an impact for a firm followed by 10 analysts than for a firm followed by one analyst. As Figure 3 shows, the number of analysts following a firm this year for a firm that was followed by two analysts last year (the 25 th percentile of the distribution of analyst following) ranges roughly from one to six analysts. For a firm followed by four analysts last year (the median of the distribution), the range is wider, roughly one to ten analysts. By sharp contrast, for a firm followed by nine analysts last year (the 75 th percentile of the distribution), the range is very wide, roughly one to 17 analysts. Therefore, in all multiple regressions that involve regressing returns on changes in analyst following, we control for the logarithm of the number of analysts following the firm last year and the interaction between the logarithm of the number of analysts following the firm last year and the change in the number of analysts this year. Furthermore, when we refer to mean 11

13 returns, we mean excess of market returns, and when we refer to returns in regressions, we mean excess of the risk free rate and we include controls for the Fama-French three factors plus momentum. Finally, we always use a firm fixed effects regression when return is a dependent variable. [Insert Figure 3 about here] Table III presents this year s changes in analyst following related to returns last year, this year, and next year. Panel A presents sample means and Panel B presents regressions of returns on change in analyst following. From Panel A, the decreaseincrease return spread is = 6.4 percentage points. From Panel B, an additional analyst this year is associated with incremental returns of 7.0, 1.7, and -4.6 percentage points last year, this year, and next year, respectively. Note that when we express relations between some variable and incremental returns, we hold constant the logarithm of analyst following last year at its mean value of about 1.5 analysts. In other words, (excess return t ) ( Δfollowing ) = + β ( ) 1. 5 / t Δfollowingt Δfollowingt followingt 1 β. ln [Insert Table III about here] Not surprisingly, we find strong evidence that more analysts follow firms that have performed well in the past. Nothing appears to attract market participants like stellar past returns. Nevertheless, it is striking that returns reverse after changes in analyst following. The interested reader can peek ahead at Section 6 where we present a battery of tests that suggest the results in Table III are robust. 5. Explanations 5.1. Investor Recognition 12

14 We have found that changes in analyst following are positively related to returns this year and negatively related to returns next year. There is a rational explanation for this. Suppose that investors only include a firm in their portfolio choice problem if they know about it, and that not all firms are known to all investors. We can think of greater investor recognition as reducing the cost of capital for the firm. Merton (1987) models a capital market equilibrium with incomplete information. His model predicts that (1) changes in investor recognition are positively related to present returns and (2) negatively related to future returns, that (3) the foregoing two relationships are more pronounced for riskier firms, and that (4) both financing and investing are increasing in changes in investor recognition. We assume that analyst following proxies for investor recognition. Accordingly, the results in Table III are consistent with Merton (1987) s predictions (1) and (2). Lehavy and Sloan (2006) also find evidence supporting Merton (1987) s predictions. They use change in institutional breadth as a proxy for investor recognition. 3 To get a sense of how our findings overlap with theirs, we replicate the results in Panel B of Table III but we add change in institutional breadth as an explanatory variable. Panel A of Table IV presents the results. A one standard deviation analyst following increase (about 2.5 analysts) is associated with incremental returns next year of percentage points. 3 Institutional breadth change is a somewhat controversial proxy for investor recognition. First, the number of investors that own a stock is a lower bound of the number of investors that know about the stock, but beyond this it is unclear how the two are related. Second, institutional breadth change may be interpreted as a noisy stock level proxy for mutual fund flows, which are a well-known proxy for investor sentiment. Conceptually at least the rational and behavioral explanations for the reduction in the cost of capital are very different (greater investor recognition versus more optimistic sentiment, respectively). 13

15 By contrast, a one standard deviation increase in change in institutional breadth (about 1.65 percent) is associated with incremental returns next year of -3.8 percentage points. Insofar as change in institutional breadth is also a proxy for investor recognition, change in analyst following and change in institutional breadth are more like complementary rather than competitive proxies for investor recognition change. [Insert Table IV about here] We continue testing Merton (1987) s predictions. To test prediction (3), we repeat our tests for predictions (1) and (2) but we separate changes in analyst following for high and low risk firms. Every year, we sort all firms in CRSP with twelve months of monthly returns data based on their annualized standard deviation of monthly returns. We then classify a given firm in a given year as high (low) risk if its standard deviation is above (below) the median standard deviation that year. Panel B of Table IV presents the results. An additional analyst this year is associated with incremental returns of 2.7 (0.6) percentage points this year for high (low) risk firms, while for next year this association is -5.3 (-3.8) percentage points for high (low) risk firms. Hence the relation between changes in analyst following and present and future returns is more pronounced for high risk firms, which is consistent with prediction (3). To test prediction (4), we test whether changes in analyst following are related to financing and investment, controlling as usual for lagged analyst following and its interaction. We measure financing with our equity issuance dummy variable. We measure investment with our acquirer dummy variable as well as change in capital expenditures. Panel C of Table IV presents the results. Change in analyst following is positively related to financing and investment, which is consistent with prediction (4). 14

16 In summary, we find that changes in analyst following are positively related to present returns and negatively related to future returns, that the foregoing two relationships are more pronounced for riskier firms, and that both financing and investing are increasing in analyst following. If analyst following is a proxy for investor recognition, Merton (1987) s predictions are confirmed Market Overreaction Investor recognition is not the only possible explanation for the return reversal after changes in analyst following. The obvious alternative explanation is that the market overreacts to changes in analyst following and eventually corrects its excesses. This explanation is consistent with the positive relation between change in analyst following and present returns and the negative relation between change in analyst following and future returns being more pronounced for more risky firms. More risky firms are harder to value and to arbitrage so they are more likely to be mispriced. It is also consistent with firms issuing more (less) equity and investing more (less) when following increases (decreases). Investors and managers alike may correctly perceive the direction of the change in the cost of capital induced by the following change but they may overshoot with the magnitude, resulting in too much equity issuance and real investment. We cannot distinguish between the rational and behavioral explanations on the basis of Merton (1987) s predictions alone. Before testing whether or not a behavioral explanation is consistent with our main results, we examine whether the market reaction to changes in analyst following is different when confirmed or contradicted by other votes of confidence. Specifically, we examine whether the magnitude of returns this year and next year is bigger (smaller) 15

17 when change in analyst following this year are confirmed (contradicted) by changes in analysts consensus recommendation this year and by changes in institutional ownership this year. The intuition behind these tests is simple. We have found that increases (decreases) in analyst following are greeted by positive (negative) returns. Womack (1996) finds that recommendation upgrades (downgrades) are greeted by positive (negative) market returns. When more of the analyst community pays attention to a firm and these analysts view the firm more favorably, there should be a greater market reaction than when these analysts view the firm less favorably. The logic for changes in institutional ownership is analogous, provided that the market responds more (less) favorably to institutional ownership increases (decreases). To implement our tests, we separate changes in analyst following into four dummy variables. The first dummy variable is for analyst following increases that are confirmed by more optimistic consensus recommendation changes ( ). The second dummy variable is for analyst following increases that are contradicted by more pessimistic consensus recommendation changes ( ). The third dummy variable is for analyst following decreases that are contradicted by more optimistic consensus recommendation changes ( ). The fourth dummy variable is for analyst following decreases that are confirmed by more pessimistic consensus recommendation changes ( ). We follow the same setup when using changes in institutional ownership. Table V presents mean excess returns this year and next year for increases and decreases in analyst following that are confirmed and contradicted by other relevant signals. First, we examine returns next year. If following increases and is confirmed by a consensus increase, incremental returns next year are -1.6 percentage points compared to 16

18 only -0.6 if contradicted by a consensus decrease. If following decreases and is confirmed by a consensus decrease, incremental returns next year are 8.4 percentage points compared to only 4.3 if contradicted by a consensus increase. If following increases and is confirmed by an ownership increase, incremental returns next year are -1.6 percentage points compared to only -1.2 if contradicted by an ownership decrease. If following decreases and is confirmed by an ownership decrease, incremental returns next year are 7.5 percentage points compared to only 2.2 if contradicted by an ownership increase. The magnitude of the return reversal is bigger (smaller) when change in analyst following are confirmed (contradicted). [Insert Table V about here] Next, we examine returns this year. If following increases (decreases) and is confirmed by a consensus increase (decrease), incremental returns this year are positive (negative). However, if following increases (decreases) and is contradicted by a consensus decrease (increase), incremental returns this year are negative (positive). The results for changes in institutional ownership tell the same story. By contrast to the results for returns next year, it appears that the consensus change effect dominates the following change effect for returns this year. Taken together, the results in Table V suggest that the market reacts more strongly to changes in analyst following that are confirmed by other votes of confidence than to changes that are contradicted. We now turn to distinguishing between rational and behavioral explanations of our main results. We ask whether or not analysts are justified in making their following decisions by fundamentals. Specifically, we examine whether or not analysts add coverage of firms that are better performers on an operating basis and drop coverage of 17

19 worse performers, both contemporaneous with and subsequent to changes in analyst following. As we have already argued, analysts should rationally add coverage of good firms and drop coverage of bad firms. We extend this analysis by separating changes in analysts following according to whether these changes are confirmed or contradicted by changes in analysts consensus recommendation and changes in institutional ownership. We wish to ensure that firms that get a double dollop vote of confidence through greater analyst attention and more glowing analyst accolades are actually better firms. We also look at how changes in analyst following are corroborated by changes in institutional ownership, i.e., by how investors are voting with their dollars. Our measures of operating performance are return on equity, sales growth, and capital expenditures. We first test how these variables this year and next year are explained by changes in analyst following this year. We then test how these variables this year and next year are explained by the same four dummy variables created for Table V based on change in analyst following and analysts consensus recommendation changes. Table VI presents the results. From Panel A, both this year and next year, profitability, growth, and investment are all monotonically higher for analyst following increases. From Panel B, with minor exceptions not only does following change in the direction of operating performance but the results are even more pronounced when we incorporate the additional information provided by consensus recommendation changes. In other words, operating performance is higher for firms for which analyst following increases than decreases and is even higher for firms for which analyst following increases and is confirmed by a consensus recommendations increase. Our findings for growth and investment are consistent with the findings of Jegadeesh, Kim, Krische, and 18

20 Lee (2004), who find that individual analysts tilt their recommendations towards glamour stocks. Our results suggest that analysts gravitate towards (away from) more (less) profitable, faster growing, and more heavily investing firms. [Insert Table VI about here] If analysts follow and recommend good firms as measured by both present and future operating performance, why do these firms turn out to be bad investments? The investor recognition explanation speaks only to changes in the cost of capital and is silent on the subject of changes in cash flows. The answer that suggests itself is that these firms are overvalued. We test the overvaluation explanation by examining whether or not firm valuations at the time of changes in analyst following are related to returns thereafter. For following increases, the return reversal should be exacerbated for firms with higher valuations because they have a longer way to fall. For following decreases, the return reversal should be moderated for firms with higher valuations because they are already richly valued. We use four valuation measures. Following Lakonishok, Shleifer, and Vishny (1994), we use book-to-market, cash flow-to-price, and sales growth, and following Lee and Swaminathan (2000), we use turnover. We form these variables as follows. We form book-to-market ratios using book value from the fiscal quarter ending during the third calendar quarter and market value from the last trading day of the calendar year. We form cash flow-to-price ratios by summing earnings before extraordinary items plus depreciation and amortization for the four fiscal quarters ending during the third calendar quarter and dividing the total by market value from the last trading day of the calendar year. We only use strictly positive cash flow-to-price ratios. We form sales growth using 19

21 sales from the fiscal quarter ending during the third calendar quarter of the current year compared to five years prior and we calculate the geometric mean growth rate. We form turnover as the mean of the monthly turnover ratios (total monthly volume divided by month end shares outstanding) during the calendar year. Next, for each of these four variables, every year, we sort all firms in Compustat into quintiles. Finally, for each of these four variables, we sort our sample firms into three groups of change in analyst following (increases, no changes, and decreases) by five valuation quintiles. For every cell, we calculate the mean excess of market return. Table VII reports the results. Before interpreting them, several pointers are in order. First, there are typically at least a few thousand observations in each cell, and any concentration in particular cells does not appear to be systematic within or across panels. Second, the mean return for each change in analyst following group is different in Panels A and D from Panels B and C because there are considerably fewer observations in the latter than the former panels. Third, in Panel D, since Lee and Swaminathan (2000) find that high turnover stocks have glamour characteristics, the progression of quintiles is reversed, running from five (glamour) to one (value). [Insert Table VII about here] We now interpret the results. For every valuation measure, for a given valuation quintile, returns are lower for following increases than for no changes and lower for no changes than for decreases. The analyst following decrease-increase spread (about five percentage points) is always statistically significant. In other words, for a given valuation level, returns reverse after following changes, which is consistent with our previous results. Moreover, in Panels A, B, and D, eight of the nine extreme value-glamour 20

22 spreads (about five percentage points) are statistically significant. (For comparison, without conditioning on change in analyst following, the t-statistics for the value-glamour premium for Panels A, B, C, and D are 5.68, 5.71, -0.29, and 8.41, respectively.) It is not surprising that results in Panel C are weak because by construction we only include firms that have at least six years of sales data, which eliminates a lot of young and growing firms, and we thus lose about 40 percent of our sample firms. The results indicate that for a given change in analyst following, returns next year are lower when valuations this year are higher, which is consistent with the market overreacting to changes in analyst following and subsequently correcting its excesses. Future returns for firms with extremely low valuations and for which following increases are similar to future returns for firms with extremely high valuations and for which following decreases. This further corroborates the market overreaction explanation. We conclude that while analysts decisions to add or drop coverage of stocks are sound in that their decisions are solidly grounded in fundamentals, market overreaction to changes in analyst following means that blindly implementing these decisions makes for a bad investment strategy. 6. Robustness Tests We subject our main result to a battery of robustness tests. We wish to ensure that our results are general rather than being mechanical or being driven by a particular or unrepresentative group of firms. To this end, we replicate the results in Panel B of Table III in various incarnations and we present the results in Table VIII. [Insert Table VIII about here] First, we run regressions of excess returns next year on changes in analyst following this year by five year intervals to test whether or not the return reversal is 21

23 consistent across time. From Panel A, the return reversal appears to be a stable phenomenon throughout our sample period. Second, we test the relation between excess returns next year and changes in analyst following this year separately by analyst following quintiles constructed using analyst following last year. Our objective is to group together firms such that firms in each group attract roughly the same amount of analyst attention. From Panel B, no matter how much or little analyst attention firms receive, returns reverse after changes in analyst following. Third, we test whether or not the return reversal is transitory or persistent by regressing returns two and three years into the future on change in analyst following this year. From Panel C, though smaller in magnitude (-2.5 and -1.6 percentage points for two and three years, respectively, rather than -4.6), the return reversal persists for several years. Fourth, we test whether or not the return reversal is a mechanical result of not controlling for the positive relation between analyst following next year and returns next year. This is a possibility since the change in analyst following is positively related to returns this year but negatively related to returns next year. We run regressions of returns next year on changes in analyst following this year for even and odd years separately, thereby necessarily destroying any mechanical relation between adjacent years. From Panel D, the return reversal persists and is clearly not generated mechanically. 4 4 We cannot control for following as usual because if we did we would have to control for following both this year and next year, which would induce perfectly multicollinearity with change in analyst following this year. 22

24 Fifth, we test whether or not it is primarily large changes in the information environment of firms that have a large impact on returns. To do this, we exclude all large changes in analyst following, namely those below the 5 th percentile (-4 analysts) and above the 95 th percentile (4 analysts), and restricting attention to firm-years during which following changes by (the absolute value of) zero to four analysts in one instance and by zero to one analysts in another. From Panel E, an additional analyst this year is associated with incremental returns of -4.6, -4.9, and -5.6 percentage points when we include all changes in following, changes of only zero to four analysts, and changes of only zero to one analyst in succession. Thus the return reversal is not primarily driven by large changes in analyst following. Finally, we briefly describe other robustness tests the results of which are not tabulated. First, lower priced stocks are harder to arbitrage so they are more likely to be mispriced and may experience larger swings in returns. To test whether this affects the return reversal, we only include stocks that have a month-end closing price of at least $5 throughout our sample period. Second, firms that go public underperform for several years thereafter, and IPOs tend to receive analyst coverage immediately after listing. To test whether this affects the return reversal, we include only firm-years for which the firm has been followed in IBES for at least five years. Third, the marginal impact of change in analyst following may be greater for younger versus older firms. To test whether this affects the return reversal, we control for the logarithm of the number of years listed. In all three cases above, the return reversal persists, though the relations between change in analyst following this year and returns this year and next year are smaller in magnitude when we exclude low priced stocks, IPOs, and younger firms, as we would expect. 23

25 7. Conclusion We examine the market reaction to sell-side analysts decisions to add or drop coverage. Firms for which analyst following increases (decreases) have higher (lower) returns during the year in which coverage changes. However, returns reverse in the next year, leading to a decrease-increase return spread of 6.4 percentage points. In contrast to the market reaction, operating performance is better for firms for which analyst following increases than for firms for which following decreases, both in the year of the change in coverage and the next year. The market reaction and operating performance results are even stronger when changes in analyst following are confirmed by changes in analysts consensus recommendation and changes in institutional ownership. Moreover, for a given change in analyst following, firms with higher valuations have lower returns next year. Glamour stocks for which analyst following increases have the worst returns next year whereas value stocks for which following decreases have the highest returns. Taken together, our results suggest that the market overreacts to changes in analyst following and subsequently corrects itself. 24

26 REFERENCES Bhushan, Ravi, 1989, Firm characteristics and analyst following, Journal of Accounting and Economics 11, Boni, Leslie, and Kent L. Womack, 2006, Analysts, industries, and price momentum, Journal of Financial and Quantitative Analysis 41, Irvine, Paul J., 2003, The incremental impact of analyst initiation of coverage, Journal of Corporate Finance 9, Jegadeesh, Narasimhan, Joonghyuk Kim, Susan D. Krische, and Charles M. C. Lee, 2004, Analyzing the analysts: When do recommendations add value?, Journal of Finance 59, Krigman, Laurie, Wayne H. Shaw, and Kent L. Womack, 2001, Why do firms switch underwriters?, Journal of Financial Economics 60, Lakonishok, Josef, Andrei Shleifer, and Robert W. Vishny, 1994, Contrarian investment, extrapolation, and risk, Journal of Finance 49, Lee, Charles M. C., and Bhaskaran Swaminathan, 2000, Price momentum and trading volume, Journal of Finance 55, Lehavy, Reuven, and Richard G. Sloan, 2006, Investor recognition and stock returns, working paper. McNichols, Maureen, and Patricia C. O Brien, 1997, Self-selection and analyst coverage, Journal of Accounting Research 35, Merton, Robert C., 1987, A simple model of capital market equilibrium with incomplete information, Journal of Finance 42,

27 Michaely, Roni, and Kent L. Womack, 1999, Conflict of interest and the credibility of underwriter analyst recommendations, Review of Financial Studies 12, O Brien, Patricia C., and Ravi Bhushan, 1990, Analyst following and institutional ownership, Journal of Accounting Research 28, Womack, Kent L., 1996, Do brokerage analysts recommendations add value?, Journal of Finance 51,

28 Appendix. Fixed Effects Regressions Our sample consists of a cross-section of firms across time, an unbalanced panel. We can approach our data in the familiar linear regression setting, y it = α + X β + v + ε, where, for each individual i, we have multiple observations, one it i it for each time period t. In our application, we examine the effect of change in analyst following across time for a given firm and for multiple firms, so the impact of change in analyst following on returns may be different from firm to firm (e.g., riskier firms may have higher returns). With observations on each individual at different time periods, the residual may be a compound residual with an individual specific component vi + ε, it where ε it is the least squares residual and v i is a residual that is constant across time and specific to individual i. For this reason, we take the fairly general approach of implementing a firm fixed effects model. Specifically, we run ordinary least squares regressions on the equation ( X it X i + X ) β + v + ( ε it ε i ε ) y it yi + y = α + +, where yi = ( 1 Ti ) t = y 1 it, y = N ( NTi ) Ti 1 i = t = y 1 1 it, and i X, X, v, ε i, and ε are analogously defined. This yields the same coefficient estimates as running an ordinary least squares regression on T i y it = α + X β + v + ε and including firm dummy variables. The standard errors from it i it ordinary least squares and firm fixed effects are identical after adjusting for the extra N-1 estimated firm means. The within R 2 s from fixed effects ( within firms, as opposed to between firms and overall ) are identical to the ordinary least squares R 2 s. Since with fixed effects we are estimating, among other things, one mean per firm, we can only include other regressors that are not constant across time. The correlation between X i 27

29 and v i is zero since v i is fixed by assumption. While fixed effects always give consistent estimates, they may not be efficient relative to random effects. However, the Hausman test for the equality of the coefficients from fixed and random effects indicates that random effects are not appropriate (p-value ), so we stick to fixed effects. 28

30 Table I Characteristics of Firms Followed by Analysts This table presents the relations between the number of analysts following a stock, the number of institutions owning a stock, and market capitalization. The sample consists of all firms that are followed by at least one analyst, both this year and the previous year. For ease of interpretation, results for a single representative year, 1994, are presented. Panel A: Number of firms by analyst following and market capitalization quintiles, 1994 only Analyst following quintile Market capitalization quintile Median number of analysts Median capitalization , ,967.8 Panel B: Number of S&P 500 firms by analyst following and market capitalization quintiles, 1994 only Analyst following quintile Market capitalization quintile Median number of analysts Median capitalization , ,578.6 Panel C: Number of firms by analyst following and institutional ownership quintiles, 1994 only Analyst following quintile Institutional ownership quintile Median number of analysts Median managers Panel D: Number of firms by institutional ownership and market capitalization quintiles, 1994 only Institutional ownership quintile Market capitalization quintile Median number of institutions Median capitalization , ,

31 Table II Calendar Year Regression of Change in Analyst Following On Its Determinants This table presents a calendar year firm fixed effects regression of the change in the number of analysts following a firm on the determinants of change in analyst following. The sample consists of all firms-years between 1984 and 2004 such that there is at least one analyst following each firm each year and the previous year. Δcap decile is measured relative to NYSE capitalization deciles. Δinstitutional breadth is the change in the number institutions that own the stock each year relative to the previous year and scaled by the total number of institutions the previous year. Δturnover percentile is measured relative to other firms listed on the same exchange as the firm. The equity issuance dummy equals one if the firm issues equity this year and zero otherwise. The acquirer dummy equals one if the firm completes an acquisition this year and zero otherwise. Return on equity is scaled by the mean of book value of equity this year and last year. Capital expenditures are scaled by the mean of total assets this year and last year. For comparability, the impact on the change in the number of analysts following a firm from a one standard deviation increase in each of the explanatory variables is also presented. Δnumber of analysts Δnumber of analysts b se(b) change from a one standard deviation increase in variable Δcap decile 0.137*** (7.60) 0.12 Δinstitutional breadth *** (25.29) 0.34 Δturnover percentile 0.006*** (7.19) 0.08 raw stock return this year 0.090*** (2.86) 0.05 raw stock return last year 0.701*** (33.67) 0.42 market return 1.534*** (21.65) 0.26 Δbook-to-market (1.47) 0.02 equity issuance dummy 0.493*** (10.63) acquirer dummy (0.61) Δreturn on equity (1.55) 0.02 Δsales growth (1.59) Δcapex 0.540** (2.41) 0.03 Constant *** (25.95) Number of firm-years Number of firms 7890 R-squared

32 Table III Calendar Year Change in Analyst Following and Returns This table presents sample means in Panel A and firm fixed effects regressions in Panel B examining change in analyst following and returns last year, this year, and next year. The sample consists of all firmsyears between 1984 and 2004 such that there is at least one analyst following each firm each year and the previous year. Panel A presents mean excess of market returns last year, this year, and next year by analyst following increases, no changes, and decreases. Panel B presents the results of regressions of returns last year, this year, and next year on change in analyst following this year. The Fama-French three factors plus momentum are included as explanatory variables in Panel B. Panel A: Mean returns conditional on change in analyst following excess return t-1 excess return t excess return t+1 Analyst following Increase No change Decrease Decrease-increase return spread t-statistic for H 0 : increase - decrease = Number of firm-years Panel B: Regressions of returns on change in analyst following excess return t-1 excess return t excess return t+1 b se(b) b se(b) b se(b) Δfollowing t 0.121*** (40.53) 0.032*** (11.88) *** (28.59) ln(following t-1 ) 0.014*** (2.77) *** (32.70) *** (34.67) Δfollowing t ln(following t-1 ) *** (28.25) *** (8.68) 0.023*** (20.63) R m -R f 1.015*** (65.56) 0.957*** (63.40) 0.933*** (60.16) SMB 0.779*** (38.57) 0.715*** (37.31) 0.687*** (34.62) HML 0.233*** (13.61) 0.233*** (14.22) 0.259*** (15.30) UML 0.059*** (2.95) ** (2.05) *** (3.19) Constant ** (2.18) 0.237*** (27.47) 0.289*** (32.26) Number of firm-years Number of firms R-squared

33 Table IV Analyst Following As Investor Recognition This table presents firm fixed effects regressions examining change in analyst following and returns last year, this year, and next year. The sample consists of all firms-years between 1984 and 2004 such that there is at least one analyst following each firm each year and the previous year. Panel A presents the results of regressions of returns this year and next year on change in analyst following this year and change in institutional breadth this year. Panel B presents the results of regressions of returns this year and next year on changes in analyst following this year separated for high risk and low risk firms, where the risk of a firm is determined by whether the firm s annualized standard deviation of monthly return this year is above or below the median for all CRSP firms this year. Panel C presents the results of regressions of equity issuance, acquisitions, and capital expenditures change on contemporaneous change in analyst following. The equity issuance dummy equals one if the firm issues equity this year and zero otherwise. The acquirer dummy equals one if the firm completes and acquisition this year and zero otherwise. The Fama-French three factors plus momentum are included as explanatory variables in all panels, though for expositional simplicity they are not reported. Panel A: Regressions of returns on change in analyst following and change in institutional breadth excess return t excess return t+1 b se(b) b se(b) Δfollowing t 0.013*** (5.33) *** (26.58) ln(following t-1 ) *** (32.31) *** (34.43) Δfollowing t ln(following t-1 ) *** (9.55) 0.023*** (20.04) Δinstitutional breadth t *** (111.83) *** (16.82) Constant 0.151*** (18.89) 0.303*** (33.12) Number of firm-years Number of firms R-squared Panel B: Regressions of returns on changes in analyst following separated for high risk and low risk firms excess return t excess return t+1 b se(b) b se(b) Δfollowing t high risk t 0.037*** (13.28) *** (29.38) Δfollowing t low risk t 0.016*** (5.04) *** (21.52) ln(following t-1 ) *** (32.97) *** (34.54) Δfollowing t ln(following t-1 ) *** (5.87) 0.021*** (18.29) Constant 0.240*** (27.82) 0.286*** (31.99) Number of firm-years Number of firms R-squared Panel C: Regressions of financing and investment on change in analyst following equity issuance dummy t acquirer dummy t Δcapex t b se(b) b se(b) b se(b) Δfollowing t 0.047*** (34.86) 0.016*** (9.94) 0.003*** (8.07) ln(following t-1 ) * (1.81) 0.034*** (11.64) *** (5.23) Δfollowing t ln(following t-1 ) *** (26.73) *** (5.48) *** (5.89) Constant 0.077*** (19.54) 0.073*** (15.50) (0.35) Number of firm-years Number of firms R-squared

34 Table V Calendar Year Changes in Analyst Following Confirmed or Contradicted By Changes in Analysts Consensus Recommendations or Changes in Institutional Ownership and Returns This table presents mean excess of market returns this year and next categorized by confirmed or contradicted change in analyst following. The sample consists of all firms-years between 1984 and 2004 such that there is at least one analyst following each firm each year and the previous year. Firms are first grouped by analyst following increases, no changes, and decreases. Then, firms are separated into strict changes by whether they are confirmed or contradicted by changes in analysts consensus recommendations in one instance and by changes in institutional ownership in another. Mean returns are also presented for increases and decreases. Pairs indicated by a, b, c, d, and e denote statistical significance at the 1% level for a difference in means test. Mean excess returns conditional on whether changes in analyst following are confirmed or contradicted by changes in analysts consensus recommendation by changes in institutional ownership This year Next year This year Next year Following increase mean return Following increase, confirmed ( ) 23.8 a,c -1.6 e 22.6 a,c -1.6 e Following increase, contradicted ( ) -0.4 a a -1.2 No following change Following decrease mean return Following decrease, contradicted ( ) 12.1 b 4.3 d 14.3 b 2.2 d Following decrease, confirmed ( ) b,c 8.4 d,e b,c 7.5 d,e Number of firm-years

35 Table VI Calendar Year Change in Analyst Following and Operating Performance This table presents sample means of operating performance measures this year and next categorized by confirmed or contradicted change in analyst following. The sample consists of all firms-years between 1984 and 2004 such that there is at least one analyst following each firm each year and the previous year. Panel A presents sample mean operating performance measures this year and next year by analyst following increases, no changes, and decreases. Panel B presents sample mean operating performance measures this year and next year by changes in analyst following this year, where increases and decreases are separated by whether they are confirmed or contradicted by analysts consensus recommendation changes. Return on equity is scaled by the mean of book value of equity this year and last year. Capital expenditures are scaled by the mean of total assets this year and last year. Pairs indicated by a, b, c, d, and e denote statistical significance at the 1% level for a difference in means test. Panel A: Mean operating performance measures conditional on change in analyst following Analyst following Return on equity Sales growth Capital expenditures This year Next year This year Next year This year Next year Increase 9.4% 7.5% 29.0% 20.8% 8.3% 7.8% No change 3.0% 2.5% 17.2% 13.9% 6.9% 6.4% Decrease 1.8% 2.5% 12.0% 9.7% 6.7% 6.1% t-statistic for H 0 : increase - decrease = Number of firm-years Panel B: Mean operating performance measures conditional on changes in analyst following confirmed or contradicted by analysts consensus recommendation changes Following increase, confirmed ( ) Following increase, contradicted ( ) return on equity t return on equity t+1 sales growth t sales growth t+1 capex t capex t a,c 7.5 d,f 35.0 c 27.8 d,f 7.7 a,c 7.7 d,f 6.5 a 4.6 d d 8.0 a 7.1 d No following change Following decrease, contradicted ( ) Following decrease, confirmed ( ) 4.5 b 5.6 e 15.0 b,c 14.3 e e -1.0 b,c -0.7 e,f 12.8 b,c 8.0 e,f 6.5 c 5.4 e,f Number of firm-years

36 Table VII Mean Returns Next Year Conditional on Change in Analyst Following This Year and Valuation Proxies This Year This table presents mean excess of market returns by fifteen categories of three change in analyst following groups and valuation quintiles. The sample consists of all firms-years between 1984 and 2004 such that there is at least one analyst following each firm each year and the previous year. Valuation quintile breakpoints are based on all Compustat firms in a given year. Sales growth is measured over the last five years. Turnover is mean monthly ratio of monthly volume to month end shares outstanding. Panel A: Mean returns next year conditional upon change in analyst following this year and book-to-market quintile this year Analyst following Book-to-market quintile Mean for change in t-statistic for H Glamour Value 0 : analyst following value = glamour Increase No change Decrease Increase minus decrease t-statistic for H 0 : increase - decrease = Number of firm-years Panel B: Mean returns next year conditional upon change in analyst following this year and cash flow-to-price quintile this year Analyst following Cash flow-to-price quintile Mean for change in t-statistic for H Glamour Value 0 : analyst following value = glamour Increase No change Decrease Increase minus decrease t-statistic for H 0 : increase - decrease = Number of firm-years

37 Panel C: Mean returns next year conditional upon change in analyst following this year and sales growth quintile this year Analyst following Sales growth quintile Mean for change in t-statistic for H Glamour Value 0 : analyst following value = glamour Increase No change Decrease Increase minus decrease t-statistic for H 0 : increase - decrease = Number of firm-years Panel D: Mean returns next year conditional upon change in analyst following this year and turnover quintile this year Analyst following Turnover quintile Mean for change in t-statistic for H Glamour Value 0 : analyst following value = glamour Increase No change Decrease Increase minus decrease t-statistic for H 0 : increase - decrease = Number of firm-years

38 Table VIII Robustness Tests This table presents various firm fixed effects robustness tests of the main result presented in Panel B of Table III. The sample consists of all firms-years between 1984 and 2004 such that there is at least one analyst following each firm each year and the previous year. Panel A presents regressions of excess of market return next year for four intervals of five years each. Panel B presents regressions of excess of market return next year on change in analyst following this year scaled by analyst following last year. Panel C presents regressions of excess of market return one, two, and three years into the future on change in analyst following this year. Panel D presents regressions of excess of market return next year on analyst following this year and next. Panel E presents regressions of excess of market returns next year on change in analyst following this year but with various restrictions on the magnitude of change in analyst following. The Fama-French three factors plus momentum are included as explanatory variables in all panels but for expositional simplicity they are not reported. Panel A: Calendar year regressions of returns on change in analyst following by five year intervals excess return t b se(b) b se(b) b se(b) b se(b) Δfollowing t *** (10.70) *** (15.69) *** (13.51) *** (19.92) ln(following t-1 ) *** (12.26) *** (16.15) *** (12.19) *** (12.52) Δfollowing t ln(following t-1 ) 0.016*** (7.70) 0.033*** (11.31) 0.030*** (10.01) 0.033*** (14.13) Constant 0.248*** (9.99) 0.462*** (16.35) 0.377*** (10.13) (0.23) Number of firm-years Number of firms R-squared Panel B: Calendar year regressions of returns on relative change in analyst following by lagged analyst following quintiles excess return t+1 Least followed Most followed b se(b) b se(b) b se(b) b se(b) b se(b) Δfollowing t /following t *** (12.93) *** (11.01) *** (13.93) *** (13.80) *** (7.16) Constant 0.031*** (3.13) 0.032** (2.24) (0.06) (0.80) 0.013* (1.91) Number of firm-years Number of firms R-squared

39 Panel C: Calendar year regressions of returns one, two, and three years into future years on change in analyst following this year excess return t+1 excess return t+2 excess return t+3 b se(b) b se(b) b se(b) Δfollowing t *** (28.59) *** (15.31) *** (8.75) ln(following t-1 ) *** (34.67) *** (19.90) *** (13.26) Δfollowing t ln(following t-1 ) 0.023*** (20.63) 0.015*** (12.02) 0.009*** (6.85) Constant 0.289*** (32.26) 0.201*** (20.59) 0.152*** (14.46) Number of firm-years Number of firms R-squared Panel D: Calendar year regression of returns next year on change in analyst following this year and next year excess return t+1 Even years only Odd years only b se(b) b se(b) Δfollowing t *** (19.85) *** (18.97) ln(following t-1 ) *** (24.91) *** (26.21) Δfollowing t ln(following t-1 ) 0.025*** (13.93) 0.022*** (14.23) Constant 0.329*** (23.10) 0.253*** (20.69) Number of firm-years Number of firms R-squared Panel E: Calendar year regressions of returns on change in analyst following by magnitude of change in analyst following excess return t+1 Δfollowing t not restricted Δfollowing t [0,1,2,3,4] Δfollowing t [0,1] b se(b) b se(b) b se(b) Δfollowing t *** (28.59) *** (25.93) *** (13.74) ln(following t-1 ) *** (34.67) *** (33.48) *** (24.16) Δfollowing t ln(following t-1 ) 0.023*** (20.63) 0.029*** (18.15) 0.039*** (8.43) Constant 0.289*** (32.26) 0.277*** (30.76) 0.210*** (20.80) Number of firm-years Number of firms R-squared

40 Figure 1-A: Size of typical firm followed by analysts each year 100,000 Market capitalization ($ million) 10,000 1,000 2, , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , Year Mean market capitalization of firms followed by analysts Mean market capitalization of S&P 500 firms 30 Figure 1-B: Distribution of analyst following each year Number of analysts Year 5th percentile Median Mean 95th percentile Figure 1. Summary statistics for analyst following each year. The sample consists of all firms-years between 1984 and 2004 such that there is at least one analyst following each firm each year and the previous year. Panel A shows each year the mean market capitalization of firms followed by analysts and of S&P 500 firms. Panel B shows each year the mean and median and 5 th and 9 th percentile of the number of analysts following firms. 39

41 Figure 2-A: Distribution of change in analyst following each year Change in number of analysts Year 5th percentile Median Mean 95th percentile Figure 2-B: Relative percentage of changes in analyst following by year 100% 90% 80% 70% Analyst following change 60% 50% 40% 30% 20% 10% 0% Year Decrease No change Increase Figure 2. Summary statistics for change in analyst following each year. The sample consists of all firms-years between 1984 and 2004 such that there is at least one analyst following each firm each year and the previous year. Panel A shows each year the mean and median and 5th and 9th percentile of the change in the number of analysts following firms. Panel B shows each year the relative percentage of increases, no changes, and decreases in the number of analysts following firms. 40

42 Figure 3-A: Number of analysts following a firm this year if 2 analysts followed the firm last year Figure 3-B: Number of analysts following a firm this year if 4 analysts followed the firm last year Figure 3-C: Number of analysts following a firm this year if 9 analysts followed the firm last year Figure 3. Number of analysts following a firm this year conditional on analyst following last year. The sample consists of all firms-years between 1984 and 2004 such that there is at least one analyst following each firm each year and the previous year. Panels A, B, and C show the distribution of the number of analysts following a firm this year conditional on 2, 4, and 9 analysts following the same firm last year, respectively, these being the 25 th, 50 th, and 75 th percentiles of the distribution of analyst following last year, respectively. 41