CAN INVESTORS PROFIT FROM THE PROPHETS? CONSENSUS ANALYST RECOMMENDATIONS AND STOCK RETURNS

Transcription

1 CAN INVESTORS PROFIT FROM THE PROPHETS? CONSENSUS ANALYST RECOMMENDATIONS AND STOCK RETURNS Brad Barber Graduate School of Management University of California, Davis Reuven Lehavy Haas School of Business University of California, Berkeley Maureen McNichols Graduate School of Business Stanford University and Brett Trueman Haas School of Business University of California, Berkeley First Draft: August 1998 Please do not cite without permission. We thank Zacks Investment Research for providing the data used in this study. Lehavy and Trueman also thank the Center for Financial Reporting and Management at the Haas School of Business for providing research support. All remaining errors are our own.

2 ABSTRACT In this paper we document that an investment strategy based on the consensus (average) recommendations of security analysts earns positive returns. For the period , a portfolio of stocks most highly recommended by analysts earned an annualized geometric mean return of 18.8 percent, while a portfolio of stocks least favorably recommended earned only 5.78 percent. (In comparison, an investment in a value-weighted market index earned an annualized geometric mean return of 14.5 percent.) Alternatively stated, purchasing stocks most highly recommended by analysts and selling short those least favorably recommended yielded a return of 102 basis points per month. The magnitude of this return is surprisingly large, and is far greater than the size effect (negative 16 basis points) and book-to-market effect (17 basis points) for the same period. Even after controlling for these two effects, as well as for price momentum, we show that the strategy of purchasing stocks most highly recommended and selling short those least favorably recommended yielded a return of 75 basis points per month. These results are robust to partitions by time period and overall market direction, and are most pronounced for small and medium-sized firms. The abnormal returns also persist when we allow a lapse of up to 15 days before acting on the investment recommendations. There is no extant theory of asset pricing that explains these results.

3 CAN INVESTORS PROFIT FROM THE PROPHETS? CONSENSUS ANALYST RECOMMENDATIONS AND STOCK RETURNS INTRODUCTION This study examines whether the publicly available recommendations of security analysts have investment value. Academic theory and Wall Street practice are clearly at odds regarding this issue. On the one hand, the semi-strong form of market efficiency posits that investors cannot trade profitably on the basis of publicly available information, such as analyst recommendations. On the other hand, research departments of brokerage houses spend billions of dollars annually on security analysis, presumably because these firms believe it can generate superior returns for their 1 clients. These observations provide a compelling empirical motivation for our inquiry and 2 distinguish our analysis from many recent studies of stock return anomalies. In contrast to many of these studies, which focus on corporate events, such as stock splits, or firm characteristics, such as recent return performance, that are not directly tied to how people invest their money, we analyze an activity security analysis that is undertaken by investment professionals at hundreds of major brokerage houses with the express purpose of improving the return performance of their clients. Our results provide surprisingly strong evidence that Wall Street may be right. For the sample period of we find that buying the stocks with the most favorable consensus 1 In recent years statistics on consensus (average) analyst recommendations have become widely available on many Internet web sites as well as on the databases of several investment information providers. See, for example, CBSMarketWatch, at and the Dow Jones Retrieval Service. The consensus analyst recommendation data usually comes from either First Call or Zacks Information Research. 2 See Fama (1998) for a review and critique of this body of work. 1

4 (average) recommendations earned an annualized geometric mean return of 18.8 percent, while buying those with the least favorable consensus recommendations earned only 5.78 percent. As a benchmark, during the same period an investment in a value-weighted market portfolio earned an annualized geometric mean return of 14.5 percent. Alternatively stated, the most highly recommended stocks outperformed the least favorably recommended ones by a strikingly large 102 basis points per month. In comparison, over the same period the book-to-market effect was a 3 mere 17 basis points, while the size effect was a negative 16 basis points per month. After controlling for market risk, size, and book-to-market effects using the Fama-French three-factor model, a portfolio comprised of the most highly recommended stocks provided an average annual abnormal return of 4.2 percent while a portfolio of the least favorably 4 recommended ones yielded an average annual abnormal return of -7.6 percent. Consequently, purchasing the securities in the top portfolio and selling short those in the lowest portfolio yielded an average annual abnormal return of 11.8 percent. Our results are robust to partitions by time period and overall market direction. They are most pronounced for small and medium-sized firms; among the few hundred largest firms we find no reliable differences between the returns of those most highly rated and those least favorably recommended. Our results also persist when we allow a lapse of up to 15 days before acting on the analysts recommendations. As such, our results provide strong positive evidence in the debate over whether security analysts 3 The size and book-to-market effects were calculated using portfolios constructed by Fama and French (1993). 4 Other return models give similar results. 2

5 recommendations have investment value. 5 These returns are gross of transaction costs, such as the bid-ask spread, brokerage commissions, and the market impact of trading. As we show, investing in highly recommended securities requires an active trading strategy with turnover rates at times in excess of 400 percent annually. After a reasonable accounting for transaction costs, active trading strategies based on the recommendations of analysts cannot reliably beat a market index. Nevertheless, the consensus recommendations remain valuable to investors who are otherwise considering buying or selling; ceteris paribus, an investor would be better off purchasing shares in firms with more favorable consensus recommendations and selling shares in those with less favorable consensus recommendations. Our analysis represents the most comprehensive study to-date of the investment performance of analysts recommendations. Using the Zacks database for the period , which includes over 360,000 recommendations from 269 brokerage houses and 4,340 analysts, we track in calendar time the investment performance of portfolios of firms grouped according to their consensus analyst recommendations. Every time an analyst is reported as initiating coverage, changing his or her rating of a firm, or dropping coverage, the consensus recommendation of the firm is recalculated and the firm moves between portfolios, if necessary. Any required portfolio rebalancing occurs at the end of the trading day. This means that investors are assumed to react to a change in a consensus recommendation at the close of trading on the 5 Among the papers which previously examined the investment performance of security analysts stock recommendations are Barber and Loeffler (1993), Bidwell (1977), Diefenbach (1972), Groth, Lewellen, Schlarbaum, and Lease (1979), Stickel (1995), and Womack (1996). Copeland and Mayers (1982) studied the investment performance of the Value Line Investment Survey while Desai and Jain (1995) analyzed the return from following Barron s annual roundtable recommendations. 3

6 day that the change took place. Consequently, any return that investors might have earned from advance knowledge of the recommendations is excluded from the return calculations. Our study is most closely related to Stickel (1995) and Womack (1996). Using the Zacks database, Stickel studies the price impact of 16,957 changes in analyst recommendations over the period. He finds that recommendation changes from sell to buy (buy to sell) were accompanied by positive (negative) returns at the time of the announcement. Further, he documents that most of the price adjustment occurred during the first 30 days after the recommendation change. Using the First Call database, Womack analyzes the impact of 1,573 changes in analyst recommendations to or from strong buy or strong sell, for the top 14 U.S. brokerage research departments during the period. He finds significantly positive (negative) returns for the buy (sell) recommendations at the time of the announcement. He also documents a post-recommendation stock price drift lasting up to one month for buys and six months for sells. Our paper differs from those of Stickel and Womack in two important respects. First, we analyze more than twenty times as many recommendations and a much longer period of time than do either Stickel or Womack. Second, we study analyst recommendations, themselves, rather than changes in their ratings. This shifts the analysis away from the immediate price reaction to analyst upgrades and downgrades and allows us to focus on the calendar-time returns earned from following strategies based on the consensus recommendations of covered firms. By doing so, we take the perspective of the investor rather than the individual analyst. The difference in these perspectives can be illustrated by considering the example of Compaq Computer, which received 11 upgrades or downgrades on March 1, In our 4

7 analysis, this firm would be represented once on that day; in Stickel's and Womack's analyses, 6 each of the upgrades and downgrades would represent an observation. Implicit in our analysis, then, is the assumption that an investor does not choose weights for each stock in his or her portfolio according to the number of strong buy recommendations the firm has received, but, rather, according to its consensus recommendation. Although an individual investor could, of course, follow such a strategy, in aggregate a stock must be held in proportion to its market capitalization, not in proportion to the number of strong buy recommendations it receives. The plan of this paper is as follows. In Section I we describe the data and our sample selection criteria. A discussion of our research design follows in Section II. In Section III we form portfolios according to consensus analyst recommendations and analyze their returns. Additional tests are performed in Section IV. In Section V we estimate the transactions costs of following the strategy of buying the most highly rated stocks and selling short those that are least favorably rated and discuss the profitability of this strategy. A summary and conclusions section ends the paper. I. THE DATA, SAMPLE SELECTION CRITERIA, AND DESCRIPTIVE STATISTICS The analyst recommendations used in this study were provided by Zacks Investment Research, who obtains its data from the written and electronic reports of brokerage houses. The recommendations encompass the period 1985 (the year that Zacks began collecting this data) 6 The clustering of upgrades or downgrades in time can also lead to event-centered returns that are crosssectionally dependent. Indeed, if the same firm receives multiple upgrades on the same day, the event-centered returns for these upgrades will be perfectly correlated. This cross-sectional dependence renders traditional statistical tests in event studies invalid. See Brav (1998), Fama (1998), and Lyon, Barber, and Tsai (1998) for further discussion of this issue. 5

8 through Each database record includes, among other items, the recommendation date, identifiers for the brokerage house issuing the recommendation and the analyst writing the report (if the analyst s identity is known), and a rating between 1 and 5. A rating of 1 reflects a strong buy recommendation, 2 a buy, 3 a hold, 4 a sell, and 5 a strong sell. This five-point scale is commonly used by analysts. If an analyst uses a different scale, Zacks converts the analyst s rating to its five-point scale. Ratings of 6 also appear in the Zacks database and signify termination of coverage. This rating will be given when an analyst announces that he or she is dropping coverage of a firm, when the analyst switches brokerage houses, or when the analyst is restricted from covering the firm for a period of time, possibly due to the initiation of investment banking activity between the analyst s brokerage house and the firm. A rating of 6 may also be assigned by Zacks if the analyst has not issued a new recommendation within the last 366 days. By doing so, Zacks is making the implicit assumption that the analyst is no longer following the firm, but that the analyst has not made a formal announcement to this effect. In some instances the date assigned to a recommendation by Zacks will be later than the date on which the recommendation becomes publicly available. This situation will arise, for example, if Zacks uses the date of publication of an analyst s written report but the analyst publicly announced his recommendation a few days earlier. In this case, investors will have been able to act on the recommendation before the date recorded by Zacks. To the extent that this is a prevalent occurrence, our tests for investment value in analysts recommendations will be less powerful. Another characteristic of the database, one that has not been explicitly acknowledged in 6

9 any prior study as far as we are aware, is that the data made available to academics does not constitute Zacks complete set of recommendations. According to an official at Zacks, some individual brokerage houses have entered into agreements that preclude their recommendations from being distributed by Zacks to anyone other than the brokerage houses clients. Consequently, the recommendations of several brokerage houses, including such large ones as Merrill Lynch, Goldman Sachs, and Donaldson, Lufkin, and Jenrette, are not part of the database that has been analyzed by academics. 7 The Zacks database contains 378,326 recommendations for the years Dropping observations for the 1,286 firms not appearing on the CRSP file leaves a final sample of 361,620 recommendations. Table I provides descriptive statistics for these recommendations. As shown in column 3, the number of firms covered by Zacks has increased steadily over the years. For the year 1996, 59.8 percent of all firms on the NYSE, ASE, or NASDAQ had at least one recommendation in the database (column 4). The market capitalization of these firms constituted 95.6 percent of the capitalization of all firms in the market (column 5). This is consistent with the conventional wisdom that analysts tend to cover larger firms, because they offer more liquidity and allow the analysts clients to more easily take large positions in the firms shares. The mean number of analysts per covered firm has increased over time, in general (column 6), while the median number has remained constant, with the exception of 1985 (column 7). From 1986 onward, the mean and median number of covered firms per analyst has also been stable (columns 8 and 9). The number of brokerage houses contributing recommendations to 7 Given the high ratings that many analysts at these brokerage firms receive in surveys of institutional investors, this omission most likely biases our tests against finding investment value in analysts recommendations. 7

10 Zacks and the number of analysts providing forecasts has steadily increased over time (columns 10 and 11). The last column of the table reports the average of all of the analyst ratings, by year. It shows a rather steady decrease over time, indicating that analysts recommendations have become more favorable. Finally, it should be noted that the year 1985 has by far the smallest number of covered firms, brokerage houses, and analysts. Because 1985 is the first year that Zacks began tracking recommendations, this finding is not at all surprising. Since the 1985 data is so sparse, though, 8 we do not include the investment returns from that year in our analysis. However, we do use the recommendations, themselves, as they are needed to calculate consensus ratings for 1986 (to the extent that the recommendations carry over to that year). A 6 x 6 transition matrix of the analysts recommendations appears in table II. Each cell {i,j} of the matrix contains two numbers. The top one is the number of observations in the database in which an analyst moved from a recommendation of i to one of j; the bottom number is the median number of days between the announcement of a recommendation of i and a revised recommendation of j. The diagonal elements of the matrix reflect reiterations of analyst recommendations. Most of the entries in this matrix are concentrated in the upper 3 x 3 cells. This is to be expected, given the conventional wisdom that analysts are reluctant to issue sell recommendations. Within this region, the bulk of the observations represent reiterations. The mean time between a recommendation and its reiteration is a little less than 300 days. This is much longer than the mean time between a recommendation and a revision by the analyst to a new rating, which is generally in the low 100-day range. 8 Our inferences are not affected by the exclusion of the 1985 returns. 8

11 The line entitled First Zacks Recommendation records the first recommendation in the database for a given analyst-company pair. Consistent with McNichols and O Brien (1998), the first recommendation is usually a buy (1 or 2), less often a hold, and rarely a sell (4 or 5). This again reflects the reluctance of analysts to issue sell recommendations. This observation is also consistent with the numbers in the last two lines of the table. Of all the recommendations in the database 47.1 percent are buys while only 5.7 percent are sells. Excluding observations with a rating of 6, buys constitute 54.1 percent of the total, while sells make up only 6.3 percent. As a general check on the accuracy of our data, we computed the average two-day announcement period return for changes in or initiations of analyst recommendations. These returns are presented in table III. Similar to the results of Stickel (1995) and Womack (1996) we find that the compound return (adjusted for size) for the day before and day of the announcement of a rating change is, in general, significantly positive for upgrades and significantly negative for downgrades. Furthermore, for the set of initial analyst-company recommendations in the database, a buy rating (1 or 2) is accompanied by a significantly positive return, while a hold or sell rating (3, 4, or 5) is associated with a significantly negative return. These significant findings provide supporting evidence for the accuracy of the recommendations and the recorded announcement dates in the Zacks database. II. RESEARCH DESIGN A. Portfolio Construction To determine whether analysts recommendations are predictive of future returns we construct calendar-time portfolios based on the consensus rating of each covered firm. The 9

12 G average analyst rating, A ij-1, for firm i on date J-1 is found by summing the individual ratings, A ijj-1, of the j = 1 to n ij-1 analysts who have outstanding recommendations for the firm on that day and dividing by n ij-1. Formally, Ā ij&1 ' n ij&1 1 j A n ijj&1. ij&1 j'1 Using these average ratings, each of the covered firms is placed into one of five portfolios as of the close of trading on date J-1. The first portfolio consists of the most highly recommended G G stocks, those for which 1#A #1.5; the second is comprised of firms for which 1.5<A #2; the ij-1 G G third contains firms for which 2<A #2.5; the fourth is comprised of firms for which 2.5<A #3; ij-1 G and the fifth portfolio consists of the least favorably recommended stocks, those for which A ij-1>3. The number of portfolios chosen and the ratings cutoffs for each, while somewhat arbitrary, are certainly reasonable. Five portfolios were chosen so as to achieve a high degree of separation across firms in the sample while retaining sufficient power for our tests. The cutoffs were set so that only the bottom portfolio contained firms whose consensus ratings corresponded to hold or sell recommendations, due to the relative infrequency of such ratings. We also ran our main analysis under the alternative specification that the top portfolio contain only firms with an average rating of one, and obtained generally similar results. 9 After determining the composition of each portfolio p as of the close of trading on date J-1, the value-weighted return for date J was calculated. Denoted by R for portfolio p, this pj ij-1 ij-1 9 Such a cutoff results in the top portfolio being comprised of firms with small analyst followings. This is because a rating of one is only possible if all the analysts covering a firm give it a strong buy recommendation. The more analysts there are, the less likely a firm will be able to obtain this rating. 10

13 return is given by: n pj&1 R pj ' j x ij&1 R ij, i'1 where: x = the market value of equity for firm i as of the close of trading on date J-1 divided by the ij-1 aggregate market capitalization of all firms in portfolio p as of the close of trading on that date, R = the return on the common stock of firm i during date J, and ij n = the number of firms in portfolio p at the close of trading on date J-1. pj-1 There are two reasons we value-weight rather than equally-weight the securities in each portfolio. First, an equal weighting of daily returns (and the implicit assumption of daily rebalancing) leads 10 to portfolio returns that are severely overstated. Second, while an individual investor could hold equal amounts of each security, in the aggregate each firm must be held in proportion to its market value. For each month in our sample period, the daily returns for each portfolio p, R, are pj compounded over the n trading days of the month to yield a monthly return, R : pt R pt ' k n J'1 (1%R pj ) & This problem arises due to the cycling over time of a firm s closing price between its bid and ask (commonly referred to as the bid-ask bounce). For a more detailed discussion see Barber and Lyon (1997), Canina, Michaely, Thaler, and Womack (1998), and Lyon, Barber, and Tsai (1998). 11

14 Our study focuses on the monthly returns for the five constructed portfolios, as well as for two additional portfolios. The first additional portfolio consists of all covered firms on each date J (those that received a rating from at least one analyst in the Zacks database on that day) and the second portfolio consists of neglected firms on that date (those firms on the CRSP daily returns 11 file that did not receive any analyst ratings that day). The composition of each of these two portfolios is recalculated every day, since firms gain or lose analyst coverage over time. B. Performance Evaluation Our analysis of portfolio performance begins with a simple calculation of market-adjusted returns. It is given by R - R for portfolio p in month t, where R is the month t return on the pt mt mt CRSP value-weighted market index. We next calculate three measures of abnormal performance for each of our constructed portfolios. First, we employ the theoretical framework of the Capital Asset Pricing Model and estimate the following monthly time-series regression: R pt! R ft ' " p % $ p (R mt! R ft ) %, pt, where: 12 R ft = the month t return on treasury bills having one month until maturity, " = the estimated CAPM intercept (Jensen's alpha), p $ = the estimated market beta, and p 11 Since the academic version of the Zacks database does not include the recommendations of all brokerage houses, it is possible that some of the neglected firms are actually covered by one or more analysts. To the extent this is true, our test for differences in returns between neglected and covered firms is less powerful. IL. 12 This return is taken from Stocks, Bonds, Bills, and Inflation, 1997 Yearbook, Ibbotson Associates, Chicago, 12

15 , = the regression error term. pt This test yields parameter estimates of " p and $ p. Second, we employ an intercept test using the three-factor model developed by Fama and French (1993). To evaluate the performance of each portfolio, we estimate the following monthly time-series regression: R pt! R ft ' " p % $ p (R mt! R ft ) % s p SMB t % h p HML t %, pt, where: SMB = the difference between the month t returns of a value-weighted portfolio of small stocks t and one of large stocks, and HML = the difference between the month t returns of a value-weighted portfolio of high book-tot 13 market stocks and one of low book-to-market stocks. The regression yields parameter estimates of ", $, s, and h. p p p p A third test includes a zero investment portfolio related to price momentum, as follows: R pt! R ft ' " p % $ p (R mt! R ft ) % s p SMB t % h p HML t % m p PMOM t %, pt. PMOM is the equally-weighted month t average return of the firms with the highest 30 percent t return over the eleven months through month t-2, less the equally-weighted month t average 13 The construction of these portfolios is discussed in detail in Fama and French (1993). We thank Ken French for providing us with this data. 13

16 14 return of the firms with the lowest 30 percent return over the eleven months through month t-2. In addition to estimates of ", $, s, and h, this regression yields a parameter estimate of m. p p p p p This specification will be referred to below as the four-characteristic model. In the analysis below we use the estimates of $, s, h, and m to provide insights into the p p p p nature of the firms in each of the portfolios. A value of $ greater (less) than one indicates that p the firms in portfolio p are, on average, riskier (less risky) than the market. A value of s greater p (less) than zero signifies a portfolio tilted toward smaller (larger) firms. A value of h greater p (less) than zero indicates a tilt toward stocks with a high (low) book-to-market ratio. Finally, a value of m greater than zero signifies a portfolio with stocks that have, on average, performed p well (poorly) in the recent past. It is important to note that our use of the Fama-French and four-characteristic models does not imply a belief that the small firm, book-to-market, and price momentum effects represent risk factors. Rather, we use these models to assess whether any superior returns that are documented are due to analysts stock-picking ability or to their choosing stocks with characteristics known to produce positive returns. C. Turnover Both the raw and risk-adjusted returns reported in the tables are gross of any trading costs arising from the bid-ask spread, brokerage commissions, and the market impact of trading. To assess the size of these costs we calculate a measure of daily turnover for each portfolio. 14 The rationale for using price momentum as a factor stems from the work of Jegadeesh and Titman (1993) who show that the strategy of buying stocks that have performed well in the recent past and selling those that have performed poorly generates significant positive returns over three to twelve month holding periods. This measure of price momentum has been used by Carhart (1997). We thank Mark Carhart for providing us with the price momentum data. 14

17 Turnover for portfolio p during trading day J is defined as the percentage of the portfolio s holdings as of the close of trading on date J-1 that has been sold off as of the close of trading on date J. That is, it is the percent of the portfolio that has been turned over into some other set of stocks during date J. Turnover is calculated by following a three-step procedure. First, for each stock i in portfolio p as of the close of trading on date J-1 we calculate the fraction it would have comprised of the portfolio at the end of trading on date J if there were no portfolio rebalancing. Denoting this fraction by G, it is given by ij G ij ' n pj&1 j i'1 x ij&1 (1%R ij ) x ij&1 (1%R ij ), where, as before, x is the market value of equity for firm i as of the close of trading on date J-1 ij-1 divided by the aggregate market capitalization of all firms in portfolio p as of the close of trading on that date. Next, G is compared to the actual fraction firm i makes up of portfolio p at the end ij of trading on date J, denoted by F, taking into account any portfolio rebalancing required as a ij result of changes in analyst recommendations. Finally, the decrease (if any) in the percentage holding of each of the date J-1 securities is summed, yielding the day s portfolio turnover. Denoted by U, it is formally given by: ij U ij ' j n pj i'1 max{g ij &F ij,0}. 15

18 Annual turnover is then calculated by multiplying U by the number of trading days in the year. ij III. PORTFOLIO CHARACTERISTICS AND RETURNS Table IV provides descriptive statistics for portfolios formed on the basis of analysts recommendations. Note first that the average number of firms in the portfolio of the least favorably ranked stocks, portfolio 5, is less than one-third that of any of the other four portfolios (column 2). This is not surprising, given the reluctance of analysts to issue sell recommendations. The average numbers of firms in the other four portfolios are roughly similar. There is considerable variation across portfolios in the average number of analysts per firm, though, ranging from a low of 2.35 for portfolio 1 to a high of 4.93 for portfolio 3 (column 3). The low number of analysts for firms in portfolio 1 may well reflect the difficulty a firm has in attaining an average rating of between 1 and 1.5 if there are many analysts covering it, and leads one to suspect that these firms are relatively small. This is confirmed by the data in column 5, which shows the market capitalization of these firms to be considerably smaller than that of the firms in portfolios 2, 3, and 4. The market capitalization of the firms in portfolio 5 is also small. This is consistent with the conventional wisdom that analysts are reluctant to issue sell recommendations for firms that might generate future investment banking business, which presumably are the larger firms. The annual turnover of each portfolio is given in column 6. It is remarkably stable across the five portfolios, varying from a low of 425 percent for portfolio 2 to a high of 476 percent for portfolio 5. These numbers are relatively high, especially when compared to an annual turnover figure of 12 percent for the portfolio of all covered firms, 70 percent for the neglected firm 16

19 portfolio, and only 7 percent for a portfolio comprised of all the firms on CRSP. These high turnover numbers are driven by the fact that, conditional on receiving coverage, a firm changes portfolios 3.81 times per year, on average. Table IV also presents the estimated coefficients for the four-characteristic model. The significant coefficients on market risk premium, SMB, and HML (columns 7-9) for portfolio 1 are indicative of small growth stocks with higher than average market risk. The significant coefficients on SMB, HML and PMOM (column 10) for portfolio 5 reflect small value firms that have performed poorly in the past. The coefficient on the market risk premium generally decreases as we move from portfolio 1 to portfolio 5 whereas the coefficient on HML increases, indicating that less favorable analyst ratings are associated with firms of lower market risk and higher book-to-market ratios. Compared to covered firms as a whole, neglected stocks are, on average, less risky, smaller, and tilted toward value. Table V presents our main results, which strongly support the hypothesis that analysts recommendations have investment value. There is a monotonic decrease in raw returns (column 2) and market-adjusted returns (column 3) as we move from more highly to less highly recommended stocks. The cumulative market-adjusted returns for the five portfolios are plotted in calendar time in figure 1. The central message of our investigation is clear from this figure: more highly recommended stocks consistently outperform less highly recommended ones. Over the entire 11 year period, portfolio 1's cumulative market-adjusted return is close to 50 percent, while portfolio 5's cumulative return is nearly -90 percent, a 140 percentage point spread. One might conjecture that the patterns in market-adjusted returns can be explained by the market risk, size, book-to-market, and price momentum characteristics of the recommended 17

20 stocks. The intercept tests from the CAPM, the Fama-French three-factor model, and the fourcharacteristic model provide strong evidence that they cannot. In every case, the intercept tests (presented in columns 4, 5, and 6 and illustrated in figure 2) indicate that more highly rated stocks have higher abnormal returns than less highly rated stocks. Furthermore, the abnormal returns for portfolios 1 and 2 are each significantly positive in all three models while the abnormal returns for portfolio 5 are significantly negative. A comparison of the abnormal returns on portfolios 1 and 5 shows the return that can be generated from a strategy of purchasing the highest ranked securities and selling short the lowest ranked ones. It ranges from a low of percent per month under the four-characteristic model to a high of percent per month using the Fama-French three-factor model. Purchasing the second most highly rated stocks, those in portfolio 2, and selling short the second least favorably rated stocks, those in portfolio 4, also produces significant, although smaller, returns. They range from percent per month under the CAPM to percent using the Fama-French threefactor model. Table V also reveals that a portfolio of all covered stocks earns positive and significant abnormal returns, while the abnormal returns of neglected stocks are negative and significant. The return to be earned from purchasing the covered firms and selling short the neglected stocks ranges from a low of percent per month using the four-characteristic model to a high of percent under the CAPM. The underperformance of neglected stocks is consistent with evidence in McNichols and O Brien (1998) that analysts tend to drop coverage of firms that they expect to do poorly, rather than retain them and issue negative comments. In contrast to our empirical findings, Arbel, Carvel and Strebel (1983) document that during the 1970's neglected 18

21 firms actually earned superior returns. There are a few possible explanations for these seemingly contradictory results. First, Arbel et. al. restricted their attention to large firms (the S&P 500), whereas our neglected firms are relatively small. Second, some of their neglected firms actually had an analyst following them. Third, they did not control for possible book-to-market effects. (As we show, neglected firms have higher book-to-market ratios.) During their sample period of , high book-to-market firms outperformed low book-to-market firms by 57 basis points per month. IV. ADDITIONAL ANALYSES In this section we partition the analysts recommendations, first according to firm size, then by subperiod, and then by overall market direction. As will be apparent, the performance results within each partition generally match those for the sample as a whole. We also consider whether positive abnormal performance can still be achieved if investors delay acting on changes in consensus analyst recommendations for a short time. A. Firm Size There are several reasons to analyze investment returns on the basis of firm size. First, to the extent that there is less information publicly available about smaller firms, we would expect the investment performance of analysts recommendations to be greater for them. Further, consistent with Shleifer and Vishny (1997) and Pontiff (1996), it is likely that investors ability to 15 arbitrage away any excess returns will be smallest for these firms. In addition, it is important to 15 Shleifer and Vishny (1997) argue that arbitrage has only a limited ability to align prices with fundamental values and that this limitation is greatest among securities with high volatility (such as small stocks). Pontiff (1996) adds that arbitrage will be limited when transaction costs are relatively high (as is again the case for small stocks). 19

22 understand the extent to which analysts recommendations can generate excess returns for larger firms as well, as they represent a greater share of the investment opportunities available in the market. In short, we find the results of our analysis to be most pronounced for small and mediumsized firms. Among the few hundred largest firms, we find no significant difference between the returns of the most highly rated stocks and those that are least favorably recommended. If return differences are attributable to a market inefficiency (a notion we discuss in greater detail in Section VI), the mispricing occurs precisely where it would be expected -- among small and medium-sized firms, where information is least available, the costs of trading are high, and arbitrage is most risky. Table VI presents the returns for our size partition. Following the criteria used by Fama and French (1993), size deciles are formed on the basis of NYSE firm-size cutoffs and are adjusted annually, in December. Each ASE and Nasdaq firm is placed in the appropriate NYSE size decile based on the market value of its equity as of the end of December. Big firms (B) are defined as those in the top three deciles, small firms (S) are those in the bottom three deciles, and medium firms (M) are those in the middle four. Of all covered stocks, the number of small firms in our sample averages 1,957 per month, the number of medium firms averages 827, and the number of big firms averages 339. Though there are relatively few large firms (about 10 percent of all covered firms and only 5 percent of all firms), they represent approximately 70 percent of the total market capitalization of all firms listed on the NYSE, ASE, and Nasdaq. Medium and small firms represent 20 and 10 percent of total market capitalization, respectively. For all return models, the most highly recommended stocks earn positive abnormal 20

23 returns, while the least favorably recommended ones earn negative abnormal returns. The difference between the returns of the most highly and least favorably recommended stocks is significantly positive for small and medium-sized firms; it is statistically insignificant for large firms in two of the three return models. In all cases, the portfolio of all covered firms earns greater abnormal returns than does the neglected firm portfolio. The difference is significant for the small firms, regardless of the return model employed. It is significant for medium-sized firms in two of the three models, but is statistically insignificant for the large firms in all cases. B. Subperiod Analysis To investigate whether our results are driven by a few years in which highly recommended stocks outperformed less favorably recommended ones, we now turn to a subperiod analysis of our sample. Figure 3 plots the monthly raw returns for the five portfolios, cumulated by year. The abnormal return from purchasing portfolio 1 and selling short portfolio 5 ranges from a low of approximately 2 percent to a high of 25 percent, and is greater than 10 percent in seven of the eleven years. The yearly abnormal return from purchasing portfolio 2 and selling short portfolio 4, while smaller, remains positive in eight of the eleven years. Purchasing the covered firms and selling short the neglected stocks generates a positive abnormal return in nine of the eleven years. This provides confirming evidence that our results are robust across time. Additional confirmation comes from a partition of our sample into two time periods, the first covering and the second covering (table VII). We estimate separate regressions for each set of years and allow the coefficients of each of the factors to differ over the two periods. The estimated intercepts are not significantly different across periods. Fewer of 21

24 them are significant, though, likely due to a reduction in the power of our tests. (We have only 60 observations for the first subperiod and 72 for the second subperiod.) The abnormal return for the least favorably recommended stocks does remain significantly negative, and the abnormal return on portfolio 1 is again significantly greater than that of portfolio 5. Additionally, the covered firms earn a significantly higher return than the neglected firms in both time periods using the Fama-French three-factor model and in one subperiod for each of the other two models. C. Bull Market and Bear Market Returns As a bull market predominated during most of our sample period of , a question naturally arises as to whether the superior investment performance of highly recommended stocks has been driven primarily by the strong market. To address this question, we partition our sample period into bull and bear market months, where a bull (bear) month is defined as one in which the 16 CRSP value-weighted market index return is positive (negative). Using this definition, 68 percent of the sample months were categorized as bull markets and 32 percent as bear markets. We estimate separate intercepts for bull and bear months, but assume the coefficients on the factors are the same. This reflects the notion that while analysts may be better at picking stocks in up markets than in down markets, a firm s association with the various factors should remain the same, particularly given that bull and bear market months are scattered throughout the sample period. These regressions yield intercepts for the covered stock portfolios that are not significantly different across bull and bear markets (table VIII). As in the subperiod partition, fewer of the 16 Our analysis is similar in spirit to that of Lakonishok, Shleifer, and Vishny (1994) who analyze the returns of value and glamour stocks in both strong and weak markets. 22

25 abnormal returns are significant. With the exception of the market-adjusted returns, which do not take into account the greater market risk of highly recommended stocks, the abnormal return on portfolio 1 remains significantly higher than that on portfolio 5 and is greater in bear than in bull markets. Interestingly, covered firms, on average, appear to do worse than neglected firms in bear markets and better in bull markets. D. Delaying Investors Reaction to Changes in Consensus Analyst Recommendations The analysis thus far has assumed that investors act on a given consensus analyst recommendation at the close of trading on the day the consensus recommendation changes. An interesting question is whether it is crucial for investors to move so quickly. To address this question we calculate the returns that investors would have earned if they waited a short time before acting on changes in consensus recommendations. Table IX documents the investment performance that can be achieved on portfolios 1 through 5, assuming waiting periods of one calendar day (Panel A), 15 calendar days (Panel B), and 30 calendar days (Panel C). For a one-day waiting period the abnormal return for portfolio 1 remains significantly positive in two of the three models, but becomes insignificant when the waiting period is extended to 30 days. In contrast, the abnormal return for portfolio 5 remains 17 significant in all cases. (These results are illustrated in figure 4.) The abnormal return from a strategy of purchasing portfolio 1 and selling short portfolio 5 remains significantly greater than zero for waiting periods of one and 15 days; it is positive in two of the three models for a 30-day waiting period. While the magnitude of these returns decreases with the waiting period, these 17 Stickel (1995) and Womack (1996) also find the post-recommendation stock drifts in their samples to be stronger for the least favorably recommended stocks. 23

26 results suggest that investors need not immediately act on analysts recommendations in order to earn a positive abnormal return. V. TRANSACTIONS COSTS All returns presented thus far have been gross of the transactions costs associated with the bid-ask spread, brokerage commissions, and the market impact of trading. Brokerage commissions are relatively small, at least for institutions, while the market impact of trading is difficult to measure. The round-trip cost of the bid-ask spread has been estimated to be 1 percent for mutual funds (Carhart (1997)) as well as for individual investors (Barber and Odean (1998)). In conjunction with the calculated turnover for each portfolio (see column 6 of table IV), this percentage can be used to assess the impact of the bid-ask spread on investment returns. (The method for computing turnover was described in Section II.C.) A bid-ask spread of 1 percent implies, for each portfolio, a total transactions cost that is at least equal to 1 percent of its annual turnover. Referring again to column 6 of table IV, the minimum annual cost associated with portfolio 1 is then 4.56 percent. This compares to an annualized abnormal gross return of 4.2 percent, using the Fama-French three-factor model. This means that an active strategy of buying the most highly recommended stocks yields an abnormal net return that is not reliably different from zero. The annual transactions cost associated with a strategy of purchasing the stocks in portfolio 1 and selling short those in portfolio 5 is 9.32 percent (the sum of 4.56 percent for portfolio 1 and 4.76 percent for portfolio 5). This compares to an annualized abnormal gross return of 11.9 percent, again using the Fama-French three-factor model. While the net return is positive in this case, the funds invested could have alternatively 24

27 been used to buy risk-free securities, yielding much more than the approximately 2.6 percent net 18 return that this strategy provides. A strategy of purchasing a portfolio of all of the covered firms and selling short the neglected stocks costs 0.82 percent annually (the sum of 0.12 percent for the covered stocks and 0.70 percent for the neglected firms). This compares to a 3.9 percent abnormal gross return, using the Fama-French model. Again, while the net return is positive, it is less than could have been earned in a risk-free security. These results notwithstanding, the consensus recommendations remain valuable to investors who are otherwise considering buying or selling. All else the same, an investor would be better off purchasing shares in firms with more favorable consensus recommendations and selling shares in those with less favorable consensus recommendations. VI. SUMMARY AND CONCLUSIONS In this paper, we document that a strategy of purchasing stocks that are most highly recommended by security analysts and selling short those that are least favorably recommended yields a return of 102 basis points per month over the period. The magnitude of this return is very large, and is far greater than the size (negative 16 basis points) and book-to-market (17 basis points) effects for the same period. Even after controlling for these effects, as well as price momentum, we document that this strategy yields a return of 75 basis points per month. 18 A strategy of investing in a single portfolio must only earn an abnormal return greater than zero to be profitable, as the abnormal return is already net of the risk-free rate. In contrast, a strategy of investing in one portfolio and selling short another must earn an abnormal return greater than the risk-free rate to be profitable. This is because the portfolio that is sold short adds back this rate to the abnormal return. Consequently, the difference between the abnormal returns of the portfolio purchased and the one sold short is no longer net of the risk-free rate. Of course, if the proceeds from the short sale were freely available to the investor to invest in a risk-free security (which they are not, in general), the benchmark return for the zero-investment strategy would be zero as well. 25

28 Our results are robust to partitions by time period and overall market direction. We were surprised by the strength and robustness of our findings, especially since there is no extant theory of asset pricing that can explain them. This leaves three potential explanations for our findings: (1) random chance (that is, data-snooping), (2) a poor model of asset pricing, or (3) market inefficiency. Many financial economists (for example Fama (1998)) view documented patterns in stock returns with great skepticism and argue that the documented anomalies are simply a result of extensive data-snooping by academics. Even the relation between stock returns and either firm size or book-to-market ratio lacks strong theoretical foundations and, as such, has been attributed by some to data-mining (see, for example, Black (1993) and MacKinlay (1995)). We are sensitive to this interpretation of our results. However, it is unlikely that they are due to random chance, for three reasons. First, Wall Street firms spend billions of dollars annually on security analysis; it is difficult to understand why they do so if analysts recommendations do not have investment value. Second, if our results were driven by mere chance, they would be an 19 exceedingly rare outcome, occurring in less than 1 of 1,600 randomly generated portfolios. Third, our results are robust to several different partitions of the data. Every test of a null hypothesis that long-run abnormal stock returns are zero is implicitly a joint test of the hypotheses that (1) these returns are zero and (2) the asset pricing model used to estimate abnormal returns is valid. Assume that the results we document are not driven by chance, but, rather, are attributable to a poor model of asset pricing. This implies that highly 19 This is based on the t-statistic of for the difference between the intercepts for portfolios 1 and 5 using the CAPM. 26