Analysts Use of Public Information and the Profitability of their Recommendation Revisions

Transcription

1 Analysts Use of Public Information and the Profitability of their Recommendation Revisions Usman Ali* This draft: December 12, 2008 ABSTRACT I examine the relationship between analysts use of public information and the performance of their stock recommendations. I construct a measure of analysts reliance on public information (RPI) based on the sensitivity of their recommendation revisions to quantitative variables that are known to predict returns. I find a strong negative relationship between analysts past RPI and their future performance. A long/short trading strategy based on recommendation revisions of low RPI analysts yields an abnormal return of over 20% per year. My results suggest that high RPI analysts underperform not only because they have little ability to uncover private information, but also because they process the publicly available information poorly. *Yale University, Department of Economics and International Center for Finance ( usman.ali@yale.edu). I thank Nick Barberis for his helpful suggestions, comments, and support. I would also like to thank James Choi, Martijn Cremers, Will Goetzmann, Jonathan Ingersoll, Owen Lamont, Antti Petajisto, Geert Rouwenhorst, Robert Shiller, Paul Tetlock, Frank Zhang, and seminar participants at Yale University and the Chicago Quantitative Alliance 15 th Annual Conference. I gratefully acknowledge financial support from the Whitebox Advisors Doctoral Fellowship and the Chicago Quantitative Alliance academic competition. All errors are my own. 1

2 A large literature examines the investment value of analyst recommendations. However, there has been surprisingly little research on identifying analysts who have superior ability to identify undervalued and overvalued stocks. In this paper, I test whether we can measure an analyst s skill by examining how much he relies on publicly available information when issuing his recommendations. There are two competing hypotheses regarding analysts use of public information and their performance. The first hypothesis is from Kacperczyk and Seru (2007), who argue that a mutual fund manager s use of public information is negatively related to the manager s skill. An analyst who is really skilled can uncover new private information so he does not need to use public information. On the other hand, a less skilled analyst is likely to base his recommendations on public information. Therefore, according to this hypothesis, there should be a negative relationship between analysts use of public information and their performance. The second hypothesis states that analysts who use public information will do better because of the strong return predictive power of public variables. A large literature in asset pricing shows that certain publicly observable stock characteristics (e.g., past return, earnings surprise, turnover) predict future returns. Therefore, analysts who have little ability to generate private information can do quite well by simply basing their recommendations on these quantitative variables. To test these hypotheses, I construct a measure of analysts reliance on public information (RPI) 1 based on the sensitivity of analysts recommendation revisions to five pieces of public information: price momentum, earnings surprise, average turnover, change in consensus recommendation, and change in consensus earnings forecast. Previous research suggests that these variables predict cross-sectional returns. I rank analysts into quintiles based on their past RPI. I then use a calendar time portfolio approach to evaluate the future performance of analysts. Specifically, for each RPI quintile, I construct long and short portfolios following the directional advice of the analysts in that quintile. Consistent with the first hypothesis, I find a strong negative relationship between analysts past reliance on public information and the profitability of their subsequent recommendations. For example, a long/short portfolio that buys stocks upgraded and short sells stocks downgraded by analysts in RPI quintile 1 yields a highly statistically significant four-factor of 1.83% per month, approximately 1 I borrow this acronym from Kacperczyk and Seru (2007). 2

3 22% per year. On the other hand, the same portfolio constructed using recommendation revisions of analysts in RPI quintile 5 generates an insignificant abnormal return of -0.45% per month. It is surprising that profitable trading strategies cannot be constructed using recommendation revisions of high RPI analysts even though these analysts tend to use quantitative variables that forecast future returns. I show that the reason why high RPI analysts underperform is that they do not follow optimal quantitative strategies. First, I show that there is a significant lag between the release of public information and high RPI analysts recommendation revisions. For example, instead of upgrading a stock immediately after the company announces a positive earnings surprise, high RPI analysts wait for about six weeks to upgrade the stock. Therefore, their recommendations fail to capture a significant amount of the post-earnings announcement drift. I show that another reason why high RPI analysts underperform is that stocks recommended by these analysts do not belong to the set of stocks for which quantitative strategies are expected to be most profitable. My results suggest that low RPI analysts either have superior ability to access information or superior ability to process information about the firms that they cover, or both. In order to determine the relative importance of these two factors, I study the effect of Regulation FD, which was implemented to curb selective disclosure by firms to a subset of market participants, on my results. If low RPI analysts better performance is solely driven by their superior access to information, then their outperformance should be limited to the pre-regulation FD period only. I show that low RPI analysts also perform better in the post-regulation FD period, suggesting that these analysts also have superior ability to process information. I also test whether the market recognizes the differences in analysts stock-picking skills by examining whether the market reaction on the day of the recommendation revision is related to analysts past reliance on public information. Controlling for a number of factors associated with the market s reaction to recommendation revisions, I find that the market fails to recognize the differences in analysts skills, as measured by analysts past reliance on public information. My results are robust to alternative measures of RPI and to the inclusion of other public variables that predict returns in the calculation of RPI. I also test for a number of alternative explanations of the results. For instance, it could be that low RPI analysts simply cover fewer firms and industries and hence generate more profitable recommendations. I show that there is virtually no variation in the number of firms and industries covered by analysts across the RPI 3

4 quintiles. In addition, I find that analyst experience, herding, and timeliness of recommendation revisions are not driving my results. The remainder of the paper is organized as follows. Section I provides a brief literature review. Section II describes the data. Section III describes the construction of the reliance on public information measure. Section IV presents the main results and Section V provides robustness checks and some extensions. Section VI concludes. I. Background and Literature Review Most of the literature on analyst recommendations has focused on analysts in aggregate. Unlike the research on mutual funds, which has proposed a vast number of performance measures to identify skilled fund managers, there has been surprisingly little research on identifying skilled analysts. Stickel (1995) and Womack (1996) find that favorable (unfavorable) revisions in analyst recommendations are accompanied by positive (negative) returns at the time of the announcement followed by a post-recommendation-revision drift in the direction of the initial price reaction. Barber et al. (2001) take a calendar-time perspective and find that the portfolio comprised of most highly recommended stocks outperforms the one comprised of least favorably recommended stocks. Jegadeesh et al. (2004) examine the source of the investment value provided by analyst stock recommendations and changes in recommendations. They find that the predictive ability of the level of recommendations is not significant, after controlling for previously documented anomalies. On the other hand, a change in consensus recommendation contains information to predict returns that is orthogonal to a number of other predictive variables. A paper that is related to mine is Kacperczyk and Seru (2007). Kacperczyk and Seru argue that a mutual fund manager s use of public information decreases with the manager s skill and they show that the sensitivity of fund managers portfolio holdings to changes in public information, as measured by changes in analysts opinions, is negatively related to various measures of fund performance and to fund flows. In some ways, my paper provides further support for their idea by showing that analysts use of public information is negatively associated with their subsequent performance. However, Kacperczyk and Seru s hypothesis cannot fully explain my results. In principle, an analyst who relies on quantitative variables could do quite well because these variables forecast future returns. I go beyond Kacperczyk and Seru by 4

5 showing that analysts reliance on quantitative signals predicts poor performance not only because these agents have little ability to uncover private information, but also because they process the publicly available information poorly. II. Data The analyst recommendations data used in this study are from the IBES Detail Recommendations file. 2 The dataset covers the period from October 1993 (the month that the first recommendation appears in the IBES database) through December The important variables in each database record are the recommendation date, an identifier for the analyst making the recommendation, an identifier for the brokerage firm to which the analyst belongs, and a rating. IBES standardizes the analyst ratings and converts them to a numerical score between 1 and 5. A rating of 1 reflects a strong buy recommendation, 2 a buy, 3 a hold, 4 an underperform, and 5 a sell. For ease of interpretation, I reverse code the ratings so that more favorable recommendations receive a higher rating (e.g., 5= strong buy, 1= sell ). I examine all common stocks (CRSP share codes 10 and 11) in the IBES database that can be matched with their respective CRSP permanent numbers. The returns and price data are from the CRSP daily and monthly stock files. The accounting data and quarterly earnings announcement dates are taken from the CRSP/COMPUSTAT merged files and analyst forecast data are from the IBES forecast files. Table I reports the descriptive statistics for the sample. The number of analysts issuing recommendations increases from 1272 in 1993 to 3480 in The number of firms covered by analysts increases from 3148 in 1993 to 5345 in 1998 before decreasing to 3720 in The mean number of analysts per covered firm has generally been increasing over time. The average number of recommendation revisions per analyst in a year ranges between 1.6 and Since previous papers have documented that recommendation revisions are more informative than recommendation levels (e.g., Jegadeesh et al. (2004)), I study recommendation revisions from one rating level to another in this paper. 4 Figure I plots the average cumulative 2 I downloaded the recommendations file in February Therefore, it is not affected by the problems documented by Ljungqvist et al. (2008). Ljungqvist et al. note that any file downloaded after 2006 is the most historically accurate file available to researchers. 3 The average number of recommendation revisions per analyst is very low in 1993 because IBES coverage is incomplete in There are a total of 186,948 recommendation changes in the sample period. 5

6 market-adjusted returns for the 3-month (63 trading days) period following the recommendation revision date. For each stock i, the cumulative market-adjusted return on day t+k, where t is the recommendation revision date, is calculated as t+k t+k j =t 1 + r ij j =t 1 + r mj where r ij and r mj are the returns on stock i and the CRSP value-weighted market index on day j, respectively. The market response to recommendation changes is considerable. The mean event-day market-adjusted return for upgrades (downgrades) is 2.1% (-3.2%) and is significant at the 1% level. Consistent with previous studies, I find that the market reacts more strongly to downward revisions than to upward revisions. The cumulative abnormal return gradually increases to 4.38% by the end of the three months for upgrades and decreases to -4.35% for downgrades. Thus the market reaction to recommendation revisions is incomplete and prices continue to reflect the information contained in the revisions for up to three months after the event date. For upgrades, most of the post-recommendation-revision drift occurs within the first 30 days. The cumulative return increases to 3.8% by day 30. The return continues to increase until day 63 but at a much slower rate. For downgrades, most of the drift occurs within the first 15 days. III. Variable Construction As stated in the introduction, there are two competing hypotheses regarding analysts use of public information and their performance. According to the first hypothesis, analysts use of public information should be negatively related to the profitability of their subsequent recommendation revisions because really skilled analysts do not need to rely on public information. On the other hand, the second hypothesis states that analysts who use public information will perform better because of the strong predictive power of public variables. To test these hypotheses, I construct a measure of analysts reliance on public information (RPI) based on the sensitivity of analysts recommendation revisions to public information in the direction favored by the public information. I use five variables to proxy for the public information available to the analyst issuing the recommendation revision: price momentum, earnings surprise, average turnover, change in the consensus recommendation of all analysts besides the analyst issuing the recommendation revision, and change in the consensus 6

7 annual earnings forecast of all analysts besides the analyst issuing the recommendation revision. Previous research suggests that all five of these variables predict returns in the cross-section. 5 A potential caveat is that these five variables might only be a subset of the available public information. However, using a smaller set will make it harder for my tests to find a relationship between RPI and subsequent performance, as analysts who rely on some other form of public information will possibly be classified as not using public information according to my model. The set of variables discussed above contains most, if not all, of the variables that predict short-term returns. Since analyst recommendation revisions have short-lasting effects on prices, it is unlikely that analysts use variables that predict long-term returns (e.g., the book-to-market ratio). Consistent with this, I find a weak negative correlation between analyst recommendation revisions and the book-to-market ratio. Nonetheless, in Section V, I show that my results are robust to the inclusion of other quantitative variables that predict returns, including the book-tomarket ratio. Another potential concern is that some analysts may use the five quantitative signals in conjunction with their private information when issuing recommendation revisions. Again, this will make it harder for my tests to find a relationship between RPI and subsequent performance, since I will not be able to separate these analysts from those who use only the quantitative signals. I construct the measure of reliance on public information as follows. First, at the beginning of each month, for each analyst j, I estimate the following regression using the entire history of the analyst s recommendation revisions: I it j = α j + β m j RETP it + β s j SURP it + β t j TURN it + β f j confor it P it + β r j conrec it + ε it, (1) where I it is a binary variable which takes a value of +1 (-1) if the recommendation revision of stock i on day t by the analyst is an upgrade (downgrade), RETP it is the cumulative marketadjusted return of stock i during months -6 through -1 preceding the month of the recommendation revision, SURP it is the cumulative market-adjusted return over the four-day window, from day -2 to day 1, around the most recent quarterly earnings announcement date of 5 Jegadeesh and Titman (1993) provide evidence of price momentum and Bernard and Thomas (1989) provide evidence of earnings momentum. Lee and Swaminathan (2000) and Datar et al. (1998) show that average turnover negatively forecasts future returns. Jegadeesh et al. (2004) find that change in consensus recommendation predicts future returns and Stickel (1991), among others, shows that earnings forecast revisions predict future returns. 7

8 stock i before the recommendation revision date 6, TURN it is the decile rank of the stock s average monthly turnover in the six months preceding the recommendation revision month 7, confor it is the change, over the previous 30 days, in the consensus annual earnings per share forecast of stock i for the current fiscal year (FY1) of all analysts, excluding the analyst issuing the recommendation revision, P it is the price of the stock two months prior to the recommendation revision date, and conrec it is the change, over the previous 30 days, in the consensus recommendation of stock i of all analysts, excluding the analyst issuing the recommendation revision. confor it and conrec it are measured at the end of day t-1 to ensure that this information is available to the analyst when he issues the recommendation revision. Consensus forecast of a stock is calculated as the average of all forecasts of that stock issued or reiterated over the previous 365 days. Similarly, consensus recommendation of a stock is calculated as the average of all recommendations of that stock issued or reiterated over the previous 365 days. 8 For a recommendation revision to be included in the regression, I require that at least one analyst, other than the revising analyst, has an active forecast and recommendation for the stock on days t-1 and t I also require that the most recent quarterly earnings announcement date of the stock before the recommendation revision date be available from COMPUSTAT. In the second step, at the beginning of each month, I construct the measure of reliance on public information for analyst j as where SD k j RPI 1 j = β m j j SD RETP + β j j s SD SURP β j j t SD TURN + β j j f SD confor + β j j r SD conrec is equal to the standard deviation of explanatory variable k in regression (1). Therefore, each of the scaled coefficients measures the partial effect of one standard deviation change in the explanatory variable. The sign of the coefficients is important since we are interested in knowing whether the analyst revises his recommendations in the direction of the 6 To ensure that the information needed to calculate the earnings surprise is available to analysts before making the recommendation revision, I require that the earnings announcement be made at least 2 days before the recommendation revision date. Chan, Jegadeesh, and Lakonishok (1996) use a similar variable to measure earnings surprise and argue that it is a fairly clean measure of surprise since it does not require a model for earnings expectations. 7 The trading volume reported for NASDAQ stocks includes inter-dealer trades and, therefore, is not comparable to the trading volume of NYSE/AMEX stocks. To adjust for this effect, I compute the decile rank of average monthly turnover separately for NASDAQ and NYSE/AMEX stocks. 8 If an analyst issues multiple forecasts (recommendations) of the same stock over the previous 365 days, then I use the analyst s most recent forecast (recommendation) in the calculation of the consensus. 9 This ensures that I have the information to calculate confor it and conrec it. 8

9 public information (e.g., upgrade after a positive earnings surprise). Positive values of RETP, SURP, confor, and conrec and low values of TURN indicate favorable public information. Thus according to this measure, analysts with high values of RPI rely heavily on public information when revising their recommendations. III. A. Relation between recommendation revisions and public information variables In this subsection, I document that, in aggregate, analyst recommendation revisions are highly correlated with my five measures of public information. I run both univariate and multivariate regressions of the recommendation revision indicator (I) on my five measures of public information (RETP, SURP, TURN, confor, and conrec). For ease of comparison of coefficients across variables, I standardize each variable by dividing the difference between the variable and its mean by its standard deviation. I pool all observations together, instead of running separate regressions for each analyst. Since analyst recommendations tend to be clustered, the observations cannot be treated as independent. Therefore, I cluster the standard errors by the recommendation revision month to allow for arbitrary correlation of revisions that occur in the same month. The results, shown in Table II, suggest that individually, each of the five variables is highly significantly correlated with analyst recommendation revisions. Thus, in aggregate, analysts tend to upgrade (downgrade) stocks with high (low) returns over the past six months, positive (negative) earnings surprise over the past quarter, low (high) turnover over the past six months, and upward (downward) revisions in earnings forecasts and recommendations by other analysts over the past one month. The results from the multivariate regression are similar, except that the coefficient on confor becomes insignificant. IV. Results At the beginning of each month, starting in January 1997, I rank analysts into quintiles based on their RPI measures at the beginning of that month. I start in January 1997 to ensure that I have a long enough history of analysts recommendation revisions to accurately compute the first ranking. The first recommendation revision in the IBES database appears in November Thus, the January 1997 ranking uses 38 months of past data on recommendation revisions. For an analyst to be included in the ranking, he must have issued at least 16 recommendation 9

10 revisions for which data on the explanatory variables in regression (1) are available. This ensures that I have at least 10 degrees of freedom to estimate regression (1). The ranking is updated at the beginning of each month to reflect the information contained in new recommendation revisions. The results of the regressions suggest that the sample exhibits significant cross-sectional variation in RPI. For example, in January 2002, which is the middle of the sample period, the median values of RPI for quintiles 1 through 5 are -0.12, 0.08, 0.26, 0.45, and The results also suggest that the RPI measure is highly persistent. At the beginning of each year, I rank analysts into quintiles based on their RPI measure. Then for each analyst in each quintile, I calculate the average quintile rank three years later. The ranking does not change much over the three year period: the mean ranking for the bottom quintile rises to 1.4 and the mean ranking for the top quintile falls to 4.6. To examine the profitability of recommendation revisions, I use a standard calendar time portfolio approach. 11 At the beginning of each month, I rank analysts into quintiles based on RPI, as described above. For each quintile, I construct two portfolios: a long portfolio which consists of stocks upgraded by analysts in that quintile in that month and a short portfolio which consists of stocks downgraded by analysts in that quintile in that month. 12 Stocks are added to portfolios at the end of the recommendation revision date. I consider two holding periods: 1.5 months (31 trading days) and 3 months (63 trading days). 13 A stock is held in a portfolio until the end of the holding period or until the end of the trading day on which an analyst in the same RPI quintile revises his recommendation of the stock, whichever of the two dates comes first. So, for example, if a stock is upgraded by an analyst in RPI quintile 1 on day τ and another analyst in the same quintile downgrades the stock on day τ +10, the stock is removed from the long portfolio and added to the short portfolio at the end of day τ If several analysts in the same RPI 10 It appears that some analysts in RPI quintile 1 go against the quantitative signals but the economic magnitude of this effect is very small. For example, using the data that is used to calculate the January 2002 ranking, I find that mean price momentum decile rank of stocks upgraded (downgraded) by RPI quintile 1 analysts is 5.6 (6.0). Similarly, the mean earnings momentum decile rank of stocks upgraded (downgraded) by these analysts is 5.4 (5.6). Also, ranking analysts based on absolute value of RPI yields very similar results. 11 This approach is similar to the one used by Barber et al. (2001). 12 A recommendation revision is classified as an upgrade (downgrade) if the new IBES rating of the analyst is below (above) his previous rating. 13 As discussed in Section II, most of the post-recommendation-revision drift occurs in the first 30 days. Therefore, we expect portfolio returns to decrease as the holding period increases. 14 The results are almost identical if stocks are held in portfolios until the end of the holding period. 10

11 quintile revise their recommendations of the same stock on the same day then the stock is added to the long (short) portfolio at the end of the day if the number of upgrades on that day is greater (less) than the number of downgrades on that day. These steps ensure that a stock is not a part of both long and short portfolios on the same date and that a stock doesn t appear multiple times in the same portfolio on the same date. After determining the composition of each of the long and short portfolios as of close of trading on date τ -1, the value-weighted return of the portfolio on day τ is calculated as r pτ = n p,τ 1 i=1 x i n p,τ 1 r iτ x j =1 j where x i is the market capitalization of firm i at the end of the month prior to the recommendation revision month, r iτ is stock i s return on day τ, and n p,τ 1 is the number of stocks in the portfolio. Finally, for each month t in the sample period, the daily returns for each portfolio are compounded over the n trading days in the month to yield a monthly return, R pt : n R pt = τ=1(1 + r pτ ) 1. This approach yields a monthly time-series of returns for each portfolio. Note that these trading strategies can be implemented in practice since they are based entirely on observables. I set a minimum liquidity threshold by disallowing trading in stocks with a closing price below $5 at the end of the month prior to the recommendation revision month. This ensures that the portfolio returns are not driven by illiquid micro-cap securities. I calculate three measures of abnormal performance for each of the portfolios. First, I employ an intercept test using the CAPM. Specifically, I estimate the following monthly timeseries regression: R pt R ft = α p + β p R mt R ft + ε pt, (2) where R ft is the monthly risk-free rate and R mt is the month t return on the CRSP valueweighted market index. French (1993): Second, I estimate the intercept using the three-factor model developed by Fama and R pt R ft = α p + β p R mt R ft + s p SMB t + p HML t + ε pt, (3) where SMB t is the month t difference in the returns of a value-weighted portfolio of small stocks and a value-weighted portfolio of large stocks, and HML t is the month t difference in the 11

12 returns of a value-weighted portfolio of high book-to-market stocks and a value-weighted portfolio of low book-to-market stocks. Finally, I add Carhart s (1997) momentum factor to regression (3): R pt R ft = α p + β p R mt R ft + s p SMB t + p HML t + u p UMD t + ε pt, (4) where UMD t is the month t difference in the returns of a portfolio of high prior return stocks and a portfolio of low prior return stocks. 15 The estimated s from these regressions are used as measures of abnormal performance. Table III shows the first main result of the paper. For each of the five RPI quintiles, I report the monthly raw returns and s of the long, short, and long/short portfolios for the two holding periods. Panel A of Table III shows the raw returns and the four-factor s, Panel B shows the Fama and French s, and Panel C shows the CAPM s. Separating analysts by their past reliance on public information induces large differences in their subsequent performance. For example, for the 1.5 month holding period, the long/short portfolio constructed using recommendation revisions of analysts in RPI quintile 1 yields a highly significant fourfactor of 1.83% per month (t-statistic = 5.5). On the other hand, the long/short portfolio constructed using recommendation revisions of analysts in RPI quintile 5 generates a statistically insignificant of -0.45% per month. The difference in the four-factor s of the two long/short portfolios is economically large (2.28% per month) and statistically significant (tstatistic = 4.1). Panels B and C of Table III show that the same result holds if we measure abnormal performance using the Fama and French or the CAPM model. For the two holding periods, the s (from all three models) of the long/short portfolios decline monotonically as the RPI quintile rank increases from 1 to 5, and the differences in the s of the bottom and top quintile long/short portfolios are positive and significant, both economically and statistically. In fact, the raw and abnormal returns of the long/short portfolios for RPI quintiles 4 and 5 are not significantly different from zero. The results in Table III also suggest that both the upgrades and downgrades of RPI quintile 1 analysts are more profitable than the upgrades and downgrades of RPI quintile 5 analysts. These results strongly support the hypothesis that analysts past reliance on public information is negatively associated with their subsequent performance. 15 I obtain the monthly factors and the risk-free rate from Ken French s website. 12

13 Table IV reports the factor loadings for the top and bottom RPI quintile long, short, and long/short portfolios for the 1.5 month holding period. 16 The results suggest that the portfolios have similar exposure to the traded factors. IV.A. Source of underperformance of high RPI analysts The results shown in Table III are surprising since they suggest that profitable trading strategies cannot be constructed using recommendation revisions of high RPI analysts even though these analysts tend to use quantitative variables that forecast future returns. One possible reason why high RPI analysts recommendations underperform is that these analysts might not update their recommendations in response to the quantitative signals in a timely fashion. To assess this possibility, I examine how quickly high RPI analysts revise their recommendations in response to the quantitative signals. Table V reports the mean and median number of calendar day lags between RPI quintile 5 analysts recommendation revisions and the most recent quarterly earnings announcement date before the recommendation revision date, the last recommendation revision of the stock over the past three months by some other analyst, and the last annual earnings forecast revision of the stock over the past three months by some other analyst. The results suggest that high RPI analysts are quite slow to update their recommendations. For example, the average delay between the recommendation revision date and the last quarterly earnings announcement is 41 days. Therefore, high RPI analysts recommendations fail to capture a significant amount of the post-earnings announcement drift. Similarly, the average number of days between a recommendation revision by a high RPI analyst and the last revision of the stock by some other analyst is 32. As shown in Figure I, most of the post-recommendation-revision drift occurs within the first month after the revision. Thus, high RPI analysts recommendations fail to capture the post-recommendation-revision drift as well. I find that another reason why high RPI analysts underperform is that stocks upgraded/downgraded by these analysts do not belong to the set of stocks for which quantitative strategies are expected to be most profitable. The median momentum decile rank of stocks upgraded (downgraded) by RPI quintile 5 analysts is 6 (4). Similarly, the median earnings 16 The factor loadings are very similar for the 3 month holding period and are available on request. 13

14 surprise decile rank of stocks upgraded (downgraded) by RPI quintile 5 analysts is 6 (5). 17 Thus high RPI analysts do not follow optimal momentum strategies since most of the price (earnings) momentum profits are concentrated among stocks belonging to extreme decile portfolios. This also explains why RPI quintile 5 portfolios do not load that heavily on the momentum factor. It is interesting to examine how well high RPI analysts would perform if they did follow optimal quantitative strategies. Obviously, there is no simple answer to this question since the quantitative variables can be combined to form a number of different strategies. To illustrate the point that public information can be used to construct profitable strategies, I consider a simple price and earnings momentum strategy. At the beginning of each month, I use a two way independent sort to rank stocks covered by RPI quintile 5 analysts into 10 x 10 partitions based on their market-adjusted returns over the past six months and their most recent earnings surprises. I then construct a long/short portfolio that buys (short sells) the stocks that belong to both the top (bottom) momentum decile and the top (bottom) earnings surprise decile. The rolling portfolios are held for six months. All stocks are value weighted in a given portfolio, and the portfolios are rebalanced every month to maintain value weights. This strategy generates an average return (three-factor ) of 1.7% (2%) per month. 18 Clearly, high RPI analysts could perform very well if they made full use of public information. IV.B. An alternative measure of RPI In this subsection, I show that my results are robust to an alternative measure of RPI. I construct this alternative measure as follows. First, at the beginning of each month, I estimate regression (1) for each analyst, as before. In the second step, at the beginning of each month, I construct the alternative measure of reliance on public information,rpi j 2, for analyst j as RPI j 2 = β j m j SE m + β j s j β j t j + β j f j + β j r j SE s SE t SE f SE r where β k j and SE k j, k = m, s, t, f, r, denote, respectively, the coefficients and standard errors from regression (1). Scaling by standard errors adjusts for noise in the estimated coefficients and makes them comparable with each other. 17 I include all common stocks with analyst coverage in the rankings. The momentum (earnings surprise) ranking is calculated at the beginning of each month (quarter). 18 Of course, once we control for the momentum factor, the abnormal return of the portfolio becomes insignificant. 14

15 I then rank analysts into quintiles based on their RPI 2 measures and construct long/short portfolios as described in the previous section. Table VI shows the raw returns and the fourfactor s of the portfolios. 19 The results show that RPI 2 is also negatively associated with subsequent performance. For example, for the 1.5 month holding period, the monthly four-factor of the long/short portfolio is 1.85% and highly statistically significant for RPI 2 quintile 1 and -0.36% and statistically insignificant for RPI 2 quintile 5, and the difference between the two is also significant, both economically and statistically. IV.C. The effect of Regulation Fair Disclosure The results presented so far suggest that analysts who rely less on public information generate more profitable recommendations. It appears that these analysts either have superior ability to access information or superior ability to analyze information about the firms that they cover, or both. In order to determine the relative importance of these two factors, I examine the difference in analysts performance before and after the implementation of Regulation Fair Disclosure. Regulation FD was implemented by the SEC in October 2000 to curb selective disclosure by firms to a subset of market participants. If the better performance of low RPI analysts is solely driven by their ability to access information about the firms they cover, we expect this superior performance to attenuate post-regulation FD. On the other hand, if low RPI analysts also have greater ability to process information, their superior performance should be observable post-regulation FD as well. Table VII shows that low RPI analysts outperform high RPI analysts in pre- and post-regulation FD periods: the difference between the abnormal returns of the bottom and top RPI 1 quintile long/short portfolios is 2.23% (t-statistic = 2.4) and 1.93% (t-statistic = 3.0) per month in the two periods, respectively. The results also indicate that low RPI 1 analysts recommendation revisions were more profitable in the pre-regulation FD period (monthly =2.16%) than in the post-regulation FD period (monthly =1.43%). 20 Therefore, the outperformance of low RPI analysts can be attributed to both superior ability to access information and superior ability to process information about the firms they cover. 19 For brevity, I only report the four-factor s. 20 As shown in Table VII, both of these conclusions hold for the alternative measure of RPI. 15

16 IV. D. Market recognition of differences in analysts skills I test whether the market recognizes the differences in analysts stock-picking skills by examining whether the market reaction on the day of the recommendation revision is related to analysts past reliance on public information. To ensure that my results are due to past reliance on public information, I control for a number of other factors associated with the market s reaction to recommendation revisions. Specifically, I estimate the following regression: ExRet it = α t + β 1 Cange ijt + β 2 Rank jt + β 3 Cange ijt Rank jt + β 4 Size it + β 5 Cange ijt Size it + β 6 EDate it + β 7 Cange ijt EDate it + β 8 BROKSize jt + β 9 Cange ijt BROKSize jt + ε it, where ExRet it is the market-adjusted return of stock i on the recommendation revision date t, Cange ijt is a dummy variable which takes a value of 1 (0) if the recommendation revision of firm i by analyst j is an upgrade (downgrade), Rank jt is the revising analyst s RPI quintile rank, measured at the beginning of the month, Size it is the log of market-capitalization of firm i, measured at the end of the previous month, EDate it is a dummy variable which takes a value of 1 if the firm announced its quarterly earnings in the five-day window centered at the recommendation revision date and 0 otherwise, and BROKSize jt is the size quintile rank of the revising analyst s brokerage firm. Brokerage size is measured as the number of analysts employed by the brokerage firm in the previous year. The unit of observation in the regression is an individual recommendation revision. Since more than one analyst can revise his recommendation of the same stock on the same day, the observations in the regression cannot be treated as independent. To deal with this dependence problem, I eliminate observations for which multiple revisions of the same firm occur on the same date. To account for arbitrary within recommendation revision date correlation across firms, I cluster the standard errors of the regression by recommendation revision date. Since Cange is a dummy variable which takes the value of 1 for upgrades, the intercept and the coefficients on Rank, Size, EDate, and BROKSize measure the effect of these variables on the market reaction for downgrades. The coefficients on Cange and the interaction terms capture the differential market reaction associated with these variables for upgrades. We are primarily interested in the coefficients on Rank and the interaction term, Cange Rank. If market participants partially recognize the differences in analysts skills, as measured by their 16

17 reliance on public information, then the coefficient on Rank (Cange Rank) should be positive (negative). In other words, market should react more strongly to recommendation revisions of analysts with low values of RPI. We expect the coefficient on Size (Cange Size) to be positive (negative); market reaction should be smaller for recommendation revisions of large firms since more information is available for large firms. Since recommendation revisions from large brokerage firms are likely to result in greater market reaction due to greater influence of these brokerages, we expect the coefficient on BROKSize (Cange BROKSize) to be negative (positive). Ivkovic and Jegadeesh (2004) show that the market reaction to recommendation revisions is stronger for revisions made around earnings announcement dates. Therefore, we expect the coefficient on EDate (Cange EDate) to be negative (positive). Table VIII presents the results of the regression. As expected, market reaction is stronger for revisions of smaller firms, revisions made by analysts of larger brokerages, and revisions made around earnings announcement dates. Surprisingly, the coefficient on the interaction term, Cange Rank, is positive and statistically significant. The results suggest that, holding all other variables constant, an upgrade by an analyst in the highest RPI quintile yields a 0.24% (=5*( ) - 1*( )) higher market-adjusted return on the recommendation revision date than an upgrade by an analyst in the lowest RPI quintile. 21 Contrary to expectation, the coefficient on Rank is negative, although it is not significantly different from zero. Thus, according to this analysis, the market fails to recognize the differences in analysts skills. V. Extensions and Robustness Although the results are consistent with my hypothesis, there are a number of other plausible explanations of the data. In this section, I conduct a series of robustness tests and extend the analysis in some other directions. 21 Note that this difference is too small to explain the stronger post-recommendation-revision drift for RPI quintile 1 analysts. As shown in Table III, stocks upgraded (downgraded) by RPI quintile 1 analysts yield significantly higher (lower) returns than those upgraded (downgraded) by RPI quintile 5 analysts. As an additional check, I construct long/short portfolios by adding stocks to portfolios at the end of the last trading day before the recommendation revision date and find that all of my results hold. The abnormal (raw) returns of the long/short portfolios decrease monotonically as RPI quintile rank increases from 1 to 5 and the difference between the bottom and top quintile portfolio returns is large and significant. For example, for the 1.5 month holding period, the difference between the four-factor s of the bottom and top RPI 1 quintile long/short portfolios is 2.02% per month (t-statistic=3.69). 17

18 V.A. Other public variables that predict returns In this subsection, I show that my results are robust to the inclusion of two other variables that predict returns: the book-to-market ratio and accruals. Fama and French (1992), among others, show that high book-to-market firms outperform low book-to-market firms. To examine the effect of the book-to-market ratio, I add it as an additional explanatory variable in regression (1): I it j = α j + β m j RETP it + β s j SURP it + β t j TURN it + β f j confor it P it + β r j conrec it + β b j BTM it + ε it, (5) where BTM it is the log of stock i s book-to-market ratio, measured at the end of the most recent quarter for which the earnings announcement was made at least two days before the recommendation revision date. I then calculate the new reliance on public information measure as RPI 3 j = β m j j SD RETP + β j j b SD BTM + β j j s SD SURP β j j t SD TURN + β j j f SD confor + β j j r SD conrec where SD k j is equal to the standard deviation of explanatory variable k in regression (5). Table IX shows that the inclusion of book-to-market ratio has no effect on the basic conclusion; low RPI 3 analysts outperform high RPI 3 analysts. 22 Sloan (1996) finds that firms with high accruals have lower future returns than firms with low accruals. He argues that the accrual component of earnings is less persistent than the cash flow component of earnings but investors fail to recognize this difference in a timely fashion. I examine the effect of accruals on my results by adding a measure of accruals to regression (1) as an explanatory variable. Accruals are calculated as ACC q = CA q,q 4 Cas q,q 4 CL q,q 4 DCL q,q 4 DT q,q 4 Dep q (TA q + TA q 4 )/2 where q is the most recent quarter for which the earnings announcement was made at least two days before the recommendation revision date, CA is the change in current assets, Cas is the change in cash/cash equivalents, CL is the change in current liabilities, DCL is the change in 22 In Table X, I report the four-factor s (or DGTW adjusted returns) of long/short portfolios constructed using recommendation revisions of analysts in RPI quintiles 1 and 5. For brevity, I report the results for the 1.5 month holding period only. 18

19 debt included in current liabilities, DT is the change in deferred taxes, Dep is depreciation and amortization, and TA is total assets. RPI 4 j = β m j I then calculate the measure of reliance on public information as j SD RETP j β Acc + β j j s SD SURP j SD Acc β j j t SD TURN + β j j f SD confor + β j j r SD conrec Table IX shows that RPI 4 is also negatively associated with the subsequent performance of analysts, although the results are weaker than those presented in Section IV, probably due to the substantial reduction in sample size. The data needed to calculate accruals are not available for all firms because of incomplete COMPUSTAT coverage. This reduces the sample size by about one-third. V.B. Characteristic adjusted returns Daniel and Titman (1997) and Daniel et al. (1997) argue that characteristics can be better predictors of future returns than factor loadings. Following Daniel et al., I compute the riskadjusted returns ( DGTW ) as the difference between the stock s raw return and the return on a value-weighted portfolio in the same size, industry-adjusted market-to-book ratio, and one year momentum quintile. The results in Table IX show that low RPI analysts recommendation revisions earn higher characteristic-adjusted returns than high RPI analysts recommendation revisions. V.C. Analysis by size Previous research suggests that quantitative strategies are more profitable among small capitalization stocks (see, for example, Hong, Lim, and Stein (2000) and Fama and French (2008)). Therefore, we might expect that, among small stocks, the trading strategies based on recommendation revisions are roughly equally profitable across RPI quintiles. To examine this possibility, I split the sample into small and large firms (defined as firms below or above NYSE median market capitalization). Table IX shows that the trading strategy based on RPI quintile 5 analysts recommendations is profitable among small stocks. But even among small stocks, RPI quintile 1 analysts recommendations are significantly more profitable: the difference between the four-factor s of the bottom and top RPI quintile long/short portfolios is statistically significant 1.49% per month. 19

20 V.D. Task Complexity, Industry specialization, Herding, and Analyst experience One possible explanation of my results is that analysts who rely less on public information simply cover fewer firms and hence generate superior recommendations. Previous research has documented that an analyst s earnings forecasting performance is negatively related to the number of firms covered by the analyst, possibly because covering a smaller set of firms allows greater attention to each covered firm. 23 To assess this possibility, I examine the relationship between the number of firms covered by an analyst during a year and the RPI rank of the analyst at the beginning of the year. Table X shows that there is little variation in the number of firms covered by analysts across RPI quintiles. The mean (median) number of firms covered by analysts in a year is about 11 (10) for all quintiles. Thus, it is unlikely that my results are due to differences in the number of firms covered by analysts. Another possibility is that low RPI analysts cover fewer industries than other analysts and thus provide more profitable recommendations because of better information about the industries they cover. I explore this issue by examining the relationship between the number of two-digit SIC code industries covered by an analyst during a year and the RPI rank of the analyst at the beginning of the year. As shown in Table X, the mean (median) number of industries covered is about 3.5 (3) for all quintiles. These results suggest that industry specialization is not driving my results. Jegadeesh and Kim (2007) show that recommendations that move away from the consensus are more profitable. It is possible that low RPI analysts simply issue recommendation revisions that deviate from the consensus (i.e., anti-herd) and, therefore, their revisions are more profitable. To assess this possibility, I examine the profitability of herding recommendation revisions (i.e., revisions that move toward the consensus). I categorize a recommendation revision as moving toward the consensus if the absolute value of deviation from consensus is smaller for the new recommendation than for the old recommendation. Consensus recommendation is calculated as the average of the most recent recommendations of analysts, excluding the revising analyst, issued over the previous 365 days. Table IX shows that low RPI analysts issue more profitable revisions than high RPI analysts even for revisions that move toward the consensus. The monthly four-factor of the long/short portfolio is 1.81% and 23 See, for example, Mikhail et al. (1997) and Clement (1999). 20