《金融投资学》教学资源（英文文献）bad news travel slowly.pdf_P6-P10

Size,Analyst Coverage,and Profitability 271 In Table II,we examine the cross-sectional determinants of analyst cov- erage.When we actually implement our trading strategies in the next sec- tion,we run a separate regression every month to create our measure of residual coverage.Because the regressions look so similar month to month, we only present one set in Table II for illustrative purposes,corresponding to December 1988,which is around the midpoint of our sample period. Again,note that in each case,the regression is run only on those stocks that are larger than the 20th percentile NYSE/AMEX breakpoint in the given month. The first point to note is that unlike some previous researchers who have run similar regressions (e.g.,Bhushan(1989)and Brennan and Hughes(1991)) we use as our left-hand side variable log(1 Analysts),rather than the raw number of analysts.We do this because we ultimately want to use the re- siduals from our analyst-coverage regressions to explain momentum,and it seems plausible that one extra analyst should matter much more in this regard if a firm has few analysts than if it has many. In Model 1,we use OLS,and the only two right-hand side variables are log (Size),where Size is current market capitalization,and a Nasdaq dummy variable.7 The size variable is clearly enormously important,generating an R2 of 0.61.In Model 2,we add 15 industry dummies to the regression.8 This has a small effect,raising the R2 to 0.63. In Models 3 and 4,we try adding the firm's book-to-market ratio.We do this because book-to-market is known to forecast returns (Fama and French (1992),Lakonishok,Shleifer,and Vishny (1994))and we want to make sure that any return-predicting power we get out of analyst coverage is not sim- ply capturing a book-to-market effect.As it turns out,the coefficient on book- to-market is positive and significant,but it adds nothing at all to the R2. Thus it is unlikely that any of the results we report below are driven by anything to do with book-to-market.?In Models 5 and 6,we undertake a similar experiment with beta.10 The coefficient on beta is positive and strongly significant,and in this case,the R2increases marginally,going from 0.61 to 0.63 when we exclude industry dummies. 7 The Nasdag dummy is the only variable whose behavior changes much over the sample period.In earlier years,it is strongly negative,which is why we include it in our baseline model.However,by the late 1980s,it is typically positive,though not always significantly so. 8 The dummies correspond to the following grouping of two-digit SIC codes:(1)SIC 01-09; (2)SIC10-14;(3)SIC15-19;(4)SIC20-21;(⑤)SIC22-23:(6)SIC24-27;(7)SIC28-32:(8) SIC33-34;(9)SIC35-39:(10)SIC40-48;(11)SIC49;(12)SIC50-52:(13)S1C53-59:(14)SIC 60-69:and(15)S1C70-79. s Even if high-coverage stocks do have higher mean returns because they have a higher loading on book-to-market,this cannot explain our central result,namely that high-coverage stocks exhibit less momentum. 10 Throughout,we calculate beta with the Scholes-Williams(1977)method,using daily re- turns and the value-weighted CRSP index in the prior calendar year.We require that 50 per. cent of single-day trade-only returns(computed using closing prices,not bid/ask averages)be available.This is the same approach used by CRSP in its NYSE/AMEX Excess Returns File

In Table II, we examine the cross-sectional determinants of analyst coverage. When we actually implement our trading strategies in the next section, we run a separate regression every month to create our measure of residual coverage. Because the regressions look so similar month to month, we only present one set in Table II for illustrative purposes, corresponding to December 1988, which is around the midpoint of our sample period. Again, note that in each case, the regression is run only on those stocks that are larger than the 20th percentile NYSE0AMEX breakpoint in the given month. The first point to note is that unlike some previous researchers who have run similar regressions ~e.g., Bhushan ~1989! and Brennan and Hughes ~1991!! we use as our left-hand side variable log~1 1 Analysts!, rather than the raw number of analysts. We do this because we ultimately want to use the residuals from our analyst-coverage regressions to explain momentum, and it seems plausible that one extra analyst should matter much more in this regard if a firm has few analysts than if it has many. In Model 1, we use OLS, and the only two right-hand side variables are log ~Size!, where Size is current market capitalization, and a Nasdaq dummy variable.7 The size variable is clearly enormously important, generating an R2 of 0.61. In Model 2, we add 15 industry dummies to the regression.8 This has a small effect, raising the R2 to 0.63. In Models 3 and 4, we try adding the firm’s book-to-market ratio. We do this because book-to-market is known to forecast returns ~Fama and French ~1992!, Lakonishok, Shleifer, and Vishny ~1994!! and we want to make sure that any return-predicting power we get out of analyst coverage is not simply capturing a book-to-market effect. As it turns out, the coefficient on bookto-market is positive and significant, but it adds nothing at all to the R2 . Thus it is unlikely that any of the results we report below are driven by anything to do with book-to-market.9 In Models 5 and 6, we undertake a similar experiment with beta.10 The coefficient on beta is positive and strongly significant, and in this case, the R2 increases marginally, going from 0.61 to 0.63 when we exclude industry dummies. 7 The Nasdaq dummy is the only variable whose behavior changes much over the sample period. In earlier years, it is strongly negative, which is why we include it in our baseline model. However, by the late 1980s, it is typically positive, though not always significantly so. 8 The dummies correspond to the following grouping of two-digit SIC codes: ~1! SIC 01–09; ~2! SIC 10–14; ~3! SIC 15–19; ~4! SIC 20–21; ~5! SIC 22–23; ~6! SIC 24–27; ~7! SIC 28–32; ~8! SIC 33–34; ~9! SIC 35–39; ~10! SIC 40–48; ~11! SIC 49; ~12! SIC 50–52; ~13! SIC 53–59; ~14! SIC 60–69; and ~15! SIC 70–79. 9 Even if high-coverage stocks do have higher mean returns because they have a higher loading on book-to-market, this cannot explain our central result, namely that high-coverage stocks exhibit less momentum. 10 Throughout, we calculate beta with the Scholes–Williams ~1977! method, using daily returns and the value-weighted CRSP index in the prior calendar year. We require that 50 percent of single-day trade-only returns ~computed using closing prices, not bid0ask averages! be available. This is the same approach used by CRSP in its NYSE0AMEX Excess Returns File. Size, Analyst Coverage, and Profitability 271

Size,Analyst Coverage,and Profitability 273 In Model 7,we add to the industry-dummy specification of Model 2 a number of variables that are considered in Brennan and Hughes(1991):1/P, where P is the price of a share;the variance of daily returns;and five years worth of annual lagged returns.Although many of the coefficients are indi- vidually significant,the overall impression is that these extra variables are not very important in explaining the variation in coverage-jointly they raise the R2 from 0.63 to 0.65.11 In Model 8,we take the baseline specification of Model 1 and add a turn- over measure,defined as the number of shares traded over the prior six months divided by total shares outstanding.(Because turnover numbers may not have the same interpretation in a dealer market,we allow the coefficient on turnover to be different for Nasdaq firms.)Turnover is significantly pos- itively correlated with coverage on all exchanges,and it raises the R2 some- what,from 0.61 to 0.64.However,with this regression,one needs to be especially careful in attaching any causal interpretation.On the one hand,it is possible that turnover causes coverage:Analysts may be more inclined to follow naturally high-turnover stocks if this makes it easier to generate bro- kerage commissions for their employers(Hayes (1996)).On the other hand, Brennan and Subrahmanyam (1995)find evidence of causality running in the other direction:More analysts reduce the adverse-selection costs of trad- ing,and thereby attract a greater volume of trade.As we argue in Sec- tion II.D below,depending on which story one believes,it may or may not make sense to control for turnover in generating our measure of residual analyst coverage. Continuing in a similar vein,Model 9 adds to the turnover measure of Model 8 another proxy for transactions costs,a dummy variable that takes on the value one if the stock in question has listed options.(About 25 percent of our sample firms have listed options in 1988,with the fraction rising to 49 percent by 1996.)As can be seen,the options-listing dummy has the expected positive sign and is statistically significant.However,unlike turn- over,it adds virtually nothing to the explanatory power of the regression- the R2 remains at 0.64,just as in Model 8. Overall,the results in Table II make it clear that although a number of other variables are significantly related to analyst coverage,firm size is by far the dominant factor.Thus,in addition to worrying about the influence of these other variables,it is also important to think about potential nonlin- earities in the relationship between log(1+Analysts)and log(Size).In this spirit,we proceed as follows.We start in Section II.B by using the simple size-based regression in Model 1 as our baseline method for generating re- 11 Interestingly,our results call into question the conclusions of Brennan and Hughes(1991), who obtain significant positive coefficients on 1/P.In our regressions,we tend to get the op- posite sign.We conjecture that this arises because we are using log(1+Analysts)on the left. hand side,rather than the raw number of analysts.Because 1/P is correlated with firm size, and because firm size is of such dominant importance,any differences in how one models the analyst-size relationship is likely to have a strong influence on the 1/P coefficient

In Model 7, we add to the industry-dummy specification of Model 2 a number of variables that are considered in Brennan and Hughes ~1991!: 10P, where P is the price of a share; the variance of daily returns; and five years’ worth of annual lagged returns. Although many of the coefficients are individually significant, the overall impression is that these extra variables are not very important in explaining the variation in coverage—jointly they raise the R2 from 0.63 to 0.65.11 In Model 8, we take the baseline specification of Model 1 and add a turnover measure, defined as the number of shares traded over the prior six months divided by total shares outstanding. ~Because turnover numbers may not have the same interpretation in a dealer market, we allow the coefficient on turnover to be different for Nasdaq firms.! Turnover is significantly positively correlated with coverage on all exchanges, and it raises the R2 somewhat, from 0.61 to 0.64. However, with this regression, one needs to be especially careful in attaching any causal interpretation. On the one hand, it is possible that turnover causes coverage: Analysts may be more inclined to follow naturally high-turnover stocks if this makes it easier to generate brokerage commissions for their employers ~Hayes ~1996!!. On the other hand, Brennan and Subrahmanyam ~1995! find evidence of causality running in the other direction: More analysts reduce the adverse-selection costs of trading, and thereby attract a greater volume of trade. As we argue in Section II.D below, depending on which story one believes, it may or may not make sense to control for turnover in generating our measure of residual analyst coverage. Continuing in a similar vein, Model 9 adds to the turnover measure of Model 8 another proxy for transactions costs, a dummy variable that takes on the value one if the stock in question has listed options. ~About 25 percent of our sample firms have listed options in 1988, with the fraction rising to 49 percent by 1996.! As can be seen, the options-listing dummy has the expected positive sign and is statistically significant. However, unlike turnover, it adds virtually nothing to the explanatory power of the regression— the R2 remains at 0.64, just as in Model 8. Overall, the results in Table II make it clear that although a number of other variables are significantly related to analyst coverage, firm size is by far the dominant factor. Thus, in addition to worrying about the influence of these other variables, it is also important to think about potential nonlinearities in the relationship between log~1 1 Analysts! and log~Size!. In this spirit, we proceed as follows. We start in Section II.B by using the simple size-based regression in Model 1 as our baseline method for generating re- 11 Interestingly, our results call into question the conclusions of Brennan and Hughes ~1991!, who obtain significant positive coefficients on 10P. In our regressions, we tend to get the opposite sign. We conjecture that this arises because we are using log~1 1 Analysts! on the lefthand side, rather than the raw number of analysts. Because 10P is correlated with firm size, and because firm size is of such dominant importance, any differences in how one models the analyst-size relationship is likely to have a strong influence on the 10P coefficient. Size, Analyst Coverage, and Profitability 273

274 The Journal of Finance sidual analyst coverage.Next,in Section II.C we rerun all of our tests sep- arately for each of the size classes(except the very smallest)in Table I.In this case,we run a separate cross-sectional analyst regression each month for firms in the 20th-40th NYSE/AMEX percentiles,for firms in the 40th- 60th percentiles,and so on.Among other things,this approach allows the relationship between log(1 +Analysts)and log(Size)to take on a piecewise linear form,hopefully correcting any deficiencies that arise from imposing an overly simple linear structure on the entire sample. Moreover,in Section II.D we also report on sensitivity checks that take into account the potential for analyst coverage to be correlated with some of the other variables considered in Table II.For example,we experiment with alternative definitions of residual coverage based on Model 2,which in- cludes the industry dummies,and Models 8 and 9,which include turnover and the options-listing dummy.Furthermore,we redo our tests in terms of beta-adjusted returns in case the pronounced relationship between beta and analyst coverage is affecting the results. II.Momentum Strategies,Cut Different Ways A.Cuts on Raw Size We begin our analysis of momentum strategies in Table III.In this table, unlike in the tables that come later,we look at the entire universe of stocks without dropping those below the 20th NYSE/AMEX percentile.In so doing,we closely follow the methodology of Jegadeesh and Titman(1993) in many respects.In particular,we focus on their preferred six-month/six- month strategy,we couch everything in terms of raw returns,and we equal- weight these returns.But there are three noteworthy differences.First, our sample period from 1980 to 1996 is more recent.Second,we do not exclude Nasdag stocks.And third,our measure of momentum differs from theirs.They sort stocks into 10 deciles according to past performance,and then measure the return differential of the most extreme deciles-which they denote by P10-P1.In contrast,we place less emphasis on the tails of the performance distribution.We sort our sample into only three parts based on past performance:P1,which includes the worst-performing 30 per- cent;P2 which includes the middle 40 percent;and P3,which includes the best-performing 30 percent.Our basic measure of momentum is then P3 -P1.This is similar to the measure used by Moskowitz (1997)and Rouwenhorst (1997). We use this alternative,broader-based measure of momentum in order to generate better signal-to-noise properties for our tests.Unlike Jegadeesh and Titman (1993),we are not so much interested in establishing the exis- tence of momentum per se,but in comparing momentum effects across sub- samples of stocks.In some cases,we look at as many as 12 subsamples, when we sort by size and residual analyst coverage simultaneously.(See Table V below.)If we also were to use 10 performance deciles,we would end

sidual analyst coverage. Next, in Section II.C we rerun all of our tests separately for each of the size classes ~except the very smallest! in Table I. In this case, we run a separate cross-sectional analyst regression each month for firms in the 20th–40th NYSE0AMEX percentiles, for firms in the 40th– 60th percentiles, and so on. Among other things, this approach allows the relationship between log~1 1 Analysts! and log~Size! to take on a piecewise linear form, hopefully correcting any deficiencies that arise from imposing an overly simple linear structure on the entire sample. Moreover, in Section II.D we also report on sensitivity checks that take into account the potential for analyst coverage to be correlated with some of the other variables considered in Table II. For example, we experiment with alternative definitions of residual coverage based on Model 2, which includes the industry dummies, and Models 8 and 9, which include turnover and the options-listing dummy. Furthermore, we redo our tests in terms of beta-adjusted returns in case the pronounced relationship between beta and analyst coverage is affecting the results. II. Momentum Strategies, Cut Different Ways A. Cuts on Raw Size We begin our analysis of momentum strategies in Table III. In this table, unlike in the tables that come later, we look at the entire universe of stocks without dropping those below the 20th NYSE0AMEX percentile. In so doing, we closely follow the methodology of Jegadeesh and Titman ~1993! in many respects. In particular, we focus on their preferred six-month0sixmonth strategy, we couch everything in terms of raw returns, and we equalweight these returns. But there are three noteworthy differences. First, our sample period from 1980 to 1996 is more recent. Second, we do not exclude Nasdaq stocks. And third, our measure of momentum differs from theirs. They sort stocks into 10 deciles according to past performance, and then measure the return differential of the most extreme deciles—which they denote by P10 2 P1. In contrast, we place less emphasis on the tails of the performance distribution. We sort our sample into only three parts based on past performance: P1, which includes the worst-performing 30 percent; P2 which includes the middle 40 percent; and P3, which includes the best-performing 30 percent. Our basic measure of momentum is then P3 2 P1. This is similar to the measure used by Moskowitz ~1997! and Rouwenhorst ~1997!. We use this alternative, broader-based measure of momentum in order to generate better signal-to-noise properties for our tests. Unlike Jegadeesh and Titman ~1993!, we are not so much interested in establishing the existence of momentum per se, but in comparing momentum effects across subsamples of stocks. In some cases, we look at as many as 12 subsamples, when we sort by size and residual analyst coverage simultaneously. ~See Table V below.! If we also were to use 10 performance deciles, we would end 274 The Journal of Finance