Determinants of Superior Stock Picking Ability

Transcription

1 Determinants of Superior Stock Picking Ability Michael B. Mikhail Fuua School of Business Duke University Box Durham, NC (919) , office (919) , fax Beverly R. Walther Kellogg School of Management Northwestern University 2001 Sheridan Road Evanston, IL (847) , office (847) , fax Xin Wang Fuua School of Business Duke University Box Durham, NC (919) , office (919) , fax Richard H. Willis Fuua School of Business Duke University Box Durham, NC (919) , office (919) , fax November 12, 2004 PRELIMINARY AND INCOMPLETE We appreciate the financial support of the Fuua School of Business at Duke University and the Kellogg School of Management at Northwestern University. Analyst recommendations are from Zacks Investment Research. Errors or omissions are our responsibility. * Corresponding author.

2 Determinants of Superior Stock Picking Ability Abstract Prior research demonstrates that analysts whose recommendation revisions earned the most (least) positive excess returns in the past continue to outperform (underperform) other analysts in the future. We extend this work by investigating the determinants of persistent stock picking ability. We find that the best analysts, based on the profitability of their stock recommendations, tend to follow fewer industries and have better resources at their disposal. Analysts classified as superior, also tend to issue their recommendations before their peers, are more likely to issue recommendations within days of a uarterly earnings announcement, and are less likely to issue revisions that skip recommendation categories. Finally, superior analysts have a greater ability to predict which firms are more likely to experience deterioration in their future performance following downgrades. The number of firms followed and the ability to distinguish between firms within an industry do not appear to play an important role in determining persistent stock picking ability. JEL classification: G11; G14; G24 Keywords: Security analysts; Stock recommendations; Stock picking ability; Trading strategy

3 1. Introduction Mikhail, Walther, and Willis (2004) document persistence in sell-side analyst stock picking ability. They find that analysts whose recommendation revisions earned the most (least) positive excess returns in the past continue to outperform (underperform) other analysts in the future. We extend this work to investigate determinants of persistent stock picking ability for sell-side security analysts. To determine what factors associated with an analyst enable her to consistently outperform her peers, we examine a broad array of possibilities related to the individual (e.g., experience, timeliness, or predictive ability), the recommendation strategy employed (e.g., industries or firms chosen), and the work environment (e.g., brokerage size). We find that the best analysts, based on the excess returns earned by their prior stock recommendation revisions, tend to follow fewer industries and have better resources at their disposal. Analysts classified as superior, also tend to issue their recommendations before their peers, are more likely to issue recommendations within days of a uarterly earnings announcement, and are less likely to issue revisions that skip recommendation categories. Finally, superior analysts have a greater ability to predict which firms are more likely to experience deterioration in their future performance following downgrades. The number of firms followed and the ability to distinguish between firms within an industry do not appear to play an important role in determining persistent stock picking ability. Recent work examining analyst recommendations has focused on specific analyst characteristics and how they affect performance (e.g., Mikhail, Walther, and Willis, 2004) and on developing a more in depth understanding of the information underlying the summary recommendations (e.g., Asuith, Mikhail, and Au, 2004). Mikhail, Walther, and Willis (2004) 1

4 investigate if sell-side security analysts exhibit relative persistence in their stock picking ability. They find that analysts whose recommendation revisions earned the most (least) positive excess returns in the past continue to outperform (underperform) other analysts in the future. We extend this literature stream by examining the determinants of persistent stock picking ability. In Section 2 we summarize prior research. Section 3 describes the sample selection criteria and data. We discuss our empirical results in Section 4. Section 5 concludes and discusses future work. 2. Prior Research We draw on two strands of prior work in developing and motivating our investigation of the determinants of persistent stock picking ability: research examining short-window and long-window buy-and-hold returns associated with sell-side security analyst recommendation revisions, and research investigating differences in individual security analyst abilities and attributes. 2.1 Recommendation profitability Most previous research examines average cross-sectional short- and long-window returns associated with revisions (e.g., hold to strong buy or buy to sell). This evidence, consistent with positive (negative) abnormal returns for new buy (sell) recommendations, has existed for over two decades. Lloyd Davies and Canes (1978) examine the market reaction to stock suggestions appearing in the Wall Street Journal s Heard on the Street column. They find an event day return of 0.93% ( 2.37%) for new buy (sell) recommendations. Womack 2

5 (1996) finds that stocks added to (removed from) strong buy lists earned size adjusted returns of 2.98% ( 1.94%) while stocks added to (removed from) strong sell lists earned size adjusted returns of 4.69% (0.32%) in the 3-day event period surrounding the release of the recommendation revision. He also finds that size-adjusted returns to upgrades are insignificant in the six-month period following the revision while size-adjusted returns to downgrades are significantly negative, averaging 9.15% (see also Jegadeesh, Kim, Krische, and Lee, 2004). Recent studies by Barber, Lehavy, McNichols, and Trueman (2001), Stickel (1995), and Womack (1996), among others, document significant price reactions in short windows containing recommendation revisions and significant buy-and-hold returns in the months following the issuance of buy or sell recommendations, but do not provide evidence on differences across analysts in the profitability of their stock recommendations. Womack (1996) documents a significant price and volume reaction during the three-day window containing recommendation changes during He focuses on extreme changes, those recommendations revised to buy (from sell or hold) and those recommendations revised to sell (from buy or hold) by 14 major U. S. brokerage firms available on First Call. His evidence indicates that the market reacts to recommendations that are revised to buy or sell, suggesting that analyst stock recommendations contain new information. Further, he finds a positive and significant buy-and-hold return in the month following the issuance of buy recommendations and a negative and significant buy-and-hold return in the six months following sell recommendations. Stickel (1995) similarly documents a significant market reaction to extreme recommendation revisions using a sample of Zacks recommendation changes during 1988 to Unlike Womack, however, Stickel concludes that the buy-and-hold returns associated 3

6 with recommendations revised to buy or sell persist for at least six months and remarks that these effects appear to be permanent. Despite this evidence, research indicates that it is difficult for an individual investor to capitalize on these excess returns once transactions costs are considered. Barber, Lehavy, McNichols, and Trueman (2001) examine the compounded monthly excess return of portfolios formed daily on the basis of the outstanding consensus analyst recommendation. They document positive buy-and-hold returns to purchasing (selling) the stocks with the most (least) favorable consensus recommendation over , but find no evidence of significant positive buy-and-hold returns from this trading strategy once turnover and transaction costs are incorporated. These results, based on the average cross-sectional excess return to analysts recommendation revisions, are consistent with recommendation changes having information content (short-window tests) and profit-generating potential (long-window tests) absent transaction costs. This strand of research, however, did not focus on identifying individual analysts stock picking abilities, an extension provided in Mikhail, Walther, and Willis (2004). They find that analysts whose recommendation revisions earned the most (least) positive excess returns in the past continue to outperform (underperform) other analysts in the future. Their study, however, does not examine why some analysts are consistently better than others the focus of this study. Next we turn to prior work that examines performance and security analyst Deleted: a cost effective trading strategy may be realized characteristics. 2.2 Security analyst characteristics 4

7 Research has investigated whether sell-side security analysts differ systematically on several dimensions of performance including forecast bias, forecast accuracy, and (to a more limited extent) recommendation profitability. The findings on forecast bias and accuracy suggest that analysts are not homogeneous, and that investors can incorporate these systematic differences to form consensus earnings forecasts with greater predictive power. For example, Butler and Lang (1991) find evidence of systematic optimism and pessimism across analysts both in the overall sample and in subsamples formed on the basis of the Value Line earnings predictability rankings. Similarly, Sinha, Brown, and Das (1997) document systematic differences across analysts in forecast accuracy (see also Jacob, Lys, and Neale, 1999). Recent research has also documented that forecast accuracy varies systematically with characteristics of the analyst, such as experience (see Clement, 1999; Jacob, Lys, and Neale, 1999; Mikhail, Walther, and Willis, 1997), membership on the Institutional Investor All American Research Team (see Stickel, 1992), or prior accuracy (see Sinha, Brown, and Das, 1997). Relying on these findings, Brown (1999) shows that these factors can be incorporated to form forecasts with greater predictive power in hold out samples. Although stock picking is a substantially different task from forecasting earnings, these results suggest, and Mikhail, Walther, and Willis (2004) demonstrate, that certain analysts may consistently issue more profitable stock recommendations than other analysts. 1 Other prior evidence on cross-sectional differences in the profitability of stock recommendations is consistent with this evidence. For example, Mikhail, Walther, and Willis (1997) document that, at the time of the revision, the market reacts more strongly to upward (but not downward) 1 Mikhail, Walther, and Willis (1999) provide Pearson and Spearman rank correlations between two measures of forecast accuracy and two measures of stock recommendation profitability. While five of the eight correlations are statistically positive, the magnitudes of these correlations are small. Based on this evidence, they conclude that forecast accuracy and stock-picking ability are not closely related. 5

8 recommendation revisions issued by more experienced analysts. Controlling for the strength of the recommendation, the magnitude of the recommendation revision, firm size, the consistency of the contemporaneous earnings forecast revision, contemporaneous earnings releases, and the existence of periodic brokerage house reports, Stickel (1995) finds a significant association between the 10-day cumulative abnormal return centered on the recommendation date and the Deleted: return in the 10-day window Deleted: the recommendation is issued analyst s reputation and the size of the brokerage house employing the analyst. Barber, Lehavy, and Trueman (2000) similarly find that the profitability of stock recommendations is associated with brokerage house size; their results indicate that buy recommendations from larger brokerage houses are more profitable than those from small brokerage houses, while sell recommendations from small brokerage houses are more profitable than those from large brokerage houses. However, Barber, Lehavy, and Trueman do not find persistence in brokerage house performance, suggesting that investors cannot earn more positive excess returns by following the recommendations issued by certain brokerage houses. Their tests, however, do not allow for persistence in analyst performance. This extension is analyzed by Mikhail, Walther and Willis (2004) see discussion in Section 2.1. In summary, prior research has found that, on average, analyst stock recommendations are profitable and that analysts stock picking abilities are persistent; however, little work exists that attempts to explain these profitability differences in terms of the individual analyst s characteristics (see Mikhail, Walther, and Willis (1997) for an exception). We investigate whether the systematic differences across analysts in terms of the profitability of their stock recommendations, as reported by Mikhail, Walther, and Willis (2004), can be explained by specific attributes of the analyst (e.g., experience), the brokerage employing her (e.g., size or specialization), the firm or industry followed, or our proxy for the use of fundamental analysis. 6

9 Evidence on the determinants of superior stock picking ability can be used by investors to maximize their returns and by brokerages in training analysts. 3. Sample To investigate the determinants of stock picking ability we use the portfolios of ranked analysts developed in Mikhail, Walther, and Willis (2004). This sample, obtained from the Zacks Investment Research database (Zacks), contains 268,170 recommendation revisions and reiterations issued by 4,923 analysts for 7,845 firms during To be included in this sample, the current and previous recommendation must be available on Zacks to determine an analyst s revision or reiteration. 3 Furthermore, the sample was restricted to observations with sufficient CRSP and Compustat data to calculate the excess return metric used to rank analysts. Also, observations that correspond to a broker merger were eliminated. 4 Deleted: at least one of Deleted: two Deleted: s Deleted: s Analyst ranks are determined based on the relative differences in excess returns accruing to recommendation revisions for the preceding one- and three-year periods. 5 Each year, analysts are placed in uintiles based on the average characteristic-adjusted excess returns from taking a 2 This database contains a uniue analyst identifying code that allows us to follow an analyst through time, regardless of his or her employer. Zacks assigns each analyst s recommendation a value from 1.0 (strong buy) to 5.0 (strong sell), with a rating of 6.0 indicating that the analyst has initiated or discontinued coverage. If the analyst provides his or her recommendation to Zacks on a different scale, Zacks converts the recommendation to a 5.0 scale. We eliminate recommendations associated with 345 analyst codes on the Zacks database corresponding to either an unidentified individual or a brokerage house (e.g., Value Line). 3 Absent a Zacks rating of 6.0, indicating initiation or discontinuance of coverage, we reuire that the previous recommendation be issued less than one year before the current recommendation to calculate an analyst s recommendation revision or reiteration. 4 It appears that when a broker merger occurs Zacks assigns a new brokerage code to the merged entity and reenters all outstanding recommendations for the merged entity. For example, on June 4, 1999 an analyst associated with broker code OK discontinues coverage that is subseuently picked up by the same analyst on the same day at broker CL. OK and CL correspond to Deutsche Bank, which completed its merger with Bankers Trust on June 4, Because it is unlikely that the recommendations associated with CL on June 4, 1999 are new recommendations, we eliminate these observations. 7

10 long (short) position in their upward (downward) revisions. 6 For the one-year performance period, we use the uintile ranking of the analyst in the prior year as the analyst s historical performance ranking. For the three-year period, we average the analyst s uintile rankings over the prior three years, and place analysts into four groups based on the average of the historical uintile rankings. The firm s daily characteristic-adjusted excess return is calculated as the firm s compounded raw return minus the value-weighted compounded return on the characteristic-sorted benchmark portfolio to which the firm belongs in that year (see Daniel, Grinblatt, Titman, and Wermers, 1997; Wermers, 2000). 7 Each ranking represents the analyst s relative performance versus all other analysts on Zacks revising their recommendations during the same time frame. The best (worst) performers based on 1- and 3-year rankings earn 3-month excess returns of 5.51% (3.25%) and 5.81% (3.63%), respectively. Specific empirical analyses reuire additional data items, which are discussed when the variables are introduced. Table 1 provides descriptive statistics on the final sample of firm-year observations. As reflected in Panel A, the recommended firms in our sample are large with mean (median) sales of $4.8 billion ($1.1 billion) and mean (median) total assets of $9.9 billion ($1.5 billion). As expected, the sample firms have a significant analyst following; the mean (median) number of analysts following the typical firm in our sample is 16.9 (15.0), compared to 4.3 (1.0) for all firms on Compustat between 1985 and Most of the recommended sample firms (65.8%) are traded on the New York Stock Exchange, with a significant number, 30.5%, traded on the 5 Multiple time frames to measure past performance and rank analysts are used since we expect that future performance differences may vary depending on the extent to which analysts revisions have earned excess returns in the past. 6 All inferences are unchanged if we place analysts in uartiles or deciles. 7 All results are unchanged if we use raw returns or size-adjusted returns. We define the size-adjusted return as the firm s compounded raw return minus the mean value-weighted compounded return on the size decile portfolio to which the firm belongs. Decile assignments are based on the firm s market capitalization at the beginning of the calendar year. 8

11 NASDAQ; the remaining sample firms, 3.7%, are traded on the AMEX or a regional exchange. These numbers compare with 22.0% and 59.5% of all firms with available data on Compustat traded on the NYSE or the NASDAQ, respectively. The industry membership of the sample firms is diverse with 73 two-digit SIC codes represented (results not tabulated). Six industries contain 5% or more of the sample observations: SIC 28, Chemicals and Allied Products, (7.4% of sample observations versus 4.2% of all Compustat firms); SIC 35, Industrial Machinery and Euipment, (6.4% versus 5.2%) SIC 36, Electronic and Other Electrical Euipment, (5.6% versus 5.0%); SIC 49, Electric, Gas, and Sanitary Services, (7.1% versus 2.1%); SIC 60, Depository Institutions, (8.4% versus 7.0%); and SIC 73, Business Services, (6.7% versus 7.0%). As summarized in Panel B, the firms tend to be strong performers. Because financial ratios vary across industries and through time, we adjust all measures in Panel B by subtracting the comparable median performance measure for the industry (based on two-digit SIC code). Our sample firms have mean (median) industry-adjusted return on assets of 2.02% (1.81%) and return on euity of 0.80% (3.41%). The sample firms also tend to have greater growth prospects than the typical firm in their industry with mean (median) industry-adjusted price-toearnings (P/E) ratios of (3.94). Mean (median) industry-adjusted book-to-market (B/M) ratios are (-0.059). As shown in Panel C, the distribution of recommendations is not eual across the buy, Deleted: each panel hold, and sell categories, consistent with prior research (e.g., Asuith, Mikhail, and Au, 2004; Barber, Lehavy, McNichols, and Trueman, 2001). Using the typical five-point scale of strong buy, buy, hold, sell, and strong sell, approximately 55% of the current recommendations are strong buys or buys (compared with 47% in Barber et al.); 39% of the current recommendations 9

12 are holds (versus 34% in Barber et al.); the incidence of sell or strong sell recommendations is infreuent, occurring approximately 4% and 2% of the time, respectively. Overall, the distribution of recommendations in our sample is similar to those in prior work. Panel D report the buy-and-hold characteristic-adjusted abnormal returns compounded over the window (t = 2, t = +60), where t = 0 is the recommendation revision date, for analysts classified as superior and inferior based on their relative performance over the prior one-year and three-year periods. Using relative performance in the previous year to determine the best (worst) analysts based on taking a long (short) position in recommendation upgrades (downgrades), the best (worst) analysts revise recommendations that earn an excess return of 5.51% (3.25%); this difference is statistically different at two-tailed p < Results using a three-year period to determine analyst relative ranking yields similar results. These findings indicate that the differences in the portfolio of recommendations in year t for analysts identified as the best and worst analysts in year t 1 (or years t 1, t 2, and t 3 in the case of the three-year ranking) are substantial. 4. Empirical results 4.1. Univariate Results Table 2 provides several univariate comparisons for superior/inferior analysts in our sample in a preliminary attempt to ascertain whether or not any systematic differences exist. Specifically, we examine several analyst attributes, analyst timeliness, and analysts predictive ability. Panel A of Table 2 examines differences in the number of industries followed, number of firms followed, experience, analysts ability to distinguish between firms they follow in the same industry and the resources at their disposal, using brokerage size as a proxy. Several 10

13 differences are readily apparent. First, the best analysts tend to follow fewer industries and firms. The best (worst) analysts, based on a 3-year ranking period, follow (17.50) firms in 6.15 (6.96) industries. These differences are significant at p-values < Qualitatively similar results are observed for analysts ranked using a 1-year period. In addition to having fewer firms and industries to follow, the top performing analysts tend to have greater resources at their disposal. The average top performing analyst works at a brokerage with approximately 50 analysts while the typical poor performing analyst works at a much smaller brokerage with 35 analysts, on average. In addition to the recommendation strategy followed (e.g., number of industries or firms followed), and the work environment (e.g., brokerage house size), we examine several attributes meant to capture analyst specific ability. To determine whether or not analysts recommendations are the result of a broad industry analysis or firm specific analysis, we examine the distribution of recommendations within industry for analysts ranked as superior and inferior. If superior analysts conduct more firm-specific analysis, we expect this in depth investigation to yield greater excess returns and to result in greater variation in the distribution of their recommendations. We use REC_STD, calculated as the standard deviation of the most current recommendations issued by analyst i for all firms in industry j (three-digit SIC code) during the prior year, to capture the variation in an analyst s recommendations. The standard deviation is scaled by the mean value of all recommendations issued that are used in the calculation. We average the scaled standard deviation metrics across all industries followed by an analyst to get our final measure. Consistent with our predictions, we find that superior analysts tend to have more variation in their recommendations, within a particular industry, than inferior analysts. 11

14 REC_STD for the best (worst) analysts is (0.339) based on 1-year ranks, statistically different at p-value <0.03; results based on 3-year ranks are ualitatively similar. Finally, we examine analyst experience. We do not find a strong positive relation between analyst experience and performance contrary to predictions from prior work based on forecasting earnings (e.g., Mikhail, Walther, and Willis, 1997). Panel B of Table 2 examines several timeliness attributes that could potentially explain the greater market reaction observed around recommendation revisions for top ranked analysts. It may be the case that analysts who generate the most abnormal returns do not possess superior ability to distinguish between firms but rather simply report new information to market participants first. We measure timeliness using three constructs. First, we construct a leaderfollower-ratio similar in spirit to the construct developed in Cooper, Day, and Lewis (2001) for forecasting earnings. For each recommendation, we identify the N preceding recommendations and N subseuent recommendations issued by other analysts for the same firm. The leaderfollower-ratio for this recommendation is defined as T 0 /T 1, where T 0 is the cumulative numbers of days the N preceding recommendations lead the recommendation of interest and T 1 is the cumulative number of days the N subseuent recommendations lag the recommendation of interest. After computing the leader-follower-ratio for every recommendation, we then calculate the average leader-follower-ratio, LEADER, for each analyst every year. 8 We find that the best analysts, based on either our 1- or 3-year ranking period, have higher LEADER ratios than the worst analysts. Using our 3-year ranking scheme, we find that the best (worst) analysts have a leader follower ratio of (4.007) significantly different at p 8 We calculate LEADER for N=1, 2, 3, and 5 all our conclusions are ualitatively similar. For brevity, we tabulate results for N=1 only. 12

15 < Qualitatively similar results are observed for analysts ranked using our 1-year performance metrics. Second, we examine the freuency with which superior (inferior) analysts issue recommendations surrounding the release of a uarterly earnings announcement (QEA). If superior analysts are timelier, we would expect a greater proportion of their recommendation revisions to occur in the 9-day window centered on the QEA. Our results are consistent with superior analysts revising their recommendations on a timelier basis. We find that superior (inferior) analysts, based on a 3-year ranking period, issue recommendations 21.6% (15.8%) of the time within our event window. This difference is significantly different at p < To examine whether or not the revisions are in response to the QEA, we collect all recommendations during (t = 5, t = +5), where t = 0 is the uarterly earnings announcement date. We then calculate the proportion of recommendation from this subsample that are issued during the period (t = 0, t = +5). We find that 84.1% (77.1%) of the recommendations issued during (t = -5, t = +5) occur on or after the QEA for the superior (inferior) analysts suggesting that the revisions are primarily reactive rather than predictive with respect to announced earnings. Qualitatively similar results are observed using our 1-year ranking period. Finally, we examine the freuency with which an analyst s revision skips a recommendation level. If superior analysts revise their recommendations on a timelier basis, we would expect their revisions to be between contiguous levels. Consistent with our expectation, we find that superior (inferior) analysts, based on a 3-year ranking period, skip a recommendation level 34.9% (41.7%) of the time in the case of upgrades and 40.0% (45.5%) in the case of downgrades. Qualitatively similar results are observed using our 1-year ranking period. 13

16 Based on the above analysis, we conclude that superior analysts are timelier. However, we are unable to determine if they are also more adept at upgrading (downgrading) firms that are likely to experience improvement (deterioration) in their future performance. Panel C of Table 2 attempts to address this issue by exploring trends in firms financial performance following an upgrade or downgrade revision. We determine a firm s financial health by computing its F-score (see Piotroski, 2000) for every fiscal uarter. The uarterly F-score is a composite result computed by summing nine individual binary signals designed to capture a firm s financial health. Four of the nine binary signals measure performance-related factors in fiscal uarter : (1) F_ROA = 1 if ROA is positive; zero otherwise. ROA is calculated as uarterly net income before extraordinary items (Compustat # 8) for uarter scaled by total assets (Compustat # 44) at the beginning of the uarter; (2) F_ ROA = 1 if ROA (=ROA ROA -4 ) is positive; zero otherwise; (3) F_CFO =1 if CFO is positive; zero otherwise. CFO is defined as cash flow from operations scaled by beginning of uarter total assets (Compustat #44). For periods after 1987, cash flow from operations is defined as Compustat #108. In the period preceding 1987, cash flow from operations is measured as the sum of Compustat items #82 (funds from operations) and #73 (change in working capital) if the Flow of Funds Statement Format Code (FFSFC) is eual to 2, 3 or 5; CFO = Compustat item #82 less Compustat item #73 if FFSFC euals one; (4) F_ACCRUAL =1 if ACCRUAL <0; zero otherwise. ACCRUAL is defined as the current uarter s net income before extraordinary items less cash flow from operations, scaled by beginning-of-the-uarter total assets, that is, ROA -CFO. Three of the nine financial signals are designed to measure changes in capital structure and the firm s ability to meet future debt service obligations: (1) F_ LEVER =1 if LEVER 14

17 (=LEVER LEVER -4 ) is negative; zero otherwise. LEVER is defined as a firm s financial leverage, calculated as the sum of the current (Compustat #45) and non-current (Compustat #51) portions of long-term debt, scaled by beginning of uarter total assets (Compustat #44); (2) F_ LIQUID = 1 if LIQUID (=LIQUID LIQUID -4 ) is positive; zero otherwise. LIQUID measures the firm s liuidity as captured by the firm s current ratio defined as current assets (Compustat #40) at the end of uarter scaled by current liabilities (Compustat #49) at the end of uarter ; (3) EQ_OFFER = 0 if Compustat item # 84 is greater than zero, one otherwise. The use of external funds is perceived as a bad signal. The final two signals are designed to measure changes in operating efficiency: (1) F_ MARGIN =1 if MARGIN (=MARGIN MARGIN -4 ) is positive; zero otherwise. MARGIN is the firm s current gross margin defined as Net sales (Compustat # 2) in uarter less the cost of goods sold (Compustat # 30) for uarter, scaled by net sales (Compustat # 2) for uarter ; (2) F_ TURN = 1 if TURN (=TURN TURN -4 ) is positive; zero otherwise. TURN is the firm s current uarter asset turnover ratio, measured as Net sales (Compustat #2) for uarter, scaled by average total assets (Compustat #44) for fiscal uarter. A firm s uarter F_SCORE is the sum of the nine binary signals discussed above: F _ SCORE = F _ ROA + F _ ROA + F _ CFO + F _ TURN + F _ LEVER + F _ ACCRUAL + F _ LIQUID + F _ MARGIN + EQ _ OFFER We find that the best analysts tend to choose firms that are slightly less financially healthy than the worst analysts. The current uarter F-Scores (QTR t ) for firms followed by the best (worst) analysts, based on our 3-year ranking, are (5.205) significantly different at p-value < To determine if analysts in our sample are reacting to prior performance shifts or are adept at upgrading (downgrading) firms that are likely to experience improvement 15

18 (deterioration) in their future performance, we examine changes in F-Scores in the uarter immediately preceding the recommendation revision (QTR -1 ) and the four uarter period following the recommendation revision (QTR +4 ). We examine upgrades and downgrades separately for this analysis. In general, we do not find any significant differences in the performance of firms chosen by superior or inferior analysts in the uarter immediately preceding the recommendation revision. The only exception observed is for downgrades in cases where analysts are ranked based on their 1-year performance. In this case, we find that inferior analysts choose firms that have prior uarter drops in financial performance that are marginally significantly worse than those firms chosen by superior analysts. The prior uarter F-score changes for firms followed by superior (inferior) analysts are ( 0.108), significantly different at p-value <0.07. This finding suggests that poorly performing analysts are downgrading firms after performance deterioration to a greater extent than the best analysts. To assess predictive ability for downgrades, we examine the changes in F-scores for the four uarters following the revision. We find that firms downgraded by superior analysts experience significantly greater deterioration in performance than those downgraded by inferior analysts. F-score changes for the four uarter period following the revision are (-0.101) for the best (worst) analysts using our 3-year ranking; similar results are observed using 1-year ranks. Results are starker for upgrades. We find that firms upgraded by superior analysts experience improvements in performance, as evidenced by changes in their F-scores of (0.082), for our 3-year (1-year) ranks but that firms upgraded by inferior analysts experience deteriorations in performance changes are ( 0.075) for our 3-year (1-year) ranks. 16

19 4.2. Multivariate Analysis We examine the determinants of persistent stock picking ability for sell-side security analysts in a multivariate setting by estimating the following logit model. We confine our estimation to only those analysts who Mikhail, Walther, and Willis (2004) find to be the best and worst; that is, we omit analysts in the middle group. We focus on the two extremes in analyst stock picking ability (i.e., the best and the worst) because in the univariate tests we compare these two extremes and investors try to avoid the worst and pick the best. This choice also should increase the power of our tests to identify determinants of superior stock picking ability. 9 β 0 + β1num_indi,t + β 2NUM_FIRMSi,t + β 3REC_STD i,t + β 4EXPER + β 5BROK_SIZEi,t + β 6LEADER i,t + β 7NEAR_QEAi,t + β8skip_upi,t Pr ob( BEST i, t = 1) = G + β 9SKIP_DOWNi,t + β10qtrt i,t + β11up_qtr -1, i, t + β12dn_qtr -1, i, t + β13up_qtr + 4, i, t + β14dn_qtr + 4, i, t i,t Where: G = BEST i,t = NUM_IND i,t = NUM_FIRMS i,t = REC_STD i,t = EXPER i,t = The cumulative logistic distribution function; An indicator variable that euals 1 if analyst i is ranked as a superior performer in year t, zero otherwise; The total number of industries (three-digit SIC code) where analyst i follows at least one firm; The total number of firms for which analyst i made at least one recommendation; The average scaled standard deviation of recommendations for the firms in the same industry followed by analyst i (see appendix for details); The average analyst i s firm-specific experience measured as the number of calendar uarters in which analyst i made recommendations; 9 In future work, we also will perform this estimation for all analysts, not just the top and bottom performers. 17

20 BROK_SIZE i,t = LEADER i,t = NEAR_QEA i,t = SKIP_UP i,t = SKIP_DOWN i,t = QTRt i, t = UP_QTR -1,τ = DN_QTR -1,τ = The total number of analysts in analyst i s brokerage firm; The average leader-follower-ratio of analyst i s recommendations; The average indictor variables for every analyst i s recommendation where the indictor variable euals one if the recommendation was issued within ( 4, +4) relative to the earnings announcement date, zero otherwise; The average indicator variables for every analyst i s recommendation upgrade where the indicator variable euals one if the recommendation upgrade skips one or more recommendation ratings; The average indictor variables for every analyst i s recommendation downgrade where the indictor variable euals one if the recommendation downgrade skips one or more recommendation ratings; The average uarterly F-Score for the firm s fiscal uarter in which analyst i s recommendation occurs (see appendix for details); The change in F-Score in the fiscal uarter immediately preceding the current fiscal uarter containing the upgrade, averaged across all analyst i s upgrades in year t (see appendix for details); The change in F-Score in the fiscal uarter immediately preceding the current fiscal uarter containing the downgrade, averaged across all analyst i s downgrades in year t (see appendix for details); UP_QTR +4,τ = The change in F-Score in the four fiscal uarter period following the current fiscal uarter containing the upgrade, averaged across all analyst i s upgrades in year t (see appendix for details); and DN_QTR +4,τ = The change in F-Score in the four fiscal uarter period following the current fiscal uarter containing the downgrade, averaged across all analyst i s downgrades in year t (see appendix for details); Table 3 provides summary results from estimating the preceding logit for our one- and three-year ranking schemes. As expected, based on our univariate results, we find that the best analysts tend to work at the largest brokerages. Our brokerage size variable is significantly positive for both our 1-year (0.0058) and 3-year (0.0111) rank samples. If larger brokerages have superior resources and greater access to management, this result is not surprising. Large 18

21 brokerages also allow analysts to become more specialized. We find that NUM_IND, representing the number of industries followed, is negatively related to the likelihood an analyst is classified as a top performer. The coefficient is statistically negative at p-value < 0.00 regardless of the ranking scheme used (1-year rank: ; 3-year rank: ). This result suggests that analysts who restrict their coverage to fewer industries are more likely to develop the expertise necessary to become a top performer. The benefits accruing to an analyst from restricting the number of industries followed do not seem to apply to the number of firms followed. Our results suggest that the number of firms that an analyst follows does not affect his chances of being classified as a superior analyst. This finding implies that once an analyst perfects his skills in a particular sector, those skills can be applied to multiple firms in that industry without significant additional cost. Similarly, the extent of variation of an analyst s recommendations within an industry, our proxy for the extent of firm-specific analysis, does not appear to affect his or her chances of being a superior performer. These results suggest that analysts, on average, may generate a significant proportion of the excess returns observed around recommendation revisions by predicting industry rather than firm performance trends. The timeliness of an analyst s recommendation revisions also plays a role in determining whether or not he or she will be a top performer. We use three constructs to capture analyst timeliness. First, we examine whether or not superior analysts tend to lead or follow their peers when issuing recommendation revisions. We find that analysts who tend to issue recommendations sooner, as proxied for by our leader-follower ratio, LEADER, are more likely to be classified as superior. The coefficients on LEADER are statistically positive at p- values <0.02 for both our 1-year (0.0246) and 3-year (0.0517) ranked samples. Our second 19

22 proxy for timeliness, NEAR_QEA, examines the freuency with which an analyst issues recommendation revisions in close proximity to the release of a QEA. We find that analysts that are more likely to issue recommendation revisions around a QEA are more likely to outperform analysts who do not. We find that NEAR_QEA is significantly positive with p- values < 0.00 regardless of the ranking metric used. Our final proxies for timeliness, SKIP_UP and SKIP_DOWN, assume that analysts whose recommendation revisions skip a recommendation category are less timely. The only evidence we find consistent with this timeliness metric is for upgrades in our 3-year ranking sample. All other skip variables are insignificant. Next, we examine the effects of a firm s financial health and an analyst s ability to predict future performance using the firm s F-score. Our logit specification includes a firm s F- score (QTR t ) for the uarter in which the recommendation occurs, as well as QTR -1, representing the change in a firm s F-Score in the uarter immediately preceding the recommendation revision and QTR +4, representing the change in a firm s F-score in the four uarter period following the recommendation revision. QTR -1 and QTR +4 are separately computed for upgrades and downgrades. Two interesting results are observed. First, superior analysts tend to follow weaker firms QTR t is significantly negative with p-values < 0.00 for both the 1-year ( ) and the 3-year ( ) rank samples. Second, superior analysts appear to have predictive ability when choosing which firms to downgrade. DN_QTR +4 is significantly negative for both our subsamples at p-values < Finally, we examine analyst experience as a predictor of persistent stock picking ability. Mikhail, Walther, and Willis (1997) document that the market reacts more strongly to upward recommendation revisions that are issued by experienced analysts. Results for our experience 20

23 metric vary depending on which sample we examine. Experience for analysts ranked based on their 1-year performance is insignificant. Using a superior / inferior analyst designation based on 3-year performance, we observe a significantly negative coefficient on experience. Although this result is counterintuitive, it is consistent with a model proposed by Chen and Cheng (2003). They argue that an inexperienced analyst will choose a higher level of effort to compensate for investors uncertainty about her ability. In summary, these results indicate that the best analysts, based on the profitability of their prior stock recommendation revisions, tend to follow fewer industries and have better resources at their disposal. Analysts classified as superior, also tend to issue their recommendation revisions before their peers, are more likely to issue revisions within days of a uarterly earnings announcement and are less likely to issue revisions that skip recommendation categories. Finally, superior analysts have a greater ability to predict which firms are more likely to experience deterioration in their future performance following downgrades. The number of firms followed and the ability to distinguish between firms within an industry do not appear to play an important role in determining persistent stock picking ability. 5. Conclusion We find that superior analysts, those individuals whose prior recommendation revisions earn the largest excess returns, follow fewer industries and have better resources at their disposal. Analysts classified as superior, also tend to issue their recommendations before their peers, are more likely to issue recommendations within days of a uarterly earnings announcement and are less likely to issue revisions that skip recommendation categories. 21

24 Finally, superior analysts have a greater ability to predict which firms are more likely to experience deterioration in their future performance following downgrades. The number of firms followed and the ability to distinguish between firms within an industry do not appear to play an important role in determining persistent stock picking ability. In future work, we will investigate whether the determinants we identify result in a profitable trading strategy. 22

25 References Asuith, P., M. Mikhail, and A. Au. Information Content of Euity Analyst Reports. Journal of Financial Economics, Forthcoming. Barber, B., R. Lehavy, and B. Trueman. Are All Brokerage Houses Created Eual? Testing for Systematic Differences in the Performance of Brokerage House Stock Recommendations. Working paper, University of California at Davis, University of Michigan, and University of California at Berkeley, March Barber, B., R. Lehavy, M. McNichols, and B. Trueman. Can Investors Profit from the Profits? Security Analyst Recommendations and Stock Returns. The Journal of Finance 56 (2001): Brown, L. Predicting Individual Analyst Earnings Forecast Accuracy. Working paper, Georgia State University, December Butler, K. and L. Lang. The Forecast Accuracy of Individual Analysts: Evidence of Systematic Optimism and Pessimism. Journal of Accounting Research 29 (1991): Chen, X. and Q. Cheng. What determines the market impact of stock recommendations? Working Paper, University of Washington, Clement, M. Analyst Forecast Accuracy: Do Ability, Resources, and Portfolio Complexity Matter? Journal of Accounting and Economics 27 (1999): Daniel, K., M. Grinblatt, S. Titman, and R. Wermers. Measuring Mutual Fund Performance with Characteristic-Based Benchmarks. The Journal of Finance 52 (1997): Jacob, J., T. Lys, and M. Neale. Expertise in Forecasting Performance of Security Analysts. Journal of Accounting and Economics 28 (1999): Jegadeesh, N., J. Kim, S. Krische, and C. Lee Analyzing the analysts: When do recommendations add value? The Journal of Finance 59: Lloyd-Davies, P. and M. Canes. Stock Prices and the Publication of Second-Hand Information. Journal of Business Vol. 51, no. 1 (1978): Mikhail, M., B. Walther, and R. Willis. Do Security Analysts Improve their Performance with Experience? Journal of Accounting Research 35 (1997): Mikhail, M., B. Walther, and R. Willis. Does Forecast Accuracy Matter to Security Analysts? The Accounting Review 74 (1999):

26 Mikhail, B., B. Walther and R. Willis. Do Security Analysts Exhibit Persistent Differences in Stock Picking Ability? The Journal of Financial Economics 74 (2004): Sinha, P., L. Brown, and S. Das. A Re-Examination of Financial Analysts Differential Earnings Forecast Accuracy. Contemporary Accounting Research 14 (1997): Stickel, S. Reputation and Performance Among Security Analysts. The Journal of Finance 47 (1992): Stickel, S. The Anatomy of the Performance of Buy and Sell Recommendations. Financial Analysts Journal 51 (1995): Wermers, R. Mutual Fund Performance: An Empirical Decomposition into Stock-Picking Talent, Style, Transactions Costs, and Expenses. The Journal of Finance 55 (2000): Womack, K. Do Brokerage Analysts Recommendations Have Investment Value? The Journal of Finance 51 (1996):

27 Table 1: Descriptive Statistics on Sample Firms Panel A: Firm size and analyst following Sample Compustat Variable Mean Median Mean Median Assets 9, , , Sales 4, , , Analyst following Panel B: Industry-adjusted financial performance Variable Mean Median Return on assets 2.02% * 1.81% * Return on euity 0.80% 3.41% * Price-to-earnings * 3.94 * Book-to-market * Panel C: Distribution of current recommendations Recommendation Percent of Sample Strong buy 29.0% Buy 26.4% Hold 39.0% Sell 3.5% Strong sell 2.1% Panel D: Mean buy-and-hold abnormal returns Classification One-year ranking Three-year ranking Worst performers 3.25%* 3.63%* Best performers 5.51%* 5.81%* Difference 2.26%* 2.18%* * Statistically different from zero at two-tailed p < 0.01 using a t-test 25

28 Table 2: Univariate Results Panel A: Analyst Attributes 1 year ranking mean values 3 year ranking mean values Variable BEST WORST p-value BEST WORST p-value NUM_IND NUM_FIRMS REC_STD EXPER BROK_SIZE Panel B: Timeliness 1 year ranking mean values 3 year ranking mean values Variable BEST WORST p-value BEST WORST p-value LEADER NEAR_QEA SKIP_UP SKIP_DOWN Panel C: Firm Fundamentals (F_SCORE) 1 year ranking mean values 3 year ranking mean values Variable BEST WORST p-value BEST WORST p-value QTR t UP_QTR UP_QTR DN_QTR DN_QTR All reported p-values are two-tailed. 26

29 Table 3: Logit Estimation Results One-Year Ranking Three-Year Ranking Variable Coefficient Pr > χ 2 Coefficient Pr > χ 2 Intercept Attributes NUM_IND NUM_FIRMS REC_STD EXPER BROK_SIZE Timeliness LEADER NEAR_QEA SKIP_UP SKIP_DOWN Fundamentals QTR t UP_QTR DN_QTR UP_QTR DN+QTR LogL 3, , N 2,687 1,219 27