Analysts’ Recommendations and the Over …fmaconferences.org/Boston/ANALYST_REC_FMA2017_FULL_PAPER.pdfRunning head: Analysts’ Recommendations and the Over-optimistic Bias Analysts’

Running head: Analysts’ Recommendations and the Over-optimistic Bias

Analysts’ Recommendations and the Over-Optimistic Bias– From the

Perspective of the Asymmetric Effectiveness

Tao Li

Bentley University

Author Note

Tao Li, Department of Mathematical Science, Bentley University

Contact: [email protected]

ANALYSTS’ RECOMMENDATIONS AND THE OVER-OPTIMISTIC BIAS 2

2

Abstract

This paper examines the effectiveness of the analysts’ recommendations on US stocks and

investigates the analysts’ over-optimistic bias through the aspect of the effectiveness asymmetry.

The event analysis approach is applied to measure the short-term effectiveness around the

recommendation announcement, and the passive portfolio construction approach is applied to

examine the long-term effectiveness in terms of the investment value. Results of both approaches

show that analysts’ recommendations are effective and the effectiveness is consistently

asymmetric between “Buy (Upgrade)” and “Sell (Downgrade)”. The asymmetric recommendation

effectiveness reflects the analysts’ over-optimistic bias that leads to the difference in the research

quality with the favorable recommendations and unfavorable recommendations. The regulatory

effort in mitigating the over-optimistic bias is addressed through the investigation of the NASD

Rule 2711 and NYSE Rule 472 (now both superseded by FINRA Rule 2241). Through the

difference-in-difference model and quantile regression model, this paper shows that the two rules

have an overall effect in reducing the over-optimistic bias since its implementation. The analysis

that captures the dynamic nature of the regulatory effect further shows that the regulatory effect

does not remain the same during the implementation period. Through a series of the difference-in-

differences models applied in the sub-period during the rule implementation and a multivariate

time-series change point detection method to identify the most significant structural changes of the

recommendation asymmetry, this paper demonstrates that the mostly effective period of the

regulatory rule is only around the initial adoption of the rules and the post-financial crisis period,

and there is little improvement of the overoptimistic bias when the economy is normal.

Keywords: Analyst Recommendations, Event Analysis, Passive Portfolio Construction,

NASD Rule 2711 and NYSE Rule 472, Difference-in-Differences Model, Structural Change


3

Analysts’ Recommendations and the Over-Optimistic Bias– From the Perspective of the

Asymmetric Effectiveness

Research analysts play an important role in the financial markets as they are believed to

have superior ability in discovering information and conveying information to various market

participants. From the market efficiency perspective, analysts smooth the procedure of

information gathering, processing, and disseminating. The influence of analysts’

recommendations on stock prices has been studied by many researchers (B. Barber, Lehavy,

McNichols, & Trueman, 2001; B. M. Barber, Lehavy, McNichols, & Trueman, 2006; B. M.

Barber, Lehavy, & Trueman, 2007, 2010; Busse, Green, & Jegadeesh, 2012; Jegadeesh & Kim,

2006, 2010). These research offers evidence to support the claim that analysts’ recommendations

are effective. Researchers also find that analysts’ recommendations usually contain over-

optimistic bias (C.-Y. Chan, Lo, & Su, 2014; Clarke, Khorana, Patel, & Rau, 2011). The concern

of the analysts’ over-optimistic bias is raised from the 2000 financial crisis that resulted from the

technology and telecommunications bubble. Many stocks that had been touted by the analysts

prior to 2000 suffered significant losses during the crisis. Since then, the analysts’ over-optimism

received increasing concerns from the press, investors, and the regulators. Both the academia and

the regulators became aware of the issue of the analysts’ conflict-of-interests. The Global Analyst

Research Settlement was proposed to address such issue and a series of regulatory rules were

enforced. Among these regulatory efforts, the adoption of the NASD Rule 2711 and the NYSE

Rule 472 (now both superseded by NASD Rule 2411) was one of the important actions to reduce

the over-optimistic bias.

This paper presents a series of research analyses to evaluate the effectiveness of analysts’

recommendations and the regulatory effort in mitigating the over-optimistic bias. The first part of


4

the research provides an answer to the question of “Whether analysts’ recommendations are

effective?” The cross-sectional event analysis approach and the passive portfolio construction

approach are utilized in the analysis. The cross-sectional event analysis approach considers the

abnormal change of the stock return around the recommendation announcement, and the passive

portfolio construction approach focuses on the investment value of the portfolios that follow

analysts’ recommendations.

In the event analysis approach, statistical significant results are obtained regarding the

abnormal return of the recommended stocks around the recommendation announcement. Two

aspects of the recommendation classification are considered: the pure rating, and the change of

rating. Based on the pure ratings, “Strong Buy/Buy” recommendations are associated with

positive expected abnormal returns on the recommended stocks and

“Hold/Underperformance/Sell” recommendations are associated with negative abnormal returns.

Based on the change of the ratings, the recommendations are further classified into three sub-

types: “initiation”, “reiteration”, and “revision”. The analysis indicates that only the initiation

and the revision actions are associated with significant abnormal return, and the reiteration of the

same rating does not result in a significant change of the abnormal returns. Conditional on the

previous rating, the magnitude of the abnormal returns are found to be consistent with the

magnitude of the rating changes.

In the passive portfolio construction approach, the investment value of the

recommendations is measured from the performance of the portfolio that follows the content of

the recommendations. This paper considers the comparison among four different portfolios.

“BUY” portfolio and “SELL” portfolio utilize the pure ratings; “UPGRADE” portfolio and

“DOWNGRADE” portfolio use the change of the ratings. “BUY” portfolio adds any stock with a


5

“Buy/Strong Buy” rating after the announcement and excludes any stock from the portfolio when

a “Sell/Underperform” rating is issued; “SELL” portfolio follows the opposite direction.

Similarly, “UPGRADE” portfolio adds stock with an “Upgrade” action or a “Buy/Strong Buy”

initiation and excludes stocks with a “Downgrade” action or “Sell/Underperform” initiations,

while “Downgrade” portfolio reverse this trading strategy. Through the performance evaluation

on the four portfolios, this paper finds that “BUY” portfolio and “UPGRADE” portfolio achieve

a significantly positive alpha, while their counterparties have significantly negative alpha.

The second part of the paper focuses on the over-optimistic bias of the analysts’

recommendation and the regulatory effort in reducing this over-optimistic bias. In the analysis of

the recommendation effectiveness, this paper also document another important finding that

regards the consistent effectiveness asymmetry across different recommendation categories. In

the event analysis, the abnormal returns are shown to be asymmetrically distributed across the

pure rating and the change of rating. Based on the pure ratings, the “Buy” effectiveness is weaker

than the “Underperform” and the “Strong Buy” is weaker than the “Sell”. Based on the change of

rating, an “Upgrade” recommendation is generally weaker than its corresponding “Downgrade”

and an initiation of “Strong Buy/Buy” is weaker than the initiation of “Underperform/Sell”. This

asymmetric effectiveness is also discovered in the passive portfolio construction approach, where

the magnitude of the negative alpha from the “SELL”/“DOWNGRADE” portfolio is much

greater than the magnitude of the positive alpha from the “BUY”/“UPGRADE” portfolio. This

consistent asymmetry of the recommendation effectiveness reflects the low quality of the

analysts’ research report that contain favorable recommendation rating, which responds to the

analysts’ over-optimistic bias. This paper contributes to the literature of analysts’

recommendations by using this asymmetric recommendation effectiveness as the measurement


6

of the over-optimistic bias instead of the asymmetric distribution between the favorable ratings

and unfavorable ratings as proposed by B. M. Barber et al. (2006); C.-Y. Chan et al. (2014); C.-

Y. Chen and Chen (2013). The regulatory effort in mitigating the over-optimistic bias is assessed

through the examination of the NYSE Rule 472 and NASD Rule 27111. A two-step analysis is

adopted to investigate the regulatory effect. The first step evaluates the overall regulatory

effectiveness in reducing the over-optimistic bias with a difference-in-differences model. In this

step, the short-term recommendation effectiveness is regressed against the indicator of the

favorable recommendations (which are subjected to the over-optimism and has inferior short-

term effectiveness), the indicator of the post-regulation period, the difference-in-differences

estimator of the regulatory effect, and covariates with the stock specific information, analysts’

specific information, and recommendation specific information. The result shows that the overall

two regulation rules have significant influence in reducing the asymmetry of the

recommendation effectiveness. The analysis also finds that the following factors are negatively

associated with the recommendation effectiveness: the market capital value of the recommended

stock, the complexity of analyst’ current task, frequency of the recommendation updates; and the

following factors are in a positive relationship with recommendation effectiveness: the stock

beta, the analysts’ experience, the size of the brokerage firm that hires the analysts, the indicator

of the major financial crisis (internet bubble and 2008 financial crisis), the analyst’s revision

frequency, and the difference between the individual rating and the consensus rating. The results

regarding the covariates are all consistent with the findings by previous literature. To perform a

1 Both the NYSE Rule 472 and NASD Rule 2711 are no longer applicable. Both rules have been

superseded by FINRA Rule 2241 Research Analysts and Research Reports from the year of 2015, which

is effective since Sep 25, 2015.


7

further robustness check of the analysis, the difference-in-differences model is conducted

through a quantile regression and the result also provides similar conclusion in a qualitative

manner.

The second step of the analysis on the regulatory effect in reducing the analysts’ over-

optimistic bias focuses on the dynamic nature of the regulation implementation. In this step, a

series of difference-in-differences model are applied to each subsequent amendments2 made to

the two rules and the incremental rule effect is examined. The results show that only the first

three amendments and the 6th amendment are associated with significant regulatory effect. The

1st amendment (2002), the 3rd amendment (2004), and the 6th amendment (2012) have the desired

regulatory effect which reduces the asymmetry of the recommendation effectiveness, while the

2nd amendment (2003) shows a deterioration of the regulatory effect in which the asymmetry is

widened but not reduced. Both the 1st amendment and the 6th amendment occurred right after the

end of the major financial crisis. The remaining amendments are made either during the crisis

period or during period with normal economic situation, in which there is no significant

incremental regulatory effect. This finding indicates that the regulatory effect is largely driven by

memory of the bad economic and analysts tend to loosen their control over the over-optimistic

bias when financial situation is well, which implies that the rules lack a durable effect over the

whole implementation period. To check the robustness of this conclusion, a multivariate time-

2 NASD Rule 2711 was adopted on April 24, 2002, and there has been 7 amendments

implemented to this rule. The 1st amendment becomes effective on June 4, 2002; the 2nd amendment

becomes effective on July 29, 2003; the 3rd amendment becomes effective on April 26, 2004; the 4th

amendment becomes effective on June 6, 2005; the 5th amendment becomes effective on Sep 27, 2006;

the 6th amendment becomes effective on April 7, 2008; and the last amendments becomes effective on

April 5th and Oct 11th in the year of 2012.

NYSE Rule 472 was amended by SR-FINRA-2011-03, which is effective on Feb 4, 2013.


8

series change point detection method is performed to identify the structural change of the

asymmetry of the recommendation effectiveness. The detected change points provide the

confirmation about the conclusion regarding the regulatory effect.

This paper contributes to the existing literature of analysts’ recommendations in three

major aspects. Firstly, this paper uses a better measurement of the over-optimistic bias to

examine the regulatory effect of NASD Rule 2711 and NYSE Rule 472. Adhering to the original

recommendation rating offered by the IBES database, this research finds the consistent pattern of

the asymmetry of the recommendation effectiveness across the rating as well as the change of

rating. Previous researchers tend to provide general conclusions regarding analysts’ effectiveness

and they group the recommendations in sub-categories into a general “Buy” or “Sell” (B. Barber

et al., 2001; B. M. Barber et al., 2006), thus the asymmetric effectiveness across the sub-

categories cannot be fully addressed. In previous literature, the over-optimistic bias is mainly

addressed through the examination of the distribution of the recommendations (B. M. Barber et

al., 2006; C.-Y. Chan et al., 2014; C.-Y. Chen & Chen, 2013). This paper offers the perspective to

view the over-optimistic bias issue through the measurement of recommendation effectiveness

instead of the recommendation counts, which brings more close attention to the economic value

of the recommendations. Secondly, this paper discovers the dynamic nature of the regulatory

effect by using data that extends to the most recent period and covers the whole period when

NYSE Rule 472 and NASD Rule 2711 was in force. Previous research that focuses on the same

topic typically uses a much shorter time span (B. M. Barber et al., 2006; Casey, 2013; C.-Y. Chan

et al., 2014; Clarke et al., 2011; Howe, Unlu, & Yan, 2009; Loh & Stulz, 2011), thus the

investigation of the incremental policy effect is not applicable. This paper provides results that

capture the dynamic mechanism of the regulatory effect in reducing the over-optimistic bias and


9

identifies the periods that both rules are most effective. As suggested by the findings in this

paper, the regulatory effect does not remain constant and the most significant effect only occurs

right after the adoption of the rule and right after the end of the major crisis. Therefore, previous

papers that adopt a short post-regulatory period may not correctly evaluate the overall regulatory

effect. Moreover, it is important for the policy makers to be aware of the actual effectiveness of

the rules and the root cause of the rule effectiveness, thus knowing when the rules are effective

and when the rules might lose the desired functionality. This paper offers a two-step empirical

analysis to investigate the most significant regulatory effect for the two rules, and yields

confirmative conclusions to each other step. As a result, findings in this paper should be of

interest to the academic researchers, policy makers, and professional practitioners alike. Lastly,

this paper considers both the event analysis approach and the passive portfolio construction

approach to analyze the recommendation effectiveness. The results from both approaches exhibit

consistent result regarding the effectiveness and the effectiveness asymmetry, thus indicating that

the conclusions are solid and robust.

The paper is organized as follows. Section 1 reviews related research. Section 2 discusses

the data, the definition of variables, and presents some descriptive statistics. Section 3 is

dedicated to the analysis of the analysts’ recommendations effectiveness from both the event

analysis approach and the portfolio construction approach. Section 4 examines the regulator

effect of NASD Rule 2711 and NYSE Rule 472 in reducing the over-optimistic bias. Section 5

provides concluding remarks.


10

1. Literature Review

It has been well documented that there is information asymmetry between the stock

market participants and the management of the associated firms (Batabyal, 2012). Consequently,

investors use analysts’ recommendation reports to obtain extra information regarding the stock

performance. Existing literature concludes that there is a significant market reaction to the

release of an analyst’s recommendation, and that analyst’s recommendations contain informative

value to the investors (B. Barber et al., 2001; B. M. Barber et al., 2010; R. Brown, Chan, & Ho,

2009; Jegadeesh & Kim, 2006).

Two branches of research in the study of analysts’ recommendations are pertinent to this

paper. The first stream of research focuses on the recommendation effectiveness, and the major

underlying hypothesis is on the market efficiency (Fama, 1998; Jegadeesh & Titman, 1993), in

which the information of the analysts’ recommendations can be transferred into the abnormal

performance of the recommended stock. The recommendation effectiveness can be measured by

both the short-term market reactions to the recommendation announcement and the long-term

investment value of following the recommendations. Event analysis is the most common

approach to examine the market reactions of the stock responding to the recommendation. R.

Brown et al. (2009) use a (-1, +1) trading day window to measure the abnormal return around the

recommendation announcement, and they find that the change of the recommendation level is the

key factor to a significant cumulative abnormal return during the event, and the predictive power

of the recommendation is also affected by the relationship between the individual analyst and the

investment banking department of the analyst’s firm, the difference between the individual rating

level and the consensus rating level (which is also referred as herding issue), and the coverage of

the analyst. Jegadeesh and Kim (2010) change the event window to 1 day, 2 days, 21 days, and


11

up to 126 days since the revision, and they find that a large stock response occurs on the date for

the revision and the informative effect of the recommendation continues to be reflected in the

market price for up to six months. In the examination of the determinants of the market reactions,

they find that market reactions are stronger for “innovative” revisions that move away from the

consensus than those that move towards it. Similarly, Loh and Stulz (2011) use a (+1, +3) post-

revision event window to study the analysts’ recommendations effectiveness, and they find that

only 10% of the prevailing recommendations have significant influence. The most influential

recommendations are found to be away-from-consensus revisions, revisions issued

contemporaneously with earnings forecasts, and recommendations issued by leading analysts. To

explore the market reactions to the recommendations across different financial markets,

Jegadeesh and Kim (2006) extend the study across all G7 countries and they find that “Buy”

recommendations are more prevalent than “Sell”s in a global setting, and US analysts appear to

have superior skills over their peer analyst in other G7 countries. Murg, Pachler, and Zeitlberger

(2014) focus on the stocks on Austria financial market, and they discover that the analysts’

recommendations effectiveness depends on the firm’s size of the stock, the difference between

the target price of the recommendation and the actual price of the stock. Jiang, Lu, and Zhu

(2014) study the Chinese financial market, and they find that unlike the US stock market, the buy

recommendations in the Chinese market generate much stronger market reactions than the sell

recommendations. They explain this phenomena by the unique nature of the Chinese market,

where a short sale is restricted and the market participants are dominated by unsophisticated

individual investors instead of institutional investors. The long-term investment value of the

recommendations is generally examined through the portfolio construction approach. B. Barber

et al. (2001) examine the profitability of the investment strategy based on the content of the


12

analysts’ recommendations. They find that portfolios that follow the analysts’ recommendations

earn a substantial abnormal, but the benefit becomes insignificant after controlling for the

transaction cost. Howe et al. (2009) construct the portfolios based on the consensus ratings, and

they find that the changes in the aggregated recommendation ratings contain information of

future earnings at both the market level and the industry level. B. M. Barber et al. (2010) focus

on the comparison of performance among the portfolio based on ratings, portfolio based on

change of ratings, and portfolio based on the mixed contents. They document the findings that all

these portfolios have an incremental predictive power on security returns, and the portfolio that

uses the mixed information generates the highest performance.

The second research stream that is pertinent to this paper investigates the regulatory

effect in reducing the analysts’ over-optimistic bias. There are many aspects showing that

analysts’ recommendations contain over-optimistic bias. For example, the over-optimistic bias

can be reflected by the unequal percentage of buy recommendations to sell recommendations.

C.-Y. Chan et al. (2014) discover that analysts are more likely to issue buy recommendations,

and analysts are also more likely to revise the ratings to hold and less likely to revise to sell if a

downgrade action has to be performed. With introduction to the NASD Rule 2711 and NYSE

Rule 472, the count of the optimistic ratings has decreased. Clarke et al. (2011) study the

distribution of buy recommendations between the year of 2000 and 2007, and their findings

suggest that the independent analysts, affiliated analysts, and unaffiliated analysts all report

fewer favorable ratings in the post-regulation period of NASD Rule 2711. However, they also

find that the recommendation downgrades become less informative for all three types of analysts

following the NASD Rule 2711 implementation. B. M. Barber et al. (2006) find the percentage

of buys decreases starting from the mid-2000, which can be at least partly attributed to the


13

implementation of NASD Rule 2711. The over-optimistic bias could also be reflected by the less

informative content in the buy recommendations. Casey (2013) compares the recommendation

performance between independent research analysts and investment banking analysts by the

examination of buy-and-hold abnormal returns. She finds that in the post-regulation period,

upgrades generates greater market reaction for both types of analysts. Loh and Stulz (2011)

examine the influential capacity of the recommendations. They find that the analysts’

recommendations become more likely to be influential after the enforcement of the NASD Rule

2711.

2. Data

The recommendation data used in this paper is obtained from IBES database and covers

the period from the 01-Jan-1998 to 31-Dec-2015. The recommendation data contains the

following information: the date of the recommendation announcement, the time of the

announcement, level of the recommendation rating, the PERMNO code of the recommended

stocks, the analyst code, and the code for the brokerage firm that hires the analyst. IBES database

applies a 5-point scale on the recommendation ratings, in which 1 denotes ‘Strong Buy’, 2

denotes ‘Buy’, 3 for ‘Hold’, 4 for ‘Underperform’, and 5 for ‘Sell’. The recommendation data is

filtered to exclude non-US stocks, stocks missing a valid CUSIP/PERMNO code, and records

with non-identifiable analysts (i.e. missing the masked code for analysts). In the case that

multiple recommendations are issued by the same analyst during the same date on the same

underlying stock, only the last announcement is kept. The data is further filtered to exclude the

stocks that do not have at least one recommendation in both the pre- and post-regulation period

of NASD Rule 2711 and NYSE Rule 472. As the result of the filtering, the data employed in the


14

analysis contains 403,543 individual recommendations, issued on 4008 different stocks, and

covered by 13,459 individual analysts from 912 different research firms. The recommendation

data is also linked to the CRSP database by matching the CUSIP/PERMNO code to obtain the

information for the historical stock return, and stock sector/industry classification.

The recommendations are further classified into “Revision”, “Initiation”, or “Reiteration”

based on the change of recommendation ratings. If there is no previous recommendation found in

the database or the most recent recommendation has been issued by the same analyst more than

360 calendar days prior to the current recommendation, then that recommendation is treated as

not being covered by that analyst and coded as “Initiation”. If the stock is being actively covered

and the level of rating for the most recent recommendation by the same analysts remains

unchanged as the current record, then this recommendation is coded as “Reiteration”, otherwise

it is coded as “Revision”. In the case of “Revision”, if previous rating has a higher numeric

rating value, then it is further classed as “Upgrade”, e.g. a change from “Sell” (5) to “Buy” (2).

Otherwise, the revision is treated as “Downgrade”. As the result of such a classification, the

recommendation data contains 206,200 “Initiations”, which amounts to 51.10% of the total

records, and 42,566 “Reiterations” (10.55%), 81,095 “Downgrades” (20.10%) and 73,682

“Upgrades” (18.26%).

Table 3 presents the frequency distribution of the recommendations based on their

current ratings and previous ratings. The proportion of the “Strong Buy/Buy” recommendation

amounts to 49% of the total records, whereas the proportion of the “Underperform/Sell” is only

about 8.5%. This asymmetric distribution of the recommendations toward “Strong Buy/Buy” is

consistent with previous findings regarding the analysts’ over-optimism in generating more

favorable ratings (B. M. Barber et al., 2006; C.-Y. Chan et al., 2014; C.-Y. Chen & Chen, 2013).


15

In the case of a recommendation change (either “Initiations” or “Revision”), the most likely

ratings that analysts want to issue is a “Hold”. This inclination to issue “Hold” is consistent with

C.-Y. Chan et al. (2014)’s findings regarding the analysts’ optimistic rating. Table 4 summarizes

the distribution of recommendations based on the recommendation actions and the level of

ratings on a year-by-year basis. The count of the recommendations remains quite stable across

each year, with the only exception for the year 2002 when a doubled amount of recommendation

counts are found. Noticeable finding is about the Buy/Sell ratio, which clearly demonstrates the

regulatory effect of NASD Rule 2711 and NYSE Rule 472 in reducing the analysts’ optimism in

terms of the asymmetric distribution. Both rules were adopted in 2002, before that the yearly

Buy/Sell ratio is as high as more than 25. After the year of 2002, this ratio dropped to around 5

and kept a decreasing trend until the year of 2010 when it bounced back to above 5. Nonetheless,

during the whole post-regulation period, the Buy/Sell ratio consistently remains at a relative low

level, and never rise above 6. As indicated by the literature review, most existing research does

not include data after 2010 (B. M. Barber et al., 2006; B. M. Barber et al., 2010; Casey, 2013; C.-

Y. Chan et al., 2014; C.-Y. Chen & Chen, 2013; Clarke et al., 2011; Loh & Stulz, 2011). This

paper finds the year of 2010 changes the monotonically decreasing trend of the Buy/Sell ratio.

Consequently, including the recommendations that are issued after 2010 should provide

additional findings regarding the study of the NASD Rule 2711 and NYSE Rule 472.

3. The Effectiveness of Analysts’ Recommendations

This section analyzes the effectiveness of analysts’ recommendations. The short-term

effectiveness of the analysts’ recommendations is measured by instantaneous market reaction

around the recommendation announcement (R. Brown et al., 2009; Jegadeesh & Kim, 2010), the


16

long-term effectiveness is measured by the profitability in the investment activity (B. Barber et

al., 2001; B. M. Barber et al., 2010; Howe et al., 2009). This paper considers both measures of

the recommendation effectiveness and applies the corresponding approaches. The cross-sectional

event analysis approach is applied to examine the instantaneous market reaction, and the

portfolio construction approach is used to examine the investment value. The following section

begins with the event analysis.

Event Analysis

The event analysis approach is widely used by many empirical researchers in finance,

economics and other business disciplines to analyze the instantaneous market reactions around

the time that an event occurred. This method is particular important in testing the market

efficiency in capital market research (Kothari & Warner, 2004).

Method

In this paper, the individual recommendation announcement denotes the event of interest.

The announcement date of is defined as t=0 for “the event”, and a (-T, +T) trading-day window is

applied to examine the abnormal returns around the event. For any individual event k with an

event window (-T, +T), both the cumulative abnormal return 𝐶𝐴𝑅𝑘(𝑇) and the buy-and-hold

abnormal return 𝐵𝐻𝐴𝑅𝑘(𝑇) are calculated to measure the market reaction associated with the

event k. The two abnormal return are calculated as follows:

𝐶𝐴𝑅𝑘(𝑇) =1

2𝑇+1∑ 𝐴𝑅𝑘,𝑡

𝑡=𝑇𝑡=−𝑇 =

1

2𝑇+1∑ (𝑅𝑘,𝑡 − 𝑅𝑏𝑒𝑛𝑐ℎ,𝑡

𝑡=𝑇𝑡=−𝑇 ) (1)

𝐵𝐻𝐴𝑅𝑘(𝑇) = ∏ (1 + 𝑅𝑘,𝑡)𝑡=𝑇𝑡=−𝑇 − ∏ (1 + 𝑅𝑏𝑒𝑛𝑐ℎ,𝑡)𝑡=𝑇

𝑡=−𝑇 (2)

where 𝐴𝑅𝑘,𝑡 is the daily abnormal return of event 𝑘 on date 𝑡, 𝑅𝑘,𝑡 is the daily return of

the event 𝑘 on date 𝑡 in the defined event window, and 𝑅𝑏𝑒𝑛𝑐ℎ,𝑡 is the benchmark return for the

recommended stock on date 𝑡 in the event window. There are four common choices for the


17

benchmark return 𝑅𝑏𝑒𝑛𝑐ℎ, which are the pure market return as given by the return of CRSP

value-weighted market index, and the equilibrium rate of return of the stock as specified by the

CAPM model (Jensen, Black, & Scholes, 1972), Fama and French (1993) 3-factor model, and

Carhart (1997) model. Equation (3) to (5) specify the three models respectively.

𝑅𝑏𝑒𝑛𝑐ℎ(𝐶𝐴𝑃𝑀)

= 𝛽(𝑅𝑚𝑘𝑡 − 𝑟𝑓) + 𝛼 + 𝑟𝑓 (3)

𝑅𝑏𝑒𝑛𝑐ℎ(𝐹𝐹)

= 𝛽1(𝑅𝑚𝑘𝑡 − 𝑟𝑓) + 𝛽2 ∗ 𝑆𝑀𝐵 + 𝛽3 ∗ 𝐻𝑀𝐿 + 𝛼 + 𝑟𝑓 (4)

𝑅𝑏𝑒𝑛𝑐ℎ(𝐶𝑎𝑟ℎ𝑎𝑟𝑡)

= 𝛽1(𝑅𝑚𝑘𝑡 − 𝑟𝑓) + 𝛽2 ∗ 𝑆𝑀𝐵 + 𝛽3 ∗ 𝐻𝑀𝐿 + 𝛽4 ∗ 𝑀𝑂𝑀 + 𝛼 + 𝑟𝑓 (5)

In the above models, 𝑅𝑚𝑘𝑡 is the rate of return for the market portfolio as given by the

CRSP value-weighted market index, and 𝑟𝑓 denotes the risk-free asset rate of return as given by

the yield of one-month Treasury bill. 𝑆𝑀𝐵 is the difference between the return on the portfolio of

“small” capitalized stocks and “big” capitalized stocks, 𝐻𝑀𝐿 is the difference between the return

on the portfolios of “high” and “low” book-to-market stocks, and 𝑀𝑂𝑀 is the difference

between the return on portfolio of past one-year “winners” and “losers”3. The parameter 𝛼

represents the excess return for the underlying asset, and parameter 𝛽s denotes the sensitivity to

each factors. This paper uses the Carhart Model to evaluate the abnormal return in the main

analysis, and similar analysis with other model specifications are performed as a robustness

check. Figure 1 demonstrates the time outline that is adopted in the event analysis. A [-60, -15]

trading-day window is used to estimate the parameters in the benchmark models. The next [-14, -

T-1] trading-day window is served as the gap between the end of estimation period and the

beginning of the event window, which is intended to prevent the contagion of any market

3 For detailed definitions, see (Carhart, 1997).


18

information in the estimation period from reaching the event window, thus avoiding potential

bias in the event analysis.

[Figure 1 here]

To assess the statistical significance of the abnormal return within any recommendation

sub-category, this paper uses Patell’s Z test (Patell, 1976), cross-sectional test (S. J. Brown &

Warner, 1985), BMP test (Boehmer, Masumeci, & Poulsen, 1991), and skewness-adjusted t-test

(Hall, 1992) to test whether the abnormal return differs significantly from zero within the [-T,+T]

event window. To perform these statistical tests, the standardized AR (SAR) and standardized

CARs (SCAR) for each trading day t in the event window are calculated as follow:

𝑆𝐴𝑅𝑘𝑡 =𝐴𝑅𝑘𝑡

√𝑉𝑎𝑟(𝜀𝐴𝑅𝑘)

(6)

𝑆𝐶𝐴𝑅𝑘 =𝐶𝐴𝑅𝑘

√𝑁𝑘∗𝑉𝑎𝑟(𝜀𝐴𝑅𝑘)

(7)

where 𝜀𝐴𝑅𝑘 is the residual from the model estimation in event k, and 𝑁𝑘 is the window

length of [-T, T] in event k. The standardization of the ARs and CARs reduces the extreme

influence of stocks with high variance in the statistical tests, and adjusts the standard error by the

forecast-error in the out-of-sample predictions of the abnormal returns in the event window.

Patell’s Z test assumes cross-sectional independence in the abnormal return as well as the

absence of the event-induced variance change during the event period. The test statistic 𝑧𝑃𝑎𝑡𝑒𝑙𝑙

follows a standard normal distribution, and it is calculated as follow:

𝑧𝑃𝑎𝑡𝑒𝑙𝑙,𝐴𝑅𝑡=

∑ 𝑆𝐴𝑅𝑘𝑡𝑀𝑘=1

√∑𝑆𝑘−𝑝−1

𝑆𝑘−𝑝−3𝑀𝑘=1

(8)


19

𝑧𝑃𝑎𝑡𝑒𝑙𝑙,𝐶𝐴𝑅 =1

𝑀∑ 𝑆𝐶𝐴𝑅𝑘

𝑀𝑘=1

1

𝑀√∑

𝑆𝑘−𝑝−1

𝑆𝑘−𝑝−3𝑀𝑘=1

(9)

where M is the total number of recommendations within the sub-category, 𝑆𝑘 is the

number of non-missing return observations in the estimation period of event k. p is the number of

explanatory variables used in the benchmark regression model. 𝑧𝑃𝑎𝑡𝑒𝑙𝑙,𝐴𝑅𝑡 is used for testing

𝐻0: 𝐴𝑅𝑡 = 0, and 𝑧𝑃𝑎𝑡𝑒𝑙𝑙,𝐶𝐴𝑅 is used for testing 𝐻0: 𝐶𝐴𝑅 = 0.

Cross-sectional test considers the change of abnormal return variance due to the event

itself, but still assumes no cross-sectional dependence in the abnormal returns. Cross-sectional

test is applicable to ARt, CAR, and BHAR. To conduct this test, the following formulas are used

to calculate the test statistics:

𝑡𝐶𝑆,𝐴𝑅𝑡=

1

𝑀∑ 𝐴𝑅𝑘𝑡

𝑀𝑘=1

√ 1

𝑀(𝑀−1)∑ [𝐴𝑅𝑘𝑡−

1

𝑀∑ 𝐴𝑅𝑘𝑡

𝑀𝑘=1 ]

2𝑀𝑘=1

(10)

𝑡𝐶𝑆,𝐶𝐴𝑅 =1

𝑀∑ 𝐶𝐴𝑅𝑘

𝑀𝑘=1

√ 1

𝑀(𝑀−1)∑ [𝐶𝐴𝑅𝑘−

1

𝑀∑ 𝐶𝐴𝑅𝑘

𝑀𝑘=1 ]

2𝑀𝑘=1

(11)

𝑡𝐶𝑆,𝐵𝐻𝐴𝑅 =1

𝑀∑ 𝐵𝐻𝐴𝑅𝑘

𝑀𝑘=1

√ 1

𝑀(𝑀−1)∑ [𝐵𝐻𝐴𝑅𝑘−

1

𝑀∑ 𝐵𝐻𝐴𝑅𝑘

𝑀𝑘=1 ]

2𝑀𝑘=1

(12)

The statistics follow a t-distribution with a degree of freedom of M-1. 𝑡𝐶𝑆,𝐴𝑅𝑡, 𝑡𝐶𝑆,𝐶𝐴𝑅,

𝑡𝐶𝑆,𝐵𝐻𝐴𝑅 are is used to test 𝐻0: 𝐴𝑅𝑡 = 0, 𝐻0: 𝐶𝐴𝑅 = 0, and 𝐻0: 𝐵𝐻𝐴𝑅 = 0, respectively.

BMP Test (Boehmer et al., 1991) addresses the violation of the assumptions on no cross-

sectional dependence and it is robust to the variance induced by the event. The calculation of the

test statistic for this test is provided as below:

𝑧𝐵𝑀𝑃,𝐴𝑅𝑡=

1

𝑀∑ 𝑆𝐴𝑅𝑘𝑡

𝑀𝑘=1

√ 1

𝑀(𝑀−1)∑ [𝑆𝐴𝑅𝑘−

1

𝑀∑ 𝑆𝐴𝑅𝑘

𝑀𝑘=1 ]

2𝑀𝑘=1

(12)


20

𝑧𝐵𝑀𝑃,𝐶𝐴𝑅 =1

𝑀∑ 𝑆𝐶𝐴𝑅𝑘𝑡

𝑀𝑘=1

√ 1

𝑀(𝑀−1)∑ [𝑆𝐶𝐴𝑅𝑘−

1

𝑀∑ 𝑆𝐶𝐴𝑅𝑘

𝑀𝑘=1 ]

2𝑀𝑘=1

(13)

Similar to Patell’s Z test, the test statistics for BMP test follows a standard normal

distribution. 𝑧𝐵𝑀𝑃,𝐴𝑅𝑡 is used to test 𝐻0: 𝐴𝑅𝑡 = 0, and 𝑧𝐵𝑀𝑃,𝐶𝐴𝑅 is used to test 𝐻0: 𝐶𝐴𝑅 = 0.

Skewness-adjusted t-Test (Hall, 1992) corrects the cross-sectional t-test for skewed

abnormal return distributions. In the long-horizon buy-and-hold returns tend to be right-skewed

(Kothari & Warner, 2004), thus resulting a skewness bias to long-horizon abnormal performance

test statistics (B. M. Barber & Lyon, 1997). To correct this skewness bias, the following

adjustment is applied to the cross-sectional t-test for 𝐻0: 𝐶𝐴𝑅 = 0, and 𝐻0: 𝐵𝐻𝐴𝑅 = 0.

𝑡𝑠𝑘𝑒𝑤,𝐶𝐴𝑅 = √𝑀 (𝑡𝐶𝑆,𝐶𝐴𝑅

√𝑀+

1

3𝛾𝐶𝐴𝑅 (

𝑡𝐶𝑆,𝐶𝐴𝑅

√𝑀)

2

+1

27𝛾𝐶𝐴𝑅

2 (𝑡𝐶𝑆,𝐶𝐴𝑅

√𝑀)

3

+1

6𝑀𝛾𝐶𝐴𝑅) (14)

𝑡𝑠𝑘𝑒𝑤,𝐵𝐻𝐴𝑅 = √𝑀 (𝑡𝐶𝑆,𝐵𝐻𝐴𝑅

√𝑀+

1

3𝛾𝐵𝐻𝐴𝑅 (

𝑡𝐶𝑆,𝐵𝐻𝐴𝑅

√𝑀)

2

+1

27𝛾𝐵𝐻𝐴𝑅

2 (𝑡𝐶𝑆,𝐵𝐻𝐴𝑅

√𝑀)

3

+1

6𝑀𝛾𝐵𝐻𝐴𝑅)(15)

where 𝛾𝐶𝐴𝑅 and 𝛾𝐵𝐻𝐴𝑅 are the unbiased skewness estimation of CAR and BHAR over the

𝑀 recommendations. 𝑡𝐶𝑆,𝐶𝐴𝑅 is specified by (11) and 𝑡𝐶𝑆,𝐵𝐻𝐴𝑅 is specified by (12), both test

statistics are asymptotically standard normal distributed.

Analysis & Result

In the main analysis, a [-3, +3] event window is applied and the Carhart Model is used to

calculate the abnormal return. Alternative event windows in [-1, +1], [-2, +2], [-4, +4], and [-5,

+5] trading-days with other benchmark model specifications are applied to provide robustness

check.

Table 5 and Figure 2 present the result of the event analysis based on the pure ratings.

According to Table 5, both “Strong Buy” and “Buy” recommendations are associated with

positive expected CAR and BHAR, and the other recommendations are associated with negative

CAR and negative BHAR. All these CAR and BHAR show statistical significance, which


21

indicates that the recommendations are effective in generating the market reaction upon the

recommendation announcement. Moreover, the magnitudes of the abnormal returns are

consistent with the level of recommendation ratings. The “Strong Buy” has higher positive

abnormal return over the “Buy”, and the “Sell” has lower negative abnormal return over the

“Underperform”. The asymmetric short-term effectiveness can be obtained in Table 6 in several

aspects. First, the magnitude of positive expected abnormal returns for the “Strong Buy/Buy”

recommendations are smaller than the magnitude of the negative expected abnormal returns in

“Sell/Underperform” recommendations. The highest rating “Strong Buy” has a 2.07% CAR,

while the lowest rating “Sell” has a -4.06% CAR. The “Buy” has a 0.73% CAR, which is smaller

in the magnitude compared to the -3.23% of the “Underperform” CAR. Secondly, the “Hold”

recommendations do not carry explicit information regarding buy or sell actions, therefore we

should anticipate a close to zero abnormal return. However, the result shows a statistical

significant negative CAR of -2.04% associated with the “Hold” recommendations, in which the

magnitude even exceeds the “Buy” recommendations. This negative “Hold” abnormal return is

consistent with the findings documented by C.-Y. Chan et al. (2014), stating that in face of the

uncertainty related to the diminishing future performance for a stock, analysts tend to issue an

ambiguous “Hold” rather than an explicit “Underperform/Sell” rating, thus leading the “Hold” to

an equivalent unfavorable ratings. Figure 2 illustrates the trend of the abnormal return. For all

types of ratings, the most significant changes of the abnormal returns occur upon the

recommendation announcement date, and there is little significant effect after the announcement

date. This finding provides the evidence that the financial market is efficient, therefore the

informative content of the analysts’ recommendations can be quickly absorbed by the market

participants and reflected by the market reactions.


22

[Figure 2 here]

Table 6 and Figure 3 provides the results of event analysis based on the change of ratings.

The first row of Table 6 presents results of initiations, the highlighted diagonals present the

reiteration, the below the diagonal cells represents the upgrade revision, and the part above the

diagonals denotes the downgrade revisions. According to previous findings regarding the change

of stock levels, the reiterations of the ratings do not provide useful information about the stock

performance, thus no significant market reactions are found (R. Brown et al., 2009). The results

in this paper confirm this finding and the only significant reiteration is “Strong Buy” while all

other reiterations are associated with nonsignificant abnormal return. Most of the upgrades are

associated with significant positive abnormal return, and most of the downgrades have

significant negative abnormal return. The only exceptions are the upgrades from “Sell” to “Buy”

and the upgrades from “Sell” to “Underperform”. Both two sub-categories suffer from the fact

that there are only a few observations within the sub-category which partially explains why the

statistic test are not significant. Conditional on the previous rating, increasing the level of change

generally increases the magnitude of the market reaction (the only exception is the downgrade

from ‘Buy’ to ‘Underperform’, which has a greater magnitude of the negative abnormal return

than the downgrade from “Buy” to “Sell”). In the case of an initiation, the event analysis results

yield similar pattern as the results in Table 5, in which all the ratings are associated with

significant abnormal return, the sign of the abnormal return remains the same, and there is

similar effectiveness asymmetry across these recommendation initiations. The pattern of the

effectiveness asymmetry can be also discovered from the categories of revisions. The magnitudes


23

of the positive abnormal return resulted from an upgrade are generally smaller than the

magnitudes of the negative abnormal return from the corresponding downgrades assuming a

reversed action, i.e. the effectiveness of an upgrade of “hold to buy” is weaker than the

effectiveness of the counterpart of downgrade from “hold to buy”. To assess whether this short-

term effectiveness asymmetry exhibits statistical significance across all the paired

recommendations, the sign of the abnormal return with unfavorable ratings is reversed and a t-

test is performed in Table 7. From the t-test result, it is clear that the asymmetry of the short-term

recommendation effectiveness is consistent and prevailing across all the paired comparisons, in

which the “Sell/Underperform” outperforms the “Strong Buy/Buy” and the Downgrades

outperform the Upgrades. Finally, Figure 3 provides confirmation that based on the change of

rating, the most significant market reactions also occur on the same date of the announcement,

and there is only minimal influence over the abnormal return observed after the announcement.

[Figure 3 here]

There are two possible explanations for the asymmetry of the short-term effectiveness

across the recommendations. One reason is that analysts are better at picking stocks with

negative performance than picking up stocks with good performance. This explanation cannot

explain why the market is also flushed with “Strong Buy/Buy”, as if analysts are more capable of

identifying underperformers there should be more “Underperform/Sell” ratings. The other reason

is that analysts are just too over-optimistic in their rating and they tend to assign favorable

ratings even if there is no solid proof that the recommended stock has superior performance, thus

the overall quality of the favorable ratings is diluted by the inclusion mediocre stocks. This paper


24

adopts this explanation as the major cause of the short-term effectiveness asymmetry and treats

this asymmetry as a proxy for the analysts’ over-optimistic bias.

Robustness check

Robustness checks are performed to assess whether the above conclusions regarding the

analysts’ recommendation effectiveness and the effectiveness asymmetry are sensitive to the

specifications of the input in the event analysis. Different benchmark models are considered, and

different length of the event window are applied. Table 8 presents the summary of the robustness

check. In the examination of the recommendation effectiveness, the signs of the expected

abnormal return are all consistent with the findings in the main analysis, in which the initiation

of “Strong Buy/Buy” and the upgrades are associated with positive abnormal returns and the

initiation of “Sell/Underperform” and downgrades are associated with negative abnormal returns.

Statistical insignificance is only found in the categories of recommendations such as

“Downgrade from Buy to Sell”, “Upgrade from Sell to Underperform”, and “Downgrade from

Underperform to Sell”, which are the sub-categories that contain few observations. In the

analysis of the effectiveness asymmetry, the sign of the t-test that identifies the underperformers

are consistent across all different benchmark models and various window lengths. The most

common cases of the insignificant result are also only found for the pair of “Downgrade from

Underperform to Sell” and “Upgrade from Sell to Underperform”. In summary, the conclusions

from the main analysis are robust against input specifications for the event analysis. To conclude

the results, analysts’ recommendations do contain significant impact over the market reaction，

and due to the over-optimistic bias, the recommendation effectiveness is consistently asymmetric


25

between the “Strong Buy/Buy” ratings and the “Underperform/Sell” ratings, and between the

“Upgrade” and “Downgrade”.

Passive Portfolio Construction Approach

In the event analysis approach, long-horizon event study produces less reliable result than

the short-horizon event (M. Y. Chen, 2014), which limits the study of the economic value of the

recommendations. To evaluate the recommendation effectiveness in a long-term period and to

capture the economic value of the recommendations, the portfolio construction approach should

be adopted instead. This approach implement a passive portfolio management strategy that

follows the trading rules based on the content of the analysts’ recommendations.

Portfolio Specifications

Like the event analysis approach, the portfolio construction approach assesses the

recommendation effectiveness in two aspects. The first aspect considers the pure ratings of the

recommendations, and the second aspect considers the change of ratings. To address the first

aspect, two long-only portfolios based on the pure ratings are constructed and the performance of

the portfolios are compared. The “BUY” portfolio adds the stocks with a “Strong Buy/Buy”

rating into the portfolio when the announcement is made. The decision to include or exclude this

recommended stock will be made upon the release of the next recommendation. If the next rating

is another “Strong Buy/Buy” or “Hold” rating, then the stock remains in the portfolio. If the next

rating is “Sell/Underperform”, then the stock will be excluded from the portfolio. Similarly, a

“SELL” portfolio adds the stocks upon the observation of the “Sell/Underperform” ratings, and

drops the stocks upon the “Strong Buy/Buy” ratings. In the case of having multiple conflicting

recommendations on the same trading day, the average of the recommendation ratings is used to


26

determine whether a “buy” or “sell” action should be taken. To assess the economic value of

recommendation revision, an “UP” portfolio and a “DOWN” portfolio are constructed in the

same way as the “BUY/SELL” portfolios. The “UP” portfolio adds the stocks with an “Upgrade”

revision and excludes the stocks with a “Downgrade” revision; while the “DOWN” portfolio

follows the opposite trading rule. An initiation or reiteration of the recommendation with a

“Strong Buy/Buy” ratings is treated as an “Upgrade” revision from an uninformed rating, and the

initiation or reiteration of “Sell/Underperform” is treated as an equivalent “Downgrade” revision.

When a “Hold” reiteration or initiation is observed, no action is made to portfolio management.

Similarly, when multiple conflicting revisions are observed during the same trading day, the

investment decision is determined by the majority opinion (i.e. if there are more upgrades than

downgrades, then it is treated as an overall upgrade).

The portfolios are managed through a market-value-weight rebalancing, in which the

weight of each stock in the portfolio is in proportion to the market capital value of that stock. As

the result of the rebalancing, the daily portfolio return on day t is computed as follows:

𝑅𝑡𝑉𝑎𝑙𝑢𝑒 = ∑ 𝑤𝑖,𝑡

𝑉𝑎𝑙𝑢𝑒𝑅𝑖,𝑡𝑁𝑖=1 (14)

where 𝑤𝑖,𝑡𝑉𝑎𝑙𝑢𝑒 is the weight of the stock i in the portfolio at day t, 𝑅𝑖,𝑡 is the daily return

for stock i at day t, and the summation is overall all the N stocks in the portfolio at day t. 𝑤𝑖,𝑡𝑉𝑎𝑙𝑢𝑒

is determined by the following equation:

𝑤𝑖,𝑡𝑉𝑎𝑙𝑢𝑒 =

𝑀𝑉𝑖,𝑡−1

∑ 𝑀𝑉𝑖,𝑡−1𝐼∈𝑃𝑜𝑟𝑡𝑓𝑜𝑙𝑖𝑜=

𝑃𝑖,𝑡−1 𝑆𝑖,𝑡−1

∑ 𝑃𝑖,𝑡−1 𝑆𝑖,𝑡−1𝐼∈𝑃𝑜𝑟𝑡𝑓𝑜𝑙𝑖𝑜 (16)

where 𝑀𝑉𝑖,𝑡−1 is the market capital value of stock 𝑖 at day 𝑡, and (𝐼 ∈ 𝑃𝑜𝑟𝑡𝑓𝑜𝑙𝑖𝑜)

indicates the set of stocks contained in the portfolio at day t-1. The market capital value is

calculated as the product of stock price 𝑃𝑖,𝑡−1 for stock i at day t-1 and 𝑆𝑖,𝑡−1 the number of

shares outstanding for the same stock at day t-1.


27

The passive portfolio construction approach further considers the effect of transaction

cost, the effect of different holding horizons, and the effect of delayed actions. To account for the

transaction costs, the bid-ask spread is used to proxy for the percentage loss due to the trading

costs and the net rate of return of the portfolio after deducting the transaction cost is calculated in

the following way:

𝑅𝑡,𝑛𝑒𝑡𝑉𝑎𝑙𝑢𝑒 = ∑ 𝑤𝑖,𝑡

𝑉𝑎𝑙𝑢𝑒𝑅𝑖,𝑡,𝑎𝑑𝑗𝑢𝑠𝑡𝑒𝑑𝑁𝑖=1 (17)

𝑅𝑖,𝑡,𝑎𝑑𝑗𝑢𝑠𝑡𝑒𝑑 = (1 −𝑏𝑖𝑑𝑖,𝑡−𝑎𝑠𝑘𝑖,𝑡

𝑝𝑖,𝑡× (

𝑡𝑢𝑟𝑛𝑜𝑣𝑒𝑟𝑖,𝑡

𝑑𝑟𝑖𝑓𝑡𝑒𝑑 𝑤𝑒𝑖𝑔ℎ𝑡𝑡𝑉𝑎𝑙𝑢𝑒)) (1 + 𝑅𝑖,𝑡) − 1 (18)

𝑑𝑟𝑖𝑓𝑡𝑒𝑑 𝑤𝑒𝑖𝑔ℎ𝑡𝑖,𝑡𝑉𝑎𝑙𝑢𝑒 =

𝑤𝑖,𝑡−1𝑉𝑎𝑙𝑢𝑒×(1+𝑟𝑡−1)

∑ 𝑤𝑖,𝑡−1𝑉𝑎𝑙𝑢𝑒×(1+𝑟𝑡−1)𝐼∈𝑃𝑜𝑟𝑡𝑓𝑜𝑙𝑖𝑜

(19)

𝑡𝑢𝑟𝑛𝑜𝑣𝑒𝑟𝑖,𝑡 = |𝑑𝑟𝑖𝑓𝑡𝑒𝑑 𝑤𝑒𝑖𝑔ℎ𝑡𝑖,𝑡𝑉𝑎𝑙𝑢𝑒 − 𝑤𝑖,𝑡

𝑉𝑎𝑙𝑢𝑒| (20)

𝑡𝑢𝑟𝑛𝑜𝑣𝑒𝑟𝑡 = ∑ 𝑡𝑢𝑟𝑛𝑜𝑣𝑒𝑟𝑖,𝑡𝐼∈𝑃𝑜𝑟𝑡𝑓𝑜𝑙𝑖𝑜 (21)

where 𝑅𝑡,𝑛𝑒𝑡𝑉𝑎𝑙𝑢𝑒 is the portfolio return on day t after the transaction costs. 𝑅𝑖,𝑡,𝑎𝑑𝑗𝑢𝑠𝑡𝑒𝑑 is the

adjusted return for stock i on day t after the transaction effect. The item of 𝑏𝑖𝑑𝑖,𝑡−𝑎𝑠𝑘𝑖,𝑡

𝑝𝑖,𝑡 represents

the bid-ask spread for stock i on day t. The two-way turnover as calculated in the equation (20)

measures the change of position for stock i on day t in the portfolio, in which the

𝑑𝑟𝑖𝑓𝑡𝑒𝑑 𝑤𝑒𝑖𝑔ℎ𝑡𝑖,𝑡𝑉𝑎𝑙𝑢𝑒 is resulted from the appreciation of the stock value for each stock on day

t-1. The portfolio turnover on day t, as described in equation (21), is given by the summation of

individual turnovers over all the stocks in the portfolio on day t.

The effect of adopting different holding horizons is considered by changing the

maximum holding horizon for each individual stocks in the portfolio. This paper considers the

specified holding horizon over 30, 60, 120, 360, 720, and infinity trading-day periods. To

implement the strategy that incorporates the specified holding horizon, the decision to exclude


28

any stock in the portfolio will be made either upon the arrival of the new recommendation

indicating a sell action, or when the duration of the stock staying in the portfolio reaches the

specified holding horizon. A short specified holding period focuses on the investment

opportunities in a relative short-term perspective and a long specified holding period focuses on

the relative long-term investment value. The short-time investment horizon strategy is also

typically associated with more frequent rebalancing and higher turnovers, thus leading to a

higher transaction cost. Therefore, the portfolio construction analysis in this paper also offers

some implications for determining the ‘optimal’ holding period that balance the benefit from

realizing the short-term gains and loss from the increasing transaction costs.

The effect of the delayed actions considers the fact that some individual investors may

have disadvantage in receiving the recommendations or acting on the recommendations. Since

the event analysis results also show that the most significant market reactions occur on the

announcement date, a delayed action in following the recommendations potentially limits the

investors’ ability to capture the most significant short-term investment opportunity. Therefore,

the consideration of the delayed actions accounts for the loss of investment value due to the

missing of the investment opportunity. B. Barber et al. (2001) considers the investors’ reactions

to the consensus ratings to be delayed by one week, half month, and one month, and they find

that the delayed actions reduce the profitability of the recommendations. In this paper, a shorter

delay is examined, which the actions are assumed to be delayed by only 1 day, 3 days, and 5

days. This assumption is more reasonable in the current financial market situation, where the

transmission of the information has been largely enhanced by the internet and other technology,

thus the investors can efficiently obtain recommendations and respond to the recommendations

timely.


29

Similar to (B. Barber et al., 2001; B. M. Barber et al., 2006; B. M. Barber et al., 2010;

Groysberg, Healy, Serafeim, & Shanthikumar, 2013), this paper uses Carhart Model to evaluate

the portfolio performance, and the Jensen’s alpha is adopted to denote the profitability of the

portfolio. This paper also reports the Sharpe ratio and the Information ratio for each portfolio in

the performance comparison.

Results and Analysis

Figure 4 presents the growth of value for the constructed portfolios up to 31-Dec-2015,

assuming a $1 initial investment at 1-Jan-1998. In this figure, all the presented portfolios assume

timely trading actions (no delayed actions), infinity holding period, and daily rebalancing.

[Figure 4 here]

Table 9 presents the performance evaluation for the 4 portfolios. The alpha, Sharpe ratio,

and Information ratio are calculated for each of the four portfolios with different delayed actions

and specified holding period, Panel A compares the “BUY” portfolio and “SELL” portfolio. The

results confirm that analysts’ recommendations are effective trading signals, however the

informative value of the recommendations does not last long. Without the transaction cost, the

“BUY” portfolio always earns positive alpha, while the “SELL” the portfolio always has a

significant negative alpha. The “BUY” portfolio alpha diminishes as the strategy increases the

specified holding period for each stock in the portfolio, which reflects the decay of the

informative value of the recommendations. Without the delayed action, the “BUY” portfolio

alphas are still positive across different specified holding period. However, when the actions to

follow the recommendation are delayed by even only 1 day, the “BUY” portfolio alphas becomes


30

insignificant. The insignificance of the alpha with the delayed action reveals that the market is

extremely efficient and most of the informative value of the recommendations are quickly

absorbed as soon as they are released by the analyst. Namely, the market impact of each

individual recommendation is only temporal. The alphas further decrease with the consideration

of the transaction cost. The proportion of the loss due to the transaction cost decreases as the

holding period increases, which can be explained by the fact that the portfolio turnover for a long

holding period is smaller than that for a short holding period. It is interesting to note that the

above results do not demonstrate an “optimal” holding period that achieves a local maximum of

the alpha. This finding would suggest that the economic value of the analysts’ recommendations

is short-term based, and the long-term profit cannot offset the loss from the trading cost. The

results on the Sharpe ratio and Information ratio shows similar pattern, the highest value for both

ratios are observed when there is no delay action and the specified holding period is short.

Furthermore, the Information ratios are positive without the consideration of the transaction cost

and become negative net of the transaction cost with delayed actions, which implies that if the

investors cannot act quickly on the recommendations signals, the recommendation based trading

strategy will not be able to outperform the passive investment strategy that simply holds a market

portfolio. It is interesting to note that, the negative alpha of the “SELL” portfolio is much larger

than the positive alpha of the “BUY” portfolio before the transaction cost, thus the asymmetry of

the recommendation effectiveness is also consistent for the portfolio construction approach.

Panel B provides the comparison between the “UP” portfolio and the “DOWN” portfolio. The

portfolio performance exhibits similar patterns as demonstrated by Panel A. The asymmetry

regarding the portfolio performance is also found in the “UP” and “DOWN” portfolio. Moreover,

the alpha, the Sharpe ratio, and the Information ratio of the “UP” portfolio outperforms those of


31

the “BUY” portfolio, which proves that investors could get more informative value when they

consider the change of recommendation rating instead of the pure rating alone.

4. Over-optimistic Bias and the Regulation Rules

Both the event analysis approach and the passive portfolio construction approach

document the consistent findings regarding the asymmetric recommendation effectiveness. This

paper attributes this effectiveness asymmetry to the analysts’ over-optimistic bias in issuing the

recommendations ratings. Analysts try to offer favorable ratings due to their conflict-of-interest

(B. M. Barber et al., 2006; L. K. Chan, Karceski, & Lakonishok, 2007; Kadan, Madureira, Wang,

& Zach, 2009) or due to the unavoidable psychological trap (Jegadeesh & Kim, 2010;

Mokoaleli‐Mokoteli, Taffler, & Agarwal, 2009). This over-optimistic bias results in the dilution

of the quality in the recommendation reports that contain favorable ratings, and leads to the

asymmetric recommendation effectiveness from both short-term perspective and long-term

perspective. The NASD Rule 2711 and NYSE Rule 472 (Now both superseded by FINRA Rule

2241) were enacted in the year of 2002 to improve the transparency of analysts’ research and to

resolve the issue of the conflict-of-interest. According to the rules, the brokerage firms have to

disclose the percentage of the rating of “buy”, “hold”, and “sell” in each research report,

therefore the sell-side analysts are implicitly forced to issue more unfavorable recommendations

and reduce the amount of favorable ratings. As the result of the regulation, favorable ratings have

become less frequent and the number of pessimistic recommendations have increased (B. M.

Barber et al., 2006; C.-Y. Chan et al., 2014; C.-Y. Chen & Chen, 2013). However it still remains

unclear whether the quality gap between the favorable ratings and unfavorable ratings has been

reduced by the regulation. If the over-optimistic bias is truly reduced, investors should not only


32

see the decrease of the favorable recommendations, but also observe the improvement of the

quality in the favorable recommendations. In this section, the short-term recommendation

effectiveness asymmetry, in terms of the abnormal return derived from the event analysis, is used

to proxy for the measurement of the over-optimistic bias and the regulatory effect of the NASD

Rule 2711 and NYSE Rule 472 is examined. The next two sections study the overall regulatory

effect, and the incremental regulatory effect from a dynamic perspective, respectively.

Overall Regulatory Effect

Difference-in-differences Model

In this section, a difference-in-differences model is applied to examine the regulatory

effect in reducing the over-optimistic bias as measured by the asymmetry of the recommendation

effectiveness. In the difference-in-differences model, the dependent variable is the short-term

recommendation effectiveness. The buy-and-hold abnormal return calculated from the Carhart

Model in a [-3, +3] event window, denoted as REC_EFF, is used. Because the initiation of

“Sell/Underperform” and all downgrades are associated with a negative expected abnormal

return, the signs are reversed for these unfavorable ratings. The reiteration of the

recommendations are typically not associated with significant abnormal returns, therefore the

reiterations are excluded in the difference-in-differences model.

Table 9 presents the analysis result. The baseline model (denoted as Model 0) only

considers 4 variables: the POST_REG, OPT_IND, and REG_DD, plus the dummy for the

recommendation sub-categories. POST_REG is a binary dummy, which indicates whether the

recommendation is issued in the post-regulation period. The post-regulation period is defined as

from Jun-04, 2002 to Dec-31, 2015, and POST_REG takes the value of 1 for recommendations


33

in the post-regulation period, otherwise 0. OPT_IND is a binary dummy that indicates whether

the recommendation is subjected to the over-optimistic bias. For favorable ratings (initiation of

“Strong Buy”, “Buy”, and upgrade revisions), the value of OPT_IND is 1, otherwise 0. The

REG_DD is the difference-in-differences estimator, which equals the multiplication of OPT_IND

and POST_REG. The REG_DD denotes the overall regulatory effect in reducing the over-

optimistic bias (asymmetry of the recommendation effectiveness), and a positive value indicates

an improvement of the quality gap and a reduction of the over-optimistic bias. Model 1 adds the

stock specific controls. FIRM_SIZE measures the market capital value of the recommended

stocks, and BETA denotes the stock beta which is calculated from the CAPM model in the event

analysis. Model 2 adds the analysts’ specific controls. EXP denotes the analyst’s working

experience, which is measured by the # of years that the analyst appears in the IBES data file.

TASK_COMP is the measurement of analyst’s current working load, which is proxied by the #

of different stocks covered by the same analyst in previous month. Model 3 adds the variable

CRISIS_IND to indicate the economic situation, which takes the value of 1 for the major

financial crisis (internet bubble, defined as the year 1999 and 2000; and the financial crisis,

defined as the period between Apr 3, 2007 and Dec 14, 2009), otherwise 0. Model 4 further

includes the recommendation specific information. INTERVAL denotes the frequency of the

reception of the recombination, which is the calendar days since the previous recommendation

issued by any analyst on the same underlying stock. REV_INTERVAL denotes the frequency of

the revision, which is the calendar days since the previous rating issued by the same analyst.

CONS_DIFF denotes the difference between the individual rating and the consensus ratings,

which measures the herding effect of the recommendation. In particular, CONS_DIFF1 uses all


34

of the previous ratings to calculate the consensus, while CONS_DIFF2 uses only the most recent

ratings in the previous 180 calendar days to calculate the consensus.

As demonstrated by the result from the difference-in-differences model, the REG_DD is

positive and exhibits statistical significance across all the four models, which implies that the

NASD Rule 2711 and NYSE Rule 472 have an overall positive effect in reducing the

recommendation effectiveness asymmetry. Therefore, this paper provides strong evidence to

support that the regulation is overall effective and mitigate the over-optimistic bias from an

economic perspective. The result also documents other important findings from the estimation of

the covariates. FIRM_SIZE is in the negative relationship with recommendation effectiveness

and the BETA is in the positive relationship with the recommendation effectiveness. This result

is consistent to the findings that influential recommendations are more likely to occur for growth

firms, small firms (Loh & Stulz, 2011), which are generally featured by low market capital

values and high betas. In the examination of the analysts’ specific controls, EXP is in a

significant positive relation with the recommendation effectiveness, which indicates the

experienced analysts are more capable for generating influential recommendations. Both

TASK_COMP and PRODUCTIVITY is in a negative relationship with the recommendation

effectiveness, which indicates that increasing analysts’ work load would dilute their effort in each

individual research report, thus decreasing the individual recommendation effectiveness.

BROKER_SIZE is in a significant positive relationship with the recommendation effectiveness,

which implies that the resource that analysts can obtain from the big brokerage firms help them

produce more effective recommendations. It is interesting to notice that the sign of the variable

CRISIS_IND is positive and the estimation also shows statistical significance. This finding

indicates that in extreme economic situation, analysts tend to achieve better recommendation


35

effectiveness. Loh and Stulz (2014) find that analysts work harder and investors rely more on

analysts in bad times; this paper confirms their conclusions. In the examination of the

recommendation specific information, the variable INTERVAL and REV_INTERVAL are both

significant and yield different sign. REV_INTERVAL is positively correlated with

recommendation effectiveness, which indicates that the analyst’s credibility will be undermined

and the recommendation effectiveness will be impaired if he/she revises the rating too frequently.

INTERVAL focuses on the interval of receiving a recommendation regardless of who issues it,

thus the negative coefficient implies that if a stock is not actively watched by any analyst (i.e. the

interval for the update is long), then the recommendation issued on that stock will have a lower

effectiveness compared to other stocks with active coverages. The variables of CONS_DIFF1

and CONS_DIFF2 yield similar results, which show a positive relation to the recommendation

effectiveness. This result confirms that market reaction is stronger for revisions that move away

from the consensus than those that move towards it (Jegadeesh & Kim, 2010; Loh & Stulz,

2011).

Quantile Regression Model

To account for the fact that the above result might be biased due to the inclusion of the

recommendations with extreme market reactions, a quantile-based regression is applied as a

robustness check in this section. The full difference-in-differences model as specified in Model

4b is adopted to run the quantile regression and Table 9 presents the result.

The quantile regression provides similar result as the difference-in-differences model

based on the mean recommendation effectiveness. The sign of the difference-in-differences

estimator, the sign of all the covariates, and the statistical significance of the estimators remain

the same, which indicates that the conclusion from the above DID model is robust, thus


36

confirming the overall effect of NASD Rule 2711 and NYSE Rule 472 in mitigating the over-

optimistic bias.

Dynamic Nature of the Regulatory Effect

Difference-in-differences Model Applied on Sub-Periods

Despite the fact that the asymmetry of the recommendation effectiveness has been overall

reduced during the post-regulation period, the regulatory effect of the two rules does not

necessarily remain the same during the whole implementation period. To capture the dynamic

change of the regulatory effect, a series of DID model are applied to each subsequent amendment

made to the two rules. For DID model, the pre-regulation period is defined as the time since last

amendment, and the post-regulation period is defined as the time till next amendment. There are

total seven amendments made to both rules, 6 for NASD Rule 2711 and 1 for NYSE Rule 472,

which leads to seven sub-models.

Table 11 presents the results of the DID models applied on the seven amendments. There

are only four amendments that are associated with significant regulatory effect: the 1st

amendment (2003), 3rd amendment (2005), and the 5th amendment (2012) are associated with the

anticipated direction; the 2nd amendment is associated with a negative regulatory effect. For all

the remaining amendments, there is no obvious reduction in the over-optimistic bias. These

results have several implications regarding the NASD Rule 2711 and NYSE Rule 472 in

eliminating the analysts’ over-optimistic bias. First, the rule effect does not remain the same

across the implementation period. The most significant rule effect is observed during the initial

adoption of the rule, which means that analysts can quickly adapt themselves to the regulation

requirements in their recommendation report. The other significant regulatory effect is observed


37

in the post-financial crisis period (the 6th amendment in 2012). This finding is also consistent

with the findings of the significance on 1st amendment (right after the internet bubble), which

implies that the memory of the extreme economic situation and the fear for the crisis are more

effective in forcing the analysts to be more cautious and less optimistic, thus reducing the

asymmetry of the recommendation effectiveness. There is no significant regulatory effect

observed on the other amendments and even a negative estimator of REG_DD is observed for

the 2nd amendment, which indicates that the over-optimistic bias is deeply inherent in analysts’

behavior. When the financial market is performing well, analysts tend to loosen their caution in

sending out the recommendations and they move back to their normal optimism, thus the

regulation rules lose the desired effect in mitigating the asymmetry of the recommendation

effectiveness.

Structural Change of the Asymmetric Recommendation Effectiveness

In this part, the dynamic regulatory effect is examined through the aspect of the structural

change of the over-optimistic bias. In the above difference-in-differences analysis, the regulatory

effect is examined on the milestones of the rules such as the rule amendments. However, the

actual improvement of the over-optimistic bias may take place either before the amendments (as

analysts successfully anticipate the regulatory action) or after the amendment (as analysts need

some time to adjust themselves to the new requirement). Moreover, the rules may have different

effect over different recommendation sub-categories. For example, a radical change of the rating

(Buy to Sell) should be more likely to be influenced by the rules than a small revision of the

ratings (Buy to Hold). The multivariate time-series change point detection technique accounts for

the inequality and asynchrony of the rule effectiveness, therefore it is suitable to examine when


38

the most significant structural changes of the over-optimistic bias take place across all different

paired recommendation sub-categories during the rule implementation.

The change point detection technique that has been applied in this paper relies on the

divergence measure proposed by Rizzo and Székely (2010); Szekely and Rizzo (2005), which

determines whether two independent random vectors are identically distributed. A hierarchical

divisive estimation of the change points is performed through iteratively applying the

nonparametric procedure for locating one single change point and adding the new detected

change point (Matteson & James, 2014). The R package “ecp” is used to perform the analysis of

the change point detection.4

The change point detection analysis begins by firstly aggregating the short-term

recommendation effectiveness for each sub-category at the year-month level. Then the spline

interpolation is performed to impute the missing value if there is no recommendation found in a

particular recommendation sub-category in that month. The difference of the recommendation

effectiveness is then calculated among each paired sub-categories as indicated in Table 7. Then

the multivariate data that represents the effectiveness difference is used as the input for detecting

the change points. Table 12 presents the result of the change point detection. The first column

records the detected change points, the second column displays the P-values associated with the

detected change points, and the third column presents the estimation of the difference-in-

differences using the corresponding change point as the to define the pre-regulation and post-

regulation period. Panel A presents the results from the data that uses the mean as the

aggregation of the recommendation effectiveness, and Panel B presents the detected change

4 For details, see (James & Matteson, 2013)


39

points from the median of the recommendation effectiveness as the aggregation method. The

result in Table 12 provides confirmation to the previous conclusions regarding the dynamic

nature of the regulatory effect, in which the most significant regulatory effect is identified in the

initial period of rule implementation and in the early post-crisis period. Both change points are

associated with a significant reduction in the over-optimistic bias, which reflects the incremental

improvement of the regulatory effect. Moreover, the estimation of the REG_DD for the second

change point (post-crisis) is smaller than that for the first change point (rule adoption), which

implies the saturation of the desired regulatory effect over the rule implementation period. Most

previous literature that studies the regulation effect of NASD Rule 2711 and NYSE Rule 472

uses data covering relatively short post-regulation period, thus their conclusions regarding the

regulatory effect are not complete. In this paper, a comprehensive investigation of the rules

across the entire implementation period indicates that rules are overall effective, however the

regulatory effect is diminishing and analysts are prone to make the same mistake in sending out

over-optimistic rating when the financial market returns to the normal situation.


40

5. Discussion and Conclusion

This paper confirms that analysts’ recommendations are effective in both generating a

significant short-term market reaction and producing a significant long-term investment value.

The consistent asymmetry of the recommendation effectiveness reflects the analysts’ over-

optimistic bias. Through the implementation of the NASD Rule 2711 and NYSE Rule 472, this

over-optimistic bias has been mitigated. However, the regulatory effect of the two rules does not

remain constant through the implementation. The rules are most effective when they are initially

introduced and shortly after the ending of financial crisis. This changing regulatory effect implies

that analysts’ over-optimism is hard to eliminate, and the primary drive for the analyst to adjust

their over-optimistic behavior is the recallability of the extreme financial crisis. When the

financial market resumes the normal situation, analysts continue to issue less effective favorable

rating. The NASD Rule 2711 and NYSE Rule 472 that regulate the analysts’ reporting behavior

are easy for the analysts to adapt, thus a much longer post-regulation period is necessary to

correctly evaluate the overall rule effectiveness.

This paper is limited by several aspects. In the application of the portfolio construction

approach to examine the long-term investment value, only the passive portfolio management

strategy is considered. However, practitioners can also take an active management strategy to

manage the portfolio, such as the implementation of the Black-litterman framework (Black &

Litterman, 1992; Idzorek, 2002; Meucci, 2010) to combine the analysts’ recommendation with

the stock return (He, Grant, & Fabre, 2013). If an active portfolio management strategy is

adopted, it might yield difference conclusions regarding the portfolio performance and the

investment value evaluation, thus providing new insight into the study of the analysts’

recommendation. Another potential limitation of this study is that analysis are based on the


41

individual recommendations instead of the consensus ratings. The consensus rating on the

underlying stock is on a continuous scale, which allows for slight change without affecting the

overall opinion of the analysts. Performing the analysis based on the consensus ratings instead of

the individual ratings may also yield different results. However, since these aspects are beyond

the scope of this paper, it will be left for the future investigation.


42

References

Barber, B., Lehavy, R., McNichols, M., & Trueman, B. (2001). Can investors profit from the

prophets? Security analyst recommendations and stock returns. The Journal of Finance,

56(2), 531-563.

Barber, B. M., Lehavy, R., McNichols, M., & Trueman, B. (2006). Buys, holds, and sells: The

distribution of investment banks’ stock ratings and the implications for the profitability of

analysts’ recommendations. Journal of accounting and Economics, 41(1), 87-117.

Barber, B. M., Lehavy, R., & Trueman, B. (2007). Comparing the stock recommendation

performance of investment banks and independent research firms. Journal of Financial

Economics, 85(2), 490-517.

Barber, B. M., Lehavy, R., & Trueman, B. (2010). Ratings changes, ratings levels, and the

predictive value of analysts’ recommendations. Financial Management, 39(2), 533-553.

Barber, B. M., & Lyon, J. D. (1997). Detecting long-run abnormal stock returns: The empirical

power and specification of test statistics. Journal of Financial Economics, 43(3), 341-

372.

Batabyal, A. A. (2012). Project financing, entrepreneurial activity, and investment in the presence

of asymmetric information. The North American Journal of Economics and Finance,

23(1), 115-122.

Black, F., & Litterman, R. (1992). Global portfolio optimization. Financial analysts journal,

48(5), 28-43.

Boehmer, E., Masumeci, J., & Poulsen, A. B. (1991). Event-study methodology under conditions

of event-induced variance. Journal of Financial Economics, 30(2), 253-272.

Brown, R., Chan, H. W., & Ho, Y. K. (2009). Analysts’ recommendations: from which signal

does the market take its lead? Review of Quantitative Finance and Accounting, 33(2), 91-

111.

Brown, S. J., & Warner, J. B. (1985). Using daily stock returns: The case of event studies.

Journal of Financial Economics, 14(1), 3-31.

Busse, J. A., Green, T. C., & Jegadeesh, N. (2012). Buy-side trades and sell-side

recommendations: Interactions and information content. Journal of Financial Markets,

15(2), 207-232.

Carhart, M. M. (1997). On persistence in mutual fund performance. The Journal of Finance,

52(1), 57-82.

Casey, R. J. (2013). Do independent research analysts issue more or less informative

recommendation revisions? Advances in Accounting, 29(1), 36-49.

Chan, C.-Y., Lo, H.-C., & Su, Y.-R. (2014). Distribution of stock ratings and analyst

recommendation revision. The North American Journal of Economics and Finance, 28,

273-286.

Chan, L. K., Karceski, J., & Lakonishok, J. (2007). Analysts' conflicts of interest and biases in

earnings forecasts. Journal of Financial and Quantitative Analysis, 42(04), 893-913.

Chen, C.-Y., & Chen, K.-C. (2013). Brokers’ Stock Ratings Distributions and the Returns from

their Stock Recommendations: Evidence post-NASD Rule 2711.

Chen, M. Y. (2014). 'I Just Did 400 Million Event Studies'–A Study of Market Model Robustness

and Deterioration in Times of Crisis. Available at SSRN 2534446.

Clarke, J. E., Khorana, A., Patel, A., & Rau, P. R. (2011). Independents’ day? Analyst behavior

surrounding the Global Settlement. Annals of Finance, 7(4), 529-547.


43

Fama, E. F. (1998). Market efficiency, long-term returns, and behavioral finance. Journal of

Financial Economics, 49(3), 283-306.

Fama, E. F., & French, K. R. (1993). Common risk factors in the returns on stocks and bonds.

Journal of Financial Economics, 33(1), 3-56.

Groysberg, B., Healy, P., Serafeim, G., & Shanthikumar, D. (2013). The stock selection and

performance of buy-side analysts. Management science, 59(5), 1062-1075.

Hall, P. (1992). On the removal of skewness by transformation. Journal of the Royal Statistical

Society. Series B (Methodological), 221-228.

He, P. W., Grant, A., & Fabre, J. (2013). Economic value of analyst recommendations in

Australia: an application of the Black–Litterman asset allocation model. Accounting &

Finance, 53(2), 441-470.

Howe, J. S., Unlu, E., & Yan, X. S. (2009). The predictive content of aggregate analyst

recommendations. Journal of Accounting Research, 47(3), 799-821.

Idzorek, T. M. (2002). A step-by-step guide to the Black-Litterman model. Forecasting expected

returns in the financial markets, 17.

James, N. A., & Matteson, D. S. (2013). ecp: An R package for nonparametric multiple change

point analysis of multivariate data. arXiv preprint arXiv:1309.3295.

Jegadeesh, N., & Kim, W. (2006). Value of analyst recommendations: International evidence.

Journal of Financial Markets, 9(3), 274-309.

Jegadeesh, N., & Kim, W. (2010). Do analysts herd? An analysis of recommendations and

market reactions. Review of Financial Studies, 23(2), 901-937.

Jegadeesh, N., & Titman, S. (1993). Returns to buying winners and selling losers: Implications

for stock market efficiency. The Journal of Finance, 48(1), 65-91.

Jensen, M. C., Black, F., & Scholes, M. S. (1972). The capital asset pricing model: Some

empirical tests.

Jiang, G. J., Lu, L., & Zhu, D. (2014). The information content of analyst recommendation

revisions—Evidence from the Chinese stock market. Pacific-Basin Finance Journal, 29,

1-17.

Kadan, O., Madureira, L., Wang, R., & Zach, T. (2009). Conflicts of interest and stock

recommendations: The effects of the global settlement and related regulations. Review of

Financial Studies, 22(10), 4189-4217.

Kothari, S., & Warner, J. B. (2004). The econometrics of event studies. Available at SSRN

608601.

Loh, R. K., & Stulz, R. M. (2011). When are analyst recommendation changes influential?

Review of Financial Studies, 24(2), 593-627.

Loh, R. K., & Stulz, R. M. (2014). Is sell-side research more valuable in bad times? : National

Bureau of Economic Research.

Matteson, D. S., & James, N. A. (2014). A nonparametric approach for multiple change point

analysis of multivariate data. Journal of the American Statistical Association, 109(505),

334-345.

Meucci, A. (2010). Black–Litterman Approach. Encyclopedia of Quantitative Finance.

Mokoaleli‐Mokoteli, T., Taffler, R. J., & Agarwal, V. (2009). Behavioural bias and conflicts of

interest in analyst stock recommendations. Journal of Business Finance & Accounting,

36(3‐4), 384-418.


44

Murg, M., Pachler, M., & Zeitlberger, A. C. (2014). The impact of analyst recommendations on

stock prices in Austria (2000–2014): evidence from a small and thinly traded market.

Central European Journal of Operations Research, 1-22.

Patell, J. M. (1976). Corporate forecasts of earnings per share and stock price behavior:

Empirical test. Journal of Accounting Research, 246-276.

Rizzo, M. L., & Székely, G. J. (2010). Disco analysis: A nonparametric extension of analysis of

variance. The Annals of Applied Statistics, 4(2), 1034-1055.

Szekely, G. J., & Rizzo, M. L. (2005). Hierarchical clustering via joint between-within distances:

Extending Ward's minimum variance method. Journal of classification, 22(2), 151-183.


Figures

Figure 1. Specification of Estimation Period and Event Window in Event-Time Analysis


46

Figure 2. Event Analysis Based on the Recommendation Rating, with a [-3, +3] trading day

window and a benchmark return calculated from the Carhart Model


47

Figure 3. Event Analysis Based on the Change of Rating, with a [-3, +3] trading day window and

a benchmark return calculated from the Carhart Model


48

Figure 4 (a). Capital Value of “BUY” portfolio and “SELL” portfolio

Figure 4 (b). Capital Value of “UP” portfolio and “DOWN” portfolio

Note: The value of all the four portfolios are calculated based on a $1 initial investment on

1998/1/8. The portfolios are designed to follow long-only strategy, and trading action is

constructed based on the assumption that there is no delayed action, and the stocks in the

portfolio will be kept if there is no updated recommendation issued. All the four portfolios use

daily market-value weighted rebalancing schema.

-

2.0

4.0

6.0

1/8

/19

98

1/8

/19

99

1/8

/20

00

1/8

/20

01

1/8

/20

02

1/8

/20

03

1/8

/20

04

1/8

/20

05

1/8

/20

06

1/8

/20

07

1/8

/20

08

1/8

/20

09

1/8

/20

10

1/8

/20

11

1/8

/20

12

1/8

/20

13

1/8

/20

14

1/8

/20

15

$ V

alue

Date

"BUY" Portfolio

$ Market

$BUY Portfolio

$ BUY Portfolio (w. cost)

-

1.0

2.0

3.0

4.0

1/8

/19

98

1/8

/19

99

1/8

/20

00

1/8

/20

01

1/8

/20

02

1/8

/20

03

1/8

/20

04

1/8

/20

05

1/8

/20

06

1/8

/20

07

1/8

/20

08

1/8

/20

09

1/8

/20

10

1/8

/20

11

1/8

/20

12

1/8

/20

13

1/8

/20

14

1/8

/20

15

$ V

alue

Date

"SELL" Portfolio

$ Market

$SELL Portfolio

$ SELL Portfolio (w. cost)

-

5.0

10.0

15.0

1/8

/19

98

1/8

/19

99

1/8

/20

00

1/8

/20

01

1/8

/20

02

1/8

/20

03

1/8

/20

04

1/8

/20

05

1/8

/20

06

1/8

/20

07

1/8

/20

08

1/8

/20

09

1/8

/20

10

1/8

/20

11

1/8

/20

12

1/8

/20

13

1/8

/20

14

1/8

/20

15

$ V

alue,

Date

"UP" Portfolio

$ Market

$UP Portfolio

$ UP Portfolio (w. cost)

-

1.0

2.0

3.0

4.0

1/8

/19

98

1/8

/19

99

1/8

/20

00

1/8

/20

01

1/8

/20

02

1/8

/20

03

1/8

/20

04

1/8

/20

05

1/8

/20

06

1/8

/20

07

1/8

/20

08

1/8

/20

09

1/8

/20

10

1/8

/20

11

1/8

/20

12

1/8

/20

13

1/8

/20

14

1/8

/20

15

$ V

alue,

Date

"DOWN" Portfolio

$ Market

$DOWN Portfolio

$ DOWN Portfolio (w. cost)


49


Tables

Table 1

Research in the Recommendation Effectiveness

DATA METHOD KEY FINDINGS

DATA SAMPLE PERIOD

B. Barber, Lehavy, McNichols, & Trueman (2001)

Zack 1985-1996 Portfolio Construction Recommendation is effective, and the effectiveness becomes insignificant after the transaction cost

Jegadeesh & Kim (2006)

IBES 1993 - 2002 Portfolio Construction + Event Analysis

The optimistic bias is prevailing in G7 countries; US analysts provide most valuable information content

Brown, Chan, & Ho (2009)

IBES (Australia stocks) 1996 - 2003 Event Analysis Level of change, difference between consensus rating and individual rating, and analysts' reputation are important drive of recommendation effectiveness

Howe, Unlu, & Yan (2009)

IBES 1994 - 2006 Portfolio Construction Changes in aggregated recommendation contain information of future earnings at both market and industry level, and has its power to predict future return

Jegadeesh & Kim (2010) IBES 1993 - 2005 Event Analysis Analysts herd towards the consensus, and herding is more likely for downgrades than for upgrades and less likely if there is large dispersion across analyst's opinions

B. M. Barber, Lehavy, & Trueman (2010)

Zack + First Call

1986 - 2006 Portfolio Construction Both rating changes and rating levels have incremental predictive power for security returns

Jiang, Lu, & Zhu (2014) CSMAR 2007 - 2011 Event Analysis The pattern of analysts' effectiveness also differs from those matured market due to is nature as an emerging market that prohibit short-sale and being dominated by individual investors.

Murg, Pachler, & Zeitlberger (2014)

individually collected data on stocks listed in ATI

2000 - 2014 Event Analysis The result using event-time analysis will not be affected by the complexity of the asset pricing model; there is no evidence showing that analysts' opinions will be more valuable during turbulent time.


Table 2

Research in the Policy Effect of NASD Rule 2711

DATA METHOD KEY FINDINGS

DATA SAMPLE PERIOD

B. M. Barber, Lehavy, McNichols, & Trueman (2006)

First Call 1996 - 2003 Portfolio Construction Approach

NASD Rule reduce the performance difference between upgrade from conservative analysts and those from innovative analysts

Chan, Lo, & Su (2014) IBES 1996 - 2010 Event Analysis Approach NASD Rule effectively reduced the number of optimistic recommendations, and stock market is less responsive to stock upgrades that are issued by analyst who are known to be overly optimistic

Loh & Stulz (2011) IBES 1993 - 2006 Event Analysis Approach Influential recommendation can be identified from several factors; and influential recommendation revisions are more likely to occur in the post NASD Rule 2711 period

Clarke, Khorana, Patel, & Rau (2011)

IBES 2000 - 2007 Event Analysis Approach Analyst from independent source, affiliated source, and unaffiliated source all issue fewer strong buys following the regulations; but the effectiveness does not have a significant improvement after the regulation

Casey (2013) IBES 1996 - 2007 Event Analysis Approach Independent research analysts are less informative than the revision by investment banking during both pre-/post-regulation period of NASD Rule 2711


52

Table 3

Frequency Distribution of the Analysts’ Recommendations based on the current and the previous rate of rating

Current Rating

Previous Rating Strong Buy Buy Hold Underperform Sell Sub Total

N/A 46,837 (11.61%)

61,239 (15.18%)

84,046 (20.83%)

10,147 (2.51%) 3,931 (.97%)

206,200 (51.1%)

Strong Buy 6,741 (1.67%)

12,668 (3.14%)

20,455 (5.07%)

422 (.1%)

531 (.13%)

40,817 (10.11%)

Buy 12,861 (3.19%)

13,516 (3.35%)

31,260 (7.75%)

1,442 (.36%)

241 (.06%)

59,320 (14.7%)

Hold 18,389 (4.56%)

26,211 (6.5%)

19,232 (4.77%)

9,383 (2.33%)

4,119 (1.02%) 77,334 (19.16%)

Underperform 336 (.08%)

1,188 (.29%)

9,002 (2.23%)

2,672 (.66%)

574 (.14%)

13,772 (3.41%)

Sell 414 (.1%)

166 (.04%)

4,617 (1.14%)

498 (.12%)

405 (.1%)

6,100 (1.51%)

Sub Total 85,578 (21.21%) 114,988 (28.49%) 168,612 (41.78%) 24,564 (6.09%) 9,801 (2.43%) 403,543 (100.%)

Note: The numbers in the parentheses represent the relative proportion of each type of recommendations to the total number of

records in the data.


53

Table 4

Distribution of Recommendations based on the Recommendation Actions and the Level of Ratings on a year-by-year basis

Recommendation Ratings

Recommendation Actions

Year Strong Buy Buy Hold Underperform Sell Initiation Reiteration Downgrade Upgrade Sub Total BUY/SELL

Ratio

1998 5,146

(28.4%) 7,194

(39.7%) 5,397

(29.8%) 219

(1.2%) 167

(.9%)

10,133 (55.9%)

1,264 (7.%)

3,650 (20.1%)

3,076 (17.%)

18,123 (4.5%) 31.97

1999 7,052

(32.%) 8,829

(40.%) 5,656

(25.6%) 330

(1.5%) 203

(.9%)

11,927 (54.%)

1,903 (8.6%)

3,865 (17.5%)

4,375 (19.8%)

22,070 (5.5%) 29.80

2000 6,970

(31.7%) 9,018

(41.%) 5,680

(25.8%) 208

(.9%) 111

(.5%)

12,271 (55.8%)

1,857 (8.4%)

4,295 (19.5%)

3,564 (16.2%)

21,987 (5.4%) 50.12

2001 6,581

(26.4%) 9,498

(38.1%) 8,239

(33.%) 389

(1.6%) 232

(.9%)

13,084 (52.5%)

2,390 (9.6%)

5,384 (21.6%)

4,081 (16.4%)

24,939 (6.2%) 25.89

2002 8,306

(19.9%) 12,431 (29.8%)

17,055 (40.8%)

3,157 (7.6%) 832

(2.%)

18,580 (44.5%)

7,035 (16.8%)

10,368 (24.8%)

5,798 (13.9%)

41,781 (10.4%) 5.20

2003 5,882

(18.1%) 7,771

(23.9%) 15,048 (46.3%)

2,698 (8.3%) 1,134 (3.5%)

12,825 (39.4%)

4,971 (15.3%)

7,790 (23.9%)

6,947 (21.4%)

32,533 (8.1%) 3.56

2004 5,665

(19.3%) 7,066

(24.1%) 13,522 (46.2%)

1,985 (6.8%) 1,040 (3.6%)

14,221 (48.6%)

2,831 (9.7%)

6,295 (21.5%)

5,931 (20.3%)

29,278 (7.3%) 4.21

2005 5,205

(20.8%) 5,857

(23.4%) 11,449 (45.8%)

1,663 (6.6%) 851

(3.4%)

12,680 (50.7%)

2,068 (8.3%)

4,917 (19.6%)

5,360 (21.4%)

25,025 (6.2%) 4.40

2006 4,404

(17.9%) 5,960

(24.3%) 11,589 (47.2%)

1,756 (7.2%) 831

(3.4%)

12,944 (52.7%)

2,402 (9.8%)

4,814 (19.6%)

4,380 (17.8%)

24,540 (6.1%) 4.01

2007 4,247

(18.6%) 5,766

(25.2%) 10,635 (46.5%)

1,426 (6.2%) 784

(3.4%)

11,518 (50.4%)

2,380 (10.4%)

4,287 (18.8%)

4,673 (20.4%)

22,858 (5.7%) 4.53

2008 4,759

(19.4%) 5,119

(20.9%) 11,349 (46.3%)

2,205 (9.%)

1,054 (4.3%)

11,401 (46.6%)

2,779 (11.3%)

5,296 (21.6%)

5,010 (20.5%)

24,486 (6.1%) 3.03

2009 4,278

(20.%) 4,700

(22.%) 9,736

(45.5%) 1,835 (8.6%)

862 (4.%)

10,418 (48.7%)

2,019 (9.4%)

4,286 (20.%)

4,688 (21.9%)

21,411 (5.3%) 3.33

2010 4,126

(21.6%) 4,794

(25.1%) 8,607

(45.%) 1,168 (6.1%)

413 (2.2%)

10,427 (54.6%)

1,798 (9.4%)

3,354 (17.6%)

3,529 (18.5%)

19,108 (4.7%) 5.64

2011 3,886

(20.3%) 5,353

(27.9%) 8,298

(43.2%) 1,316 (6.9%)

334 (1.7%)

10,196 (53.1%)

1,782 (9.3%)

3,406 (17.8%)

3,803 (19.8%)

19,187 (4.8%) 5.60

2012 2,796

(16.%) 4,610

(26.4%) 8,305

(47.6%) 1,452 (8.3%)

289 (1.7%)

9,321

(53.4%) 1,628

(9.3%) 3,546

(20.3%) 2,957

(16.9%)

17,452 (4.3%) 4.25

2013 2,276

(15.9%) 3,928

(27.4%) 6,803

(47.4%) 1,069 (7.4%)

282 (2.%)

8,700

(60.6%) 1,338

(9.3%) 2,205

(15.4%) 2,115

(14.7%)

14,358 (3.6%) 4.59

2014 2,127

(16.7%) 3,813

(30.%) 5,780

(45.4%) 816

(6.4%) 186

(1.5%)

8,055 (63.3%)

1,203 (9.5%)

1,709 (13.4%)

1,755 (13.8%)

12,722 (3.2%) 5.93

2015 1,872

(16.%) 3,281

(28.1%) 5,464

(46.8%) 872

(7.5%) 196

(1.7%)

7,499 (64.2%)

918 (7.9%)

1,628 (13.9%)

1,640 (14.%)

11,685 (2.9%) 4.82

Sub Total

85,578 (21.2%)

114,988 (28.5%)

168,612 (41.8%)

24,564 (6.1%) 9,801 (2.4%)

206,200 (51.1%)

42,566 (10.5%)

81,095 (20.1%)

73,682 (18.3%)

403,543 (100.%) 5.84

Note: The numbers in the parentheses in the last column of “Sub Total” represent the proportion of the total recommendations counts

in each year to the total number of records in the data. The numbers in the parentheses in the other columns represent proportion of the


54

recommendations counts in the sub-category to the total count of recommendations in the corresponding year. The last column

represents ratio of the count of “Strong Buy/Buy” to the count of “Sell/Underperform”.


55

Table 5

Event Analysis Based on Level of Rating

Level of Ratings

Summary Strong Buy Buy Hold Underperform Sell

CAR [-3,3] 2.07% .73% -2.04% -3.23% -4.06%

CS T; Skew-adj CS T 53.33***; 58.16*** 21.21***; 21.35*** -65.66***; -65.33*** -36.94***; -34.99*** -26.51***; -27.66***

Patell Z; BMP Z 56.64***; 75.64*** 30.31***; 41.81*** -57.93***; -102.91*** -37.93***; -68.27*** -25.83***; -49.44***

BHAR [-3,3] 2.01% .67% -2.05% -3.24% -4.07%

CS T; Skew-adj CS T 49.57***; 57.81*** 19.23***; 19.46*** -66.86***; -60.6*** -36.79***; -20.19*** -27.5***; -24.96***

n 81,047 108,996 158,116 22,464 9,280

Note: The calculation of the cumulative abnormal return (CAR) and buy-and-hold abnormal return (BHAR) is based on a [-3, +3]

trading day window, the benchmark return is calculated from a Carhart Model. For CAR, the table reports the cross-sectional t-test

(CS T), the skewness adjusted cross-sectional t-test (Skew-adj CS T), Patell Z, and BMP Z. For BHAR, only the test-statistic of CS T

and Skew-adj CS T is presented.

*** indicates a P-Value <0.01 for a two-tailed test

** indicates a p-value <0.05 for a two-tailed test

* indicates a p-value <0.1 for a two-tailed test


56

Table 6

Event Analysis Based on the Change of Level of Rating

Updated Rating

Previous Rating

Summary Strong Buy Buy Hold Underperform Sell

None

CAR [-3,3]

1.73% .61% -1.79% -2.9% -3.45%

CS T; Skew-adj CS T 35.03***; 37.28*** 13.83***; 13.72*** -45.39***; -45.55*** -22.31***; -17.08*** -16.33***; -18.4***

Patell Z; BMP Z 47.73***; 38.79*** 27.74***; 21.35*** -67.46***; -40.21*** -39.43***; -22.82*** -29.09***; -15.07***

BHAR [-3,3] 1.66% .54% -1.81% -2.85% -3.46%

CS T; Skew-adj CS T 32.34***; 36.18*** 11.94***; 11.61*** -46.41***; -43.33*** -20.22***; -6.26*** -17.15***; -18.49***

n 44,247 57,906 78,824 9,253 3,723

Strong Buy

CAR [-3,3] -.3% -3.5% -4.67% -8.41% -8.17% CS T; Skew-adj CS T -2.12**; -2.13** -27.67***; -29.36*** -45.13***; -48.79*** -9.29***; -11.5*** -8.71***; -10.27*** Patell Z; BMP Z -2.08**; -1.79* -41.41***; -27.43*** -83.16***; -40.41*** -21.22***; -8.63*** -19.75***; -6.92*** BHAR [-3,3] -.43% -3.56% -4.62% -8.16% -8.16% CS T; Skew-adj CS T -2.93***; -2.95*** -29.02***; -31.14*** -47.08***; -47.31*** -9.81***; -11.23*** -8.92***; -10.35*** n 6,036 12,015 19,696 401 509

Buy

CAR [-3,3] 2.35% -.11% -4.3% -5.63% -4.7% CS T; Skew-adj CS T 20.43***; 21.17*** -1.19; -1.19 -52.15***; -55.42*** -10.69***; -11.*** -4.52***; -5.1*** Patell Z; BMP Z 25.84***; 18.79*** -2.3**; -2.02** -90.61***; -45.68*** -25.72***; -9.34*** -6.75***; -3.87*** BHAR [-3,3] 2.3% -.18% -4.3% -5.59% -4.77% CS T; Skew-adj CS T 19.18***; 21.02*** -2.**; -1.99** -54.56***; -55.39*** -11.31***; -10.89*** -4.71***; -5.17*** n 12,250 12,636 29,294 1,350 233

Hold

CAR [-3,3] 3.44% 3.27% .06% -4.1% -4.69% CS T; Skew-adj CS T 41.27***; 58.26*** 46.35***; 59.96*** .87; .87 -29.27***; -32.51*** -18.94***; -18.19*** Patell Z; BMP Z 63.23***; 40.74*** 71.05***; 47.42*** .6; .53 -53.75***; -29.85*** -37.78***; -20.29*** BHAR [-3,3] 3.41% 3.26% .01% -4.17% -4.7% CS T; Skew-adj CS T 38.46***; 38.46*** 43.4***; 43.4*** .2; .2 -31.13***; -31.13*** -19.71***; -19.71*** n 17,790 25,133 17,222 8,568 3,974

Under-perform

CAR (-3,3) 4.27% 4.25% 2.65% .14% -1.41% CS T; Skew-adj CS T 6.72***; 7.12*** 10.56***; 14.52*** 18.81***; 23.11*** .71; .71 -2.51**; -2.7*** Patell Z; BMP Z 10.53***; 7.15*** 18.99***; 11.74*** 36.8***; 19.79*** -1.47; -1.31 -.62; -.46 BHAR (-3,3) 4.34% 4.31% 2.64% .05% -1.57% CS T; Skew-adj CS T 6.66***; 7.25*** 10.19***; 14.43*** 17.36***; 25.69*** .24; .25 -2.88***; -2.93*** n 322 1,145 8,588 2,501 459

Sell

CAR [-3,3] 4.15% 1.33% 2.67% -.04% -.8% CS T; Skew-adj CS T 6.72***; 6.88*** 1.01; 1.06 12.71***; 16.97*** -.07; -.07 -1.22; -1.15 Patell Z; BMP Z 8.72***; 4.92*** 1.43; 1.03 25.48***; 13.75*** -1.03; -.79 -2.26**; -2.13** BHAR [-3,3] 4.12% 1.52% 2.7% -.11% -.68% CS T; Skew-adj CS T 6.67***; 7.08*** 1.05; 1.14 12.02***; 19.61*** -.24; -.24 -.96; -.85 n 402 161 4,492 391 382


57

Note: The calculation of the cumulative abnormal return (CAR) and buy-and-hold abnormal return (BHAR) is based on a [-3, +3]

trading day window, the benchmark return is calculated from a Carhart Model. For CAR, the table reports the cross-sectional t-test

(CS T), the skewness adjusted cross-sectional t-test (Skew-adj CS T), Patell Z, and BMP Z. For BHAR, only the test-statistic of CS T

and Skew-adj CS T is presented.

*** indicates a P-Value <0.01 for a two-tailed test

** indicates a p-value <0.05 for a two-tailed test

* indicates a p-value <0.1 for a two-tailed test


58

Table 7

T-Test of the Performance Asymmetry

Panel A: Performance Asymmetry Based on Pure Rating BHAR[-3,+3] CAR[-3,+3] SCAR[-3,+3]

"Strong Buy" vs. "Sell" -8.62*** -7.93*** -8.45*** "Buy" vs. "Underperform" -15.6*** -16.68*** -18.41***

Panel B: Performance Asymmetry Based on Revision

BHAR[-3,+3] CAR[-3,+3] SCAR[-3,+3]

Initiation: "Strong Buy" vs. "Sell" -13.43*** -12.6*** -12.96*** Initiation: "Buy" vs. "Underperform" -27.09*** -26.67*** -27.52*** Downgrade vs. Upgrade ("Strong Buy"-"Buy") 7.32*** 6.74*** 5.17*** Downgrade vs. Upgrade ("Strong Buy"-"Hold") 9.14*** 9.26*** 4.25*** Downgrade vs. Upgrade ("Strong Buy"-"Underperform") 3.62*** 3.74*** 2.78*** Downgrade vs. Upgrade ("Strong Buy"-"Sell") 3.66*** 3.57*** 1.88* Downgrade vs. Upgrade ("Buy"-"Hold") 9.53*** 9.53*** 4.9*** Downgrade vs. Upgrade ("Buy"-"Underperform") 1.97** 2.09** 1.05 Downgrade vs. Upgrade ("Buy"-"Sell") 1.84* 2.01** 2.48** Downgrade vs. Upgrade ("Hold"-"Underperform") 7.53*** 7.3*** 7.55*** Downgrade vs. Upgrade ("Hold"-"Sell") 6.13*** 6.2*** 6.22*** Downgrade vs. Upgrade ("Underperform"-"Sell") 2.3** 1.95* 1.81*

Note: The abnormal returns are calculated from the Carhart Model. For the comparison based on pure ratings, the sign of the

“Underperform/Sell” recommendations are reversed. For the comparison based on change of ratings, the sign of initiations of

“Underperform/Sell” and the sign of downgrades are reversed to reflect the different direction of the expected abnormal return. The t-

test are based on the assumption of unequal variance between two samples.


59

Table 8

Robustness Check of Event-time Analysis

Result of Robustness for the Event Time Analysis

Effectiveness Performance Asymmetry

EVENT WINDOW

BENCHMARK MODEL

Buy (+) Sell (-)

Upgrade(+) Downgrade(-)

Case of Insignificance Upgrade

underperform Downgrades

Strong Buy / Buy underperform

Sell / Underperform Case of Insignificance

[-1, +1]

Market Return Yes Yes Sell -> Buy Yes Yes Underperform <->Sell

CAPM Yes Yes Underperform -> Sell Yes Yes Underperform <->Sell

Fama-French Yes Yes Sell -> Underperform Sell -> Buy

Yes Yes Underperform <->Sell

Carhart Yes Yes Sell -> Underperform Sell -> Buy

Yes Yes Underperform <->Sell

[-2, +2]

Market Return Yes Yes Sell -> Underperform Sell -> Buy

Yes Yes Underperform <-> Sell

CAPM Yes Yes Sell -> Buy Yes Yes Underperform <-> Sell

Fama-French Yes Yes Sell -> Underperform Sell -> Buy Underperform -> Sell



Yes Yes Underperform<->Sell

[-3, +3]

Market Return Yes Yes Sell -> Underperform Sell -> Buy Underperform -> Sell


CAPM Yes Yes Sell -> Underperform Sell -> Buy Underperform -> Sell





Yes Yes N/A

[-4, +4]

Market Return Yes Yes Sell -> Buy Yes Yes Buy <-> Underperform


Yes Yes Buy <-> Underperform Underperform <-> Sell


Yes Yes Buy<->Underperform


Yes Yes N/A


60

[-5, +5]

Market Return Yes Yes Sell -> Underperform Sell -> Buy Underperform -> Sell

Yes Yes Buy<-> Sell Strong Buy <-> Buy


Yes Yes Buy<->Underperform Buy<-> Sell Underperform <-> Sell


Yes Yes Buy<->Underperform


Yes Yes N/A

Note: The recommendation effectiveness is determined by whether the abnormal return in the specified event window is significant at

0.1 for a two-tailed test. The performance asymmetry is determined by the t-test to examine whether the average of the magnitude for

the paired recommendations are different than zero at 0.1 significant level.


61

Table 8

Portfolio Evaluation Based on Different Holding Period and Delayed Actions

Panel A: "BUY" Portfolio Performance "SELL" Portfolio Performance

Annual Alpha (Without Transaction Cost / Net of Transaction Cost) Annual Alpha (Without Transaction Cost / Net of Transaction Cost)

Hold Period No Delay 1-Day Delay 3-Day Delay 5-Day Delay No Delay 1-Day Delay 3-Day Delay 5-Day Delay

30 7.3%*** / 3.1%*** 2.4%** / -1.8% .4% / -3.7%*** .3% / -3.9%*** -26.7%*** / -41.3%*** -9.9%* / -24.9%*** -2.6% / -16.7%*** -1.6% / -16.9%***

60 4.8%*** / 2.3%*** 1.8%** / -.7% .3% / -2.2%*** .2% / -2.3%*** -19.6%*** / -27.7%*** -6.% / -15.%*** -1.8% / -9.7%*** -.3% / -9.2%**

120 3.6%*** / 1.9%*** 1.4%* / -.3% .3% / -1.4% .2% / -1.4% -16.6%*** / -23.8%*** -4.7% / -12.8%*** -1.2% / -8.4%** .2% / -7.7%*

360 2.3%*** / 1.3% .7% / -.3% -.1% / -1.% .% / -1.% -14.5%*** / -21.3%*** -3.6% / -11.1%*** -.3% / -7.1%* 1.% / -6.4%

720 2.1%*** / 1.4%* .7% / -.1% .% / -.8% .1% / -.8% -14.3%*** / -21.1%*** -3.6% / -11.%*** -.2% / -7.%* 1.% / -6.4%

infinity 2.2%*** / 1.5%** .8% / .% .1% / -.7% .2% / -.6% -14.2%*** / -21.%*** -3.5% / -11.%*** -.2% / -7.%* 1.% / -6.4%

Sharpe Ratio (Without Transaction Cost / Net of Transaction Cost) Sharpe Ratio (Without Transaction Cost / Net of Transaction Cost)


30 .66 / .46 .42 / .22 .32 / .13 .31 / .12 -.84 / -1.32 -.14 / -.71 .16 / -.41 .21 / -.42

60 .54 / .43 .4 / .28 .32 / .2 .32 / .2 -.57 / -.94 .01 / -.38 .19 / -.16 .25 / -.14

120 .49 / .42 .38 / .31 .33 / .25 .33 / .25 -.46 / -.8 .07 / -.3 .22 / -.11 .28 / -.08

360 .43 / .39 .35 / .31 .31 / .27 .32 / .27 -.38 / -.7 .12 / -.23 .26 / -.05 .32 / -.02

720 .42 / .39 .35 / .32 .32 / .28 .32 / .28 -.37 / -.69 .12 / -.22 .27 / -.05 .32 / -.02

infinity .43 / .4 .36 / .33 .32 / .29 .33 / .29 -.37 / -.69 .12 / -.22 .27 / -.05 .32 / -.02

Information Ratio (Without Transaction Cost / Net of Transaction Cost) Information Ratio (Without Transaction Cost / Net of Transaction Cost)


30 2.52 / 1.05 .84 / -.62 .14 / -1.28 .08 / -1.34 -2.24 / -2.45 -.82 / -1.49 -.21 / -1.08 -.13 / -1.09

60 2.2 / 1.04 .81 / -.3 .11 / -.97 .07 / -1.02 -1.93 / -2.66 -.59 / -1.41 -.19 / -.97 -.05 / -.92

120 2.04 / 1.02 .79 / -.1 .18 / -.65 .17 / -.67 -1.81 / -2.5 -.51 / -1.33 -.16 / -.92 -.01 / -.84

360 1.53 / .77 .48 / -.13 . / -.54 .06 / -.51 -1.64 / -2.31 -.41 / -1.18 -.06 / -.8 .07 / -.73

720 1.52 / .81 .51 / -.04 .01 / -.46 .07 / -.44 -1.62 / -2.3 -.4 / -1.18 -.06 / -.8 .07 / -.73

infinity 1.63 / .9 .6 / .06 .11 / -.34 .15 / -.32 -1.62 / -2.29 -.4 / -1.18 -.06 / -.8 .07 / -.73


62

Panel B: "UP" Portfolio Performance "DOWN" Portfolio Performance

Annual Alpha (No Cost / Net) Annual Alpha (No Cost / Net)


30 13.9%*** / 9.%*** 2.3%** / -2.5%** .3% / -4.3%*** .5% / -4.2%*** -32.1%*** / -39.7%*** -8.3%*** / -16.6%*** -1.3% / -9.6%*** -.2% / -8.5%***

60 10.5%*** / 7.2%*** 2.%*** / -1.2% .% / -3.1%*** .2% / -2.9%*** -25.6%*** / -31.5%*** -6.7%*** / -13.2%*** -1.4% / -8.%*** -.2% / -6.8%***

120 9.1%*** / 6.6%*** 1.8%*** / -.6% .2% / -2.2%*** .1% / -2.3%*** -22.7%*** / -28.%*** -5.7%*** / -11.6%*** -1.1% / -7.%*** .3% / -5.7%***

360 8.1%*** / 6.%*** 1.7%*** / -.3% .2% / -1.9%*** .1% / -2.%*** -21.4%*** / -26.5%*** -5.3%*** / -10.9%*** -.9% / -6.6%*** .4% / -5.3%***

720 8.1%*** / 5.9%*** 1.8%*** / -.3% .2% / -1.8%*** .1% / -1.9%*** -21.2%*** / -26.3%*** -5.2%*** / -10.8%*** -.9% / -6.5%*** .4% / -5.3%***

infinity 8.%*** / 5.9%*** 1.8%*** / -.2% .3% / -1.8%*** .2% / -1.9%*** -21.1%*** / -26.2%*** -5.2%*** / -10.8%*** -.9% / -6.5%*** .4% / -5.3%***

Sharpe Ratio (No Cost / Net) Sharpe Ratio (No Cost / Net)


30 .98 / .75 .41 / .19 .32 / .1 .33 / .1 -1.21 / -1.57 -.12 / -.51 .2 / -.18 .25 / -.13

60 .82 / .67 .4 / .26 .31 / .16 .32 / .17 -.93 / -1.22 -.04 / -.35 .21 / -.1 .26 / -.05

120 .76 / .65 .4 / .29 .33 / .21 .32 / .21 -.8 / -1.07 . / -.28 .22 / -.06 .29 / .

360 .72 / .63 .4 / .31 .33 / .23 .32 / .23 -.74 / -1. .03 / -.24 .24 / -.03 .3 / .03

720 .72 / .63 .4 / .31 .33 / .23 .33 / .23 -.73 / -.99 .03 / -.24 .24 / -.03 .3 / .03

infinity .72 / .62 .41 / .31 .33 / .24 .33 / .23 -.73 / -.99 .03 / -.24 .24 / -.03 .3 / .03

Information Ratio (No Cost / Net) Information Ratio (No Cost / Net)


30 4.6 / 3. .88 / -.98 .12 / -1.69 .17 / -1.63 -5.28 / -6.26 -1.57 / -2.95 -.34 / -1.73 -.15 / -1.55

60 4.47 / 2.96 1. / -.55 .03 / -1.44 .1 / -1.33 -4.98 / -5.78 -1.48 / -2.71 -.4 / -1.66 -.16 / -1.44

120 4.46 / 3.06 1.1 / -.28 .18 / -1.09 .14 / -1.08 -4.76 / -5.51 -1.36 / -2.53 -.36 / -1.56 -.07 / -1.31

360 4.48 / 2.99 1.16 / -.14 .16 / -.98 .13 / -.98 -4.63 / -5.35 -1.3 / -2.44 -.32 / -1.49 -.03 / -1.24

720 4.49 / 2.98 1.2 / -.1 .23 / -.91 .16 / -.94 -4.62 / -5.33 -1.3 / -2.43 -.32 / -1.49 -.03 / -1.24

infinity 4.48 / 2.96 1.22 / -.09 .24 / -.9 .19 / -.92 -4.61 / -5.33 -1.29 / -2.44 -.31 / -1.5 -.02 / -1.24

Note: For each portfolio, the daily alpha is calculated from Carhart model, and the annualized alpha is obtained by multiplying the

daily alpha by 252 (assuming a 252 trading days/year). Transforming the Sharpe ratio and Information ratio is performed by

multiplying the daily Sharpe Ratio and daily Information Ratio by √252 .

*** indicates a P-Value <0.01 for a two-tailed test; ** indicates a p-value <0.05 for a two-tailed test; * indicates a p-value <0.1 for a

two-tailed test.


63

Table 9

Difference-in-differences Regression on the Short-term Recommendation Effectiveness

Model 0 Model 1 Model 2 Model 3 Model 4a Model 4b

REG_DD 0.03756*** 0.03750*** 0.03783*** 0.03813*** 0.03617*** 0.03432***

(0.0017) (0.0017) (0.0017) (0.0017) (0.0018) (0.0018)

POST_REG -0.03052*** -0.02883*** -0.02840*** -0.02725*** -0.02804*** -0.02737***

(0.0017) (0.0016) (0.0016) (0.0016) (0.0016) (0.0016)

OPT_IND -0.04498*** -0.04363*** -0.04368*** -0.04400*** -0.03650*** -0.03830***

(0.0017) (0.0016) (0.0016) (0.0016) (0.0017) (0.0017)

FIRM_SIZE -0.00686*** -0.00719*** -0.00720*** -0.00874*** -0.00871***

(0.0002) (0.0002) (0.0002) (0.0003) (0.0003)

BETA 0.00567*** 0.00583*** 0.00597*** 0.00837*** 0.00838***

(0.0005) (0.0005) (0.0005) (0.0006) (0.0006)

EXP 0.00050*** 0.00050*** 0.00039*** 0.00038***

(0.0001) (0.0001) (0.0001) (0.0001)

TASK_COMP -0.00025 -0.00025 -0.00069* -0.00069*

(0.0002) (0.0002) (0.0004) (0.0004)

PRODUCTIVITY -0.00000 -0.00000 -0.00004* -0.00004*

(0.0000) (0.0000) (0.0000) (0.0000)

BROKER_SIZE 0.00008*** 0.00008*** 0.00010*** 0.00010***

(0.0000) (0.0000) (0.0000) (0.0000)

CRISIS_IND 0.00558*** 0.00742*** 0.00728***

(0.0009) (0.0009) (0.0009)

INTERVAL -0.00004*** -0.00005***

(0.0000) (0.0000)

REV_INTERVAL 0.00003*** 0.00003***

(0.0000) (0.0000)

CONS_DIFF1 0.01237***

(0.0007)

CONS_DIFF2 0.00890***

(0.0006)

_cons 0.05758*** 0.10137*** 0.09794*** 0.09553*** 0.08164*** 0.08895***

(0.0017) (0.0023) (0.0023) (0.0023) (0.0033) (0.0033)

N 265463 265232 265232 265232 197427 197427


64

Note:

*** 99% significance level; ** 95% significance level; * 90% significance level.

Dependent Variable:

REC_EFF: Short-term recommendation effectiveness, measured by the buy-and-hold abnormal return from the Carhart model in a [-

3,+3] trading day window

Independent Variables:

REG_DD: DID estimator of the regulatory effect in reducing the over-optimistic bias

POST_REG: Indicator of the post-regulation period

OPT_IND: Indicator of the favorable ratings that are subjected to the over-optimistic bias

FIRM_SIZE: The market capital value of the recommended stock, measured in $M.

BETA: The beta value of the recommended stock from the CAPM model

EXP: The working experience of the analyst, measured as the # of years in the IBES data file.

TASK_COMP: The analyst’s current task complexity, measured as the # of covered stocks in previous month

PRODUCTIVITY: The analyst’s productivity, measured as the average # of recommendation issued per year from previous record

BROKER_SIZE: The size of the brokerage firm that the analyst is employed, measured as the # of analysts hired by the same

brokerage firm in previous year.

CRISIS_IND: The indicator for the major financial crisis (internet bubble: in the year from 1999 to 2000; global financial crisis:

from Apr 3, 2007 to Dec 14, 2009).

INTERVAL: The # of days since the previous recommendation issued by any analyst on the same underlying stock

REV_INTERVAL: The # of days since the previous recommendation issued by the same analyst on the same underlying stock.

CONS_DIFF1: The difference between the individual rating and the consensus ratings (consensus rating is calculated by aggregating

all the previous issuances)

CONS_DIFF2: The difference between the individual rating and the consensus ratings (consensus rating is calculated by aggregating

the previous issuances within 180 calendar days)


65

Table 10

Quantile based Difference-in-differences Regression on the Short-term Recommendation Effectiveness

Dependent Variable:

REC_EFF Coef. Std.Err t P>|t| [95% Conf. Interval]

REG_DD 0.008159 0.000786 10.38 0 0.0066182 0.009699

POST_REG -0.00622 0.000691 -8.99 0 -0.0075704 -0.00486

OPT_IND -0.00897 0.00042 -21.34 0 -0.0097954 -0.00815

FIRM_SIZE -0.00495 9.18E-05 -53.91 0 -0.0051274 -0.00477

BETA 0.007364 0.000295 24.97 0 0.0067857 0.007942

EXP 0.000325 3.73E-05 8.71 0 0.0002518 0.000398

TASK_COMP -0.0009 2.09E-05 -43.2 0 -0.0009452 -0.00086

PRODUCTIVITY -4.57E-06 4.01E-06 -1.14 0.254 -0.0000124 3.29E-06

BROKER_SIZE 0.000068 3.61E-06 18.84 0 0.0000609 0.000075

CRISIS_IND 0.004918 0.0005 9.84 0 0.0039383 0.005898

INTERVAL -3.4E-05 1.53E-06 -22.06 0 -0.0000368 -3.1E-05

REV_INTERVAL 1.74E-05 2.09E-06 8.35 0 0.0000133 2.15E-05

CONS_DIFF2 0.005487 0.000265 20.7 0 0.0049673 0.006006

_CON 0.042497 0.001388 30.62 0 0.0397767 0.045217

Note:

Number of observation used in the median =197427

Raw sum of deviations is 7862.897 (about 0.02165664)

Min sum of deviations is 7736.266, and the pseudo R2 = 0.0161


66

Table 11.

Difference-in-differences Regression on the Short-term Recommendation Effectiveness for Each Amendments to the Rules

Dependent Variable:

REC_EFF

Amendment

(2003)

Amendment

(2004)

Amendment

(2005)

Amendment

(2006)

Amendment

(2008)

Amendment

(2012)

Amendment

(2013)

REG_DD 0.01161*** -0.01096*** 0.00996*** 0.00102 0.00131 0.00420** -0.00019 -0.0029 -0.0026 -0.0024 -0.0025 -0.0024 -0.0021 -0.0041

POST_REG -0.01084*** 0.00357* -0.00793*** -0.00136 0.00137 0.00333* 0.00359 -0.0021 -0.002 -0.002 -0.0025 -0.0021 -0.0018 -0.0034

OPT_IND -0.01228*** -0.00212 -0.01440*** -0.00534*** -0.00501*** -0.00387** -0.00004 -0.0023 -0.002 -0.0018 -0.0017 -0.0019 -0.0015 -0.0037

FIRM_SIZE -0.00917*** -0.00907*** -0.00989*** -0.00829*** -0.00784*** -0.00766*** -0.00733*** -0.0006 -0.0005 -0.0005 -0.0005 -0.0005 -0.0004 -0.0006

BETA 0.01249*** 0.00993*** 0.00547*** 0.00112 0.00606*** 0.00694*** 0.00658*** -0.0014 -0.0011 -0.0011 -0.0012 -0.0012 -0.0012 -0.0018

EXP 0.00073** 0.00078*** 0.00084*** 0.00081*** 0.00065*** 0.00052*** 0.00032*** -0.0003 -0.0002 -0.0002 -0.0002 -0.0001 -0.0001 -0.0001

TASK_COMP -0.00210*** -0.00021** -0.00020** -0.00108*** -0.00132*** -0.00145*** -0.00133*** -0.0002 -0.0001 -0.0001 -0.0001 -0.0002 -0.0001 -0.0001

PRODUCTIVITY 0.00002 -0.00005* -0.00001 0.00004** -0.00002 -0.00005** -0.00003 0 0 0 0 0 0 0

BROKER_SIZE 0.00005*** 0.00006*** 0.00006*** 0.00007*** 0.00007*** 0.00005*** 0.00002 0 0 0 0 0 0 0

INTERVAL -0.00008*** -0.00002 -0.00005*** -0.00003* -0.00002 -0.00002* -0.00002 0 0 0 0 0 0 0

REV_INTERVAL 0.00002** 0.00003*** 0.00003*** 0.00003*** 0.00002*** 0.00002*** 0.00001 0 0 0 0 0 0 0

CONS_DIFF2 0.01345*** 0.00526*** 0.00541*** 0.00753*** 0.00740*** 0.00554*** 0.00328** -0.0013 -0.0012 -0.0011 -0.0013 -0.0013 -0.001 -0.0013

Constant 0.06756*** 0.06485*** 0.08786*** 0.07416*** 0.05723*** 0.06105*** 0.06706***

-0.007 -0.0062 -0.0058 -0.0066 -0.0064 -0.0058 -0.0086

N 33023 27153 31172 33916 64824 66048 19266

Note: the estimation is displayed as b/se. *** 99% significance level; ** 95% significance level; * 90% significance level.


67

Table 12

Structural Change of the Recommendation Effectiveness Difference

Panel A: Change Points Detection Using Mean Recommendation Effectiveness

Date P-Value REG_DD

2002-Nov 0.002 0.02952***

2009-July 0.002 0.01358***

Panel B: Change Points Detection Using Median Recommendation Effectiveness

Date P-Value REG_DD

2002-Nov 0.002 0.03027***

2009-Aug 0.002 0.01171***

Analysts’ Recommendations and the Over …fmaconferences.org/Boston/ANALYST_REC_FMA2017_FULL_PAPER.pdfRunning head: Analysts’ Recommendations and the Over-optimistic Bias Analysts’

Documents