1
Recommendation Revisions, Analyst Characteristics and
Investment Value
FLORIS VAN HALDER a∗
Master’s Thesis MSc BA - Finance
Faculty of Economics and Business, Rijksuniversiteit Groningen
_________________________________________________________________________________________________________________
ABSTRACT
Using the event study methodology, this study analyzes the impact of financial analysts’ stock recommendation revisions on stock prices immediately surrounding the announcement day. Additionally, by means of a random effects model this study focuses on which analyst characteristics influence the size of the impact following such a revision. Analyzing 1,040 revisions of 53 research departments for the 25 currently listed AEX stocks over the time-period between 2007 and April 2010, there is evidence that stock recommendation revisions contain investment value. The significant positive (negative) abnormal returns on the announcement day for upgrades (downgrades) range from 0.38% to 0.42% (-0.43% to -0.53%). The analyst- and revision-characteristics significant in explaining this return differ between up- and downgrades. For upgrades these characteristics are; target price, upgrade, frequency and NL stock analysts. For downgrades, next to target price and downgrade, IBD, stocks and banks are significant as well.
___________________________________________________________________________ Keywords: Analyst Reports & Characteristics, Stock Recommendation Revisions, Event Study, Panel Data
JEL-Classification: C23, G14, G24
Supervisor: Dhr. A.J. Meesters
a E-mail-address: Floris.van.Halder@morganstanley.com
2 TABLE OF CONTENTS
I. INTRODUCTION ... 3
II. LITERATURE REVIEW ... 6
A.1. Theoretical Framework ... 6
A.2. Stock Recommendation Revisions ... 7
A.2.1. US Empirical Evidence ... 8
A.2.2. Non-US Empirical Evidence ... 10
A.2.3. Other Empirical Evidence ... 13
A.3. Analyst Characteristics ... 13
A.3.1. Overview of Characteristics ... 14
A.3.2. Analyst Experience ... 14
A.3.3. Firm Size ... 15
A.3.4. Portfolio Complexity ... 16
A.3.5. Nationality ... 17
A.3.6. Investment Banking Department ... 18
A.3.7. Frequency ... 19
A.3.8. Size of Revision ... 19
III. DATA ... 20
IV. METHODOLOGY ... 23
C.1. Event Study ... 24
C.1.1. Mean Adjusted Return Model ... 25
C.1.2. Market Adjusted Return Model ... 25
C.1.3. Market & Risk Adjusted Return Model ... 25
C.1.4. GARCH Market & Risk Adjusted Return Model ... 26
C.1.5. Calculation of Test Statistics ... 26
C.1.6. Descriptive Statistics Event Study Models ... 28
C.2. Regression... 28
C.2.1. Dependent Variable ... 29
C.2.2. Independent Variables ... 30
V. RESULTS ... 33
D.1. Event Study Results ... 33
D.1.1. Event Study Results - Upgrades ... 33
D.1.2. Event Study Results - Downgrades ... 36
D.2. Regression Analysis Results ... 38
D.2.1 Regression Analysis Results - Upgrades ... 40
D.2.2 Regression Analysis Results - Downgrades ... 43
VI. CONCLUSION ... 47
REFERENCES ... 49
3 I. INTRODUCTION
Over the last several decades, numerous studies have been conducted involving analyst reports and more specifically the respective recommendations. Many banks and other financial related institutions have departments dedicated to the analysis of stocks. Some banks are generalists and analyze stocks globally across all sectors (i.e. global investment banks such as Goldman Sachs and Credit Suisse), while other banks are more specialized. They tend to primarily focus on niches such as certain countries and/or sectors (i.e. Kempen & Co primarily focuses on the Netherlands, real-estate, bio-pharma, utilities and small-cap stocks).
However, all brokerage departments have a similar goal; distributing research reports and the inherent recommendations to investors. Michaely and Womack (1999) define the brokerage department tasks as follows: “The analyst's specific information dissemination tasks can be categorized as (i) gathering new information on the industry or individual stock from customers, suppliers, and firm managers; (ii) analyzing these data and forming earnings estimates and recommendations; and (iii) presenting recommendations and financial models to buy-side customers in presentations and written reports.” Through the supply of these presentations and reports, brokerage departments aim to not only improve the returns of their clients but also optimize their own income (i.e. commissions and spreads)
This study examines whether it is possible for investors to profit from publicly available stock recommendations displayed in research reports. Academic theory and common practice do not give one unambiguous answer to this question. On the one hand, according to the theory of market efficiency (Fama; 1970, 1991), investors should not be able to trade profitably based on publicly available information. On the other hand, enormous amounts of resources are spent on stock analysis in order to persuade investors to buy or sell certain stocks. Grossman and Stiglitz (1980) underline this by indicating that market prices cannot fully incorporate all available information, since financial analysts then would not benefit from their costly information gathering activities. In a world, which is competitive and rational, these activities should be rewarded by corresponding expected proceeds (i.e. advisory fees, trading commissions, etc.). Likewise, investors would merely compensate financial analysts if the expected gains are greater than the costs. These gains are most likely to result from excess stock returns following financial analysts’ recommendation revisions.
4
profitable strategies do exist when trading on analysts’ recommendation revisions (e.g. Jegadeesh and Kim; 2006). On the other hand, similar research utilizing different data periods, databases and countries has indicated that one is not able to trade profitably on analysts’ recommendation revisions (e.g. Cowles, 1933; Diefenbach, 1972; and Bidwell ,1977).
Clearly, no single answer to the question whether stock recommendations contain investment value has been found in the literature. Much of the research performed involves the US, this paper aims to provide new evidence by examining stock recommendations regarding Dutch stocks listed on the AEX stock exchange in the period of 2007 to April 2010. The stock price reaction following a recommendation revision is calculated by making use of the event study methodology as described by Brown and Warner (1980, 1985) and MacKinlay (1997).
Additionally, this paper aims to provide new evidence on which analyst characteristics have the greatest impact on stock prices following stock recommendation revisions. So far, literature has relatively neglected this subject. It has mainly been studied in an indirect way, namely through the accuracy of estimation. Both theory and empirical research demonstrate results, which indicate that the accuracy of estimation of revisions is positively associated with the size of the revision (Abarbanell et al., 1995; and Gleason and Lee, 2003). This study contributes to the literature by analyzing which analyst characteristics impact the stock price following a recommendation revision in a direct way. I construct a regression with the abnormal returns (AR) of the event study as dependent variables and analyst characteristics identified in the literature as independent variables (e.g. analyst experience, size of research department, number of stocks followed, nationality and portfolio complexity).
5
accuracy of estimation following a recommendation revision in the US, are significant outside the US as well.
This study is the first to incorporate practitioners’ data from Bloomberg. Both revision dates as well as analyst characteristics are sourced from Bloomberg. Bloomberg allows characteristics not yet studied in the literature (target price and size of revision) to be incorporated in this study. Besides, data homogeneity is ensured since all data is sourced from one database. Moreover, it is of interest to compare the results found by this paper with studies which used other data sources (e.g. I/B/E/S, First Call, Zachs).
This research is of special interest since it covers stock recommendation revisions disclosed recently. Over the last decennium, confidence in analyst reports has deteriorated (e.g. Tully, 2001; and Kahn, 2002). Media reports imply that research analysts are not completely unbiased and distribute buy recommendations to create goodwillwith prospective (investment banking) clients. Empirical evidence supports these media reports by showing that analysts hardly issue negative recommendations (e.g. Jegadeesh et al.; 2004). Additionally, research analysts covering investment banking clients distribute more favorable reports as compared to other analysts (e.g. Michaely and Womack; 1999). Stock exchange authorities globally have taken measures in order to re-establish the confidence in research reports. The most well-known measure taken was at April 28, 2003 when the Global Settlement1 was agreed to address conflicts of interests within investment banks. Among the actions taken was the installation of Chinese walls in order to separate the banking and research departments. Additionally, financial institutions were compelled to provide independent stock analysis to their clients in order to “ensure that individual investors get access to objective investment advice”1. As a result, pre-2003 research on recommendation revisions may currently no longer hold.
This paper consists of the following sections. Firstly, an overview of the literature is provided. Secondly, the data is introduced and the methodology described. Thirdly, the results are presented and interpreted. Lastly, the conclusion is given including a discussion of limitations and potential areas of further research.
6 II. LITERATURE REVIEW
This section consists of three sub-sections. Firstly, the theoretical framework how financial analysts operate in the economic environment is described. Secondly, the empirical evidence regarding stock recommendation revisions and the corresponding impact on the stock price is analyzed. Thirdly, I provide an overview of analyst characteristics influencing the magnitude of the stock price reaction following a recommendation revision.
A.1. Theoretical Framework
This section provides the theoretical framework which describes the functioning of financial analysts. As discussed in the introduction, one of the main activities of financial analysts is to provide recommendations regarding stocks.In order to issue a recommendation, analysts assess all information regarding a company. They then typically input this information in a financial model in order to come up with an intrinsic firm value.Clearly, this value depends on many factors which are subjective. The dividend discount model is frequently used in economic theory to value a company. In this model the intrinsic value of a firm is calculated by discounting the future dividends by the discount rate.There would be no disagreement over intrinsic value if there would be no uncertainty. However, dividend and discount rates are not certain and analysts and investors base their valuations on expectations. Generally, these expectations are based on all available information. However, the amount of information affecting the valuation of a firm is vast.
7
When analyzing financial analysts, especially the second condition has important implications. Financial analysts seem to be unable to make a profit when all available information is costlessly available to all market participants. Grossman and Stiglitz (1976, 1980) provide an expanded view on the market efficiency hypothesis with the assumption of information search being costly. Fama (1991) indeed points out that there are information costs, but he assumes these to equal zero since this avoids “the messy problem of deciding what are reasonable information … costs.” Grossman and Stiglitz (1980) reason that information is costly to search and process, therefore firms spend vast amounts of resources on research departments. In conjunction, they note that stock markets cannot reflect all available information since financial analysts would then not be compensated for their resource-intensive analyzing activities. In a world, which is competitive and rational, corresponding expected proceeds (i.e. advisory fees, trading commissions, etc.) should result from these activities. Similarly, investors would merely compensate financial analysts if the expected gains were greater than the costs. These gains are most likely to result from excess stock returns following financial analysts’ recommendation revisions.
A.2. Stock Recommendation Revisions
8
The following sub-sections summarize the empirical evidence regarding recommendation revisions for US and non-US data. Table 1 provides an extensive summary of the empirical results.
A.2.1. US Empirical Evidence
The first research regarding the abilities of financial analysts was published in the 1930’s. Using a sample of 16 US financial services companies, Cowles (1933) does not find stock-picking abilities. While analyzing a period of 4½ years, Cowles (1933) only finds 6 out of 16 firms to be successful in picking stocks while the overall average return of recommended stocks is negative.
Aligned with Cowles’ (1933) results, Diefenbach (1972) and Bidwell (1977) find similar results using a different research period. They also doubt the predictive power of stock price forecasters. Using a similar method as Cowles (1933), Diefenbach (1972) finds a mean underperformance of -0.40% for the 12-months following a buy recommendation. However, while analyzing sell recommendations, he does find these to contain investment value with a mean underperformance of -12.6%. He does not give any specific reason for this difference but he does point out buy recommendations to outnumber sells by 26-to-1. Bidwell (1977) also finds research reports to not claim any investment value in a research period of 3 years directly following Diefenbach (1972).
However, more recent research does find brokerage reports to contain investment value and profitable trading strategies to exist. One explanation for the shift in opinion according to Bidwell (1977) could be the structural changes which occurred in the brokerage industry at March 1, 1975. At this date, fully negotiable commission rates were introduced for brokerage departments. As of then, research departments have been forced to improve the quality in order to validate a high commission rate.
9
Womack (1996) adds to Stickel (1995) by showing that the announcement effects are also statistically significant for the dataset he utilizes. Additionally, he shows that for buy recommendations the post-event drift is relatively small (but statistically significant) and short-term while for sell recommendation it is substantial and medium-term (up to 6 months). These non-rapidly mean-reverting returns suggest that revisions contain additional information for which analysts have to be rewarded.
On the other hand, Francis and Soffer (1997) take a somewhat different approach. Most research focuses on stock returns following revisions of earnings forecasts or recommendations. However, Francis and Soffer (1997) examine the interaction of these revisions. They find investors to place more weight on earnings forecast revisions in buy recommendations. They explain this result by noting that the incentives for analysts to issue buy recommendations cause investors to attach more value to other information included in a buy recommendation. A feature researched by Francis and Soffer (1997) not included in most other research, is the inclusion of reiterations. As part of their study they also focus on immediate stock price reactions to revisions of recommendations (not taking into account earnings forecast revisions). For all buy recommendations (upgrades to and reiterations), they find a mean significant stock price reaction of +0.85% for the cumulative abnormal return of
days -1 to +1. For sell recommendations, they only find a mean significant reaction of -4.76% for downgrades to sell. The median reaction of -0.44% for all hold recommendations
(upgrades to, downgrades to and reiterations) is found to be significant. The sign is consistent with investors considering hold recommendations as sell recommendations.
10
market-adjusted returns for all the different upgrades and downgrades (i.e. from strong-buy to buy, to hold, to sell and to strong-sell, etc.). The announcement effect matrix depicts results similar to Stickel (1995) and Womack (1996). They document significant positive (negative) returns for the three days surrounding the announcement of recommendation upgrades (downgrades). The returns are asymmetric with a bigger impact for downgrades (-1.28%) as compared to upgrades (0.76%).
A.2.2. Non-US Empirical Evidence
The literature discussed so far involves solely US data. Bulk of the research has been conducted in the US, mainly caused due to lack of data in the rest of the world. More recently, focus has shifted towards European and global cross-country studies which focus on a spectrum of countries.
Not surprisingly, Canada was one of the first countries to receive attention from researchers. Bjerring et al. (1983) analyze stock recommendations with an extensive research design. They elaborately analyze recommendations of one Canadian brokerage house. Not only do they analyze the immediate effect of a recommendation with an event-time perspective, but also the medium- and long-term effects with both event- and calendar-time perspectives. Besides, the inclusion of both Canadian- and US-listed stocks makes comparative analyses possible. When analyzing the immediate effect, Bjerring et al (1983) find that only Canadian stocks show a positive significant weekly return following a recommendation during week 0. However, the result for US stocks could be biased since it only comprises 26 observations versus 66 Canadian observations. Looking at the medium- and long-term effects of both the event- and calendar-time approaches, indicate significant results for all groups of recommended stocks (i) Canadian, (ii) US and (iii) Canadian and US. When taking into account transaction costs, the results remain significant. In sum, Bjerring et al. (1983) show that the medium- and long-term stock price reactions following a recommendation of US and Canadian stocks are significant, while the immediate price reaction is only significant for Canadian stocks.
11
One of the first to conduct a large-scale cross-country study is Au (2005). The study analyzes 16 European countries and spans 220,000 recommendations made in the time-period of November 1993 to April 2004. Although his study mainly focuses on the 60-day cumulative abnormal return, he does find the immediate effect (5-day cumulative abnormal return) to be statistically significant for European countries. The cumulative abnormal return for the most favorable recommendations quintile is 0.2% for the 5-day period. The subsequent post-event drift brings the cumulative abnormal return to 0.6% for the 60-day period. The least favorable recommendations quintile is symmetric to the most favorable, as was also shown by Liu et al (1990). Au (2005) confirms the results of US research (e.g. Womack, 1996; and Barber et al., 2001) for an European dataset in a time-period following Reg FD1.
Building on previous study, Jegadeesh and Kim (2006) added to Au (2005) by not only examining European countries but a worldwide sample; G7 countries. By using event-time methodology, they find significant stock price reactions around the announcement day for all countries except Italy. Additionally, the largest immediate price reaction and largest post-event drift (6 months) are found in the US (1.76% and 4.75% for upgrades and -3.19% and -6.20% for downgrades). Of the other countries, only Japan experiences a post-event drift (4.21%) following upgrades. On the other hand, all countries but Japan and Britain experience a post-event drift following downgrades (-2.20% to -5.86%). By analyzing a set of securities which are followed by analysts around the globe, Jegadeesh and Kim (2006) suggest that the outperformance by US analysts is most probably caused by superior skill at detecting under/overvalued securities. According to Jegadeesh and Kim (2006), the high salaries and bonuses in the US relative to other countries cause this superior analysis. The calendar-time methodology generally confirms the results found by the event-time methodology. The announcement effect is analyzed by incorporating stocks into a portfolio with a 1-day delay. The 1-day delay leads to an average loss of abnormal return of 0.83%, which can be interpreted as the announcement effect. The 1-month abnormal returns following inclusion into the portfolio are significant for all G7 countries (ranging from 1.27% to 5.79%) except for Italy.
12
Table 1: Literature Review Regarding Stock Returns Following Stock Recommendations
% Return on Pre-Recommendation Price
Literature Year Coun
-try # FA 1
# REC1 Methodology1
Pre-event t=0
Post-event
Analyst Reports (US Evidence)
Cowles 1933 US 16 7,500 Buy-Portfolio n.a. n.a. -1.422
Diefenbach 1972 US 24 1,209 Buy-Portfolio n.a. n.a. -0.42
6 46 Sell-Portfolio n.a. n.a. -12.62
Stickel 1995 US 1,510 8,790 Buy-Event 0.653 0.903 1.813
1,510 8,167 Sell-Event -1.063 -0.803 -1.343
Womack 1996 US 144 694 Buy-Event n.a. 2.984 0.094
144 209 Sell-Event n.a. -4.694 -9.154
Francis & Soffer 1997 US 515 576 Buy-Event n.a. 1.285 n.a.
515 576 Sell-Event n.a. -4.765 n.a.
Barber, Lehavy, McNichols & Trueman 2001 US 4,340 361,620 Buy-Portfolio n.a. 0.766 4.136 4,340 361,620 Sell-Portfolio n.a. -1.286 -4.916
Analyst Reports (Global Evidence)
Bjerring, Lakonishok & Vermaelen 1983 CAN7 17 92 Buy-Event 0.327 1.497 4.097
17 92 Buy-Portfolio n.a. n.a. 0.277
Von Nandelstadh 2003 FI 633 4,592 Buy-Portfolio n.a. n.a. 6.22
Au 2005 EU 5008 220,000 Buy-Portfolio n.a. 0.28 0.68
Jegadeesh & Kim 2006 G7 1,890 87,966 Buy-Event 0.649 1.769 4.759
1,890 103,208 Sell-Event -2.129 -3.199 -6.209 Buy-Portfolio n.a. 3.399 1.109
Other Empirical Evidence
Lloyd-Davies & Canes 1978 US 110 597 Buy-Event 1.3010 0.9210 0.9210
110 188 Sell-Event 0.2810 -2.3710 -2.5510
Liu, Smith & Syed 1990 US 110 566 Buy-Event 2.0610 1.5410 0.8510
110 286 Sell-Event -1.8510 -1.9910 -2.4510
Beneish 1991 US 110 286 Buy-Event 1.3110 1.0110 1.2810
110 118 Sell-Event -1.0610 -1.0010 -1.4110
1
# FA = number of financial analysts, # REC = number of recommendations. Methodology refers to either portfolio (calendar-time) strategies or event methodology (event-time analysis)
2
Annualized returns
3
Pre-event refers to days (-10,-1), t=0 to days (0,+10) and post-event to days (0,+60)
4
# FA refers to the number of brokerage houses, t=0 to days (0,+3) and post-event to days (0,+180)
5
# FA refers to the number of brokerage houses and t=0 to days (-1,+1)
6 Long-term: annual return after controlling for market risk, size, book-to-market, and price momentum effects. t=0: refers to days (-1,+1).
7 # FA reflects number of brokerage houses. Data involves combination of US and Canadian stocks, but one Canadian brokerage house. 1st
row refers to event-time perspective with pre-event being days (-7,-1), t=0 being days (0, +7) and post-event being days (0, +266). 2nd row refers to calendar-time perspective where the return represents a weekly return..
8 #FA refers to number of brokerage houses. t=0 refers to (0,+5) and post-event to (0,+60). Returns for sell-portfolios are symmetric; a
trading strategy on both buy- and sell-portfolios thus leads to abnormal returns of 0.4% and 1.2%.
9 Returns depicted are for US. For event methodology: the other 6 countries analyzed also experience significant immediate returns (ranging from 0.16% to 0.46% for upgrades and -0.18% to -0.45% for downgrades) except for Italy (0.04% for upgrades and -0.09% for downgrades). Pre-event relates to days (-10,-1) and post-event to days (0,+132). For portfolio methodology: the other 6 countries analyzed also experience positive returns (ranging from 0.24% to 0.60% for t=0 and 0.43% to 1.62% for post-event) except for Italy (0.13% for t=0 and 0.29% for post-event). t=0 relates to day 0 and post-event to days (0,180).
10
13
A.2.3. Other Empirical Evidence
Next to the literature concerning research reports, much research has been conducted analyzing recommendations published in other sources; most notably recommendations regarding US stocks in the Wall Street Journal (“Heard on the Street” column). Using event study methodology, Lloyd-Davies and Canes (1978) find significant abnormal returns on the announcement day of a recommendation (0.92% for buy and -2.37% for sell recommendations). Subsequently, similar research by Liu et al. (1990) and Beneish (1991) confirm and extend these results for more recent time periods. Next to abnormal returns on the announcement day, they also find significant returns on the two days preceding the announcement. Additionally, Liu et al. (1990) find the size of the reaction to be symmetric for buy and sell recommendations. Over a 3-day event window (-2, 0), they find the cumulative abnormal return for upgrades (downgrades) to be 2.82% (-3.63%) with the largest return on day 0; 1.54% (-1.99%). For the same event window, Beneish (1991) finds similar results: cumulative abnormal return of 1.70% (-2.43%) for upgrades (downgrades) with the largest return occurring on day 0; 1.01% (-1.00%).
After a thorough analysis of the literature, I have formulated the following hypothesis:
H1 Recommendation revision hypothesis: significant stock price reactions occur immediately following analysts’ recommendation revisions of Dutch-listed companies
A.3. Analyst Characteristics
14
A.3.1. Overview of Characteristics
Von Nandelstadh (2003) mentions two sets of characteristics which influence the accuracy of estimation; (i) company and (ii) analyst characteristics. My research will focus on the second set of characteristics; analyst characteristics. Additionally, I have added a set of characteristics, best classified as revision characteristics (table 2: H7 and H8), which have been relatively neglected by the prevailing literature. Von Nandelstadh (2003) classifies analysts’ experience and size of the broker firm under analyst characteristics. I have identified several other characteristics in the literature which I have added to this research design. Table 2 summarizes the hypotheses which I test through regression analyses (H2 to H8). The following subsections discuss the literature which led to the formation of these hypotheses.
Table 2: Overview of Hypotheses and Expected Signs (Following Revision Upgrades)
Hypotheses Exp. Sign
Event Study
H1 Recommendation Revision +/-
Regression (Analyst Characteristics Impacting Stock Price Following Recommendation Revision)
H2 Analyst Experience +
H3 Firm Size +
H4 Portfolio Complexity -
H5 Nationality +
H6 Investment Banking Department -
H7 Frequency -
H8 Size of Revision +
A.3.2. Analyst Experience
Experience in the literature is mostly defined as the number of months between the date of the revision and the analysts’ first recorded recommendation. They then make a distinction between firm-specific experience (time-period between revision of particular stock and first recorded recommendation on the same stock) and general experience (time-period between revision of stock and first-ever recorded recommendation).
15
While studying countries outside the US, Clement at al. (2000) find that for Japan and Germany, experience does not appear to have a significant influence on forecast accuracy as was concluded from US data. They justify their finding by noting that the employer-employee relationship in Japan and Germany is greater as compared to the US, and therefore employees do not have to show superior skills in order to maintain their positions.
Lastly, Bolliger (2004) finds contradicting results with respect to experience. Firm-specific experience is positively correlated with forecast accuracy while general experience is found to be unassociated with accuracy. He reasons that little in Europe is known about the compensation drivers of financial analysts. In the US, forecast accuracy is highly linked to salary, while in Europe it could be the case that this relation is less pronounced. Analysts then do not have the incentive to improve their accuracy.
H2 Analyst experience hypothesis: analyst experience is positively related to the stock price impact immediately following analysts’ recommendation revisions of Dutch-listed companies
A.3.3. Firm Size
In the literature the size of a firm is calculated by summing the number of research analysts employed by a firm, alternatively they make use of dummy variables or size quintiles. The general hypothesis is that large firms are associated with more resources, e.g. better support services, access to larger number of databases, etc.
16
conflicts of interest (to be discussed in section A.3.6.). Large brokers in the US have an impact 1.84% larger as compared to small brokers while this difference is smaller but statistically significant in the non-US countries; 0.11%.
Evidence from Clement et al. (2000) supports the non-US evidence found by Jegadeesh et al. (2006). They find firm size to impact forecast accuracy in Canada and the UK but not in Germany and Japan. They explain this by pointing out the difference between common-law (Canada and UK) and civil-law countries (Japan and Germany). In common-law countries, financial analysts have an advantage as equity is more widely used as a source of capital. In these countries, analysts have more access to private communication with firm management and are thus able to provide more accurate forecasts.
In a similar research, Bolliger (2004) finds no difference between forecast accuracy of analysts working for large or small firms for 14 European countries. He explains this result by noting that smaller banks focus on less countries, industries and companies. Since he also finds accuracy to decline with complexity, the effect of increased accuracy of larger banks could be offset by increased complexity.
H3 Firm size hypothesis: firm size is positively related to the stock price impact immediately following analysts’ recommendation revisions of Dutch-listed companies
A.3.4. Portfolio Complexity
Next to employer size, Clement (1999) and Jacob et al. (1999) also study the number of firms and industries followed by an analyst. Both find evidence that the number of firms and industries followed is negatively associated with forecast accuracy. They reason that number of firms and industries followed can be seen as a proxy for portfolio complexity. Larger portfolios lead to less focus on each company and thus to less accurate forecasts. In a follow-up research, Clement and Tse (2003) reconfirm their findings for a different data-period.
17
H4 Portfolio complexity hypothesis: number of stocks followed by analyst is negatively related to the stock price impact immediately following analysts’ recommendation revisions of Dutch-listed companies
A.3.5. Nationality
Geography (domestic versus non-domestic, local versus remote) of the analyst is an internationally extensively researched characteristic. Generally, it is hypothesized that local researchers have an informational advantage caused by; a.o. comprehension of local culture and language, lower cost of gathering of information, top management proximity, etc. Malloy (2005) provides evidence for the US that local analysts are significantly more accurate and have more influence following revisions. The difference in accuracy ranges from $0.03 to $0.14 per share and difference in impact for sell revisions ranges from -0.13% to -0.16% and for buy revisions from +0.09% to +0.10%.
The evidence across countries is rather mixed with Bacmann and Bolliger (2001) and Chang (2009) finding foreign analysts to perform better. While analyzing Latin-American emerging countries, Bacman and Bolliger (2001) find foreign analysts to outperform local analysts. Chang (2009) finds a different but also counterintuitive result. For Taiwan he finds foreign financial analysts, who have a local establishment, to be superior to foreign financial analysts without local presence and domestic analysts. An explanation for these results could be the superiority of developed countries’ financial analysts.
Research conducted in Western economies is aligned with the hypothesis of local researchers to possess an informational advantage over foreign researchers. Bolliger (2001) analyzes 14 European countries and finds an accuracy advantage of local analysts. Orpurt (2004) extensively analyzes home-country analysts´ forecast accuracies in Europe. He finds local analysts to issue more accurate forecasts, with German and Dutch analysts having the biggest amount of accuracy advantages. Orpurt (2004) reasons that relatively limited financial disclosure in the Netherlands and Germany as compared to other European countries could potentially generate a bigger local analyst advantage.
18
A.3.6. Investment Banking Department
Investment banks have received a great deal of attention in the international literature regarding recommendation reports. The consensus is that research departments which operate within the same firm as investment banking departments provide overly-optimistic recommendations. The reasoning is that IBD research departments are better off not issuing negative recommendations (sell/hold) in order to stay in favor with a company and its management. Issuing negative recommendation could potentially lead to loss of investment banking revenues. An internal memo from Morgan Stanley1 underlines this: “Our objective … is to adopt a policy, fully understood by the entire firm, including the Research Department, that we do not make negative or controversial comments about our clients as a matter of sound business practice.”
Especially in the period before 2003, pre-Global Settlement, this conflict-of-interest played a crucial role. Barber et al (2007) research the period from 1996 to 2003 and find evidence of investment banks issuing overly-optimistic reports. They find buy recommendations of independent research firms to outperform those of investment banks by nearly 8% annualized. When researching recommendations on companies with which investment banks have a relationship (i.e. advisory mandate), results become more evident.
Numerous other papers confirm the results found by Barber et al. (2007) for different datasets (time periods, countries, etc.). Cliff (2004) finds for the US in the time-period 1994 to 2003, investment banks to publish inferior recommendations (except for sell recommendations). Conflicts-of-interest are the main cause for this. Michaely and Womack (1999) find similar results for the US for a slightly earlier time-period of 1990 to 1991. On the announcement day the excess return is 2.7% for underwriter analysts and 4.4% for non-underwriter analysts.
H6 Investment banking department hypothesis: research operating within the same firm as investment banking departments, is negatively related to the stock price impact immediately following analysts’ recommendation revisions of Dutch-listed companies
1
19
A.3.7. Frequency
Next to the previous hypotheses, which were formed while analyzing the literature, I have added two hypotheses which have not been researched in a similar research design; frequency and size of revision. An analyst typically reiterates a previous recommendation before revising. Frequency refers to the number of times an analyst reiterates his previous recommendation before revising. A recommendation is normally made for a certain period of time. Francis and Soffer (1997) find that 95% of the analysts explicitly or implicitly state a time-frame for which the recommendation holds in their report. I hypothesize that analysts typically issue many reiterations when they are working for an infamous bank. By issuing many reiterations, they try to put themselves on the top of the lists of the different databases. Additionally, superior analysts are expected to only issue a revision when there is substantial ground to do so. This automatically leads to a smaller amount of recommendations for superior analysts.
H7 Frequency hypothesis: issuing many reiterations before going over to a revision is negatively related to the stock price impact immediately following analysts’ recommendation revisions of Dutch-listed companies
A.3.8. Size of Revision
The second hypothesis which I have added to my research is the size of revision. It follows automatically that analysts express great confidence when they issue a substantial revision (i.e. from strong sell to strong buy, high target price). A larger revision will hence lead to a bigger stock price impact.
H8 Size of revision hypothesis: size of a revision is positively related to the stock price impact immediately following analysts’ recommendation revisions of Dutch-listed companies
20
Table 3: Literature Review Regarding Analyst Characteristics Impacting Stock Price Reaction (Forecast Accuracy) Following Recommendation Revisions
Hypothesis Literature Year Sign
H2: Analyst Experience Mikhail, Walther & Willis 1997 +
Clement 1999 +
Clement, Rees & Swanson 2000 +/-
Bolliger 2004 +
H3: Firm Size Stickel 1995 +
Clement 1999 +
Jacob, Lys & Neale 1999 +
Clement, Rees & Swanson 2000 +/-
Bolliger 2004 +/-
Jegadeesh & Kim 2006 +
H4: Portfolio Complexity Clement 1999 -
Jacob, Lys & Neale 1999 -
Clement and Tse 2003 -
Bolliger 2004 -
H5: Nationality Bacmann and Bolliger 2001 -
Bolliger 2001 +
Orpurt 2004 +
Malloy 2005 +
Chang 2009 -
H6: Investment Banking Department Michaely & Womack 1999 -
Cliff 2004 -
Barber, Lehavy & Trueman 2007 -
H7: Frequency n/a n/a E(-)*
H8: Size of Revision n/a n/a E(+)*
* E(..): Expected
III. DATA
In this section I introduce the data. Firstly, I describe the data source and selection criteria used to select the data. Secondly, I provide descriptive statistics regarding the recommendations followed by descriptive statistics of the variables used in the regression.
21
study on the period January 2007 to April 2010, chosen due to data completeness and accuracy. The initial sample consisted out of 1,172 revisions, however after evaluation of the data 1,040 events are included in the sample. One analyst which had 37 revisions was deleted from the sample since he was covering 237 stocks. Clearly, this was an outlier since the average amount of stocks covered was 12 (see table 6). Secondly, through steps 2, 3 and 4, I have further evaluated the sample by looking at the target prices issued in the analyst reports. 50 events did not include a target price, therefore I did not incorporate these in my study. Furthermore, one expects an analyst to issue a target price above or approximate to the market price following an upgrade (vice-versa for downgrades). However, some events displayed a market price higher (lower) than a target price following an upgrade (downgrade). This can be explained by the fact that i.e. sell-hold (buy-hold) revisions which are classified as upgrades (downgrades) may still have a target price underneath (above) the market price. Bloomberg is a system in which input is done manually, therefore errors in the databases could occur. Hence, I deleted the extreme cases of above-described phenomenon from the sample through steps 3 and 4. In step 5 I have analyzed the frequency of revisions by dividing the number of recommendations made by an analyst for a stock by the number of months up to the revision of the same stock. On average, analysts issue less than one recommendation per month before revising a recommendation. To ensure data homogeneity, I corrected for an outlier in the database which had an average of 9 recommendations per month. Lastly, I deleted conflicting revisions of the same analyst on the same stock with an overlapping event window; i.e. analyst Y who issues an upgrade on stock X on t = 0 followed by a downgrade on t = 3. These conflicting events are not incorporated since it is unlikely that analysts revise their recommendations within the event window (i.e. 5 days). Recommendations are usually issued for the medium-/long-term (Francis and Soffer; 1997). Again, errors in the recording of recommendations in the database could be the cause of this. Secondly, conflicting events lead to a distortion of the results since opposing effects then occur within the event window. An overview of the selection criteria are listed in table 4.
Table 4: Selection Criteria Utilized in Order to Evaluate Dataset
Step Selection Criteria # of Recommendation Revisions
Initial Sample 1,172
1 Number of Stocks Covered by Analyst =< 40 1,135
2 Inclusion of Target Price 1,085
3 Target Over Market Price Following Upgrade >= -30% 1,079
4 Target Over Market Price Following Downgrade =< +30% 1,049
5 Frequency (# recommendations up to revision/ # months) =< 5.0 1,048
6 No conflicting events in analysts’ event window 1,040
22
The 1,040 recommendation revisions of currently AEX-listed stocks, over the period January 2007 to April 2010, are split up between 554 upgrades and 486 downgrades. This result is in-line with empirical evidence which show that analysts tend to issue more upgrades than downgrades (e.g. Womack, 1996; and Jegadeesh et al, 2004). The difference is rather small (see table 5) while previous research showed significantly larger differences; e.g. Diefenbach (1972) with buy-to-sell of 26-to-1 and Womack (1996) 7-to-1. This smaller difference could be the cause of the actions taken by the government in order to separate the research and investment banking departments. Jegadeesh and Kim (2006) indeed find that sell recommendations have become more common since 2003, sell recommendations comprise 18% of their sample while the percentage is 21% in my sample. Additionally, more sell and hold recommendations are expected in a negative market sentiment; the AEX declined by 31% over the research period. The recommendations mostly comprise of post-Lehman bankruptcy data (2007: 133, 2008: 252, 2009: 475 and 2010: 180).
Analysts use different classifications in order to rank recommendations. The most-widely used classification is the sell-hold-buy scale, therefore I have chosen to rank the recommendations according to this. This allows me to identify two different types of upgrades (downgrades): (i) two-step upgrade (downgrade) from sell-to-buy (buy-to-sell) and (ii) one-step upgrade (downgrade) from either sell-to-hold or buy (buy-to-hold or hold-to-sell). My sample consists of 100 two-step upgrades, 454 one-step upgrades, 90 two-step downgrades, and 396 one-step downgrades. Table 5 provides an overview of the split-up between up- and downgrades.
Table 5: Overview of Up- and Downgrades in the Research Sample and Corresponding Difference between Target and Market Price
# of Revisions
% of Total
% Difference of Target over Market Price Mean Median Maximum Minimum
23
The 1,040 recommendation revisions are issued by 53 banks and independent research firms, 304 different analysts and all 25 currently AEX-listed stocks are involved. Appendix A provides an overview of the financial institutions whose recommendation revisions are included in the sample. Additionally, it gives information about the amount of researchers covering the Dutch market and amount of revisions per bank included in the sample. Appendix B depicts the AEX-stocks included in the research and the corresponding amount of recommendation revisions
Table 6 displays an overview of the descriptive statistics of the analyst- and revision-specific characteristics studied in this research.
Table 6: Descriptive Statistics of the Sample (Used in the Regression Analysis)
Variable Mean Median Max. Min. Std. dev.
Continuous Variables
Analysts Covering NL Stocks at Bank (#) 8.76 9.00 15.00 1.00 3.57
Total Revisions at Bank (#) 34.24 34.00 69.00 1.00 17.13
Target over Market Price (%) 10.49% 9.84% 102.43% -49.92% 16.36%
Frequency (# / month) 0.87 0.67 5.00 0.11 0.77
Stocks Covered (#) 12.32 11.00 39.00 1.00 5.64
NL Stocks Covered (#) 6.36 3.00 39.00 1.00 7.56
Previous Banks (#) 0.69 0.00 4.00 0.00 0.87
General Experience (months) 59.42 64.00 133.00 1.00 33.75
Firm-Specific Experience (months) 28.18 19.00 133.00 1.00 25.01
Binary Variables
Dutch Nationality 0.39 0.00 1.00 0.00 0.49
Investment Banking Division 0.82 1.00 1.00 0.00 0.39
Analysts Covering NL Stocks at Bank is the total number of analysts currently working at a bank following Dutch-listed stocks. Total Revisions at Bank is the total number of revisions on Dutch-listed stocks of all analysts working at a particular bank. Target over Market Price is the percentage difference of the target price minus previous day closing divided by the previous day closing price. Frequency is the number of recommendations prior to a revision divided by the number of months up to a revision. Stocks Covered is the total amount of stocks covered by the analyst. NL Stocks Covered is the total amount of Dutch-listed stocks covered by the analyst. Previous Banks is the number of previous banks where an analyst was employed. General Experience is the number of months an analyst has been analyzing stocks. Firm-Specific Experience is the number of months an analyst has been analyzing a particular stock. Dutch Nationality indicates with a 1 whether an analyst is Dutch or with a 0 whether he is non-Dutch. Investment Banking Division indicates with a 1 whether an analyst is employed at a bank with an investment banking department or with a 0 if he is employed at an independent research firm.
IV. METHODOLOGY
24
regression in order to measure the impact of certain analyst- and revision-characteristics on the stock price following the announcement date. For both the event study and regression, two separate analyses will be performed; (i) upgrades and (ii) downgrades.
C.1. Event Study
The event study methodology can be divided into some important steps. Firstly, one tests whether abnormal returns (ARit) are present in the event window. The event window is
the time-frame where one expects an impact to occur. Subsequently, one needs to compare the observed returns in the event window (Rit) to the normal return. This normal return is
calculated over the estimation period (E(Rit)). In my study I have made use of an estimation
period of 200 days as suggested by Brown and Warner (1980, 1985), who note that this time-frame is sufficient to capture normal returns. Next to this, I make use of an event window of 8 days (-2, 5), similar as e.g. Liu et al. (1990) and Beneish (1991) who find that abnormal returns are concentrated on days -1 to 1. A potential information leakage effect is captured in, as recommendation revisions could be spilled or leaked before the announcement day (t = 0). The announcement day is the day that an analyst revises his or her recommendation of a stock. Figure 1 depicts the estimation and event window periods.
Figure 1: Graphical Depiction of the Estimation Period and Event Window
All returns in this research are continuously compounded, where Rit is the return of
stock i at day t, Pit is the stock price on day t and Pit-1 the stock price on day t-1.
1 ln − = it it it P P R (1)
Mathematically the event study methodology can be expressed as follows: )
( it it
it R E R
AR = − (2)
In order to increase the robustness of this research I use four different models to calculate the normal returns of each revision (E(Rit)). The first three models are extensively
described by Brown and Warner (1980, 1985): (i) mean adjusted return, (ii) market adjusted return and (iii) market & risk adjusted return model. The fourth model (iv) GARCH market &
Event Window Estimation Period
25
risk adjusted return model, is a GARCH corrected version of the market & risk adjusted model.
C.1.1. Mean Adjusted Return Model
First, the mean adjusted return model can be defined as follows:
it i it
R =
µ
+ε
(3)The observed return of stock i on day t (Rit) consists of the average return of stock i in
the estimation period (µi) and the error-term for stock i on day t (εit).
Secondly, the abnormal returns are calculated by taking the difference of the observed return of stock i on day t and the average of the returns during the estimation period. The error-term of stock i on day t thus can be interpreted as the abnormal return of stock i on day t.
C.1.2. Market Adjusted Return Model
The specification of the market adjusted return model incorporates market movements of market m on day t (Rmt). The AEX-index is utilized as market-index since this index reflects the broad Dutch stock market and it can be assumed that Dutch individual stock returns are highly linked to the returns of this index:
it mt
it R
R = +
ε
(4)This equation is a simplified version of equation (5) used for the market & risk adjusted return model. However, in this model systematic risk (α) is assumed to be non-existent while systematic risk (β) is assumed to be equal to 1 and thus reflecting market movements (Rmt).
Subsequently, the abnormal returns are calculated by taking the difference between the observed return of stock i on day t and the return of the market-index m on day t. Again, the error-term of stock i on day t (εit) then can be interpreted as the abnormal return of stock i on
day t
C.1.3. Market & Risk Adjusted Return Model
26 it mt i i it R R =
α
ˆ +β
ˆ +ε
, where var(ε )=σ2 it (5)Subsequently, the abnormal returns are calculated similarly as in previous two models and again equals the error-term of stock i on day t (εit).
C.1.4. GARCH Market & Risk Adjusted Return Model
Previous studies (e.g. Corhay and Tourani Rad, 1994) indicate time dependence to be present in analyzing daily stock returns. The assumed constant variance of the error terms as indicated by equation (5) is therefore likely not to lead to efficient parameters and consistent test statistics and a correction should therefore be made. I make use of a GARCH (1,1) model to estimate the parameters more efficiently as suggested by Corhay and Tourani Red (1996), who note that including more lags does not lead to an improved fit. The formulation of the GARCH corrected model stays equal to the market & risk adjusted return model as in equation (5), only the variance is defined differently in order to allow for time dependence:
2 1 , , 2 2 , 1 , 0 2 1 , ) var( = = + + − − i it i i it it
σ
α
α
ε
itα
σ
ε
(6)C.1.5. Calculation of Test Statistics
The abnormal returns are calculated separately for each revision i at day t for the four different models. In order to provide evidence for the general impact of a revision, I aggregate and average the separate abnormal returns for each revision with N being the total number of revisions. In doing so, I find an average abnormal return for each day in the event window:
∑
= = N i t it t N AR AAR 1 (7)In case the AARt substantially deviates from zero, stock recommendation revisions on this particular day have a significant influence. To calculate whether the average abnormal returns are significant in the days of the event window I make use of an ordinary t-test:
) 1 , 0 ( ~ ) var( 12 t AAR AAR t= t (8)
Where the variance of the average abnormal return is calculated as follows:
∑
= = N i t i N AAR 1 2 2 1 ) var( σε (9)27
in the event window shows a significant pattern. These cumulative average abnormal returns are calculated by adding the average abnormal returns of days in the event window (τt) for varying intervals:
∑
= = 2 1 ) , ( 1 2 τ τ τ ττ
τ
AARi CAAR (10)Subsequently, I also conduct t-tests on these cumulative abnormal returns to test for significance of stock recommendation revisions over a certain interval:
~ (0,1) )) , ( var( ) , ( 2 1 2 1 2 1 t CAAR CAAR t
τ
τ
τ
τ
= (11)Where the variance is calculated as follows:
∑
= = 2 1 ) var( )) , ( var( 1 2 τ τ τ ττ
τ
AARi CAAR (12)I make use of two-sided t-tests since the literature is at odds whether the stock price reaction following a recommendation revision is positive or negative.
Additional to the parametric t-test, I carry out a non-parametric Wilcoxon Signed-Rank test. Brown and Warner (1980, 1985) point out that solely analyzing data with a t-test can lead to a distorted picture, since a t-test assumes normality of the data. Additionally, by conducting a non-parametric test, robustness of the results can be enhanced.
The Wilcoxon tests the null-hypothesis whether the median of the average abnormal returns is equal to zero. The test is conducted for each day in the event window. First, the Z-value is calculated by taking the difference between the observed return on day i in the event window (Yi) and the median of the returns on this day (Xi):
i i
i Y X
Z = − (13)
Subsequently, I rank the differences by their absolute value (Ri). A 1 is assigned for
the smallest value followed by a 2 for the next value, etc.
i
i rankZ
R = (14)
28
∑
= = − N i i i R R 1 for Zi ≤0 (16)The Wilcoxon Signed-Rank test statistic (W+) is the smaller of these two summations. ) , min( + − = + i i R R W (17)
C.1.6. Descriptive Statistics Event Study Models
Appendix C provides an overview of the descriptive statistics of the average abnormal returns in the estimation period for the four different models. In order to make use of the parametric t-test, the data need to be normally distributed. The Jarque-Berra test statistic is calculated in order to measure normality. This test uses the skewness (s) and kurtosis (k). The skewness needs to be close to 0 while the kurtosis needs to be close to 3, in order to indicate normality. The data are normal when the null-hypothesis of normality of the Jarque-Berra test cannot be rejected. Appendix C shows that all models can be tested with parametric tests, with exception of the mean model for upgrades. The significant negative skewness indicates a bigger median than mean, and thus a right-skewed distribution. Additionally, the high kurtosis indicates a leptokurtic distribution, which is caused by a relatively high amount of observations in the tails and around the mean of the distribution.
C.2. Regression
29 it i it UPG TARGET FREQ STOCKS STOCKS NL NAT IBD FREQ BANKS EXP G EXP F REV BANK FA BANK AAR C
ε
β
β
β
β
β
β
β
β
β
β
β
β
β
β
+ + + + + + + + + + + + + + = 13 12 11 10 9 8 7 6 5 4 3 2 1 * * * * * _ * * * * * _ * _ * _ * _ ) ( (18)The advantage of the fixed and random effects model over the OLS regressions is that the intercept βi can differentiate per analyst i as opposed to an OLS regression where the intercept β0 is constant. These varying intercepts βi are able to capture the unobserved analyst characteristics (i.e. gender, sector specialization, etc.), which are not included as independent characteristics in the regression. The main difference between the random and fixed effects model, is that with fixed effects estimation the intercepts βi are independent and do not have any restrictions. The intercepts βi of the random effects model on the other hand are drawn from a normal distribution. Due to this difference, one can only include independent variables which fluctuate per analyst for the fixed effects model. Of the 12 variables included in equation (18), only 5 meet this criteria, i.e. F_EXP, G_EXP, FREQ, TARGET and UPG. This leads to the construction of the following fixed effects model:
it i it UPG TARGET FREQ EXP G EXP F AAR C
ε
β
β
β
β
β
β
+ + + + + + = 5 4 3 2 1 * * * * _ * _ ) ( (19)Additionally, I have incorporated heteroskadistic autocorrelated robust standard-errors as recommended by White (1980). These standard-errors exclude problems of heteroskedasticity and autocorrelation from the estimated regression.
The following sub-sections discuss the dependent and independent variables.
C.2.1. Dependent Variable
30
C.2.2. Independent Variables Experience Hypothesis
In the literature two forms of experience are analyzed; (i) firm-specific and (ii) general experience. In this paper I proxy for both firm-specific (F_EXP) and general experience (G_EXP). For general experience, I take the difference between the date of the relevant recommendation revision and the first-ever recorded recommendation in Bloomberg, similar as to e.g. Clement (1999). Additionally, I have generated a new proxy for general experience; number of banks where the analyst was previously employed (BANKS). I assume that analyst who have worked at multiple banks have more experience.
Firm-specific experience is proxied in a way similar to general experience: I take the difference between the date of the firms’ relevant recommendation revision and the first recommendation regarding this firm made in Bloomberg.
Literature (e.g. Mikhail, 1997; and Clement, 1999) has indicated experience to have a positive effect on the size of the impact following a revision. I also expect that learning-by-doing effects are present and investors to value analysts’ experience. Investors will hence react stronger to a revision of a highly experienced analyst as compared to a novice analyst.
Firm Size Hypothesis
Size of the brokerage department is a difficult to quantify variable, since the bulk of the research departments do not openly disclose the amount of resources or researchers. I use two proxies for research department size. By summing the amount of researchers covering the Dutch market, similar as e.g. Stickel (1995), I construct a variable which quantifies the size of the Dutch franchise within the research firm: BANK_FA. Secondly, I construct a new proxy by adding the revisions regarding Dutch stocks for each bank (BANK_REV).
As was found in the literature (e.g. Stickel, 1995; and Jacob et al., 1999), research department size is expected to have a positive effect on the size of the impact following a revision. Larger firms are associated with more resources and thus superior information.
Portfolio Complexity Hypothesis
31
complexity. They find that complexity reduces the accuracy (and thus impact) of recommendation revisions. Analysts cannot focus on a selected amount of stocks but instead divert their attention over multiple stocks.
Nationality Hypothesis
Similar as to e.g. Orpurt (2004) and Malloy (2005), a dummy variable is used in order to indicate Dutch nationality (NAT). A 1 indicates Dutch nationality while a 0 resembles foreign nationality. Orpurt (2004) and Malloy (2005) indicate that local presence is associated with an informational advantage. Domestic researchers are more up-to-date about domestic news and domestic companies plus they are accustomed to the local language and culture. Especially in Europe the differences are more pronounced due to the multitude of different countries, cultures and accounting policies. Investors are therefore expected to react stronger to revisions of Dutch researchers.
Investment Banking Department Hypothesis
Another dummy variable is used in order to classify whether a research department operates within an investment bank (IBD). A 1 denotes a research department which has close affiliations with an investment banking department. Similar research uses a comparable dummy variable (e.g. Michaely and Womack, 1999).
The effect of this variable on the magnitude of the stock impact following a revision is widely researched. It is found that an investment banking department has a negative impact on the accuracy and stock price impact (e.g. Cliff, 2004; and Barber et al., 2007). Research firms which operate closely with its investment banking departments tend to issue overly-optimistic recommendations in order to keep a close relationship with firms. This close relationship will allow for potential future investment banking fees. However, most research regarding the effect of investment banks on research reports is rather outdated. Since the Global Settlement in 2003, connections between the two departments have been fading. Through this research, I analyze whether indeed a clear separation between the two departments is achieved.
Frequency Hypothesis
32
conducted regarding the impact of this variable on the size of the stock return following a revision. Usually, recommendations are issued for a period of 6 to 12 months (Francis and Soffer; 1997). I expect analysts which tend to change their recommendations regarding stocks continuously (and thus reflecting high frequency) to work at the least reputable research firms. By making many reiterations, they try to establish a name for themselves. Besides, I expect the most-notable researchers to only revise when necessary and thus reflect low frequency.
Size of Revision Hypothesis
So far, the size of the revision has not been researched in a research setup similar to mine. I incorporate two different proxies for size; target price (TARGET) and upgrade (UPG/DGD).
Bloomberg provides a comprehensive overview of recommendations including target prices. I expect that a big difference between the target price and market price indicates a high amount of analyst confidence to which investors react strongly. TARGET is calculated by dividing the difference between the target and market price by the market price.
Secondly, I include the upgrade variable in the regression. Previous literature did not make a clear distinction between one- and two-step revisions. However, I have included a dummy variable which indicate a two-step revision. I expect these revisions to impact stock returns more heavily than one-step upgrades.
Table 7 provides an overview of the hypotheses, variables and measures used to test the hypotheses (signs reflect expected coefficient signs following upgrades).
Table 7: Overview of the Hypotheses, Variables and Corresponding Measures Used to Test the Hypotheses and the Expected Coefficient Signs
Hypotheses Variable Measure Exp.
Sign
H2 Experience G_EXP # of Months General Experience +
F_EXP # of Months Firm-Specific Experience +
BANKS # of Banks Previously Worked +
H3 Firm Size BANK_FA # of Analysts Covering NL Market at Bank +
BANK_REV # of NL Revisions at Bank +
H4 Portfolio Complexity STOCKS # of Stocks Covered by Analyst -
NL_STOCKS # of NL Stocks Covered by Analyst -
H5 Nationality NAT Dutch Nationality (1 = Yes) +
H6 IBD IBD Operating under Investment Bank (1 = Yes) -
H7 Frequency FREQ # of Reiterations / Months to Revision -
H8 Size of Revision TARGET Target – Market / Market Price +
33 V. RESULTS
This section is divided in two main subsections. Firstly, I depict and describe the results of the event study. Secondly, I give descriptions of the results of the regression analysis. Both subsections are further divided into up- and downgrades.
D.1. Event Study Results
D.1.1. Event Study Results - Upgrades
Table 8 depicts the event study results following upgrades for the 4 models, both the parametric t-test and non-parametric W+ test statistics are depicted for each model.
Table 8: Daily Average Abnormal Returns and Corresponding Significance Levels Following Stock Recommendation Upgrades for the Mean Adjusted, Market Adjusted, Market & Risk
Adjusted and Market & Risk GARCH Adjusted Return Model1
Mean Market Market & Risk GARCH
Day AAR t-value W+ AAR t-value W+ AAR t-value W+ AAR t-value W+ -2 -0.05% -1.10 0.31 -0.06% -1.75 0.49 -0.02% -0.62 0.06 -0.02% -0.68 0.00 -1 0.13% 2.60** 2.10* 0.16% 4.71** 2.53* 0.18% 5.77** 3.42** 0.18% 5.58** 3.44** 0 0.42%a 8.52** 7.36** 0.38%a 11.54** 7.81** 0.40%a 12.59** 8.37** 0.39%a 12.41** 8.33** 1 0.16% 3.31** 1.37 0.12% 3.75** 0.60 0.17% 5.37** 1.60 0.17% 5.34** 1.63 2 0.01% 0.11 0.07 0.01% 0.32 0.36 0.03% 0.81 0.72 0.03% 0.79 0.79 3 -0.12% -2.48* 0.06 -0.11% -3.23** 1.74 -0.07% -2.36* 0.90 -0.08% -2.51* 0.99 4 0.06% 1.13 1.58 0.02% 0.54 0.74 0.01% 0.34 0.21 0.01% 0.20 0.25 5 0.00% 0.08 0.71 0.03% 0.85 0.05 0.01% 0.32 0.38 0.01% 0.32 0.23
* and ** indicate statistical significance at the 5% and 1% level respectively (two-tailed test)
a indicate whether the AAR on day 0 significantly differs from day -1 at the 1% level (one-tailed test)2
Surprisingly, statistic significant results occur prior to the announcement day. On day -1 there already is a highly significant positive return of between 0.13% and 0.18% prior to an upgrade. Most other researchers (e.g. Stickel, 1995; and Jegadeesh and Kim, 2006) also find pre-announcement day effects and point out that these are potentially caused by leakage effects, where the market picks up signals of an upcoming revisions upgrade.
1
The results (and respective significance levels) do not change in a significant way when filtering out 58 overlapping events, i.e. only the revision of analyst A is taken into account when analysts A, B and C covering stock X issue revisions on days 0, 2 and 4 respectively.
2
Test-statistic calculated as follows:
