Master Thesis Finance
Value of Sell-side Analyst Recommendations:
Impact on the US Market Controlled for event-induced Variance & Cross-correlation By Bart van Steenhoven
Abstract
Using traditional event study methodology developed by Brown and Warner (1981, 1985) I find consistent with earlier research a tremendous impact from analyst recommendations on stock prices on the US market. I am the first researcher to control for event-induced variance using the event study method of Boehmer, Musumeci & Poulsen (1991). I find that an issued sell recommendation is no longer of significant effect. Further testing using the Kolari & Pynnönen (2010) method to control for a level of cross-correlation yields the result that the effects of analyst recommendations is no longer significant if cross-correlation is present. A subsample is created using a set of full service investment bank recommendations. The BMP (1991) method the results contradict the results from the previous mixed diversified sample of recommendations. Using the KP (2010) method only for some small event windows the returns remain significant for small levels of cross-correlation. This research adds to the existing literature that whilst the influence of recommendation changes is heavily overstated.
JEL classification: G24, G14, G15, G10, C18, C12
Keywords: sell-side security analysts, event study methodology, brokerage house reputation, capital markets, cumulative abnormal returns, analyst recommendations
Author: Bart Frederik Cuno van Steenhoven Mail: bartvansteenhoven@hotmail.com Phone: +31646213681
Student number: S1882910
Place and date: Groningen, 26/06/2015 Supervisor: J.J. Bosma
1. Introduction
‘’Real knowledge is to know the extent of one's ignorance’’ -Confucius This paper investigates the value of a sell-side analyst1 recommendation on a firm’s stock in the US market over the time span 2003 to 2013. Sell-side security analysts provide information to aid decisions of investors to sell, hold or buy a certain stock. Usually analysts cover a range of stocks within certain industries and try to provide guidance in the direction of the share price movement over a certain time span. For an investment bank, having a reputable analyst who knows the industry can also help to enable secure other types of banking activities such as M&A activities or underwriting the firm in raising equity capital from the market. Analyst recommendations are built up in two ways: a recommendation of an analyst usually responds to news events that affects the company and provides an outlook of the company’s future cash flows and valuation.
Investors react strongly to recommendation upgrades or downgrades, whether or not they are coincidental with other corporate news (Womack & Michaely, 2005). Analyst recommendations have an important perceived information content. Over the years various researchers have provided evidence that announcements of changes in recommendation levels have a significant impact on a firm’s stock price in such a way that the stock price followed the direction of the analyst recommendation. So far any empirical research that is published to date finds that the market responds positively to upward changes in recommendation level and negative to downward changes in the level of recommendation (Womack & Michaely, 2005; Jegadeesh, Kim, Krische & Lee, 2004; Stickel,1995; Barber, Lehavy, McNichols & Trueman, 2001).
Analyst recommendations do contain an element of bias towards being favorable. Earlier papers document that analysts rarely issue sell or strong sell recommendations. Jegadeesh, Kim, Krische & Lee (2004) find in their fifteen year sample that only 5% of the issued recommendations are sell recommendations. Furthermore the research of Lin & McNichols (1998) and Michaely & Womack (1999) indicate that an analyst who works a for top-tier investment bank issues more positive recommendations than other analysts. A possible explanation for this effect is that these banks rely on their underwriting businesses and need positive recommendations to secure new equity issues by the covered firms. Barber, Lehavy, McNichols & Trueman (2001), Jegadeesh et al. (2004), and Boni & Womack(2003) all find that the stocks with a buy recommendation outperform the stocks with a sell recommendation over a period of one year and further. This indicates that an investor who adjusts
1 ‘’Sell-side analysts” are securities analysts employed by banks and brokerage firms (e.g. selling parties from
brokerage transactions)
his/her portfolio of stocks according to the recommendation changes of the analysts can yield positive returns of their portfolio following the reports. The analyst can provide an added value to investors by better interpreting and analyzing publicly available information that is not fully understood by the market.
The existing literature studies the effects of analyst recommendations on stock prices according to the traditional event study methodology developed by Brown and Warner (1980,1985). Although this method is the main method of event study methodology to detect abnormal stock returns, a growing number of researchers have improved the traditional event study methodology methods leading to more reliable results(Charest,1978; Collins and Dent, 1984; Ball and Torous 1988; Corrado ,1989 and Boehmer, Musumeci & Poulsen, 1991). The improved methods help researchers to evade the detection of abnormal performance when none is present. The main issue in overestimating the results comes from the variance-estimation for the abnormal returns. The variance is not correctly measured because of an underestimation in de standard deviation of the cross-sectional distribution of cumulative abnormal returns. Brown and Warners (1980,1985) do not provide in their methodology not a solution in dealing with event-induced variance. Event-induced variance occurs when the variance during the event window exceeds the variance over the estimation period. The estimation window is the period before the event window that is used to estimate the normal returns of the firms in the event window. Several papers have confirmed this effect of event-induced variance (Beaver, 1968; Patell, 1976; Dann, 1981; Kalay and Lowenstein,1985; Rosenstein & Wyatt, 1990). However not a single researcher documents the possible occurrence of event-induced variance in the event studies of analyst recommendations.
Macro-economic, government policies or industry specific factors can generate simultaneous movements in stock returns of different firms. The modern portfolio theory is based on the fact that there is correlation between stock returns. If a certain company reports or is affected by positive news, it is likely that another firm is also affected by this news event, in a positive or negative direction. An investor who strives for a portfolio that has a certain amount of risk should reduce the amount of correlation by diversification between the selected stocks to optimize the return for the desired level of risk (Fama & French, 1992). This reduces the covariance between the stock in the portfolio and therefore the variance of the portfolio. The degree of cross-dependence decreases in the effectiveness of the risk-adjustment approach and increases in the homogeneity of the sample firms examined (Kothari & warner, 1997). This means that the stock behaves in similar fashion, or in case of a negative relation move into opposite direction. This effect would be the greatest if the event dates are clustered or a lot of firms are from the same industry in the sample (industry
effects). With regards to recommendation changes it is inevitable that most of the event dates are similar for each stock. This creates cross-correlation in the abnormal returns from the recommendation changes. Usually an analyst recommendation is changed for four reasons. If new quarterly earnings statements are released most of the analysts update their recommendation the day after the release of the quarterly earnings . Most companies release their quarterly earnings in the same period. The event date is therefore non-random and the events are clustered around a common calendar period. Some analyst recommendations changes occur because of a corporate event such as a stock repurchase or merger announcement. In this case multiple analysts from different firms react to this event. These kind of corporate events occur in waves linked to the economy’s overall performance. Roughly every 5 years there is a major increase in M&A activity, and therefore an increase in releases of analyst recommendations. Another reason for a change in recommendation level is the occurrence of a worldwide economic event, such as the financial crisis of 2008. Here especially stocks from the financial industry received a recommendation change (mostly negative). Lastly analyst recommendation changes can be the result of extensive research on the stock from an analyst and may be not directly linked to one certain news event. In this the event date of release is random and less affected by cross-correlation in the abnormal returns. Overrepresentation of firms with the same industry affects the amount of cross-correlation as well. Industries with a cluster, where a small amount of firms control the market experience this effect the most (King, 1966). The initial sample needs to be filtered of recommendation changes that are affected by these other events. If another event is in the event window of the recommendation, the effect measured cannot be solely attributable to the recommendation change. A news event can already be priced in before the release of the recommendation change. Whilst Barber, Lehavy,
McNichols, & Trueman (2003 & 2010) and Bamber(1987) control for the earnings announcement
around recommendation changes other relevant news to analyst recommendation changes is not taken into account. The assumption that there is no cross-correlation between the abnormal return is very hard to defend considering the factors mentioned above. For this reason Kolari & Pynnönen (2010) developed a testing method that can provide more powerful test statistics because it can cope with levels of cross-correlation.
On this basis, it is possible that previous studies may have overstated the effect of analyst recommendations on stock prices. This is due the fact that the basic assumptions of the traditional event study are not met with regard to analyst recommendations. Another possibility is that the recommendation is shortly before or after a major news event that on its self also affects the stock price. So far no research has taken into account the effect of overlapping events or researched the effect with the improved event study methodology.
Based on the observations mentioned the research question is formed as follows:
Are sell-side analyst stock recommendations still of value under more advanced event study method testing when controlling for event-induced variance and cross-correlation?
In order to answer this question I investigate all the stocks listed on the NYSE, AMEX, and NASDAQ and in particular the stocks covered by sell-side analysts. I examine the abnormal returns under the 4-factor model of Fama and French (1993) and Carhart (1997). Furthermore, I evaluate the impact of the recommendations when earnings announcements, heavy industry changes, economic and firm specific changes are filtered from the sample. The main adjustment in the research framework comes from controlling for event-induced variance and to see if the outcome differs if there is a level of cross-correlation between the abnormal returns. The objective of the study is to see when applying these methods if there is still a significant effect attributable to recommendation changes. This is reflected in a firms’ abnormal returns around the event dates.
This paper contributes to the existing literature by examining the true effect of analyst recommendations to see if the effects still hold under new improved event study methods. It is the first of its kind to implement this strict data restrictions and using the better test measurements. This paper is of interest to sell-side analysts, policy makers, governments, researchers and practitioners because it draws a more accurate conclusion on the value of the analyst recommendation to a firms’ stock price.
The remainder of this paper is structured as follows: the second section is attributed to the literature review that provides an overview of the relevant studies done on the subject of the value of analyst recommendations and event study methodology. Section three describes the methodology and the data used in order to be able to carry out the research. Section four provides an overview of the results and critically evaluates the empirical results and possible explanations for the results are presented. The fifth and final section contains the conclusion, which answers the main question. This section also provides a discussion on the restrictions found and a proposal for a follow-up research/study.
2. Theoretical framework
This section explains the theory behind the traditional event study methodology, the improved insights and significance testing gained over the years by various researchers. In the second part, the current research on the impact of analyst recommendations will be discussed. The third part will combine the theory on event study methodology and analyst recommendations.
2.1 Event study methodology
Event studies are the primary methodology used to assess the effect that the occurrence of an event has an impact on the returns of a firm’s common stock price. The methodology is based on that the market is semi-strong efficient and reflects all new publicly information. This assumes that new information brought to the market will have an impact on the stock price of the affected firm. Investors react to the new information but cannot benefit above the return of the market using this new information (Fama, 1969, 1970).
The most common methodology is developed by Brown and Warner (1980) using monthly stock returns. They developed a return-generating model and testing procedures, to identify the presence of abnormal returns under a variety of event conditions. In 1984 they extended their analysis using daily stock returns. The use of daily data and knowing the exact event date adds to the power of their model because the timing of the event and the stock price reaction is more accurate. Their own research and the research of Berry, Gallinger and Henderson (1990) points out that daily returns are non-normally distributed. However the abnormal daily returns exhibit a (nearly) normal distribution such that non-parametric models(data is not drawn from a probability distribution but is ranked) do not add power (Brown and Warner 1980, 1985 ; Berry, Gallinger and Henderson, 1990). Dyckman, Philbrick & Stephan (1984) examine the serial dependence of abnormal returns, cross-sectional dependence of abnormal returns, and the increase in variance of returns around events for the effects on the interpretation of hypothesis tests for the model of Brown and Warner (1980, 1985). They find consistent with Brown and Warner (1980, 1985) that the model is adequate in measuring effect attributable to the researched events. However the model can lack statistical power in the case of event-induced variance and event date clustering. Event date clustering by industry or time appears in general to reduce the ability of traditional methods to detect abnormal performance. Overall they support that the event study method of Brown and Warner (1980, 1985) yields test results that have power to support the tested hypotheses.
The market model says that the return on a security depends on the return on the market portfolio and the extent of the security's responsiveness as measured by beta. To enhance the predicting power of the market model Fama and French (1993) developed the three-factor model, which was
further modified by Carhart (1997) to incorporate the momentum factor. The three-factor model was a response to the CAPM model to measure returns (improvement of the market model). The CAPM assumes a linear relationship between the expected return of a stock and its beta, which captures the stocks risk as opposed to the market. Also the time value of the investment is taken into account (the risk free rate). Fama and French (1993) discovered that average excess portfolio returns are influenced by three factors: the abnormal return of the market, the difference between abnormal returns of a small market capitalization stock portfolio minus the abnormal returns of large market capitalization stocks portfolio (SMB) and the abnormal returns of high price-to-book stocks minus low book stocks (HML). They find that small cap stocks with a low price-to-book ratio (value stocks) outperform large cap stocks with a high price-to-price-to-book ratio (growth stocks). The first kind of stocks do however have a higher perceived risk, but this is captured by the beta. Carhart (1997) added the momentum factor to the three-factor model. He finds that stock who were past winners continue to increase in value and stocks who were past losers continue to decline in value. This four-factor model has a far higher accuracy in measuring returns as opposed to the market model. Brav & Gompers (1997) find that incorporating these four factors do not truly represent the total measure of risk or an indication of market inefficiency. However in an event study, the size, book-to-market, and momentum factors are used to provide a more adequate measure of abnormal returns and to isolate the incremental impact of the stock returns. The four factors are applicable to all the stocks in the sample, not only the firms who experience the proposed event. Therefore the usage of the model in event study is one of the most adequate and accurate methods for determining the returns. The 4-factor model is used to generate the expected returns for the event and estimation window periods.
Jovanovic & Fox (2010) discovered that during times of high volatility in the market, many firms show a significant abnormal return. This means that the state of the economy affected the sample as a whole and findings that a researches tries to attribute to another event as opposed to the financial crisis is overstated. This leads to a possible type 1 error by falsely accepting the hypotheses. A basic assumption in traditional event study methodology is that the abnormal returns are cross-sectional uncorrelated. The presence of cross-sectional correlation in the specific security excess returns (Brown and Warner, 1980) can affect the power of statistical tests if the method of variance estimation assumes cross-sectional independence of returns. If positive cross-sectional correlation is present, such tests result in inappropriately high rejection rates of the null hypothesis whether the null is true or false. This assumption is valid when the event day is not common to the firms. Even in the case when the event day is common, if the firms are not from the same industry, Brown and
Warner (1982, 1985) show that use of the market model to derive the abnormal return reduces the inter-correlations nearly to zero and, hence, can be ignored in the analysis. However, firms within the same industry share similar properties and can influence each other. Certain events such as regulation or news event can impact the industry as a whole or can cause spill over from one firm to comparable industry firms. Therefore, assuming that there is no correlation between the stocks in the sample would yield over powered test results. Therefore this needs to be taken into account to correctly interpret the effects of the events.
Boehmer et al. (1991) provide evidence that their standardized cross-sectional test (requiring an estimation window) exhibits a comparable size, but is more powerful. Beaver (1968) and Patel (1976) provides evidence that the variance shifts contemporaneous with the financial event. Event-induced variance occurs when the variance during the event window exceeds the variance over the estimation period. Boehmer et al. (1991) estimate the cross-sectional variance of the standardized abnormal returns and define a t-statistic (BMP t statistic) as where 𝑆𝑆𝑆𝑆 is the (cross-sectional) standard deviation of the standardized abnormal returns. The advantage of the BMP (1991) method is that the method weighs individual observations by the inverse of the standard deviation. This implies that more volatile observations get less weight in the averaging than the less volatile and hence more reliable observations. Particularly relevant to the present study, methods based on standardized abnormal returns have been found to outperform those based on non-standardized returns. The BMP statistic has gained popularity because it is found to be more robust with respect to possible volatility changes associated with the event. The BMP (1991) approach reduces the no-volatility-impact, and estimates the (common) event-day-volatility cross-sectional with the usual sample standard deviation. To account for the dependence across firms' average residuals, in event time, Brown and Warner (1980) suggest that the standard deviation of average residuals should be estimated from the time series of the average abnormal returns over the estimation period. The cumulative abnormal returns from the firms should be divided by the standard deviation from the estimation window. This adjustment results that each cumulative abnormal return has a variance of one and is therefore equal for each event in the sample. ‘’However, when the event days are clustered, the standardized abnormal returns are potentially correlated, which can bias the volatility estimates in both cases’’ (Kolari & Pynnönen, 2010).
Kolari & Pynnönen (2010) have demonstrated via simulation that, using the traditional standardized return test statistics, even moderate cross-sectional correlation in an event study causes substantial over-rejection of the null hypothesis. The adjusted BMP statistic they propose is robust against levels of cross-correlation that are present in the sample. However, in order to get reliable results with the
BMP (1991) method, the dataset must include a sufficient amount of firms for the cross-sectional volatility estimation. They utilize scaled (or standardized) abnormal returns and propose a modfied t-test statistic from the BMP method (1991) that takes into account both cross-correlation and inflation of event-date variance. This will further enhance the quality of the t-statistic in order to make sure that the null hypotheses can be accepted or rejected. In the methodology section the different formulas for the event study methods and their significance tests are presented.
2.2.1 Current literature on the impact of analyst recommendations on stocks
Information is costly to process. Brokerage firms spend a lot of resources analyzing stocks and trying to persuade investors that certain stocks are more or less attractive than others. Sell-side security analysts make investment recommendations to individual and institutional investors. By making a stock recommendation, financial analysts express their expectation about the relative near-term return performance of a given firm.Grossman and Stiglitz (1980) observe that market prices cannot perfectly reflect all available information. This is where an analyst of a brokerage firm steps in, to gather information of a firm and to provide specialized information about the firm. Brokerage research builds on factual sources of firm-specific information such as annual reports, earnings announcements or industry events. The recommendation based on these facts tries to predict the firms future performance and therefore is speculative in nature. These analysts produce corporate earnings forecasts, write reports on individual companies, provide industry and sector analyses, and issue stock recommendations. A skilled analyst is either able to access private information or has an outstanding ability to interpret publicly available information.
Womack(1996) was one of the first researchers that was able to work with real time recommendation data from First Call, a database that is able to provide recommendation changes from the 14 most prominent U.S. brokers on domestic stocks. Womack finds that recommendations by the large nationally known brokerage firms are predominantly issued on well-followed, large-capitalization stocks. Furthermore, he finds that recommendation changes to a buy or strong buy level in a 3-day event window was over 3% positive. A stock where the recommendation was downscaled to a underperform or sell recommendation has a price drop of over 4.5%. Because they are issued seldom, the value to the investor is of greater significance than a buy recommendation. On the longer-term returns following unfavorable recommendations (sell or underperform) appear to include both accurate stock selection and industry selection components. Because a sell recommendation is issued seldom, the stock most likely has already a hampered reputation and for the analyst to maintain a credible reputation must comply with the market response.
Barber, Lehavy, McNichols, and Trueman (2001) also support the findings of Womack (1996) that sell-side analysts’ stock recommendations have significant value. They document that stocks with a more favorable consensus (average) recommendations outperformed those with less favorable recommendations. A portfolio comprised of the most highly recommended stocks, for example, generated an average annual market-adjusted return of 4% percent while a portfolio of the least favorably recommended ones yielded an average annual market-adjusted return of -9%. They have, however, incentives to distort this information upward. Positive recommendations are more likely to generate trading commissions than negative recommendations, given short-selling constraints (Malmendier & Shanthikumar, 2004).
Barber, Lehavy, McNichols & Truman (2002) control for the 4-factor model for the stock returns as proposed by Fama & French (1992) & Carhart (1997). Theycalculate each portfolio’s (portfolios are sorted by the different levels of recommendations) abnormal return, controlling for the return expected on the portfolio given the beta, size, book-to-market ratio, and price momentum of each of its component stocks. Similar to the conclusion drawn for market-adjusted returns, they find
the most highly recommended stocks earned a higher average annual abnormal return than did the least favorably recommended stocks. The most highly recommended stocks were generally small, with low book-to-market ratios (growth stocks), while the least favorably recommended stocks, although also small, had high book-to-market ratios (value stocks).
Jegadeesh, Kim, Krische & Lee (2004) examine the relation of the recommendations to other available public information. They focus on variables that prior studies show have some predictive power for future returns, and critically evaluate the investment value of these recommendations in light of the other signals. They find that analysts prefer high momentum stocks and growth stocks. This is positively correlated with momentum, similar to the findings of Barber, Lehavy, McNichols & Truman (2002). One interpretation of their finding is that recommendation changes capture qualitative aspects of a firm's operations (e.g., managerial abilities, strategic alliances, intangible assets, or other growth opportunities) . Their evidence is at least consistent with the analysts' claim that they bring new information to market. From an investment perspective, their results suggest analyst recommendations play a dual role in the price formation process. On the one hand, analysts seem more occupied with growth and glamour stocks. To the extent that their opinion affects public sentiment, this evidence is consistent with the view that they contribute to noise trading in the market. On the other hand, these findings suggest analyst recommendations can still play a useful role in investment strategies and to provide a bundled interpretation of the available market information.
The value of an analyst recommendation is dependent on several factors than just the timing of the recommendation and the information in the report itself. Earlier research proved that the value of an analyst recommendation is affected by many variables. Womack (1996) and Ryan & Taffler (2006) find that the value of an analyst recommendation is greater if the firm has a smaller market capitalization. Smaller firms are in general covered by fewer analysts, therefore one’s opinion will be more influential. The implication is that information about smaller firms is gathered and processed less frequently, the impact of any single information release is greater and, therefore, both buy and sell recommendations should have a greater price impact for smaller firms than larger firms. Also they support the current literature that the abnormal returns are higher if the recommendation change is contemporaneous with an earnings announcement.
One can wonder if all analyst recommendations contribute to a better market efficiency, if there are so many other factors involved what determines the value of an analyst recommendation. In general an analyst recommendation helps the market interpret the available information and the report can bring new information for the market. An analyst can create clarity in all the different information that is provided over a certain stock. It gives a one direction opinion on the companies prospected future which gives investor a clear decision on whether the company is a good or bad investment. As a result fewer noisy signals are misinterpreted for information, resulting in less noise in price fluctuations (Schutte & Unlu, 2007). In this context, noise is the price variation resulting from trades on noisy signals that investors misinterpret for information (Black (1986). This makes the actual reason for a shift in a company’s stock price harder to identify. An analyst has the power to reduce noise in the market by accumulating all news events around the company.
However some analysts may have distorted incentives. Barber et al. (2007) find that analysts employed by investment banks provide less profitable buy recommendations than analysts employed by independent research firms. Recommendation profitability can also differ based on corporate events such as earnings announcements or a policy industry related change.
2.2.2 Controversy about the neutrality of analyst recommendations
Lead and co-underwriter analysts’ growth forecasts and recommendations are significantly more favorable than those made by unaffiliated analysts, although their earnings forecasts are not generally greater (Lyn & Nichols, 1998). Conflicts between the desire of corporate finance to complete transactions and the need of brokerage analysts to enhance and protect their reputations are most likely come to play during an IPO process. This market is very lucrative to the investment
banking industry. Second implicit in the relationship between the firm and the underwriter is to follow the newly issued stock and provide (positive) analyst coverage. This coverage is of essence to new firms to enhance their value and knowledge about the firm and industry to institutional investors (Michaely & Womack, 1999).
Ljungqvist, Marston, Wilhelm (2006) document that analysts are willing to sacrifice their objectivity when they are pressured from investment bankers who work at the same firm. They find that a trade-off exists between the analyst's career concerns (the cost of jeopardizing his/her reputation) and the incentives bankers may have provided analysts to bias their recommendations. The research confirms that analysts are more aggressive when a potential fee income was at stake. Notable is that the most reputable analysts are less prone to this aggressive behavior. Having the leading industry analyst is a perk that will generate enough underwriting activities and losing this position will reflect in a lower utility overall.
2.3.1 Event studies and analyst recommendations
Analysts often write reports on days of firm-specific news, and recommendation changes on such days are more likely to be favorable if the news is positive (Loh & Stulz, 2010). Though the traditional event study method reduces or even eliminates the impact of confounding news on the average abnormal return, it does so only when news and the probability of occurrence of the event are uncorrelated. In the case of analysts, there is no reason to believe that this condition holds. It is therefore important to construct a sample of recommendation changes where the impact of confounding firm-specific news is minimized. Not surprisingly, eliminating firm-specific news days by implementing the data restrictions reduces the stock-price reaction to analyst recommendation changes, but the average stock-price reaction remains statistically significant. Because analyst recommendations are based on corporate and industry information, a recommendation change is closely tied to these events. A strict selection in event dates must take place to select the changes in recommendation levels that are not based on a widespread or firms event.
With regards to event-induced volatility, it is not uncommon for an event to cause an increase in the cross-sectional dispersion of a stocks’ returns. Moreover, the degree of dispersion may vary tremendously from one firm to the next . If researchers fail to control for these varying degrees of dispersion across firms, they will generally observe a variance increase on the event day (Seiler, 2000). Such varying firm effects lead to an increase in measured cross-sectional dispersion that actually reflects the failure to control for all relevant return influencing factors. If the variance is understated, as has been found to be the case using traditional event study methodology, the null
hypothesis of zero abnormal returns is rejected more frequently than it should, even though the average abnormal return is significantly close to zero. Hence, many studies may appear to find abnormal performance due to a failure to consider event-induced variance. As stated in the introduction, not a single study about the value of analyst recommendation controls for event-induced variance, however a certain level of cross-correlation is likely present in event studies with same industry firms. This research will show the effects on the significance levels when levels of cross-correlation are taken into account.
2.3.2 Presence of momentum in stock returns following recommendation changes
Using the 4-factor model to estimate the expected return of a stock provides valuable insights following from the coefficients used to determine the expected return of a stock. The momentum factor added by Carhart (1997) is of particular interest. This coefficient controls for momentum in stock prices. Momentum in a stock is described as the tendency for the stock price to continue to rise in value if it was already rising and continue to decline in value if it already was declining.
Whilst Carhart (1997) acknowledges that momentum is present in the stocks, he does not provide an explanation to the occurrence of the phenomenon. Chan, Jegadeesh & Lanishok (1996) provide evidence that momentum can be explained through a delayed market response to new information. The late response to new information creates a drift in the firm’s stock future returns which is of importance for at least 6 months. An investor can obtain abnormal profits by swiftly responding to trends in the market (Hong & Stein, 1998).
Security analysts tend do not directly incorporate new negative news about their covered firms in their forecasts. Analysts tend to provide to optimistic forecasts, but downscale their forecasts towards normal expectations over time (Klein, 1990). This can be linked to the bias in analyst recommendations and their incentives to not jeopardize the relationship with the firm’s management because of other banking activities. They initially do not adjust their earnings or recommendation forecast, but wait for more confirmation that the outlook needs to be negatively adjusted. Investors who rely on the interpretation from analysts to new events have a delayed response because of this wait. This phenomenon adds to momentum in the stock price.
Finding that momentum is present in the sample of an event study research imposes severe biases in the eventual outcome of the results. The aggregation of past winners and past losers in two portfolios and finding that the portfolio moves accordingly in their past direction means that the cross -sectional variation in expected returns is influenced by momentum. Using the event study methodology of Kolari & Pynnönen (2010), which test takes into account a certain level of
correlation between the stocks creates a more powerful test than previous popular event study methods. In case that the coefficient from the momentum factor in the 4-factor model significantly deviate from zero a researcher must take into account that the result are influenced by momentum in stock prices.
3. Sample selection
The initial sample consists of all the recommendations on US stocks by analysts that are recorded in the IBES database from January 1, 2001 to December 31, 2013. The IBES database captures recommendation data over 40,000 firms from 900 brokerage houses. The IBES database has over 33 data items and detailed information about the issued recommendations. It is the most comprehensive and used database for portfolio managers, analysts and researchers. The time span is chosen to counter the effect of non-normality in daily excess returns. The mean excess return in a cross-section of securities converges to normality as the number of sample securities increases (Brown & Warner, 1985). The analysts in the sample are analysts who work for independent brokerages or (full-service) investment banks globally. Each event in the database represents the issuance of a recommendation by a particular brokerage firm analyst over a specific company. Recommendations in the IBES database are coded as follows: 1 = strong buy, 2 = buy, 3 = hold, 4 = sell, 5 = strong sell. The hold recommendations for are not taken into the research sample due to the fact that the analyst believes that the stock will keep pace but will not outperform or underperform the market. Therefore the value of the recommendation is around the performance of the benchmark index and will not yield an abnormal return that is representative to measure an effect. Stickel (1995) find that there is no greater significant influence in recommendations that skip a level. Therefore only the level from the issued recommendation will be taken into the research.
Analyst recommendations rely on the earnings announcements to make their recommendations and interpret the information given in the quarterly reports. The information of the quarterly earnings is available in the market and affects the share price positive or negative if the earnings are not in line with the forecast. The post earnings announcement drift in the share price, as a reaction to the new provided information is a well-documented phenomenon. The post earnings announcement drift literature shows that stocks with positive earnings surprises tend to earn positive abnormal returns and stocks with negative earnings surprises tend to earn negative abnormal returns (Bernard and Thomas, 1989). This implies that the abnormal return around an analyst recommendation that is close to an earnings announcement will be greater, but is incorrectly measured to attribute the effect to the recommendation. Chari, Jagganathan & Ofer (1988) find significant effects of the post earnings announcement drift up to 5 days before and after the earnings announcement. Bamber(1987) controls for abnormal daily trading volume around quarterly earnings announcements and finds significance up to 5 days. Therefore to control for these both effects any recommendation within 10 days before or after the earnings announcement is left out of the sample. 13,237 events (8.2% of the total sample) do not meet these criteria and are left out of the sample.
The dates of the recommendation as stated with the earnings announcements may be not random, but due to the news event that shocks in share prices will occur without an analyst recommendation. As a second control measure whether within an industry an announcement occurs on the same date, these recommendations will be left out of the sample. The sample shows that certain brokerage firms adjust their analyst recommendation on a particular date, for example on 01/06/2008 Merril Lynch adjusts for 54 of their covered bank stocks to either strong buy or underperform. It is most likely that a macroeconomic factor, industry event or policy change affected the recommendation levels of the banks.
Some brokerage firms adjust their recommendation for a particular stock on the same day. This usually happens if there is new information provided by the company or the case of a news event. Therefore as a third control measure whether a particular stock receives a recommendation of at least two different firms on the same date the recommendation will be left out the sample. For example HCA Holdings receives on 19/04/2011 from 19 different brokerage houses a strong buy or buy recommendation, which indicates that there is a news event that affects the outlook of all the brokerage houses. 93,755 events (63.1 % of the sample after the first control measure) after the second and third measure are implemented do not meet this criteria and are left out. A total 54,726 events remain.
Figure 1: Distribution of the recommendations and the total amount of recommendations per year
4. Methodology
The data sample is divided into four groups based on the recommendation level given by the broker. The abnormal return is the actual ex post return of the security over the event window minus the normal return of the firm over the event window (Mackinlay, 1997).
In the formulas used to explain the methodology different types 𝑡𝑡 are used to distinguish between estimation period and event period. The estimation period is the period before the event window to estimate the normal returns in absence of the event. The event window is the period surrounding the event where the returns will deviate from normal levels. The following figure provides the explanation of the symbols for time in the event study:
Figure 2: Time sequence of an event study
The estimation window used to determine the normal returns before is defined as the time 𝑇𝑇0 until 𝑇𝑇1. The event window is defined as 𝑇𝑇1until 𝑇𝑇2. τ is the event date. 𝑇𝑇1and 𝑇𝑇2can vary according the length of the event window that is chosen by the researcher.
4.1 Traditional event study methodology
For 𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑖𝑖 and event 𝑑𝑑𝑑𝑑𝑡𝑡𝑑𝑑𝑡𝑡 the abnormal return is:
𝐴𝐴𝐴𝐴𝑖𝑖,𝑡𝑡= 𝐴𝐴𝑖𝑖,𝑡𝑡− 𝐸𝐸(𝐴𝐴𝑖𝑖,𝑡𝑡 ∣ 𝑋𝑋𝑡𝑡) (1)
where 𝐴𝐴𝐴𝐴𝑖𝑖,𝑡𝑡, 𝐴𝐴𝑖𝑖,𝑡𝑡 and 𝐸𝐸� 𝐴𝐴𝑖𝑖,𝑡𝑡∣∣ 𝑋𝑋𝑡𝑡� are the abnormal, actual, and normal returns respectively for time period 𝑡𝑡 in the event window. 𝑋𝑋𝑡𝑡 represents the return estimated by the 4-factor model. For the market return the CRSP index is used that entails all value weighted returns form the stocks listed on the NYSE, AMEX and NASDAQ. The returns in this research are continuously compounded, this improves the normality of the return distribution.
The following 4-factor daily time-series regression model is used to obtain the parameters needed to estimate the normal returns over the estimation window:
Where:
𝐴𝐴𝑖𝑖,𝑡𝑡 = the return of 𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑖𝑖 at day 𝑡𝑡
𝐴𝐴𝑅𝑅𝑅𝑅𝑖𝑖,𝑡𝑡 = the abnormal return on the market at day 𝑡𝑡, defined as the return from the market minus the risk free rate. The risk free rate is the one-month treasury bill rate.
𝑆𝑆𝑅𝑅𝑆𝑆𝑖𝑖,𝑡𝑡= the average return on three small portfolios minus the average return on the three big portfolios entailing all the stocks from the all NYSE, AMEX, and NASDAQ. The 𝑆𝑆𝑅𝑅𝑆𝑆𝑖𝑖,𝑡𝑡 factor should capture the abnormal return of small over big stocks.
𝐻𝐻𝑅𝑅𝐻𝐻𝑖𝑖,𝑡𝑡= the average return on the two value portfolios minus the average return on the two growth portfolios entailing all the stocks from the all NYSE, AMEX, and NASDAQ. The 𝐻𝐻𝑅𝑅𝐻𝐻𝑖𝑖,𝑡𝑡 factor should capture the abnormal return of stock with a high market-to-book ratio over stocks with a low market-to-book ratio.
𝑊𝑊𝑅𝑅𝐻𝐻𝑖𝑖,𝑡𝑡= the average return on a past winner stocks portfolio minus the average return on the on a portfolio of pas losers stocks entailing all the stocks from the all NYSE, AMEX, and NASDAQ the difference between the month t returns of an equally-weighted portfolio of past stock market winners and one of past losers. The 𝑊𝑊𝑅𝑅𝐻𝐻𝑖𝑖,𝑡𝑡 factor should capture momentum in a stock. 𝜀𝜀𝑖𝑖,𝑡𝑡= the regression error term
In addition to providing an estimate of the daily abnormal return on each firm i, this regression yields the coefficient estimates 𝛼𝛼𝑖𝑖 𝛽𝛽𝑖𝑖 , 𝑠𝑠𝑖𝑖, ℎ𝑖𝑖 and 𝜔𝜔𝑖𝑖. These estimates provide insights into the nature of the firms in sample. A value of 𝛽𝛽𝑖𝑖 greater (less) than one means that the firms in given a certain recommendation are, on average, riskier (less risky) than the market. A value of 𝑠𝑠𝑖𝑖 greater (less) than zero given a certain recommendation tilted toward smaller (larger) firms. A value of ℎ𝑖𝑖 greater (less) than zero indicates a tilt toward stocks with a higher (lower) book-to-market ratio, conventionally thought of as value (growth) stocks. Finally, a value of 𝜔𝜔𝑖𝑖. greater (less) than zero signifies a portfolio comprised, on average, of stocks that have performed well (poorly) in the recent past. A positive coefficient means that the stock was a past winner and a negative coefficient means that the stock was a past loser.
The abnormal returns are summed into cumulative abnormal returns (CAR). A cumulative abnormal return is necessary to accommodate the multiple period event window. CAR(𝑇𝑇1, 𝑇𝑇2) is defined as the sample CAR from τ where 𝑇𝑇1< τ ≤ 𝑇𝑇2. The CAR from (𝑇𝑇1, 𝑇𝑇2) is the sum of the included abnormal returns. The estimation periods in amount of trading days around the event are: (-1,1),(-2,2) ,(-3,3) ,(-5,5) ,(-10,10) and (-20,20). The estimation window is set on (-200,-21).
In order to calculate the cumulative abnormal returns for the event window the daily abnormal returns are summed to (𝑇𝑇1, 𝑇𝑇2):
𝐶𝐶𝐴𝐴𝐴𝐴i,(𝑇𝑇1,𝑇𝑇2) = ∑ 𝐴𝐴𝐴𝐴𝑖𝑖,𝑡𝑡 𝑇𝑇2
𝑡𝑡=𝑇𝑇1 (3)
The event windows and estimation periods are set according to the periods of MacKinlay (1997). The estimation periods should be as big as possible and should be clean of any other relevant events. With regards to analysts recommendation, it is inevitable that other recommendations are given in the estimation period. Therefore as extra measure the forecast error is taken into account to better be able to determine the normal returns. The event windows in this study are clean of other recommendations recorded in the IBES database. Therefore the dataset is adjusted accordingly with the control measures.
Hereafter the CAR’s are averaged to determine the cumulative average abnormal returns. The cumulative average abnormal returns give the mean of the effect of the various recommendation levels. The cumulative abnormal return is calculated as follows:
𝐶𝐶𝐴𝐴𝐴𝐴������(𝑇𝑇1,𝑇𝑇2) = 1
𝑁𝑁∑ 𝐶𝐶𝐴𝐴𝐴𝐴𝑖𝑖,(𝑇𝑇1,𝑇𝑇2) 𝑡𝑡2
𝑖𝑖=1 (4)
For determining significance under traditional event study method testing a time-series T-test is created where 𝑇𝑇𝑡𝑡𝑖𝑖𝑡𝑡𝑡𝑡 is defined as:
𝑇𝑇𝑡𝑡𝑖𝑖𝑡𝑡𝑡𝑡 = 𝐶𝐶𝐶𝐶𝐶𝐶(𝑇𝑇1,𝑇𝑇2) ��������������� (𝑡𝑡2−𝑡𝑡1+1)
1
2 ∗ 𝑆𝑆𝑆𝑆𝐴𝐴𝐴𝐴𝐴𝐴𝑡𝑡 (5)
𝑆𝑆𝑆𝑆𝐶𝐶𝐶𝐶𝐶𝐶𝑡𝑡 is the standard deviation over the average abnormal return at time t and is defined as
follows: 𝑆𝑆𝑆𝑆𝐶𝐶𝐶𝐶𝐶𝐶𝑡𝑡= �𝑣𝑣𝑑𝑑𝑓𝑓𝐶𝐶𝐶𝐶𝐶𝐶𝑡𝑡 (6.1) 𝑣𝑣𝑑𝑑𝑓𝑓𝐶𝐶𝐶𝐶𝐶𝐶𝑡𝑡= 1 𝑀𝑀−2∑ [𝐴𝐴𝐴𝐴𝐴𝐴𝑡𝑡− 1 𝑀𝑀 𝑇𝑇1 𝑡𝑡=𝑇𝑇1+1 ∑ (𝐴𝐴𝐴𝐴𝐴𝐴𝑡𝑡 𝑇𝑇1 𝑡𝑡=𝑇𝑇1+1 )]2 (6.2)
Here 𝐴𝐴𝐴𝐴𝐴𝐴𝑡𝑡 is defined as the average abnormal return at time 𝑡𝑡. M is defined as the number of non-missing returns.
The event window returns are also predicted from the estimation window returns. As earlier stated, the probability of a recommendation change is not stable over time. This imposes a forecast error because the event window returns are predicted from values that are out of sample. Patel (1976) adjusts the standard error by the forecast-error to counter this problem. The forecast error
adjustment also takes into account the amount of non-missing returns in the estimation window. The standard deviation in the event window is corrected as following:
𝑆𝑆𝑆𝑆𝐶𝐶𝐶𝐶𝐶𝐶𝑡𝑡= 𝑆𝑆𝑆𝑆�𝐶𝐶𝐶𝐶𝐶𝐶𝐸𝐸𝐸𝐸𝑇𝑇,𝑡𝑡∗ �1 + 1 𝑀𝑀+
�𝐶𝐶𝑚𝑚,(𝑇𝑇1,𝑇𝑇2)−𝐶𝐶�𝑚𝑚,𝐸𝐸𝐸𝐸𝑇𝑇�2
∑𝐸𝐸𝐸𝐸𝑇𝑇𝑡𝑡=𝑇𝑇1�𝐶𝐶𝑚𝑚,𝑡𝑡−𝐶𝐶�𝑚𝑚,𝐸𝐸𝐸𝐸𝑇𝑇�2 (7)
Here 𝑆𝑆𝑆𝑆�𝐶𝐶𝐶𝐶𝐶𝐶𝐸𝐸𝐸𝐸𝑇𝑇,𝑡𝑡 is defined as the standard deviation during the estimation window at time 𝑡𝑡. M is defined as the number of non-missing returns. 𝐴𝐴𝑡𝑡 is the return and 𝐴𝐴�𝑡𝑡,𝐸𝐸𝑆𝑆𝑇𝑇 is the average market return during the estimation period.
4.2 Event study methodology following Boehmer, Musumeci and Poulsen (1991)
Boehmer, Musumeci and Poulsen (1991) developed a test that incorporates variance information from both the estimation and the event periods. They propose a small adjustment to the cross sectional T-test. They standardize the cumulative abnormal returns (CSAR) to make them robust against event-induced variance increases in stock returns.
The CSAR is calculated as follows:
𝐶𝐶𝑆𝑆𝐴𝐴𝐴𝐴𝑖𝑖,(𝑇𝑇1,𝑇𝑇2)= 𝐶𝐶𝐶𝐶𝐶𝐶𝑖𝑖,(𝑇𝑇1,𝑇𝑇2) 𝑆𝑆𝑆𝑆𝐸𝐸𝐸𝐸𝑇𝑇,𝐶𝐶𝐴𝐴𝐴𝐴𝑖𝑖 ∗�1+𝑀𝑀1+�𝐴𝐴𝑚𝑚,(𝑇𝑇1,𝑇𝑇2)−𝐴𝐴�𝑚𝑚,𝐸𝐸𝐸𝐸𝑇𝑇� 2 ∑𝐸𝐸𝐸𝐸𝑇𝑇𝑡𝑡=𝑇𝑇1�𝐴𝐴𝑚𝑚,𝑡𝑡−𝐴𝐴�𝑚𝑚,𝐸𝐸𝐸𝐸𝑇𝑇�2 (8)
Here the denominator consists of the standard deviation of the estimation period from the CAR adjusted for the forecast error. The cross-sectional average of 𝐶𝐶𝑆𝑆𝐴𝐴𝐴𝐴𝑖𝑖,(𝑇𝑇1,𝑇𝑇2) is calculated as follows:
𝐶𝐶𝑆𝑆𝐴𝐴𝐴𝐴�������(𝑇𝑇1,𝑇𝑇2)= 1
𝑛𝑛∑𝑁𝑁𝑖𝑖=1𝐶𝐶𝑆𝑆𝐴𝐴𝐴𝐴𝑖𝑖(𝑇𝑇1,𝑇𝑇2) (9)
The standard deviation of 𝐶𝐶𝑆𝑆𝐴𝐴𝐴𝐴�������(𝑇𝑇1,𝑇𝑇2) is estimated from the cross section of event-window abnormal returns:
𝑆𝑆𝑆𝑆(𝐶𝐶𝑆𝑆𝐴𝐴𝐴𝐴�������) = �𝑁𝑁(𝑁𝑁−1)1 (∑𝑖𝑖=1𝑁𝑁 𝐶𝐶𝑆𝑆𝐴𝐴𝐴𝐴𝑖𝑖,(𝑇𝑇1,𝑇𝑇2)− 𝐶𝐶𝑆𝑆𝐴𝐴𝐴𝐴�������(𝑇𝑇1,𝑇𝑇2))2 (10) The standardized cross-sectional test according to the BMP method statistic is defined as follows:
𝑇𝑇𝐵𝐵𝐵𝐵𝑡𝑡ℎ𝑡𝑡𝑡𝑡𝑚𝑚 =𝐶𝐶𝑆𝑆𝐶𝐶𝐶𝐶��������𝑆𝑆𝑆𝑆 (𝑇𝑇1,𝑇𝑇2)
(𝐶𝐶𝐸𝐸𝐴𝐴𝐴𝐴��������) (11)
4.3 Event study methodology following Kolari & & Pynnönen (2010)
Kolari & & Pynnönen (2010) developed a test where an average level of cross-correlation between the abnormal returns is taken into account. The methodology builds forward on the test statistic from BMP. The adjusted standardized cross-sectional test statistic as described by Kolari & Pynnönen (2010) for the null hypothesis that the cumulative average abnormal return is equal to zero is:
𝑇𝑇𝐵𝐵𝐵𝐵𝑡𝑡ℎ𝑡𝑡𝑡𝑡𝑚𝑚 𝑎𝑎𝑎𝑎𝑎𝑎.= 𝑇𝑇𝐵𝐵𝐵𝐵𝑡𝑡ℎ𝑡𝑡𝑡𝑡𝑚𝑚�1+(𝑁𝑁−1)𝜌𝜌�1−𝜌𝜌� (12)
𝜌𝜌̅ is an average level of cross-correlation that is manually chosen. The level of cross-correlation can vary from roughly 0.01 to 0.99. A negative cross-correlation cannot be taken into account. N is the amount of events in the sample.
5. Results
5.1 Results from the traditional event study
Table 1 shows the results of the regular event study adjusted for the 4-factor model. Figure 3 gives a representation of the cumulative average abnormal returns for the different levels of recommendation. The results are in line with the previous literature and are on all levels significant at the 1% level using the time-series T-statistic to estimate if the CAARs significant differ from zero. Also in line with the previous literature is that the sell recommendation has the biggest effect on the stock price, this attributable due to the fact that sell recommendations are rarely issued. Even with the hard data restrictions I use in this research the impact of an analyst recommendation on a firm’s stock is of significant value.
The regression coefficients using the 4-factor model provides me with the characteristics of the stocks that have a certain recommendation level. Table 2 shows the average coefficients and their significance level. The average betas of all the recommendation levels of the stocks are close to one, we do see that the sell recommendation level has the highest level of risk and the strong buy the lowest level of risk indicated from the beta. The positive coefficients in the SMB column for all the recommendation levels indicate that they are tilted towards small stocks. The positive coefficients in the HML column indicate that the analysts have a tendency to cover value stocks. In the WML column, the positive coefficient for the positive recommendation shows that these stocks are past winners. The stocks with a negative recommendation are previous losers. The findings are consistent with earlier research (Barber et al, 2004). The momentum coefficients significantly differ from zero, this means that momentum is present in the stocks.
Figure 3: Cumulative average abnormal returns for the different levels of recommendation. The Y-axis indicates the cumulative average abnormal returns over the event window of -20 to 20 trading days. 0 is the recommendation day.
Table 1
Cumulative average abnormal returns (CAAR) and their standard deviations (SD) following the Brown & Warner (1985) methodology to test for significance
The event window is the amount of trading days before and after the event date. N is the amount of recommendations per recommendation level. SBUY is a strong buy recommendation and UPRF is a underperform recommendation . To test for significance a time series T-test is used (formula 5). ***, ** and * indicates significance at α= 0.01, α= 0.05 & α=0.1 levels respectively (2-sided).
Table 2
Coefficient average estimates of the actual returns following the 4-factor model for determining returns
The mean is the mean of the coefficients from the stock returns The * denotes that the coefficient significantly deviates from zero using a one-sample T-test at the 1% level (two-sided). The % unequal to 0 gives the percentage of coefficients that truly deviate from zero. The medians do not differ strongly from the means. The intercepts are not tested for significance, this beyond the scope of this research. N is the amount of events per recommendation level.
SBUY N= 20162 BUY N= 25055 UPRF N= 6523 SELL N= 6483
Event window CAAR SD CAAR SD CAAR SD CAAR SD
(-20,20) 0.012 0.045 *** 0.013 0.042 *** -0.019 0.002 *** -0.036 0.045 *** (-10,10) 0.016 0.002 *** 0.013 0.035 *** -0.026 0.002 *** -0.040 0.038 *** (-5,5) 0.017 0.001 *** 0.014 0.030 *** -0.028 0.001 *** -0.042 0.032 *** (-3,3) 0.018 0.001 *** 0.013 0.027 *** -0.029 0.001 *** -0.042 0.029 *** (-2,2) 0.018 0.001 *** 0.013 0.025 *** -0.030 0.001 *** -0.042 0.027 *** (-1,1) 0.017 0.001 *** 0.013 0.022 *** -0.029 0.001 *** -0.039 0.023 *** (0,0) 0.013 0.000 *** 0.009 0.017 *** -0.020 0.000 *** -0.030 0.018 *** N
Strong buy (mean) 0.000 0.813 0.458 0.105 0.020 20162
5.2 Results from the event study using the Boehmer, Musumeci and Poulsen (BMP) method to test for significance
Table 3
Cumulative average abnormal returns (CAAR) and their standard deviations (SD) following the BMP (1991) methodology to test for significance
CSAR is the Cumulative standardized average abnormal returns and SD is the corresponding standard deviation. The event window is the amount of trading days before and after the event date. N is the amount of recommendations per recommendation level. SBUY is a strong buy recommendation and UPRF is a underperform recommendation . To test for significance a BMP T-test is used(formula 11). ***, ** and * indicates significance at α= 0.01, α= 0.05 & α=0.1 level respectively(2-sided).
Table 3 shows the results of the event study adjusted for the 4 factor model using the Boehmer, Musumeci and Poulsen (BMP) method. The CAARs and standard errors are now robust to event-induced variance. I find significant results for all event windows for a strong buy and underperform recommendation. A buy recommendation remains of relevant impact for only up to three days before and after the recommendation. An interesting result is that the strong sell recommendation has no significant impact in CSARs only for the longest event window of 20 days before and after the event. An explanation can be possibly found in the classic finance theory. Given the fact that sell recommendation are rarely given because of the value destruction in the relationship between the brokerage firm and the covered firm one can assume that the company’s current state can be already classified as highly volatile and risky before the recommendation is issued. The movement of the equity value of a firm with a sell recommendation is bounded by zero. Therefore, limiting the downward potential, but the variance of the firm has no such limitation (Sharpe, 1964). The systematic risk is no longer on the capital market line for the stock, and regardless of the recommendation the stock was already not relevant to the investor, because the stock in not an efficient investment. Other stocks are of preference for an investor for making an efficient
SBUY N= 20162 BUY N= 25055 UPRF N= 6523 SELL N= 6483
Event window CSAR SD CSAR SD CSAR SD CSAR SD
(-20,20) 1.647 0.552 *** 1.980 2.878 -1.518 0.620 ** -2.866 1.657 * (-10,10) 2.232 0.906 ** 2.113 78.266 -2.093 0.867 ** -3.221 2.163 (-5,5) 2.457 0.663 *** 2.184 1.396 -2.254 0.922 ** -3.390 2.768 (-3,3) 2.531 0.674 *** 2.102 0.951 ** -2.367 0.967 ** -3.398 3.009 (-2,2) 2.548 0.587 *** 2.079 0.962 ** -2.407 0.983 ** -3.406 3.406 (-1,1) 2.433 0.622 *** 1.983 0.674 *** -2.311 0.944 ** -3.164 3.164 (0,0) 1.825 0.331 *** 1.474 0.607 ** -1.626 0.664 ** -2.424 2.424
Standardized cumulative abnormal returns and their standard deviations
investment portfolio with the best utility. Meaning that the sell recommendation does not provide an added value to the investor because he is already aware of the situation without the need of a sell recommendation.
The information about the high-perceived risk that is not optimal is already incorporated in the stock price. The news what caused the shift to a sell recommendation could be already incorporated in the stock. This also supports the theory of Klein(1990) that analysts are reluctant to issue a sell recommendation and wait until further evidence of a bad forecast is brought to the market. Lui, Markov & Tamayo (2007) also support this explanation. They find that risk explains almost 50% of the incremental price jump followed by a recommendation change trough a shift in idiosyncratic risk, leverage, book to market, accounting losses and earnings quality. These kind of shifts in levels could easily be spotted and interpreted by news sources other than analysts.
A behavioral explanation is found the herding theory, this kind of behavior of analysts is proven by Loh & Stulz (2010). If a sell recommendation is given and has great influence on the firm’s share price, it is most likely that other analysts will follow. Herding, which is the tendency for individuals to mimic the actions (rational or irrational) of a larger group. Individually, however, most people would not necessarily make the same choice. Most likely if a notable analyst issues a sell recommendation others follow, however the biggest price drift is noted in the first recommendation. Desai, Liang and Singh(2000) find that stocks with one or more all-star analyst recommendation outperformed the recommendations but if multiple analysts revised on the event day the performance was less strong. They find that lower publication-day reaction to stocks recommended by several all-star analyst may reflect that the information is already incorporated in the firms’ stock price. This means that the first sell recommendation that moves away from the consensus can exhibit a significant effect. The other that follow to the sell recommendation do not have a tremendous impact on the share price. Therefore there is no abnormal return measured for these event dates.
A possible explanation found in the incentive perspective comes from Abarbanell & Lehavy (2002). They study the effects between earnings management and analyst recommendations. They find that in the case of a large negative recommendation the company adjust theirs earnings management (create a large cash reserve) to decrease the sensitivity in the ex-ante stock price. Firms with a large positive recommendation exhibit the reversed effect. They also support the research of Klein (1990), where analysts have an incentive to wait with adjusting their forecasts to bad news, to prevent the weakening of relationships between the firm and the brokerage.
5.3 Results from the event study using the Kolari & & Pynnönen method for levels of average cross-correlation
Table 4
Different P-values for the Kolari & Pynnönen (2010) method for average correlation for a strong buy recommendation
The initial average correlation level is 0, which gives the same level of significance as the values in table 3. The event window is the amount of trading days before and after the event date. The other values correspond to the average cross-correlation between the stocks. The T-statistic is determined using formula 12 and with this statistics the 2-tailed p-values are derived.
Tables 4 shows the various p-levels for the recommendations when there is a certain level of cross-correlation between the abnormal returns. The results show immediately that even with the slightest level of cross-correlation between the abnormal returns all the results will yield to insignificant test results and the CAARs do not statistically differ from zero. Bernard (1987) finds average cross-correlations of 0.04 between industries in 1984 in daily stock returns for 1080 firms. Thus, although the 4-factor model captures a large share of the zero return contemporaneous correlation, the remaining relatively small correlation still materially biases the significance levels with even moderate sample sizes. The results of this paper appear to yield that the effect of analyst recommendations on the stocks is heavily overestimated due to the fact that ignoring even small (average) correlation may substantially bias the distributional properties of the test statistics via underestimation of the true (abnormal) return variability.
SBUY Event window 0 0.01 0.05 0.1 0.15 0.2 (-20,20) 0.003 0.835 0.927 0.950 0.960 0.966 (-10,10) 0.014 0.863 0.940 0.959 0.967 0.972 (-5,5) 0.000 0.796 0.910 0.938 0.950 0.958 (-3,3) 0.000 0.793 0.908 0.937 0.950 0.958 (-2,2) 0.000 0.762 0.894 0.927 0.942 0.951 (-1,1) 0.000 0.784 0.904 0.934 0.948 0.956 (0,0) 0.000 0.700 0.866 0.907 0.926 0.938
Average correlation BUY
Event window 0 0.01 0.05 0.1 0.15 0.2 (-20,20) 0.491 0.966 0.985 0.990 0.992 0.993 (-10,10) 0.978 0.999 0.999 1.000 1.000 1.000 (-5,5) 0.118 0.922 0.966 0.976 0.981 0.984 (-3,3) 0.027 0.890 0.951 0.967 0.973 0.978 (-2,2) 0.031 0.892 0.953 0.967 0.974 0.978 (-1,1) 0.003 0.853 0.935 0.956 0.965 0.970 (0,0) 0.015 0.879 0.947 0.963 0.971 0.976 Average correlation UPRF Event window 0 0.01 0.05 0.1 0.15 0.2 (-20,20) 0.014 0.765 0.895 0.928 0.942 0.952 (-10,10) 0.016 0.768 0.896 0.929 0.943 0.952 (-5,5) 0.015 0.765 0.895 0.928 0.943 0.952 (-3,3) 0.014 0.765 0.895 0.928 0.942 0.952 (-2,2) 0.014 0.765 0.895 0.928 0.942 0.952 (-1,1) 0.014 0.765 0.895 0.928 0.942 0.952 (0,0) 0.014 0.765 0.895 0.928 0.942 0.952
Average correlation SELL
5.4 Results for the top tier full service investment bank recommendations
In the current research setting the analyst recommendations are stated to have the same character and are all of equal importance. Previous research (Loh & Stulz, 2010) points out the link of the power of a certain recommendation connected to the status of the analyst and its corresponding firm.
Another explanation for the insignificant results if the CARs are standardized is the fact that not all recommendations create the same amount of usable information and investors do not respond to these recommendations. The characteristics of the analyst and the firm they work is of importance in the value of the recommendation. Smaller firm analysts tend to make more bolder statements for a company’s performance outlook in order to create more buzz for their reports, star analysts from the big investment firms create more influential recommendations that will affect the stock price of the firms (Loh & Stulz, 2010). Also the experience of an analyst is an important factor to take into account, the more experience an analyst has the greater the influence on absolute value of analyst due to a better accuracy in forecasts (Mikhail, Walther & Willis, 1997). However this effect cannot be measured in my research setup due to unavailability of the required data. Also analysts have the tendency to herd, therefore a recommendation that is away from the consensus leads to a bigger price shock according to the direction of the recommendation. Jegadeesh & Kim (2004) find that analysts from larger brokerages, analysts following stocks with smaller dispersion across recommendations, and analysts who make less frequent revisions are more likely to herd. Furthermore the belief in hot hand comes in to play. When an analyst makes an influential recommendation, which is accurate, the market will more likely follow the next recommendation made by the analyst (Loh & Stulz, 2010).
I decide to run the event study again but now only using the recommendations from the top 3 most influential or biggest sell-side analyst firms with the same event date restrictions mentioned in the sample selection paragraph. Initial sample consisted of the following banks: Citibank, Nomura, J.P. Morgan, Merril Lynch, Goldman Sachs and Lazard. After the event date restrictions the recommendations of Nomura, Merril Lynch and Lazard remain. The analysts of the top 3 firms are most likely to have perceived experience, have a star analyst (or the firm) status, are less likely to be prone of herding behavior and therefore have bigger value to the market. Table 5 gives the outcome of the CSARs and their errors. The results deviate from the previous sample in the fact that for (nearly) all event windows the recommendations are significant except the strong buy recommendations.
Table 5
Cumulative standardized average abnormal returns (CSAR) and their standard deviations (SD) for the for the top 3 analyst sell-side firms (select sample) following the BMP (1991) to test for significance
The event window is the amount of trading days before and after the event date. N is the amount of recommendations per recommendation level. SBUY is a strong buy recommendation and UPRF is a underperform recommendation . To test for significance a BMP T-test is used(formula 11). ***, ** and * indicates significance at α= 0.01, α= 0.05 & α=0.1 level.
The results can be explained using the discount theory of Malmendier & Shanthikumar (2004). They find that large (institutional) traders account for the upward bias in the recommendations. Therefore they discount the value of a buy and strong buy recommendation. A fully rational investor should discount the recommendations even more if the analysts are affiliated with the underwriter of an issuer. The banks in the sample are full service investment banks and one can assume that they have close relationship with the covered firms through the capital market offerings (IPO’s, SPO’s, debt notes), other investment bank activities (M&A) or traditional relationship banking. Strong buy recommendations from analyst underwriter firms are often given when a firm goes public. An IPO does not solely need an optimistic analyst who is positive about the firms outlook. The underwriter needs to create stock liquidity and buyers for the stock. The reputation of a bank, the relation it has with large institutional investors and leading industry analysts are the keys to acquiring new underwriting business. If an analyst writes a recommendation on a stock where the firm is also underwriter the recommendation is heavy positively biased, because of the pressure into making sure the IPO will be a success and profitable for the firm and underwriter. Not a single analyst underwriter will place weak forecast outlook and recommendation of a firm that it guides into going public. It is held that, in such cases, a severe conflict of interest exists, and the market may thus be inclined to discount recommendations from affiliated analysts (Bradley, Jordan & Ritter, 2006).
28 SBUY N= 2112 BUY N= 3259 UPRF N= 1851 SELL N= 190
Event window CSAR SD CSAR SD CSAR SD CSAR SD
(-20,20) 0.841 0.595 1.718 0.859 ** -0.805 1.024 -0.288 0.143 ** (-10,10) 0.744 0.527 1.370 0.711 * -1.187 0.900 *** -0.401 0.156 ** (-5,5) 0.850 0.601 1.284 0.669 * -1.175 0.337 *** -0.624 0.162 *** (-3,3) 0.836 0.591 1.205 0.630 * -1.218 0.195 *** -0.675 0.143 *** (-2,2) 0.786 0.556 1.142 0.600 * -1.230 0.514 *** -0.656 0.120 *** (-1,1) 0.726 0.513 0.988 0.525 * -1.265 0.456 *** -0.583 0.099 *** (0,0) 0.542 0.383 0.725 0.419 * -0.925 0.075 *** -0.369 0.059 ***
During IPO’s the latter of the equity for sale is reserved for the large institutional investors, so almost all-trading activity will come from the large institutional investors. This explains why the buy recommendation is significant under the BMP method, because the trading activity would be more dispersed between different types of investors. McNichols, O’Brien & Pamukcu (2006) confirm that investors tend to discount the recommendations of affiliated analysts, furthermore they find that affiliated recommendations do not discriminate between good and bad IPO stocks, but unaffiliated recommendations generally arrive too late to provide useful trading advice.
Reversely this could be an explanation why the sell recommendation stays significant if a bulge bracket full service investment bank issues it. A sell recommendation is negative for the relationship between the covered firm and the investment bank; it could be value destructing for the other services other than brokerage. In case the analyst issues it, it must be of great importance to the brokerage clients or institutional investors if an investment bank is willing to sacrifice the relationship.
Table 6
Different P-values for the Kolari & & Pynnönen (2010) method for average correlation for a the different levels of recommendation.
SBUY is a strong buy recommendation, UPRF is a underperform recommendation. The initial average correlation level is 0, which gives the same level of significance as the values in table 5. The event window is the amount of trading days before and after the event date. The other values correspond to the average cross-correlation between the stocks. The T-statistic is determined using formula 9 and with this statistics the 2-tailed p-values are derived.
