Analysts’ stock recommendations’ value: a comparison of the US and European markets

Размещено на http://www.allbest.ru/

The Government of the Russian Federation

Federal State Autonomous Institution for Higher Professional Education National Research University Higher School of Economics

St. Petersburg Branch

St. Petersburg School of Economics and Management

Master's dissertation

Analysts' stock recommendations' value: a comparison of the us and european markets

Ugolnikov Vladislav Alekseevich

Saint Petersburg 2019



Abstract

Do the sell-side analysts' recommendations worth trusting and bring value to the investors? This topical question was discussed in many previous papers but most of them were related to a single market instead of a comparative analysis of two large and developed stock exchanges. In this master thesis, such recommendations are analyzed across the LSE and the NYSE and considering a period between 1 January 2016 and 31 April 2019. The main objective of the paper is to investigate the analysts' recommendations' value, and to determine on which market the analysts have more predictive power that can be defined as an extent of a stock price's reaction around a particular recommendation what leads to abnormal returns of the security. Thus, with the usage of an event study analysis, it was found that both American and British markets are feasible to be outperformed by the analysts but it is impossible to highlight any of the markets as they behave almost identically around positive, neutral and negative recommendations. Considering the level of a preceding recommendation was also emphasized in this paper, and it does not strikingly change the picture. From a linear regression model, it was found that investors react stronger to negative and neutral recommendations for larger companies rather than smaller ones on the both stock exchanges.

Key words: analysts' recommendations, recommendations' value, stock market reaction, LSE, NYSE.



Table of Contents

Introduction

1. Theoretical background

2. The US and the UK markets' summary

3. Hypotheses development

3.1 Reaction to pure recommendations

3.2 Reaction to changes in previously published recommendation levels

4. Data and methodology

4.1 Data and sample description

4.2 General event study's methodology description

4.3 Testing the first hypothesis

4.4 Testing the second hypothesis

4.5 Determining other factors of the markets' reaction

5. Results and discussion

5.1 Results from testing the first hypothesis

5.2 Results from testing the second hypothesis

5.3 Results of determining other factors affecting the markets' reaction

Conclusion

Annotated bibliography

Appendix



Introduction

Global expenses on stock markets' analytics exceed millions of dollars annually, and the amount is still growing because of more and more investors desire to obtain information that allows them to get surplus returns on securities' investments. In a perfectly efficient market, such costs would have no sense as the quotes already reflected all available information at that moment. However, if the market is not completely efficient, the majority are likely to spend their time and money to obtain knowledge that might help outperform the market. If to compare with investors, a part of which are spontaneous ones, sell-side analysts who offer their recommendations to clients are supposed to possess more resources such as inside news and greater financial expertise to conduct a fundamental analysis of business environment of a company and, in particular, its stocks. In addition, many brokerage houses obtain stocks in their investment portfolios that often become an object of future recommendations. Therefore, analysts are more likely to have unique and privileged information that is expected to shift stock prices when a report is published. Usually, these reports contain different kinds of information, but one of the key points is an opinion about future price changes' direction. In fact, this is a strong suggestion to investors what to do with a particular stock: to sell, to hold, or to buy if this stock is not included in an investor's portfolio yet. Therefore, an ability of more accurate identification of undervalued (`buy' recommendations) and overvalued (`sell' recommendations) stocks is one of essential resources of possible analysts' outperformance.

However, an accuracy of analysts' reports and a subsequent market reaction are rather complicated and depend on what factors are under consideration in a particular research. Different methodology and used data might lead to controversial results as well. Usually, authors who dedicate their papers to developed markets choose several factors and try to identify their impact [Keasler, McNeil, 2010; Pogozheva, 2013; Ishigami, Takeda, 2018].

Nevertheless, to the author's best knowledge, few frequently cited and well-conducted research is focused on such an issue as a comparison of analysts from the point of their recommendations' value across two developed stock exchanges, for example, Rounaghi and Zadeh (2016). This is an essential issue because suggestions about whether sell-side analysts' recommendations bring value to their clients depend on various factors: from analytical tools to collected data. We have to accept the risk that an evidence of the recommendations' usefulness (or uselessness) was obtained due to the methodology's imperfection. However, when two or more markets are investigated in terms of the same methodology, similar data and other common assumptions simultaneously, the results are likely to be more representative and reliable. Thus, this study includes both the point of a separate analysis of the markets and their comparison. Since then, this paper is expected to shorten the gap in scientific literature, and to bring some novelty to the issue of the markets' comparison from the point of analysts' stock recommendations' value, and provide with additional evidence related to the research question.

This paper is dedicated to a comparison of the United States and the European markets. The markets are worth discussing as they, generally, can be characterized as those that have huge trading volumes, developed monetary and banking infrastructure, and widely spread various derivative tools. The main objective of the paper is to investigate the analysts' recommendations' value related to the US and the European stock markets, and to determine on which market the analysts have more predictive power. The term `predictive power' mostly relate to whether a stock price moves in line with a particular recommendation what leads to abnormal returns of the security. Thus, the greater these abnormal returns caused by the recommendation, the greater the recommendation's value and the analysts' predictive power.

The suggestion whether American and European analysts that are expected to possess a unique information are able to precisely define overrated and underrated securities and to give valuable recommendations is going to be an object of thorough analysis in the following paper. Although testing the efficient market hypotheses is not a direct goal of this research, the obtained results might be applicable to a discussion of this side of the question. Besides other possible practical implications, this study is aimed to provide investors with a piece of support whether such kinds of analysts are worth trusting.

The body of this paper can be divided into two parts. The first part includes theoretical base of the research and review of the former literature. The second part consists of hypotheses development, sample and methodology description and discussion of the results.

As the sources of information, the international periodical and citation databases such as EBSCO, Springer, SCOPUS, and Passport GMID were used. All financial data and historical analysts' recommendations were gathered from Thomson Reuters Eikon.



1. Theoretical background

The two most essential indicators that investors take into consideration while choosing a stock market to trade on are prices and volumes. These figures together usually represent the trading activity which shows how the investors are interested in earning money through managing their investment portfolio. In general, a trading volume can be defined as how many securities change hands and are traded over a particular period and in an exact place. However, it seems more reasonable to observe this information in connection with stock prices as a number of shares traded is not representative itself. A combination of these figures is called `dollar trading volume', and is calculated as a stock number that were traded multiplied by its money value. The greater the dollar traded volume of the market, the more liquid it tends to be. Malkiel (1992) indirectly describes a relationship between the number of a market's participants, liquidity and the market's level of efficiency. He also outlines three statements of an efficient market, but the most applicable to this paper is that the stock market is efficient if securities' prices do not change when information becomes available to all participants. According to Fama (1991), cited by Okulov (2010), depending on what type of information is implied there, the three forms of the market efficiency can be determined. Firstly, the weak form relates to historical prices of securities. Secondly, the semi-strong form is devoted to publicly available information including the historical quotes. Thirdly, the strong form includes the previous two types of information and considers private and privileged information.

Higher liquidity makes it easier to trade and frequently shows the threshold level of entering the market for ordinary investors. It seems evident that the lower this threshold, the more investors participate, and this is what makes the market more transparent and somehow allocates the valuable information between the traders. Therefore, we can assume the securities markets with larger dollar trading volumes are expected to be more developed. In addition to this, the gap between the investors' awareness of fundamental information related to the stocks, main tendencies and coming business environmental changes becomes shorter. This logic is consistent with the findings received by Murg and Zeitlberger (2014) who came to a suggestion that the relation between the market's size and the strength of its reaction to an information's diffusion is reverse.

On balance, in the conditions of higher dollar trading volumes and the market's liquidity, the ability of the participants to extract abnormal returns and to outperform the other players declines. This assumption mostly relates to large and well-known stock exchanges, such as the New York Stock Exchange and the London Stock Exchange, as well. Looking ahead, the LSE is the largest stock exchange in Europe and has a lot in common with the American markets what is going to be discussed in the next section. That is why this stock exchange is chosen to be compared with the largest American one.

The mentioned goal is formulated basing on the existing literature on related topics. Most of them are dedicated to the questions whether analysts possess privileged knowledge, and how different markets react to dissemination of such information. They investigate the strength, direction and duration of the investors' response, as well.

Although in previous studies there are several methodology tools used, there is an existing agreement that the application of event study analysis in absence of serious data clustering and release of relevant recommendations allows to reach reliable results in short-term periods. However, as for long-term effects of more than several months, the findings are controversial [Kraemer, 2016]. Event study tools provides with an ability to capture abnormal returns of securities and abnormal trading volumes in connection with considered events, and to determine their significance level. Traditional method of event study was thoroughly discussed in the paper of Khotari and Warner (2006) and consists of the following procedures:

1. Choosing a model that describes daily return of a security in a period without significant events and determining the length of an observation window.

2. Determining the length of an event window in terms of the research objectives and calculating abnormal returns based on estimated parameters in the model from observation window.

3. Accumulating abnormal returns across the whole event window in order to extract the effect from random fluctuations of returns.

4. Standardizing the calculated cumulative abnormal returns for a group of events across the whole sample.

5. Testing the hypothesis of a difference between these standardized abnormal returns from zero and determining the level of significance.

Authors of studies of investment value of analyst recommendations and the market reaction to these publications (Womack, 1996; Keasler, McNeil, 2010; Murg, et al, 2014) acknowledged the essence of the event study analysis that allows tracking the reactions of the stock market to certain events and news arising from the activities of market participants. As it was mentioned earlier, the core characteristic of the investors' response is abnormal returns and cumulative abnormal returns that occur around the event. In those studies, the core events that are to be reacted to are analysts' recommendations' releases. In other words, the moment when privileged information becomes publicly available and spreads among potential and actual investors is taken into consideration.

The whole range of research of market reactions to analyst recommendations can be divided according to the properties of methodology of the event study analysis they use. The properties are closely connected with the aforementioned steps, and correspond to such peculiarities as lengths of observation and event windows, approach to calculating normal stock returns, and further investigation of factors that affect them.

A sufficient factor influencing the reached results is the length of estimation and event windows. During the estimation window, usual behavior of stock prices and, consequently, returns are estimated; whereas in the event window a reaction itself is analyzed. Traditionally, the usage of longer event windows (up to several years) makes sense only in cases of significant and rare releases of internal information, when the effect is expected to be noticeable for a long period reflecting its complexity. The most outstanding examples of such events are mergers and acquisitions initiation, or decisions about dividend policies [Khotari, Warner, 2006]. Generally, such events as stock recommendations' publications appear enough regular and are a normal practice in the life of stock exchanges and other securities markets. Therefore, the mostly spread length of the event window that appears in related previous studies is up to five or ten days before and after the issuance date. In the pioneer paper of this kind delivered by Davies and Canes (1978) an influence of information published in different analytical journals was investigated. They found some evidence that `buy' recommendations cause the stock prices' growth in a publication date, and vice versa. However, the idea of more durable market reaction became a justification for the usage of longer event windows. In the papers of Beneish (1991), Palmon, et al (1994), and Womack (1996), the authors desired to capture subsequent prices drifts and possible fluctuations. They analyzed event periods of different length from one to six months, and found that after a significant stock prices' jump on the event date they inevitably come back to their pre-event levels. “In the edge of the century, due to the Internet emergence, valuable information became more accessible for usual investors, and new impulse was given to the research of a market's information efficiency” [Okulov, 2010, p.8]. These hypotheses were successfully tested in further studies such as Keasler and McNeil (2010), Murg, et al (2014) and Suliga (2016), where the results appeared to be similar to those found in previous papers, and the presence of recommendations' short-term effect on a market and subsequent reverse prices drift were confirmed.

Further, as for the length of an estimation window the researchers' opinions differ as it is closely connected with a model of determining normal returns for a particular stock. In a previously mentioned study of Beneish (1991), the author estimated normal returns for each stock in his sample over 300 days. What makes this paper outstanding is that he calculates the regression model parameters both before and after the event window of 60 days, so that it is possible to consider and mitigate other informational injections outside a recommendation date. In the study delivered by Pogozheva (2013), the estimation window's length is 120 days, what is justified by rather small number of recommendations' available and instability on the Russian stock market. Nevertheless, such long observation periods are only applicable when a company's strategy and its business risks remain the same what arises some doubts in terms of emerging and rapidly developing Russian business environment. In their paper related to thin Austrian market, Murg and Zeitlberger (2014) found that, according to GARCH model analysis, the most reasonable observation window is 30 days, what gives the highest explanatory power of regression model. The same observation period is used in further research of Suliga (2016) dedicated to Polish market, as well.

Commonly, abnormal returns are the determined as a difference between the returns around a significant event and the returns that are expected to be without such kind of events. Generally, they can be found by the following formula [Pogozheva, 2013]:

,

- an actual return of a security i in a moment t,

- a normal return of a security i in a moment t,

- a random variable that characterizes the stock's i abnormal return in a moment t.

For the normal returns' estimation in terms of event study analysis a large choice of models exists starting from the simplest one of average stock returns to complex multiple factor market models. Every model has its advantages and disadvantages related to estimation accuracy and biases of results what plays an important role in the findings' reliability.

Firstly, the simplest method of finding normal returns implies a calculation of average stock returns for a particular period before an event window. Theoretically, “although the model is easy for understanding and realization, in case of a very short event window it may give enough significant results due to the lower returns' variance in a couple-of-days event window” [Pogozheva, 2013, p.39]. However, to the authors' best knowledge, this model was not used in previous studies dedicated to a market's efficiency and capturing its response to some events due to the model's incompatibility with the set research objectives.

Secondly, one of the widely used methods is a simple market model that considers changes of normal returns during an event window. Among others models, this model appears in studies of Okulov (2010) and Yang and Chen (2013). It seems to be better than the model of average stock returns since it allows excluding the market's returns and reducing the abnormal returns' variance. The logic behind this is that the lower the abnormal returns' variance, the easier to capture the influence of an event precisely. In a study of Yang and Chen (2013), the researchers used this simple market model in order to find abnormal returns and then investigate how analysts' reputation affect the market response. They were managed to provide evidence that the better the reputation of a brokerage house, the stronger the investors reaction. The model was also used in the study delivered by Okulov (2010) where the efficient market hypothesis for the Russian blue-chip stock market was tested. However, this study has a sufficient limitation because the analyzed stock had noticeable share in the applied market index. One of the latest papers where this approach is used is Pogozheva (2013), where, among other issues, she tried to compare results received from this model with those obtained from capital asset pricing model (CAPM) we are going to discuss next. As it was expected, the results across these two models are similar, and the simple one-factor market model shows its competitiveness among more complicated models.

Thirdly, one of the main assumptions of CAPM is stable linear relationship between the returns of a particular stock and a considered market index. In turn, this model appears in several prior studies such as Pogozheva (2013), Murg and Zeitlberger (2014), and Suliga (2016) that investigate a market's reaction to a particular type of events. As a benchmark of a market portfolio, the aforementioned authors often used country-related stock market indices, where the analyzed companies are traded.

Fourthly, the three-factor Fama-French models also appears in previous papers related to capturing a market's response to analyst recommendations and many other significant events. Actually, this model presents an expanded form of CAPM by considering firm size and value risk factors for stocks [Fama, French, 1993]. On the one hand, since the Fama-French three-factor model considers stocks' tendencies of outperformance and includes two additional risk factors, it seems to be more flexible and adjusted compared to other previous approaches. On the other hand, one of the model's usage limitations is difficulty to access and gather huge data arrays in order to obtain necessary parameters' values. One of the resources of such data is a personal scientific web page of the authors where they frequently update the recalculated their model's parameters and many other factors for different periods and markets. Nevertheless, the list of data presented there is not full, and the information for rarely studied emerging or stand-alone markets is not available from this source.

There are enough preceding studies about a market reaction to analyst recommendations where the Fama-French model for identification of normal stocks' returns. In his already mentioned study dedicated to the US market, Womack (1996) applied this model but with some insufficient adjustments what helped him to capture prices drift after the event. In turn, Keasler and McNeil (2010), besides risk factors described by Fama and French (1993), included in their model another parameter of a return on a high-prior stocks portfolio, and finally documented that analysts are able to outperform the US market, and such recommendations are useful for usual investors. This study was reperformed one year later by Palmon, et al (2009), and the authors found that recommendations with references to management or related to M&A announcements lead to sharper investors' response. As for the most recent papers, Ishigami and Takeda (2018) followed the classical approach to the three-factor Fama-French model. This study brought some evidence to a suggestion that the Japanese market significantly reacts to target stock prices' forecasts considering both local and foreign analysts and a number of published reports on the same day. In fact, this is what boils down to the Japanese market's inefficiency around a recommendation emergence.

As it was already stated, to the author's best knowledge, there are few studies devoted to a comparison of two or more markets in terms of their reaction to analyst recommendations and, thus, their efficiency around a publication date. The majority of such papers consider only one particular market but it raises some doubts about practical implications of such findings for investors. One of the latest and mostly cited research that analyzed two markets simultaneously was conducted by Rounaghi and Zadeh (2016). However, instead of classical event study approach, the authors used ARMA model for prediction LSE and NYSE stock returns, and found, from the point of ability to forecast returns, that both markets are efficient, and participants that are more aware cannot outperform them.

Nevertheless, the findings diverge with those previous conclusions made in papers of Womack (1996), Palmon, et al (2009), Keasler and McNeil (2010), and Pogozheva (2013) related to the American and British stock markets separately, as the authors were boiled to reject the strongest form of the market's efficiency. Therefore, further study from the point of comparison of analysts' stock recommendations on the UK and the US markets with the usage of event study analysis is expected to add new evidence to the issue.



2. The US and the UK markets' summary

The USA and the UK are the most prominent representatives of so-called developed countries, and can be characterized by fair treatment of minority shareholders, efficient trading mechanism, visibility, timely trade reporting process, and strong stock market regulatory authorities. As for such authorities, the Securities and Exchange Commission (SEC) for the US market and Securities and the Investment Board (SIB) for the British market are the main regulatory services that develop and controls issuers' alignment with regulations [Hansen, 1995]. Moreover, what differs the markets from emerging and undeveloped ones is a sufficient variety of brokerage houses and other mutual funds that, among other advantages, simplify entering the markets for investors.

As for stock exchanges operating in the United States and the United Kingdom, the largest ones for each country are the New York Stock Exchange and the London Stock Exchange. The New York Stock Exchange (NYSE) is located in New York City and is the largest securities exchange in the world by the total market capitalization of companies listed there. According to the World Federation of Exchanges' reports, the London Stock Exchange (LSE) is the largest European stock exchange.

Almost all stock exchanges have their related market indices that present the general market performance and are considered as the main indicators, which reflect global business environmental changes and an implicit measure of an economy's health. Although many market indices are calculated for the LSE and the NYSE, in this paper FTSE 100 and S&P 100 are emphasized, and their returns were taken as benchmarks for the US and UK stocks. These indices include the most liquid stocks from different economy sectors what provides with an ability of an asymptotically optimally diversified portfolio depending on the stocks' returns and risks for both active and passive traders.

3. Hypotheses development

3.1 Reaction to pure recommendations

To summarize findings from the section about literature related to the topic, it seems reasonable to outline the main concerns about what might influence the market reaction to stock recommendations and consequently to attributes to the analysts' recommendations' value.

Before an analyst, or a brokerage house, release their recommendations, they gather all information they have and conduct a fundamental and technical analysis of stocks, as well as their demand and supply among investors, so that it becomes clear whether the stock is underrated or overrated. If analysts are enough precise, the stock price under the market pressure is expected to move in the corresponding direction. Although the logic behind for the stock prices' reaction to `buy' and `sell' analysts' recommendations is enough clear, the picture of prices movements after `hold' recommendations is ambiguous, and mostly depend on side factors such as, for instance, investors' personal behavior, or whether this neutral recommendation is a result of a downgrade/upgrade of the stock's rating. This aspect is disclosed a little bit further. Although the US and the UK stock markets are assumed to be efficient and rather transparent, basing on aforementioned previous studies (Womack, 1996; Keasler and McNeil, 2010; Pogozheva, 2013), they are expected to react to such recommendations. Thus, as of now, along with Murg, et al (2014), the first hypothesis can be formulated as follows:

H1: The NYSE and the LSE stocks experience significant abnormal returns when new analysts' recommendations are released.

3.2 Reaction to changes in previously published recommendation levels

A good analyst always observes the subsequent stock price movements. The best practice for analysts is to capture the moment when under the market pressure mentioned above the actual stock price shifts further than its expected fair value, especially when the market usually shows imperfect level of efficiency. In this moment, a change in a recommendation level is to appear in order to outperform the market. The idea of more robust market's response to stocks' upgrades and downgrades than a recommendation alone was present in several prior papers such as Jegadeesh, et al (2006), Murg, et al (2014), and Suliga (2016). There is also an assumption that ordinary investors are likely to interpret the switching from `buy' to `sell' recommendation (or vice versa) as coming significant changes in a company's business climate as well. Thus, it seems natural that the greater the change in a recommendation level compared to the previous one, the stronger the market reaction is expected to be. For instance, switching from `buy' to `sell' is expected to show greater explanatory power in comparison with switching from `hold' to `sell'. The other situation is a repetition of the prior recommendation, when analysts feel that the current stock price still has not reached its fair value, and change the forecasted target price. However, in this case, the stock prices' shift is expected to be lower rather than in a situation when the recommendation direction changes, as many investors are likely to already act after the first analysts' signal, and are less perceptive to such maintaining reports. The main concern in this paper is to investigate whether analysts that give recommendations related to the NYSE and the LSE are enough qualified, sensitive or possess privileged information than investors have. This is what corresponds to the second research hypothesis that is split into two sub-hypotheses that appeared in Murg, et al (2014) in similar form:

H21: The stock prices do not significantly react to stock recommendations' reiterations.

H22: The greater the change in a stock recommendations' level, the stronger the markets' reaction.



4. Data and methodology

4.1 Data and sample description

The initial sample of historical recommendations consisted of 1881 events in total for 168 companies traded at the LSE and the NYSE in a period between 1 January 2016 and 31 April 2019. This data array was downloaded from Thomson Reuters Eikon, an information terminal that provides its users with historical and current quotes, events, opinions and forecasts. However, not all of the events appeared to contain all necessary details about a recommendation, and, thus were included in the further research. The steps of reducing the initial number of recommendations for future analysis are the following:

1) Deleting those recommendations where the name of an analyst or a brokerage house was not disclosed. To the author's point of view, if the name is undisclosed, the source of such recommendations is expected not to be enough reliable or these recommendations are random and non-systematic, thus, there is an assumption that the market would not consider such recommendations.

2) Deleting those events, where the stocks' target prices were absent, as this leads to a problem of determining a connection between the significance of the recommendations and the markets' response.

3) Omitting those recommendations, which forecasted target prices change the current quote by less than 5%. This is likely to allow concentrating only on those events that influence the stock returns with more probability.

4) There was a necessity of reducing the companies' sample by a liquidity criterion. It was discovered that, although the chosen indices are expected to include highly liquid shares, several companies had poor trading volumes during an observation window, what biases their normal returns identification.

5) If there were several unidirectional recommendations during a period less than five trading days what corresponds to a length of an event window in terms of event study analysis, only the first of such recommendations was taken into account.

6) If such recommendations were ambivalent, and their event windows overlapped, these events were skipped in order to avoid the biasness of abnormal returns.

On balance, the final numbers of recommendations in the sample were cut down on to 806 events related to 148 companies. On the following Table 1 and Table 2, the collected data is summarized.

Table 1. Allocation of companies in the sample by sectors.

	NYSE	LSE
Mining & Oil & Gas	6	5
FMCG	8	10
Financials	13	12
Communication Services & Media	4	6
Health Care & Pharmaceuticals	12	4
Industrials	13	6
Consumer Discretionary	7	13
Others	14	16
Total	77	72

Table 2. Allocation of recommendations in the sample.

Stock exchange	Positive recommendations	Neutral recommendations	Negative recommendations	Total
NYSE	346	125	42	513
LSE	157	85	51	293
Total	503	210	93

As it can be noticed, although the total number of companies included in the sample, as well as their distribution across the business sectors, is similar, the difference in total number of recommendations between the two stock changes is sufficient. Another point is that there is a skewness to positive recommendations on both markets. This might be explained by either a desire of analysts to support the markets and their aversion to giving negative recommendations, or by their confidence in general market growths.

4.2 General event study's methodology description

In order to test the strength and direction of the markets' reaction to analysts' recommendations, and, thus, to determine whether these recommendations bring value to investors, it was decided to use the event study approach in this paper. The whole array of collected data is split into groups by the market, which a stock recommendation is related to, and by a level of these recommendations.

Firstly, classical one-factor market model is chosen to determine expected returns for a particular stock in an estimation window. Unlike the model of average returns, this model is more sensitive estimation approach as it takes into consideration such exogenous factors as the market index fluctuations. Moreover, the usage of more sophisticated models (for instance, multiple-factor Fama-French model [Fama, French, 1993]), is not necessary as it requires expanding the observation window due to its complexity. Otherwise, in case of the narrow window and, thus, insufficient number of observations, the estimated parameters are likely to be not enough significant, and the model loses its explanatory power.

Secondly, along with methodology from Murg, et al (2014) and Suliga (2016), the estimation window lasts for 30 trading days. A justification of such window is that, on the one hand, it is enough long to purely define a particular stock's beta, and, on the other hand, the general riskiness of the company's business is unlikely to change during this period.

Thirdly, as it was already mentioned, market reaction to analysts' recommendations is supposed to be short due to the events' frequency. Thus, in this paper an event window of 5 days is considered. The usage of such short event window allows mitigating the effect from other informational `splashes' in an issuer's life, and to omit price fluctuations that are not connected with an analyzed event. Moreover, there is a suggestion that “long-horizon tests are highly susceptible to the joint-test problem, and have lower power than the shorter ones” [Kothari, Warner, 2006, p.8]. Therefore, if is a date of a recommendation release, then in a framework of the event study analysis, a period of an observation window is […], and the event window is […]. Based on Okulov (2010), the event window starts from the day before a publication, so that it becomes feasible to capture probable information leakage and investors' pre-event reaction.

4.3 Testing the first hypothesis

The first step of testing the strength and direction of the markets' response is calculating historical daily returns of each stock in the sample and daily returns of the S&P100 and FTSE100 indices during the considered period using the following formula:

,

After that, abnormal returns that are widely used as a measure of a market's reaction to events is calculated. For each recommendation in the sample, companies' normal returns were estimated in an observation window with the usage of a simple one-factor market model. Since the observation window lasts for 30 days, the model can be presented as follows:

,

where , is a daily returns of a market index related to observed stock exchange, and parameters and is estimated parameters obtained from a regression model. Further, the companies' daily abnormal returns during whole period of estimation and event windows are identified:

,

where is an expected normal daily company's returns that are determined by putting in the market model actual market index returns, and previously estimated and from the equation 5.2. The formula for expected daily returns of a particular company can be written in the following form:

,

Then, the whole sample of recommendations is split into those that relate to the LSE and the NYSE, and allocated by a recommendations' direction. The main concern in this paper is to determine whether these average abnormal returns (AAR) for each day in the event windows across the stock exchanges and levels of recommendations significantly differs from zero. Thus, for each AAR the statistical hypothesis are presented as follows:

,

,

where .

In order to apply parametric cross-sectional Student's t-test, it is necessary to make sure that AAR for positive, neutral and negative recommendations for both stock exchanges has similar-to-normal distribution (please, see Appendix A). However, perfect normal distribution is not a crucial point in an application of parametric tests because, according to the Central Limit Theorem, large number of observations in a sample allows not to consider the normality [Kothari, Warner, 2006, p.33]. The formula for calculating the cross-sectional t-statistics looks like in the paper of Shen (2014):

,

where corresponds to the standard deviation across firms at the day :

,

After this, the calculated t-statistics is compared with the table value of considering two-tailed parameter and corresponding number of degrees of freedom. Actually, this is how we obtain the p-value of the statistics and the probability of . The results from testing the significance of AARs during the event window are presented in the Table 4 (for the LSE) and Table 5 (for the NYSE) in the section 6 “Results and discussion”.

Besides the AAR for each day in the event window, the average cumulative abnormal returns (ACAR) during the whole event windows are calculated for the two markets. This makes possible to consider not only those days, where significant response is captured, but also a general markets' reaction to published recommendations. In fact, the steps for ACARs' levels of significance' identification is similar to procedures related to AARs although the ACAR for a particular stock is a sum of AAR for each day from to in the event window. To sum up, the following statistical hypotheses are developed for all recommendations' levels and the stock exchanges:

,

,

The t-statistics is calculated similarly to equation 7.6 [Shen, 2014]:

,

where is a standard deviation:

,

For determining a significance of difference between the ACAR for the NYSE and the LSE, the Student's two-tailed test with expected zero-mean difference is employed. Please, see the results of the NYSE and the LSE cumulative reactions and their comparison in the Table 6.

4.4 Testing the second hypothesis

As for the second hypothesis, the logic is the same as for the first one despite an additional emphasize on a preceding level of recommendation. We have three-leveled system of published recommendations: positive, neutral, and negative. Therefore, if to consider the feasible changes in the levels, we get nine possible outcomes: repeated positive, neutral > positive, negative > positive, positive > neutral, repeated neutral, negative > neutral, positive > negative, neutral > negative, repeated negative. As it can be noticed from the Table 5 and Table 6 in the next section, the markets response similarly to all types of recommendations. Therefore, for testing the second hypothesis it was decided not to split the recommendations into group based on the stock exchanges because this allows enlarging the sample for the analysis and, thus, obtaining more reliable and representative results. The following Table 3 shows the summary of such recommendations' numbers.

Table 3. Allocation of recommendations for the second hypothesis.

	To positive recommendations	To neutral recommendations	To negative recommendations	Total
From positive	297	94	38	429
From neutral	84	58	24	166
From negative	31	17	25	73
Total	412	169	87

4.5 Determining other factors of the markets' reaction

Firstly, according to Banz (1981), “smaller firms have higher risk-adjusted returns than bigger ones”. Another related point is that, since smaller companies are less transparent and rarely highlighted in news, the difference in awareness between ordinary investors and analytic agencies is likely to increase [Murg, et al, 2014]. This makes people to rely on the recommendations more, and as a result boosts the market reaction. Although a detached research hypothesis related to this issue is not stated to be tested, this study intends to take into consideration this point by conducting an additional regression model analysis with an implementation of some adjustments to existing models from prior studies. Therefore, we expect the stronger market's response to recommendations about smaller companies. Secondly, as the analysts' recommendations very often include not only a detached advice whether to buy or to sell a particular stock, but also a supporting opinions about the stock's target price, investors before their trading actions take into account these information in couple. Thus, this issue is worth considering in a regression model. The last but not least factor is already analyzed extent to which the recommendation level is changed.

As a depending variable, is included in the model. The main reason is that, according to results from testing the first and the second hypotheses, the most significant market reaction is likely to appear in an event day and lasts for two days after it.

Despite several adjustments, the regression model is built based on Suliga (2016), and can be presented as follows:

,

where:

· is a variable that describes the extent of a change in a recommendation level. It has the value of 2 when a recommendation is upgraded from negative to neutral, value -2 - for downgrades from positive to negative, value 1 - upgrades from neutral to positive or from negative to neutral, value -1 - downgrades from positive to neutral or from neutral to negative. A zero-value of this variable stands for the first recommendations for each stock in the sample, and for their repeated levels.

· is a variable that describes the related difference between the forecasted and current stock price. It is calculated with simple formula similar to calculation of stock returns:

,

· is a variable that determines a size of a company, and is calculated as a natural logarithm from the average company's annual sales. This variable is calculated in USD, so for companies traded on the LSE the sales volumes were converted from GBP using the average currency rates for considered period.

· , and are the dummy variables that are valued as 1 if a particular recommendation is positive, neutral of negative, respectively; otherwise, they are zero.

If for the first two variables the logic is transparent, the last three variables were multiplied by the company's size. To the author's point of view, this adjustment provides with an ability to capture a detached effect of a company's size to such recommendations without splitting the sample to three separate models related to positive, neutral and negative ones. In order to get closer to normal distribution of the explained variable, the outliers are to be deleted from the sample using the box and whisker plot.



5. Results and discussion

5.1 Results from testing the first hypothesis

Table 4. Abnormal returns around a recommendation date for the LSE (2016-2019)

	Positive recommendations	Neutral recommendations	Negative recommendations
	157 events	85 events	51 events
t	AAR	p-value (t-statistics)	AAR	p-value (t-statistics)	AAR	p-value (t-statistics)
3	0.00178	0.1215	-0.00304	0.1072	0.00046	0.8681
2	0.00145	0.1854	0.00191	0.3260	0.00035	0.8744
1	0.00729***	0.0000	0.002527	0.1732	-0.00114	0.5956
0	0.00835***	0.0000	-0.00174	0.3005	-0.01073***	0.0012
-1	0.00072	0.6040	-0.00250	0.2735	0.00014	0.9520

Table 5. Abnormal returns around a recommendation date for the NYSE (2016-2019)

	Positive recommendations	Neutral recommendations	Negative recommendations
	346 events	125 events	42 events
t	AAR	p-value (t-statistics)	AAR	p-value (t-statistics)	AAR	p-value (t-statistics)
3	-0.00004	0.9686	-0.00017	0.9105	-0.00417	0.1013
2	0.00031	0.7513	-0.00129	0.4654	0.00176	0.4893
1	0.00644***	0.0000	0.00240	0.1254	-0.00543	0.1127
0	0.00791***	0.0000	0.00044	0.7904	-0.00688***	0.0088
-1	0.00101	0.1947	-0.00232	0.1178	0.00164	0.5594

From the tables above we can obtain a piece of evidence that both markets almost identically react to recommendations released by analysts. As for positive recommendations, the significant response is captured on a publication date and the day after it. The direction of the reactions matches the level of a recommendation given, and the recommendation is 99% likely to cause significant positive abnormal returns for the analyzed stocks. Although the situation around negative recommendations on the both markets is almost the same as significant negative abnormal returns are identified on the event day, the investors' reaction seems to be shorter compared to the positive recommendations and lasts for one day. These results are partly in line with those reached by Pogozheva (2013), Murg and Zeitlberger (2014), and Suliga (2016). In these papers, the authors managed to find sufficient two-days effect for both positive and negative recommendations starting from the date of an event. As for the neutral recommendations, no significant markets' response around the publication date is discovered. However, in studies delivered by Suliga (2016), and Ishigami and Takeda (2018), it was concluded that neutral recommendations are expected to be negatively perceived by investors, and to cause short-term negative abnormal returns. If to have a look at the results for the day before an event, they do not overcome 90% level of significance, what boils down to a conclusion that information's leakages have low probability on these markets. For the days and the reaction is generally insignificant as well.

To summarize, we have to accept statistical null hypotheses of zero average abnormal returns for the days , and for both markets across all recommendation levels, plus the other days for neutral recommendations. At the same time, we have to accept the alternative statistical hypotheses for the days and for positive, and for the day for negative recommendations on the NYSE and the LSE, and to conclude the direct connection between the levels of recommendations given the directions of the abnormal returns shifts.

On the Table 6, the results of testing the significance of average cumulative abnormal returns calculated for the whole event windows. The findings totally correspond to the previously discussed results. It was determined that ACARs for both markets overcome the necessary level of significance (90% for the negative recommendations and 99% for the positive ones). The directions of the markets' response match the directions of the related recommendations. One of the main points is to compare the ACARs caused by the events between the LSE and the NYSE. Based on a conducted Student's t-test, we are to suggest that the differences between the markets across all recommendation levels do not significantly exceed zero in absolute value. Although one can notice a little skewness of the reaction on the LSE to positive recommendations, and, vice versa, on the NYSE to negative ones, this differences are statistically insignificant. Therefore, the stock exchanges respond to such events identically and with almost equal strengths. value price stock return

Table 6. Comparison of LSE and NYSE ACAR around recommendations.

	LSE	NYSE	Difference
	Mean ACAR [-1;3]	p-value	Mean ACAR [-1;3]	p-value	Mean ACAR [-1;3]	p-value
Positive	0.02008***	0.0000	0.01563***	0.0000	0.00446	0.2689
Neutral	-0.00285	0.5240	-0.00095	0.7904	-0.00190	0.7401
Negative	-0.01091*	0.0914	-0.01309**	0.0372	0.00218	0.8109

5.2 Results from testing the second hypothesis

The Tables 7-9 present the average abnormal returns for all possible cases that were determined in the previous section. Firstly, have a look at positive recommendations taking into consideration levels of the preceding ones on the Table 7. In common, when a recommendation switches from a negative or neutral to a positive level, as well as when a positive recommendation is repeated, this leads to a positive abnormal returns. This corresponds to general significant markets' response to a single positive recommendation observed previously. However, here we have one controversial point. Despite the sharpest reaction is captured in a case of prior negative level, this reaction is shorter than in cases of preceding neutral and positive recommendations. In those situations, the significant response lasts from a publication date until one day after, whereas for negative ones the reaction is captured only on the event day. This might be explained by the fact that sufficient stock upgrades are enough rare situations, and investors are likely to be cautious to such changes. The other suggestion is that such striking recommendations' changes are frequently given by independent and not so famous brokerage houses that are less afraid of accepting risks of a probable mistake, and the investors considers them more carefully.

...