Do capital markets discriminate. against analysts with foreign sounding names?

Transcription

1 Do capital markets discriminate against analysts with foreign sounding names? Alexander Kerl 1 / Nils Schäfer Abstract Based on a cross-country dataset including the most developed capital markets we examine to what extent the informativeness of sell-side analyst reports depends on analysts characteristics such as analyst location and, new within this strand of the literature, the sound and familiarity of the analysts name. In line with the literature, our results show that local analysts, i.e. analysts that are close to the company, issue forecasts that trigger a higher market response following recommendation updates and earnings and target price revisions. Similarly, the accuracy of their forecasts is much higher. However, we can show within our study that market participants additionally react differently to analysts recommendations, depending on the sound and familiarity of the analysts name. Investors seem to follow financial analysts more if the analysts name seems familiar, relative to the cultural and ethnical background of the company that the analyst is covering. In contrast, our results show that forecasts issued by analyst with foreign sounding names trigger a much weaker stock price reaction. Similarly, those forecasts are also much less accurate, relative to forecasts of analysts with familiar names. Our results are first evidence that investors value the cultural proximity between analysts and companies apart from pure location effects. Keywords: analyst reports, forecast, cultural proximity, foreign sounding name, local analyst, foreign analyst JEL-classification: G14; G15; G18; G24; G32 1 Corresponding author s Alexander.Kerl@wirtschaft.uni-giessen.de. All authors are with the Department of Financial Services, University of Giessen, Licher Str. 74, Giessen, Germany. 1

2 1. Introduction Past research suggests that geographic and cultural proximity between investors and companies affect portfolio choices and performance. Coval and Moskowitz (2001), for example, demonstrate that mutual fund managers in the US possess information advantages about local companies and thus, earn abnormal returns. In a related study, Choe et al. (2001) document for the Korean stock market that domestic investors have private information advantages compared to foreign investors. Grinblatt and Keloharju (2001) reveal that investors are more likely to trade stocks of companies whose manager share the same cultural background as the investor or companies which report their financial statements in the native language of the investor. In a recent working paper, Kumar et al. (2012) show that US mutual fund managers with foreign sounding names have significantly lower fund flows compared to managers with typical English sounding names. Intrigued by the aforementioned findings, we investigate in this study what influence the location as well as the cultural/ethnic background of a financial analyst has on the informational value of his/her research. Financial analysts act as information intermediaries between companies and investors and thus, play an important role on the financial markets. Despite the growing attention to cultural and geographic aspects in finance literature, the number of studies that are directly related to our topic is limited and they primarily deal with the geographical location of analysts. Malloy (2005), for example, shows that analysts in the US who are closely located to the covered company issue more accurate earnings forecasts than analysts from more distant cities. Taking on an international perspective, Bae et al. (2008) find that analysts being resident in the same country as the followed company are more precise in their earnings forecasts than non-resident analysts. To the best of our knowledge, there are only two studies providing evidence about the impact of culture on analyst research, each focusing on Chinese analysts and Chinese companies. Jia 2

3 et al. (2013) demonstrate that earnings forecast revisions by local Chinese analysts are more likely to be influential among Chinese investors than among foreign investors. Du et al. (2013) focus exclusively on companies that are headquartered in China but listed on a US stock exchange. According to their findings, Chinese analysts who work in the US produce more accurate earnings forecasts than their non-chinese counterparts. Furthermore, they reveal that upward revisions of earnings forecasts and stock upgrades by non-local Chinese analysts exert a stronger impact on stock prices. Our study contributes to the described research streams by using a multi-country sample. In concrete, we base our analysis on 729,126 analyst reports for companies from France, Germany, Italy, Japan, Spain, Switzerland, United Kingdom and the United States for a period from 2004 until Taken together, these countries account for approximately 56 % of the world market capitalization. 2 We extract from each analyst report available to us, the name of the brokerage house employing the analyst as well as the full name of the analyst. We derive the location of the analyst from the country in which the broker is headquartered. From the full name of the analyst we infer to the most likely ethnic origin, with the help of a name-based classification scheme. Then we use both information and match them with the countries in which the covered companies are located. To analyze the impact of location on analyst research, we compare the informational value of research reports by local analysts and non-local analysts. Analogous to Bacmann and Bolliger (2003), we define a local analyst as an analyst who is working for a brokerage house that is headquartered in the same country as the covered company. Where this is not the case, we classify an analyst as a non-local analyst. To examine the impact of the ethnic background on analyst research, we compare the informational value of research reports by domestic and foreign analysts. A domestic analyst is defined as an analyst whose ethnic background 2 The information is based on Bloomberg as per June

4 corresponds to the country in which the covered company is located. In the opposite case we refer to a foreign analyst. To assess the informational value of analyst reports we use two proxies. On the one hand we compare the short-term market reaction to stock recommendation changes, earnings forecast revisions and target price revisions. On the other hand we compare forecast errors with respect to future earnings and future stock prices. In contrast to most related work, we explicitly use two measures of informativeness to get a more comprehensive overall view. Moreover, we consider target prices that were neglected in analyst-related research for a long time. We assume that the stronger the short-term market reaction and the lower the forecast errors, the more informative are analyst reports. The analysis of the short-term market reaction reveals a number of interesting findings. First, cumulative abnormal returns in response to stock recommendation downgrades as well as earnings forecast and target price revisions by local analysts are significantly higher compared to the ones of non-local analysts. Both Malloy (2005) and Bae et al. (2008) document that local analysts have better access to company-related information. Our results suggest that investors are aware of the information advantage and thus pay more attention to local analysts. However, for stock recommendation upgrades, we find the diametrically opposite result, indicating a stronger market reaction to non-local analysts. We argue that local analysts use stock upgrades to support the investment banking business of the respective brokerage house (see e.g., Lai and Teo, 2008). Investors are aware of the conflicts of interest and thus tend to follow stock upgrades by non-local analysts. Second, cumulative abnormal returns following earnings forecast and target price revisions suggest that domestic analysts exert a stronger impact on stock prices than foreign analysts. With respect to stock recommendation changes, we find weak evidence for a stronger market 4

5 reaction to downgrades of domestic analysts but no evidence for a different market reaction to upgrades at all. Splitting our sample according to the location of the analysts provides us with further insights. Within the group of non-local analysts, domestic analysts consistently outperform foreign analysts. For the subsample of local analysts, we measure incremental abnormal returns associated with target price and earnings forecast revisions by domestic analysts. Hence, our results reveal that investors attribute more information value to analyst reports of domestic analysts irrespective of their location. There are several potential explanations for this finding. For example, according to Du et al. (2013), investors may value the cultural proximity between analysts and companies. Following Kumar et al. (2012), it is also conceivable that investors discriminate against foreign analysts because of social biases such as in-group favoritism and stereotyping. Another argument put forward by Jia et al. (2013) is that investors are faced with attention constraints and thus, focus exclusively on sources of information that seem familiar to them, namely reports by domestic analysts. The results from the forecast error analysis comply with our previous findings. We find that forecast errors of local analysts for both earnings and target prices are consistently lower compared with non-local analysts. Similarly, for the overall sample, we detect earnings and target price forecasts by domestic analysts to be more accurate than those of foreign analysts. However, for the subsample of non-local analysts, forecast errors do not differ significantly between domestic and foreign analysts. In contrast, for the subsample of local analysts, we reveal a significantly higher forecast accuracy of domestic analysts. Thus, we can confirm the results of prior literature showing that distance is negatively related to the precision of analysts forecasts (see e.g., Orpurt, 2004; Bae et al., 2008). Our results suggest that the location of analysts is a more important factor for forecast accuracy compared to the ethnic origin of the analysts. Nevertheless, we detect the lowest earnings and target price forecast errors among domestic analysts who work for a local broker. 5

6 The findings above have practical implications for both brokerage houses and investors. It might be advisable for a broker to recruit analysts who have cultural or ethnic ties with the covered companies or to establish local subsidiaries in the respective countries. Investors, on the other hand, should follow reports by local domestic analysts since their forecasts are more accurate and lead to higher abnormal returns. The structure of this study is organized as follows: Section 2 provides more comprehensive overview about the related literature and establishes the main research hypotheses. In Section 3, the data set, the variables and the name-based ethnicity classification scheme are introduced. Section 4 presents the empirical results of both the analyses of market reaction and forecast errors. In Section 5, the robustness of our results using alternative ethnicity classifications is tested. Finally, Section 6, concludes the study. 2. Related literature and research hypotheses In recent years, a few studies have already examined the impact of analysts location on forecast accuracy and/or market reaction. Conroy et al. (1997) were among the first to address this issue. Their results show that analysts who work for a Japanese broker produce more accurate earnings forecasts for Japanese companies than analysts working for a US broker. Employing a sample of companies from Latin America, Bacmann and Bolliger (2003) find the opposite result indicating that earnings forecast errors of non-local analysts are lower than those of local analysts. Moreover, compared to local analysts, they reveal higher abnormal returns in response to earnings revisions of non-local analysts. The authors classify an analyst as local or non-local according to the country of origin of the brokerage house. In contrast, Orpurt (2004) defines an analyst as local in dependence of whether the analyst is resident in the same country as the covered company or not. With respect to earnings forecast accuracy, he finds a significant local analyst advantage for five out of seven European countries. Bae et al. (2008) compare earnings forecast accuracy between local analysts, expatriates (local 6

7 analysts working for a foreign broker) and non-local analysts for an extensive sample of 32 countries. The results of the authors show that both local analysts and expatriates are more accurate than non-local analysts. When comparing local analysts with expatriates, the authors find no significant differences in forecast accuracy. Focusing on stock recommendations for Taiwanese companies, Chang (2010) documents an outperformance, as measured by the market reaction, of expatriates compared to both non-local analysts and local analysts. According to Chang (2010), it is important for an analyst to have both physical presence in the country of the covered company as well as a global network of knowledge and resources. Most of the cited studies conclude that geographical proximity between analysts and covered companies leads to better forecasts. As a core argument for this finding, it is often stated that local analysts benefit from information advantages as it is easier for local analysts to gather information from company s stakeholders (e.g. customers, employees and suppliers) and monitor the company s activities as well as the market environment (Bae et al. 2008). In addition, local analysts are more likely to meet members of company s management during one-to-one interviews compared to foreign analysts who mostly rely on conference calls in which private information is rarely disclosed (Malloy 2005). Cohen et al. (2010) show that social networks between analysts and senior executives stemming from a common educational background are associated with a valuable information flow which can lead to a superior quality of research. Thus, one could assume that such networks are more widespread amongst local analysts than among non-local analysts. Another disadvantage concerning nonlocal analysts could be a limited knowledge of local financial accounting standards. Tan et al. (2011) for example reveal in their study that non-local analysts earnings accuracy increased for countries which adopted the International Financial Reporting Standards. Conversely, non-local analysts issue less accurate earnings forecasts when local accounting standards are in force. 7

8 Cultural or ethnic aspects have been largely neglected in analyst-related literature. Recently, two studies appeared providing first evidence on how culture affects analysts research. Jia et al. (2013) take advantage of the segmented stock market in China and analyze whether Chinese and foreign investors react differently to earnings revisions published by Chinese analysts. Their results show that earnings revisions by Chinese analysts are more likely to be influential for Chinese investors whereas earnings revisions by foreign analysts are more likely to be influential for foreign investors. The authors argue that investors must choose from a large number of analyst reports. Due to attention constraints Chinese investors prefer to follow analyst reports by Chinese analysts since they are more familiar or trustworthy from their perspective. In a related study, Du et al. (2013) find that non-local Chinese analysts produce more accurate earnings forecasts for Chinese companies that are traded in the US than their non-local foreign colleagues. Moreover, they identify a stronger market reaction to earnings revisions of the non-local Chinese analysts. Since both groups of analysts are nonlocal and only differ in their ethnicity, the authors conclude that not only the geographic location is a determining factor for analyst s research quality but also their cultural understanding (e.g. language skills and familiarity with customs and practices). Both Jia et al. (2013) and Du et al. (2013) use analysts names to detect whether an analyst has Chinese origin or not. There is an interesting strand of literature dealing with name-based discrimination. In an early study, Razran (1950) asked participants of an experiment to rate photos (e.g. in terms of beauty and intelligence) of persons who optically could not be assigned to a specific ethnicity. Two month later the experiment was repeated but each photo was attached a name that is typical for a specific ethnic group. Compared to the first rating, the same participants often evaluated the photos differently, so that the author concludes that names can be triggers for ethnic stereotyping. More recent studies show that name-based discrimination is still a topic 8

9 of debate. Cotton et al. (2007) detect that a person s name has a considerable impact on how the person is perceived by others (e.g. in terms of likeability). Their results suggest that persons with common English names have the better hiring chances on the US labor market compared to Russian and African-American names. Carpusor and Loges (2006) confirm similar results for the US housing market. In a field experiment, the authors sent s with varying names to lessors offering vacant apartments. Interested parties with common English names received significantly more positive responses than persons with names of other ethnic origin. More related to our study, Kumar et al. (2012) find that in the US mutual fund managers with foreign sounding names have lower fund flows compared to fund managers with English names. Since their result cannot be traced back to differences in performance or investment profiles, the authors argue that investors tend to discriminate against managers with foreign sounding names due to social biases such as ethnic stereotypes. Based on the literature presented above and the definitions mentioned in the introduction we derive the following three hypotheses: H1: The market reaction to research reports published by local analysts is stronger compared to those published by non-local analysts. H2: The market reaction to research reports published by domestic analysts is stronger compared to those published by foreign analysts. H3: Forecasts of local analysts (domestic analysts) are more accurate than those of nonlocal analysts (foreign analysts). 9

10 3. Data sample 3.1. Analyst report data The primary dataset for this study is provided by the FactSet Research System and contains analysts report information for 4,410 companies from eight major stock markets, for the period from August 2004 to November The countries included in our sample are France, Germany, Italy, Japan, Spain, Switzerland, the United Kingdom and the United States. We concentrate our study on the analysts information that is most crucial for investment decisions, namely earnings per share forecasts, target price forecasts as well as stock recommendations. For an analyst report to be included in our sample we require that each of these three measures is jointly available. Following Arand et al. (2013), we exclude companies from our sample which are covered by less than three analysts in a given year and whose stock price is less than one US dollar at the publication day of the analyst s report. Furthermore, we drop observations with a price-to-book ratio smaller or equal to zero. 3 Market reaction variables Our first proxy for the informational value of an analyst report is the short-term market reaction to stock recommendation changes and forecast revisions of earnings and target prices. To reveal shifts in an analyst s opinion about a company s market prospects we compare the information of the current research report with the information from the previous research report. Recommendation changes are identified on the basis of a five-step scale ranging from strong sell to strong buy. We define dummy variables to indicate whether an analyst s stock recommendation is an upgrade (R_Up), a reiteration (R_0) or a downgrade (R_Down) compared to the previous recommendation by the same analyst for the same stock. Earnings revisions (EPS_Rev) are defined as percentage difference between an analyst s 3 These steps are taken to avoid that our results are influenced by illiquid, small stocks or unusual large bid-ask spreads. 10

11 current earnings per share forecast for a given stock and the previous earnings per share forecast for the same stock. Similarly, target price revisions (TP_Rev) are defined as percentage change in an analyst s target price forecast for a given stock. Byard and Shaw (2003) argue that reports conveying stale information adversely affect the correct inference of analyst-related variables. Hence, we calculate recommendation changes and forecast revisions only, if the time elapsed between the current report and the previous report does not exceed 90 days. To mitigate the influence of outliers, we cut off the 1% and the 100% percentile of both, earnings revisions and target price revisions. Extreme revisions are likely caused by coding errors. For measuring market reaction, we calculate abnormal stock returns for an event window of five days centered on the publication day of the analyst report. We draw the necessary data on daily stock returns from Datastream. At first, we use a standard market model, proposed by Brown and Warner (1985), to calculate the expected stock returns. The estimation window ranges from day -250 to day -11, relative to the publication day of the analyst report. The abnormal stock returns then result from the difference between the actual stock returns and the expected stock returns. Finally, we aggregate the abnormal stock returns over the event window and receive five-day cumulative abnormal returns (). We exclude observations representing the 1% and the 100% percentiles of. Forecast error variables Besides market reaction we consider forecast accuracy as a second proxy for the informational value of an analyst report. For both earnings forecasts and target price forecasts, we calculate analyst s absolute and relative forecasts errors. We define the relative earnings forecast error (EPS_Error_rel) as percentage deviation of an analyst s earnings per share forecast for a given year, from the actual realized earnings per share in the same year. The 11

12 relative target price forecast error (TP_Error_rel) on the other hand is defined as percentage deviation of an analyst s target price forecast from the actual stock price at the end of the 12- month forecasting period. To account for outliers, we truncate the 1st and 100th percentile of EPS_Error_rel and TP_Error_rel. 4 The absolute earnings forecast error (EPS_Error_abs) and the absolute target price forecast error (TP_Error_abs) are obtained by taking the absolute value of the respective relative forecast errors. Since the value range of the absolute forecast errors is - by definition - limited downwards towards zero, we only drop the 100% percentile. Whenever an analyst's forecast falls below the actual value, the relative forecast errors become negative indicating a higher accuracy. The absolute forecast errors in contrast become positive regardless of whether the actual value exceeds the forecast or falls below it. Thus, per definition, the absolute forecast error will be at least as high as the relative forecast error. As pointed out by Kerl (2011), the relative forecast error is more relevant for investors, whereas the absolute forecast error is a better proxy for an analyst s forecasting abilities. 5 Several studies have shown that analysts tend to publish overly optimistic forecasts (see, e.g., Chopra, 1998; Eastwood and Nutt, 1999). It can be assumed that analyst s optimism is positively related to analyst s forecast errors. Thus, we calculate for both earnings forecasts and target prices, respectively, a variable controlling for analyst s optimism. The earnings yield (EPS_Yield) is measured as the ratio of the earnings per share forecast to the current stock price at the day the earnings forecast was published. As regards optimism in analyst s target prices we follow Bonini et al. (2010) and calculate the implicit return (Implicit_Return). For this, we subtract the current stock price from the target price and divide the result by the current stock price. A positive value for Implicit_Return implies, for example, that an analyst expects an increase in the current stock price over the period forecasted. The higher EPS_Yield and Implicit_Return are, the higher is the analyst s optimism. 4 Bradshaw et al. (2012) control in a similar way for outliers. 12

13 [Insert Table 1 about here] Table 1 shows the number of analyst reports that meet our selection criteria, for each, country and year. In total our sample comprises of 729,126 analyst reports, of which almost half is attributable to the US (389,396). The number of reports has steadily increased from 848 in 2004 to 254,834 in 2005 due to a growing number of brokerage houses cooperating with FactSet. Unfortunately, given a lack of data, the number of observations for the forecast error analysis is reduced to 687,907 (target prices) and 482,691 (earnings forecasts), respectively Name-based ethnicity classification In order to test our second hypothesis, we need to know the ethnic origin of each analyst. Unfortunately, our dataset does not provide direct information in this regard. The only information available to draw any conclusions about the ethnic origin of an analyst is his/her full name. We utilize this information and develop a name-related scoring system that assigns each analyst from our sample to an ethnic group. The use of names to classify persons into distinctive ethnic groups is not uncommon. For example, studies from the field of medicine often use similar approaches. 7 Although our dataset contains analyst reports for companies from eight different countries, we can only categorize six corresponding ethnic groups. As the UK and the US share a common language, no reasonable name-related distinction is possible. Thus, we group analysts from both of these countries as English. Similarly, German analysts and the majority of their Swiss colleagues are grouped as German. The remaining four ethnicities are French, Italian, Japanese and Spanish. All analysts whose ethnicity cannot be determined or is clearly different from the previously mentioned groups are assigned to a residual group named as Others. 6 With respect to the earnings forecast errors, we include all forecasts for which actual earnings were available until the end of See Mateos (2007) for a comprehensive review of name-based ethnicity classification schemes. 13

14 The name-related scoring system is based upon a comparison between the list of our analysts full names and several reference lists, containing first names and surnames which are common or typical for a specific ethnic group. At first, we divide each analyst s full name into several parts, namely: first name, middle name, surname prefix (e.g. nobiliary particles), surname and second surname. 8 The decomposition of analysts full names into these individual parts helps us to extract as much information as possible for determining the ethnic background of each analyst. Apart from an analyst s full name, we also have information about the broker the analyst works for. We figure out in which language area each broker is headquartered and also evaluate this information, although to a lesser extent than the analyst s full name. Next, we gather data on first names and surnames, in order to compile reference lists for comparison purposes. Data on names, at least for some of the predefined ethnic groups, can be obtained from national statistic offices, social insurance agencies and similar institutions. 9 Noteworthy here is particularly the list of Frequently Occurring Surnames from Census 2000, published by the US Census Bureau. The list contains almost 150,000 surnames that were recorded at least 100 times in US census A useful feature of the census list is the linkage between demographic attributes and surnames frequency, enabling us to extract English-sounding surnames and Spanish-sounding surnames. 10 Unfortunately, official data on names is not published for each country or each ethnic group we defined. However, to obtain reference lists for those countries / ethnicities as well, we use unofficial data from websites dealing with genealogy and family history. Some of these websites derive surname frequencies at country level from telephone listings or from personal information of 8 To avoid any misunderstandings, consider the following full name of an analyst as an example: Jan Willem Van Kranenburg. The individual components of the full name are: Jan (first name), Willem (middle name), Van (surname prefix), Kranenburg (surname). An example for a double-barreled name is Echanove- Hernandez : Echanove (surname), Hernandez (second surname). 9 Other useful sources for name s data are name-related studies or projects with university background in general. For instance, the Brigham Young University offers instructional guidelines for analyzing historical records from several European countries. Some of these instructional guides contain lists of typical first names, surnames as well as surname prefixes. 10 For more details see 14

15 registered members. With regard to first names, official data sources often neglect diminutive forms and short forms. 11 Extending our reference lists with first names from unofficial data sources allows us to account for variations in first names in order to make a more precise assignment. A complete list of each of the sources that were used to compile the reference lists can be seen in the Appendix (A1). Altogether, the reference lists contain 16,397 first names and 264,788 surnames (including double entries). On the other hand, we have 8,777 analysts with 2,605 different first names and 6,602 different surnames. Using MS-Excel, we finally compare each part of an analyst s full name with the corresponding full name s part on each reference list. If an analyst s full name contains a middle name (second surname) it is compared with each first name (surname) presented on the reference lists. Whenever there is a match between a part of an analyst s full name and a corresponding name on our reference lists, an analyst receives two points for that particular ethnic group. For example, if an analyst has a first name, a middle name and a surname he/she can be awarded a maximum of six points for that ethnic group, since the full name consists of three different parts and each part is awarded two points. Consequently, an analyst having only a single first name and a single surname can score in maximum four points. In order to reach the full score for a specific ethnic group, each part of an analyst s full name must appear on the reference list of the same ethnic group. An analyst can automatically score one extra point if he/she works for a broker that is headquartered in a language area which corresponds to one of the predefined ethnic groups. If, for example, an analyst is employed by an Italian broker, he/she scores one extra point with respect to the Italian ethnicity. 12 Many first names 11 The first name Bill, e.g., occurs 23 times in our dataset, is typical English-sounding and derived from William. While William appears in official statistics, its short from Bill does not appear. 12 We are aware of the fact that the location of a broker is no direct indication for the ethnicity of an analyst. However, due to the enormous variety of names, we cannot assign each single first name as well as each surname to a specific ethnicity. We consider broker s location merely as additional clue for the most likely ethnicity of an analyst. In no case an analyst is assigned to a specific ethnic group on the basis of the location of 15

16 and many surnames are widespread among more than one ethnic group. Thus, in most cases, an analyst scores points for more than one ethnicity. An analyst is finally assigned to that ethnic group in which the highest score is achieved, unless the maximum score is less than three. The last restriction ensures that an analyst is not assigned to a specific ethnic group solely based upon a single piece of information. Instead, the analyst is automatically classified to Others. In rare cases, where the maximum score is equal to three, we make manual changes due to misleading assignments. Sometimes, the maximum score was ambiguous and an analyst could have been assigned to several ethnic groups. In this case, we check manually, to establish the country where the surname is most common. 13 The following figure illustrates the functioning of the name-based classification scheme on the basis of five examples: [Insert Figure 1 about here] 3.3. Measures of analysts location and analysts ethnic origin After each analyst has been assigned to an ethnicity and we have figured out in which country each broker is headquartered, we match both sets of information with the location of the companies covered. For this purpose, we define two different dummy variables that are of prime importance for testing our three hypotheses. First, we define a dummy variable (LocalBroker) equal to the value of one if an analyst works for a broker that is headquartered in the same country as the covered company, and zero otherwise. If LocalBroker is equal to one (zero) we denote the analyst as local analyst ( non-local analysts ) and the brokerage house as local broker ( foreign broker ). 14 Second, we create a dummy variable (Ethnicity) that takes on the value one if the ethnicity (proxied by the analyst s full name) of an analyst the respective broker. We use the information about broker s location exclusively in conjunction with namerelated information that is weighted more strongly. 13 We used the following website: to check surnames frequency per country. 14 Unfortunately we do not know exactly in which country an analyst is actually located. In line with Conroy et al. (1997) and Bacman and Bolliger (2003), we use the country in which the broker is headquartered as best possible approximation for analyst s location. It is reasonable to assume that in the majority of cases the location of the analyst coincides with the headquarter of the broker. 16

17 corresponds to the country in which the company covered is located, and zero otherwise. For instance, if an analyst classified as Spanish covers a company located in Spain (Japan), Ethnicity is equal to one (zero). In this context, it is irrelevant whether the analyst is employed by a Spanish broker or not. For analysts classified as English ( German ) Ethnicity equals one if either a company from the UK or the US (Germany or Switzerland) is covered. As already mentioned in the introduction, we denote an analyst whose ethnicity corresponds to the country of the covered company as domestic analyst. Conversely, if an analyst s ethnicity and a company s country do not fit together, the analyst is denoted as a foreign analyst. [Insert Table 2 about here] Table 2 briefly summarizes the information from this section. Panel A gives information about the number of analysts per ethnicity and about the number of brokers and companies per country. According to our name-based classification scheme, the majority of analysts are English (4,803). Similarly, most brokerage houses and companies are located either in the US or in the UK. Panel B shows the resulting number of reports from our sample (broken down by country) that are attributable to local and non-local analysts and domestic and foreign analysts, respectively. The first two columns of Panel B reveal that the number of analyst reports is almost evenly distributed between local brokers (333,551) and foreign brokers (395,575). Nevertheless, there are considerable differences between the individual countries. For example, analysts of Japanese brokers cover mainly non-japanese companies, whereas analysts of US brokers focus to a greater extent on US based companies. The last two columns of Panel B indicate that domestic analysts account for 72.4% or 528,151 of the total reports, whereas the number of reports by foreign analysts is 200,975 or 27.6%. 17

18 3.4. Control variables In addition to the above mentioned variables, we deploy a set of control variables that could also influence an analyst s research quality. We distinguish between control variables at analyst-, broker- and company level. Bolliger (2004) analyzes how analysts individual characteristics affect forecast accuracy. He finds a negative relation between the number of countries an analyst follows and the precision of his/her earnings forecasts. We proxy analyst s portfolio complexity by using the number of companies (Workload_Companies) and the number of countries (Workload_Countries) an analyst covers in a given year. Moreover, we control for analyst s reputation. Stickel (1992) shows that analysts from the All-American Research Team provide more accurate earnings forecasts and exert stronger impact on stock prices than other analysts. We use analyst rankings published by Thomson Reuter / StarMine to account for the influence of reputation. StarMine awards financial analysts for their estimation- and stock picking-skills. The rankings are published regularly in the financial press and are thereby open to a wide public audience. We define a dummy variable equal to one (StarAnalyst) if the report was issued by an analyst who was named in the rankings of StarMine in the year prior to the publication of the report and zero otherwise. With regard to the broker level, Clement (1999) emphasizes that analysts from large brokers can draw on better administrative and technical resources and may even have better access to private management information. Our proxy for broker size is the number of employed analysts per broker and per year (BrokerSize). At the company level we control for size as well by using the natural logarithm of the market capitalization (CompanySize) on the day prior to the publication date of the analyst s report. According to Garcia-Meca and Sanchez- Ballesta (2006), large companies pursue a more transparent and a more extensive corporate disclosure policy than small companies. For this reason it can be assumed that analyst s 18

19 research quality is positively correlated with the size of the company. Another control variable that is closely linked with company size is the number of analysts covering a specific company in a given year. Lys and Soo (1995) consider analyst coverage as an additional proxy for the information environment of a company as well as a proxy for the competition between analysts. The greater the number of analysts covering a given company, the higher are potential learning effects from other reports. The simultaneous use of analyst coverage and company size is problematic due to a high correlation between both variables. We follow the approach of Hong et al. (2000) and run an auxiliary regression of analyst coverage on CompanySize and use the residuals (Coverage) for our analysis. As reported by Ivkovic and Jegadeesh (2004), earnings forecasts and stock recommendations are revised more frequently after companies announced their earnings. Major company events can have a considerable impact on the stock market, making it difficult to distinguish the impact of an analyst report that was preceded by such an event. To absorb the impact of major company events, we therefore define a dummy variable (EventImpact) equal to one, if an analyst report was published between the second and the sixth day after the announcement of a major company event and zero otherwise. 15 Moreover, we include the price-to-book ratio (PTBV), at the day prior to publication, in order to account for differences between value stocks and glamour stocks. Finally, we control for prior stock return (Performance) and prior stock volatility (Volatility), each measured relative to the publication day of the analyst report. The stock return is measured over the previous three-month period, whereas the volatility is measured over the previous one-year period. The calculation of the volatility is based on the standard deviation of daily returns. An overview of all variables used in this study can be found in the appendix (A2). 15 We consider the following company events: earnings calls, earnings releases, general meetings. 19

20 3.5. Summary statistic and correlation matrix [Insert Table 3 about here] Table 3 shows summary statistics for the dependent variables (Panel A) and for the independent variables (Panel B), defined in the sections above. As shown in Panel A, the mean market reaction reflected by five-days s is quite small for the overall sample, as positive and negative returns balance each other out. However, the minimum and the maximum of -20% and +17.4% suggest that analyst reports can exert a material impact on stock prices. The mean with respect to EPS_Error_rel indicates that the actual earnings on average fell short to analyst s forecasts by 8.6%. Since the absolute earnings forecast error (EPS_Error_abs) penalizes each deviation from the actual earnings, the higher mean of 20.6% is less surprising. As a comparison, Loh and Mian (2006) define the absolute earnings forecast error in a similar manner and display a mean of 22.6%. The mean relative target price forecast error and the mean absolute target price forecast error (TP_Error_rel, TP_Error_abs) are 19.6% and 37.7%, respectively. Bradshaw et al. (2013) detect a slightly higher absolute forecast error for target prices of 45%. Concerning the independent variables in Panel B, we find consistent with Ivkovich and Jegadeesh (2004) that earnings forecast revisions are more frequent than recommendation changes. The number of downgrades (R_Down) at 37,922, is slightly higher than the number of upgrades (R_Up), of 34,911. Analysts revise their earnings forecasts (EPS_Rev) on average by 1.1%, whereas the average target price change (TP_Rev) amounts to 0.9%. For comparison purposes, Asquith et al. (2005) report for a different sample period ( ) an average earnings (target price) revision of -1.0% (2.8%). With regard to the remaining control variables, we find among others, that an analyst on average covers 13 companies from 2 different countries per year. The number of employed 20

21 analysts per broker ranges from 1 analyst up to 258 analysts. A company is covered on average by 16 different analysts per year. The positive mean and median values for EPS_Yield and Implicit_Return support the finding that analysts create optimistic forecasts. For instance, the mean concerning Implicit_Return suggests that on average target price forecast are associated with an increase in the stock price of 6.7%. [Insert Table 4 about here] Table 4 displays pairwise correlation coefficients for the main variables that are used in the following analysis. The highest correlation among two variables used simultaneously in a regression model is between Performance and TP_Rev. The correlation coefficient is 0.42, indicating that analysts obviously adjust their target prices in response to prior stocks returns. As 0.42 is substantially less than perfect correlation and given that the remaining correlation coefficients are even lower, the problem of multicollinearity should not be an issue in our analysis. 4. Empirical results 4.1. Analysis of short-term market reaction (i) Univariate results According to our first (second) hypothesis, we assume that local (domestic) analysts outperform non-local (foreign) analysts in terms of market reaction. To gain a first impression whether the hypotheses are correct, we calculate for each analyst group mean cumulative abnormal returns in response to stock recommendation changes, earnings forecast revisions and target price revisions. Panel A in Table 5 contrasts mean s between local analysts and non-local analysts, whereas in Panel B, domestic analysts are compared with foreign analysts. In Panel C and Panel D, the sample is restricted to local brokers and foreign brokers, in order to compare whether differences in market reaction between domestic and foreign analysts 21

22 vary according to the location of the analysts. In general, we find that stock upgrades, upward revisions of both earnings forecasts and target prices, are on average, accompanied by positive s, whereas stock downgrades and both downward revisions of earnings forecasts and target prices are accompanied by negative s. [Insert Table 5 about here] As shown by Panel A, local analysts clearly outperform non-local analysts. Except for the stock upgrades (R_Up), the mean s are higher (in absolute terms) within the group of local analysts. For example, the average for an upward revision of the target price (TP_Rev_pos) of a local analyst is 1.41 % and just 0.96 % for a non-local analyst. The mean difference of 0.44 percentage points (pp) is statistically significant at the 1 % level. A target price revision in the opposite direction (TP_Rev_neg) leads on average to a of % in case of a local analyst and to % in case of a non-local analyst. Compared to target price revisions, the differences in s in response to earnings revisions are slightly smaller but nevertheless, highly significant. Turning to Panel B, we see that the market reaction is consistently stronger within the group of domestic analysts. Upgrades by domestic analysts lead on average to positive s, which are approximately 0.18 pp higher than those of foreign analysts. Downgrades (R_Down) in turn, lead to negative s, which on average are 0.16 pp lower than those of foreign analysts. Revisions of earnings forecasts, regardless of the direction, lead in absolute terms, to on average of 0.30 % higher s, when issued by a local analyst. With regard to target price revisions, the difference in market reaction between domestic and foreign analysts amounts to 0.35 % in absolute terms. The results in Panel C and Panel D reveal that domestic analysts outperform foreign analysts, irrespective of their location. The only notable exceptions are insignificant 22

23 differences in s for R_Up and R_Down within the subsample of local brokers (Panel C). Up to this point, the results appear to be in accordance with the hypotheses mentioned above. (ii) Multivariate results We next want to examine whether the results from the univariate analysis hold when we simultaneously account for a variety of explanatory variables that may also affect the market reaction. The starting point for the multivariate analysis is the following regression model: = β 0 + β 1 R_Up + β 2 R_Down + β 3 EPS_Rev + β 4 TP_Rev + β 5 Ethnicity (1) + β 6 LocalBroker + β 7 CompanySize + β 8 PTBV + β 9 BrokerSize + β 10 Workload_Companies + β 11 Workload_Countries + β 12 Performance + β 13 EventImpact + β 14 Volatility + β 15 StarAnalyst + β 16 Coverage + ε Throughout this chapter we use five-day cumulative abnormal returns () as the dependent variable. On the right-hand side of Equation (1), we add upgrades (R_Up), and downgrades (R_Down) as well as earnings forecast (EPS_Rev) and target price (TP_Rev) revisions. Based on prior literature we expect these variables to have a strong impact on the market reaction of investors (see, e.g. Asquith et al., 2005). We further add our set of control variables to rule out that our results are driven by differences in analyst, broker or company characteristics. Since we are not interested in measuring general differences in market reaction between countries and industry sectors or sample years, we incorporate dummy variables into our regression models to capture such fixed effects. 16 To ensure correct statistical inference, we calculate robust standard errors (White 1980). Based on Eq. (1) we specify an additional regression model which includes interactions between LocalBroker and R_Up, R_Down, EPS_Rev and TP_Rev, respectively. In equation form, the model can be written as follows: 16 The industry dummies are based on Datastream s INDM2 classification scheme comprising ten different sectors. 23

24 = β 0 + β 1 R_Up + β 2 R_Down + β 3 EPS_Rev + β 4 TP_Rev + β 5 R_Up LocalBroker (2) + β 6 R_Down LocalBroker + β 7 EPS_Rev LocalBroker + β 8 TP_Rev LocalBroker + β 9 Ethnicity + β 10 LocalBroker + + ε With the help of interaction terms we can easily measure differences in market reaction between local analysts and non-local analysts. For example, the interaction term R_Up LocalBroker from Eq. (2) measures the difference in between stock upgrades by local analysts and non-local analysts. If the interaction term is significant and takes on the same sign as the stand-alone variable (R_Up), this indicates a stronger market reaction to local analyst s upgrades. The stand-alone coefficient on R_Up can be interpreted as the change in s to stock upgrades by non-local analysts. Similarly, we specify a third model in which we interact Ethnicity instead of LocalBroker. The corresponding regression equation is: = β 0 + β 1 R_Up + β 2 R_Down + β 3 EPS_Rev + β 4 TP_Rev + β 5 R_Up Ethnicity (3) + β 6 R_Down Ethnicity + β 7 EPS_Rev Ethnicity + β 8 TP_Rev Ethnicity + β 9 Ethnicity + β 10 LocalBroker + + ε In contrast to Eq. (2), the interactions in Eq. (3) now capture differences in market reaction between domestic analysts and foreign analysts. The estimation results of the three models are displayed successively in columns (1) to (3), in Table 6. In column (4) and (5) we present reestimates of Eq. (3), for the subsamples of local brokers and foreign brokers, respectively. [Insert Table 6 about here] Refering to the base model in column (1), we find in line, with prior literature, positive coefficients on R_Up, EPS_Rev and TP_Rev. As we expected, the coefficient on R_Down is negative. All four variables are significant at the 1 % level, implying that investors regard each piece of information contained in an analyst s report as being valuable. Analogous to Asquith et al. (2005), we measure in absolute terms, a stronger market reaction to target price 24