Local versus foreign analysts’ forecast accuracy: does herding matter?

The aim of this paper is to compare the information and resource endowments possessed by diﬀerent analyst types, classiﬁed by both location and employment within the context of seven emerging Asian markets. Our results show that local analysts are more accurate than expatriate and global analysts when we consider all earnings forecasts. However, when we control for the segregated herding behaviour of analysts, we ﬁnd that herding local forecasts are one of the least accurate compared to other herding forecasts. By contrast, bold local (that is, non-herding local) forecasts are more accurate than all other bold forecasts. This suggests that the information endowment of bold local analysts is superior to the information and resource endowments of bold expatriate analysts. We show that the superior accuracy of bold local forecasts does not stem from business group aﬃliations, investment banking relationships, demand for local analysts’ services or the specialisation of analysts vis-` a-vis countries or sectors. We consistently ﬁnd that bold local analysts are better at assessing the earnings of the ﬁrm they forecast. Our results show that the prior documented advantage of local analysts in terms of forecasting accuracy is driven by the bold local analysts, with herding locals diluting this eﬀect. To the best of our knowledge, this is the ﬁrst study to explore the segregated herding behaviour of local, expatriate and global analysts, and its impact on relative forecast accuracy across analyst types.


Introduction
There is considerable debate regarding the relative quality of earnings forecasts produced by local and foreign analysts (see, e.g., Conroy et al., 1997;Bacmann and Bolliger, 2003;Bolliger, 2004;Orpurt, 2004;Bae et al., 2008b;Comiran and Siriviriyakul, 2019). 1 The majority of studies in the field argue that resident analysts (those located in the same country as the local firm) are more accurate than non-resident analysts.This is driven primarily by their local advantage in terms of their understanding of local customs, cultures, language and the relationship with local firms (the information endowment factor). 2 By contrast, non-resident analysts, especially globally operating analysts such as UBS, Merrill Lynch and JP Morgan Chase, benefit from superior skills at processing and extrapolating pertinent information when forecasting earnings that are enhanced by in-house training programmes and the team/network of professionals they work with (the resource endowment factor).However, to the best of our knowledge, no prior studies in the field have taken into account the fact that a specific earnings forecast can be formed simply by mimicking other forecasts (that is, by herding).The failure to account for this behaviour prevents the researcher from testing the true forecasting abilities of analysts and accurately comparing the information versus resource endowment factors.To address this concern, we control for the segregated herding behaviour of analysts and focus on bold (non-herding) forecasts as they are more likely to reflect the true ability of analysts to collect and process extant information.
1 Analyst types are defined differently in prior studies.Some studies such as Orpurt (2004) define analysts according to geographical location, so that local analysts are those employed in a brokerage house located inside a particular country, regardless of the origin of the brokerage house.Other studies, such as Conroy et al., (1997) and Bacmann and Bolliger (2003) for example, define analysts according to the origin of brokerage houses, so that local analysts are those employed in a local brokerage house.In this paper we take into account the location as well as employment effects when categorising analysts.Specifically, as stated in Appendix II of the paper, 'local' analysts are defined as analysts employed in a brokerage house located inside a particular country in which the covered firm's equity is traded.Resident foreign analysts are referred to as 'expatriate' analysts, and are employed in a foreign brokerage house operating in a country where the covered firm's equity is traded.Non-resident foreign analysts are referred to as 'global' analysts who work outside the country in which the firm's equity is traded and forecast earnings for local firms.The term 'resident' analysts refers to local and expatriate analysts, and the term 'foreign' analysts refers to expatriate and global analysts.The terms 'relative quality of earnings forecasts' or 'relative accuracy of earnings forecasts' are used to refer to the difference across analyst types.
There is well-documented evidence that analysts tend to herd when making earnings forecasts or stock recommendations (see, e.g., Welch, 2000;Gleason and Lee, 2003;Clement and Tse, 2005;Jegadeesh and Kim, 2010;Lee and Lee, 2015;Frijns and Huynh, 2018;Lin, 2018), and that such behaviour leads to improvements in the quality of forecasts (see, e.g., Ke and Yu, 2006;Salamouris and Muradoglu, 2010;Mira and Taylor, 2011). 3However, there are no studies in the field that examine whether the propensity to herd can be explained by the location and employment characteristics of analysts.Also, it is not clear whether this potential herding behaviour affects the accuracy of forecasts made by different analyst types.This knowledge will help investors to make more informed investment decisions by possibly relying on non-herding forecasts produced by a specific analyst type (e.g., Chang, 2010).It will also further assist brokerage houses to make strategically important and capitalintensive decisions on whether to enter new markets and the role of behavioural factors in the forecasting activity of the analysts they employ.Furthermore, this will advance the understanding of the academic community on the determinants of herding when producing forecasts and ultimately the role of different analyst types in disseminating firm-specific information.
This paper first explores whether the propensity to herd is explained by the location and employment characteristics of analysts and then it tests whether the segregated herding behaviour across different analyst types affects the accuracy of earnings forecasts produced by these analysts.We argue that bold forecasts reflect more accurately the information and resource endowments possessed by different analyst types.Hence, in order to perform an accurate comparison on the endowments held by different analyst types, we place greater emphasis on bold forecasts as herding forecasts are formed by mimicking the forecasts of others.
This study focuses on Asian countries as we believe that emerging countries possess a number of distinguishing features that makes them an ideal laboratory to explore the impact of herding behaviour on the accuracy of analysts' forecasts.First, as pointed out by Hsieh et al., (2011), the information environment in emerging markets is relatively opaque mainly due to weak reporting requirements, lower accounting standards, lax enforcement of regulations and costly information acquisition.These factors may force market participants, including financial analysts, to herd on the information of others. 43 The literature regarding whether (non-herding) bold forecasts are on average more accurate than herding forecasts provides mixed findings.Clement and Tse (2005) and Huang et al., (2017), for example, use US data and find that bold forecasts are more accurate than herding forecasts.However, Salamouris and Muradoglu (2010) and Mira and Taylor (2011) use UK data and show that bold forecasts are less accurate than herding forecasts.To the best of our knowledge, there are no prior studies exploring the impact of boldness on the accuracy of forecasts in an international or Asian setting.
Second, prior evidence also suggests the role of financial analysts in terms of explaining stock returns is far more prominent in Asia compared to the US.Covrig and Low (2005), for example, explore the role of financial analysts in equity valuation in Japan and show that the incremental contribution of analysts' forecasts to explain stock returns is 405 percent (in terms of the percentage increase in adjusted-R 2 ) over the accounting information.This compares to the 41 percent previously reported using US evidence by Amir et al., (1999).This indicates that the skill and expertise of analysts are more valuable in markets with poor financial disclosure, such as Asian markets.Third, Asia is one of the fastest growing economic centres with financial markets representing trillions of dollars (Walsh, 2014).Any market that large will attract the interest of international investors.Therefore, knowledge of the determinants of the relative quality of earnings forecasts produced by different analyst types vis-à-vis their proximity to the forecasted markets is important to investors.And, finally, there is only limited empirical work on herding behaviour in emerging marketssurprising given the greater tendency of investors and analysts to herd in such markets.In a review study on herding behaviour in financial markets, Bikhchandani and Sharma (2001) acknowledge the need for more empirical work exploring the herding behaviour of investors and financial analysts in emerging markets.This concern is also echoed by Dang and Lin (2016) when researching the herding behaviour of investors with heterogeneous information in emerging markets.The lack of attention is particularly surprising given the growing importance of emerging markets in the global economy (e.g., Kaminsky et al., 2001).
The main objective of this paper is to test the relative accuracy of different analyst types by taking into consideration their propensity to herd.We categorise analyst types as local, expatriate and global (see Section 3.2.3 and Appendix II for detailed definitions of each analyst type), and address the following three issues.First, we re-examine the relative accuracy of forecasts produced by the local, expatriate and global analysts (e.g., Bae et al., 2008b).Second, we identify whether some types of analysts are more likely to herd than other analysts.To the best of our knowledge, very few attempts have been made, if any, to explore the segregated herding behaviour of local, expatriate and global analysts.Indeed, while previous evidence suggests that analysts herd (see, e.g., Clement and Tse, 2005;Jegadeesh and Kim, 2010;Huang et al., 2017), it is not clear whether this behaviour tends to vary across different analyst types vis-à-vis their proximity to the forecasted market.Finally, we explore the relative accuracy of analysts but only focusing on bold forecasts as these reflect the true information and resource endowments possessed by the different analyst types.
The findings of this study show that local analysts are the most accurate analyst type in terms of their earnings forecasts, and expatriate analysts in emerging Asian markets are less likely to herd compared to other analysts.However, when the herding behaviour is controlled for, the herding local analysts emerge as one of the least accurate analyst types whereas bold local analysts are the most accurate analyst type.This indicates that bold local analysts have a higher information advantage than even bold expatriate analysts.It also shows that the prior documented advantage of local analysts in terms of forecasting accuracy is driven by the bold local analysts, with herding locals diluting this effect.The informational advantage of bold locals is not explained by the business group affiliations, investment banking relationships, demand for local analysts' services or the specialisation of analysts vis-à-vis countries or sectors.We consistently show that bold local analysts are better at assessing the earnings of the firms they forecast.
To the best of our knowledge this is one of the first studies to explore the relative propensity of analysts to herd by analyst types.Moreover, it is one of the first studies to rationalise the relative accuracy of analysts by taking into account behavioural factors.Analysts' forecasts have a decisive role in the stock market and we believe it is vital to be aware of the propensity of analysts to herd and how this tendency affects the accuracy of their forecasts.This evidence contributes to a better understanding by academics and practitioners of the determinants of the accuracy of the forecasts produced by analysts.
The rest of the paper is organised as follows.Section 2 provides a brief review of the literature and develops the hypotheses.Section 3 describes the data and the models and Section 4 presents the empirical results.Section 5 concludes the paper.

Hypotheses
We first examine the relative accuracy of local versus foreign analysts.Prior studies in the field find support for the geographic information asymmetry hypothesis (hereafter GIAH) according to which geographically proximate analysts produce more accurate forecasts due to their geographical advantage.For instance, Orpurt (2004) finds evidence that resident analysts (i.e., local and expatriate in terms of our definition) are more accurate compared to nonresident foreign analysts (i.e., global in terms of our definition) in seven European countries.Also, Conroy et al., (1997) examine the relative forecast accuracy of local analysts in Japan (Japanese brokerage houses) compared with expatriate analysts in Japan (Western brokerage houses operating in Japan), and find that local analysts produce more accurate forecasts than expatriate analysts.They argue that Japanese houses have 'a better gauge on year-end information to be announced by the firm they forecast' (p.30).However, Bacmann and Bolliger (2003) analyse the performance of local analysts for seven Latin American countries and find that foreign analysts outperform local analysts in these markets.
More recently, using an international dataset (32 non-US countries covering the markets that we consider), Bae et al., (2008b) find that there is no significant difference in the accuracy of forecasts produced by local and expatriate analysts.Nevertheless, the authors document that resident analysts (i.e., local and expatriate in terms of our definition) produce more accurate forecasts than non-resident forecasters.Sonney (2009) further tests the GIAH and shows that both geographical proximity between analysts and the firm they cover as well as superior knowledge about the country-specific factors are significant determinants of analysts' forecasting accuracy.Hence, there is strong but not unanimous support for the GIAH when exploring analysts' forecasting accuracy.
Before testing our main hypothesis, we re-examine the relative forecast accuracy of different analyst types using Asian markets.Given the mixed evidence in prior studies, the unknown relative advantage between the information and resource endowment factors in Asia, and the highly possible impact of herding behaviour, it is not clear which type of analyst is more accurate.We test the following null hypothesis: On average, there is no difference in forecast accuracy between local, expatriate and global analysts.
Next, in line with the main aim of this study, we focus on the relative forecast accuracy of different analyst types conditional on their herding behaviour.While the herding literature is vast (see, e.g., Devenow and Welch, 1996;Bikhchandani and Sharma, 2001;Hirshleifer and Teoh, 2003, for reviews of the theoretical and empirical issues associated with such behaviour), few, if any, studies have examined this issue within the context of analyst types.This is surprising as local analysts have good incentives to herd.Specifically, their lack of resource endowments may force them to mimic the behaviour of others.In so doing they may be able to produce more accurate forecasts (e.g., Mira and Taylor, 2011).In addition, local analysts in emerging markets are less confident compared to foreign analysts when making earnings forecasts.Orpurt (2004) argues that local analysts are more likely to be penalised for inaccurate earnings forecasts in terms of their career prospects and reputation.A further pressure comes from the market itself.Chang (2010) shows that local institutional traders in Taiwan herd around the expatriate recommendations at all horizons.If the local analysts are aware of this, they might be tempted to mimic the earnings forecasts made by what the market considers to be 'most reliable' and prevent their forecasts from deviating too much from the consensus forecasts.Given the reputational and market pressures, as well as the resource versus information endowments, we expect that the relative accuracy of different analyst types is affected by their herding behaviour, and we need to disentangle these effects.
In the spirit of the reputational principal-agent models of Scharfstein and Stein (1990) and Trueman (1994), we categorise the analysts into two types; those that possess and make use of their information and/or resource endowments and those that do not have these advantages or find it beneficial to ignore their endowments and issue forecasts close to the prevailing consensus forecast. 5The herding type behaviour is consistent with the information cascade models of Bikhchandani et al., (1992) and Welch (1992), where an economic agent with a small private information endowment imitates the actions of a superior agent (as revealed by the consensus forecast), as they cannot directly observe the private information endowment of the superior agent.
In the context of our study, in order to disentangle the effect of herding behaviour on the relative accuracy of different analyst types, we control for analysts' herding behaviour and focus on bold forecasts as the latter reflect the information and resource endowments possessed by analysts whereas the herding forecasts are formed by simply mimicking the forecasts of others.Our analysis is unique and differs from the traditional rationale for the relative performance of different analyst types.Specifically, previous studies explain the superior performance of resident analysts using the information advantage argument, whereby local analysts have an information advantage due to their familiarity with local languages, customs and cultures, and/or due to cheaper collection of information using their human network (Conroy et al., 1997;Orpurt, 2004;Bae et al., 2008b). 6In this study, however, we argue that local analysts' information advantage can only be directly compared with other analysts' information and/or resource-based advantage after controlling for their herding behaviourachieved by focusing on bold forecasts that reflect the true endowments possessed by analysts. 7 Based on our argument that expatriate analysts possess both information and resource endowment factors compared to a single-factor holding of local and global analysts, we expect bold expatriate analysts to produce more accurate forecasts compared to other bold forecasts.However, based on the reputational principal-agent models of Scharfstein and Stein (1990) and Trueman (1994) and the cascade models of Bikhchandani et al., (1992) and Welch (1992), we believe 5 This categorisation of behaviour can also be made based on the anti-herding and herding behaviour; see Sharma and Bikhchandani (2001) for an overview of the theoretical and empirical issues associated with such behaviour.

6
The superior performance of local analysts may also be explained by career-concern incentives.This is because, compared to global analysts, local analysts may be penalised more by their employer and investors for inaccurate forecasts.Hence, given the external information advantage of local analysts, their incentive to produce accurate earnings forecasts may be larger in order to secure their job.Orpurt (2004) documents that career-concern incentives, including reputation, compensation and security, may induce local analysts to produce superior earnings forecasts.In other words, to accurately compare the information and resources endowments possessed by analysts and differentiate analysts' ability in terms of collecting and processing information, we need to focus on bold forecasts only as the herding forecasts are formed by simply mimicking the forecasts of others.that bold local analysts possess and make use of a greater (private) informational advantage compared to bold expatriate analysts despite the fact that both types are resident analysts.Given the lack of prior evidence in the field and mixed theoretical expectations, we do not predict who is more accurate and instead we test the following null hypothesis: What factors and theories explain the potential differences in terms of relative forecasting accuracy by analyst types?Past intuition and evidence in the field show that the information advantage of bold local analysts may stem from at least four sources, such as business group affiliations, investment banking relationships, demand for local analysts' services and specialisation of analysts vis-à-vis countries or sectors.
Business group affiliations (chaebols) is a common feature in our sample of Asian countries.According to the information-sharing hypothesis (e.g., Lim and Jung, 2012) affiliated analysts tend to have a better understanding of the firms they produce forecasts for because of their proprietary information, personal ties and working relations with the employees, suppliers and customers of member firms.These closer relations may help local affiliated analysts to assess better firm-specific risks and other firm characteristics resulting in accurate forecasts.Hence, affiliated bold local forecasts may have an information advantage over other bold forecasts because they benefit from private information which is not easily available to non-affiliated bold analysts. 8It is possible that the differences between the accuracy of bold local forecasts and other bold forecasts is greater in the case of business group affiliations.
Investment banking relationships may be another reason for a greater information endowment possessed by bold locals.Lai and Teo (2008) show that Asian firms usually choose local underwriters when issuing new equity.Bold locals may have better access to firm-specific information to firms with investment banking ties (e.g., Jacob et al., 2008), which ultimately results in more accurate forecasts.Analysts working for lead or co-lead underwriters benefit from better access to management and richer private information. 9We argue that the differences between the accuracy of bold local forecasts and other bold forecasts are increasing with the investment banking relationships.

8
A downside to this argument is that affiliations create incentives that may lead to conflict of interest and ultimately biased forecasts.For example, prior studies found evidence of tunnelling and propping in business groups (e.g., Bae et al., 2008a) as well as risk sharing within business groups (Khanna and Yafeh, 2005).
Next, the demand for local analysts' services may result in greater information endowment possessed by local analysts due to greater efforts and funds allocated to their activity.Using an international sample (Asian countries included), Comiran and Siriviriyakul (2019) show that the local advantage in terms of forecasting accuracy disappears in American depositary receipt (ADR) firms.In terms of our hypothesis, this means that bold locals may produce more accurate forecasts for the non-ADR firms.This accuracy advantage may be driven by the greater interest of local investors in the local firms and ultimately higher demand for the services provided by the local analysts.Hence, we expect that the differences between the accuracy of bold local forecasts and other bold forecasts are increasing with the demand for local analysts' services.
Finally, the specialisation of analysts vis-à-vis countries or sectors may result in greater relative accuracy of forecasts for certain analyst types.Sonney (2009) shows that the organisational structure of brokerage houses affects the accuracy of the forecasts.He finds that country-specialised analysts are more accurate than sector-specialised analysts and this advantage stems from superior knowledge of country-specific factors, as well as the geographical proximity between analysts and the firms they follow.Hence, the advantage of the bold local forecasts may stem from their country specialisation, because analysts employed by local brokerage houses cover fewer countries and sectors (e.g., Bae et al., 2008b).This leaves the possibility that local brokerage houses are country specialists and this specialisation drives the difference between the accuracy of bold local forecasts and other bold forecasts.
All four potential explanations discussed above indicate that if we disentangle the effect of the herding behaviour on the relative accuracy of forecasts by analyst type, we expect bold locals to have a greater information endowment advantage compared to other bold forecasts.However, given the lack of prior evidence in the field and mixed theoretical expectations, this issue still remains an empirical question.
Finally, although it is not formally hypothesised, comparing the relative accuracy of herding forecasts between different analyst types is also of interest.If all herding analysts mimic other forecasts randomly, the accuracy of different analyst types is expected to be the same.However, the different analyst types may refer to different sets of prior earnings forecasts.Anecdotal evidence, for example, indicates that local analysts usually refer to the local databases containing forecasts made only by the analysts employed by the local brokerage houses. 10If this is the case, the herding local forecasts may be based on a sub-set of prior forecasts made by the local analysts only, whereas the herding expatriate forecasts may be based on a full set of prior forecasts.This setting is in line with our argument that expatriate analysts benefit from both information and resource endowments, including more comprehensive databases, whereas local analysts benefit from an information advantage only.As a consequence, the accuracy of forecasts may be systematically different for different analyst types with herding expatriates ultimately producing more accurate herding forecasts.

Sample selection
The analysis is based on earnings forecasts from the International Edition of the Institutional Brokers Estimate System (I/B/E/S) Detail History files.11Specifically, for each forecast we have the name of the company the forecast is made for, the forecast and actual earnings per share (EPS) and, respectively, their forecasting and reported date.We retain in the sample all the forecasts made by a particular analyst instead of focusing just on either beginning or end-of-year forecasts made by an individual analyst for a firm (Hong et al., 2000).This will allow us to capture the likely scenario that an individual analyst produces a mixture of bold and herding forecasts throughout the year.A vital requirement for this study is to know the name of the brokerage house an analyst works for in order to classify them into local, expatriate and global analysts.This information is available up to 2006 after which I/B/E/S reviewed their policy whereby, in order to protect the providers of specific forecasts, the name of the providers of forecasts is no longer made public.12Hence, given the data availability, we classify all the brokerage houses included in the sample prior to 2006 and then extrapolate this information post-2006. 13ae et al., (2008b) examine a large set of countries (32 non-US countries) and demonstrate that resident analysts are more accurate compared to global analysts and local analysts are as accurate as expatriate analysts.However, such comprehensive studies may overlook the idiosyncratic behaviour of analysts within a particular market environment.This suggests that focusing on a particular geographical region will alleviate some of the (behavioural) inconsistencies encountered in the previous studies.In order to control for the political, economic and cultural conditions, we purposefully focus on only one (important) economic region such as Asia.
For a country to be included in the sample, and given the restrictions in the data availability, we require it to be an emerging country in Asia with at least 100,000 annual earnings forecasts in the I/B/E/S database until 2006 fiscal year end. 14This leads to data covering seven countries: Hong Kong, India, Korea, Malaysia, Singapore, Taiwan and Thailand.We then set the sample period from January 1994 to December 2016. 15These selection criteria resulted in a sample of 1,040,416 forecasts.As noted in Appendix I, the sample of observations is reduced to 981,151 after dropping forecasts with missing control variables, such as forecasting horizon, firm size, number of industries, companies, countries, general and firm experience and brokerage size.From these we remove 10,931 forecasts with a horizon greater than 400 days (see Huang et al., 2017) and another 11,443 forecasts with missing actual EPS, forecasted EPS and/or a mean actual forecast error of all forecasts for firm j in year T equal to zero. 16As a result of these restrictions, the number of observations in the sample drops to 958,777.
We use boldness as our measure of non-herding (see details in Section 3.2.2).Forecasts are regarded as bold if they deviate by a large margin from the consensus forecast immediately prior to the analyst's forecast.In order to calculate the consensus forecast for an observation, we need at least two forecasts by different analysts in the 90 days prior to an analyst's forecast.Applying this criterion leads to 120,560 forecasts dropping out of the sample.Next, we classify the brokerage house into local, expatriate or global.Appendix I reports the sample composition separately for the 1994 to 2006 and 2007 to 2016 sub-periods to account for the changes in the disclosure related to the providers of forecasts by I/B/E/S.We have lost only 260 forecasts due to insufficient information regarding the classification of the brokerage houses as local, expatriate or global in the sub-sample of forecasts made between 1994 and 2006.Yet, as expected, the attrition rate is much higher in the 2007-2016 sub-period resulting in a drop of 115,685 observations.Finally, the three continuous variables (the forecast accuracy measure, the boldness measure and firm size) are winsorised at one percent each in the top and bottom tails. 17The final sample comprises 722,272 observations.18

Variables
We test the validity of our hypotheses using forecast accuracy as the dependent variable.Analyst-type variables, boldness and the respective interaction terms between analyst-type variables and boldness represent our test variables, or independent variables of interest.In all the regressions we also control for forecast-specific characteristics (horizon), firm-specific characteristics (market value of the firm), and analyst-specific characteristics (forecasting frequency, number of industries, companies and countries followed by an analyst each year, firm and general experience of an analyst as well as the size of the brokerage house an analyst works for).We further examine the validity of our second hypothesis by controlling for the implementation of the countryspecific Regulation Fair Disclosure, affiliation of the brokerage houses with the firms, the demand for local analysts' services and specialisation of analysts.Next, we confine attention to the forecast accuracy, boldness and analyst-type variables, with a description of the control variables provided in Appendix II.

Forecast accuracy
The forecast accuracy measure is given by the proportional mean absolute forecast error (PMAFE ijt ).This is defined as the ratio of the difference between the absolute forecast error made by analyst i for firm j at time t (AFE ijt ) and the mean absolute forecast error of all forecasts for firm j in year T (MAFE jT ), to the mean absolute forecast error.That is, where Here A jT is the actual earnings reported by firm j at the fiscal year end T. Following Bae et al., (2008b), in order to facilitate the interpretation of the calculated variable, we multiply it by minus one.Thus, the larger PMAFE ijt is, the more accurate the forecast.

Boldness
We use boldness (BOLD ijt ), which is the proportional mean deviation from the consensus, as our measure of herding.BOLD ijt is defined as the difference between the deviation of a forecast made by analyst i for firm j at time t from the consensus forecast (DC ijt ) and the mean deviation of all forecasts for firm j in the fiscal year T (MDC jT ), divided by the mean deviation.That is, where Here F ijt is the earnings forecast produced by analyst i for firm j at time t, whereas C jt is the consensus forecast calculated using forecasts by different analysts for firm j within 90 days of the forecast date t. n jT is the number of forecasts within firm j and fiscal year T. Note that we use the deviation from the consensus proportional to the mean deviation in order to ensure comparability across different countries and companies.Country indices are suppressed in all the above equations.
In practical terms, boldness is calculated as follows.First, the consensus forecast at time t when an analyst produces the forecast is calculated as the average of all forecasts made by other analysts in the 90 days prior to the forecast.Second, the forecast deviation from the consensus is then taken as the absolute value of the difference between the forecast and the consensus.Finally, after calculating the deviations for all the forecasts made for firm j in a fiscal year T, its mean is computed and then used to calculate the proportional mean deviation from the consensus (BOLD ijt ).This measure of boldness is motivated by Hong et al.,'s (2000) boldness score measure and by the definition of herding. 19Higher values of BOLD ijt correspond to more deviation from the consensus (bolder forecasts).In order to make a clear distinction between bold and herding forecasts, we generate three dummy variables such as DBOLD md , DBOLD t1t3 and DBOLD q1q4 .These variables equal one if BOLD ijt is greater than its median, the third tertile, and the fourth quartile respectively.These variables are equal to zero if BOLD ijt is equal to or smaller than its median, the first tertile, and the first quartile respectively.A forecast is defined as bold if DBOLD md , DBOLD t1t3 and DBOLD q1q4 , respectively, equal one and a herding forecast otherwise. 20If our boldness measure is related to forecast accuracy, the expectation is that if we compare the top and bottom of the BOLD ijt distribution only, we should observe more pronounced results compared to the more conservative approach when using the median as a cut-off point.

Analyst types
Following Bae et al., (2008b) and Chang (2010), we take into account both location and employment effects when categorising analysts.Specifically, local analysts are defined as analysts employed by a local brokerage house located inside a particular country in which the covered firm's equity is traded.Expatriate analysts are those working for a foreign brokerage house operating in a country where the covered firm's equity is traded.Finally, global analysts are defined as analysts who work outside the country in which the firm's equity is traded.We categorise each brokerage house in the sample into local, expatriate or global using at least one of the following methods: (i) screening the information on the brokerage houses' websites, (ii) emailing each country's relevant authorities/regulators that possess information about the relevant brokerage houses, and (iii) asking various professional acquaintances working in the financial industry in one of the countries in the sample to classify the brokerage houses. 21In most cases, a particular classification is cross-verified across at least two of these sources.In order to test our hypotheses, we generate two dummy variables, such as local and expatriate dummy variables.Local is a dummy variable which equals one if an analyst is local, and zero if it is expatriate or global.Expatriate equals one if the analyst is expatriate, and zero for local and global analysts.The impact of the global analysts is captured in the respective intercepts of the regressions.

Models
The H 1 0 hypothesis focuses on the unconditional difference in the relative forecast accuracy of analyst types and it is tested using an ordinary least squares (OLS) regression.The dependent variable in the OLS regression is the proportional mean absolute forecast error (PMAFE ijt ).The model includes the dummy variables for local and expatriate analysts and the control variables described in Appendix II; specifically: where Z ðnÞ ijt is the nth control variable, and ϵ ijt is a suitably defined error term.In Equation ( 5), γ 1 (γ 2 ) is the difference in PMAFE ijt between local (expatriate) and global analysts, and γ 1 -γ 2 is the difference in PMAFE ijt between local and expatriate analysts.
In line with the prior approach in the field (e.g., Clement, 1999;Malloy, 2005;Bradley et al., 2017), all the independent variables, except firm size, which is a firm-specific characteristic, are demeaned by the firm-year values.The standard errors are heteroscedasticity-consistent and double-clustered at the firm and analyst levels.
The H 2 0 hypothesis is then tested by augmenting Equation ( 5) such that our herding measure and the interaction terms between the herding measure and analyst types are added.Specifically, the augmented model takes the following form: An indicator variable DBOLD is used in Equation ( 6) rather than the continuous variable BOLD as our aim is to examine the differential effect of herding versus non-herding forecasts on forecast accuracy. 22Using DBOLD also alleviates the possible noise in the measurement of BOLD.

Sample composition
Table 1 reports the distribution of the sample across countries (Panel A) and time (Panel B).Panel A shows that Hong Kong is the most represented country in the sample with 160,492 out of a total of 722,272 forecasts (22.2 percent).This is followed by India (16.2 percent) and Korea (15.4 percent).The least represented country in the sample is Thailand with a total of 69,259 forecasts (9.6 percent).The 722,272 forecasts are made by 12,317 individual analysts employed by 279 brokerage houses producing forecasts for 4,430 firms across seven Asian countries.Expatriate analysts are the most representative in the sample with a total of 7,645 analysts being classified as expatriates (50.4 percent), 5,833 analysts as local (38.5 percent) and the remaining 1,690 analysts as global (11.4 percent).

Main results
Table 2 presents the results from the univariate analysis and provides the mean and median values of all variables used, and investigates whether there is any significant difference across analyst types in terms of forecast accuracy and their herding behaviour.The results reported in Table 2 show that expatriate analysts are significantly more accurate than local and global analysts, albeit the mean difference in the forecast accuracy between local and expatriate analysts is insignificant.The mean results support previous evidence in Bae et al., (2008b) that resident (i.e., local and expatriate) analysts are more We test hypothesis H 1 0 using an OLS regression of forecast accuracy on analyst classification dummies and a set of forecast-, firm-and analyst-specific characteristics.Table 3 reports the OLS regression results in Panel A and the Ftests of the coefficient differences in Panel B. The results reported in the first column of Table 3 show that local analysts are more accurate than expatriate analysts (positive γ 1 -γ 2 ) and also global analysts (positive γ 1 ) at the 5 percent level of significance or better.Thus, we do not find support for our first null hypothesis that, on average, there is no difference in the forecast accuracy between local, expatriate and global analysts.By contrast, our results show that locals are the most accurate analyst type and expatriate analysts are as accurate as global analysts (γ 2 is positive but insignificant).This result suggests that the information endowment factor possessed by the local analysts is superior compared to the information endowment factor possessed by the expatriate analysts.This means that location is not the only factor determining the information endowment possessed by the local and expatriate analysts.Local analysts are unconditionally more accurate than expatriate analysts despite both analyst types operating locally.In terms of the control variables, we find that the accuracy of the forecasts improves as the forecasting horizon shortens and also for smaller firms (Clement, 1999;Bolliger, 2004;Bae et al., 2008b;Hilary and Hsu, 2013).The accuracy of the forecasts increases with the number 23  Bae et al., (2008b) find that local and expatriate analysts are more accurate than global analysts.The coefficients and t-values reported in their study for local and expatriate analysts are very similar to each other.They use a sample of 32 countries, including the seven countries we look at, that are located across the world (Europe, Asia, Latin America and Australia).However, as discussed in the introduction, the difference between the resident and non-resident analysts seems to be influenced by the countries/geographical regions included in the sample.For example, local analysts in Europe are found to be more accurate than non-local (foreign) analysts (Orpurt, 2004).On the other hand, Bacmann and Bolligier (2003) find that local analysts are less accurate that foreign analysts in Latin American countries.Hence, it is vital to focus on a particular region (in our case Asia) to alleviate some of the behavioural inconsistencies encountered in the previous studies.of companies followed by an analyst and the size of the brokerage house that employs the analyst (Clement, 1999).However, analysts revising their forecasts more frequently and following firms in more countries are less accurate.
Next, although it is not formally hypothesised, we also test whether the herding behaviour varies across analyst type.Specifically, we estimate an OLS regression with a continuous dependent variable BOLD and three logistic regressions with binary dependent variables such as DBOLD md , DBOLD t1t3 or DBOLD q1q4 .Using DBOLD md (DBOLD t1t3 and DBOLD q1q4 ) the sample is partitioned according to the value of the boldness measure so that half (one third and one quarter) of observations with higher BOLD values are coded as bold forecasts and the remaining half (one third below the first tertile and one quarter below the fourth quartile) of observations are coded as herding forecasts, respectively.All four models include the dummy variables for local and expatriate analysts and the control variables used in Equation ( 5).Similar to the regression results reported in column (1) of Table 3, all the independent variables, except firm size which is a firm-specific characteristic, are demeaned by the firm-year values.The standard errors are heteroscedasticity-consistent and double-clustered at the firm and analyst levels. 24The results reported in columns ( 2)-( 5) of Table 3 show that local (negative γ 1 -γ 2 ) and global (positive γ 2 ) analysts are more likely to herd compared to the expatriate analysts and the results are qualitatively similar across the four measures of herding.In line with the univariate results, the coefficient γ 1 is insignificant showing that the herding tendency of local analysts is not significantly different compared to global analysts.Local and global analysts are as likely to herd and both of them are more likely to herd compared to the expatriate analysts.These results suggest that the lack of resource (information) endowment factors forces local (global) analysts to mimic other forecasts more than expatriate analysts.The lower propensity of expatriate analysts to herd is not surprising as they possess both information and resource endowment factors.In terms of the control variables, the coefficients on the forecast horizon, the number of countries followed by an analyst, analyst's general experience and also the size of the brokerage house employing the analyst are consistently positive and significant across all four models, indicating that analysts at the early stage of the forecasting horizon, employed by a large brokerage house, with more general experience and covering more countries produce bolder forecasts.However, analysts following larger firms and more companies produce less bold forecasts.
Hypothesis H 2 0 is tested by running OLS regressions of forecast accuracy on analyst classification dummies, the interaction terms between the latter and our herding measure and a set of forecast-, firm-and analyst-specific  The first column of this table reports the OLS regression results (with the respective t-values in parentheses) of the analyst forecast accuracy (PMAFE) on dummy variables signifying the local and expatriate analysts and a set of firm-, forecast-and analyst-specific characteristics.Hence, in column (1) we test the first null hypothesis that, on average, there is no difference in forecast accuracy between local, expatriate and global analysts.Column (2) reports the OLS regression results using the continuous measure of boldness (BOLD) as a dependent variable on a set of dummy variables signifying the local and expatriate analysts and a set of firm-, forecast-and analyst-specific characteristics.Columns (3)-( 5) report the logistic regression results using DBOLD md , DBOLD t1t3 and DBOLD q1q4 as dependent variables (with z-values in parentheses).DBOLD md (DBOLD t1t3 and DBOLD q1q4 ) is a firm-year-forecast-specific indicator, which equals one if BOLD, the proportional mean deviation from the consensus of a forecast, is greater than its median (the third tertile and fourth quartile respectively), and zero if smaller than its median (first tertile and first quartile respectively).All the variables, except Ln (MV), are demeaned by the firm-year values.See Appendix II for variable definitions.The t-values presented in parentheses are heteroscedasticity consistent and the standard errors are clustered by firm and analyst.H 0 : characteristics.Table 4 reports the OLS regression results and the F-tests of the coefficients (and coefficient differences) obtained from these regressions using DBOLD md , DBOLD t1t3 and DBOLD q1q4 as alternative measures of boldness.
The results reported in Panels A and B of Table 4 test the conditional difference in the relative forecast accuracy of analyst types.Now that the interaction terms are added in the regressions, a clear picture emerges.The results show that local analysts are the most accurate analyst type when focusing on the bold forecasts across all three measures of performance.Specifically, H 0 : γ 1 + γ 4 = γ 2 + γ 5 (bold local versus bold expatriate) and H 0 : γ 1 + γ 4 = 0 (bold local versus bold global) are rejected at the 1 percent level in Panel B of Table 4.We cannot reject H 0 : γ 2 + γ 5 = 0 at the 10 percent level of significance which suggests that expatriate bold forecasts are as accurate as global bold forecasts.The potential sources of the greater informational advantage of bold locals are discussed later in this section.In line with our expectations, we also note that the coefficients on the interaction terms between BOLD and Local reported in column (1) of Table 4 are smaller compared to the respective coefficients reported in columns ( 2) and ( 3).This is because we focus on the top and bottom of the BOLD distribution in columns ( 2) and (3).Hence, DBOLD appears to accurately capture bold and herding forecasts.It also shows that forecast accuracy is explained by the categorisation of forecasts into bold and non-bold and also by the actual magnitude of boldness.Overall, we show that despite the fact that both local and expatriate analysts work locally, the information advantage of bold local analysts is still bigger compared to that of bold expatriate analysts and bold global analysts when controlling for the behavioural factors.Nevertheless, our results also show that controlling for the herding behaviour yields different results when comparing the unconditional and conditional differences in the relative forecast accuracy of expatriate and global forecasts.Specifically, we find that unconditionally expatriates are more accurate than globals but this difference disappears in the conditional setting (i.e., bold expatriates are as accurate as bold globals).Hence, we reject H 2 0 that analysts' herding behaviour does not condition the relative accuracy of the earnings forecasts of local, expatriate and global analysts. 25 Although, given the aim of this study, we focus on the relative accuracy of bold forecasts, the relative forecast accuracy of herding forecasts between analyst types shows that expatriate analysts produce the most accurate herding forecasts.Specifically, H 0 : γ 1 = γ 2 , H 0 : γ 1 = 0 and H 0 : γ 2 = 0 in Table 4 report 25 We also estimate separate regressions for the crisis period (i.e., 1997-1998 and 2008) and non-crisis period; untabulated results show that the evidence in the non-crisis period is largely consistent with our main findings.However, the results in the crisis period show that bold global forecasts are the most accurate, suggesting that the resource endowment factor may dominate the information endowment factor when capital markets are unstable.the coefficient difference and significance of herding forecasts between local and expatriate, local and global, and expatriate and global, respectively.From these we can conclude that the herding forecasts produced by the expatriate analysts are the most accurate, whereas the herding forecasts produced by the local analysts are as accurate as the herding forecasts of global analysts.This is somewhat surprising as mimicking other forecasts randomly should make no difference between analyst types.One possible explanation is that different analyst types refer to different sets of prior forecasts when producing earnings forecasts.Anecdotal evidence, for example, indicates that local analysts usually refer to databases containing forecasts made only by the local brokerage houses.Thus, the relative accuracy advantage of herding expatriate forecasts compared to other herding forecasts may stem from the fact that forecasts produced by the herding expatriates are based on a full set of prior forecasts whereas the herding locals may have access to a sub-set of prior forecasts produced by the local analysts only.As a result, the herding expatriate forecasts are the most accurate herding forecasts.This finding also shows that the prior documented advantage of local analysts in terms of forecasting accuracy may be understated as herding locals are diluting this effect.
Finally, H 0 : γ 3 + γ 4 = 0, H 0 : γ 3 + γ 5 = 0andH 0 : γ 3 = 0 in Table 4 report the coefficient difference and significance between the bold forecasts and herding forecasts of local, expatriate and global, respectively.As expected, they are all negative and significant, indicating that bold forecasts are significantly less accurate than herding forecasts on average (Salamouris and Muradoglu, 2010;Mira and Taylor, 2011).The accuracy of all bold forecasts has increased after the implementation of the CS-Reg FD (γ 3before − γ 3after is negative and significant at the 5 percent level of significance) but we still find that bold locals are more accurate than other bold forecasts with no significant differences across the two periods ((γ 1+ γ 4 ) before − (γ 1+ γ 4 ) after is insignificant).Hence, business group affiliations do not explain our results.

Investment banking relationships
Next, we test whether the greater accuracy reported for bold locals compared to other bold forecasts stems from investment banking relationships.For this purpose we re-estimate Equation ( 6) augmented by including past underwriter and future underwriter dummy variables and using DBOLD md and DBOLD q1q4 as measures of herding. 27We create two types of underwriter dummies: fiveyear and more than five-year underwriter dummies. 28The five-year (greater than five year) past underwriter dummy equals one if the analyst belongs to the lead or co-lead underwriter for the firm's IPO or SEO which took place no more (more) than five years ago, and zero otherwise.The respective future underwriter dummies are defined in a similar fashion.
Results reported in Table 6 show that the regression coefficients on the future underwriter dummy variables are positive and significant at the 10 percent level or better across both horizons and measures of boldness, except the greater than five years pre-IPO/SEO when using DBOLD q1q4 as a measure of boldness.At the same time, the past underwriter dummy variables are insignificant across all horizons and measures of boldness.All the interaction terms between the affiliation dummies and our test variables are insignificant suggesting that underwriter affiliation does not explain the relative difference in the accuracy of forecasts by analyst type.Despite the explanatory power of the underwriter dummies, the regression coefficient on the bold local is still positive and significant at the 5 percent level across all the regressions reported in Table 6.A comparison of the respective coefficients on the bold local dummy variable reported in Tables 4 and 6 shows that future and past underwriter dummies do not have an impact on the difference between the bold forecasts.Hence, the difference between the accuracy of bold locals compared to other bold forecasts is not explained by investment banking relationships.

Demand for local analysts' services
To test this explanation, we generate a dummy variable which takes a value of one for the ADR stocks in our sample and zero otherwise. 29This dummy variable is then included in Equation ( 6) and also used to divide the sample into ADR and non-ADR firms, enabling a sub-sample analysis.The results reported in column (1) of Table 7 show that the ADR dummy is positive and significant which indicates that earnings forecasts made for ADR firms are more accurate compared to other forecasts.Further sub-sample analysis reported in columns ( 2) and (3) of Panel A as well as tests of coefficients (and their differences) in Panel B provide support for the findings that local analysts lose their advantage for ADR firms but they are still more accurate compared to other bold forecasts for the non-ADR firms.
To further test the argument that the advantage of bold locals is driven by the level of interest in local firms, we capture the latter using the proportion of analysts following a firm in a particular year who are local (PrLocalAnalysts).A higher percentage indicates higher interest in local firms.The regression results reported in column (4) of Table 7 show that the regression coefficient on the proportion of local analysts is negative and insignificant, and the DBOLD × Local coefficient is still positive and significant.When interacted with our variable of interest in column (5), we find that all the coefficients on the interaction terms are insignificant with the exception of the DBOLD × PrLocalAnalysts coefficient which is positive and significant at the 1 percent level.This shows that bold forecasts made for firms with a higher (2) (3) (4)   This table replicates our main regression model in the presence of dummy variables controlling for the past and future investment affiliations with the firm.As a consequence, these regressions are based on 928 firms for which we have the underwriter information from January 1989 to June 2020.The dependent variable in all the regressions is the analyst forecast accuracy (PMAFE).Boldness is measured using DBOLDmd in columns (1)-( 4) and DBOLD q1q4 in columns ( 5)-( 8).The affiliation of the underwriter with the firm is measured up to five years as well as more than five years pre-or post-IPOs.The five years past affiliation dummy equals one if the analyst belongs to the lead or co-lead underwriter for the firm's IPO or SEO which took place no more than five years ago, and zero otherwise.The greater than five years past affiliation dummy equals one if the analyst belongs to the lead or co-lead underwriter for the firm's IPO or SEO which took place more than five years ago, and zero otherwise.All the variables, except Ln (MV), are demeaned by the firm-year values.See Appendix II for variable definitions.The t-values presented in parentheses are heteroscedasticity consistent and the standard errors are clustered by firm and analyst.*** , ** and * denote significance at the 1, 5 and 10 percent levels, respectively (two-tailed test).
proportion of local analysts are more accurate compared to herding forecasts and this is the case for both local and expatriate analysts.A similar conclusion is drawn when partitioning the sample based on the median value of the proportion of locals in columns ( 6) and ( 7) of Table 7. 30 These results show that the accuracy advantage of bold locals is not explained by the level of interest in local firms.

Specialisation of analysts vis-à-vis countries or sectors
To test for this potential effect, we generate three dummy variables which capture the specialisation of the financial analysts, such as country, sector or absolute specialists. 31Untabulated results show that sector specialisation is rare in local brokerage houses, whereas this is a more frequent occurrence in the case of global brokerage houses. 32Next, we include the firm-year demeaned country, sector and absolute specialisation dummy variables in Equation ( 6).The results tabulated in column (1) of Panel A in Table 8 show that the regression coefficients on the country, sector and absolute specialisation dummy variables are all insignificant and the regression coefficient on DBOLD × Local is still positive and significant at the 1 percent level.We further re-run our main regression model by partitioning the sample into subsamples based on the specialisation of analysts.The results reported in columns ( 2)-( 7) of Panel A and also the three columns in Panel B show that we still find support for our results across all the sub-samples and there is no difference in the coefficients on the DBOLD × Local across the sub-samples ((γ 1+ γ 4 ) yes −  (γ 1+ γ 4 ) no are insignificant).Overall, our results are robust when controlling for the specialism of local, expatriate and global analysts.

Other potential explanations
Prior studies in the field identify other potential sources of the accuracy advantage of local analysts.First, Tan et al. (2001) show that mandatory IFRS adoption improved foreign analyst forecast accuracy.Hence, bold locals may be more accurate compared to other bold forecasts because they are better at understanding the national accounting standards.To test this argument, we rerun Equation ( 6) restricted to the post-2008 period, when firms in our sample are not required to report their financial statements using the national accounting standards.Untabulated results show that the regression coefficient on the bold locals is 0.038 and significant at the 5 percent level, which is comparable to the coefficient reported in column (2) of Table 4.This suggests that the advantage of bold locals is not driven by better understanding of the national accounting standards.
Second, bold local forecasts may be more accurate as they are produced closer to the announcement day.Although we control for the forecasting horizon in our main regression, to tease out this effect we interact forecasting horizon with the bold dummy, bold local and bold expatriate dummy variables and include these interaction terms in Equation ( 6).Untabulated regression results show that our findings are not explained by the forecasting horizon. 33 Third, Mayew (2008), using the transcripts from earnings conference calls, shows that managers discriminate among analysts by allowing more management access to more favourable analysts.In our case, it is possible that managers favour local analysts because they produce more favourable/ optimistic forecasts.Untabulated results, however, show that bold locals are less optimistic compared to other bold forecasts.These results are not in line with the discrimination of certain analyst types.
Finally, the accuracy advantage of bold locals may stem from the fact that local analysts are better than expatriate or global analysts.By design, analysts in our sample can produce bold as well as herding forecasts.Hence, if locals are consistently better than expatriates or globals, the herding local forecasts should also be more accurate than other herding forecasts.However, the results reported in Table 4 indicate that this is not the case: herding locals are as accurate as herding globals.Thus, our results cannot be explained by consistently better-quality forecasts produced by the local analysts. 33Untabulated regression results show that on average bold forecasts are more accurate for longer horizons but there is no difference between the local and expatriate bolds in this regard.This table replicates our main regression model when controlling for the specialisation of analysts as country, sector and absolute specialists.Analysts' specialisation is captured via three dummy variables which take a value of one if the analyst is classified as a country, sector or absolute specialist, and zero otherwise.We classify analysts as country (sector) specialists if their country (sector) Herfindahl Index (HI) is larger than 0.90 and their sector (country) HI is smaller than 0.90.Analysts with country and sector HIs above 0.90 are classified as absolute specialists.The dependent variable in all the regressions is the analyst forecast accuracy (PMAFE).Boldness is measured using DBOLDmd in all the regressions.The regression results reported in column (1) are based on the entire sample of 722,272 observations.The remaining columns report the regression results for the sub-sample analysis for the three categories of analysts.For example, the regression results reported in column ( 2

Discussion and conclusion
This paper examines whether analysts' herding behaviour has a significant impact on their forecasting accuracy.To start, we test our first null hypothesis by re-examining the relative forecast accuracy of local, expatriate and global analysts using a sample of seven Asian countries.Then, we briefly explore whether the propensity of analysts to herd varies by analyst type.In this paper we argue that in order to compare the relative forecasting accuracy of different analyst types, one should control for the herding behaviour of analysts and ultimately use the bold earnings forecasts as a barometer to evaluate the segregated forecasting ability of analysts.This is because herding forecasts are produced by simply mimicking other forecasts and do not reflect the true endowments possessed by different analyst types.Consequently, we finally test our second null hypothesis by examining the relative forecast accuracy of bold forecasts by analyst type.To the best of our knowledge this is the first study to investigate the relative forecast accuracy of local, expatriate and global analysts by taking into account analysts' herding behaviour.
Our results suggest that local analysts are the most accurate analyst type when accounting for their herding behaviour.Bold forecasts produced by the local analysts are more accurate compared to other bold forecasts.These results are consistent across different measures of boldness, samples and periods of time.This suggests that the information advantage of bold local analysts is bigger than that of bold expatriate analysts despite the fact that local and expatriate analysts are both resident analysts.We test four competing explanations aimed at identifying the source of the accuracy advantage of bold locals, such as business group affiliations, investment banking relationships, demand for local analysts' services and specialisation of analysts vis-à-vis countries or sectors.The results show that none of these hypotheses explain the accuracy advantage of bold local forecasts.
These results echo the conclusion drawn by Conroy et al., (1997) that local analysts in Japan, for whatever reason, have a better gauge of the information announced by the firm they forecast.Brown et al., (2015, p. 3) show that 'private communication with management is a more important input to analysts' earnings forecasts and stock recommendations than (their own) primary research, recent earnings performance, and recent 10-K and 10-Q reports'.It is possible that the four sources of information advantage considered in this paper do not capture this 'private communication channels' with the management of the firm.Future research potentially could explore the role of gender in producing bolder forecasts 34 (Kumar, 2010) and explore 34  Kumar (2010) shows, in the context of US firms, that female analysts issue bolder and more accurate forecasts and their accuracy increases with the market segments with lower female concentration.The author concludes that female analysts have better than average skill due to self-selection and financial markets recognise their superior abilities.further the exact source of information used by the analyst types.For example, Hu et al., (2008) show that information sources have a significant impact on analysts' information comprehension, analysing abilities and job quality in China.Specifically, analysts tend to exhibit greater information comprehension and better job quality when they conduct more company-level surveys.In addition, anecdotal evidence indicates that local brokerage houses may use different databases compared to foreign analysts when forecasting earnings.This type of analysis is beyond the scope of this paper.Importantly though, we show for the first time that herding behaviour of analysts is present in the Asian markets and this behaviour affects the accuracy of earnings forecasts.Therefore, it is vital for investors, regulators and academics to be cognisant of these potential biases, especially given the growing relevance and appeal of Asian emerging markets.It also offers useful evidence to brokerage houses when deciding about their location and the role of behavioural factors in the earnings forecasting of financial analysts.Our evidence shows that location matters but this needs to be considered in conjunction with behavioural factors.

I (continued) Variable Description
BrkSize it Number of analysts employed by the brokerage house in each year, adjusted by firm-year averages CS-Reg FD Dummy A dummy variable which equals one for the post-implementation period of the respective versions of Regulation Fair Disclosure in each country and zero otherwise.

Past Underwriter
A dummy variable which equals one if the analyst belongs to the lead or co-lead underwriter for the firm's past IPO or SEO, and zero otherwise.This dummy variable is calculated for two periods: for IPO or SEO which took place no more than five years ago and more than 5 year ago Future Underwriter A dummy variable which equals one if the analyst belongs to the lead or co-lead underwriter for the firm's future IPO or SEO, and zero otherwise.This dummy variable is calculated for two periods: for IPO or SEO which took place no more than five years and more than five year ago after the issue of the forecast by the analyst ADR Dummy A dummy variable which takes a value of one if I/B/E/S forecasted earnings are reported in US dollars and zero otherwise PrLocal Analysts The proportion of analysts following a firm in a particular year who are local Country/Sector/Absolute Specialist Dummy Analysts' specialisation is captured via three dummy variables which take a value of one if the analyst is classified as a country, sector or absolute specialist, and zero otherwise, adjusted by firm-year averages.We classify analysts as country (sector) specialists if their country (sector) Herfindahl Index (HI) is larger than 0.90 and their sector (country) HI is smaller than 0.90.Analysts with country and sector HIs above 0.90 are classified as absolute specialists.See section 3.1 in Sonney (2009) for further details on the calculation of HIs 7 ' herding behaviour does not condition the relative accuracy of the forecasts of local, expatriate and global analysts.
the coefficient difference between γ 1 and γ 2 and the corresponding F(χ 2 )-test results, with p-values in brackets.*** , ** and * denote significance at the 1, 5 and 10 percent levels, respectively (two-tailed test).© 2021 The Authors.Accounting & Finance published by John Wiley & Sons Australia, Ltd on behalf of Accounting and Finance Association of Australia and New Zealand 1166 Y.-S.Choi et al./Accounting & Finance 62 (2022) 1143-1185

©
2021 The Authors.Accounting & Finance published by John Wiley & Sons Australia, Ltd on behalf of Accounting and Finance Association of Australia and New Zealand Y.-S.Choi et al./Accounting & Finance 62 (2022) 1143-1185 OLS parameter estimates (with t-values in parentheses) of the analyst forecast accuracy (PMAFE) on dummy variables signifying the local and expatriate analysts, boldness indicator (DBOLD), and a set of firm-, forecast-and analyst-specific characteristics.Panel B reports results testing the coefficient restrictions based on F-tests, with p-values in brackets.All the variables, except Ln (MV), are demeaned by the firm-year values.See Appendix II for variable definitions.The t-values presented in round parentheses are heteroscedasticity consistent and the standard errors are clustered by firm and analyst.*** , ** and * denote significance at the 1, 5 and 10 percent levels, respectively (two-tailed test).© 2021 The Authors.Accounting & Finance published by John Wiley & Sons Australia, Ltd on behalf of Accounting and Finance Association of Australia and New Zealand and further tests of coefficients in Panel B.
) are based on the sub-sample of forecasts made by country specialists whereas the results in column (3) are based on the forecasts made by the noncountry specialists.All the variables, except Ln (MV), are demeaned by the firm-year values.See Appendix II for variable definitions.The tvalues presented in parentheses are heteroscedasticity consistent and the standard errors are clustered by firm and analyst.*** , ** and * denote significance at the 1, 5 and 10 percent levels, respectively (two-tailed test).© 2021 The Authors.Accounting & Finance published by John Wiley & Sons Australia, Ltd on behalf of Accounting and Finance Association of Australia and New Zealand Y.-S.Choi et al./Accounting & Finance 62 (2022) 1143-1185

©
2021 The Authors.Accounting & Finance published by John Wiley & Sons Australia, Ltd on behalf of Accounting and Finance Association of Australia and New Zealand © 2021 The Authors.Accounting & Finance published by John Wiley & Sons Australia, Ltd on behalf of Accounting and Finance Association of Australia and New Zealand Y.-S.Choi et al./Accounting & Finance 62 (2022) 1143-1185This table shows the number of forecasts, firms, brokerage houses and the number of local, expatriate and global analysts in each country (Panel A) and each year (Panel B).As a brokerage house (an analyst) can conduct research for several countries and for several years, the number of brokerage houses in the sample is different from the sum of brokerage houses (analysts) across seven countries and across sample periods.accuratethanglobalanalysts, and local and expatriate analysts are equally accurate.23Interms of herding behaviour, the BOLD statistics show that global analysts produce forecasts with the smallest deviation from the consensus, indicating that they herd more.The median difference in boldness for local and global analysts is insignificant suggesting that both local and global analysts herd more compared to the expatriate analysts.Consequently, this evidence shows that locals are more likely to herd compared to expatriates.The Pearson correlation coefficient (not tabulated) between forecast accuracy (PMAFE) and boldness (BOLD) is negative and significant at the 1 percent level (Pearson = −0.306).This leaves open the possibility that the relative forecast accuracy of different analyst types is distorted by the existence of herding forecasts.This issue is investigated next in the context of multivariate analysis.
© 2021 The Authors.Accounting & Finance published by John Wiley & Sons Australia, Ltd on behalf of Accounting and Finance Association of Australia and New Zealand Y.-S.Choi et al./Accounting & Finance 62 (2022) 1143-1185 This table contains the mean and median values of the variables used in the study, and test results for mean and median differences between different analyst types, such as Local, Expatriate and Global.The significance of the differences in the mean and median values (across the analyst types) is represented with p-values, by using Cochran t-test statistics and Wilcoxon rank sum test statistics (approximated z-statistics), respectively.See Appendix II for the definition of variables.

Table 5
OLS regression results: forecast accuracy controlling for business group affiliations

Table 5
This table replicates our main regression model in the presence of a dummy variable controlling for the implementation of the country-specific Regulation Fair Disclosure (CS-Reg FD).CS-Reg FD is a dummy variable which equals one for the post-implementation period of the respective versions of Regulation Fair Disclosure in each country and zero otherwise.The dependent variable in all the regressions is the analyst forecast accuracy (PMAFE), withDBOLDmd used to measure boldness.The regression results reported in the first two columns are based on the entire sample of 722,272 observations, whereas the last two columns report the robustness results using a sub-sample analysis.All the variables, except Ln (MV), are demeaned by the firm-year values.See Appendix II for variable definitions.The tvalues presented in parentheses are heteroscedasticity consistent and the standard errors are clustered by firm and analyst.*** , ** and * denote significance at the 1, 5 and 10 percent levels, respectively (two-tailed test).© 2021 The Authors.Accounting & Finance published by John Wiley & Sons Australia, Ltd on behalf of Accounting and Finance Association of Australia and New Zealand

Table 7
OLS regression results: forecast accuracy controlling for the demand for local analyst services © 2021 The Authors.Accounting & Finance published by John Wiley & Sons Australia, Ltd on behalf of Accounting and Finance Association of Australia and New Zealand Y.-S.Choi et al./Accounting & Finance 62 (2022) 1143-1185

Table 7
This table replicates our main regression model when controlling for the demand for local analyst services.ADR Dummy takes a value of one if I/ B/E/S forecasted earnings are reported in US dollars and zero otherwise.PrLocalAnalysts is the proportion of analysts following a firm in a particular year who are local.The dependent variable in all the regressions is the analyst forecast accuracy (PMAFE).Boldness is measured using DBOLDmd in all the regressions.The regression results reported in the first column are based on the entire sample of 722,272 observations.The sample is then partitioned into ADR and non-ADR firms in columns (2) and (3).The regression results reported in columns (4)-(7) are estimated on our sample after excluding firm-years with zero local analysts following the firm.The last two columns report the robustness results using a sub-sample analysis.All the variables, except Ln (MV), are demeaned by the firm-year values.See Appendix II for variable definitions.The tvalues presented in parentheses are heteroscedasticity consistent and the standard errors are clustered by firm and analyst.© 2021 The Authors.Accounting & Finance published by John Wiley & Sons Australia, Ltd on behalf of Accounting and Finance Association of Australia and New Zealand 1178 Y.-S.Choi et al./Accounting & Finance 62 (2022) 1143-1185 *** , ** and * denote significance at the 1, 5 and 10 percent levels, respectively (two-tailed test).

Table 8
OLS regression results: forecast accuracy controlling for specialisation of analysts