On the Structure of Analyst Research Portfolios and Forecast Accuracy

Article (PDF Available)inJournal of Accounting Research 47(4):867-909 · September 2009with64 Reads
DOI: 10.1111/j.1475-679X.2009.00338.x · Source: RePEc
Abstract
ABSTRACT This study provides insights into the forces and constraints that shape analyst research coverage along country and sector dimensions and the impact of the structure of an analyst's portfolio on forecast accuracy. We find that analyst specialization by country and sector is sensitive to the extent to which firms "within" a country or sector and firms "across" country-sectors are exposed to common economic forces, the potential for revenue generation, and broker culture. Our tests indicate that existing research on the relation between analyst portfolio structure and forecast accuracy may suffer from an endogeneity bias. We use our analysis of analyst specialization to develop controls for this bias. Once we employ these controls, we find that country diversification is associated with superior forecast accuracy. However, the relation between sector diversification and forecast accuracy is context-specific. Specifically, sector diversification enhances forecast accuracy in an international context, while it detracts from forecast accuracy in a domestic U.S. context. Copyright (c), University of Chicago on behalf of the Accounting Research Center, 2009.
On the Structure of Analyst Research Portfolios and Forecast Accuracy
Omesh Kini
Robinson College of Business
Georgia State University
e-mail: okini@gsu.edu
Shehzad Mian
Goizueta Business School
Emory University
e-mail: shehzad_mian@bus.emory.edu
Michael Rebello
A.B. Freeman School of Business
Tulane University
e-mail: mrebello@tulane.edu
Anand Venkateswaran
College of Business
Northeastern University
e-mail: anand@neu.edu
Latest Draft: November 2006
We would like to thank Sudipta Basu, George Benston, Lawrence D. Brown, Diane Denis, Karl
Lins, Mike Lemmon, Daphne Lui, Stan Markov, Grace Pownall, Raghu Rau, Jeremy Stein, and
Gregory Waymire. Discussions with Ashish Bhutani (Lazard), Mathew Chapin (UBS Warburg),
Devashish Chopra (formerly with Citi-SSB Asset Management), and Dan Meade (formerly with
Credit Suisse First Boston) have given us valuable insights from a practitioner’s perspective. We
thank IBES for providing us permission to use their database and Ajay Negi of IBES for
patiently answering all our questions regarding IBES International. We acknowledge valuable
research support from Terry Bennett and Ron Harris. The usual disclaimer applies.
On the Structure of Analyst Research Portfolios and Forecast Accuracy
ABSTRACT
This paper provides insights into the determinants of the structure of analyst research portfolios,
and then investigates the impact of portfolio structure on forecast accuracy. Specifically, we
study the forces and constraints that shape analyst research coverage along country and sector
dimensions. We find that analyst specialization is sensitive to the extent to which firms within a
country or sector are exposed to common economic forces, the potential for revenue generation,
and broker culture. Further, in contrast to the existing research on the relationship between
analyst portfolio structure and forecast accuracy, we find that complex research portfolios
promote forecast accuracy once we control for the endogeneity in portfolio choice. Overall, our
results are consistent with brokerages organizing analyst research coverage in a pragmatic
manner. While maintaining a focus on the economic viability of the research production
process, the organization of brokerage research is consistent with producing high quality
research.
JEL Classification: G24; G15; L23
Keywords: International Analyst Coverage, Brokerage Organization, Forecast Accuracy, Capital
Markets
On the Structure of Analyst Research Portfolios and Forecast Accuracy
Recent research has documented that U.S. analysts frequently cover stocks from
multiple sectors, and international analysts often follow stocks belonging to both different
countries and sectors (see, e.g., Clement and Tse (2005), Boni and Womack (2005), and Sonney
(2005)). Recognizing that this diversity in research portfolios can affect analysts’ research
output, researchers have attempted to assess its effect on analyst performance as captured by
forecast accuracy (see, e.g., Clement (1999), and Clement, Rees, and Swanson (2003)). Existing
studies, however, have treated the structure of analysts’ research portfolios as an exogenous
variable which may cause their results to be subject to a selection-bias given that analyst
research portfolios are the result of choices made by analysts and their employing brokerages
(see, e.g., Clement (1999), and Gilson, Healy, Noe, and Palepu (2001)).
In this paper, we examine the impact of portfolio structure on analyst forecast accuracy
while controlling for selection biases resulting from portfolio choices made by
analysts/brokerages. To accomplish this, we first examine the determinants of one aspect of
analyst organization—the choice between either single- versus multi-country coverage or
single- versus multi-sector coverage. We then use this analysis as a means of controlling for
self-selection in our examination of the relation between forecast accuracy and portfolio
diversity in the country and sector dimensions.
We employ a comprehensive sample of international (non-U.S.) analysts from 53
countries over the period 1996 through 2002. We focus on international analysts for the
following reasons. First, there is limited evidence on both the determinants of research portfolio
structure and forecast accuracy for international analysts. This omission in the literature
becomes all the more glaring once we consider the fact that international analysts account for
2
approximately 60% of all analysts in the IBES database. Second, by using international analysts
we add a dimension that would be missing in an analysis of only U.S. analysts since analysts in
our sample can diversify by following firms either across countries or sectors.
To examine the choice of single- versus multi-country (single- versus multi-sector)
coverage, we estimate binomial logistic regressions to explain specialization of analyst
portfolios by country (sector). We find that increased commonalities among stocks in a country,
as proxied by common variation in returns, increase the likelihood that analysts will specialize
by country. Similarly, increased commonalities among stocks in a sector promote sector
specialization. We also find that higher country equity market capitalization increases the
likelihood of country specialization, while a larger total equity market capitalization of a sector
in a particular country (country-sector market capitalization) increases the likelihood that
analysts will cover firms in a single sector. Finally, we find that international analysts
employed by a U.S. brokerage are more likely to specialize by sector and less likely to specialize
by country.
We then consider the impact of portfolio choice on forecast accuracy. To establish a
benchmark, we first follow the extant literature and treat portfolio structure as an exogenous
variable in our tests. We find that neither country diversification nor sector diversification has
an effect on forecast accuracy. These results are similar to those documented by Clement, Rees,
and Swanson (2003) in their investigation of Canadian, German, Japanese, and U.K. analysts.
Next, we investigate the relation between forecast accuracy and portfolio organization by
employing a two-stage Heckman methodology that controls for research portfolio choice.
1
The
results from this analysis support our central thesis that self-selection can have a significant
1
See, for example, Heckman (1978; 1979), Heckman and Robb (1986), and Greene (2000) for a general
approach to estimate a causal parameter in the face of selection bias. Also see, Li and Prabhala (2005) for a
survey on the use of self-selection models in finance research.
3
impact on analyst forecast accuracy. Further, once we employ the Heckman procedure to
control for endogeneity in analyst portfolio choices, we find that diversified portfolios boost
analysts’ forecast accuracy. This is the case whether we consider portfolio diversification across
either countries or sectors. These findings stand in contrast to the results in earlier studies
which document that, when analyst organization is treated as an exogenous variable, portfolio
diversification either has no impact or detracts from forecast accuracy (see, e.g., such as those of
Clement (1999), Clement, Rees, and Swanson (2003), and Clement and Tse (2005)).
Our explanation for these results is as follows: A focused portfolio should enable an
analyst to enjoy economies of scale in information acquisition and production. The time and
effort savings resulting from these scale economies could be transferred to gaining a more
thorough understanding of the companies she follows and, thus, enable the analyst to construct
more accurate forecasts. By switching to a diversified portfolio, an analyst will lose some scale
economies but gain exposure to complementary information that is needed to research
companies in other countries or sectors. By accessing and analyzing this complementary
information, the analyst may obtain a more complete picture of the economic forces at play, and
thereby generate more accurate forecasts. The benefit from accessing complementary
information is likely to dominate the resulting loss in scale economies when the possibility of
harnessing scale economies by maintaining a focused portfolio is relatively low. Thus, analysts
will tend to follow diversified portfolios when the cost of diversification, i.e., the loss of scale
economies, is relatively low. This is precisely what our results seem to suggest as (i) the
diversified portfolios in our sample boost forecast accuracy and (ii) analysts tend to follow
diversified portfolios precisely when the common variation in returns, our proxy for the
presence of scale economies, is relatively low for stocks belonging to the same country (sector).
4
Our results provide several new insights into the organization structure of brokerages
and the influence of this structure on analyst performance. First, our results support the notion
that brokerages organize their research production in a pragmatic fashion. Consistent with
revenue generation being essential for the existence of sell-side research, we find that the
decision to specialize by country or sector is determined by the potential for investment
banking business and/or brokerage commissions.
2
However, we also find that brokerages
structure their research production processes in a manner that is conducive to producing
quality research by allowing analysts to follow diversified portfolios and consequently sacrifice
the economies of scale in information acquisition and production resulting from this
diversification precisely when the cost of the lost scale economies is relatively low. This ensures
that, on average, diversified research portfolios outperform focused research portfolios.
Our paper relates to and has implications for several streams of research. It is closely
related to the literature that examines how the organizational structures of financial
intermediaries evolve in response to their information processing environment. For instance,
Berger et al. (2005) find evidence that the nature and availability of information regarding firms
leads large banks to focus on lending to large firms, while smaller banks target smaller firms.
They attribute this difference in focus across banks of differing sizes to the cross-sectional
relationship between the mix of hard and soft information and firm size. In a similar vein, we
document that specialization of sell-side analyst research portfolios is influenced by information
availability on the firms that analysts research. When efficiency in gathering and processing
country-specific (sector-specific) information is increasingly vital, analysts tend to focus their
coverage on an individual country (sector).
2
Chen (2003), for example, argues that the separation of investment banking from sell-side research to
solve the conflict of interest problem may make sell-side research infeasible; thereby diminishing
information production in the economy and reducing social welfare.
5
Recent studies have shown that analysts’ forecast accuracy is systematically related to
the geographical proximity of analysts to the companies they research. For example, Bacmann
and Bolliger (2001) provide evidence indicating that foreign analysts provide more accurate
forecasts than local analysts. In contrast, Bae, Stulz, and Tan (2005), Bolliger (2001), Malloy
(2005), and Chang (2003) document that locally based analysts make more accurate forecasts
than their counterparts who are based abroad.
3
Our analysis complements this literature as it
focuses on the “proximity”, as measured by the extent of exposure to common economic factors,
of the companies covered by the analysts to each other rather than the geographical proximity
of the analysts to the companies they cover. This difference in perspective provides us with a
rationale that can potentially reconcile the conflicting results in the above cited studies. For
example, if a good understanding of sector factors is crucial to forecasting accurately, a foreign
analyst who focuses on a sector may have an advantage over a local analyst who focuses on
local firms across multiple sectors. The reverse would be true if an understanding of country
(local) factors is the key determinant of forecast accuracy. This suggests that comparisons of the
performance of local and foreign analysts should control for differences in the industrial
structure among countries.
Finally, and most importantly, we directly contribute to the literature examining the role
of analyst portfolio complexity on forecast accuracy by examining the forces and constraints
that drive analysts to select complex portfolios (see, for example, Clement (1999) and Clement,
Rees, and Swanson (2003)). In doing so, we provide insight into the cross-sectional variation in
analysts’ costs and benefits from diversifying across sectors and countries. Further, our analysis
3
There exists a parallel literature on geography and investment performance. Kang and Stulz (1997),
Choe, Kho, and Stulz (2000), and Coval and Moskowitz (2001) find that proximity improves performance,
while Grinblatt and Keloharju (2000) and Seasholes (2004) document the opposite result.
6
highlights the need to control for self-selection in portfolio choice when studying the relation
between forecast accuracy and analyst portfolio structure.
The remainder of the paper is organized as follows: Section I provides the underlying
hypotheses that are tested in this paper. Section II contains a description of our data sources and
sample selection screens. Section III contains our analysis of the determinants of analysts’
portfolio choice along both country sector dimensions. In Section IV, we examine the impact of
portfolio complexity on forecast accuracy. Section V presents results of tests designed to
provide further insight into our central results and to assess their robustness. The paper
concludes with some final observations in Section VI.
I. Development of Testable Hypotheses
In this section, we first develop hypotheses related to the determinants of analyst
organization. The predicted relations with analyst organization are summarized in Table I.
Subsequently, we develop hypotheses relating analyst organization to performance as
measured by the accuracy of analysts’ current fiscal year-end forecasts.
I.A. Determinants of Analysts Research Portfolio Structure
Information Efficiency Hypothesis. Sets of firms operate in common markets and, thus, common
forces influence their costs and revenues. These commonalities across firms allow an analyst
who has studied a firm to expend less time and effort to determine the value of other firms
exposed to the same economic forces. Thus, by focusing her attention on a set of firms that are
strongly influenced by a common set of forces, an analyst is able to harness economies of scale
in the acquisition and production of information. By taking advantage of these scale economies,
an analyst may be able to study firms in greater depth and thus produce more accurate earnings
forecasts.
7
Less focused portfolios may also promote forecast accuracy because a less focused
portfolio, while resulting in the loss of scale economies in information acquisition and
production, may expose the analyst to alternate sources and types of (complementary)
information regarding the firms she covers. For example, an analyst covering utility companies
may benefit from the information and knowledge gained from the coverage of the oil and
natural gas industry. Similarly, an analyst covering the ship manufacturing industry may be
better positioned to assess the size of the potential order books of ship manufacturers if she also
covers firms in the shipping industry.
The extent of focus in analyst portfolios is likely to be influenced by this tradeoff
between the relative information-based benefits of focus and diversity. At the margin, however,
an increase in commonalities will strengthen the incentives for an analyst to specialize. Thus,
we hypothesize that an analyst is more likely to confine her research portfolio to a set of firms as
the strength of commonalities across the set of firms increases. We focus on two sources of
commonality—shared country of domicile and shared industrial sector membership—and
hypothesize that an analyst will tend to pick a country-focused portfolio as country-based
commonalities increase. Similarly, as sector-based commonalities increase, the likelihood of a
sector-focused portfolio improves. In our analysis, we decompose equity market returns in a
manner similar to that in Heston and Rouwenhorst (1994) and Griffin and Karolyi (1998) to
estimate the extent of commonalities between firms belonging to the same country
(VARCTYTOT or VARCTYSYS) or sector (VARSECTOT or VARSECSYS). A description of the
construction of these variables is presented in the next section.
Revenue Generation Hypothesis. Market size may also influence the structure of an analyst’s
research portfolio. Analysts have to ensure that the information they produce generates
sufficient revenues for their employer. These revenues take the form of brokering commissions
8
and investment banking fees, both of which are a function of the size of an analyst’s portfolio.
Thus, a smaller market is likely to increase the cost of focusing coverage on the market and,
thus, discourage an analyst from restricting her entire coverage to the market. Thus, we
hypothesize that specialization is more likely when market size is larger.
4
In our analysis, the natural logarithm of equity market capitalization (CTYMV) serves as
a proxy for market size. The likelihood of country specialization, therefore, will increase with
the country’s equity market capitalization. Global sector market capitalizations are large and,
thus, are unlikely to be binding constraints on portfolio choice. Sector capitalizations within a
country are, however, relatively small, and are likely to act as constraints on sector
specialization for analysts who choose to restrict their coverage of firms to a single country.
Thus, the likelihood of sector specialization will increase with the natural logarithm of the
market capitalization of a sector within a country, CTYSECMV.
Broker Culture Hypothesis. It is argued that there are two well-established and distinct research
models—the European and U.S. models—that have evolved in the context of very different
domestic markets (see, e.g., Rubino (2003)). U.S. broking firms are reputed to have attempted to
export their sector-focused organization structure as they have expanded internationally,
encouraging their analysts to focus on a single sector. Conversely, European firms are reputed
to lean toward a country-focused organization, biasing their analysts towards focusing their
research coverage on a single country (See, for example, Reuters Institutional Investor Research
Group (2002)). Analyst research portfolios are likely to reflect the organizational culture and
resources of their employing brokers. Therefore, we hypothesize that an analyst’s research
focus is influenced by her employer’s traditional research organizational structure. Analysts
4
This reasoning echoes Stigler (1951) who identifies the extent of the market as a limiting factor on
specialization.
9
employed by European brokerages will display a greater tendency to specialize by country,
while analysts employed by U.S. brokerages are more likely to specialize by sector. In our
analysis, we use a broker’s domicile as a proxy for its research culture and traditional
organization structure. More specifically, we employ a dummy variable, BRORIGIN that takes
the value of one if the broker is a U.S broker, and zero otherwise.
Brokers can also vary in the opportunities that they afford an analyst to specialize. We
hypothesize that the opportunities for specialization are positively related to broker size.
5
Larger brokers cover a larger number of stocks and are, therefore, less likely to constrain an
analyst portfolio away from specialization along any given dimension. We use the number of
analysts employed by the brokerage firm, BRSIZE, as a proxy for broker size.
Analyst Experience Hypothesis. Just as a pecking order theory of capital structure argues that a
firm’s capital structure is the culmination of its history of security issuance choices, analyst
portfolios may be the result of a history of portfolio decisions. Experienced analysts are more
likely to have changed employers and their past employers may have had different research
cultures. As a result, they are more likely to maintain research portfolios that contain diverse
stocks. Further, more experienced analysts by definition have been in the industry for a longer
period of time, and, as such, are likely to have demonstrated superior ability in the past to be
still counted amongst the survivors in the industry.
6
Since following stocks across sectors
and/or countries may require the ability to conduct complex analyses, brokerages will assign
diverse research portfolios to these analysts in keeping with their higher perceived ability.
5
Larger professional firms, such as auditors and law firms, are often associated with more specialization
in their practice (see, for example, DeAngelo (1981), Garicano and Hubbard (2005), and Zhou and Elder
(2003))
6
Consistent with this line of argument, Hong and Kubik (2003) and Clement and Tse (2005) find that U.S.
analysts who make more accurate earnings forecasts face better employment outcomes. Kini et al. (2006)
document similar results for international analysts.
10
Further, if more experienced analysts indeed have greater ability then we should also observe a
positive relation between forecast accuracy and analyst experience. Our proxy for analyst
experience, ANALEXP, is the number of days between when an analyst first appears in IBES
until her last forecast for a given fiscal year.
I.B. Portfolio Structure and Forecast Accuracy
As we have argued earlier, the structure of an analyst’s portfolio is likely to impact her
performance as measured by the accuracy of her current fiscal year-end earnings forecasts. A
focused portfolio should enable an analyst to enjoy economies of scale in information
acquisition and production, which in turn, can enable the analyst to make more accurate
forecasts. Alternatively, by diversifying her portfolio, an analyst gains exposure to
complementary information. By incorporating this complementary information in earnings
forecasts, the analyst may be able to improve her accuracy. Since the economies of scale and
informational complementarities arguments affect forecast accuracy in opposite directions, the
predicted relation between forecast accuracy and analyst specialization is ambiguous.
Existing research examines the relation between forecast accuracy and analyst portfolio
complexity by estimating regressions that treat the structure of the analysts’ portfolios as being
exogenously determined. However, the structure of an analyst’s portfolio may be the result of
choices made by the analyst’s employing broker and/or the analyst herself. Further, there may
be latent factors that affect both the structure of analyst portfolios and their forecasting ability.
For example, it is possible that brokerages prefer to assign diversified portfolios to analysts who
are also responsible for supervising other analysts, placing these supervising analysts in a better
position to evaluate and monitor the activities of analysts covering different types of firms.
These administrative duties, however, may adversely affect the forecasting abilities of the
11
supervisors. Thus, to the extent that the structure of an analyst’s portfolio is determined by
choices made by the analyst, tests of the relationship between the structure of an analyst’s
portfolio and her forecast accuracy should account for the endogeneity of the portfolio structure.
In our analysis, we do so by using the Heckman methodology that allows us to account for the
self-selection bias.
II. Description of Data Sources, Analyst Portfolios, and Factor Extraction
A. Data Sources for Analyst Portfolios and Firm Information
The primary sources of data for this study are IBES International, Datastream International,
and Nelson’s Directory of Investment Research (Nelson’s). From the Details file of IBES International,
we obtain the list of all analysts and stocks that they cover over our sample period (January 1,
1996 through December 31, 2002).
7
We divide our sample period into one-year intervals for
computation of analyst portfolios and other variables. Next, we assign each stock to a unique
country and sector pair for each year in our sample in accordance with the IBES Sector Industry
Group file.
8
We supplement IBES with Nelson’s to identify the employing broker for every
7
Analysts are excluded from our sample if either their name is missing, or IBES did not assign them an
analyst code, or IBES assigned analyst names that are names for industry groups like Health Care or
Pharmaceuticals. Our conversations with Mr. Ajay Negi at IBES suggest that when broker firms are small
or analyst turnover is high, they sometimes report their earnings forecasts using sector names rather than
individual analyst names. Another possible explanation is that these types of names are used for a sector
or country team.
8
We deleted firms that belong to the “Miscellaneous/Undesignated” sector in IBES because, despite
research into these firms, we could not easily classify them into a specific IBES sector. In some instances,
firms are assigned by IBES either to multiple sectors, multiple countries, or both. For firms assigned by
IBES to multiple countries, we checked other sources to identify the location of the company’s
headquarters and accordingly assigned each firm to a particular country. We classify firms that IBES
assigned to multiple sectors to the sector to which IBES assigned them for the majority of any given year
in our sample period.
12
analyst.
9
We then construct a research portfolio for every analyst that includes all stocks for
which they issued any forecast (quarterly, semi-annual, current-year annual, following-year
annual, long-term earnings growth, etc.) during the year in consideration. Our measure of
performance is, however, based solely on the accuracy in annual earnings forecasts for the
current fiscal year. Analyst forecasts included in our sample correspond to the last forecast
issued by an analyst at least one month prior to the fiscal year end, and earnings realizations for
each fiscal year for stocks in our sample are obtained from the IBES database. Furthermore, we
trim the sample at the 1 percentile and 99 percentile levels of raw forecast accuracy. We obtain
stock prices, number of shares outstanding, stock returns, and equity market capitalization data
from Datastream International.
10
The resulting sample consists of 19,379 analysts following
11,292 stocks over the seven-year sample period across 53 countries and 11 sectors, resulting in
47,049 analyst-year observations.
B. Description of Analyst Portfolios
Table II provides summary statistics for our analyst sample. The mean (median) number
of countries followed by an analyst is 1.46 (1.00) indicating that the analysts’ coverage appears
to be predominantly focused on a single country. Analysts however tend to cover multiple
9
In instances in which IBES assigns multiple broker codes to the same broker, we consolidate all broker
codes for a given broker using a unique broker code. These unique broker codes are then employed as
the basis for all broker-analyst linkages examined in our paper. Even after the multiplicity of broker codes
are dealt with, some analysts were assigned to between 2 and 6 brokers by IBES. This can either be due to
changes in employment or because IBES continued linking the analyst with previous employers. To
resolve multiple broker associations, we assign analysts to the broker who employed the analyst for the
maximum period during a two-year window around the period within which the analyst’s affiliation is
an issue. Thus, an analyst who appears in the sample for the year 1998 is assigned to the broker with
whom he was employed for the longest time between January 1, 1998 and December 31, 1999.
10
We delete firms from the sample if their average market capitalization was less than zero, if there was
no price, market value, or returns data for them during the sample period, or if they were classified in the
“Miscellaneous/Undesignated” sector in IBES. We also delete analyst-stock-year observations if they
had missing forecast accuracy information.
13
sectors and the mean (median) number of sectors being followed is 2.36 (2.00). The mean
(median) number of firms followed by analysts is 8.41 (6.00), and the mean (median) number of
days of analyst experience, as measured by the number of days from when the analyst first
appears on IBES to the time period in which the analyst makes a forecast for a firm in our
sample during a given fiscal year, is 841 (606) days.
To gain an understanding of how analysts organize their research portfolios, we develop
two classifications for the research coverage of each analyst. First, each analyst is classified
either as a single-country analyst or multi-country analyst. Second, each analyst is classified as
a single sector or multi-sector analyst. Panels B and C of Table II describe analyst portfolios
conditional on the two classifications described above. The vast majority (78.00%) of our
observations are for single-country analysts. In contrast, only about 40% of the analysts in our
sample cover firms in a single sector. Single-country analysts tend to follow fewer firms (mean
value of 8.27 versus 8.91) than multi-country analysts. Analyst experience is approximately
four months lower for single country analysts. A similar pattern is apparent when examining
analyst portfolios based on the split between single sector and multi-sector coverage—analysts
who cover multiple sectors tend to follow more firms and have greater experience.
C. Description of Variables Hypothesized to Impact Portfolio Choice
Note that, while each analyst-stock-year observation in our sample is unique, each
analyst enters our sample multiple times based on the number of stocks she follows. This allows
for the possibility that regression estimates employing analyst-stock-year observations may be
influenced disproportionately by analysts who cover larger portfolios. Consequently, we
employ analyst-year observations in the subsequent analysis. To arrive at the sample that is
employed to test our hypotheses, for each time period, every variable is averaged across stocks
covered by the analyst. For example, if x
ijt
is a variable of interest for analyst i and stock j in time
14
period t, in our analysis we employ the average value, x
it
, of this variable across stocks covered
by the analyst for time period t.
These averages of variables across stocks covered by an analyst for a period are taken in
two ways. First, we average each variable across all stocks covered by the analyst in a given
period. The second method for averaging takes into consideration the fact that, while an analyst
may cover a diverse set of stocks, her primary responsibility may be producing research on a set
of homogeneous “core” stocks that belong to a single sector within a given country. Thus, we
also average across stocks belonging to the “core” of an analyst’s portfolio during that period,
where core stocks are stocks that belong to the unique country-sector combination that accounts
for the largest portion of the analyst’s research portfolio in that year. All other stocks are
considered non-core stocks. For all the results reported in the paper, the core of an analyst’s
portfolio is the country-sector that accounts for the largest number of stocks in the analyst’s
portfolio (CORE). The sectors defined above are based on IBES International’s 11 sectors.
11
In Table III, we present summary statistics for all the variables employed in modeling
analyst organization. The first two rows present statistics for our measures of exposure to
common economic forces for firms belonging to the same country, while the next two rows
present statistics for our measures of exposure to common economic forces for firms that belong
to the same sector. These proxies for the exposure to common economic forces are based on the
common variations in stock returns along country (sector) dimensions. Following Heston and
Rouwenhorst (1994) and Griffin and Karolyi (1998) we decompose the variance of the weekly
returns for each firm i in country j and sector k,
2
i
σ
, as follows:
11
In results not reported in the paper, we replicate our entire analysis by defining the core of an analyst’s
portfolio as the country-sector that accounts for the largest fraction of the market capitalization of the
analyst’s portfolio. Our results are not sensitive to these alternative definitions of an analyst’s core.
15
22222
ikjGi
εγβ
σσσσσ
+++= ,
where
2
G
σ
represents the variation in the global factor,
2
j
β
σ
(
2
k
γ
σ
) represents the variation in the
country (sector) factor for the country j (sector k) to which firm i belongs, and
2
i
ε
σ
represents the
variance of the firm-specific component of returns. The details of the decomposition
methodology are presented in Appendix A.
We develop two proxies each for the influence of country and sector factors on firm
returns. The proxies for the influence of the country factor on a firm’s returns are normalized
values of
2
j
β
σ
for each firm. VARCTYTOT is obtained by normalizing the weekly return
variation on the country factor by the total weekly firm return variability
(
)/
22
ij
σσ
β
.VARCTYSYS is obtained by normalizing the weekly return variation on the country
factor by the sum of the variances of the weekly returns on the global, country, and sector
factors
)/((
2222
kjGj
γββ
σσσσ
++
), i.e., the proportion of total “systematic risk” attributable to the
country factor. The first row in Table III indicates that, on average, the country factor accounts
for 19.9% of total variation in the returns on core stocks in an analyst’s portfolio. On the other
hand, when the average value of VARCTYTOT is computed using all stocks in the analyst’s
portfolio, the country factor accounts for approximately 20.1% of return variation. Similarly, the
mean value of VARCTYSYS is 55.3% when only stocks belonging to the core of an analyst’s
portfolio are considered and 55.7% when all stocks in an analyst’s portfolio are considered.
The variable VARSECTOT (VARSECSYS) is the sector counterpart of VARCTYTOT
(VARCTYSYS) in that it measures the proportion of total risk (systematic risk) of firm i
attributable to the sector factor. From Table III, it is clear that the sector factor accounts for a
smaller fraction of variation in stock returns than does the country factor. When core stocks in
16
an analyst’s portfolio are considered, the variable VARSECTOT has a mean (median) value of
6.6% (4.3%), while VARSECSYS has a mean (median) value of 32.0% (19.5%). Both variables
have similar mean values when all stocks belonging to the analyst’s portfolio are considered.
Every stock in an analyst’s portfolio belongs to a particular country. The variable
CTYMV represents the mean value of the country market capitalization for stocks in the
analyst’s portfolio. Similarly, CTYSECMV represents the mean value of the country-sector
market capitalization for stocks in the analyst’s portfolio. The figures reported in Table III
suggest that market capitalizations are larger for core countries and core country-sectors. The
mean value for BRORIGIN indicates that 24.3% of analysts in our sample are employed by
brokerages of U.S. origin. BRSIZE is a count of the number of analysts employed by the broker
that employs a given analyst. The mean value of BRSIZE is approximately 156 and its median
value is 107, suggesting that a significant fraction of the analysts in our sample are employed by
relatively large brokers. Our measure of experience, GENEXP represents the number of days
from when the analyst first appears on IBES to the time period in which the analyst makes a
forecast for a firm in our sample during a given fiscal year. The mean and median values of this
variable suggest that, on average, analysts in our sample have been employed in the profession
for approximately two years. Note that the mean (median) values for BRSIZE, GENEXP, and
BRORIGIN are the same for all stocks and core stocks in analysts’ portfolios since they are
measured at the analyst level.
The remaining variables in the table are employed as controls in our examination of
analyst portfolios. We now turn to variables employed as controls for the information
environment in which analysts operate. VARRESTOT, (
22
/
ii
σσ
ε
) the proportion of idiosyncratic
variation in stock returns, allows us to control for the relative strength of factors other than
country and sector factor that might influence the complexity of an analyst’s task. FIRMMV, the
17
average market capitalization of an individual stock covered by an analyst in a given year, is
employed as a proxy for the disclosure environment. From Table III, it appears that the market
capitalizations of firms in the core of an analyst portfolio are similar in size to non-core firms
suggesting that the disclosure environment may not vary dramatically between the core and
non-core portion of an analyst’s portfolio.
ACCTG INDX, is an index of country-level disclosure regulation. A higher value for this
index indicates greater information disclosure in that country. It is a count of how many of 90
standard accounting items are included in the annual reports of firms in a country. We
obtained this information from International Accounting and Auditing Trends published by the
Center for International Financial Analysis and Research. Table III suggests that firms in an
analyst’s portfolio belong to countries that, on average, require approximately 69 out of the 90
possible accounting items in annual reports and that these disclosure requirements are similar
for core and non-core firms.
Improved safeguards for investors should attract investors who initiate trades based on
fundamentals (Morck, Yeung, and Yu (2000)), and thereby enhance the value of analyst
research.
12
To control for the level of minority shareholder protection in a country, we include a
variable ANTIDIR RIGHTS in our regressions. The source for this data is La Porta et al. (1997)
and the variable represents the average value across an analyst’s portfolio of the La Porta et al.
variable.
13
The mean (median) value for this variable is approximately 3.4 (4.0) out of a
12
Minority protection will tend to encourage greater ownership dispersion and, therefore, greater capital
market liquidity. Enhanced liquidity is likely to generate greater trading volume and higher commission
income. This effect will render more focused portfolios economically viable.
13
ANTIDIR RIGHTS indicates how many of the following mechanisms prevail in a given country: (1)
shareholders are allowed to mail their proxy vote (2) shareholders are not stipulated to submit their
shares before a general shareholders’ meeting (3) cumulative voting is allowed (4) there is a system in
place to address issues pertaining to oppressed minority shareholders (5) the minimum ownership
18
maximum of five mechanisms to protect minority shareholder rights for core stocks and all
stocks in an analyst’s portfolio. The variable, CTYMVGDP is a proxy for the degree of capital
market development of the country in which the firm is domiciled, and is our final control
variable for the level of disclosure. It is the ratio of stock market capitalization to GDP, and has
been employed to capture informational availability by other researchers (See, Rajan and
Zingales (1998)). The data to compute this variable for each year is obtained from various issues
of the World Stock Exchange Fact Book. The mean (median) values for CTYMVGDP appear to be
similar for core and non-core stocks in an analyst’s portfolio.
III. Analyst Organization
In this section, we present results from our logistic regression analysis modeling analyst
portfolio choices. First, we examine the determinants of single country versus multiple country
research portfolios. We then present results from our logistic regression analysis modeling the
determinants of single sector versus multiple sector research portfolios. These tests employ
analyst-year level observations. As a consequence, while an analyst appears only once for each
time period, an analyst may enter the sample several times over our sample period. Since
analysts may appear in the sample multiple times, it opens up the possibility that forecast errors
may be correlated and t-statistics may be overstated because of the “cluster sample” problem
(see, e.g., Wooldridge (2002)). To control for this issue, we employ adjusted standard errors that
account for the possible correlations between forecast errors for an analyst. Our adjustment,
however, is made under the assumption that forecast errors across analysts are independent
threshold for a shareholder to call an extra-ordinary shareholders’ meeting is no greater than ten percent.
Thus, this index will lie between zero and five.
19
(see, e.g., Huber (1967), Rogers (1993), White (1980), and Wooldridge (2002)).
14
In addition, in
our analysis of this panel data, we control for year-fixed effects in the estimated regressions.
15
One way of looking at the choice of an analyst to diversify across countries and/or
sectors is to ask the question what factors cause her to follow stocks outside of her core country-
sector, i.e., stocks in countries outside her core country or stocks in sectors outside her core
sector. We attempt to answer this question by both studying the impact on portfolio choice of (i)
the characteristics of just the stocks in her core country-sector and (ii) the characteristics of all
the stocks in her research portfolio. Thus, each model of analyst portfolio choice is estimated
two ways. One estimate employs independent variables values that are averaged across all
stocks in an analyst’s core while the second estimate employs independent variable values that
are averages across an analyst’s entire portfolio.
III.A. Analysis of Single versus Multiple Country Coverage
Table IV presents tests of our hypotheses regarding analyst specialization by country.
We employ a binomial logistic regression approach to model the probability of being organized
as a multiple-country versus single-country analyst.
16
The first two models reported in the
table are identical with the exception of the variable that represents the country factor--
VARCTYTOT is employed in the first model and VARCTYSYS in the second model. All
independent variables employed in these two regressions are average values across all stocks in
the analyst’s core country-sector for a given period. The third and fourth models reported in
14
See Gleason and Lee (2003) for a similar approach to dealing with cluster sample problems. Further,
we also employ adjusted standard errors that account for the possible correlations between forecast errors
for analysts that work for the same brokerage. Our results are qualitatively similar to those reported in
this section.
15
Since we use analyst-year observations, we cannot control for country and sector fixed effects since
many analysts cover stocks across countries and/or sectors.
16
We obtain similar results when we use the number of countries followed by an analyst rather than a
dummy variable to indicate multi-country versus single-country coverage as the dependent variable in
our estimated regressions. We do not report these results in the paper for purposes of brevity.
20
the table are also identical to each other again with the exception of the variable that represents
the country factor. They differ from the first two models in one important respect—the
independent variables in these regressions represent averages for an analyst’s entire portfolio
for a given period. For each of the four models, we also report the change in implied probability
of following a multi-country portfolio as an independent variable in the logistic regression
changes from quartile 1 (25
th
percentile value) to quartile 4 (75
th
percentile value), holding all
other variables constant at their mean values.
Across all four models, the coefficients associated with the country factor are negative
and significant at the one percent level, indicating that a higher country factor decreases the
likelihood of multi-country coverage. Since a higher value for the country factor implies greater
commonalities across firms in the same country, this evidence is consistent with the information
efficiency hypothesis. The economic significance of this variable can be gauged by examining
the reported change in implied probability that is associated with the parameter estimates. We
find that as the country factor rises from its 25
th
percentile value to its 75
th
percentile value, the
probability of country specialization increases between 0.036 (3.6%) and 0.119 (11.9%).
The revenue generation hypothesis—specialization is more likely if a market is relatively
large—is supported by our results as the parameter estimate associated with the natural
logarithm of the country market capitalization variable (LCTYMV) is negative and significant in
all four models. In fact, LCTYMV appears to have the second highest economic significance
among all the variables. Specifically, increasing LCTYMV from its 25
th
percentile value to its
75
th
percentile value decreases (increases) the probability of country diversification
(specialization) by an amount ranging from 0.084 (8.4%) to 0.180 (18.0%).
The parameter estimates associated with broker origin (BRORIGIN) are positive and
significant. This evidence suggests that employment by brokers domiciled in the U.S. increases
21
the likelihood of multi-country coverage. A change in the value of the BRORIGIN dummy
from zero to one increases the probability of multi-country coverage by approximately 11%,
suggesting that the economic significance of this variable is relatively high. We also find that
likelihood of country specialization is negatively related to broker size (BRSIZE). This result
runs counter to our predictions regarding this variable. A possible explanation for this result is
that larger brokerages try to appeal to a variety of clients. Some of these clients may have an
organizational structure that is more compatible with the utilization of multi-country sector-
oriented research making it more likely that analysts will cover multiple countries if employed
by a large broker (see, for example, Rubino (2003)). Consistent with the analyst experience
hypothesis, we find that more experienced analysts are significantly more likely to diversify
their coverage across countries. An increase in analyst experience from quartile one to quartile
four diminishes the probability of country specialization by approximately 0.03 (3.0%), across
all four models.
With the exception of the proportion of idiosyncratic variation in stock returns
(VARRESTOT) and the extent of accounting disclosure (ACCTG INDX), the coefficients on the
control variables are consistent across all models and are significant at the one percent level.
Recall that we use the natural logarithm of firm equity market capitalization (LFIRMMV) and
capital market development (CTYMVGDP) as control variables that capture different aspects of
the availability of information. The coefficients associated with both these variables are positive,
implying that increased informational availability raises the likelihood of multi-country
coverage. Among these variables, firm size has the greatest economic significance as evidenced
by the fact that an increase in LFIRMMV from quartile one to quartile four increases the
probability of multiple country coverage by an amount ranging from 0.098 (9.8%) to 0.133
(13.3%). These results parallel those of Petersen and Rajan (2002) who show that the
22
geographical distance between banks and their clients changes in response to banks’
information processing capabilities. In a similar vein, we find that larger firms and firms
located in countries with better disclosure environments are more likely to be followed by
multi-country analysts. These results are consistent with the notion that improved information
availability, because of better capital market development or larger firm size, reduces the need
for geographical proximity between analysts and the stocks they cover.
Our measure of minority shareholder protection, ANTIDIR RIGHTS has a negative
coefficient and an increase in this variable from its 25
th
percentile value to its 75
th
percentile
value reduces the probability of multiple country coverage from 0.127 (12.7%) to 0.176 (17.6%).
Better minority shareholder protection may result in higher commission income for the
brokerage due to greater interest shown by longer-term investors, i.e., investors who are more
likely to value analyst research, thereby making smaller research portfolios more economically
viable and specialization more likely.
III. B. Analysis of Single versus Multiple Sector Analysts
Table V presents tests of our hypotheses regarding analyst specialization by sector.
17
The first two models reported in the table are identical with the exception of the variable that
represents the sector factor--VARSECTOT in the first model and VARSECSYS in the second
model. All the independent variables employed in these two regressions are average values of
the independent variables listed in the first column across all stocks in the analyst’s core
country-sector for a given period. The third and fourth models reported in the table are also
identical to each other again with the exception of the variable that represents the sector factor.
17
We obtain similar results when we use the number of sectors followed by an analyst rather than a
dummy variable to indicate multi-sector versus single-sector coverage as the dependent variable in our
estimated regressions. We do not report these results in the paper for purposes of brevity.
23
They differ from the first two models in one important respect—the independent variables in
these regressions represent averages for all stocks in an analyst’s portfolio for a given period. To
study the economic significance of our results, we also provide information on changes in
implied probability. Overall, the economic significance of explanatory variables in our sector
regressions is relatively low compared with their power in the country regressions.
Consistent with the information efficiency hypothesis, the parameter estimates
associated with the sector factor are negative in all four models. However, they are only
significant at conventional levels in Models 2, 3 and, 4. These results are consistent with the
notion that a higher sector factor decreases the likelihood of multi-sector coverage. We find
that increases in the sector factor from its 25
th
percentile value to its 75
th
percentile value only
decreases (increases) the probability of sector diversification (specialization) by between 0.008
(0.8%) and 0.050 (5.0%). The relatively low variation of the sector factor documented in Table III
may explain these relatively small changes in implied probabilities.
The revenue generation hypothesis is supported by our results as the parameter
estimate associated with the natural logarithm of country-sector market capitalization variable
(LCTYSECMV) is negative and significant in all four models. Increases in LCTYSECMV from
quartile one to quartile four raises the probability of sector specialization by an amount ranging
from 0.045 (4.5%) to 0.081 (8.1%).
The parameter estimates associated with BRORIGIN are negative and significant. The
change in implied probability suggests that employment by brokers domiciled in the U.S.
increases the likelihood of single-sector coverage by approximately 0.15 (15.0%). We also find
that likelihood of sector specialization is positively related to broker size, with an increase in
BRSIZE from the median value of quartile one to median of quartile four boosting the
probability of sector specialization by approximately 0.06 (6.0%). These results are consistent
24
with our broker culture hypothesis. We find that more experienced analysts are significantly
less likely to be specialized, and an increase in analyst experience from its 25
th
percentile value
to its 75
th
percentile value diminishes the probability of sector specialization by approximately
0.075 (7.5%) across all four models. This result is consistent with the analyst experience
hypothesis.
All the control variables are significant at conventional levels. The coefficient associated
with VARRESTOT is negative and significant in all four models. This result is consistent with
the argument that greater idiosyncratic risk forces an analyst to spend more time and effort to
effectively cover a stock, thereby leading to more focused coverage. The likelihood of multiple
sector coverage is negatively associated with the natural logarithm of equity market
capitalization of the firm (LFIRMMV) and accounting index (ACCTG INDX). Once again, firm
size seems to have the greatest explanatory power as an increase in LFIRMMV from quartile
one to quartile four reduces the probability of multiple sector coverage by an amount ranging
from 0.138 (13.8%) to 0.142 (14.2%). We believe that these relations are consistent with the
existence of gains to sector specialization with improved disclosure. For example, improved
disclosure allows for analysts to make more meaningful intra-industry comparisons because
they have access to more than just basic financial statement information. We also find that
specialization by sector is more likely for countries with a higher score for ANTIDIR RIGHTS,
i.e., countries that offer greater protection for minority shareholders, consistent with our earlier
argument that increased shareholder protection will encourage analyst specialization. Finally,
greater capital market development is conducive to multi-sector coverage.
In summary, the information efficiency, revenue generation, broker culture, and analyst
experience hypotheses all explain, in varying degrees, the choice of analysts/brokerages to
follow stocks across countries and sectors. In addition, the inclusion of information on non-core
25
stocks does not provide any more insights regarding portfolio choice than just an examination
of how core characteristics impact the portfolio diversification decision.
IV. Forecast Accuracy and Portfolio Choice
In this section, we investigate whether the structure of analysts’ research portfolios
affects their earnings forecast accuracy. Once again, because we employ analyst-year
observations in this analysis, an analyst may appear in our sample in more than one year. We,
therefore, employ standard errors that account for the possibility that observations for an
analyst across years may be correlated in the analysis that follows (see, for example, Gleason
and Lee (2003)).
18
Prior to addressing the above issue, we will first present summary statistics
for variables, other than those related to portfolio choice since they were presented earlier in
Table III, which the extant literature has argued have an impact on forecast accuracy.
IV.A. Description of Forecast Accuracy and Control Variables
As is the case in Clement and Tse (2005), we transform all the variables to conduct our
examination of the relationship between forecast accuracy and portfolio choice. We undertake
these transformations for two reasons. First, the transformed variables control for firm effects by
ensuring that every variable is a relative measure for a given stock. Second, since every
transformed variable takes on a value between 0 and 1, it ensures that the coefficients in our
estimated regressions are comparable across variables.
With the exception of our measures of firm size and forecast accuracy, each raw
observation, x
ijt
, for analyst i covering stock j at time t, is transformed into a “relative” value rx
ijt
,
as follows:
18
In additional robustness tests, we also employ standard errors that account for the possibility that the
forecasts for analysts employed by the same brokerage are correlated. The results are qualitatively
similar.
26
}{}{
}{
jtjt
jtijt
ijt
xMinxMax
xMinx
rx
= ,
where Min{x
jt
} and Max{x
jt
} represent the minimum and maximum values for the variable x for
stock j during fiscal year t. These “relative” values are then averaged across stocks in the
analyst’s portfolio.
Our measure of analyst forecast accuracy, ACCURACY
ijt
represents the absolute error of
analyst i’s forecast of stock j’s earnings in year t. Consistent with the literature, only the last
forecast issued for the current fiscal year by an analyst at least one month prior to the fiscal year
end is employed. To ensure that increasing accuracy of forecasts results in a higher value for the
relative variable, the relative value of ACCURACY
ijt
(rx
ijt
), is constructed from each raw
observation, x
ijt
, using the following procedure:
}{}{
}{
jtjt
ijtjt
ijt
xMinxMax
xxMax
rx
= .
Once again, these “relative” values are then averaged across stocks in the analyst’s portfolio.
To ensure that we can construct these relative variables, we only consider firms that are
covered by at least two analysts. Note that this relative value of the forecast accuracy captures
the analyst’s relative forecast accuracy for a given firm. Note also that normalizing the absolute
forecast error by either the actual earnings for the fiscal year or by a constant price will not
affect the computed values for the scaled accuracy variable.
Summary statistics for the variables employed to model forecast accuracy are presented
in Table VI. The statistics presented pertain to the “raw” values of the variables so as to give the
reader a better feel for the sample characteristics. Note that all values reported in the table
represent averages of the respective raw variables across all stocks in an analyst’s portfolio for a
given fiscal year. The variable DAYSELAPSED
ijt
measures the length of time in days between
27
the last earnings forecast by any analyst of stock j’s fiscal year t earnings and analyst i’s forecast
of fiscal year earnings. The mean (median) value for DAYSELAPSED
it
when averaged across all
stocks in an analyst’s portfolio is 8.47 (6.00) days. Statistics for FORHOR
it
, which measures the
average number of days from the date on which analyst i forecast fiscal year t earnings for
stocks in her portfolio and the last day of fiscal year t, indicate that the average forecast is issued
between three and four months prior to the fiscal year end (mean 119.45 days and median
107.60 days), and that seventy five percent of the forecasts are issued no more than five months
prior to the fiscal year end. A mean of 2.26 and a median of 2.00 for FORFREQ
it
, which capture
the average number of times analyst i issues forecasts for portfolio stocks during fiscal year t,
suggest that analysts tend to issue two forecasts per year for the stocks they cover, with more
active coverage (above the 75
th
percentile) resulting in three or more forecasts per year for a
stock.
The next two variables describe characteristics of the analyst. The mean (median) value
for GENEXP
it
is 841.05 (606.00) indicating that analysts in IBES International have been following
stocks and reporting to IBES for approximately two years. The mean (median) number of
stocks covered by an analyst, COMPANIES
it
is 8.41 (6.00). The last two variables employed in
our analysis describe characteristics of the stocks covered by the analysts and the brokerages
employing the analysts. The mean (median) value of $1,221.70 million ($1,163.28 million) for
FIRMMV
it
, the average U.S. dollar value of the equity market capitalization of stocks in analyst
i’s portfolio in year t, suggests that most of the stocks in our sample have relatively large market
capitalizations.
19
This is not surprising given that our sample includes only stocks that are
19
We include FIRMMV as a control in our tests to control for systematic differences across firms in
information availability, analyst incentives, and the level of competition between analysts, as these factors
could induce systematic variation in the distributions of the forecast accuracy of analysts across firms.
28
actively covered by analysts in IBES. Given that the 75
th
percentile of the variable is $3,379.92
million, a comparison with the mean value for the variable suggests that FIRMMV is relatively
skewed. To account for this skewness, we employ the natural logarithm of FIRMMV in our
subsequent analysis. BRSIZE
it
is a count of the number of analysts employed by analyst i’s
employing broker in year t. On average, brokers employ 155.93 analysts with the smaller
brokerages (25
th
percentile or lower) employing under 21 analysts and the larger brokerages
employing over 279 analysts (75
th
percentile or higher).
IV.B. Impact of Portfolio Choice on Forecast Accuracy
Table VII presents regression results that examine the relation between analyst forecast
accuracy and portfolio diversity along the country dimension. Model 1 establishes a benchmark
against which we compare our results on the relationship between forecast accuracy and
portfolio diversity across countries when we control for the endogeneity of portfolio choice. In
estimating Model 1, we treat analyst organization as an exogenous variable. The statistically
insignificant coefficients associated with the variable DUMCTY, a dummy variable that takes
the value of 1 if an analyst covers stocks from multiple countries and 0 otherwise, in Model 1
suggests that diversification of coverage across countries has no impact on forecast accuracy.
This result supports the notion that the benefits of information complementarities realized
offset the loss of scale economies when analysts diversify their coverage across countries.
Models 2 through 5 are estimated under the assumption that country diversification is
endogenously determined. We account for the country diversification choice by employing the
maximum-likelihood estimation procedure to implement a two-stage Heckman treatment effect
methodology. Here the binary self-selection first stage equation is modeled as a probit. Each of
these models employs one of the models presented in Table IV as the basis for the country
29
diversification decision. Model 2 in Table VII employs Model 1 from Table IV to model country
diversification. Similarly, Model 3 employs Model 2 from Table IV, Model 4 employs Model 3
from Table IV, and Model 5 employs Model 4 from Table IV. The positive and statistically
significant coefficient associated with the variable DUMCTY in Models 2 through 5 supports the
notion that multi-country research portfolios improve forecast accuracy once the country
diversification choice is treated as endogenous.
In all the estimated Heckman models, the coefficient associated with the Inverse Mills
Ratio is significantly negative at the 1% level. This negative coefficient implies that brokerages
assign multi-country research portfolios to analysts who, for a reason that is not captured by the
variables in our forecast accuracy model, are not as likely to forecast accurately.
20
Our results,
therefore, indicate that the performance of these analysts is boosted because they are assigned
multiple-country portfolios. Alternatively, these results imply that the performance of these
analysts would have been worse if they were assigned single-country research portfolios.
21
The coefficients associated with the control variables in this regression are consistent
across all the regressions presented in the table. Further, these coefficients are generally
consistent with the evidence presented in prior investigations into the forecast accuracy of
analysts. The coefficient estimate associated with RDAYSELAPSED indicates that forecasts that
20
For example, as we argued earlier, it is possible that brokers prefer to assign diversified portfolios to
analysts who are also responsible for supervising other analysts because the supervisors can better
supervise analysts covering different sorts of firms if they themselves cover a variety of firms. Further,
these administrative duties may adversely affect the forecasting abilities of the supervisors. An
alternative explanation for our result is that supervisors, knowing that their performance may suffer
because of their supervisory responsibilities, pick portfolios that help limit the drop off in their
performance by selecting performance boosting complex portfolios when they are possible to put
together.
21
Our findings are analogous to finding students who take SAT preparation classes are inherently less
capable than students who do not take these classes, but that, by taking the classes, their SAT scores are
boosted. Consequently, by not controlling for this self-selection bias, we may find that SAT preparation
classes have a non-positive impact on SAT scores.
30
are clustered together tend to be more accurate. This result is consistent with the findings of
Clement and Tse (2005). The variable RFORHOR has a significantly negative coefficient,
consistent with prior research showing that earnings forecasts closer to the fiscal year end are
more accurate (see, e.g., O’Brien, 1988). This result supports the idea that forecasts made later
in the fiscal year benefit from the availability of more information and, as a result, tend to be
more accurate. The positive coefficient associated with RFORFREQ, a proxy for analyst effort,
suggests that analysts who work harder produce more accurate results. Jacobs, Lys, and Neale
(1999) and Clement (1999) document similar relationships between forecast accuracy with
forecast horizon and forecast frequency.
Analyst experience displays a tendency to detract from forecast accuracy as evidenced
by the negative and statistically significant coefficient for RGENEXP.
22
Our results are generally
consistent with prior evidence on the relationship between forecast accuracy and analyst
experience. For example, Hong, Kubik, and Solomon (2000) find that forecast accuracy
deteriorates with analyst general experience. We had earlier documented that analysts with
greater experience are assigned diversified portfolios. Here we find that greater general
experience reduces forecast accuracy. Thus, it appears that analysts with greater experience do
not have higher ability. We speculate that the responsibilities of these analysts may have
evolved over time in that generating more accurate forecast may be just one small part of their
job description. It is possible that these analysts are now being evaluated more for their ability
to generate revenues for their brokerages and/or have added supervisory responsibilities.
Hong, Kubik, and Solomon (2000) find that more experienced analysts are less likely to face
22
Note that the variable RGENEXP is different from the variable GENEXP that is employed in the first
stage of the Heckman procedure and thus its coefficient can be interpreted without accounting for the
value of the coefficient of GENEXP in the choice model.
31
unfavorable job outcomes for poorer forecast accuracy than less experience analysts. This result
is consistent with the interpretation that the market for analysts recognizes these other
responsibilities and, as a result, does not penalize experienced analysts as much for poor
performance. In turn, these analysts respond to the incentives provided by the market for
analysts.
Our results with respect to RBRSIZE suggest that the likely increased availability of
resources for analysts working for larger brokers does not translate into more accurate
forecasts.
23
We also find that following a larger number of stocks enhances analyst forecast
accuracy. Finally, our results suggest that forecasts for larger firms tend to be more accurate.
There are a number of potential explanations for this phenomenon, not least of which are the
likelihood of improved informational availability for larger firms, greater competition among
analysts covering larger firms, and larger rewards for analysts who forecast well for larger firms.
Table VIII presents results from regressions estimated to study the impact of coverage
diversification across sectors on forecast accuracy. In Model 1, we examine whether there is any
difference in forecast accuracy between analysts who focus their coverage entirely on one sector
and analysts who cover stocks from multiple sectors. In this case, analyst organization is treated
as an exogenous variable. Once again, the statistically insignificant coefficients associated with
the variable, DUMSEC, a dummy variable that takes the value of 1 if an analyst covers stocks
from multiple sectors and 0 otherwise, suggests that diversification of coverage across sectors
has no impact on forecast accuracy. This result supports the notion that the benefits of
23
This result differs from that reported in Clement (1999) who employs a dichotomous classification of
brokerages. In his paper, in contrast to our use of a firm level ranking of brokerages, broker size is
captured by a dummy variable that takes the value one when the broker is in the highest decile in a
global ranking of brokers based on the number of analysts employed in that year. When we employ a
broker size classification similar to that in Clement (1999), we also find that the relation between this
broker size dummy and forecast accuracy is significantly positive, supporting the hypothesis that the
larger brokerages are better able to support analysts. The remaining results remain unchanged.
32
information complementarities realized offset the loss of scale economies when analysts
diversify their coverage across sectors.
Models 2 through 5 also examine the impact of sector diversification on forecast
accuracy. Unlike Model 1, however, they are estimated under the assumption that sector
diversification is endogenously determined. We account for the sector diversification choice by
employing a two-stage Heckman methodology. Each of these models employs one of the
models presented in Table V as the model for the sector diversification decision. Model 2 in
Table VIII employs Model 1 from Table V to model sector diversification. Similarly, Model 3
employs Model 2 from Table V, Model 4 employs Model 3 from Table V, and Model 5 employs
Model 4 from Table V. The positive and statistically significant coefficient associated with the
variable DUMSEC in Models 2 through 5 supports the notion that multi-sector research
portfolios provide more accurate forecasts than single-sector research portfolios once the sector
diversification choice is treated as an endogenous variable. Here too the coefficients associated
with the Inverse Mills Ratios are negative in all estimated Heckman models but significantly
negative at conventional levels in Models 2, 4, and 5. The negative coefficient on the Inverse
Mills Ratio implies that analysts who follow multi-sector research portfolios are those who are
not likely to forecast accurately. Taken together, these results imply that if multi-sector analysts
are assigned to follow single-sector research portfolios, their performance would suffer. Note,
however, that results are weaker than those in Table VII. The coefficients associated with the
control variables in this regression are consistent across all the regressions presented in the table
and are similar to those in Table VII.
Overall, our results indicate that portfolio diversity in the country or sector dimension
enhances forecast accuracy once we account for the fact that portfolio choice is endogenously
determined. These findings are consistent with the notion that brokerages rationally assign
33
analysts to follow diversified portfolios as long as the benefits arising from informational
complementarities dominate the costs that result from any decrease in economies of scale in
information production. The natural question that arises is then why all research portfolios are
not diversified either along country and sector lines. The answer may simply lie in the fact the
opportunities to exploit informational complementarities may be limited, and where they exist,
brokerages organize analyst research to take advantage of them. These opportunities will be
greater when the common variation in returns, our proxy for the presence of scale economies, is
relatively low for stocks belonging to the same country (sector).
34
V. Extensions
In this section we describe results from additional tests designed to provide further
insights into the relationship between analyst organization and forecast accuracy. First, we
discuss results from a switching regression method to assess the impact of multi-country and
multi-sector portfolios on forecast accuracy. The next two sets of tests described in this section
are designed to assess the impact of country (sector) diversification while controlling for the
sectoral (country) structure of analyst portfolios. All the robustness tests support our earlier
conclusions. Regardless of the nature of the test employed, switching to multi-country coverage
appears to boost forecast accuracy. The evidence on the positive effects of a switch to multi-
sector coverage is somewhat weaker but most of the tests described below suggest that
diversifying coverage across sectors tends to boost performance. For purposes of brevity, the
results from these tests are not reported in tables in the paper.
V.A. Switching Regressions
In earlier sections, the tests employed to asses the impact of analyst organization on
forecast accuracy are valid under the assumption that analyst organization only affects the
intercept terms in the regression and that the sensitivity of forecast accuracy to other factors
such as forecast horizon and experience is independent of analyst organization. However, it is
possible that this assumption is inappropriate. If so, the effect of analyst organization on
forecast accuracy can be assessed using a switching regression technique. The only difference
between the switching regression technique and the methodology employed thus far is that
multiple regressions, one for each form of analyst organization, are estimated to assess the
relationship between organization and forecast accuracy. More specifically, as in the tests
presented earlier, we first model the choice of analyst portfolios using probit regressions. These
35
estimates are then employed to estimate the Inverse Mills ratio (IMR) for each observation.
Finally, in the second stage of the estimation, a separate OLS regression is estimated for each
type of analyst organization, i.e., for each form of analyst organization, o, we estimate the
following regression:
ACCURACY
o
=
β
o1
+
β
o2
X
2
+
β
o3
X
3
+…….
β
on
X
n
+
γ
o
IMR +
ε
The effect of switching from analyst organization type a to analyst organization type b is then
assessed by computing the Average Treatment Effect (ATE), which ignores the effect of the
Inverse Mills Ratio and measures the difference in the predicted accuracy across the entire
sample of the two forms of analyst organization, i.e.,
ATE
ab
=E{(
β
a1
-
β
b1
)+ (
β
a2
-
β
b2
)
X
2
+(
β
a3
-
β
b3
)
X
3
+…….(
β
an
-
β
bn
)
X
n
}
Note that, because our focus is on the effect of portfolio structure on forecast accuracy of the
average analyst, we use the Average Treatment Effect and not the Treatment Effect for the Treated
(see, e.g., Hamilton and Nickerson (2003), Heckman and Robb (1986), Rubin (1974), and Rubin
(1978) for a distinction between these two effects). Estimating the statistical significance of the
average treatment effect using parametric methods is not feasible because we need an estimate
from the accuracy regressions of the covariance between the error terms for each type of analyst
organization (σ
12
) to compute its standard error. The parameter σ
12
is, however, not estimable in
the above kinds of switching regression models (see, for example, Maddala, 1983). We,
therefore, resort to non-parametric methods to compute the significance of the average
treatment effects.
All results from these switching regression tests of our hypotheses are similar to those
reported earlier—complex portfolios tend to boost forecast accuracy, whether the comparison is
made between the effect of diversification across sectors or diversification across countries. For
example, the average treatment effect is 0.0259 for country diversification when we use the
36
independent variables in Model 4 of Table IV to model the choice of analyst portfolio along the
country dimension. We find that 96.7% of the treatment effects evaluated across all analysts in
the sample are non-negative. This proportion is significantly different from 50.0% at the 1%
level of significance. Similarly, the average treatment effect for sector diversification is 0.0270
when we use the independent variables in Model 4 of Table V to model the choice of analyst
portfolio along the sector dimension. Here, we find that 98.8% of the treatment effects
evaluated across all analysts in the sample are non-negative, and that this proportion is
significantly different from 50.0% at the 1% level of significance.
V.B. Binomial Tests Employing Finer Classification of Organization Structure
The tests described earlier compare the impact on forecast accuracy of switching from
single country (sector) to multi-country (sector) coverage. Note, however, that some of the
analysts who cover stocks from a single country (sector) may cover stocks from multiple sectors
(countries) while analysts who cover stocks from multiple countries (sectors) may only focus on
a single sector (country). To the extent that changes in sectoral/country coverage can affect
performance and there is a systematic relationship between these coverage decisions, tests that
assess the impact of country (sector) diversification while controlling for the sector (country)
structure of analysts portfolios may result in more accurate conclusions regarding the effect of
organization structure on forecast accuracy.
This prompted us to run two sets of robustness tests. First, we employed the Heckman
methodology described earlier on appropriate subsets of our sample to estimate the effect of
switching from (i) single-country single-sector (SCSS) coverage to single-country multi-sector
(SCMS) coverage, (ii) single-country single-sector (SCSS) coverage to multi-country single-
sector (MCSS) coverage, and (iii) single-country single-sector (SCSS) coverage to multi-country
37
multi-sector (MCMS) coverage. The results from these tests support our earlier conclusions that
diversifying coverage across countries and sectors results in improved forecast accuracy.
V.C. Multinomial Tests Employing Finer Classification of Organization Structure
In the final set of tests, we extend the switching regression methodology described
earlier to examine the effect on forecast accuracy of the following four forms of analyst
organization: SCSS, SCMS, MCSS, and MCMS. To estimate the average treatment effects of
these four organizational forms, we first use the variables employed earlier to estimate
multinomial logistic regressions that model analyst choices among these four organizational
forms with SCSS coverage as the base group. Coefficient estimates associated with the variables
measuring commonalities within countries and sectors, country and sector sizes, and broker
origin conform closely both in terms of sign and statistical significance with our earlier
estimates of coefficients associated with these variables. The four regressions estimating
forecast accuracy employ the Inverse Mills Ratios generated from the multinomial regression to
control for the endogeneity in analyst organization.
24
In all cases, estimates of the average
treatment effects obtained from pairs of these forecast accuracy regressions (SCSS vs. SCMS,
SCSS vs. MCSS, and SCSS vs. MCMS) suggest that diversification across countries helps
improve forecast accuracy. The evidence on the positive effects of diversification across sectors
is somewhat weaker but generally suggests that diversification across sectors also helps
improve forecast accuracy.
24
Lee (1982) demonstrates how non-normal unobservables implied by the multinomial regressions can be
transformed into normal variables, thus, easily allowing us to construct the inverse Mills ratios. His approach
provides us with a simple method to account for multinomial choices.
38
VI. Conclusions
A widely held belief among financial economists is that analysts focus their research
coverage on individual sectors. For example, Ross, Westerfield, and Jaffe (2005, p. 337) contend
that security analysts, “…are employed by brokerage houses to follow the companies in
individual industries. For example, an analyst for a particular brokerage house might follow all
the firms in, say, the auto industry.” While this sectoral approach may largely reflect the
manner in which analyst research is organized in the U.S., we find rich cross-sectional variation
in analyst research portfolios along both country and sector dimensions for analysts covered by
International IBES. In fact, this variation reflects a long-standing debate among practitioners on
whether international analyst specialization should occur along country or sector dimensions
(see Rudd (1989)). We shed light on this debate by conducting an in-depth study of the
economic forces and constraints that determine the structure of the research portfolios of
international analysts.
We develop four non-mutually exclusive hypotheses relating to analyst research
portfolios. First, the information efficiency hypothesis posits that the strength of commonalities
between firms within a country or sector determine the nature of specialization. Second, the
revenue generation hypothesis posits that the choice between focused versus broad coverage of
stocks by analysts is determined by opportunities to generate revenue streams for the brokerage.
Next, the broker culture hypothesis predicts that the observed cross-sectional variation in analyst
specialization reflects broker culture in terms of how brokerages have traditionally organized
their research in their home economies. Finally, the analyst experience hypothesis states that more
experienced analysts, either due to following different types of stocks over the course of their
careers or because of their greater ability, will tend to cover more diverse portfolios.
39
Our empirical tests provide broad support for all these hypotheses. Specifically,
consistent with the information efficiency hypothesis, we find that the probability that analysts
specialize by country is higher as the country factor rises, and the likelihood that they focus on a
sector increases as the sector factor strengthens. In line with the revenue generation hypothesis, we
find that higher country equity market capitalization increases the likelihood of country
specialization; while a larger total equity market capitalization of a sector in a particular country
(country-sector market capitalization) increases the likelihood that analysts will cover firms in a
single sector. In support of the broker culture hypothesis, we document that analysts employed by
U.S. brokers are less likely to specialize by country and more likely to specialize by sector.
Finally, consistent with the analyst experience hypothesis, we find that more experienced analysts
have less focused research portfolios.
When we examine the relation between forecast accuracy and portfolio organization
using a two-stage Heckman methodology to control for the endogeneity in the structure of
analyst research portfolios, we find that diversified portfolios promote forecast accuracy. This
is the case when we consider portfolio diversification across either countries or sectors. In
contrast, Clement (1999) documents a negative relation between forecast accuracy and measures
of portfolio complexity for a sample of U.S. analysts and Clement, Rees, and Swanson (2003)
find no relation between forecast accuracy and measures of portfolio complexity for a sample of
international analysts. In both these studies, the structures of analyst portfolios are treated as
being exogenously determined. Thus, our analysis indicates that a simple comparison of
forecast accuracy across analysts with different portfolio structures without controlling for self-
selection in portfolio choice appears to understate differences in forecast accuracy. Our finding
that diversified research portfolios outperform focused research portfolios is consistent with the
notion that brokerages rationally assign analysts to follow diversified portfolios as long as the
40
marginal benefits arising from informational complementarities dominate the marginal costs
that result from any decrease in economies of scale in information production. We, therefore,
conclude that brokerages structure analyst research portfolios in a manner that is conducive to
enhancing forecast accuracy.
Our results provide a fresh perspective on the scandals that have linked various conflicts
of interest faced by analysts to the quality of their research. The general impression is that
brokerage research and recommendations are entirely driven by investment-banking related
incentives and have little relation to fundamental values. We weigh in on this issue not only by
examining how the production of research is structured but by also studying the impact of
portfolio choice on brokerage output. We do find that the potential for investment banking
business and/or brokerage commissions impacts the manner in which analyst research is
organized. However, we also find that the structure of the production process of research in
brokerages is consistent with producing good quality research.
41
REFERENCES
Bacmann, J., and G. Bolliger, 2001, Who are the best? Local versus foreign analysts on the Latin
American stock markets, University of Neuchatel and FAME working paper.
Bae, K.H., R. Stulz, and H. Tan, 2005, Do local analysts know more? A cross-country study of
the performance of local analysts and foreign analysts, Ohio State University and
Queen’s University working paper.
Berger, A.N, N.H. Miller, M.A. Petersen, R.G. Rajan, and J.C. Stein, 2005, Does function follow
organizational form? Evidence from the lending practices of large and small firms,
Journal of Financial Economics 76, 237-269.
Bolliger, G., 2001, The characteristics of individual analysts’ forecasts in Europe, FAME working
paper.
Boni, L., and K.L. Womack, 2005, Analysts, industries, and price momentum, University of New
Mexico working paper.
Chang, C., 2003, Information footholds: Expatriate analysts in an emerging market, University
of California at Berkeley working paper.
Choe, H., C. Kho, and R.M. Stulz, 2000, Do domestic investors have more valuable information
about individual stocks than foreign investors? Review of Financial Studies 18, 795-829.
Clement, M.B., 1999, Analyst forecast accuracy: Do ability, resources, and portfolio complexity
matter? Journal of Accounting and Economics 27, 285-303.
Clement, M.B., L. Rees, and E.P. Swanson, 2003, The influence of culture and corporate
governance on the characteristics that distinguish superior analysts, Journal of Accounting,
Auditing, and Finance 18, 593-618.
Clement, M.B., and S.Y. Tse, 2005, Financial analyst characteristics and herding behavior in
forecasting, Journal of Finance 60, 307-341.
Chen, Z., 2003, Buy-side and sell-side: The industrial organization of information production in
securities industry, Wharton School working paper.
Coval, J.D., and T.J. Moskowitz, 2001, The geography of investment: Informed trading and asset
prices, Journal of Political Economy 109, 811-841.
DeAngelo, L., 1981, Auditor size and Auditor quality, Journal of Accounting and Economics 3,183-
199.
Garicano, L., and T. N. Hubbard, 2005, Specialization, firms and markets: The division of labor
within and between law firms, University of Chicago working paper.
42
Gilson, S.C., P.M. Healy, C.F. Noe, and Krishna G. Palepu (2001), Analyst specialization and
conglomerate stock breakups, Journal of Accounting Research 39, 565-582.
Gleason, C., and C. Lee, 2003, Analyst forecast revisions and market price discovery, Accounting
Review 78, 193-225.
Greene, W.H., 2000, Econometric Analysis 5
th
Edition, NJ: Prentice Hall, Inc.
Griffin, J.M., and A.G. Karolyi, 1998, Another look at the role of the industrial structure of
markets for international diversification strategies, Journal of Financial Economics 50, 351-
373.
Grinblatt, M., and M. Keloharju, 2000, The investment behavior and performance of various
types of investors: A study of Finland’s unique data set, Journal of Financial Economics 55,
43-67.
Hamilton, B.H., and J.A. Nickerson, 2003, Correcting for endogeneity in strategic management
research, Strategic Organization 1, 51-78.
Heckman, J., 1978, Dummy endogenous variables in a simultaneous equations system,
Econometrica 46, 931-961.
Heckman, J., 1979, Sample selection bias as a specification error, Econometrica 47, 153-161.
Heckman, J., and R. Robb, 1986, Alternative methods for solving the problem of selection bias
in evaluating the impact of treatments on outcomes, In Drawing inferences from self-
selected samples, edited by H. Wainer, 63-107, Mahwah, NJ: Lawrence Erlbaum Associates.
Heston, H.L., and G.K. Rouwenhorst, 1994, Does industrial structure explain the benefits of
international diversification? Journal of Financial Economics 36, 3-27.
Hong, H., and J.D. Kubik, 2003, Analyzing the analyst: Career concerns and biased forecasts,
Journal of Finance 58, 313-351.
Hong, H., J.D. Kubik, and A. Solomon, 2000, Security analysts’ career concerns and herding of
earnings forecasts, RAND Journal of Economics 31, 121-144.
Huber, P., 1967, The behavior of maximum likelihood estimators under non standard conditions,
In Procedures of the Fifth Annual Berkeley Symposium on Mathematical Statistics and
Probability, Volume 1, edited by L.M. LeCam and J. Neyman, 221-233, Berkeley, CA:
University of California Press.
International Accounting and Auditing Trends, Princeton, NJ: Center for International Financial
Analysis and Research, 1995.
Jacob, T., T. Lys, and M. Neale, 1999, Expertise in forecasting performance of security analysts,
Journal of Accounting and Economics 28, 51-82.
43
Kang, J., and R.M. Stulz, 1997, Why is there a home bias? An analysis of foreign portfolio equity
ownership in Japan, Journal of Financial Economics 46, 2-28.
Kennedy, P., 1986, Interpreting dummy variables, Review of Economics and Statistics 68, 174-175.
Kini, O, S.L. Mian, M.J. Rebello, and A. Venkateswaran, 2006, Core specialization, forecast
accuracy, and the market for international analysts, Georgia State University working
paper.
La Porta, R., F. Lopez-de-Silanes, A. Shleifer, and R.W. Vishny, 1997, Legal determinants of
external finance, Journal of Finance 52, 1131-1150.
Lee, L.F., 1982, Some approaches to the correction of selectivity bias, Review of Economic Studies
49, 355-372.
Li, K., and N. Prabhala, 2005, Self-selection models in corporate finance, University of Maryland
working paper.
Maddala, G.S., 1983, Limited-dependent and qualitative variables in econometrics, Cambridge:
Cambridge University Press.
Malloy, C.J., 2005, The geography of equity analysis, Journal of Finance 60, 719-755.
Morck, R., B. Yeung, and W. Yu, 2000, The information content of stock markets: Why do
emerging markets have synchronous stock price movements? Journal of Financial
Economics 58, 215-260.
O’Brien, P., 1988, Analysts’ forecasts as earnings expectations, Journal of Accounting and
Economics 10, 53-85.
Petersen, M.A., and R.G. Rajan, 2002, Does distance still matter? The information revolution in
small business lending, Journal of Finance 57, 2533-2570.
Rajan, R.G., and L. Zingales, 1998, Financial dependence and growth, American Economic Review
88, 559-586.
Reuters Institutional Investor Survey, Institutional Investor Research Group, 2002.
Rogers, W.H., 1993, Regression standard errors in clustered samples, Stata Technical Bulletin 13,
19-23.
Ross, S.A., R.W. Westerfield, and J. Jaffe, Corporate Finance 7
th
edition, New York, NY: McGraw-
Hill Irwin, 2005.
Rubin, D., 1974, Estimating causal effects of treatments in randomized and nonrandomized
studies, Journal of Educational Psychology 66, 688-701.
44
Rubin, D., 1978, Bayesian inference for causal effects, The Annals of Statistics 6, 34-58.
Rubino, J., 2003, “The new global industry analysis,” CFA Magazine, July-August, 46-47.
Rudd, A., 1989, Industry versus country correlation, Journal of Portfolio Management 15, 79-88.
Seasholes, M., 2004, Re-examining information asymmetries in emerging markets, University of
California at Berkeley working paper.
Sonney, F., 2005, Sector versus country specialization and financial analysts’ performance,
University of Neuchatel and FAME working paper.
Stigler, G., 1951, Division of labor is limited by the extent of the market, Journal of Political
Economy 59, 185-193.
White, H, 1980, A heteroskedasticity-consistent covariance matrix and a direct test for
heteroskedasticity, Econometrica 48, 817-838.
Wooldridge, J.M., 2002, Econometric Analysis of Cross Section and Panel Data, Cambridge, MA:
MIT Press.
World Stock Exchange Fact Book, Morris Plains, NJ: Electronic Commerce, Inc., 1996-2003.
Zhou, J. and Ronald Elder, 2003, Audit firm size, industry specialization, and earnings
management by IPO firms, State University of New York at Binghampton and Syracuse
University working paper.
45
Table I
Determinants of Analyst Research Portfolio Structures
This table presents the predicted relations between the probability of country/sector diversification and various independent
variables under our four hypotheses. The variables employed to explain analyst diversification by country (sector) include the
following variables: VARCTYTOT (VARSECTOT) the proportion of total weekly return variability for each firm that is attributable
to the movements in the weekly returns on the country (sector) factor. VARCTYSYS (VARSECSYS) is the ratio of the variance of
the weekly returns on the country (sector) factor to the sum of the variances of the weekly returns on the global, country, and
sector factors. CTYMV (CTYSECMV) is the stock market capitalization of the country (country-sector) of the firm. BRSIZE is
measured as the number of analysts employed by the broker. BRORIGIN is a dummy variable that takes the value of one if the
broker is of U.S. origin, and is zero otherwise. ANALEXP is the number of days elapsed between the first ever forecast made by an
analyst and her latest forecast for that particular fiscal year on International IBES.
Hypothesis Proxy Variable Hypothesized sign
Country
diversification
Sector
diversification
Information efficiency hypothesis
Country factor (VARCTYTOT and
VARCTYSYS)
(-) n.a.
Sector factor (VARSECTOT and
VARSECSYS)
n.a. (-)
Revenue generation hypothesis
Country market capitalization
(CTYMV)
(-) n.a.
Country-sector market capitalization
(CTYSECMV)
n.a. (-)
Broker culture hypothesis
Broker origin (BRORIGIN) (+) (-)
Broker size (BRSIZE) (-) (-)
Analyst experience hypothesis
Number of days on IBES (GENEXP) (+) (+)
46
Table II
This table presents mean and median values of characteristics of analyst research portfolios for a sample of 19,379 analysts covering 11,292 firms between 1996 and 2002.
The characteristics described include the number of countries in which firms followed by an analyst are domiciled, the number of sectors in which firms followed by an
analyst operate based on the SIG codes on International IBES, the number of firms followed, and analyst experience. This data is presented in three panels. Panel A contains
a description of analyst portfolios for our entire sample of analysts. Panel B presents this information for analysts whose entire research portfolio consists of firms
domiciled in a single country (SC) and those covering firms in multiple countries (MC). In Panel C descriptions are for single-sector analysts (SS) and multiple sector
analysts (MS). We report t-statistics (Wilcoxon z-statistics) for differences between mean (median) characteristics across analyst groups in each panel.
a
,
b
, and
c
indicate
significance at the 0.01 0.05, and 0.10 levels, respectively.
Anal
y
st
g
roup
Number of
anal
st
ears
Percenta
g
e of
total
Number of countries
followed b
y
anal
y
s
t
Number of sectors
followed b
y
anal
y
s
t
Number of firms followed
b
y
anal
y
st
Anal
y
st experience
(da
y
s)
Mean Median Mean Median Mean Median Mean Median
Panel A: All
A
LL
47,049 100 1.46 1.00 2.36 2.00 8.41 6.00 841 606
Panel B: By country
SC
36,697 78.00 1.00 1.00 2.40 2.00 8.27 6.00 814 569
MC
10,352 22.00 3.07 3.00 2.22 2.00 8.91 8.00 934 726
t-stat (z-stat)
350.00
a
214.92
a
-11.09
a
-4.99
a
9.71
a
24.93
a
14.53
a
18.66
a
Panel C: B
y
sector
SS
18,714 39.78 1.43 1.00 1.00 1.00 5.47 3.00 781 496
M
S
28,335 60.22 1.47 1.00 3.26 3.00 10.35 9.00 880 672
t-stat (z-stat)
4.22
a
5.11
a
280.00
a
191.84
a
94.82
a
100.37
a
13.98
a
25.58
a
47
Table III
Descriptive Statistics on Variables that are Hypothesized to Impact Portfolio Structure
The sample consists of 47,049 analyst-year observations over the period 1996-2002. Each of the explanatory variables described
below is measured in two ways. First, their values are averaged only for the stocks in the core country-sector in the analyst’s
portfolio (Core selection). Second, their values are averaged across all stocks in the analyst’s portfolio (Overall selection). The
variables employed to explain analyst specialization include the proportion of total weekly return variability for each firm that
is attributable to the movements in the weekly returns on the country factor (VARCTYTOT), the ratio of the variance of the
weekly returns on the country factor to the sum of the variances of the weekly returns on the global, country, and sector factors
(VARCTYSYS), the proportion of total weekly return variability for each firm that is attributable to the movements in the weekly
returns on the sector factor (VARSECTOT), and the ratio of the variance of the weekly returns on the sector factor to the sum of
the variances of the weekly returns on the global, country, and sector factors (VARSECSYS). CTYMV is the stock market
capitalization of the country in which the firm is domiciled. CTYSECMV is the country-sector market capitalization for the stock.
BRORIGIN is a dummy variable that takes the value of one if the broker is of U.S. origin, and is zero otherwise. BRSIZE is
measured as the number of analysts employed by the broker. GENEXP is the number of days elapsed between the first ever
forecast made by an analyst and his/her latest forecast during a given year on International IBES. VARRESTOT is the residual
variance of the firm divided by its total variance, where both are estimated using weekly returns. FIRMMV represents the firm
equity market capitalization. Following LLSV (1997), we also employ proxies for the levels of information disclosure (ACCTG
INDX), minority shareholder protection (ANTIDIR RIGHTS), and capital market development (CTYMVGDP)
.
Variable Core Selection Overall Selection
Mean Median Mean Median
VARCTYTOT
0.199 0.148 0.201 0.156
VARCTYSYS
0.553 0.540 0.557 0.543
VARSECTOT
0.066 0.043 0.066 0.049
VARSECSYS
0.320 0.195 0.324 0.242
CTYMV ($ m.) 458,630.429 530,194.925 437,136.158 479,260.706
CTYSECMV ($ m.) 51,534.151 62,255.365 46,536.862 53,369.783
BRORIGIN (proportion) 0.243 0.000 0.243 0.000
BRSIZE (# of analysts) 155.925 107.000 155.925 107.000
GENEXP (days) 841.053 606.000 841.053 606.000
FIRMMV ($ m.) 1,208.337 1,152.859 1,221.702 1,163.281
ACCTG INDX
68.773 69.000 68.548 69.000
ANTIDIR RIGHTS
3.407 4.000 3.372 4.000
CTYMVGDP
0.839 0.429 0.803 0.451
48
Table IV
Binomial Logistic Regression Analysis for Single Country versus Multiple Country Analysts
The dependent variable is a dummy variable, DUMCTY, that takes the value of one if the analyst is a multiple country
analyst, and is zero otherwise. Our unit of analysis is the analyst-year combination. Under the core selection models,
all the independent variables for a year are computed by averaging the value of the underlying variable across all
stocks in an analyst’s core country-sector, while under the overall selection models, all the independent variables for a
year are computed by averaging the value of the underlying variable across all stocks in an analyst’s portfolio. The
variables employed to explain analyst specialization by country include the proportion of total weekly return
variability for each firm that is attributable to the movements in the weekly returns on the country factor (VARCTYTOT)
and the ratio of the variance of the weekly returns on the country factor to the sum of the variances of the weekly
returns on the global, country, and sector factors (VARCTYSYS). LCTYMV is the natural logarithm of stock market
capitalization of the country the firm in which the firm is domiciled. BRORIGIN is a dummy variable that takes the
value of one if the broker is of U.S. origin, and is zero otherwise. BRSIZE is measured as the number of analysts
employed by the broker. GENEXP is the number of years elapsed between the first ever forecast made by an analyst
and his/her latest forecast during our 1995-1996 period on International IBES. VARRESTOT is the residual variance of
the firm divided by its total variance, where both are estimated using weekly returns. LFIRMMV represents the natural
logarithm of firm equity market capitalization. Following LLSV (1997), we also employ proxies for the levels of
information disclosure (ACCTG INDX), minority shareholder protection (ANTIDIR RIGHTS) and capital market
development (CTYMVGDP). The change in implied probability is computed by varying the independent variable in the
logistic regression from quartile 1 (25
th
percentile value) to quartile 4 (75
th
percentile value), holding all other variables
constant at their mean values. The p-values are in parentheses and reflect standard errors that are corrected for the
cluster sample problem arising from multiple observations across years for each analyst.
a
,
b
, and
c
indicate significance
at the 0.01, 0.05, and 0.10 levels, respectively.
Core Selection Overall Selection
Variable Model 1
(
N = 44,671
)
Model 2
(
N = 44,671
)
Model 3
(
N = 45,984
)
Model 4
(
N = 45,984
)
Coeff. Implied
Prob.
Coeff. Implied
Prob.
Coeff. Implied
Prob.
Coeff. Implied
Prob.
Intercept 0.615
a
(0.01)
n.a. 4.080
a
(0.00)
n.a. 2.407
a
(0.00)
n.a. 5.909
a
(0.00)
n.a.
Year Dummies
Yes n.a. Yes n.a. Yes n.a. Yes n.a.
VARCTYTOT
-1.328
a
(0.00)
-0.036 -1.794
a
(0.00)
-0.052
VARCTYSYS
-2.036
a
(0.00)
-0.115 -2.264
a
(0.00)
-0.119
LCTYMV
-0.330
a
(0.00)
-0.084 -0.504
a
(0.00)
-0.134 -0.524
a
(0.00)
-0.141 -0.704
a
(0.00)
-0.180
BRORIGIN
0.661
a
(0.00)
0.100 0.694
a
(0.00)
0.110 0.688
a
(0.00)
0.111 0.724
a
(0.00)
0.110
BRSIZE
0.001
a
(0.00)
0.035 0.002
a
(0.00)
0.072 0.002
a
(0.00)
0.074 0.002
a
(0.00)
0.069
GENEXP (x
10
2
)
0.024
a
(0.00)
0.029 0.022
a
(0.00)
0.028 0.025
a
(0.00)
0.032 0.024
a
(0.00)
0.029
VARRESTOT
-0.068
(0.
11
)
-0.003 0.050
(0.
24
)
0.002 -0.104
c
(0.0
8
)
-0.005 0.083
c
(0.0
7
)
0.004
LFIRMMV
0.343
a
(0.00)
0.098 0.330
a
(0.00)
0.099 0.449
a
(0.00)
0.133 0.431
a
(0.00)
0.119
ACCTG INDX
0.000
(0.9
6
)
0.000 -0.006
(0.
39
)
-0.009 0.001
(0.
68
)
0.002 -0.004
(0.
27
)
-0.006
ANTIDIR
RIGHTS
-0.440
a
(0.00)
-0.176 -0.357
a
(0.00)
-0.149 -0.484
a
(0.00)
-0.167 -0.398
a
(0.00)
-0.127
CTYMVGDP
0.535
a
(0.00)
0.050 0.557
a
(0.00)
0.055 0.563
a
(0.00)
0.057 0.585
a
(0.00)
0.055
χ
2
3,219.84
a
(0.00)
3,370.16
a
(0.00)
3,748.28
a
(0.00)
3,908.84
a
(0.00)
Pseudo R
2
(%) 14.63 15.75 17.21 18.37
49
Table V
Binomial Logistic Regression Analysis for Single Sector versus Multiple Sector Analysts
The dependent variable is a dummy variable, DUMSEC, that takes the value of one if the analyst is a multiple sector
analyst, and is zero otherwise. Our unit of analysis is the analyst-year combination. Under the core selection models,
all the independent variables for a year are computed by averaging the value of the underlying variable across all
stocks in an analyst’s core country-sector, while under the overall selection models, all the independent variables for a
year are computed by averaging the value of the underlying variable across all stocks in an analyst’s portfolio. The
variables employed to explain analyst specialization by sector include the proportion of total weekly return variability
for each firm that is attributable to the movements in the weekly returns on the sector factor (VARSECTOT) and the
ratio of the variance of the weekly returns on the sector factor to the sum of the variances of the weekly returns on the
global, country, and sector factors (VARSECSYS). LCTYSECMV is the stock market capitalization of the sector for the
country in which the firm is domiciled. BRORIGIN is a dummy variable that takes the value of one if the broker is of
European origin, and is zero otherwise. BRSIZE is measured as the number of analysts employed by the broker.
GENEXP is the number of years elapsed between the first ever forecast made by an analyst and his/her latest forecast
during our 1995-1996 period on International IBES. VARRESTOT is the residual variance of the firm divided by its total
variance, where both are estimated using weekly returns. LFIRMMV represents the natural logarithm of the firm
equity market capitalization. Following LLSV (1997), we also employ proxies for the levels of information disclosure
(ACCTG INDX), minority shareholder protection (ANTIDIR RIGHTS) and capital market development (CTYMVGDP).
The change in implied probability is computed by varying the independent variable in the logistic regression from
quartile 1 (25th percentile value) to quartile 4 (75th percentile value), holding all other variables constant at their mean
values. The p-values are in parentheses and reflect standard errors that are corrected for the cluster sample problem
arising from multiple observations across years for each analyst.
a
,
b
, and
c
indicate significance at the 0.01, 0.05, and
0.10 levels, respectively.
Core Selection Overall Selection
Variable Model 1
(
N = 44,671
)
Model 2
(
N = 44,671
)
Model 3
(
N = 45,984
)
Model 4
(
N = 45,984
)
Coeff. Implied
Prob.
Coeff. Implied
Prob.
Coeff. Implied
Prob.
Coeff. Implied
Prob.
Intercept 4.167
a
(0.00)
n.a. 4.465
a
(0.00)
n.a. 4.809
a
(0.00)
n.a. 5.074
a
(0.00)
n.a.
Year Dummies
Yes n.a. Yes n.a. Yes n.a. Yes n.a.
VARSECTOT
-0.425
(0.
15
)
-0.008 -0.778
b
(0.0
2
)
-0.013
VARSECSYS
-0.659
a
(0.00)
-0.050 -0.698
a
(0.00)
-0.049
LCTYSECMV
-0.084
a
(0.00)
-0.045 -0.107
a
(0.00)
-0.057 -0.139
a
(0.00)
-0.070 -0.159
a
(0.00)
-0.081
BRORIGIN
-0.622
a
(0.00)
-0.153 -0.610
a
(0.00)
-0.150 -0.626
a
(0.00)
-0.153 -0.617
a
(0.00)
-0.151
BRSIZE
-0.001
a
(0.00)
-0.062 -0.001
a
(0.00)
-0.062 -0.001
a
(0.00)
-0.061 -0.001
a
(0.00)
-0.062
GEXP (x 10
2
) 0.034
a
(0.00)
0.076 0.035
a
(0.00)
0.077 0.034
a
(0.00)
0.075 0.034
a
(0.00)
0.075
VARRESTOT
-0.143
a
(0.
00)
-0.011 -0.166
a
(0.00)
-0.013 -0.162
a
(0.00)
-0.012 -0.153
b
(0.04)
-0.012
LFIRMMV
-0.278
a
(0.00)
-0.142 -0.273
a
(0.00)
-0.139 -0.285
a
(0.00)
-0.138 -0.283
a
(0.00)
-0.139
ACCTG INDX
-0.006
a
(0.00)
-0.016 -0.006
(0.00)
-0.016 -0.005
b
(0.0
2
)
-0.013 -0.006
b
(0.00)
-0.016
ANTIDIR
RIGHTS
-0.135
a
(0.00)
-0.098 -0.124
a
(0.00)
-0.090 -0.136
a
(0.00)
-0.071 -0.126
a
(0.0
1
)
-0.066
CTYMVGDP
0.103
a
(0.00)
0.020 0.109
a
(0.00)
0.021 0.107
a
(0.00)
0.021 0.109
a
(0.00)
0.021
χ
2
1,878.03
a
(0.00)
1,976.75
a
(0.00)
1,987.29
a
(0.00)
2,074.84
a
(0.00)
Pseudo-R2 (%) 6.74 7.08 7.20 7.50
50
Table VI
Descriptive Statistics on Raw Control Variables in Accuracy Regressions
The sample consists of 47,049 analyst-year observations over the period 1996-2002. All the control variables are
computed for each year by averaging the value of the underlying variable across all stocks in an analyst’s portfolio.
DAYSELAPSED measures the length of time in days between the last earnings forecast by any analyst of firm j’s
fiscal year t earnings and analyst i’s forecast of fiscal year earnings. FORHOR measures the number of days from the
date on which analyst i’s forecast fiscal year t earnings for firm j and the last day of fiscal year t. FORFREQ is a proxy
for the intensity with which an analyst covers a firm. It is estimated by the number of times analyst i issues forecasts
for firm j during fiscal year t. GENEXP measures the number of days between an analyst’s first forecast in the IBES
database and her last forecast for year t. BRSIZE
represents the number of analysts working for analyst i’s employer
during fiscal year t in which she issues a forecast of firm j’s earnings. COMPANIES
is a count of the number of stocks
in analyst i’s research portfolio in year t. The final control variable FIRMMV
is the equity market capitalization of
firm j during fiscal year t during which it is covered by analyst i.
Variable 25
th
Percentile Mean Median 75
th
Percentile
DAYSELAPSED (days) 2.667 8.472 6.000 11.500
FORHOR (days) 75.667 119.453 107.600 149.600
FORFREQ
1.333 2.264 2.000 3.000
GENEXP (days) 271.000 841.053 606.000 1,184.000
BRSIZE (# of analysts) 21.000 155.925 107.000 279.000
COMPANIES (# of stocks) 3.000 8.409 6.000 11.000
FIRMMV ($ m.) 436.217 1,221.702 1,163.281 3,379.929
51
Table VII
Forecast Accuracy for Multi- versus Single-Country Analysts
This table presents fixed effects regression models with average relative forecast accuracy (ACCURACY) of analyst i for
fiscal year t as the dependent variable for a sample of analyst-year observations made over the period 1996-2002. The
average scaled accuracy is computed across all the stocks in the analyst’s portfolio. Model 1 estimates an ordinary least
squares regression. Models 2 through 5 estimate the same relations but employ the Heckman’s two-stage procedure to
control for the endogeneity in analyst portfolio choice (multi- versus single-country analyst) in the first stage. These
models differ only in the first-stage selection model employed. In Models 2 and 3, the first-stage selection models
employ independent variables that are averaged across all stocks in an analyst’s core with VARCTYTOT (VARCTYSYS)
as the proxy for the country factor in Model 2 (3). In Models 4 and 5, the first-stage selection models employ
independent variables that are averaged across all stocks in an analyst’s portfolio with VARCTYTOT (VARCTYSYS) as
the proxy for the country factor in Model 4 (5). In all these models, we control for year fixed effects. DUMCTY is a
dummy variable that takes the value one if the analyst follows stocks in more than one country; otherwise it takes the
value zero. DAYSELAPSED measures the length of time in days between the last earnings forecast by any analyst of
firm j’s fiscal year t earnings and analyst i’s forecast of fiscal year earnings. RFORHOR measures the number of days
from the date on which analyst i’s forecast fiscal year t earnings for firm j and the last day of fiscal year t. RFORFREQ is
a proxy for the intensity with which an analyst covers a firm. It is estimated by the number of times analyst i issues
forecasts for firm j during fiscal year t. GENEXP measures the number of days between an analyst’s first forecast in the
IBES database and her last forecast for year t. BRSIZE
represents the number of analysts working for analyst i’s
employer during fiscal year t in which she issues a forecast of firm j’s earnings. COMPANIES
is a count of the number
of stocks in analyst i’s research portfolio in year t. The final control variable LFIRMMV is the natural logarithm of
average equity market capitalization of firm j during fiscal year t during which it is covered by analyst i. All control
variables are relative with the exception of LFIRMMV and, as such, have the prefix “R” preceding their names. The p-
values are in parentheses and reflect standard errors that are corrected for the cluster sample problem arising from
multiple observations across years for each analyst. Subscripts have been dropped in the table below for ease of
presentation.
a
,
b
, and
c
indicate significance at the 0.01, 0.05, and 0.10 levels, respectively.
Core Selection Overall Selection
Model 1 Model 2 Model 3 Model 4 Model 5
OLS
(N = 45,947)
Heckman
(VARCTYTOT)
(N 43 794)
Heckman
(VARCTYSYS)
(N 43 794)
Heckman
(VARCTYTOT)
(N 44 999)
Heckman
(VARCTYSYS)
(N 45 207)
Intercept 0.601
a
(0.00)
0.624
a
(0.00)
0.621
a
(0.00)
0.634
a
(0.00)
0.630
a
(0.00)
Year Dummies
Yes Yes Yes Yes Yes
DUMCTY
-0.003
(0.26)
0.068
a
(0.00)
0.060
a
(0.00)
0.081
a
(0.00)
0.070
a
(0.00)
RDAYSELAPSED
-0.070
a
(0.00)
-0.078
a
(0.00)
-0.077
a
(0.00)
-0.073
a
(0.00)
-0.073
a
(0.00)
RFORHOR
-0.231
a
(0.00)
-0.234
a
(0.00)
-0.234
a
(0.00)
-0.234
a
(0.00)
-0.234
a
(0.00)
RFORFREQ
0.039
a
(0.00)
0.040
a
(0.00)
0.039
a
(0.00)
0.041
a
(0.00)
0.041
a
(0.00)
RGENEXP
-0.015
a
(0.00)
-0.022
a
(0.00)
-0.021
a
(0.00)
-0.023
a
(0.00)
-0.021
a
(0.00)
RBRSIZE
-0.009
a
(0.00)
-0.016
a
(0.00)
-0.015
a
(0.00)
-0.017
a
(0.00)
-0.016
a
(0.00)
RCOMPANIES
0.006
(0.19)
0.012
a
(0.00)
0.012
a
(0.00)
0.013
a
(0.00)
0.012
a
(0.00)
LFIRMMV
0.015
a
(0.00)
0.011
a
(0.00)
0.011
a
(0.00)
0.009
a
(0.00)
0.010
a
(0.00)
Adj. R
2
(%) 9.20
χ
2
3,143.14
a
(0.00)
3,144.43
a
(0.00)
3,224.78
a
(0.00)
3,212.20
a
(0.00)
Inverse Mills
Ratio
-0.044
a
(
0.00
)
-0.040
a
(
0.00
)
-0.054
a
(
0.00
)
-0.048
a
(
0.00
)
52
Table VIII
Forecast Accuracy for Multi- versus Single-Sector Analysts
This table presents fixed effects regression models with average relative forecast accuracy (ACCURACY
it
) of analyst i for
fiscal year t as the dependent variable for a sample of analyst-year observations made over the period 1996-2002.
Model 1 estimates an ordinary least squares regression. Models 2 through 5 estimate the same relations but employ the
Heckman’s two-stage procedure to control for the endogeneity in analyst portfolio choice (multi- versus single-sector
analyst) in the first stage. These models differ only in the first-stage selection model employed. In Models 2 and 3, the
first-stage selection models employ independent variables that are averaged across all stocks in an analyst’s core with
VARSECTOT (VARSECSYS) as the proxy for the country factor in Model 2 (3). In Models 4 and 5, the first-stage
selection models employ independent variables that are averaged across all stocks in an analyst’s portfolio with
VARSECTOT (VARSECSYS) as the proxy for the country factor in Model 4 (5). DUMSEC is a dummy variable that
takes the value one if the analyst follows stocks in more than one country; otherwise it takes the value zero.
DAYSELAPSED
measures the length of time in days between the last earnings forecast by any analyst of firm j’s fiscal
year t earnings and analyst i’s forecast of fiscal year earnings. FORHOR measures the number of days from the date on
which analyst i’s forecast fiscal year t earnings for firm j and the last day of fiscal year t. FORFREQ is a proxy for the
intensity with which an analyst covers a firm. It is estimated by the number of times analyst i issues forecasts for firm j
during fiscal year t. GENEXP measures the number of days between an analyst’s first forecast in the IBES database and
her last forecast for year t. BRSIZE
represents the number of analysts working for analyst i’s employer during fiscal year
t in which she issues a forecast of firm j’s earnings. COMPANIES
is a count of the number of stocks in analyst i’s
research portfolio in year t. The final control variable LFIRMMV is the natural logarithm of average equity market
capitalization of firm j during fiscal year t during which it is covered by analyst i. All control variables are relative with
the exception of LFIRMMV and, as such, have the prefix “R” preceding their names. The p-values are in parentheses
and reflect standard errors that are corrected for the cluster sample problem arising from multiple observations across
years for each analyst. Subscripts have been dropped in the table below for ease of presentation.
a
,
b
, and
c
indicate
significance at the 0.01, 0.05, and 0.10 levels, respectively.
Core Selection Overall Selection
Model 1 Model 2 Model 3 Model 4 Model 5
OLS
(N =45,947)
Heckman
(VARSECTOT)
(
N = 43
,
794
)
Heckman
(VARSECSYS)
(
N = 43
,
794
)
Heckman
(VARSECTOT)
(
N = 44
,
999
)
Heckman
(VARSECSYS)
(
N = 44
,
999
)
Intercept 0.600
a
(0.00)
0.584
a
(0.00)
0.590
a
(0.00)
0.575
a
(0.00)
0.582
a
(0.00)
Year Dummies
Yes Yes Yes Yes Yes
DUMSEC
0.002
(0.38)
0.017
b
(
0.02)
0.012
c
(0.10)
0.024
a
(0.00)
0.019
b
(0.01)
RDAYSELAPSED
-0.070
a
(0.00)
-0.078
a
(0.00)
-0.078
a
(0.00)
-0.074
a
(0.00)
-0.074
a
(0.00)
RFORHOR
-0.231
a
(0.00)
-0.232
a
(0.00)
-0.232
a
(0.00)
-0.232
a
(0.00)
-0.232
a
(0.00)
RFORFREQ
0.039
a
(0.00)
0.039
a
(0.
00)
0.039
a
(0.00)
0.040
a
(0.00)
0.040
a
(0.00)
RGENEXP
-0.015
a
(0.00)
-0.019
a
(0.00)
-0.018
a
(0.00)
-0.019
a
(0.00)
-0.018
a
(0.00)
RBRSIZE
-0.009
a
(0.00)
-0.008
b
(0.02)
-0.008
b
(0.01)
-0.007
b
(0.04)
-0.008
b
(0.03)
RCOMPANIES
0.004
(0.42)
0.006
(0.16)
0.006
(0.17)
0.005
(0.26)
0.005
(0.27)
LFIRMMV
0.015
a
(0.00)
0.016
a
(0.00)
0.015
a
(0.00)
0.016
a
(0.00)
0.016
a
(0.00)
Adj. R
2
(%) 9.20
χ
2
3,137.44
a
(0.00)
3,135.25
a
(0.00)
3,176.67
a
(0.00)
3,174.19
a
(0.00)
Inverse Mills
Ratio
-0.010
b
(
0.02
)
-0.007
(
0.15
)
-0.014
a
(
0.00
)
-0.011
b
(
0.02
)
53
Appendix A: Extraction of Country and Sector Factors
We followed the methodology described in Heston and Rouwenhorst (1994) to extract
country and sector factors. The underlying model for weekly stock returns for security i that
belongs to country j and sector k is assumed to be as follows:
itktjttit
R
εγβα
+++=
,
where
t
α
is the base level of return in period t,
jt
β
and
kt
γ
are the country and sector effects
respectively, and
it
ε
is the firm specific component of returns. The firm-specific component of
returns has zero mean and finite variance and is uncorrelated across securities.
Our data consists of weekly returns for firms in 45 countries and 11 sectors.
25
For each of
the 104 weeks in our sample period, we estimated the following cross-sectional dummy variable
regression using weighted least squares:
,.................
1010221144442211 ititititititittit
SSSCCCR
ε
γ
γ
γ
β
β
β
α
+
+
+
+
+
+
+++=
where
)(
ikij
SC is dummy variable that takes the value of one if the firm belongs to country j
(sector k), and zero otherwise. The weights employed in the regressions are the market
percentage capitalizations of the securities at the beginning of each week.
The number of country (sector) dummies in the regression is one less than the number of
countries (sectors) in our sample. Following Kennedy (1986) and Heston and Rouwenhorst
(1994), the country and sector factors for the remaining country and sector were estimated using
the following two restrictions:
25
To decompose returns, we used a subset of the Datastream Master file that imposes additional screens to
capture either obvious data entry errors or stocks that trade very infrequently. Firms were dropped from
the sample if they did not meet the following screens: (i) returns and market capitalization data were
available for less than half the sample period and (ii) more than half the available market capitalization
and returns data had zero values. The weekly returns data on the resulting sample of firms was used to
extract global, country, and sector factors.
54
,0;0
11
1
45
1
==
== k
ktkt
j
jtjt
ww
γβ
where w
jt
(w
kt
) is the ratio of the market capitalization of all firms in country j (sector k) and the
market capitalization of all the firms in the sample in week t. This procedure enabled us to
construct a time-series of 104 weekly returns on global, sector, and country factors.
26
26
As a check, we replicate Table 3 in Heston and Rouwenhorst (1994) for a sample consisting of 12 European
countries and 9 sectors. Specifically, we decompose the variance of the excess return of a country index into a
pure country factor and a component that is the variance of the sum of the 9 sector factors. We also decompose
the variance of the sector index return into a pure sector effect and sum of the 12 country effects. These
decompositions yield results similar to what they report in their paper.
    • "F-tests on the equality of the coefficients on CHAIN_ANALYST and IND indicate that the difference is statistically significant (p , 0.01) in all regression specifications, consistent with analysts following supply chain specialization generating significantly more accurate and less optimistically biased forecasts for the supplier firms than industry specialists following the same supplier firms (H3). The signs and significance levels for the effects of other analyst and firm characteristics on forecast performance are generally consistent with those reported in prior research (e.g., Mikhail et al. 1997; Clement 1999; Jacob et al. 1999; Kini et al. 2009; Sonney 2009), except for the unexpectedly positive, but insignificant, coefficient on DM_logNFIRM in the ACCURACY regression. In terms of the economic significance of these results, the coefficient of 0.0517 on CHAIN_ ANALYST the ACCURACY regression in column (1) ofTable 4, for example, suggests that holding various firm, analyst, and forecast characteristics constant, supply chain analysts are 0.05 percent more accurate than NONE ¼ 1 analysts, which represents 34.5 percent (0.0517 percent/0.15 "
    [Show abstract] [Hide abstract] ABSTRACT: This study examines the antecedents and consequences of analysts choosing to become supply chain analysts (i.e., analysts following both a supplier and its major customer). We find that information complementarities between firms in the same supply chain, between a supplier firm and its industry peer firms, and between the supplier’s major customer and other firms in analysts’ portfolio affect their supply chain specialization decision. The potential revenues supplier firms generate for analysts’ brokerage houses also significantly affect this decision. While supply chain analysts achieve superior forecast performance compared to non-supply chain analysts for supplier firms, they provide lower-quality forecasts for other firms in their portfolios. These findings suggest that analysts allocate resources strategically. Our results are robust to techniques designed to address the potential endogeneity of analysts’ supply chain portfolio choices
    Full-text · Dataset · Feb 2016 · The Accounting Review
    • "Palepu (1985) and Rumelt (1982) utilize complete data to investigate the same issue and reach a conclusion similar to that presented by Gort. In terms of literature on product diversification of a firm and earnings forecast accuracy, Kini et al. (2009) indicate that the relation between sector diversification and forecast accuracy is context‐ specific. In an international context such as European Zone, forecast accuracy may increase with sector diversification, whereas in the U.S., analyst may generate higher accuracy when focus on a specific industry. "
    Full-text · Article · Dec 2015
    • "Following prior studies, all analyst and forecast characteristics are adjusted by subtracting the mean value of the variable over the firm-year to eliminate firm-year effects (e.g., Clement 1999). characteristics on forecast performance are generally consistent with those reported in prior research (e.g., Mikhail et al. 1997; Clement 1999; Jacob et al. 1999; Kini et al. 2009; Sonney 2009), except for the unexpectedly positive but insignificant coefficient on DM_logNFIRM in the ACCURACY regression. [Insert TABLE 4 about here] In terms of the economic significance of these results, the coefficient of 0.0517 in the ACCURACY regression (column (1) ofTable 4), for example, suggests that holding various firm, analyst, and forecast characteristics constant, supply chain analysts are 0.05% more accurate than NONE = 1 analysts, which represents 34.5% (0.0517%/0.15%=34.5%) of the mean forecast accuracy of NONE =1 analysts for the supplier firms. "
    [Show abstract] [Hide abstract] ABSTRACT: This study examines the antecedents and consequences of analysts choosing to become supply chain analysts (i.e., analysts following both a supplier and its major customer). We find that information complementarities between firms in the same supply chain, between a supplier firm and its industry peer firms, and between the supplier’s major customer and other firms in analysts’ portfolio affect their supply chain specialization decision. The potential revenues supplier firms generate for analysts’ brokerage houses also significantly affect this decision. While supply chain analysts achieve superior forecast performance compared to non-supply chain analysts for supplier firms, they provide lower quality forecasts for other firms in their portfolios. These findings suggest that analysts allocate resources strategically. Our results are robust to techniques designed to address the potential endogeneity of analysts’ supply chain portfolio choices.
    Full-text · Article · Sep 2015
Show more