chinese walls in german banks - tu wien
TRANSCRIPT
Chinese Walls in German Banks�
Alfred Lehar�
Department ofBusiness Studies
University of Vienna
Otto Randl�
Department ofBusiness Studies
University of Vienna
September 6, 2001
�We gratefully acknowledge the contribution of I/B/E/S International Inc. for providing earnings per share
forecast data, available through the Institutional Brokers Estimate System. This data has been provided as partof a broad academic program to encourage earnings expectations research. We thankfully acknowledge financialsupport from the Austrian National Bank (ONB) under the Jubilaumsfonds grant number 8523. We thank HelmutElsinger, Evelyn Hayden, Yishay Yafeh and Josef Zechner for helpful comments and Eva Smolen for excellentresearch assistance.�
Department of Business Studies, University of Vienna, Brunner Straße 72, A-1210 Vienna, Austria. e-mail:[email protected], Tel:+43 - 1 - 4277 38077, Fax: +43 - 1 - 4277 38074�
Department of Business Studies, University of Vienna, Brunner Straße 72, A-1210 Vienna, Austria. e-mail:[email protected], Tel:+43 - 1 - 4277 38081, Fax: +43 - 1 - 4277 38074
Chinese Walls in German Banks
ABSTRACT
In a universal banking system, information advantages but also conflicts of interest may
arise within a bank. We analyze the extent of these effects in a specific setting. Analysts in
a bank’s research department regularly issue reports on firms having no relationship to the
bank as well as on firms where the bank has a strategic interest. Officially banks have to
establish a so-called Chinese Wall around the research department to allow the analysts to
work independently, and to avoid the flow of insiders’ information. Using ownership data
and analysts’ earnings per share forecasts on German companies from 1994 to 1999, we
test for both informational advantages and possible conflicts of interests. Our findings are
consistent with banks having superior knowledge on firms where they own equity stakes.
We also find evidence for conflicts of interest. These can arise from pressure from clients
to announce favorable forecasts and from a bank’s self interest to increase the market
value of its equity stake in the firm. Interestingly, we find that conflicts of interest are less
pronounced for large equity holdings, leading to the interpretation that client pressure
more than bank’s self interest with respect to equity valuation is the important driving
force of overoptimistic forecasts.
JEL Codes: G19, G21, G24
Keywords: Chinese walls; analyst forecasts; universal banking
1
I. Introduction
In contrast to the Anglo-American capital markets, banks in Germany are not restricted in their
activities. They participate in loan and deposit markets as well an in the securities markets,
where they act as brokers, dealers and hold long term equity stakes in non-financial firms.
Proponents of the concept of universal banking argue, that banks can now offer their clients an
extensive range of financial services, which may result in a very intense bank-client relation-
ship. This tight cooperation of a firm with a bank, the ”Hausbank”, may mitigate information
asymmetries and prove to be mutually beneficial. The main drawback of a universal banking
framework as it is implemented in Germany is, that conflicts of interest are more likely to arise
between the different departments of the bank and their respective customers.
In this paper we want to examine the existence of conflicts of interest in a specific setting.
Analysts in the banks research department regularly issue reports on firms having no rela-
tionship to the bank as well as on firms which are customers of the banks or where the bank
has a strategic interest (e.g. equity stake). Officially banks have to establish a Chinese Wall
around the research department for two reasons. First to shield analysts and allow them to
work independent, and second to ensure that analysts do not incorporate confidential insider
information in their reports. While the second point is quite obvious, let us have a closer look
at the first issue. Several groups might exert pressure to influence the analyst’s report. Loan
officers might be concerned to loose lending volume to rival banks after a downgrade in an
analysts report, members of the investment banking division can fear the loss of future deals,
and the bank itself might be afraid of its equity stakes loosing value.
In our paper we want to examine how well built these Chinese Walls are and in what
direction information flows are more likely to occur. We therefore test whether housebanks’
analysts publish biased recommendations, indicating that pressure is put on analysts, and we
also explore, if the housebanks’ analysts reports are more accurate than the reports from rival
banks, which would be indicative of inside information.
2
There are two streams of literature related to our research. The first one examines a pos-
sible conflict of interest in analyst forecasts.1 Michaely and Womack (1999) find that under-
writer analyst recommendations for IPO firms are less accurate and positively biased. Dunbar,
Hwang, and Shastri (1999) also support the hypothesis of conflict of interest for underwriters
buy recommendations made shortly after the IPO. However, non-initial buy recommendations
by underwriters lead to positive stock market reactions, indicating superior information of
these analysts. Lin and McNichols (1998) show that analysts affiliated with investment banks
underwriting a seasoned equity offering issue more favorable recommendations, while their
forecast of subsequent earnings is not biased.
Most of these studies focus only on conflicts of interests within investment banks. The
situation might be different for banks operating under an universal banking paradigm, which
is examined in the second branch of literature related to our research.2 Analyzing data from the
pre-Glass-Steagall area when commercial banks as well as investment houses were allowed
to engage in underwriting, Kroszner and Rajan (1994) show that bank underwritten securities
performed better and defaulted less than investment bank underwritten issues. Puri (1996) and
Gande, Puri, Saunders, and Walter (1997) find evidence that commercial banks could sell the
issues they underwrote at a higher price, indicating a certification role for commercial banks. 3
Gompers and Lerner (1999), Hamao and Hoshi (2000) and Klein and Zoeller (2001) examine
the venture capital, the Japanese bond market and German IPOs, repectively. Using recent
data they find that securities underwritten by bank affiliates are issued at a discount, possibly
1Other studies examine forecast accurracy (Brown and Rozeff (1978), Stickel (1992)), analyst coverage(Barth, Kasznik, and McNichols (forthcoming)), analyst herding and career concerns (Hong, Kubik, andSolomon (2000), Hong and Kubik (2001), Welch (2000)), the profitability of analysts’ recommendations (Wom-ack (1996), Barber, Lehavy, McNichols, and Trueman (2001)), institutional ownership (Ackert and Athanassakos(2000)), and biases in analyst forecasts due to over- and underreaction (Easterwood and Nutt (1999)) or due tobetter access to firm management (Lim (2001)). Some papers providing evidence on analyst behavior outside theU.S. are Chang, Kahanna, and Palepu (2000), Bolliger (2001), and Capstaff, Paudyal, and Rees (1998)
2In most countries with universal banking, also the corporate governance structures differ substantially fromthe U.S. This is also the case in Germany. The majority of German publicly listed companies has a dominantshareholder, and banks exert substantial influence in the governance process, as reported by Franks and Mayer(2000) and Boehmer (1999). Thus, close ties exist in two levels: first, through the heavy weight of banks in thegovernance structure, and second, because of the German housebank system.
3Using data on bond issues in the UK, Hebb and Fraser (2001) finds no evidence of conflict of interest.
3
because of a conflict of interest. Ber, Yafeh, and Yosha (2001) find conflict of interest in the
Israeli IPO market and Narayanan, Rangan, and Rangan (2001) find banks to commit against
opportunistic behavior by only engaging as co-managers in a syndicate.
This paper contributes to the literature by analyzing analyst forecast ability in an environ-
ment with tight relationships between corporations and banks. Our main findings are consis-
tent with an information advantage of banks which such a relationship. We also find evidence
for a conflict of interest within banks owning equity stakes of up to 25 percent in the compa-
nies for which they make earnings forecasts. For these banks the main incentive for eventually
suppressing bad news will generally come from client pressure. We do not find evidence
for a conflict of interest within banks owning larger equity stakes, leading to the conclusion
that a bank’s self interest to increase the value of its equity stake is not an important factor
influencing forecasts.
The rest of the paper is composed as follows: Section II derives the testable hypotheses,
Section III describes the sample, the results of the empirical analysis are presented in Section
IV, additional robustness checks are implemented in Section V and Section VI concludes.
II. Hypotheses
It is very hard to define a good measure for the intensity of bank relationships in Germany,
since they are often of multiple nature. Firms borrow money from banks, who in turn hold
equity of these firms, have seats on the supervisory boards,4 and exert a lot of influence through
proxy votes in the shareholders’ meeting.5 We chose ownership to measure the degree of
relationship between a bank and a firm for two reasons. First it is well defined and it is
possible to collect a consistent data set and second, because it is a conservative measure. A
4The ”Aufsichtsrat” of German firms consists of representatives of the firm owners and the employees and isappointed by the shareholder meeting, see Becht and Boehmer (1999) for an excellent survey on the institutionalfeatures of the German governance system.
5See e.g. Franks and Mayer (2000) for a detailed analysis.
4
higher equity stake will imply a higher portion of votes in the firm’s shareholder meetings and
a higher number of seats on the management board. The actual ownership in a company also
often underestimates the actual direct control rights, that a bank has on a firm. The classical,
pyramid type governance structure in corporate Germany allows banks to control companies
even with holding a small equity stake. For example a bank may hold 51% of company A,
which in turn holds 51% of company B. The bank then controls B but only holds 26% of
B’s equity. Thus if we find that the actual equity ownership has a significant influence on the
analysts’ behavior we will expect more of that bias to be there in the economy.
We define an analyst being adjunct to a firm as an analyst whose broker firm owns a sig-
nificant stake in that firm. To capture the effect of substantial ownership, we discriminate
between large equity stakes (25% or above) and small equity stakes (below 25%). A stake-
holder of more than 25% has significant influence under German corporate law as he may
e.g. block a merger or liquidation, inhibit the dismissal of members of the supervisory board,
prevent the issuance of new shares, or hold back any changes in the firms charter.
If the Chinese Walls are leak we can test for influences in two directions. First the analyst
might be pressured to issue biased reports in order not to jeopardize the banks client relation-
ships or lower the value of the bank’s equity stake, and second the analyst may benefit from
insider information and publish more accurate estimates of future firm performance. There-
fore we test, if the reports of adjunct analysts contain a bias relative to the analysts in the
control group as well as for the precision of the analyst’s estimates. Finally, as we will explain
in more detail in section IV, we also test, whether an individual analyst’s forecast contains
superior information relative to the consensus estimate, i.e. it is an improvement relative to
the consensus.
When there is support for the ”Conflict of interest hypothesis”, we will find adjunct an-
alyst forecasts to be upward biased. With respect to the consensus forecasts, we should find
adjunct analysts to report their information mainly in those situations, where the consensus
underestimates actual earnings per share. Table I summarizes these main hypothesis.
5
Table IMain Hypothesis
To test for conflict of interest, we look whether forecasts that are issued by brokers owning an equitystake are biased relative to the consensus. Under this hypothesis, forecasts by those brokers containrelevant information mainly if the consensus underestimates actual earnings. Superior informationshould result in more precise forecasts. We expect these brokers to contribute valuable informationregardless of the consensus estimate.
Hypothesis Bias Precision Information contentConsensus Consensusoverestimates underestimates
Conflict of Interest upward No YesSuperior Information superior Yes Yes
As a refinement we will distinguish, whether the conflict of interest arises because of client
pressure or stems from the bank’s self interest. The amount of influence, that firms have on the
bank will depend on the fraction of equity owned by the bank. When banks are the dominant
shareholders, the management can not credibly threaten the bank. The only reason to publish
overoptimistic reports is to increase the value of the bank’s equity stake. When the bank does
not have a dominant share in the company, the firms’ management will be more independent
and optimistic reports are more likely to be published to maintain a beneficial bank-client
relationship. We will therefore also test, whether analysts of banks with large stakes behave
differently than their colleagues of banks with minor stakes.
If adjunct brokers have superior information, we will expect their forecasts to be more pre-
cise than the reports from rival banks’ analysts, whereas the bias should not necessarily differ
from the control group.6 Informed analysts will generally improve the consensus forecast.
That is, if the consensus currently overestimates earnings, we expect the informed broker to
issue a relatively lower estimate, whereas an underestimating consensus will be pushed up-
wards. We refer to this theory as the ”Superior information hypothesis”.
6As the bias may be positive or negative in different years, a smaller bias will not necessarily imply a higherdegree of precision.
6
III. Data
We examine the period from 1994 to 1999. The data needed for our analysis consists of
two building blocks: ownership data and analyst forecasts. We collect data of equity stakes
in publicly traded German companies for nine large German banks known to be actively in-
volved in both strategic stock investments and analyst forecast activities. These banks are Bay-
erische Landesbank, BHF Bank, Commerzbank, Deutsche Bank, DG Bank, Dresdner Bank,
HypoVereinsbank, Nord/LB and WestLB.7 For wholly owned brokerage subsidiaries of these
nine banks, we assume the same bank-firm relationship between the broker and a firm as be-
tween the bank which is parent company and the firm. We start collection of ownership data in
1993, using banks’ annual reports where they disclose major equity stakes in other companies.
HypoVereinsbank has come into existence in 1998 only, before this point in time we use data
for Bayerische Hypotheken- und Wechselbank and Bayerische Vereinsbank separately. We
cross-check our data with the database wer gehort zu wem?, which is published regularly by
Commerzbank. We focus on the percentage of equity in a specific company owned by one
of the above mentioned banks. We include direct as well as indirect equity holdings in our
analysis.8 From these ownership data, we construct dummy variables reflecting equity stakes
up to 25 percent and ownership in stakes at least as large as 25 percent.9
For analyst forecasts, we use the I/B/E/S International Detail History database, as of Au-
gust 2000. This database collects analyst forecasts on a forecast by forecast basis (as opposed
to consensus forecasts). We use annual earnings per share (EPS) forecasts only. For each
7Based on balance sheet information from the year 2000 (for Nord/LB 1999) and the reports of the DeutscheBundesbank the banks in our sample have a market share (total assets) of 18,8%. They own equity stakes of114,7 billion Euro (book values) which is more than 88% of all equity owned by German banks.
8If bank A owns 50 percent of company B which in turn owns 50 percent of company C, a 25 percent equitystake of bank A in company C is reported.
9Boehmer (1999) estimates that for a typical German bank the value of the loan portfolio to a typical Germancorporation exceeds the value of the equity stake by a factor between 7 and 20, which might induce banks toact in order to maximize debt value and not equity value. This will be less pronounced in cases of large equityownerships. The use of dummy variables instead of the actual stakes is justified by the observation that banks’officially reported shareholdings frequently do not reflect the true extent of control exercised (see Lehmann andWeigand (2000)). Franks and Mayer (2000) also use a cutoff point of 25 percent in their analysis of corporategovernance in Germany.
7
forecast, I/B/E/S provides numerous details, from which we select the broker who made the
forecast, the value of the forecast made, the estimate date, the forecast period end date (defined
as end of the fiscal year for which the forecast is made) and the later reported, actually real-
ized value. We exclude forecasts made by Boerse online, Das Wertpapier and Going Public
Media, as these companies are publishing houses or information providers and therefore not
comparable to stock brokers or banks. We group our data with respect to the time lag between
the forecast and the forecast period end date. We define short term forecasts as those being
made from 3 months before to 3 months after the forecast period end date. Forecasts made
up to three months after the period end date are included, because actual earnings per share
for a given year are typically reported several months after the fiscal year ends. Medium term
forecasts are being made from four months up to one year before the period ends; and finally
long term forecasts are those being made longer than one year, but no longer than three years
before the end of the forecast period. This restriction to forecasts up to three year is made
because very long term forecasts exhibit considerable noise and are rare. During the sample
period, numerous companies and analysts switched from reporting in DEM to EUR. We have
converted all forecasts and actual earnings per share into EUR.
While it is clear to identify the sample of forecasts from adjunct analysts, there are several
possibilities to identify the correct control sample. One possibility would be to use all forecasts
where such an ownership relation does not exist. However, this would lead to an extremely
unbalanced sample. Therefore, forecasts of adjunct analysts will be compared with two control
groups of forecasts.
First, a comparison can be made with forecasts for the same set of firms issued by non-
adjunct brokers. To construct this sample, we select all companies where at least one of the
five above mentioned banks has an equity stake. For these firms, we use forecasts from all
brokers that do not have an equity stake as control for the adjunct analysts. The sample that
includes this comparison group will be referred to as all banks sample. This first sample could
8
suffer from a bias if the nine brokers for which we have collected ownership in firms exhibit
some kind of specialness.
Second, a relevant comparison group are forecasts of banks for companies where they do
not own equity stakes. Therefore, we select a second sample including all forecasts by the
nine banks for all German companies. If forecasts are to be found different for corporations
where the analysts’ houses own equity stakes, this can not be only due to the special nature of
the banks making the forecasts, as only forecasts by these banks are used as a control group.
This dataset will be referred to as all firms sample. This sample could suffer from a bias
if the companies where bank equity investment exists have special characteristics which e.g.
make it harder to issue correct forecasts. Implementing the analyses for both samples will be
sufficient to distinguish whether any results might be driven by a sample selection bias. Figure
1 illustrates the composition of the samples. 10
We have followed the procedure described below to construct the datasets used for empir-
ical analysis. For the all banks sample, we use all forecasts from the I/B/E/S data which have
been made for the group of German companies where one of the nine banks mentioned above
owns an equity stake and issues at least one forecast. We restrict the sample to annual earn-
ings per share forecasts made for the years from 1994 to 1999. We then drop all observations
where actual earnings per share are not available. We furthermore eliminate all observations
concerning forecasts made later than three months after and earlier than three years before the
forecast period end date. We also purge the dataset from observations where the forecast EPS
is lower than -1000 Euro, and where the currency is neither DEM nor EURO. This leaves us
with 32663 observations. For the all firms sample, we use all forecasts for German companies
issued by the nine mentioned banks. We then follow the same procedure as for the all banks
sample, which leaves us with 36952 observations in the all firms sample.11
10Note that even in the area at the bottom left not all observations are forecasts made by an adjunct broker, aseach one of the banks (and its broker subsidiaries) generally only owns equity stakes in a few on the 46 companieswhere at least one bank is owner.
11Following this procedure, we eliminate approximately 20 percent of the initial samples.
9
Figure 1. Samples Used for Analysis.
������������ ����������������������� ��� � !#"$&%�' � !)(*�+!,�-� $ � . �/������� �102� 3,4-� $ 3,�5" 4�6 78� ��9/� 4 $&%8' � !)(
:;<=> ?@A
B CD EFG HI JK
L M NO PHNQ EG
RCSD EFG HI JK NPL M NO PHNQ EGK I T
UWVXVZY\[^]`_badce[^f ghVjikllmnopqrtspul�v
IV. Empirical Analysis
In the literature, there exist several approaches for the correct measurement of the forecast
error. As we will see from the summary statistics, the forecast error, measured in percent of
the actual value, is a noisy variable. Especially if actual earnings are close to zero, already a
small deviation in monetary terms can lead to a huge percentage error. This problem can be
solved by using the difference between forecast value and actual value for the forecast error.
This definition, however, would be misleading as pure size effects would influence the results
and as the definition is not invariant to stock splits. Another approach would be to scale the
forecast error by the stock price. Theoretically this is appealing as size effects as well as huge
outliers stemming from earnings close to zero could be avoided. However, the stock price is
10
not independent from earnings and earnings forecasts. The use of historic stock quotes, say
3 years before end of forecast period, brings data problems (availability of price history for
young firms, furthermore history of stock splits would be needed); using more recent data
brings the risk of endogeneity problems.12
We have therefore decided to use the percent forecast error, defined as
����������� ��� ���������������������� ����������� � (1)
and the absolute percent forecast error��������� � ����� �
, where��� �������
is the individual
forecast, and�����������
the actual value.
To be able to assess the quality of individual forecast errors, we need an appropriate bench-
mark. We therefore calculate the consensus forecast prevailing at time � as the median of all
individual forecasts issued or explicitly confirmed as accurate between the period of three
months and two days before � .13 We define the percent forecast error of the consensus forecast
(�������
) and the absolute percent forecast error of the consensus (���������
) in an analogous
way to individual forecast errors.
To limit the problems arising from outliers, we use robust estimation techniques. As the
median is less influenced by outliers than the mean, we explain the median forecast error in
our regressions. 14 In the context of earnings forecasts the use of the median is not uncommon,
and is frequently reported in consensus forecasts.
12Some authors, e.g. Easterwood and Nutt (1999) scale errors with stock prices, but do not encounter theproblems outlined above because they only consider forecasts issued during a short time interval.
13We calculate the consensus forecast before assigning forecasts to the all firms or the all banks subsample,using all forecasts except those issued by Boerse online, Das Wertpapier and Going Public Media. Note thatfor calculation of consensus forecasts, we make use of the whole rectangle of Figure 1. We then assign to eachforecast in one of the two samples the consensus forecast prevailing at the time when the individual forecast wasissued.
14Using median regression, we want to make sure that our results are not influenced by the way outliers areeliminated (see also Ahmed, Lobo, and Zhang (2001)). Gu and Wu (2001) find empirical evidence that analystsforecast median EPS rather than mean EPS.
11
A. Summary Statistics
Table II shows the summary statistics for the various variables. Inspection of the panel refer-
ring to the all banks sample in Table II indicates that forecasts and realized values of Earnings
per Shares (EPS) are indeed noisy variables. The median forecast EPS is 1.48 Euro, while
the actual EPS over the period is only 1.18 Euro. The median forecast error is 2.17 percent.
This shows that analysts tend to overestimate earnings, a feature documented for instance by
Chopra (1998). A measure for forecast variability is the median absolute forecast error, which
equals 22 percent. The median time difference between a forecast and the forecast period end
date is 0.88 years (which is 320 days). 5.1 percent of forecasts are made by a broker who
owns a small equity stake, and additional 0.8 percent of forecasts are made by brokers who
own equity stakes larger than 25 percent. The median size of a broker is 37.8 percent, where
size of broker is defined as the percentage of German firms covered by a particular broker to
the total number of German firms in the I/B/E/S international database in a specific year. The
median coverage is 57.1 percent, where coverage is the percentage of brokers in the I/B/E/S
international database that issue at least one forecast for a specific firm in a year, within all
brokers that issue at least one forecast for at least one German firm. The summary statistics
for these variables in the second panel, referring to the all firms sample, are very similar. The
median forecast EPS is 1.69 Euro, the median actual EPS 1.16 Euro. The forecast error is con-
siderably higher, with a median of 24 percent. Approximately 5 percent of forecasts are made
by adjunct brokers. The median size of brokers is larger (46 percent), in line with the sample
selection strategy of investigating forecasts by the nine large German banks mentioned. The
median coverage of the firms in this sample is lower, due to the larger number of firms (and
therefore inclusion of smaller companies) compared to the all banks sample.
12
Table IISummary Statistics
Forecast and actual earnings per share (EPS) are given in Euro. The forecast error and the absoluteforecast error are stated in percent. Time difference, measured in years, is the difference between thedate when a forecast is made and the last day of the period for which the forecast is made (period enddate). Small stake dummy is one if a broker owns a stake less than 25 percent in the company for whichthe forecast is made and zero otherwise. Similarly, the large stake dummy is one when a broker ownsa stake of at least 25 percent in the company whose EPS are forecast. Size of broker is the number ofGerman companies in the I/B/E/S database covered by a broker divided by the total number of Germancompanies in the database, calculated annually. Coverage is the number of brokers issuing forecasts fora company divided by the total number of brokers issuing forecasts for German companies. Coverageis calculated for each year in the sample.
All Banks Sample. ( ������������
)Variable Mean St. Dev. Median Min MaxForecast EPS 2.323 4.468 1.483 -114.274 45.830Actual EPS -0.001 11.705 1.176 -104.805 26.587Forecast error 81.226 1392.726 2.169 -9179.248 71480.86Absolute error 115.007 1390.344 22.125 0 71480.86Consensus error 94.484 1506.164 2.161 -5899.513 56142.11Consensus abs. error 119.264 1504.404 22.283 0 56142.11Time Difference 0.877 0.627 0.843 -0.249 2.645Small stake dummy 0.051 0.220 0 0 1Large stake dummy 0.008 0.091 0 0 1Size of broker 0.378 0.149 0.385 0.002 0.736Coverage 0.571 0.179 0.589 0.011 0.870
All Firms Sample. ( ����������
)Variable Mean St. Dev. Median Min MaxForecast EPS 3.337 6.997 1.687 -105.837 143.162Actual EPS 1.426 9.800 1.161 -104.805 109.060Forecast error 139.498 12069.34 24.096 -87651.16 607437.3Absolute error 162.144 12041.01 39.702 0 607437.3Consensus error 157.401 1379.021 27.481 -5899.513 56142.11Consensus abs. error 174.865 1376.916 40.525 0 56142.11Time Difference 0.879 0.637 0.851 -0.249 2.678Small stake dummy 0.045 0.208 0 0 1Large stake dummy 0.007 0.085 0 0 1Size of broker 0.450 0.106 0.455 0.002 0.687Coverage 0.508 0.207 0.517 0.014 0.926
13
B. Intensity of Forecast Activity
Table III compares the median number of forecasts issued by adjunct and non-adjunct brokers.
The first number in each row of the table is the median number of forecasts per broker for a
given company over the respective horizon. The second number in parentheses is the number
of independent observations used for calculation of these medians. From both samples can be
seen, that the number of forecasts issued by a broker is increasing with the forecast horizon. As
longer intervals are chosen for more long term forecasts, this is not surprising. However, it can
be seen that adjunct analysts issue forecasts more frequently than their non-adjunct colleagues.
This can clearly be seen in both subsamples for all horizons in the case of adjunct brokers with
small equity stakes up to 25 percent. While non-adjunct brokers issue a median number of 3
short term, 6 medium term and 7 long term forecasts, adjunct brokers with small stakes issue
4 short term, 7 medium term and 9 long term forecasts. Brokers with very large equity stakes
seem more reluctant to issue frequent short term forecasts. The median number of forecasts
in the all firms control sample is 2 short term, 7 medium term and 12 long term forecasts.
These numbers show that adjunct brokers are slightly more active in issuing forecasts than
independent ones; but the difference is not huge.
C. Explaining the Forecast Error
To see whether there is any systematic difference of forecasts of adjunct brokers to the control
group, we report the forecast error within a grid of forecast horizon and size of the equity
stake. Table IV reports the median forecast error for short term, medium term and long term
forecasts, distinguishing between large, small and no equity stake. We can see that the forecast
error increases quickly with the forecast horizon for all groups of brokers. The median fore-
cast error for long term forecasts is approximately 50 percent for adjunct brokers owning large
stakes and -5 percent for adjunct brokers owning small stakes. In the all banks control group
the median error is near 5 percent, in the all firms sample approximately 44 percent. While the
14
Table IIIMedian Number of Forecasts per Broker.
This Table shows the median number of forecasts per broker within a given time interval (short term,medium term or long term) for a specific company. The number in parentheses is the number ofindependent observations used to calculate the median.
Both Samples.Forecast horizon
short term medium term long term(-3 to +3 months) (4 to 12 months) (1 to 3 years)
no equity stake 3 6 7all banks sample (1557) (1727) (1691)no equity stake 3 6 7all firms sample (1789) (1933) (1726)adjunct brokers 4 7 9small stakes ( � ����
) (60) (79) (76)adjunct brokers 2 7 12large stakes ( � ����
) (11) (13) (9)
extent of the error is smaller for shorter horizons, the general picture is similar. This first com-
parison shows important differences between the various groups. However, interpretation has
to be cautious due to two main reasons. First, the sample of firms for which forecasts are made
is different for every row in table IV. Comparisons between rows are therefore problematic.
To avoid that the analysis is driven by market optimism for a specific set of firms, we need to
know the forecasts of adjunct brokers relative to the market opinion. Therefore, in the subse-
quent analysis, we will explain the precision and possible bias of individual forecasts relative
to the consensus forecast prevailing at the time when an individual forecast has been made.
The consensus reflects the public available information at the time prior to the forecast and
therefore allows us to evaluate conditional biases or precision measures. Second, any univari-
ate analysis neglects the influence of potentially important explanatory variables. Therefore,
we will now turn to a multivariate analysis to further explore the relationship between the
forecast error and the explanatory variables.
15
Table IVMedian Forecast Error.
This Table shows the median forecast error by equity stake groups and time horizons (short term,medium term or long term). The number in parentheses is the number of independent observationsused to calculate the median.
Both Samples.Forecast horizon
short term medium term long term(-3 to +3 months) (4 to 12 months) (1 to 3 years)
no equity stake 0.000 2.437 5.042all banks sample (5984) (11574) (13163)no equity stake 9.091 21.429 43.722all firms sample (7063) (12970) (14977)adjunct brokers -2.109 -2.339 -5.772small stakes ( � ����
) (287) (626) (758)adjunct brokers 2.273 11.711 50.000large stakes ( � ����
) (47) (90) (134)
As the inspection of the summary statistics indicated, many variables are skewed (the
mean is large off the median) and the data include numerous extreme realizations. To reduce
the impact of these outliers, we explain the median forecast error, conditional on a number
of explanatory variables. This median regression is a special case of quantile regression,
where the 50 percent quantile is explained. The estimator is also called LAD (least absolute
deviation), as it minimizes the sum of absolute residuals from the estimated values.
The choice of the specification can be explained as follows. First, we want to explain
the potential bias in forecasts made by adjunct brokers. Therefore, we explain the individual
forecast error by the error of the consensus forecast (�������
) prevailing at this point of time,
plus control variables as the time difference (����� �
) between the forecast and the period
end date, the coverage of the firm (��� ��� �������
), the size of broker ( � ��� � ), and dummies
stating whether a forecast has been made by adjunct brokers ( � � ��� ��� and � � ��� ��
for small and large stakes respectively). We calculate the consensus forecast prevailing at the
16
time of an individual forecast as the median of all forecasts issued for the same firm and period
within a time window of three months and three days before the individual brokers’s forecast.
����� � ������� ������������ ����� ����� ��� ��� ����������� � ��� ������ � � ��� �������� � � ��� �� ���� (2)
We run this regression in the all banks sample to see whether for the same firms, forecasts
by adjunct brokers are systematically different. We also run this regression setup in the all
firms sample to see whether the same brokers behave differently for firms in which they have
a significant equity stake than for non-related firms. To test the second hypotheses related to
an information advantage, we run similar regressions to explain the absolute forecast error.
Table V reports the median regression results. Not surprisingly, the error of the prevailing
consensus forecast is an important factor influencing the error of individual forecasts, receiv-
ing a weight of 92.5 percent of the estimate of individual forecast errors. Although time
effects are already partially incorporated in the consensus forecast error included as a regres-
sor, the time difference between the issue date of an individual forecast and the company’s
fiscal period end date is highly significant. Individual forecast errors are higher relative to
the consensus for longer forecast horizons. Note that the focus of these regressions is on the
bias of forecasts respectively their optimism / pessimism relative to the consensus. Higher er-
rors do not necessarily mean less precise estimates, as the consensus could also underestimate
actual earnings. However, in most cases the consensus already overestimates, and a higher
forecast relative to the consensus would then also translate into a less precise forecast. More
intense coverage of a firm reduces the median individual forecast error. The forecast errors
of large brokers are smaller than those of their smaller competitors. This is consistent with
previous studies, e.g. Lim (2001). Of interest for our main hypothesis, the coefficients on
the adjunct broker dummies are positive. Brokers with a small stake in a firm issue forecasts
17
with a 0.85 percentage points higher forecast error relative to the consensus forecast than their
independent colleagues. Their views on earnings are more optimistic in an economically and
statistically significant way. The coefficient on brokers with large stakes is positive, but not
statistically significant.
The results from the all banks sample can not in all respects be confirmed by the all firms
sample. Here, larger brokers (out of the reduced number of brokers in this sample) exhibit
higher forecast errors, and the coefficients of the adjunct broker dummies have both negative
sign, although not statistically significantly different from zero. The sample selection in the
all firms sample allows to tell whether the same brokers behave in a different way for firms
where they own equity stakes and others where they do not. However, the regression is not
conclusive on this issue.
Together, the results of these regressions only weakly support the conflict of interest hy-
pothesis. The conflict of interest seems to stem from client pressure rather than self interest,
as significant effects are only found for analyst houses owning small stakes.
To analyze the superior information hypothesis, we want to explain the absolute forecast
error��������� � ����� �
by the following regression:
������� � � ����� �������������� ����� � ���� ��� ��� ����������� � ��� � ���� � � ��� �������� � � ��� �� ��� (3)
where���������
is the percentage absolute forecast error of the consensus forecast. The
results are presented in Table VI.
Here, supporting evidence is strong for both samples. Again, the absolute error of the
consensus forecast plays a decisive role in explaining the magnitude of individual absolute
forecast errors. The longer the time horizon, the less exact forecasts are. The coefficients
18
Table VMedian Regression Explaining the Forecast Error.
This regression explains the percent forecast error of individual brokers’ forecasts. Regressors arethe error of the consensus forecast, the time difference in years between the day of the forecast andthe forecast period end date, the coverage of the firm by brokers in I/B/E/S, the size of the brokerforecasting measured as market share, and dummy variables that take the value of one if the brokerforecasting owns a small (up to 25 %) respectively large (more than 25 %) equity stake in the companywhose earnings per share are to be forecast.
All Banks Sample.Median RegressionNumber of obs = 32663 Pseudo
� � � � � ������Coeff. t-stat
��� � � �Error of consensus estimate 0.925
�100 0.000
Time difference 0.727 5.95 0.000Coverage -2.371 -5.46 0.000Size of broker -1.154 -2.23 0.026Small stake dummy 0.851 2.47 0.014Large stake dummy 0.576 0.69 0.488Constant 1.368 3.94 0.000
All Firms Sample.Median RegressionNumber of obs = 36952 Pseudo
� � � � ��� � �Coeff. t-stat
��� � � �Error of consensus estimate 0.928
�100 0.000
Time difference 1.181 9.80 0.000Coverage -0.749 -2.00 0.045Size of broker 3.031 4.23 0.000Small stake dummy -0.514 -1.42 0.157Large stake dummy -1.105 -1.25 0.210Constant -0.532 -1.33 0.184
19
Table VIMedian Regression Explaining the Absolute Forecast Error.
This regression explains the percent absolute forecast error of individual brokers’ forecasts. Regressorsare the absolute error of the consensus forecast, the time difference in years between the day of theforecast and the forecast period end date, the coverage of the firm by brokers in I/B/E/S, the size ofthe broker forecasting measured as market share, and dummy variables that take the value of one if thebroker forecasting owns a small (up to 25 %) respectively large (more than 25 %) equity stake in thecompany whose earnings per share are to be forecast.
All Banks Sample.Median RegressionNumber of obs = 32663 Pseudo
� � � � � ������Coeff. t-stat
��� � � �Absolute error of consensus estimate 0.925
�100 0.000
Time difference 1.649 16.88 0.000Coverage -0.093 -0.27 0.790Size of broker -3.153 -7.61 0.000Small stake dummy -1.305 -4.73 0.000Large stake dummy -0.305 -0.46 0.646Constant 1.591 5.73 0.000
All Firms Sample.Median RegressionNumber of obs = 36952 Pseudo
� � � � ��� � ��Coeff. t-stat
��� � � �Absolute error of consensus estimate 0.922
�100 0.000
Time difference 1.755 15.72 0.000Coverage -2.139 -6.16 0.000Size of broker -1.664 -2.51 0.012Small stake dummy -2.343 -6.96 0.000Large stake dummy -1.117 -1.37 0.172Constant 3.135 8.44 0.000
20
on coverage show a different picture in the two samples. While high coverage is important
to explain higher precision in the all firms sample, the coefficient is close to zero in the all
banks sample. Note however, that fewer firms are included in the all banks sample and all
of these are partly owned by a big German bank. It seems that in this case, the number of
brokers following the firm does not proxy for the degree of information available to the uni-
verse of brokers. Valid for both samples, larger brokers are able to achieve higher precision
than smaller ones. The evidence is consistent with the hypothesis that adjunct brokers possess
superior information, as they issue more precise forecasts relative to the consensus than non-
adjunct brokers. The magnitude of the difference is statistically and economically significant
only for adjunct brokers with small equity stakes. In the all banks sample, the relative advan-
tage of adjunct brokers with small stakes results in a absolute forecast errors 1.3 percentage
points lower relative to the consensus, compared to non-adjunct brokers. In the all firms sam-
ple, the information advantage of adjunct brokers with small stakes results in a 2.3 percentage
points lower forecast error. For adjunct brokers with large stakes, the effect is smaller and
statistically not different from zero. Together, the regressions give support to the information
advantage hypothesis. The fact that adjunct brokers with large stakes do not seem to have an
information advantage could also be due to the smaller number of datapoints for this type of
broker, making inference noisier. Alternatively, as ownership data is publicly available, these
brokers might be cautious to release inside information too early to the public to avoid adverse
reputation effects.
The evidence obtained by the regressions on the error and absolute error of earnings fore-
casts is clearly consistent with the superior information hypothesis: Forecasts made by adjunct
brokers, in particular by those with relatively small stakes, are in general more precise than
those made by the average analyst. The evidence is less obvious on conflict of interest issues.
While there is weak support for the client pressure hypothesis, the self interest hypothesis can
not be confirmed.
21
These results seem to indicate that there are leakages in Chinese Walls. Adjunct brokers
face a tradeoff from using this information. Giving their superior knowledge to their cus-
tomers in form of better EPS forecasts could result in a comparative advantage in this segment
of universal banking. However, there might also arise costs from negative effects on busi-
ness relationships. To explore whether adjunct analysts tend to publish or retain information
in particular situations respectively, we now try to explore the information content of new
forecasts.
D. Strategic Behavior
If adjunct brokers possess superior information, they have still to decide whether to use their
information advantage or not. Analysts working for these brokerage houses will then have
conflicting interests. First, for reputation, remuneration and carreer concerns they will try to
issue as precise forecasts as possible. In particular, they should avoid issuing wrong forecasts.
Second, they might have to bend under interests of the bank which in many cases could mean
issuing optimistic forecasts. A possible way out of this conflict could be to report their superior
information to the public only if the consensus appears to be too pessimistic. This behavior
could reconcile the interests of having average forecasts with higher precision, while still more
optimistic than forecasts of their independent competitors.
We divide therefore the sample into two supsamples: A first subsample, where the forecast
error of the consensus estimate������� � � , i.e. the consensus overestimates actual EPS, and
a second subsample where the consensus strictly underestimates EPS (������� � � ). We then
run regressions based on equation 3 for both subsamples.15
The findings presented in Table VII support the hypothesis of strategic behavior. While
adjunct brokers with small stakes appear to possess superior information, they do not always
15We also run logistic regressions on a dummy variable that indicates increased precision relative to the con-sensus. Among the regressors, we use dummies indicating overoptimism of the consensus. While neglectinginformation about the extent of increased precision of forecasts, the logit regression still gives qualitatively sim-ilar results as the methodology presented hereafter.
22
Table VIIMedian Regression Explaining Strategic Behavior.
This regression explains the percent absolute forecast error of individual brokers’ forecasts. The sampleis split into two subsamples, according to the sign of the forecast error of the consensus estimate, wherea zero error is assigned to the positive subsample. Regressors are the absolute error of the consensusforecast, the time difference in years between the day of the forecast and the forecast period end date,the coverage of the firm by brokers in I/B/E/S, the size of the broker forecasting measured as marketshare, and dummy variables that take the value of one if the broker forecasting owns a small (up to 25%) respectively large (more than 25 %) equity stake in the company whose earnings per share are to beforecast.
All Banks Sample.������� � � ������� � ����� � � � �
��� �����Median Regressions Pseudo-
� � � � � � � ��� Pseudo-� � � � � ����
Coeff. t-stat��� � � � Coeff. t-stat
��� � � �Absolute error of consensus 0.925
�100 0.000 0.908
�100 0.000
Time difference 2.756 13.94 0.000 1.023 9.87 0.000Coverage -0.016 -0.02 0.981 0.779 1.95 0.051Size of broker -3.824 -4.50 0.000 -2.446 -5.64 0.000Small stake dummy -0.404 -0.66 0.510 -1.505 -5.60 0.000Large stake dummy -0.664 -0.58 0.562 -0.463 -0.50 0.614Constant 1.474 2.72 0.007 1.185 3.77 0.000
All Firms Sample.������� � � ������� � ��� �� � �
����� �� �
Median Regressions Pseudo-� � � � ��� ����� Pseudo-
� � � � � � � � Coeff. t-stat
��� � � � Coeff. t-stat��� � � �
Absolute error of consensus 0.917�
100 0.000 0.930�
100 0.000Time difference 2.424 12.09 0.000 0.723 6.29 0.000Coverage -2.067 -3.27 0.001 -1.137 -3.26 0.001Size of broker 0.779 0.66 0.512 -3.589 -5.24 0.000Small stake dummy -2.288 -3.03 0.002 -1.535 -5.83 0.000Large stake dummy -2.051 -1.44 0.150 -0.072 -0.08 0.935Constant 2.951 4.43 0.000 2.780 7.28 0.000
23
convey this information to the public. From the regression using the all banks sample can be
seen that the conditional median absolute forecast error of adjunct brokers with small stakes is
1.5 percentage points below the estimate for independent brokers if the consensus estimate of
EPS is too low. However, no evidence for superior information can be found if the consensus
estimate of EPS is overoptimistc. In the all firms sample, the regressions indicate superior
information regardless of the sign of the consensus forecast error. However, in this sample
adjunct brokers issue three times more forecasts if the consensus underestimates earnings
than in the opposite case. Adjunct brokers with small stakes issue 8.3% of all forecasts made
when the consensus underestimates actual EPS. If the consensus forecast is too high, the
weight of adjunct brokers reduces to 2.8%.16 If these brokers observe that the consensus is
overoptimistic, they seem to hide among other brokers. Depending on the choice of the control
group, this hiding appears to take place primarily either by issuing forecasts that mimic their
competitors (this is the interpretation suggested by the findings of the all banks sample) or by
avoiding to make forecasts (as indicated by the all firms sample). The results described above
are statistically significant for adjunct brokers with small stakes only, giving support to the
client pressure hypothesis.
V. Robustness Checks
We check our results for robustness along three lines. First, we investigate the effect of in-
clusion of firm characteristics in the regression on our results. Second, we use alternative
definitions for the bias relative to the consensus and for the precision of estimates. In addition,
we then use a different methodology for our regressions. Third, we investigate stock market
reactions to EPS forecasts. All robustness checks are undertaken for the all banks sample only,
as it was only possible to obtain data on company characteristics for this subsample.
16In the all banks sample, these percentages are 6.3 and 4.1 respectively.
24
A. Firm Characteristics
Numerous studies have found that the precision of analyst forecasts is related to company fun-
damentals. However, it is not obvious why public available company data should influence
the relative bias or the relative precision of individual forecasts relative to a consensus fore-
cast. To check our empirical results for robustness, we have rerun the median regressions on
forecast errors and absolute forecast errors including company specific data frequently used
in studies of analyst forecast quality. We have chosen two variables potentially reflecting the
degree of uncertainty over a firm’s earnings. First, total assets, because it is likely that there is
more information available over large firms. Second, the book-to-market ratio. High book-to-
market ratios mean that the market places low growth expectations into a firm, rather relying
on real assets and current earnings than growth options to value a firm. It should be easier to
forecast earnings for this type of firms. As forecasts are forward looking (with a median of ap-
proximately 10 months), we have decided to use 2 years lagged values to reduce any possible
feedback effects. We obtain the data for these variables from the Global Vantage database.
As these additional variables do not alter the interpretation of the above results, we only
report the median regression on the absolute forecast error for brevity. Note that the sample
size is slightly reduced relative to the regressions omitting these additional control variables.
This is due to some missing balance sheet data. The augmented regression confirm the more
parsimonious specification. The coefficient on size is significant, meaning that indeed absolute
forecast errors for larger firms are smaller. The coefficient of the book-to-market ratio is
statistically not different from zero. The coefficient of the small stake dummy is very close to
the specification before: adjunct brokers make more precise forecasts. While the coefficient
of the large stake dummy has changed the sign, it is still statistically not different from zero.
25
Table VIIIMedian Regression Explaining the Absolute Forecast Error.
This regression explains the percent absolute forecast error of individual brokers’ forecasts. Regressorsare the absolute error of the consensus forecast, the time difference in year between the day of theforecast and the forecast period end date, the coverage of the firm by brokers in I/B/E/S, the size ofthe broker forecasting measured as market share, and dummy variables that take the value of one if thebroker forecasting owns a significant small (up to 25 %) respectively large (more than 25 %) equitystake in the company whose earnings per share are to be forecast. Total assets are two-years laggedvalues in million Euro. Book-to-market are two years lagged values of the book-to-market ratio.
All Banks Sample.Median RegressionNumber of obs = 30608 Pseudo
� � � � � ��� � �Coeff. t-stat
��� � � �Absolute error of consensus estimate 0.925
� ���0.000
Time difference 1.692 15.66 0.000Coverage -0.112 -0.27 0.784Size of broker -3.317 -7.19 0.000Small stake dummy -1.052 -3.45 0.001Large stake dummy 0.078 0.11 0.916Total assets 0.013 -6.06 0.000Book-to-market ratio 0.056 0.25 0.803Constant 1.898 5.47 0.000
26
B. Alternative Error Definition and Regression Methodology
In Section IV we have defined the error and absolute error of forecasts relative to ex post re-
alized values. This approach has - among others - the advantages of straightforward interpre-
tation, avoidance of scaling effects due to stock splits, and comparability with the literature.
The main drawback is the large number of outliers, requiring the use of robust estimation
techniques. However, when thinking about the bias of an analysts’ forecasts relative to other
analysts, one might ask the question ”How many standard deviations is this forecast away
from the consensus?” To formalize this approach, let us define the innovation as deviation
from the consensus, measured in standard deviations, as
����� ��� ��� ������� ������� ���������� � �� ��� � (4)
where��� �������
is the individual forecast,����� � the consensus forecast prevailing at
the time when the individual forecast was issued, and � the standard deviation of��� ���������
����� � over the 3-months time window for which the consensus forecast was calculated.
The measure����� �
allows us to express the magnitude of the deviation of a brokers
forecast from the consensus in terms of typical deviations in forecasts for a specific stock.
Using the all banks sample, we first explore whether these innovations differ systematically
among brokers. We therefore run a regression on the variable innovation. Here, we exclude all
observations leading to an absolute innovation larger than 4, which allows us to proceed with
OLS regression. The results IX indicate again, that adjunct brokers with small stakes tend to
make more optimistic forecasts in a statistically significant way.
As in the previous analysis, we split the sample into subsamples along the consensus over-
or underestimating actual EPS. Also by this methdology we find that forecasts of adjunct
27
Table IXRegression Estimates Explaining Innovation of Forecasts.
Dependent variable is the innovation ������� in forecasts, measured as forecast minus consensus fore-cast, divided by the standard deviation of forecast minus consensus forecast. Regressors are the timedifference in years between the issue date of a forecast and the forecast period end date, the coverageof a company measured as percentage of all brokers covering German companies, the size of the brokermeasured as percentage of German companies covered, and dummies indicating whether a forecast wasmade by an adjunct broker with a small respectively large stake in the company.
All Banks Sample.Least Squares RegressionNumber of obs. 31601 Adjusted
� � � � � �� ���Coeff. t-stat.
��� � � �Time difference 0.028 2.90 0.004Coverage 0.031 0.90 0.368Size of broker -0.180 -4.44 0.000Small Stake Dummy 0.165 6.11 0.000Large Stake Dummy -0.066 -1.00 0.318Constant -0.034 -1.24 0.214
28
brokers exceed those of independent ones especially when the consensus underestimates earn-
ings.17
C. Stock Market Reaction
If adjunct brokers indeed behave differently when expressing views on companies, efficient
stock markets should factor any potential bias out. We test whether the stock markets reacts in
a different way to forecasts by independent and adjunct brokers by regressing 11 day cumula-
tive returns on the innovation measure calculated for each forecast.18 We distinguish among
the different types of brokers.
The results indicate that stock markets generally do not react to earnings per share fore-
casts that are more positive than the median of existing forecasts. However, stock markets
do significantly react to pessimistic forecasts. The point estimates suggest that the market
reaction is more severe for adjunct brokers. While a forecast that is one standard deviation be-
low the consensus reduces the stock price by approximately half a percent if this pessimistic
forecast originates from an independent analyst, the stock price depresses by one percent for
adjunct analysts with small stakes and by 2.7 percent for adjunct analysts with large stakes.
The market seems to take the view that the situation is really bad if even adjunct brokers are
pessimistic. The finding that pessimistic forecasts are more credible if made by adjunct bro-
kers is consistent with both the information advantage and the conflict of interest hypotheses.
17We have omitted the supplementary regressions and a more detailed analysis for brevity. To check forrobustness of the analysis of the precision of forecasts, we have also analysed the measure �
�������������������� ��������������������������
by ordinary regression. Again, the results lead to the same interpretation as the analysis in Section IV. Theomitted Tables are available from the authors by request.
18We report cumulative returns. We also analyze cumulative excess returns. This does not change our conclu-sions.
29
Table XRegression Estimates Explaining Market Reaction to Forecasts.
Least squares regression explaining market reactions to forecasts. Dependent variable is the 11-daycumulative stock price return centered around the day of the forecast. Innovation is the difference be-tween the new forecast and the consensus forecast prevailing at the time of the forecast, scaled by thestandard deviation of the difference between individual forecasts and consensus forecasts in the timeinterval three months prior to the forecast date. Innovation - independent equals the innovation forall forecasts made by independent brokers and zero otherwise. Innovation - small stakes equals theinnovation of the forecast for adjunct brokers with small stakes and zero otherwise. Innovation - largestakes equals the innovation of the forecast for adjunct brokers with large stakes and zero otherwise.The sample is split along the cases of positive (first panel) and negative (second panel) innovations. Ob-servations where the innovations is larger in absolute terms than 4 (standard deviations) are excluded.
All Banks Sample. Positive Innovations.Least Squares RegressionNumber of obs. 10318 Adjusted
� � ��� � � �����Coeff. t-stat.
��� � � �Innovation - independent -0.0005 -0.23 0.818Innovation - small stakes 0.0024 0.43 0.666Innovation - large stakes -0.0031 -0.14 0.885Constant 0.0125 5.89 0.000
All Banks Sample. Negative Innovations.Least Squares RegressionNumber of obs. 11098 Adjusted
� � � � � ������Coeff. t-stat.
��� � � �Innovation - independent 0.0057 5.20 0.000Innovation - small stakes 0.0100 2.77 0.006Innovation - large stakes 0.0272 3.51 0.000Constant 0.0076 6.41 0.000
30
VI. Conclusion
We have examined the existence of spill overs of superior information and possible conflicts
of interests within universal banks in the specific setting of analyst forecasts of earnings per
share. While there are so-called Chinese Walls in place in universal banks, which should pre-
vent both types of influence on analysts to occur, we find evidence for leakages through which
both superior information and pressure on analysts find their way to research departments. The
relationship between banks and their corporate customers in Germany is often best described
by the term of a housebank relationship. Frequently, these housebanks even dominate corpo-
rations in shareholder meetings through large equity stakes and additional proxy votes from
small shareholders. According to the size of the stake in a company, and whether the total
investment has rather characteristics of a loan or a stock portfolio, we expect bank behavior
to be different. For the purpose of our analysis, we distinguish therefore between small and
large equity stakes. Our findings give support to both the superior information hypothesis and
the conflict of interest hypothesis. According to our results, there is evidence for an informa-
tion advantage of banks owning equity stakes, regardless of the size of the equity stake. This
superior information is flowing at least partly to analysts in research departments despite of
Chinese Walls. This is demonstrated by evidence that the conditional median absolute forecast
error is smaller for adjunct brokers. Furthermore, both brokers owning small and large stakes
are better able than their independent colleagues to make earnings per share forecasts that are
an improvement relative to the consensus estimate. Second, we also find support for the con-
flict of interest hypothesis, demonstrated by asymmetries of analyst behavior with respect to
the consensus forecast over- oder underestimating actual earnings per share. Adjunct brokers
with small equity stakes are relatively more likely to announce forecasts that are an improve-
ment to the consensus if the consensus underestimates actual earnings. This finding suggests
that the conflict of interest originates rather from client pressure than the banks self interest.
Consistent with the literature on conflict of interest in the investment banking industry, clients
31
who have established business relationships with other departments of the bank are unlikely
to have to suffer from unfavorable forecasts.
32
References
Ackert, L.F., and G. Athanassakos, 2000, A Simultaneous Equation Analysis of Analysts’
Forecast Bias and Institutional Ownership, Federal Reserve Bank of Atlanta, working paper
2000-5.
Ahmed, A.S., G.J. Lobo, and X. Zhang, 2001, Do analysts under-react to bad news and over-
react to good news?, Working Paper.
Barber, B., R. Lehavy, M. McNichols, and B. Trueman, 2001, Can Investors Profit from the
Prophets? Security Analyst Recommendations and Stock Returns, Journal of Finance 56,
531–563.
Barth, M.E., R. Kasznik, and M.F. McNichols, forthcoming, Analyst Coverage and Intangible
Assets, Journal of Accounting Research.
Becht, M., and E. Boehmer, 1999, Transparency of Ownership and Control in Germany, work-
ing paper.
Ber, H., Y. Yafeh, and O. Yosha, 2001, Conflict of interest in Universal Banking: Bank lending,
Stock Underwriting, and Fund Management, Journal Of Monetary Economics 47, 189–
218.
Boehmer, E., 1999, Corporate governance in Germany: Institutional background and empiri-
cal results, working paper.
Bolliger, G., 2001, The Characteristics of Individual Analysts’ Forecasts in Europe, working
paper.
Brown, L.D., and M.S. Rozeff, 1978, The superiority of analyst forecasts as measures of
expectations: Evidence from earnings, Journal of Finance 33, 1–16.
Capstaff, J., K. Paudyal, and W. Rees, 1998, Analysts Forecasts of German Firms’ Earnings:
a Comparative Analysis, Journal of International Financial Management and Accounting
9, 83–116.
33
Chang, J.J., T. Kahanna, and K.G. Palepu, 2000, Analyst Activity around the world, working
paper.
Chopra, V.K., 1998, Why so much error in analysts earnings forecasts?, Financial Analysts
Journal 54, 35–43.
Dunbar, C., C. Hwang, and K. Shastri, 1999, Underwriter Analyst Recommendations: Conflict
of Interest or Rush to Judgement, working paper.
Easterwood, J.C., and S.R. Nutt, 1999, Inefficiency in Analysts’ Earnings Forecasts: System-
atic Misreaction or Systematic Optimism, Journal of Finance 54, 1777–1797.
Franks, J., and C. Mayer, 2000, Ownership and Control of German Companies, working paper.
Gande, A., M. Puri, A. Saunders, and I. Walter, 1997, Bank underwriting of debt securities:
modern evidence, Review of Financial Studies 10, 1175–1202.
Gompers, P., and J. Lerner, 1999, Conflict of Interest in the Issuance of Public Securities:
Evidence from Venture Capital, Journal of Law and Economics 62, 1–28.
Gu, Z., and J.S. Wu, 2001, Earnings Skewness and Analyst Forecast Bias, working paper.
Hamao, Y., and T. Hoshi, 2000, Bank Underwriting of Corporate Bonds: Evidence from Japan
After the Financial Reform of 1993, Working Paper, University of Southern California.
Hebb, G.M., and D.R. Fraser, 2001, Conflict of Interest in Commercial Bank Security Under-
writings: United Kingdom Evidence, Working Paper.
Hong, H., and J.D. Kubik, 2001, Measuring the Career Concerns of Security Analysts: Earn-
ings Forecast Accuracy, Job Separations and Brokerage House Status, working paper.
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.
Klein, P.G., and K. Zoeller, 2001, Universal-Bank Underwriting and Conflicts of Interest:
Evidence from German Initial Public Offerings, Working Paper.
Kroszner, R., and R. Rajan, 1994, Is the Glass-Steagall Act justified? A Study of the U.S.
experience with universal Banking before 1933, American Economic Review 84, 810–832.
34
Lehmann, E., and J. Weigand, 2000, Does the Governed Corporation Perform Better? Gover-
nance Structures and Corporate Performance in Germany, working paper.
Lim, T., 2001, Rationality and Analysts’ Forecast Bias, Journal of Finance 56, 369–385.
Lin, H.-W., and M. McNichols, 1998, Underwriting Relationships and Analysts’ Earnings
Forecasts and Investment Recommendations, Journal of Accounting and Economics 25,
101–127.
Michaely, R., and K. Womack, 1999, Conflict of Interest and the Credibility of Underwriter
Analyst Recomendations, Review of Financial Studies 12, 653–686.
Narayanan, R.P., K.P. Rangan, and K.K. Rangan, 2001, Commercial Bank Underwriting: Con-
flict of Interest and Credible Commitment, Working Paper.
Puri, M., 1996, Commercial banks in investment banking: Conflict of interest or certification
role?, Journal of Financial Economics 40, 373–401.
Stickel, S., 1992, Reputation and performance among security analysts, Journal of Finance
47, 1811–1836.
Welch, I., 2000, Herding among security analysts, Journal of Financial Economics 58, 369–
396.
Womack, K.L., 1996, Do brokerage analysts recommendations have investment value?, Jour-
nal of Finance 51, 137–167.
35