Prof. Dr. Wolfgang König, Prof. Dr.-Ing. Ralf ... - E-Finance Lab

Ergebnisband des E-Finance Lab
Top-5-Publikationen je Layer des 1. Halbjahres 2012

Inhaltsverzeichnis
Layer 1: IT-Infrastructure: Service Systems in E-Finance
(Prof. Dr. Wolfgang König, Prof. Dr.-Ing. Ralf Steinmetz,
Prof. Dr. Roman Beck)
� Beck R., Weber S., Gregory R. (2012):
Theory-Generating Design Science Research
In: Information Systems Frontiers - online first, DOI: 10.1007/s10796-012-9342-4.
� Beck R., Schott K. (2012):
The Interplay of Project Control and Interorganizational Learning: Mitigating Effects
on Cultural Differences in Global, Multisource ISD Outsourcing Projects
In: Business & Information Systems Engineering - online first, DOI:10.1007/s12599-
012-0217-5.
� Lampe U., Siebenhaar M., Papageorgiou A., Schuller D., Steinmetz R. (2012):
Maximizing Cloud Provider Profit from Equilibrium Price Auctions
In: Proceedings of the 5 th International Conference on Cloud Computing (CLOUD
2012), Honolulu, Hawaii, USA.
� Schuller D., Lampe U., Schulte S., Eckert J., Steinmetz R. (2012):
Cost-driven Optimization of Complex Service-based Workflows for Stochastic QoS
Parameters
In: Proceedings of the 9 th International Conference on Web Services (ICWS 2012),
Honolulu, Hawaii, USA.
� Wolf M., Beck R., Pahlke I. (2012):
Mindfully Resisting the Bandwagon – Reconceptualising IT Innovation Assimilation in
Highly Turbulent Environments
In: Journal of Information Technology (JIT) - Online first, DOI:10.1057/jit.2012.13.
Layer 2: E-Financial Markets & Market Infrastructures
(Prof. Dr. Peter Gomber)
� Gomber P., Pujol G., Wranik A. (2012):
Best Execution Implementation and Broker Policies in Fragmented European Equity
Markets
In: International Review of Business Research Papers, Vol. 8, Issue 2, 144-162.
� Haferkorn M., Lutat M., Zimmermann K. (2012):
The Effect of Single-Stock Circuit Breakers on the Quality of Fragmented Markets
In: FinanceCom 2012, Barcelona, Spain.
� Lattemann C., Loos P., Johannes G., Burghof H., Breuer A., Gomber P.,
Krogmann M., Nagel J., Riess R., Riordan R., Zajonz R. (2012):
High Frequency Trading - Costs and Benefits in Securities Trading and its Necessity
of Regulations
In: Business & Information Systems Engineering, Vol. 4, Issue 2, 93-108.

� Siering, M. (2012):
Investigating the Market Impact of Media Sentiment and Investor Attention
In: FinanceCom 2012; Barcelona, Spain.
� Weber M.C., Wondrak C. (2012):
Measuring the Influence of Project Characteristics on Optimal Software Project
Granularity
In: Proceedings of the 20 th European Conference on Information Systems (ECIS),
Barcelona, Spain.
Layer 3: Customer Management in E-Finance
(Prof. Dr. Andreas Hackethal, Prof. Dr. Bernd Skiera, Prof.
Dr. Oliver Hinz)
� Bhattacharya U., Andreas H., Simon K., Benjamin L., Steffen M. (2012):
Is unbiased financial advice to retail investors sufficient? Answers from a large field
study
In: Review of Financial Studies, Vol. 25, 975-1032.
� Hackethal A., Michael H., Tullio J. (2012):
Financial advisors: A case of babysitters?
In: Journal of Banking & Finance, Vol. 36, 509-524.
� Schmitt P., Skiera B., Van den Bulte C. (2011):
Referral Programs and Customer Value
In: Journal of Marketing, Vol. 75, Jan., 46-59.
� Hinz O., Skiera B., Barrot C., Becker J. (2011):
An Empirical Comparison of Seeding Strategies for Viral Marketing
In: Journal of Marketing, Vol. 75, Nov., 55-71.
� Skiera B., Bermes M., Horn L. (2011):
Customer Equity Sustainability Ratio: A New Metric for Assessing a Firm’s Future
Orientation
In: Journal of Marketing, Vol. 75, May, 118-131.
Die Autoren bedanken sich herzlich für die Unterstützung aller ihrer Arbeiten durch
das E-Finance Lab!

Layer 1:
IT-Infrastructure: Service Systems in E-Finance
(Prof. Dr. Wolfgang König, Prof. Dr.-Ing. Ralf
Steinmetz, Prof. Dr. Roman Beck)
� Beck R., Weber S., Gregory R. (2012):
Theory-Generating Design Science Research
In: Information Systems Frontiers - online first.
� Beck R., Schott K. (2012):
The Interplay of Project Control and Interorganizational Learning: Mitigating
Effects on Cultural Differences in Global, Multisource ISD Outsourcing Projects
In: Business & Information Systems Engineering - online first.
� Lampe U., Siebenhaar M., Papageorgiou A., Schuller D., Steinmetz R.
(2012):
Maximizing Cloud Provider Profit from Equilibrium Price Auctions
In: Proceedings of the 5 th International Conference on Cloud Computing
(CLOUD 2012), Honolulu, Hawaii, USA.
� Schuller D., Lampe U., Schulte S., Eckert J., Steinmetz R. (2012):
Cost-driven Optimization of Complex Service-based Workflows for Stochastic
QoS Parameters
In: Proceedings of the 9 th International Conference on Web Services (ICWS
2012), Honolulu, Hawaii, USA.
� Wolf M., Beck R., Pahlke I. (2012):
Mindfully Resisting the Bandwagon – Reconceptualising IT Innovation
Assimilation in Highly Turbulent Environments
In: Journal of Information Technology (JIT) - Online first.

Inf Syst Front
DOI 10.1007/s10796-012-9342-4
Theory-generating design science research
Roman Beck & Sven Weber & Robert Wayne Gregory
# Springer Science+Business Media, LLC 2012
Abstract A frequently mentioned challenge in design
science research (DSR) is the generation of novel theory
above and beyond information technology artefacts.
This article analyzes the DSR process and extends
established frameworks for theory generation to exemplify
improvements to theory generation through methods
of grounded theory development. On a conceptual
base, we developed a theory-generating DSR approach
which integrates methods of grounded theory development
with established DSR methodology. This combination
enables a design theorist to generate theoretical
knowledge that extends the applicable knowledge base.
We do not elaborate this combination on a meta-level,
but rather provide a process model for researchers in
form of an extension of a well-known DSR model to
combine both methods in a pluralistic research design.
With this suggested research approach, scholars can
draw theoretical insights from analytical abstractions
and can improve the development of IT artefacts in a
structured way to avoid failure or repair loops.
R. Beck : S. Weber (*)
Institute of Information Systems, Goethe University Frankfurt,
Grüneburgplatz 1,
60323 Frankfurt am Main, Germany
e-mail: svweber@wiwi.uni-frankfurt.de
R. Beck
e-mail: rbeck@wiwi.uni-frankfurt.de
R. W. Gregory
University of Göttingen,
Platz der Göttinger Sieben 5,
37073 Göttingen, Germany
R. W. Gregory
e-mail: gregory@wiwi.uni-goettingen.de
Keywords Design science research . Behavioural science
research . Grounded theory method . Pluralistic research
method
1 Introduction
In recent years, design science research (DSR) has attracted
more information systems (IS) research attention, because it
deals with the generation of information technology (IT)
artefacts and their evaluation, which is just as important to
IS research as is research into the impacts of IS (Benbasat
and Zmund 2003; Hevneretal.2004). In general, DSR
provides models and guidelines to researchers that enable
them to create, improve, and evaluate IT artefacts (Hevner et
al. 2004; Holmström et al. 2009; March and Smith 1995;
Weber et al. 2012). Yet many existing DSR research
attempts pay less attention in generating an original theoretical
contribution that goes beyond problem-solving IT artefacts
(Carlsson 2006; Gregory and Muntermann 2011;Hevner
et al. 2004; Winter 2008).
We therefore investigate potential avenues for combining
both aims, using a conceptual approach based on an extensive
analysis of DSR and design theory literature. As a result
of our analysis, we propose a theory-generating DSR approach
that is informed by behavioural science elements
used for theorizing, such as techniques from the grounded
theory method (GTM) (e.g., Glaser 1978; Glaser1998;
Glaser and Strauss 1967). Therefore, our proposed approach
closes some methodological gaps in DSR to achieve insights
at a higher level of analytical abstraction (Yadav 2010).
Certain techniques from GTM are remarkably appropriate,
in that they generate additional theoretical insights in DSR.
The proposed theory-generating DSR approach combines

elements from DSR and GTM for simultaneous problem
solving and theory generation.
Calls for more pluralistic research in IS (e.g., Mingers
2001) have prompted some prior attempts to combine different
methods. For example, Goldkuhl (2004) uses behavioural
science techniques in a DSR approach and suggests
that three types of grounding (internal, empirical, and theoretical)
enhance DSR as a means to generate practical
knowledge. Holmström et al. (2009) accumulates DSR with
a second research cycle that features the development of
substantive and formal theory (the focus of GTM) to contribute
to the knowledge base, but they also use behavioural
science elements to develop an explanatory, theory-testing
approach instead of a means to develop new theoretical
insights. Kuechler and Vaishnavi (2008a, b) address theory
development and theorizing in the context of DSR but
without combining any specific behavioural science research
method. A more advanced approach is presented by
Sein et al. (2011), who integrate DSR with action research in
one model. Their model argues that design science needs to
go beyond the traditional focus on the IT artefact and recognize
its organizational embeddedness. By integrating design
science with action research lenses, the authors are able
to emphasize the organizational interventions through IT
artefact deployment as well as the learning from this interventions
in terms of contributions to the knowledge base. In
this critical reflection and learning step, which is somewhat
disconnected from the traditional build and evaluate DSR
cycle, the researchers abstract the key theoretical findings
that go beyond solving an individual problem. Finally, some
other studies combine DSR with action research (e.g., Allen
et al. 2000), ethnography (e.g., Baskerville and Stage 2001),
or soft design science (Baskerville et al. 2009) withthe
goals of improving theoretical abstraction and knowledge
generation within a DSR project.
In summary, existing research frameworks guided towards
the simultaneous use of DSR and behavioural science but fall
short to provide specific processes for how to conduct a multi
method DSR project (Carlsson 2006). Yet both Goldkuhl
(2004) and Holmström et al. (2009) concludethatDSR
and behavioural science may coexist within a single
project but rather as an additional and independent step.
In this paper, we extend this view and develop a process
model for theory-generating DSR that relies on the simultaneous
and not separated use of both DSR and GTM elements.
The structure of the remainder of this article follows the
building blocks of conceptual research described by Yadav
(2010, p. 3). In the next two sections, we provide an overview
of the two research approaches (DSR and GTM)
before we outline how design theory components map onto
the DSR approach. In the subsequent section, we present a
detailed description of the theory-generating DSR approach,
followed by an overview on theory-generating DSR. We
also detail how the IS design theory components relate to
our proposed approach and provide a first illustration how
theory-generating DSR can be applied in DSR projects. We
conclude with a discussion section and an outlook for further
research.
2 Design science research
Inf Syst Front
The primary focus of IS research is the IT artefact (Benbasat
and Zmund 2003), whereas DSR centres on the design and
creation of the artificial (Simon 1969), especially IT artefacts
(in the form of a construct, model, method, instantiation, or
combination thereof; (March and Smith 1995)). Therefore,
DSR encompasses the creation of an innovative construct that
has not existed before and can be used to serve human purposes
(March and Smith 1995). Its historical origins stem
from engineering (Au 2001; March and Smith 1995), but
DSR also relates to several other academic disciplines, such
as architectural science and computer science. A common
ground across these disciplines is a dual focus on both the
practical application of the IT artefact and the scientific abstraction
and learning (Baskerville 2008). In this context,
theorizing in DSR is worthy of discussion and at least as
important as problem-solving itself (Lee et al. 2011).
Most DSR researchers also recognize that the developed
IT artefact can provide theoretical contributions, provided
that key DSR guidelines are followed (Gregor 2002, 2006)
and so-called kernel theories, drawn from the natural and
social sciences through a creative translation process, are
applied (Hevner et al. 2004; Markus et al. 2002; Walls et al.
1992). Derived principles and concepts of such kernel theories
may provide a basis for advancing DSR by specifying
requirements and generating an IT artefact (Walls et al.
1992). However, Gregor and Jones (2007) explorethat
theorization is a key goal in theory-generating DSR which
mayeventuallyleadtoanISdesigntheory.Therefore,
problem solutions and knowledge contributions can be derived
from either existing kernel theories and IT artefacts or
the development of a new IT artefact and subsequent theorization
(Holmström et al. 2009).
In the DSR process, the initial step must be the search for a
problem that has practical relevance (Hevner et al. 2004). In
other words, “a DSR approach seeks a solution to a real-world
problem of interest to practice” (Kuechler and Vaishnavi
2008a, p. 492), which requires a differentiation between products
(IT artefact) and processes (activities that lead to the IT
artefact) in the DSR cycle (March and Smith 1995; Walls et al.
1992). The product outcome is inevitably embedded in
some place, time, and community and must undergo
theorization to meet innovative and progressive demands
(Orlikowski and Iacono 2001).

Inf Syst Front
The DSR process in turn consists of two basic processes:
building and then evaluating the IT artefact (Baskerville et
al. 2009; Hevner and March 2003; March and Smith 1995).
In the building process, a sequence of activities aims to
produce “something new,” then in the evaluation process,
the created IT artefact undergoes evaluation to produce
feedback and generate new knowledge about the problem.
The newly generated insights serve to improve the quality of
the IT artefact and the design process itself (Hevner et al.
2004). These processes take place partly in parallel and
involve multiple iterations, which enables the IT artefact to
be generated such that it fully satisfies the researchers and
practitioners who later make use of it (Markus et al. 2002).
In total then, DSR offers a rigorous and meaningful contribution
to practice and theory, in the form of an IT artefact
and its evaluation (Gregor 2006).
3 Grounded theory method
The supporting component for theory-generation in our
model is GTM. Since it was introduced to sociology by
Glaser and Strauss (1967), the method has been widely
developed and applied in various disciplines (Bryant and
Charmaz 2007a; Urquhart2007). In a grounded theory
study, the focus lies on the discovery or generation of theory
grounded on empirical evidence (Glaser and Strauss 1967).
A grounded theory study considers the process of discovering
concepts and categories and the relationships between
them (Bryant and Charmaz 2007b), with the end result of a
substantive or grounded theory (Glaser 2007).
Over time, GTM has evolved in different directions. For
example, in the mixed methods approach, GTM combines
with other techniques and methods, such as case study
research (Eisenhardt 1989; Eisenhardt and Graebner 2007).
The most predominant approaches are those propagated by
the originators, Glaser and Strauss, who offer “Glaserian”
and “Strausserian” forms of grounded theory (Ketokivi and
Mantere 2010). This distinction also has been called the
emerging versus forcing debate. The proponents of Glaserian
grounded theory place more emphasis on the principle of
emergence; grounded theory should emerge from the data
and existing theories or concepts, and coding schemes
should not be forced onto the data (Udo 2005).
The term “grounded theory” refers to both the end product
and the process. Conducting grounded theory research
involves various techniques prescribed by GTM (Glaser
1998; Strauss and Corbin 1990), including—among others
—theoretical sampling and the constant comparative method
(Suddaby 2006). Theoretical sampling means that
insights from the initial data collection and analysis guide
subsequent data collection and analysis, so the grounded
theory can emerge over time through iterative cycles of
deeply intertwined data collection and analysis. Urquhart
et al. (2010) provide an in-depth discussion of this key
grounded theory concept and refer to this as deciding on
analytic grounds where to sample from next. Over time,
researchers achieve theoretical saturation because additional
data collection and analysis efforts do not yield new findings
(Eisenhardt 1989). In the constant comparative method, the
researcher constantly compares instances of data labelled as
a particular category with other instances of data in the same
category as a means to substantiate these categories and
build theory. This method is discussed at length by Urquhart
et al. (2010) and the authors refer to this as an important tool
for exposing generated analytical insights to rigorous scrutiny
and building theory. Thereby, all kinds of “slices” of data
(e.g., primary data such as qualitative interviews, secondary
data such as documentation and extant literature) are used to
reach higher levels of abstraction and advance conceptualization.
The relations identified among the categories and the
theoretical integration lead to the grounded theory.
Grounded theory method can in principle be used within a
wide range of epistemological stances and research
approaches (Bryant and Charmaz 2007a; Madill et al. 2000).
It provides the ability to generate an in-depth understanding of
a phenomena by analyzing qualitatively the collected data
from different sources, sorting it into consistent categories,
and emerging a grounded insights of it (Urquhart et al. 2010).
For an overview of the process of grounded theory development,
we recommend, for example, Fernandez (2004). In the
past, Weedman (2008) used this multi-epistemological nature
of GTM to analyze an IT artefact design project (called
Sequoia) with elements of GTM. While GTM is never mentioned,
her analyzing techniques were very similar to theoretical
sampling and the constant comparison method. In
addition, Baskerville and Pries-Heje (1999) combinedaction
research with grounded theory to add rigor and reliability to
the theory formulation process of action research. As an
contrary example, Arazy et al. (2010) used qualitative methods
to evaluate and test a social recommender systems that
utilizes data regarding users’ social relationships by filtering
relevant information to users. However, this approach is more
on theory testing and not building. Hence, GTM, as a behavioural
method provides suitable instruments for an in-depth
understanding of the problem space and its environmental
factors to theorize and not only test in DSR.
4 IS design theory and design science research
On the one hand, DSR provides scholarly contributions based
on the design, evaluation, generalization, and/or theorization
of the IT artefact (Gregor 2006; Gregor and Jones 2007;
Orlikowski and Iacono 2001). On the other hand, it provides
a practical contribution to practitioners in the form of the useful

IT artefact (Hevner et al. 2004). Hence, both the IS design
theories as well as extant DSR models have to be taken into
account to develop a new theory-generating DSR approach.
4.1 IS design theory
As a theoretical lens, our literature review focused on fundamental
components of IS design theory. Our initial DSR
understanding pushed us toward Walls et al. (1992) who laid
the basic foundations for developing a precise understanding
about the nature and anatomy of a design theory. They developed
seven components of an IS design theory guided by prior
literature (e.g., Dubin 1978; Nagel1961) and separated them
into product and process components.
The product components encompass meta-requirements,
meta-design, existing kernel theories governing the requirements,
and a test of whether the meta-design satisfies the
meta-requirements. Most existing DSR publications spend
little time considering the requirements or assume they are
already specified (e.g., Aalst and Kumar 2003; Abbasi and
Table 1 Eight components of IS design theory in different DSR approaches
IS Design Theory
Component
Description Comparison with Existing
DSR Frameworks
1. Purpose and scope “What the system is for,” that is, the set
of meta-requirements that specifies the
type of artefact to which the theory
applies and also defines the scope or
boundaries of theory
2. Constructs Representations of entities of interest in
the theory (design sub-parts)
3. Principle of form
and function
The abstract blueprint or architecture
that describes an IT artefact, whether
a product or method/intervention
4. Artefact mutability Changes in the state of the artefact
anticipated by theory; that is, what
degree of artefact change is
encompassed by theory?
5. Testable
propositions
6. Justification
knowledge
7. Principles of
implementation
8. Expository
instantiation
Chen 2008; Umapathy et al. 2008; for further information
please see the Appendix). Because requirement specification
is a central part of developing an IT artefact, as well as
of subsequent satisfaction with its functionality, we regard
this requirements specification phase critical for DSR. In
addition, a lot of information becomes available from a
system when we consider its purpose, the problem it
aims to solve, and the overall intentions behind building
it. Therefore, we searched for an appropriate theoretical
lens to integrate requirements into our theory-generating
DSR approach.
The process components of Walls et al. (1992) represent
the design method for the IT artefact and integrate existing
kernel theories that govern the design process. They also
describe the underlying knowledge for the design process
and guide research projects.
In turn, using the components of design theory identified
by Walls et al. (1992), Gregor and Jones (2007) suggest
eight components that an IS design theory should include.
We build on these eight components to analyze existing
Meta-requirements
(Walls et al. 1992)
Partly covered by the design cycle:
build and evaluate (March and
Smith 1995)
Different frameworks provide
guidelines and recommendations
for the design process (e.g.,
Hevner et al. 2004)
Partly covered by theorizing about
the IT artefact (Orlikowski and
Iacono 2001)
Truth statements about the design theory Evaluation of whether the design
satisfies requirements (Markus
et al. 2002)
The underlying knowledge or theory
from natural or social sciences that
gives a basis and explanation for the
design (kernel theories)
Description of processes for
implementing the theory (either
product or method) in specific
contexts
A physical implementation of the IT
artefact that can help represent the
theory as an expository device and
for the purposes of testing
Underlying kernel theories
(Walls et al. 1992)
Partly covered by the people
involved in a given or changing
environment (Hevner et al. 2004)
A viable IT artefact in the form of
a construct, model, method, or
instantiation (Hevner et al. 2004)
Inf Syst Front
Improvements Required
for Theory-Generating DSR
A more detailed requirements
analysis increases our
fundamental understanding of
the IT artefact and its sub-parts
Detailed theoretical analysis
of the IT artefact and its design
process to adapt to changing
environments and allow for
evolution over time
Evaluation and theory generation
through theory building from
behavioural science (e.g.,
Holmström et al. 2009)
Extend underlying kernel theories
with behavioural science aspects
(Gregor and Jones 2007)
A detailed analysis of influential
factors (e.g., interaction between
involved people)

Inf Syst Front
DSR frameworks and determine if they include all the
components needed to provide theoretical insights (see
Table 1). In the first and second columns of Table 1, we
outline each required design theory component and then
provide information about existing DSR frameworks. Finally,
we illustrate ways to overcome the challenges we have
identified to generate additional theoretical insights from
DSR. Although many IS design theory components have
been addressed, others might be improved or strengthened
by behavioural science elements, particularly to formulate
and evaluate the theoretical insights derived from IT artefact
development and use.
4.2 Extant DSR models
To develop our theory-generating DSR approach, we
searched for a established DSR model that could serve as
a basis for our approach. Several scholarly contributions
focus on the epistemological positioning of DSR. For example,
Winter (2008) differentiates design science from
design research: The former encompasses reflection on and
guidance for the IT artefact construction and evaluation
process while the latter encompasses the creation and evaluation
of a specific IT artefact. Hevner et al. (2004) focus on
the process of IT artefact development. Thus, IS research
appears influenced by both the environment (people,
organizations, existing technology) and the knowledge
base (Hevner et al. 2004). This classification is closely
related to Winter’s (2008) point of view where an IT
artefact results from the constant iteration of refinement
and assessment. In this regard, Hevner et al. (2004), suggest
seven guidelines for conducting DSR, but many other
examples of DSR frameworks and models exist, including
Peffers et al. (2008), Pries-Heje and Baskerville (2008),
March and Smith (1995), and Nunamaker et al. (1991). These
DSR frameworks build the foundation for many researchers
conducting DSR but do not provide explicit prescriptions how
to achieve theoretical abstraction from an IT artefact in a
structured way.
A somewhat different approach is provided by Vaishnavi
and Kuechler (2008) as well as Kuechler and Vaishnavi
(2008a, b) who propose a design cycle for DSR that address
theory development and theorizing. Moreover, they provide
a general model describing each process step in DSR as
illustrated in Fig. 1.
According to the model of Vaishnavi and Kuechler
(2008) as well as Kuechler and Vaishnavi (2008a, b), problem
awareness is the starting point of DSR which is
reflected by an initial proposal depicting the problem that
has to be solved. This step is similar to Hevner et al. (2004)
and their demand for an initial problem that has to be solved.
The next phase is the suggestion phase where it is tested if
the formulated proposal can be transferred into a tentative
Knowledge
Flows
Process
Steps
Awareness of
Problem
Suggestion
Development
Evaluation
Conclusion
Outputs
Proposal
Tentative
Design
Artefact
Performance
Measures
Results
Fig. 1 General design cycle of DSR (Kuechler and Vaishnavi 2008a,
b; Vaishnavi and Kuechler 2008)
design or not. Thereafter, the IT artefact is developed and
evaluated and conclusions are drawn from the IT artefact as
a result of the problem solving process. However, the whole
process is highly repetitive in the sense that every process
step can lead to a repetition and improvement of prior steps
through the knowledge flows depicted in Fig. 1.
Departing from the described DSR model, we used its
basic elements to structure and develop our theorygenerating
DSR approach which incorporates the findings
from Table 1, as illustrated in the following.
5 Theory-generating design science research
On the basis of prior findings and identified components in
DSR design theory development, as well as the ways they
might be improved with behavioural science elements, we
developed our theory-generating DSR approach, which
combines DSR and GTM techniques, as illustrated in
Fig. 2. The proposed approach extends the general design
cycle of DSR from Kuechler and Vaishnavi (2008a, b).
Therefore, the first steps encompass typical methods used
already in DSR: the awareness of the problem as well as to
develop and evaluate the IT artefact. These steps are complemented
and specified by supportive GTM techniques,
such as detailed analyses of the tentative design and requirements,
theoretical sampling, and the constant comparative
technique. A subsequent theory-generation step which is an
extension of Kuechler and Vaishnavi’s (2008a, b) model
encompass the simultaneous use of DSR and GTM to
generate additional theoretical insights, such as by collecting
additional data from the application of the IT artefact in
an appropriate environment. A complementary research
approach can create new insights and results in the conclusion
step to be integrated into existing knowledge, in
addition to the developed IT artefact. The loop back to
the awareness of the problem to refine the tentative design

Fig. 2 Process of generating
additional scholarly
contributions by theorizing in
DSR
and the requirements therefore closes. Moreover, these new
insights and knowledge provide a theoretical basis for
upcoming DSR projects.
In the following sections, we discuss in detail the different
steps of our proposed theory-generating DSR approach
(Fig. 2).
5.1 Awareness of problem and suggestion
The awareness of the problem and the initial definition of
the research lens is critical, in the sense that the underlying
theories of different research approaches (e.g., design or
behavioural science) justify and explain the researcher’s
design process (Gregor and Jones 2007; Kuechler and
Vaishnavi 2008a, b). Unfortunately, serious methodological
problems can occur in this step, for example, a researcher
with a strict GTM lens could primarily focus on exploring
the theoretical implications behind the requirements for the
IT artefact and pay less attention to problem-solving. In
contrast, a DSR researcher might deemphasize important
behavioural issues by focusing only on the actual problem.
Theory-generating DSR therefore tries to address both practical
relevance and theoretical contributions. In contrast to
Kuechler and Vaishnavi's (2008a, b) model, we merged the
awareness of the problem and the suggestion in one step. As
they mention themselves, both complement each other and
are highly intertwined (Kuechler and Vaishnavi 2008a, b).
Hence, both can be combined with our theory-generating
approach in one step to explore the requirements of the IT
artefact.
5.1.1 Theoretical sampling
Inf Syst Front
A first tentative design of the IT artefact and well-defined
requirements are key for problem solving, so this step is
heavily influenced by the real-world problem as it tries to
determine the proper requirements and entities of interest
with regard to the IT artefact (Gregor and Jones 2007).
Researchers therefore need to focus on the actual problem,
but they cannot lose perspective on potentially emergent
problems and must actively integrate solutions into the
development of the IT artefact. By focusing on both goals
simultaneously, researchers can establish a basis for a satisfying
IT artefact. Theoretical sampling in GTM means that
insights from initial data collection and analysis efforts
guide subsequent ones. In other words, understanding of
the phenomena (or business requirements) emerges over
time through iterative cycles of data collection and analysis

Inf Syst Front
that are deeply intertwined (Glaser 1978). Urquhart et al.
(2010) define theoretical sampling as “deciding on analytic
grounds where to sample from next” (p. 371). Thereby,
theoretical sampling assists the classification of data, the characterization
of relationships between data, and to clarify these
relationships. Without this method, it is nearly impossible for
the researcher to generate theoretical insights about the phenomenon
(Urquhart et al. 2010). Accordingly, subsequent data
collection and analysis efforts over time must build upon one
another for cumulative generation of theoretical insights. Over
time, researchers reach a kind of saturation, such that additional
data collection and analysis efforts do not yield any new
findings (Eisenhardt 1989). Moreover, theoretical sampling
increases the researchers’ flexible ability to look for new
aspects that emerge during the process. It also provides a
means to consider potential problems with the IT artefact in
advance, thus enhancing the likelihood of creating innovative
and progressive IT artefacts for testing.
5.1.2 Slices of data
The slices of data to be collected and analyzed relate to the
environment, including people, the organization, and technology,
as well as the knowledge base or extant literature
(Hevner et al. 2004). However, in DSR, slices of data
mainly refer to extant literature or models (e.g., Aalst and
Kumar 2003), collected data (e.g., Albert et al. 2004), or
kernel theories (e.g., Markus et al. 2002) (for a detailed
discussion, see the Appendix). The inclusion of extant literature
alongside the empirical data is a technique widely
adopted by IS grounded theorists (e.g., Fernandez 2004;
Levina and Vaast 2005) as a means to raise the overall
analysis to a higher conceptual level. In addition, the process
of systematic data collection and analysis helps to
identify and clarify the requirements and entities of interest
(Gregor and Jones 2007), which provide the basis for the
subsequent development or adaptation of an IT artefact. This
process is inherently iterative, in that data collection and
analysis on the one hand and IT artefact creation and evaluation
on the other hand are deeply intertwined.
5.1.3 Tentative design and requirements
A first tentative design and well-defined requirements are
extremely important to DSR, because they ensure the IT
artefact’s relevance to a real-world problem (Gregor and
Jones 2007; Hevner et al. 2004; Pries-Heje and Baskerville
2008; Walls et al. 1992). In our theory-generating DSR
approach, these requirements and entities emerge from the
intertwined process of theoretical sampling and the collection
of slices of data. Most DSR projects integrate the IT
artefact’s requirements as an essential step (e.g., Aalst and
Kumar 2003; Albert et al. 2004; Markus et al. 2002; for a
detailed discussion please see the Appendix), because without
clearly defined requirements, the IT artefact will not be
useful and cannot offer a satisfying solution.
5.2 Development
The tentative design and the explored requirements lead to
the development and creation of the IT artefact which is the
outcome of this step (Kuechler and Vaishnavi 2008a, b).
Because DSR has its roots in engineering and the science of
the artificial (Au 2001; Baskerville 2008; Hevneretal.
2004; McKay and Marshall 2005), the creation and evaluation
of the IT artefact are frequently at the focus of researcher’s
attention. They must exist in any theory-generating
DSR approach. In addition, DSR provides additional theoretical
contributions by representing both an expository
device and the purpose of the testing (Gregor and Jones
2007), as we describe next.
5.3 Evaluation
In the evaluation step, the researcher must evaluate if the IT
artefact meets the requirements and solves the real-world
problem (Gregor and Jones 2007), because “What works
and doesn’t work will evolve over time based upon feedback
and learning from applying the ideas and analyzing the
results” (Basili 1996). This step tests if the criteria that were
explored in the awareness of the problem and suggestion step
are accomplished. The outcome of this step is represented by
performance measures for the IT artefact (Kuechler and Vaishnavi
2008a, b). Both the creation and evaluation of the IT
artefact are highly intertwined. Without a problem solution,
the preceding steps must be repeated, until the evaluation
actually indicates a satisfactory problem solution. Thus,
theory-generating DSR includes a refinement cycle of data
collection, development, and creation. For traditional
approaches, a solution signals the end (Aalst and Kumar
2003; Umapathyetal.2008), such that the contribution to
the knowledge base is the development and evaluation of the
IT artefact (Gregor 2006; Hevneretal.2004). However,
theory-generating DSR goes beyond that border.
5.4 Theory-generation
Beside the simultaneous use of GTM and DSR in the
awareness and suggestion step, theory-generating DSR enables
the researcher to conduct an additional theorygeneration
step. In our model, this step represents an extension
of Kuechler and Vaishnavi’s (2008a, b). It enables
additional theoretical insights, beyond the developed IT
artefact and therefore represents an intertwined process of
problem solving and theorizing. In particular, these additional
insights present the outcomes of this extension. It also

uncovers the potential of GTM to offer valuable research
advice, e.g., techniques such as theoretical sampling and
constant comparative method, to a growing area of interest
in IS research that focuses on the IT artefact, as demanded by
leading scholars (e.g., Hevner et al. 2004; Orlikowskiand
Iacono 2001). GTM focuses on generating a behavioural
understanding of the focal phenomenon and making a theoretical
contribution to the knowledge base, whereby in the
prior steps, DSR solved a real-world problem and developed
the basis for this analysis; the IT artefact.
While GTM enables DSR to offer an additional theoretical
contribution to the knowledge base beyond the IT artefact
(Hevner et al. 2004), DSR enables GTM to bring the IT
artefact to the centre of scholarly attention (Orlikowski and
Iacono 2001).
5.4.1 Additional theoretical sampling
Several additional steps provide new insights about the usage
and performance of the IT artefact. A created IT artefact can
change users’ behaviours and expectations, which would
define new requirements for follow-up projects and improvements
of the IT artefact. In this sense, the approach has created
a reusable contribution to the knowledge base. Weedman
(2008) thus describes a DSR project focusing on the evaluation
and theory-generating part supported by GTM
techniques.
Additional theoretical sampling involves further data collection,
after the creation of the IT artefact, with support
from GTM. These additional data enable the researchers to
explore the performance, usability, and assimilation of the
IT artefact. This step increases understanding of the usage of
the IT artefact and provides a means to explore changes in
people’s behaviour after they use the IT artefact.
Again, the data collection and analysis guide any subsequent
data collection. Therefore, the understanding of the
phenomena, the usage of the IT artefact, and its impact on
people’s behaviour emerge over time, through iterative cycles
of data collection and analysis (Glaser 1978). Over time, the
researcher achieves theoretical saturation (Eisenhardt 1989).
This extremely important step in theory-generating DSR
bridges the gap between design and behavioural sciences
(Holmström et al. 2009; Orlikowski and Iacono 2001). Some
prior research used additional theoretical sampling but unfortunately
only to evaluate the created IT artefact, without
clarifying the connection between the approaches (e.g., Albert
et al. 2004; Markusetal.2002).
5.4.2 Additional slices of data
Additional data may come from, though are not limited to,
the application of the IT artefact in the appropriate environment,
applicable knowledge from existing literature, and
results from the IT artefact evaluation. The overall analysis
moves up to a higher conceptual level (Levina and Vaast
2005). Furthermore, DSR provides unique data that cannot
be used in an ordinary GTM approach, in the form of socalled
throw-away prototypes (e.g., Markus et al. 2002) and
data from IT artefact testing. Not only does the usage of an
IT artefact depend on its environment, but IT artefacts explicitly
can change their environment and people’s behaviours
and thus the requirements for future IT artefacts.
However, throw-away prototypes offer only instantiations
as IT artefacts. Thus, in theory-generating DSR, additional
slices of data can contain more than literature, including
prototypes and other innovative data.
5.4.3 Categories
An essential step that condenses core categories during the
analysis involves the structured process of coding and analysis
(Glaser 1978). Different models for coding exist, but in the
Glaserian version, the researcher starts with so-called open
coding, such that he or she groups indicators from the data or
initial IT artefacts into concepts and then categories over time,
when it becomes clear which themes are of central interest to
yield the desired theoretical insights. The coding process itself
evolves and changes over time to become more selective, such
that prior conceptualizations and codes guide subsequent
steps. To generate these categories (and their properties) from
the collected data, researchers undergo several iterations of
coding and analysis and employ constant comparison (Glaser
1978). In this sense, constant comparison is the process of
constantly comparing instances of slices of data with other
instances of data to generate the categories. This comparison
can also be applied to define and refine existing categories. It
contributes to the development of theory by structuring the
analytic properties of the data and categories (Urquhart et al.
2010). For instance, Markus et al. (2002) use throw-away
prototypes as sorted categories, refined over such iteration
steps. Thus, researchers can reveal any clear fit of the IT
artefacts with primary or secondary data, including people’s
changed behaviours. The integration of extant literature into
the coding and conceptualization process can enrich the process
of creating and exploring categories. As a result, the
researcher must not “force” additional theoretical insights
onto the data. The additional data collection and analysis
continues until a point of theoretical saturation. This point is
reached when the researcher believes further data will not lead
to any more insights.
5.4.4 Additional theoretical insight
Inf Syst Front
Condensing the emerged categories allows for analyses of
the relationships among them and thus increases insights
(Glaser 1978). The end result, such as an added contribution

Inf Syst Front
to the domain of study in the form of grounded theory about
the developed IT artefact, expands the knowledge base (e.g.,
Weedman (2008) contributes to social science theory). After
identifying and defining the core categories, researchers
must assess the relationships among them to generate additional
theoretical insights. An important step to achieve the
final theoretical contribution of this study entails extensive
comparisons across the generated insights and prior published
work in the same domain. The emergent contribution
and its theoretical insights then can be integrated into
follow-up DSR projects as new requirements to consider.
5.5 Conclusion
To encourage a cumulative research tradition and benefit
from existing problem solutions, the results of the proposed
research approach should be used in further projects or
research cycles, integrated as an existing knowledge base
and slice of data in initial iterations.
In summary, theory-generating DSR integrates simultaneously
techniques from DSR and GTM to construct an IT
artefact and undertake an evaluation through conceptualizations
and theory building (Hevner et al. 2004; Winter 2008).
Prior DSR research has contributed to the knowledge base
with designs and actions affiliated with the IT artefact itself
(Gregor 2006). We see further potential derived from the
design and action processes. Our intertwined theorygenerating
approach provides a process model for such
research, though the sequence of the procedure naturally
must be defined by the researcher and his or her intentions.
That is, theory-generating DSR is highly dynamic and
should lead to a mutable IT artefact that can be easily
adapted to changing environments (Gregor and Jones 2007).
6 Evaluation of our approach to theory-generating
design science research
To evaluate the developed theory-generating DSR approach,
we present first a theoretical evaluation in which we describe
how the developed model fits to the IS design theory
components of Gregor and Jones (2007) and second, illustrate
with the help of an exemplary case of the extant DSR
literature how some of the theory-generating DSR elements
can be applied.
6.1 Theoretical evaluation
In Table 2 we provide a comparison of the proposed steps
and activities against the eight IS design theory components
provided by Gregor and Jones (2007). Beyond descriptions
of our developed theory-generating DSR approach, we offer
information about the activity in and legitimation for each
step. Finally, we illustrate how the IS design theory components
of Gregor and Jones (2007) relate to the steps of
theory-generating DSR.
The missing design theory component not addressed by
our theory-generating DSR approach is component 4 from
Gregor and Jones (2007): artefact mutability. Despite this
apparent gap, we believe the overall process of generating
additional theoretical insights from DSR finding leads to a
more mutable IT artefact that can be adapted easily to
changing environments. Furthermore, the detailed analysis
of the IT artefact’s requirements and its functionality in a
given or changing environment makes the IT artefact more
accountable, flexible, and mutable; it is not necessarily
embedded in some specific place, time, or community
(Orlikowski and Iacono 2001). Therefore, though our approach
does not explicitly deal with IT artefact mutability,
which may be regarded as a limitation, we believe it
addresses artefact mutability implicitly, through the process
of iterating IT artefact cycles (problem solving, theoretical
saturation) as the IT artefact goes through repeated cycles of
change and reflections. However, because our theorygenerating
DSR approach aims to achieve new insights,
testing artefact mutability with an upfront chosen kernel
theory admittedly is not possible.
6.2 Illustration of our model through an exemplary case
Weedman (2008) analyzed a design project from a behavioural
and social science perspective. This design project
was about the implementation of ‘Sequoia 2000’ which
represents an interactive information system based on the
data-handling needs in global change science. Earth scientists
need to analyze a huge mass of data of different types,
e.g., satellite data, text, raster, and vector with Sequoia 2000
as the platform that allows to analyze and share the data
among them. The development of Sequoia 2000 can be used
to illustrate several steps of our model even though they are
not explicitly followed by Weedman (2008).
Weedman (2008) entered the field and positioned their
research among others on the design theories of Gregor and
Jones (2007). She conducted an extensive literature review
on DSR and design theories to be informed (awareness) of
the problem. However, their problem was grounded in the
need for a collaboration and analysis tool for earth scientists.
The requirements for the huge data-handling problem were
officially stated as: vastly increased storage, much faster
networking, visualization techniques for modelling, and a
new database management system to replace the existing
flat file system. The expectation towards the system was that
it would change the way how global change researchers
work, for example by allowing much more data transfer
than in the past based on increased storage and network
speed. However, initial theoretical sampling, such as

Table 2 Eight components of IS design theory in different DSR approaches
Step Derived from Activity Reason Supported Design Theory
Components
Awareness of problem
and suggestion
DSR (Kuechler and
Vaishnavi 2008a, b;
Peffers et al. 2008)
Theoretical sampling GTM (Glaser 1978) Systematic collection and
analysis of data
Slices of data GTM (Fernandez
2004; Levina and
Vaast 2005)
Tentative design and
requirements
DSR (e.g., Hevner et
al. 2004)
Development DSR (Hevner et al.
2004; Kuechler and
Vaishnavi 2008a, b)
Evaluation DSR (Basili 1996;
Kuechler and
Vaishnavi 2008a, b)
Additional theoretical
sampling
Additional slices of
data
interviews with the prospective users, pointed out that their
understanding of the problem was somewhat different. They
thought that it would only provide them with more advanced
hard- and software, without however triggering a fundamental
change in earth science. For instance, a programmer working
for earth scientists pointed out that he does not understand that
this project is beneficial for him. To meet these moving
targets, a questionnaire was sent out to all participants and
several interviews were conducted. Thereby, the participants
and the collected data represent environmental factors as slices
of data. For instance, the interviewees were asked about
potential problems with Sequoia 2000. They pointed out that
some problems with the handling of the technology could
occur. The interviewees based these statements on experiences
Definition of research lens Need for appropriate
combination of design and
behavioural science
Identification of necessary
slices of data for both
DSR and behavioural
understanding
Shaping requirements and
business needs
Development of the IT
artefact
Provides the means to identify
requirements and a basis for
generating theory from an
IT artefact
Allows for combination of
requirement identification/
business needs and
theoretical conceptualization
Fosters the relevance of the
problem solution
Strengthens the practical
contribution of the DSR
finding
Evaluation of the IT artefact Confirms whether the
problem is solved
GTM (Glaser 1978) Collection and analysis of
additional data based on
IT artefact embedded in
the environment
GTM (Fernandez
2004; Levina and
Vaast 2005)
Identification of additional
slices of data for
behavioural understanding
Categories GTM (Glaser 1978) Generation of conceptual
categories (and properties)
from the intertwined data
Conclusion DSR (Hevner et al.
2004; Kuechler and
Vaishnavi 2008a, b)
collection and analysis
Embedding theoretical
insights into DSR
literature
Extends the scope of the
analysis beyond business
needs/requirements toward
a behavioural understanding
based on preliminary
theoretical sampling
Strengthens the theoretical
conceptualization process
within the DSR process
Forces DSR researchers to
conceptualize and reach
theoretical abstraction
Closes the loop by stimulating
further DSR research that
builds on the generated
theoretical findings
Inf Syst Front
Component 6: Underlying
theories to justify and
explain the design
Component 2: Through a
detailed analysis, the
entities of interest can be
derived and clearly
communicated
Component 1: Defines
what the system is for
Component 8: The
physical implementation
of the IT artefact
Component 5: Evaluates
whether the IT artefact
satisfies the requirements
Component 7: Helps
implement additional
theoretical insights
about the IT artefact
through additional data
collection and analysis
Component 3: An
additional theoretical
contribution of the IT
artefact and its
behaviours can function
as a blueprint for followup
projects
with the programmers who emphasized that assistance in the
use of the technology was not within scope. As a consequence,
the development was adapted to this insight and earth
scientists were included prior to the implementation. This
implied that the potential user became accustomed to the
technology which resembles an intertwined process of data
collection and analysis to refine the tentative design and the
requirements of the technology.
After solving these problems, Sequoia 2000 was developed
and evaluated stepwise. The first prototypes were fed
with data from the earth scientists to test the performance.
Several problems occurred in this process. For instance, the
network went down several times or the file system broke
down because of the huge mass of data. As a consequence,

Inf Syst Front
additional theoretical sampling and slices of data were
collected from the participants. The programmers saw the
earth scientists involved in the build and evaluate step as a
kind of test bed for early systems testing and evaluation.
However, the earth scientists saw their role as customers not
being interested in acting as test persons but rather receive a
stable and ready-to-use technology that enables them to
conduct their research. The understanding of the earth scientists
as customers rather than as participants in the design
reinforced their research goals as being more important than
the design goals. Hence, Sequoia 2000 had to go through
several other improvement loops because of additional
deliverables that were not fixed in the initial requirements
analysis.
The entire Sequoia project can be seen as an example of
including behavioural science aspects into the design and
development of a technology. This project generated different
conclusions and thereby additions to the knowledge
base. In particular, Weedman (2008) developed two different
types of contributions: First, she developed an explanation
of how the meta-requirements of handling large or
massive amounts of data in the domain of earth science
could be matched with different kinds of solution components,
including a cross-disciplinary metadata schema and
associated chunking strategy for database management,
principles of data visualization, and the principle of data
reusability for running different analysis tasks. According to
Baskerville and Pries-Heje (2010), this type of contribution
resembles an ‘explanatory design theory’, i.e. the design
product (Walls et al. 1992). However, the second and probably
more important contribution of Weedman (2008) was the
development of categories and relationships between them
regarding the design process of solving ill-defined or wicked
problems. Such a ‘design process theory’ (Baskerville and
Pries-Heje 2010) is represented in the following Fig. 3.
Accordingly, different categories emerged from
Weedman’s (2008) analysis which are related to the core
Involvement of
users as partners
Collaboration
Reflective designeruser
conversation
Interdisciplinary
cooperation
Mutual interests
and benefits
Different viewpoints
and knowledge
Incentives/
motivation
Fig. 3 Derived design process theory drawn from Weedman (2008)
notion of collaborative DSR projects with ‘reflective
designer-user conversations’ at its centre. Adapted from
Schön (1983, 1992), it represents the idea that designers
and users engage in interactive conversations about both
problem formulation and solving. Such a conversation
stimulates a kind of reflection-in-action and shared understanding
between the designers and users which
leads to enhanced design outcomes. This central aspect
of the collaboration is leveraged and enabled through
three further concepts. First, designers and users are
motivated by placing mutual interest and benefits at
the centre of team composition. According to Weedman
(2008), this can be further leveraged through appropriate
incentive structures. Second, the concept of interdisciplinary
cooperation provides different viewpoints and
knowledge into the design process. Therewith, different
knowledge types to be combined, including domain expert
knowledge and design knowledge. Finally, users are involved
as partners in the design process to stimulate collaboration
with designers. In summary—drawing from Weedman (2008)
—a design process theory can be derived that addresses
motivational (i.e. mutual interests and benefits), cognitive (i.e.
interdisciplinary knowledge), and behavioural (i.e.
designer-user collaboration) aspects.
7 Discussion
While DSR has become an established area of research
within IS, it still lacks the maturity of more accepted
research approaches in terms of set of research instruments
used or evaluation of theoretical contributions made. Since the
seminal work of Hevner et al. (2004) there is an even stronger
debate how DSR can be combined with other research
methods to increase its rigor.
Theory-generating DSR focuses on the creation and evaluation
of theoretical abstractions from an IT artefact by combining
elements of constructive research with key concepts
and techniques from interpretative research, thus allowing for
grounded theorizing. Our suggested approach extends the
general design cycle for DSR proposed by Kuechler and
Vaishnavi’s (2008a, b). Different from Holmström et al.
(2009) and Sein et al. (2011) we do not include behavioural
science as an additional and independent component into DSR
but rather develop a process model that enables a simultaneous
and highly intertwined use of both. In so doing, we can
combine the development of an IT artefact that solves a class
of real-world problems in a new and innovative way with
rigorous theorizing, such as about how the newly developed
IT artefact interacts with its environment.
In particular, the awareness of the problem and the suggestion
step can be enhanced through a detailed analysis of
the environment and the extant knowledge base. Through

theoretical sampling, slices of data, and the constant comparison
method researchers may be able to explore the
problem and the phenomenon more precisely and analyze
potential failures in the development in advance. A tentative
design and the requirements of the IT artefact are more
detailed through this refined step in theory-generating
DSR. In addition, the extension of theory-generation after
the evaluation of the IT artefact enables researchers to create
an additional theoretical insight, e.g., from the application of
the IT artefact in the appropriate environment, applicable
knowledge from existing literature, or results from the IT
artefact evaluation. This additional theoretical component
offers another research outcome, but it also complements
the underlying DSR approach and thereby meets the
requirements for a IS design theory, as specified by Gregor
and Jones (2007). As an exemplary case from extant literature,
we used Weedman (2008) who used different behavioural
science elements to analyze their developed IT
artefact in more detail.
From a scientific point of view, theory-generating DSR
focuses on the IT artefact and its improvement, as well as
offering additional theoretical insights based on the use of the
IT artefact. Thus, its reusable contribution adds to existing
knowledge bases. Furthermore, it designs and creates IT artefacts
as means to discover new knowledge (Baskerville et al.
2009), which distinguishes theory-generating DSR from, e.g.,
action research, which aims to create change in an organizational
setting and studies the subsequent effects (Baskerville et
al. 2009). Moreover, action researchers take an action and
apply extant theories within the course of the research project
(Coglan and Coughlan 2002; McKay and Marshall 2001). It
thus explores the research project from an internal perspective;
the action researcher works with the people directly affected
by the action or who have the potential to influence the action
in their environment (Avison et al. 1999). In contrast, a theorygenerating
DSR researcher observes phenomena and interacts
only within the scientific environment. Accordingly, theorygenerating
DSR provides findings about the potential
improvements to an IT artefact, and the researcher simply
observes.
8 Conclusion
We have developed and presented a conceptual approach
for a complementary use of DSR and behavioural
science research elements, on the basis of Kuechler
and Vaishnavi’s (2008a, b) model and comparison with
components necessary for IS design theories. For our initial
theory development, we used interrelations to combine previously
unconnected bodies of knowledge (DSR and GTM). In
the course of this combination, we identified gaps in their
conceptualizations and added missing components to an
already existing DSR model from Kuechler and Vaishnavi’s
(2008a, b). However, the proposed approach is a general
process model, rather than a strict recipe, and has not yet been
challenged or tested through application to actual DSR
projects.
According to Hevner et al. (2004), theoretical contributions
to the knowledge base represent an important and
necessary part of DSR. We therefore propose the combination
of DSR and GTM to unite design and behavioural
aspects. In particular, we illustrate how design and behavioural
research elements combine effectively in a pluralistic
research design, which responds to calls to find potential
research improvements (e.g., Mingers 2001).
Future research in this direction should evaluate whether
such a combined research method is applicable or not in
DSR projects. The used behavioural science method in this
paper uses elements from GTM which is only one out of
many methods in the field. Hence, this paper presents a new
attempt to combine design with behavioural research. Future
research should evaluate further the simultaneous use of
behavioural and DSR methods into a pluralistic research
design.
In summary, for a scientific discipline, generating contributions
to its knowledge base is at least as important as
solving real-world problems. We strongly believe that the
proposed theory-generating DSR approach as one possible
combination of design and behavioural science can support
this goal by providing a process model for researchers who
strive to follow and accomplish this aim in a research
setting. From a theoretical perspective, this model can provide
a fruitful contribution to the DSR community and
expect to be very interesting to researchers focusing on
DSR from a methodological point of view. From a practical
perspective, managers could use the developed model for
their DSR projects and thereby improve the development of
prototypes in a structured way to avoid failure or repair
loops.
Appendix: Review of DSR Articles
Inf Syst Front
In the course of our literature research, we found several
articles beside Weedman (2008) that offer support for various
steps of our theory-generating DSR approach. Each
article depicts a usage of DSR methods. Many articles
present the IT artefact as an instantiation, though as Hevner
et al. (2004) state: “IT artefacts can also be represented by
constructs, models, methods, instantiations, or a combination
thereof” (see also, March and Smith 1995). In Table 3,
we list articles that discuss typical examples of DSR and that
provided some basic ideas for our theory-generating DSR
approach.

Inf Syst Front
Table 3 Literature using DSR as a research approach
Article Developed IT-Artefact Description
Aalst and Kumar (2003) Interorganizational routing
language based on XML
Abbasi and Chen (2008) CyberGate: A design framework
and system for text analysis of
computer-mediated
References
communication
Albert et al. (2004) GIST: A model of the design and
management of content and
interactivity of customer-centric
Web sites
Umapathy et al. (2008) A model for knowledge reuse for
designing enterprise integration
solutions
Markus et al. (2002) A design theory for systems that
support emergent knowledge
processes
Aalst, W., & Kumar, A. (2003). XML-based schema definition for
support of interorganizational workflow. Inform Syst Res, 14(1),
23–46.
Abbasi, A., & Chen, H. (2008). CyberGate: a design framework and
system for text analysis of computer-mediated communication.
MIS Quarterly, 32(4), 811–837.
Albert, T. C., Goes, P. B., & Gupta, A. (2004). GIST: a model for design
and management of content and interactivity of customer-centric
web sites. MIS Quarterly, 28(2), 161–182.
Allen, D. K., Colligan, D., Finnie, A., et al. (2000). Trust, power and
interorganizational information systems: the case of the electronic
trading community TransLease. Inform Syst Res, 10(1), 21–40.
Arazy, O., Kumar, N., & Shapira, B. (2010). A theory-driven design
framework for social recommender systems. J Assoc Inform Syst
Online, 11(9), 455–490.
Au, Y. A. (2001). Design Science I: the role of design science in electronic
commerce research. Comm Assoc Inform Syst, 7(1), 1–15.
Avison, D., Lau, F., Myers, M., et al. (1999). Action research. Comm
ACM, 42(1), 94–97.
Basili VR (1996) The role of experimentation in software engineering:
past, current, and future. Proceedings of the International
Conference on Information Systems (ICIS), Ohio, USA.
Baskerville, R. (2008). What design science is not. Eur J Inform Syst,
17(5), 441–443.
The authors describe a standard problem-oriented approach. They use
DSR methods accidentally to create a new interorganizational routing
language. They test this language with a prototype implementation.
The article depicts the typical engineering and problem-oriented
approach from which DSR is derived.
The authors conduct a procedure analogous to Aalst and Kumar’s
(2003) and develop CyberGate, an IT artefact for text analysis
techniques.
The authors describe a standard problem-oriented approach. The IT
artefact is derived from existing models, methods, and additional
data. The requirements are obtained from an analysis of the existing
environment. The authors conduct an elaborate evaluation of the IT
artefact by collecting interview and survey data among users to get a
deeper insight on its usability. Moreover, they conduct a gap analysis
to identify potential gaps and missing values in their created model.
Unfortunately, they do not use the data they gathered to improve the
IT artefact.
The authors describe a standard problem-oriented approach. The
requirements of the IT artefact are described by an analysis of the
existing environment. The evaluation of the IT artefact involves a
comparison of the old IT artefact with the new one.
The authors use DSR to create their IT artefact without mentioning it.
Their research offers a closely intertwined combination of DSR and
behavioural science. The authors build their IT artefact on existing
kernel theories and conduct interviews to ensure the usability of their
IT artefact. Unfortunately, these steps were not sequential. Several
iterations of building and testing the IT artefact were conducted with
so-called throw-away prototypes. The article provides extremely
fruitful suggestions for our theory-generating DSR approach.
Baskerville, R., & Pries-Heje, J. (1999). Grounded action research: a
method for understanding IT in practice. Account Manag Inform
Tech, 9(1), 1–23.
Baskerville, R., & Pries-Heje, J. (2010). Explanatory design theory.
Business & Information Systems Engineering, 2(5), 271–282.
Baskerville, R., & Stage, J. (2001). Accommodating emergent work
practices: ethnographic choice of method fragments. In B.
Fitzgerald, N. Russo, & J. I. DeGross (Eds.), Realigning research
and practice: the social and organizational perspectives. Boston:
Kluwer Academic Publishers.
Baskerville R, Pries-Heje J, Venable J (2009) Soft design science
methodology. Proceedings of Design Science Research in Information
Systems and Technology (DESRIST), Philadelphia, USA.
Benbasat, I., & Zmund, R. W. (2003). The identity crisis within the IS
discipline: defining and communicating the discipline’s core
properties. MIS Quarterly, 27(2), 183–194.
Bryant, A., & Charmaz, K. (2007a). Grounded theory in historical perspective:
an epistemological account. In A. Bryant & K. Charmaz
(Eds.), The SAGE handbook of grounded theory. London: SAGE
Publications.
Bryant, A., & Charmaz, K. (2007b). Grounded theory research: methods
and practices. In A. Bryant & K. Charmaz (Eds.), The SAGE
handbook of grounded theory. London: SAGE Publications.
Carlsson SA (2006) Towards an information systems design research
framework: a critical realist perspective. Proceedings of Design
Science Research in Information Systems and Technology
(DESRIST), Claremont, USA.

Coglan, D., & Coughlan, P. (2002). Action research for operations
management. Int J Oper Prod Manag, 22(2), 220–240.
Dubin, R. (1978). Theory building. New York: Free Press.
Eisenhardt, K. M. (1989). Building theories from case study research.
Acad Manage Rev, 14(4), 532–550.
Eisenhardt, K. M., & Graebner, M. E. (2007). Theory building from
cases: opportunities and challenges. Acad Manage J, 50(1), 25–32.
Fernandez WD (2004) The grounded theory method and case study
data in IS research: issues and design. Proceedings of Information
Systems Foundations Workshop: Constructing and Criticising,
Canberra, Australia.
Glaser, B. G. (1978). Theoretical sensitivity: advances in the methodology
of grounded theory. Mill Valley: Sociology Press.
Glaser, B. G. (1998). Doing grounded theory: issues and discussions.
Mill Valley: Sociology Press.
Glaser, B. G. (2007). Doing formal theory. London: SAGE
Publications.
Glaser, B. G., & Strauss, A. L. (1967). The discovery of grounded
theory: strategies for qualitative research. Chicago: Aldine
Publishing Company.
Goldkuhl, G. (2004). Design theories in information systems—a need
for multi-grounding. J Inform Tech Theor Appl, 6(2), 59–72.
Gregor, S. (2002). Design theory in information systems. Aust J Inform
Syst, 10(1), 14–22.
Gregor, S. (2006). The nature of theory in information systems. MIS
Quarterly, 30(3), 611–642.
Gregor, S., & Jones, D. (2007). The anatomy of a design theory. J
Assoc Inform Syst Online, 8(5), 312–335.
Gregory, R. W., & Muntermann, J. (2011) Theorizing in design science
research: inductive versus deductive approaches. In: Thirty second
international conference on information systems, Shanghai,
China.
Hevner, A. R., & March, S. T. (2003). The information systems
research cycle. IT Systems Perspective, 36(11), 111–113.
Hevner, A., March, S., Park, J., et al. (2004). Design science in
information systems research. MIS Quarterly, 28(1), 75–105.
Holmström, J. B., Ketokivi, M., & Hameri, A. P. (2009). Bridging
practice and theory: a design science approach. Decision Sci, 40
(1), 65–87.
Ketokivi, M., & Mantere, S. (2010). Two strategies for inductive
reasoning in organizational research. Acad Manage Rev, 35(2),
315–333.
Kuechler, B., & Vaishnavi, V. (2008a). On theory development in
design science research: anatomy of a research project. Eur J
Inform Syst, 17(5), 489–504.
Kuechler B, Vaishnavi V (2008b) Theory development in design
science research: anatomy of a research project. In Proceedings
of Design Science Research in Information Systems and Technology
(DESRIST), Atlanta, GA.
Lee JS, Pries-Heje J, Baskerville R (2011) Theorizing in design science
research. 6th International Conference on Design Science Research
in Information Systems and Technology, Milwaukee, WI.
Levina, N., & Vaast, E. (2005). The emergence of boundary spanning
competence in practice: implications for implementation and use
of information systems. MIS Quarterly, 29(2), 335–363.
Madill, A., Jordan, A., & Shirley, C. (2000). Objectivity and reliability
in qualitative analysis: realist, contextualist and radical constructionist
epistemologies. Br J Psychol, 91(1), 1–20.
March, S. T., & Smith, G. (1995). Design and natural science research
on information technology. Decis Support Syst, 15(4), 251–266.
Markus, M. L., Majchrzak, A., & Gasser, L. (2002). A design theory
for systems that support emergent knowledge processes. MIS
Quarterly, 26(3), 179–212.
Inf Syst Front
McKay, J., & Marshall, P. (2001). The dual imperatives of action
research. Inform Tech People, 14(1), 46–59.
McKay J, Marshall P (2005) A review of design science in information
systems. Proceedings of the Australasian Conference on Information
Systems (ACIS), Sydney, Australia.
Mingers, J. (2001). Combining IS research methods: towards a pluralist
methodology. Inform Syst Res, 12(3), 240–259.
Nagel, E. (1961). The structure of science. New York: Harcourt, Brace
& World.
Nunamaker, J. F., Chen, M., & Purdin, T. D. M. (1991). Systems
development in information systems research. J Manag Inform
Syst, 7(3), 89–106.
Orlikowski, W. J., & Iacono, C. S. (2001). Research commentary:
desperately seeking the ‘IT’ in IT research—a call to theorizing
the IT artifact. Inform Syst Res, 12(2), 121–134.
Peffers, K., Tuunanen, T., Rothenberger, M. A., et al. (2008). A design
science research methodology for information systems research. J
Manag Inform Syst, 24(3), 45–77.
Pries-Heje, J., & Baskerville, R. (2008). The design theory nexus. MIS
Quarterly, 32(4), 731–755.
Schön, D. A. (1983). The reflective practitioner: how professionals
think in action. New York City: Basic Books.
Schön, D. A. (1992). Designing as reflective conversation with the
materials of a design situation. Knowl Base Syst, 5(1), 3–14.
Sein, M. K., Henfridsson, O., Puaro, S., et al. (2011). Action design
research. MIS Quarterly, 35(1), 37–56.
Simon, H. A. (1969). The science of the artificial. Cambridge: MIT
Press.
Strauss, A., & Corbin, J. (1990). Basics of qualitative research: grounded
theory procedures and techniques. London: Sage Publications.
Suddaby, R. (2006). From the editors: what grounded theory is not.
Acad Manage J, 49(4), 633–642.
Udo K (2005) ‘Emergence’ vs. ‘Forcing’ of empirical data? A crucial
problem of ‘grounded theory’ reconsidered. Forum Qualitative
Social Research, 6(2), Art. 27.
Umapathy, K., Purao, S., & Barton, R. R. (2008). Designing enterprise
integration solutions: effectively. Eur J Inform Syst, 17(5), 518–
527.
Urquhart, C. (2007). The evolving nature of grounded theory method:
the case of the information systems discipline. In A. Bryant & K.
Charmaz (Eds.), The SAGE handbook of grounded theory. London:
SAGE Publications.
Urquhart, C., Lehmann, H., & Myers, M. D. (2010). Putting the
‘theory’ back into grounded theory: guidelines for grounded theory
studies in information systems. Inform Syst J, 20(4), 357–381.
Vaishnavi, V., & Kuechler, W. (2008). Design science research methods
and patterns—innovating information and communication
technology. Boca Reton: Auerbach Publications Taylor & Francis
Group.
Walls, J. G., Widmeyer, G. R., & El Sawy, O. A. (1992). Building an
information system design theory for vigilant EIS. Inform Syst
Res, 3(1), 36–59.
Weber, S., Beck, R., & Gregory, R. W. (2012) Combining design
science and design research perspectives - Findings of three
prototyping projects. In: 45 Hawaii international conference on
system sciences, Wailea, Maui, HI.
Weedman, J. (2008). Client as designer in collaborative design science
research projects: what does social science design theory tell us?
Eur J Inform Syst, 17(5), 476–488.
Winter, R. (2008). Design science research in Europe. Eur J Inform
Syst, 17(5), 470–475.
Yadav, M. S. (2010). The decline of conceptual articles and
implications for knowledge development. JMarket,74(1), 1–19.

Inf Syst Front
Prof. Dr. Roman Beck is the E-Finance and Services Science
Chair at Goethe University, Frankfurt, Germany. His research
focuses on the creation, sourcing, and management of knowledge
intensive IT services and has been published in journals such as
Communications of AIS, Electronic Markets, IEEE Transactions
on Software Engineering, Information Systems Frontiers, Information
Technology & People, and others.
Sven Weber is a research assistant and doctoral student at the
E-Finance Lab at Goethe University, Frankfurt, Germany. His research
focuses on the combination of design science and behavioural science
research and the management of communication systems. His research
has been published at several IS conferences (e.g., AMCIS, DESRIST,
HICSS).
Dr. Robert Wayne Gregory is an Assistant Professor at the Chair
of Electronic Finance and Digital Markets at the University of
Göttingen. He holds a PhD in Information Systems from University
of Frankfurt and has a focus on both behavioural and design
science-oriented research. Robert has been involved in a wide
range of research projects that have been conducted in collaboration with
industry, especially with financial service providers and banks. His
research has been published in outlets such as Information Technology
& People and ICIS.

BISE – RESEARCH PAPER
The Interplay of Project Control
and Interorganizational Learning: Mitigating Effects
on Cultural Differences in Global, Multisource ISD
Outsourcing Projects
The study examines how to mitigate the cultural differences inherent in global, multisource,
information systems development outsourcing projects. Its main finding is that the
influence of informal control and interorganizational learning on formal control does not
remain constant. Rather, it changes over time, from providing operational information to
reducing formal management efforts. In turn, transparency created through formal
management mechanisms provides room for effective informal control mechanisms and
interorganizational learning. This interplay supports the mitigation of cultural differences
through the harmonization of work-related values and practices.
DOI 10.1007/s12599-012-0217-5
The Authors
Prof. Dr. Roman Beck
E-Finance Lab & Institute of
Information Systems
J.W. Goethe University
House of Finance Grueneburgplatz 1
60323 Frankfurt am Main
Germany
rbeck@wiwi.uni-frankfurt.de
url: http://www.servicesscience.de
Dr. Katharina Schott (�)
Institute of Information Systems
J.W. Goethe University
House of Finance Grueneburgplatz 1
60323 Frankfurt am Main
Germany
katharina.schott@googlemail.com
Received: 2011-06-26
Accepted: 2012-02-14
Accepted after two revisions by Prof.
Leidner.
Business & Information Systems Engineering
This article is also available in German
in print and via http://www.
wirtschaftsinformatik.de: Schott K,
Beck R (2012) Das Zusammenspiel
von Projektsteuerung und interorganisationalem
Lernen und
dessen Effekt auf kulturelle Unterschiede
in globalen ISE-Outsourcing-
Projekten mit mehreren Dienstleistern.
WIRTSCHAFTSINFORMATIK. doi:
10.1007/s11576-012-0324-4.
© Gabler Verlag 2012
1Introduction
Worldwide, companies benefit from sophisticated
sourcing strategies that rely
on both near- and offshore destinations.
Unlike offshore outsourcing, nearshore
outsourcing aims to mitigate offshorespecific
challenges, such as significant
time zone differences or language barriers,
as well as to exploit nearshorespecific
advantages, such as closer interactions
through geographic proximity
(e.g., Meyer and Stobbe 2007). Internationally
operating vendors thus increasingly
take advantage of hybrid global delivery
models and organize service delivery
across off-, near-, and onshore locations
(Willcocks et al. 2007). Yet organizing
smooth global service delivery
remains challenging, especially in constellations
in which client companies
deal with multiple, globally distributed
vendors on a single project. For example,
in global, multisource projects
for information systems development
(ISD), both national cultures and multiple
organizational cultures must converge.
The perceived cultural distance between
client and vendor thus increases,
which requires more integrated management
approaches to address greater demand
for communication and coordination
(Carmel and Agarwal 2001; Hildenbrand
et al. 2007).
Prior research in the global ISD
and outsourcing domains spans multiple
streams and thus focuses on various
management aspects. Therefore, a
large part of empirical research in global
IS outsourcing analyzes cultural differences
on the national, organizational,
and individual level and how to deal
with these differences from a project
management point of view (e.g., David
et al. 2008; Iacovou and Nakatsu 2008;
Winkler et al. 2007). In this context, it
has been shown that in particular control
mechanisms as well as interorganizational
learning contribute to the mitigation
of cultural differences (Gregory
2010b).
Literature in the global ISD and outsourcing
domains covers numerous studies
considering control issues, such as the
role of contracts (e.g., Gopal and Sivaramakrishnan
2008; LacityandWillcocks
1998) or other formal control mechanisms
(e.g., Tiwana 2008), as well as

BISE – RESEARCH PAPER
the input of informal control mechanisms
(Holmström Olsson et al. 2008).
Also, there are several studies focusing
on interorganizational learning, mainly
in relation to cross-cultural (Nicholson
and Sahay 2001; Vlaar et al. 2008; Walsham
2002) and knowledge (e.g., Kotlarsky
et al. 2008; Leonardi and Bailey
2008; Nicholson and Sahay 2004) issues.
However, while these studies cover
control mechanisms and interorganizational
learning separately, the existing literature
to the best of our knowledge
has not investigated in more detail on
the relationship between control mechanisms
and interorganizational learning
against the background of cultural differences
(Gregory 2010b), in particular
in an analysis that features both the
client’s and multiple vendors’ perspectives.
Seeking to address this gap, we adopt
an exploratory research design and analyze
the interplay of formal and informal
control with interorganizational learning
in global, multisource, ISD outsourcing
projects. Accordingly, we seek to answer
the following research question:
How do formal and informal control
mechanisms and interorganizational
learning interact and contribute
to the mitigation of cultural
differences in global, multisource,
ISD outsourcing projects?
To answer this question, we apply an exploratory
case study design with a global,
multisource, ISD outsourcing project initiated
by a large German financial institute
as the research object. The objective
of this project was to reengineer the
financial institute’s online banking systems,
and the project team included approximately
100 people from five different
organizations (client and four vendors),
distributed across Germany, Spain,
Brazil, and India.
We begin our account of this project
and the findings based on it by presenting
theoretical foundations for our understanding
of formal and informal control,
as well as interorganizational learning.
After we describe the case study, the underlying
research methodology, and the
analysis results, we discuss the findings in
the light of previous literature. The final
section summarizes key findings and provides
some implications for research and
practice.
2 Theoretical Background:
Control Dynamics and
Interorganizational Learning
Prior studies on domestic IS outsourcing
emphasize the general importance
of social capabilities, which not only facilitate
intra-organizational cooperation
but also foster mutual trust and performance
among client organizations and
external IS vendors (Dibbern et al. 2003).
However, a recent review of IS outsourcing
practices reveals that global IS outsourcing
projects actually cope with even
more challenges (Lacity et al. 2009). In
the global context of ISD outsourcing,
the need for social competency is significant
because individual participants
must cope with the global sourcingspecific
distance between the client and
vendor. This so called client-vendor distance
encompasses not only geographic,
economic, and political distance but also
and in particular cultural differences
(Vogt et al. 2009;Winkleretal.2007).
In this regard governance can be conducted
by the joint reliance on different
governance or control mechanisms such
as informal and formal control mechanisms
for the governance of economic
transaction (Adler 2001; Bradach and Eccles
1989; Cannon et al. 2000). In the
global IS outsourcing domain it has been
shown that successful control balancing
using combinations of different control
modes is a promising way to manage cultural
differences inherent in such projects
(Gregory 2010a).
Literature on control suggests two basic
modes: formal and social control.
Formal control involves the establishment
and application of codified rules,
goals, and procedures to define, monitor,
and evaluate performance (Das and Teng
2001). They usually involve explicit information
transfers and include for example
formal reporting guidelines and frequent
meetings between key representatives
(Inkpen and Currall 2004). Prior
research in global IS outsourcing has
employed a control perspective to show
how the contract between the client and
the vendor is enforced via formal behavior
and outcome control; accordingly,
the client controls the behavior and outcomes
of his vendor (Choudhury and
Sabherwal 2003). Social control involves
trust-based mechanisms and operates via
the development of shared values and
norms (Das and Teng 2001). According
to Das and Teng (1998), the key difference
between formal and social control
is that “formal control is more of a strict
evaluation of performance while social
control is about dealing with people.”
(p. 501). Examples for social controls include
socialization, training, and spontaneous
interactions between representatives
of the exchange partners (Das and
Teng 1998). In the global IS outsourcing
literature, social control has most often
been conceptualized as informal control
(Choudhury and Sabherwal 2003; Kirsch
1996).
An expanding body of research notes
control dynamics and examines how
and why control modes change across
project phases. For example, Choudhury
and Sabherwal (2003) analyzetheevolution
of control portfolios in information
systems development (ISD) outsourcing
projects and reveal several factors that influence
the choice and evolution of the
control mechanisms. To extend such research,
Kirsch (2004) describes how control
modes change during three project
phases of internal global IS deployment
projects. Changes in the control mode
appear triggered by factors from three
categories: project context, stakeholder
context, and global context.
Besides control, also interorganizational
learning has been shown to contribute
to the mitigation of cultural differences
in global IS outsourcing projects.
On the one hand, interorganizational
learning comprises the accumulation of
relevant business, functional, and clientspecific
knowledge, as vendors must accumulate
business knowledge about their
client’s application domain, functional
knowledge about the client’s IT infrastructure
and systems, and specific recognition
of its functional requirements and
processes (Dibbern et al. 2008). On the
other hand, interorganizational learning
pertains the important topic of knowledge
transfer in global IS outsourcing
projects. The existing literature repeatedly
demonstrates positive effects of successful
knowledge transfer (e.g., Kotlarsky
and Oshri 2005; Nicholson and
Sahay 2004; Oshrietal.2007; Rottman
2008), and many researchers cite a lack
of knowledge transfer as a major drawback
for global ISD outsourcing projects
(e.g., David et al. 2008; Dibbernetal.
2008; Gupta and Raval 1999;Kliem2004;
Leonardi and Bailey 2008). And third,
interorganizational learning in a global
context also refers to cross-cultural issues
and the development of cultural intelligence,
defined as “a form of firmlevel
capability in functioning effectively
Business & Information Systems Engineering

Table 1 Overview of the multisourcing portfolio
Organization Area of responsibility Locations involved
ARCHITECT • Definition of architectural framework
• Implementation controls
Two locations in Germany
IMPLEMENT • Functional design Four locations in Spain
• Technical design
• Implementation
One location in Brazil
SCREEN • End-user front-end design One location in Germany
TEST • Software test One location in India
in culturally diverse situations” (Ang and
Inkpen 2008, p. 338), as both the client
and the vendor must negotiate differences
in their values and work practices
and learn how to adapt for the project to
succeed (e.g., Carmel and Agarwal 2001;
Krishna et al. 2004; Levina and Vaast
2008).
In summary, both control mechanisms
as well as interorganizational learning
have been shown to individually contribute
to the mitigation of cultural differences.
However, there remains little
understanding of the interplay of formal
and informal control mechanisms
with interorganizational learning or how
the interaction might contribute to mitigate
cultural differences in global, multisource,
ISD outsourcing projects. Furthermore,
most of the preceding studies
address dyadic client-vendor relationships,
whereas the specific challenges of
multisource projects remain unconsidered.
3 Research Methods
As outlined, we lack sufficient knowledge
about the interplay of control and
learning in global, multisource, IS outsourcing
projects. Therefore, to increase
this understanding, as well as to reveal
the interactive effects on cultural differences,
the present qualitative research
features an in-depth, exploratory, singlecase
study (Stebbins 2001;Yin2003). The
subsequent sections describe the underlying
case (3.1) as well as the procedures
used for data collection (3.2) and analysis
(3.3). A chronology of the overall research
process is presented at the end of
this chapter.
3.1 Case Description
The primary unit of analysis was a global
ISD outsourcing project to reengineer a
Business & Information Systems Engineering
financial institution’s online banking system.
The project was initiated because the
old system’s technology required a high
degree of costly expertise, and its maintenance
was set to expire soon. Thus,
the financial institution (BANK) decided
to migrate its system to a new technology.
To develop this new system, BANK
applied a multisourcing strategy to reduce
the risks of dependence. Therefore,
it included four vendors in the project,
as summarized in Table 1. The project
started in October 2008 and finished
in December 2009, successfully and on
time.
ARCHITECT, a German boutique consulting
firm, designed the architectural
framework of the new online banking
system. IMPLEMENT was a leading international
IT vendor for the financial
services sector and operated and maintained
BANK’s old online banking system.
During the reengineering project,
IMPLEMENT had the responsibility to
create the functional and technical design
documents and implement the new
online banking system. To provide these
services at the required quality and cost
levels, IMPLEMENT chose a global delivery
model that involved four locations
in Spain and a captive center in Brazil.
Another vendor, SCREEN, specialized in
web development and was based in Germany.
It was responsible for the frontend
screen design. Finally, a large IT vendor
with international operations, TEST
was responsible for the software testing,
conducted in a testing facility in India.
Thus, the project featured both nearshore
and offshore outsourcing.
Thus BANK’s sourcing strategy for this
project featured not only a global context
but also the involvement of multiple
vendors. In addition to national cultural
differences resulting from the geographically
distributed setting, the various organizational
cultures played major roles
in the project. These national and organizational
cultural differences became
BISE – RESEARCH PAPER
especially visible in work-related values
and practices, as the following examples
indicate.
In particular, BANK’s project manager
recognized the differences on the
national level:
There are also cultural aspects influencing
the cooperation. Due to
thetimepressure,werealizedeven
more that our Spanish colleagues
have another understanding of milestones
and time planning. During
the work day, they sometimes spend
two hours for a coffee break; this is
very different from Germany. With
detailed progress tracking, we managed
to at least communicate our
expectations regarding timing and
quality very clearly. This helped a
lot.
His colleague, a sub-project manager,
also cited differences that arose when
working simultaneously with project
teams from Spain and India:
When it comes to communication, it
is important to differentiate whether
you are talking with a Spanish colleague
or with somebody from India.
When talking to an Indian colleague,
you have to specify 100 %
what you expect, then you also get
100 %. When talking to a Spanish
colleague, you have to specify what
you do not want to get what you
want.
Beyond national cultural differences, organizational
cultural differences became
visible in the newly established intervendor
cooperation. All the vendors previously
had been involved in projects
with BANK, but they had not interacted
with one another in a multivendor setting
before. Against this background, different
work practices came to the surface
and challenged the cooperation. For
example, ARCHITECT was represented
by a team of five experienced experts
who followed an onshore delivery model.
IMPLEMENT’s 50-person implementation
team spread across five geographic
locations and possessed various competences
and practice levels. ARCHITECT
thus initially had trouble understanding
the challenges that faced IMPLEMENT’s
large, distributed project team, including
the need to scale the work to a
distributed team of developers in both
Brazil and Spain. A project manager from
ARCHITECT explained:

BISE – RESEARCH PAPER
Table 2 Demographics of
interview partners
Table 3 Distribution of interviews across organizations
Organization Overall work experience
(average)
BANK (9 informants) 15 years 7 years
ARCHITECT (2 informants) 6 years 6 years
IMPLEMENT (9 informants) 14 years 9 years
SCREEN (2 informants) 7 years 5 years
Organization Organizational level/role in the project Number of people
interviewed
BANK (Client) Top Management 2 2
Project & Sub-Project Management 5 6
Project Team 2 2
ARCHITECT (Vendor) Top Management 1 1
Project & Sub-Project Management 1 1
Project Team 0 0
IMPLEMENT (Vendor) Top Management 3 4
Project & Sub-Project Management 5 6
Project Team 1 1
SCREEN (Vendor) Top Management 1 1
Project & Sub-Project Management 1 1
Project Team 0 0
Total 22 25
Software architecture has a quite
comprehensive character; if you
want to understand, you need to see
the overall picture, whereas IMPLE-
MENT follows a “split the task and
distribute the sub-tasks to the developers”
mode. The developers never
see the big picture.
IMPLEMENT’s managers confirmed this
discrepancy in work practices and explained
that their global delivery model
required them to look at the system at
large first, and then split up the task into
preferably self-contained sub-tasks. Next
they could match the sub-tasks with the
competence profiles of different developers
to distribute the work effectively.
Therefore, the distributed teams worked
mostly independently – as was crucial for
the geographically distributed setup.
3.2 Data Collection
The data collection phase comprised
both interviews (primary data) and documents
generated during the course
of the project (secondary data). Thus,
we supplemented and triangulated our
interview data with project presentations,
tracking sheets, status reports, and
lessons-learned documents to create a
rich data set. In the course of the primary
data collection (taking place in July
and August 2009 as well as in November
and early December 2009), we conducted
25 interviews with 22 respondents
at both client and vendor locations.
We present the demographic information
about these interview partners in
Table 2.
Except for one conference call with a
respondent located in Brazil, all the interviews
took place in person: 15 in Germany
and 9 in Spain. Each interview
lasted between one and two hours, such
that we obtained more than 38 hours
of interviews. Because BANK’s corporate
policy did not permit recordings of any
interviews (with either BANK’s or the
vendors’ employees), we took extensive
notes, ultimately producing more than
130 pages of write-up notes.
Time spent working
for the company (average)
Number of interviews
conducted
In terms of the distribution of interviews
across organizations, differences
resulted from different team sizes for the
parties involved. For example, ARCHI-
TECT’s project team included five experienced
experts who were located onshore,
whereas IMPLEMENT’s team at
times consisted of more than 50 people
with various competencies and practice
levels who worked in geographically distributed
areas, including both near- and
offshore locations. Table 3 clarifies the
distribution of interviewees.
The interview partners represented different
organizational levels in BANK and
three of the four vendor companies, with
various profiles and roles that implied
different skill and knowledge levels. We
unfortunately did not have an opportunity
to interview any representatives from
TEST, despite repeated inquiries. The diversity
of respondents ensures that our
study features various organizational perspectives,
hierarchical perspectives, and
professional perspectives. The interviews
were conversational in nature, conducted
Business & Information Systems Engineering

using an interview guideline with semistructured
questions. We transcribed our
field notes after each interview session
and used these notes to identify appropriate
questions for subsequent interviews.
Thus, we refined the interview
questions multiple times during the
course of the data collection and analysis,
especially when we realized needs
for additional information to confirm
emerging themes or substantiate initial
findings.
3.3 Data Analysis
The data analysis began with coding of
the interview write-up notes. We identified,
named, and categorized phenomena
related to our research question,
through comparisons of the interviews
with one another and the available secondary
data. The preliminary codes included
concepts such as “project setup”
or “initialization phase.” After we identified
main conceptual themes from the interview
data, according to their high frequency,
we again compared all the interview
data with the conceptual themes to
find additional quotes or parallel statements
from other interviewees. Thus, we
substantiated our findings and managed
to identify the main management mechanisms
and learning issues from the primary
data. Near the end of the analysis
process, we compared our findings with
the available literature, to conceptualize
the emerging themes in our data. During
this comparison, we elaborated on the
identified themes by developing the descriptive
categories into more meaningful
notions at a higher level of abstraction.
Thus the research team engaged in
intensive considerations throughout the
processtoensurenoexistingtheorywas
forced onto the data. For the coding and
conceptualization process, we used AT-
LAS.ti. Table 4 illustrates the chronology
of the overall research process.
4Findings
From this exploratory case study, two key
findings emerged. The first finding pertains
the interplay of formal and informal
controls and interorganizational learning
in global, multisource, ISD outsourcing
projects (Sect. 4.1), while the second
finding relates to the mitigating effect of
these mechanisms on cultural differences
in such projects (Sect. 4.2). In the following,
we explain how these findings arose,
Business & Information Systems Engineering
describe them, and support the findings
with illustrative empirical quotes from
the case study interviews.
4.1 Finding 1: Interplay of Formal and
Informal Controls and
Interorganizational Learning
We start the description of this finding
with the influence of formal controls on
informal control and interorganizational
learning: our data suggest that implementing
formal control mechanisms encourages
the emergence of both informal
control mechanisms and interorganizational
learning processes because the formal
controls create transparency about
the essential project parameters.
In the focal project for example, the
reengineering had tight deadlines because
the software support for the old
system’s underlying technology was set
to expire. Therefore, tight project management
was of particular importance,
and BANK used an in-depth work breakdown
structure, or traceability matrix, as
a formal control mechanism. The traceability
matrix originally came from the
project plan, which was compiled by all
involved parties. To cope with emerging
differences in quality perceptions and accuracy,
the parties enhanced the work
breakdown structure to reflect work tasks
with a very high level of detail. According
to project manager,
Our clear focus was on tight control
of the project’s overall progress to ensure
that all involved parties were
on track, working toward a successful
overall service delivery. Therefore,
we used a very detailed progress
sheet that showed for each single
component when it had to be created,
by whom, and according to
which quality measures. Moreover,
the sheet enabled the project team to
identify and handle a series of delays
resulting from significant task dependencies
due to the involvement of
multiple vendors. As a consequence,
we had to track more than 1000
milestones and interfaces.
This traceability matrix fulfilled its function
as a formal control mechanism because
it helped keep track of the overall
project progress and quickly identified
plan variances and the need for countermeasures.
Moreover, it influenced the
emergence of informal control mechanisms
and interorganizational learning.
Specifically, by using the in-depth
BISE – RESEARCH PAPER
traceability matrix, the involved parties
gained detailed insights into their different
areas of responsibility, the associated
work portfolios, and the resulting
tasks for their sub-teams; they
also recognized cross-organizational interdependencies.
The interviewees indicated
that such transparency significantly
contributed to the lack of conflict
or even bargaining about responsibilities
and task assignments. Rather,
the project team could direct its primary
focus toward establishing a multisource
cooperation, in terms of both growing
as a team and establishing an integrated
service delivery process. A project manager
from one of the vendor companies
explained the absence of bargaining as
follows:
It is essential to fully understand the
project’s objectives and its planning
and to always give them top priority.
This may also mean that we as a
vendor have to concede at a certain
point. But because we are concentrated
on the benefit of the overall
project goals, we accept this without
discussion.
Although BANK and the vendor companies
had never interacted in a multisourcing
setting before, they developed
a joint mindset with shared norms and
values, including a cross-organizational
team spirit and absolute goal orientation.
Thus, the formal control created
room for informal controls to
emerge.
The emergence of the joint mindset
(i.e., an informal control mechanism) is
perhaps clearest in IMPLEMENT’s altered
attitude. When the project started,
IMPLEMENT expressed a different selfconcept
than the other vendors because
of its long and intense prior cooperation
with BANK. This vendor was accustomed
to a great deal of autonomy in its implementation
activities, as well as limited
control by BANK, the client.
In this focal project though, the multisourcing
constellation moved certain
tasks that IMPLEMENT had previously
performed, such as the architectural
framework and software testing, to other
vendors. Furthermore, IMPLEMENT’s
design and implementation activities no
longer were controlled by BANK but
instead by another, partly competing
vendor (ARCHITECT). Thus, IMPLE-
MENT’s project team needed to undergo
a change in mindset about not only

BISE – RESEARCH PAPER
Table 4 Chronology of research process
No Process step Main features Additional description Result
1 Literature
analysis
2 Identification
of research
case
3 Data
collection
4 Creation of
interview
notes
• Identification of relevant literature
streams
• In-depth analysis of literature within
these streams
• Definition of central requirements
for the phenomenon
• Selection of suitable case, in
cooperation with industry partner
• Identification of interviewees
• Interviews
• Collection of secondary data
• Creation of clean copies
• Complement interview notes with
recalled details
• Addition of comments
5 Data analysis • Open coding
• Grouping codes into categories and
identifying major conceptual themes
• Refinement of concepts
6 Validation of
findings
• Discussion of researchers’
interpretations with selected interview
partners
the content (i.e., learning the new software
architecture) but also their company’s
role and responsibilities in the
multiparty cooperation. IMPLEMENT’s
project manager described the change,
and the team’s goal orientation, as follows:
For sure, we would prefer to have
the responsibility for the framework,
as it is part of our portfolio, but the
client has decided differently..., so
now we are jointly responsible for
this project. As a consequence, we
have to prioritize the project goals
more than our goals as a service
delivery company to create a winwin
situation. This regularly also involves
overlooking the respective organizational
affiliation and thinking
of an integrated project team
with a joint goal.
• Literature streams covered:
–GlobalISsourcing
– IS project management
• Initial interviewees suggested by senior
management; further interviewees
identified by the initial interviewees
• 25 in-depth interviews in Germany and
Spain
Intensivenotetaking
• Access to documents generated during
thecourseoftheproject
• Descriptions of the atmosphere during
the interview and emotions of interview
partners
• Reading of transcripts and documents
and highlighting of descriptions
associated with the research question
• Central criterion: frequency of
mentions
• Repeated comparisons of concepts with
interview data
• In the late analysis process, comparison
of concepts with relevant literature
• Presentation of major findings and
assessments of robustness, according to
interviewees
This adoption of a corporate projectoriented
mindset appeared among the
other vendors too, as the following
quote from ARCHITECT’s senior manager
confirmed:
In my opinion, the core team did a
very good job concerning the multiparty
cooperation. I sensed a broad
willingness to act informally instead
of insisting upon contractually
agreed details and defined responsibility
areas. In my experience,
this is the fundamental ingredient
that in the end makes projects
successful.
With regard to the effect of formal controls
on interorganizational learning, we
turn to an example from the ramp-up
phase, when the vendors practiced and
harmonized their cooperation and seized
on opportunities to adjust their mutual
• Motivation
• Research gap
• Specific research question
• Theoretical background
• Requirements list
• Specific unit of analysis
for study
• In total: 22 interviewees
• In total: Approximately 38
hours of interviews
• Project tracking sheets,
project presentations, status
reports, lessons learned
• In total: Approximately
130 pages of interview notes
• Initial code list
• Initial concepts
• Final concepts
expectations and needs. By practicing
and reinforcing the allocation of tasks
and required interactions during a fixed
period at the beginning of the project,
BANK ensured that the involved parties
understood and accepted their own
roles and responsibilities, as well as those
of the other parties. All the involved
parties perceived this effort as valuable.
For example, IMPLEMENT’s project lead
explained:
We were very interested in getting
this project up and running fast
in order to prove our capabilities
and establish a trust-based relationship
with the other vendors. [...]
Our objective was to first understand
their organizational cultures
and then partly adapt ourselves to
them in order to foster a smooth
integrated service delivery.
Business & Information Systems Engineering

In global, multisource, ISD outsourcing
projects, the responsibility for service delivery
is distributed, so cooperation between
vendors must be established to
ensure that individual service deliveries
from the different areas of responsibility
intertwine and lead to a successful overall
service delivery. A member of BANK’s
project team summarized this essential
phase:
In the course of this phase, we practiced
and evaluated, based on selected
business transactions, how the
overall service delivery had been set
up and how the performance was
in terms of process and outcome
quality.
The initialization of the multiparty cooperation
also revealed performance deficits
that could be attributed to national
and organizational cultural differences.
Such differences became manifest as divergent
working modes, including values
(e.g., sense of quality) and practices
(e.g., knowledge transfer approaches).
The project team actively addressed any
conflicts or errors and used them as
learning tools to improve subsequent
interactions.
We now describe the influence of informal
controls and interorganizational
learning on formal controls. It emerged
from our data that in fact this influence
is two-part. In the short-term, both
informal control and learning mechanisms
generated valuable operational information
that enabled the parties to adjust
their formal control mechanisms on
a granular level. Then in the mid- to
long-term, the informal control mechanisms
and interorganizational learning
processes contributed to lessen the need
for formal controls. That is, our data
indicated that the impact changes over
time. In the following, we provide two
examples to describe this effect.
During the ramp-up phase, as we
noted previously, the vendors practiced
and adjusted their cooperation to ensure
a smoothly integrated service delivery.
The parties therefore identified functional,
process, and technological issues
and noted performance deficits, which
enabled them to deduce important operational
information and then sharpen
their project tracking (i.e., short-term effect
on formal control). A project team
member explained:
From my point of view, the rampup
phase at the beginning of the
Business & Information Systems Engineering
project was very important and reasonable.
As we had to deliver very
early, we had to deal with problems
very early as well. As a consequence,
the impact of these problems
could be minimized through
rescheduling, and further mitigation
measures could be initiated.
At the beginning of the project, BANK
strongly encouraged cooperation among
the vendors to reinforce their roles and
responsibilities and foster a stable working
mode. However, this coordination
and control effort gradually decreased,
replaced by self-organizing mechanisms
within the increasingly well-established
multivendor cooperation (i.e., mid-tolong
term effect on formal control).
Thus the reduction of formal management
effort resulted from interorganizational
learning processes, as described by
a project manager from IMPLEMENT:
At the beginning, [BANK] arranged
formal meetings managed by
BANK, but later, there was more
and more direct interaction between
the vendor companies. BANK was
not the driver but was always informed
to sustain transparency.
4.2 Finding 2: Mitigating Effects on
Cultural Differences
We now describe the mitigating effect
of the interplay between control mechanisms
and interorganizational learning
on cultural differences in global, multisource,
ISD outsourcing projects. It
emerged from our data that these mechanisms
can help partners overcome national
and organizational cultural differences
by harmonizing their varied workrelated
values and practices. That is, our
data indicated that the integrated use
of formal and informal control and interorganizational
learning dominate or
overrule patterns of behavior which are
rooted in national or organizational cultural
differences; thus, differences in national
and organizational cultures become
less salient and the occurrence of
culture-induced conflict declines. In the
following, we describe this finding in
more detail.
In the focal project, the interplay of formal
and informal controls, together with
interorganizational learning, helped mitigate
the risks associated with national
and organizational cultural differences.
By creating transparency about each individual
deliverable (i.e., through the use
BISE – RESEARCH PAPER
of the formal control mechanism “traceability
matrix”), the parties assimilated
their divergent assessments of quality and
accuracy, at least to some extent. One
project manager for the vendor stated:
The traceability matrix helped to
mitigate the risks caused by cultural
differences by defining clear roles
and responsibilities, supporting the
identification of interdependencies,
and specifying our joint deliverables.
In the following illustrative quote, another
Spanish project team member explained
the above-mentioned cultural
differences between Germany and Spain
from a Spanish perspective:
When problems occur, we would expect
the customer to be near you,
helping you, offering his support to
jointly solve the problem regardless
of the timeline.
However, in the course of the cooperation,
it turned out that
When you deliver to Germans, you
have to deliver absolutely on time
and bug-free. The quality expectations
are high. They [German colleagues
from BANK and other vendors]
mainly insist on on-time delivery
with defined quality; they stick
more to their plan. Thus, it was very
helpful that there was this detailed
tracking tool, as we could see the status
and our forthcoming tasks at any
time.
In parallel, divergent norms and values
were being renegotiated and consolidated
through the use of informal
control and interorganizational learning
processes. For example, while the
development of a corporate, projectoriented
mindset (i.e., informal management
mechanism) helped mitigate organizational
cultural differences by establishing
a project culture, driven by team
spirit and goal orientation, instances of
interorganizational learning enabled the
project team to cope with emerging differences
in work-related practices. As one
of our senior-level interview partners
commented:
From my perspective, the main reason
for the success of this project
is the fact that the project members
from our organization and the
involved vendor organizations were
always willing to pursue the goals
of the project in a very collaborative
way.

BISE – RESEARCH PAPER
Table 5 Elements of conceptual framework
Underlying reasons Methods used Main results achieved
• Multi-party cooperation
(from both, a national and
organizational perspective)
• Differences in work
practices
• Multi-party cooperation
(from both, a national and
organizational perspective)
• Differences in work-related
values
• Multi-party cooperation
(from both, a national and
organizational perspective)
• Differences in work
practices and values
Categories Concepts
• Formal
control
• Informal
control
• Interorganizational
learning
A project manager from the client organization
explained in greater detail how
this worked:
We talked about different perceptions,
divergent understandings, issues
resulting from cultural differences
and different communication
styles, and reflected jointly on the
positive and negative episodes of
the cooperation between the vendors
within our project, constantly
aiming at further improving this
cooperation.
In summary, the interplay of formal and
informal control and interorganizational
learning enabled the project team to
harmonize their divergent work-related
values, norms, practices, and expectations,
which resulted from national and
organizational cultural differences.
The findings described above are summarized
in Table 5.
5 Conclusion and Research
Implications
With the goal of increasing our understanding
of ways to mitigate cultural
differences in global, multisource, ISD
outsourcing projects, we apply an exploratory
single-case study design. In
turn, we can detail how formal and informal
control mechanisms and interorganizational
learning interact; furthermore,
our data show that these interactions
help mitigating cultural differences
• Traceability matrix (used as example in text)
• Joint, template-based status reporting (not
described in text)
CMMI review process (not described in text)
• Development of joint, goal-oriented
mindset (used as example in text)
• Socialization activities, e.g., joint dinner
(not described in text)
• Stimulation of teammate interaction (not
described in text)
• Initialization phase (used as example in
text)
• Use of joint structurization tools such as
issue log, wiki, etc. (not described in text)
in such projects. To achieve this benefit
though, the implementation of formal
controls appears essential because it prepares
the project for the emergence of informal
controls and interorganizational
learning. Project partners also should anticipate
changes in the effects of informal
control and interorganizational learning
over time, shifting from feedback and
information that support the design of
formal controls (short term) to an actually
reduced need for and use of formal
controls (mid- to long term). Together,
these mechanisms can help partners
overcome national and organizational
cultural differences by harmonizing
their varied work-related values and
practices.
With these findings, this study contributes
to global IS outsourcing literature
and provides a clearer understanding
of ways to deal with cultural differences
in global, multisource, ISD outsourcing
projects. Furthermore, we contribute
to research into control dynamics
in global IS projects: this study illustrates
that changes in control modes across
project phases can be triggered by external
factors (e.g., project context, stakeholder
context, and global context as revealed
by Kirsch 2004), but also through
the interplay of control modes within a
single project’s control portfolio. Finally,
the detailed case analysis offers implications
for global sourcing practices. The
potentialtoreduceformalcontrolsoffers
an important benefit for project managers
of global, multisource, ISD out-
• Transparency regarding roles and
responsibilities of each party
• Absence of negotiations and bargaining
• Growing as a team
• Development of joint project culture
• Establishing an integrated service
delivery process
sourcing projects, as well as a likely reduction
of the high management overhead
costs normally associated with such
projects.
However, we also note several limitations.
First, we did not have the opportunity
to interview any representatives from
one of the vendors, TEST, which was responsible
for the software testing in a
new factory in India. Second, the rampupphaseforthetestfactorydidnotwork
out as planned, causing major problems
for the project and its multiple vendors
during the testing phase. In this case, the
parties seemingly should have suffered
great conflict, resulting from cultural differences.
Surprisingly though, the cooperation
was characterized by harmony,
perhaps because each of the vendors had
a prior business relationship with BANK.
Further research should examine multisource,
ISD outsourcing relationships
in the context of newly composed vendor
portfolios to examine this proposed
explanation. Third, our results are specific
to large, technology reengineering
projects in the German financial services
industry. Accordingly, we encourage researcherstocontinuetostudytheposed
research question in other contexts and
settings. Fourth, our analysis of the interplay
of formal and informal management
mechanisms and learning in global, multisource,
ISD outsourcing projects could
be extended to identify further aspects
and reveal an increasingly differentiated
picture of the interaction.
Business & Information Systems Engineering

References
Adler PS (2001) Market, hierarchy, and trust:
the knowledge economy and the future
of capitalism. Organization Science
12(2):215–234
Ang S, Inkpen AC (2008) Cultural intelligence
and offshore outsourcing success: a framework
of firm-level intercultural capability.
Decision Sciences 39(3):337–358
Bradach JL, Eccles RG (1989) Price, authority,
and trust: from ideal types to plural forms.
Annual Review of Sociology 15(1):97–118
Cannon J, Achrol R, Gundlach G (2000) Contracts,
norms, and plural form governance.
Journal of the Academy of Marketing Science
28(2):180–194
Carmel E, Agarwal R (2001) Tactical approaches
for alleviating distance in global
software development. IEEE Software
18(2):22–29
Choudhury V, Sabherwal R (2003) Portfolios
of control in outsourced software development
projects. Information Systems
Research 14(3):291–314
Das TK, Teng B-S (1998) Between trust and
control: developing confidence in partner
cooperation in alliances. Academy of
Management Review 23(3):491–512
Das TK, Teng B-S (2001) Trust, control, and
risk in strategic alliances: an integrated
framework. Organization Studies 22(2):251
David CG, Chand D, Newell S, Resende-Santos
J (2008) Integrated collaboration across distributed
sites: the perils of process and the
promise of practice. Journal of Information
Technology 23(1):44–54
Dibbern J, Heinzl A, Leibbrandt S (2003) Interpretation
des Sourcings der Informationsverarbeitung.
Hintergründe und Grenzen
ökonomischer Einflussgrößen. WIRT-
SCHAFTSINFORMATIK 45(5):533–540
Dibbern J, Winkler J, Heinzl A (2008) Explaining
variations in client extra costs between
software projects offshored to India. MIS
Quarterly 32(2):333–366
Gopal A, Sivaramakrishnan K (2008) On vendor
preferences for contract types in offshoresoftwareprojects:thecaseoffixed
price vs time and materials contracts. Information
Systems Research 19(2):202–220
Gregory R (2010a) The plural governance perspective:
opportunities for is outsourcing
research. In: Proc 18th European conference
on information systems (ECIS 2010),
Pretoria
Gregory R (2010b) Review of the IS offshoring
literature: the role of cross-cultural differences
and management practices. In: Proc
18th European conference on information
systems (ECIS 2010), Pretoria
Gupta U, Raval V (1999) Critical success factors
for anchoring offshore projects. Information
Strategy 15(2):21–27
Hildenbrand T, Rothlauf F, Heinzl A (2007)
Ansätze zur kollaborativen Softwareerstellung.
WIRTSCHAFTSINFORMATIK
49(Sonderheft):72–80
Holmström Olsson H, Ó Conchúir E, Ågerfalk
PJ, Fitzgerald B (2008) Two-stage offshoring:
an investigation of the Irish bridge.
MIS Quarterly 32(2):257–279
Iacovou CL, Nakatsu R (2008) A risk profile
of offshore-outsourced development
projects. Communications of the ACM
51(6):89–94
Inkpen AC, Currall SC (2004) The coevolution
of trust, control, and learning in joint ventures.
Organization Science 15(5):586–599
Kirsch LJ (1996) The management of complex
tasks in organizations: controlling the sys-
Business & Information Systems Engineering
tems development process. Organization
Science 7(1):1–21
Kirsch LJ (2004) Deploying common systems
globally: the dynamics of control. Information
Systems Research 15(4):374–395
Kliem R (2004) Managing the risks of offshore
it development projects. Information
Systems Management 21(3):22–27
Kotlarsky J, Oshri I (2005) Social ties, knowledge
sharing and successful collaboration
in globally distributed system development
projects. European Journal of Information
Systems 14(1):37–48
Kotlarsky J, van Fenema PC, Willcocks LP
(2008) Developing a knowledge-based
perspective on coordination: the case of
global software projects. Information &
Management 45(2):96–108
Krishna S, Sahay S, Walsham G (2004) Managing
cross-cultural issues in global software
outsourcing. Communications of the ACM
47(4):62–66
Lacity MC, Willcocks LP (1998) An empirical
investigation of information technology
sourcing practices: lessons from experience.
MIS Quarterly 22(3):363–408
Lacity M, Khan SA, Willcocks L (2009) A review
of the IT outsourcing literature: insights for
practice. Journal of Strategic Information
Systems 18(3):30–146
Leonardi PM, Bailey DE (2008) Transformational
technologies and the creation of
new work practices: making implicit knowledge
explicit in task-based offshoring. MIS
Quarterly 32(2):411–436
Levina N, Vaast E (2008) Innovating or doing
as told? Status differences and overlapping
boundaries in offshore collaboration. MIS
Quarterly 32(2):307–332
Meyer T, Stobbe A (2007) Welche “Standorte”
wählen deutsche Unternehmen? WIRT-
SCHAFTSINFORMATIK 49(Sonderheft):81–
89
Nicholson B, Sahay S (2001) Some political
and cultural issues in the globalisation
of software development: case experience
from Britain and India. Information
and Organization 11(1):25–43
Nicholson B, Sahay S (2004) Embedded
knowledge and offshore software development.
Information and Organization
14(4):329–365
Oshri I, Kotlarsky J, Willcocks L (2007) Managing
dispersed expertise in it offshore
outsourcing: lessons from Tata consultancy
services. MIS Quarterly Executive 6(2):53–
65
Rottman JW (2008) Successful knowledge
transfer within offshore supplier networks:
a case study exploring social capital in
strategic alliances. Journal of Information
Technology 23(1):31–43
Stebbins RA (2001) Exploratory research in
the social sciences. Sage, Thousand Oaks
Tiwana A (2008) Does technological modularity
substitute for control? A study
of alliance performance in software outsourcing.
Strategic Management Journal
29(7):769–780
Vlaar PWL, van Fenema PC, Tiwari V (2008)
Cocreating understanding and value in distributed
work: how members of onsite
and offshore vendor teams give, make,
demand, and break sense. MIS Quarterly
32(2):227–255
Vogt K, Gregory R, Beck R (2009) Measuring
client-vendor distance in global
outsourcing relationships: a conceptual
model.In:Proc9.InternationaleTagung
Wirtschaftsinformatik, Vienna
Abstract
Roman Beck, Katharina Schott
BISE – RESEARCH PAPER
The Interplay of Project Control
and Interorganizational
Learning: Mitigating Effects
on Cultural Differences
in Global, Multisource ISD
Outsourcing Projects
Research into global, multisource, information
systems development outsourcing
projects has uncovered management
challenges, including cultural
differences on multiple levels. While
control mechanisms and interorganizational
learning have been shown to
contribute to the mitigation of cultural
differences in such projects, a gap persists
regarding the effect of the interplay
between these mechanisms. This
study employs an exploratory singlecase
study design to analyze how formal
and informal control mechanisms
and interorganizational learning interact
and thus contribute to the mitigation
of cultural differences in global,
multisource, information systems development
outsourcing projects. With
the key finding that the influence of informal
controls and interorganizational
learning on formal controls changes
over time, this research helps expand
the domain of control dynamics in global
IS projects. This study also contributes
to literature on ways to handle cultural
differences in global, multisource, IS
outsourcing projects.
Keywords: Cultural differences, Formal
control, Informal control, Interorganizational
learning, IS outsourcing,
Multisourcing, Global information systems
development

BISE – RESEARCH PAPER
Walsham G (2002) Cross-cultural software
production and use: a structurational analysis.
MIS Quarterly 26(4):359–380
Willcocks LP, Lacity MC, Cullen S (2007) Information
technology sourcing research:
critique, lessons and prospects. In: Proc
13th Americas conference on information
systems, Keystone
Winkler J, Dibbern J, Heinzl A (2007) Der
Einfluss kultureller Unterschiede beim
IT-Offshoring. WIRTSCHAFTSINFORMATIK
49(2):95–103
Yin R (2003) Case study research – design and
methods. Sage, Thousand Oaks
Business & Information Systems Engineering

Maximizing Cloud Provider Profit from Equilibrium Price Auctions [PRE-PRINT]
Ulrich Lampe, Melanie Siebenhaar, Apostolos Papageorgiou, Dieter Schuller, Ralf Steinmetz
Multimedia Communications Lab (KOM)
Technische Universität Darmstadt
Darmstadt, Germany
Email: ulrich.lampe@KOM.tu-darmstadt.de
Abstract—Auctioning constitutes a market-driven scheme
for the allocation of cloud-based computing capacities. It is
practically applied today in the context of Infrastructure as a
Service offers, specifically, virtual machines. However, the
maximization of auction profits poses a challenging task for
the cloud provider, because it involves the concurrent
determination of equilibrium prices and distribution of
virtual machine instances to the underlying physical hosts in
the data center. In the work at hand, we propose an optimal
approach, based on linear programming, as well as a
heuristic approach to tackle this Equilibrium Price Auction
Allocation Problem (EPAAP). Through an evaluation based on
realistic data, we show the practical applicability and benefits
of our contributions. Specifically, we find that the heuristic
approach reduces the average computation time to solve an
EPAAP by more than 99.9%, but still maintains a favorable
average solution quality of 96.7% in terms of cloud provider
profit, compared to the optimal approach.
Keywords-Cloud Computing; IaaS; Equilibrium; Auction;
Allocation; Optimization; Optimal; Heuristic
A. Motivation
I. INTRODUCTION
In the past few years, cloud computing has gained
tremendous interest both among practitioners and
researchers. One of the essential ideas of this novel
paradigm consists in the provision of Information
Technology (IT) in a utility-like manner [1]. In the context
of an envisioned cloud computing market, the decision
whether to supply IT capacities in-house or lease them from
the cloud is – aside from strategical considerations – largely
determined by the price [2]. One potential instrument to
achieve a market-based pricing and thus, efficient allocation
of computing capacities, are auctions [3]. In this respect,
Amazon Web Services 1 has not only been one of the
pioneers in the cloud computing domain, but also among
the first to employ auctions as a complement to traditional
fixed-price schemes.
Specifically, in the Spot Instances system 2 , cloud users
submit bids to Amazon for Infrastructure as a Service (IaaS)
offers in the form of Virtual Machines (VMs). Each bid
states the desired number of instances and the maximum
willingness to pay for a specific VM type. At a periodic
1 http://aws.amazon.com/
2 http://aws.amazon.com/ec2/spot-instances/
time interval, Amazon determines an equilibrium price for
each type. Users whose bids exceed (or meet) the
equilibrium price are (partially) served with the desired
instances. All users pay the identical equilibrium price –
rather than their respective bid price – per instance.
While many researchers have previously assumed that the
price-setting mechanism is market-driven, i. e., determined
by supply and demand [3]–[5], more recent research
indicates that this is not the case. According to Agmon
Ben-Yehuda et al. [6], the publicly posted equilibrium
prices are likely selected at random from a small predefined
interval. In the opinion of the aforementioned authors, this
indicates a low demand in the Spot Instance system at
present.
However, it is safe to assume that with increasing
acceptance and utilization of cloud computing offers,
auctions will gain in popularity for the efficient allocation
of computing capacities. Based on this notion, in the work
at hand we examine how cloud providers can maximize
their profit using equilibrium price auctions.
B. Research Problem
When applying equilibrium price auctions for the allocation
of capacities, a cloud provider faces two distinct yet linked
challenges. First, the provider has to determine specific
equilibrium prices for each offered VM type. Second, the
provider has to distribute the VM instances, which have
been requested in the served (or satisfied) bids, among the
Physical Machines (PM) instances in her/his data center. In
this process, the cloud provider will commonly pursue the
objective of maximizing profit. This profit is given by the
difference between the revenue from the served bids and the
operating costs of the PM instances. In the remainder of
this paper, the combination of both challenges is referred to
as Equilibrium Price Auction Allocation Problem (EPAAP).
We assume that the cloud provider operates a limited
number of PM types, which provide restricted resource
supplies. These supplies are specified in terms of certain
resource types, e. g., processor power or memory. Of each
PM type, a restricted number of instances are available in
the cloud provider’s data center(s). These PM instances may
be independently powered on or off. Each active instance

leads to a fixed operating cost due to, e. g., idle power
consumption and maintenance demands.
We further assume that the cloud provider offers a
predefined set of VM types, which exhibit certain resource
demands. Due to these resource demands, each individual
VM instance imposes additional variable operating costs on
the PM instance that hosts it, due to, e. g., increased power
consumption.
Lastly, we assume that in accordance with the Spot
Instances system, the cloud users may continuously submit
or cancel bids. In contrast, the allocation process (i. e., the
pricing of VM types and distribution of VM instances) is
only periodically conducted by the cloud provider.
The remainder of this paper is structured as follows: In
Section II, we introduce a formal notation for the EPAAP
and subsequently describe two optimization approaches, an
optimal and a heuristic approach. Section III describes our
evaluation procedure and the obtained results. In the
subsequent Section IV, we provide an overview of related
work. Lastly, Section V concludes the paper with a
summary and outlook on future work.
A. Formal Notations
II. OPTIMIZATION APPROACHES
As a basis for the optimization approaches that are presented
in the following subsections, we introduce a formal notation
for the EPAAP. First, we define the basic entities:
• V ⊂ N: Set of offered VM types.
• P ⊂ N: Set of available PM types.
• R ⊂ N: Set of regarded resource types.
Based on the previous definitions, the characteristics of and
relations between the basic entities can be further specified:
• RDvr ∈ R + : Resource demand of VM type v ∈ V for
resource type r ∈ R.
• RSpr ∈ R + : Resource supply of PM type p ∈ P for
resource type r ∈ R.
• CFp ∈ R + : Fixed operating cost of PM type p ∈ P
per utilized instance.
• CVpv ∈ R + : Variable operating cost of PM type
p ∈ P per hosted instance of VM type v ∈ V .
• np ∈ N: Available instances of PM type p ∈ P in the
cloud provider’s data center(s).
The bids that have been submitted by the cloud users are
formalized using the following constructs:
• B ⊂ N: Set of submitted bids.
• Wb ∈ R + : Specified price for bid b ∈ B, i. e.,
maximum willingness to pay.
• Tb ∈ V : Specified VM type for bid b ∈ B.
For reasons of simplicity and without loss of generality, we
assume that each bid b ∈ B specifies an individual request
for one VM instance. Thus, if a user bids for η ∈ N VM
instances of the same type at an identical price – as it is the
x111
y11
Bid B1 = B 1 1
T1 = 1
W1 = 0.20
PM
Instance
P1, j = 1
y12
Bid B4 = B 1 2
T4 = 1
W4 = 0.15
PM
Instance
P1, j = 2
Bids for
VM Type 1 (B 1 )
PM Type 1 (P1)
y21
Bid B2 = B 2 1
T2 = 2
W2 = 0.25
PM
Instance
P2, j = 1
y22
Bid B3 = B 2 2
T3 = 2
W3 = 0.20
x4** x2** x3**
x122
x121
x*11
x112 x*12 x*21
PM
Instance
P2, j = 2
z11 z12 z21 z22
x*22
Bids for
VM Type 2 (B 2 )
Figure 1. Schematic overview of the optimization model, depicting the
decision variables (in bold) and most relevant entities. A portion of the
links has solely been sketched (in gray) to improve readability.
case in the Spot Instances system – this results in η
individual bids in the set B.
Lastly, for reasons of convenience, we infer the following
definitions from the previous specifications:
• B v ⊆ B: Set of submitted bids for VM type v ∈ V .
• W v ⊆ W : Set of prices for the bids in set B v .
• mv ∈ N: Overall number of submitted bids for VM
type v ∈ V .
Without loss of generality, we establish that the bids in the
set B are given in monotonically decreasing order of the
corresponding prices. That is, it holds that Wb ≥ Wb ′ for
b, b ′ ∈ B, b < b ′ . The same applies for the respective
subsets of bids, i. e., Bv .
B. Optimal Allocation Approach
To compute an optimal solution to the EPAAP, we transfer
the problem definition from Subsection I-B into a
mathematical optimization model. The result is given in
Model 1 and will be explained in detail in the following. In
order to promote an easier interpretation of the model, we
first discuss the significance of the decision variables.
In Equation 12, x, y, and z are defined as binary decision
variables. x is the primary decision variable, whereas y and
z can be considered auxiliary decision variables.
Specifically, xbpj indicates whether the bid b has been
served and assigned to a PM instance of type p with the
running index j or not. yvk indicates whether for a VM
type v, exactly the first k bids are served or not. Lastly, zpj
represents whether a PM instance of type p with the
running index j is utilized, i. e., powered on, or not. An
overview of the optimization model, which highlights the
relations between the decision variables and the most
important entities, is depicted in Figure 1.
Equation 1 specifies the objective of the optimization model,
namely the maximization of profit. The profit comprises
PM Type 2 (P2)

three components. The first component is the revenue that is
generated through the served bids. For that matter, yvk also
indicates whether the k-th bid for VM type v corresponds
to the equilibrium price or not. The second component is
the fixed operating cost of all utilized PM instances, while
the third component represents the additional variable
operating cost due to the hosted VM instances.
Equations 2 and 3 link the decision variables x and y and x
and z respectively. Equation 4 assures that the resource
demands of all VM instances are met by the resource
supplies of the respective PM instance that host them.
Equation 5 guarantees that an individual bid cannot be
served more than once. Equation 6 ensures that solely one
equilibrium price for each VM type can be set.
Equation 7, for performance reasons, restricts the solution
space by preferring PM instances of the same type with a
lower running index over those with a higher running index.
Equation 8, also for performance reasons, excludes
dominated solutions from the solution space. A solution is
referred to as dominated if another solution exists that
would yield a higher or equal revenue, even if a smaller
number of bids for a specific VM type is served.
Equations 9 through 11 define valid running indices for VM
and PM instances. The definition of the latter is based on
two observations: First, the number of utilized PM instances
cannot exceed the available number of instances of this
type, np. Second, in the theoretical case that all bids were
served and the requested VM instances were each assigned
to an individual PM instance of the same type, no more
than |B| instances would be required.
As can be seen, Model 1 constitutes a Linear Program (LP),
or more specifically, Binary Integer Program (BIP). This
class of optimization problems can be solved using
well-known methods from the field of Operations Research,
most notably, the Branch and Bound (B&B) algorithm [7].
While the B&B algorithm can be very efficient in some
cases, it is still based on the principle of enumeration, i. e.,
in the worst case, all potential solutions have to be
examined [8].
Specifically, for a BIP, the solution space grows
exponentially with the number of decision variables. As can
be observed from Model 1 (notably, Equations 1 and 12),
the number of decision variables increases quadratically
with the number of bids and linearly with the number of
PM types. Accordingly, the computational complexity of the
optimal allocation approach is exponential and corresponds
to O(2 |B|2 ∗|P | ).
C. Heuristic Allocation Approach
For real-life application scenarios involving thousands of
bids, the optimal allocation approach may be problematic
due to its exponential growth in computational complexity.
Thus, we have developed a heuristic approach that trades
Model 1 Optimal Allocation Model
Maximize P r(x, y, z) = �
− �
p∈P,j∈Jp
k ∗ yvk ≤
zpj ∗ CFpj −
�
1≤i≤k,p∈P,j∈Jp
v∈V,k∈Kv
�
b∈B,p∈P,j∈Jp
subject to
yvk ∗ k ∗ W v k
xbpj ∗ CVpTb
(1)
xB v i pj ∀v ∈ V, k ∈ Kv (2)
zpj ≥ xbpj ∀b ∈ B, p ∈ P, j ∈ Jp (3)
�
xbpj ∗ RDTbr ≤ RSpr ∀p ∈ P, j ∈ Jp, r ∈ R (4)
b∈B
Jp =
�
p∈P,j∈Jp
�
k∈Kv
xbpj ≤ 1 ∀b ∈ B (5)
yvk ≤ 1 ∀v ∈ V (6)
zpj ≥ zpj ′ ∀p ∈ P, j ∈ Jp, j ′ ∈ Jp, j < j ′
yvk = 0 if k ∗ W v k ≤ k ′ ∗ W v k ′
∀k ∈ Kv, k ′ ∈ Kv, k > k ′
Kv =
�
�
{1, ..., mv}
∅
if mv > 0
else
{1, ..., min(np, |B|)} if np > 0
∅ else
(7)
(8)
∀v ∈ V (9)
∀p ∈ P (10)
Jp, Kv ⊂ N (11)
xbpj ∈ {0, 1} ∀b ∈ B, p ∈ P, j ∈ Jp
yvk ∈ {0, 1} ∀v ∈ V, k ∈ Kv (12)
zpj ∈ {0, 1} ∀p ∈ P, j ∈ Jp
reductions in computation time against potentially
sub-optimal solutions.
The principle idea is to initially determine the equilibrium
prices for all VM types, such that the expected profit from
the served bids is maximized. Subsequently, these served
bids – or more specifically, the VM instances that have
been requested in these bids – are cost-efficiently distributed
across the physical hosts. Accordingly, the approach is split
into two phases, VM pricing and VM distribution.
The procedure for the second phase is inspired by a
heuristic for the distribution of software services across VM
instances, which we have introduced in our previous work

[9]. This heuristic, in turn, adapts concepts that are
frequently applied for solving the well-known Knapsack
problem [7]. For additional details, we refer to our previous
publication.
The pseudo code for the two phases is provided in Listing 1
and 2. In accordance with the previous subsection, xbpj
represents the main binary decision variable. Qv ∈ R +
denotes the equilibrium price for each VM type v. Lastly,
the set S ⊆ B contains the bids that will be served.
In the first phase, VM pricing, we initially estimate the cost
of serving an individual bid for each VM type v. This
serving cost CSvp of a VM type v on each PM type p
corresponds to the partial fixed and additional variable
operating cost of a respective PM instance. The partial fixed
operating cost, in this context, is the fixed operating cost of
a PM instance, multiplied by the ratio between the resource
demands of VM type v and the resource supplies of PM
type p. The individual serving costs are aggregated into an
weighted average serving cost CSv across all suitable PM
types for each VM type v, using the respective weights gvp.
Suitable, in this respect, means that a PM type can host a
VM type subject to the given resource constraints (lines
3-13). On the basis of these serving costs and the initial bid
prices, we determine a favorable number ˆ kv ∈ N of bids for
each VM type v that should be served. The number is
considered favorable if the expected profit Ev ∈ R, i. e., the
difference between the revenue from the bids and the
expected serving costs, becomes maximal. In the same step,
the equilibrium price Qv for each VM type v is inferred.
The served bids are stored in the set S, which, in
accordance with the set B, we assume to be ordered by
monotonically decreasing bid prices (lines 14-21). The
corresponding VM instances are considered for distribution
in the following phase.
In the second phase, VM distribution, we first initialize an
instance count jp ∈ N for each PM type p ∈ P (lines 1-3).
Subsequently, the following packing process is conducted:
We scan the list of PM types to determine a favorite ˆp ∈ P .
For this purpose, we initially create a packing list Lp ⊆ S
for each type p, unless the maximum number of instances
np of this type has already been reached. A packing list
constitutes a set of bids that could be hosted by a new
instance of the regarded PM type. For that matter, we scan
the list of served bids S. For each bid b, we check whether
it could – in addition to the current packing list Lp – be
hosted by a new instance of type p, based on the given
resource constraints. If the bid b meets this condition, it is
added to the packing list. If not, the associated VM type Tb
is added to an ignore list Ip ⊆ V , and bids of the identical
type are ignored during the remainder of the packing list
creation (lines 8-17). In the following, we compute the
utility Up of the current PM type p. Utility is defined as the
ratio between the revenue from the packing list Lp and the
resulting fixed and variable operating costs of a new PM
Listing 1 Algorithm for VM Pricing
1: S ← ∅
2: for all v ∈ V do
3: gv ← 0, CSv ← 0
4: for all p ∈ P do
5: if p.canHost(v) = true then
6: gvp ← min(np, |B|)
7:
8:
gv ← gv + gvp
CSvp ← CVpv + CFp ∗ 1
9:
�
|R| r∈R
CSv ← CSv + gvp ∗ CSvp
10: end if
11: end for
12: if gv > 0 then
13:
14:
CSv ← CSv/gv
kv
ˆ ← 0, Êv ← 0, Qv ← ∞
15:
16:
for k = 1 → mv do
Ev ← k ∗ (W v 17:
k − CSv)
if Ev > Êv then
18: Êv ← Ev, ˆ kv ← k, Qv ← W v 19:
k
end if
20: end for
21: S ← S ∪ {Bv 1 , ..., Bv 22:
kv ˆ }
end if
23: end for
� RDvr
RSpr
instance. If a PM type exhibits higher utility than the
current favorite ˆp, it is stored as new favorite (lines 18-21).
After all PM types have been scanned and a favorite has
been identified, all bids in the packing list Lˆp are assigned
to a new instance of type ˆp with the previously incremented
index jˆp. The bids are also removed from the set S (lines
26-29). The packing process is repeated until all served bids
have been assigned or no suitable (i. e., favorite) PM
instance remains for assignment.
In terms of computational complexity, the heuristic has
substantial advantages over the optimal approach: As it can
be observed from Listing 1, the computational complexity
of the first phase grows linearly with the number of PM
types and VM types. The latter is commonly substantially
smaller than the number of bids and thus negligible. For the
second phase, according to Listing 2, the maximum number
of iterations corresponds to the number of bids. In each
iteration, the computational complexity relates linearly to
the number of bids and PM types again. Thus, the
computational complexity of the heuristic allocation
approach is polynomial and corresponds to O(|B| 2 ∗ |P |).
III. EVALUATION
Both optimization approaches from the previous section
have been implemented in a prototypical Java program. In
order to solve the optimization model for the first allocation
approach, we map it into a programmatic representation
�

Listing 2 Algorithm for VM Distribution
1: for all p ∈ P do
2: jp ← 0
3: end for
4: repeat
5: ˆp ← ∅, Û ← 0
6: for all p ∈ P do
7: if jp < np then
8: Lp ← ∅, Ip ← ∅
9: for all b ∈ S do
10: if Tb /∈ Ip then
11: if p.canHost(Lp ∪ b) = true then
12: Lp ← Lp ∪ b
13: else
14: Ip ← Ip ∪ Tb
15: end if
16: end if
17: end for
18: Up ←
� �
b∈Lp QTb
19:
20:
if Up > Û then
Û ← Up, ˆp ← p
21: end if
22: end if
23: end for
24: if ˆp �= ∅ then
25: jˆp ← jˆp + 1
26: for all b ∈ Lˆp do
27:
28:
xbvj ← 1 ˆp
end for
29: S ← S \ Lˆp
30: end if
31: until S = ∅ ∨ ˆp = ∅
� �
/ CFp + �
b∈Lp CVpTb
�
using the JavaILP framework 3 . As actual solvers, the
commercial IBM CPLEX Optimizer 4 and the free lpsolve 5
frameworks may be employed, with the first constituting the
default choice in our evaluation.
A. Approach and Methodology
The aim of our evaluation lies in a quantitative assessment
of the two optimization approaches. Thus, the evaluation
complements the brief and solely qualitative analyses of
computational complexity from Subsections II-B and II-C.
Our focus lies on two metrics that are of practical relevance
in the context of auction-based VM allocation: First, the
metric computation time demonstrates the overall scalability
of the approaches. It also expresses the delay that is
introduced into the allocation process through the
3 http://javailp.sourceforge.net/
4 http://www.ibm.com/software/integration/optimization/cplex-optimizer/
5 http://sourceforge.net/projects/lpsolve/
application of the optimization approaches. Second, the
metric profit represents the absolute solution quality of the
approaches. It thus expresses the utility of an optimized
allocation to the cloud provider in monetary terms.
For the evaluation, we have created 18 distinct classes of
EPAAPs. Each class contains 100 individual problems.
Across the different classes, we vary the problem dimension
with respect to the regarded number of bids and PM types.
The number of VM types and resource types are fixed
across all classes. Each problem is randomly generated
based on realistic data, which has been obtained as
described in the following.
For the data of the VM types in the evaluation, we use the
specifications provided by Amazon Web Services for its
Elastic Computing Cloud (EC2) and Spot Instances offers 6 .
Excluding special-purpose and non-deterministic types
(namely, Cluster and Micro), we infer eight different VM
types. Each of these types exhibits specific resource
demands with respect to three resource types, namely
processor (CPU), memory (RAM), and storage (HDD).
Unfortunately, to the best of our knowledge, Amazon Web
Services has not published detailed information about the
PMs in its data centers to date. However, empirical results
by an industry researcher indicate that the most resource
intensive VM types are run on dedicated PMs [10]. Based
on this notion, the specifications of the High-Memory
Quadruple Extra Large VM type, which exhibits the highest
demands for each resource type, is assumed as baseline for
the definition of five different PM types. Based on
calculations by Walker [2], the fixed operating costs of
these PM types are conservatively estimated to range from
$0.20 to $0.40 per hour. The specific figure for each PM
type is correlated with its resource supply.
The bid prices that the cloud users submit to the cloud
provider were modeled using specific distribution functions
Fv for each VM type v. For the choice of these distribution
functions, we reason as follows: The lowest observed bid
price, αv, will most likely correspond to the lowest
permissible bid price, i. e., αv = 0.01. Given that the users
act rational, the highest observed bid price, γv, will not
exceed the price Ov ∈ R of a so-called On-Demand VM
instance, which imposes a static usage fee, but essentially
offers guaranteed availability in return. Thus, we assume
γv = Ov − 0.01. Lastly, according to empirical findings by
Wee [5] and statements by Amazon Web Services [11], the
average savings from Amazon Spot Instances – compared to
On-Demand instances – amount to 52.3% and between 50%
and 66% respectively. Thus, we assume that the most
frequently observed bid price, βv, corresponds to 50% of
the On-Demand price, i. e., βv = 0.5 ∗ Ov. The underlying
notion is that the cloud users would like to maximize their
savings from the auction mechanism, but also maintain a
6 http://aws.amazon.com/en/ec2/#instance

easonable chance of actually being allocated VM instances.
Based on the previous reasoning, we believe that the use of
a triangle distribution, i. e., Fv ∼ T r(αv, βv, γv), best
reflects a realistic bidding behavior.
Based on past research by Greenberg et al. [12] and Barroso
and Hölzle [13], we estimate that the additional variable
operating cost of a fully utilized server amounts to about
25% of its fixed operating cost. As a first approximation, we
further assume that the cost of hosting a certain VM type
linearly relates to the resource utilization on the underlying
PM. Accordingly, the variable operating costs of each VM
type on every PM type are determined using Equation 13.
CVpv = 0.25∗CFp∗ 1
|R| ∗�
RDvr
RSpr
r∈R
B. Results and Discussion
∀p ∈ P, v ∈ V (13)
Following the problem generation, each EPAAP was solved
using both optimization approaches, optimal and heuristic.
In the process, the respective computation time and
resulting profit for the cloud provider were recorded. In
order to complete the evaluation within an acceptable
amount of time, a timeout period of 5 minutes (i. e., 300
seconds) was imposed per problem and approach. The
evaluation was conducted on a desktop computer with an
Intel Core 2 Quad Q9450 processor and 4 GB of memory,
operating under Microsoft Windows 7.
Table I provides an overview of the evaluated EPAAP
classes. As it has been previously outlined, the number of
bids (dB) and PM types (dP ) was varied for each problem
class, whereas the number of VM types and resource types
was fixed (dV = 8, dR = 3). The table also indicates the
percentage of problems that could be solved within the
specified timeout period by each optimization approach.
Only those problems that could be solved by both
approaches served as sample for the evaluation results
provided hereafter. Please note that problem classes E and
F constitute an exception, because they were exclusively
solved using the heuristic approach in order to show its
applicability to large-scale problems.
To begin with, Figure 2 depicts the absolute average
computation times per problem across all considered
problem classes. In accordance with the qualitative
discussion, the computation times for the optimal approach
grow roughly exponentially with the number of bids. The
effect of an increasing number of PM types is less
accentuated, but still well observable. For the problem
classes involving 20 bids (B1−4), the absolute computation
time of the optimal approach reaches the magnitude order
of one second. For problem classes that involve 30 bids or
more (C1−4 and D1−4), the computation time reaches and
even exceeds the magnitude order of ten seconds.
Accordingly, a substantial share of problems cannot be
solved at all within the specified timeout period (cf.
Table I
OVERVIEW OF EVALUATED EPAAP CLASSES AND SHARE OF SOLVED
PROBLEMS PER CLASS.
Problem class Solved problems (%)
ID dP dB Opt. Heur. Both
A1 1 10 100 100 100
A2 1 20 100 100 100
A3 1 30 100 100 100
A4 1 40 100 100 100
B1 2 10 100 100 100
B2 2 20 100 100 100
B3 2 30 100 100 100
B4 2 40 100 100 100
C1 3 10 95 100 95
C2 3 20 97 100 97
C3 3 30 85 100 85
C4 3 40 80 100 80
D1 4 10 84 100 84
D2 4 20 72 100 72
D3 4 30 54 100 54
D4 4 40 46 100 46
E 4 1000 – 100 –
F 4 10000 – 100 –
Table I). In contrast, for the heuristic approach, the average
computation time does not exceed the magnitude order of
one millisecond up to problem class D. For problem classes
E and F , the average computation time lies in the
magnitude order of ten milliseconds and one second
respectively. In general, the increase with growing problem
size is rather moderate and roughly corresponds to the
square of the number of bids. In accordance, the timeout is
not of practical relevance to the heuristic approach.
The performance difference between the two approaches is
further highlighted by Figure 3, which depicts the ratio of
computation time between the heuristic and optimal
approach. For the smallest problem classes involving 10
bids (A1−4), the ratio amounts to approximately 0.2% or
less. This is equivalent to a reduction of more than 99.8% in
computation time. For the larger problem classes involving
additional bids, the ratio converges toward 0, indicating
reductions of more than 99.9% by the heuristic approach.
The results for the second metric, profit, are given in
Figure 4. As can be seen, the heuristic approach achieves
favorable and consistent results, ranging between about
95.2% and 97.6% compared to the optimal solution. On
average across all classes, the figure corresponds to
approximately 96.7%. That means, due to the application of
the heuristic allocation approach, the cloud provider would
incur an average reduction in profit of 3.3% compared to
the optimal solution.
In summary, the results indicate that the computation of an
optimal solution to the EPAAP is difficult to achieve under
practical conditions. A cloud provider will usually receive

Computation time [ms]
100000
10000
1000
100
10
1
0.1
A1:1;10
A2:2;10
A3:3;10
A4:4;10
B1:1;20
B2:2;20
B3:3;20
B4:4;20
C1:1;30
C2:2;30
C3:3;30
C4:4;30
EPAAP class (ID:dP ;dB)
Optimal approach
Heuristic approach
Figure 2. Mean absolute computation times per problem for both
optimization approaches.
Ratio of computation time [%]
0.25
0.2
0.15
0.1
0.05
0
A1:1;10
A2:2;10
A3:3;10
A4:4;10
B1:1;20
B2:2;20
B3:3;20
B4:4;20
C1:1;30
C2:2;30
C3:3;30
C4:4;30
EPAAP class (ID:dP ;dB)
D1:1;40
D1:1;40
D2:2;40
D3:3;40
D2:2;40
D4:4;40
Heuristic / Optimal approach
Figure 3. Ratios of computation times between both optimization
approaches (based on micro-average).
Ratio of profit [%]
100
98
96
94
92
90
A1:1;10
A2:2;10
A3:3;10
A4:4;10
B1:1;20
B2:2;20
B3:3;20
B4:4;20
C1:1;30
C2:2;30
C3:3;30
C4:4;30
EPAAP class (ID:dP ;dB)
D1:1;40
D2:2;40
D3:3;40
D3:3;40
E:4;1000
D4:4;40
Heuristic / Optimal approach
Figure 4. Ratios of profits between both optimization approaches (based
on micro-average).
D4:4;40
F:4;10000
- All -
- All -
thousands of bids that have to be regarded in the allocation
process. At the same time, the cloud provider underlies
stringent time constraints, given that the timespan between
accepting the last bids in an auction period and the
announcement of the resulting equilibrium prices and VM
allocations should be minimal. For such application
scenarios, our proposed heuristic approach presents a viable
option. It achieves substantial reductions in computation
time, but also retains a favorable solution quality in terms
of profit for the cloud provider.
IV. RELATED WORK
To the best of our knowledge, we are the first to
scientifically examine the Equilibrium Price Auction
Allocation Problem, i. e., the challenge of concurrently
pricing and distributing VM instances based on an auction
scheme. However, in the broader context of cloud
computing, a substantial amount of work has been
conducted with respect to related topics. In the following,
we focus on a set of papers that we consider representative
for each topic area.
Breitgand et al. [14], for instance, have proposed an
optimization model and corresponding heuristic for the
distribution of VM instances on physical hosts through a
cloud provider. The authors take into account various
constraints, including resource demands and supplies, and
also permit the definition of different objectives, including
profit maximization. However, their research does not
address the aspect of pricing the VMs in an auction-based
setting.
Korupolu et al. [15] have proposed an optimization scheme
for the placement of applications, which comprise compute
and storage components, in data centers. In accordance with
our work, their heuristic approach is inspired by the
Knapsack problem. However, the work of Korupolu et al.
does not involve an auction-based pricing mechanism.
Zaman and Grosu [3] have examined the allocation of VM
instances to physical machines based on combinatorial
auctions, where cloud users bid for arbitrary bundles of VM
instances. The authors propose multiple heuristic allocation
strategies, which are evaluated with respect to different
objectives, including maximization of revenues. In contrast
to our work, the prices of identical VM types may be
discriminated between different users, which is not the case
in equilibrium price auctions. Furthermore, the authors do
not explicitly regard the distribution among physical hosts
under resource constraints.
Zaman and Grosu [16] have additionally addressed the issue
of auction-based VM allocation in a more recent paper.
However, the focus of this work lies on bidding strategies
for the cloud user, rather than optimization approaches for
the cloud provider.
Lin et al. [17], in accordance with our work, have proposed
an allocation mechanism for second-price (i. e., equilibrium

price) auctions. However, the authors solely focus on the
optimal pricing of resources, but do not consider the
concurrent distribution of VM instances across physical
hosts.
Özer and Özturan [18] have proposed an optimal approach,
as well as different heuristics for the allocation of grid and
cloud resources. In contrast to our work, the authors assume
combinatorial auctions, where users submit bids for bundles
of resources, which results in a different pricing approach.
In addition, Özer and Özturan do not consider the
distribution to physical hosts as part of the allocation
process.
In our own previous research [9], we have examined the
Software Service Distribution Problem. This challenge
concerns the cost-minimal distribution of Software as a
Service instances across leased VM instances under
resource constraints. We have presented an optimal, as well
as heuristic solution approach. However, these approaches
only address the distribution process. They do not cover the
concurrent pricing of entities, which constitutes a major
challenge in the EPAAP.
V. SUMMARY AND OUTLOOK
In the work at hand, we have introduced the Equilibrium
Price Auction Allocation Problem (EPAAP), a challenge in
the context of cloud computing. This problem concerns
cloud providers and involves the concurrent pricing and
distribution of virtual machines across physical machines
based on equilibrium price auctions.
As first major contribution, we have introduced a
mathematical formulation of the EPAAP as binary integer
program. This model serves as the basis of an optimal
allocation approach, which permits the computation of
profit-maximizing solutions to the EPAAP. As second major
contribution, given the computational complexity of the
optimal approach, we have developed a heuristic approach.
This heuristic trades substantial reductions in computation
time against small reductions in overall profit.
Through an evaluation based on realistic data from the
cloud computing domain, we have demonstrated the
practical applicability of our approaches. Specifically, we
have shown that the heuristic is able to achieve reductions
in computation time of more than 99.9% compared to an
optimal approach. At the same time, it achieves profits that
correspond to about 96.7% of the optimal solution on
average. Thus, our work is not only of scientific interest, but
can also provide a foundation for the practical application of
equilibrium price auctions in the cloud computing domain.
In our future work, we will aim at the further improvement
of the heuristic approach with respect to the solution quality.
In addition, we plan to extend the proposed approaches
such that the specific requirements of an optimization across
multiple subsequent auction periods, such as the live
migration of virtual machines, are supported.
ACKNOWLEDGMENTS
This work has partly been sponsored by E-Finance Lab
e. V., Frankfurt a.M., Germany (www.efinancelab.de).
REFERENCES
[1] R. Buyya, C. Yeo, S. Venugopal, J. Broberg, and I. Brandic,
“Cloud Computing and Emerging IT Platforms: Vision, Hype,
and Reality for Delivering Computing as the 5th Utility,”
Future Generation Computer Systems, vol. 25, no. 6, pp.
599–616, 2009.
[2] E. Walker, “The Real Cost of a CPU Hour,” Computer,
vol. 42, no. 4, pp. 35–41, 2009.
[3] S. Zaman and D. Grosu, “Combinatorial Auction-Based
Allocation of Virtual Machine Instances in Clouds,” in IEEE
2nd Int. Conf. on Cloud Computing Technology and Science,
2010.
[4] R. T. Bahman Javadi and R. Buyya, “Statistical Modeling of
Spot Instance Prices in Public Cloud Environments,” in 4th
IEEE/ACM Int. Conf. on Utility and Cloud Computing, 2011.
[5] S. Wee, “Debunking Real-Time Pricing in Cloud Computing,”
in 11th IEEE/ACM Int. Symp. on Cluster, Cloud and Grid
Computing, 2011.
[6] O. Agmon Ben-Yehuda, M. Ben-Yehuda, A. Schuster, and
D. Tsafrir, “Deconstructing Amazon EC2 Spot Instance
Pricing,” in IEEE 3rd Int. Conf. on Cloud Computing
Technology and Science, 2011.
[7] W. Domschke and A. Drexl, Einführung in Operations
Research, 6th ed. Springer, 2004, in German.
[8] F. Hillier and G. Lieberman, Introduction to Operations
Research, 8th ed. McGraw-Hill, 2005.
[9] U. Lampe, T. Mayer, J. Hiemer, D. Schuller, and
R. Steinmetz, “Enabling Cost-Efficient Software Service
Distribution in Infrastructure Clouds at Run Time,” in 2011
IEEE Int. Conf. on Service Oriented Computing &
Applications, 2011.
[10] H. Liu, “Amazon’s Physical Hardware and EC2 Compute
Unit,” http://huanliu.wordpress.com/2010/06/14/
amazons-physical-hardware-and-ec2-compute-unit/, last
accessed May 10, 2012.
[11] Amazon Web Services LLC, “Using Amazon EC2 Spot
Instances for Scientific Computing,”
http://aws.amazon.com/en/ec2/spot-and-science/, last accessed
May 10, 2012.
[12] A. Greenberg, J. Hamilton, D. Maltz, and P. Patel, “The Cost
of a Cloud: Research Problems in Data Center Networks,”
ACM SIGCOMM Computer Communication Review, vol. 39,
no. 1, pp. 68–73, 2008.
[13] L. Barroso and U. Hölzle, “The Case for Energy-Proportional
Computing,” Computer, vol. 40, no. 12, pp. 33–37, 2007.
[14] D. Breitgand, A. Maraschini, and J. Tordsson, “Policy-Driven
Service Placement Optimization in Federated Clouds (TR
H-0299),” IBM Research, Tech. Rep., 2011.

[15] M. Korupolu, A. Singh, and B. Bamba, “Coupled Placement
in Modern Data Centers,” in 2009 IEEE Int. Symp. on
Parallel & Distributed Processing, 2009.
[16] S. Zaman and D. Grosu, “Efficient Bidding for Virtual
Machine Instances in Clouds,” in 2011 IEEE Int. Conf. on
Cloud Computing, 2011.
[17] W. Lin, G. Lin, and H. Wei, “Dynamic Auction Mechanism
for Cloud Resource Allocation,” in 10th IEEE/ACM Int. Conf.
on Cluster, Cloud and Grid Computing, 2010.
[18] A. Özer and C. Özturan, “An Auction Based Mathematical
Model and Heuristics for Resource Co-Allocation Problem in
Grids and Clouds,” in 5th Int. Conf. on Soft Computing,
Computing with Words and Perceptions in System Analysis,
Decision and Control, 2009.

Cost-driven Optimization of Complex Service-based Workflows
for Stochastic QoS Parameters
Dieter Schuller, Ulrich Lampe, Julian Eckert, Ralf Steinmetz
Multimedia Communications Lab (KOM)
Technische Universität Darmstadt
Darmstadt, Germany
Email: {firstname.lastname}@KOM.tu-darmstadt.de
Abstract—The challenge of optimally selecting services from
a set of functionally appropriate ones under Quality of Service
(QoS) constraints – the Service Selection Problem – has been
extensively addressed in the literature based on deterministic
parameters. In practice, however, Quality of Service QoS
parameters rather follow a stochastic distribution. In the
work at hand, we present an integrated approach which
addresses the Service Selection Problem for complex workflows
in conjunction with stochastic Quality of Service parameters.
Accounting for penalty cost which accrue due to Quality
of Service violations, our approach reduces the impact of
stochastic QoS behavior on total cost significantly.
Keywords-Service Selection, Stochastic Quality of Service,
Optimization, Simulation
I. INTRODUCTION
The selection of services from a set of appropriate ones
that are able to provide the required functionality and thereby
best meeting cost and Quality of Service (QoS) requirements
– the Service Selection Problem (SSP) – is widely
recognized in the literature, e.g., [1]–[4]. The optimization
of the SSP is thereby based on deterministic QoS values.
A solution to the SSP describes an execution plan, i.e., an
assignment of services to certain tasks of a workflow, which
satisfy the mentioned cost and QoS constraints.
But QoS, e.g., the response time of a service or its
availability, is not always deterministic in reality. Due to
network latency or server load, response times of services
may change dynamically. I.e., when the execution of the
computed execution plan actually takes place, the perceived
QoS might differ from the expected QoS which has previously
been used for the calculation of the execution plan.
Thus, although having computed an optimal solution to the
SSP during design time which satisfies the constraints, it still
is possible that these constraints are violated during runtime.
The work at hand addresses this issue. Based on a service
broker scenario, which is presented in Section II, we describe
how QoS violations due to stochastic QoS behavior negatively
impacts total cost. In order to account for this impact
of stochastic QoS parameters, we present an integrated
approach comprising an optimization, a simulation, and an
adaptation step.
Stefan Schulte
Distributed Systems Group
Vienna University of Technology
Vienna, Austria
Email: s.schulte@infosys.tuwien.ac.at
During the optimization step, we compute an optimal
solution to the SSP, i.e., an optimal execution plan, satisfying
the QoS constraints based on the deterministic QoS values
denoted by the respective service providers. In the simulation
step, we observe the expected runtime behavior of the
computed execution plan in terms of QoS. This way, we can
assess potentially occurring constraint violations. We thereby
assume that violating QoS constraints is penalized, i.e.,
penalty fees become due in addition to service invocation
cost. According to the results of the simulation, we apply a
greedy adaptation heuristic in order to reduce the impact of
potentially occurring constraint violations.
The remainder of this work is structured as follows. In
Section II, we present a motivating scenario, which will be
used throughout the paper. In Section III, we describe our
solution to the SSP, which is based on our previous work
in [5]. The applied simulation approach is presented in Section
IV. Based on the simulation results, we apply our greedy
adaptation heuristic which is described in Section V and
evaluated in Section VI. Finally, after having distinguished
our approach from related work in Section VII, we draw
conclusions and discuss future work in Section VIII.
II. SCENARIO
In this section, we present a scenario, which is used as
an example in the work at hand in order to illustrate the
impact of stochastic QoS parameters. The application of our
approach is not limited to this scenario.
Imagine a service broker who receives requests from its
customers. Paying the broker a fixed amount of money,
the customers require certain tasks and workflows, respectively,
to be executed. For this, they provide the broker
with a document which specifies the required tasks from
a functional perspective and indicates the ordering of the
tasks, i.e., the structure of the workflow. One of the broker’s
customers asks for instance for the workflow in Figure 1.
The process steps P Si thereby indicate the tasks which have
to be accomplished. Each task can be executed by a single
service.

PS1
0.7
0.3
X
PS5
PS3
PS2
PS7
Figure 1: Example Workflow (in BPMN)
In addition to these functional requirements, the customers
also specify their QoS needs regarding the execution of the
respective workflow. For this, they provide restrictions in the
form of upper or lower bounds for specific QoS parameters,
the so-called Service Level Objectives (SLOs). With this
information, the broker tries to select those services among
functionally appropriate ones which satisfy the customers’
QoS requirements, as a violation of these requirements will
be penalized by the broker’s customers. Having selected respective
services, the broker pays and invokes these services
in order to execute the customers’ tasks and workflows, respectively.
If the customers’ QoS requirements are violated,
penalty fees become due, which also have to be paid by
the broker. In order to reach an optimized decision, the
broker models the selection of services as an optimization
problem aiming at minimizing invocation cost and satisfying
QoS constraints. I.e., the broker formulates an SSP, which
is described in the following Section III.
PS4
PS6
III. SERVICE SELECTION PROBLEM
In order to formulate an SSP and therewith to compute
an execution plan, it is necessary to aggregate the QoS and
cost values of eligible candidate services according to the
regarded workflow structures. For this, we specify a system
model in Section III-A and discuss respective aggregation
functions in Section III-B. Finally, we utilize the presented
aggregation functions and provide an optimization model in
Section III-C.
A. System Model
In this section, we describe the system model utilized in
the work at hand. The set of all tasks of a workflow is
labeled with I, i ∈ I = {1, ..., i # }. The set of services
appropriate to accomplish a certain task i is labeled with Ji,
j ∈ Ji = {1, ..., j #
i }. The decision variables xij ∈ {0, 1}
indicate whether a service j is selected to accomplish task i.
According to our running example, we consider cost c
(charge for invoking a service in cent applying a pay-peruse
pricing model), response time r (time elapsed between
invoking a service and receiving its response), availability
a (probability that a service is available), and throughput d
(number of requests a service is able to serve within a certain
time interval). These parameters – in fact, even a subset of
these parameters – are sufficient to cover the aggregation
types summation, multiplication, and min/max-operator (cf.
X
PS8
Section III-B). Further QoS parameters can be integrated
into the optimization problem straightforwardly. Bounds for
these parameters are labeled with bc, br, ba, bd.
Regarding branchings, we label the set of paths with
L, in which l ∈ L = {1, ..., l # } indicates the respective
path number. Referring to the workflow in Figure 1, there
are two paths l within the AND-block, thus L = {1, 2}.
Different sets of paths will be distinguished by utilizing
additional indices, i.e., La, Lx for AND/XOR. We refer to
them as branching La, Lx. The tasks within a branching are
covered by the set IL ⊆ I, whereas Il ⊆ IL represents
the set of tasks within path l. We label the set of the
remaining tasks, which are not located within a branching,
with Is = I \ (Il | l ∈ L). Utilizing this system model, we
develop aggregation functions in the following.
B. Aggregation Functions
As previously stated, it is necessary to aggregate the
QoS and cost values of candidate services according to
regarded workflow patterns in order to compute the overall
cost and QoS for a specific workflow. Regarding our example
scenario, this actually is a prerequisite for comparing
workflow QoS with respective bounds issued by the broker’s
customers. As previously stated, violation of the customers’
bounds will result in additional penalty cost. Thus, the
broker has a high interest in making sure that the customers’
bounds are satisfied. For this, the broker performs a worstcase
analysis as opposed to a best-case or average-case
analysis. While probabilities for the execution of certain
paths of a branching would be considered in an averagecase
analysis, the worst (best) paths of each branching in
terms of aggregated cost and QoS are considered in a worstcase
(best-case) analysis for the computation of an optimal
solution to the SSP. Respective aggregation functions for
sequences, AND-blocks, and XOR-blocks are indicated in
Table I.
Table I: Worst-Case Aggregation Functions
Sequence AND-block XOR-block
� �
� �
� �
rijxij max rijxij max rijxij
l∈L
l∈L
i∈Is � j∈J �i
� �i∈I
l j∈J � i
i∈I �l
j∈J �i
aijxij
aijxij min aijxij
l∈L
i∈Is j∈J �i
l∈L i∈Il j∈Ji i∈Il j∈Ji min dijxij min
i∈Is
l∈L
j∈Ji (min
�
dijxij) min
i∈Il l∈L
j∈Ji (min
�
dijxij)
i∈Il � � � � �
� j∈J �i
cijxij
cijxij max cijxij
l∈L
i∈Is j∈Ji l∈L i∈Il j∈Ji i∈Il j∈Ji For a sequence, the cost and QoS values of all services
selected to accomplish certain tasks i have to be aggregated
according to the respective aggregation type, e.g., summed
up for cost. Regarding AND-blocks, it has to be noted that
for response time r only the path with the highest aggregated
response time – the critical path – requires consideration, as
the tasks within the different paths are executed in parallel.

Regarding the other non-functional parameters, the cost
and QoS values of all services have to be aggregated as
all selected services are executed in the end (similar to a
sequence). For XOR-blocks, where only one of the potential
paths is executed, we consider the path with the worst
aggregated cost and QoS values according the respective
aggregation types as we pursue a worst-case analysis.
In the work at hand, we stick to these patterns for the
sake of simplicity. The interested reader may refer to our
former work in [5] for further structured workflow patterns
(OR-blocks, Repeat Loops) as well as for unstructured patterns
of Directed Acyclic Graphs and respective aggregation
functions, which could have additionally been utilized.
Applying the presented aggregation functions, we formulate
the SSP as an optimization problem in the following.
C. Optimization Model
In this section, we formulate the SSP in Model 1 –
accounting for the example workflow in Figure 1. For
this, we specify an objective function in (1) and a set of
restrictions in (2)–(10) by applying the aggregation functions
from Table I.
Model 1 Service Selection Problem for Example Workflow
Objective Function
minimize � �
cijxij + �
(c ′ l + � �
cijxij) (1)
so that
i∈Is j∈Ji
i∈Is j∈Ji
l∈La
l∈La
i∈Il j∈Ji
� �
rijxij + max(r
l∈La
i∈Is j∈Ji
′ l + � �
rijxij) ≤ br (2)
i∈Il j∈Ji
( � �
aijxij) · ( �
(a ′ l · � �
aijxij)) ≥ ba (3)
min(min
i∈Is
�
i∈Il j∈Ji
dijxij, min(d
l∈La
j∈Ji
′ l, min
i∈Il
j∈Ji
�
dijxij))) ≥ bd (4)
max (
lx∈Lx
� �
rijxij) = r
i∈Ilx j∈Ji
′ l ∀l ∈ La| interlaced (5)
min (
lx∈Lx
� �
aijxij) = a
i∈Ilx j∈Ji
′ l ∀l ∈ La| interlaced (6)
max
lx∈Lx
(min(
�
dijxij)) = d ′ l ∀l ∈ La| interlaced (7)
i∈Ilx
j∈Ji
max (
lx∈Lx
� �
cijxij) = c
i∈Ilx j∈Ji
′ l ∀l ∈ La| interlaced (8)
�
xij = 1 ∀i ∈ I (9)
j∈Ji
xij ∈ {0, 1} ∀i ∈ I, ∀j ∈ Ji (10)
Note that the considered workflow in Figure 1 contains
an XOR-block within the AND-block. As the aggregation
functions indicated in Table I assume the tasks within splits
and corresponding joins to be arranged sequentially (cf. [6]),
we have to account for this interlaced structure. Referring to
our former work in [6], we abstract from the interlacing and
insert additional variables r ′ l , c′ l , a′ l , d′ l
and their respective
aggregation functions (i.e., for XOR-blocks in this case)
into Model 1. As the broker receives a fixed amount of
money from his/her customers for satisfying their needs,
(s)he maximizes his/her profit by selecting and invoking
the services with minimal cost. Thus, the objective function
in (1) aims at minimizing service invocation cost. The
customers’ restrictions on cost and QoS are indicated in (2)–
(7). In (9), we make sure that only one service is selected
for each task, and (10) represents the integrality condition
for the decision variables.
Model 1 constitutes a non-linear optimization problem as
it contains non-linear aggregations of decision variables xij,
i.e., multiplication and min/max-operator in (2), (3), (4) for
instance. We transform it into a linear one by linearizing the
non-linear restrictions in (3)–(7). Due to space restrictions,
we omit describing the linearization in the work at hand
and refer the interested reader to our former work in [5],
[6]. The optimal solution to the obtained linear optimization
problem can then be computed by applying Integer Linear
Programming (ILP) techniques from the field of Operations
Research [7].
Thus, having computed an optimal execution plan minimizing
service invocation cost and (presumably) satisfying
the customers’ constraints, the broker would apply this
solution and invoke the respective services while assuming
that they actually hold the constraints. But as the invoked
services may show a different behavior during runtime than
expected during design time, the application of the computed
execution plan could lead to violations of the customers’
constraints. This would result in additional cost for the
broker as, in this case, penalty fees will become due.
In order to assess the impact of stochastic QoS values,
we propose to perform an additional simulation step, which
is described in the following Section IV.
A. General Concept
IV. SIMULATION
The simulation approach presented in the work at hand
is substantially inspired by findings from the domain of
project management. In our former work in [8], we outlined
the conceptual similarities between workflows and project
networks, specifically, generalized activity networks [9]: a
complex workflow can easily be interpreted as a project
network, which consists of comparable entities, such as tasks
and branches.
Based on the work in [10], we argued in [8] that simulation
provides the best means to assess the risk of breaking

given QoS constraints in practical application. Referring to
[10], deterministic methods regularly fail to correctly quantify
such risks, specifically if large and complex networks
and workflows, respectively, are concerned.
Thus, regarding our example scenario, an optimization
approach based on deterministic QoS values does not capture
the risk of violating QoS constraints. For instance, if a
selected service experiences unusually high demand, it may
not be able to provide a certain response time, even if a
corresponding bound has been guaranteed by the service
provider. Thus, in turn, an execution plan may not satisfy
the QoS requirements in some cases. The existence of
substantial QoS fluctuations – specifically with respect to
Web service response times – has been empirically shown,
for instance by Rosario et al. [11] and Miede et al. [12]. In
the work at hand, we quantify these inherent risks through
simulation, i.e., repeated “virtual” execution of a workflow.
B. Stochastic QoS Parameters
As previously stated, it is assumed in most related approaches
addressing the SSP that a “hard” – deterministic
– QoS guarantee is specified for each service candidate and
QoS parameter. In order to perform a simulation, we further
assume that probabilistic QoS specifications are available.
In accordance with the notation in Section III-A, Rij, Aij,
and Dij represent random variables for the QoS parameters
response time, availability, and throughput of service j
for task i. These random variables may follow arbitrary
probability distributions. For instance, the response time of
a service may be modeled using a normal distribution, i.e.,
Rij ∼ N(100, 20). For an overview of common probability
distributions, we refer to corresponding compendiums [13].
Probability distributions can be deduced in at least two
principle ways: First, they may be explicitly provided by a
service provider as part of contracted guarantees in terms of
Service Level Agreements (SLA), following the idea of “soft
contracts” [11]. Second, they may be inferred by mining the
monitoring data from past service executions.
C. Execution of the Simulation
For the actual simulation, we virtually execute the previously
computed execution plan a predefined number of
times. In each iteration, we draw a specific realization of
each QoS parameter for each selected service, based on
the given random variables Rij, Aij, and Dij. In addition,
depending on the branching probabilities of XOR-splits, we
draw random variables in order to determine which paths
are actually executed in the current iteration. The realized
values of the individual services are aggregated afterwards
according to the respective workflow structure applying the
aggregation functions from Table I. This way, we compute
the overall QoS values for the whole workflow.
Based on the workflow in Figure 1, we provide an
example in Figure 2. It represents an execution plan where
r11 = 100
R11 ~ N(95, 5)
S11
br = 430
0.7
0.3
S51
S21
X X
S31
r31 = 116
R31 ~ N(110, 7)
r51 = 111
R51 ~ N(100, 12)
r21 = 221
R21 ~ N(220, 23)
S71
r71 = 224
R71 ~ N(210, 15)
S41
r41 = 107
R41 ~ N(95, 13)
S61
r61 = 112
R61 ~ N(105, 8)
r81 = 106
R81 ~ N(100, 7)
Figure 2: Workflow execution plan with services’ QoS
guarantees, serving as input to the simulation.
possible, appropriate services have been identified for the
realization of all tasks. Each service is associated with
deterministic QoS bounds, as specified by respective service
providers (boxes with light-gray background). Further,
probabilistic QoS specifications – as observed by the broker
– are indicated (boxes with dark-gray background). Note
that we assume conservative service providers as the deterministic
QoS values specified by the service providers are
higher than the expected values. For simplicity reasons, we
only incorporate the QoS parameter response time in this
example. As can be quickly validated, the workflow will
meet the QoS constraint, namely br = 430, according to the
given deterministic QoS values.
However, performing a simulation with 10, 000 iterations,
i.e., drawing 10, 000 corresponding realizations of
the regarded QoS parameter response time, reveals that the
specified QoS constraint will be violated in approximately
15% of all executions – although we assumed conservative
service providers. Figure 3 indicates the respective aggregated
response time for the workflow. The QoS violation
can be attributed to one or more services exceeding their
deterministic QoS guarantees, which cannot be sufficiently
captured by the initial service selection and, thus, will result
in potentially severe penalities for the broker. In order to
reduce the impact of stochastic QoS behavior and therewith
the accruing penalties, the boker applies our greedy adaptation
heutistic, which is discribed in the following Section V.
V. GREEDY ADAPTATION HEURISTIC
In the previous section, we have outlined how a simulation
step may reveal the risk of violating certain QoS constraints
in practice. This knowledge can be exploited in order to minimize
the risk of penalties and therewith the total cost for the
broker in our example scenario comprising cost for invoking
the selected services and cost according to expected penalty
fees. For this, we present a greedy adaptation heuristic in
this section aiming at reducing the total cost for the broker.
In this context, adaptation denotes excluding and replacing,
respectively, those services from the formerly computed
execution plan for other functionally appropriate services,
S81

Cumulative probability
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
350 360 370 380 390 400 410 420 430 440 450 460 470
Workflow response time
Figure 3: Cumulative probability function of the workflow’s
response time, according to the simulation.
which cause high penalty cost due to unexpected runtime behavior.
One may argue that we could have also performed an
adaptation by trying to improve QoS for the services of the
computed execution plan, i.e., asking the respective service
providers to either enhance resource deployment in terms
of, e.g., main storage and CPU, or to (generally) improve
the implementation of the service providers’ services. As
both is out of the broker’s control sphere, we do not focus
on nor account for such possibilities. We further assume
SLAs to be fixed and do not account for SLA negotiations,
neither between the broker and its customers, nor between
the broker and respective service providers – which could
also be seen as an alternative adaptation in order to tackle
QoS violations – as (optimal and efficient) SLA negotiation
is a research topic on its own and would go beyond the
scope of the paper at hand.
Regarding penalty cost, we assume linear penalty fees per
unit of QoS violation, e.g., cents per second the execution
took longer than restricted by the respective bound, or cent
per percent point the availability was lower than restricted.
It would also be imaginable to assume variable penalty fees,
which increase quadratically or exponentially with the size
of the violation. As this would only influence the calculation
of the actual total penalty cost depending on the chosen
penalty model but does not change our approach, we stick
to linear penalty fees for the sake of simplicity.
Our greedy adaptation heuristic is indicated in Algorithm
1, using pseudocode. The heuristic is split into four
steps. First, we determine the total cost of the current solution.
Second, we compute the critical QoS parameter, i.e.,
the QoS parameter which causes the highest penalty cost, as
this parameter might bear the highest penalty cost savings.
Third, we determine for which of the tasks i ∈ I a potential
improvement of the respective critical QoS parameter would
be highest. Fourth, we finally perform the adaptation.
More details on these steps are provided in the following.
Referring to Algorithm 1, we compute an optimal solu-
tion cs, which represents the current solution, based on
deterministic QoS values, perform the previously mentioned
simulation step, and calculate the QoS violation in lines
2 to 4. In order to determine the total cost of the current
solution, we aggregate the invocation and accruing penalty
cost as indicated in line 5. By computing and comparing the
penalty cost for each QoS parameter, we determine critical
QoS parameter in lines 7 to 13. In order to compute the
critical task, we compute in line 16 the potential benefit for
each task i. In this context, the potential benefit of a task
i corresponds to the highest possible reduction in standard
deviation σ. Referring to (11), we compute the difference
of the selected service’s σs to the σj values of the other
candidate services of task i.
benefit = max(σs
− σj) (11)
j∈Ji
The highest possible reduction in σ, which corresponds
to the highest possible reduction in uncertainty and risk
regarding stochastic QoS parameters, thereby represents the
potential (absolute) benefit for that task. In line 17, we
compute the weight ω for task i. For this, we divide the
number of the simulation runs where task i has been virtually
executed and the restriction for the critical QoS parameter
has been violated by the total number of runs. As indicated
in lines 18 to 20, the task with the highest weighted benefit
becomes the critical task, which is required in order to
perform the actual adaptation.
As stated earlier, our adaptation comprises excluding and
replacing services while other adaptation techniques and
mechanisms also would have been possible and supported by
our heuristic. Insofar, our approach is flexible and extensible
for supporting further adaptation techniques. We determine
the currently selected service js in line 24. We ban in lines
25 to 27 all services of the critical task, which have a
negative benefit, i.e., whose σ and therewith the risk of QoS
violation is larger than or equal to the σjs
of the currently
selected service js. This way, we adapt the list of available
candidate services for the critical task.
Utilizing this adapted list, we rerun the optimization and
obtain a new execution plan, which is optimal under the
adapted circumstances. Afterwards, we conduct the simulation
step again and calculate both the invocation and penalty
cost for the lastly computed execution plan (cf. the first
step). By comparing the total cost of this new solution with
the previous – formerly known as current best – solution,
we determine whether applying our adaptation heuristic has
been advantageous from the broker’s point of view. Via a
parameter greed, we specify the algorithm’s degree of greed,
i.e., whether and how often the described steps are repeated
as long as past iterations reduced total cost. Via a further
parameter anneal, we control for allowing worse solutions
temporarily as starting point for a continuous application of
the algorithm. An evaluation for this approach is presented
in the following Section VI.

Algorithm 1 Greedy Adaptation Heuristic
1: //First step – determine current cost
2: cs = computeCurrentSolution();
3: sim = simulate(cs);
4: v = computeQoSViolation(sim);
5: totalCost = computeInvCost(cs) + computePanalty(v);
6: //Second step – determine the critical QoS parameter
7: for all p ∈ QoSparameters do
8: penaltyCost = computePenalty(p);
9: if penaltyCost ≥ highestP enaltyCost then
10: highestP enaltyCost = penaltyCost;
11: criticalQoSparameter = p;
12: end if
13: end for
14: //Third step – determine the critical task
15: for all i ∈ I do
16: benefit = computeBenefit(i);
17: ω = computeWeight(i);
18: if ω · benefit ≥ highestBenefit then
19: highestBenefit = ω · benefit;
20: criticalT ask = i;
21: end if
22: end for
23: //Fourth step – perform the adaptation
24: js = getSelectedServiceOf(i);
25: for all j ∈ Ji do
26: if σj ≥ σjs then
27: setBanned(j);
28: end if
29: end for
VI. EVALUATION
As a proof of concept, we implemented our greedy adaptation
heuristic. For the computation of an optimal solution to
the SSP based on deterministic values, we utilized the linear
programming solver CPLEX 1 . In this section, we evaluate
the impact of QoS violation on total cost with respect to our
broker scenario. For this, we conducted a set of experiments
in order to assess the effects of different configurations
regarding the adaptation parameters greed and anneal, which
allow to control for the number of iterations the algorithm
performs in order to achieve improved solutions and for the
number of thereby temporarily accepted worse solutions.
In the following, we describe our experimentation setup.
We consider the workflow in Figure 1, which contains P S1-
P S8 indicating tasks i ∈ {1, ..., 8}. In order to determine
QoS values of the respective services, we draw realizations
of the random variables Rij, Aij, and Dij. We assume
these random variables to follow a normal distribution, i.e.,
Rij ∼ N(µr, σr), Aij ∼ N(µa, σa), Dij ∼ N(µd, σd).
1 http://www.ibm.com/software/integration/optimization/cplex-optimizer/
Table II: Stochastic QoS
QoS P S2, P S7 P S1, P S3 − P S6, P S8
Rij
Aij
Dij
cij
µr ∼ U(160, 240)
σr ∼ U(0, 40)
µa ∼ U(0.92, 1.0)
σa ∼ U(0.0, 0.08)
µd ∼ U(80, 120)
σd ∼ U(0, 30)
U(0.8, 1.2) · (40+
(0.03 · (µd − µr)) · µ 2 a )
µr ∼ U(80, 120)
σr ∼ U(0, 20)
µa ∼ U(0.94, 0.98)
σa ∼ U(0.0, 0.04)
µd ∼ U(80, 120)
σd ∼ U(0, 30)
U(0.8, 1.2) · (20+
(0.03 · (µd − µr)) · µ 2 a )
We could have equally utilized other distribution functions
or inferred the respective distributions by mining the
monitoring data from past service executions, which we
actually envisage in our future work, but we stick to normal
distributions in the work at hand for the sake of simplicity.
The parameterization of the random variables Rij, Aij, and
Dij is indicated in Table II. We assume that the invocation
cost of a service partly depends on its QoS, i.e., good QoS
values in terms of low response time r, high availability a,
and high throughput d result in higher invocation cost. For
this, as also indicated in Table II, we compute the invocation
cost of a service according to its QoS, utilizing an additional,
uniform distributed random variable U(a, b).
In order to assess the impact of the number of beneficial
iteration steps – as indicated by the parameter greed –
on total cost and computation time, we varied greed in
Figure 4a and Figure 4d from 0 to 20 step two, utilizing
anneal = 4. Regarding the influence of the number of
temporarily allowed worse solutions, we varied the anneal
parameter in Figure 4b from 0 to 16 step two, utilizing a
fixed greed value of 10. For these experiments, we set the
violation cost to 10% of the respective service invocation
cost – per unit of QoS violation (cf. Section V). In Figure 4c,
we account for different penalty cost by varying the penalty
cost percentage from 0% to 20%, utilizing greed = 10,
anneal = 4. Finaly, in Figure 4d, the computation time for
computing respective solutions is indicated. The experiments
were performed on an Intel Core 2 Quad processor at
2.66 GHz, 4 GB RAM, running Microsoft Windows 7.
The evaluation results show that the application of our
greedy adaptation heuristic to the considered service broker
scenario leads to a cost reduction of 9 ct to 12.5 ct in relation
to total cost of 142.5 ct for the whole workflow, which corresponds
to a reduction of 6% to 8.5%. Thus, the broker could
save 6% to 8.5% of the total cost. But, Figure 4d reveals
that reducing the cost actually “costs” computation time –
up to 10 times as much. As Figure 4a indicates, additional
reduction in total cost decreases with additional adapatation
steps, i.e., higher greed values. Therefore, utilizing a greed
value of 4 to 6, for which the computation time is roughly
6 times higher, seems to be a good compromise between
cost reduction and additional computation time. Also values
greater than 4 for anneal do not improve the cost reduction

Total Cost (in ct)
165
160
155
150
145
140
135
130
No adapatation
Adaptation heuristic
125
0 5 10
Greed
15 20
(a) Impact of greed
Total Cost (in ct)
165
160
155
150
145
140
135
130
No adapatation
Adaptation heuristic
125
0 2 4 6 8 10 12 14 16
Annealing
(b) Impact of anneal
significantly. Thus, using the parameterization greed = 6,
anneal = 4, which leads to a cost reduction of 7.3% and a
6-times magnified computation time, appears sensible.
Having described our approach, we discuss related approaches
in the following Section VII.
VII. RELATED WORK
As previously stated, the SSP is widely recognized in
the literature. A survey of current approaches can be found
in [4]. In principle, current approaches can be divided into
two categories: heuristic approaches which try to find rather
good solutions within a reduced amount of computation
time, e.g., [2], [14], [15], and approaches aiming at finding
an optimal solution to the SSP, e.g., [1], [3], [16]. All those
approaches presume that the utilized QoS parameters are
deterministic. Relevant related work in the area of stochastic
QoS parameter, however, is rather sparse.
In their work in [11], Rosario et al. consider probabilistic
QoS values, but not for the purpose of services selection.
They rather focus on SLA and contract composition, respectively,
using soft probabilistic contracts as already stated in
Section IV-A. The authors in [17] pay insofar attention to
stochastic QoS parameters as they try to achieve an accurate
prediction of QoS values based on historic data which
then is utilized for service selection rather than predefined
values guaranteed by service providers. In [18], Hwang
et al. utilize Probability Mass Functions (PMFs) for QoS
instead of deterministic values. They describe approaches for
aggregating the PMFs of single services. The authors thereby
utilize a preselected set of services with discrete PMFs, i.e.,
they do not perform service selection based on PMFs, but
rather aim at computing and estimating QoS for servicebased
workflows. In contrast to this, our approach targets
service selection accounting for stochastic QoS parameters.
Cardellini et al. consider stochastic QoS parameters for the
SSP insofar as they use an α-percentile (with α = 95%)
for the QoS parameter response time [19]. Accordingly,
instead of utilizing deterministic response time values for the
optimization, the authors integrate a restriction demanding
the probability of violating the bound for response time to
be lower or equal to 1 − α, i.e., 1 − 0.95 = 5%. Projected to
our broker scenario, this means that the broker can assume
Total Cost (in ct)
165
160
155
150
145
140
135
130
Figure 4: Evaluation Results
No adapatation
Adaptation heuristic
125
0 0.05 0.1
Penalty Fee
0.15 0.2
(c) Impact of penalty cost
Computation Time (in msec)
900
800
700
600
500
400
300
200
100
0
No adapatation
Adaptation heuristic
0 5 10
Greed
15 20
(d) Computation time
satisfying the respective bound with a probability of 95%.
But Cardellini et al. thereby fail to account for the arising
penalty cost due to QoS violations in the remaining 5% of
the cases. Our approach, on the other hand, proceeds one
step beyond and considers the impact of QoS violations in
terms of penalty cost. Depending on the ratio between invocation
and penalty cost, it could be beneficial for the broker
selecting rather cheap services, which statistically cause QoS
violations more often, and paying the penalty cost rather
than selecting expensive services with low probabilities of
violating QoS constraints. Thus, our approach enables the
broker to select those services that bear the lowest cost.
In the approach in [20], which probably comes closest to
ours, Leitner et al. assume a fixed service composition and
a fixed set of possible adaptations to improve the service
composition in terms of, e.g., “utilize Express Shipping”
instead of “ utilize Standard Shipping”. The aim is to select
and apply those adaptations that minimize cost comprising
invocation cost, penalty cost for QoS violation, and cost for
applied adaptations, which aim at avoiding QoS violation. If
we abstract from the term adaptation and interpret available
adaptations as alternative services, then Leitner et al. are
solving a SSP with the aim of minimizing total cost, at which
penalty cost for QoS violation are considered as well. In
order to account for non-deterministic QoS behavior during
runtime, the authors utilize a predictor component which
predicts prospective QoS values and therewith expected QoS
violation. Thus, Leitner et al. estimate the impact of stochastic
QoS behavior for each service separately and perform a
optimization with these estimated, deterministic QoS values
allowing for QoS violations and accounting for their impact
on total cost. Our approach, however, goes one step further,
as we do not consider the stochastic QoS behavior of the
services independently of each other, but account for the
whole workflow during our simulation step. Thus, potential
reverse QoS deviations of different services from expected
behavior can be considered, which is not possible in [20] due
to their isolated consideration of expected QoS per service,
independently of other services.
In summary, our approach extends related work as it
considers, on the one hand, the impact of QoS violation in
terms of accruing penalty costs. On the other hand, we do not

only regard isolated stochastic QoS behavior for individual
services, but account for probably compensating reverse
QoS deviations of different services. Thus, we consider the
impact of stochastic QoS behavior for the whole workflow.
Conclusions are drawn in the following Section VIII.
VIII. CONCLUSION
The SSP is widely recognized in the literature and has
been discussed in several scientific papers – based on deterministic
QoS parameters. In the work at hand, we addressed
the SSP in conjunction with stochastic QoS parameters
which has been considered as yet only insufficiently in the
literature. For this, we presented an integrated approach
comprising an optimization, a simulation, and an adaptation
step which aims at reducing the impact of stochastic QoS
behavior on total cost. The evaluation shows that the application
of our approach leads to a cost reduction up to 8.5%,
utilizing the described service broker scenario. Thus, the
actual, absolute level of cost reduction depends on the concrete
paramerization and the regarded scenario. For this, we
will extend the evaluation in our future work by considering
further workflow structures and QoS distribution functions as
well as different degrees of conservative, deterministic QoS
values issued by service providers. In addition, we focus on
improving the scalability of our greedy adaptation heuristic.
ACKNOWLEDGMENT
This work is supported in part by the Commission of the
European Union within the ADVENTURE FP7-ICT project
(Grant agreement no. 285220) and by E-Finance Lab e. V.,
Frankfurt am Main, Germany (http://www.efinancelab.com).
REFERENCES
[1] D. Ardagna and B. Pernici, “Adaptive service composition in
flexible processes,” IEEE Transactions on Software Engineering
(TSE), vol. 33, no. 6, pp. 369–384, 2007.
[2] D. A. Menascé, E. Casalicchio, and V. K. Dubey, “A Heuristic
Approach to Optimal Service Selection in Service Oriented
Architectures,” in Workshop on Software and Performance
(WOSP). ACM, 2008, pp. 13–24.
[3] A. F. M. Huang, C.-W. Lan, and S. J. H. Yang, “An Optimal
QoS-based Web Service Selection Scheme,” Information
Sciences (ISCI), vol. 179, no. 19, pp. 3309–3322, 2009.
[4] A. Strunk, “QoS-Aware Service Composition: A Survey,” in
European Conf. on Web Services (ECOWS). IEEE Computer
Society, 2010, pp. 67–74.
[5] D. Schuller, A. Polyvyanyy, L. García-Bañuelos, and
S. Schulte, “Optimization of Complex QoS-Aware Service
Compositions,” in Int. Conf. on Service Oriented Computing
(ICSOC). Springer, 2011, pp. 452–466.
[6] D. Schuller, A. Miede, J. Eckert, U. Lampe, A. Papageorgiou,
and R. Steinmetz, “QoS-based Optimization of Service Compositions
for Complex Workflows,” in Int. Conf. on Service
Oriented Computing (ICSOC). Springer, 2010, pp. 641–648.
[7] H. Taha, Operations Research – An Introduction, 8th ed.
Pearson Prentice Hall, London, 2007.
[8] U. Lampe, D. Schuller, J. Eckert, and R. Steinmetz, “Optimizing
Service Selection for Probabilistic QoS Attributes,”
in Workshop on Architecture, Concepts and Technologies for
Service Oriented Computing (ACT4SOC). SciTePress, 2010,
pp. 42–51.
[9] C. W. Dawson and R. J. Dawson, “Generalised Activity-onthe-Node
Networks for Managing Uncertainty in Projects,”
Int. Journal of Project Management, vol. 13, no. 6, pp. 353–
362, 1995.
[10] R. J. Schonberger, “Why Projects are ’always’ late: a Rationale
based on Manual Simulation of a PERT/CPM Network,”
Interfaces, vol. 11, no. 5, pp. 66–70, 1981.
[11] S. Rosario, A. Benveniste, S. Haar, and C. Jard, “Probabilistic
QoS and Soft Contracts for Transaction-Based Web Services
Orchestrations,” IEEE Transactions on Services Computing
(TSC), vol. 1, no. 4, pp. 187–200, 2008.
[12] A. Miede, U. Lampe, D. Schuller, J. Eckert, and R. Steinmetz,
“Evaluating the QoS Impact of Web Service Anonymity,” in
European Conf. on Web Services (ECOWS). IEEE Computer
Society, 2010, pp. 75–82.
[13] M. McLaughlin, “A Compendium of Common Probability
Distributions,” http://www.causascientia.org/math stat/Dists/
Compendium.html, 2001.
[14] F. Lécué, “Optimizing QoS-Aware Semantic Web Service
Composition,” in Int. Semantic Web Conf. (ISWC). Springer,
2009, pp. 375–391.
[15] N. B. Mabrouk, N. Georgantas, and V. Issarny, “A Semantic
End-to-End QoS Model for Dynamic Service Oriented Environments,”
in Workshop on Principles of Engineering Serviceoriented
Systems (PESOS). IEEE Computer Society, 2009,
pp. 34–41.
[16] L. Zeng, B. Benatallah, A. H. H. Ngu, M. Dumas,
J. Kalagnanam, and H. Chang, “QoS-Aware Middleware for
Web Services Composition,” IEEE Transactions on Software
Engineering (TSE), vol. 30, no. 5, pp. 311–327, 2004.
[17] M. Li, J. Huai, and H. Guo, “An Adaptive Web Services
Selection Method Based on the QoS Prediction Mechanism,”
in Int. Joint Conf. on Web Intelligence and Intelligent Agent
Technology (WI-IAT). ACM, 2009, pp. 395–402.
[18] S.-Y. Hwang, H. Wang, J. Tang, and J. Srivastava, “A probabilistic
approach to modeling and estimating the qos of webservices-based
workflows,” Information Sciences (ISCI), vol.
177, no. 23, pp. 5484–5503, 2007.
[19] A. Bellucci, V. Cardellini, V. D. Valerio, and S. Iannucci, “A
Scalable and Highly Available Brokering Service for SLA-
Based Composite Services,” in Int. Conf. on Service Oriented
Computing (ICSOC). Springer, 2010, pp. 527–541.
[20] P. Leitner, W. Hummer, and S. Dustdar, “Cost-Based Optimization
of Service Compositions,” IEEE Transactions on
Services Computing (TSC), 2011, prePrint.

Research Article
Mindfully resisting the bandwagon:
reconceptualising IT innovation assimilation
in highly turbulent environments
Martin Wolf, Roman Beck, Immanuel Pahlke
Institute of Information Systems, Goethe University Frankfurt, Frankfurt am Main, Germany
Correspondence:
M Wolf, Institute of Information Systems, Goethe University Frankfurt, Grüneburgplatz 1, Frankfurt am Main 60323,
Germany.
Tel: þ 49 (0)69 798 33862;
Fax: þ 49 (0)69 798 33910;
E-mail: mwolf@wiwi.uni-frankfurt.de
Abstract
Environmental turbulence (ET), as exemplified by the recent financial crisis between 2007
and 2009, leads to a high degree of uncertainty, and fosters mimicry and resulting
bandwagon phenomena in information technology (IT) innovation assimilation processes.
In these highly turbulent environments, ‘mindless’ IT innovation assimilation by
participating organizations plays a major role in the manifestation and facilitation of
mimetic influences. Even in less turbulent economic cycles, highly turbulent industries
such as the financial services industry have to deal with demanding IT innovation
assimilation processes, and are exposed to varying levels of ET and mimicry. Drawing
upon the theory of dynamic capabilities, organizational mindfulness (OM) is one viable
means to mitigate the potentially negative consequences of mimetic behaviour. Here,
mindful organizations are more successful in overcoming situations of high dynamism, and
sometimes are even able to exploit them. So far, little empirical research has been
conducted to quantify the influence of OM in scenarios of high dynamism and mimicry. On
the basis of 302 complete responses from senior IT managers in the financial services
industry from the Anglo-Saxon countries (the United States, Canada and the United
Kingdom), this research contributes to a deeper understanding of the interaction of OM
with institutional pressures against the background of ET.
Journal of Information Technology advance online publication, 5 June 2012;
doi:10.1057/jit.2012.13
Keywords: organizational mindfulness; IT innovation assimilation; institutional theory; bandwagon
phenomena; IT business value; environmental turbulence
Introduction
Economic scenarios of high dynamism and volatility,
such as the financial crisis between 2007 and 2009,
demand a firm’s continuous technological adaptation
to retain a competitive position and comply with regulatory
influences (Barua et al., 2004; Sugumaran et al., 2008; Dong
et al., 2009). In particular, scenarios of high uncertainty
resulting from above-average volatility facilitate the emergence
of mimicry and bandwagon phenomena among
competing organizations that, in turn, might eventually
negatively affect the realization of business value (Fiol
and O’Connor, 2003; Swanson and Ramiller, 2004). For
instance, uncertainty resulting from incomplete information
on future market developments most likely leads to a
Journal of Information Technology (2012), 1–23
& 2012 JIT Palgrave Macmillan All rights reserved 0268-3962/12
palgrave-journals.com/jit/
situation where organizations inconsiderately follow a
so-called ‘best practice approach’. By doing so, they may
join an information technology (IT) innovation assimilation
bandwagon generated by prior adopters. In such
situations, organizations tend to justify their decisions with
the consensus of the ‘herd’, rather than an environmentally
aligned and customized-IT innovation strategy (Fiol and
O’Connor, 2003). In this context, organizational mindfulness
(OM) is assumed to be an effective means to identify
and accommodate changes facilitated by the market, as well
as to actively resist bandwagon phenomena that might
otherwise negatively affect the generation of (IT-induced)
business value (Fiol and O’Connor, 2003; Swanson and

2
Mindfully resisting the bandwagon M Wolf et al
Ramiller, 2004; Butler and Gray, 2006). Bandwagons are
defined as diffusion processes that are reflected by the
individual or organizational adoption of an idea, technique,
technology or product solely as a result of the number of
organizations that have already adopted it (Abrahamson
and Rosenkopf, 1990).
As one specific instance of a complex and thus
demanding IT innovation, the study at hand analyses the
assimilation of Grid-based architectures (Foster and
Kesselman, 1999; Buyya et al., 2009). In essence, Grid-based
architectures serve to meet the volatile IT resource and
IT service demands of organizations in highly turbulent
environments (Hackenbroch and Henneberger, 2007).
Following the seminal definition by Foster (2002), a Grid
is a system that coordinates IT resources that are not
subject to centralized control; uses standards, open
protocols and interfaces; and delivers non-trivial qualities
of service. Accordingly, Grid computing enables heterogeneous
and geographically dispersed IT resources to be
virtually shared and accessed across an industry, organization
or workgroup. Increasingly, large-scale enterprise
applications are no longer running on dedicated, centralized
computing facilities. Instead, they operate on heterogeneous
Grid resources that may span multiple
administrative units across different locations within an
organization (Strong, 2005). Depicting the continuous
growth of the associated market for virtualized IT resource
services, the market research company IDC (2010) projects
that, with regard to revenues, the market will grow by
27.4% per annum up to 2014. The report estimates that
worldwide revenues from virtualized IT resource provisioning
will grow from US$16 billion in 2009 to $55.5 billion
in 2014 (IDC, 2010). Recently, the core concepts of Grid
computing have transitioned to the domain of Cloud
computing (Foster et al., 2008; Weinhardt et al., 2009),
which represents a paradigmatic change in the dynamic
provisioning of IT resources (Gartner, 2010).
In general, IT innovations can be classified, according to
Swanson (1994), into Type I innovations, which are purely
technologically driven (e.g., database systems), and Type II
innovations, which involve the technological support of
administrative tasks (e.g., payroll or human resources
systems). Since Grid-based architectures will eventually
foster the development of advanced data processing, data
mining and (inter-)organizational collaboration capabilities,
they have the potential to generating strategic value
through integration with firms’ core business model. They
can therefore be classified as Type III innovations, which
potentially can affect the entire business strategy, and thus
are especially susceptible to bandwagon phenomena
because of their strategic complexity. Despite their strategic
potential, Type III innovation studies are still rarely
covered by the existing literature. In particular, there has
been little empirical research to explicate and quantify the
interplay of mimetic pressure (MP) and resulting bandwagon
phenomena influencing IT innovation assimilation that
stem from environmental turbulence (ET) (Fiol and
O’Connor, 2003; Pavlou and El Sawy, 2006). Furthermore,
few studies have focused on dynamic capabilities, such as
OM, which help to mitigate negative consequences of
bandwagons. OM is defined by a firm’s ‘rich awareness of
discriminatory detail and a capacity for action’ (Weick
et al., 1999:37). OM is one means by which organizations
can successfully overcome critical events, and master
scenarios of high uncertainty, such as complex IT innovation
processes. Such processes exhibit a high probability
of negative consequences, such as significant monetary
losses.
In the context of IT innovation assimilation, OM is
assumed to foster the identification of and resistance to
purely mimetic IT assimilation behaviour (Fiol and
O’Connor, 2003; Swanson and Ramiller, 2004). Moreover,
as an outgrowth of a dynamic capability, OM might help
organizations to cope better with ET and the resulting
high uncertainty (Weick and Sutcliffe, 2007), eventually
leading to above-average IT-based business value generation.
Despite several calls to integrate the interplay between
OM and MP in IT innovation assimilation processes
(Fichman, 2004; Swanson and Ramiller, 2004), to the best
of our knowledge this issue has not been addressed so
far. Accordingly, our research conceptualizes this interplay
as one instance of the institutional pressures driving the
assimilation of IT innovation and OM against the background
of a highly turbulent environment, such as the
recent financial crisis. The research model draws on
institutional theory to account for mimetic influences,
and on the theory of the dynamic capabilities of the firm to
integrate OM.
The remainder of the paper is structured as follows.
First, the research questions and their relevance are
introduced in the context of the overall research model.
Second, the theoretical background that informed the
research model is depicted, and the related hypotheses
are introduced. Then the research method used for the
field study conducted among senior Anglo-Saxon IT
decision makers and the results of partial least squares
(PLS) analyses of 302 complete responses are depicted.
Finally, the paper concludes by illustrating the contributions
of this analysis, and highlighting some further
research opportunities.
Interplay between institutional pressure and mindfulness in
highly turbulent environments
Our study analysed the Anglo-Saxon financial services
industry during the 2007–2009 financial crisis, as a recent
example of a highly turbulent environment. During this
extraordinary period the financial industry exhibited a high
extent of market volatility and uncertainty about further
market developments. This uncertainty resulted from rapid
changes in the market (e.g., rapid developments and
uncertainties in the corporate bond, mortgage and derivative
markets), and from concurrent technological demand
(resulting from new regulatory requirements, such as the
Basel II accord), which are subsumed by the concept of
ET (Pavlou and El Sawy, 2006). During this period, the
highly turbulent market required organizations and financial
services providers in particular, to assimilate IT
innovations that were suitable to deal with these rapid
changes. Concurrently, 172 US American banks failed,
resulting in systemic bandwagon phenomena that exacerbated
the extent of uncertainty and ET (FDIC, 2012). Even
in times of low market volatility, the financial services
industry is exposed to an above-average extent of

uncertainty, eventually leading to mimicry and bandwagon
phenomena among competitors (Ang and Cummings,
1997; Zhu et al., 2004). This uncertainty about future
market developments and current market conditions might
have serious negative effects on the generation and
realization of business value from IT innovation assimilation.
For instance, the adoption of an inadequate riskmanagement
system might restrict the extent of regulatory
compliance and thus impair the generation of IT-based
business value. Consequently, our first guiding research
question is:
RQ1: How does environmental turbulence affect the
influence of mimetic pressure and the realization of
business value from IT innovation assimilation?
Even in highly turbulent environments, some firms
overcome related challenges more effectively than their
competitors, or are potentially able to exploit them. This
significant difference in IT business value generation can be
attributed to advanced (dynamic) capabilities to align the
IT innovation assimilation process to environmental contingencies.
Organizations that exhibit enhanced dynamic
capabilities (i.e., OM) can identify impending changes in
the market earlier, and are able to derive highly contextualized
IT innovation strategies. In addition, they can
forecast and evaluate the consequences of the ensuing
bandwagon phenomena. The second guiding research
objective was therefore to assess the differences between
rather mindful and less mindful firms in channelling MP to
IT innovation assimilation processes against the background
of ET. To achieve a holistic perspective on IT-based
business value generation, we explicitly focused on the
whole assimilation process, from initiation to routinization
(Fichman, 2001), and thus omitted the non-assimilating
firms, which had no expertise of Grid-based infrastructures.
Consequently, we explore:
RQ2: How do rather mindful and less mindful firms
differ from each other with regard to the influence of
mimetic pressure, and to the impact of environmental
turbulence on IT-based business value generation stemming
from its assimilation?
Mimetic Pressure
(MP)
Coercive Pressure
(CP)
Normative Pressure
(NP)
Mindfully resisting the bandwagon M Wolf et al
significant influence for both groups
stronger influence for the more mindful
than the less mindful group
stronger influence for the less mindful
than the more mindful group
H1, H3
Figure 1 IT innovation assimilation in highly turbulent environments.
H2
Top Management
Support (TMS)
To address RQ1 and RQ2, the paper proposes the
research model depicted in Figure 1.
The required dynamic capabilities encompass reasoning
to actively resist prevailing bandwagon phenomena that
stem solely from MP, rather than from rational decisionmaking.
As far as the mediating agencies are concerned, we
conceptualize the influence of institutional pressure on top
management support (TMS) as the core human agency for
channelling MP on the IT innovation assimilation process
(Liang et al., 2007) against the moderating influence of ET.
On the basis of 302 complete responses from organizations
operating in a highly turbulent industry (i.e., the Anglo-
Saxon financial services industry), our group comparisons
of rather mindful and less mindful organizations indicate
that OM indeed has a mitigating impact on the influence of
MP and bandwagon phenomena in such an environment.
Moreover, rather mindful organizations are able to benefit
from ET, whereas less mindful firms are negatively affected
with regard to business process performance (BPP).
ET as driver of MP in IT innovation assimilation initiatives
ET encompasses rapid environmental changes resulting
from technological developments and changing market
preferences (Pavlou and El Sawy, 2006; Buganza et al.,
2009). Such changes eventually lead to uncertainty and
unpredictability in market demand, consumer requirements
and competitor strategies (Jap, 2001). Manifestations
of ET (Buganza et al., 2009) eventually foster the emergence
of uncertainty and mimicry among organizations, because
of incomplete information about further market developments
(Ang and Cummings, 1997; Fiol and O’Connor,
2003). In order to mitigate such uncertainty, organizations
are prone to imitate the behaviour of successful competitors
in terms of their strategies for IT innovation
assimilation (Swanson and Ramiller, 2004; Pavlou and El
Sawy, 2006). Institutional theory provides one perspective
to account for the rise of mimicry in highly turbulent
environments (Meyer and Rowan, 1977; Ang and
Cummings, 1997; Zhu et al., 2004; Liang et al., 2007).
According to DiMaggio and Powell (1983), three different
types of institutional pressure can be distinguished:
mimetic, coercive and normative. In essence, a highly
Environmental
Turbulence (ET)
H5
Grid Assimilation
(ASSM)
H4
Controls
Grid Infrastructure Capabilities (GIC)
Grid Technology Integration (GTI)
Earliness of Grid Adoption (TIME)
Firm Size (SIZE)
Business Process
Performance (BPP)
3

4
Mindfully resisting the bandwagon M Wolf et al
turbulent environment, such as the financial services sector,
makes an industry susceptible to mimetic behaviour (Fiol
and O’Connor, 2003; Swanson and Ramiller, 2004). Thus,
even if the consequences and benefits of an IT innovation
are poorly understood, MP fosters its assimilation, if
adopting firms are perceived as successful. By contrast,
coercive pressure (CP) arises from societal expectations in
a broader sense, where firms have to conform to expectations,
policies or regulation from government, from
customers or from the competitive environment. For
example, one consequence of the financial crisis was tighter
regulation of the financial services industry. Finally,
normative pressure (NP) arises from the ongoing process
of professionalization, which is further enforced by close
collaboration with suppliers, business partners and governmental
promotion. Since the focus of our study is on the
interplay between mimicry and OM with regard to IT-based
business value generation, we conceptualized normative
and CPs as controls to account for other confounding
institutional influences.
In our case, mimicry is of special importance, particularly
since organizations in the financial sector are exposed
to an uncertain and hyper-competitive environment
(Ang and Cummings, 1997). Consequently, structural
and behavioural changes in firms are driven equally by
an inherent organizational need for legitimacy and by
considerations of competitive advantage and hidden
efficiency potentials (Meyer and Rowan, 1977; DiMaggio
and Powell, 1983). In some scenarios, information on the
number and type of first adopters might outweigh the
importance of the characteristics of the actual innovation
(Meyer and Rowan, 1977; Aldrich and Fiol, 1994). Under
the high-ambiguity conditions that prevail in highly
turbulent environments, bandwagon pressure or MP is
likely to be higher (Rosenkopf and Abrahamson, 1999).
In the long run, this is likely to lead to organizational
isomorphism and homogeneity (Meyer and Rowan, 1977;
DiMaggio and Powell, 1983).
Accordingly, we hypothesize:
H1: Mimetic pressure drives top management to support
IT innovation assimilation.
H2: A higher extent of environmental turbulence
strengthens the influence of mimetic pressure on top
management support for IT innovation assimilation.
OM as dynamic capability to accommodate conditions of high
uncertainty
The success of organizations that achieve and sustain a
superior position in highly turbulent markets can be
attributed, in particular, to their timely responsiveness
(Teece et al., 1997). In this context, responsiveness can be
characterized by the capability of continuous product
innovation, complemented by management capability to
coordinate and redeploy internal and external competences
effectively with regard to identified changes in the
environment (Teece et al., 1997). These characteristics are
subsumed as part of the concept of dynamic capabilities
(Teece et al., 1997), which captures a firm’s ability to tailor
decisions and actions to environmental conditions, and
identify mutual interdependencies (Lawrence and Lorsch,
1967).
In general, dynamic capabilities are defined as the
‘capacity to renew competences so as to achieve congruence
with the changing business environment’ and the ability to
‘appropriately adapt, integrate, and reconfigure internal
and external organizational skills, resources, and functional
competences to match the requirements of a changing
environment’ (Teece et al., 1997:515). These capabilities
are critical in highly turbulent environments (Ang and
Cummings, 1997). Here, accurate perception of the
environment facilitates a better fit between decision making
and the organizational context (Tenbrunsel et al., 1996).
Moreover, dynamic capabilities foster the ability to identify
new opportunities, to (re-)design business processes
effectively and efficiently to foster business value generation
(Teece et al., 1997), and to resist bandwagon
phenomena that are inappropriate for the organizational
context (Swanson and Ramiller, 2004). Highly turbulent
environments, as exemplified by the financial services
industry during the financial crisis, are characterized by a
high extent of uncertainty, which demands a complex
sense-making capability to safeguard organizational
performance (McGill et al., 1993). Basically, the microfoundations
of these sense-making capabilities can be
attributed to scanning and interpretation activities (Thomas
et al., 1993). Scanning is defined as ‘searching the external
environment to identify important events or issues that
might affect an organisation’ (Thomas et al., 1993:241). The
scanning process is particularly critical for top management,
which is usually exposed to more information than it can
process in its decision making (Mintzberg, 1973). For
instance, top management has to anticipate the consequences
of a technological paradigm shift, such as the
on-demand IT resource provisioning facilitated by Grid
computing. In contrast, interpretation is defined as ‘development
or application of ways of comprehending the
meaning of information – it entails the fitting of information
into some structure for understanding and action’ (Thomas
et al., 1993:241). In the context of this research, interpretation
involves the derivation of IT innovation strategies
that maximize utilization of the benefits of Grid computing
for the assimilating firm, and foster awareness of the
inherent risks.
In a first step, accurate and discriminant perception
(scanning) and mindful evaluation (interpretation) of the
environmental conditions explicitly allows for cognitive
dissonance, which is assumed to be a core prerequisite for
resisting bandwagon phenomena (Festinger, 1967; Fiol and
O’Connor, 2003). Cognitive dissonance allows decision
makers to notice contradicting information that might shed
different light on future and past decisions (Festinger,
1967). One concept that particularly explicates this
cognitive state of heightened awareness and nuanced
decision making is that of individual mindfulness (Langer,
1989; Langer and Moldoveanu, 2000), which was eventually
transferred by Weick et al. (1999) to the organizational
level of high-reliability organizations (HROs). Analogous to
financial services providers, these HROs, such as nuclear
power plants and naval aircraft carriers, have to deal with
unexpected events in an environment characterized by
extreme turbulence, where error is omnipresent, and is

Mindfully resisting the bandwagon M Wolf et al
most likely to have far-reaching consequences. Thus,
these idiosyncratic requirements demand a sense-making
capability that fosters continuous alignment with environmental
conditions (Johns, 2006). Therefore, such organizational
sense-making capabilities are not only important for
the HRO domain, but also enable firms to stay efficient in
highly turbulent environments, such as the financial
market. During economic crises, a high extent of uncertainty
and turbulence is channelled into the organizational
IT innovation assimilation process (Fiol and
O’Connor, 2003; Butler and Gray, 2006). Here, contextual
information has to be gathered and aligned with the
behaviour and decisions of competitors. In these scenarios,
a mindful evaluation of contextual information improves a
firm’s ability to resist or intentionally follow potential
bandwagon phenomena to improve the business value
gained from IT innovation assimilation (Fiol and O’Connor,
2003). In sum, prior research has conceptualized mindfulness
as both a cognitive capability of non-algorithmic
thinking that can be consciously influenced by external
stimuli, such as cognitive training sessions (Langer and
Moldoveanu, 2000), and a static cognitive style that is stable
and immutable (Sternberg, 2000). In the following, we
assume that mindfulness is a cognitive capability that can
be actively stimulated, and thus either evolves or deteriorates
over time. According to the conceptualization of Weick
and Sutcliffe (2007), OM is formed of five complementary
cognitive dimensions that reflect the sub-dimensions of
dynamic capabilities (i.e., scanning and interpretation):
preoccupation with failure, reluctance to simplify, sensitivity
to operations, commitment to resilience and deference to
expertise.
Preoccupation with failure defines a firm’s ability to learn
from experiences in close-call situations, and to encourage
proactive reporting and definition of mistakes (scanning).
Consequently, it improves the firm’s learning capability,
which is essential in crises (interpretation). The firm’s
ability to ground its decision making on a more complete
and nuanced picture of its operations, rather than draw
from existing categories or solutions without considering
contextual characteristics, is encompassed by its reluctance
to simplify. Firms that exhibit a high extent of reluctance
tosimplifyaremorelikelytoconsider(scanning) and
integrate contradictory findings in their reasoning, and
thus explicitly allow the interpretation of past practices as
wrong (interpretation) (Fiol and O’Connor, 2003). In
addition, such firms are more likely to attend to details
based on organizational conditions (scanning), and exhibit
the ability to extract the value of the information with
respect to given circumstances (interpretation) (Fiol and
O’Connor, 2003). The firm’s attentiveness and situational
awareness to its operational front line are captured by its
sensitivity to operations. With a high degree of sensitivity to
operations, system anomalies, such as data anomalies,
can be isolated while they are still tractable. This leads to a
real-time processing of information (scanning) that can
subsequently be considered for further decisions (interpretation).
Commitment to resilience is defined by the
firm’s ability to detect, contain and bounce back from
inevitable errors to a dynamically stable state that facilitates
the reconfiguration and transformation of business
processes. Firms that have a well-developed commitment to
resilience are more likely to learn through experimentation
(scanning), and are thus able to interpret unusual and
unexpected results in a more nuanced way (interpretation)
(Fiol and O’Connor, 2003). Finally, the firm’s preference for
systematically delegating decision-making processes to the
most experienced employee, regardless of his or her hierarchical
rank, defines the degree of deference to expertise.
In essence, OM eventually leads to a state of high
situational awareness and self-control. Accordingly, we
propose that OM helps to actively resist bandwagon
phenomena (i.e., MP) in situations where mimicry is not
advantageous for the assimilating firm, and to make
effective sense of a high degree of ET. Consequently, we
hypothesize:
H3: Rather mindful organizations are less affected by
the impact of mimetic pressure on the IT innovation
assimilation process compared with less mindful
organizations.
Consequences of mindful IT innovation assimilation at the
business process level
In contrast to the isolated decision to invest in an IT
innovation (adoption), IT innovation assimilation is the
continuous (and thus more holistic) organizational transition
from a stage of technological initiation to a stage of
routine utilization of the adopted IT innovation (Fichman,
2001). Here, IT innovation assimilation is a process that
can be characterized by seven stages (Fichman, 2001): in
Stage 1 (initiation), an organization becomes aware of an IT
innovation. The organization then decides to invest in this
IT innovation (Stage 2: adoption), and to implement it
(Stage 3: implementation). In successful IT innovation
assimilation, the IT innovation is actively used in business
processes (Stage 4: occasional usage to Stage 7: routinized
usage).
A business process level perspective was chosen as the
unit of analysis for conceptualizing the progress of IT
assimilation and business value generation, since IT
investments are supposed to affect, first, the performance
of specific business processes (Davamanirajan et al., 2006).
Prior research suggests that a firm encompasses 18 key
processes that are crucial for overall firm performance
(Davenport, 1993). In order to identify the key business
processes that are primarily influenced by Grid assimilation
(ASSM) in the financial services industry, several expert
interviews were conducted with IT executives. The interviews
revealed that asset management, risk management
and new product development are especially crucial
processes for the financial services industry in this context.
For instance, in 2008 Bank of America’s asset management
process accounted for 10% of revenue but contributed
35% to the net income (SEC, 2008), with growing projected
potential for the future. Grid-based infrastructures foster
assessment of the performance of asset allocation decisions
in a more timely manner, thanks to concurrent processing
(Hackenbroch and Henneberger, 2007). Further evidence of
the relevance of the risk management process was provided
by the Risk Management Association (RMA) in 2010.
Among 75 global firms being interviewed by the RMA,
56.8% rated the data quality of their risk management
5

6
Mindfully resisting the bandwagon M Wolf et al
process as average or worse (RMA, 2010). The perception of
the risk management professionals was that an improved
data quality and risk management process would lead to a
more timely identification of emerging problems, and more
efficient capital allocation and utilization. Accordingly, the
risk management process is essential, and vital in improving
scanning capabilities, driven mainly by: (1) the pressure
from regulators for better control of financial risks; (2) the
globalization of financial markets, which has led to
exposure to more sources of risk; and (3) technological
advances that have made enterprise-wide risk management
possible (Jorion, 2006). Grid-based infrastructures enable
organizations to conduct more sophisticated risk management
calculations, which lead to a gain of economic capital
for follow-up businesses (Hackenbroch and Henneberger,
2007). The need for continuous enhancement of the new
product development process as a vital response capability
is created mainly by fast-changing customer needs, which
force financial services providers to provide highly customized
financial products on demand. In such demanding
environments, mindful organizations are able to realize
more business value from IT innovation assimilation
(Swanson and Ramiller, 2004). Mindful decision makers
are able to efficiently identify growth options (to realize
growth from an IT investment), deferral options (to
postpone an IT investment until there is more information),
learning options (to build on information gained
from an initial investment) and staging options (to
structure an IT innovation assimilation process into more
manageable, separate process steps), which facilitate the
realization of business value from IT innovation assimilation
(Goswami et al., 2008). Here, OM manifests as
contextually nuanced reasoning that facilitates the development
of both an expanded scanning capability and an
interpretation capability with regard to the utilized IT
innovation (Goswami et al., 2008). Mindful organizations
will therefore gather more relevant information, and are
likely to be in a better position to recognize the various
options that arise from IT innovation assimilation. In
addition, OM eventually leads to an improved IT business
alignment with regard to the fit of the IT innovation and
overarching business objectives, and thus contributes to
firm performance (Valorinta, 2009). Consequently, we
hypothesize:
H4: Rather mindful firms are able to realize a higher
extent of business process performance resulting from IT
innovation assimilation in the risk management, asset
management and new product development process than
less mindful firms.
As an analytical perspective, real options pricing is
especially appropriate for assessing IT innovation investment
decisions under conditions of high uncertainty (Dixit
and Pindyck, 1994), such as those that prevailed during the
financial crisis. Mindfulness here facilitates differentiated
reasoning with regard to the options that arise from IT
innovation assimilation, based on the organization’s own
facts and specifics (Goswami et al., 2008). In highly
turbulent environments, the increased scanning and
interpretation capabilities that characterize mindfulness
lead to entrepreneurial alertness and the generation of
digital options that, in turn, eventually enhance the
generation of competitive actions and above-average firm
performance (Sambamurthy et al., 2003). Hence, we
hypothesize the following:
H5: Rather mindful organizations are able to benefit
from a high extent of environmental turbulence, whereas
less mindful organizations are likely to be negatively
affected by it.
Controls
IT innovation assimilation processes are subject to
various other organizational influences (Fichman, 2001;
Zhu et al., 2006b). To minimize the confounding impact of
spurious correlation, we included Grid infrastructure
capability (GIC) and Grid technology integration (GTI)
adapted from Zhu et al. (2006b), firm size (SIZE), and
earliness of Grid adoption (TIME) measured by years
elapsed since first adoption (Fichman, 2001), as control
variables for ASSM as our instance of an IT innovation and
the realized business value through process performance
improvements to account for differences among financial
services providers.
GIC captures the firm’s technical capability resulting
from its having access to distributed computing power and
purpose-specific technologies (e.g., a high-capacity, lowlatency
network) within the firm. GTI refers to a set of
investigation, evaluation and refinement activities aimed at
creating a match between technological options and the
application context (Iansiti, 1998). Consistent with the prior
literature (e.g., Zhu and Kraemer, 2005; Zhu et al., 2006a),
we assume that higher levels of GIC and GTI drive ASSM
positively in the three identified key business processes.
Moreover, higher levels of GIC and GTI constitute a vital
prerequisite to benefiting from Grid infrastructure, resulting
in performance improvements.
Prior studies suggest that smaller firms are more flexible
with regard to innovative technologies (Zhu and Kraemer,
2005), but there is also an opposite theoretical view, based
on the fact that larger firms are assumed to exhibit slack
resources, which facilitate innovation diffusion (Tornatzky
and Fleischer, 1990; Rogers, 1995). However, in general we
assume that at least a certain size of firm is required for a
Grid infrastructure to be implemented and utilized in a
reasonable manner, since there has to be a significant
number of IT resources (e.g., servers) that can be
interconnected and virtualized. Consequently, we assessed
responses only from financial services providers with more
than 1000 employees. Among these, we hypothesize that the
smaller firms are more likely to assimilate and profit from
Grid technology, thanks to their higher openness to
innovation. Finally, Grid experience reflects the fact that
firms that initiated Grid implementation activities earlier
had more time to reach later stages of IT innovation
assimilation and create potential benefits (the so-called
time lag effect).
Empirical study
In order to validate the research model, we conducted a
questionnaire-based quantitative field study. The study was
aimed at senior IT decision makers working for financial

Mindfully resisting the bandwagon M Wolf et al
institutions located in the Anglo-Saxon countries (the
United States, Canada and the United Kingdom) with more
than 1000 employees. The research model was analysed
using PLS, a component-based structural equation modelling
(SEM) technique that concurrently assesses the
psychometric properties of the measurement scales and
the strength of the hypothesized relationships. We deemed
this method appropriate for our research for several
reasons. First, PLS handles measurement errors in exogenous
variables better than other methods, such as multivariate
regression (Chin, 1998). Second, component-based
SEM approaches require fewer distributional assumptions
with regard to the sample data (Cassel et al., 1999); in the
early stages of measurement instrument development and
theory testing, in particular, little is known about the
distributional characteristics of the observed variables.
Third, PLS can accommodate both exploratory and
explanatory analyses, which is particularly suitable for
our research context. Although PLS is often used for theory
testing, it can provide first evidence for new relationships
that are relevant for subsequent in-depth testing (Chin
et al., 2003). Generally, the PLS approach is predictionoriented
(Chin, 1998) and estimates latent variables as exact
linear combinations of the observed measures (Wold,
1982), which is advantageous, because theory generation
is as important as theory testing.
Data collection and sample profile
A survey instrument was developed to collect the quantitative
data required for model and hypothesis testing. On
behalf of the authors, 2866 potential participants of a
business panel were invited by a large international market
research company to respond to the survey from August
until September 2009. The date of invitation, the date of
participation and the user ID were recorded to ensure that
each panellist completed the online survey only once. In
total, 782 responses were returned, indicating a response
rate of 27.3%. Since the study was aimed at Gridassimilating
organizations, the study participants were
asked to indicate whether or not they had already
assimilated Grid technology in at least one of the analysed
processes. Responses from non-Grid assimilators, and
responses that exhibited missing values, were removed.
Thus, of the 782 completed responses, 349 were non-Grid
assimilators and 131 questionnaires exhibited missing
values and thus were excluded, which resulted in a sample
of 302 complete responses.
With respect to our research questions, the limited focus
of our study on assimilating organizations is deemed
appropriate, for the following reasons. First, the study at
hand seeks to improve our understanding of the determinants
of the assimilation stages of complex (i.e., Type III) IT
innovations with respect to different levels of ET, interrelated
institutional pressures and dynamic capabilities, such as
OM. According to this research goal, we investigate the
impact of variations in the assimilation stages of an IT
innovation (e.g., Zhu and Kraemer, 2005; Liang et al., 2007)
on BPP. We thereby follow the reasoning of Teo et al.
(2003) that isomorphic processes influence all assimilation
stages equally and eventually lead to further advance of
organizations in the process. Second, assessing the impact of
a Type III innovation on the process level (as our unit
of analysis) requires an approach that focuses on how
technology characteristics enhance specific business processes
(Davamanirajan et al., 2006). Thus, only those firms
that already had made sense of this technology (i.e., were at
least already in the initiation stage), and had decided how
to draw upon corresponding functionalities, were assumed
to be able to realize benefits for the related business
processes. In essence, we assume that Grid-based IT
infrastructures have the potential to meet the volatile IT
resource and service demands of organizations in highly
turbulent environments. Consequently, we conducted a
group comparison analysis, based on key performance
indicators, between organizations that had and had not
assimilated Grid-based infrastructures. In particular, we
included an additional question in our survey asking for
the name of the respondent’s firm. On the basis of this
information, we were able to extract different key
performance indicators for over 110 North American
Organizations (including 39 non-assimilators and 71
assimilators) from the BankScope database (30,000 companies
listed) and the COMPUSTAT database (from
Wharton Research Data Services; 25,000 companies listed).
These databases encompass an extensive set of banks’
financial statements, ratings and intelligence reports.
For empirical investigation of the groups of assimilators
and non-assimilators, we conducted several one-way
analyses of variance (ANOVA), since this is a common
approach to test for differences in indicators’ mean values
(Box et al., 2005). In response to the need for statistical
robustness, we also consider the results of Kruskal–Wallis
rank-sum tests as the non-parametric version of ANOVA,
and a generalized form of the Mann–Whitney approach
(Kruskal and Wallis, 1952). The results depicted in
Table A6 in the Appendix suggest that, at a 10% level of
significance, firms that had assimilated Grid-based infrastructures
for at least one of the identified key business
process (i.e., asset management, risk management and new
product development) were good as, or even better than,
non-assimilating organizations with respect to asset quality,
capital adequacy, profitability and efficiency and other
liquidity indicators. Accordingly, we assume that nonassimilators
have no business purpose for utilizing such
an IT infrastructure, or they are not mindfully unaware of
what the technology has to offer.
Further details of the sample profile are depicted in
Table 1.
In the data-gathering stage, we followed a key informant
approach (Bagozzi and Phillips, 1991) to collect data on
Grid-related business processes of financial services providers.
In IS research, this approach is often used to gather
information on organizational factors, especially in the
context of the business value of IT (Tanriverdi, 2005).
However, if respondents and their context differ substantially
from those who do not respond, then the respondents
and the corresponding empirical findings may not be
generalizable to the population (Miller and Smith, 1983).
In order to address this issue of non-response bias, we
compared the first 25% of responses with the last 25% (Sivo
et al., 2006). Thereby, we follow Armstrong and Overton
(1977), who argue that people responding in later waves
can be assumed to be proxies for non-respondents. In this
7

8
Table 1 Sample profile
Mindfully resisting the bandwagon M Wolf et al
Country Number of employees
United States 189 (62.6%) 1001–5000 37 (12.3%)
Canada 10 (3.3%) 5001–10,000 39 (12.9%)
United Kingdom 103 (34.1%) 10,001–50,000 74 (24.5%)
50,000+ 152 (50.3%)
Respondent’s position Year of first Grid adoption
CTO, COO, CIO 43 (14.2%) o2000 18 (6.0%)
Chief systems architect 15 (5.0%) 2000–2001 20 (6.6%)
Other senior IT decision maker 244 (80.8%) 2002–2003 16 (5.3%)
2004–2005 38 (12.6%)
2006–2007 93 (30.8%)
2008–2009 117 (38.7%)
regard, we treated the first 25% of the responses received
as early respondents, and the last 25% of the responses as
late respondents. We conducted both ANOVAs (parametric)
and Kruskal–Wallis (non-parametric) tests to
compare the responses of early and late respondents for
all survey questions. We also compared firm differences
with respect to key performance indicators extracted from
archival data. The results depicted in Table A7 in the
Appendix indicate that, at a 10% level of significance,
almost no differences were found between the identified
organizations in terms of asset quality, capital adequacy,
profitability and efficiency or liquidity. In essence, the
findings indicate that the responses of early and late
respondents do not differ significantly, and thus indicate
no substantial influence of non-response bias in our study.
Research instrument: securing content validity
Overall, our research model encompasses five reflective
constructs (MP, CP, NP, TMS, ET) and two formative
constructs (ASSM, BPP). All measures were informed by
the extant literature, and were adapted to the Grid context
where necessary (see Tables A1–A3 in the Appendix). To
ensure content validity of the utilized measures, several
expert interviews were conducted, and the survey instrument
was made available to a panel of judges of both
practitioners and academics (Straub, 1989; Straub et al.,
2004).
For the ASSM construct, a 7-item Guttmann scale was
used to capture an organization’s current ASSM stage. This
scale was based on prior research on the assimilation of
software process innovations (Fichman, 2001), and on the
assimilation of electronic procurement innovations (Rai
et al., 2009). The respondents were requested to identify the
current stage of ASSM for their risk management, asset
management and new product development processes.
As already outlined, these three processes take particular
advantage of a high-performance Grid infrastructure,
and were identified as being especially appropriate and
vital for the financial services industry. Consequently, the
measurement items of the assimilation construct focused
on Grid-related activities in these processes.
BPP was conceptualized as a dependent variable for each
of the three key business processes, to capture the business
value generation momentum of ASSM. As Karimi et al.
(2007a) propose in the domain of enterprise resource
planning assimilation, process efficiency, process effectiveness
and process flexibility reflect the overall BPP
construct. Process efficiency reflects the extent to which
the use of IT innovation assimilation reduces operational
costs and decreases the input/output conversion ratio, and
process effectiveness is the extent to which IT innovation
assimilation provides improved functionality, and enhances
the quality of users’ work. The extent to which IT innovation
assimilation provides firms with more flexibility
in response to changing business environments defines
the process flexibility dimension. Since successful business
processes are characterized by the presence of
efficiency, effectiveness and flexibility, which are not
mutually interchangeable (Adler et al., 1999), we assume
that the sub-dimensions of the BPP construct co-vary to a
high extent (Karimi et al., 2007a, b). BPP was therefore
operationalized as a second-order reflective model, and a
two-stage approach was utilized (Yi and Davis, 2003) to
model and include this construct as a dependent variable.
The latent variable scores of the second-order construct
were extracted in an initial analysis for each of the three
identified business processes that were subsequently used
as formative indicators for the estimation of the overall
research model.
In order to distinguish between firms with respect to
their OM, we operationalized an OM score based on a
second-order, Type II model. We therefore started off with
wordings by Weick and Sutcliffe (2001) and Knight (2004),
which we subsequently refined and adapted to create an
initial pool of 25 items for the five distinct dimensions of
OM (see Table A4 in the Appendix). To ensure content
validity, we followed a two-staged approach. First, we
conducted four expert interviews, which resulted in minor
refinement, and the exclusion of two ambiguous items
(Straub, 1989; Straub et al., 2004). Second, we conducted
two rounds of unstructured and structured Q-sorting with
different participants (four doctoral students and four
experts from the financial services industry) (Moore and
Benbasat, 1991). There was strong inter-judge reliability,
and the required Cohen’s k of all constructs met the
criterion of 0.65 (see Moore and Benbasat, 1991). Finally,
we pre-tested the instrument before final rollout in a

small-scale setting with 37 respondents (2 chief operating
officers, 5 chief technology officers, 4 chief information
officers, 5 chief systems architects and 21 senior IT decision
makers). Insufficient construct reliability meant that we
had to drop 10 of the initial 25 (reflective) OM items.
Because of the reflective second-order model specification
(and thus assuming that items are interchangeable within
their associated dimension), we deemed the impact of the
dropped items on content validity as minor. Table A4 in
the Appendix presents the final measurement scales of the
different dimensions of OM.
Analysis and results
The results for the PLS estimation were obtained from
SmartPLS (Version 2.0 M3; Ringle et al., 2005). The
measurement and structural models were first evaluated
separately for the rather mindful and less mindful groups,
before the between-group comparisons were conducted.
Consistent with Chin (1998), we utilized a 500 bootstrap
sampling technique to test for the significance of the path
estimates, factor loadings and weights.
A variety of means for assessing between-group comparisons
exist, including ANOVA and moderated regression
(e.g., Sharma et al., 1981). In the realm of SEM, several
means for testing between-group differences in these
second-generation techniques have been developed (Qureshi
and Compeau, 2009). In our particular case, we assess the
between-group differences with PLS, as a component-based
SEM approach, by following the widely used approach
introduced by Chin (2000). This approach involves
estimating model parameters for each group separately,
using SmartPLS, and then performing a between-group test
of significant differences (e.g., Keil et al., 2000; Venkatesh
and Morris, 2000; Zhu et al., 2006a; Hsieh et al., 2008).
We partition the sample on the basis of a median split of
the calculated OM score (mean ¼ 5.29; median ¼ 5.41;
Table 2 Descriptive statistics, and validity and reliability criteria
Mindfully resisting the bandwagon M Wolf et al
SD ¼ 1.03), resulting in a group of less mindful organizations
(n ¼ 152) and a group of rather mindful organizations
(n ¼ 150). In order to test for significant differences
between the groups, we conducted two-sample t-tests on
the estimated coefficients obtained from the bootstrap
processes for each group (Chin, 2000; Sarstedt et al., 2011).
According to Chin (2000), this procedure for comparing
multiple groups with pairwise t-tests is subject to several
assumptions about the data and the measurement
model: (1) each nested model considered has to reach an
acceptable goodness of fit; (2) there should be measurement
invariance; and (3) the data should not deviate significantly
from normality. In order to address the first two issues
we examine the construct reliability and construct validity
for each nested model separately. The results are given in
Tables 2 and 3, and indicate that both nested models
achieve an appropriate goodness of fit, and support the
assumption of measurement invariance (i.e., no significant
differences of the item loadings for each construct). To test
for normal distribution of the estimations across all
bootstrapping samples we performed tests of normality
based on skewness and kurtosis (see Table A5 in the
Appendix) and visually inspected q–q plots (Chambers
et al., 1983; D’Agostino et al., 1990). Finally, we conducted
a Levene test statistic (Levene, 1960) to assess whether the
parameters’ standard deviations differed significant across
the groups, and then applied the appropriate t-test statistics
according to the definitions of Chin (2000) and Sarstedt
et al. (2011).
In response to the need for robustness in statistical
analysis, and with respect to the non-normality of some
estimations (see Table A5 in the Appendix), we also applied
the non-parametric bootstrapping approaches proposed by
Henseler et al. (2009), Sarstedt et al. (2011) and Henseler
(2012) for multigroup analyses in PLS. These approaches
estimate the probability of path differences based on the
outcomes of the bootstrap processes for each group fitting
OM high Mean SD AVE CR a MP CP NP TMS ET ASSM a BPP a
MP 4.90 1.24 0.89 0.96 0.94 0.94
CP 5.23 1.24 0.75 0.90 0.83 0.53 0.87
NP 4.88 1.22 0.74 0.89 0.82 0.04 0.71 0.86
TMS 5.65 1.08 0.80 0.95 0.94 0.19 0.36 0.40 0.89
ET 5.87 0.97 0.56 0.91 0.88 0.24 0.39 0.39 0.23 0.75
ASSM a 4.69 1.86 NA NA NA 0.31 0.33 0.43 0.41 0.22 NA
BPP a 4.21 2.08 NA NA NA 0.21 0.28 0.34 0.23 0.28 0.50 0.94
OM low Mean SD AVE CR a MP CP NP TMS ET ASSMa BPPa MP 4.44 1.02 0.83 0.94 0.90 0.91
CP 4.55 1.13 0.66 0.85 0.76 0.49 0.81
NP 4.16 1.01 0.65 0.84 0.72 0.15 0.32 0.81
TMS 4.22 1.27 0.80 0.95 0.94 0.33 0.29 0.25 0.89
ET 5.41 0.91 0.55 0.85 0.81 0.15 0.19 0.00 0.04 0.74
ASSM a
3.96 1.77 NA NA NA 0.15 0.13 0.21 0.33 0.04 NA
BPPa 3.04 2.12 NA NA NA 0.04 0.13 0.05 0.31 0.09 0.44 0.93
a Formative measure.
Mean value (Mean), standard deviation (SD), average variance extracted (AVE), composite reliability (CR), Cronbach’s a, correlations
among constructs (off-diagonal) and square roots of average variance extracted (diagonal).
9

10
Table 3 Item loadings and cross-loadings
Mindfully resisting the bandwagon M Wolf et al
OM high MP CP NP TMS ET ASSM a BPP a OM low MP CP NP TMS ET ASSM a BPP a
MP1 0.93 0.53 0.41 0.18 0.16 0.24 0.10 MP1 0.87 0.38 0.15 0.25 0.18 0.18 0.12
MP2 0.95 0.44 0.44 0.18 0.25 0.31 0.21 MP2 0.95 0.48 0.17 0.27 0.10 0.15 0.08
MP3 0.95 0.50 0.49 0.19 0.26 0.32 0.15 MP3 0.91 0.48 0.11 0.26 0.14 0.07 0.07
CP1 0.48 0.80 0.67 0.25 0.26 0.37 0.28 CP1 0.29 0.64 0.33 0.09 0.16 0.15 0.07
CP2 0.44 0.90 0.59 0.38 0.32 0.21 0.18 CP2 0.36 0.93 0.30 0.31 0.22 0.17 0.13
CP3 0.46 0.88 0.60 0.29 0.42 0.30 0.29 CP3 0.56 0.84 0.22 0.22 0.17 0.03 0.11
NP1 0.39 0.56 0.86 0.29 0.32 0.34 0.26 NP1 0.15 0.29 0.81 0.22 0.08 0.18 0.08
NP2 0.43 0.62 0.91 0.40 0.37 0.33 0.29 NP2 0.22 0.24 0.82 0.18 0.04 0.24 0.06
NP3 0.39 0.66 0.80 0.31 0.31 0.46 0.34 NP3 0.01 0.24 0.77 0.20 0.05 0.09 0.01
TMS1 0.09 0.28 0.29 0.87 0.20 0.35 0.24 TMS1 0.22 0.26 0.17 0.88 0.01 0.28 0.25
TMS2 0.11 0.33 0.38 0.91 0.25 0.34 0.25 TMS2 0.25 0.23 0.18 0.92 0.08 0.36 0.24
TMS3 0.23 0.39 0.38 0.91 0.27 0.36 0.18 TMS3 0.33 0.26 0.21 0.90 0.10 0.32 0.33
TMS4 0.14 0.29 0.34 0.94 0.18 0.39 0.20 TMS4 0.25 0.27 0.27 0.93 0.04 0.30 0.32
TMS5 0.28 0.33 0.35 0.84 0.12 0.41 0.17 TMS5 0.23 0.28 0.28 0.83 0.06 0.20 0.23
ET1 0.06 0.21 0.22 0.11 0.71 0.01 0.15 ET1 0.05 0.19 0.04 0.04 0.80 0.13 0.07
ET2 0.24 0.33 0.34 0.22 0.80 0.10 0.16 ET2 0.03 0.19 0.09 0.08 0.63 0.01 0.02
ET3 0.23 0.27 0.36 0.19 0.79 0.19 0.19 ET3 0.20 0.27 0.14 0.09 0.60 0.07 0.03
ET4 0.23 0.31 0.27 0.08 0.73 0.08 0.13 ET4 0.12 0.34 0.12 0.10 0.59 0.04 0.04
ET5 0.16 0.31 0.27 0.18 0.82 0.24 0.27 ET5 0.20 0.14 0.04 0.04 0.64 0.01 0.09
ET6 0.18 0.34 0.34 0.26 0.81 0.31 0.34 ET6 0.10 0.03 0.00 0.04 0.77 0.01 0.09
ET7 0.11 0.28 0.21 0.06 0.67 0.06 0.21 ET7 0.21 0.14 0.03 0.01 0.64 0.01 0.08
ASSM1a 0.22 0.24 0.32 0.38 0.13 0.80 0.58 ASSM1a 0.08 0.03 0.11 0.23 0.03 0.71 0.46
ASSM2 a
0.28 0.32 0.38 0.28 0.26 0.80 0.33 ASSM2 a
0.11 0.12 0.18 0.28 0.02 0.86 0.35
ASSM3a 0.26 0.26 0.37 0.35 0.17 0.86 0.31 ASSM3a 0.19 0.19 0.21 0.27 0.06 0.79 0.20
BPP1a 0.33 0.33 0.41 0.34 0.25 0.68 0.89 BPP1a 0.08 0.15 0.27 0.11 0.60 0.19 0.90
BPP2 a
0.20 0.30 0.37 0.26 0.31 0.46 0.69 BPP2 a
0.11 0.14 0.04 0.30 0.38 0.05 0.57
BPP3a 0.22 0.26 0.38 0.24 0.34 0.54 0.77 BPP3a 0.15 0.14 0.12 0.22 0.55 0.18 0.87
a Formative measure.
Bold values refers to the associated construct, e.g., CP1 to CP3 refer to the CP construct.
the PLS path modelling’s distribution-free characteristic.
In particular, the approach proposed by Henseler et al.
(2009) and Henseler (2012) reflects a Wilcoxon rank-sum
test applied to the bootstrap values corrected for the
original parameter values. In contrast, the approach
proposed by Sarstedt et al. (2011) investigates the overlapping
of bias-corrected bootstrap confidence intervals.
In particular, we follow the so-called percentile method
(Efron, 1981) for constructing appropriate bootstrapping
confidence intervals accounting for the distribution’s
skewness. We applied both non-parametric tests to the
subsamples of our study, and derived P-values that are
reported in Table 5 and are discussed in the corresponding
section.
Validation of the measurement model
In order to validate the measurement model, the psychometric
properties of all scales were assessed within the
context of the structural model by examining the construct
reliability and construct validity (Chin, 1998; Petter et al.,
2007).
Construct reliability refers to the internal consistency of
the measurement model (Straub et al., 2004), or the degree
to which items are free of systematic error, and yield
consistent results. As our results in Table 2 indicate, the
values of average variance extracted (AVE) of all constructs
are above the recommended threshold of 0.5 (Fornell,
1992). Consequently, more than 50% of measurement
variance is explained by the constructs. Moreover, aggregate
measures of the degree of inter-correlations among
measurement items, such as the composite reliability or
Cronbach’s a (a), exceed the recommended threshold of
0.7 (Nunnally, 1978; Hair et al., 1998), suggesting internal
consistency among the reflective measurement items. As far
as the formative constructs (ASSM, BPP) are concerned,
Table 4 shows that all weights are above 0.2, and almost
all coefficients are statistically significant (Chin, 1998),
indicating the relevance and consistency of the formative
indicators used to measure these constructs. To ensure
content validity, we decided to keep all indicators in our
sample for the subsequent analyses (Petter et al., 2007).
Furthermore, the estimated variance inflation factors (VIF)
for the ASSM and BPP items (see Table 4) are all below 3.3
(tolerance40.30), indicating no serious concern of multicollinearity
(Petter et al., 2007).
Construct validity captures whether the indicators of a
construct measure what they are supposed to measure,
from a psychometric perspective. Hence, the parameter
estimates of the relationships between constructs and their
indicators are assessed for consistency. Construct validity
encompasses an assessment of both convergent validity and

Mindfully resisting the bandwagon M Wolf et al
Table 4 Coefficients, t-values and VIF of exogenous variable TMS and formative measures ASSM and BPP
OM high Coefficient t-value VIF Tolerance OM low Coefficient t-value VIF Tolerance
MP 0.01 0.21 1.44 0.70 MP 0.23 2.64 1.40 0.71
CP 0.14 1.31 2.36 0.42 CP 0.12 1.38 1.68 0.60
NP 0.28 2.38 2.22 0.45 NP 0.14 1.82 1.17 0.85
ET 0.08 1.05 1.27 0.78 ET 0.01 0.11 1.09 0.91
ET MP 0.23 2.63 1.20 0.83 ET MP 0.32 4.83 1.10 0.91
ASSM1 0.24 1.62 1.64 0.61 ASSM1 0.27 2.01 1.77 0.57
ASSM2 0.28 1.93 1.56 0.64 ASSM2 0.50 2.66 1.73 0.58
ASSM3 0.67 2.85 1.43 0.70 ASSM3 0.46 2.22 1.15 0.87
BPP1 0.62 2.64 1.49 0.67 BPP1 0.53 2.37 1.73 0.58
BPP2 0.33 1.84 1.48 0.68 BPP2 0.19 1.54 1.64 0.61
BPP3 0.28 1.65 1.30 0.77 BPP3 0.48 2.29 1.24 0.81
discriminant validity (Campbell and Fiske, 1959). The test
for convergent validity is relevant only for reflective
measures (MacKenzie et al., 2005), and determines whether
the indicators of latent constructs that theoretically should
be related are actually observed to be related. Loch et al.
(2003) propose that the existence of significant interindicator
and indicator-to-construct correlations suggests
convergent validity; our results in Table 3 show that almost
all loadings of the reflective constructs are greater than the
recommended threshold of 0.707 (Chin, 1998), such that
there exists more shared variance between the construct
and its indicators than error variance, and the measurement
items used are adequate for measuring the assigned
constructs. Except for single items of the CP (low
mindfulness group) and ET (both groups) constructs, all
other items meet the recommended threshold. Consistently,
the results in Table 3 reveal that the items load significantly
higher on their own construct than on any other construct
(cross-loadings). Furthermore, the corresponding loadings
are higher than 0.5, suggesting the appropriateness of the
items (Hulland, 1999). Consequently, we deemed this issue
as minor for the measurement model quality.
To assess the discriminant validity of the model specification,
we tested whether indicators of latent constructs
that are theoretically unrelated are indeed unrelated
with regard to the parameter estimates. In this context,
MacKenzie et al. (2005) propose an approach appropriate
for evaluating the discriminant validity of formative and
reflective measures, which analyses whether the interconstruct
correlations are relatively low. The discriminant
validity for the reflective constructs can be assessed by (1)
analysing the cross-loadings and (2) assessing the Fornell–
Larcker criterion. The cross-loadings indicate that each
indicator exhibits a higher loading on its assigned construct
than on the other constructs (Henseler et al., 2009), see
Table 3. The Fornell–Larcker criterion (Fornell and Larcker,
1981) postulates that a construct has to share more variance
with its assigned indicators than with any other construct,
as assessed by the relationships between the inter-construct
correlations and the square roots of the AVE scores. From a
psychometric perspective, the square root of the AVE for
each construct should exceed the inter-construct correlations
involving the construct (Fornell and Larcker, 1981).
As depicted in Table 2, the square root of each AVE score is
11
greater than the correlations between the construct and any
other construct, which indicates satisfactory discriminant
validity.
Validation of the structural model
Since our analyses of the measurement model indicate
convergent and discriminant validity, and since all
indicators meet the established reliability and validity
criteria, the presented measures were used to test the
structural model and the associated hypotheses. The
moderating effects of ET were operationalized and estimated
following the procedure proposed by Chin et al.
(2003). Accordingly, we first reduced multicollinearity by
standardizing all indicators reflecting the predictor and
moderator constructs to a mean of 0 and variance of 1. This
step also allows for an easier interpretation of the path
coefficient for the predictor variable. The path coefficient
reflects the effect expected at the mean value of the
moderator variable, which is 0. Second, using the standardized
indicators of the predictor and moderator variables,
we generated product indicators to reflect the latent
interaction variables. Third, we utilized PLS to estimate
the dependent variable TMS, with interaction terms as
explanatory variables. Considering the potential high
inter-correlations among the main effects and interaction
terms, we also assessed the confounding influence of
multicollinearity statistically according to the VIF. As we
show in Table 4, the VIF scores of the explanatory variables
of TMS are lower than the recommended level of 3.3
(tolerance 4 0.30), which indicates the absence of multicollinearity
(Petter et al., 2007).
In Table 5, the validation results for the rather mindful
firms and less mindful firms are depicted. We compared
two nested models for the dependent variables ASSM and
BPP to check for robustness of our results: Model 1 was the
baseline model with the control factors only, and Model 2
was the full model with all main effects, interaction effects
and control effects. These models are fully nested, so that
the difference in explanatory power allows a valid model
comparison in terms of effect sizes. The explanatory power
of the structural model is measured by the squared
multiple correlations (R 2 ) of the dependent variables (Chin
et al., 2003).

12
Table 5 Empirical results
Relationship OM low (n ¼ 152) OM high (n ¼ 150) Group comparison
Basic model
(controls only)
Mindfully resisting the bandwagon M Wolf et al
Full
model
Basic model
(controls only)
As the results in Table 5 show, we find strong support
for four of our five hypotheses (H1, H2, H3 and H5). Our
analyses indicate that the moderating effect of ETs on the
influence of MP on TMS for IT innovation assimilation
initiatives and resulting IT business value differs for rather
mindful and less mindful firms. If we consider the effect of
MP on TMS for IT innovation assimilation initiatives, the
results indicate that the coefficient of the relation of MP on
TMS is not significant for the group of rather mindful firms
(bMP-TMS ¼ 0.02, P40.05). However, the empirical results
clearly suggest a significant positive effect of MP on TMS
for the group of less mindful firms (b MP-TMS ¼ 0.23,
Po0.01), which is positively moderated by ET (bET MP-TMS ¼
0.32, Po0.01), indicating mimicry due to less-developed
mindfulness in turbulent environments as a source of
uncertainty. This effect decreases (bET MP-TMS ¼ 0.21,
Po0.01) in the group of rather mindful organizations.
Here, the interaction term (ET MP) indicates the strength
of the theoretical relationship, and provides an estimate
of the extent to which ET influences support for IT innovation
assimilation. Moreover, by estimating the differences
in the relevant IT innovation assimilation–performance
Full
model
Coefficient
(full model)
Significance
(parametric;
Chin, 2000)
Significance
(non-parametric test)
Henseler et al.
(2009), Henseler
(2012)
Sarstedt
et al.
(2011)
CP 0.23** 0.12* 0.17* 0.14 +0.02 0.86 0.45 40.10
NP 0.17* 0.14* 0.27** 0.28** +0.14 0.18 0.10 o0.10
MP — 0.23** — 0.02 0.21 0.04 0.03 o0.05
ET — 0.01 — 0.08 +0.07 0.48 0.24 40.10
ET MP — 0.32** — 0.21** 0.11 0.34 0.12 o0.10
R2 (TMS) 0.11 0.24 0.17 0.23 — — — —
DR2 — +0.13 — +0.06 — — — —
GIC 0.10 0.05 0.03 0.04 0.09 0.96 0.47 40.10
GTI 0.27** 0.32** 0.38** 0.30** 0.02 0.72 0.37 40.10
TIME 0.09 0.08 0.13* 0.10 +0.02 0.83 0.41 40.10
SIZE 0.22** 0.20** 0.03 0.03 +0.23 o0.01 o0.01 o0.025
TMS — 0.31** — 0.32** +0.01 0.81 0.45 40.10
R 2 (ASSM) 0.19 0.29 0.15 0.24 — — — —
DR2 — +0.10 — +0.09 — — — —
GIC 0.09 0.06 0.02 0.04 0.10 0.53 0.26 40.10
GTI 0.17** 0.08 0.28** 0.02 0.10 0.68 0.32 40.10
TIME 0.13 0.04 0.08 0.03 0.01 0.86 0.46 40.10
SIZE 0.20** 0.06 0.05 0.02 +0.08 0.28 0.14 o0.05
ASSM — 0.61** — 0.68** +0.07 0.56 0.29 40.10
ET — 0.17** — 0.21** +0.38 o0.01 o0.01 o0.025
R 2 (BPP) 0.13 0.46 0.10 0.56 — — — —
DR2 — +0.33 — +0.46 — — — —
P-values: *Po0.05; **Po0.01.
The P-values, and the corresponding significance level for the group comparison where calculated, are based on bootstrap samples of
1000 observations.
Bold values refers to the associated construct, e.g., CP1 to CP3 refer to the CP construct.
coefficients for the rather mindful and less mindful firms,
the impact of this dynamic capability can be assessed. The
results of the group comparison indicate that the strength
of the direct link between ASSM and BPP increases in the
group of mindful organizations (Db ASSM-BPP ¼þ0.07).
Moreover, the results indicate that the effect of ET on
BPP turns from a negative effect for the less mindful group
(b ET-BPP ¼ 0.17, Po0.01) to a positive effect for the rather
mindful firms (bET-BPP ¼ 0.21, Po0.01).
With regard to the control variables, we note that SIZE
seems to be negatively related to ASSM within the group of
less mindful organizations (b SIZE-ASSM ¼ 0.20, Po0.01),
whereas, in particular, GTI is positively related to the
assimilation stage (ASSM) in both groups (bSIZE-ASSM ¼
0.30/0.32, Po0.01). The other control variables are insignificant
with respect to the effect on ASSM and BPP. In
sum, the estimates clearly indicate the robustness of the
hypothesized relations when we control for other influence
factors (GIC, GTI, SIZE, TIME). Furthermore, assessment
of the coefficients of determination (R 2 ) indicates that
the hypothesized main effects and positive moderation of
ET contribute substantially to the explanatory power of our

esearch model. The R 2 scores for the dependent variable in
the full model are close to 0.30 for ASSM and 0.50 for the
BPP, indicating that our full model explains a moderate
amount of variance of the dependent variable (Chin, 1998).
In particular, the explained variance increases by about
10% for ASSM and 35% for BPP for our full model,
compared with the baseline model containing only the
control factors. This indicates the strength of the theoretical
relationships found in our analysis, and thereby provides a
solid estimation of the degree to which ET, as a source of
uncertainties, encourages mimetic behaviour and influences
both the adoption decision and assimilation process
and the corresponding business value creation in our
investigation.
Since we obtained our survey data from key respondents
in IT departments of financial services providers, and all selfreported
data can potentially be confounded by common
method bias (Podsakoff et al., 2003), we used several means
to assess and minimize the presence of this potential threat.
First, our measurement instrument contains different scales
to reduce scale commonality (Podsakoff et al., 2003). After
data collection, we conducted Harman’s one-factor test
(Podsakoff and Organ, 1986) to control for single-respondent
bias. An unrotated principal components factor analysis
combining independent and dependent variables indicates
that the identified principal factors account for only 22.3%
of the covariance among all constructs, well below the 50%
rule-of-thumb cut-off (Podsakoff and Organ, 1986). Next,
following the recommendation of Podsakoff et al. (2003) and
the analytical procedure used by Liang et al. (2007), we
added a common method factor to the PLS measurement
model. The indicators of all constructs were associated
reflectively with this method factor. Then each indicator’s
variances substantively explained by the principal construct
and by the method factor were computed. The results
show that (1) only one out of the 26 method loadings is
significant, and (2) whereas the average substantively
explained variance for an indicator is 0.716, the commonmethod-based
variance is only 0.0026. In addition, results of
the structural models demonstrated different levels of
significance for path coefficients. On the basis of the results
Table 6 Hypothesis testing results
Mindfully resisting the bandwagon M Wolf et al
of these tests, we concluded that there is no serious concern
of common method bias for this study.
Discussion
Key findings
In total, the results support four of the five hypotheses
(see Table 6), and suggest that four relationships differ
between the rather mindful group and the less mindful
group (see Table 5). Consistent with Ang and Cummings
(1997) and Liang et al. (2007), our results emphasize that,
in particular, it is MP that drives top management to
support IT innovation assimilation initiatives (H1) (see
Figure 2). The behaviour of successful competitors is likely
to initiate a new bandwagon that seduces other firms in the
same market to join it without considering their firmspecific
circumstances. Our results indicate that the
influence of MP (as the primary origin of bandwagons)
on TMS is indeed strengthened by a highly turbulent
environment, such as the financial services industry (H2).
Thereby, we underpin the conceptual work by Rosenkopf
and Abrahamson (1999), who note that ET eventually
fosters uncertainty and resulting mimicry. In sum, the
underlying mechanisms seem to be similar in highly
turbulent environments. In essence, our results reveal
mechanisms of mimicry and herding behaviour that are
potentially present in highly turbulent industries in general.
With the distinction between rather mindful and less
mindful firms we provide a more nuanced perspective
on IT innovation assimilation and its exposure to ET, as
well as the consequences of MP (H3, H4, H5). By doing so,
we address calls by Swanson and Ramiller (2004), Butler
and Gray (2006) as well as Fichman (2004) to assess the
role of OM as a firm’s capability to resist misguiding
environmental influences due to strict organization-specific
reasoning, nuanced appreciation of the environment and
contextualized derivation of the action repertoire. We find
that OM mitigates the positive moderating impact of ET on
the influence of TMS (H2). Consequently, in rather mindful
firms top management is less likely to be affected by the MP
Type of hypothesis Hypothesis Support
Same impact across groups H1: Mimetic pressure drives the management to support IT innovation
assimilation
H2: A higher extent of environmental turbulence strengthens the
influence of mimetic pressure on top management support for IT
innovation assimilation
Differential impact
H3: Rather mindful organizations are less affected by the impact of
across groups
mimetic pressure on the IT innovation assimilation process compared
with less mindful organizations
H4: Rather mindful firms are able to realize a higher extent of business
process performance resulting from IT innovation assimilation in the
risk management, asset management and new product development
process than less mindful firms
H5: Rather mindful organizations are able to benefit from a high extent
of environmental turbulence, whereas less mindful organizations are
likely to be negatively affected by it
Supported
Supported
Supported
Not
supported
Supported
13

14
0.10
**
0.10
**
ns
0.10
Path coefficient for the more mindful group
Path coefficient for the less mindful group
Not significant
0.10
R values for the more mindful (left)
and the less mindful (right) group
Mimetic
Coercive
Normative
H1, H3
caused by ET. Moreover, the direct influence of MP on TMS
may even vanish in rather mindful firms. This can be
attributed to the capability of reflection in action, which is
assumed to be especially developed in mindful organizations
(Jordan et al., 2009). Reflection in action is defined by
the ability to actively learn and realign from prior and
current experiences, and in particular from critical,
‘transformative’ change as initiated by bandwagons.
Transformative change continuously challenges internalized
ways of thinking, and therefore facilitates active
resistance to bandwagons that are not appropriate for a
firm-specific context.
Interestingly, our results suggest that highly turbulent
environments can be even an opportunity for rather
mindful firms with regard to BPP (H4). In detail, our
research indicates that rather mindful firms can benefit
from ET at the business process level, whereas less mindful
firms are negatively affected by rapid environmental
changes with regard to BPP. This surprising result can be
explained by the increased scanning and interpretation
capabilities that rather mindful firms exhibit, which enable
them to operate proactively in highly turbulent markets,
rather than being passively driven by them (Teece, 2007).
In contrast, less mindful firms are influenced by the
institutional environment and prevailing turbulence, and
therefore are more susceptible to bandwagon phenomena.
This surprising result can be explained by the increased
scanning and interpretation capabilities that rather mindful
firms exhibit that enable them to operate proactively in
highly turbulent markets (Teece, 2007). In the case of
Grid computing, mindful firms are likely to identify the
potential of this IT innovation to meet the varying IT
resource demand in highly turbulent environments. In
contrast, less mindful firms are driven by the institutional
environment and prevailing turbulence and therefore are
more susceptible to bandwagon phenomena. As far as the
IT innovation assimilation process is concerned, less
mindful firms are likely to assimilate it even if the IT
innovation does not meet their specific requirements.
ns
0.23
**
ns
0.12
*
0.28
**
0.23
**
Figure 2 Results of the group comparison (low vs high OM).
P-values: *Po0.05; **Po0.01.
Mindfully resisting the bandwagon M Wolf et al
H2
0.21
**
0.32
**
Environmental
Turbulence
Top Management
Grid Assimilation
H4
Business Process
Support 0.23 0.24
0.32
0.24 0.29
0.68
Performance 0.56 0.46
**
**
0.31
0.61
**
**
H5
0.21
**
-0.17
**
Controls
Grid Infrastructure Capabilities (GIC)
Grid Technology Integration (GTI)
Earliness of Grid Adoption (TIME)
Firm Size (SIZE)
Finally, we found descriptive evidence that rather mindful
firms can realize a higher level of business value from IT
innovation assimilation at the business process level than
can less mindful firms (H5). Nevertheless, with regard to
inferential statistics, we could not find a significant
difference. This can be explained by the fact that the
realization of IT-based business value is highly contingent,
and thus cannot be reduced solely to the IT innovation
assimilation process itself. This would also explain why the
direct effect of ET on business value generation is relatively
strong and significant.
Implications for research and practice
The theoretical contribution of the depicted research is
twofold. First, it contributes to the assimilation and
diffusion of innovations theory by assessing the role of
OM in mitigating the negative influences that potentially
may arise from pure mimicry in the IT innovation
assimilation process. By doing so, it sheds light on the
micro-foundations of institutionalization and its interplay
with the dynamic capabilities of the firm, which help to
identify these influences and make the exposed firm aware
of them. The firm is therefore enabled to consciously follow
and frame a bandwagon or consciously resist it, rather than
be passively driven by it. Our research thereby addresses
recent calls to extend the nomological net of IT innovation
assimilation as well as influencing and attenuating sources
of institutionalization (e.g., Ang and Cummings, 1997; Fiol
and O’Connor, 2003; Swanson and Ramiller, 2004). Second,
our results indicate that OM is one viable means to realize a
higher extent of IT business value against the background
of a highly turbulent environment. Mindful firms are even
potentially able to profit from ET. Consistent with the
theory of dynamic capabilities, the discriminant perception
of environmental change (scanning) and a response to it
(interpretation) reflect an increased entrepreneurial alertness
and the results of an above-average generation of
digital options that are particularly valuable in a rapidly

Mindfully resisting the bandwagon M Wolf et al
changing environment. Surprisingly, we find evidence that
an environment of high turbulence can be an opportunity
for rather mindful firms, whereas it reflects a threat for less
mindful firms.
In addition, our study on IT innovation assimilation
(i.e., ASSM) has different implications for decision-makers
of organizations with a high reliance on the availability and
dependability of IT systems (such as financial services
providers). First, senior IT decision-makers need to be
aware of the level of participation and mindfulness that is
required in the process of assimilating IT innovations
(Liang et al., 2007). More importantly, they need to take
into account that, with increased ET, they are exposed to
higher MP. Because of uncertainty (e.g., incomplete
information about future developments), firms tend to
restructure themselves using other successful competitors
in their market as role models. The awareness of this
relation can help to improve scanning capabilities (e.g.,
through better decision support systems) and initiate a
mindful decision-making process to identify contextually
appropriate IT innovation decisions. In this regard, OM
as a dynamic capability and the corresponding complementary
cognitive dimensions may serve as an overall
organizational concept to ensure successful assimilation of
IT innovation and business value generation in highly
turbulent environments. OM can thereby be assumed to be
one means to identify and accommodate changes facilitated
by the market, and to resist arising bandwagon phenomena
that might otherwise negatively affect the generation of ITinduced
business value. In particular, mindfulness training
sessions could empower employees to develop increased
scanning and interpretation capabilities.
Our research also informs the work of financial
regulatory authorities, which should take a closer look at
the mindful use of IT systems within the financial services
industry. For example, one lesson learnt for regulators
could be to enforce a stricter regime enforcing standards
among less mindful organizations, or to be less supportive
in crisis situation, in contrast with rather mindful
competitors, in order to encourage the development of a
rather mindful culture within the industry.
Limitations and future research
Despite the rich findings, our study has some limitations that
suggest avenues for future research. In our study, we expand
the nomological network of the assimilation and diffusion of
innovation theory by assessing the role of OM as a dynamic
capability mitigating the negative influences that potentially
may arise from pure mimicry. We thereby expand the
operationalization of cognitive organizational capabilities by
aligning it to dynamic capabilities theory in the light of ETs.
However, our study emphasizes the top management level as
the sole human agency for transferring institutional pressures
to the IT innovation assimilation process, although the extant
literature suggests that the middle management level could
be equally important for transferring such pressures (Liang
et al., 2007). Accordingly, further research might find our
research model a valuable starting point to integrate the
middle management level as well.
In addition, future research can investigate how
other organizational capabilities influence the innovation
15
assimilation process in turbulent environments. The consequences
of increasing ETs demand organizational sensemaking
and responsiveness to safeguard organizational
performance. Accordingly, companies can be seen as sensemaking
units that are stimulated by ET, and which are
constantly challenged to identify the appropriate contextual
response (McGill et al., 1993). From a theoretical perspective,
and with regard to the existing research in behavioural
economics (e.g., Kahneman, 2003), future research should
integrate and investigate factors to account for further
cognitive capabilities and limitations, the amount and
quality of accessible information, and other restrictions that
influence an economic-rational decision-making process.
In this regard, our theorizing is driven by an institutional
perspective rather than by an adaptive learning perspective,
although human beings (and therefore top managers) are
constantly engaged in active sensing, responding and
learning processes within different environmental conditions
(Overby et al., 2006). Specifically, we focus on the
relatively understudied relationship between uncertainties
resulting from ETs and mimetic isomorphism that constitutes
one of the core arguments in institutional theory
(DiMaggio and Powell, 1983). Since our current sample
encompasses organizational data, further research could
explore the underlying processes at the individual level.
Accordingly, future studies may employ a process view
of isomorphism, with a special focus on the early stage
of the innovation assimilation process, by investigating
individual-level data from organizational decision makers
facing critical make-or-buy decisions. Such longitudinal
investigations could address recent calls by researchers
for increased attention to the mechanisms and microfoundations
of institutional theory (Powell and Colyvas,
2008; Battilana et al., 2009), which have received only
limited attention in the existing literature.
From a methodological perspective, our sample consists
of firms with more than 1000 employees, and thus
potentially limits the generalizability of the results. Moreover,
the responses of our survey represent the IT delivery
side, which might overemphasize a technologically driven
perspective. Dyadic data matching responses from the IT
department with responses from the business department
could overcome this limitation, and provide further
insights into the process of assimilating IT innovation.
Finally, our analyses focus on the financial services
industry, which might be idiosyncratic with respect to an
above-average level of ET. Concurrently, our findings
represent an avenue for further research that could draw
a more nuanced picture of the role of ETs and necessary
dynamic capabilities in the innovation assimilation process.
Conclusion
In essence, ‘mindless’ IT innovation assimilation in highly
dynamic environments, such as the financial services
industry, can eventually lead to the emergence of bandwagons
and a loss in BPP. A well-developed organizational
scanning and interpretation ability is required to comprehend
the implications of complex IT innovations and the
relevant IT innovation decisions that implement them.
In contrast, not all IT departments are able to ‘mindfully’

16
derive the strategic implications of IT innovations, such as
Grid computing.
Our results indicate that OM is one way to contrast
IT innovation strategies that eventually serve as a ‘fire
accelerant’ with a mindful mitigation of bandwagon
influences and the accompanying precise risk assessment.
This mindfulness in the assimilation of IT innovation might
not be restricted to the originating firm; it might also alert
other, less mindful firms, since the overall market volatility
is likely to be mitigated. Mindfulness analyses at an interorganizational
or industry-level perspective could be a
fruitful area for future research. In the end, OM could be
propagated by the same means as the financial crisis itself:
MP. A majority of mindful firms can force their less
mindful competitors to follow.
References
Mindfully resisting the bandwagon M Wolf et al
Abrahamson, E. and Rosenkopf, L. (1990). When Do Bandwagon Diffusions
Roll? How Far Do They Go? And When Do They Roll Backwards? A
Computer Simulation, Academy of Management Best Paper Proceedings
50(1): 155–159.
Adler, P.S., Glodoftas, B. and Levine, D.L. (1999). Flexibility Versus Efficiency?
A Case Study of Model Changeovers in the Toyota Production System,
Organization Science 10(1): 43–68.
Aldrich, H.E. and Fiol, C.M. (1994). Fools Rush In? The Institutional Context of
Industry Creation, Academy of Management Review 19(1): 645–670.
Ang, S. and Cummings, L.L. (1997). Strategic Response to Institutional
Influences on Information Systems Outsourcing, Organization Science 8(3):
235–256.
Armstrong, J. and Overton, T. (1977). Estimating Non-Response Bias in Mail
Surveys, Journal of Marketing Research 14(8): 396–402.
Bagozzi, R.P. and Phillips, L.W. (1991). Assessing Construct Validity in
Organizational Research, Administrative Science Quarterly 36(3): 421–458.
Barua, A., Konana, P., Whinston, A.B. and Fang, Y. (2004). An Empirical
Investigation of Net-Enabled Business Value, MIS Quarterly 28(4):
585–620.
Battilana, J., Leca, B. and Boxenbaum, E. (2009). How Actors Change
Institutions: Towards a theory of institutional entrepreneurship, Academy of
Management Annals 3(1): 65–107.
Box, G.E.P., Hunter, J.S. and Hunter, W.G. (2005). Statistics for Experimenters:
Design, innovation, and discovery, 2nd edn, New York: Wiley-Interscience.
Buganza, T., Dell’Era, C. and Verganti, R. (2009). Exploring the Relationships
Between Product Development and Environmental Turbulence: The case of
mobile TLC services, Journal of Product Innovation Management 26(3):
308–321.
Butler, B.S. and Gray, P.H. (2006). Reliability, Mindfulness, and Information
Systems, MIS Quarterly 30(2): 211–224.
Buyya, R., Yeo, C.S., Venugopal, S., Broberg, J. and Brandic, I. (2009). Cloud
Computing and Emerging IT Platforms: Vision, hype, and reality for
delivering computing as the 5th utility, Future Generation Computer Systems
25(6): 599–616.
Campbell, D.T. and Fiske, D.W. (1959). Convergent and Discriminant
Validation by the Multitrait-Multimethod Matrix, Psychological Bulletin
56(1): 81–105.
Cassel, C., Hackl, P. and Westlund, A. (1999). Robustness of Partial
Least-Squares Method for Estimating Latent Variable Quality Structures,
Journal
of Applied Statistics 26(4): 435–446.
Chambers, J.M., Cleveland, W.S., Kleiner, B. and Tukey, P.A. (1983). Graphical
Methods for Data Analysis, Belmont, CA: Wadsworth.
Chin, W.W. (1998). The Partial Least Squares Approach to Structural Equation
Modeling, in G.A. Marcoulides (ed.) Modern Methods for Business Research,
Hillsdale, NY: Lawrence Erlbaum Associates, pp. 295–336.
Chin, W.W. (2000). Frequently asked questions: Partial least squares and
PLS-graphs, [WWW document] http://disc-nt.cba.uh.edu/chin/plsfaq/
plsfaq.htm (accessed 2 September 2011).
Chin, W.W., Marcolin, B.L. and Newsted, P.R. (2003). A Partial Least Squares
Latent Variable Modeling Approach for Measuring Interaction Effects:
Results from a Monte Carlo simulation study and an electronic-mail
emotion/adoption study, Information Systems Research 14(2): 189–217.
D’Agostino, R.B., Belanger, A. and D’Agostino Jr, R.B. (1990). A Suggestion for
Using Powerful and Informative Tests of Normality, American Statistician
44(1): 316–332.
Davamanirajan, P., Kauffman, R.J., Kriebel, C.H. and Mukhopadhyay, T.
(2006). Systems Design, Process Performance, and Economic Outcomes in
International Banking, Journal of Management Information Systems 23(2):
65–90.
Davenport, T. (1993). Process Innovation: Reengineering work through IT,
Boston, MA: Harvard Business School.
DiMaggio, P.W. and Powell, W.W. (1983). The Iron Cage Revisited:
Institutional isomorphism and collective rationality in organizational fields,
American Sociological Review 48(2): 147–160.
Dixit, A.K. and Pindyck, R.S. (1994). Investment under Uncertainty, Princeton,
NJ: Princeton University Press.
Dong, S., Xu, S.X. and Zhu, K.X. (2009). Information Technology in Supply
Chains: The value of IT-enabled resources under competition, Information
Systems Research 20(1): 18–32.
Efron, B. (1981). Nonparametric standard errors and confidence intervals,
Canadian Journal of Statistics 9(2): 139–172.
FDIC (2012). Failed Bank List, [WWW document] http://www.fdic.gov/bank/
individual/failed/banklist.html (accessed: 18 April 2012).
Festinger, L. (1967). A Theory of Cognitive Dissonance, Evanston, IL: Row
Peterson.
Fichman, R.G. (2001). The Role of Aggregation in the Measurement of
IT-related Organizational Innovation, MIS Quarterly 25(4): 427–455.
Fichman, R.G. (2004). Going Beyond the Dominant Paradigm for Information
Technology Innovation Research: Emerging concepts and methods, Journal
of the Association for Information Systems 5(8): 314–355.
Fiol, C.M. and O’Connor, E.J. (2003). Waking Up! Mindfulness in the
Face of Bandwagons, Academy of Management Review 28(1): 54–70.
Fornell, C. (1992). A National Customer Satisfaction Barometer: The Swedish
experience, Journal of Marketing 56(1): 6–21.
Fornell, C. and Larcker, D.F. (1981). Evaluating Structural Equation Models
with Unobservable Variables and Measurement Error: Algebra and statistics,
Journal of Marketing Research 18(1): 382–388.
Foster, I. (2002). What is the Grid? A Three Point Checklist, GRID Today 1(6):
22–25.
Foster, I. and Kesselman, C. (1999). The Grid: Blueprint for a new computing
infrastructure, San Francisco, CA: Morgan Kaufmann.
Foster, I., Zhao, Y., Raicu, I. and Lu, S. (2008). Cloud Computing and
Grid Computing 360-Degree Compared, in Proceedings of the IEEE Grid
Computing Environments Workshop (Austin, TX, USA, 2008); San Diego,
CA: IEEE Computer Society Press.
Gartner (2010). Forecast: Public Cloud Services, Worldwide and Regions,
Industry Sectors, 2009–2014, [WWW document] http://www.gartner
.com/DisplayDocument?ref ¼ clientFriendlyUrl&id ¼ 1378513 (accessed
29 March 2011).
Goswami, S., Teo, H.H. and Chan, H.C. (2008). Real Options from RFID
Adoption: The role of institutions and managerial mindfulness, in
Proceedings of the 29th Annual International Conference on Information
Systems (ICIS) (Paris, France, 2008); Atlanta, GA: Association for
Information Systems.
Hackenbroch, W. and Henneberger, M. (2007). Grid Computing Infrastructures
and their Value for Risk Management, in Proceedings of the 40th Annual
Hawaii International Conference on System Sciences (HICSS) (Waikoloa,
Hawaii, 2007); San Diego, CA: IEEE Computer Society Press.
Hair, J., Anderson, R., Tatham, R. and Black, W. (1998). Multivariate Data
Analysis with Readings, 4th edn, Englewood Cliffs, NJ: Prentice Hall.
Henseler, J. (2012). PLS-MGA: A non-parametric approach to partial least
squares-based multi-group analysis, in W.A. Gaul, A. Geyer-Schulz,
L. Schmidt-Thieme and J. Kunze (eds.) Challenges at the Interface of
Data Analysis, Computer Science, and Optimization, Berlin, Heidelberg:
Springer, pp. 495–501.
Henseler, J., Ringle, C.M. and Sinkovics, R.R. (2009). The Use of Partial
Least Squares Path Modeling in International Marketing, in R.R. Sinkovics
and P.B. Ghauri (eds.) Advances in International Marketing, Vol. 20, Bingley:
Emerald JAI Press, pp. 277–319.

Mindfully resisting the bandwagon M Wolf et al
Hsieh, J.J.P.-A., Rai, A. and Keil, M. (2008). Understanding Digital Inequality:
Comparing continued use behavioral models of the socio-economically
advantaged and disadvantaged, MIS Quarterly 32(1): 97–126.
Hulland, J. (1999). Use of Partial Least Squares (PLS) in Strategic Management
Research: A review of four recent studies, Strategic Management Journal
20(2): 195–204.
Iansiti, M. (1998). Technology Integration: Making critical choices in a
dynamic world, Boston, MA: Harvard Business Press.
IDC (2010). Through 2014 Public IT Cloud Services Will Grow at More Than
Five Times the Rate of Traditional IT Products, New IDC Research Finds
[WWW document] http://www.idc. com/about/viewpressrelease. jsp?
containerId ¼ prUS22393210&sectionId ¼ null&elementId ¼ null&
pageType ¼ SYNOPSIS (accessed 18 April 2012).
Jap, S.D. (2001). Perspectives on Joint Competitive Advantages in Buyer–
Supplier Relationships, International Journal of Research in Marketing
18(1–2): 19–35.
Jaworski, B.J. and Kohli, A.K. (1993). Market Orientation: Antecedents and
consequences, Journal of Marketing 57(7): 53–70.
Johns, G. (2006). The Essential Impact of Context on Organizational Behavior,
Academy of Management Review 31(2): 386–408.
Jordan, S., Messner, M. and Becker, A. (2009). Reflection and Mindfulness in
Organizations: Rationales and possibilities for integration, Management
Learning 40(4): 465–473.
Jorion, P. (2006). Value at Risk, 3rd edn, Burr Ridge, IL: Richard D. Irwin.
Kahneman, D. (2003). Maps of Bounded Rationality: Psychology for behavioral
economics, American Economic Review 93(5): 1449–1475.
Karimi, J., Somers, T.M. and Bhattacherjee, A. (2007a). The Impact of ERP
Implementation on Business Process Outcomes: A factor-based study,
Journal of Management Information Systems 24(1): 101–134.
Karimi, J., Somers, T.M. and Bhattacherjee, A. (2007b). The Role of
Information Systems Resources in ERP Capability Building and Business
Process Outcomes, Journal of Management Information Systems 24(2):
221–260.
Keil, M., Tuunainen, V. and Wassenaar, A. (2000). A Cross-Cultural Study on
Escalation of Commitment Behaviors in Software Projects, MIS Quarterly
24(2): 299–325.
Knight, A.P. (2004). Measuring Collective Mindfulness and Exploring its
Nomological Network, Unpublished Master thesis, University of Maryland,
College Park.
Kruskal, W.H. and Wallis, W.A. (1952). Use of Ranks in One-Criterion
Variance Analysis, Journal of the American Statistical Association 47(260):
583–621.
Langer, E.J. (1989). Minding Matters: The consequences of mindlessness–
mindfulness, San Diego, CA: Academic Press.
Langer, E.J. and Moldoveanu, M. (2000). The Construct of Mindfulness, Journal
of Social Issues 56(1): 1–9.
Lawrence, P.R. and Lorsch, J.W. (1967). Organization and Environment, Irwin,
IL: Homewood.
Levene, H. (1960). Robust Tests for Equality of Variances, in I. Olkin (ed.)
Contributions to Probability and Statistics, Palo Alto, CA: Stanford University
Press, pp. 278–292.
Liang, H., Nilesh, S., Hu, Q. and Xue, Y. (2007). Assimilation of Enterprise
Systems: The effect of institutional pressures and the mediating role of top
management, MIS Quarterly 31(1): 59–87.
Loch, K.D., Straub, D.W. and Kamel, S. (2003). Diffusing the Internet in the
Arab World: The role of social norms and technological culturation, IEEE
Transactions on Engineering Management 50(1): 45–63.
MacKenzie, S.B., Podsakoff, P.M. and Jarvis, C.B. (2005). The Problem of
Measurement Model Misspecification in Behavioral and Organizational
Research and Some Recommended Solutions, Journal of Applied Psychology
90(4): 710–730.
McGill, A.R., Johnson, M.D. and Bantel, K.A. (1993). Cognitive Complexity
and Conformity: The effects on performance in a turbulent environment,
Academy of Management Proceedings 75(3): 379–383.
Meyer, J. and Rowan, B. (1977). Institutional Organizations: Formal
structure as myth and ceremony, American Journal of Sociology 83(2):
340–363.
Miller, L. and Smith, K. (1983). Handling Non-Response Issues, Journal of
Extension 21(9–10): 45–50.
Mintzberg, H. (1973). The Nature of Managerial Work, New York:
Harper & Row.
17
Moore, G.C. and Benbasat, I. (1991). Development of an Instrument to Measure
the Perceptions of Adopting an Information Technology Innovation,
Information Systems Research 2(3): 192–222.
Nunnally, J.C. (1978). Psychometric Theory, New York: McGraw-Hill.
Overby, E., Bharadwaj, A.S. and Sambamurthy, V. (2006). Enterprise Agility
and the Enabling Role of Information Technology, European Journal of
Information Systems 15(2): 120–131.
Pavlou, P.A. and El Sawy, O.A. (2006). From IT Leveraging Competence to
Competitive Advantage in Turbulent Environments: The case of new product
development, Information Systems Research 17(3): 198–227.
Petter, S., Straub, D. and Rai, A. (2007). Specifying Formative Constructs in
Information Systems Research, MIS Quarterly 31(4): 623–656.
Podsakoff, P.M., MacKenzie, S.B., Jeong-Yeon, L. and Podsakoff, N.P. (2003).
Common Method Biases in Behavioral Research: A critical review of the
literature and recommended remedies, Journal of Applied Psychology 88(5):
879–903.
Podsakoff, P.M. and Organ, D.W. (1986). Self-reports in Organizational
Research: Problems and prospects, Journal of Management 12(4):
531–544.
Powell, W.W. and Colyvas, J.A. (2008). Microfoundations of Institutional
Theory, in R. Greenwood, C. Oliver, K. Sahlin and R. Suddaby
(eds.) Handbook of Organizational Institutionalism, London: Sage,
pp. 276–298.
Qureshi, I. and Compeau, D. (2009). Assessing Between-Group Differences in
Information Systems Research: A comparison of covariance- and
component-based SEM, MIS Quarterly 33(1): 197–214.
Ragu-Nathan, B.S., Apigian, C.H., Ragu-Nathan, T.S. and Tu, Q. (2004). A Path
Analytic Study of the Effect of Top Management Support for Information
Systems Performance, Omega 32(6): 459–471.
Rai, A., Brown, P. and Tang, X. (2009). Organizational Assimilation of
Electronic Procurement Innovations, Journal of Management Information
Systems 26(1): 257–296.
Ringle, C.M., Wende, S. and Will, A. (2005). SmartPLS 2.0 (beta), [WWW
document] http://www.smartpls.de (accessed 28 April 2011).
RMA (2010). Data Quality: Is the glass half empty or half full?, [WWW
document] http://ebiz.rmahq.org/eBusPPRO/Default.aspx?TabId=
55&ProductId=30500 (accessed 28 April 2010).
Rogers, E.M. (1995). Diffusion of Innovations, New York: The Free Press.
Rosenkopf, L. and Abrahamson, E. (1999). Modeling Reputational and
Informational Influences in Threshold Models of Bandwagon Innovation
Diffusion, Computational and Mathematical Organization Theory 5(1):
361–384.
Sambamurthy, V., Bharadwaj, A. and Grover, V. (2003). Shaping Agility
through Digital Options: Reconceptualizing the role of information
technology in contemporary firms, MIS Quarterly 27(2): 237–263.
Sarstedt, M., Henseler, J. and Ringle, C.M. (2011). Multi-Group Analysis in
Partial Least Squares (PLS) Path Modeling: Alternative methods and
empirical results, Advances in International Marketing 22(1): 195–218.
SEC (2008). Bank of America Corporation Form 10-K, [WWW document]
http://www.sec.gov/archives/edgar/data/70858/000119312509041126/
D10K.htm (accessed 28 April 2010).
Sharma, S., Durand, R. and Gur-Arie, O. (1981). Identification and
Analysis of Moderator Variables, Journal of Marketing Research 18(3):
291–299.
Sivo, S.A., Saunders, C., Chang, Q. and Jiang, J.J. (2006). How Low Should You
Go? Low Response Rates and the Validity of Inference in IS Questionnaire
Research, Journal of the Association for Information Systems 7(6):
351–414.
Sternberg, R.J. (2000). Images of Mindfulness, Journal of Social Issues
56(1): 11–26.
Straub, D. (1989). Validating Instruments in MIS Research, MIS Quarterly
13(2): 147–169.
Straub, D., Boudreau, M.-C. and Gefen, D. (2004). Validation Guidelines for IS
Positivist Research, Communications of the Association for Information
Systems 13(24): 380–427.
Strong, P. (2005). Enterprise Grid Computing, ACM Queue 3(6): 50–59.
Sugumaran, V., Tanniru, M. and Storey, V.C. (2008). A Knowledge-Based
Framework for Extracting Components in Agile Systems Development,
Information Technology and Management 9(1): 37–53.
Swanson, E.B. (1994). Information Systems Innovation among Organizations,
Management Science 40(9): 1069–1092.

18
Mindfully resisting the bandwagon M Wolf et al
Swanson, E.B. and Ramiller, N.C. (2004). Innovating Mindfully with
Information Technology, MIS Quarterly 28(4): 553–583.
Tanriverdi, H. (2005). Information Technology Relatedness, Knowledge
Management Capability, and Performance of Multibusiness Firms, MIS
Quarterly 29(2): 311–334.
Teece, D.J. (2007). Explicating Dynamic Capabilities: The nature and
microfoundations of (sustainable) enterprise performance, Strategic
Management Journal 28(13): 1319–1350.
Teece, D.J., Pisano, G. and Shuen, A. (1997). Dynamic Capabilities and Strategic
Management, Strategic Management Journal 18(7): 509–533.
Tenbrunsel, A.E., Galvin, T.L., Neale, M.A. and Bazerman, M.H. (1996).
Cognitions in Organizations, in S.R. Clegg, C. Hardy and W.R. Nord (eds.)
Handbook of Organization Studies, Thousand Oaks, CA: Sage,
pp. 313–337.
Teo, H.H., Wei, K.K. and Benbasat, I. (2003). Predicting Intention to Adopt
Interorganizational Linkages: An institutional perspective, MIS Quarterly
27(1): 19–49.
Thomas, J.B., Clark, S.M. and Gioia, D.A. (1993). Strategic Sensemaking and
Organizational Performance: Linkages among scanning, interpretation,
action, and outcomes, Academy of Management Journal 36(1): 239–270.
Tornatzky, L.G. and Fleischer, M. (1990). The Processes of Technological
Innovation, Massachusetts/Toronto: Lexington Books.
Valorinta, M. (2009). Information Technology and Mindfulness in
Organizations, Industrial and Corporate Change 18(5): 963–997.
Venkatesh, V. and Morris, M.G. (2000). Why Don’t Men Ever Stop to Ask for
Directions? Gender, Social Influence, and Their Role in Technology
Acceptance and Usage Behavior, MIS Quarterly 24(1): 115–139.
Weick, K.E. and Sutcliffe, K.M. (2001). Managing the Unexpected:
Assuring high performance in an age of complexity, San Francisco, CA:
Jossey-Bass.
Weick, K.E. and Sutcliffe, K.M. (2007). Managing the Unexpected: Resilient
performance in an age of uncertainty, San Francisco, CA: Jossey-Bass.
Weick, K.E., Sutcliffe, K.M. and Obstfeld, D. (1999). Organizing for High
Reliability: Processes of collective mindfulness, Research in Organizational
Behavior 21(1): 81–123.
Weinhardt, C., Anandasivam, A., Blau, B., Borissov, N., Meinl, T., Michalik, W.
and Stößer, J. (2009). Cloud Computing: A classification, business models,
and research directions, Business and Information Systems Engineering 1(5):
391–399.
Wold, H. (1982). Soft Modeling: The basic design and some extensions,
in K.G. Jöreskog and H. Wold (eds.) Systems under Indirect
Observation: Causality, structure, prediction, Amsterdam: North-Holland,
pp. 1–54.
Yi, M.Y. and Davis, F.D. (2003). Developing and Validating an Observational
Learning Model of Computer Software Training and Skill Acquisition,
Information Systems Research 14(2): 146–169.
Zhu, K., Dong, S., Xu, S.X. and Kraemer, K.L. (2006a). Innovation Diffusion
in Global Contexts: Determinants of post-adoption digital transformation
of European companies, European Journal of Information Systems 15(6):
601–616.
Zhu, K. and Kraemer, K.L. (2005). Post-Adoption Variations in Usage
and Value of E-Business by Organizations: Cross-country evidence
from the retail industry, Information Systems Research 16(1):
61–84.
Zhu, K., Kraemer, K.L. and Xu, S. (2006b). The Process of Innovation
Assimilation by Firms in Different Countries: A technology diffusion
perspective on e-business, Management Science 52(10): 1557–1576.
Zhu, K., Kraemer, K.L., Xu, S. and Dedrick, J. (2004). Information Technology
Payoff in E-Business Environments: An international perspective on value
creation of e-business in the financial services industry, Journal of
Management Information Systems 21(1): 17–54.
About the author
Martin Wolf is a Research Associate at the Chair of
Business Administration, esp. Information Management
and the E-Finance Lab at Goethe University, Frankfurt,
Germany. Previously, he earned a diploma in IS and finance
from the University of Mannheim, Germany and the
University of Waterloo, Canada as well as a Ph.D. from
Goethe University, Germany. He gained industry expertise
as part of his assignment in the financial services analytics
division at SAP. His current research focuses on bounded
rationality concepts and their influence in complex
decision-making situations, such as IT innovation assimilation
processes. His articles have been published in
journals such as Business & Information Systems Engineering
and the Australasian Journal of Information Systems as
well as several conference proceedings such as International
Conference on Information Systems and the European
Conference on Information Systems.
Roman Beck is an Assistant Professor and the E-Finance
and Services Science Chair at Goethe University in
Frankfurt, Germany, where he also earned his Ph.D. His
research in services science focuses on the role of IT
services sourcing, services management and services
engineering with a special focus on IS outsourcing, social
media and virtualization. From a theoretical perspective, he
is interested in institutional logics of organizations, control
balancing, as well as organizational mindfulness and
awareness. He serves as Senior Editor for the Journal of
Information Technology Theory and Application and has
published over 100 conference papers and articles in
journals such as Communications of the AIS, Information
Technology and People, Scandinavian Journal of Information
Systems, IEEE Transactions on Software Engineering,
Communications of the ACM, Business and Information
Systems Engineering, and others.
Immanuel Pahlke is a Ph.D. candidate at the E-Finance and
Services Science Chair of the Institute of Information
Systems at Goethe University, Frankfurt, Germany. He has
a diploma in IS and business economics, which he received
from the Technical University in Darmstadt, Germany. He
also had worked for a management consulting firm in the
financial services sector and currently works as a research
assistant at the E-Finance Lab. His research interests are in
the areas of end user empowerment and knowledge
management specifically with a focus on enterprise social
media platforms. His articles have been published in
journals such as Business & Information Systems Engineering
and Zeitschrift für Betriebswirtschaftslehre as well as
several conference proceedings such as International
Conference on Information Systems, European Conference
on Information Systems.

Appendix
Table A1 Measurement items
Mindfully resisting the bandwagon M Wolf et al
MP (reflective measures) 7-point Likert (1 ¼ strongly disagree; 7 ¼ strongly agree)
Source: Liang et al. (2007)
MP1 Our main competitors who have adopted Grid technology have greatly benefited
MP2 Our main competitors who have adopted Grid technology are favourably perceived
by others in the same industry
MP3 Our main competitors who have adopted Grid technology are favourably perceived
by their suppliers and customers
CP (reflective measures) 7-point Likert (1 ¼ strongly disagree; 7 ¼ strongly agree)
Source: Liang et al. (2007)
CP1 The increasing regulatory pressure requires our firm to use Grid technology
CP2 The increasing customer demand requires our firm to use Grid technology
CP3 The competitive conditions require our firm to use Grid technology
NP (reflective measures) 7-point Likert (1 ¼ strongly disagree; 7 ¼ strongly agree)
Source: Liang et al. (2007)
NP1 Our firm’s IT services providers have already adopted Grid technology
NP2 Our firm’s business partners have already adopted Grid technology
NP3 The government’s promotion of IT influences our firm to use Grid technology
TMS (reflective measures) 7-point Likert (1 ¼ strongly disagree; 7 ¼ strongly agree)
Source: Ragu-Nathan et al. (2004)
TMS1 Top management understands the importance of our Grid infrastructures
TMS2 Top management supports our Grid development projects
TMS3 Top management considers our Grid infrastructures as a strategic resource
TMS4 Top management understands the benefits of our Grid infrastructures
TMS5 Top management keeps the pressure on operating units to use our Grid
infrastructures
ASSM (formative measures) 7-level Guttman (see Table A2)
Source: Rai et al. (2009)
One item for each of the three key business processes (asset management, risk management and new product development
process)
ET (reflective measures) 7-point Likert (1 ¼ strongly disagree; 7 ¼ strongly agree)
Sources: Pavlou and El Sawy (2006); Jaworski and Kohli (1993)
ET1 The environment in our industry is continuously changing
ET2 Environmental changes in our industry are very difficult to forecast
ET3 The technology in our industry is changing rapidly
ET4 In our kind of business, customers’ product preferences change a lot over time
ET5 Marketing practices in our product area are constantly changing
ET6 New product introductions are very frequent in our market
ET7 There are many competitors in our market
BPP (formative measures) Extracted latent variable scores (see Table A3)
Sources: Karimi et al. (2007a, b); expert interviews
One item for each of the three key business processes (asset management, risk management and new product development
process)
Control: GIC (reflective measures) 5-point Likert (0%; 25%; 50%; 75%; 100%)
Sources: Zhu and Kraemer (2005); Zhu et al. (2006a, b)
Please estimate the percentage of your firm’s business applications that access y
GIC1 Distributed computing power
GIC2 High-capacity, low-latency networks
GIC3 Service-oriented architectures
GIC4 Computer clusters
GIC5 Distributed databases
19

20
Table A1 Continued
Control: GTI (reflective measures) 5-point Likert (0%; 25%; 50%; 75%; 100%)
Sources: Zhu and Kraemer (2005); Zhu et al. (2006a, b)
Please estimate the percentage of your firm’s business applications that y
GTI1 Are integrated in a service-oriented architecture
GTI2 Are part of an enterprise application integration (EAI) infrastructure
GTI3 Access computing resources from external IT services providers
GTI4 Access cloud computing resources (e.g., Amazon’s EC2)
GTI5 Are executed in a virtual machine
Further controls
Open survey questions and secondary data
(one-item measures)
Sources: Rogers (1995); Fichman (2001)
TIME Years elapsed since the first Grid adoption
SIZE Number of employees (worldwide)
Table A2 Guttman scale for Grid assimilation (ASSM)
Assimilation stage Criteria to enter the assimilation stage Survey items
1. Awareness Key decision makers are aware of Grid
technology
2. Interest The organization is committed to actively learn
more about Grid technology for its PROCESS
3. Evaluation/trial The organization has acquired specific
innovation-related products and has initiated
evaluation or trial for its PROCESS
4. Commitment The organization has committed to use Grid
technology in a significant way for its PROCESS
5. Limited
deployment
6. Partial
deployment
7. General
deployment
Mindfully resisting the bandwagon M Wolf et al
The organization has established a programme of
regular, but limited, use of Grid technology for
part of its PROCESS
The organization has established a programme of
regular, but limited, use of Grid technology for
their PROCESS
The organization has reached a state where Grid
technology is substantially used for its PROCESS
Are you aware of initial or prior Grid-related
activities at site?
Are you aware of plans to use a Grid
environment for PROCESS within the next 12
months?
Is any Grid environment for PROCESS
currently being evaluated or trialled?
Are any Grid application development
projects for PROCESS planned, in progress,
implemented or cancelled?
Are more than 5% but less than 25% of the
business applications for PROCESS running
on a Grid?
Are more than 25% but less than 50% of the
business applications for PROCESS running
on a Grid?
PROCESS ¼ Asset management process/risk management process/new product development process.
Are more than 50% of the business
applications for PROCESS running on a Grid?

Mindfully resisting the bandwagon M Wolf et al
Table A3 Measurement scales for business processes performance second-order construct
BPP (reflective second-order) 7-point Likert (1 ¼ strongly disagree; 7 ¼ strongly agree)
Sources: Karimi et al. (2007a, b)
Process efficiency (reflective measures)
EFC1 Grid implementation has improved the efficiency of our PROCESS
EFC2 Grid implementation has lowered our costs in PROCESS
EFC3 Grid implementation has decreased the time-to-market of new financial products
due to an improved PROCESS
Process effectiveness (reflective
measures)
EFT1 Grid implementation has improved our effectiveness in PROCESS
EFT2 The functionalities of the Grid adequately meet the requirements of managing
PROCESS
EFT3 Grid implementation has improved our quality of managing PROCESS
Process flexibility (reflective measures)
FLX1 Grid implementation has given us more ways to customize our PROCESS
FLX2 Grid implementation has made our firm more agile due to an improved
PROCESS
FLX3 Grid implementation has made us more adaptive to a changing business
environment due to an improved PROCESS
FLX4 Grid implementation has improved the flexibility of our PROCESS
PROCESS ¼ Asset management process/risk management process/new product development process.
Table A4 Measurement scales for organizational mindfulness score
7-point Likert (1 ¼ strongly disagree; 7 ¼ strongly agree)
Reluctance to simplify
(reflective measures)
RS1 People are encouraged to express different views of the world
RS2 People listen carefully; it is rare that anyone’s view is dismissed
RS3 We appreciate sceptics
Commitment to resilience
(reflective measures)
CR1 People are encouraged to limit any negative consequences so that the firm can continue
operations in case of a mistake
CR2 People are known for their ability to use their knowledge in novel ways
CR3 People have a number of informal contacts that they sometimes use to solve problems
Deference to expertise
(reflective measures)
DE1 If something out of the ordinary happens, people know who has the expertise to respond
DE2 In this organization, the people most qualified to make decisions make them
DE3 People in this organization value expertise and experience over hierarchical rank
Operational sensitivity
(reflective measures)
OS1 Should problems occur, someone with the authority to act is always accessible and available,
especially to people on the front lines
OS2 During an average day, people come into enough contact with each other to build a clear picture
of the current situation
OS3 We have access to resources if unexpected surprises occur
Preoccupation with failure
(reflective measures)
PF1 We treat near misses and errors as information about the health of our system, and try to learn
from them
PF2 People are inclined to report mistakes that could have significant consequences, even if nobody
notices
PF3 People feel free to talk to superiors about problems
21

22
Table A5 Tests for equality of variance and normality among bootstrapping estimations
Relationship Leven’s test (test for equality of variance) Skewness and kurtosis test for normality
OM low (n ¼ 152) OM high (n ¼ 150)
CP-TMS 43.29 (0.00) 20.91 (0.00) 21.47 (0.00)
NP-TMS 71.45 (0.00) 5.87 (0.05) 5.97 (0.05)
MP-TMS 59.21 (0.00) 54.88 (0.00) 13.18 (0.00)
ET-TMS 17.06 (0.00) 0.12 (0.94) 44.57 (0.00)
ET MP-TMS 43.27 (0.00) 6.23 (0.04) 0.42 (0.81)
GIC-ASSM 1.17 (0.28) 21.12 (0.00) 22.83 (0.00)
GTI-ASSM 7.19 (0.01) 4.25 (0.12) 2.43 (0.30)
TIME-ASSM 0.30 (0.58) 20.34 (0.00) 30.36 (0.00)
SIZE-ASSM 32.71 (0.00) 62.27 (0.00) 3.04 (0.22)
TMS-ASSM 50.07 (0.00) 8.56 (0.01) 0.59 (0.75)
GIC-BPP 7.31 (0.01) 43.18 (0.00) 73.47 (0.00)
GTI-BPP 3.53 (0.06) 62.39 (0.00) 29.52 (0.00)
TIME-BPP 18.06 (0.00) 68.06 (0.00) 39.6 (0.00)
SIZE-BPP 11.56 (0.00) 53(0.00) 33.62 (0.00)
ASSM-BPP 1.12 (0.29) 1.17 (0.56) 1.32 (0.52)
ET-BPP 5.75 (0.02) 20.91 (0.00) 22.68 (0.00)
Leven’s test: P-values in parentheses measure the probability of inequality of variance among the two subgroups.
Skewness/Kurtosis test: if the P-value in parentheses is below 0.05, then the null hypothesis (i.e., that the corresponding variable is
normally distributed) is rejected at the 0.05 significance level.
Table A6 Tests for non-response bias based on archival and self-reported data
Indicator Definition and rationale nnon assimilators nassimilators ANOVA
(parametric)
Asset quality
Loan loss reserve/
gross loans
Net charge off/
average gross loans
Reserve for losses arising from risk
exposure, expressed as ratio to total
loans. The higher the ratio. The
poorer will be the quality of the loan
portfolio
Reflects the percentage of today’s
loans in comparison with loans that
have been written off. The lower this
figure the better
Capital structure
Equity/total assets A higher ratio reflects the ability of
the bank to withstand losses
Equity/net loans This ratio measures the equity
cushion available to mitigate losses
in the loan book
Equity/costs of
short-term funding
Mindfully resisting the bandwagon M Wolf et al
Measures the amount of permanent
funding relative to short-term,
potentially volatile funding
Operational performance
Price/earnings ratio This ratio represents investors’
collective opinion regarding a firm’s
future risk and opportunity mix
Earnings per share The portion of a company’s profit
allocated to each outstanding share
of common stock
Kruskal–Wallis
(non-parametric)
39 69 0.21 (0.08) 13.61 (0.02)
39 69 0.33 (0.13) 12.04 (0.06)
39 70 1.49 (0.04) 11.16 (0.08)
39 70 10.72 (0.32) 5.97 (0.34)
39 69 1.49 (0.25) 8.51 (0.18)
15 20 6.67 (0.09) 3.15 (0.23)
15 23 2.45 (0.23) 5.45 (0.11)

Table A6 Continued
Indicator Definition and rationale nnon assimilators nassimilators ANOVA
(parametric)
Net interest margin The higher this ratio, the cheaper
the funding or the higher the margin
the bank is commanding. Higher
margins and profitability are
desirable as long as the asset quality
is being maintained
Non op items and
taxes/average assets
Return on average
assets
Measure of the operating
performance of the bank before tax
and unusual items
Ratio shows how profitable
accompany’s assets are in generating
revenue
Liquidity
Net loans/total assets This liquidity ratio indicates what
percentage of the assets of the bank
are tied up in loans. The higher this
ratio, the less liquid the bank will be
Net loans/costs of
short-term funding
This loans-to-deposit ratio is a
measure of liquidity for which
higher figures denotes lower
liquidity
Kruskal–Wallis
(non-parametric)
39 71 0.21 (0.52) 0.41 (0.05)
39 71 0.13 (0.26) 7.67 (0.22)
39 71 0.01 (0.98) 0.76 (0.91)
39 70 1.74 (0.64) 1,1 (0.86)
39 70 12.29 (0.09) 3.21 (0.61)
All values are calculated based on archival data for 2008, extracted from the BankScope and COMPUSTAT databases.
Definitions of the asset quality, capital structure, operational performance and liquidity indicators were obtained from the BankScope and
COMPUSTAT databases.
ANOVA with Bonferroni test: mean (assimilators) – mean (non-assimilators), with P-values in brackets.
Kruskal–Wallis tests: rank mean (assimilators) – rank mean (non-assimilators), with P-values in brackets.
Table A7 Tests for non-response bias, based on archival and self-reported data
Group comparison: archival data
(n early, n late–between 10 and 30 observations)
Mindfully resisting the bandwagon M Wolf et al
Indicator ANOVA
(parametric)
Kruskal–Wallis
(non-parametric)
Group comparison: self-reported data
(n early ¼ 75, n late ¼ 75)
Question a ANOVA
(parametric)
Kruskal–Wallis
(non-parametric)
Asset quality MP1 0.26 (0.20) 7.94 (0.26)
Loan loss reserve/gross loans 0.52 (0.27) 3.48 (0.39) CP1 0.23 (0.39) 3.94 (0.58)
Net charge off /average gross loans 0.35 (0.32) 3.18 (0.43) NP1 0.16 (0.49) 1.94 (0.78)
Capital structure TMS1 0.28 (0.26) 7.26 (0.30)
Equity/total assets 1.42 (0.24) 4.53 (0.29) ASSM1 0.25 (0.45) 7.70 (0.27)
Equity/net loans 5.25 (0.77) 6.05 (0.15) ET1 0.15 (0.47) 1.86 (0.79)
Operations EFC1 0.11 (0.67) 3.47 (0.56)
Price/earnings ratio 6.57 (0.15) 2.48 (0.32) EFT1 0.29 (0.14) 8.25 (0.18)
Earnings per share 2.05 (0.17) 2.48 (0.33) FLX1 0.18 (0.44) 3.05 (0.61)
Net interest margin 0.45 (0.41) 1.66 (0.69) GIC1 0.08 (0.66) 2.42 (0.72)
Liquidity GTI1 0.07 (0.69) 5,54 (0.43)
Net loans/total assets 3.49 (0.49) 1.81 (0.66) TIME 0.68 (0.20) 12.89 (0.07)
Net loans/costs of short-term Funding 3.48 (0.55) 2.27 (0.59) SIZE 2437 (0.32) 8.68 (0.22)
a Because of space restrictions, we report only the results of the first question (i.e., indicator) for each construct.
ANOVA with Bonferroni test: mean (late responses) – mean (early responses), with P-values in brackets.
Kruskal–Wallis tests: rank mean (late responses) – rank mean (early responses), with P-values in brackets.
Archival data for 2008 were extracted from the BankScope and COMPUSTAT databases.
23

Layer 2:
E-Financial Markets & Market Infrastructures
(Prof. Dr. Peter Gomber)
� Gomber P., Pujol G., Wranik A. (2012):
Best Execution Implementation and Broker Policies in Fragmented European
Equity Markets
In: International Review of Business Research Papers, Vol. 8, Issue 2, 144-
162.
� Haferkorn M., Lutat M., Zimmermann K. (2012):
The Effect of Single-Stock Circuit Breakers on the Quality of Fragmented
Markets
In: FinanceCom 2012, Barcelona, Spain.
� Lattemann C., Loos P., Johannes G., Burghof H., Breuer A., Gomber P.,
Krogmann M., Nagel J., Riess R., Riordan R., Zajonz R. (2012):
High Frequency Trading - Costs and Benefits in Securities Trading and its
Necessity of Regulations
In: Business & Information Systems Engineering, Vol. 4, Issue 2, 93-108.
� Siering, M. (2012):
Investigating the Market Impact of Media Sentiment and Investor Attention
In: FinanceCom 2012; Barcelona, Spain.
� Weber M.C., Wondrak C. (2012):
Measuring the Influence of Project Characteristics on Optimal Software Project
Granularity
In: Proceedings of the 20 th European Conference on Information Systems
(ECIS), Barcelona, Spain.

Best Execution Implementation and Broker Policies in Fragmented
European Equity Markets
Peter Gomber Gregor Pujol Adrian Wranik1
From November 2007, the “Markets in Financial Instruments
Directive” (MiFID) has to be applied by investment firms and
regulated markets in Europe. Investment firms are obliged to make
provisions including processes and IT systems for order routing to
achieve the best possible result for clients in order execution. We
empirically investigate the implementation of best execution
obligations applying a longitudinal analysis of best execution policies
in 2008 and 2009 respectively. The European trading landscape has
changed as competition between established exchanges and new
trading venues has increased significantly. In both studies, 75
policies of German investment firms were analyzed to investigate
how best execution obligations have been implemented and whether
market fragmentation has been considered.
Field of Research: Financial Service and Banking Regulation, European Financial Markets.
1. Introduction
From November 2007, the “Markets in Financial Instruments Directive” (MiFID) has to be
applied by investment firms and regulated markets when providing investment services in
Europe. The central innovations of MiFID are the new classification of trading venues
(regulated markets, multilateral trading facilities (MTF), systematic internalisers), the definition
of “best execution” at a European level and transparency regulations for OTC-trading.
Investment firms are obliged to make adequate provisions including processes and IT systems
for order routing (“best execution arrangements”) to achieve the best possible result and to
disclose sufficient information of the most important measures to clients (“best execution
policies”). Although “best execution” and the associated duties constitute a legal obligation in
the relationship between clients and investment firms, at the economic level this topic also
decisively affects the interface between investment firms and execution venues.
Peter Gomber, Gregor Pujol, Adrian Wranik, Chair of e-Finance, E-Financelab, Goethe-University Frankfurt, Germany
email: {gomber | pujol | wranik} @wiwi.uni-frankfurt.de

Best execution is discussed from different perspectives in literature, e.g. how best execution
can be realized and measured. Macey and O’Hara (1996) analyzed legal and economic
aspects of the duty of best execution and recommended that best execution for a particular
trade is best achieved through competition between trading venues. However, McCleskey
(2004) suggested that best execution should be subject to regulation, as investors are not
capable to evaluate execution quality due to limited access to appropriate information. A
number of papers examined costs as a key aspect involved with best execution for a single
trading venue (Roll, 1984; Stoll, 1989) as well as between different markets (Huang & Stoll,
1996; De Jong et al., 1993).
Recent research increasingly focuses on the benefits of technology such as smart order
routing systems to achieve best execution. Foucault and Menkveld (2008) studied the
competition for order flow and concluded that transaction costs could be reduced if market
participants adopted smart order routing. Ende et al. (2009) developed a methodology to
assess advantages of dynamic routing and quantified the economic benefits of smart order
routing technology for European equities.
After the enforcement of MiFID, Hengelbrock and Theissen (2009) studied the market entry of
Turquoise, finding that the new MTF does not provide lower execution costs than primary
markets. Riordan et al. (2010) compared market quality of the London Stock Exchange (LSE)
against a number of MTFs. While quoted spreads are lower on the LSE, implicit transaction
costs measured as effective spreads are on average smaller on Chi-X, BATS, and Turquoise.
With MiFID, the European Regulator particularly intends to harmonize best execution
requirements on a European level. The new regime is characterized by a large diversity of
influencing factors (financial instruments, execution venues) that investment firms have to
consider in order execution processes and IT systems.
Earlier studies among 200 investment firms in Germany revealed that for most German
financial institutions MiFID is more of a regulatory burden than a chance to leverage
competitive potentials (Gomber et al., 2007). However, 32% of the investment firms answered
that competitive differentiation can be achieved through the design of best execution policies
and considered this aspect as having the highest chances of all services connected with
MiFID.
Thus, the purpose of this paper is to analyze the evolution of the implementation of best
execution processes and applied IT systems as documented in best execution policies after
MiFID entered into force. Different policies of German investment firms are analyzed and
compared over time. The study also checks how far the institutions' earlier assessment of the
competitive potential of order execution is reflected in best execution policies.
The motivation for a longitudinal analysis, i.e. the collection and analysis of the best execution
policies both in 2008 (right after the initial implementation of MiFID) and also in 2009 is twofold:
First, MiFID mandates a review of the investment firms’ policies at least on an annual
basis or whenever major changes occur, i.e. we can assure that investment firms investigated
and adapted their policies between the two points of analysis. Second, the European trading
landscape has experienced an intensified competition after the enforcement of the MiFID that
unfolded its impact on the market shares among trading venues between 2008 and 2009,
especially with the advent of new trading venues (MTF) offering serious alternatives for best
execution.

The next chapter provides some background on developments of European financial markets.
Chapter 3 presents the research questions and sample data. The main results are reported in
chapter 4. Chapter 5 concludes.
2. MiFID impact on European trading venues
MiFID has triggered a new competitive environment for equity trading and services. With the
proliferation of new MTFs such as Chi-X, Turquoise or BATS Europe, the number of trading
venues has substantially increased offering a wider range of trading options.
Two key trends can be observed: First, order flow is increasingly directed away from
incumbent exchanges towards new MTFs. Second, cost structures for trading and post-trading
activities in the European marketplace have changed significantly.
Due to MiFID, market shares have been shaken up and trading volumes have recognizably
shifted among execution venues. Figure 1 illustrates the changes for three important European
equity indices over time. While prior to MiFID domestic markets held market shares close to
100%, new trading platforms gradually have gained importance, e.g. for FTSE-100 the LSE’s
market share accounts for just 50%.
Figure 1: Market shares of major European indices over time (Fidessa, 2010)
DAX30
CAC40
FTSE100
July 2008
1st July 2008
1 MiFID study
st MiFID study
Xetra
Chi-X
Paris
Chi-X
LSE
Chi-X
FTSE100
4%
4%
July 2009
2nd July 2009
2 MiFID study
nd MiFID study
Xetra
Chi-X
Turquoise
Bats Europe
Others
Paris
Chi-X
Amsterdam
Bats Europe
Turquoise
Others
LSE
Chi-X
Turquoise
Bats Europe
Others
April 2010
Xetra
Chi-X
Bats Europe
Turquoise
Others
Paris
Chi-X
Amsterdam
Bats Europe
Turquoise
Others
LSE
Chi-X
Bats Europe
Turquoise
Nasdaq Europe
In 2009, the European Commission commissioned a study (Oxera, 2009) to monitor fee
developments across 18 European markets. One key result addresses the explicit costs for
trading services that on average have decreased along the entire trading value chain.
Others
time

Figure 2: Change in explicit costs over time (2006-2008) (Oxera, 2009)
80
60
40
20
0
-20
-40
-60
-80
-100
Trading platforms Central Counterparties Central Securities
(CCPs) Depositories (CSD)
�� Major financial centres: Germany, France, UK, Italy, Switzerland and Spain
�� Secondary financial centres: Belgium, Luxemburg, Netherland, Norway, Poland and Sweden
▲ Other financial centres: Denmark, Greece, Ireland, Austria, Portugal und Tschech Republic
Figure 2 highlights changes for different groups of European markets from 2006 to 2008.
Across all groups, average fees charged by trading platforms as well as central counterparty
clearing fees have significantly reduced. Fees charged by Central Securities Depositories do
not reveal a systematic trend.
3. Research questions and data
MiFID does not prescribe how investment firms have to implement best execution obligations.
On the one hand they need to establish adequate internal provisions for business processes
and IT (best execution arrangements), on the other hand sufficient information regarding these
arrangements has to be communicated to clients (best execution policies). As only latter
information is publicly disclosed, the subsequent analysis focuses on publicly available data
(best execution policies).
The new European concept of best execution in securities trading and its practical
implementation raise a number of important research questions: (i) How do firms implement
best execution obligations from a process and IT perspective and which criteria are used for
order routing? (ii) Do best execution policies implement the mandatory legal requirements of
MiFID and do they reflect increased competition among trading venues over time? (iii) Which

execution venues are preferred and how does this preference evolve due to the changed
competitive landscape?
The paper takes four steps to address these research questions:
First, it analyses which option the investment firms selects to implement best execution
requirements. It is investigated whether a static rule framework applying historical data only is
used to determine the venues for order execution or whether a dynamic order routing process
is applied that relies on sophisticated order routing software and real time market data to
determine which execution venue is able to provide the best result at the time of order
submission. This analysis of static versus dynamic best execution policies is directly linked to
the criteria an investment firm takes into account for the selection of an execution venue and
therefore these criteria are analyzed in this context. Second, the best execution policies are
investigated as a whole and compared with each other regarding the level of fulfillment of the
mandatory legal requirements over time. Third, the policies were analyzed with regard to the
existence of a ranking for specific client categories or classes of instruments.
As MiFID had to be transformed into individual national legislations by all European member
states, a consistent view can only be achieved for one member state. For both studies
Germany was selected as it represents the largest European economy. For the purpose of
investigating the changes of the best execution implementation and the reaction of investment
firms on the recent developments in the European trading landscape both studies use an
identical sample of firms’ best execution policies in Q2/2008 and Q3/2009 respectively.
Various channels were used for obtaining the policies, e.g. by data collection over the Internet,
by email or telephone contacts.
The study is based on the 100 largest German financial institutions in terms of total assets in
2006 (Karsch, 2007) and the 15 largest online brokers according to number of security
accounts (Kundisch & Holtmann, 2008). The list of the 100 largest institutions was adjusted by
removing the companies which do not provide investment services reducing the number to 63.
Since in 2008, three best execution policies were not made available, the final sample covers
75 best execution policies.
In the following, results of the data analysis in 2009 are presented and compared to the
findings in 2008. In subsequent figures and tables, the results of the data analysis in 2009 are
marked in bold, the comparative values of 2008 are indicated in parentheses. While in the text,
percentage values always refer to the total sample size (75 policies), absolute figures are used
to explain findings of a sub-sample.
4. Results
4.1 Implementation of Best Execution Processes
Best execution policies are a major part of the provisions which firms must make in order to
ensure that they can regularly execute orders in the best possible manner. The most important
legal requirements of MiFID implementation in Germany are specified in the German
Securities Trading Act (WpHG) i and the Ordinance Specifying Rules of Conduct and
Organisation Requirements for Investment Firms (WpDVerOV). ii

Although MiFID lacks guidance on how to implement the best execution obligation, two basic
concepts can be distinguished: Firms may apply a static approach, i.e. the decision to route
client orders to a particular venue is based on a pre-defined rule framework taking into account
different criteria such as client category (retail or professional client), order types and sizes and
classes of instruments (shares, bonds, derivatives). The Committee of European Securities
Regulators (CESR) suggests a minimum level of differentiation (CESR, 2007) by distinguishing
between different client categories and classes of instruments. The result of such a rule
framework typically leads to one particular venue that provides the best possible result for a
particular client category, class of financial instrument or a combination of both based on
historical data.
Alternatively, the appropriate execution venue is selected by applying a dynamic approach
which is an implementation that exceeds MiFID minimum legal requirements: Each order is
treated in an individual manner considering real time market data for the order routing
decision. Such provisions enable a real-time evaluation and support a dynamic allocation of
the individual order to the venue offering the best conditions at the time of order entry.
Considering the increased competition in Europe such a real time comparison allows an
optimized selection between venues improving execution quality but it obviously leads to
higher costs of implementation.
Figure 3: Static vs. dynamic best execution approach
Static
approach
Exchange Exchange Exchange Exchange Exchange Exchange Exchange AA
AA
A AA
A
Buy Buy 100 100 100 Daimler
Daimler
Daimler
Client Client order order order order order order
order order order order order
- Instrument: Daimler
- Index: DAX
- Quantity: 100
- Buy/Sell: Buy
Pre Pre-defined
Pre Pre-defined
Pre
Pre defined defined rule rule
rule framework framework framework framework
framework
framework
Domestic Domestic/ Domestic Domestic/ Domestic Instrument
Instrument Execution
Execution
Category
Category
abroad
abroad class
class class class venue
venue
Shares domestic DAX A
Shares domestic MDAX B
… …
Buy 100 @ 40,80 €
Exchange B
Buy 100 @ 40,70 €
Buy 100 @ 40,70 €
Routing Routing routines
routines
routines
Smart Smart Order Order Order Order Order Order
Order
Routing
Routing
System
System
MTF A
real real-time real real-time real real real real time time market market
market data
data
data
Dynamic
approach
Buy 100 @ 40,75 €
Figure 3 illustrates both concepts. While in the static approach (dashed box) client orders are
routed to the best execution venue (here: Exchange A) based on a pre-defined rule framework,
the dynamic approach (solid box) applies real-time market data in the decision process. Based

on routing routines of the smart order routing system the best venue (here: Exchange B) is
detected at order entry and subsequently the order is submitted to that venue for execution.
Relevant Criteria and Implementation of best execution processes
For both concepts, the relevant criteria for achieving the best possible result are derived
directly from §33a (2) WpHG, in accordance to which “in particular the prices of the financial
instruments, the costs involved in order execution, the speed, the probability of execution and
the processing of the order, as well as the scope and type of order” must be taken into account
as criteria. It was checked whether the relevant criteria had been weighted in the best
execution policies (Table 1).
No weighting can be recognized in 8.0% (2008: 10.7%) (absolute: 6 (2008: 8)) of the policies.
In two cases no details are provided, in the other four cases the criteria have not been
prioritized. Weighting of relevant criteria can be recognized in 92.0% (89.3%) of policies
(absolute: 69 (67)). In 13 (12) policies percentage values (e.g. price: 80%, external costs: 20%)
are named for individual criteria; in 56 (55) policies a ranking, e.g. price priority over speed,
can be observed. In both studies only one policy provides a recognizable ranking of criteria
and uses real-time market data to identify the execution venue providing the best result for the
individual order.
Table 1: Recognizability of the weighting of the relevant criteria and implementation of best
execution processes
Number of policies evaluated 75 (75)
No details 2 (2)
No ranking of criteria recognizable 4 (6)
Ranking of the criteria recognizable and
dynamic order execution (real time)
1 (1)
Ranking of the criteria recognizable 55 (54)
Percentage weighting of the criteria 13 (12)
No recognizable
weighting
Recognizable
weighting
6 (8)
69 (67)
Concerning the different implementation options for best execution processes and systems,
Table 1 shows an important result: For both data analyses (2008 and 2009) it turns out that an
overwhelming majority of firms prefers a static implementation approach for best execution and
does not utilize the competitive potential stated in recent studies.
4.2 Legal Requirements for the Best Execution Policies
According to §33a (6) No. 1 WpHG an investment firm must inform “its clients of its best
execution policy before providing investment services for the first time and obtain the clients'
acceptance of its policy.”
Specific minimum requirements with regard to the scope and design of the contents are linked
to this obligation specified in §33a (5) and (6) WpHG in conjunction with §11 (4) WpDVerOV:
investment firms must show how they aim to achieve the best possible result for order
execution in various categories of financial instruments and the crucial factors for selecting a

venue must be named, and finally a list containing at least those venues must be provided
which can be considered as consistently achieving the best possible results.
Content of the law
Description of the weighting defined for
the relevant criteria
or
description of the method which is used
for weighting
Details of the various execution venues
for each category of financial instrument
Details of the crucial factors for selecting
an execution venue
List of the principal execution venues
which consistently achieve the best
possible result when executing client
orders
Information for private clients in the case
of orders with client instructions
Releasing the investment firms from the
obligation to execute an order in
accordance with the best execution policy
Figure 4: Legal requirements
Policies: 2009: 75 (2008): 75
No details
13 (12)
% - value
56 (55) 4(6) 2 (2)
ranking
method
Yes No
71 (71) 4 (4)
73 (73)
65 (64)
10 (11)
2 (2)
71 (71) 4 (4)
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
Figure 4 shows the results for the legal requirements which. It displays the requirements and
the levels of fulfilment.
Description of the Weighting Implemented or Description of the Method
§ 33a (6) No. 1 WpHG in conjunction with §11 (4) No. 1 WpDVerOV requires that policies must
contain either a “description of the weighting implemented for the relevant criteria to achieve
the best possible result” or “a description of the method which is used for this weighting.” In the
analysis a distinction is made between three examination criteria (percentage, ranking,
method). Policies which contain at least information on a particular procedure are
acknowledged to have a method (e.g. “the bank assumes that their clients would like to
achieve the best price”). In the policies both analysed in 2009 and 2008, this applies for nearly
all investment firms (97.3%) except for two which did not provide any details.
It was investigated whether the weighting defined was expressed as a percentage value or by
a particular ranking. In 17.3% (16.0%) of the best execution policies concrete percentage
values are specified for individual criteria (e.g. price: 80%; external costs: 20%), in 74.4%
(73.3%) a ranking, e.g. price has priority over speed, is provided.
Details of Execution Venues for each Category of Financial Instrument
§33a (5) No. 1 WpHG (first half sentence) requires “details of the various execution venues
with regard to each category of financial instrument.” This provision specifies that a category
must be assigned to the venue(s) when a policy is created. The following variants of policies
exist: either precisely one venue is specified for one category or several venues are specified

for one category or several venues are specified for multiple categories simultaneously. In
94.6% of the policies an assignment of category to venue can be recognized.
Details of the Crucial Factors for Selecting an Execution Venue
In addition in accordance with §33a (5) No. 1 WpHG, “the crucial factors for selecting an
execution venue” must also be specified. In the analysis, information was taken into account
whether documents reveal that the firms used various factors for the evaluation (e.g. “in
particular the recognizable factors price and costs which arise through execution at an
execution venue are used for the evaluation.”). Like in 2008, nearly all policies, i.e. 97.3%,
provided comprehensive details.
List of the Principal Execution Venues
A further obligation derives from §33a (5) No. 2 WpHG in conjunction with §11 (4) No. 2
WpDVerOV according to which a “list of the principal execution venues [�] at which the
investment firms can consistently achieve the best possible results when executing client
orders” must be contained in the policies. In 86.7 % (85.3%) of the policies examined in 2009
this list is provided either as a text list or as a table in the appendix.
Information for Private Clients about Execution in accordance with Instructions
Finally, in accordance with §33a (6) No. 2 WpHG in conjunction with §11 (4) No. 3 WpDVerOV
firms must inform “private clients expressly that when instructed by the client the investment
firm will execute the order according to these client instructions and will therefore not be
obliged to execute the order in accordance with its best execution policy to achieve the best
possible result.” Both in the 2009 and the 2008 analysis, this information is clearly emphasized
in most execution policies, i.e. in 94.7%.
4.3 Analysis of the Execution Venues Specified
Finally, details regarding the ranking of the venues were examined. Obviously, from a
competitive view, this ranking is highly relevant for execution venues. Therefore, it is of interest
how policies list and rank the different venues. Table 2 shows the distribution of the
nominations for various securities groups, classified according to whether they can be traded
domestically or abroad. It is noticeable that firms primarily prefer abstract and summarizing
descriptions to document their choice of a venue (e.g. domestic execution venue, foreign
exchange) instead of naming specific venues. In addition to the abstract details relating to the
venues, the table also includes special cases, such as forwarding, fixed-price business and
also obligations to provide instructions. A concrete venue is named only in every fourth policy
(26.6%; 2008: 24%).
Table 2: Execution venues for different securities groups

Tradable
Tradable
domestically
domestically
Shares Shares Bonds
Bonds
Certified
Certified
derivatives
derivatives
Investment
Investment
shares
shares
ETF
ETF
Other
Other
securities
securities
Other
Other
financial
financial
instruments
instruments
Domestic
execution venue
42 (43) 42 (43) 41 (42) 5 (5) 4 (4) 11 (10) 4 (4)
Domestic floor
trading system
2 (4) 4 (5) 4 (5) 1 (1) 4 (4) 0 (0) 0 (0)
Domestic
exchange
7 (6) 11 (10) 12 (11) 6 (6) 9 (9) 5 (5) 6 (6)
Domestic home
exchange
0 (0) 3 (2) 2 (2) 0 (0) 0 (0) 0 (1) 1 (1)
Others (e.g.
forwarding)
1 (1) 3 (4) 4 (5) 1 (1) 0 (0) 1 (1) 1 (1)
Fixed-price
business
1 (1) 2 (3) 0 (1) 0 (0) 0 (0) 2 (3) 0 (0)
Instructions 0 (0) 1 (1) 1 (1) 46 (46) 2 (2) 34 (34) 35 (35)
No details 2 (2) 1 (1) 3 (3) 13 (13) 49 (48) 19 (19) 9 (9)
Not possible 0 0 (0) (0) 0 0 (0) (0) 0 0 (0) (0) 2 2 (2) (2) 0 0 (1) (1) 0 0 (0) (0) 1 1 (1)
(1)
Specific
execution venue
20 (18) 8 (6) 8 (5) 1 (1) 7 (7) 3 (3) 18* (18)
Total 75 (75) 75 (75) 75 (75) 75 (75) 75 (75) 75 (75) 75 (75)
Tradable
abroad
Shares Bonds
Certified
derivatives
Investment
shares
ETF
Other
securities
Other
financial
instruments
Foreign
execution venue
2 (2) 5 (5) 3 (3) 1 (1) 1 (1) 2 (2) 13 (12)
Foreign exchange 27 (27) 18 (17) 17 (17) 0 (0) 0 (0) 11 (10) 1 (1)
Others (e.g.
forwarding)
4 (4) 8 (6) 8 (8) 2 (2) 2 (2) 2 (2) 2 (2)
Fixed-price
business
1 (2) 1 (1) 0 (0) 0 (0) 0 (0) 0 (1) 1 (2)
Instructions 36 (35) 34 (36) 34 (34) 45 (46) 45 (46) 34 (34) 34 (34)
No details 4 (4) 9 (10) 13 (13) 25 (24) 25 (24) 26 (26) 23 (23)
Not possible 0 (0) 0 (0) 0 (0) 2 (2) 2 (2) 0 (0) 1 (1)
Specific
execution venue
1 (1) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0)
Total 75 (75) 75 (75) 75 (75) 75 (75) 75 (75) 75 (75) 75 (75)
* 16 (16) nominations here for Eurex
For the categories of financial instruments which can be traded domestically, the abstract label
“domestic execution venue” is chosen most frequently for the first three securities groups
(shares, bonds, certified derivatives). For investment shares, other securities and other
instruments orders are mostly accepted with explicit instructions only. Firms most frequently
provide - with 20 (18) nominations - a specific execution venue for the securities group shares.
As in 2008, the securities group other financial instruments was named 18 times, thereof 16
nominations for Eurex. The term “foreign exchange” is frequently used for financial instruments
which can be traded abroad; with one exception, no specific details are provided here. It is
noticeable that for all securities groups order executions in most cases are only possible with
instructions or no details about the execution venues are provided at all.
Due to the few specific details about the execution venues and their ranking, a more detailed
analysis was only performed for the securities group shares.
In Table 4, 20 (18) policies are listed which at least named one specific venue for this
securities group and documented a recognizable ranking.
Table 3: Segmenting of the securities group shares

Segment Segment DAX DAX 30
30
Other Other DAX
DAX
(MDAX, (MDAX, TECDAX,
TECDAX,
SDAX)
SDAX)
EUROSTOXX EUROSTOXX 50,
50,
DJ DJ STOXX STOXX 40,
40,
NASDAQ NASDAQ 100
100
Other Other domestic domestic shares
shares
Execution venue Freq. Rank 1 Rank 2 Freq. Rank 1 Rank 2 Freq. Rank 1 Rank 2 Freq. Rank 1 Rank 2
Xetra-Best 4 (4) 4 (4) 0 (0) 4 (4) 4 (4) 0 (0) 3 (3) 3 (3) 0 (0) 1 (1) 1 (1) 0 (0)
Xetra 11 (14) 6 (8) 5 (5) 13 (14) 6 (8) 7 (5) 8 (10) 4 (5) 4 (4) 9 (11) 5 (6) 4 (4)
Berlin 1 (2) 0 (0) 1 (1) 1 (2) 0 (0) 1 (1) 2 (2) 1 (1) 1 (0) 1 (2) 0 (0) 1 (1)
Düsseldorf 3 (2) 0 (0) 3 (1) 1 (2) 0 (0) 1 (1) 2 (2) 1 (1) 1 (0) 1 (2) 0 (0) 1 (1)
Frankfurt 5 (4) 2 (0) 3 (3) 5 (4) 2 (0) 3 (3) 5 (4) 2 (1) 3 (2) 6 (4) 3 (2) 3 (2)
Hamburg 3 (2) 3 (1) 0 (0) 1 (2) 1 (1) 0 (0) 3 (2) 3 (2) 0 (0) 1 (2) 1 (2) 0 (0)
Hannover 2 (1) 2 (0) 0 (0) 0 (1) 0 (0) 0 (0) 0 (1) 0 (1) 0 (0) 0 (1) 0 (1) 0 (0)
Munich 1 (2) 0 (0) 1 (2) 1 (2) 0 (0) 1 (2) 1 (2) 0 (0) 1 (2) 1 (2) 0 (0) 1 (2)
Stuttgart 1 (2) 0 (0) 1 (1) 3 (2) 2 (0) 1 (1) 4 (2) 1 (1) 3 (0) 3 (2) 2 (0) 1 (1)
Tradegate 1 (1) 1 (1) 0 (0) 1 (1) 1 (1) 0 (0) 1 (1) 1 (1) 0 (0) 1 (1) 1 (1) 0 (0)
OTC 1 (1) 1 (1) 0 (0) 1 (1) 1 (1) 0 (0) 1 (1) 1 (1) 0 (0) 1 (1) 1 (1) 0 (0)
Domestic floor
trading system
3 (3) 0 (0) 3 (3) 3 (3) 0 (0) 3 (3) 4 (4) 1 (1) 3 (3) 5 (4) 2 (2) 3 (2)
Domestic home
exchange
4 (2) 0 (0) 4 (2) 4 (2) 0 (0) 4 (2) 3 (2) 0 (0) 3 (2) 3 (1) 0 (0) 3 (2)
Fixed-price
business
3 (3) 3 (3) 0 (0) 3 (3) 3 (3) 0 (0) 3 (3) 3 (3) 0 (0) 3 (3) 3 (3) 0 (0)
No details 3 Execution venue Freq. Rank 1 Rank 2 Freq. Rank 1 Rank 2 Freq. Rank 1 Rank 2 Freq. Rank 1 Rank 2
Xetra-Best 4 (4) 4 (4) 0 (0) 4 (4) 4 (4) 0 (0) 3 (3) 3 (3) 0 (0) 1 (1) 1 (1) 0 (0)
Xetra 11 (14) 6 (8) 5 (5) 13 (14) 6 (8) 7 (5) 8 (10) 4 (5) 4 (4) 9 (11) 5 (6) 4 (4)
Berlin 1 (2) 0 (0) 1 (1) 1 (2) 0 (0) 1 (1) 2 (2) 1 (1) 1 (0) 1 (2) 0 (0) 1 (1)
Düsseldorf 3 (2) 0 (0) 3 (1) 1 (2) 0 (0) 1 (1) 2 (2) 1 (1) 1 (0) 1 (2) 0 (0) 1 (1)
Frankfurt 5 (4) 2 (0) 3 (3) 5 (4) 2 (0) 3 (3) 5 (4) 2 (1) 3 (2) 6 (4) 3 (2) 3 (2)
Hamburg 3 (2) 3 (1) 0 (0) 1 (2) 1 (1) 0 (0) 3 (2) 3 (2) 0 (0) 1 (2) 1 (2) 0 (0)
Hannover 2 (1) 2 (0) 0 (0) 0 (1) 0 (0) 0 (0) 0 (1) 0 (1) 0 (0) 0 (1) 0 (1) 0 (0)
Munich 1 (2) 0 (0) 1 (2) 1 (2) 0 (0) 1 (2) 1 (2) 0 (0) 1 (2) 1 (2) 0 (0) 1 (2)
Stuttgart 1 (2) 0 (0) 1 (1) 3 (2) 2 (0) 1 (1) 4 (2) 1 (1) 3 (0) 3 (2) 2 (0) 1 (1)
Tradegate 1 (1) 1 (1) 0 (0) 1 (1) 1 (1) 0 (0) 1 (1) 1 (1) 0 (0) 1 (1) 1 (1) 0 (0)
OTC 1 (1) 1 (1) 0 (0) 1 (1) 1 (1) 0 (0) 1 (1) 1 (1) 0 (0) 1 (1) 1 (1) 0 (0)
Domestic floor
trading system
3 (3) 0 (0) 3 (3) 3 (3) 0 (0) 3 (3) 4 (4) 1 (1) 3 (3) 5 (4) 2 (2) 3 (2)
Domestic home
exchange
4 (2) 0 (0) 4 (2) 4 (2) 0 (0) 4 (2) 3 (2) 0 (0) 3 (2) 3 (1) 0 (0) 3 (2)
Fixed-price
business
3 (3) 3 (3) 0 (0) 3 (3) 3 (3) 0 (0) 3 (3) 3 (3) 0 (0) 3 (3) 3 (3) 0 (0)
No details 2 (3) 0 (0) 2 (3) 2 (3) 0 (0) 2 (3) 6 (6) 2 (1) 4 (5) 7 (6) 1 (0) 6 (6)
Total 22¹ (18) 23² (21)² 20 (18) 23² (21)² 23² (22)¹² 23² (18) 20 (19)¹ 23² (21)²
1 The execution venues Hamburg / Hanover were counted twice in Rank 1
2 The execution venues Frankfurt/ Stuttgart/ Düsseldorf/ Berlin were counted four times in Rank 1 or 2
3 Some best execution policies do not contain specific details for every segment: for example, in the segment
„EUROSTOXX 50“ the two (one) nomination(s) for Rank 1 mean(s) that two (one) investment firm provided no details
about order execution in this segment, but did document a ranking for the other segments (e.g. DAX 30). As regards the
four (five) nominations for Rank 2: the best execution policies only envisage one specific execution venue being ranked
first, but no further information is providedabout alternative execution venues in the following ranks.
3 2 (3) 0 (0) 2 (3) 2 (3) 0 (0) 2 (3) 6 (6) 2 (1) 4 (5) 7 (6) 1 (0) 6 (6)
Total 22¹ (18) 23² (21)² 20 (18) 23² (21)² 23² (22)¹² 23² (18) 20 (19)¹ 23² (21)²
1 The execution venues Hamburg / Hanover were counted twice in Rank 1
2 The execution venues Frankfurt/ Stuttgart/ Düsseldorf/ Berlin were counted four times in Rank 1 or 2
3 Some best execution policies do not contain specific details for every segment: for example, in the segment
„EUROSTOXX 50“ the two (one) nomination(s) for Rank 1 mean(s) that two (one) investment firm provided no details
about order execution in this segment, but did document a ranking for the other segments (e.g. DAX 30). As regards the
four (five) nominations for Rank 2: the best execution policies only envisage one specific execution venue being ranked
first, but no further information is providedabout alternative execution venues in the following ranks.
Specifically, the segments iii DAX30, other DAX, EUROSTOXX 50, DJ STOXX 40, NASDAQ
100 and other domestic shares were analyzed.
The venue mentioned most frequently in all segments is Xetra. In some policies precisely one
venue is prioritized and occupies rank 1; in these cases no specific nomination for rank 2
exists - a “domestic floor trading system” or “domestic home exchange” was ranked second.
This is why these abstractly mentioned execution venues are also contained in the table. As
execution venues of equal rank are also named twice or four times in the policies, these
multiple nominations result in a value greater than 20 (18) in the bottom line. In three execution
policies order execution as fixed-price business is placed in Rank 1. iv It is noticeable that - like
in 2008 - regional exchanges are rarely placed in rank 1.
The analysis reveals a remarkable finding: The substantial changes of the European trading
landscape triggered by MiFID are not reflected at all in any best execution policy analyzed in
the data of 2009. This becomes apparent in the minor changes identified in Table 4 and the
fact that no new venue (MTF) is considered in any best execution policy.
5. Summary and Outlook
This paper compares the implementation of European best execution obligations in Germany
based on analyses of best execution policies in Q2/2008 and Q3/2009. The analysis aims at
identifying changes in the firm’s policies as a result of the mandatory annual reviews and
investigates whether recent developments in the trading landscape are reflected.
We analysed 75 policies of the 100 largest financial institutions and of the 15 largest online
brokers in Germany. 59% (44) policies remain unchanged compared to 2008. 41% (31) of the
firms provided an updated version of their policy.
The key message of the study is that the results of the two data analyses are largely in line
and no substantial changes can be identified although the European securities markets were

affected by major changes concerning the competitive landscape and the number of available
venues. In both years, the policies show that the minimum legal requirements have
recognizably been implemented. Like in 2008, in 2009 only one policy mentions the use of a
dynamic order execution approach, i.e. applying real time market data to achieve best
execution. Furthermore, a significant heterogeneity can be recognized between the policies:
some are extremely comprehensive and describe the selected procedure in great detail, while
others are limited to minimum details and are not very meaningful for clients.
This also applies for the details about the venues: Only in 26.6% (24.0 %), i.e. in approximately
every fourth policy, the venues are named specifically or a ranking is provided.
In earlier studies (Gomber et al., 2007), best execution principles are most frequently named
as key competitive factor. However, in summarizing it must be noted that the use of policies as
competitive instrument cannot be recognized in a large majority of German financial
institutions. This is surprising given the intensified competition between incumbent and new
trading venues resulting in market shares shifts and cost reductions. These changes are not
reflected in any policy so far. Orders are still primarily executed on domestic markets or (if only
tradable abroad) on the foreign exchange respectively. Furthermore, MTFs are not considered
at all in any best execution policy.
As policy implication of our research and as an input for the currently ongoing MiFID review of
the EU Commission (with a final new regulation scheduled for 2012), it must be noted that
these changes might possibly be implemented by firms internally but they are not listed
explicitly in the policies. Nevertheless, from the client's viewpoint it is desirable that these
analyses and their concrete results should be communicated in a more transparent form so
that the policies can function not just as regulatory necessity but also as driver of competition
between firms and also between venues. For future research, both analyses serve as a basis
for benchmarking best execution policies over longer time periods, e.g. to monitor whether
firms change their processes and systems from a static towards a dynamic approach for order
routing and execution. Also, studies regarding the client benefits of offering access to the new
MTFs or analyses regarding the MiFID perception from a client perspective and its potential
implications in terms of changes in reputation and performance of investment firms represent
challenging research topics.
6. References
CESR 2007. Best Execution under MiFID, Questions & Answers, Ref CEST/07-320.
De Jong, F., Nijman, T. and Roell, A. 1993. A comparison of the cost of trading French shares
on the Paris Bourse and on Seaq International, European Economic Review, 39, 7, pp.
1277-1301.
Ende, B., Gomber, P. and Lutat, M. 2009. Smart Order Routing Technology in the New
European Equity Trading Landscape, Software Services for e-Business and e-Society,
9th IFIP WG 6.1 Conference, I3E 2009; Proceedings, pp. 197-209, Springer, Boston.
Ende, B., Gomber, P., Lutat, M. and Weber, M.C. 2010. A methodology to assess the benefits
of smart order routing, Proceedings of IFIP I3E 2010 conference.
Fidessa 2010. Fidessa Fragmentation Index May 2008 until April 2010,
http://fragmentation.fidessa.com, accessed 04th May 2010.

Foucault, T. and Menkveld, A. 2008. Competition for Order Flow and Smart Order Routing
Systems, Journal of Finance, 63, 1, pp. 119-158.
Gomber, P. and Chlistalla, M. 2008. Implementing MiFID by European execution venues –
Between threat and opportunity, Journal of Trading, Spring 2008, 3, 2, pp. 18-28.
Gomber, P., Chlistalla, M., Gsell, M. and Pujol, G., Steenbergen, J. 2007. Umsetzung der
MiFID in Deutschland - Empirische Studien zu Status Quo und Entwicklung der MiFID-
Readiness der deutschen Finanzindustrie, Books on Demand.
Hengelbrock, J. and Theissen, E. 2009. Fourteen at One Blow: The Market Entry of Turquoise
(December 31, 2009). Available at SSRN: http://ssrn.com/abstract=1570646.
Huang, R.D. and Stoll, H.R. 1996. Dealer versus auction markets: A paired comparison of
execution costs on NASDAQ and the NYSE, Journal of Financial Economics, 41, 3, pp.
313-357.
Karsch, W. 2007. Top 100 der deutschen Kreditwirtschaft: Auf Wachstumskurs, Die Bank,
2007, 8, pp. 34-37.
Kundisch, D. and Holtmann, C. 2008. Competition of Retail Trading Venues – Onlinebrokerage
and Security Markets in Germany. in: F. Schlottmann, D. Seese, C. Weinhardt, (Eds.),
Handbook on Information Technology in Finance, pp. 171-192, Springer, Berlin.
Macey, J. and O’Hara, M. 1996. The Law and Economics of Best Execution, Journal of
Financial Intermediation, 6, 3, pp. 188-223.
McCleskey, S. 2004. Achieving Market Integration - Best Execution, Fragmentation and the
Free Flow of Capital, Butterworth-Heinemann, Elsevier.
Oxera 2009. Monitoring prices, costs and volumes of trading and post-trading services, report
prepared for European Commission DG Internal Market and Services, July 2009.
Riordan, R., Storkenmaier, A. and Wagener, M. 2010. Fragmentation, Competition and Market
Quality: A Post-MiFID Analysis (June 18, 2010), available at SSRN:
http://ssrn.com/abstract=1626711.
Roll, R. 1984. A simple implicit measure of the bid-ask spread in an efficient market, Journal of
Finance, 39, 4, pp. 1127-1139.
Stoll, H.R. 1989. Inferring the components of the bid-ask spread: theory and empirical tests,
Journal of Finance, 44, 1, pp. 115-134.
i The WpHG in the version of 01 November 2007, amended by the law for implementing the Markets in Financial
Instruments Directive (DIR 2004/39/EG, MiFID) and the implementing directive (DIR 2006/73/EG) of the Commission (law
for implementing the financial markets directive) of 16 July 2007, Federal Law Gazette I 2007, 1330 of 19 July 2007
ii Ordinance Specifying Rules of Conduct and Organisation Requirements for Investment Firms (WpDVerOV) of 20 July
2007, Federal Law Gazette I 2007, p. 1432 of 23 July 2007
iii A segment is divided into three columns: the frequency (freq.) specifies how often an execution venue was named in total.
Rank 1 and Rank 2 specify the number of times the execution venue concerned was nominated for this rank. The total of
Rank 1 and Rank 2 does not necessarily match the frequency because nominations from Rank 3 and above were taken into
account, but for reasons of clarity have not been included in this table. Thus, for example, Xetra for the DAX 30 values is
listed six (eight) times in Rank 1 and five (five) times in Rank 2; two (one) further nomination(s) were (was) registered, but
with regard to the prioritization this was counted in one of the lower ranks.
iv In these cases investment firms offer order execution primarily as fixed-price business. However, if fixed-price business
does not come about, the order is directed to a concrete execution venue (e.g. Xetra) which occupies Rank 2.

The Effect of Single-Stock Circuit Breakers on
the Quality of Fragmented Markets
Abstract: Since the May 6 th , 2010 flash crash in the U.S., appropriate
measures ensuring safe, fair and reliable markets become more relevant from
the perspective of investors and regulators. Circuit breakers in various forms
are already implemented for individual markets to ensure price continuity and
prevent potential market failure and crash scenarios. However, coordinated
inter-market safeguards have hardly been adopted, but are essential in a
fragmented environment to prevent situations, where main markets halt trading
but stock prices continue to decline as traders migrate to satellite markets. The
objective of this paper is to study empirically the impact of circuit breakers in
a single market and an inter-market setup. We find a decline in market
volatility after the trading halt at the home and satellite market, but at the cost
of higher spreads. Moreover, the satellite market’s quality and price discovery
during CBs are sorely afflicted and only restore as the other market restarts
trading.
Keywords: Circuit Breaker, Electronic Trading, Exchanges, Market
Coordination, Market Fragmentation
1. Introduction
Fragmentation of investors’ order flow has been a long-time phenomenon in U.S.
equity markets. Competition in European equity markets started in 2007 due to the
introduction of the Markets in Financial Instruments Directive (MiFID) that enabled
new venues, called multilateral trading facilities (MTF), to compete with established
exchanges. Additionally, new technologies such as high frequency trading and smart
order routing emerged on both sides of the Atlantic taking a significant share of the
total trading volume. Against the background of these developments and the May 6 th ,
2010 flash crash in the U.S., appropriate measures to ensure safe, fair and reliable
electronic markets are becoming more relevant. Circuit breakers (CBs) in various
forms are already implemented. However, coordinated measures between different
market centers seem to become more and more relevant in the light of those recent
developments. Academic research on the coordination of CBs in fragmented markets
is still scarce, but would provide relevant input to market participants, exchange
operators and regulators. In this paper we fill the gap by empirically investigating

CBs and their effects on liquidity, volatility and price discovery in a single market
first and eventually for an inter-market case. This enables us to determine (i) how
current CBs help to improve market quality for investors at the home and satellite
market and (ii) if trading activity migrates from a market on halt to a satellite venue,
like it is predicted in theoretical work and (iii) if this satellite venue is able to discover
an effective price during the halt of the home market. Particularly, the second might
pose a systemic risk to the European trading landscape if volatility shocks in one
market would allow price cascades to continue in satellite trading venues, while the
latter imposes risk to the integrity and reliability of fragmented markets in the case of
single stock CB.
The remainder of this paper is structured as follows. In the next section we will
provide a review of academic papers related to our work indicating the need for more
empirical research on inter-market CBs. Section 3 describes our dataset and provides
some descriptive statistics for our single- and inter-market analysis of CBs while the
next section presents our findings within the inter-market context. Eventually,
conclusions are given in section 5.
2. Related Literature
Analyses of CBs in the single market case can be found widely in academic literature.
Most theoretical models address the question if CBs are useful to protect market
participants against extensive market volatility. [1] show how CBs may help to
overcome some of the informational distortion problems caused by volume shocks in
continuous trading and thereby help to improve the market's ability to absorb large
volume shocks, i.e. abnormal large quantities of orders. The authors propose a
temporarily switch to an alternative transactional mechanism in order to trade
immediacy provided by continuous trading for information allocation in auctions.[2]
prove in a model-based approach that CBs lessen the order implementation risk which
happens in times of substantial volatility. [3] exhibits that CBs can be understood as a
substitute for margin requirements as the down- and upside risk are limited to the
corridor of the CB. In contrast to the beneficial effects described before many
theoretical works on CB suggest that market activity is only delayed [4] and CB cause
a magnet effect towards the threshold [5]. While the previous mentioned theoretical
models on CBs deal with markets in general the following empirical studies focus on
equity markets. The central question which most empirical studies tackle is how
volatility is affected by CBs. Most of the studies conclude that CBs are not helping in
decreasing volatility [6]. [7] does not find any support of the hypothesis that CBs help
the market to calm down. [8] and likewise [9] observed a spillover effect of the
volatility to the near future after a trading halt was put in place. [10] found no
significant impact of CB’s on the volatility in the equity market of Bangladesh. [11]
proved that consecutive halts on the Taiwan Stock Exchange dampen the volatility
better than single and closing halts. Further a decreasing volatility could be observed
on Korea Stock Exchange via a portfolio-based approach [12]. While academic

esearch on CBs in the case of a single market is quite extensive, research on the
coordination of CBs in fragmented markets is still scarce. Focusing on inter-venue
effects, [13] analyzes the utility of CBs implemented in a two-market perspective.
The first market is modeled as a dominant market with a relatively high liquidity and
the second market attracts only a small volume during trading hours. The authors
show that - in the case of a trading halt in the dominant market - liquidity as well as
price variability will shift towards the satellite market. This leads to a negation of any
beneficial effects intended by the implementation of CBs on the primary market.
Besides these results, the authors acknowledge that if a CB is triggered on a
coordinated basis across venues, price variability decreases at the cost of declining
liquidity. [14] conclude in an argumentative approach that uncoordinated CBs will
more harm the market than help due to higher volatility and a rising demand in
liquidity on the non-halting markets. He suggests that a better coordination across
venues is strongly needed to ensure the effectiveness of such mechanisms.
The first empirical study concerning the two market case is provided by [15]. In their
research paper, they address market activities during NYSE halts at the NASDAQ,
which is regarded as an OTC market. They find evidence that even if trading volume
declines, volatility spikes significantly during these times. Due to the nature of
volatility, which gives traders no arbitraging opportunity to capitalize on this
situation, they conclude that a halt should not be mandatory for both trading locations.
The most recent and comprehensive research on this matter was presented by [16]
which is focused on delayed openings in the U.S. Their dataset consists of 2,461 halts
in 1,055 stocks at the NYSE between 2002 and 2005. During these halts, trading at
other venues was allowed. They find an increase of off-NYSE trading within the
observation period and a significant contribution by off-NYSE trades to the price
discovery process. Further they suggest that off-NYSE trades dampen the abnormal
post-halt volatility and spreads. This leads them to the conclusion that continued
trading may be beneficial to the market even at higher spreads. Research on CBs in
two or more market case is scarce and empirical work only addresses the U.S. market
system. This is especially unfavorable as empirical findings can be hardly compared
to other markets due to the specialties of the U.S. trading and post-trading system.
Noteworthy in the context is the Regulation NMS and its trade-trough-rule (also
known as order protection rule). This rule prohibits market places to execute trades
and forces them to pass through or cancel the order if the price is below/over the
National Best Bid and Best Offer (NBBO). Further the implementation of the MiFID
in Europe initiated a fragmentation process which makes the research in this topic
necessary. Due to the lack of the European market place analysis, we are among the
first who provide empirical findings for the European trading landscape within the
single market and inter-market case.

3. Data Setup
The absence of a mandatory inter-market coordination between MTF and regulated
markets (RM) in the European market system potentially induces multiple scenarios
for market quality shifts before, during and after the CB. To systematically illustrate
these changes in market quality and trading behavior during a home market CB, we
provide an empirical analysis of CBs effects. First, we study impact of CBs on the
securities’ home market and the securities’ satellite market. Therefore we are
analyzing market quality before and after the CB. This gives us the opportunity to
evaluate CBs in a single market scenario, i.e. to determine its “circuit breaking”
ability. Further, trading at the satellite market during the home markets halt will be
assessed in order to evaluate changes in market conditions through the home market’s
halt. Second, we will take a closer look at coordination effects between both markets
in order to analyze trading migration and price discovery during the CB. We are
among the first who provide insight to inter-market coordination effects of CBs,
especially in illustrating each market’s role in the price determination process. Our
sample is thereby structured as follows. We focus at instruments traded on multiple
venues. In the case of a home market’s CB, trading at the satellite venue is assessed in
the time interval of the CB triggered trading halt.
TABLE 1: SUMMARY STATISTICS OF THE SAMPLE
Number of Circuit Breakers 464
Number of Instruments 27
Average (median) number of CB per Instrument 17.19 (13.00)
Maximum CBs per Instrument 73
Minimum CBs per Instrument 4
Average (median) duration in minutes 02:16 (02:16)
Table 1 illustrates the aggregated descriptive statistics for our sample. Based on the
German Blue Chips DAX 30 index, we analyze CBs in the year 2009 at Deutsche
Boerse’s electronic order book Xetra (home market). As satellite trading venue, we
choose the most relevant MTF for German Blue Chips in respect of trading volume -
the London based Chi-X MTF. The 2009 scenario provided relative calm market
conditions, after the 2008 financial crisis which was accompanied by tremendous
market turmoil. Due to a volatility flag, we are able to identify all CBs during
continuous trading over the given time period. The sample consists of the
millisecond-precise start and end point of each interruption on Xetra. The total sample
size consists of 522 single stock CBs in the DAX 30 instruments. Modifications
within the DAX 30 index composition in 2009 and Volkswagen’s volatile price
movement caused by the merger attempt by Porsche made it necessary to exclude 3
instruments (58 CBs). This leads to a total sample of 464 CBs in 2009 distributed
over 27 stocks. We retrieved reference data from the Thomson-Reuters Data Scope
Tick History covering all Chi-X DAX 30 trades in the respective time horizon. Since

the CB data includes the start and end point of each interruption, time and date
mapping will reveal trading activity on Chi-X during the home markets halt due to a
CB. Based on our data, in 16% of all CBs we observe no trades on Chi-X, i.e. in 84%
of all cases at least one trade occurred. Chi-X will generally not halt their electronic
order book in case of a reference market’s CB. Nevertheless, Chi-X provides no
information whether trading was halted during the home markets interruption [17]. In
order to get consistent analysis and due to the competitive structure of the European
trading market, we have to assume that Chi-X will continue trading whenever the
trading halt is not triggered by a regulatory intervention.
4. Results
Single Market Case
A trade price reaching either the last price’s dynamic threshold or the last auction
price’s static threshold will trigger a CB and the trading phase will either halt or
switch to auction mode. Thereby, this 2 minute interruption is meant to calm down
unreasonable price movement caused by information overload. At first, we want to
determine if CBs are actually capable in calming down above-average trading activity
by analyzing market quality parameters before and after a CB on the home market. To
do so, we look at price volatility around the CB as proposed by [11]. We measure
price volatility in two ways. High variability in asset prices indicates large uncertainty
about the value of the underlying asset, thus alienating an investor’s valuation and
potentially resulting in incorrect investment decisions when price variability is high
[18]. We obtain standard deviation of trade prices to account for the average prices’
waviness around its mean. On the other hand we use a high to low ratio in order to
take a closer look at the prices’ maximum deviation, i.e. the amount of mispricing.
Both values outline different characteristics of volatility, so both are included in the
analysis. Further, we calculate the relative spread, i.e. the relative differences between
the best bid and best ask quote. This allows us to measure the risk premium market
participants require for being exposed to market risk while submitting orders to the
order book. [18] relates relative spread with overall market liquidity and
effectiveness, therefore we also calculate the relative spread before and after the CB.
All measures are calculated before the CB was triggered and after continuous trading
restarted. We choose a short term 2 minutes interval, a medium term 5 minutes
interval and a long term 10 minutes interval. These intervals where chosen, because
the average CB in our sample interrupts trading for 02:16 minutes, we find it
unreasonable to check for market quality changes in a future distance, since these
changes could not be directly related to the CB. Pre- and post-CB market quality
parameters are compared via Wilcoxon sign rank test for equality of the pre- and post-
CB sample [19]. Table 2 depicts results for the 2 minute interval, due to space
constrains the 5 and 10 minute interval could not be included, results however remain
the same.

TABLE 2: MARKET QUALITY PARAMETER BEFORE AND AFTER THE CB ON
XETRA
2min standard deviation
2min high low ratio
2min relative spread (in Bps)
Before After Z-Value
0.064
(0.035)
0.010
(0.007)
19.33
(11.77)
0.058
(0.035)
0.010
(0.006)
22.06
(12.65)
2.811***
1.362
-4.730***
Average (median) market quality parameters before and after the CB on Xetra for all DAX 30
instruments. Z-Values for Wilcoxon sign rank test under the null that samples are drawn from
the same population. Significance level are 1% (***), 5% (**) and 10% (*).
The results illustrate a highly significant drop in the market volatility parameters after
the CB. Besides the 2min high low ratio, we find all standard deviations as well as
high low ratios on a lower level within the 2 minutes, the 5 minutes and the 10
minutes interval, indicating calmer trading conditions after the CB in the short- as
well as in the long-run. This comes by the cost of a higher implicit risk premium
which is required for market participants after the CB, i.e. a higher spread level. We
find the relative spread significantly increased in all analyzed intervals revealing an
adoption of the overall market premium after the CB. First intuition for this market
change one would suspect a decline in trading activity to be the driver of this shift.
[20], [21], [22], and [23], among others, document a positive relation between
volatility and trading volume. Therefore we look at the respective number of trades in
each measurement period. By comparing the pre and post-CB trading activity for each
instrument this assumption is rejected, since Wilcoxon sign rank test indicates an
systematical increase in the number of trades after the CB within each observed
period(results omitted due to space constrains). Calmer market volatility is achieved
although trading continues on a higher activity level. Therefore, we conclude that CBs
succeed to calm a home market’s hastiness by lowering market volatility after the CB.
Further, this effect comes at the price of a higher spread level, which will be paid by
the market participants trading after a CB.
Within the current European regulatory environment alternative exchanges are
allowed to continue trading even if the home market is halted due to a CB. In this case
the question for trading efficiency within this period rises as inter-market coordination
is not possible. We address this question by analyzing the above presented market
quality parameters for the satellite market.

Again, we compare previous and posterior market quality levels using the same
approach we analyzed the home market. Additionally, we measure standard deviation,
high low ratio and the relative spread during the home markets interruption in order to
provide insight to trading quality in this period. Therefore, we analyze CB trading in
relation to the previous and subsequent CB period. As the following section will
illustrate, during the CB the trading activity on Chi-X is massively reduced, so if we
compare volatility on Chi-X during the CB with pre- or post-CB intervals of the same
length, the results will be seriously biased as volatility will be significantly lower due
to the reduced trading volume. Therefore, we recalculate the volatility measures on a
per-trade manner. This way we compare values on an equal number of trades and do
not suffer from a systematic bias. The before, during and after samples are tested by
non-parametric Wilcoxon test for equality of medians, testing for the null that all
samples are retrieved from the same population. Since every sample is tested twice,
alpha levels are corrected by Bonferroni’s approach (while a given alpha value may
be appropriate for each individual comparison, it is not for the set of all comparisons).
In order to avoid a lot of spurious positives, the alpha value needs to be lowered to
account for the number of comparisons being performed [24]. Table 3 summarizes the
results of the before and after comparison as well as the between comparison
approach, again 5 and 10 minute intervals where omitted due to space limitations but
show identical results.
TABLE 3: MARKET QUALITY PARAMETER BEFORE, DURING AND AFTER THE CB
ON CHI-X
Before During After Z-Value
2min standard deviation
0.063
(0.037)
0.052
(0.030)
4.631***
2min high low ratio
0.009
(0.005)
0.008
(0.004)
4.162***
2min relative spread (in Bps)
45.87
(11.48)
28.10
(12.51)
-3.824***
standard deviation
0.027
(0.016)
0.054
(0.017)
0.019
(0.011)
high low ratio
0.012
(0.002)
0.011
(0.003)
0.005
(0.002)
relative spread (in Bps)
50.20
(13.19)
133.22
(50.83)
37.76
(14.99)
Average (median) market quality parameters before, during and after the CB on Chi-X for all
DAX 30 instruments Z-Values for Wilcoxon sign rank under the null that samples are drawn
from the same population. Significance level are 1% (***), 5% (**) and 10% (*).

The results indicate pre- and post-CB market conditions on the satellite market to be
calmer, by the meaning of a significantly reduced standard deviation as well as a
lower degree of maximum price deviation. Looking at the relative spread, the home
market’s findings are supported, as the satellite market also demands a higher risk
premium for market risk after the CB. Again, we check the trading activity before and
after the CB in order to find possible explanation in a reduced trading activity. Results
indicate an increase in trading activity after the CB contradicting the before
mentioned findings of related studies – again we find market condition calmer
although trading activity increases. By looking at the time period during the home
market’s CB, shown in the lower area of Table 4, we find prices during the
interruption to be undergone by a high degree of uncertainty, compared to the period
short before and short after. Standard deviation as well as the high low ratio exceeds
prior and past trading levels significantly (Table 4). We also find the relative spread
significantly increased. Market participants demanded a higher level of compensation
for market risk as in the period the home market was active. The relative spread
during the CB is on average three times as high as it is on average within the next ten
minutes, and two times as high, as it was on average ten minutes before.
Acknowledging, that this comparison controls for different trading volumes, we
summarize a massive disturbance within the satellite market’s trading quality during
the home market halt. While afterwards, with the home market proceeding with
continuous trading, market quality is reassuring to a level, lower than the pre-CB
period. Again this effect is accompanied by a higher spread level after the restart of
the home market.
TABLE 4: WILCOXON SIGN RANK TEST FOR THE EQUALITY OF MARKET
QUALITY PARAMETERS BEFORE, DURING AND AFTER THE CB
H0: MQPpre CB = MQPduring CB vs. H0: MQPpost CB = MQPduring CB vs.
H1: MQPpre CB ≠ MQPduring CB H1: MQPpost CB ≠ MQPduring CB
standard deviation -3.353*** -6.099***
high low ratio -3.949*** -4.740***
relative spread -13.459*** -12.773***
Z-Values and significance values for Wilcoxon sign rank test for the equality of the pre-CB and
during-CB market quality parameters (MQP), as well as the post-CB and during-CB sample.
Bonferroni modified significance levels are 0.5% (***), 2.5% (**) and 5% (*) for each test.
While previous analysis mainly focused on the independent, venue specific aspect of
the CB’s impact on market quality, the following chapter will highlight inter-market
coordination of trading behavior and price adjustments. In the first place, we analyze
trading behavior on Chi-X during the CB. This issue is motivated by theoretical
models about trading migration during CBs and the results will essentially determine
the following approach to ask for price efficiency on Chi-X during the CB. Therefore,
in the second step, we investigate the satellites market’s ability to uncover an efficient
price in the absent of the home market.

Inter-Market Case
Referring to [13], volatility shocks would allow price cascades to continue on satellite
trading venues and negated any positive CB effect on the home market if traders
migrate to satellite trading venues in order to continue trading. As of 2012 the
European market system does not force systematic halts in case of a single venue CB
halt, this scenario could occur in the Europe system. In order to control for abnormal
trading behavior on Chi-X, i.e. a significant change in the number of trades, we
calculated “ordinary” trading statistics within a symmetrical 30 trading day interval of
the event. For each of the 30 trading days we gathered number of trades for each
instrument of the corresponding CB time span. This way we can distinguish whether
the trading behavior within the halt could be considered “abnormal” or “ordinary”.
The length of the trading day intervals represent a short and long term view, in order
to control for short term effects shortly after/before the CB and a long term “ordinary”
trading behavior. Due to the extreme variability within the reference samples,
parametric testing is impractical as hypothesis for normality within the data is
rejected. To strengthen our implication, we perform a regression analysis. Thereby,
the number of trades on Chi-X during the home market’s CB are regressed on the on
the 3 day (15 day) median number of trades of the reference period. If there is actual
trading migration toward the satellite market, we would expect the coefficient to be
significantly larger than one. On the other side, the retreat from trading would be
indicated by a coefficient significantly smaller one. We perform ordinary least square
regression, suppressing the intercept and clustering for each instrument. If the null
hypothesis, that the slope equals one, could not be refused, we conclude that there is
no difference between the number of trades during the CBs and the reference period.
In both cases, the null is rejected on a high significant level, therefore our assumption
is strengthened that traders retreat from trading on the satellite market, if the home
market halts (Table 5).
TABLE 5: TRADING MIGRATION REGRESSION
Coefficient
Clustered
Std. Err.
T-value R 2
+/- 3 day median 0.608 0.148 -2.66*** 0.3389
+/- 15 day median 0.691 0.166 -1.86** 0.3242
Trading migration regression to determine trading migration to Chi-X during the CB. The
number of trades on Chi-X during the CBs is regressed on the 3 day (15 day) median number of
trades. T-values result from a t-test under the null hypothesis of the coefficient being one.
Significance level are 1% (***), 5% (**) and 10% (*).
In the previous chapter we empirically show that with the home markets eruption the
satellite market suffers significantly in market quality, i.e. volatility rises. While there
is no trading migration within the CB interval, trading activity is actually decreasing
below the expected “ordinary” level. The Question rises, if the satellite market retains

his ability to determine an efficient price during the home market’s CB and therefore
still contributes to price determination, or, if the satellite market systematically fails in
this ability. Therefore we look at inter-market price coordination after the CB, in
order to analyze possible systematic shifts between the markets. By looking at the
price coordination after the CB, we try to investigate which price, i.e. the home
market’s auction price or the prices revealed on the satellite market, remains the most
relevant for further trading. In order to answer this question, we start by comparing
price levels on both markets in the time before and after the CB. Since a trade by
trade analysis is unreasonable due to the un-harmonized trade occurrence, we
calculate a venue’s per second average price and therefore rely on a per-second
comparison of both markets. We do not expect both average prices to be at the same
level, but by looking at this difference ratio before and after the CB, we hope to
acquire possible systematic shifts, either finding the difference to be significantly low
or higher the time the home market returns to continuous trading.
14 Bps
12 Bps
10 Bps
8 Bps
6 Bps
4 Bps
2 Bps
0 Bps
Circuit Breaker
Average Median
Figure 1: Average and median differences within an eight second interval before and after the
CB
Figure 1 illustrates the average and median differences within an eight second interval
before and after the CB. It is obvious that after the CB the difference between both
venues is broader and short after approaches to the normal, i.e. pre-CB level. Table 6
summarizes Wilcoxon sign rank test for the differences between each post-CB
differences and the median pre-CB level. The results show that four seconds after a
CB are needed for the two venues to get to the pre-CB difference level. Therefore we

assume that coordination on either one or both markets is needed after the CB in order
to achieve a normal level in price difference on both markets. This interesting fact
raises a new question: Which price is the most relevant for future trading?
TABLE 6: TRADE PRICE DIFFERENCES BETWEEN XETRA AND CHI-X BEFORE AND
AFTER THE CB
Relative differences in Bps
Z-Value
mean (median)
Pre-CB level 8.96 (1.20) -
CB + 1 second period 11.80 (2.14) -5.340***
CB + 2 second period 10.84 (1.69) -3.798***
CB + 3 second period 9.84 (1.63) -2.154**
CB + 4 second period 9.71 (1.34) -1.335
CB + 5 second period 9.37 (1.33) -1.611
CB + 6 second period 9.12 (1.26) -1.657
Relative differences in % between the home and satellite market after the CB. Z-Value contains
statistics for a Wilcoxon sign rank test with the null hypothesis that samples are obtained from
the same population. Significance level are 1% (***), 5% (**) and 10% (*).
We assume two possible scenarios for the post-CB price coordination. On the one
hand, if only one price, either the home market’s auction price or the satellite markets
last price during the CB, provides superior information, we would expect the other
venues price to approach to this price level as both markets continue trading. On the
other hand, if both prices contain different and relevant information about the future
price level, we would expect both venues’ prices to adjust to a newer level. By taking
a future reference price for each market, we regress both reference prices on the home
market’s auction price as well as the satellite market’s last price during the CB. If
there is one price with superior information, we would expect this price to be crucial
in explaining both markets’ reference prices. Since the home markets auction price is
highly correlated with the satellite market’s last price, we expect all coefficients to
equal one, so a simple test for differences might not indicate structural differences.
Instead, we systematically compare regression accuracy and model efficiency in order
to reach a conclusion. If one price provides more explanatory power or less deviation
than the other, we expect this model to be superior. Regression accuracy is measured
in root-mean-square error (RMSE), i.e. the standard error of the estimate in the
regression analysis, model explanatory power is measured in Akaike information
criterion. The regressions take the following form:

P = β , ∗ P + ε ,
P = β , ∗ P + φ , j ∈ auction, satellite!
Where Pj is either the home markets auction price or the satellite market’s last price
during the CB, PXetra and PChi-X are the venue specific reference prices, in our case the
ten minute average price after the CB. Again we cluster for instruments.
TABLE 7: POST-CB PRICE DETERMINATION REGRESSION
Home Market’s
price determination
(P )
(1)
Satellite Market’s
price determination
(P )
Home market’s auction price (P $% &') 0.205 (-138.91) 0.203 (-146.43)
Satellite market’s first post-CB price (P ( )) ) 0.406 (432.99) 0.408 (436.23)
Xetra and Chi-X reference prices are regressed on the home markets auction price and the last
trade price on Chi-X during the CB. The table shows RMSE (AIC) for each regression.
We expect both, the home markets auction price and the satellite market’s last price
during the CB, to be good predictors for both future reference prices, since price
deviation within ten minutes is limited. By looking at the individual information
criteria and the RMSE, we interestingly find the Home market’s auction price to be of
superior value for both markets future prices than the satellite markets price (Table 7).
Thus, even the home market’s auction price explains the future Chi-X price more
efficiently than the satellite market’s last price during the CB does. We repeat our
regression by estimating the auction price’s and satellite price’s contribution together.
We are aware, that multicollinearity may inflate our standard errors, but if the home
market’s auction price systematically delivers additional explanatory power next to
the information both prices share, the auction price should indicate significance while
the last price during the CB should not. Regression is as follows:
P * = β +,* ∗ P $% &' + β ,,* ∗ P ( )) + ε * , - ∈ ./012, 3ℎ- − .! (2)
Pi denotes either the home or satellite markets 10 minutes average reference price,
Pauction and Psatellite are the home markets auction price and the satellite market’s last
price during the CB.

TABLE 8: COMBINED POST-CB PRICE DETERMINATION REGRESSION
Home Market’s
price determination
(P )
Satellite Market’s
price determination
(P )
Home market’s auction price (P $% &') 0.913*** 0.920***
Satellite market’s first post-CB price (P ( )) ) 0.087 0.080
Xetra and Chi-X reference prices are regressed on the home markets auction price and the last
trade price on Chi-X during the CB. Table shows coefficients and significance levels 1% (***),
5% (**) and 10% (*).
Table 8 supports our findings, the home markets auction price contains additional
information in predicting both market’s future prices and therefore turns out to be
significant. The satellite market’s price information are therefore redundant, i.e. not
significant. This effect is only possible, if the satellite market’s last price during the
CB experienced a systematic shift towards the home markets price level after the CB.
Therefore, we conclude that the home market’s auction price is the dominant after a
home markets CB.
In summary, the significantly reduced trading activity during the CB compared to
regular trading activity within the reference days levels the spread maker makers
demand to continue quoting during the CB. Interestingly, the reduced trading activity
does not calm down price volatility. We find standard deviation as well as maximum
price uncertainty dramatically increased. This also impacts price determination at the
satellite market. The results indicate that the price revelation process on the satellite
market is seriously restricted during the CB. Although trading during the CB is not
totally misguided, only with the restart of the home market and the return of the
traders, the satellite market regains the ability to effectively reveal prices and
approaches to the home market’s price level.
5. Conclusion
While the U.S. trading landscape has been fragmented since the late 1990s, the
trading industry in Europe only recently experienced a sea change from regulatory
and technological developments. Regulatory changes in the form of the Markets in
Financial Instruments Directive (MiFID) laid the foundation for more competition
between execution venues in late 2007, while trading was traditionally consolidated in
national exchanges before. Since then, a multitude of Multilateral Trading Facilities
MTFs emerged offering pan-European trading. It is common to exchanges and
satellite trading platforms to apply some form of CB to ensure price continuity.
Despite the numerous implementations of CBs in European primary markets, a
coordination of CBs is completely missing. As shown in some theoretical models, this

situation could lead to price-falling-cascades if traders migrate from a market on halt
to satellite venues which keep their systems open for investors’ order flow. Against
this background, we analyzed trading of German blue chip stocks in two markets, i.e.
Xetra and the most relevant European MTF Chi-X. We particularly investigated those
periods when the Xetra trading system was on halt and Chi-X sustained its trading
services.
We find price volatility to decline significantly in both markets after a CB, while
trading activity remains at a high level. In contrast to this, relative spreads do not
revert to their pre-CB levels, but even increase after a trading halt. Therefore we
conclude, that CBs do “break the circuit”, but there is no such thing as a free lunch.
The price is paid by a higher relative spread. Moreover, we find trading in the satellite
market to be sorely afflicted during CBs, while price volatility and relative spreads
are tremendously increased on the satellite market. Further, we find no trading
migration between venues in the number of trades. Instead, the number of trades
significantly drops below an anticipated normal level during a stock’s trade
interruption. Evidently, traders retreat from the satellite venue. In respect to previous
academic work on the issue of CBs we reject the finding of the inter-venue migration
model from [13] in case of non-coordinated market safeguards. A potential
explanation for our empirical observation is that investors are reluctant to trade when
the dominant market is absent as a liquidity pool due to a higher anticipated risk.
Additionally, momentum traders and statistical arbitrage strategies depend on intermarket
price differences. These types of market participants also have no incentive to
continue trading, resulting in an elimination of another significant part of the original
order flow. Our findings are in line with [25] which investigate liquidity providing
strategies during extreme market movements and provide another explanation for
investors’ market withdrawal. Implicit trading costs become very high and make
trading during these periods unattractive to investors. This might encourage a
negative impact on the satellite market’s price discovery process. We observe
abnormally wide price differences between home and satellite market after the CB,
making price coordination with the home market inevitable. By comparing satellite
and home market reference prices, we find the satellite market price level to adjust
towards the home market as trading activity raises after the home markets CB,
indicating the home market’s auction price to be the dominant price for future trading.
Our findings are particularly interesting for regulators, who might support instrumentbased
market-wide trading halts to be necessary to ensure market quality in a
fragmented landscape, as well as traders, adapting trading behavior in times of market
stress.
References
1. Greenwald, B., Stein, J.C.: Transactional Risk, Market Crashes, and the Role of Circuit
Breakers. The Journal of Business 64 (4), pp. 443--462 (1991)
2. Kodres, L. E., O'Brien, D. P.: The Existence of Pareto-Superior Price Limits. The American
Economic Review 84 (4), pp. 919--932 (1994)

3. Brennan, M.: A theory of price limits in futures markets. Journal of Financial Economics 16
(2), pp. 213--233 (1986)
4. Coursey, D. L., Dyl, E. A.: Price limits, trading suspensions, and the adjustment of prices to
new information. Review of Futures Markets 9 (2), pp. 342--360 (1990)
5. Ackert, L., Church, B., Jayaraman, N.: An experimental study of circuit breakers: The effect
of mandated market closures and temporary halts on market behavior. Journal of Financial
Markets 4 (2), pp. 185--208 (2001)
6. Kim, Y. H., Yang, J. J.: What Makes Circuit Breakers Attractive to Financial Markets? A
Survey. Financial Markets, Institutions & Instruments 13 (3), pp. 109--146 (2004)
7. Chen, Y.-M.:X: Price limits and stock market volatility in Taiwan. Pacific-Basin Finance
Journal 1 (2), pp. 139--153 (1993)
8. Kim, K. A., Rhee, S. G.: Price Limit Performance: Evidence from the Tokyo Stock
Exchange. The Journal of Finance 52 (2), pp. 885--901 (1997)
9. Bildik, R., Gülay, G.: Are price limits effective? Evidence from the Istanbul Stock
Exchange. Journal of Financial Research 29 (3), pp. 383--403 (2006)
10. Hassan, M. K., Islam, A. M., Basher, S. A.: Market Efficiency, Time-Varying Volatility and
Equity Returns in Bangladesh Stock Market. Working Paper (2000)
11. Kim, Y. H., Yang, J. J.: The effect of price limits on intraday volatility and information
asymmetry. Pacific-Basin Finance Journal 16 (5), pp. 522--538 (2008)
12. Lee, S.-B., Kim, K.-J.: The effect of price limits on stock price volatility: Empirical
evidence in Korea. Journal of Business Finance & Accounting 22 (2), pp. 257--267 (2006)
13. Subrahmanyam, A.: Circuit breakers and market volatility: A theoretical perspective.
Journal of Finance 49 (1), pp. 237--254 (1994)
14. Morris, C. S.: Coordinating Circuit Breakers in Stock and Futures markets. FRB Kansas
City - Economic Review 75 (2), pp. 35--48 (1990)
15. Fabozzi, F. J., Ma, C. K.: The over-the-counter market and New York Stock Exchange
trading. The Financial Review 23 (4), pp. 427--437 (1988)
16. Chakrabarty, B., Corwin, S. A., Panayides, M. A.: When a halt is not a halt: An analysis of
off-NYSE trading during NYSE market closures. Journal of Financial Intermedation 20 (3),
pp. 361--386 (2011)
17. Chi-X: Europe Guidance Note 1; “ERRONEOUS TRADES”. Retrieved October 20, 2011,
from http://www.chi-xeurope.com/chi-x-pdf-downloads/guidance-notes-1.-8-final-
(clean).pdf (2011)
18. Harris, L.: A transaction data study of weekly and intradaily patterns in stock returns.
Journal of Financial Economics 16 (1), pp. 99--117 (1986)
19. Wilcoxon, F.: Individual Comparisons by Ranking Methods. Biometrics Bulletin 6 (1), pp.
80--83 (1945)
20. French, K. R., Roll, R.: Stock Return Variances: The Arrival of Information and the
Reaction of Traders. Journal of Financial Economics 17, pp. 5--26 (1986)
21. Karpoff, J. M.: The Relation between Price Changes and Trading Volume: A Survey.
Journal of Financial and Quantitative Analysis 22 (1), pp. 109--126 (1987)
22. Schwert, W.: Why Does Stock Market Volatility Change Over Time? The Journal of
Finance 44 (5), pp. 115--153 (1989)
23. Stoll, H. R., Whaley, R. E.: The Dynamics of Stock Index and Stock Index Futures Returns.
Journal of Financial and Quantitative Analysis 25 (4), pp. 441--468 (1990)
24. Bonferroni, C. E.: Teoria statistica delle classi e calcolo delle probabilit. Pubblicazioni del R
Istituto Superiore di Scienze Economiche e Commerciali di Firenze 8, pp. 3--62 (1936)
25. Goldstein, M., Kavajecz, K.: Trading strategies during circuit breakers and extreme market
movement. Journal of Financial Markets 7 (3), pp. 301--333 (2004)

High Frequency Trading
Costs and Benefits in Securities Trading and its Necessity
of Regulations
DOI 10.1007/s12599-012-0205-9
The Authors
Prof. Dr. Christoph Lattemann
School of Humanities and Social
Sciences
Jacobs University Bremen
28759 Bremen
Germany
c.lattemann@jacobs-university.de
Prof. Dr. Peter Loos (�)
IWi at DFKI
Saarland University
66123 Saarbrücken
Germany
loos@iwi.uni-sb.de
Dr. Johannes Gomolka
Tempelhove Research
10117 Berlin
Germany
Prof. Dr. Hans-Peter Burghof
Arne Breuer
Prof. Dr. Peter Gomber
Michael Krogmann
Dr. Joachim Nagel
Rainer Riess
Prof. Dr. Ryan Riordan
Dr. Rafael Zajonz
Published online: 2012-03-06
This article is also available in German
in print and via http://www.
wirtschaftsinformatik.de: Lattemann
C, Loos P, Gomolka J, Burghof
H-P, Breuer A, Gomber P, Krogmann
M, Nagel J, Riess R, Riordan
R, Zajonz R (2012) High Frequency
Trading. Kosten und Nutzen im
Wertpapierhandel und Notwendigkeit
der Marktregulierung. WIRT-
SCHAFTSINFORMATIK. doi: 10.1007/
s11576-012-0311-9.
© Gabler Verlag 2012
1Introduction
Recent publications reveal that high frequency
trading (HFT) is responsible for
10 to 70 per cent of the order volume in
stock and derivatives trading (Gomber et
al. 2011; Hendershott and Riordan 2011;
Zhang 2010). This observation leads to
a controversial debate over positive and
negative implications of HFT for the liquidity
and efficiency of electronic securities
markets and over the costs and
benefits of and needs for market regulation.
Currently the European Union
(EU) is considering the introduction of
a financial transaction tax to curtail the
harmful effects of HFT strategies. The
consideration behind this market policy
is based on the assumption that the
very short-term oriented HFT trading
strategies lead to market frictions. This
current discourse shows that the arguing
parties do not homogeneously define
HFT. Reasons for this are the proponents’
different but intertwined perspectives,
which lead to new unanswered
questions in numerous subjects of expertise.
From a macroeconomic point of
viewthequestionarisesifHFTconstrains
or supports the allocation function of financial
markets. Capital market research
and information management research
raise questions about the future form
of intermediation in securities trading
and the coming architecture of markets,
about the HFT’s impact on liquidity and
about price volatility. Financial authorities
and regulators discuss whether HFT
has a stabilizing or destabilizing function
on financial systems and how a future
regulation should be shaped.
This collection of articles shall help to
develop a common definition of HFT
and contribute to the ongoing discussions.
To that end we have collected articles
from representatives of information
systems, business management, the
Deutsche Bundesbank and the Deutsche
Boerse AG. The following scientists and
practitioners participated in the discussion
(in alphabetical order):
BISE – DISCUSSION
� Prof. Dr. Hans-Peter Burghof and Arne
Breuer, Chair of Business Economics,
especially Banking and Financial Services,
University of Hohenheim, Germany.
� Prof. Dr. Peter Gomber, Chair of Business
Economics, especially e-Finance,
Johann Wolfgang Goethe-University
of Frankfurt, Germany.
� Dr. Joachim Nagel, Member of the
Board of Directors, and Dr. Rafael Zajonz,
Central Market Analysis, Portfolio,
Deutsche Bundesbank, Frankfurt,
Germany.
� Rainer Riess, Managing Director of the
Frankfurter Wertpapierbörse (FWB),
and Michael Krogmann, Executive
Vice President of Xetra Market Development
of Deutsche Börse AG, Frankfurt,
Germany.
� Prof. Dr. Ryan Riordan, Institute for
Information Systems and Management,
Karlsruhe Institute of Technology
(KIT), Karlsruhe, Germany.
HFT is a part of algorithmic trading. Gomolka
(2011) defines algorithmic trading
as the processing and/or execution
of trading strategies by the means
of intelligent electronic solution routines
(known as algorithms). Thus algorithmic
trading encompasses computersupported
trading as well as computergenerated
sell-side and buy-side market
transactions. Algorithmic trading strategies
can be both short-term and longterm
oriented.
In general, HFT is defined as real-time
computer-generated decision making in
financial trading, without human interference
and based on automatized order
generation and order management. HFT
encompasses short-term trading strategies,
which – in extreme cases – operate
in the range of microseconds using minimal
price differences. HFT thus results in
minimal profit margins per transactions
and exhibits very short holding periods of
securities positions.
However, HFT definitions vary and
various properties of HFT are not
consistently discussed in the literature.
Business & Information Systems Engineering 2|2012 93

BISE – DISCUSSION
Aldridge’s (2009) definition of HFT holding
periods range from milliseconds to
one day. Durbin (2010) on the other
hand describes HFT as trading strategies,
which covers seconds or milliseconds
only. According to Brogaard (2010),
HFT is extremely short-term buying or
selling with the intention to profit from
minimal price fluctuations.
Further characteristics are often mentioned
but are not always included in
HFT definitions, such as the exclusive usage
by professional/institutional investors
in proprietary trading, real-time data
processing and direct market access (Dacorogna
et al. 2001). Another controversial
issue is the avoidance of overnight
risk (Aldridge 2009). Other definitions
underline the role of HFT as financial
intermediary (Hendershott and Riordan
2011) or try to find differences among
the implemented trading strategies (Ye
2011).
On the basis of the broad HFT definition
given before the authors in this article
will shed light on the following questions:
(1) How does HFT influence the
liquidity and efficiency of electronic securities
markets? (2) What are the costs and
benefits of, and what are the needs for a
HFT regulation?
Peter Gomber analyzes HFT from a
market microstructure perspective, and
finds HFT to be a central element of the
value creation chain in securities trading.
As part of the value creation chain,
HFT contributes to increased efficiency
and reduced explicit and implicit transaction
costs. In his eyes, regulation of HFT
could lead to dramatic changes in market
behavior, while an inappropriate regulation
might even be counterproductive for
market quality. Gomber sees the problems
for profound research on HFT in the
lack of data available for empirical studies.
Again this leads to adverse effects in
discussions of the topic in the public, in
the media, and with regulators.
Ryan Riordan also looks at HFT from
the perspective of market microstructure
and interprets HFT as one form of technological
financial intermediation which
contributes to the efficiency of operations
in exchange trading. In his eyes,
HFT plays an important role in the process
of price formation and influences the
size of transaction costs in securities trading.
According to him, one cannot yet
say whether HFT will have a positive or
a negative impact on the capital markets.
However, he sees major advantages in a
highly technologized market. It is no alternative
for him to turn back the wheels
and return to a backward oriented, artificially
slowed, regulated trading, which is
based on human intermediation.
Rainer Riess and Michael Krogmann
describe HFT as the highest evolutionary
level of securities trading. In their
opinion HFT leads to faster processing
of information, to an increase in liquidity,
and thus added values for the
overall economy. The authors describe
how HFT is currently technically realized
and integrated into trading operations
at the exchange, and deduct their
arguments accordingly. From the point of
view of Deutsche Börse, HFT is mainly
used by institutional investors in proprietary
trading and focuses on highly
liquid stocks. The authors correlate the
rise of HFT with a continuous improvement
of the electronic trading system XE-
TRA, which – from the point of view
of Deutsche Börse – benefits all market
participants in the same way. In the eyes
of Riess and Krogmann, a future regulation
of HFT should primarily focus on
equal chances of competition in the EUarea,
in order to create “a level playing
field“. From the point of view of Deutsche
Börse, it is necessary not only to implement
security mechanisms on the side of
exchanges but also with HFT-firms.
Arne Breuer and Hans-Peter Burghof
also recognize that, due to HFT, information
can be processed more perfectly
and faster than ever before. They look at
the topic from the perspective of financial
economics. This point of view leads
them to believe that more and faster information
does not necessarily lead to
a correct determination of the intrinsic
value of financial instruments. Rather
HFT processes short-term information,
which primarily is made of short-term
volume and short-term time series data,
and thus does not contribute to the evaluation
of the intrinsic values. The authors
vote for a stricter regulation of HFT.
However, before this can be done, more
analyses should be conducted. For this,
more data are necessary.
Finally, Joachim Nagel und Rafael Zajonz
argue from the perspective of regulators.
A blanket judgment on HFT is
from the regulators’ point of view neither
adequate nor would it lead to improvements
of the regulatory framework
regarding transparency, stability, and efficiency.
The impact of HFT on the efficiency
of securities trading is – due to the
absence of a scientific discussion – still
unclear for the regulators. The possibility
to destabilize the market due to HFT in
volatile market situations is regarded as
critical but should be looked into in detail.
From the point of view of the authors
“market friendly” strategies exist, a fact
which can be judged positively. But there
are also ”unfriendly strategies“, which –
from their perspective – can be categorized
as potentially harmful. In the center
of their article, the authors formulate
the wish that this complex topic may
be discussed more intensely by the scientific
community in the future, in order
to better understand which fundamental,
regulatory measures should be applied to
HFT.
If you would like to comment
on this topic or another article of
the journal Business & Information
Systems Engineering, please send
your contribution (max. 2 pages) to
the editor-in-chief, Prof. Hans Ulrich
Buhl, University of Augsburg,
Hans-Ulrich.Buhl@wiwi.uni-augsburg.de.
Prof. Dr. Christoph Lattemann
School of Humanities and Social
Sciences
Jacobs University Bremen
Prof. Dr. Peter Loos
IWi at DFKI
Saarland University
Dr. Johannes Gomolka
Tempelhove Research
2 High Frequency Trading
Regulation Required at a
Reasonable Level
It is uncommon for a specific subject
in the field of securities trading and ITinnovation
to draw as much public attention
as high frequency trading (HFT) has
been doing in recent months. Merely a
special field for a small group of experts
prior to 2010, it is now a frequent part
of the general news coverage. Against the
background of the recent debt crisis, the
current volatility and market turmoil as
well as the “US Flash Crash” on May 6,
2010 lead to this extreme attention. Several
parties attempt to exert pressure on
politics and regulation by making HFT
responsible for that crisis and the high
market volatility. In reaction to the aforementioned
incidents and to the subsequent
public discussions, the regulatory
94 Business & Information Systems Engineering 2|2012

authorities of international financial centers
have debated the adoption of various
regulatory measures and now propose
regulatory procedures, which currently
substantiate especially in Europe
and will presumably be approved in 2012
in the context of the revision of the Markets
in Financial Instruments Directive
(MiFID).
Basically, the trading strategies based
on HFT can be subdivided into active
and passive ones. Whereas passive
strategies provide other market participants
with trading opportunities in terms
of quotes and limit orders (e.g. electronic
market making), active strategies
primarily attempt to exploit imbalances
of asset prices in fragmented markets
(e.g. primary market and multilateral
trading facilities), discrepancies in
valuation between different asset classes
(e.g. between derivatives and their underlyings)
or deviances of current asset
valuations compared to historical
correlations (e.g. in the so-called pairs
trading) immediately after the emergence
of these trading/arbitrage opportunities.
The emerging academic literature,
which analyzes the effects of HFT based
strategies on market quality, shows
mostly positive impact (for a systematic
outline of academic research concerning
HFT see Gomber et al. 2011).
Regarding price discovery, liquidity and
volatility, most studies discover positive
effects of HFT. Only a few publications
indicate that HFT can increase the adverse
selection problem under specific
circumstances, and in the case of the “US
Flash Crash” another survey (Kirilenko
et al. 2011) reveals that HFT can increase
volatility.
The growing market efficiency and a
reduction of explicit and implicit transaction
costs triggered by HFT is an obvious
issue particularly for those market
participants who used to capitalize on
intermediary services and broad bid/ask
spreads in a formerly less efficient and
less liquid trading environment. In contrast
to off-exchange trading via internalization
and so-called dark pools, i.e.
non-transparent execution venues, HFT
market-making strategies on lit markets
face relevant adverse selection costs as
they provide liquidity on the market
without knowing their counterparties.
Within their internalization systems and
dark pools in the OTC field, banks and
brokers are aware of their counterparties’
identities and can benefit from this information.
Contrary to this, HFTs in lit markets
are not aware of the toxicity of their
counterparts and are – analogous to market
makers – exposed to the problem of
adverse selection.
Inappropriate regulation of HFT based
strategies or an impact on HFT business
models due to excessive burdens might
turn out to be counterproductive and
lead to unforeseeable consequences for
the quality of markets. However, abusive
strategies have to be combated effectively
by the regulators. Particularly the analysis
of the “US Flash Crash” with its discussed
solution approaches can hardly
be transferred to the European situation,
since the issues related to the “US Flash
Crash” primarily result from the US market
structure. In Europe, where a more
flexible best execution regime is implemented
and a share-by-share volatility
safeguard regime has been in place for
two decades, no market quality problems
related to HFT have been documented so
far. Therefore, a European approach to
the subject matter is required, and Europe
should be cautious about addressing
and fixing a problem that exists in a
different market structure and thus creating
risks for market efficiency and market
quality.
Any regulatory interventions in Europe
should try to preserve the benefits of HFT
while mitigating the risks as far as possible
by assuring that (i) HFT firms are
able to provide documentation on their
algorithms upon authorities’ request and
to conduct back-testing, (ii) markets are
capable of handling peaks in trading activity
and apply safeguards to react to
technical issues of their members’ algorithms,
(iii) a diversity of trading strategies
prevails to prevent systemic risks,
(iv) co-location and proximity services
are implemented on a level playing field,
(v) regulators have a complete overview
of the possible systemic risks which could
be triggered by HFT, and have employees
who have the knowledge and the tools
to assess the impact of the trading algorithms
on market quality and the associated
risks. Furthermore, it is crucial
that market places in a fragmented environment
develop coordinated safeguards
und circuit breakers, which mirror the
HFT reality and enable all market participants
to react adequately even in market
stress.
Regulatory proposals demanding continuous
liquidity provision by HFT in the
sense of market marking obligations or
BISE – DISCUSSION
minimum quote lifetimes miss the mark
and are not suitable to improve market
stability or market integrity. They rather
contribute to a decrease in market quality
and higher transaction costs.
At first sight, demanding obligations
for HFTs to provide quotes seems an appropriate
measure to tackle the problem
of a sudden liquidity withdrawal. However,
it is highly doubtful whether any
rule can force market makers to buy in
the face of overwhelming selling pressure.
In such a situation they might rather
take the risk of being fined for not fulfilling
their obligations. Many HFT strategies
are characterized by rapid closing
of built-up positions to minimize risk.
Hence, an obligation to provide liquidity
and thereby risk capital is in sharp contrast
to many HFT business models. Due
to the significant regulatory costs those
obligations would potentially lead to a
retreat from the market and thus to a
notable loss of liquidity.
Also a minimum order lifetime, which
at first glance appears to be useful to
avoid fast order submissions and immediate
cancellations, would lead to a
significant change in market behavior.
Market participants are then no longer
able to react quickly and adequately to
market-exogenous information (e.g. adhoc
news) and the necessity to keep an
order in the order book presents a free
option for other market participants. Besides,
the existence of minimum order
lifetimes would lead to an implementation
of trading strategies capitalizing on
the“lockin”oforders.HFTwouldanticipate
the accompanied risks and costs
and attempt to compensate these costs
with higher spreads, which again would
have negative effects on market quality.
In this debate it should not be neglected
that speed is the key tool for HFTs’ risk
management.
HFT is an important factor in markets
that are driven by sophisticated technology
on all layers of the trading value
chain. However, discussions on this topic
often lack sufficient and precise information.
A remarkable gap between the results
of academic research on HFT and
its perceived impact on markets in public,
media and regulatory discussions (European
Commission 2010) can be observed.
Here, the provision of granular and reliable
data by the industry can assist empirical
research at the interface of finance
and IS to provide important contributions
to a reasonable regulation of HFT.
Business & Information Systems Engineering 2|2012 95

BISE – DISCUSSION
This regulation should eventually minimize
the inherent risks of the technology
in question without hindering the
indisputably existing positive effects for
market quality.
Prof. Dr. Peter Gomber
University of Frankfurt
E-Finance Lab
3 High Frequency Trading (HFT) –
ANewIntermediary
Financial markets require intermediaries
to provide liquidity and immediacy for
other participants. These intermediaries,
often called market makers or specialists,
were often afforded special status
and located on the trading floor, or close
to the trading mechanism of exchanges.
The automation of financial markets has
increased their trading capacity and intermediaries
have expanded their use of
technology. This has resulted in a reduced
role for traditional human market makersandledtotheriseofanewintermediary,
referred to as high frequency traders
(HFTs).
This development has been made possible
by the technological innovations in
recent years. HFT strategies usually make
use of the high speed technologies to
build up and unwind positions within
milli- and microseconds. Prerequisites
for this development were the reduction
of system latency and the increase
of computing power and data processing
capabilities of computers. Next to the
large investments in HFT, exchanges have
also invested large amounts of money in
their IT infrastructure. For example, the
costs of a high-speed connection between
Chicago and New York are estimated
around $200,000 per mile (Forbes 2010).
The question remains whether these investments
are justified with regard to the
increase of overall market quality and
welfare that results from higher HFT
activity on the market.
Like traditional intermediaries HFTs
hold little inventory, have short holding
periods, and trade often. Unlike traditional
intermediaries, however, HFTs
are not granted preferential access to the
market not available to others and they
employ advanced and innovative technology
to intermediate trading. Withoutsuchprivileges,thereisnoclearbasis
for imposing the traditional obligations
of market makers on HFT. The substantial,
largely negative media coverage
of HFT and the so called “flash crash”
on May 6, 2010 raise significant interest
and concerns about the role HFT play in
the stability and price efficiency of financial
markets. The predominantly negative
coverage seems mostly unfounded.
Overall, HFTs’ impact is similar to
that of other intermediaries and speculators.
Speculators can improve price efficiency
by obtaining more information
on prices and by trading against pricing
errors. Manipulative strategies and
predatory trading could decrease price
efficiency. Reducing pricing errors improves
the efficiency of prices. HFTs’ informational
advantage, which is driven
by the technology they employ, is shortterm.
It is unclear whether or not this
short-term information and intraday reductions
of pricing errors facilitate better
financial decisions and resource allocations
by firms and investors. If the shortterm
information – that HFTs price in –
would not otherwise become public microseconds
later, HFT clearly plays an important
role (Hendershott and Riordan
2011). It would be an important positive
role of smaller pricing errors if these corresponded
to lower implicit transaction
costs by long-term investors.
One important point left unaddressed
thus far is whether or not HFTs engage in
manipulative or predatory trading. Their
use of technology may allow HFTs to manipulate
prices at speeds that are undetectable
by slower traders. A manipulative
strategy might be the ignition of a
price movement in one direction only in
order to trade on the opposite side of the
market as proposed by the SEC (2010)
and therefore cause significant pricing
errors. As is frequently done, one can
argue whether the underlying problem
of possible manipulation lies with the
manipulator or the market participant
who is manipulated. In the SEC example,
the passive manipulation could not
succeediftherewerenopricematching.
The manipulation stories could be tested
with more detailed data identifying each
market participant’s orders, trading, and
positions in all markets.
Despitethestrongevidenceofthepositive
role of HFT for the efficiency of
price determination and trading costs
(Hendershott et al. 2011; Brogaard 2010;
Zhang and Riordan 2011), regulators and
the media are certain that they must be
regulated.Itis,however,unclearandalso
debatable how we should regulate HFT.
Assuming that some, or most, of their
activities contribute positively to liquidity
and price efficiency, which parts of
their trading should we regulate? There
are controversially discussed suggestions
to restrict HFTs’ mostly passive trading or
to enforce a minimum order life on limit
orders. Restricting HFTs’ ability to trade
actively necessarily impedes their ability
to manage the risks associated with intermediation.
This may lead to less intermediation
and lower liquidity. Imposing
minimum order lives on limit orders
may also negatively impact HFTs’
ability to manage trading risks during
volatile market periods that existed before
HFT dominated the equity market.
Finally, the discussions of US and European
regulation should take into account
specific differences of both markets. Despite
the high market fragmentation, the
European market has maintained a comparably
high degree of efficiency. This is
also due to the help of HFTs. They make
use of arbitrage strategies to dissolve existing
price deviations within seconds
which results in an interconnectedness of
European markets.
A final point is a more general one
on technology investments. HFTs must
makealargeandlong-terminvestmentin
technology, both hardware and software.
This investment in technology seems to
have to paid-off both for HFTs and the
equity markets. If regulation were to
negatively impact the returns on investments
in HFT technologies by reducing
the profitability of intermediation, fewer
firms will be willing to invest in these
technologies. This may lead to a situation
in which one or two highly specialized
firms dominate intermediation,
which ultimately leads to less competition,
lower liquidity and reduced priceefficiency.
Competition, ease of market
entry and the use of specialized and innovative
technology seem to be the best
guarantors of market stability.
It is hard to imagine a situation in
which HFTs are able to artificially manipulate
prices for longer periods of
time given the intense competition other
HFTs. HFTs are one type of intermediary.
When thinking about the role HFT plays
in markets it is natural to try to compare
the new market structure to the previous
market structure. Some primary differences
are that there is free entry into
HFT, HFTs do not have a designated role
with special privileges, and HFTs do not
have special obligations. When considering
the optimal industrial organization of
the intermediation sector, which includes
96 Business & Information Systems Engineering 2|2012

egulation, market structure, technology
and incumbency, HFT more closely resembles
a highly competitive environment
than traditional market structures.
A central question is whether there were
benefits of the more highly regulated and
less technology intensive intermediation
sector which outweigh the costs of lower
innovation and higher entry costs typically
associated with regulation. The answer
to this question seems thus far to be
a resounding “no”.
Prof. Dr. Ryan Riordan
Karlsruhe Institute of Technology
4 High Frequency Trading – An
Exchange Operator’s Perspective
4.1 High Frequency Trading – Myth and
Reality
On 2010-09-30, the U.S. Securities and
Exchange Commission (SEC) and the
Commodity Futures Trading Commission
(CFTC) (2010) issuedajointreport
showing that the so-called “flash crash”,
a sequence of events which made prices
plunge throughout the US stock market,
was caused by an incorrectly programmed
trading algorithm of a traditional
investment company which did
not use high frequency trading (HFT).
Nevertheless, HFT has gained massive
public attention ever since. The news media,
as well scientists and regulatory authorities,
are busy discussing and analyzing
the effect of HFT on the global capital
markets. While the public perception
of HFT is largely critical – and driven by
headlines demanding a HFT ban or, at
least, strict regulation – scientific analysis
comes to rather different conclusions (see
Gomber’s discussion above). According
to Brogaard’s (2010) study of HFT, blaming
HFT for the US flash crash is not
the only popular fallacy regarding the
role of HFT in securities trading. Brogaard’s
analysis of NASDAQ data showed
thatfor65%ofthetimeHFTaccounted
for the best bid and ask quotes. Also,
Broogardfoundnoevidencesuggesting
that HFT firms systematically engage in
market abuse, e.g. by illegally taking advantage
of information about client orders,
the so-called “front running”. Since
HFT firms are proprietary traders, they
do not have any clients. Generally, scientific
analysis did not find a correlation
between HFT and market abuse.
The Netherlands Authority for the Financial
Markets (AFM 2010) considersHFT
as a legitimate trading method which is
not market abusive under normal circumstances.
According to Gomber, academic
papers mostly could not find evidence
for negative effects of HFT on
market quality. Moreover, the Germanybased
Karlsruhe Institute of Technology
(KIT) concluded their study based on
analysis of NASDAQ data with the finding
that HFT even worked as a buffer
against plunging stock prices during the
crisis years 2008 and 2009 (Zhang and
Riordan 2011).
4.2 Insights of an Exchange Operator
We live in a technology-driven society,
continuously striving to further improve
and advance the achievement potential
of our economy as well as of nearly
every aspect in our everyday life: can
anyone imagine a commercial flight today
without the aid of an autopilot, or
modern microsurgery without robotics?
These advancements are by no means
ends in themselves but serve a greater
good. Just the same goes for the ever
increasing speed in securities trading –
a development which leads to continuously
improving general market quality
and also to more efficient risk management
for every market participant. The
faster the market data transmission, the
faster investors are able to adapt to ongoing
market developments. This does not
only have a very positive effect on the
safety in securities trading but also on
transaction cost: faster trading leads to
tighter spreads and, therefore, to higher
liquidity. The implicit transaction costs
of every securities trade are determined
mainly by liquidity and account for up
to 80 percent of the overall transaction
costs, while the explicit transaction costs
– commissions, fees, taxes – are of minor
significance. With this in mind, Deutsche
Börse started long before the advent of
HFT to improve the trading infrastructure
of its electronic trading platform Xetra,
especially in view of ever decreasing
systemic latency. At the same time,
Deutsche Börse further developed the security
mechanisms and technologies respectively
adapted them to the increasing
demands of a more and more sophisticated
and faster trading system, one of
them being the very effective instrument
of the volatility interruption, introduced
in 1999. This security mechanism is used
in extremely volatile market phases and
BISE – DISCUSSION
leads to higher price stability: whenever
an indicative price is outside the price
range – which is pre-defined for every security
traded on Xetra – a volatility interruption
will be initiated around the
reference price.
While continuously advancing the
technical infrastructure, Deutsche Börse
expanded its offer of individually selectable
bandwidths for market participants
connected to Xetra from
512 Kbit/sec up to 2 Mbit/sec for their
Values API interfaces. In 2008, for Xetra
members requiring even faster market
data transmission and more order
book depth, an additional interface with
a bandwidth of 1 Gbit/sec was implemented,
called Enhanced Broadcast Solution
respectively Enhanced Transaction
Solution. Today, bandwidths of up to 10
GBit/sec are available. With the introduction
of the so-called “non-persistent”
orders in 2009, Deutsche Börse further
enabled Xetra members to optimize their
response times to price changes thanks
to even faster data processing. “Nonpersistent”
orders are not saved in exchange
systems and are thus designed not
be executed after volatility interruptions.
In late 2011 Deutsche Börse complemented
its connectivity portfolio with
the FIX (Financial Information Exchange)
gateway. Market participants usingthisprotocolnowareabletoconnect
to Xetra far more easily.
However, there was one latency factor
left that even the most sophisticated technology
could not overcome: the propagation
delay due to physical distance. For
every 100 km which a market participant’s
trading engine and the trading system
of Xetra are physically apart, transaction
latency increases by 1 msec approximately.
This could mean a true competitive
disadvantage for market participants
relying on ultra low latency. Deutsche
Börse addressed this growing market demand
by introducing its proximity services
in 2006. By placing the trading engine
of distant Xetra members not only
virtually but physically close to the exchange
back end – a process called colocation
– the travel time of the market
data could be drastically reduced. Today,
141 Xetra members take advantage
of Deutsche Börse’s co-location offer.
Thanks to a continuously perfected
trading infrastructure and the introduction
of proximity services, Deutsche
Börse has not only remained competitive
on an international level but has also prepared
Xetra optimally for the needs of
Business & Information Systems Engineering 2|2012 97

BISE – DISCUSSION
HFT firms. Over the last few years, systemic
latency on Xetra has been further
reduced notwithstanding a dramatic increase
of technical transactions – an advantage
to all market participants alike:
a fair, equal access to Xetra and the preand
post-trade transparency characteristic
of a regulated exchange make sure
that every investor enjoys all advantages
Deutsche Börse’s trading platform has to
offer.
While being a minority, HFT firms
nevertheless play an important role in
improving the order book quality on Xetra,
e.g. by bundling the very heterogenic
order flow. There are three organized
forms of HFT on Xetra: the proprietary
trading of investment firms, hedge
funds, and proprietary trading companies.
Two types of trading prevail: first of
all, the so-called electronic liquidity provision.
In this case, HFT firms act as voluntary
market makers, adding liquidity
to a multitude of securities. The second
type of HFT on Xetra is called statistical
arbitrage which leads to improved price
discovery. Both types of HFT account
for tighter spreads and, ultimately, improved
market efficiency on Xetra. So far,
Deutsche Börse could find no evidence of
HFT having lead to destabilizing markets
during periods of market turmoil, e.g. by
strengthening trends. During the highly
volatile market phase in August 2011, the
trading volume on Xetra increased temporarily
to 107 million transactions on
one single day. Despite up to 30 volatility
interruptions, the average transaction
processing took only 0.4 msec longer
than usual. System availability was guaranteed
at all times, Xetra members did
not have to face any restrictions, let alone
system failure. Deutsche Börse’s market
security mechanisms made sure that all
trading activities could be executed properly
and continuously while price stability
was guaranteed even during market
turmoil.
Thus, Deutsche Börse succeeded in advancing
the Xetra infrastructure in terms
of continuously decreasing systemic latency
and, at the same time, met the
permanently increasing needs regarding
safety and speed of its trading system
even before the term HFT came up.
4.3 Regulatory Recommendations
Within a national economy it is the explicit
function of a securities exchange to
facilitate the most efficient employment
of capital, ensuring best possible corporate
financing and re-financing. HFT, as
it is today, supports faster processing of
economically relevant data and leads to
higher liquidity in the trading of company
shares. Thanks to a stable, highperformance
trading system, Deutsche
Börse was able to integrate HFT successfully
and to use the positive effects of
HFT to improve overall market quality.
This would not have been possible without
Deutsche Börse’s principle of equal
access and a fair set of rules applying to
every market participant trading on Xetraalike.Fromaregulatoryperspective
– and keeping MiFID’s ultimate goal of
creating an EU-wide “level playing field”
in mind – comprehensive rules regarding
HFT definitely would make sense.
Therefore, Deutsche Börse supports all
measures to enhance transparency, e.g.
the complete registration of all market
participants and a full recording of all
their trading activities – traditional trading
and HFT alike. The Deutsche Börse
(2011) has come to the conclusion that
regulatory intervention in HFT must not
hurt the proven positive effect on market
quality HFT has to offer. In particular,
the variety of HFT strategies should
be preserved, as systemic risk should be
prevented. To achieve these goals, HFT
firms themselves may have to implement
security mechanisms – just as exchange
operators as Deutsche Börse already have.
Whichever regulatory rules may be implemented
in the end, the regulators will
have to make sure that these rules apply
to every European market and to every
market participant in Europe to the very
same extent.
Rainer Riess
Michael Krogmann
Deutsche Börse AG
5 Paradigm Change Through
Algorithmic Trading
5.1 Introduction
Algorithmic trading nowadays often accounts
for more than half of trade volume
and order volume at large stock
exchanges. Its net effects are generally
found positive by researchers. Only few
voices from the scientific community –
more,however,fromtraders–pointout
negative effects of algorithmic trading. A
notable difference lies between empirical
findings – that usually find positive effects
– on the one hand, and some theoretical
works and especially the sentiment
of traders, who often express their frustration
about their computerized counterparts,
on the other hand.
5.2 Availability of Data
Most scientific studies about algorithmic
trading share one fundamental problem:
data about algorithmic trading are
scarce. As one of the few stock exchanges,
Deutsche Börse had for some
time quite reliable data on algorithmic
trading. Their “Automated Trading
Program“ (ATP), which was in effect
from 2007 to early 2009, enabled them
to distinguish between algorithmic orders
and human ones (Deutsche Börse
2009). Hendershott and Riordan (2011),
Gsell (2009), Groth (2009), and Maurer
and Schäfer (2011) analyze such datasets
which contain flags for orders placed
within the ATP environment. Their research
questions differ, but they all more
or less conclude that the overall effect of
algorithmic trading is positive.
A fundamental critique of such analyses
is that algorithms usually work well
in “normal” markets and then show the
often-found positive effects. The models
that algorithms base on are abstractions
of reality and must fail to reflect it in its
entirety. If a market situation is not part
of the possibility space of the model, several
options are possible: The algorithm
halts trading and waits until the market is
“normal” again, thereby facing the risk to
generate possibly considerable losses. Another
option is to continue trading using
the usual model, thus failing to trade optimally
and possibly worsening the situation.
Since the flash crash on May 6, 2010,
there have been repeated miniature flash
crashes that did not affect the whole market
but only individual stocks. For both
phenomena, algorithms are blamed to be
the cause of the market irregularities.
However, an effective approach to regulation
should base on well-established
results. A lot of work has to be done
here. Above all, the insufficient availability
of appropriate data confines scientific
progress. The deduction of the effect
of algorithmic trading on the market
from anonymous order book data can
only be very rudimentary. In our current
work, we attempt to find a way to analyze
algorithmic trading activity whilst
only using anonymous order book data
98 Business & Information Systems Engineering 2|2012

(Breuer and Burghof 2011). A mandatory
flagging of algorithmic orders would
be desirable. Only then would it be possible
to independently analyze algorithmic
trading from many points of view
and estimate the effect on the market.
The restrictive handling of historic ATP
data by Deutsche Börse does not build
confidence but could increase the probability
that the sentiment towards AT is
influenced by irrational fears.
5.3 Information Efficiency
Recent studies (Hendershott and Riordan
2011; Gsell 2009; Groth 2009; Maurer
and Schäfer 2011) analyze rather shortterm
aspects of market microstructure in
an AT environment. Indeed, its existence
alters behavioral incentives of other market
participants fundamentally and in the
long run. It is apparent that algorithms
process new information ever faster and
– assuming normal market conditions –
probably calculate its price impact better
than humans. It is still to be seen, however,
how accurate trading algorithms
process information without slow human
traders monitoring them. Sometimes, the
superfast processing of news can be undesirable.
An example for this is the news
about the bankruptcy of United Airways.
The airline’s stock price plummeted until
it became clear that the news was already
a couple months old. Because the
possibility to extract yields from new information
has a very short and decreasing
half-life, systems tend to react hastily
and without challenging the information.
Especially in delicate market situations,
rumors can develop a destructive power.
The effect that is likely to be most
important has however escaped scientific
analysis so far. Capital markets are a
highly efficient instrument of capital allocation,
especially because a large number
of actors feed information into the
price via their trading activity. This information
comes from various sources;
it may be obtained haphazardly or with
some effort. Algorithmic trading uncovers
trade activity which is caused by that
information and uses this knowledge to
pocket a considerable part of the information
yield. The better these algorithms
work, the less money the informed person
will make out of this information. In
the long run, this could mean that the
costly generation of information turns
unprofitable, and in an extreme case even
the trade based on incidentally obtained
information does not pay anymore.
In such a hypothetical market, ever less
information is traded ever more perfectly
and faster. The market draws nearer and
nearer the weak form of market efficiency
(Fama 1970) or eventually even the semistrong
form of market efficiency. At the
same time, it moves away from the strong
form of market efficiency, because the incentive
to feed information into the market
becomes considerably less powerful.
It is this very effect that traders witness
when they trade against algorithms. They
know that information-based strategies
are detected rapidly and thwarted by appropriate
front-running strategies (Biais
et al. 2010; Cvitanic and Kirilenko 2010).
Surly, there is still a need for theoretical
as well as empirical analysis here, because
duetothesethoughts,theusefulnessof
algorithmic trading is subject to scrutiny.
5.4 Regulation and Regulatory
Instruments
Regulatory considerations have to distinguish
between the different types of algorithms.
Limit orders which are bogus orders
or part of quote-stuffing techniques
have to be considered under the light of
laws against market manipulation (e.g.,
§20a (1) No. 2 of the German Securities
Trading Act [WpHG]). Other strategies
improve the price quality by arbitraging
prices and equalizing them across different
trading venues. Because of the market
power of algorithms, there is the risk
that overly mechanic thinking and potent
algorithms may perturb the price formation
process. Naturally it would be desirable
to capture the positive effects of
algorithmic trading and to dampen the
potentially negative ones. There may be
more than one way to reach this aim.
A simple ban of algorithmic trading, as
sometimes demanded by certain political
circles, cannot serve to reach this difficult
aim. This would mean to also destroy
many preferable effects of algorithmic
trading. Of course, a distinction of
algorithmic and “normal” trading is not
easy. And certainly market participants
would program algorithms that operate
inthegrayareatohidetheirtruenature.
Currently, regulatory bodies are discussing
possible means (Dombert 2011).
The often contemplated lower boundary
for limit order lifetimes is regarded sceptically.
The comprehensible reason is that
an efficient risk management of orders
would be drastically complicated – especially,
but not exclusively, in volatile
BISE – DISCUSSION
markets. Dombert (2011)proposesanalternative
that is worth discussing. With
an order-transaction-ratio, the number
of orders divided by the number of transactions
would have to remain above some
exogenous constant.
In our view, a European regulatory
framework is desirable that defines the
playground for all market participants.
Within this framework, it should be left
to the trading venues how they wish to
treat algorithmic trading in the context
of their business model. Then it would
be up to them if they wanted to attract
algorithmic trading or to limit it in specific
market conditions. Such a “menuapproach”
leaves it in essence to the individual
trader if he or she wishes to
face the competition from algorithms
with all their positive and negative effects
or evade them by trading on trading
venues with appropriate restrictions that
apply always or under specific market
conditions.
5.5 Conclusion
As long as algorithms operate in the
dark, there is a profound uncertainty
about the effect of their activities. Therefore,
algorithmic trading is partly in contradiction
to fundamental principles of
stock exchanges: bringing buyers and
sellers together in a transparent manner.
On stock exchanges, trust is paramount.
The opacity of algorithmic trading –
as comprehensible it may be from the
point of view of their operators – undermines
this principle. Currently, there
is no level playing field. However, it
is equally important to enable technical
progress, which algorithmic trading
with its high-quality information processing
definitely is. An improved availability
of data and associated scientific
research can help to develop reasonable
regulatory frameworks for algorithmic
trading. With the increasing importance
of this way of trading in mind, there is
less and less reason to doubt that the implementation
of appropriate regulatory
frameworks should have a high priority.
Arne Breuer
Prof. Dr. Hans-Peter Burghof
University Hohenheim
6 High Frequency Trading –
ACentralBankView
The capital markets are currently at an
important juncture in their development.
Business & Information Systems Engineering 2|2012 99

BISE – DISCUSSION
Roughly half of all stock and foreign exchange
trades conducted on the major
exchanges are no longer initiated by human
traders; instead, they are the product
of computer algorithms that are able
to analyze large volumes of data and initiatehundredsofordersinfractionsof
a second. Humans are increasingly being
eliminated from the immediate decisionmaking
process relating to the sale and
purchase of assets and being replaced by
software programs.
The speed with which orders are executed
has become to be the most important
factor and is now measured in milliand
microseconds. New practices such as
“co-location” or “quote stuffing” – placing
huge quantities of buy or sell orders
which the instigator intends to cancel almost
immediately before they are executed
– have become important instruments
in the battle for the most rapid order
execution. Fundamental data on the
value of the respective securities or currencies
are of no, or only subordinate,
importance for HFT algorithms.
In HFT, positions are usually held for
between a number of milliseconds and
several hours. In today’s high-speed markets,
the scales are no longer tipped in
favor of the investor who is best able to
assess the true value of an asset, but of
the investor able to trade fastest. True
investments are becoming increasingly
rare.
Since the “flash crash” of May 6, 2010
(a roughly 15-minute phase of unusual
and irrational volatility on the New York
Stock Exchange), HFT has been called
to the attention not only of the general
public but also of regulators and central
banks.
Numerous observers regard HFT as a
new technical means of executing existing
trading strategy rather than a strategy
in its own right. Advantages in terms
of speed have, they say, always been an
essential component of many successful
trading strategies. Seen from this perspective,
HFT is not a completely new
phenomenon, but rather a technical evolution
of the securities markets. HFT
should be regarded merely as an overarching
term covering a multitude of different
fields of use. Among the many tactics,
several of the most important are based
on providing liquidity in stock market
trading (market making). Others can be
included under the category “statistical
arbitrage” and use algorithms to swiftly
identify and exploit profitable trading
opportunities based on price data. Others
belong to a category known as liquidity
detection, in which traders try to
seek out hidden large orders in order
books. Many critics term this “predatory
trading”, and it is suspected of being
unfair and potentially damaging to the
market.
Against this complex background, any
assessment of HFT and all discussion
relating to potential regulation should,
where possible, be limited to the underlying
HFT strategy. From a central
bank perspective, a sweeping judgment
on HFT is therefore neither appropriate,
nor would it serve to improve the
regulatory framework for transparency,
stability and efficiency. That means that
both the advantages and disadvantages of
HFT need to be evaluated very specifically.
Statements that HFT is in general
either good or bad for the market should
therefore be viewed with caution.
HFT players and exchange operators
are at pains to stress that overall HFT
perceptibly improves market liquidity
and the efficiency of price discovery
(McEachern Gibbs 2009). The majority
of investors benefit from reduced bid/ask
spreads – a common measure of liquidity,
they say. This statement is backed
up by several scientific studies (Gomber
et al. 2011). However, there is increasing
evidence to suggest that, especially
in very volatile market situations, HFT
could prove problematic and could additionally
destabilize the market (Brogaard
2010). This must be investigated and, if
found to be true, regulators must step in
to limit the risks for the financial system.
The flash crash demonstrated that the
liquidity generated by HFT market makers,
which usually keeps transaction costs
low, may suddenly evaporate in difficult
market phases (NANEX 2010). Unlike
regular “human” market makers, who are
obliged to remain in the market even in
times of extremely volatile prices, HFT
tradersaregenerallynotboundbysuch
constraints. In good times, HFT traders
therefore crowd out normal market makers
and often even perform their role better,
to the advantage of all market players.
In difficult markets, however, there is a
risk that trading flows could collapse with
all the attendant problems for the market
as a whole, as HFT players withdraw.
To many market participants, the narrower
bid/ask spreads and higher trading
volume generated by HFT therefore
only represent “sham liquidity”. For this
reason there have been calls from various
quarters to oblige HFT market makers to
remain in the market even in times of
high volatility, similar to the obligations
imposed on normal market makers (EC
2010). In other words, they should start
to take some responsibility for the markets
which they have, to date, merely used
to their advantage from their superior
position.
From a regulatory perspective, HFT
has proven problematic not only in these
rare but dramatic high volatility events,
but also in daily trading activities. While
bid/ask spreads have dropped significantly
in recent years thanks to HFT
market makers, the average period for
which such players hold positions has
dropped sharply. According to a study on
the flash crash, most HFT market makers
close out their positions after no more
than roughly 10 seconds (Kirilenko et
al. 2011). That means that the stabilizing
effect in the event of heightened market
volatility exerted by “normal” market
makers has given way to a “hot potato
effect”, where falling shares are merely
passed around at lightning speed.
As HFT has become more widespread,
the number of buy and sell orders has increased
dramatically in recent years. The
tactic known as quote stuffing, which is
used by several HFT algorithms, is particularly
problematic. For reasons of tradingstrategy,theHFtraderplacesalarge
number of orders per second, only to
cancel them again almost immediately
before execution. The very high cancellation
rate this causes leads to a marked
divergence between apparent market liquidity
and actual trading volume. An investor
placing an order in response to
a bid or ask is therefore often unable
to carry out the transaction at the limit
shown. Although the explicit transaction
costs appear low, the implied costs may
be much higher. Apparent market liquidity
and the size of bid/ask spreads
are therefore not by themselves reliable
indicators of market liquidity and
efficiency.
An analysis of 1,172 trading days on
the New York Stock Exchange from 2007-
01-01 to 2011-09-14 that was carried out
recently by the research firm NANEX
showed that there were just 35 billion real
transactions for 535 billion quotes. The
quotes-to-trades ratio needed to generate
US$ 10,000 in real transaction volume
moved from roughly 6–7 at the beginning
of 2007 to 60–80 in mid-2011.
Higher figures indicate a less efficient
market: more information is required to
100 Business & Information Systems Engineering 2|2012

achieve the same trading volume. Sudden
and dramatic spikes in the number
of quotes are increasingly being observed
for individual US stocks, with individual
HFT algorithms generating several
tens of thousands of quotes per second
for several seconds. Such bursts of activity
are frequently accompanied by what
are known as “mini flash crashes”, where
securities lose 20%, 40% or even more
than 50% of their value in a space of seconds
for no fundamental reason, only to
recover shortly afterwards. For instance,
according to the SEC, the United States
has witnessed more than 100 such inexplicable
crashes since mid-2010 which
are suspected of being caused by HFT
algorithms.
Sending bids or asks is similar to sending
spam email: both are virtually free for
the sender, but not for the recipient. Forwarding
and processing such large volumes
of data causes a lot of problems
and high costs for exchanges and market
participants. Systems are often overloaded,
which is seen by many observers
as one of the causes of the flash crash. To
make matters worse, certain HFT algorithms
send some of these quotes only to
cause other traders or algorithms to act in
a certain way, which they can, in turn, exploit.
As a consequence, an ever increasing
number of institutional investors are
transferring their transactions away from
normal exchanges to “dark pools”, where
it is usually more difficult to make a profit
in HFT.
The above-described criticisms intend
to show that HFT is a controversial issue,
requiring an exact analysis of the
details. In addition to “market friendly”
strategies that regulators regard as positive
for the market – for instance, statistical
arbitrage – there are also “unfriendly”
strategies that are seen as worrying. Others
are basically welcome but when actually
applied on the market entail problems
and dangers which should be eliminated.
HFT market making is just such
an example.
When considering the ultimate question
of whether there is a correlation
between HFT and market efficiency, it
should be borne in mind that market efficiency
mainly means that the price of
an asset adjusts to fundamental changes
in its value rapidly. It is not immediately
clear how HFT algorithms can contribute
to that, as decisions are based only on
the status of the order book in the last
few seconds or indicators based on technical
analysis. A block trade of 10,000
shares between two well-informed large
investors represents true price discovery
on the market. By contrast, shifting 100
shares back and forth between two HFT
algorithms in innumerous times makes
no equivalent contribution to trading efficiency,
even if this takes place at impressive
speed. A market that is mainly
dominated by HFT is also a market where
most orders have lost all connection to
fundamental factors. And this correlation
between price and fundamental value is
what should, in the main, determine the
quality of a market.
References
To: Section 1
Dr. Joachim Nagel
Dr. Rafael Zajonz
Deutsche Bundesbank
Aldridge I (2009) High-frequency trading.
A practical guide to algorithmic strategies.
Wiley, Hoboken
Brogaard JA (2010) High frequency trading
and volatility. Working paper, Northwestern
University, Chicago
Dacorogna MM, Gencay R, Müller U, Olsen
RB, Pictet OV (2001) An introduction to
high-frequency finance. Academic Press,
San Diego
Durbin M (2010) All about high frequency
trading – the easy way to get started.
McGraw Hill, New York
Gomber P, Arndt B, Lutat M, Uhle T (2011)
High-frequency trading. Working paper,
Goethe-Universität Frankfurt
Gomolka J (2011) Algorithmic Trading: Analyse
von computergesteuerten Prozessen
im Wertpapierhandel unter Verwendung
der Multifaktorenregression. Dissertation,
Potsdam, Universitätsverlag Potsdam
Hendershott T, Riordan R (2011) Algorithmic
trading and information. Working paper,
University of California (Berkeley), Karlsruhe
Institute of Technology
Ye G (2011) High frequency trading models +
website. Wiley, Hoboken
Zhang F (2010) High-frequency trading,
stock volatility, and price discovery. http://
ssrn.com/abstract=1691679. Accessed
2012-01-30
To: Section 2
European Commission (2010) European
Commission directorate general internal
market and services. Public consultation:
review of the markets in financial
instruments directive (MiFID). http://ec.
europa.eu/internal_market/consultations/
docs/2010/mifid/consultation_paper_en.
pdf. Accessed 2011-01-19
Gomber P, Arndt B, Lutat M, Uhle T (2011)
High frequency trading. Working paper,
Universität Frankfurt
Kirilenko A, Kyle A, Samadi M, Tuzun T (2011)
The flash crash: the impact of high frequency
trading on an electronic market.
Working paper
To: Section 3
BISE – DISCUSSION
Brogaard JA (2010) High frequency trading
and volatility. Working paper, Northwestern
University, Chicago
Forbes (2010) Wall Street’s speed war. 2010-
09-27
Hendershott T, Jones CM, Menkveld AJ (2011)
Does algorithmic trading improve liquidity?
J Finance 66(1):1–33
Hendershott T, Riordan R (2011) High frequency
trading and price discovery. Working
paper, University of California (Berkeley),
Karlsruhe Institute of Technology
Securities and Exchange Commission (SEC)
(2010) Concept release on equity market
structure. Release No. 34-61358, File
No. S7-02-10
Zhang S, Riordan R (2011) Technology and
market quality: the case of high frequency
trading. In: Proc ECIS 2011
To: Section 4
Deutsche Börse AG (2011) Several internal papers
and statistics. August 2011, Eschborn
Brogaard JA (2010) High frequency trading
and volatility. Working paper, Northwestern
University, Chicago
AFM/Netherlands Authority for the Financial
Markets (2010) High frequency trading
– the application of advanced trading
technology in the European marketplace
SEC and CFTC (2010) Findings regarding the
market events of May 6, 2010
Zhang S, Riordan R (2011) Technology and
market quality: the case of high frequency
trading. In: Proc ECIS 2011
To: Section 5
Biais B, Foucault T, Moinas S (2010) Equilibrium
algorithmic trading. Working paper,
Toulouse School of Economics
Breuer A, Burghof H-P (2011) An empirical
analysis of order dynamics in a highfrequency
trading environment. Unpublished
working paper, Universität Hohenheim
Cvitanic J, Kirilenko A (2010) High frequency
traders and asset prices. Unpublished
working paper, CFTC
Deutsche Börse (2009) Price list for the utilization
of the trading system of FWB Frankfurt
Stock Exchange. Deutsche Börse AG
Dombert AR (2011) Hochfrequenzhandel:
Nutzen und Risiken neuer Technologien
auf den Finanzmärkten. Kreditwesen 18:
16–18
Fama E (1970) Efficient capital markets: a review
of theory and empirical work. J Finance
25(2):383–417
Gsell M (2009) Algorithmic activity on Xetra.
J Trading 4(3):74–86
Groth S (2009) Algorithmic trading engines
and liquidity contribution: the blurring
of “traditional” definitions. In: Godart C,
Gronau N, Sharma S, Canals S (eds) Proc 9th
IFIP conference on e-business, e-services,
and e-society (I3E 2009)
Hendershott T, Riordan R (2011) High frequency
trading and price discovery. Working
paper, University of California (Berkeley),
Karlsruhe Institute of Technology
Business & Information Systems Engineering 2|2012 101

BISE – DISCUSSION
Maurer K-O, Schäfer C (2011) Analysis of binary
trading patterns in Xetra. J Trading
6(1):46–60
To: Section 6
McEachern Gibbs C (2009) HFT is market
making evolved, says NYSE exec.
http://www.advancedtrading.com/
managingthedesk/219501268. Accessed
2012-01-30
Gomber P, Arndt B, Lutat M, Uhle T (2011)
High frequency trading. Working paper,
Universität Frankfurt
Brogaard JA (2010) High frequency trading
and volatility. Working paper, Northwestern
University, Chicago
NANEX (2010) Analysis of the “flash
crash”. http://www.nanex.net/20100506/
FlashCrashAnalysis_CompleteText.html.
Accessed 2012-01-30
European Commission (2010) European
Commission directorate general internal
market and services. Public consultation:
review of the markets in financial
instruments directive (MiFID). http://ec.
europa.eu/internal_market/consultations/
docs/2010/mifid/consultation_paper_en.
pdf. Accessed 2011-01-19
Kirilenko A, Kyle A, Samadi M, Tuzun T (2011)
The flash crash: the impact of high frequency
trading on an electronic market.
Working paper
102 Business & Information Systems Engineering 2|2012

Investigating the Market Impact of
Media Sentiment and Investor Attention
Michael Siering 1
1 Goethe-University Frankfurt, Grüneburgplatz 1, 60323 Frankfurt, Germany,
siering@wiwi.uni-frankfurt.de
Abstract. Media sentiment has been shown to be related to stock returns.
However, one prerequisite for this influence has not been taken into account
yet: the question whether investors actually pay attention to news and the
related financial instruments. Within this study, we close this research gap by
examining the interplay between media sentiment and investor attention.
Thereby, we find that the positive impact of media sentiment on returns is
increased when investor attention is high. Furthermore, we evaluate whether
these variables can be used to forecast future market movements. Although our
results reveal that the obtained forecasting accuracy cannot be achieved by
chance, we conclude that further information has to be included in the
forecasting model to obtain satisfying results.
Keywords: Media Sentiment, Investor Attention, Behavioral Finance.
1 Introduction
Within recent years, the impact of media sentiment on financial markets has been of
great interest. Recent studies have found that sentiment expressed in traditional
mainstream media like newspapers or social media such as blogs and message boards
is related to stock returns [1, 2]. Additionally, the question whether investors are
aware of financial news has also been examined in order to explain anomalies like the
slow incorporation of new information into stock prices on Fridays [3]. Furthermore,
the impact of an increased number of irrelevant news on the processing of relevant
ones has been examined as well [4].
However, to the best of our knowledge, previous research has neglected the
interplay between investor attention and media sentiment as well as the following
influence on financial markets. Nevertheless, it seems intuitive to take the interaction
between both variables into account: the prerequisite for investors being influenced by
media sentiment is that investors actually notice the news articles published within the
media. Consequently, it can be assumed that a higher level of investor attention leads
to an increased media sentiment impact on financial markets.
Furthermore, the contemporaneous influence of media sentiment on financial
markets still remains underexplored: Recent studies investigate the impact of media
sentiment with at least a one day delay [2, 5]. In contrast, we show that many news
articles are published during the day. Since new information contained within these

news articles is usually processed within short periods of time [6], it also seems
reasonable to investigate the contemporaneous impact of media sentiment.
We contribute to the literature in several ways. At first, we investigate the
contemporaneous impact of media sentiment and investor attention on stock returns.
Second, we investigate the interplay of media sentiment as well as investor attention
and its impact on the following stock market reactions. Third, we forecast future
returns taking into account those variables.
Based on an analysis of the media sentiment related to the Dow Jones Industrial
Average (DJIA), we confirm the impact of media sentiment on DJIA returns and
enhance the previous understanding related to the fact that media sentiment already
has an influence on financial markets at the same day when the corresponding news
are published. Additionally, we find that positive media sentiment has an increased
influence on DJIA returns when investor attention is high. Related to forecasting
future DJIA returns, we conclude that the accuracy that can be obtained is higher than
results that are achieved by chance. However, additional information has to be
included in the forecasting model to obtain satisfying results.
The remainder of this paper is structured as follows. At first, we present related
work concerning the influence of media sentiment and investor attention on financial
markets. Thereafter, we outline the data used within our study, derive our sentiment
measure and describe the proxy for quantifying investor attention. Next, we consider
the joint impact of media sentiment and investor attention on DJIA returns.
Thereafter, we evaluate whether future market movements can be forecasted by
taking into account these variables. Finally, we conclude this paper.
2 Related Work
2.1 Media Sentiment and Financial Markets
Investors are considered to decide to trade because of two general reasons. First, they
take into account new fundamental information like dividend announcements or
management decisions [7]. Second, they are also considered to rely on expectations
that do not follow rational rules [8]. For instance, these expectations can be based on
the advice of “financial gurus” [9] or simply on the sentiment prevailing in the media
that causes them to be overconfident in making the right decisions [10]. In this
context, sentiment expressed in the media covers opinions, expectations or beliefs of
market participants towards certain companies or towards certain financial
instruments [11]. If many investors take media sentiment into account, have similar
(irrational) expectations and follow each other, this can influence stock prices [12].
Recent research has provided evidence for these assumptions. It has been shown that
sentiment expressed in media has an impact on investors’ decision-making activities
and thus affects several financial variables. Consequently, investors act according to
their expectations and buy or sell the respective financial instrument.
Different studies investigate the impact of sentiment expressed in traditional media
like newspapers. Within this context, [5] analyzes a daily Wall Street Journal column
and finds that high media pessimism leads to a decline in market prices. Additionally,

an abnormal high or low level of pessimism is supposed to predict high trading
volumes. A similar study is conducted by [2]. On a daily basis, they analyze the news
stories published in the Wall Street Journal as well as in the Dow Jones News Service
and confirm that stock prices are related to media sentiment. In contrast, [13] evaluate
the sentiment of corporate disclosures (i.e. 10-K company reports) and find that,
compared to general approaches, domain-specific sentiment measures are more
appropriate for sentiment detection within this document type. Furthermore, they
confirm the relation of sentiment and several financial variables.
Compared to these studies, another stream of research focuses on the impact of
sentiment expressed in social media. A seminal study that investigates the impact of
sentiment on a stock level is presented by [1], who collect and analyze messages
posted on two finance message boards. The authors find that disagreement in
sentiment among the messages leads to an increase in trading volume. Additionally,
they observe that the number of messages posted during a day can help to predict the
stock returns during the following day. [14] follow a similar approach and investigate
messages which are published on stock message boards, too. However, they focus on
an index rather than stock level. Thereby, they determine the sentiment for every
message whereas these messages are then used to calculate an overall sentiment
index. [14] find that the level of this sentiment index has explanatory power for the
level of the corresponding stock index. In contrary to this result, they only find weak
evidence that the sentiment concerning individual stocks can forecast daily stock price
movements. Apart from these results, recent studies have found a link between the
sentiment prevailing in microblogging services like twitter and financial markets [15].
The studies presented above provide evidence that sentiment expressed in news
articles or message board postings is related to stock returns. However, these studies
mainly focus on the long-term effect of media sentiment on financial variables.
Instead of taking into account contemporaneous effects, stock returns are related to
the previous days’ media sentiment. However, due to the fact that new information is
often processed within minutes rather than days [6] it seems possible that the related
sentiment has a contemporaneous effect as well.
2.2 Investor Attention and Financial Markets
In recent years, several financial market anomalies have been investigated
theoretically and empirically, such as underreaction and overreaction to financial
news, the influence of weekdays on investors’ reactions as well as the impact of
advertisements on investors’ decisions. Many of these anomalies have been attributed
to the level of investor attention, i.e. the question whether investors are aware of the
current market situation or not. Thus, there is a large number of studies investigating
which instruments receive attention, how corporate advertising impacts the level of
investor attention and how investors pay attention to news published by firms or by
the media in general. Concerning the question which financial instruments are of
interest for different groups of investors, [16] examine how individual and
institutional investors react to “attention-grabbing” stocks. Thereby, they find that
individual investors especially pay attention to stocks which are discussed within the
media, exhibit high abnormal trading volumes and high returns.

Next to the question which financial instruments gain attention in general, another
stream of research investigates how stock recommendations published within the
media influence investor attention. Within this context, it has been found that trading
volumes increase after a stock is discussed within television [17]. Additionally, it has
been figured out that a firm’s advertising expenses lead to an increased number of
individual investors buying a stock [18]. In this case, a spillover effect of
advertisements can be measured: although a firm advertises its product and intends to
increase the product related attention, there is also a higher interest related to its
stocks. These results are confirmed by [19]. Furthermore, [19] find that
advertisements lead to an increase in stock returns in the contemporary year, but they
also note that stock returns decrease in the following year. Considering the company
size, they find that this effect is larger for small firms [19].
Next to stock recommendations and advertisements, investor attention is also
influenced by ordinary news published within the media. In this context, [20] analyze
the market reactions on news of economically linked firms. In this context, they find
that news are incorporated slower when they are not directly related to the firm under
investigation but deal with an economically linked firm. This effect is attributed to a
small degree of investor attention. Additionally, a study by [4] investigates whether
the amount of news articles published within the same period of time has an impact
on market reactions. [4] find that an increased amount of unimportant news decreases
the investors’ reactions to relevant news. Thereby, price and volume reactions are
lower and the post-announcement adjustment to the news is stronger. Thus, as a
result, an increased number of news articles published in the same period is said to
reduce investor attention towards specific news items. Considering investor attention
on different days of the week, [3] find that the response to earnings announcements is
slower on Fridays compared to the remaining days of the week. This effect is
attributed to a lower level of investor attention on Fridays, whereas investors are said
to be more distracted because of the following weekend. Similar results concerning
limited investor attention on Fridays are also reported by [21]. Additionally, [22]
provide evidence that the level of investor attention is related to stock returns.
Based on these studies, it can be noted that the level of investor attention has an
influence on the market reactions following the publication of financial news. In this
context, investors who are aware of the news articles published are confronted with
the corresponding sentiment. As follows, it is more likely that their expectations and
trading decisions are influenced by media sentiment. Thus, a joint effect of both
variables on financial markets can be expected. However, previous studies have not
focused on this specific relation of media sentiment and investor attention.
3 Research Methodology
3.1 Measuring Media Sentiment
In general, sentiment analysis encompasses the investigation of documents like news
articles, message board postings or product reviews in order to determine their tone
concerning a certain topic [23, 24]. There are two broad strategies to perform

sentiment analysis: it can be distinguished between supervised and unsupervised
approaches [25]. Supervised approaches require a dataset composed of documents
that are manually labeled according to the respective sentiment. After several preprocessing
steps, this dataset is used to train machine learning classifiers like naive
bayes or support vector machines. During the training phase, the classifiers search for
patterns within the documents. These patterns can thereafter be used to determine the
sentiment of further documents or sentences. In contrast, unsupervised approaches
rely on external knowledge such as predefined dictionaries providing lists of words
that are connected with a positive or negative sentiment. These word lists are usually
created manually with a couple of precoded terms and are used to determine a
sentiment measure [26].
Within our study, we decide to follow an unsupervised dictionary-based approach
which determines the sentiment taking into account a dictionary containing sentiment
bearing words [26]. This is appropriate because dictionary-based approaches have
proven to be very promising within the financial domain [2, 5, 13]. In contrast,
applying a supervised machine learning-based approach would require a manually
labeled dataset for training whereas manual labeling would be time-consuming and
error-prone.
For unsupervised approaches, different dictionaries are available that contain
positive and negative expressions. Within this study, we make use of the Harvard-IV-
4 dictionary. This dictionary has often been applied in the financial context [2, 5].
Since we analyze general financial news articles rather than specific corporate
disclosures, we make use of this dictionary instead of using the specific dictionary
proposed by [13] which was suggested for the analysis of corporate disclosures.
To calculate a daily sentiment index, we first determine the sentiment of each
document. Accordingly, we obtain the occurrences of positive and negative words by
comparing each news article with the positive and negative word lists. To take
negations into account, we follow [13] and reverse the interpretation of a word if it is
preceded by a negation so that positive words are counted as negative and vice versa.
Thereafter, we adapt a document-level sentiment polarity measure which determines
the direction of the sentiment (i.e. ranging from negative to positive) as well as its
strength [2, 27]:
The measure defined in equation 1 takes into account the number of positive words
pos as well as the number of negative words neg, calculated as described above. If a
document contains neither positive nor negative words, sentdoc is defined as zero. In
line with [2], this measure assumes that all positive and negative words are equally
important, i.e. no weights are assigned to certain words.
As a next step, we determine a daily sentiment index by aggregating the documentlevel
sentiment on a daily basis. Therefore, we calculate the average of sentdoc. In the
following, the resulting daily sentiment index sent that takes into account the
sentiment related to the DJIA is used to investigate the research questions of our
study.
(1)

3.2 Measuring Investor Attention
Within previous studies, different approaches for measuring investor attention have
been proposed. In general, it can be distinguished between indirect and direct
measures of investor attention. On the one hand, a large body of literature deals with
indirect measures of investor attention. Exemplary proxies used are unusual trading
volumes or returns as well as the number of news articles published per day [4, 16,
19, 20]. In this case, it is assumed that large trading volumes or extreme returns
indicate that investors are extremely aware of a stock and respond more timely to new
information, i.e. they trade this stock. As follows, these measures can be denoted as
ex post measures of investor attention. In contrast, the number of news articles per
day can be seen as an ex ante measure of investor attention: an increased amount of
news articles is assumed to lead to an increased amount of investor attention related to
the corresponding financial instruments [16]. However, an increase in these indirect
measures only expresses the results of investors buying or selling a stock (ex post
measures) or a general increase in media attention (ex ante measures). In contrast,
these indirect measures do not indicate whether investors are interested in a financial
instrument or whether the news articles in the media are actually noticed by them at
all [16].
In consequence, [22] provide a seminal paper about the direct measurement of
investor attention, i.e. the measurement of investor attention without relying on tradebased
proxies or the number of news articles published. Instead, they propose to take
into account the amount of web searches related to the company under investigation,
assuming that investors being interested in a financial instrument also search for
related information. In this case, [22] make use of Google’s search volume index
(SVI). Thereby, they find that this measure is correlated with indirect proxies for
investor attention but that it encompasses investor attention in a more timely way.
Furthermore, they note that this measure is especially suited to cover retail investor
attention [22]. Since studies from other domains have already proven the applicability
of the amount of search queries to forecast housing sales, car sales or the outbreak of
influenza (e.g. [28]), we also decide to use SVI as a direct measure for investor
attention.
The SVI can be obtained via Google Trends for different search terms and for
different time horizons (beginning with January 2004). However, SVI is only
displayed for search terms that received attention exceeding a certain (unknown)
threshold. Thus, identifying the correct search term to cover investor attention can be
seen as a crucial step. Within our study, we decide to take the SVI related to the
search term “DJIA” into account to represent investor attention related to the DJIA.
An alternative would have been to download the SVI for each constituent of the DJIA
separately. However, in this case, several problems would arise: As already noted by
[22], some company names are ambiguous (e.g. searching for Kraft). Using ticker
symbols instead could be an alternative, however, there are also some pitfalls in this
case. At first, SVI is not available for every ticker symbol and second, some ticker
symbols are ambiguous, too. For example, searching for “T” as ticker symbol for
AT&T also leads to results related to T-Mobile, “HD” for Home Depot could also be
interpreted as a search for the technical abbreviation “high definition” (as in HD-
DVD) and the same applies to “BA” (Bank of America) which can also be an

abbreviation for British Airways. In these cases, the SVI would not cover the
corresponding ticker symbol and would be inappropriate to measure investor
attention. Thus, we decide to use the SVI for “DJIA” as a proxy for DJIA investor
attention. In contrast to [22], who make use of the weekly SVI, we take into account
the daily SVI in order to measure contemporaneous effects.
3.3 Dataset Acquisition
Within this study, we consider three data sources. First, we acquire financial news
articles in order to determine the media sentiment index. Therefore, we make use of
news articles published by Dow Jones Newswires (DJNS). Second, we download the
SVI related to the DJIA from Google trends. Third, we acquire the corresponding
DJIA closing prices and trading volumes from Yahoo! Finance.
The news articles by DJNS are accessed via the application programming interface
provided by Interactive Data. Thereby, we search for all news articles that are tagged
by DJNS to deal with the constituents of the DJIA. We see DJNS as a representative
source for financial news since DJNS is a major news provider that publishes
financial news throughout the day and whose news are accessed by a large audience
[5]. As revealed by a manual review of the news articles at hand, the assigned labels
are too broad: news articles are already tagged to be related to a certain company
when they mostly deal with its competitors. Thus, we only include those news articles
into our analysis that contain the corresponding search term within the headline. Due
to licence terms, we were able to request all news articles from 2011/01/01 until
2012/02/29, so that 292 trading days could be analyzed. In total, the news article
dataset obtained for this study consists of 13,696 news articles. Thereby, the dataset
consists of different news categories. First, 6,454 regular financial news articles are
included. Second, 7,176 news articles are included that explicitly deal with corporate
disclosures and press releases. Finally, there are 66 news articles included in the
dataset that contain analyst opinions. Thus, our news article dataset covers the full
spectrum of articles that is available within a regular financial newspaper.
As already discussed above, some ticker symbols and company names are
ambiguous. As a result, the SVI cannot be acquired with an adequate accuracy for
each ticker symbol separately. Thus, we decided to acquire the daily SVI for the
search term “DJIA” from Google trends to measure investor attention. In this context,
the SVI can be downloaded relative to the beginning of the corresponding month or
relative to the beginning of the year 2004. Since the first option does not allow to
compare the SVI across several months, we have chosen to download the SVI for our
sample period relative to the search volume in 2004. Finally, we acquire the DJIA
prices and trading volumes for the sample period from Yahoo! Finance.

4 Empirical Results
4.1 Descriptive Results
At first, we consider the daily number of news articles related to the DJIA and its
constituents published by Dow Jones Newswires. Taking into account the daily
distribution as indicated in Fig. 1, it first can be noted that the number of news articles
published per day is not constant over time. Instead, a much smaller amount of news
articles is published on weekends as compared to the rest of the week.
Number of News Articles
3500
3000
2500
2000
1500
1000
500
0
Mon Tue Wed Thu Fri Sat Sun
Day
Fig. 1. Total Number of News Articles Published per Day / per Hour
Second, it can be observed that the number of news articles published from Monday
to Thursday is relatively constant, except from a peak on Tuesday. It is notable that
the number of news articles on Friday is smaller than during the remaining trade days.
This result may be attributed to the fact that the general number of financial news
issued by firms is smaller on Fridays, as already reported by [3], [29] as well as [30].
Thereby, these related studies find that on Fridays, a lower fraction of earnings
announcements is published. As follows, the number of news articles published
dealing with these events is smaller, too. Another explanation could be that next to
investor inattention, also journalists are distracted on Fridays because of the following
weekend.
Next, considering the time of day when news articles are published, it can be noted
that a large peak can be found in the morning at the start of the trading hours (all
times reported in Eastern Standard Time) and a small peak can be found at the end of
the trading hours. Since the news articles published by Dow Jones Newswires are
delivered electronically, a lot of information (and related sentiment) is released during
the day after the newspapers have been printed in the morning.
Furthermore, we consider the amount of Google searches for the Keyword “DJIA”.
In this case, we find abnormal high search volumes in August 2011 (see Fig. 2).
These high levels occur simultaneously with a decline of the DJIA caused by weak
economic perspectives. Thus, we control for this abnormal movement within our
further analyses. Additionally, Fig. 3 shows that investor attention is low on
weekends. Thus, investors are distracted on Saturdays and Sundays [3].
Number of News Articles
1400
1200
1000
800
600
400
200
0
1 AM 3 AM 5 AM 7 AM 9 AM 11 AM 1 PM 3 PM 5 PM 7 PM 9 PM 11 PM
Time

SVI
25
20
15
10
5
0
20110101
20110116
20110131
20110215
20110302
20110317
20110401
20110416
20110501
20110516
20110531
20110615
20110630
20110715
Fig. 2. Daily raw SVI Time Series
Finally, Table 1 presents the means, standard deviations and correlations of the main
variables of interest. Since stock returns are measured as a percentage change and are
only available for trading days, we do not include the raw measures of dowt, sentt and
SVIt. Instead, we calculate the percentage change of sentiment and investor attention
on trading days, report these values in the following table and include them in our
further analyses. Non-trading days are excluded. As Table 1 shows, we find small
positive correlations, except for SVI. In this case, SVI is negatively correlated to
DJIA returns. Furthermore, we find a small negative correlation of SVI and sent.
However, this correlation is not statistically significant (p = 0.84).
Table 1. Means, Standard Deviations and Correlations
Mean Std. Dev. dj sent sent x SVI SVI
dj 0.00046 0.00073 1.0000
sent 0.04286 0.02078 0.0754 1.0000
sent x SVI -0.00017 0.00290 0.0657 0.0066 1.0000
SVI 0.01162 0.00935 -0.3205*** -0.0118 0.1380** 1.0000
* / ** / *** = significant at a 10% / 5% / 1% level of significance
4.2 Impact of Media Sentiment and Investor Attention on Stock Returns
To investigate the impact of media sentiment and investor attention on DJIA returns,
we regress the DJIA returns (dowt) on our daily sentiment measure (sentt), investor
attention (SVIt) as well as the moderating effect taking into account both variables
(sentt x SVIt):
Within Equation 2, εt denotes the error term. Additionally, Controlst stands for
several control variables that are also included to ensure that the results are not biased
because of further effects possibly influencing stock returns: To control for day
patterns of stock returns and for the January effect that can cause abnormal stock
returns, we include dummy variables for the different trading days as well as for
20110730
Day
20110814
20110829
20110913
20110928
20111013
20111028
20111112
20111127
20111212
20111227
20120111
20120126
20120210
20120225
(2)

January [5]. Furthermore, to account for the developments within August 2011, we
also include a dummy variable for this month. Additionally, we include variables for
past volatility 1 , previous trading volume 2 as well as previous DJIA returns up to five
lags [5]. Within the regression, we use heteroscedasticity- and autocorrelationconsistent
standard errors [32]. The results of the regression are denoted in Table 2. In
order to test for multicollinearity, the variance inflation factor was calculated for each
independent variable. Thereby, no multicollinearity was detected since the highest
score of 2.17 is below common thresholds of 4 and 10 [33].
Table 2. Impact of Media Sentiment and Investor Attention on DJIA Returns
Coefficient Standard Error
sent 0.0033198** (0.0015118)
sent x SVI 0.0323227* (0.0187595)
SVI -0.0284493*** (0.0068986)
* / ** / *** = significant at a 10% / 5% / 1% level of significance
At first, the results confirm the impact of media sentiment on stock prices. As
indicated by a positive coefficient for the sentiment measure, we can note that an
increase in media sentiment leads to an increase in the corresponding DJIA return.
Thereby, the coefficient is significant at a 5% level of significance.
Considering the joint impact of investor attention and media sentiment on DJIA
returns, we also find a positive relationship which is significant at a 10% level of
significance. In this context, the positive effect of media sentiment on DJIA returns is
increased when investor attention is high. Although an increased SVI does not imply
that investors actually read the news articles published via DJNS, it can be noted that
an augmented interest in the corresponding topic (i.e. the DJIA) prevails. Since the
news articles at hand are published on several websites as well, it is more likely that
the news articles are actually read by investors’ searching for information via Google.
As follows, more investors are confronted with a certain level of media sentiment and
consequently, their trading decisions are influenced.
Interestingly, the sole impact of investor attention on DJIA returns is negative,
whereas the coefficient is significant at a 1% level of significance. At a first sign, this
result contradicts previous research. In this context, [22] find a positive relationship of
investor attention (measured by the amount of Google searches) and stock returns.
However, [22] show that, when controlling for market capitalization, the positive
price pressure is only present among the smaller half of their stock sample.
Furthermore, in their study, an interaction effect of market capitalization and SVI has
a negative impact on returns [22]. Since the constituents of the DJIA have a high
market capitalization and are analyzed on an aggregated level, these results do not
contradict previous studies. Considering the control variables, the results remain
robust when including a dummy variable for August 2011. Thereby, the day-of-week
1 Thereby, the approach proposed by [5] is followed: to account for past volatility, the daily
returns of the DJIA are demeaned to obtain a residual, this residual is squared and the past
60-day moving average is subtracted.
2 Specifically, the detrended logvolume is used as proposed by [31].

dummy variables as well as the dummy variable for January have no significant
influence, whereas few of the lagged control variables for previous returns, volatility
and trading volumes have a significant influence (not reported in Table 2 due to space
constraints).
5 Predicting Bidirectional Market Movements based on Media
Sentiment and Investor Attention
5.1 General Setup
In the following, we investigate whether the influence of media sentiment and
investor attention on DJIA returns can be taken into account to forecast future market
movements. Thereby, we focus on predicting DJIA returns by means of machine
learning techniques. In this case, machine learning techniques are advantageous
because of two main reasons. At first, the evaluation becomes more reliable since
evaluation methodologies like 10-fold cross validation can be used [34]. With this
respect, 10-fold cross validation offers the possibility to use an increased number of
data items for evaluating the trained model. Second, machine learning classifiers like
Support Vector Machine (SVM) are also suitable to cover nonlinear relationships
within the data which may improve forecasting results [35].
For predicting DJIA returns, we make use of supervised learning and train a
machine learning classifier with labeled training data in order to find patterns within
the data that can serve for future predictions. Thus, every observation (i.e. each
trading day) is labeled according to the corresponding DJIA return. Thereby, we
assign two labels: the first label is assigned according to the contemporaneous DJIA
return (T0), the second label is related to the one day ahead return (T0+1). We follow
previous studies and focus on two classes [36]: the class negative is assigned if the
corresponding DJIA return is lower than zero, otherwise, the class positive is
assigned. In total, for T0 and T1 forecasts, 161 (131) observations are labeled as
positive (negative).
Within this study, we make use of a Support Vector Machine (SVM) classifier
since SVMs have proven to be a good choice for financial forecasting [37]. Thereby,
the same input variables have been used that were already defined in section 4.2, i.e.
media sentiment, investor attention, the interaction term as well as the control
variables.
5.2 Evaluation
To evaluate the proposed machine learning setup, we make use of 10-fold stratified
cross validation [34]. In this case, the whole dataset is split into 10 subsets, whereas
each subset is used k-1 times for classifier training and once for classifier testing. At
the end of each iteration, a contingency table is created that contains the number of
true positives (TP), false positives (FP), true negatives (TN) and false negatives (FN).
Finally, a global contingency table is created by summing up the different

contingency tables (micro-averaging) [38]. Based on this global contingency table,
different performance metrics are calculated. Thereby, we focus on accuracy,
indicating the percentage of cases classified correctly as well as Precision, Recall and
the F1-measure [39, 40]. These metrics are defined as follows:
(3)
(4)
The results of our evaluation for the contemporaneous as well as the T0+1 forecasts
are depicted in Table 3. Thereby, a SVM classifier using a radial basis function kernel
(as proposed by [41]) has been made use of. The corresponding parameters have been
selected via grid search 3 .
Table 3. Forecasting DJIA Market Movements Using Media Sentiment and Investor Attention
Class: positive Class: negative
Forecast Accuracy Precision Recall F 1 Precision Recall F 1
T 0 58.55 58.40 86.34 69.67 59.26 24.43 34.60
T 0+1 58.56 59.71 76.40 67.03 55.81 36.64 44.24
All values are given as percentages.
Related to the predictability of DJIA market movements based on media sentiment
and investor attention, it can be observed that the obtained results are better than
results being achieved just by chance. This is evidenced by the fact that the precision
scores are above 50% in all cases. Additionally, contemporaneous market movements
as well as the returns of the following day can be predicted with similar accuracies of
58.55% and 58.56% respectively. Considering the class recall, we find that recall of
the class positive is higher than the recall of the class negative which can be attributed
to the class distribution within our sample: 55% of all observations are labeled as
positive and, as follows, the SVM is trained respectively.
However, taking the economic value of these results into account, an accuracy of
below 60% cannot be considered as promising. Thus, using only these structured
variables as input data for a decision model can hardly be seen as a source for
significant profits. Instead, many cases are classified incorrectly. This may be
attributed to the noisy nature of financial markets and to the fact that the decision
model does not take into account the textual information published within the news
articles under investigation. As a consequence, investor sentiment and investor
attention should not be used solely to forecast market movements. Instead, they
should be incorporated in existing forecasting models to improve forecasting results.
3 we followed the procedure proposed by [41] and evaluated the proposed values for C, a
penalty parameter and γ, a parameter of the radial basis function. For To+1, C = 512 and γ =
2 -15 lead to the best results. In the case of T 0, C= 32 and γ = 2 -15 were selected.
(5)
(6)

5 Summary and Conclusion
In recent years, the impact of media sentiment on financial variables like stock prices
has been of great interest. However, one crucial prerequisite of relating media
sentiment to financial variables has not been taken into account: current studies do not
consider whether the news articles expressing sentiment are actually noticed by
investors. As a consequence, we examine the interplay between media sentiment and
investor attention in order to investigate the joint impact of both variables on Dow
Jones Industrial Average returns.
Based on an empirical analysis of the sentiment expressed within 13,696 financial
news articles, we find that higher investor attention increases the impact of media
sentiment on DJIA returns. Thus, when investors actually pay attention to a financial
instrument, the impact of media sentiment on these financial variables is higher.
Furthermore, this effect is already measured at the same day rather than with a delay
of several days. If the variables under investigation are used to forecast the DJIA
returns, it can be observed that the results obtained are higher than results being
achieved just by chance. However, there are still many cases which are classified
incorrectly. As follows, further (unstructured) information has to be incorporated
within the decision model in order to improve forecasting results. For example,
textual inputs or technical indicators may also be considered to incorporate the
information published as well as current market trends [42, 43].
Within our study, media sentiment and investor attention are measured on a daily
basis. As a consequence, our study does not cover intraday effects of media sentiment
and investor attention on stock returns. In this context, the intraday stock price impact
of media sentiment may be measured by taking into account tick-by-tick trading data.
However, since SVI is only available on a daily basis, we are aware of the limitation
that actually, intraday effects of both variables are not covered and should be
investigated as soon as an intraday SVI is available. Additionally, within further
research, the interplay between media sentiment and investor attention as well as its
impact on financial variables should also be examined at a stock level. Furthermore,
since the effect of investor attention among small-capitalized stocks has found to be
higher [22], less frequently traded stocks will also be incorporated in our analysis.
Finally, contemporary research reports that traditional news media and social media
are interconnected. In this case, topics that are discussed within newspapers are also
talked about within blogs [44]. Thus, the discussions within social media could also
be analyzed in order to develop a more fine-grained indicator for measuring investor
attention at a topic level.
Acknowledgements. The research leading to these results has received funding from
the European Community's Seventh Framework Programme (FP7/2007-2013) within
the context of the Project FIRST, Large scale information extraction and integration
infrastructure for supporting financial decision making, under grant agreement n.
257928.

References
1. Antweiler, W., Frank, M.Z.: Is All That Talk Just Noise? The Information Content of
Internet Stock Message Boards. The Journal of Finance 59, 3, 1259–1294 (2004)
2. Tetlock, P.C., Saar-Tsechansky, M., Macskassy, S.: More Than Words: Quantifying
Language to Measure Firms' Fundamentals. The Journal of Finance 63, 3, 1437–1467
(2008)
3. Dellavigna, S., Pollet, J.M.: Investor Inattention and Friday Earnings Announcements. The
Journal of Finance 64, 2, 709–749 (2009)
4. Hirshleifer, D., Lim, S.S., Teoh, S.H.: Driven to Distraction: Extraneous Events and
Underreaction to Earnings News. The Journal of Finance 64, 5, 2289–2325 (2009)
5. Tetlock, P.C.: Giving Content to Investor Sentiment: The Role of Media in the Stock
Market. The Journal of Finance 62, 3, 1139–1168 (2007)
6. Muntermann, J., Guettler, A.: Intraday stock price effects of ad hoc disclosures: the German
case. Journal of International Financial Markets, Institutions and Money 17, 1, 1–24 (2007)
7. Shleifer, A., Summers, L.H.: The Noise Trader Approach to Finance. The Journal of
Economic Perspectives 4, 2, 19–33 (1990)
8. Black, F.: Noise. The Journal of Finance 41, 3, 529–543 (1986)
9. Shleifer, A., "Inefficient markets. An introduction to behavioral finance". Oxford Univ.
Press, Oxford (2000)
10. de Bondt, W.F. M.: A portrait of the individual investor. European Economic Review 42, 3-
5, 831–844 (1998)
11. Brown, G.W., Cliff, M.T.: Investor sentiment and the near-term stock market. Journal of
Empirical Finance 11, 1, 1–27 (2004)
12. Shiller, R.J.: Stock Prices and Social Dynamics. Brookings Papers on Economic Activity
1984, 2, 457–510 (1984)
13. Loughran, T., McDonald, B.: When Is a Liability Not a Liability? Textual Analysis,
Dictionaries, and 10-Ks. The Journal of Finance 66, 1, 35–65 (2011)
14. Das, S.R., Chen, M.Y.: Yahoo! for Amazon: Sentiment Extraction from Small Talk on the
Web. Management Science 53, 9, 1375–1388 (2007)
15. Bollen, J., Mao, H., Zeng, X.-J.: Twitter mood predicts the stock market. Journal of
Computational Science 2, 1, 1–8 (2011)
16. Barber, B.M., Odean, T.: All That Glitters: The Effect of Attention and News on the
Buying Behavior of Individual and Institutional Investors. Review of Financial Studies 21,
2, 785–818 (2008)
17. Busse, J.A., Green, T.C.: Market efficiency in real time. Journal of Financial Economics
65, 3, 415–437 (2002)
18. Lou, D.: Attracting investor attention through advertising. Discussion paper 644. Financial
Markets Group, London School of Economics and Political Science, London, UK. (2009)
19. Chemmanur, T.J., Yan, A.: Advertising, Attention, and Stock Returns. Working Paper.
Available at SSRN: http://ssrn.com/abstract=1340605 (2009)
20. Cohen, L., Frazzini, A.: Economic Links and Predictable Returns. The Journal of Finance
63, 4, 1977–2011 (2008)
21. Louis, H., Sun, A.: Investor Inattention and the Market Reaction to Merger
Announcements. Management Science 56, 10, 1781–1793 (2010)
22. Da, Z., Engelberg, J., Gao, P.: In Search of Attention. The Journal of Finance 66, 5, 1461–
1499 (2011)
23. Pang, B., Lee, L., Vaithyanathan, S.: Thumbs up? Sentiment Classification using Machine
Learning Techniques. Proceedings of the Conference on Empirical Methods in Natural
Language Processing, Philadelphia, Pennsylvania, USA, 79–86 (2002)
24. Pang, B., Lee, L.: Opinion Mining and Sentiment Analysis. Foundations and Trends in
Information Retrieval 2, 1-2, 1–135 (2008)

25. Zhou, L., Chaovalit, P.: Ontology-supported polarity mining. Journal of the American
Society for Information Science and Technology 59, 1, 98–110 (2008)
26. Thelwall, M., Buckley, K., Paltoglou, G.: Sentiment in Twitter events. Journal of the
American Society for Information Science and Technology 62, 2, 406–418 (2011)
27. Zhang, W., Skiena, S.: Trading Strategies To Exploit Blog and News Sentiment.
Proceedings of the 4th International AAAI Conference on Weblogs and Social Media,
Washington, DC, USA (2010)
28. Ginsberg, J., Mohebbi, M.H., Patel, R.S., Brammer, L., Smolinski, M.S., Brilliant, L.:
Detecting influenza epidemics using search engine query data. Nature 457, 7232, 1012–
1014 (2009)
29. Penman, S.H.: The distribution of earnings news over time and seasonalities in aggregate
stock returns. Journal of Financial Economics 18, 2, 199–228 (1987)
30. Patell, J.M., Wolfson, M.A.: The intraday speed of adjustment of stock prices to earnings
and dividend announcements. Journal of Financial Economics 13, 2, 223–252 (1984)
31. Campbell, J.Y., Grossman, S.J., Wang, J.: Trading Volume and Serial Correlation in Stock
Returns. The Quarterly Journal of Economics 108, 4, 905–39 (1993)
32. Newey, W.K., West, K.D.: A Simple, Positive Semi-Definite, Heteroskedasticity and
Autocorrelation Consistent Covariance Matrix. Econometrica 55, 3, 703–708 (1987)
33. O'Brien, R.M.: A Caution Regarding Rules of Thumb for Variance Inflation Factors.
Quality & Quantity 41, 673–690 (2007)
34. Mitchell, T., "Machine learning". McGraw-Hill, London (1997)
35. Dhar, V., Stein, R., "Intelligent decision support methods. The science of knowledge
work". Prentice Hall, Upper Saddle River, NJ (1997)
36. Groth, S.S., Muntermann, J.: Supporting Investment Management Processes with Machine
Learning Techniques. Proceedings of the 9th Internationale Tagung Wirtschaftsinformatik,
vol. 2, Vienna, Austria, 275–284 (2009)
37. Tay, F.E. H., Cao, L.: Application of support vector machines in financial time series
forecasting. Omega 29, 4, 309–317 (2001)
38. Chau, M., Chen, H.: A machine learning approach to web page filtering using content and
structure analysis. Decision Support Systems 44, 2, 482–494 (2008)
39. Hotho, A., Nürnberger, A., Paaß, G.: A Brief Survey of Text Mining. GLDV Journal for
Computational Linguistics 20, 1, 19–62 (2005)
40. van Rijsbergen, C.J., "Information retrieval", 2nd ed. Butterworths, London (1979)
41. Hsu, C.W., Chang, C.C., Lin, C.J.: A practical guide to support vector classification.
National Taiwan University, http://www.csie.ntu.edu.tw/~cjlin/papers/guide/guide.pdf
(2011-10-16) (2003)
42. Mittermayer, M.-A.: Forecasting Intraday Stock Price Trends with Text Mining
Techniques. Proceedings of the 37th Hawaii International Conference on System Sciences,
Big Island, Hawaii, USA (2004)
43. Lam, M.: Neural network techniques for financial performance prediction: integrating
fundamental and technical analysis: Data mining for financial decision making. Decision
Support Systems 37, 4, 567–581 (2004)
44. Leskovec, J., Backstrom, L., Kleinberg, J.: Meme-tracking and the Dynamics of the News
Cycle. Proceedings of the 15th International Conference on Knowledge Discovery and Data
Mining (2009)

MEASURING THE INFLUENCE OF PROJECT
CHARACTERISTICS ON OPTIMAL SOFTWARE PROJECT
GRANULARITY
Weber, Moritz Christian, Goethe University Frankfurt, Campus Westend, Grüneburgplatz 1,
60329 Frankfurt, Germany, moweber@wiwi.uni-frankfurt.de
Wondrak, Carola, Goethe University Frankfurt, Campus Westend, Grüneburgplatz 1,
60329 Frankfurt, Germany, cwondrak@wiwi.uni-frankfurt.de
Abstract
With a well-structured design of reusable software, it is possible to save costs in future software
development cycles and generate added value. Nowadays it is difficult to quantify which grade of
project structuring enhances its reusability. This is the question this paper aims to examine, based on
the financial Portfolio Selection theory and Value at Risk methods that can be adapted to the field of
software project structuring. Using these methods, the value of the reusable source code can be
evaluated and the risks of reimplementation be compared with alternative code structures. So we can
quantify and measure the effects of the software project structure on the risks of reimplementation.
This model is applied to 27 Github open source software projects and enables us to investigate the
influences of the project characteristics on the best possible software project granularity.
Keywords: Project Management, IT Project Portfolio, Software Granularity, Value at Risk.

1 Motivation
Static and monolithic information systems mark the beginning of enterprise IT infrastructures (Krafzig
et al., 2004). With the emergence of spreading network structures, these inflexible infrastructures are
split into subsystems and managed as interrelated parts of an IT portfolio (Erl, 2005). ‘The foundation
of an IT portfolio is the firm’s information technology’ (Weill and Vitale, 2002). This is — among IT
components and human infrastructures — defined by shared IT services and standard applications
(Weill and Vitale, 2002). Not only software development, but also general strategy research reflects
synergies and diversifications of portfolios. These allow the generation of additional positive
economic values gained by the merging and restructuring of financial portfolios (Amihud and Lev,
1981; Christensen and Montgomery, 1981; Farjoun, 1998). As the realization of synergistic economies
is crucial, it can be seen as a third pillar next to financial and vertical economics (Hill and Hoskisson,
1987). Markowitz’s Portfolio Selection theory (1952) defines a theoretical framework for portfolio
diversification and synergies which can be extended (Myers 1982) towards strategic planning and has
the potential to be applied in other research domains (Rockafellar and Uryasev, 1999). As IT projects
are also aggregated in portfolios, this theory supports evidence as to how software portfolios should be
structured.
Synergies typically depend on the size and structure of the portfolio (Eckbo, 1983; Bradley et al.,
1983). Two kinds of synergies can be distinguished for IT portfolios (Cho and Shaw, 2009;
Tanriverdi, 2005; 2006): On the one hand bundling two components together allows sub-additive cost
savings within similar processes (Teece, 1980). On the other hand, two interrelated components can
actually enable an additional, super-additive benefit at the same time as increasing the effort for both
(Milgrom and Roberts, 1995). Economies of scope enhance the potential of those synergies (Teece,
1980; 1982) and the ability to diversify a portfolio (Willig, 1978). This gives reason to expect lower
risk costs and higher generated added value (Panzar and Willig, 1981), also for IT project portfolios.
Particularly the risk stemming from size and the structure of a portfolio can be expressed in
granularity. In the following sections granularity measurement is defined as ’quantifying the
contribution of [...] concentrations to portfolio risk’ (Lütkebohmert and Gordy, 2007). Unlike the
monolithic enterprise IT infrastructures at the beginning, it has not yet been investigated how granular
and diversified a software portfolio should be from an economical point of view (Friedl, 2011). It is
said that two thirds of all software projects fail or are not finished within the fixed requirements or the
given time. This is also due to organizational failures in software development (Standish, 2011). That
is why we ask the following research questions:
� How granular should software project structures be?
� How can project-granularity levels be measured and compared?
� How can an optimum level of granularity be determined at present, so that good reusability
and lower change risks can be expected?
� Which project characteristics have a critical influence on the optimum of granularity?
In the following section, we reflect the Portfolio Selection theory and the Value at Risk method (2.1)
as theoretical framework to value portfolio risks. Related work on IT portfolio risk, project
management (2.2) and granularity analysis (2.3) let us derive research hypotheses on risk costs, project
characteristics and granularity from literature. To analyze these hypotheses, we adopted Portfolio
Selection and Value at Risk to measure quantitatively risk costs in software project data (3.1 and 3.2).
Alternative levels of project granularities are simulated by three methods (3.3) that are applied to
software version control data for the 27 most observed Github (2011) open source projects. The Value
at Risk is measured for each method and each simulated project granularity level. Accordingly, a
regression setup is developed to explain the influences of project characteristics on optimal software
project granularity (3.4). As a result alternative project granularities are selected to show and discuss

corresponding changes of risk costs (4.1, 4.2 and 4.3). Changes in the project characteristics are
measured over time for all projects and each version. These changes are taken to explain their
influence to the granularity level and the minimal risk costs of software projects (4.4). Finally we
summarize our results, highlight limitations and provide an outlook on future research (5).
2 Related work
As it is often claimed that software projects are structured like portfolios with component assets, we
align our methodology with the Portfolio Selection theory (Markowitz, 1952) and the risk
measurement approach called Value at Risk (Wolke, 2008) in Section 2.1. This economic perspective
is applied to IT project portfolio management and risk management to link the theory to the scope of
information systems research (Section 2.2). It is often assumed that software development projects –
as a subset of IT projects – have a high risk of failure (Standish, 2011). That is why we supplement our
review of related work by recent studies on structuring of software development projects and
analyzing the quality of software granularity (Section 2.3).
2.1 Portfolio Selection and Value at Risk
A software project contains several components that are linked with each other. This is similar to a
financial portfolio where each portfolio component generates a value contribution. Since software can
be a substantial long-term investment, it is important to select portfolio components in the most
efficient way, so that a high level of reusability and lower maintenance costs can be anticipated.
Credit portfolio
short
fall
short
fall
short
fall
short
fall
likely
VaR
95%
short
fall
likely
VaR
95%
short
fall
likely
VaR
95%
risky positive very
positive
Changes
risky positive very
positive
Changes
risky positive very
positive
Changes
How many credits of which size are risk optimal?
Select the portfolio
granularity with the
lowest shortfall risk
Figure 1. Portfolio diversification and risk in the context of correlation effects
Granularity
optimal
Value at Risk
A theoretic-based approach for the selection of components of a financial portfolio is the Portfolio
Selection theory developed by Harry Markowitz (1952). It outlines how to structure an optimal
portfolio with minimal risks by efficiently dividing it into partial investments (so called
diversification). Thereby, interrelated influences between the components are taken into account and
assembled in a single structure, so that the total risk of value depreciation is minimized (so called
correlation). The value depreciation can be estimated using the Value at Risk (VaR) approach
(Albrecht and Maurer, 2005) (see Figure 1). It ‘can be applied to any risk-factor model of portfolio [...]
risk, and [is ...] a very general framework’ (Lütkebohmert and Gordy, 2007). Hereby, the maximum
value depreciation is a value that will not be exceeded in a given time period with a given probability
(Wilkens, 2001; Wolke, 2008). The VaR approach calculates not only the expected value of losses
(Expected Loss), but also explicitly outlines the unexpected losses (Campbell et al., 2001). Like this, it
enables us to include the extreme cases of value depreciation in our estimations.
This concept can be applied to any single evaluation, but also to multiple investment assets and entire
investment portfolios. If there are multiple investments with considerable risks in one portfolio,
additional synergy effects need to be taken into account (Wiedemann, 2002). Value fluctuations of
these components could be both chances and risks for the whole portfolio. These risks can be

diversified by combining components which compensate the interrelated value depreciation risks. The
goal is to find synergies that compensate the risks among themselves (Rockafellar and Uryasev, 1999).
Through systematic validation of different portfolio structures, it is possible to assess and valuate the
risk of value depreciation using the VaR approach. In the case of inefficient combinations, the so
called ‘clustering risks’ can emerge (Huschens and Stahl, 2004). This kind of insufficient
granularisation leads to the danger of a higher, diversified VaR (Albrecht and Maurer, 2005). The
Portfolio Selection theory allows evaluating the efficient selection of portfolio components, so that the
risk costs of value depreciation will decline for the future period. Rockafellar and Uryasev (1999)
explicitly suggest that the VaR concept should be applied as methodology in other areas. Outside of
the financial research domain we adopt this for managing the portfolio risk of IT projects.
2.2 IT portfolio risk and project management
Cho and Shaw (2009) model and simulate the influences of synergies on the IT portfolio selection by
applying the Markowitz theorem and the VaR approach to the research scope of IT portfolios. They
find that not only portfolio returns, but also portfolio risks need to be taken into account. Moreover,
they argue that only few studies investigated the influences of IT synergy on portfolio risks. Benaroch
(2002; 2006) analyzes IT risks and their influence on large portfolios of IT investments. Strong
evidence is found that IT risk management is largely driven by intuitions, which leads to suboptimal
and counterproductive results. Tanriverdi and Ruefli (2004) state that synergies of IT projects also
influence the risk of those. A failure of one IT portfolio component may cause failures in other parts of
the portfolio, which may increase the risks at higher levels of diversification (Tanriverdi, 2005; 2006).
This indicates that the following hypothesis on software project granularity should be rejectable:
Hypothesis H10: Risk costs have no crucial influence on optimal software project granularity.
The organization of software development project portfolios can be improved by a well coordinated
project management. Lichter and Mandl-Striegnitz (1999) show that most software development
projects fail because of inappropriate organizational settings and not because of the technical
implementations. In addition, inefficient processes may cause design errors and an unexpected
increase in project costs. A rule of thumb on how the cost of a single present design error influences
future development and refactoring expenditures is given by so called ‘Rule of Ten’ (Pfeifer, 2001). It
estimates that an undiscovered design error multiplies the required cost of rectification by ten for each
following period. It is possible to recognize design errors earlier by using efficient project
management and verifiable key figures (Lichter and Mandl-Striegnitz, 1999).
Especially the reuse of previously implemented software artefacts is an opportunity to reduce costs
and work effort in multiperiodic software projects. Necessarily the design of software artefacts should
not be too complex and problems with complex interdependence between the artefacts should be taken
into account (Schirmer and Zimmermann, 2008). This involves both, dependencies between different
parts of the software and a good kind of granularity. More specific, special attention must be paid to
project characteristics, the structure of the portfolio and the selection of subprojects (Schirmer and
Zimmermann, 2008). So we expect to reject the following null hypothesis:
Hypothesis H20: Project characteristics have no influence on optimal software project granularity.
2.3 Granularity analysis
Different approaches analyze optimal software project structures (Krafzig et al., 2004, p.18). Thereby,
efforts are being made to quantify the quality of structures by measuring the utility and monetary value
to software structuring decisions: Erradi et al. (2006) propose a quantitative ranking with numerical
steps up to 10, which considers the level of dependencies between the project components.
Furthermore, they criticize the lack of theory that would support the findings of correct granularity for
maximal component reuse and lower project risks (Erradi et al., 2006). Alternative research strategies
evaluate the outsourcing and awareness potential of distributed projects and identify project

structuring candidates by an economical analysis (Klose et al., 2007). In their empirical analysis
Braunwarth and Friedl (2010) took the internal cost allocation of a company as a benchmark and
simulate potential structures of granularity. In addition of this study, Friedl classified granularity
approaches based on a literature review and gave advice for designs of good granularity on typical
factors of influence in software development projects (Friedl, 2011). Katzmarik supports the
granularity decision with a micro-economical approach, in which the project structures are modeled as
‘Software as a Service’-products (SaaS) and evaluated (Katzmarzik, 2011). Therefore, we assume that
the following null hypothesis cannot be supported:
Hypothesis H30: Optimal software IT project granularity cannot be estimated empirically.
3 Methodology
In the following section the VaR concept is applied to software development projects by using version
change data from software control systems. The software project structures as well as the historical
variance and covariance of source code lengths are measured. Through three different methods
alternative granularity structures are simulated and the level of minimal change expenditure risk is
determined by VaR valuation. Finally, this minimal level is regressed against the given project
characteristics to measure the influences of the project parameters on the optimal level of granularity.
3.1 Assumptions
In accordance with the Portfolio Selection theory (Markowitz, 1952) it is assumed that each portfolio
can be diversified by a requirement-optimal or risk-minimal allocation, if a value can be determined.
Modern software engineering approaches separate classes, services and other entities into single files
(Erl, 2005). Therefore, we assume that each change of a file is an implicit change of a portfolio
component. Agile software development practitioners (like in open source projects) claim that each
change of a software project implements a single functional requirement or bug fix, which is
documented in our case by the version control system. For simplification, it is assumed that the change
expenditure of a file is highly correlated with the number of changed and unchanged lines of code
(LOC). It is further assumed that the kind of development and refactoring is anchored in the project
characteristics and that it stays similar in the history of a project. That is why historical change data is
a good estimate of future change risks. The current size of a source code file is taken as present value.
Using historical recorded project development costs, the LOC calculation can easily be transformed
into a utility measure by applying COCOMO methodology (Katzmarzik, 2011, p. 23).
3.2 Data aggregation and measuring method
The data of software version control systems is used to measure direct development structures and
characteristics of software projects. Thereby, the historical sizes of each file as well as realized
software changes are measured. So the extent of each requirement implementation or bug fix can be
precisely determined. Each requirement implementation can affect multiple file changes. Like Grigore
and Rosenkranz (2011), Turek et al. (2010) and Wierzbicki et al. (2010) we measure the residual value
of reusable source code: This is the number of document lines which are created by one author and
remain in the document after the document revision (adds/removes). This residual value of reusable
code is identified for each file of every requirement implementation (patch). This value is saved with
the residual values of the corresponding files of each change (see Figure 2 ‘Data aggregation’).
The VaR concept is applied (Wiedemann, 2002) to identify at which structural level the residual value
of reusable source code remains the highest possible. This determines the loss that is not exceeded in
the covered period of time with a given probability. Thus the calculation also takes historical
fluctuations and correlative dependencies between the components of a portfolio into account. Here,
the fluctuations of the residual value of reusable code (determined in the data aggregation) are used.
For all pairwise changed files, a covariance value is calculated and aggregated in a covariance matrix

(Wiedemann, 2002). The risk value is calculated by matrix multiplications of the file lengths at the
current point in time and the variance-covariance matrix of relative residual values of reusable source
code. Finally, the risk quantile for the given probability is determined (Wilkens, 2001; Wolke, 2008)
(see Figure 2 ’Measuring method’).
Version control system
Data aggregation Measuring method Simulation
c
a
b
Residual value of reusable source
code
=
Lines per file – all changes
t0 t1 t2 t3
b
a
Original
file
Covariance of value
fluctuations
t0 t1 t2 t3
Covariance
Reusable source code
- +
New version
after revison
divVaR = LOC*cov(a,b,..)*(LOC)’
Maximum loss of residual value with
given probability in a period
Original corr.
(tar method)
a b c d
1
10 -1
10 -2
.
.
e f g h
Strong corr.
(corr method)
a
e
b
d
Increasing granularity
Search for optimal granularity by
valuing the minimal diversified
Value at Risk
c
Weak corr.
(rand methode)
a,c,g,h,..
1 2 3 4 5 6 7 8
The worst granularity level is normalized to 1 for validation of
the ten spot rule
Figure 2. Data aggregation, measuring method and simulation of the analysis
3.3 Simulation
The granularity of every historic grown project is given implicitly by the project structure. Therefore,
a diversified VaR can be calculated directly. It quantifies risk of reimplementation for the given
granularity structure. To evaluate the risk of alternative granularity levels, variants with comparable
structures must be derived. A sufficient effective method (hereinafter ‘tar method’) starts with the
lowest folder level and aggregates all files per folder into a single file. Afterwards this file is moved to
the superior folder (Lee et al., 2007). This way, granularity variants can be generated which keep the
same basic structure. In a simplification of Lee’s approach, the common tar algorithm is recursively
applied (FSF, 2006). With source code packages and functional structuring of folders the dependency
structures and conditional changes remain unaltered even for extensive aggregations. The most coarse
and presumable worst granularity generated by this method, is to copy all source code into a single
file. For each aggregation level, the size of files and changes as well as the covariance are remeasured
and a risk value is calculated using the VaR approach.
The results of tar methodology are validated by methodological triangulation using two additional
control methods. The first validation method (‘corr method’) selects an initial file from the project for
each simulation run (Blobel and Lohrmann, 1998). Starting with this file, dependently changed files
are recursively added. This is repeated until the number of files fits the amount of files of the
corresponding granularity level of the tar method. The second method (‘rand method’) selects random
files, but does not consider correlation structures. Thus, the first control method simulates subprojects
of the entire project with a strong correlation. The second method simulates the omission of these
correlations. Huschens and Stahl (2004) show that changes in the correlation structures influence the
relative evaluation of granularities only slightly. The structure of the entire project is approximated by
averaging the partial results (Blobel and Lohrmann, 1998).
The validation methods yield good comparative values. But it should be kept in mind that these
methods are too aggressive and too coagulative in their simulation/estimation, due to their random
selection and partial simulation. As a result, badly diversified and extreme project structures can be
generated. This is why these two validation methods are used for triangulation reasons only and not
for measuring the absolute risk levels of the relative residual value of the reusable source code. To
facilitate comparison — also with the ‘Rule of Ten’ (Pfeifer, 2001) — we start at the granularity with
Risk
costs
Granularity

the highest VaR for each method. Subsequent all VaRs are normalized to this initial value of 1. The
normalized value of 1 is usually the case, if just one file contains all source code (Katzmarzik, 2011,
p.29) (see Figure 2 ‘Simulation’).
3.4 Regression setup
To investigate our research hypotheses and the influences of given project characteristics on the
optimal project structure, we regress typical project key figures on the optimal project granularity of
each release version (see Equation 1). As the time and the amount of version releases is not
synchronized between the 27 projects, we build a cross-sectional dataset containing the interversion
differences in optimal granularity (OptGranu) and the risk costs at optimal granularity (RiskCosts).
Additionally typical project characteristics like size units (Katzmarzik, 2011, p.8) and time measures
(Pfeifer, 2001) are added. These are number of patches (Patches), average file changes per patch
(AvgPFiles), number of files (Files), total lines of code (LOC) and duration of the version release
(Time) are added. To harmonize sizes of increasing and decreasing changes, all differences are
logarithmized. Moreover, we added control variables (Controls) for each project to identify project
specific influences.
The sample is tested for a typical structural bias. Considering the assumably close dependency of the
exogenous variables, no multicollinearity is observed. Since heteroscedasticity of the residuals
occurred in the original model, robust standard errors are applied.
4 Results
4.1 Descriptive results
Due to easy accessibility and publishability our data sample consists of top ranked Github open source
projects. However, this setup can also be easily applied for commercial projects like SOA, SaaS or
proprietary software. Github ranks the interest in projects by counting the number of observers
following the project and the number of projects that are build upon this project. We selected 27
projects and ensured that the sample consists of equally distributed programming paradigm styles
(scripting, imperative/functional and object-oriented), execution styles (native, byte code, interpreted)
and software category (webserver/ -framework, database/ libraries, system/ languages). The
occurrences of each language, the type and paradigms of projects within this sample can be seen in
Table 1:
Exec native bytecode interpet Lang c c++ java objc erlang python ruby js
27 9 9 9 27 6 3 3 3 3 2 2 5
Categ web db/lib sys/lan Type sys lang db lib webserver webframe
27 9 9 9 27 6 3 7 2 2 7
Style oo script im/fun Parad oo script imp func
27 9 9 9 27 9 9 6 3
Table 1. Descriptive statistics of the selected sample of projects
Overall, we extracted 1625 observations (version tags). Some projects like ‘postgres’ have a long
version history, others like ‘nu’ displayed a history of continuous integration into a single master
version. As we want to observe the influence of project characteristics on the optimal project
granularity, we accept also these extreme project settings. Project-specific influences are tested with
the aid of control variables in the final regression. Descriptive figures are listed in Table 2 and 3.
(1)

Name
CassandraClojure
Couch
DB
Django
Ejabberd
Git
Gitx
Httpd
jQuery
Memcached
Mongo
DB
Node
JS
Language java java erl python erl c objc c js c c++ c++ c erl
Type db lang db webfr sys sys sys webse webfr db db sys lang lib
Paradigm oo oo func script func imp oo imp script imp oo oo imp func
Versions 54 3 5 23 28 296 11 174 63 30 104 105 1 6
Fold. Avg. 204 39 70 3152 36 178 87 187 32 23 83 359 103 2109
Files Avg. 870 250 512 4087 336 2242 330 2905 155 135 1703 4968 379 9754
Table 2. Overview of project characteristics (part1)
Due to page limitations and a better understanding of the calculation results, we highlight the findings
for the popular web framework (Ruby on) Rails as an example. It shows average figures in its project
statistics and the calculation yields similar patterns and results for the rest of 26 projects. In the final
regression all 1625 observations are taken as single events. However, for clarity we aggregate all the
change data of the 116 versions in Rails and calculated the all-time VaR for each granularity level.
Name Phone-
gap
Postgres
Prototype
Rails
Redis
Ruby
Scriptaculous
Solr
Sparkle
Three
20
Tornado
Language js c js ruby c ruby js java objc objc python c++ js
Type webfr db webfr webfr db lang webfr db sys lib webse sys webfr
Paradigm script imp script script imp script script oo oo oo script oo script
Versions 5 260 11 116 61 73 3 6 4 6 9 157 22
Folders Avg. 76 380 47 777 45 459 25 1392 49 205 57 56 1587
Files Avg. 113 3826 130 2055 275 3460 111 4647 297 1159 185 1646 4358
Table 3. Overview of project characteristics (part2)
Figure 3 illustrates that starting with a normalized VaR of 1, the VaR of all three methods declines
with an increase in granularity levels. As expected, the control methods (rand and corr) calculate
structures with higher VaRs. The tar method shows a similar slope to the ‘Rule of Ten’, but at the
granularity level of 5 the VaRs stabilize at a stationary level. All three methods indicate the optimal
VaR at a granularity level of 7. Even though the ‘Rule of Ten’ does not depend on the project
characteristics of Rails, it approximates the tar graph to the granularity of 5 in the stationary region.
Consequently, the control methods cannot predict correct risk cost but they estimate the same optimal
granularity level as the tar method. The ‘Rule of Ten’ cannot predict the level of granularity correctly,
yet it indicates a good benchmark for the risk costs predicted by the VaR.
4.2 Rule of Ten as cost measure for incorrect granularity
In general, the slope of the graphs shows that the theoretical ‘Rule of Ten’ underestimates the risk
costs, with increasing levels of granularity. Additionally, it can be easily observed that the ‘Rule of
Ten’ approximates the risk cost value of the tar method best towards the optimum. Simultaneously, the
project shows that this rule of thumb noticeably underestimates the risk costs between the levels of the
worst and optimal granularity. This is especially the case when multiple levels of granularity have
measurable similar risk costs. Moreover, it can be argued that the validation methods simulate more
aggressively, coarsely and result in worse diversified portfolios than the ‘Rule of Ten’ and the tar
method. It is in line with the results of other empirical studies that even the randomized validation
methods confirm the minimum risk costs found by the tar method (Huschens and Stahl, 2004).
Comparing with the ‘Rule of Ten’, it can be argued that the tar method estimated the risk costs best,
Nu
V8
OTP
YUI3

ut also the validation methods determine the same level of granularity as optimal, when the absolute
risk cost values are negligible.
Figure 3. Analysis of Rails project results
4.3 Marginal utility of granularity
The data gives insights into the risk cost decline with an increase in granularity and subsequent
stabilization at a stationary level. It can be intuitively understood that projects are structured with best
practice methodologies and iterative approaches (Erl, 2005) until a manageable software foundation is
formed. Coevally, it can be seen at the stationary levels of the project that more or less adequate level
of granularity can already be discovered at lower levels of granularity. The marginal utility of
additional granularity towards the optimum is declining due to complexity of coordination between the
components. This can be explained by the effect of overspecification which is mentioned in literature
(Millard et al., 2009). If software structures are initially designed for good reusability, but then never
reused, then they might result in an inefficient overspecification and overhead. The measurement of all
projects shows, that a slight over- and underspecification can be tolerable. On the one hand the risk
cost increase slightly around the area of optimal granularity, while this increase is much smaller then
at lower or even monolithical granularity.
4.4 Influence of project characteristics on optimal granularity
Figure 3 illustrates that the implicit ‘Rule of Ten’ assumption of a log-log dependency of risk costs
and granularity is too unprecise. That is why we regress all 1625 version observations of the 27
projects as a cross-sectional dataset of first differences of log and investigate influences of additional
project characteristics. Results of the regression analysis explain changes of risk costs at optimal
granularity. The number of patches, files per patch and total number of files have a highly significant
influence on the change of optimal granularity. Even the influence of LOC changes is significant at
least at a 5% level and time at a 10% level of significance. While risk costs, patches, files and LOC
changes have a positive impact on the changes in level of optimal granularity, changes of files per
patch and time have a negative influence. As all project-specific control variables do not show any
significance, they are removed from the regression equation and the final results (see Table 4).
All measured project characteristics are significant and allow us to reject null hypothesis H20 that there
are no influences by the project characteristics. As the impact of chances in risk costs shows the
highest t-value null hypothesis H10 is rejected as well. The whole regression model has an adjusted Rsquared
value of 54.05% for the analyzed cross-sectional dataset. Appling an F-test for the significant
influence of all coefficients of the model yields an F-Value of 273.62, which is highly significant for
1625 observations and enables us to reject also null hypothesis H30. From a project management point
of view, this gives evidence that increases in project parameters like project risk costs, number of
patches, files and total lines of code cause a need of more granular structuring. Risk cost influence has
the highest significance of the measured project characteristics. If the number of files per patch
increases in a current version, it indicates too granular project structures and results in a need for a less

granular structuring. A shortage of release cycles and version snapshots shows similar results: if the
time between two releases decreases, the need for more granular and more diversified project
structures increases.
Const Risk Costs Patches Files/Patch Files LOC Time
Coefficient -0.0118 0.0229 0.0885 -0.3272 1.0451 0.1976 -6.5603
Std. Error 0.0135 0.0011 0.0083 0.0397 0.3231 0.0775 3.9047
t-Value -0.8728 20.8817 10.6532 -8.2349 3.2348 2.5486 -1.6801
P-Value 0.3829

from project source code data. The deducted model has an explanatory power of 54%, which means
46% of the observed influences are caused by influence that cannot identified with the facts from data
set. The data set contains only open source projects. The measurement methods are just able to
simulate the granularity between 1 and the given project granularity. Even with the deducted model an
initial project version snapshot is needed to determine a good level of granularisation. In further
research, we will address these limitations and deepen our analysis of the relation between risk costs
and optimal granularity as these have the highest significance in our regression model. Depending on
data accessibility and publishability we plan to shift our data scope from open source to SOA, SaaS
and proprietary software projects to support structuring decisions in commercial software projects. We
also want to investigate the dynamics of granularity over time to understand how previous versions
influence the current optimal granularity.
Finally, concerning the research questions, it is has been shown that the optimal granularity structure
of software projects, can be measured at present. A method has been presented that enables us to
measure and compare the quality of software project granularities and the influence on the optimal
level of granularity. Consequently, it becomes possible to estimate quantitatively how granular
software projects should be structured and how this granularity must be adjusted when specific project
parameters change. We thankfully acknowledge the support of Martin Haferkorn and the E-Finance
Lab, Frankfurt for this work.
Literature
Albrecht, P. and Maurer, R. (2005). Investment- und Risikomanagement: Modelle, Methoden,
Anwendungen. Schäffer-Poeschel, 2. Edition.
Amihud, Y. and Lev, B. (1981). Risk reduction as managerial motive for conglomerate mergers. Bell
Journal of Economics (12), pp. 605-617.
Benaroch, M.(2002). Managing information technology investment risk: A real options perspective.
Journal of management information systems, (19:2), pp. 43–84.
Benaroch, M., Lichtenstein, Y. and Robinson, K. (2006). Real options in information technology risk
management: an empirical validation of risk-option relationships. MISQ, (30:4), pp. 827 – 864.
Blobel, V. and Lohrmann, E. (1998). Statistische und num. Methoden der Datenanalyse; Teubner.
Bradley, M., Desai, A. and Kim, E.H. (1983). The rationale behind interfirm tender offers: information
or synergy. Journal of Financial Economics (11), pp. 182-206.
Braunwarth, K. and Friedl, B. (2010). Towards a financially optimal design of IT services. ICIS 2010
Proceedings, Saint Louis.
Campbell, R., Huisman, R. and Koedijik, K. (2001). Optimal portfolio selection in a Value-at-Risk
framework. Journal of Banking& Finance 25.
Christensen, H.A. and Montgomery, C.A. (1981). Corporate economic performance: Diversification
strategy versus market structure. Strategic Management Journal (2:4), pp. 327-343.
Cho, W. and Shaw, M. J. (2009). Does IT Synergy Matter in IT Portfolio Selection? ICIS 2009
Proceedings. Paper 160.
Eckbo, B. E. (1983). Horizontal mergers, collusion, and stockholder wealth. Journal of Financial
Economics (11), pp. 241-274.
Erl,T. (2005). Service-Oriented Architecture: Concepts, Technology & Design, Prentice Hall.
Erradi, A., Anand, S. and Kulkarni, N. (2006). SOAF: An Architectural Framework for Service
Definition and Realization.
Farjoun, M. (1998). The independent and joint effects of the skill and physical bases of relatedness.
Strategic Management Journal (19:7), pp. 611-630.
FSF (2006). Free Software Foundation - TAR. Accessed August 6 th 2011,
http://www.gnu.org/software/tar/tar.html.
Friedl, B. (2011). Zur optimalen Granularität von IT-Services - Eine Analyse relevanter ökonomischer
Einflussfaktoren, 10. Int. Tagung Wirtschaftsinformatik, Zürich.
Github (2011). Github Social Coding. Accessed November 9 th 2011, https://github.com/repositories.

Grigore, M. and Rosenkranz, C. (2011). Increasing the Willingness to Collaborate Online: Analysis of
Sentiment-Driven Interactions in Peer Content Production. ICIS 2011 Proceedings.
Hill, C.W.L. and Hoskisson, R.E. (1987). Strategy and structure in the multiproduct firm. Academy of
Management Review, (12:2), pp. 331-341.
Huschens, S. and Stahl, G. (2004). Granularität dominiert Korrelation, Risknews 06/04.
Katzmarzik (2011). Product differentiation for SaaS-Providers, BISE, 01/2011, pp. 19-31.
Klose, K., Knackstedt, R. and Beverungen, D. (2007). Identification of Services - A Stakeholder-based
approach to SOA development and its application in the area of production planning, ECIS.
Krafzig, D., Banke, K. and Slama, D. (2004). Enterprise SOA: Service Oriented Architecture Best
Practices, Prentice Hall International.
Lee, J., Seo, Y. and Lee, J. (2007). Method and appartus for merging data objects, European Patent
EP181840A2.
Lichter, H. and Mandl-Striegnitz, P. (1999). Defizite im Software- Projektmanagement - Erfahrungen
aus einer industriellen Studie, Informatique 5.
Lütkebohmert, E. and Gordy, M. (2007). Granularity adjustment for Basel II. No 2007,01, Discussion
Paper Series 2: Banking and Financial Studies, Deutsche Bundesbank, Research Centre.
Markowitz, H. (1952). Portfolio Selection: Journal of Finance, p. 77-91.
Millard, D. E., Howard, Y., Abbas, N., Davis, H. C., Gilbert, L., Wills, G. B. and Walters, R. J.
(2009). Pragmatic web service design: An agile approach with the service responsibility and
interaction design method. Computer Science.
Milgrom P. and Roberts, J. (1995). Complementarities and fit Strategy, structure, and organizational
change in manufacturing. Journal of Accounting and Economics (19), pp. 179 -208.
Myers, S.C. (1984). Finance theory and financial strategy. Interfaces (4), pp. 26-37.
Pfeifer, T. (2001). Qualitätsmanagement: Strategien, Methoden, Techniken, Hanser Fachb., 3. Aufl.
Rockafellar, R. and Uryasev, S. (1999). Conditional value-at-risk for general loss distributions. Journal
of Banking & Finance, Volume 26, Issue 7, Pages 1443-1471
Schirmer, I. and Zimmermann, K. (2008). Visualisierung von Projektportfolios zur Unterstützung des
Architekturmanagements, Universität Hamburg.
Standish (2011). The Standish Group CHAOS Manifesto 2011. Accessed June 9, 2011,
http://standishgroup.com/newsroom/chaos_manifesto_2011.php.
Panzar, J. C. and Willig, R. D. (1981). Economies of scope. American Eco. Rev., (71:2), pp. 268-272.
Tanriverdi, H. and Ruefli, T. W. (2004). The Role of Information Technology in Risk/Return
Relations of Firms. Journal of the Association for IS, Vol. 5, No. 11-12, pp. 421-447.
Tanriverdi, H. (2005). Information Technology Relatedness Knowledge Management Capability, and
Performance of Multibusiness Firms. MIS Quarterly (29:2), pp. 311-334.
Tanriverdi, H. (2006) Performance Effects of Information Technology Synergies in Multibusiness
Firms. MIS Quarterly (30:1), pp. 57-77.
Teece, D.J. (1980). Economies of scope & the scope of the enterprise. Journal of Economic Behavior
and Organization (1:3), pp. 223-247.
Teece, D. J. (1982). Towards an economic theory of the multiproduct firm. Journal of Economic
Behavior and Organization (3), pp. 39-63.
Turek, P., Wierzbicki, A. and Nielek, R. (2010). WikiTeams: Evaluating Teamwork in Wikipedia. In:
Proceedings of the WebSci10.
Weill, P. and Vitale, M. (2002). What IT infrastructure capabilities are needed to implement ebusiness
models? MIS Quarterly, (1:1), pp.17-34.
Wiedemann, A. (2002). Messung von Zinsrisiken mit dem Value-at-Risk-Konzept, WISU 12, pp.
1548-1553.
Wilkens, M. (2001). Basel II- Berücksichtigung von Diversifikationseffekten im Kreditportfolio durch
das Granularity Adjustment, Zeitschrift für das gesamte Kreditwesen, 12; pp. 670-676.
Willig, R. (1978). Technology and market structure. American Economic Review (69), pp. 346-351.
Wolke, T. (2008). Risikomanagement; Oldenbourg Wissenschaftsverlag.
Wierzbicki, A., Turek, P. and Nielek, R. (2010). Learning about team collaboration from Wikipedia
edit history. International Symposium on Wikis and Open Collaboration.

Layer 3:
Customer Management in E-Finance
(Prof. Dr. Andreas Hackethal, Prof. Dr. Bernd Skiera,
Prof. Dr. Oliver Hinz)
� Bhattacharya U., Andreas H., Simon K., Benjamin L., Steffen M. (2012):
Is unbiased financial advice to retail investors sufficient? Answers from a large
field study
In: Review of Financial Studies, Vol. 25, 975-1032.
� Hackethal A., Michael H., Tullio J. (2012):
Financial advisors: A case of babysitters?
In: Journal of Banking & Finance, Vol. 36, 509-524.
� Schmitt P., Skiera B., Van den Bulte C. (2011):
Referral Programs and Customer Value
In: Journal of Marketing, Vol. 75, Jan., 46-59.
� Hinz O., Skiera B., Barrot C., Becker J. (2011):
An Empirical Comparison of Seeding Strategies for Viral Marketing
In: Journal of Marketing, Vol. 75, Nov., 55-71.
� Skiera B., Bermes M., Horn L. (2011):
Customer Equity Sustainability Ratio: A New Metric for Assessing a Firm’s
Future Orientation
In: Journal of Marketing, Vol. 75, May, 118-131.

Is Unbiased Financial Advice to Retail
Investors Sufficient? Answers from a Large
Field Study
Utpal Bhattacharya
Indiana University
Andreas Hackethal
Goethe University, Frankfurt
Simon Kaesler
Goethe University, Frankfurt
Benjamin Loos
Goethe University, Frankfurt
Steffen Meyer
Goethe University, Frankfurt
Working with one of the largest brokerages in Germany, we record what happens when
unbiased investment advice is offered to a random set of approximately 8,000 active retail
customers out of the brokerage’s several hundred thousand retail customers. We find that
investors who most need the financial advice are least likely to obtain it. The investors who
do obtain the advice (about 5%), however, hardly follow the advice and do not improve
their portfolio efficiency by much. Overall, our results imply that the mere availability of
unbiased financial advice is a necessary but not sufficient condition for benefiting retail
investors. (JEL D14, G11, G24, G28)
Hwa is thet mei thet hors wettrien the him self nule drinken? [Who
can give water to the horse that will not drink of its own accord?]
—Oldest English proverb, first recorded in 1175, compiled by
Heywood (1546)
This research would not have been possible without the collaboration of a large German bank. We thank this
bank and all its employees who helped us. For their comments, we also thank John Campbell, Daniel Dorn, Eitan
Goldman, Arvid Hoffmann, Craig Holden, Gur Huberman, Roman Inderst, Narasimhan Jegadeesh, Artashes
Karapetyan, Min Kim, Adair Morse, Antoinette Schoar, Jay Shanken, Andrei Shleifer, Paolo Sodini, Meir Statman,
Noah Stoffman, and Brian Wolfe, as well as the conference and seminar participants at Emory, Frankfurt,
Harvard Business School, Henley, National Chengchi University, National Sun Yat-sen University, National
Taiwan University, Wisconsin, the EFA conference in Stockholm, the European Retail Investment Conference in
Stuttgart, and the FIRS conference in Sydney for their comments. Send correspondence to Utpal Bhattacharya,
Kelley School of Business, Indiana University, Bloomington, IN 47405-1701; telephone: (812) 855-3413; fax:
(812) 855-5875. E-mail: ubhattac@indiana.edu.
c○ The Author 2012. Published by Oxford University Press on behalf of The Society for Financial Studies.
All rights reserved. For Permissions, please e-mail: journals.permissions@oup.com.
doi:10.1093/rfs/hhr127 Advance Access publication January 12, 2012
Downloaded from
http://rfs.oxfordjournals.org/ at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

The Review of Financial Studies / v 25 n 4 2012
There is a large and growing body of literature in household finance documenting
how retail investors make serious investment mistakes by deviating
from the prescriptions of normative finance. 1,2 As Campbell (2006) notes,
a number of potential remedies have sprung up over the years to resolve
these investment mistakes, which include financial education, default options,
and regulation by governments. An important remedy is financial advice
provided by professionals. Providing financial advice is big business. In the
United States, the financial planning and advice industry was estimated at
US$39 billion in 2010. 3 The Investment Company Institute (2007) notes
that over 80% of respondents state that they obtain financial advice from
professional advisors or other sources in the United States. In Germany, a
survey among retail investors indicates that more than 80% of investors consult
a financial advisor. 4
However, financial advice is offered by third parties and may be biased. The
theoretical literature notes that the information asymmetry between investor
and advisor may provide camouflage to advisors who act in their own interests
and to the detriment of their clients. 5 Recent empirical evidence supports the
agency conflict claims formalized in the above theoretical models. 6 A survey
1 This literature is too vast to cover here. We thus offer a sample of important findings. The majority of households
do not even participate in the stock market (Guiso, Haliassos, and Jappelli 2002), despite the large equity
premium that exists (Mehra and Prescott 1985; Dimson, Marsh, and Staunton 2007). The few households
that do participate in equity markets hold underdiversified portfolios (Blume and Friend 1975; Kelly 1995;
Goetzmann and Kumar 2008). Underdiversification, with regard to geographical diversification, is particularly
acute—investors are found to exhibit both a home bias and a preference for local stocks (French and Poterba
1991; Lewis 1999; Cooper and Kaplanis 1994; Huberman 2001; Zhu 2002; Ahearne, Griever, and Warnock
2004; Calvet, Campbell, and Sodini 2007). Research observes inertia that results in individual investors making
insufficient portfolio adjustments in response to general market movements (Agnew, Balduzzi, and Sundén
2003; Calvet, Campbell, and Sodini 2009a; Madrian and Shea 2001). Investors trade too much because they
are overconfident (Odean 1999; Barber and Odean 2000; Deaves, Lüders, and Luo 2003). Investors tend to sell
winners too early and hold on to losers too long, which is an investment mistake called the disposition effect
(Shefrin and Statman 1985; Odean 1998; Frazzini 2006). Investors are fixated on cognitive reference points
(Bhattacharya, Holden, and Jacobsen, forthcoming).
2 Barber and Odean (2000) find that overconfidence leads to substantial return decreases after the deduction of
transaction costs. Looking at the aggregate portfolio of individual Taiwanese investors, Barber et al. (2009)
document an annualized loss of 3.8%. Calvet, Campbell, and Sodini (2007) measure the cost of nonparticipation
and underdiversification and report a substantial loss compared with the world market for Swedish households.
Malkiel (1995) and Carhart (1997) show that even mutual funds underperform the market, net of expenses.
Barras, Scaillet, and Wermers (2010) and Fama and French (2010) find no evidence that the average mutual
fund produces alphas that cover its costs.
3 http://www.ibisworld.com/industry/default.aspx?indid=1316.
4 Two-thirds say that they obtain financial advice from their main bank, whereas about a fifth obtain advice from
an independent financial advisor (DABbank 2004).
5 See, e.g., Bolton, Freixas, and Shapiro (2007), Carlin (2009), Inderst and Ottaviani (2009), Carlin and Gervais
(2009), Stoughton, Wu, and Zechner (2011), and Carlin and Manso (2011).
6 Mutual funds in the United States that are sold through brokered channels underperform. Those funds that
provide higher fees are sold more heavily, which, in turn, negatively affects returns (Bergstresser, Chalmers,
and Tufano 2009; Chen, Kubik, and Hong 2011; Edelen, Evans, and Kadlec 2008). Hackethal, Haliassos, and
Jappelli (2012) find that individual investors whose accounts are run by, or in consultation with, a financial
advisor achieve lower returns. Additionally, in an audit study, Mullainathan, Nöth, and Schoar (2009) conclude
that financial advisors seem to be aggravating the existing biases of their investors. In addition, recommendations
made by research analysts who are compromised through an incentive scheme are shown to have only very
limited potential for value enhancement (Womack 1996; Metrick 1999; Barber and Odean 2001).
976
Downloaded from
http://rfs.oxfordjournals.org/ at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

Is Unbiased Advice Sufficient?
of European Union (EU) members by the CFA Institute (2009) revealed that
64% of respondents believe that the prevailing fee structure is intended to steer
sales instead of serve the customer.
If retail investors make serious investment mistakes, and if investment
advice, which is provided to investors by professional advisors as a means
to ameliorate these mistakes, is conflicted and biased, it would seem that
retail investors would benefit if they could instead receive unbiased and
theoretically sound professional advice. This supply-side cure to improve
the portfolio efficiency of retail investors should work. Regulators certainly
think so. The plethora of financial market reforms that are instituted in major
financial markets—like the Restoring American Financial Stability Act of
2010 (also known as the Dodd–Frank Act) in the United States or the Markets
in Financial Instruments Directive in the EU, which are devised mainly to clean
up the conflicts of interest and improve disclosures in the financial sector—are
essentially supply-side remedies.
This article attempts to answer an important public policy question: will
investors really benefit from only supply-side remedies? To be specific, can
unbiased financial advice steer retail investors toward efficient portfolios?
In this study, we worked with one of the largest brokerages in Germany.
For the first time, the brokerage offered financial advice to a random set
of approximately 8,000 active retail customers out of their several hundred
thousand retail customers. The advice was free and was aimed at improving
portfolio efficiency. The advice was unbiased—i.e., it was free of monetary
incentives for the brokerage and was generated by an algorithm that was
designed to improve portfolio efficiency—and, as we will later document,
the advice does significantly improve portfolio efficiency ex post. We have
demographic data on all retail customers who obtain the advice and who
do not obtain the advice. For both groups, we also possess customers’ daily
trading records for a number of years before the advice is offered and up to
ten months after the advice is offered. We can answer some key questions
because of this, including which types of customers accept the offer, whether
the advice received is followed, whether portfolio efficiency improves for the
average advisee who follows the advice or does not follow the advice, and if
investors who most need the financial advice are least likely to obtain it.
In trying to answer the above questions, this article explores the demand
side of financial advice. To the best of our knowledge, this is the first study to
link the recommendations of advisors with actual customer behavior after the
advice is given.
We arrive at several interesting findings. First, those who accept the offer
(5%) are more likely to be male, older, richer, more financially sophisticated,
and have a longer relationship with the brokerage. Second, of those who accept
the offer, the advice is hardly followed. Third, though portfolio efficiency
hardly improves for the average advisee, it does improve for the average
advisee who follows the advice. Fourth, it seems that investors who most need
977
Downloaded from
http://rfs.oxfordjournals.org/ at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

The Review of Financial Studies / v 25 n 4 2012
the financial advice are least likely to obtain it. Overall, our results imply that
the mere availability of unbiased and theoretically sound financial advice is a
necessary but not sufficient condition for benefiting retail customers. As the
adage goes, you can lead a horse to water, but you can’t make it drink. 7
The construction of our study allows for a greater understanding of the
factors that contribute to a person opting to obtain and then to follow financial
advice. A probit test gives us a good handle to understand why advice is
not being sought. Such reasons include 1) lack of financial sophistication, as
measured by poor past portfolio performance (has many interpretations); 2) a
desire to not increase tax payments; and 3) lack of familiarity and/or trust, as
measured by the length of relationship with the brokerage. It does seem from
the first result that the clients who most need (least need) the financial advice
are the ones who are least likely to obtain it (most likely to obtain it). This
conclusion is later buttressed by a more formal test.
The results of a regression about why clients do not follow the advice once
they get it is not so extensive. Wealthier investors and those with lower-risk
portfolio values tend to follow the advice more often. Though we have few
positive results, we can rule out many obvious suspects, like trust or financial
sophistication or clients not following advice because the advice asked for a
dramatic increase in investments. We suspect the lack of results, with respect to
the other variables in this test, is due to a lack of power; there is little variation
in the dependent variable: most clients who opt for advice do not follow the
advice. The other reason could be that a systematic cause may not exist to
explain why people do not follow advice. 8
To summarize, the contribution our article makes is to highlight the centrality
of the demand-side problem—unbiased financial advice is useless unless
it is followed—and to recognize the limitations of regulations in dealing with
this demand-side problem. How can regulation convince a person to follow
unbiased and sound financial advice, when we do not completely understand
why people follow advice? Though the answer to this question is beyond the
7 Campbell (2006), and especially Campbell et al. (2011), recognize that consumers need financial protection
because of supply-side problems and also demand-side problems. Using the latest research from behavioral
economics, Campbell et al. (2011) make a very powerful argument that many consumers do not have the ability
to understand complex financial products in an age in which they have to make most of their own financial
decisions. Benartzi and Thaler (2004) document demand-side problems in the context of savings decisions. They
offer possible solutions by taking into account behavioral factors when designing savings plans. They develop a
choice architecture system called “Save More Tomorrow” (SMarT), which is designed to help people commit in
advance to defined contribution increases in pension plans. In an experiment on payday loan borrowers, Bertrand
and Morse (2009) examine the cognitive limitations of these borrowers and focus on solutions. They find that
information disclosure that aims at “de-biasing” is effective.
8 A multitude of reasons are posited in the literature that may explain why people may not follow advice.
These incude bounded rationality (Kahneman 2003) and procrastination that leads to inertia (Samuelson and
Zeckhauser 1988; Laibson 1997; O’Donoghue and Rabin 1999). Other studies, such as Barber and Odean
(2000), find that investors tend to be overconfident, and overconfident individuals tend not to follow advice.
Other influences on the propensity to opt for and follow financial advice may be social interaction (Hong,
Kubik, and Stein 2005) and financial literacy (Christelis, Jappelli, and Padula 2010). More articles are discussed
later.
978
Downloaded from
http://rfs.oxfordjournals.org/ at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

Is Unbiased Advice Sufficient?
scope of this article, our article does cast some doubt on some of the reasons
trotted out to answer this question: noncomprehension of the financial advice,
mistrust of advisor, or simply inertia. 9
The analysis is organized as follows. Section 1 describes our field study,
including details of the offer made to the retail customers by the brokerage
and the particulars of the advice. We explain how the recommendations are
generated and argue that the advice is unbiased and theoretically sound.
Section 2 details the raw data and describes the methodology used to estimate
portfolio efficiency and the degree to which investors follow the advice.
Section 3 gives some revealing, descriptive statistics. Section 4 examines
which retail customers are most likely to choose advice. Section 5 examines
who follows the advice. Section 6 explores the portfolio efficiency of the
customers after receiving the advice. Section 7 analyzes which customers
would most benefit from the advice. Section 8 concludes.
1. Field Study
1.1 Overview
The brokerage we work with was originally founded as a direct bank. Its
focus was on offering brokerage services via telephone and the Internet,
and, over time, it evolved into a full service bank, providing clients with
brokerage and banking services as well as advice on mortgages. However,
the bank never offered investment advice to their clients; all client trades
were self-directed. This changed in 2009. In order to retain existing customers
and attract new ones, the brokerage set out to introduce a financial advisory
business. As a new entrant to the investment advisory market, the brokerage
designed a financial advice model that was distinctly different from those
offered by traditional retail banks. First, the financial advice offered would
not be conflicted; i.e., recommendations would be independent of product
issuers. Second, the financial advice would not be discretionary advice from an
individual advisor but rather recommendations produced by an optimizer that
improves portfolio efficiency. The optimizer would use primarily exchangetraded
funds (ETFs) and mutual funds to increase diversification within and
across asset classes, both domestic and foreign. Third, in order to ensure
and signal the objectivity of its financial advice, the bank would avoid any
9 The same issues are faced in medicine as well. Research on patients’ adherence to medical advice has
been conducted for decades. A meta-analysis of fifty years of research in this field finds that “the average
nonadherence rate is 24.8%.” The reasons for not following a doctor’s advice are that patients think they know
more than the doctor, depression, lack of social support, or simply because they misunderstand or forget what
they have been told. Adherence increases with more circumscribed regimens, as well as education and income,
but not as a function of demographic characteristics (such as gender or age) or the severity of illnesses (DiMatteo
2004). Another finding in the medical literature, which parallels our finding, is that people who most (least) need
to go to their doctor go less (more) often.
979
Downloaded from
http://rfs.oxfordjournals.org/ at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

The Review of Financial Studies / v 25 n 4 2012
incentive problems by not charging commission on trades that were based on
the recommendations offered. Fourth, during a test phase from which our data
originate, the advice was free of charge.
1.2 Details of the offer
About 8,000 active customers were randomly selected from the brokerage’s
several hundred thousand customers. Active customers are defined as customers
with an account volume of at least e25,000, who have made at least
three trades over the past twelve months and are between eighteen and eighty
years of age. A total of 8,195 customers were offered the advice. In early May
2009, an e-mail was sent to the selected customers’ banking account inboxes
(not private e-mail). In this e-mail, the new advisory service was advertised to
be objective. It was mentioned that 1) the recommendations would be systemgenerated
and independent of product issuers; 2) no commissions, overt or
covert, would be charged for trades that were based on the recommendations;
and 3) the advice would be free during the test phase. Customers were told
that at the end of the test phase, the free advisory service would be terminated
automatically. It was also made clear to the customers that there would be no
obligation to make any transactions that were based on the recommendations
given. Thus, there would be no risk of unintended future commitment for
the customer. If customers did not respond to the offer, a follow-up phone
call was initiated in which an advisor explained the offer again and answered
questions.
1.3 Details of the advice
All customers who opt to receive the free advice from the brokerage are
assigned to our treatment group. All customers who decline the offer form our
control group. Note that this is not a random assignment. However, our basic
empirical methodology—difference in difference—somewhat ameliorates this
shortcoming of our field study.
Every person in the treatment group was contacted by an advisor to
schedule an initial call. This call was used to gather additional demographic
information (e.g., job and household size) and wealth proxies (income and
total financial wealth, including cash and other assets). Risk preferences were
mainly solicited by asking advisees to select between six categories that ranged
from “safe” to “opportunity” as their investment philosophy. There were no
replies for the two most risk-averse categories; we are therefore left with
four levels of risk aversion. Based on demographic data and customer inputs,
the brokerage calculated a risk capacity score that determined the maximum
possible level of risk a client should be exposed to in the recommended
portfolio.
This risk capacity score is the main input for forming customer-specific
recommendations that enhance portfolio efficiency. The customer received
980
Downloaded from
http://rfs.oxfordjournals.org/
at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

Is Unbiased Advice Sufficient?
detailed documentation (see Appendix A for a disguised example of a detailed
recommendation) that included the following information:
• Description of the idea of diversification by investing in different asset
classes and markets
• Explanation of important concepts (e.g., volatility, mean-variance efficiency,
and the Sharpe ratio)
• Analysis of the existing portfolio (historical and expected risk and return
profile)
• Analysis of the recommended portfolio (list of securities and risk and return
profile compared with the existing portfolio)
• Inclusion of client’s requests into recommendations (e.g., securities to
retain)
• Consideration of tax advantages by keeping old investments in the portfolio
• List of trades that are necessary to realize the recommended portfolio
• Fact sheet for each security on the recommended list
In addition to the detailed written documentation that was sent via e-mail, an
advisor also explained the recommendations to the customer over the phone.
1.4 Recommendations
The bank’s recommendations are generated by a mean-variance optimizer that,
based on the original framework of Markowitz (1952), focuses on portfolio
efficiency. The household finance literature indicates that retail investors make
mistakes by holding underdiversified portfolios. This underdiversification is
typically not linked to investment skill (Goetzmann and Kumar 2008). If a
portfolio optimizer improves diversification, it will consequently add value for
a typical retail investor.
It is relatively unimportant which optimization method is applied, as long
as it increases diversification. DeMiguel, Garlappi, and Uppal (2009) show
that more sophisticated techniques are not significantly better than a naïve 1/N
portfolio strategy. To build efficient portfolios, however, it is always desirable
that the expected return estimates are not biased by any past extreme return
realizations. Two precautionary measures are taken. First, the optimizer uses
a shrinkage factor, as proposed by Jorion (1986) and Michaud (1998). The
shrinkage factor is implemented by using a security’s deviation from the longrun
average return of securities with a comparable level of risk. Second, and
more importantly, the optimizer is set up in such a way that it selects from a set
of only eighty securities, predominantly ETFs and/or mutual funds. For such
highly diversified portfolios, Holden (2009) notes that the potential effect of
past idiosyncratic realizations is minimized. Volatilities are estimated by using
981
Downloaded from
http://rfs.oxfordjournals.org/ at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

The Review of Financial Studies / v 25 n 4 2012
past volatilities. The risk capacity score of the customer is the final input into
the optimizer.
Even though only eighty securities are considered in the basic investment
opportunity set of the optimizer, the potential opportunity set of the optimizer
is much larger because it is capable of considering almost all securities held
by the investors in our sample. Additionally, the optimization is subject to
some constraints and side conditions, such as the maximum weight on an
asset class due to a client’s wishes or risk capacity, maximum number of
securities, minimum weight on a single security, and short-sale constraints.
It is also possible to define other constraints, such as the number of securities
retained from the existing portfolio. As shown in Table A2 in Appendix B,
about 25% of the value of the original portfolio is retained on average. This is
done for two reasons: first, to retain some securities in which customers have
a tax advantage, and second, to increase the chance that the investors will act
on the recommendations. This implies that the recommended portfolios are
different across investors because investors differ in their risk capacity score,
prior portfolio allocations, side constraints, and the point in time at which they
receive their recommendations.
Despite these constraints, Table 1 (panels A–C) shows that the optimizer
is indeed able to improve the diversification of the portfolios along three
important dimensions. First, Table 1, panel A, shows that investments in single
stocks are reduced from 53% to 27%. Clients are advised to invest 67% in
well-diversified ETFs and mutual funds. Recall that the bank has no incentive
to push certain funds, since no commissions are charged. Second, Table 1,
panel B, shows that the average recommended portfolio is more diversified in
different asset classes than is the average existing portfolio before the advice.
The share of equity is reduced from 73% to 59%, while the share of fixed
income and real estate securities is increased from 7% to 23%. Likewise, the
average share of commodities increased from 1% to 13%. Third, Table 1,
panel C, shows that international diversification is strongly enhanced by the
recommendations. Prior to the advice, investors hold about 52% of their equity
in German securities; the recommendations suggest holding 30%.
Table 1, panel D, shows the size of the recommended portfolios compared
with their other portfolios for each advisee. As can be seen, the median
recommended portfolio has exactly the same size as the original portfolio. This
implies that most of our clients are advised not to increase or decrease their
investments in risky assets. This table also shows that the mean recommended
portfolio is 19% larger than the original portfolio. This is because, by default,
the optimizer matched the size of the recommended portfolio with the size of
the advisee’s original portfolio, but the advisee could request a larger (smaller)
size of the recommended portfolio if he or she wanted to invest (divest). Even
those investors who wanted to invest considerably more had sufficient financial
assets to cover this net investment, as shown by the ratio of the recommended
portfolio divided by total financial wealth (median 48%; even the ninety-fifth
982
Downloaded from
http://rfs.oxfordjournals.org/
at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

Is Unbiased Advice Sufficient?
Table 1
Advisees’ original portfolios versus recommended portfolios
Portfolio share
Instrument Original (%) Recommended (%)
Single stocks 53 27
Funds 26 57
ETFs 4 10
Single bonds 4 0
Others 13 6
Total (%) 100 100
Portfolio share
Asset class Original (%) Recommended (%)
Equity 73 59
Fixed income 6 15
Commodities 1 13
Real estate 1 8
Others 19 5
Total (%) 100 100
Equity share
Region Original (%) Recommended (%)
Germany 52 30
Europe 14 24
North America 12 10
Asia Pacific 9 18
World 8 9
Central and eastern Europe 2 7
Other emerging markets 3 2
Total (%) 100 100
Percentile
Standard
Ratio Mean Median 5th 25th 75th 95th deviation N
Recommended portfolio over 1.19 1.00 0.92 0.99 1.04 1.88 0.91 380
original portfolio
Recommended portfolio over 0.90 0.90 0.43 0.75 0.99 1.29 0.43 365
financial wealth with the brokerage
Recommended portfolio over 0.51 0.48 0.15 0.32 0.67 0.97 0.26 318
total financial wealth
Recommended portfolio over 0.28 0.21 0.05 0.12 0.40 0.77 0.23 318
total wealth
Panel A shows the portfolio share by instrument of the advisees’ original portfolio compared with the portfolio
recommended to the advisees at the time of the recommendation. Panel B shows the portfolio share by asset class
of the advisees’ original portfolio compared with the portfolio recommended to the advisees at the time of the
recommendation. Panel C provides a regional breakdown of the equity share for the advisees’ original portfolio
compared with the portfolio recommended to the advisees at the time of the recommendation. Panel D compares
the size of the advisees’ recommended portfolios with the size of other portfolios. The original portfolio of an
advisee is the size of the portfolio held at the time of the recommendation. Financial wealth with the brokerage
is the advisee’s original portfolio plus the value of cash at the brokerage. Total financial wealth is the advisee’s
original portfolio plus the value of cash at the brokerage plus an estimate of financial assets held outside the
brokerage (this estimate is provided by the client). Total wealth is the advisee’s original portfolio plus the value
of cash at the brokerage plus an estimate of financial and nonfinancial assets held outside the brokerage (this
estimate is provided by the client). All sizes are measured in euros at the time of the recommendation. Different
counts of observations are due to data availability of certain variables (five customers had no risky portfolios at
the time of the recommendations).
983
Downloaded from
http://rfs.oxfordjournals.org/ at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

The Review of Financial Studies / v 25 n 4 2012
Table 2
Average asset class shares of recommended portfolios by advisees’ risk aversion
Risk aversion
Asset class 1 = Highest 2 3 4 = Lowest
Equity 38 43 53 63
Fixed income 27 20 14 7
Money market 14 15 13 10
Commodities 5 9 10 13
Real estate 14 10 7 5
Other 2 4 3 3
Total 100 100 100 100
N 20 95 172 55
Table 2 shows the average asset class share within advisees’ recommended portfolios for each level of risk
aversion.
percentile of the distribution had enough financial assets to cover the required
net investment with a ratio of 97%) in Table 1, panel D. As our clients may
not follow advice because they are asked to increase their investments in
risky securities, we use the ratio of the recommended and original portfolio
as another independent variable in our test to discover why clients do or do not
choose to follow the advice they opted for. We will find later that this variable
has no significant effect on this decision.
Table 1, panel D, also shows that the median recommended portfolio is 90%
of the financial wealth held with the brokerage (the other 10% is cash), is 48%
of the total financial wealth, and is 21% of total wealth. These large numbers
imply that the risky portfolio investment held in this brokerage is not “play
money.”
Table 2 provides evidence that the optimizer also aligns each client’s risk
capacity with the riskiness of his or her recommended portfolio. From risk
class 1 (highly risk-averse) to risk class 4 (least risk-averse), the share of equity
and commodities monotonically increases, while the share of fixed income,
money market, and real estate investments declines.
Overall, given the constraints, it seems as if the optimizer manages to
recommend well-diversified portfolios to each investor.
1.5 Timeline
The offer was sent to a random sample of 8,195 customers who were drawn
from a population of several hundred thousand customers in early May 2009.
This is our “event date,” which we index as t = 0 on the event timeline.
We collect demographic data on our sample of 8,195 customers. It is an
unbalanced panel. Only 5,952 customers were with the brokerage in September
2005. The other 2,243 joined afterward but before May 2009. The period
between September 2005 and May 2009 is called the “pre-advice” period.
Thus, the pre-advice period is t = −44 months to t = 0. The 8,195 sample
984
Downloaded from
http://rfs.oxfordjournals.org/ at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

Is Unbiased Advice Sufficient?
Figure 1
Timeline
The sequence of events in the field study (dates are always at the beginning of the respective month).
customers could join the treatment group between May and October 2009. The
first free recommendation was given on May 13, 2009, and the last customer
opted in during the last week of September 2009. This is why we define the
“post-advice” period as after September 2009. The “post-advice” period lasts
six months, ending on March 31, 2010. Figure 1 illustrates the timeline.
2. Data and Methods
2.1 Data collected
The first part of the data set consists of demographics for the entire random
sample group of 8,195 customers. Table 3 shows the collected data.
These data include gender, age, and microgeographic status, as well as
time-invariant account information, such as the account opening date. The
microgeographic status measures the average wealth level of people living in
Table 3
Data collected
Type of data Type of data Dates available
Client demographics Gender Time-invariant
Date of birth (measure of age) Time-invariant
Microgeographic status (measure of wealth) Time-invariant
Total financial and nonfinancial wealth (advisees only) Time-invariant
Risk aversion Time-invariant
Portfolio characteristics Actual position statements Monthly
Actual transactions Daily
Recommended securities On day of recommendation
Recommended transactions On day of recommendation
Cash On day of recommendation
Number of trades Daily
Account opening date (measure of length of relationship) Time invariant
Market data Carhart (1997) four factors on broad domestic index
−Market factor Weekly
−Small minus big (SMB) Weekly
−High minus low (HML) Weekly
−Momentum factor Weekly
Individual security returns Daily
Table 3 summarizes the data collected during the course of the study. Client demographics and portfolio
characteristics have been provided by the brokerage. The record date is July 2010. Market data are taken from the
Thomson Reuters Datastream. The third column reports the availability of time-series data and their frequency.
985
Downloaded from
http://rfs.oxfordjournals.org/ at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

The Review of Financial Studies / v 25 n 4 2012
a given micro-area (on a street level). It has nine categories, with nine being
the wealthiest. This variable is provided by a specialized data service provider
that uses several factors (such as house type and size, dominant car brands,
rent per square meter, and unemployment rate) to construct this variable. For
our multivariate tests, we further group this variable into three categories:
low (1–3), medium (4–6), and high wealth (7–9). The account opening date
allows us to compute the length of the relationship between a customer and the
brokerage. The total financial and nonfinancial wealth (e.g., real estate) held
by the client outside the brokerage is an estimate given to us only by clients
who opted to take the advice.
The second part of the data set includes portfolio characteristics for each
customer. Portfolio characteristics include position statements, transactions,
and transfers of holdings from other portfolios. Position statements are on a
security-by-security level and are taken from the beginning of each month.
We obtain the International Security Identification Number (ISIN), the number
of securities held per ISIN, and the respective euro-value for each position.
Transactions and transfers are recorded on a daily basis. For each transaction,
we know the ISIN, trade volume, and transaction price. We use this to calculate
portfolio turnover, as in Barber and Odean (2001), as well as trades per
month. We also have information on the cash accounts of each customer at
t = 0, enabling us to calculate the risky share as the risky portfolio value
divided by financial wealth with the brokerage (risky portfolio value plus cash
value).
Furthermore, for the customers who opted to receive advice, we have data
on when a customer received advice and what exactly the customer was
recommended to buy and sell. This permits us to calculate four different daily
return series for each customer in the treatment group (customers who opted
to obtain advice): actual investment returns in the pre-advice period, actual investment
returns in the post-advice period, buy and hold investment returns in
the post-advice period if the portfolio had not changed from the day before the
advice was given, and investment returns on the recommended portfolio in the
post-advice period if advice had been followed exactly. The data also allow us
to calculate two different daily return series for each customer from the control
group (customers who opted not to obtain advice): actual investment returns
in the pre-advice period and actual investment returns in the post-advice
period.
The third part of the data set contains market data from Thomson Reuters
Datastream. Sample customers hold and trade a total of 46,361 securities
over the observation period. Thomson Reuters Datastream covers 97% of
these securities. A number of our tests use the Carhart (1997) model to
estimate risk-adjusted returns. Therefore, we also produce weekly return
series for the following four factors: the country market factor (MKT),
small minus big (SMB), high minus low (HML), and the momentum factor
(MOM).
986
Downloaded from
http://rfs.oxfordjournals.org/
at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

Is Unbiased Advice Sufficient?
2.2 Return calculations
The brokerage data, in conjunction with the market data from Thomson
Reuters Datastream, allow us to compute daily portfolio returns and daily
holdings on a security-by-security level.
To do this, we first infer daily holdings from monthly position statements,
security transactions, and account transfers. We have end-of-day holdings
for the last day in every month. To obtain the next end-of-day holdings, we
multiply the end-of-day value of each holding by the corresponding price
return (excluding dividends but taking into account any capital actions) for that
security. These holdings are then properly adjusted for any sales, purchases,
and account transfers that occurred on that same day. We repeat this procedure
for every security and investor for each trading day in a given month. The
holdings on the last day of each month are then reconciled with the true
holdings obtained from the brokerage. 10
Second, we compute daily portfolio returns as the weighted average of the
returns of all securities held, purchased, or sold by the investor on that day. We
use total return data (including dividends) for securities without transactions on
that day. For securities that are either purchased or sold, we take into account
exact transaction prices to compute returns. We weight each security’s return to
calculate investors’ daily portfolio returns. All holdings and sales are weighted
by using values in euros on the basis of the previous day’s closing prices. All
purchases are weighted by using the transaction value in euros.
To compute hypothetical returns for portfolios recommended by the broker,
we follow the same procedure and assume that purchase and sale transactions
occur at the price that prevails at the end of the day on which the recommendation
was given. 11
For our regressions, we cumulate all daily portfolio returns into weekly
returns. Portfolio excess returns are weekly portfolio returns minus the riskfree
rate, which we assume to be equal to the three-month EURIBOR. We
regress this excess return on the four factors used by Carhart (1997),
R j,w − R f,w = a j,w + β j,w ∙ � �
Rm,w − R f,w + s j,w ∙ SM Bw
+ h j,w ∙ H M Lw + m j,w ∙ M O Mw + ε j,w, (1)
where R j,w is the return on investor j’s portfolio in week w, R f w is the threemonth
EURIBOR rate in week w, Rm,w is the return in week w on a broad
domestic stock market index (MKT), SMBw and HMLw are the returns for the
size and value-growth portfolios, in accordance with Fama and French (1993),
10 The deviations between inferred and actual holdings are negligible.
11 We can compute net returns for actual transactions because we know all transaction costs. However, since we
cannot compute these net returns for the benchmark “buy and hold” portfolios or the “recommended” portfolios,
all our analysis is done with bid-ask adjusted gross returns (computed at bid and ask prices). Net returns for
actual portfolios are, on average, 1.5% lower per year than their corresponding gross returns.
987
Downloaded from
http://rfs.oxfordjournals.org/ at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

The Review of Financial Studies / v 25 n 4 2012
in week w, and MOMw is the one-year momentum return from Carhart (1997)
in week w.
The intercept (alpha) in regression Equation (1) is our measure for riskadjusted
portfolio return. For robustness, we also calculate Jensen’s one-factor
alphas for both the domestic stock market as well as the MSCI World in euros.
All results that we report later are robust to using these measures. Note that we
estimate alphas only in the pre-advice stage, where we have forty-four months
of data. We do not estimate alphas in the post-advice period, since we only have
six months of data. We do, however, estimate measures of portfolio efficiency,
as described in the next subsection.
2.3 Measures of portfolio efficiency
We first focus on measuring diversification in a portfolio. To do so, we
use two primary measures—the Herfindahl-Hirschman index (HHI) and the
idiosyncratic variance share—and a secondary measure—home bias. The
HHI is a commonly accepted and simple measure of diversification (Dorn,
Huberman, and Sengmueller 2008; Ivkovic, Sialm, and Weisbenner 2008).
It is calculated by summing the squared portfolio weights of all securities.
Therefore, it follows that the lower the HHI, the better the diversification.
We follow Dorn, Huberman, and Sengmueller (2008) in assuming that if the
security is a fund, the fund consists of one hundred equally weighted positions.
We use idiosyncratic risk in investor portfolios as a measure of diversifiable
risk. For that purpose, we take the variance of the residuals from regression
Equation (1) and, as in Calvet, Campbell, and Sodini (2007), divide it by the
total variance of the dependent variable from the same regression. As this share
increases, the investor bears more diversifiable idiosyncratic risk. Therefore, it
follows that the lower this ratio, the better the diversification. We calculate the
home bias as the percentage of equity in German companies out of total equity.
Therefore, it follows that the lower the home bias, the better the diversification.
We next focus on two portfolio performance metrics: the Sharpe ratio
and a manipulation-proof performance measure (MPPM). The Sharpe ratio
(Sharpe 1966) is a commonly accepted measure of risk-adjusted portfolio
performance. It is calculated by dividing the portfolio excess return by the
portfolio return’s standard deviation. The larger the Sharpe ratio, the better
the portfolio performance. Nevertheless, the Sharpe ratio is subject to several
shortcomings, which Goetzmann et al. (2007) address by proposing a new
Manipulation Proof Performance Measure (MPPM). We follow their MPPM
formula and, similar to other studies (e.g., Deuskar et al. 2011), use the values
2, 3, and 4 for the risk aversion coefficient ρ. We report results for only ρ =
3, as all our results are qualitatively unaltered with ρ = 2 or 4. The larger the
MPPM, the better the portfolio performance. 12
12 The MPPM is not defined for returns of negative 100% (e.g., when an investor holds only one security that
defaults on a specific day and loses all its value). These observations in our data (about 0.001% of the total
number of our observations) are set to missing.
988
Downloaded from
http://rfs.oxfordjournals.org/ at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

Is Unbiased Advice Sufficient?
2.4 Degree of following
To measure the extent to which investors who opted into the advisory model
actually follow the recommendations, we construct a variable that captures the
degree of following the advice:
Degree of following j,d
=
�i=1 N Euro| j, i, d �
actual j, i, drecommended
� i=1
N Euro| j, i, d actual + � i=1
N Euro| j, i, d recommended − � i=1
N Euro| j, i, d actual
� ,
j, i, drecommended
where j denotes the investor, i indicates the specific security, d indexes the
trading day, and Euro is the value in euros that investor j holds in security i on
trading day d. The numerator is the sum in euros of all overlapping securities
(i.e., of those securities that occur in both the actual and the recommended
portfolio). The denominator is the value of the actual portfolio plus the value
of the recommended portfolio, less the overlap. Thus, the degree of following is
actually a ratio between the intersection of the two sets and the union of the two
sets, where the two sets are the actual portfolio and the recommended portfolio.
The ratio can only take on values between zero and one. The ratio is one if a
customer fully follows the advice and zero if the actual and recommended
portfolio do not share a single security.
Table 4 uses a real recommendation to illustrate how this metric works.
The degree of following variable is a measure of how closely advice is
implemented at a particular point in time. We also calculate the change in the
degree of following from the day the advice is given to each day in the period
from t = 5 to t = 11. This change in the degree of following is an exact measure
of the client’s efforts to implement the advice; if the measure is increasing, it
is implied that the advisee is acting in accordance with the recommendations
of the advice.
Our measure considers both the buy and sell sides of following advice.
If a security in the advisee’s portfolio is not included in the recommended
portfolio, this suggests that the investor has been advised to sell this security.
It is possible, however, that the advisee sells a security for reasons other than
following the advice (e.g., liquidity or tax motives). As a check for robustness,
we construct an alternative measure of the degree of following. This measure
considers only the buy side and is simply the share of the recommended
portfolio that the investor holds at any time. Appendix A shows that all results
remain the same qualitatively when using this alternative measure.
3. Descriptive Statistics
3.1 Statistics on clients and portfolios
Table 5 provides summary statistics. It divides the sample group into customers
who opt to obtain the free advice and customers who opt not to obtain the free
(2)
989
Downloaded from
http://rfs.oxfordjournals.org/ at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

The Review of Financial Studies / v 25 n 4 2012
990
Table 4
Illustration of the degree of following
Original portfolio (t = 0) Recommended portfolio (t = 0) Required Buy Sell Keep
e e action e e e
Deutsche Bank 8,646 −→ Sell 8,646
HSBC Indian Equity Fund 3,622 −→ Sell 3,622
Raiffeisen CEE Equity Fund 2,792 −→ Sell 2,792
HSBC BRIC Equity Fund 1,862 −→ Sell 1,862
Commerzbank 439 −→ Sell 439
E.ON 2,523 E.ON AG 681 −→ Decrease 1,842 681
H&M 3,543 H&M 3,543 −→ Keep 3,543
Comstage ETF EONIA 4,072 −→ Buy 4,072
Schroder ISF Europa Corporate Bond Fund 3,883 −→ Buy 3,883
Allianz Pimco Europazins Bond Fund 3,799 −→ Buy 3,799
Allianz Pimco Corporate Bond Europa Fund 2,434 −→ Buy 2,434
UBS Lux Bond Fund 1,751 −→ Buy 1,751
Grundbesitz Europa Real Estate Fund 1,470 −→ Buy 1,470
Pictet EM Fund 1,247 −→ Buy 1,247
Allianz RCM Small Cap Fund 808 −→ Buy 808
Total 23,426 23,688 19,463 19,202 4,224
Table 4 provides an actual example of the measure that we call the degree of following. Securities E.ON and H&M overlap between the actual and the recommended portfolio. The value
of the overlap is 4,224 euros (e681 in E.ON and e3,543 in H&M). We calculate the degree of following as the overlap between the two portfolios divided by the value of assets in euros in
the actual portfolio plus assets in the recommended portfolio, less the overlap (here: 4,224 / (23,426 + 23, 688 − 4,224) = 10%).
Downloaded from
http://rfs.oxfordjournals.org/ at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

Is Unbiased Advice Sufficient?
Table 5
Summary statistics
Obtain advice Not obtain advice t-test
Data variable Measurement units Mean Median N Mean Median N P-value
Client demographics
Gender Dummy = 1 if male 91.4 100.0 385 81.3 100.0 7,810 .00
Age Years 52.9 52.0 385 49.0 47.0 7,810 .00
Wealth Microgeoraphic status 6.6 7.0 340 6.3 6.0 6,847 .00
Portfolio characteristics
Length of relationship with the bank Years since account opening 9.1 9.7 385 7.4 8.8 7,810 .00
Risky portfolio value at t = 0 e thousands 70.8 46.3 369 45.3 31.6 7,116 .00
Cash at t = 0 e thousands 20.7 10.1 369 21.7 13.2 7,116 .52
Risky share at t = 0 % 75.9 85.4 369 66.1 73.9 7,116 .00
Average trades from t = −44 to t = 0 Trades per months 2.4 1.4 385 1.9 1.0 7,810 .01
Average portfolio turnover from t = −44 to t = 0 %, monthly 5.8 3.6 378 7.5 4.0 7,745 .00
Disposition effect from t = − 44 to t = 0 PGR - PGL 10.2 5.3 385 12.3 2.7 7,810 .13
Share of tax-free assets at t = 0 % 85.8 96.9 369 85.5 98.9 7,081 .82
Portfolio performance measures
Raw returns
From t = −44 to t = 0 (actual portfolios) %, annualized −5.3 −4.7 316 −7.0 −5.4 5,232 .13
From t = 5 to t = 11 (actual portfolios) %, annualized 21.2 0.2 382 17.0 20.2 7,157 .21
From t = 5 to t = 11 (buy and hold portfolios) %, annualized 18.4 18.5 384 .68a From t = 5 to t = 11 (recommended portfolios) %, annualized 24.8 22.6 383 .02a Standard deviation
From t = −44 to t = 0 (actual portfolios) %, annualized 25.8 23.8 316 30.4 26.5 5,232 .00
From t = 5 to t = 11 (actual portfolios) %, annualized 15.0 14.0 382 21.2 17.9 7,157 .00
From t = 5 to t = 11 (buy and hold portfolios) %, annualized 14.5 14.0 384 .00a From t = 5 to t = 11 (recommended portfolios) %, annualized 9.6 8.9 383 .00a Sharpe ratio
From t = −44 to t = 0 (actual portfolios) % −4.1 −4.8 316 −4.0 −4.3 5,231 .69
From t = 5 to t = 11 (actual portfolios) % 21.8 22.5 382 15.8 16.5 7,137 .00
From t = 5 to t = 11 (buy and hold portfolios) % 13.9 18.7 384 .02a From t = 5 to t = 11 (recommended portfolios) % 35.2 35.4 384 .00a (continued)
991
Downloaded from
http://rfs.oxfordjournals.org/
at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

The Review of Financial Studies / v 25 n 4 2012
992
Table 5
Continued
Obtain advice Not obtain advice t-test
Data variable Measurement units Mean Median N Mean Median N P-value
MPPM
From t = −44 to t = 0 (actual portfolios) % −16.0 −13.3 316 −22.3 −14.6 5,232 .06
From t = 5 to t = 11 (actual portfolios) % 17.3 17.5 382 9.4 16.0 7,157 .03
From t = 5 to t = 11 (buy and hold portfolios) % 14.4 15.0 384 .18a From t = 5 to t = 11 (recommended portfolios) % 22.4 21.2 384 .00a Four-factor alpha
From t = −44 to t = 0 (actual portfolios) %, annualized −6.0 −5.6 316 −8.5 −6.3 5,231 .03
Four-factor beta
From t = −44 to t = 0 (actual portfolios) 0.8 0.8 316 0.9 0.9 5,231 .21
From t = 5 to t = 11 (actual portfolios) 0.5 0.5 381 0.7 0.7 7,080 .00
From t = 5 to t = 11 (buy and hold portfolios) 0.6 0.6 384 .00a From t = 5 to t = 11 (recommended portfolios) 0.4 0.4 384 .00a Portfolio diversification measures
HHI
At t = 0 (original portfolios) % 12.0 7.6 369 20.3 10.8 7,100 .00
At t = 11 (actual portfolios) % 10.4 4.8 377 19.8 9.8 7,014 .00
At t = 11 (buy and hold portfolios) % 9.7 4.7 384 .00a At t = 11 (recommended portfolios) % 2.9 1.6 384 .00a Idiosyncratic variance share
From t = −44 to t = 0 (actual portfolios) % 36.3 30.6 316 39.5 33.9 5,231 .01
From t = 5 to t = 11 (actual portfolios) % 29.6 21.9 381 32.5 23.7 7,080 .03
From t = 5 to t = 11 (buy and hold portfolios) % 30.0 19.6 384 .06a From t = 5 to t = 11 (recommended portfolios) % 21.2 18.1 384 .00a (continued)
Downloaded from
http://rfs.oxfordjournals.org/
at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

Is Unbiased Advice Sufficient?
Table 5
Continued
Obtain advice Not obtain advice t-test
Data variable Measurement units Mean Median N Mean Median N P-value
Home bias
At t = 0 (original portfolios) % Of equity in Germany 51.2 51.7 369 55.2 58.2 7,116 .03
At t = 11 (actual portfolios) % Of equity in Germany 43.9 41.3 378 51.1 52.2 7,177 .00
At t = 11 (buy and hold portfolios) % Of equity in Germany 45.6 42.5 384 .00a At t = 11 (recommended portfolios) % Of equity in Germany 30.3 28.7 384 .00a aActual portfolios used for non-advised clients
Table 5 reports summary statistics on client demographics, portfolio characteristics, portfolio performance measures, and portfolio diversification measures. The columns “Obtain advice”
and “Not obtain advice” present means, medians, and number of observations in regard to the respective clients in each group. The last column reports p-values of a t-test on a difference
of means. Client demographics comprise statistics on the share of male clients (Gender), the age of clients (Age), and the wealth of a client measured by the microgeographic status rating,
one through nine, by an external agency (Wealth). Portfolio characteristics comprise statistics on the number of years the client has been with the bank (Length of relationship), the risky
portfolio value of the customer (Risky portfolio value at t = 0), the amount of cash held with this brokerage (Cash at t = 0), the proportion of risky assets held with this brokerage (Risky
share at t = 0), the number of trades per month (Average trades from t = −44 to t = 0), the average portfolio turnover per month (Average portfolio turnover from t = − 44 to t = 0), the
difference between the proportion of realized gains and losses (Disposition effect from t = − 44 to t = 0), and the proportion of tax-free assets (Share of tax-free assets at t = 0). Portfolio
performance measures comprise statistics on raw returns, standard deviations, Sharpe ratios, Manipulation Proof Performance Measures (MPPM), four-factor alphas, and four-factor betas
for the period before the advice was offered, t = −44 to t = 0, and the period after the advice was offered, t = 5 to t = 11. Portfolio diversification measures comprise statistics on the
Herfindahl-Hirschmann index (HHI), the idiosyncratic variance share (Idiosyncratic variance share), and the share of domestic equity (Home bias). Alpha, beta, and idiosyncratic risk share
stem from the application of a Carhart (1997) four-factor model that is calibrated for Germany.
993
Downloaded from
http://rfs.oxfordjournals.org/
at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

The Review of Financial Studies / v 25 n 4 2012
advice. P-values of t-tests from our tests for the equality of variables across
these two groups are provided in the last column.
Table 5 shows that 91% of the customers who accepted the offer are male,
compared with 81% in the control group. The mean age is also slightly
higher (52.9 years vs. 49.0 years), as is the wealth level measured by the
microgeographic status (6.6 vs. 6.3). This indicates that the customers who
accept the offer are more likely to be male, older in age, and richer. The
customers who accept the offer have had a longer relationship with the bank
(9.1 years vs. 7.4 years). Portfolio characteristics are also significantly different
for the two subgroups of our sample. The customers who accept the offer have
a higher-risk portfolio value at t = 0 (e70,800 vs. e45,300), a higher share
of risky assets (75.9% vs. 66.1%), more trades per month (2.4 vs. 1.9), and
lower portfolio turnover (5.8% vs. 7.5%). 13 We also calculate the disposition
effect for each investor during the pre-advice period by applying the Odean
(1998) method. Differences between advisees and non-advisees exist but are
not significant.
For those customers who have an account from September 2005 to May
2009—the pre-advice period—we calculate average daily returns of their
investment portfolios, the standard deviations of these returns, and four-factor
alphas. The raw returns are not significantly higher for the customers who
accept the offer than they are for the customers who do not accept the offer.
The standard deviations of returns are significantly lower for the customers
who accept the offer than they are for the customers who do not accept the
offer. The alphas are significantly higher for the customers who accept the
offer than they are for customers who do not accept the offer, though both their
alphas are significantly negative. The Sharpe ratios and the betas are similar.
However, MPPM is significantly higher for the customers who accept the offer
than it is for customers who do not accept the offer, though for both groups
the MPPMs are negative. All of this evidence suggests that the customers who
accept the offer are likely to be more financially sophisticated.
We also notice that diversification (as measured by lower HHI, lower
idiosyncratic risk share, and lower home bias) is significantly higher for the
customers who accept the offer than it is for the customers who do not accept
the offer. This further confirms that the customers who accept the offer are
likely to be more financially sophisticated. All the above results are confirmed
in Section 4 by multivariate tests.
It is important to mention that the largely negative alpha estimates in the
pre-advice period tell us that all our customers, regardless of whether they
accept the offer, significantly underperform the benchmark index. Similarly,
13 In comparison with official statistics provided by Deutsche Bundesbank (2010) and Deutsches Aktieninstitut
(2009), the investors in our sample have about the same age (fifty years) but are more likely to be male and
richer. The larger portfolio and large cash values signal that these brokerage accounts do not represent “play
money” (Goetzmann and Kumar 2008). In line with Calvet, Campbell, and Sodini (2007), our investors also
hold portfolios with high idiosyncratic volatility shares.
994
Downloaded from
http://rfs.oxfordjournals.org/ at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

Is Unbiased Advice Sufficient?
high idiosyncratic risk shares, high HHIs, and high home bias show significant
potential for improvement of diversification. Regardless of whether investors
are self-directed or follow some outside advice, they could benefit from
unbiased and theoretically sound advice, especially the ones who were doing
relatively worse.
We notice that the improvement in actual raw returns from the pre-advice
period to the post-advice period for the customers who accept the offer (−5.3%
to 21.2%) does not seem to be much different from the customers who do
not accept the offer (−7.0% to 17.0%). The drop in standard deviations from
the pre-advice period to the post-advice period for the customers who accept
the offer (25.8% to 15.0%) also does not seem to be much different from the
customers who do not accept the offer (30.4% to 21.2%). A multivariate test
later confirms that obtaining advice does not improve diversification: there is
no significant decrease in HHI or the idiosyncratic risk of advisees’ portfolios
compared with the non-advisees’ portfolios. The results, with respect to
portfolio performance, are mixed: there is a slight increase in Sharpe ratio but
no increase in MPPM of advisees’ portfolios compared with the non-advisees’
portfolios. Taken together, this suggests that the average advisee does not much
benefit from the advice.
There are several explanations to answer the questions of why the average
advisee does not much benefit from the advice. It could be that the advice is not
sound or that the average advisee does not follow the advice. To check whether
the advice is sound, we perform two simple univariate tests: we compare the
recommended portfolios with the investors’ actual portfolios and their buy and
hold portfolios.
Table 5 shows these results. The recommended portfolios perform much better
than the actual portfolios of the advisees in the post-advice period: a return
of 24.8% vs. 21.2%, a standard deviation of 9.6% vs. 15.0%, a Sharpe ratio of
35.2% vs. 21.8%, an MPPM of 22.4% vs. 17.3%, an HHI of 2.9% vs. 10.4%,
an idiosyncratic risk share of 21.2% vs. 29.6%, and a home bias of 30.2% vs.
43.9%. It is also noteworthy that the recommended portfolios perform much
better than the buy and hold portfolios of the advisees in the post-advice period:
a return of 24.8% vs. 18.4%, a standard deviation of 9.6% vs. 14.5%, a Sharpe
ratio of 35.2% vs. 13.9%, an MPPM of 22.4% vs. 14.4%, an HHI of 2.9%
vs. 9.7%, an idiosyncratic risk share of 21.2% vs. 30.0%, and a home bias of
30.2% vs. 45.5%. These results suggest that the financial advice was sound and
that the average advisee does not follow the advice. If he or she had followed
the recommendations, he or she would have improved his or her investment
performance. Formal tests later conducted confirm all these initial findings.
3.2 Statistics on the degree of following the advice
We now give descriptive statistics on the measure of degree of following to give
a sense of how many investors follow the advice once they elect to receive it.
995
Downloaded from
http://rfs.oxfordjournals.org/ at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

The Review of Financial Studies / v 25 n 4 2012
Figure 2
Initial degree of following
Panel A: The cross-sectional distribution of the degree of following when the first recommendation was given.
Panel B: The cross-sectional distribution of the average degree of following between time t = advice start and
t = 11 months. Panel C: Increase in degree of following from the time of the first recommendation to t = 11.
Figure 2, panel A, gives the distribution of the degree of following on the
day the recommendation is received by an investor. This figure shows that the
recommended portfolio is very different from the average advisee’s existing
portfolio. For about one in five investors, there is essentially no overlap, and
for about half of them, overlap is less than 20%. In fact, no one’s existing
portfolio coincides with his or her recommended portfolio.
Figure 2, panel B, gives the distribution of the average degree of following
between t = advice start through t = 11 months, which is the “post-advice”
period, for investors who choose to receive advice. Compared with Figure 2,
panel A, this figure indicates that some mass of the distribution shifted to the
right. This implies that some investors follow the advice in the “post-advice”
period and increase the degree of following. However, the distributions shown
in Figure 2, panels A and B, are not very different from each other, suggesting
that, in fact, few investors follow the advice.
Figure 2, panel C, gives the distribution of the increase in the degree of
following from the date the investor receives the advice to t = 11. As the big
mass is at zero, it tells us that most investors do not follow the advice. In
fact, some mass is in the negative zone, indicating that some investors even
996
Downloaded from
http://rfs.oxfordjournals.org/
at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

Is Unbiased Advice Sufficient?
disregard the advice, i.e., sell securities they are recommended to keep and/or
buy securities that are not recommended. However, more investors follow the
advice and then disregard it. (Though, there are relatively few investors who do
either of these.) Figure 2, panel C, looks similar when we measure the increase
in the degree of following from the date of the recommendation to ten days
later, twenty days later, thirty days later, or use the average increase over the
entire post-advice period.
Table 6 gives the descriptive statistics on distributions of the degree of
following advice taken at different points in time.
We see in this table that the mean degree of following is 15.6% on the day of
the recommendation for the average advisee. This means that, on average, the
majority of the positions in the original portfolio have to be changed. This is
not an onerous task because the average advisee’s portfolio already experiences
an annual turnover of approximately 70%.
For the average advisee, the mean degree of following increases to 21.6%
ten days after the advice, to 24.4% twenty days after the advice, and to 25.4%
thirty days after the advice. However, the mean degree of following is lower
(16.4%) at the end of the post-advice period, suggesting that, although they
may initially act in accordance with the advice, they do not stick to it. This also
explains why the mean degree of following for the entire post-advice period is
only 21.1%. An important statistic in this table is the number of investors who,
at least partially, follow the advice. Of the 385 who opt to receive advice, 260
(385 – 125) do not follow it, as measured at the end of the post-advice period.
The 125 investors who do follow the advice, as can be seen in Figure 2, panel
C, follow it very little.
Our conclusions from Figure 2 and Table 6 are straightforward. Recommended
portfolios are very different from the average advisee’s actual
portfolios, and the average advisee does not much heed the financial advice.
We do not believe that this is due to transaction costs because, as previously
stated, investors frequently turn over their portfolios.
Table 6
Summary statistics for the degree of following
t =11 months Average over
(end of measurement
measurement period (t = 5 to
t = Advice start t = 10 days t = 20 days t = 30 days period) t = 11 months)
Degree of following
Mean (%) 15.6 21.6 24.4 25.4 16.4 21.1
Median (%) 10.8 14.2 15.5 16.2 10.4 14.7
Number of followers n/a 98 141 158 125 157
Observations 385 385 385 385 385 385
Table 6 reports summary statistics for the degree of following. The table reports the mean degree of following,
the median degree of following, the number of followers, and observations for six distinct time periods (upon
receiving the first recommendation, after ten days, after twenty days, after thirty days, and at the end of the
measurement period and the average over the entire measurement period). These time intervals are specific to
each investor in the sample. Number of followers is defined as the number of investors who increase their degree
of following until that point in time, at least marginally.
997
Downloaded from
http://rfs.oxfordjournals.org/ at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

The Review of Financial Studies / v 25 n 4 2012
4. Who Chooses to Obtain Advice?
By October 2009, a total of 385 customers, out of the 8,195 customers that
were offered free and unbiased financial advice, elected to accept the offer.
This constitutes a little less than 5% of the customers who were contacted.
A total of thirty-eight investors joined in May 2009, 146 in June 2009, seventythree
in July 2009, and the remaining 128 joined later.
We now formally examine those who choose to receive advice. Table 7
reports the results of a probit test, where the dependent variable is set to one
if a client opted to receive financial advice and zero otherwise. 14 We make the
following conclusions.
Though our sample is predominantly male, older, and rich, the clients who
opt to receive free financial advice are more likely to be male, older, richer
(to be precise, the effect comes from low-wealth clients who choose not to
obtain advice), and have a higher risky share in their portfolio. They are more
active traders, as measured by the number of trades per month, but have less
portfolio turnover, relative to their risky portfolio values. Their alphas are
higher, their HHI are lower, their home bias is lower, and their disposition
effect bias is lower. Their long-term raw returns are higher, but their short-term
raw returns are lower. They also have more experience with the brokerage,
where experience is measured by the length of the relationship between the
client and the brokerage. A lower share of tax-free assets, especially when
controlling for diversification, is positively correlated with opting for advice. 15
Age and wealth are linked to financial sophistication in the literature (Calvet,
Campbell, and Sodini 2007; Calvet, Campbell, and Sodini 2009b), though
being male is not linked to financial sophistication (Barber and Odean 2001).
Higher alphas, lower HHI, lower home bias, and lower disposition effect bias
are linked to financial sophistication. So, why do the financially sophisticated
opt for advice? It is possible that these investors became financially sophisticated
because of financial literacy training—see Carlin and Robinson (2010)—
and so they have a positive view of advice. A second hypothesis is plausible:
elderly and financially more sophisticated men are more likely to be defrauded
(NASD 2006) or older investors are more likely to make erroneous financial
investment decisions (Agarwal et al. 2009); opting for financial advice by these
groups could also be interpreted as an attempt to address these disadvantages.
14 In order to address the issue of commonality among the recommendations, all our results in Tables 7 and 8 are
also estimated by using a cluster robust regression analysis, with risk class being the cluster variable. All nonadvised
clients were grouped into one class. Results remain qualitatively unaltered. We also notice that investors
opted for the advice at different points in time. To deal with this, we also use clustered standard errors in Tables
7 and 8, with week of opting being the cluster variable. Again, results remain qualitatively unaltered.
15 Poterba and Samwick (2003) suggest that taxes may influence an investor’s trading decisions. A tax law change
in Germany made it favorable to hold assets that were in existing portfolios at the end of 2008 because they
would remain tax-free if held for more than one year. We calculate the share of tax-free assets in the portfolio
at the time of the offer. As this share increases, propensity to seek financial advice may decrease. However,
advisees sell, on average, 57.5% of their tax-free assets until the end of our measurement period. So, they lose
most of their tax advantage.
998
Downloaded from
http://rfs.oxfordjournals.org/ at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

Is Unbiased Advice Sufficient?
Table 7
Who opts for advice? A probit test
Dummy advice
Dependent variable (1) (2) (3) (4) (5) (6) (7)
Dummy male 0.361∗∗∗ 0.278∗∗∗ 0.285∗∗∗ 0.288∗∗∗ 0.291∗∗∗ 0.290∗∗∗ 0.257∗∗∗ (0.000) (0.001) (0.002) (0.002) (0.000) (0.000) (0.007)
Age 0.009∗∗∗ 0.008∗∗∗ 0.011∗∗∗ 0.010∗∗∗ 0.008∗∗∗ 0.008∗∗∗ 0.011∗∗∗ (0.000) (0.001) (0.000) (0.000) (0.000) (0.000) (0.000)
Dummy low wealth −0.255∗ −0.267∗∗ −0.177 −0.183 −0.264∗ −0.271∗∗ −0.200
(0.059) (0.047) (0.213) (0.197) (0.050) (0.043) (0.181)
Dummy high wealth 0.048 0.062 0.047 0.047 0.065 0.067 0.030
(0.353) (0.233) (0.413) (0.417) (0.217) (0.201) (0.615)
Log portfolio value (t =0) 0.144∗∗∗ 0.130∗∗∗ 0.130∗∗∗ 0.131∗∗∗ 0.102∗∗∗ 0.130∗∗∗ 0.138∗∗∗ (0.000) (0.001) (0.004) (0.005) (0.008) (0.001) (0.002)
Length of relationship 0.063∗∗∗ 0.098∗∗∗ 0.098∗∗∗ 0.068∗∗∗ 0.064∗∗∗ 0.102∗∗∗ (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
Trades per month (t = −44 to t =0) 0.042∗∗∗ 0.043∗∗∗ 0.042∗∗∗ 0.028∗∗∗ 0.037∗∗∗ 0.042∗∗∗ (0.000) (0.000) (0.000) (0.008) (0.000) (0.001)
Portfolio turnover (t = −44 to t =0) −1.104∗∗ −1.042∗ −1.030∗ −0.450 −0.966∗∗ −0.589
(0.010) (0.080) (0.095) (0.310) (0.024) (0.427)
Disposition effect (t = −44 to t =0) −0.163∗ −0.208∗∗ −0.203∗∗ −0.204∗∗ −0.154∗ −0.208∗∗ (0.065) (0.031) (0.035) (0.027) (0.079) (0.039)
Risky share (t =0) −0.103 −0.192 −0.201 −0.082 −0.103 −0.268∗∗ (0.353) (0.115) (0.101) (0.468) (0.356) (0.029)
Share of tax-free assets (t = 0) −0.194∗ −0.219∗ −0.235∗ −0.287∗∗ −0.217∗ −0.223
(0.085) (0.094) (0.075) (0.017) (0.055) (0.151)
Alpha (t = −44 to t = 0) 21.836∗∗ (0.012)
Idiosyncratic variance share (t = −44 to t = 0) −0.189
(0.241)
HHI (t = 0) −0.883∗∗∗ (0.000)
(continued)
999
Downloaded from
http://rfs.oxfordjournals.org/
at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

The Review of Financial Studies / v 25 n 4 2012
1000
Table 7
Continued
Dummy advice
Dependent variable (1) (2) (3) (4) (5) (6) (7)
Home bias (t =0) −0.197∗∗∗ (0.008)
Long-term raw return (t = −44 to t = −1) 27.917∗∗ (0.036)
Short-term raw return (t = −1 to t =0) −2.750∗ (0.100)
Constant −3.937∗∗∗ −3.923∗∗∗ −4.327∗∗∗ −4.266∗∗∗ −3.513∗∗∗ −3.850∗∗∗ −4.287∗∗∗ (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
Observations 7,450 7,449 5,520 5,520 7,432 7,449 5,056
Pseudo-R2 0.0381 0.0668 0.0677 0.0664 0.0778 0.0691 0.0682
Table 7 reports probit estimates of the participation in the advisory model offered by the brokerage. Clients are set equal to one once they obtain the advice. For the estimation of the
probit model, we include the following independent variables: a dummy that is equal to one if a client is male (Dummy male), the age of a client (Age), a dummy that is equal to one if a
client falls into categories one to three of a microgeographic status rating by an external agency (Dummy low wealth), a dummy that is equal to one if a client falls into categories seven to
nine of the microgeographic status (Dummy high wealth), the risky portfolio value of the customer (Log portfolio value), the number of years the client has been with the bank (Length of
relationship), the number of trades per month (Trades per month), the average portfolio turnover per month (Portfolio turnover), the difference between the proportion of realized gains and
losses (Disposition effect), the proportion of risky assets in the account (Risky share), the proportion of tax-free assets (Share of tax-free assets), the weekly alpha of a particular customer
before opting for financial advice (Alpha), the idiosyncratic variance share (Idiosyncratic variance share), the Herfindahl-Hirschman index (HHI), the share of domestic equity (Home
bias), the raw return from the beginning of observation until one month prior to the offer of financial advice (Long-term raw return), and the raw return of the month before the offer of
financial advice (Short-term raw return). Alpha and idiosyncratic risk share stem from the application of a Carhart (1997) four-factor model that is calibrated for Germany. P-values are in
parentheses. Pseudo-R2 values and number of observations are reported. *** denotes significance at the 1% level or less, ** significance at 5% or less, and * significance at 10% or less.
Heteroscedasticity robust standard errors are used. Standard errors shown are not clustered, but results remain qualitatively unaltered when clustering them by advice week or risk aversion.
Different counts of observations are due to data availability of certain variables (see Table 5); results are robust to using the lowest common denominator.
Downloaded from
http://rfs.oxfordjournals.org/ at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

Is Unbiased Advice Sufficient?
It is also possible that it is not financial sophistication but a third hypothesis
that explains the results: regret. Investors who have performed poorly in the
past have regret. They are reluctant to obtain advice and to acknowledge past
mistakes. They prefer inaction (ignore advice and be wrong, which comes
at a lower psychic cost) to action (obtain advice, change portfolio, and risk
being wrong, which comes at a higher psychic cost). Carlin and Robinson
(1999) have a good example of this happening in Las Vegas. This regret
hypothesis is, however, weakened by our finding that investors with lower
short-term returns opt for advice more often. Furthermore, we computed a
variable, proportion of paper losses, in an advisee’s portfolio when the advice
was offered, and this variable seemed to have no effect. A fourth hypothesis
could be the “once bitten, twice shy” hypothesis: investors who had followed
past advice and had done well want to receive more advice, whereas investors
who had followed past advice and had not done well do not want to receive
more advice. However, any past advice received by these investors could not
have come from this brokerage, as this is the first time that the brokerage
has offered advice. A fifth hypothesis could be that our clients are optimistic,
perhaps extremely optimistic. Puri and Robinson (2007) show that optimists
invest more in individual stocks, and extreme optimists display imprudent
financial behavior (like not wanting advice). While this could be true, we
do not have any psychological metrics on our clients and cannot test this
hypothesis. A sixth hypothesis could be that investors receive advice from
other sources. Even if they receive outside advice, the outside advice must
either be bad or not followed because we find portfolios to be largely inefficient
before, as well as after, the advice is offered. A seventh hypothesis could be
that the offer of free and unbiased financial advice is just regarded as spam.
Again, the evidence suggests otherwise. The e-mail that contained the offer
was an official message sent to the inbox of the banking account. This inbox
is never crowded because these e-mails have a predetermined maturity. Also,
if there is no reaction to the initial e-mail, a personal phone call follows.
To summarize, we do have a fairly good handle on why advice is not being
sought: lack of financial sophistication, as measured by poor past portfolio
performance (has many interpretations), a desire to not increase tax payments,
and lack of experience, familiarity, and/or trust, as measured by the length
of relationship with the brokerage. It does seem from the first result that the
clients who most need (do not need) the financial advice are the least likely to
obtain it (most likely to obtain it). We will explore this hypothesis more fully
in Section 7.
5. Who Chooses to Follow the Advice If They Obtain It?
We now formally examine who chooses to follow the advice once they obtain
it. Recall that advisees rarely follow the advice, and when they do, it is only
followed to a small extent. Table 8 reports the results of an ordinary least
1001
Downloaded from
http://rfs.oxfordjournals.org/ at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

The Review of Financial Studies / v 25 n 4 2012
Table 8
Who chooses to follow advice?
1002
Increase in degree of following
Dependent variable (1) (2) (3) (4) (5) (6) (7)
Dummy male −0.080 −0.076 −0.055 −0.062 −0.079 −0.080 −0.048
(0.122) (0.131) (0.298) (0.238) (0.119) (0.115) (0.374)
Age 0.000 0.000 0.000 0.000 0.000 0.000 0.000
(0.535) (0.826) (0.995) (0.967) (0.872) (0.925) (0.797)
Dummy low wealth −0.041∗∗ −0.054∗∗ −0.063∗∗∗ −0.059∗∗ −0.053∗∗ −0.052∗∗ −0.062∗∗ (0.046) (0.031) (0.007) (0.018) (0.035) (0.049) (0.011)
Dummy high wealth 0.045∗∗ 0.044∗ 0.038 0.038 0.044∗ 0.043∗ 0.033
(0.039) (0.051) (0.120) (0.115) (0.051) (0.055) (0.182)
Log portfolio value (t = 0) −0.060∗∗∗ −0.065∗∗∗ −0.060∗∗∗ −0.056∗∗∗ −0.064∗∗∗ −0.066∗∗∗ −0.056∗∗∗ (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
Length of relationship 0.003 0.006 0.006 0.002 0.003 0.005
(0.398) (0.156) (0.188) (0.459) (0.425) (0.257)
Trades per month (t = −44 to t = 0) 0.001 0.002 0.003 0.002 0.002 0.003
(0.751) (0.636) (0.408) (0.614) (0.618) (0.528)
Portfolio turnover (t = −44 to t = 0) −0.025 −0.052 −0.186 −0.054 −0.051 −0.304
(0.906) (0.842) (0.470) (0.806) (0.811) (0.413)
Disposition effect (t = −44 to t = 0) −0.027 −0.013 0.002 −0.026 −0.030 0.032
(0.579) (0.798) (0.960) (0.591) (0.538) (0.518)
Risky share (t = 0) 0.002 −0.014 −0.018 −0.001 0.003 −0.002
(0.973) (0.753) (0.684) (0.976) (0.945) (0.970)
Share of tax-free assets (t = 0) 0.051 0.054 0.084 0.057 0.057 0.083∗
(0.240) (0.272) (0.126) (0.200) (0.188) (0.056)
Recommended portfolio over original portfolio −0.003 −0.002 −0.003 −0.005 −0.003 0.014
(0.829) (0.841) (0.830) (0.728) (0.785) (0.464)
Alpha (t = −44 to t = 0) 6.160
(0.155)
Idiosyncratic variance share (t = −44 to t = 0) 0.089
(0.152)
HHI (t = 0) 0.061
(0.534)
(continued)
Downloaded from
http://rfs.oxfordjournals.org/
at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

Is Unbiased Advice Sufficient?
Table 8
Continued
Increase in degree of following
Dependent variable (1) (2) (3) (4) (5) (6) (7)
Home bias (t = 0) 0.031
(0.382)
Long-term raw return (t = −44 to t = −1) 8.195
(0.169)
Short-term raw return (t = −1 to t = 0) −0.872
(0.401)
Constant 0.754∗∗∗ 0.752∗∗∗ 0.670∗∗∗ 0.580∗∗∗ 0.741∗∗∗ 0.751∗∗∗ 0.594∗∗∗ (0.000) (0.000) (0.000) (0.001) (0.000) (0.000) (0.001)
Observations 369 365 313 313 365 365 297
R2 0.132 0.139 0.138 0.139 0.141 0.141 0.131
Table 8 reports OLS estimates of the coefficients related to an increase in the degree of following measure from the date the investor received advice to t = 20 days (Increase in degree
of following). For the estimation of the model, we include the following independent variables: a dummy that is equal to one if a client is male (Dummy male), the age of a client (Age),
a dummy that is equal to one if a client falls into categories one to three of a microgeographic status rating by an external agency (Dummy low wealth), a dummy that is equal to one if a
client falls into categories seven to nine of the microgeographic status (Dummy high wealth), the risky portfolio value of the customer (Log portfolio value), the number of years the client
has been with the bank (Length of relationship), the number of trades per month (Trades per month), the average portfolio turnover per month (Portfolio turnover), the difference between
the proportion of realized gains and losses (Disposition effect), the proportion of risky assets in the account (Risky share), the proportion of tax-free assets (Share of tax-free assets), a ratio
of the required net investment (Recommended portfolio/original portfolio), the weekly alpha of a particular customer before opting for financial advice (Alpha), the idiosyncratic variance
share (Idiosyncratic variance share), the Herfindahl-Hirschman index (HHI), the share of domestic equity (Home bias), the raw return from the beginning of observation until one month
prior to the offer of financial advice (Long-term raw return), and the raw return of the month before the offer of financial advice (Short-term raw return). Alpha and idiosyncratic risk share
stem from the application of a Carhart (1997) four-factor model that is calibrated for Germany. P-values are in parentheses. R2 values and number of observations are reported. *** denotes
significance at 1% or less, ** significance at 5% or less, and * significance at 10% or less. Heteroscedasticity robust standard errors are used. Standard errors shown are not clustered, but
results remain qualitatively unaltered when clustering them by advice week or risk aversion. Different counts of observations are due to data availability of certain variables (see Table 5);
results are robust to using the lowest common denominator.
1003
Downloaded from
http://rfs.oxfordjournals.org/ at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

The Review of Financial Studies / v 25 n 4 2012
squares (OLS) regression, where the dependent variable is the increase in the
degree of following the advice. We find that wealthier investors (as measured
by the microgeographic status) and investors with lower risky portfolio values
tend to follow the advice more. Interestingly, though investors with higher
portfolio value opt for the advice, investors with lower portfolio value follow
it more. One interpretation of this seeming contradiction is that people with
lower portfolio value have, among the richer clients, relatively less experience
with risky assets and may therefore follow advice more.
The more interesting observation in Table 8 is the set of nonresults. None
of the other variables have a significant effect on the degree to which advice is
followed, including the disposition effect. The next paragraph discusses more
nonresults.
First, it is conceivable that investors do not pay attention to advice because
they consider their accounts in the brokerage as just “play money.” The
evidence suggests otherwise. As can be seen in Table 1, panel D, the median
original portfolio of the client has the same size as his median recommended
portfolio which, in turn, is about 48% of his total financial wealth. This
suggests that we are looking at investors’ main accounts and that their financial
wealth is predominantly held by this brokerage. 16 Additionally, in comparison
with official statistics provided by Deutsche Bundesbank (2010) and Deutsches
Aktieninstitut (2009), the average investor in our sample has a much higher
account value than the average retail investor in the population.
Second, it could be that the advice is not followed because investors are
asked to increase their investments in risky assets. However, our research
design recommended an increase only if the investor asked for an increase.
Even these investors had sufficient financial assets to cover this net investment,
as shown by the ratio of the recommended portfolio divided by total financial
wealth (median 48%; even the ninety-fifth percentile of the distribution had
enough financial assets to cover the required net investment with a ratio of
97%) in Table 1, panel D. Moreover, we use the ratio of the recommended
portfolio size to the original portfolio size as another independent variable in
Table 8. As can be seen, this variable has no significant effect on the decision
of whether or not to follow advice.
Third, one could argue that the advice was short-lived and therefore ignored.
Again, the evidence suggests otherwise. Our clients trade often; their average
holding period is only a little more than one year. The advice does not
seem short-lived compared with their trading horizons. Moreover, even if the
free advice ceased after a test phase, there should be no utility loss from
investing in an efficient portfolio. Fourth, it is conceivable that investors may
ignore the advice merely because they are not sophisticated. However, our
results show that subjects who choose to receive the advice are actually more
16 As pensions in Germany are provided by the state and the employer, such a concentration of savings in one
brokerage does not seem excessive.
1004
Downloaded from
http://rfs.oxfordjournals.org/ at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

Is Unbiased Advice Sufficient?
sophisticated, as measured by their past investment performance, than are their
peers who do not accept the offer. Fifth, investors may also choose to ignore
the advice because they do not trust the brokerage, but our results also show
that customers who opt for the advice have had a long relationship with the
brokerage (on average 9.1 years), even significantly longer than non-advisees.
Sixth, is inertia causing them to ignore the advice? Subjects do not seem
to be inert because they trade actively after the recommendation was given,
though they do not follow the recommendation. Seventh, it is also unlikely
that the investors are simply too busy to heed the advice, considering that they
trade actively. Eighth, it could be that clients follow advice only after a certain
time. We recomputed our dependent variable in Table 8—degree of following
from t = 0 to t = 20—for various other time durations. Our results do not
qualitatively change. Ninth, it could be that clients do not buy the stocks that
they are unfamiliar with. We defined a stock to be familiar if an advisee had
traded it in the past and then computed the proportion of stocks in the recommended
portfolio that an advisee was familiar with. This variable also had no
effect in Table 8. Tenth, advisees may not heed advice because they are too
risk-averse. Though the recommendations were tailored for their risk aversion,
we checked whether this variable had an effect in Table 8. The answer is no.
To summarize, our results cannot pinpoint why financial advice is ignored
after it is sought. The only results we have are that wealthier investors and
investors with lower risky portfolio values tend to follow the advice more. We
suspect the lack of results, with respect to the other variables in this test, is due
to a lack of power; there is little variation in the dependent variable. As seen
in Figure 2, panel C, most clients who opt for advice do not follow the advice.
The other reason could be that a systematic cause may not exist to explain why
people do not follow advice.
6. Does the Advice Benefit the Advisee?
We now come to the most critical issue addressed by our study—whether advisees
benefit from financial advice. Many previous studies have documented
that financial advice does not benefit the advisee, but the reason that is given as
the root cause of the finding is that the advice is conflicted. Our research design
ensured that the financial advice was unbiased, un-conflicted, and theoretically
sound ex ante. Was this true ex post, too?
6.1 Is the financial advice sound?
From the description of the data set and the timeline, it follows that the natural
research design to investigate the quality of financial advice is to examine its
effect by using a difference-in-difference methodology. This requires calculating
the improvement of portfolio efficiency between the pre-advice period
and the post-advice period for the treatment group (clients who opt to receive
advice) if they had fully followed the advice and comparing this difference
1005
Downloaded from
http://rfs.oxfordjournals.org/ at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

The Review of Financial Studies / v 25 n 4 2012
with the improvement of portfolio efficiency between the pre-advice period
and the post-advice period for the control group (clients who opt not to receive
advice). To do this, we define a recommended portfolio as the portfolio that the
treatment group would hold had they completely followed the advice given.
We use two measures of diversification—HHI and the share of idiosyncratic
risk (the part of the risk that is uncompensated)—and two measures of portfolio
performance—the Sharpe ratio and MPPM—as our four measures of portfolio
efficiency. The HHI for the pre-advice period is computed from the actual
portfolios of both the treatment group and the control group at t = 0. The
HHI for the post-advice period for the treatment group is computed on the
recommended portfolios of the treatment group at t = 11. The HHI for the postadvice
period for the control group is computed from their actual portfolios
at t = 11. The share of idiosyncratic variance in the pre-advice period is
computed by running the regression in Equation (1) on the actual portfolios of
both the treatment group and the control group in the period from September
2005 to May 2009. The share of idiosyncratic variance in the post-advice
period for the treatment group is computed by running the regression in
Equation (1) on the recommended portfolios of the treatment group in the
period from October 2009 to April 2010. This computation is possible because
the date and the details of each of the recommendations are known. The share
of idiosyncratic risk in the post-advice period for the control group is computed
by running the regression in Equation (1) on the actual portfolios of the control
group in the period from October 2009 to April 2010. Sharpe ratios and MPPM
in the pre-advice period are computed for the actual portfolios of both the
treatment group and the control group in the period from September 2005 to
May 2009. Sharpe ratios and MPPM in the post-advice period for the treatment
group are computed for the recommended portfolios of the treatment group in
the period from October 2009 to April 2010. Sharpe ratios and MPPM in the
post-advice period for the control group are computed for the actual portfolios
of the control group in the period from October 2009 to April 2010.
If the decrease (improvement) in the HHI or share of idiosyncratic variance
from the pre-advice period to the post-advice period for the treatment group
is greater than those of the control group, it would suggest that the advice is
theoretically sound. If the increase (improvement) in the Sharpe ratio or MPPM
from the pre-advice period to the post-advice period for the treatment group is
greater than those of the control group, it would suggest that the advice is
theoretically sound.
We run four sets of OLS regressions. The dependent variable in our first
set of OLS regressions is the decrease in HHI from the pre-advice period
to the post-advice period. The dependent variable in our second set of OLS
regressions is the decrease in the share of idiosyncratic variance from the preadvice
period to the post-advice period. The dependent variable in our third
set of OLS regressions is the increase in Sharpe ratio from the pre-advice
period to the post-advice period. The dependent variable in our fourth set
1006
Downloaded from
http://rfs.oxfordjournals.org/
at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

Is Unbiased Advice Sufficient?
of OLS regressions is the increase in MPPM from the pre-advice period to
the post-advice period. The main independent variable of interest in all the
four sets of OLS regressions is a “dummy advice” variable set to one for the
treatment group and zero for the control group. Table 9 reports the results of
these regressions.
The coefficient on the “dummy advice” variable is positive and statistically
significant in all four sets of regressions. This implies that if advisees had
fully followed the advice, they would have improved their portfolio efficiency
(decreased the HHI and share of idiosyncratic risk in their portfolios and
increased the Sharpe ratios and MPPM of their portfolios). Thus, the advice
does improve portfolio efficiency.
6.2 Does advice benefit the average advisee?
We use the same research design as described in the previous subsection, with
one important change. Instead of computing the HHI, share of idiosyncratic
risk, Sharpe ratio, and MPPM in the post-advice period for the treatment group
by using their recommended portfolios, we instead use their actual portfolios.
The coefficient on the “dummy advice” variable is statistically insignificant
in the HHI and the idiosyncratic risk share cases but is positively significant
in the Sharpe ratio case and is positively significant in one out of four
models in the MPPM case. Comparing the coefficients in these latter significant
cases with their counterparts in Table 8 reveals that the coefficients are
much smaller (of the order of one-third to one-half). All this evidence implies
that the average advisee does not improve portfolio diversification from the
advice but does improve portfolio performance, though very modestly.
Why does the average advisee not much benefit? Given that the advice
was theoretically sound, it follows that the average advisee did not much
benefit because he or she did not follow the advice. We provided evidence
that the average advisee does not follow the advice in subsection 3.2. We now
investigate whether advice benefits the advisee if he or she partially follows
the advice. To do this, we run the same regressions we ran for Table 10, with
an additional independent variable: the variable measures the increase in the
degree of following the advice between the time a client first received the
advice and the end of the post-advice period (see Figure 2, panel C). Recall
that this variable is positive (negative) if the advisee is partially or fully acting
on (acting against) the advice. Table 11 shows the results.
The coefficient on the “increase in the degree of following” measure is
statistically significant if we use HHI as a diversification measure but is
statistically insignificant if we use the idiosyncratic risk share as our measure
of diversification. The coefficient on the “increase in the degree of following”
measure is statistically significant if we use Sharpe ratio as a portfolio
performance measure but is statistically insignificant if we use MPPM as
a portfolio performance measure. Our results thus show that even partially
following the advice would have improved the efficiency of the portfolio.
1007
Downloaded from
http://rfs.oxfordjournals.org/ at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

The Review of Financial Studies / v 25 n 4 2012
Table 9
Does financial advice improve diversification and portfolio performance?
Decrease in HHI Decrease in idiosyncratic variance share
1008
Dependent variable (1) (2) (3) (4) (5) (6) (7) (8)
Dummy advice 0.088∗∗∗ 0.091∗∗∗ 0.091∗∗∗ 0.091∗∗∗ 0.082∗∗∗ 0.087∗∗∗ 0.086∗∗∗ 0.086∗∗∗ (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
Dummy male −0.000 −0.000 −0.000 0.007 0.005 0.004
(0.993) (0.983) (0.959) (0.309) (0.523) (0.549)
Age −0.000 −0.000 −0.000 0.000 0.000 0.000
(1.000) (0.948) (0.871) (0.687) (0.765) (0.796)
Dummy low wealth 0.003 0.004 0.004 −0.006 −0.008 −0.008
(0.714) (0.691) (0.693) (0.618) (0.530) (0.525)
Dummy high wealth 0.003 0.003 0.002 0.008 0.009 0.009
(0.430) (0.499) (0.558) (0.144) (0.116) (0.115)
Log portfolio value (t = −44) −0.001 0.002 0.002 −0.018∗∗∗ −0.020∗∗∗ −0.019∗∗∗ (0.698) (0.266) (0.292) (0.000) (0.000) (0.000)
Length of relationship 0.000 0.000 −0.001 −0.001
(0.673) (0.753) (0.257) (0.257)
Trades per month (t = −44 to t = 0) −0.000 −0.000 0.003∗∗ 0.003∗∗ (0.918) (0.938) (0.042) (0.038)
Portfolio turnover (t = −44 to t = 0) −0.007 −0.002 0.113 0.098
(0.924) (0.977) (0.103) (0.155)
Disposition effect (t = −44 to t = 0) −0.009 −0.009 0.012 0.014
(0.282) (0.287) (0.228) (0.165)
Risky share (t = 0) −0.032∗∗∗ −0.032∗∗∗ 0.038∗∗∗ 0.039∗∗∗ (0.000) (0.000) (0.000) (0.000)
Share of tax-free assets (t = 0) −0.020 −0.019 −0.046∗∗∗ −0.045∗∗ (0.266) (0.301) (0.009) (0.012)
Alpha (t = −44 to t = 0) 0.167 −1.549
(0.901) (0.158)
Constant 0.002 0.006 0.018 0.020 0.072∗∗∗ 0.241∗∗∗ 0.274∗∗∗ 0.267∗∗∗ (0.225) (0.732) (0.437) (0.400) (0.000) (0.000) (0.000) (0.000)
Observations 7,255 5,426 5,403 5,402 5,453 5,450 5,444 5,444
R2 0.016 0.021 0.027 0.027 0.009 0.028 0.039 0.040
(continued)
Downloaded from
http://rfs.oxfordjournals.org/
at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

Is Unbiased Advice Sufficient?
Table 9
Continued
Increase in Sharpe ratio Increase in MPPM
Dependent variable (9) (10) (11) (12) (13) (14) (15) (16)
Dummy advice 0.193∗∗∗ 0.193∗∗∗ 0.189∗∗∗ 0.190∗∗∗ 0.053∗∗∗ 0.057∗∗∗ 0.041∗∗ 0.083∗∗∗ (0.000) (0.000) (0.000) (0.000) (0.003) (0.002) (0.016) (0.000)
Dummy male −0.020∗∗∗ −0.015∗∗∗ −0.015∗∗∗ 0.015 0.002 −0.016
(0.000) (0.006) (0.004) (0.392) (0.908) (0.317)
Age −0.001∗∗∗ −0.000∗ −0.000∗ −0.001 −0.001 −0.001∗∗ (0.004) (0.075) (0.068) (0.245) (0.209) (0.050)
Dummy low wealth 0.013 0.011 0.011 −0.006 −0.001 −0.008
(0.181) (0.246) (0.260) (0.836) (0.962) (0.745)
Dummy high wealth 0.003 0.004 0.004 0.023 0.024 0.028
(0.456) (0.320) (0.276) (0.352) (0.306) (0.197)
Log portfolio value (t = −44) 0.010∗∗∗ 0.003∗ 0.004∗∗ −0.008 −0.018 0.003
(0.000) (0.062) (0.021) (0.460) (0.161) (0.839)
Length of relationship −0.001 −0.001 0.001 0.002
(0.161) (0.195) (0.816) (0.535)
Trades per month (t = −44 to t = 0) 0.011∗∗∗ 0.011∗∗∗ 0.017 0.020∗∗
(0.000) (0.000) (0.147) (0.022)
Portfolio turnover (t = −44 to t = 0) −0.513∗∗∗ −0.548∗∗∗ 1.108∗∗ 0.034
(0.000) (0.000) (0.029) (0.906)
Disposition effect (t = −44 to t = 0) 0.008 0.012 −0.232∗∗∗ −0.110∗∗∗ (0.307) (0.131) (0.001) (0.000)
Risky share (t = 0) 0.008 0.009 0.085 0.107
(0.252) (0.169) (0.215) (0.107)
(continued)
1009
Downloaded from
http://rfs.oxfordjournals.org/
at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

The Review of Financial Studies / v 25 n 4 2012
1010
Table 9
Continued
Increase in Sharpe ratio Increase in MPPM
Dependent variable (9) (10) (11) (12) (13) (14) (15) (16)
Share of tax-free assets (t = 0) 0.064∗∗∗ 0.065∗∗∗ 0.072 0.154∗∗ (0.000) (0.000) (0.396) (0.038)
Alpha (t = −44 to t = 0) −3.510∗∗∗ −93.799∗∗∗ (0.000) (0.000)
Constant 0.199∗∗∗ 0.138∗ ∗ ∗ 0.146∗∗∗ 0.131∗∗∗ 0.330∗∗∗ 0.429∗ ∗ ∗ 0.365∗∗∗ −0.036
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.003) (0.804)
Observations 5,479 5,476 5,470 5,469 5,487 5,484 5,478 5,477
R2 0.079 0.091 0.141 0.148 0.000 0.001 0.019 0.204
Table 9 reports OLS estimates of the coefficients related to a decrease in HHI (models 1 to 4), a decrease in idiosyncratic variance share (models 5 to 8), an increase in Sharpe ratio
(models 9 to 12), and an increase in MPPM (models 13 to 16). HHI and portfolio returns are calculated by using the assumption that investors had fully followed the recommendations
(recommended portfolios). The focus of the table is on the variable Dummy advice that is equal to one if a client opts for financial advice. Additionally, the model controls for several other
independent variables: a dummy that is equal to one if a client is male (Dummy male), the age of a client (Age), a dummy that is equal to one if a client falls into categories one to three of a
microgeographic status rating by an external agency (Dummy low wealth), a dummy that is equal to one if a client falls into categories seven to nine of the microgeographic status (Dummy
high wealth), the risky portfolio value of the customer (Log portfolio value), the number of years the client has been with the bank (Length of relationship), the number of trades per month
(Trades per month), the average portfolio turnover per month (Portfolio turnover), the difference between the proportion of realized gains and losses (Disposition effect), the proportion of
risky assets in the account (Risky share), the proportion of tax-free assets (Share of tax-free assets), and the weekly alpha of a particular customer before opting for financial advice (Alpha).
Alpha and idiosyncratic variance share stem from the application of a Carhart (1997) four-factor model that is calibrated for Germany. P-values are in parentheses. R2 values and number
of observations are reported. *** denotes significance at 1% or less, ** significance at 5% or less, and * significance at 10% or less. Heteroscedasticity robust standard errors are used.
Different counts of observations are due to data availability of certain variables (see Table 5); results are robust to using the lowest common denominator.
Downloaded from
http://rfs.oxfordjournals.org/ at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

Is Unbiased Advice Sufficient?
Table 10
Does the average advisee benefit?
Decrease in HHI Decrease in idiosyncratic variance share
Dependent variable (1) (2) (3) (4) (5) (6) (7) (8)
Dummy advice 0.013 0.011 0.010 0.011 −0.003 0.002 0.001 0.002
(0.144) (0.298) (0.321) (0.304) (0.807) (0.886) (0.907) (0.858)
Dummy male −0.001 −0.000 −0.000 0.005 0.003 0.003
(0.901) (0.961) (0.940) (0.475) (0.637) (0.667)
Age −0.000 −0.000 −0.000 0.000 0.000 0.000
(0.598) (0.619) (0.553) (0.635) (0.647) (0.676)
Dummy low wealth 0.003 0.004 0.004 −0.005 −0.006 −0.006
(0.713) (0.689) (0.691) (0.713) (0.608) (0.603)
Dummy high wealth 0.002 0.001 0.001 0.007 0.008 0.008
(0.681) (0.743) (0.814) (0.189) (0.153) (0.151)
Log portfolio value (t = −44) 0.000 0.003 0.003 −0.017∗∗∗ −0.019∗∗∗ −0.019∗∗∗ (0.899) (0.131) (0.151) (0.000) (0.000) (0.000)
Length of relationship 0.000 0.000 −0.002 −0.002
(0.756) (0.839) (0.126) (0.126)
Trades per month (t = −44 to t = 0) 0.001 0.001 0.003∗∗ 0.003∗∗ (0.390) (0.376) (0.017) (0.015)
Portfolio turnover (t = −44 to t = 0) −0.050 −0.045 0.077 0.062
(0.464) (0.519) (0.266) (0.371)
Disposition effect (t = −44 to t = 0) −0.009 −0.009 0.016 0.017∗ (0.284) (0.283) (0.132) (0.089)
Risky share (t = 0) −0.036∗∗∗ −0.036∗∗∗ 0.035∗∗∗ 0.036∗∗∗ (0.000) (0.000) (0.000) (0.000)
Share of tax-free assets (t = 0) −0.017 −0.015 −0.037∗∗ −0.035∗∗ (0.367) (0.407) (0.039) (0.048)
Alpha (t = −44 to t = 0) 0.243 −1.611
(0.857) (0.144)
Constant 0.002 0.003 0.017 0.019 0.072∗∗∗ 0.232∗∗∗ 0.265∗∗∗ 0.258∗∗∗ (0.225) (0.853) (0.479) (0.432) (0.000) (0.000) (0.000) (0.000)
Observations 7,251 5,422 5,399 5,398 5,453 5,450 5,444 5,444
R2 0.000 0.000 0.007 0.007 0.000 0.017 0.026 0.027
(continued)
1011
Downloaded from
http://rfs.oxfordjournals.org/
at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

The Review of Financial Studies / v 25 n 4 2012
1012
Table 10
Continued
Increase in Sharpe ratio Increase in MPPM
Dependent variable (9) (10) (11) (12) (13) (14) (15) (16)
Dummy advice 0.060∗∗∗ 0.060∗∗∗ 0.056∗∗∗ 0.057∗∗∗ 0.010 0.014 −0.001 0.040∗∗ (0.000) (0.000) (0.000) (0.000) (0.614) (0.473) (0.938) (0.027)
Dummy male −0.021∗∗∗ −0.015∗∗∗ −0.016∗∗∗ 0.015 0.002 −0.016
(0.000) (0.005) (0.003) (0.384) (0.899) (0.326)
Age −0.001∗∗∗ −0.000∗ −0.000∗ −0.001 −0.001 −0.001∗∗ (0.002) (0.056) (0.051) (0.231) (0.198) (0.046)
Dummy low wealth 0.011 0.009 0.008 −0.007 −0.003 −0.010
(0.288) (0.377) (0.395) (0.788) (0.914) (0.699)
Dummy high wealth 0.003 0.004 0.004 0.022 0.024 0.027
(0.513) (0.364) (0.317) (0.365) (0.318) (0.207)
Log portfolio value (t = −44) 0.010∗∗∗ 0.004∗∗ 0.005∗∗ −0.008 −0.018 0.003
(0.000) (0.042) (0.014) (0.449) (0.161) (0.841)
Length of relationship −0.001 −0.001 0.001 0.002
(0.139) (0.169) (0.815) (0.539)
Trades per month (t = −44 to t = 0) 0.011∗∗∗ 0.011∗∗∗ 0.018 0.020∗∗ (0.000) (0.000) (0.143) (0.021)
Portfolio turnover (t = −44 to t = 0) −0.520∗∗∗ −0.555∗∗∗ 1.116∗∗ 0.043
(0.000) (0.000) (0.028) (0.882)
Disposition effect (t = −44 to t = 0) 0.008 0.012 −0.233∗∗∗ −0.112∗∗∗ (0.322) (0.144) (0.001) (0.000)
Risky share (t = 0) 0.004 0.006 0.082 0.104
(0.537) (0.401) (0.234) (0.119)
(continued)
Downloaded from
http://rfs.oxfordjournals.org/
at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

Is Unbiased Advice Sufficient?
Table 10
Continued
Increase in Sharpe ratio Increase in MPPM
Dependent variable (9) (10) (11) (12) (13) (14) (15) (16)
Share of tax-free assets (t = 0) 0.063∗∗∗ 0.065∗∗∗ 0.074 0.156∗∗ (0.000) (0.000) (0.381) (0.035)
Alpha (t = −44 to t = 0) −3.417∗∗∗ −93.756∗∗∗ (0.000) (0.000)
Constant 0.199∗∗∗ 0.138∗∗∗ 0.148∗∗∗ 0.133∗∗∗ 0.330∗∗∗ 0.433∗∗∗ 0.367∗∗∗ −0.034
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.003) (0.813)
Observations 5,479 5,476 5,470 5,469 5,487 5,484 5,478 5,477
R2 0.008 0.021 0.074 0.081 0.000 0.001 0.019 0.203
Table 10 reports OLS estimates of the coefficients related to a decrease in HHI (models 1 to 4) and a decrease in idiosyncratic variance share (models 5 to 8), an increase in Sharpe ratio
(models 9 to 12) and an increase in MPPM (models 13 to 16). HHI and portfolio returns are calculated based on the actual portfolios of investors. The focus of the table is on the variable
Dummy advice that is equal to one if a client opts for financial advice. Additionally, the model controls for the following independent variables: a dummy variable that is equal to one if a
client is male (Dummy male), the age of a client (Age), a dummy that is equal to one if a client falls into categories one to three of a microgeographic status rating by an external agency
(Dummy low wealth), a dummy that is equal to one if a client falls into categories seven to nine of the microgeographic status (Dummy high wealth), the risky portfolio value of the customer
(Log portfolio value), the number of years the client has been with the bank (Length of relationship), the number of trades per month (Trades per month), the average portfolio turnover per
month (Portfolio turnover), the difference between the proportion of realized gains and losses (Disposition effect), the proportion of risky assets in the account (Risky share), the proportion
of tax-free assets (Share of tax-free assets), and the weekly alpha of a particular customer before opting for financial advice (Alpha). Alpha and idiosyncratic variance share stem from the
application of a Carhart (1997) four-factor model that is calibrated for Germany. P-values are in parentheses. R2 values and number of observations are reported. *** denotes significance
at 1% or less, ** significance at 5% or less, and * significance at 10% or less. Heteroscedasticity robust standard errors are used. Different counts of observations are due to data availability
of certain variables (see Table 5); results are robust to using the lowest common denominator.
1013
Downloaded from
http://rfs.oxfordjournals.org/ at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

The Review of Financial Studies / v 25 n 4 2012
Table 11
Does the average advisee benefit if the advice is partially followed?
Decrease in HHI Decrease in idiosyncratic variance share
Dependent variable (1) (2) (3) (4) (5) (6) (7) (8)
Improvement of degree of following 0.136∗∗∗ 0.146∗∗∗ 0.147∗∗∗ 0.147∗∗∗ 0.113 0.103 0.119 0.118
(0.002) (0.003) (0.003) (0.004) (0.122) (0.150) (0.100) (0.104)
Dummy advice 0.012 0.009 0.008 0.009 −0.004 0.001 0.000 0.001
(0.177) (0.392) (0.422) (0.403) (0.727) (0.964) (0.998) (0.947)
Dummy male −0.001 −0.000 −0.000 0.005 0.003 0.003
(0.906) (0.960) (0.939) (0.479) (0.645) (0.675)
Age −0.000 −0.000 −0.000 0.000 0.000 0.000
(0.625) (0.647) (0.580) (0.619) (0.628) (0.657)
Dummy low wealth 0.003 0.004 0.004 −0.004 −0.006 −0.006
(0.699) (0.676) (0.678) (0.720) (0.616) (0.610)
Dummy high wealth 0.002 0.001 0.001 0.007 0.008 0.008
(0.709) (0.771) (0.843) (0.195) (0.158) (0.157)
Log portfolio value (t = −44) 0.000 0.003 0.003 −0.017∗∗∗ −0.019∗∗∗ −0.019∗∗∗ (0.876) (0.126) (0.146) (0.000) (0.000) (0.000)
Length of relationship 0.000 0.000 −0.002 −0.002
(0.772) (0.855) (0.121) (0.121)
Trades per month (t = −44 to t = 0) 0.001 0.001 0.003∗∗ 0.003∗∗ (0.354) (0.341) (0.016) (0.014)
Portfolio turnover (t = −44 to t = 0) −0.050 −0.045 0.077 0.062
(0.463) (0.519) (0.266) (0.370)
Disposition effect (t = −44 to t = 0) −0.009 −0.009 0.015 0.017∗
(0.272) (0.270) (0.137) (0.093)
Risky share (t = 0) −0.036∗∗∗ −0.036∗∗∗ 0.036∗∗∗ 0.036∗∗∗ (0.000) (0.000) (0.000) (0.000)
Share of tax-free assets (t = 0) −0.017 −0.016 −0.037∗∗ −0.036∗∗ (0.342) (0.380) (0.036) (0.044)
Alpha (t = −44 to t = 0) 0.259 −1.603
(0.848) (0.147)
Constant 0.002 0.002 0.017 0.019 0.072∗∗∗ 0.232∗∗∗ 0.265∗∗∗ 0.258∗∗∗ (0.225) (0.890) (0.478) (0.430) (0.000) (0.000) (0.000) (0.000)
Observations 7,251 5,422 5,399 5,398 5,453 5,450 5,444 5,444
R2 0.001 0.002 0.008 0.009 0.001 0.017 0.027 0.028
(continued)
1014
Downloaded from
http://rfs.oxfordjournals.org/
at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

Is Unbiased Advice Sufficient?
Table 11
Continued
Increase in Sharpe ratio Increase in MPPM
Dependent variable (9) (10) (11) (12) (13) (14) (15) (16)
Improvement of degree of following 0.193∗∗∗ 0.197∗∗∗ 0.187∗∗∗ 0.185∗∗∗ 0.018 0.006 0.085 0.026
(0.001) (0.001) (0.001) (0.002) (0.752) (0.913) (0.199) (0.645)
Dummy advice 0.058∗∗∗ 0.058∗∗∗ 0.054∗∗∗ 0.055∗∗∗ 0.010 0.014 −0.002 0.040∗∗ (0.000) (0.000) (0.000) (0.000) (0.622) (0.477) (0.898) (0.030)
Dummy male −0.021∗∗∗ −0.015∗∗∗ −0.016∗∗∗ 0.015 0.002 −0.016
(0.000) (0.004) (0.003) (0.384) (0.902) (0.325)
Age −0.001∗∗∗ −0.000∗ −0.000∗ −0.001 −0.001 −0.001∗∗ (0.003) (0.063) (0.058) (0.232) (0.200) (0.047)
Dummy low wealth 0.011 0.009 0.009 −0.007 −0.003 −0.010
(0.278) (0.365) (0.383) (0.789) (0.917) (0.700)
Dummy high wealth 0.003 0.004 0.004 0.022 0.024 0.027
(0.539) (0.383) (0.334) (0.365) (0.320) (0.207)
Log portfolio value (t = −44) 0.010∗∗∗ 0.004∗∗ 0.005∗∗ −0.008 −0.018 0.003
(0.000) (0.039) (0.013) (0.449) (0.161) (0.840)
Length of relationship −0.001 −0.001 0.001 0.002
(0.128) (0.155) (0.818) (0.540)
Trades per month (t = −44 to t = 0) 0.011∗∗∗ 0.011∗∗∗ 0.018 0.020∗∗ (0.000) (0.000) (0.142) (0.021)
Portfolio turnover (t = −44 to t = 0) −0.520∗∗∗ −0.555∗∗∗ 1.116∗∗ 0.043
(0.000) (0.000) (0.028) (0.882)
Disposition effect (t = −44 to t = 0) 0.008 0.012 −0.233∗∗∗ −0.112∗∗∗ (0.342) (0.156) (0.001) (0.000)
Risky share (t =0) 0.004 0.006 0.082 0.104
(0.522) (0.389) (0.234) (0.118)
Share of tax-free assets (t = 0) 0.062∗∗∗ 0.064∗∗∗ 0.074 0.156∗∗ (0.000) (0.000) (0.384) (0.036)
(continued)
1015
Downloaded from
http://rfs.oxfordjournals.org/
at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

The Review of Financial Studies / v 25 n 4 2012
1016
Table 11
Continued
Increase in Sharpe ratio Increase in MPPM
Dependent variable (9) (10) (11) (12) (13) (14) (15) (16)
Alpha (t = −44 to t = 0) −3.405∗∗∗ −93.754∗∗∗ (0.000) (0.000)
Constant 0.199∗∗∗ 0.137∗∗∗ 0.148∗∗∗ 0.133∗∗∗ 0.330∗∗∗ 0.433∗∗∗ 0.367∗∗∗ −0.034
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.003) (0.813)
Observations 5,479 5,476 5,470 5,469 5,487 5,484 5,478 5,477
R2 0.011 0.023 0.077 0.084 0.000 0.001 0.019 0.203
Table 11 reports OLS estimates of the coefficients related to a decrease in HHI (models 1 to 4) and a decrease in idiosyncratic variance share (models 5 to 8), an increase in Sharpe ratio
(models 9 to 12) and an increase in MPPM (models 13 to 16). HHI and portfolio returns are calculated based on the actual portfolios of investors. The focus of the table is on the variable
Improvement of degree of following, which shows the improvement of the degree of following from the time of the first recommendation to t = 11 and takes the value zero for non-advised
clients. Dummy advice is equal to one if a client opts for financial advice. Additionally, the model controls for the following independent variables: a dummy variable that is equal to one if
a client is male (Dummy male), the age of a client (Age), a dummy that is equal to one if a client falls into categories one to three of a microgeographic status rating by an external agency
(Dummy low wealth), a dummy that is equal to one if a client falls into categories seven to nine of the microgeographic status (Dummy high wealth), the risky portfolio value of the customer
(Log portfolio value), the number of years the client has been with the bank (Length of relationship), the number of trades per month (Trades per month), the average portfolio turnover per
month (Portfolio turnover), the difference between the proportion of realized gains and losses (Disposition effect), the proportion of risky assets in the account (Risky share), the proportion
of tax-free assets (Share of tax-free assets), and the weekly alpha of a particular customer before opting for financial advice (Alpha). Alpha and idiosyncratic variance share stem from the
application of a Carhart (1997) four-factor model that is calibrated for Germany. P-values are in parentheses. R2 values and number of observations are reported. *** denotes significance
at 1% or less, ** significance at 5% or less, and * significance at 10% or less. Heteroscedasticity robust standard errors are used. Different counts of observations are due to data availability
of certain variables (see Table 5); results are robust to using the lowest common denominator.
Downloaded from
http://rfs.oxfordjournals.org/ at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

Is Unbiased Advice Sufficient?
7. Who Would Most Benefit from the Advice?
We now explore an important public policy question: are the individuals who
are most likely to benefit from financial advice the ones least likely to obtain
the advice, and are the persons least likely to benefit from financial advice most
likely to obtain and follow the advice? Section 4 indicates that this seems to
be the case; the evidence showed that the more (less) financially sophisticated
investors are, the more (less) likely they are to obtain advice.
We revisit the probit regression that was run in order to generate the results
in Table 7, i.e., the test that investigates those who elect to receive financial
advice. We take probit model 5 in Table 7 because it has the highest pseudo-
R 2 value and can be estimated for 7,432 customers. We then use the coefficient
estimates from this regression to predict the 5% of the 7,432 clients with
the highest probability of obtaining the advice. 17 We define those 5% (372
investors) as “predicted to obtain advice” and the remaining 95% (7,060
investors) as “not predicted to obtain advice.” Of the former group, sixty-two
opted for advice and 310 did not. Of the latter group, 307 opted for advice and
6,753 did not. Thus, we can test whether those persons predicted to be less
likely to obtain the advice are the ones who benefit more from the advice (i.e.,
whether the 307 benefited more than the sixty-two).
The number of observations in these subgroups is lower than 372 and 7,060
because of data availability. We now run the same regression that generated
the results in Table 9 for both these subgroups. The first subgroup consists of
clients who are predicted to obtain advice. Group size is 372. Of these, sixtytwo
took advice and have the indicator variable “dummy advice” turned on to
equal one, and 310 did not take advice and have the indicator variable “dummy
advice” turned on to equal zero. The second subgroup is clients who are not
predicted to obtain advice. The group size is 7,060. Of these, 307 took advice
and have the indicator variable “dummy advice” turned on to equal one, and
6,753 did not take advice and have the indicator variable “dummy advice”
turned on to equal zero. Table 12 reports the results for these regressions.
The coefficient on the “dummy advice” variable is 0.042 (positive and
statistically significant) in the HHI regression for the clients who are predicted
to opt for the advice and did opt for it. The coefficient on the “dummy advice”
variable is 0.099 (positive and statistically significant) in the HHI regression
for the clients who are predicted to not obtain the advice but did obtain
it. The difference between these two coefficients is positive and statistically
significant.
The coefficient on the “dummy advice” variable is 0.066 (positive and
statistically significant) in the idiosyncratic risk regression for the clients who
are predicted to opt for the advice and did opt for it. The coefficient on the
“dummy advice” variable is 0.077 (positive and statistically significant) in the
17 Results are robust to different cutoff points, including 10%, 20%, or even 50%, with highest likelihood to opt
for advice.
1017
Downloaded from
http://rfs.oxfordjournals.org/ at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

The Review of Financial Studies / v 25 n 4 2012
1018
Table 12
Who would benefit most from advice?
Decrease in HHI Decrease in idiosyncratic variance share
Predicted to Not predicted to Predicted to Not predicted to
Dependent variable obtain advice obtain advice P-value obtain advice obtain advice P-value
Dummy advice 0.042∗∗∗ 0.099∗∗∗ .00 0.066∗∗ 0.077∗∗∗ .73
(0.000) (0.000) (0.013) (0.000)
Constant −0.007∗ 0.003 0.112∗∗∗ 0.070∗∗∗ (0.089) (0.135) (0.000) (0.000)
Observations 366 6,859 359 5,085
R2 0.056 0.017 0.018 0.007
Increase in Sharpe ratio Increase in MPPM
Predicted to Not predicted to Predicted to Not predicted to
Dependent variable obtain advice obtain advice P-value obtain advice obtain advice P-value
Dummy advice 0.147∗∗∗ 0.194∗∗∗ .00 −0.034 0.053∗∗∗ .18
(0.000) (0.000) (0.579) (0.006)
Constant 0.250∗∗∗ 0.196∗∗∗ 0.439∗∗∗ 0.324∗∗∗ (0.000) (0.000) (0.000) (0.000)
Observations 359 5,110 359 5,118
R2 0.179 0.069 0.000 0.000
Table 12 reports coefficient estimates of the variable “dummy advice” when the tests of Table 9 are conducted on subgroups. The first subgroup consists of clients who are predicted to
obtain advice. These are 5% of the 7,432 clients with the highest probability of obtaining the advice predictions as given in model 5 of Table 7. Their size is 372. Of these, sixty-two took
advice and so have the indicator variable “dummy advice” turned on to equal one, and 310 did not take advice and so have the indicator variable “dummy advice” turned on to equal zero.
The second subgroup consists of clients who are not predicted to obtain advice. These are 95% of the 7,432 clients with the lowest probability of obtaining the advice predictions as given
in model 5 of Table 7. Their size is 7,060. Of these, 307 took advice and so have the indicator variable “dummy advice” turned on to equal one, and 6,753 did not take advice and so have
the indicator variable “dummy advice” turned on to equal zero. R2 values and number of observations are reported. *** denotes significance at 1% or less, ** significance at 5% or less,
and * significance at 10% or less. Heteroscedasticity robust standard errors are used. The third column reports the p-value of a test of equality of the coefficients for Dummy advice using
seemingly unrelated estimation. Different counts of observations are due to data availability of certain variables (see Table 5); results are robust to using the lowest common denominator.
Downloaded from
http://rfs.oxfordjournals.org/ at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

Is Unbiased Advice Sufficient?
idiosyncratic risk regression for the clients who are predicted to not obtain
the advice but did obtain it. The difference between these two coefficients is
positive but not statistically significant.
The coefficient on the “dummy advice” variable is 0.147 (positive and
statistically significant) in the Sharpe ratio regression for the clients who are
predicted to opt for the advice and did opt for it. The coefficient on the “dummy
advice” variable is 0.194 (positive and statistically significant) in the Sharpe
ratio regression for the clients who are predicted to not obtain the advice but
did obtain it. The difference between these two coefficients is positive and
statistically significant.
The coefficient on the “dummy advice” variable is -0.034 (negative but
not statistically significant) in the MPPM regression for the clients who are
predicted to opt for the advice and did opt for it. The coefficient on the “dummy
advice” variable is 0.053 (positive and statistically significant) in the MPPM
regression for the clients who are predicted to not obtain the advice but did
obtain it. The difference between these two coefficients is positive but not
statistically significant.
The important result here is that the decrease in HHI or the increase in the
Sharpe ratio is significantly greater for clients predicted to not obtain the advice
but who obtained it than for clients who are predicted to obtain the advice and
did obtain it. As seen in Section 4, the former group includes the investors who
are less financially sophisticated than are the latter group. We come to the same
conclusion: those who most need (do not need) the financial advice are least
(most) likely to obtain it.
8. Conclusion
Can unbiased financial advice steer retail investors toward efficient portfolios?
To answer this question, we work with one of the biggest brokerages in
Germany and offer advice that is unbiased and theoretically sound. First, we
find that those who accept the offer (5%) are more likely to be male, older,
wealthier, more financially sophisticated, and have a longer relationship with
the brokerage. Second, of those who accept the offer, the advice is hardly
followed. Third, though the average advisee’s portfolio efficiency hardly
improves, the average advisee who follows the advice does see an increase
in portfolio efficiency. Fourth, it seems that investors who most need the
financial advice are least likely to obtain it. Overall, our results imply that the
mere availability of unbiased financial advice is a necessary but not sufficient
condition for benefiting retail customers. As the adage goes, you can lead a
horse to water, but you can’t make it drink.
At a time when protecting financial consumers has risen to the top of
the regulatory agenda in many countries, the results of this article provide a
basis for skepticism of supply-side solutions imposed by regulators. However,
examining the demand side of financial advice raises more questions than
1019
Downloaded from
http://rfs.oxfordjournals.org/ at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

The Review of Financial Studies / v 25 n 4 2012
answers. We learn that even honest and sound financial services are useless,
unless the customer actually follows them. In addition, it seems even more
difficult to reach people who most need financial advice. Thus, if financial
economists are to develop remedies to correct widespread investment mistakes
made by retail investors, this study indicates that much additional research
must be done to answer the question of why people do or do not follow
advice. This must be taken into account when trying to help people make
asset allocation decisions. A promising start has been made by Benartzi and
Thaler (2004, 2007) who, after taking behavioral factors into account, develop
sophisticated savings plans.
Experimenting with ways to offer advice is another useful avenue to explore.
For example, in our study, the advice would require people to turn over, on
average, 75% of their portfolios, since investors’ existing portfolios are largely
inefficient. Investors in our sample may have found it too complicated or too
cumbersome to implement the full list of recommendations, though they did
turn over 70% of their portfolios every year during the pre-advice period. Although
the information given to our advisees is extensive and clear, it may not
be much different from other, less theoretically anchored sources of investment
advice. Perhaps future settings could therefore seek to build greater trust with
advisees. 18 To conclude, much more needs to be done to understand why and
how financial advice is actually followed. What makes the horse drink?
Appendix A: Disguised Example of Advice That Was Sent to Advisees
1. Description of the idea of diversification, explanation of important concepts, intuitive
explanation of the portfolio optimization methodology, and discussion of tax implications
2. Analysis of the existing portfolio
3. Analysis of recommended (optimized) portfolio 19
18 Bonaccio and Dalal (2006) conduct a literature review to document the determinants of all effective advice, not
just financial advice. A variety of characteristics of the advice are examined, including the distance of the advice
from the original opinion (Yaniv 2004); whether the advice was paid for (Gino 2008); whether advice is didactic
or just offers information about choices (Bonaccio and Dalal 2010); and many other aspects. Statman (2010)
explores what investors really want.
19 Recall that the goal of the optimizer is to enhance portfolio efficiency by improving diversification. This may
improve portfolios along two dimensions: either the risk for a given level of return decreases or the expected
return for a given level of risk increases. The brokerage communicated in this report changes along both
dimensions.
1020
Downloaded from
http://rfs.oxfordjournals.org/ at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

Is Unbiased Advice Sufficient?
Model calculation of possible portfolio development in 5 years
Lower upper
Current value Expected value bound (5%) bound (95%)
Existing portfolio
Recommended (optimized)
1,000 e 1,700 e 1,100 e 2,500 e
portfolio 1,000 e 2,200 e 1,000 e 4,100 e
Model calculation assumes constant riek/return estimates
4. The client’s investment requests
5. List of necessary transactions
6. Fact sheet for each security on recommendation list
Appendix B: Measure for the Degree of Following
Our degree of following measures both the buy and sell sides of following advice. If a security
in the advisee’s portfolio is not included in the recommended portfolio, this suggests that the
investor has been advised to sell this security. It is, however, possible that the advisee sells a
security for reasons other than following the advice (e.g., liquidity or tax motives). As a check for
robustness, we construct an alternative measure of the degree of following. This measure considers
only the buy side and is simply the share of the recommended portfolio that the investor holds at
any time.
The tables below that use this alternative measure of the degree of following (Tables A1–A4)
are the counterparts of the tables in the text (Table 4, Table 6, Table 8, and Table 11).
The figures below that use this alternative measure of the degree of following (Figure A1,
panels A–C) are the counterparts of the figures in the text (Figure 2, panels A–C).
1021
Downloaded from
http://rfs.oxfordjournals.org/ at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

The Review of Financial Studies / v 25 n 4 2012
1022
Table A1
Illustration of the alternative degree of following (buy side only). Counterpart of Table 4
Required Keep Buy
Original portfolio (t = 0) e Recommended portfolio (t = 0) e action e e
⎫
Deutsche Bank 8,646 → Sell
HSBC Indian Equity Fund 3,622 → Sell ⎪⎬
Ignored for
Raiffeisen CEE Equity Fund 2,792 → Sell measuring
HSBC BRIC Equity Fund 1,862 → Sell ⎪⎭ alternative “degree
Commerzbank 439 → Sell of following advice”
⎫
E.ON 2,523 E.ON 681 → Keep/Decrease 681
H&M 3,543 H&M 3,543 → Keep 3,543
Comstage ETF EONIA 4,072 → Buy 4,072
Schroder ISF Europa Corporate Bond Fund 3,883 → Buy 3,883⎪⎬ Considered for
Allianz Pimco Europazins Bond Fund 3,799 → Buy 3,799 measuring
Allianz Pimco Corporate Bond Europa Fund 2,434 → Buy 2,434 alternative “degree
UBS Lux Bond Fund 1,751 → Buy 1,751 of following advice”
Grundbesitz Europa Real Estate Fund 1,470 → Buy 1,470
Pictet EM Fund 1,247 → Buy 1,247⎪⎭
Allianz RCM Small Cap Fund 808 → Buy 808
Total 23,426 23,688 4,224 19,463
Table A1 provides an actual example of our alternative measure of degree of following, which considers only the assets the advisee is recommended to buy. Securities E.ON and H&M
overlap between the actual and the recommended portfolio (E.ON only partially). The value of the overlap in this case is e4,224 (e681 in E.ON and e3,543 in H&M). We calculate the
degree of following for each day as the euro-value of recommended assets in the investor’s original portfolio divided by the value of assets in euros in the actual portfolio (here: 4,224 /
23,688 = 18%).
Downloaded from
http://rfs.oxfordjournals.org/
at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

Is Unbiased Advice Sufficient?
Table A2
Summary statistics of the alternative degree of following (buy side only). Counterpart of Table 6
t = 11 months Average
(end of between t =
measurement advice start and
t = advice start t = 10 days t = 20 days t = 30 days period) t = 11 months
Degree of following
Mean (%) 25.0 32.7 36.0 37.8 26.3 32.9
Median (%) 21.4 26.3 27.4 28.9 21.0 28.1
Number of followers n/a 99 142 169 148 175
Observations 385 385 385 385 385 385
Table A2 reports summary statistics of the alternative degree of following. The table reports mean alternative degree of following, median alternative degree of following, number of
followers, and observations for six distinct time periods (upon receiving the first recommendation, after ten days, after twenty days, and at the end of the post-advice period and the average
over the entire post-advice period). These time intervals are specific to each investor in the sample. Number of followers is defined as the number of investors who increase their alternative
degree of following, at least marginally.
1023
Downloaded from
http://rfs.oxfordjournals.org/ at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

The Review of Financial Studies / v 25 n 4 2012
Table A3
Who chooses to follow advice? (buy side only) Counterpart of Table 8
1024
Increase in alternative degree of following (buy side only)
Dependent variable (1) (2) (3) (4) (5) (6) (7)
Dummy male −0.063 −0.062 −0.041 −0.049 −0.066 −0.065 −0.028
(0.223) (0.226) (0.444) (0.361) (0.197) (0.203) (0.605)
Age 0.000 0.000 0.000 0.000 0.000 0.000 0.000
(0.619) (0.746) (0.896) (0.871) (0.807) (0.825) (0.664)
Dummy low wealth −0.048∗ −0.061∗∗ −0.071∗∗ −0.066∗∗ −0.059∗ −0.058∗ −0.062∗∗ (0.075) (0.046) (0.016) (0.033) (0.051) (0.064) (0.048)
Dummy high wealth 0.046∗ 0.042 0.031 0.031 0.042 0.041 0.030
(0.073) (0.114) (0.266) (0.259) (0.114) (0.121) (0.305)
Log portfolio value (t =0) −0.068∗∗∗ −0.069∗∗∗ −0.065∗∗∗ −0.061∗∗∗ −0.068∗∗∗ −0.070∗∗∗ −0.059∗∗∗ (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
Length of relationship 0.002 0.004 0.004 0.001 0.002 0.003
(0.685) (0.454) (0.509) (0.786) (0.712) (0.635)
Trades per month (t = −44 to t =0) 0.002 0.002 0.003 0.004 0.003 0.001
(0.639) (0.729) (0.482) (0.472) (0.537) (0.804)
Portfolio turnover (t = −44 to t =0) −0.151 −0.017 −0.169 −0.198 −0.176 −0.116
(0.590) (0.955) (0.569) (0.504) (0.537) (0.800)
Disposition effect (t = −44 to t =0) −0.007 0.016 0.033 −0.006 −0.010 0.062
(0.899) (0.782) (0.563) (0.917) (0.853) (0.288)
Risky share (t =0) −0.016 −0.036 −0.041 −0.021 −0.014 −0.013
(0.764) (0.513) (0.461) (0.712) (0.787) (0.837)
Share of tax-free assets (t =0) 0.008 0.058 0.092 0.017 0.013 0.080
(0.910) (0.310) (0.160) (0.816) (0.854) (0.141)
Recommended portfolio over original portfolio 0.001 0.000 0.000 −0.002 0.000 0.026
(0.951) (0.985) (0.992) (0.904) (0.987) (0.245)
Alpha (t = −44 to t =0) 7.230
(0.161)
Idiosyncratic variance share (t = −44 to t =0) 0.100
(0.202)
HHI (t =0) 0.099
(0.483)
Home bias (t =0) 0.030
(0.470)
(continued)
Downloaded from
http://rfs.oxfordjournals.org/
at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

Is Unbiased Advice Sufficient?
Table A3
Continued
Increase in alternative degree of following (buy side only)
Dependent variable (1) (2) (3) (4) (5) (6) (7)
Long-term raw return (t = −44 to t = −1) 8.582
(0.208)
Short-term raw return (t = −1 to t = 0) −1.995∗ (0.093)
Constant 0.842∗∗∗ 0.859∗∗∗ 0.759∗∗∗ 0.656∗∗∗ 0.842∗∗∗ 0.859∗∗∗ 0.668∗∗∗ (0.000) (0.000) (0.000) (0.002) (0.000) (0.000) (0.000)
Observations 369 365 313 313 365 365 297
R2 0.118 0.115 0.126 0.127 0.118 0.117 0.123
Table A3 reports OLS estimates of the coefficients related to an increase in the alternative degree of following measure from the date the investor received advice to t = 20 days (Increase
in the alternative degree of following). For the estimation of the model, we include the following independent variables: a dummy that is equal to one if a client is male (Dummy male), the
age of a client (Age), a dummy that is equal to one if a client falls into categories one to three of a microgeographic status rating by an external agency (Dummy low wealth), a dummy that
is equal to one if a client falls into categories seven to nine of the microgeographic status (Dummy high wealth), the risky portfolio value of the customer (Log portfolio value), the number
of years the client has been with the bank (Length of relationship), the number of trades per month (Trades per month), the average portfolio turnover per month (Portfolio turnover), the
difference between the proportion of realized gains and losses (Disposition effect), the proportion of risky assets in the account (Risky share), the proportion of tax-free assets (Share of
tax-free assets), a ratio for the required net investment (Recommended portfolio/original portfolio), the weekly alpha of a particular customer before opting for financial advice (Alpha),
the idiosyncratic variance share (Idiosyncratic variance share), the Herfindahl-Hirschman index (HHI), the share of domestic equity (Home bias), the raw return from the beginning of
observation until one month prior to the offer of financial advice (Long-term raw return), and the raw return of the month before the offer of financial advice (Short-term raw return).
Alpha and idiosyncratic risk share stem from the application of a Carhart (1997) four-factor model that is calibrated for Germany. P-values are in parentheses. R2 values and number of
observations are reported. *** denotes significance at 1% or less, ** significance at 5% or less, and * significance at 10% or less. Heteroscedasticity robust standard errors are used. Standard
errors shown are not clustered, but results remain qualitatively unaltered when clustering them by advice week or risk aversion. Different counts of observations are due to data availability
of certain variables (see Table 5); results are robust to using the lowest common denominator.
1025
Downloaded from
http://rfs.oxfordjournals.org/ at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

The Review of Financial Studies / v 25 n 4 2012
Table A4
Does the average advisee benefit if partially followed? (buy side only) Counterpart of Table 11
Decrease in HHI Decrease in idiosyncratic variance share
1026
Dependent variable (1) (2) (3) (4) (5) (6) (7) (8)
Improvement of alternative degree of following 0.176∗∗∗ 0.197∗∗∗ 0.196∗∗∗ 0.196∗∗∗ 0.139∗∗ 0.134∗∗ 0.149∗∗ 0.148∗∗ (0.001) (0.001) (0.001) (0.001) (0.025) (0.029) (0.016) (0.017)
Dummy advice 0.011 0.007 0.006 0.007 −0.005 −0.001 −0.001 −0.000
(0.218) (0.506) (0.543) (0.521) (0.661) (0.965) (0.925) (0.976)
Dummy male −0.001 −0.000 −0.001 0.005 0.003 0.003
(0.874) (0.924) (0.903) (0.492) (0.662) (0.693)
Age −0.000 −0.000 −0.000 0.000 0.000 0.000
(0.626) (0.649) (0.582) (0.621) (0.630) (0.658)
Dummy low wealth 0.004 0.004 0.004 −0.004 −0.006 −0.006
(0.688) (0.666) (0.668) (0.724) (0.619) (0.613)
Dummy high wealth 0.001 0.001 0.001 0.007 0.008 0.008
(0.738) (0.799) (0.872) (0.203) (0.165) (0.164)
Log portfolio value (t = −44) 0.000 0.003 0.003 −0.017∗∗∗ −0.019∗∗∗ −0.019∗∗∗ (0.878) (0.131) (0.152) (0.000) (0.000) (0.000)
Length of relationship 0.000 0.000 −0.002 −0.002
(0.777) (0.861) (0.119) (0.119)
Trades per month (t = −44 to t =0) 0.001 0.001 0.003∗∗ 0.003∗∗ (0.321) (0.310) (0.014) (0.013)
Portfolio turnover (t = −44 to t =0) −0.051 −0.045 0.077 0.061
(0.456) (0.513) (0.269) (0.373)
Disposition effect (t = −44 to t =0) −0.009 −0.009 0.015 0.017∗ (0.268) (0.265) (0.140) (0.095)
Risky share (t =0) −0.035∗∗∗ −0.036∗∗∗ 0.036∗∗∗ 0.036∗∗∗ (0.000) (0.000) (0.000) (0.000)
Share of tax-free assets (t =0) −0.018 −0.017 −0.038∗∗ −0.036∗∗ (0.326) (0.362) (0.034) (0.042)
Alpha (t = −44 to t =0) 0.269 −1.596
(0.842) (0.149)
Constant 0.002 0.003 0.018 0.020 0.072∗∗∗ 0.232∗∗∗ 0.265∗∗∗ 0.258∗∗∗ (0.225) (0.879) (0.456) (0.408) (0.000) (0.000) (0.000) (0.000)
Observations 7,251 5,422 5,399 5,398 5,453 5,450 5,444 5,444
R2 0.003 0.005 0.011 0.011 0.001 0.018 0.028 0.028
(continued)
Downloaded from
http://rfs.oxfordjournals.org/
at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

Is Unbiased Advice Sufficient?
Table A4
Continued
Increase in Sharpe ratio Increase in MPPM
Dependent variable (9) (10) (11) (12) (13) (14) (15) (16)
Improvement of alternative degree of following 0.131∗∗∗ 0.134∗∗∗ 0.128∗∗ 0.127∗∗ 0.027 0.018 0.088 0.031
(0.007) (0.008) (0.012) (0.014) (0.625) (0.743) (0.145) (0.553)
Dummy advice 0.058∗∗∗ 0.058∗∗∗ 0.054∗∗∗ 0.055∗∗∗ 0.009 0.014 −0.003 0.040∗∗ (0.000) (0.000) (0.000) (0.000) (0.631) (0.484) (0.876) (0.031)
Dummy male −0.021∗∗∗ −0.015∗∗∗ −0.016∗∗∗ 0.015 0.002 −0.016
(0.000) (0.004) (0.003) (0.385) (0.906) (0.324)
Age −0.001∗∗∗ −0.000∗ −0.000∗ −0.001 −0.001 −0.001∗∗ (0.003) (0.059) (0.054) (0.232) (0.199) (0.047)
Dummy low wealth 0.011 0.009 0.008 −0.007 −0.003 −0.010
(0.280) (0.368) (0.385) (0.789) (0.917) (0.700)
Dummy high wealth 0.003 0.004 0.004 0.022 0.024 0.027
(0.545) (0.389) (0.339) (0.365) (0.321) (0.208)
Log portfolio value (t = −44) 0.010∗∗∗ 0.004∗∗ 0.005∗∗ −0.008 −0.018 0.003
(0.000) (0.041) (0.014) (0.449) (0.161) (0.841)
Length of relationship −0.001 −0.001 0.001 0.002
(0.130) (0.159) (0.818) (0.540)
Trades per month (t = −44 to t =0) 0.011∗∗∗ 0.011∗∗∗ 0.018 0.020∗∗ (0.000) (0.000) (0.142) (0.021)
Portfolio turnover (t = −44 to t =0) −0.521∗∗∗ −0.555∗∗∗ 1.116∗∗ 0.043
(0.000) (0.000) (0.028) (0.882)
Disposition effect (t = −44 to t =0) 0.008 0.012 −0.233∗∗∗ −0.112∗∗∗ (0.338) (0.154) (0.001) (0.000)
Risky share (t =0) 0.005 0.006 0.082 0.104
(0.505) (0.375) (0.233) (0.118)
(continued)
1027
Downloaded from
http://rfs.oxfordjournals.org/
at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

The Review of Financial Studies / v 25 n 4 2012
1028
Table A4
Continued
Increase in Sharpe ratio Increase in MPPM
Dependent variable (9) (10) (11) (12) (13) (14) (15) (16)
Share of tax-free assets (t =0) 0.063∗∗∗ 0.064∗∗∗ 0.074 0.156∗∗ (0.000) (0.000) (0.385) (0.036)
Alpha (t = −44 to t =0) −3.405∗∗∗ −93.753∗∗∗ (0.000) (0.000)
Constant 0.199∗∗∗ 0.138∗∗∗ 0.149∗∗∗ 0.133∗∗∗ 0.330∗∗∗ 0.433∗∗∗ 0.367∗∗∗ −0.034
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.003) (0.814)
Observations 5,479 5,476 5,470 5,469 5,487 5,484 5,478 5,477
R2 0.010 0.022 0.076 0.083 0.000 0.001 0.019 0.203
Table A4 reports OLS estimates of the coefficients related to a decrease in HHI (models 1 to 4) and a decrease in idiosyncratic variance share (models 5 to 8), an increase in Sharpe ratio
(models 9 to 12) and an increase in MPPM (models 13 to 16). HHI and portfolio returns are calculated based on the actual portfolios of investors. The focus of the table is on the variable
Improvement of the alternative degree of following, which shows the improvement of the alternative degree of following from the time of the first recommendation to t = 11 and takes the
value zero for non-advised clients. Dummy advice is equal to one if a client opts for financial advice. Additionally, the model controls for the following independent variables: a dummy
variable that is equal to one if a client is male (Dummy male), the age of a client (Age), a dummy that is equal to one if a client falls into categories one to three of a microgeographic status
rating by an external agency (Dummy low wealth), a dummy that is equal to one if a client falls into categories seven to nine of the microgeographic status (Dummy high wealth), the risky
portfolio value of the customer (Log portfolio value), the number of years the client has been with the bank (Length of relationship), the number of trades per month (Trades per month), the
average portfolio turnover per month (Portfolio turnover), the difference between the proportion of realized gains and losses (Disposition effect), the proportion of risky assets in the account
(Risky share), the proportion of tax-free assets (Share of tax-free assets), and the weekly alpha of a particular customer before opting for financial advice (Alpha). Alpha and idiosyncratic
variance share stem from the application of a Carhart (1997) four-factor model that is calibrated for Germany. P-values are in parentheses. R2 values and number of observations are
reported. *** denotes significance at 1% or less, ** significance at 5% or less, and * significance at 10% or less. Heteroscedasticity robust standard errors are used. Different counts of
observations are due to data availability of certain variables (see Table 5); results are robust to using the lowest common denominator.
Downloaded from
http://rfs.oxfordjournals.org/ at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

Is Unbiased Advice Sufficient?
Figure A1
Initial alternative degree of following (buy side only). Counterparts of Figure 2
Panel A: The cross-sectional distribution of the degree of following when the first recommendation was given.
Panel B: The cross-sectional distribution of the average degree of following between time t = advice start and
t = 11 months. Panel C: Increase in degree of following from the time of the first recommendation to t = 11.
References
Agarwal, S., J. Driscoll, X. Gabaix, and D. Laibson. 2009. The Age of Reason: Financial Decisions over the
Life-cycle with Implications for Regulation. Brooking Papers on Economic Activity Fall:51–117.
Agnew, J., P. Balduzzi, and A. Sundén. 2003. Portfolio Choice and Trading in a Large 401(k) Plan. American
Economic Review 93:193–215.
Ahearne, A. G., W. L. Griever, and F. E. Warnock. 2004. Information Costs and Home Bias: An Analysis of US
Holdings of Foreign Securities. Journal of International Economics 62:331–36.
Barber, B. M., Y.-T. Lee, Y.-J. Liu, and T. Odean. 2009. Just How Much Do Individual Investors Lose by
Trading? Review of Financial Studies 22:609–32.
Barber, B. M., R. Lehavy, M. McNichols, and B. Trueman. 2001. Can Investors Profit from the Prophets?
Security Analysts Recommendations and Stock Returns. Journal of Finance 56:531–63.
Barber, B. M., and T. Odean. 2000. Trading Is Hazardous to Your Wealth: The Common Stock Investment
Performance of Individual Investors. Journal of Finance 55:773–806.
———. 2001. Boys Will Be Boys: Gender, Overconfidence, and Common Stock Investment. Quarterly Journal
of Economics 116:261–92.
Barras, L., O. Scaillet, and R. Wermers. 2010. False Discoveries in Mutual Fund Performance: Measuring Luck
in Estimated Alphas. Journal of Finance 65:179–216.
Benartzi, S., and R. Thaler. 2004. Save More Tomorrow: Using Behavioral Economics to Increase Employee
Saving. Journal of Political Economy 112:164–87.
1029
Downloaded from
http://rfs.oxfordjournals.org/ at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

The Review of Financial Studies / v 25 n 4 2012
———. 2007. Heuristics and Biases in Retirement Savings Behavior. Journal of Economic Perspectives 21:
81–104.
Bergstresser, D., J. M. R. Chalmers, and P. Tufano. 2009. Assessing the Costs and Benefits of Brokers in the
Mutual Fund Industry. Review of Financial Studies 22:4129–4156.
Bertrand, M., and A. Morse. 2009. Information Disclosure, Cognitive Biases, and Payday Borrowing. Working
Paper, Milton Friedman Institute for Research in Economics.
Bhattacharya, U., C. Holden, and S. Jacobsen. Forthcoming. Penny Wise, Dollar Foolish: Threshold Trigger
Effects in Security Trading. Management Science.
Blume, M., and I. Friend. 1975. The Allocation of Wealth to Risky Assets—The Asset Structure of Individual
Portfolios and Some Implications for Utility Functions. Journal of Finance 30:585–603.
Bolton, P., X. Freixas, and J. Shapiro. 2007. Conflicts of Interest, Information Provision, and Competition in the
Financial Services Industry. Journal of Financial Economics 85:297–330.
Bonaccio, S., and R. S. Dalal. 2006. Advice Taking and Decision-making: An Integrative Literature Review, and
Implications for the Organizational Sciences. Organizational Behavior and Human Decision Processes 101:
127–51.
———. 2010. What Types of Advice Do Decision-makers Prefer? Organizational Behavior and Human
Decision Processes 112:11–23.
Calvet, L. E., J. Y. Campbell, and P. Sodini. 2007. Down or Out: Assessing the Welfare Costs of Household
Investment Mistakes. Journal of Political Economy 115:707–47.
———. 2009a. Fight or Flight? Portfolio Rebalancing by Individual Investors. Quarterly Journal of Economics
124:301–48.
———. 2009b. Measuring the Financial Sophistication of Households. American Economic Review 99:393–98.
Campbell, J. Y. 2006. Household Finance. Journal of Finance 61:1553–1604.
Campbell, J. Y., H. Jackson, B. C. Madrian, and P. Tufano. 2011. Consumer Financial Protection. Journal of
Economic Perspectives 25:91–114.
Carhart, M. M. 1997. On Persistence in Mutual Fund Performance. Journal of Finance 52:57–82.
Carlin, B. I. 2009. Strategic Price Complexity in Retail Financial Markets. Journal of Financial Economics
91:278–87.
Carlin, B. I., and S. Gervais. 2009. Legal Protection in Retail Financial Markets. Working Paper, University of
California and Duke University.
Carlin, B. I., and D. T. Robinson. 2009. Fear and Loathing in Las Vegas: Evidence from Blackjack Tables.
Judgment and Decision Making 4:385–96.
———. 2010. What Does Financial Literacy Training Teach Us? Working Paper, National Bureau of Economic
Research.
Carlin, B., and G. Manso. 2011. Obfuscation, Learning, and the Evolution of Investor Sophistication. Review of
Financial Studies 24:754–85.
CFA Institute. 2009. European Union Member Poll on Retail Investment Products. Summary Report.
Chen, J., J. Kubik, and H. Hong. 2011. Outsourcing Mutual Fund Management: Firm Boundaries, Incentives,
and Performance. Working Paper, University of California, Syracuse University, and Princeton University.
Christelis, D., T. Jappelli, and M. Padula. 2010. Cognitive Abilities and Portfolio Choice. European Economic
Review 54:18–38.
Cooper, I., and E. Kaplanis. 1994. Home Bias in Equity Portfolios, Inflation Hedging, and International Capital
Market Equilibrium. Review of Financial Studies 7:45–60.
1030
Downloaded from
http://rfs.oxfordjournals.org/ at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

Is Unbiased Advice Sufficient?
DABbank. 2004. Faszination Wertpapier: Fakten und Hintergründe zum Anlegerverhalten in Deutschland. DAB-
Bank: Munich.
Deaves, R., E. Lüders, and G. Y. Luo. 2003. An Experimental t-test of the Impact of Overconfidence and Gender
on Trading Activity. Review of Finance 13:555–75.
DeMiguel, V., L. Garlappi, and R. Uppal. 2009. Optimal versus Naive Diversification: How Inefficient Is the
1/N Portfolio Strategy? Review of Financial Studies 22:1915–1953.
Deuskar, P., J. M. Pollet, Z. J. Wang, and L. Zheng. 2011. The Good or the Bad? Which Mutual Fund Managers
Join Hedge Funds? Review of Financial Studies 24:3008–24.
Deutsche Bundesbank. 2010. Wertpapierdepots März 2010. Frankfurt am Main.
Deutsches Aktieninstitut. 2009. DAI Factbook 2009: Statistiken, Analysen und Graphiken zu Aktionären,
Aktiengesellschaften und Börsen. Frankfurt am Main.
DiMatteo, M. R. 2004. Variations in Patient’s Adherence to Medical Recommendations: A Quantitative Review
of 50 Years of Research. Medical Care 42:200–209.
Dimson, E., P. Marsh, and M. Staunton. 2007. The Worldwide Equity Premium: A Smaller Puzzle. In R. Mehra
(ed.), Handbook of Investments: Equity Risk Premium, pp. 467–514. Amsterdam: Elsevier.
Dorn, D., G. Huberman, and P. Sengmueller. 2008. Correlated Trading and Returns. Journal of Finance 63:
885–920.
Edelen, R. M., R. Evans, and G. B. Kadlec. 2008. What Do Soft-dollars Buy? Performance, Expense Shifting,
Agency Costs. Working Paper, AFA San Francisco Meetings.
Fama, E. F., and K. R. French. 1993. Common Risk Factors in the Returns on Stocks and Bonds. Journal of
Financial Economics 33:3–56.
——. 2010. Luck versus Skill in the Cross-section of Mutual Fund Returns. Journal of Finance 65:1915–1947.
Frazzini, A. 2006. The Disposition Effect and Underreaction to News. Journal of Finance 61:2017–2046.
French, K., and J. Poterba. 1991. Investor Diversification and International Equity Markets. American Economic
Review 81:222–26.
Gino, F. 2008. Do We Listen to Advice Just Because We Paid for It? The Impact of Advice Cost on Its Use.
Organizational Behavior and Human Decision Processes 107:234–45.
Goetzmann, W., J. Ingersoll, M. Spiegel, and I. Welch. 2007. Portfolio Performance Manipulation and
Manipulation-proof Performance Measures. Review of Financial Studies 20:1503–1546.
Goetzmann, W., and A. Kumar. 2008. Equity Portfolio Diversification. Review of Finance 12:433–63.
Guiso, L., M. Haliassos, and T. Jappelli. 2002. Household Portfolios. Cambridge, MA: MIT Press.
Hackethal, A., M. Haliassos, and T. Jappelli. 2012. Financial Advisors: A Case of Babysitters? Journal of
Banking & Finance 36:509–24.
Heywood, J. 1546. A Dialogue Conteinyng the Nomber in Effect of all the Prouerbes in the Englishe Tongue.
Classic Books.
Holden, C. 2009. Excel Modeling and Estimation, 3rd ed. Upper Saddle River, NJ: Pearson Prentice Hall.
Hong, H., J. D. Kubik, and J. C. Stein. 2005. Thy Neighbor’s Portfolio: Word-of-mouth Effects in the Holdings
and Trades of Money Managers. Journal of Finance 60:2801–2824.
Huberman, G. 2001. Familiarity Breeds Investment. Review of Financial Studies 14:659–80.
Inderst, R., and M. Ottaviani. 2009. How (Not) to Pay for Advice: A Framework for Consumer Protection.
Working Paper, Goethe University Frankfurt and Northwestern University.
Investment Company Institute. 2007. Why Do Mutual Fund Investors Use Professional Financial Advisors?
Research Fundamentals 16:1–8.
1031
Downloaded from
http://rfs.oxfordjournals.org/ at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

The Review of Financial Studies / v 25 n 4 2012
Ivkovic, Z., C. Sialm, and S. Weisbenner. 2008. Portfolio Concentration and the Performance of Individual
Investors. Journal of Financial and Quantitative Analysis 43:613–55.
Jorion, P. 1986. Bayes-Stein Estimation for Portfolio Analysis. Journal of Financial and Quantitative Analysis
21:279–92.
Kahneman, D. 2003. Maps of Bounded Rationality: Psychology for Behavioral Economics. American Economic
Review 93:1449–1475.
Kelly, M. 1995. All Their Eggs in One Basket: Portfolio Diversification of US Households. Journal of Economic
Behavior and Organization 27:87–96.
Laibson, D. 1997. Golden Eggs and Hyperbolic Discounting. Quarterly Journal of Economics 112:443–77.
Lewis, K. K. 1999. Trying to Explain Home Bias in Equities and Consumption. Journal of Economic Literature
37:571–608.
Madrian, B. C., and D. F. Shea. 2001. The Power of Suggestion: Inertia in 401(k) Participation and Savings
Behavior. Quarterly Journal of Economics 116:1149–1187.
Malkiel, B. 1995. Returns from Investing in Equity Mutual Funds 1971 to 1991. Journal of Finance 50:549–72.
Markowitz, H. M. 1952. Portfolio Selection. Journal of Finance 7:77–91.
Mehra, R., and E. C. Prescott. 1985. The Equity Premium: A Puzzle. Journal of Monetary Economics 15:145–61.
Metrick, A. 1999. Performance Evaluation with Transactions Data: The Stock Selection of Investment
Newsletters. Journal of Finance 54:1743–1774.
Michaud, R. 1998. Efficient Asset Management: A Practical Guide to Stock Portfolio Optimization and Asset
Allocation. New York: Oxford University Press.
Mullainathan, S., M. Nöth, and A. Schoar. 2009. The Market for Financial Advice: An Audit Study. Working
Paper, AFA Denver Meeting.
NASD Investor Education Foundation. 2006. Investor Fraud Study: Final Report.
O’Donoghue, T., and M. Rabin. 1999. Doing It Now or Later. American Economic Review 89:103–24.
Odean, T. 1998. Are Investors Reluctant to Realize Their Losses? Journal of Finance 53:1775–1798.
———. 1999. Do Investors Trade Too Much? American Economic Review 89:1278–1298.
Poterba, J. M., and A. A. Samwick. 2003. Taxation and Household Portfolio Composition: US Evidence from
the 1980s and 1990s. Journal of Public Economics 87:5–38.
Puri, M., and D. T. Robinson. 2007. Optimism and Economic Choice. Journal of Financial Economics 86:71–99.
Samuelson, W., and R. J. Zeckhauser. 1988. Status Quo Bias in Decision-making. Journal of Risk and
Uncertainty 1:7–59.
Sharpe, W. F. 1966. Mutual Fund Performance. Journal of Business 39:119–38.
Shefrin, H., and M. Statman. 1985. The Disposition to Sell Winners Too Early and Ride Losers Too Long:
Theory and Evidence. Journal of Finance 40:777–90.
Statman, M. 2010. What Investors Really Want. New York: McGraw-Hill.
Stoughton, N. M., Y. Wu, and J. Zechner. 2011. Intermediated Investment Management. Journal of Finance
66:947–80.
Womack, K. L. 1996. Do Brokerage Analysts’ Recommendations Have Investment Value? Journal of Finance
51:137–67.
Yaniv, I. 2004. Receiving Other People’s Advice: Influence and Benefit. Organizational Behavior and Human
Decision Processes 93:1–13.
Zhu, N. 2002. The Local Bias of Individual Investors. Working Paper, Yale.
1032
Downloaded from
http://rfs.oxfordjournals.org/ at Universitaetsbibliothek Johann Christian Senckenberg on July 3, 2012

Financial advisors: A case of babysitters?
Andreas Hackethal a , Michael Haliassos a,c,d , Tullio Jappelli b,c,e,⇑
a House of Finance, Goethe University Frankfurt, Grueneburgplatz 1, D-60323 Frankfurt am Main, Germany
b Department of Economics, University of Naples Federico II, Via Cinzia 45, 80126 Naples, Italy
c CEPR, Centre for Economic Policy Research, London, United Kingdom
d CFS, Center for Financial Studies, Frankfurt, Germany
e CSEF, Centre for Studies in Economics and Finance, Naples, Italy
article info
Article history:
Received 1 March 2011
Accepted 31 August 2011
Available online 8 September 2011
JEL classification:
G1
E2
D8
Keywords:
Financial advice
Portfolio choice
Household finance
1. Introduction
abstract
In recent years households have increased their exposure to
financial risk taking, partly in response to the demographic transition
and increased responsibility for retirement financing. Recent
research points to differential financial literacy and sophistication
across households, creating the potential for important distributional
consequences of these developments (Campbell, 2006; Lusardi
and Mitchell, 2007).
In principle, financial advisors could ameliorate consequences
of differential ability to handle finances by improving returns
and ensuring greater risk diversification among less sophisticated
households. Indeed, delegation of portfolio decisions to advisors
opens up economies of scale in portfolio management and information
acquisition, because advisors can spread information
acquisition costs among many investors. Such economies of scale,
as well as possibly superior financial practices of advisors, create
the potential for individual investors to improve portfolio performance
by delegating financial decisions. But delegation entails
⇑ Corresponding author at: Department of Economics, University of Naples
Federico II, Via Cinzia 45, 80126 Naples, Italy. Tel.: +39 081 675042; fax: +39 081
675014.
E-mail addresses: hackethal@gbs.uni-frankfurt.de (A. Hackethal), haliassos@wi
wi.uni-frankfurt.de (M. Haliassos), tullio.jappelli@unina.it (T. Jappelli).
0378-4266/$ - see front matter Ó 2011 Elsevier B.V. All rights reserved.
doi:10.1016/j.jbankfin.2011.08.008
Journal of Banking & Finance 36 (2012) 509–524
Contents lists available at SciVerse ScienceDirect
Journal of Banking & Finance
journal homepage: www.elsevier.com/locate/jbf
We use two data sets, one from a large brokerage and another from a major bank, to ask: (i) whether
financial advisors are more likely to be matched with poorer, uninformed investors or with richer and
experienced investors; (ii) how advised accounts actually perform relative to self-managed accounts;
(iii) whether the contribution of independent and bank advisors is similar. We find that advised accounts
offer on average lower net returns and inferior risk-return tradeoffs (Sharpe ratios). Trading costs contribute
to outcomes, as advised accounts feature higher turnover, consistent with commissions being the
main source of advisor income. Results are robust to controlling for investor and local area characteristics.
The results apply with stronger force to bank advisors than to independent financial advisors, consistent
with greater limitations on bank advisory services.
Ó 2011 Elsevier B.V. All rights reserved.
costs in terms of commissions and fees, and might give rise to
agency problems between advisors and firms and between advisors
and poorly informed customers, as shown by Inderst and Ottaviani
(2009): on the one hand they need to sell financial products
and on the other they need to advise customers on what is best for
them to do. 1 The notion that financial advisors tend to be used by
less informed or unsophisticated investors who could be easily misled
by them, underlies much of the existing literature on financial literacy,
the possible role of financial advice and the case for regulation
of financial advisors.
In this paper, we examine three questions. First, we ask whether
financial advisors tend to be matched with poorer, uninformed
investors or rather with richer, experienced investors with higher
opportunity cost of time. Second, how brokerage accounts run by
individuals without financial advisors actually perform compared
to accounts run by (or in consultation with) financial advisors.
Third, whether the estimated contribution of financial advisors is
persistent across different advisory models such as independent
financial advisors (IFA) and bank financial advisors (BFA). Direct
performance comparisons are made possible by two unique
administrative data sets: one from a large German brokerage firm
that allows its clients choice of whether to run their accounts
1 Stulz and Mehran (2007) review the existing empirical literature on the nature
and implications of various conflicts mainly focusing on analysts.

510 A. Hackethal et al. / Journal of Banking & Finance 36 (2012) 509–524
themselves or with the guidance of an independent financial advisor;
and another from a large German commercial bank that offers
(optional) advice to its customers with investment accounts. The
answers we obtain provide a novel perspective on the role of financial
advice in individual portfolio performance.
Our first data set, from the online brokerage, tracks accounts of
32,751 randomly selected individual customers. Our second data
set contains full data of 4447 clients of a German branch-based
commercial bank. Both data sets cover 34 months, from January
2003 to October 2005. In many respects, discussed in this paper,
descriptive statistics for both samples paint a very similar picture
of the role of financial advisors as econometric analysis that controls
for investor and region characteristics.
We find that involvement of financial advisors lowers portfolio
returns net of direct cost, worsens risk-return profiles, as measured
by the Sharpe ratio; and increases account turnover and investment
in mutual funds, consistent with incentives built into the
commission structure of both types of financial advisors. If anything,
negative advisory effects on portfolio performance are even
stronger for BFAs than for IFAs. This is consistent with greater limitations
faced by BFAs in the range of products they offer and in the
way they can confer financial advice to clients.
Regression analysis of who delegates portfolio decisions presents
a further twist. It suggests that advisors are matched with richer,
older, more experienced, self-employed, female investors
rather than with poorer, younger, inexperienced and male ones.
In this respect, advisors are similar to babysitters: babysitters are
matched with well-to-do parents, they perform a service that parents
themselves could do better, they charge for it, but observed
child achievement is not boosted by babysitters but by positive
characteristics of the family. No issues of regulating babysitters
emerge, however, because the nature of the activity and the contribution
is known to all parties involved.
The paper is organized as follows. In Section 2 we discuss the
role of financial advice in overcoming investors’ informational constraints
and their incentives in handling financial portfolios in view
of relevant existing literature. Section 3 describes the brokerage
and the commercial bank data sets, the measures that we use to
characterize portfolio performance, and the estimation procedure.
Section 4 compares records of account performance with and without
involvement of financial advisors. Section 5 studies econometrically
the role of investors’ characteristics in determining which
investors are matched with financial advisors. Section 6 reports
regression estimates of the effects of independent financial advisors
on account performance, return variance, Sharpe ratios, trading
frequency, turnover, and diversification. Section 7 presents similar
results from the second sample on bank financial advisors. Section
8 concludes.
2. The role of financial advice
There is a limited but budding theoretical literature on the possible
role of financial advisors. Current theoretical work but also
policy debate on financial regulation seem to be based on the idea
that financial advisors know what is good for individual customers
but have an incentive to misrepresent this and to take advantage of
their customers, who are typically uninformed and cannot figure
out the poor quality of advice. Regulation is then needed to make
sure that this conflict of interests is dealt with.
In a recent pioneering paper, Inderst and Ottaviani (2009) analyze
‘misselling’, i.e. the practice of misdirecting clients into buying
a financial product that is not suitable for them. Their model
emphasizes the internal agency problem between the firm and
its sales agents. The agency problem is complicated by the fact that
sales agents perform the dual task of prospecting for customers
and of providing adequate advice to them on whether to buy a particular
product. As a consequence, higher sales incentives will increase
the likelihood that sales agents sell unsuitable products to
customers. If this occurs, there is a probability with which the firm
receives a complaint and has to pay a fine. To avoid misselling the
firm can set internal suitability standards for advising customers
and exert costly monitoring to verify compliance with these standards.
The standards implemented by the firm in equilibrium are
increasing in the fine (or equivalently in the reputation damage),
the transparency of the incentive scheme, and in the effectiveness
of monitoring, but they are decreasing in the sales incentives and
the private cost for the agent to investigate the match between
product and customer.
There are two relevant implications for our study. First, sales
incentives can lead financial advisors to systematically recommend
unsuitable products to their clients that entail suboptimal outcomes
on the client side. Second, due to agency costs from multitasking
and monitoring, a firm employing sales agents (such as
BFAs) would be expected to choose lower standards than an entrepreneur
(IFA). Our findings below are quite consistent with these
predictions and provide two further insights: (i) advisors may affect
portfolio outcomes not only by recommending unsuitable
products but also by encouraging excessive trading; and (ii) the notion
that advisors have an edge over their clients need not refer solely
to unsophisticated clients, but also to experienced but
inattentive ones who fail to monitor advisors and the outcome of
their activities effectively.
The empirical literature on financial advice has so far mostly focused
on whether professional analysts and advisors have an informational
advantage to contribute to individual investors when it
comes to predicting stock price movements. Ever since Cowles
(1933), there have been questions regarding the ability of stock
market forecasters and analysts to predict and reveal movements
in the stock market. 2
For example, Womack (1996) examines stock price movements
following ‘buy’ or ‘sell’ recommendations by fourteen major US
brokerage firms. He documents significant price and volume reactions
in the direction of the recommendation, especially for new
‘sell’ recommendations. He concludes that there is value to these
recommendations viewed as returns to information search costs.
However, new ‘buy’ recommendations occur seven times more often
than ‘sell’ recommendations, suggesting that brokers are reluctant
to issue sell recommendations, both in order to avoid harming
potential investment banking relationships and to maintain future
information flows from managers.
Metrick (1999) analyzes a database of recommendations of 153
investment newsletters and finds no evidence that newsletters
have superior stock-selection skill. Average abnormal returns are
close to zero; and even the performance of the best newsletters
seems to be driven more by luck then by skill. In related work,
Anderson and Martinez (2008) examine abnormal returns around
stock recommendations by Swedish brokers. A sizeable share of
abnormal profits results from transactions before the recorded recommendation
date, suggesting that tipping of customers may be
taking place. However, given the small size of these abnormal profits
(only 0.04% in yearly performance of total Swedish equity fund
assets under management), the authors wonder whether clients
are fully compensated for the costs of commissions charged by
brokers.
Barber et al. (2001) explicitly take into account trading costs
from following analyst recommendations. They analyze abnormal
gross and net returns that would result from purchasing (selling
2 Early studies include Barber and Loeffler (1993) on The Wall Street Journal’s
Dartboard column and Desai and Jain (1995) on ‘‘Superstar’’ money managers in
Barron’s.

short) stocks with the most (least) favorable consensus recommendations,
in conjunction with daily portfolio rebalancing and a
timely response to recommendation changes. Although they find
that such strategies would yield annual abnormal gross returns
greater than 4%, they also show that abnormal net returns are
not statistically significant. Bergstresser et al. (2009) compare performance
of mutual fund ‘classes’ distinguished by their distribution
channel: directly sold to investors versus sold through
brokers, with correspondingly different fee structures. They find
that funds sold through brokers offer inferior returns, even before
the distribution fee, no superior aggregate market timing ability,
and exhibit the same return-chasing behavior as observed among
direct-channel funds. Finally, more sales are directed to funds with
larger distribution fees.
Our reading of the literature on informational contributions of
analysts or brokers to direct stockholding is that these may be
present but unlikely to be exploitable by individuals given the
trading costs they entail. Therefore, in a world in which financial
advisors solely provided security selection advice we would expect
the effect of financial advice on abnormal portfolio returns to be
around zero on average after transaction costs.
However, some researchers take a different angle and point out
that, even if professional advisors do not have superior information
that is exploitable for the normal trading within an individual
account, they may be less likely to exhibit behavioral biases that
hurt account performance. They could thus help either by running
the account themselves or by encouraging investors to behave
appropriately.
A behavioral bias that has received considerable attention is the
‘disposition effect’, i.e. the tendency of some individuals to sell
winners and keep losers when it comes to direct stockholding
(Odean, 1998). Shapira and Venezia (2001) found that the disposition
effect is significantly less pronounced among professional
than among self-directed investors. Well trained advisors could
therefore aid their clients in reducing the disposition effect, potentially
enhancing risk adjusted portfolio returns.
Advisors might also be simply able to moderate trading activity
(Campbell and Viceira, 2003). Barber and Odean (2000) show that
some investors trade excessively in brokerage accounts, suffering
transactions costs that result in significantly lower returns. Such
behavior is often attributed to overconfidence, especially pronounced
among male investors (Odean, 1999; Barber and Odean,
2001). Shu et al. (2004) analyze the returns on common stock
investments by 52,649 accounts at a brokerage house in Taiwan
for 45 months ending in September 2001. They find a U-shaped
turnover and performance relation rather than the monotonic
one predicted by overconfidence: the most frequent traders in
the top turnover quintile perform better than investors in the middle
three quintiles. Other behavioral biases have been found to
influence some individual investors. For instance, Venezia et al.
(2011) focus on herding, and also document that professional
investors herd less than amateurs, while Grinblatt and Keloharju
(2001) suggest that investors trade on the basis of past returns, reference
prices, or the size of holding period gain or loss.
While the list of potential behavioral biases can grow longer, an
important question – consistent with our approach in this paper –
remains as to whether individuals who exhibit such biases are
likely to make use of professional investors. For example, Guiso
and Jappelli (2006) argued that overconfidence (i.e. the disposition
of investors to overstate the value of their private information) reduces
their propensity to seek advice. Indeed, the Barber and
Odean data come from a discount broker that does not offer advice.
Even if overconfident traders approach financial advisors, one
might wonder whether financial advisors who earn sales commissions
would actually discourage them from executing too many
trades without some incentive scheme.
A. Hackethal et al. / Journal of Banking & Finance 36 (2012) 509–524 511
On the other hand, financial advisors may help correct behavioral
biases or investment mistakes when such correction is
aligned with their interests. A case in point is diversification. A
number of empirical studies find that many individual investors
hold undiversified portfolios (see e.g. Campbell, 2006; Goetzmann
and Kumar, 2008). Financial advisors who earn commissions for
selling mutual funds have an incentive to promote such sales and
through them diversification of their client’s accounts. Shapira
and Venezia (2001) find that the number of different stocks and
the number of transactions per year is about three times as high
for accounts managed by professionals than for self-managed accounts.
This finding indicates that financial advisors might promote
diversification through single stocks if they participate in
brokerage commissions. 3
Taken together, literature suggests an ambiguous effect of
financial advice on net returns and risk profiles of client portfolios.
Although it seems to be rather unlikely that advisors enhance portfolio
performance through informational contributions they might
in fact improve the risk-return profile by ironing out behavioral
biases of their clients. Of course, such positive effects must exceed
the cost of advice in order to yield an overall positive effect.
Our paper takes a direct approach to the issue of the role and
contribution of financial advisors. Recognizing both the potential
informational advantage and the potential contribution of professional
investors to controlling behavioral biases and correcting
investment mistakes, it compares directly portfolio returns (net
of transactions costs) and portfolio risk levels that investors actually
accomplish on their own versus what they accomplish with
the guidance of a financial advisor. It does so with reference to
portfolios actually chosen and adjusted by investors, which include
directly held stocks, bonds, and mutual funds; and it accounts for a
number of investors’ and region characteristics observable in our
data and for how they influence the tendency to use a financial
advisor. Moreover, we are able to measure the effects of advice
across two distinct advisory models (IFAs and BFAs).
3. Data, measurement, and estimation
3.1. Data on independent financial advisors
The first data set we use is administrative information from a
large German brokerage firm. It covers the investments of 32,751
randomly selected customers. They all had an active account with
the brokerage firm over the sample period from January 2003 to
October 2005. If customers opened multiple accounts we consolidated
them into one single account.
For each sampled customer we have information on date of
birth, gender, marital status, profession (including status as employed
or self-employed), zip-code of place of residence, nationality,
and self-reported security-trading experience in years. 4 All
information was collected by the brokerage firm on the date of account
opening and updated according to new information that the
firm has obtained from the customer in the interim.
On average (not excluding account owners aged under 18),
sample customers held 38.6% of their account volume in the form
of equity mutual funds, 47.4% in the form of single stocks (28%
3 Moreover, Shapira and Venezia (2001) find that the round trip performance of
professionally managed accounts is slightly (and for some specifications also
statistically) higher than that of independently run accounts. The discrepancy to
our own results is likely due to the fact that they focus on round trip returns for stock
investments in one particular year, whereas we measure total portfolio returns for a
longer time period, and in addition control for individual investor characteristics.
4 Self-reported trading experience is reported on a scale with intervals equal to 5
years. We construct a variable that has the interval midpoints as values and then add
the number of years that elapsed since account opening to approximate total trading
experience at the beginning of our observation period.

512 A. Hackethal et al. / Journal of Banking & Finance 36 (2012) 509–524
thereof in German stocks), 2.4% in the form of bond mutual funds,
3.8% in the form of single bonds and the remainder in the form of
structured investment certificates, warrants, and other securities.
Our administrative data set includes a variable that indicates
whether a given brokerage customer is also a client of an IFA
who registered with the brokerage firm. We know from the brokerage
firm that, typically, advised customers were brought to the
brokerage by IFAs. About 90% of IFAs registered with the brokerage
are former employees of commercial banks advising customers on
investment accounts. They decided to leave the bank and become
independent, thereby offering lower costs than banks and greater
choice of financial products. Thus, they were able to persuade
many of their former customers at the bank to transfer investment
accounts to the brokerage firm. The remaining 10% of IFAs in our
sample are not former bank employees but they instead joined a
larger team of IFAs directly and built up their own customer base,
again drawing mostly from former bank customers.
At the time of account opening, IFAs had typically obtained a
client mandate to place orders on behalf of the client. We do not
have information on which clients fully delegate trading decisions
to their IFAs and which only consult their IFAs for guidance and
then place trades themselves. The brokerage firm offers several
compensation schemes to IFAs. Only for a negligible fraction of
IFAs are revenues dependent solely on assets under management.
More than 90% of IFAs generate at least a portion of revenues from
trades, such as sales commissions. In the case of mutual funds the
commission is a function of the upfront load the brokerage firm
earns from the fund producer.
Of the customers in our sample, 12.8% consult IFAs registered
with the brokerage firm. More than half of these customers are
IFAs’ former banking clients, with the remaining half (typically also
former bank customers) having been acquired over the years, most
importantly through existing customers’ referrals. We cannot rule
out that customers coded as not using an IFA obtain professional
advice from outside advisors. This is, however, rather unlikely because
such outside advisors do not participate in the fees and commissions
paid by the client to the brokerage firm and must
therefore charge their services on top of the full brokerage fees
and commissions.
Table 1 shows descriptive statistics of the total brokerage sample
and of the two sub-samples distinguished by whether sample
customers were advised by IFAs or not, after dropping accounts
that report age of account owner below 18. 5 As shown in the table,
77.8% of account owners were male, and 47.9% married. Overall,
86.1% were employed (excluding public servants) and 13.2% were
self-employed, with the remaining 0.7% being public servants, retirees,
housewives or students. Average trading experience as of January
2003 was 9.34 years. Among IFA-assisted customers, men are
underrepresented relative to their share in the overall account owner
pool, and so are married owners (as indicated by the corresponding
t-statistics and p-values in the last two columns of Table 1). Older
owners (above 50) are overrepresented, and advised customers have
on average more years of experience and larger initial size of
accounts.
Table 1 also reports performance figures for the brokerage accounts.
As indicated by the t-statistics and associated p-values,
raw returns, abnormal returns (see Section 3.3 for definitions),
5 These are typically accounts run by parents on behalf of their children.
Specifically, 796 investors in our original sample were younger than 18 on September
5, 2006, and the youngest investor in that sample was just under 6 years old. Tax
advantages for parents arise because during the observation period there was a per
person threshold level of interest or dividend income above which capital income tax
needed to be paid. We have also run the regressions including investors under 18, but
our results were hardly affected in terms of sign, significance, and even size of
estimates, except for small changes in the estimates for age categories.
Table 1
Descriptive statistics for the brokerage sample (IFAs).
Selfmanaged
accounts
sample
mean
Accounts
run by
IFA
sample
mean
T-test for
difference
in means
All
accounts
sample
mean
All
accounts
standard
deviation
Dependent variables
Log monthly
returns
0.0101 0.0063 24.70 *
0.0097 0.0089
Jensen’s alpha 0.0098 0.0061 23.72 *
0.0093 0.0091
Alpha – four
factor
model
0.0093 0.0055 22.31 *
0.0088 0.0100
Variance of
monthly
returns
0.0032 0.0019 28.39 *
0.0031 0.0027
Sharpe ratio 0.2229 0.1916 11.73 *
0.2189 0.1585
No. of trades/
account
volume in
‘000
0.0861 0.0884 0.87 0.0864 0.2609
Monthly
turnover
rate
0.0405 0.0895 33.38 *
0.0468 0.0865
Share of
directly
held stocks
0.5777 0.2000 58.70 *
0.5295 0.3838
Control variables
Male 0.7925 0.6739 15.37 *
0.7774 0.4160
Married 0.4812 0.4636 0.92 0.4790 0.4996
Employed
(excluding
public
servants)
0.8655 0.8334 4.93 *
0.8614 0.3455
Selfemployed
0.1280 0.1577 5.10 *
0.1318 0.3383
Experience 9.3415 11.1535 16.27 *
9.5684 6.2182
18 6 Age 6 30 0.0101 0.0415 3.43 *
0.0462 0.2100
30 < Age 6 40 0.0098 0.1180 18.35 *
0.2409 0.4276
40 < Age 6 50 0.0093 0.2680 8.66 *
0.3346 0.4719
50 < Age 6 60 0.0530 0.2287 5.00 *
0.1995 0.3997
Age > 60 0.0708 0.3437 28.11 *
0.1787 0.3831
Log account
volume in
2003
9.1588 10.2823 42.98 *
9.3023 1.4917
Observations 25,173 3686 28,321 28,321
The t-test refers to a test of the null hypothesis that the mean of the sample with
self-managed accounts equals the mean of the sample with accounts run by an
independent financial advisor (IFA).
* The two means are statistically different from each other at the 1% level.
raw return variance and Sharpe ratios are on average significantly
lower for IFA-assisted than for self-directed customers.
All reported return figures are monthly and net of any transactions
costs and provisions charged by the brokerage on its own account
or on behalf of the IFA. 6 Transaction costs and provisions are
divided between the brokerage and IFA, with the bank typically
earning roughly 30 basis points for transaction fees, account maintenance,
and front loads, leaving about 170 basis points for the IFA.
There is a minority of advisors who follow a different business model:
instead of earning front loads, they forward those to their clients
and earn an extra fee as a percentage of account volume. As this
6 Although we only observe net returns in our data and therefore cannot directly
measure transaction cost, we know from the data provider that the brokerage and the
IFA combined earn typically 100–200 basis points on clients with account volume
greater than 50,000 Euros. For smaller accounts, this number is typically in the
neighborhood of 200 basis points, although it can be as high as 300–500 basis points,
due to front loads (principally observable in the data set) and kick-backs (not
observable) from mutual funds.

extra fee is not run through the bank, it is not observed by us and it is
not taken into account in computing returns and other measures of
performance net of costs. Since we obtain negative effects of IFAs on
account performance in econometric estimations below, the resulting
understatement of costs in these cases, if anything, strengthens
our findings on the role of IFAs.
The monthly position statements list for each item the type of
security (e.g. stocks, bonds, mutual funds, etc.), the number of
securities, and the market value per security at month end. At
the start of the sample period (January 2003), average annual account
volume was 10,963 Euro. We compute monthly turnover
by dividing the combined transaction value of all purchase transactions
for a given month by the average of beginning-of-month and
end-of-month account volume. Average monthly turnover is 4.7%
in our sample, but about double of this for advised customers.
3.2. Data on bank financial advisors
In order to compare our findings across different advisory models,
we also consider a second data set of investment accounts, this
time from a large German commercial bank that offers optional advice
to its customers through its bank employees assigned to this
task. Unlike the online brokerage that likely attracts a selected
sample of the German population interested in trading, the bank
has a wide network of branches that reach a broad cross section
of the German population. This data set consists of 10,434 randomly
selected customers observed over a 34-month period, from
Table 2
Descriptive statistics for the bank sample (BFAs).
Selfmanaged
accounts
sample
mean
Accounts
run by
BFAs
sample
mean
T-test for
difference
in means
All
accounts
sample
mean
All
accounts
standard
deviation
Dependent variables
Log monthly
returns
0.0076 0.0040 11.59 0.0054 0.0101
Variance of
monthly
returns
0.0045 0.0046 0.43 0.0046 0.0130
Sharpe ratio 0.4252 0.2662 12.23 0.3253 0.4020
Monthly
turnover
rate
0.0680 0.0520 8.34 0.0579 0.1103
Share of
directly
held stocks
0.2975 0.1188 20.07 0.1853 0.3009
Control variables
Male 0.5102 0.4350 4.86 0.4630 0.4986
Employed
(excluding
public
servants)
0.4413 0.3501 6.06 0.3840 0.4864
Executive
employee
0.0284 0.0257 0.54 0.0267 0.1613
Housewife 0.0665 0.1034 4.22 0.0897 0.2858
Retired 0.1577 0.2205 5.05 0.1972 0.3979
18 6 Age 6 30 0.1203 0.0920 2.92 0.1020 0.3033
30 < Age 6 40 0.1644 0.0941 6.88 0.1203 0.3253
40 < Age 6 50 0.1922 0.1385 4.80 0.1585 0.3652
50 < Age 6 60 0.1753 0.1736 0.22 0.1742 0.3793
Age > 60 0.3476 0.5016 10.05 0.4443 0.4969
Log account
volume in
2003
9.0002 9.7146 10.63 9.4489 2.1695
Observations 1648 2792 4440 4440
The t-test refers to a test of the null hypothesis that the mean of the sample with
self-managed accounts equals the mean of the sample with accounts run by a bank
financial advisor (BFA).
A. Hackethal et al. / Journal of Banking & Finance 36 (2012) 509–524 513
January 2003 to October 2005. For 4447 of those, we have detailed
information on whether particular trades were executed following
consultation with a bank financial advisor (a bank employee) or
without such consultation. Accordingly, we construct a dummy
variable for bank financial advisor use (BFA) that takes the value
of 1 if the customer has consulted with a BFA at least once during
the observation period and 0 if the customer never consulted a BFA
during the period. For comparability’s sake, we match these accounts
to the same regions as in the brokerage data set and use
the same regional variables and (virtually) the same set of account
owner characteristics as for the brokerage sample.
Table 2 presents descriptive statistics for the bank sample.
Again we distinguish between self-managed accounts and accounts
that are at least partly managed by an advisor. Male account
owners are in a minority in this sample (46.3%) and they are underrepresented
(again as indicated by the corresponding t-statistics in
the table) among those customers who consulted a bank financial
advisor before executing some trade(s). As expected, retirees and
housewives are much more strongly represented in the bank sample
than in the brokerage sample. They comprise just fewer than
30% of the observations and they are overrepresented among advised
customers. The majority of account owners are at least
50 years old, and those above 60 are overrepresented among advised
customers. The average account volume at the start of
2003 was slightly higher in this sample, namely 12,694 Euro. The
average monthly turnover rate was 5.8%, somewhat higher than
in the brokerage sample, and smaller for advised customers than
for those who never consulted a BFA.
Finally, Table 2 reports performance figures for the bank accounts,
showing that raw returns and Sharpe ratios are on average
significantly lower for BFA-assisted customers than for self-directed
customers. Given the different composition and advisory models
of the two samples, it will be interesting to see if findings on the
contribution of financial advice to account performance persist (or
differ) across samples.
3.3. Measuring account performance
In this paper we are interested in the effect of financial advice
on portfolio performance and portfolio risk and in particular on
abnormal returns. In order to compute monthly portfolio returns,
we assume as in Dietz (1968) that all transactions occur in the
middle of a given month:
VE VB CF
R ¼
VB þ 0:5 CF
where VE is the value of the portfolio at end of month including
earned dividends and coupons, VB the market value of the portfolio
at beginning of month, and CF is the net cash flow for month t from
purchases (enter positively) and sales of securities(enter negatively)
at transaction prices
Monthly returns from (1) are winsorized by treating returns
that fall into the first or the 100th percentile as missing values. 7
We construct log returns and use them and the standard regression
model in (2) to estimate abnormal (log) returns for each portfolio
based on CAPM.
rp;t rf ;t ¼ ap þ b pðrM;t rf ;tÞþep;t ð2Þ
where ap is the estimated abnormal return (Jensen’s Alpha) for portfolio
p, b p is estimated market beta for portfolio p, r M,t is log return
7 Extreme monthly return observations were treated as missing (and not set to the
upper/lower boundary that would be customary for winsorization) because (a) they
most likely represent erroneous data, and (b) we do not lose customers but just single
months. As a consequence, some customers have only 33 instead of 34 monthly
return observations.
ð1Þ

514 A. Hackethal et al. / Journal of Banking & Finance 36 (2012) 509–524
of the Euro-denominated MSCI-World Index in month t, rf,t is log return
on the 1-month Euribor, e p,t is the error term of regression for
portfolio p.
In order to test robustness of our results to the way abnormal
returns are computed, we also present results for an alternative
estimate of excess returns based on a four-factor model proposed
by Carhart (1997) to measure portfolio performance. The model
is specified as follows:
ri;t ¼ ai þ b 1iRm;t þ b 2iSMBt þ b 3iHMLt þ b 4iMOMt þ eit
where the intercept ai measures risk-adjusted monthly abnormal
portfolio returns, ri;t denotes monthly excess returns on portfolio i
relative to the risk-free rate which is captured by monthly returns
on the JP Morgan 3 Month Euro Cash Index, R m,t denotes the excess
return on the market portfolio which we approximate by the comprehensive
German CDAX Performance Index, SMB t, HML t, and
MOMt correspond to monthly returns on size, value premium and
momentum portfolios. The size portfolio return (SMB) is approximated
by the difference in monthly returns on the small cap SDAX
index and the large cap DAX 30 index. The book-to-market portfolio
return (HML) is approximated by the return difference between the
MSCI Germany Value Index and the MSCI Germany Growth Index.
Finally, the momentum portfolio return (MOM) is the difference in
monthly returns between a group of stocks with recent above-average
returns and another group of stocks with recent below-average
returns. The group with above-average returns is defined as the top
30% of stocks from the CDAX index over the past 11 months and the
below-average group contains the lowest 30% of stocks from the
same index over the same time period.
4. Performance record of financial advisors
For many brokerage clients, a natural first step towards deciding
whether to use an IFA or not would be to compare the historical performance
of accounts run with IFA involvement and those run without
it. Similarly, bank customers would like to know if those who have
contacted bank financial advisors have done better on average than
those who did not. Even in the absence of official records (indeed neither
the broker or the bank compute or print portfolio performance
0 .05 .1 .15
ð3Þ
records for their clients), prospective clients may still be influenced
by the experiences of existing clients through word of mouth.
Fig. 1 plots histograms of average monthly log returns over our
observation period for brokerage accounts that were self-managed
and for those run with IFA input. Self-managed accounts exhibit a
more symmetric distribution, while advised accounts show higher
mass at the lower end of returns. Table 1 shows monthly logarithmic
returns. The sample mean log monthly return on IFA accounts
over this period is actually lower than that of self-managed accounts:
0.63% versus 1.01%. This corresponds to a difference in annual
rates of return of 5% points (7.9% for advised customers versus
12.9% for those who invested alone). Even though the brokerage
house itself neither collected nor published such statistics, the difference
seems rather hard to miss. Table 1 confirms that IFA accounts
are also characterized by lower abnormal returns than
self-managed accounts, regardless of whether we use a single-factor
model based on the MSCI-World Index or whether we use a
four-factor model based on German stock data.
These lower returns offered by IFAs are combined with lower
average variance of portfolio returns raising the possibility that
they simply reflect an efficient risk-return tradeoff. Strikingly,
however, the sample average of the Sharpe ratio on advised accounts
is also lower than that on self-run brokerage accounts, suggesting
that advisees ‘paid’ on average a higher cost (in terms of
returns) to attain lower risk than what was available to self-managed
accounts. Fig. 2 shows that the distribution of total portfolio
variance under IFAs is ‘squeezed’ towards values closer to zero
compared to what is produced by individuals managing their accounts,
but Fig. 3 shows a much greater heterogeneity in Sharpe ratios
among advised customers than among the rest.
Comparison of IFA and non-IFA accounts also shows a rather
small difference in frequency of trades across the two types of accounts,
but a much more pronounced one when average portfolio
turnover (which is sensitive to the size of purchases) is considered:
the average turnover rate is more than double for IFA accounts.
Looking at Figs. 4 and 5, both measures tend to be clustered closer
to zero for self-managed accounts. In other words, IFAs get commission
based on the volume of purchases and tend to exhibit
greater purchases than individual clients on average. IFA accounts
Self-managed Run by Financial Advisor
-.03 -.02 -.01 0 .01 .02 .03 -.03 -.02 -.01 0 .01 .02 .03
Monthly Returns
Fig. 1. The distributions of log monthly returns.

0 .1 .2
0 .05 .1
0 .005 .01 .015 0 .005 .01 .015
Variance of Monthly Returns
tend also to be larger, and are therefore associated with larger positions
and trades.
Finally, IFA accounts tend to exhibit far greater diversification
than those run by individuals alone. The average share of directly
held stocks among self-managed accounts is just under 60%, while
that for IFA accounts is about 20%. This seems consistent with
incentives to sell mutual funds that IFAs have.
Table 2 presents a similar comparison for bank customers who
have used the advice of bank employees prior to making trades versus
those who have not. Accounts of customers who have resorted to
bank advisors exhibit on average lower returns, comparable variance,
much lower Sharpe ratios, and smaller shares of directly held
stocks than those who did not approach the benchmark. Unlike what
we found for brokerage clients, turnover rates of those who made
use of bank advice were on average lower than of those who did not.
A. Hackethal et al. / Journal of Banking & Finance 36 (2012) 509–524 515
Self-managed Run by Financial Advisor
Fig. 2. The distributions of the variance of monthly returns.
Self-managed Run by Financial Advisor
-.2 0 .2 .4 .6 .8 -.2 0 .2 .4 .6 .8
Sharpe ratio
Fig. 3. The distribution of the Sharpe ratio.
All in all, performance records of IFAs and BFAs during this sample
period do not appear favorable towards advised accounts, especially
in terms of the risk-return tradeoff offered. The deeper
question is, of course, whether these differences are due to financial
advisors themselves or to the customers they tend to attract.
It is to this household finance question that we now turn, focusing
first on IFAs and then on BFAs.
5. Who has a financial advisor?
We first consider which client characteristics of the brokerage
firm or bank contribute to the client’s account being run with
advisor input. A priori, it may be that advisors tend to be matched
with younger, less experienced and less wealthy investors, who

516 A. Hackethal et al. / Journal of Banking & Finance 36 (2012) 509–524
Fraction
0 .2 .4 .6 .8
Fraction
0 .5
0 .5 1 1.5 2 0 .5 1 1.5 2
Number of Trades per Year / '000 of Account Volume
need them most; or that they are matched with older, more experienced
and wealthier investors who can pay them most.
Table 3 reports linear probability regressions of whether the
client makes use of an IFA or a BFA, respectively. 8 We control for
time-invariant characteristics (such as availability and cost of financial
advice and characteristics of investor pools) in the region
(columns 1 and 3) and zip code (columns 2 and 4). 9 We see that,
8 Results from probit models (omitting zip-code dummies) deliver very similar
results to the linear probability models, and are available on request.
9 The German Zip Code (Postleitzahlen) is a five digit number consisting of the wider
area that is placed on the thousandth position and the postal district (the unit, tenth
and hundredth positions). Regressions in columns 1, 3, 4 and 6 include dummies for
broader regions (Dresden, Berlin, Hamburg, Hannover, Dusseldorf, Bonn, Frankfurt,
Stuttgart, Munich and Nuremberg), while regressions in columns 2 and 5 include
dummies at the zip code level. There are 5652 zip code dummies for the brokerage
sample and 646 for the bank sample.
Self-managed Run by Financial Advisor
Fig. 4. The distribution of number of trades (per ‘000 account volume).
Self-managed Run by Financial Advisor
0 .1 .2 .3 .4 .5 .6 0 .1 .2 .3 .4 .5 .6
Turnover Rate
Fig. 5. The distribution of the monthly turnover rate.
given other characteristics, males are less likely to use an advisor,
consistent with the view that males tend to have more (over)confidence.
Older clients (over 50) have a significantly greater probability
than investors between 18 and 30 of using an advisor, by about 10%
points in both samples. Wealthier brokerage clients, as proxied by
the beginning-of-period account size to minimize endogeneity problems,
are significantly more likely to use IFA or BFA.
Married clients are less likely to use an IFA, controlling for other
factors, probably because spouses can be used as sounding boards.
An extra year of self-reported experience with the relevant financial
products increases the probability of using an IFA. We do not
have information on the marital status and trading experience of
bank customers, but their professional status is statistically
insignificant.
Overall our regressions show that advisors are more likely to be
matched with wealthier (as measured by account volume), older,

Table 3
The determinants of having the account run by an IFA or a BFA.
Brokerage sample (IFAs) Bank sample(BFAs)
(1) (2) (3) (4)
Male 0.062 ***
0.061 ***
0.049 ***
0.038 **
(12.06) (10.92) (3.14) (2.15)
Employee 0.057 **
0.061 **
0.054 ***
0.039 *
(2.44) (2.52) (2.93) (1.91)
30 < Age 6 40 0.031 ***
0.025 **
0.050 0.060
(3.32) (2.40) (1.44) (1.49)
40 < Age 6 50 0.004 0.004 0.013 0.021
(0.36) (0.35) (0.38) (0.57)
50 < Age 6 60 0.023 **
0.021 *
0.053 *
0.042
(2.20) (1.81) (1.67) (1.16)
Age > 60 0.088 ***
0.088 ***
0.112 ***
0.089 ***
(7.53) (6.71) (3.87) (2.73)
Log account volume 0.045 ***
0.044 ***
0.027 ***
0.034 ***
(22.07) (20.19) (7.51) (7.92)
Self-employed 0.060 **
0.064 **
(2.48) (2.55)
Experience/100 0.159 ***
0.148 ***
(3.45) (3.03)
Married 0.023 ***
0.025 ***
(5.61) (5.10)
Executive 0.003 0.032
(0.06) (0.59)
Housewife 0.002 0.016
(0.06) (0.49)
Retired 0.025 0.009
(1.11) (0.34)
Constant 0.358 ***
0.316 ***
0.366 ***
0.303 ***
(12.58) (10.34) (7.68) (6.58)
Observations 28,321 28,321 4440 4440
R-squared 0.09 0.37 0.05 0.23
Zip code dummies No Yes No Yes
The table reports estimates from a linear probability model of having an Independent
Financial Advisor or a bank financial advisor. Log account volume is measured
in January 2003. All regressions include regional dummies (absorbed by zip code
dummies where present). Asymptotic standard errors corrected for clustering at the
zip code level are reported in parentheses.
* Statistical significance at the 10% level.
** Statistical significance at the 5% level.
*** Statistical significance at the 1% level.
more experienced, single, and female investors. Such investors
have better reasons to want to delegate to advisors, such as high
opportunity cost or low inclination to spend time managing investments,
as well as sizeable wealth. The results are remarkably consistent
across the two samples, and robust to inclusion of zip code
dummies that control for unobserved factors at the local level.
Since IFAs and BFAs earn more on wealthy clients with high opportunity
costs of time, they seem to go for the big players who have a
lot to invest, rather than for the younger, smaller, inexperienced
investors who have a lot to learn.
6. Independent financial advisors and portfolio performance
We now turn to how IFA use affects account performance once
we control for client characteristics. An important estimation issue
is omitted variable bias: unobserved factors may simultaneously
affect the probability of using an advisor as well as account performance.
For instance, our data do not report willingness to undertake
financial risk: more risk averse clients may be inclined to consult an
advisor and to invest in a safer portfolio, thus influencing account
returns and variance. Other factors could also influence both advisor
use and account performance: financial literacy and sophistication;
attitudes developed in formative years, e.g. through parental influence
or observation of others in the parental social circle; social attitudes,
such as trust in others, that have been shown to influence
portfolio composition (e.g. participation in stockholding) and dele-
A. Hackethal et al. / Journal of Banking & Finance 36 (2012) 509–524 517
gation to a financial advisor. In order to attenuate this problem, we
control for as many possible factors that are observed in our data
set as well as for regional dummies and a finer classification of zip
code dummies.
A second issue is the potential endogeneity of the choice of consulting
a financial advisor. Investors with low performing portfolios
may be induced to use a financial advisor by the media or
specific campaigns that advisor-assisted portfolios perform better.
A negative correlation between advisor consultation and, say, returns,
might therefore be driven by the effectiveness of these campaigns,
rather than by a negative role of financial advisors per se. In
order to handle this possibility, we present (in Appendix A) instrumental
variable estimates using as instrument the local GDP share
of financial services. Results are consistent with our OLS findings
below.
6.1. IFA effect on portfolio returns
Table 4 presents OLS estimates regarding the influence of IFA
use on raw net returns, and on abnormal net returns, constructed
on the basis of a single-factor and of a four-factor model in the
spirit of Carhart (1997). Columns 1 and 2 report estimated effects
on average portfolio returns, controlling for investor characteristics
and regional dummies. Model (2) adds zip code dummies as controls
for time-invariant local characteristics. IFA effects are almost
identical in both models: negative and statistically significant at
the 1% level, implying that IFA use reduces monthly log returns
by roughly 0.4% points. Thus, the lower returns for advised accounts
in descriptive statistics survive controls for personal and regional
characteristics.
Even if IFAs reduce raw returns, they might still be found to create
value by increasing risk adjusted returns. Columns 3 and 4 in
Table 4 report OLS regressions for alphas from a model with the return
on the MSCI world index as the single factor (denoted Jensen’s
alpha). The IFA contribution is again negative and of similar magnitude
as for raw returns, once characteristics of the account owner
and region are taken into account. In columns 5 and 6, we examine
robustness with respect to using the four-factor model for German
stock markets outlined above. The strongly statistically significant
negative effect of IFA use is observed regardless of whether we use
a single or a four-factor model, and its size is remarkably similar
with the other models and with the descriptive statistics from Table
1.
Across all models, male gender is found to detract from account
returns, consistent with the literature on overconfidence. Years of
experience tend to contribute to higher total return, albeit by a
small estimated amount. This is consistent with recent studies
indicating that the magnitude of investment mistakes decreases
with sophistication and experience. 10
Findings in this section imply that involvement of IFAs with
brokerage accounts tends to reduce both raw and abnormal returns,
even after investor and area characteristics are taken into account.
Our results are consistent with the cost of financial advice
exceeding, on average, any benefits from informational contributions.
Importantly, this does not necessarily imply that (some) IFAs
engage in misselling or that all IFAs give uniformly bad advice.
10 For example, Feng and Seasholes (2005) ask whether investor sophistication and
trading experience eliminate behavioral biases, such as the disposition effect, using
data from the PR of China. They proxy sophistication mainly by the number of trading
rights (indicating the number of methods to trade) and an indicator of initial portfolio
diversification, both at the start of the observation period. Experience is proxied by the
number of positions taken by investor i up until date t, a time-varying covariate. They
conclude that sophistication and experience eliminate the reluctance to realize losses,
but only reduce the propensity to realize gains by 37%. See also Grinblatt and
Keloharju (2001), Feng and Seasholes (2005), and Lusardi and Mitchell (2007).

518 A. Hackethal et al. / Journal of Banking & Finance 36 (2012) 509–524
Table 4
The determinants of portfolio returns in the brokerage sample.
Log returns Jensen’s alpha Alpha 4 factor model
(1) (2) (3) (4) (5) (6)
Financial advisor (IFA) 0.004 ***
0.004 ***
0.004 ***
0.004 ***
0.004 ***
0.004 ***
(26.45) (18.33) (26.88) (18.22) (27.03) (17.11)
Male 0.001 ***
0.000 ***
0.000 ***
0.000 ***
0.001 ***
0.001 ***
(4.34) (3.14) (3.95) (2.99) (5.37) (4.17)
Married 0.000 0.000 0.000 0.000 0.000 **
0.000
(0.79) (0.22) (1.16) (0.43) (2.12) (1.49)
Employee 0.001 0.001 0.001 0.001 0.001 *
0.001 *
(1.57) (1.43) (1.47) (1.39) (1.95) (1.72)
Self-employed 0.001 *
0.001 *
0.001 *
0.001 *
0.001 **
0.002 **
(1.91) (1.77) (1.86) (1.74) (2.35) (2.11)
Experience/100 0.002 **
0.001 0.001 0.000 0.001 0.001
(2.28) (0.98) (1.58) (0.43) (1.31) (0.53)
30 < Age 6 40 0.000 *
0.001 *
0.000 *
0.001 *
0.000 0.000
(1.84) (1.70) (1.86) (1.71) (0.89) (1.07)
40 < Age 6 50 0.000 0.000 0.000 0.000 0.000 0.000
(0.32) (0.37) (0.46) (0.43) (1.36) (0.97)
50 < Age 6 60 0.000 0.000 0.000 0.000 0.000 0.000
(0.15) (0.43) (0.18) (0.48) (1.11) (0.33)
Age > 60 0.001 *
0.000 0.000 0.000 0.001 ***
0.001
(1.80) (0.95) (1.64) (0.79) (2.63) (1.58)
Log account volume 0.000 ***
0.000 ***
0.000 ***
0.000 ***
0.000 ***
0.000 ***
(6.92) (4.81) (6.68) (4.62) (7.51) (5.24)
Constant 0.008 ***
0.008 ***
0.008 ***
0.008 ***
0.007 ***
0.008 ***
(11.92) (9.78) (11.24) (9.36) (9.48) (7.90)
Observations 28,321 28,321 28,321 28,321 28,321 28,321
R-squared 0.03 0.23 0.02 0.22 0.02 0.22
Zip code dummies No Yes No Yes No Yes
Log account volume is measured in January 2003. All regressions include regional dummies (absorbed by zip code dummies where present). Asymptotic t-statistics corrected
for clustering at the zip code level are reported in parentheses.
* Statistical significance at the 10% level.
** Statistical significance at the 5% level.
*** Statistical significance at the 1% level.
6.2. IFA effect on variance of returns and Sharpe ratio
The finding that IFAs tend to lower both raw and abnormal account
returns, given investor characteristics, need not be negative,
if IFAs ensure that clients are exposed to smaller portfolio risk.
Descriptive statistics above seem to be pointing in this direction.
We therefore turn next to the effect of IFA involvement on variance
of monthly portfolio returns and on the risk-return tradeoff as captured
by the Sharpe ratio. Table 5 reports our findings.
Column 1 reports OLS results for a model with regional dummies,
whereas column 2 reports results when zip-code dummies
are included. In both models, IFAs reduce portfolio risk in line with
descriptive results. Being male, inexperienced, single, self-employed,
and with a smaller account all contribute significantly to
higher total portfolio risk. Results on control variables are intuitive.
For example, larger accounts should allow more diversification,
and we indeed find below that they tend to have smaller portfolio
shares in directly held stocks.
Set against the negative effects of IFAs on account returns, their
moderating effect on variance raises the question of whether IFAs
help achieve an efficient risk-return tradeoff. We present two
regressions (columns 3 and 4), one with regional and the other
with zip code dummies. Both show a statistically significant negative
effect of IFAs on Sharpe ratios. Male gender contributes to inferior
risk-return tradeoffs, consistent with overconfidence; while
more experienced or married investors tend to achieve better
tradeoffs. Interestingly, self-employed and older clients are seen
to have a tendency to expose themselves to more risk than what
is efficient for a given increase in expected return. Finally, wealthier
investors tend to achieve better risk-return tradeoffs, presumably
by exploiting economies of scale in asset management. We
conclude from Table 5 that IFAs tend to reduce portfolio risk but
do not compensate sufficiently for lower returns: IFA use decreases
ex post portfolio efficiency.
6.3. IFA effect on trading, turnover, and diversification
What type of behavior underlies our results on returns and risk?
The fact that IFAs earn commissions mainly when the account
owner purchases mutual funds creates an incentive for them to
encourage fund purchases. The first two columns of Table 6 examine
the effect of IFA on the number of purchases per month scaled
by account volume. These exclude account transactions from corporate
actions, periodic saving plan investments and portfolio
transfers, so as to be more directly linked to the IFA incentives to
sell specific financial instruments.
Our results imply a negative effect of IFAs on the standardized
number of purchases. Purchases result in transactions costs and
could contribute to lower net returns, but it appears that the negative
effect of IFAs on net returns reported above does not result
simply from an increased frequency of purchases. The regression
does confirm the positive role of male gender found in other studies
(see above). Financial experience is estimated to reduce the
number of purchases, consistent with Dorn and Huberman
(2005) finding that respondents with longer investment experience
trade less, but the effect is not statistically significant. Account
holders between 40 and 60 are significantly more likely to
engage in purchases than other age groups. Subject to the
provision interpreting age effects, this finding is consistent with
them being in the asset accumulation phase, prior to entering
retirement.
Although we do not find a simple channel through frequency of
trading, this does not mean that IFAs do not respond to incentives
offered by commissions. It is useful to recall that commissions are

Table 5
The determinants of portfolio return variance and Sharpe ratios in the brokerage sample.
Variance of monthly returns Sharpe ratio
(1) (2) (3) (4)
Financial advisor (IFAs) 0.001 ***
0.001 ***
0.056 ***
0.052 ***
(14.74) (9.95) (13.87) (10.92)
Male 0.001 ***
0.001 ***
0.022 ***
0.021 ***
(15.82) (13.11) (10.31) (7.85)
Married 0.000 ***
0.000 ***
0.006 ***
0.006 ***
(8.01) (6.27) (3.26) (2.62)
Employee 0.000 *
0.000 0.021 **
0.017
(1.77) (0.24) (2.10) (1.39)
Self-employed 0.001 ***
0.001 ***
0.033 ***
0.029 **
(5.18) (2.93) (3.22) (2.21)
Experience/100 0.000 0.000 0.056 ***
0.043 **
(0.50) (0.47) (3.74) (2.31)
30 < Age 6 40 0.000 *
0.000 0.001 0.002
(1.77) (1.19) (0.17) (0.42)
40 < Age 6 50 0.000 ***
0.000 ***
0.012 ***
0.013 **
(5.46) (3.99) (2.59) (2.33)
50 < Age 6 60 0.000 ***
0.000 ***
0.014 ***
0.013 **
(5.84) (4.21) (2.77) (2.19)
Age > 60 0.001 ***
0.000 ***
0.018 ***
0.015 **
(6.22) (3.99) (3.58) (2.45)
Log account volume 0.001 ***
0.001 ***
0.023 ***
0.021 ***
(39.58) (31.94) (26.96) (20.47)
Constant 0.007 ***
0.008 ***
0.066 ***
0.064 ***
(38.35) (32.36) (5.22) (4.12)
Observations 28,321 28,321 28,321 28,321
R-squared 0.12 0.31 0.05 0.25
Zip code dummies No Yes No Yes
Log account volume is measured in January 2003. All regressions include regional dummies (absorbed by zip code dummies where present).
Asymptotic t-statistics corrected for clustering at the zip code level are reported in parentheses.
* Statistical significance at the 10% level.
** Statistical significance at the 5% level.
*** Statistical significance at the 1% level.
linked to the size, and not merely to the frequency of purchases.
The third and fourth columns of Table 6 show a positive and
strongly statistically significant effect of IFAs on average account
turnover. This could be part of the explanation for why IFAs contribute
negatively to portfolio returns. 11 Again, males are more
likely to have larger account turnover and more experienced investors
are less likely to turn over their portfolio frequently. Younger
investors, between 30 and 60 years of age, are estimated to have
higher purchase turnovers, as they actively expand their portfolios.
A different perspective on the role of IFAs applies to encouraging
diversification. Columns 5 and 6 of Table 6 report a negative IFA
effect on the average share of directly held stocks in the account,
even after characteristics of account holders and areas are controlled
for. 12 This finding is consistent both with the descriptive statistics
at the start of the paper and the incentive of IFAs to sell
mutual funds. It is also one channel through which the reduction
in portfolio variance that we found in Table 5 is likely to be accomplished
by IFAs.
Controlling for other factors, males tend to put larger shares of
their account in directly held stocks, suggesting overconfidence in
portfolio behavior, in addition to the gender effects on frequency of
trading and on the volume of purchases. 13 Interestingly, experience
tends to lower the share of directly held stocks, dampening overcon-
11 Higher turnover might be motivated simply by commissions but also by an
incentive of IFAs to justify their fees by rebalancing client portfolios (see e.g.
Lakonishok et al., 1992).
12 Since the share of single stocks is bound between zero and one, we also run probit
regressions. Results deliver very similar results to the OLS estimates, and are available
on request.
13 Being married tends to have the opposite effect, presumably because more people
are at risk and maybe vocal in encouraging diversification. Employees and selfemployed
account owners tend to invest more in directly held stocks, probably
because of their increased social interactions.
A. Hackethal et al. / Journal of Banking & Finance 36 (2012) 509–524 519
fidence rather than encouraging account owners to manage direct
investments in stocks. The conclusion from the regression analysis
is that IFAs seem to boost the volume of purchases, while reducing
the fraction of the account invested in directly held stocks.
7. Bank financial advisors and portfolio performance
Given the rather striking nature of the estimated contribution of
IFAs to portfolio performance, the question arises as to whether
our results are specific to brokerage accounts, e.g. because of selectivity
into these types of accounts, or because financial advisors are
independent and not accountable to the financial institution. For
this reason, we consider a second data set of investment accounts,
this time from a large German commercial bank.
Our discussions with the brokerage and the commercial bank
suggest that there are important similarities and differences between
incentives facing IFAs and BFAs. For example, upfront loads
for mutual funds, a key component of any incentive scheme, are
typically fixed by the mutual fund producer and therefore identical
for all sales organizations. Although the bank does not funnel all
commissions through to its BFAs, it gives powerful non-monetary
incentives to its sales force through its sales control system. 14 On
the other hand, IFAs are not subject to the constraints imposed by
banks on BFAs. In fact, many banks not only narrow down the menu
of financial products offered to investors, but also provide extra
incentives for their agents to advise clients to purchase funds or
structured products produced by the bank itself or by one of its
subsidiaries. 15 We expect the negative association between BFA
14
Another similarity refers to legal fines for any detected misselling. Since they are
a function of the loss to the client they should be identical across IFAs and banks.
15
Yoong and Hung (2009) extend Ottaviani and Inderst (2009) to address this kind
of self-dealing.

520 A. Hackethal et al. / Journal of Banking & Finance 36 (2012) 509–524
Table 6
The determinants of trading, turnover and diversification in the brokerage sample.
Number of trades Turnover rate Share of single stocks
(1) (2) (3) (4) (5) (6)
Financial advisor (IFA) 0.008 ***
0.009 **
0.062 ***
0.053 ***
0.353 ***
0.341 ***
(2.59) (2.14) (16.77) (15.21) (47.19) (37.72)
Male 0.035 ***
0.036 ***
0.018 ***
0.017 ***
0.077 ***
0.079 ***
(16.04) (12.32) (16.10) (13.15) (14.71) (12.37)
Married 0.001 0.001 0.001 0.000 0.027 ***
0.029 ***
(0.32) (0.19) (1.37) (0.14) (6.03) (5.03)
Employee 0.010 0.015 0.001 0.001 0.082 ***
0.068 **
(0.80) (0.89) (0.19) (0.26) (2.94) (1.99)
Self-employed 0.006 0.012 0.004 0.005 0.128 ***
0.108 ***
(0.46) (0.71) (0.99) (0.91) (4.47) (3.10)
Experience/100 0.020 0.003 0.073 ***
0.054 ***
0.571 ***
0.556 ***
(1.07) (0.12) (8.58) (5.42) (15.22) (11.91)
30 < Age 6 40 0.005 0.004 0.004 **
0.003 0.000 0.001
(1.39) (0.80) (2.05) (1.17) (0.03) (0.04)
40 < Age 6 50 0.015 ***
0.016 ***
0.010 ***
0.009 ***
0.028 **
0.026 *
(4.02) (2.95) (4.29) (3.44) (2.50) (1.88)
50 < Age 6 60 0.024 ***
0.026 ***
0.016 ***
0.014 ***
0.061 ***
0.064 ***
(5.35) (4.28) (6.53) (4.78) (5.06) (4.33)
Age > 60 0.020 ***
0.021 ***
0.011 ***
0.011 ***
0.077 ***
0.072 ***
(4.11) (3.17) (4.28) (3.50) (6.27) (4.74)
Log account volume 0.015 ***
0.016 ***
0.009 ***
0.007 ***
0.027 ***
0.030 ***
(13.23) (10.75) (15.95) (12.03) (17.05) (14.69)
Constant 0.097 ***
0.093 ***
0.101 ***
0.089 ***
0.689 ***
0.735 ***
(6.21) (4.57) (16.08) (12.43) (21.07) (18.62)
Observations 28,303 28,303 28,321 28,321 28,321 28,321
R-squared 0.02 0.21 0.07 0.32 0.14 0.32
Zip code dummies No Yes No Yes No Yes
Number of trades is expressed as a fraction of account volume in ‘000. Log account volume is measured in January 2003. All regressions include regional dummies (absorbed
by zip code dummies where present). Asymptotic t-statistics corrected for clustering at the zip code level are reported in parentheses.
* Statistical significance at the 10% level.
** Statistical significance at the 5% level.
*** Statistical significance at the 1% level.
Table 7
The determinants of portfolio returns, variance and Sharpe ratio in the bank sample.
Log monthly returns Variance of monthly returns Sharpe ratio
(1) (2) (3) (4) (5) (6)
Financial advisor (BFA) 0.003 ***
0.003 ***
0.001 ***
0.001 *
0.180 ***
0.178 ***
(9.20) (8.07) (2.80) (1.92) (12.20) (10.20)
Male 0.001 **
0.001 *
0.001 ***
0.001 ***
0.031 **
0.033 **
(2.18) (1.75) (3.05) (2.63) (2.37) (2.13)
Employee 0.001 ***
0.001 *
0.001 0.001 0.008 0.005
(2.59) (1.88) (1.49) (0.92) (0.47) (0.25)
Executive 0.001 0.002 0.002 0.002 0.013 0.030
(1.37) (1.57) (1.46) (1.33) (0.31) (0.65)
Housewife 0.001 0.001 *
0.000 0.000 0.008 0.020
(1.58) (1.70) (0.19) (0.14) (0.34) (0.72)
Retired 0.000 0.000 0.000 0.000 0.010 0.001
(0.27) (0.48) (0.22) (0.58) (0.53) (0.04)
30 < Age 6 40 0.001 *
0.001 0.000 0.000 0.034 0.029
(1.86) (1.49) (0.09) (0.14) (1.21) (0.86)
40 < Age 6 50 0.000 0.000 0.002 *
0.001 0.023 0.010
(0.41) (0.32) (1.87) (1.46) (0.91) (0.33)
50 < Age 6 60 0.000 0.001 0.000 0.001 0.023 0.015
(0.64) (1.08) (0.52) (0.65) (0.81) (0.46)
Age > 60 0.002 **
0.002 **
0.001 0.001 0.028 0.012
(2.58) (2.33) (0.88) (1.21) (1.10) (0.40)
Log account volume 0.001 ***
0.000 ***
0.001 ***
0.001 ***
0.023 ***
0.024 ***
(5.44) (3.81) (8.04) (7.15) (5.08) (4.85)
Constant 0.001 0.003 **
0.013 ***
0.013 ***
0.197 ***
0.210 ***
(1.01) (2.18) (8.47) (8.58) (4.33) (4.48)
Observations 4440 4440 4440 4440 4440 4440
R-squared 0.06 0.20 0.04 0.19 0.05 0.19
Zip code dummies No Yes No Yes No Yes
Log account volume is measured in January 2003. All regressions include regional dummies (absorbed by zip code dummies where present). Asymptotic t-statistics corrected
for clustering at the zip code level are reported in parentheses.
* Statistical significance at the 10% level.
** Statistical significance at the 5% level.
*** Statistical significance at the 1% level.

Table 8
The determinants of turnover and diversification in the bank sample.
Turnover rate Share of single stocks
(1) (2) (3) (4)
Financial advisor (BFA) 0.099 ***
0.121 ***
0.158 ***
0.148 ***
(6.11) (5.91) (16.31) (13.22)
Male 0.072 ***
0.075 ***
0.056 ***
0.055 ***
(4.72) (4.37) (5.88) (4.85)
Employee 0.041 **
0.019 0.045 ***
0.037 ***
(1.98) (0.82) (3.66) (2.59)
Executive 0.147 **
0.137 **
0.046 0.054 *
(2.43) (2.02) (1.63) (1.72)
Housewife 0.035 0.029 0.010 0.004
(1.07) (0.76) (0.62) (0.19)
Retired 0.038 0.023 0.016 0.019
(1.54) (0.84) (1.16) (1.16)
30 < Age 6 40 0.011 0.046 0.072 ***
0.079 ***
(0.38) (1.32) (3.69) (3.52)
40 < Age 6 50 0.016 0.038 0.088 ***
0.099 ***
(0.56) (1.19) (4.69) (4.45)
50 < Age 6 60 0.051 *
0.069 **
0.053 ***
0.061 ***
(1.87) (2.14) (3.04) (3.02)
Age > 60 0.012 0.016 0.014 0.029
(0.47) (0.55) (0.85) (1.58)
Log account volume 0.065 ***
0.062 ***
0.006 ***
0.008 ***
(14.85) (12.42) (2.70) (2.90)
Constant 0.339 ***
0.368 ***
0.275 ***
0.267 ***
(6.53) (7.07) (10.04) (9.73)
Observations 4440 4440 4440 4440
R-squared 0.10 0.26 0.13 0.26
Zip code dummies No Yes No Yes
Log account volume is measured in January 2003. All regressions include regional
dummies (absorbed by zip code dummies where present). Asymptotic t-statistics
corrected for clustering at the zip code level are reported in parentheses.
* Statistical significance at the 10% level.
** Statistical significance at the 5% level.
*** Statistical significance at the 1% level.
use and portfolio returns to be even stronger than in the brokerage
sample.
As with the brokerage sample, we introduce regional and zip
code dummies to capture unobserved heterogeneity. Columns 1
and 2 in Table 7 present OLS results on raw returns, where use of
a BFA is seen to have a statistically significant negative effect.
According to this model, BFAs reduce monthly log returns by
0.3% points, slightly less than IFAs in Table 4 ( 0.4%). However, unlike
IFAs, who were found to reduce overall portfolio risk, BFAs are
found in columns 3 and 4 of Table 7 to increase total portfolio risk.
Given the negative BFA effect on returns and their positive effect
on total risk, we expect a strong negative effect on Sharpe ratios,
and this is confirmed by the last two columns in Table 7. 16 This
pronounced negative effect of BFAs is consistent with our conjecture
from above and with Inderst and Ottaviani (2009) who posit that
advisory standards should be lower for BFAs than for IFAs because
the latter face no internal agency conflicts with costly monitoring.
Consistent with results on IFAs, males exhibit lower returns and
riskier portfolios. The initial size of the account contributes to
higher returns, in levels or normalized by risk, and to lower portfolio
variance. Columns 1 and 2 of Table 8 point to higher turnover
rates (based on purchases) for BFA accounts. The estimated BFA
coefficients are larger than the corresponding IFA ones (Table 6),
suggesting lower advisory standards for BFAs than for IFAs.
Finally, regressions reported in columns 5 and 6 of Table 8 confirm
that BFAs, as indeed IFAs, tend to push towards investing in
mutual funds, consistent with their compensation incentives. All
in all, our analysis of the bank sample produces remarkably consis-
16 The coefficients of the BFA dummy are negative ( 0.18), with values considerably
larger than the corresponding IFA coefficients in Table 5 ( 0.05 in columns 3 and 4).
A. Hackethal et al. / Journal of Banking & Finance 36 (2012) 509–524 521
tent results with the brokerage sample and points to systematic
negative effects of financial advisors rather than to statistical flukes
or sample peculiarities.
8. Conclusions
We investigate who tends to use a financial advisor, whether
investors tend to produce better account performance on their
own rather than with the help of financial advisors, whether results
depend on the advisory model (IFA versus BFA), and whether
they can be traced to trading behavior and security choice. We also
examine robustness of our findings with respect to asset pricing
model, dummies for wider regions or zip codes, control variables,
and estimation procedure (OLS versus IV).
Our first data set tracks accounts of a major brokerage firm,
some of which are run with the help of an independent financial
advisor (IFA). Sample statistics and regression analysis show that
advisors tend to be matched with wealthier, older, more experienced,
and female investors rather than with poorer, younger
and inexperienced ones. Our second data set comes from a major
commercial bank with branches throughout the country. We find
that also bank clients who tend to consult a bank employee prior
to executing a trade are older, wealthier and more likely to be
female.
Descriptive statistics as well as regression analysis that controls
for investors’ characteristics and characteristics of the region of the
account paint a very similar picture of the role of IFA and BFAs in
account performance. In both samples, advised accounts offer lower
returns than those run by similar investors without advisor input.
Although IFA use reduces total portfolio risk, it still reduces
ex post Sharpe ratios significantly. BFAs increase portfolio variance,
lowering Sharpe ratios even more.
Trading costs and associated commissions earned by both IFAs
and banks certainly contribute to these outcomes, since we find
that advised accounts feature higher portfolio turnover (though
not necessarily more frequent trading) relative to self-managed accounts.
Consistent with their remuneration incentives, financial
advisors tend to encourage lower account shares in directly held
stocks. Robustness analysis suggests that our results on the negative
role of IFAs are not an artifact of endogeneity between account
performance and advisor use, nor of the way we adjust portfolio
net returns for systematic risk.
Our results provide a new perspective on the role of financial
advisors that might be useful for theoretical and policy analysis
of their conflicting incentives, their likely effects, and the need to
regulate them. Based on our findings, it should not be taken for
granted that financial advisors provide their services to small,
young investors typically identified as in need of investment guidance.
Indeed, the opposite is true both for the broker and for the
bank data we consider. The finding stands to reason: financial
advisors with commission-based incomes naturally prefer to devote
time to customers likely to trade on a bigger scale. However,
it also creates doubts as to how viable financial advice is as a solution
to the problem of limited financial literacy in the population.
In view of the rapidly growing literature on investment mistakes,
providing financial advice to inexperienced, naïve investors could
be an alternative to trying to educate them in financial matters,
but financial advisor incentives and tendencies of inexperienced
clients might result in relatively few matches. Other alternatives,
such as simpler products and carefully designed default options,
may be more promising than currently existing forms of financial
advice in averting negative distributional consequences.
Our findings imply that many financial advisors end up collecting
more in fees and commissions than any monetary value they
add to the account. This raises the further question of whether

522 A. Hackethal et al. / Journal of Banking & Finance 36 (2012) 509–524
advisors overcharge and should be regulated. While the case for
regulation seems much clearer when advisors are matched with
inexperienced investors, negative effects appear even when the
tendency is for experienced investors to be using an advisor. In
such cases, it may be that investors are inattentive and fail to monitor
the advisors effectively; or that they face high opportunity
costs of running accounts by themselves and are willing to pay a
luxury premium to have their advisors run their accounts. What
distinguishes these two cases is customer awareness of the financial
advisor incentives and effects. Even if regulation is not warranted
in both cases, transparency and information on the role
and outcome of financial advice seem crucial. Moreover, this need
is not limited to naïve, inexperienced customers but extends also
to older, experienced ones. Thus, questionnaires on investor experience,
such as those dictated by MIFID (the EU directive aimed at
increasing financial markets transparency and competition),
should not waive the need for information regarding incentives,
ensuing conflicts of interest, and the outcomes of professional
financial advice in terms of portfolio returns and risks, even for
experienced investors.
Acknowledgements
We are grateful to the following individuals for very helpful
comments and suggestions: the editor, an anonymous referee, Raquel
Carrasco, Guenter Franke, Dimitris Georgarakos, Luigi Guiso,
Jose Martinez, Pedro Mira, Steven Stern, John Rust; participants
in the November 2008 Fundación Ramón Areces conference on
Population Ageing and its Economic Consequences in Madrid, the
2009 SIFR-NETSPAR Pension workshop in Stockholm, the 2009 IM-
AEF conference in Ioannina, the 2009 CSEF-IGIER Symposium in
Economics and Institutions in Capri, the SAVE Conference in Mannheim,
the Bank of Spain 2009 Household Finance Conference; and
in the seminar series at the Universities of Exeter, Konstanz, Leicester,
Piraeus, and Central European University in Budapest. We
would like to thank Ralph Bluethgen and Yigitcan Karabulut for
outstanding research assistance and the two financial institutions
for providing us with the administrative data. Haliassos acknowledges
research funding from the German Research Foundation
(Deutsche Forschungsgemeinschaft) under the project Regulating
Retail Finance. Jappelli acknowledges funding from the Italian Ministry
of Universities and Research.
Appendix A. Endogeneity of financial advice
Although we define as advised portfolios those portfolios that
are continuously assisted from 2003 to 2006, if performance is persistent
over time one may suspect that portfolio performance actually
induced the choice of the advisor. As mentioned in Section 6,in
such a case, OLS regressions are problematic. In this appendix, we
carry out IV estimation to examine the robustness of our main
findings to possible endogeneity of financial advice. Finding suitable
instruments in our context is not easy, and the robustness
exercise described below, based on regional variation, is indicative
and unavoidably rests on the validity of the identification
assumption.
We choose to exploit regional variability in the share of financial
services over GDP, and merge our two data sets with administrative
data available from the German Federal Statistical Office, which provides
a broad set of structural data on some 500 local areas. The system
of German zip codes is more granular than the regional grid
used by the Federal Statistical Office. Accordingly, we map the zip
codes for customer accounts into the regional grid of the Statistical
Office by assuming that zip-codes from the same region share the
same structural characteristics.
We motivate the use of the GDP share of financial services by
reference to the potential role that the density of financial services
plays in reducing the cost of gathering financial information: the
greater such density, the more likely it is that investors are able
to gather information from local sources more cheaply, substituting
for financial advice. Our identification assumption is that regional
proximity with financial intermediaries affects account
performance by facilitating the matching of account holders to
financial advisors, but not directly. The instrument is collinear with
the zip code dummies, which cannot be used in the estimation.
Note however that in the IV estimates we still use wider regional
dummies.
Table A1 reports the first-stage estimates for IFA and for BFA,
and shows that being located in an area with a larger density of
intermediaries reduces the probability of having an account run
by an IFA or a BFA. In both cases, the coefficients are statistically
different from zero (at the 1% and 5% level, respectively).
The IV estimates for portfolio performance are reported in Table
A2. They are based on a standard IV estimator (with a linear probability
model in the first stage) and standard errors adjusted for
clustering at the local area level. With only one instrument, the
model is exactly identified and we cannot provide a test of overidentification
restrictions. However, we do find that the instrument
has statistically significant impact on use of IFA (the F-statistic is
reported in the last row of Table A2). The signs of the instrumented
Table A1
First-stage results: the determinants of having the account run by an IFA or a BFA.
Brokerage sample
(IFAs)
Male 0.063 ***
Bank sample
(BFAs)
0.049 ***
(12.32) (3.16)
Employee 0.058 **
0.048 ***
(2.49) (2.61)
30 < Age 6 40 0.030 ***
0.049
(3.11) (1.41)
40 < Age 6 50 0.002 0.011
(0.25) (0.33)
50 < Age 6 60 0.023 **
0.055 *
(2.21) (1.74)
Age > 60 0.089 ***
0.113 ***
(7.62) (3.89)
Log account volume 0.046 ***
0.028 ***
(22.60) (7.73)
Self-employed 0.062 **
(2.56)
Experience/100 0.165 ***
(3.58)
Married 0.025 ***
(6.06)
Executive 0.001
(0.01)
Housewife 0.005
(0.18)
Retired 0.022
(0.95)
Financial services/regional
GDP
0.238 ***
0.324 **
(6.58) (2.53)
Constant 0.301 ***
0.446 ***
(9.85) (7.54)
Observations 28,321 4440
R-squared 0.10 0.05
The table reports estimates from a linear probability model of having an independent
financial advisor or a bank financial advisor. Log account volume is measured
in January 2003. The regressions include regional dummies. Asymptotic standard
errors corrected for clustering at the zip code level are reported in parentheses.
* Statistical significance at the 10% level.
** Statistical significance at the 5% level.
*** Statistical significance at the 1% level.

Table A2
Instrumental variable regressions for the brokerage and bank sample.
Brokerage sample (IFAs) Bank sample (BFAs)
Log monthly returns Variance of monthly returns Sharpe ratio Log monthly returns Variance of monthly returns Sharpe ratio
(1) (2) (3) (4) (5) (6)
Financial advisor 0.012 ***
0.002 **
0.166 ***
-0.023 ***
0.003 0.529 *
(3.68) (2.54) (3.27) (2.60) (0.34) (1.73)
Male 0.001 ***
0.000 ***
0.029 ***
0.000 0.001 **
0.048 **
(4.20) (6.99) (7.64) (0.38) (2.27) (2.30)
Employee 0.000 0.000 **
0.015 0.000 0.001 0.026
(0.54) (2.14) (1.31) (0.12) (1.17) (1.12)
30 < Age 6 40 0.000 0.000 0.003 0.000 0.000 0.017
(0.80) (1.08) (0.53) (0.43) (0.18) (0.49)
40 < Age 6 50 0.000 0.000 ***
0.013 ***
0.000 0.002 *
0.019
(0.44) (5.31) (2.59) (0.06) (1.89) (0.64)
50 < Age 6 60 0.000 0.000 ***
0.011 **
0.001 0.000 0.041
(0.69) (5.97) (2.10) (0.59) (0.32) (1.16)
Age > 60 0.000 0.001 ***
0.009 0.000 0.001 0.067
(0.44) (5.57) (1.21) (0.34) (0.37) (1.45)
Log account volume 0.001 ***
0.000 ***
0.028 ***
0.001 ***
0.001 ***
0.032 ***
(4.66) (11.75) (11.60) (4.37) (4.09) (3.42)
Self-employed 0.001 0.001 ***
0.027 **
(0.86) (5.23) (2.31)
Experience/100 0.003 ***
0.000 0.073 ***
(2.97) (0.39) (4.24)
Married 0.000 0.000 ***
0.004
(0.58) (7.58) (1.53)
Executive 0.001 0.002 0.014
(0.95) (1.47) (0.32)
Housewife 0.001 0.000 0.009
(1.13) (0.18) (0.34)
Retired 0.001 0.000 0.019
(0.90) (0.13) (0.83)
Constant 0.005 ***
0.007 ***
0.027 0.009 **
0.012 ***
0.325 ***
(4.02) (18.28) (1.20) (2.31) (3.28) (2.72)
Observations 28,321 28,321 28,321 4440 4440 4440
Cragg–Donald F-stat. 43.30 43.30 43.30 11.23 11.23 11.23
The instrument is the GDP ratio of the financial asset share at the zip code level. The first stage regressions are reported in column 1 (brokerage sample) and column 2 (bank
sample) of Table A1. Log account volume is measured in January 2003. All regressions include regional dummies. Asymptotic t-statistics corrected for clustering at the zip
code level are reported in parentheses.
* Statistical significance at the 10% level.
** Statistical significance at the 5% level.
*** Statistical significance at the 1% level.
advisor effects on returns, variance and Sharpe ratios remain qualitatively
unchanged as compared to the OLS results, for both the
brokerage and bank sample. In particular, we find that advised accounts
yield lower returns and Sharpe ratios, and that the negative
effects are larger in the bank sample. Moreover, factors such as
being female, experienced, and wealthier still contribute to higher
returns, lower variance and higher Sharpe ratios.
References
Anderson, A., Martinez, J.V., 2008. Vice or advice? Profits from brokerage firm
trading around recommendation revision dates. Working paper, Stockholm
University.
Barber, B.M., Loeffler, D., 1993. The ‘Dartboard’ column: second-hand information
and price pressure. Journal of Financial and Quantitative Analysis 28, 273–284.
Barber, B.M., Odean, T., 2000. Trading is hazardous to your wealth: the common
stock investment performance of individual investors. Journal of Finance 55,
773–806.
Barber, B.M., Odean, T., 2001. Boys will be boys: gender, overconfidence, and
common stock investment. Quarterly Journal of Economics 116, 261–292.
Barber, B.M., Lehavy, R., McNichols, M., Trueman, B., 2001. Can investors profit from
the prophets? Security analysts recommendations and stock returns. Journal of
Finance 56, 531–563.
Bergstresser, D., Chalmers, J., Tufano, P., 2009. Assessing the costs and benefits of
brokers: a preliminary analysis of the mutual fund industry. Review of Financial
Studies 22, 4129–4156.
Campbell, J.Y., 2006. Household finance. Journal of Finance 61, 1553–1604.
Campbell, J.Y., Viceira, L.M., 2003. Strategic Asset Allocation – Portfolio Choice for
Long-Term Investors. Oxford University Press, New York.
A. Hackethal et al. / Journal of Banking & Finance 36 (2012) 509–524 523
Carhart, M.M., 1997. On the persistence in mutual fund performance. Journal of
Finance 52, 57–82.
Cowles, A., 1933. Can stock market forecasters forecast? Econometrica 1, 309–324.
Desai, H., Jain, P.C., 1995. An analysis of the recommendations of the ‘Superstar’ money
managers at Barron’s annual roundtable. Journal of Finance 50, 1257–1274.
Dietz, P.O., 1968. Components of a measurement model: rate of return, risk, and
timing. Journal of Finance 23, 267–275.
Dorn, D., Huberman, G., 2005. Talk and action: what individual investors say and
what they do. Review of Finance 9, 437–481.
Feng, L., Seasholes, M., 2005. Do investor sophistication and trading experience
eliminate behavioral biases in financial markets? Review of Finance 9, 305–351.
Goetzmann, W.N., Kumar, A., 2008. Equity portfolio diversification. Review of
Finance 12, 433–463.
Grinblatt, M., Keloharju, M., 2001. What makes investors trade? Journal of Finance
56, 589–616.
Guiso, L., Jappelli, T., 2006. Information acquisition and portfolio performance. CSEF
Working Paper No. 167.
Inderst, R., Ottaviani, M., 2009. Misselling through agents. American Economic
Review 99, 883–908.
Lakonishok, J., Shleifer, A., Vishny, R.W., Hart, O., Perry, G.L., 1992. The structure and
performance of the money management industry. Brookings Papers on
Economic Activity: Microeconomics, 339–391.
Lusardi, A., Mitchell, O.S., 2007. Baby boomer retirement security: the roles of
planning, financial literacy, and housing wealth. Journal of Monetary Economics
54, 205–224.
Metrick, A., 1999. Performance evaluation with transactions data: the stock
selection of investment newsletters. Journal of Finance 54, 1743–1774.
Odean, T., 1998. Are investors reluctant to realize their losses? Journal of Finance
53, 1775–1798.
Odean, T., 1999. Do investors trade too much? American Economic Review 89,
1279–1298.
Shapira, Z., Venezia, I., 2001. Patterns of behavior of professionally managed and
independent investors. Journal of Banking and Finance 25, 1573–1587.

524 A. Hackethal et al. / Journal of Banking & Finance 36 (2012) 509–524
Shu, P., Chiu, S., Chen, H., Yeh, Y., 2004. Does trading improve individual investor
performance? Review of Quantitative Finance and Accounting 22, 199–217.
Stulz, R.M., Mehran, H., 2007. The economics of conflicts of interest in financial
institutions. Journal of Financial Economics 85, 267–296.
Womack, K.L., 1996. Do brokerage analysts recommendations have investment
value? Journal of Finance 51, 137–167.
Venezia, I., Nashikkar, A., Shapira, Z., 2011. Firm specific and macro herding by
professional and amateur investors and their effects on market volatility.
Journal of Banking and Finance 35, 1599–1609.
Yoong, J., Hung A., 2009. Self-dealing and compensation for financial advisors. RAND
Working Paper Series, WR-713.

Philipp Schmitt, Bernd Skiera, & Christophe Van den Bulte
© 2011, American Marketing Association
ISSN: 0022-2429 (print), 1547-7185 (electronic)
Referral Programs and
Customer Value
Referral programs have become a popular way to acquire customers. Yet there is no evidence to date that
customers acquired through such programs are more valuable than other customers. The authors address this gap
and investigate the extent to which referred customers are more profitable and more loyal. Tracking approximately
10,000 customers of a leading German bank for almost three years, the authors find that referred customers (1)
have a higher contribution margin, though this difference erodes over time; (2) have a higher retention rate, and
this difference persists over time; and (3) are more valuable in both the short and the long run. The average value
of a referred customer is at least 16% higher than that of a nonreferred customer with similar demographics and
time of acquisition. However, the size of the value differential varies across customer segments; therefore, firms
should use a selective approach for their referral programs.
Keywords: customer referral programs, customer loyalty, customer value, customer management, word of mouth,
social networks
Word of mouth (WOM) has reemerged as an important
marketing phenomenon, and its use as a customer
acquisition method has begun to attract
renewed interest (e.g., Godes and Mayzlin 2009; Iyengar,
Van den Bulte, and Valente 2011). Traditionally, WOM’s
appeal has been in the belief that it is cheaper than other
acquisition methods. A few recent studies have documented
that customers acquired through WOM also tend to churn
less than customers acquired through traditional channels
and that they tend to bring in additional customers through
their own WOM (Choi 2009; Trusov, Bucklin, and Pauwels
2009; Villanueva, Yoo, and Hanssens 2008). Villanueva,
Yoo, and Hanssens (2008) further suggest that customers
acquired through WOM can generate more revenue for the
firm than customers acquired through traditional marketing
efforts.
From a managerial point of view, these findings are
encouraging and suggest purposely stimulating WOM to
acquire more customers. However, there are concerns that
firm-stimulated WOM may be substantially less effective
than organic WOM in generating valuable customers
(Trusov, Bucklin, and Pauwels 2009; Van den Bulte 2010)
because (1) targeted prospects may be suspicious of stimu-
Philipp Schmitt is a doctoral student (e-mail: pschmitt@wiwi.uni-frankfurt.
de), and Bernd Skiera is Professor of Marketing and Member of the Board
of the E-Finance Lab at the House of Finance (e-mail: skiera@skiera.de),
School of Business and Economics, Goethe University Frankfurt.
Christophe Van den Bulte is Associate Professor of Marketing, the Wharton
School, University of Pennsylvania (e-mail: vdbulte@wharton.upenn.
edu). The authors thank the management of a company that wants to
remain anonymous for making the data available and Christian Barrot,
Jonah Berger, Xavier Drèze, Peter Fader, Jeanette Heiligenthal, Gary
Lilien, Renana Peres, Jochen Reiner, Christian Schulze, Russell Winer,
Ezra Zuckerman, and the anonymous JM reviewers for providing comments
on previous drafts of this article.
46
lated WOM efforts; (2) such efforts often involve a monetary
reward for the referrer, who, as a result, may seem less
trustworthy; (3) programs providing economic benefits tend
not to be sustainable (Lewis 2006); (4) unlike organic
WOM, stimulated WOM is not free, raising questions about
cost effectiveness; and (5) stimulated WOM is prone to
abuse by opportunistic referrers.
The uncertainty about the benefits of stimulated WOM
in customer acquisition is frustrating for managers facing
demands to increase their marketing return on investment
and considering whether to use this method. Our study
addresses this managerial issue by investigating the value of
customers acquired through stimulated WOM and comparing
it with the value of customers acquired through other
methods. We do so in the context of a specific WOM marketing
practice that is gaining prominence, namely, referral
programs in which the firm rewards existing customers for
bringing in new customers. Although these programs are
typically viewed as an attractive way to acquire customers,
their benefits are often viewed to be their targetability and
cost effectiveness (Mummert 2000). We broaden this view
by assessing the value of customers acquired through these
types of programs.
Specifically, we answer four questions: (1) Are customers
acquired through a referral program more valuable
than other customers? (2) Is the difference in customer
value large enough to cover the costs of such stimulated
WOM customer acquisition efforts? (3) Are customers
acquired through a referral program more valuable because
they generate higher margins, exhibit higher retention, or
both? and (4) Do differences in margins and retention
remain stable, or do they erode? The answers to the last two
questions provide deeper insight into what might be driving
the value differential.
We answer these four questions using panel data on all
5181 customers that a leading German bank acquired
Journal of Marketing
Vol. 75 (January 2011), 46 –59

through its referral program (referred customers) between
January 2006 and December 2006 and a random sample of
4633 customers the same bank acquired through other
methods (nonreferred customers) over the same period. For
both groups of customers, we track profitability (measured
as contribution margin) and loyalty (measured as retention)
at the individual level from the date of acquisition until September
2008. The total observation period spans 33 months.
We use two metrics of customer value: (1) the present value
of the actually observed contribution margins realized
within the data window and (2) the expected present value
over a period of six years from the day of acquisition.
Although our study is limited to a single research site, as is
common for studies that require rich and confidential data,
the methodology and findings are of broad interest. Customer
referral programs are gaining popularity in many
industries, including financial services, hotels, automobiles,
newspapers, and contact lenses (Ryu and Feick 2007).
We make the following contributions: First, we provide
empirical evidence that a referral program, a form of stimulated
WOM, is an attractive way to acquire customers.
Referred customers exhibit higher contribution margins,
retention, and customer value. Second, building on our finding
that differences in contribution margin erode over time
whereas those in retention do not, we document that
referred customers are more valuable in both the short and
the long run. Third, we show that the referral effect need not
be present in every customer segment. Finally, we illustrate
how the type of analysis we conduct enables firms to calculate
the return on investment and the upper bound for the
reward in their customer referral programs.
We proceed by describing referral programs and developing
our hypotheses. A description of the research setting,
the data, and the model specifications follows. Then, we
report the results. Finally, we discuss implications for practice,
the limitations, and opportunities for further research.
Customer Referral Programs
Customer referral programs are a form of stimulated WOM
that provides incentives to existing customers to bring in
new customers. An important requirement for such programs
is that individual purchase or service histories are
available so the firm can ascertain whether a referred customer
is indeed a new rather than an existing or a former
customer.
Referral programs have three distinctive characteristics.
First, they are deliberately initiated, actively managed, and
continuously controlled by the firm, which is impossible or
very difficult with organic WOM activities such as spontaneous
customer conversations and blogs. Second, the key
idea is to use the social connections of existing customers
with noncustomers to convert the latter. Third, to make this
conversion happen, the firm offers the existing customer a
reward for bringing in new customers.
Although leveraging the social ties of customers with
noncustomers to acquire the latter is not unique to customer
referral programs, the three distinctive characteristics of
these programs set them apart from other forms of networkbased
marketing (Van den Bulte and Wuyts 2007). Unlike
organic WOM, the firm actively manages and monitors
referral programs. Unlike most forms of buzz and viral marketing,
the source of social influence is limited to existing
customers rather than anyone who knows about the brand or
event. Unlike multilevel marketing, existing customers are
rewarded only for bringing in new customers. They do not
perform any other sales function (e.g., hosting parties) and
do not generate any income as a function of subsequent
sales. Consequently, referral programs do not carry the
stigma of exploiting social ties for mercantile purposes, as
multilevel marketing does (Biggart 1989).
In most referral programs, the reward is given regardless
of how long the new referred customers stay with the firm.
Such programs are prone to abuse by customers. Although
the firm does not commit to accept every referral, the incentive
structure combined with imperfect screening by the
firm creates the potential for abuse in which existing customers
are rewarded for referring low-quality customers.
This kind of moral hazard is less likely to occur with WOM
campaigns that do not involve monetary rewards conditional
on customer recruitment.
Existing studies of customer referral programs have
provided guidance about when rewards should be offered
(Biyalogorsky, Gerstner, and Libai 2001; Kornish and Li
2010), have quantified the impact of rewards and tie
strength on referral likelihood (Ryu and Feick 2007; Wirtz
and Chew 2002), and have quantified the monetary value of
making a referral (Helm 2003; Kumar, Petersen, and Leone
2007, 2010). The key managerial question of the (differential)
value of customers acquired through referral programs
has not yet been addressed.
Hypotheses
Because referral programs are a customer acquisition
method, an important metric to assess their effectiveness is
the value of the customers they acquire. Additional insights
come from investigating differences between referred and
nonreferred customers in contribution margins and retention
rates, the two main components of customer value (e.g.,
Gupta and Zeithaml 2006; Wiesel, Skiera, and Villanueva
2008).
Our hypotheses regarding these customer metrics of
managerial interest are informed by prior work in economics
and sociology on employee referral (e.g., Coverdill 1998;
Rees 1966), especially the work of Fernandez, Castilla, and
Moore (2000), Neckerman and Fernandez (2003), and
Castilla (2005) on the quality of employee referral programs.
These studies show that the benefits of such programs are
realized through distinct mechanisms, of which better
matching and social enrichment appear particularly relevant
to marketers. Better matching is the phenomenon that referrals
fit with the firm better than nonreferrals, and social
enrichment is the phenomenon that the relationship of the
referral to the firm is enriched by the presence of a common
third party (i.e., the referrer).
Customer and employees referral programs are likely to
be subject to similar mechanisms because they share the
three distinctive characteristics of having active management,
using the social connections of existing contacts, and
Referral Programs and Customer Value / 47

offering rewards with the risk of abuse. Selecting a new
employer or bank both are high-involvement decisions
involving a fair amount of uncertainty. Although some basic
banking products, such as checking accounts, are well
known to most customers, the wider set of financial services
that banks provide are considered experience goods rather
than search goods (e.g., Bolton, Freixas, and Shapiro 2007;
Parasuraman, Zeithaml, and Berry 1985). The recurrent
losses of many private investors indicate that many people
are not very skilled at assessing complex bank offerings.
We use the better matching and social enrichment
mechanisms to develop and motivate our hypotheses. However,
our goal is to document managerially relevant differences
in contribution margin, retention, and customer value
rather than to test those specific mechanisms. The mechanisms
are only possible explanations for the differences we
document.
Differences in Contribution Margin
Several characteristics of social dynamics in human networks
(e.g., Van den Bulte and Wuyts 2007) imply that
referred customers may match up with the firm better than
other newly acquired customers. The first is reciprocity.
Because referring customers receive a reward, norms of reciprocity
may make nonopportunistic customers feel obliged
to bring in new customers who they think may be valuable
to the firm (Gouldner 1960). This process explains Neckerman
and Fernandez’s (2003) finding that referrals for which
the referrer claimed a fee had lower turnover than referrals
for which no fee was claimed. The second social dynamic
likely to be at work is triadic balance. If the main function
of the program is simply to nudge customers into making
referrals without much consideration for the monetary
reward (Thaler and Sunstein 2008), principles of triadic balance
will make existing customers more likely to bring in
others who they believe would match well with what the
firm has to offer. A third social dynamic likely to be at work
is homophily—the tendency for people to interact with
people like them. Whereas reciprocity and triadic balance
imply that referrers are diligent and active in screening and
matching peers to firms, homophily implies that customers
are likely to refer others who are similar to themselves.
Because existing customers have an above-average chance
of being a good match (otherwise, they would not be customers),
firms may benefit from referral programs through
“passive” homophily-based matching rather than only
deliberate “active” screening-based matching by referrers
(Kornish and Li 2010; Montgomery 1991; Rees 1966).
Acquisition through referral can also result in informational
advantages, making referred customers more profitable
than other customers. Referred customers are likely
to have discussed the firm’s offerings with their referrer. As
a result, they are likely to use its products more extensively
than novice customers who take a more cautious approach
in building involvement. Informational advantages to the
firm can also result if people refer others similar to themselves
on dimensions that are relevant to the enjoyment of
the product or service but are not immediately observable to
the firm (Kornish and Li 2010). Examples for financial services
include risk aversion and a sense of fiscal responsibil-
48 / Journal of Marketing, January 2011
ity. In these situations, the firm can make inferences from
the observed behavior of the referrers about the products in
which the referred customers will be most interested (e.g.,
Guseva 2008). As a result, the firm is able to serve the
referred customer in a tailored way early on, something that
takes time to learn for other newly acquired customers.
Because of this informational advantage, the firm should be
able to generate a higher contribution margin from referred
customers at the beginning of the relationship.
However, the advantages of better matching and better
information should gradually vanish. As nonreferred customers
accumulate experience with the firm, they should
become equally well informed about the firm’s offerings
and procedures. Likewise, the firm should be able to use the
purchase and service history of the nonreferred customers
to serve them better. Furthermore, nonreferred customers
who are not a good match for the firm are more likely to
leave. Consequently, both revenues and costs of referred
and nonreferred customer should converge, eliminating the
difference in contribution margin over time. Thus:
H 1: (a) The average contribution margin of a customer acquired
through a referral program is higher than that of a customer
acquired through other methods, but (b) this difference
becomes smaller over time.
Differences in Retention
Social enrichment is another mechanism that may increase
the value of referred customers. The argument is that the
relationship with the firm is enriched because a family
member or friend is a customer of the same firm (Castilla
2005; Fernandez, Castilla, and Moore 2000). Having a person
close to oneself in a similar position (i.e., being a customer
of the same firm) should increase the person’s trust in
the firm and strengthen his or her emotional bond with it, as
both balance theory and social closure theory predict (Van
den Bulte and Wuyts 2007). This prediction is also consistent
with findings that customers reflecting on their affect toward
a firm mention friends who are customers with the same
firm (Yim, Tse, and Chan 2008). Such relationships should
be particularly relevant in a banking context, in which emotions
and trust play important roles in the customer–firm
relationship (e.g., Edwards and Day 2005; Fleming, Coffman,
and Harter 2005). In short, referred customers are
likely to have a stronger sense of commitment and attachment
to the firm. This implies that referred customers are
less likely to churn than nonreferred customers, provided
that their referrer does not churn either. The latter condition
is likely to hold: Referrers typically have a greater longterm
likelihood of staying, which is why intention to refer is
frequently used as an indicator of loyalty (Gupta and Zeithaml
2006).
Although the informational advantage of a referred customer
decreases over time as direct experience substitutes
for social learning, there is no reason to expect erosion in
the affective bonding underlying the social enrichment
mechanism. Consequently, the erosion of the differential
expected in contribution margin need not apply to retention.
Therefore, we state the following:

H 2: (a) The average retention of a customer acquired through a
referral program is higher than that of a customer acquired
through other methods, and (b) this difference does not
become smaller over time.
Differences in Customer Value
If H1 and H2 hold and if the erosion of contribution margins
does not outweigh the initial difference in margins and the
persisting difference in retention, the following should hold
as well:
H 3: The average value of a customer acquired through a referral
program is higher than that of a customer acquired
through other methods.
H 3 can hold even when H 1 and H 2 do not, because it is possible
for the differences in both margins and retention to be
small but for their combined effect to be sizable and significant.
Another reason to test H 3 on customer value, in addition
to H 1 and H 2 on margins and retention, is that customer
value is what managers should care about most.
Although we base our prediction on sound theoretical
arguments, the posited effects are not as obvious as they
may seem because there are several competing forces at
work. First, the prospect of earning a referral fee may
induce referrers to pressure their peers to become customers
without giving much consideration to whether the firm
actually matches their peers’ needs. Second, the relationship
between the referred customer and his or her referrer, which
is necessary for social enrichment to operate, comes with an
inherent risk: When referrers defect, the customers they
brought in may become more likely to leave as well.
Although it seems unlikely that referrers as a whole are
more prone to churn than the average customer, the risk of
contagious defection should not be ignored altogether.
Third, an abuse of the referral program by customers who
are interested only in the monetary reward is probably the
most important reason for practitioners’ skepticism. This is
illustrated by TiVo’s termination of its referral program as a
result of frequent abuses (ZatzNotFunny 2008).
Support for our hypotheses would allow us to conclude
that the positive effects outweigh the negative ones. In addition,
the empirical analysis provides not only a test of the
hypotheses but also an estimate of the size of the customer
value differential. Firms can use the latter to put a maximum
value on the reward to be paid out as part of their
referral program.
Methods
Research Setting
We use data from a leading German bank, whose name we
do not divulge for confidentiality reasons. The data capture
all customers acquired through the bank’s referral program
between January 2006 and December 2006 and a random
sample of customers acquired through other methods (e.g.,
direct mail, advertising) over the same period. The latter
group may include customers affected by organic WOM for
which the bank did not pay any fee. To the extent that the
value of customers acquired through organic WOM is equal
to or greater than that of customers acquired through the
referral program, our results underestimate the value differential
between WOM and non-WOM customers. Regardless,
we correctly estimate the value differential between
customers acquired through the referral program and all
other customers for whom no referral fee was paid.
The observation period spans from January 2006 to
September 2008 (33 months), and the data on each individual
customer include the day of acquisition, the day of leaving
the bank (if applicable), the contribution margin in each
year, and some demographics. In total, we have data on
5181 referred and 4633 nonreferred customers. Because the
referral program was used only in a business-to-consumer
context, all customers are individual people.
The referral program was communicated to existing
customers through staff and flyers in the branches and
through mailings. 1 The procedure was straightforward:
Every existing customer who brought in a new customer
received a reward of 25 euros in the form of a voucher that
could be used at several well-known German retailers. 2
Except for opening an account, the referred customer did not
need to meet any prerequisites (e.g., minimum amount of
assets, minimum stay) for the referrer to receive the reward.
In addition, 2006 was not an unusual year in terms of
customer acquisition, and the bank’s management was confident
that findings about customers acquired in 2006 would
be applicable to customers acquired in earlier and later
years. Proprietary information of the bank shows that its
customers are similar to those of other leading European
banks. Regarding the usage of its referral program and the
response of its customers to it, no differences with other
firms are apparent either.
Dependent Variables
Daily contribution margin is the individual contribution
margin on a daily basis. It is the total contribution margin
the customer generates in the observation period, divided by
the total number of days the customer was with the bank
(duration). This per diem scaling ensures the comparability
of the contribution margin of customers with different
observed (and possible censored) durations. The contribution
margin equals revenue (interest and fees) less direct
costs (e.g., interest expenses, sales commissions, brokerage,
trading costs). The acquisition costs are not subtracted from
the contribution margin. We also compute a time-varying
version of daily contribution margin by dividing the contribution
margin generated by the customer in a particular
year (2006, 2007, 2008) by the number of days the customer
was with the bank in that year.
1These mailings went to the referring customers. Mailings to
which the nonreferred customers responded were sent directly to
them.
2Although confidentiality concerns preclude us from reporting
the average cost of acquisition for referral and nonreferral customers,
we can report that the total acquisition cost for referred
customers (including not only the referral fee but also the additional
administrative costs of record keeping, paying out, and so
on) is on average approximately 20 euros lower than that for nonreferred
customers acquired through mailings.
Referral Programs and Customer Value / 49

Duration is the total number of days the customer was
observed to be with the bank. It is the basis for analyzing
retention.
We calculate two measures of customer value, one
based only on observed data and the other based on both
observed data and predictions. Observed customer value is
the present value of all actual contribution margins the customer
generated during the whole observation period (e.g.,
33 months for retained customers acquired in January
2006). This metric is affected by both contribution margin
and retention because a customer generates no margins after
leaving the bank. Our second metric, customer lifetime
value, is the present value of all contribution margins, both
actual and predicted, of the customer within the six-year
span following the day of acquisition. 3 For customers who
churned during the observation period, customer lifetime
value equals observed customer value because their predicted
value is 0. The formulas are as follows:
(1) Customer Lifetime Value i = Observed Customer Value i
( 2)
Observed Customer Valuei =
+ Predicted Customer Value i,
Duri
∑
s = 1
( 3)
Predicted Customer Valuei = δi
OM
( 1 + r)
is
s 12
, and
where OM is is the observed monthly contribution margin of
customer i in the sth month after acquisition (calculated
from the observed annual contribution margin and the
observed duration), Dur i is the customer’s observed lifetime
with the bank in months, d i is a dummy censoring variable
that indicates whether the customer was still with the bank
by the end of the observation period, PM is is the predicted
monthly contribution margin of customer i in the sth month
after acquisition, PA is is the predicted probability that customer
i is still “alive” (i.e., with the bank) in that month,
and r is the firm-specific annual discount rate of 11.5%. 4
The present value reflects what the customer is worth at the
day of acquisition.
Independent Variables
The independent variable of central interest is referral program,
a binary indicator that takes the value 1 for referred
customers (i.e., those who were acquired through the referral
program) and 0 for nonreferred customers. To improve the
comparability of referred and nonreferred customers, we control
for the demographic variables age, sex (dummy variable,
s
72
∑
= Duri + 1
50 / Journal of Marketing, January 2011
PMis × PAis,
( 1 + r)
s 12
3This way, we do not need to predict margins and retention rates
beyond four years after the end of the data window, and the resulting
customer lifetime values are unlikely to be overly affected by
forecasting error (Kumar and Shah 2009).
4We base the discount rate on the capital asset pricing model.
We assume a risk-free interest rate of 4.25% (Svensson 1994), a
5% market risk rate based on the Institute of German Accountants,
and a firm-specific beta of 1.45 based on Thomson Financial
Datastream.
with women coded as 1 and men coded as 0), and marital
status (dummy variables for married, divorced/separated,
single, and widowed, with no answer as the base category).
We also control for the customer’s month of acquisition (11
dummy variables for each month, with December 2006 as
the base category).
To assess the robustness of the difference in customer
value, we also conduct separate analyses for the two key
segments of the bank: retail customers with standard financial
needs and nonretail customers with significant assets or
requiring more sophisticated financial advice. This segmentation
scheme the bank uses is based on a comprehensive
analysis of both financial data (e.g., assets invested with the
bank, monthly checking account balance) and demographic
information (e.g., profession, place of residence). The segments
form the basis for all strategic customer-related decisions
of the bank.
Descriptive Statistics
The sample includes several customers with an extremely
high daily contribution margin that is up to 80 standard
deviations above the mean and median. Such extreme data
points can influence comparisons of means and regression
results, so we purify the data using the standard DFBETA
and DFFIT criteria to identify influence points (Belsley,
Kuh, and Welsch 1980). This procedure led to the deletion
of 172 referred customers (3.3% of the original 5181
referred customers) and 147 nonreferred customers (3.2%
of the original 4633 nonreferred customers). As we report in
the subsection “Robustness Checks,” testing the hypotheses
without deleting the influence points results in larger differences
and provides stronger support for the hypotheses. Yet
the size estimates obtained without the influence points are
more informative.
Table 1 presents the means, standard deviations, and the
correlation matrix for the purified sample of 9495 customers.
As the nonzero correlations between the referral
program variable and the various demographic and time of
acquisition variables indicate, the groups of referred and
nonreferred customers are not perfectly matched on demographics
and time of acquisition. Thus, it is desirable to
control for these differences.
Figure 1 plots the average daily contribution margin for
the referred and nonreferred customers of the purified sample,
for 2006, 2007, and 2008. The pattern is encouraging.
Referred customers tend to generate higher margins, and
the margins tend to erode more quickly among referred customers,
such that the margin differential is narrower in 2008
than in 2006 (8 cents versus 18 cents per day). Similarly, as
Figure 2 shows, after about a year, the retention rate of
referred customers is higher, and this is the case regardless
of duration. However, controlling for differences in demographics
and time of acquisition is necessary to draw conclusions
more confidently.
Statistical Models
To estimate the difference in contribution margin between
acquisition modes (H1a), we use a regression model with
the following specification:

TAble 1
Descriptive Statistics
Units M SD 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
1. Referral
program 0–1 .53 .50 1.00
2. Observed
customer
value Euros 210.66 336.15 .02 1.00
3. Customer
lifetime value Euros 255.75 338.95 .01 1.00 1.00
4. Daily
contribution Euros
margin per day .33 .50 .04 .98 .98 1.00
5. Duration Days 751.05 119.48 –.17 .18 .21 .09 1.00
6. Age Years 42.90 17.47 –.20 .10 .09 .10 –.01 1.00
7. Female 0–1 .54 .50 .07 .01 .01 .01 .01 .05 1.00
8. Married 0–1 .39 .49 –.15 –.02 –.03 –.02 –.03 .43 .01 1.00
9. Single 0–1 .44 .50 .16 –.05 –.04 –.05 .01 –.65 –.10 –.70 1.00
10. Divorced 0–1 .10 .30 .00 .03 .03 .03 .02 .13 .06 –.26 –.29 1.00
11. Widowed 0–1 .05 .22 –.05 .11 .11 .10 .03 .36 .14 –.18 –.20 –.07 1.00
12. Acquired
January 2006 0–1 .03 .17 –.17 .07 .08 .03 .31 .02 .00 .01 –.03 –.00 .04 1.00
13. Acquired
February 2006 0–1 .03 .18 –.18 .02 .03 –.01 .27 .05 –.02 .04 –.04 –.00 .01 –.03 1.00
14. Acquired
March 2006 0–1 .06 .24 –.18 .04 .04 .01 .29 .07 –.01 .05 –.04 –.00 .00 –.04 –.05 1.00
15. Acquired
April 2006 0–1 .06 .23 .02 .04 .04 .02 .24 –.01 –.00 –.03 .02 .02 .01 –.04 –.05 –.06 1.00
16. Acquired
May 2006 0–1 .07 .26 .03 .03 .04 .01 .22 –.02 –.01 –.01 .02 –.01 –.02 –.05 –.05 –.07 –.07 1.00
17. Acquired
June 2006 0–1 .08 .28 –.01 .02 .02 .01 .14 .02 .02 –.01 –.01 .00 .04 –.05 –.06 –.08 –.07 –.08 1.00
18. Acquired
July 2006 0–1 .10 .30 .00 .01 .01 .01 .08 .02 –.00 .00 –.02 .03 –.00 –.06 –.06 –.09 –.08 –.09 –.10 1.00
19. Acquired
August 2006 0–1 .11 .31 .06 –.00 –.00 –.00 –.01 –.08 .00 –.06 .05 .02 –.02 –.06 –.06 –.09 –.08 –.09 –.10 –.12 1.00
20. Acquired
September 2006 0–1 .08 .27 .07 .00 .00 .01 –.08 –.06 .01 –.03 .04 –.00 –.02 –.05 –.06 –.08 –.07 –.08 –.09 –.10 –.10 1.00
21. Acquired
October 2006 0–1 .12 .33 .04 –.03 –.04 –.01 –.22 .01 .01 .02 –.00 –.01 –.01 –.07 –.07 –.09 –.09 –.10 –.11 –.13 –.13 –.11 1.00
22. Acquired
November 2006 0–1 .12 .33 .04 –.05 –.06 –.02 –.31 –.00 –.03 .01 –.01 –.01 –.00 –.07 –.07 –.09 –.09 –.10 –.11 –.13 –.13 –.11 –.14 1.00
23. Nonretail
customers 0–1 .12 .32 –.03 .27 .27 .27 –.00 .03 .00 .01 –.01 –.03 .00 .02 –.00 .01 .01 –.04 .00 –.01 –.02 –.02 –.03 .00
Notes: N = 9495. All correlations with absolute value of .02 or higher are significant at the 5% level. Differences in observed duration across customers are strongly affected by differences in the
month of acquisition. As a result, the zero-order correlations of duration with other variables that are also correlated with month of acquisition can be misleading. For example, the correlation
between duration and referral program changes from –.17 to .03 after we control for month of acquisition.
Referral Programs and Customer Value / 51

Daily Contribution Margin (€)
Probability
FIGURe 1
Average Values of Daily Contribution Margin for
Referred and Nonreferred Customers by Year
.600
.500
.400
.300
.200
.100
.496
.315
.331
.350
∑
( ) DCM = α + β RP + β X + ε ,
4 i 1 i k
k = 2
ik i
where i indexes the customer, DCM is daily contribution
margin over the observation period, RP is the binary variable
representing the referral program, X represents the control
variables, and the errors e i have a mean of zero and are
independent of the included covariates. We use ordinary least
squares to estimate the coefficients and compute Huber–
White heteroskedasticity-consistent standard errors (Breusch–
Pagan test, p < .001). The size of our sample implies that
we do not need to assume that the random shocks are normally
distributed for statistical inference using t- and F-statistics
(e.g., Wooldridge 2002, pp. 167–71).
To test whether difference in margin decreases the longer
the customer has been with the bank (H 1b), we use a fixedeffects
specification estimated with ordinary least squares:
( ) DCM = α + β T + β RP × T + η + ε ,
5 it i 2 it 3 i it t it
52 / Journal of Marketing, January 2011
Referred
Nonreferred
.269
.189
2006 2007
Year
2008
FIGURe 2
Probability That Referred and Nonreferred
Customers Have Remained with the Firm
(Kaplan–Meier estimates of Survivor Functions)
1.00
.95
.90
.85
.80
.75
Referred Nonreferred
0 200 400 600 800 1000
Days Since Acquisition
Notes: Customers were able to leave immediately after joining, but
only a handful did so. The earliest defection took place after
64 days, and only 27 customers left within the first year of
joining.
where i indexes the customer; t indexes the year (t = 1, 2, 3);
DCM it is the daily contribution margin of customer i in year
t (i.e., the total contribution margin generated by customer i
in year t divided by the number of days the customer was
with the firm during year t); T it is the cumulative number of
days customer i had been with the bank by the end of year t;
h t is a year-specific fixed effect; and the customer-specific
intercepts a i are not constrained to follow any specific distribution
but capture all individual-specific, time-invariant
differences, including the effect of acquisition through the
referral program (RP) and that of the control variables X.
The errors e it have a mean of zero and are independent of
the covariates. The b 3 coefficient captures the proper interaction
effect because the b 1 effect of RP is now captured
through the fixed effects. Again, we use the robust
Huber–White standard errors (Breusch–Pagan test, p < .001).
To assess the difference in retention between acquisition
modes, we use the Cox proportional hazard model. This
enables us to analyze right-censored duration data and to
exploit the fine-grained measurement of churn at the daily
level without imposing any restriction on how the average
churn rate evolves over time. Furthermore, the nonparametric
baseline hazard makes the model robust to unobserved
heterogeneity in all but extreme cases (e.g., Meyer 1990).
We can represent the model to test H 2a as follows:
∑
( ) ln[ h ( t)] = α( t) + β RP + β X ,
6 i 1 i k ik
k = 2
where i indexes the customer, t indexes the amount of days
since the customer joined the bank, h i(t) is the hazard rate
for the customer’s defection, and a(t) is the log of the nonparametric
baseline hazard common across all customers.
To test whether the difference in churn propensity changes
over time (H 2b), we extend Equation 6 with the interaction
between RP i and ln(t). The latter is also a test of whether
the RP effect meets the proportionality assumption (e.g.,
Blossfeld, Hamerle, and Mayer 1989), but we use it here to
test a hypothesis of substantive interest.
To test H 3 and assess the difference in customer value,
we again use the regression model in Equation 4, but now
with observed customer value as the dependent variable.
Theoretical claims can be subjected to empirical validation
or refutation only by comparing hypothesized effects with
actual data. As a result, the truth content of H 3 can be validly
tested using the observed customer value as the dependent
variable but not customer lifetime value, which itself is based
on predictions. Still, given the right censoring of our data
and the hypothesized erosion of the margin differential over
time, it is informative also to perform a similar analysis with
the six-year customer lifetime value as the dependent variable.
To calculate the predicted values entering the customer lifetime
value metric, we use (1) predicted annual contribution
margins from a fixed-effects model, such as that specified
in Equation 5, but one in which we allow all parameters to
vary between referred and nonreferred customers, and (2)
predicted annual survival rates from a parametric Weibull
hazard model because the nonparametric baseline hazard of
the Cox model does not allow for forecasts. 5

Results
Is the Contribution Margin of Referred Customers
Higher?
In accordance with H1a, referred customers are, on average,
4.5 cents per day more profitable than other customers
(Mann–Whitney test, p < .001). The difference is even larger
after we control for differences in customer demographics
and time of acquisition, variables on which the two groups
of customers are not perfectly matched. The first column of
Table 2 reports the coefficients of Equation 4, controlling
for age, sex, marital status, and month of acquisition.
Whereas the average contribution margin of nonreferred
customers in our sample is 30 cents per day, customers
acquired through the referral program have a margin that is
7.6 cents per day or 27.74 euros per year higher (p < .001),
an increase of approximately 25%. Among the covariates,
older age and being widowed are associated with a higher
contribution margin, whereas being married is associated
with a lower contribution margin. The pattern in the
monthly coefficients suggests that the bank was more successful
in acquiring profitable customers in some months
than in others. The R-square is low, indicating that important
elements other than acquisition method, acquisition
time, and demographics drive customer profitability.
Although the large unexplained variance depresses the
power of statistical tests and thus makes it more difficult to
reject the null hypothesis, H1a is strongly supported.
Does the Contribution Margin of Referred
Customers Remain Higher?
H1b predicts that the difference in contribution margin
erodes the longer a customer stays with the bank. The
results support this. Column 2 of Table 2 reports the coefficients
of the fixed-effects model in Equation 5. There is a
significant, negative interaction between referral program
and the number of days the customer has been with the
bank. The difference in daily contribution margin between
referred and nonreferred customers decreases by 23.1 cents
per 1000 days, or 8.4 cents per year.
The individual-level fixed effects (intercepts) in the
model capture the expected daily contribution margin when
the included covariates equal zero (i.e., on the day of acquisition
in 2006). Regression of these 9495 fixed-effects estimates
on the referral program and control variables indicates
that a referred customer has an expected contribution
margin on the first day of joining the firm that is 19.8 cents
higher than that of a nonreferred customer with similar
demographics and time of acquisition. 6 This implies that
5In-sample parameter estimates from the Cox and Weibull models
are nearly identical. The reason for using the Cox model in
testing the hypotheses is the absence of a restrictive parametric
assumption on the duration dependence.
6This difference in daily contribution margin of 19.8 cents is
close but not identical to the 18 cents difference between the two
groups of customers in 2006, shown in Figure 1. The small disparity
between the two values occurs because the former controls for
differences in demographics and time of acquisition, whereas the
latter does not. Another reason for the disparity is that the former
is the difference on the day of acquisition, whereas the latter is the
difference on an average day in 2006.
the expected contribution margin advantage of a referred
customer disappears after 857 days (.198/.000231), or
approximately 29 months after the customer joined the
bank.
Is the Retention of Referred Customers Higher?
To test whether the retention rate is higher for referred than
for nonreferred customers (H2a), we use the Cox proportional
hazard model specified in Equation 6. The results
reveal that the association between referral program and
churn (i.e., nonretention) is significantly negative and sizable.
Use of only referral program as the explanatory
variable shows that at any point in time, customers acquired
through the referral program who are still with the firm are
approximately 13% less likely to defect than nonreferred
customers who are still with the firm. After we control for
differences in demographics and month of acquisition (see
Column 3 of Table 2), the effect size increases to approximately
–18% [exp(–.198) – 1]. This multiplicative effect of
18% is relative to a baseline hazard that is small. As the survival
curves in Figure 2 show, the probability of being an
active customer (i.e., a nonchurning customer) after 33
months is 82.0% for referred customers and 79.2% for nonreferred
customers. Age is associated with a higher churn
rate, whereas the opposite holds for being widowed. There
is also a trend in the monthly coefficients, indicating that
customers acquired late in 2006 (especially in September
and later) exhibit more churn than those acquired earlier.
This trend is a cohort effect and not duration dependence,
which is captured in the nonparametric baseline hazard.
Does the Retention of Referred Customers
Remain Higher?
We also assess whether the difference in retention varies
over the customer’s lifetime (H2b). To do so, we extend the
Cox model with an interaction between the referral program
variable and the natural logarithm of the customer’s duration
with the bank (see Column 4 of Table 2). The interaction
is not significant, and the model fit does not improve
significantly (p > .05). So, although there is an eroding difference
between referred and nonreferred customers in contribution
margin, there is no such erosion for customer
retention. 7
Are Referred Customers More Valuable?
Using the observed customer value, we find that, on average,
referred customers are 18 euros more valuable
(Mann–Whitney test, p < .001). After we control for demographics
and month of acquisition, the difference increases
to 49 euros (Column 5 of Table 2; p < .001). A referred cus-
7Note that in the model with the interaction term included, the
coefficient of referral program (.917, p > .05) is not the average
difference between referred and nonreferred customers anymore
but rather the difference between the two groups on the day of
acquisition (i.e., when the interaction variable Log(duration)
equals 0; see Irwin and McClelland 2001). Thus, the insignificant
coefficient of referral program in the model including the interaction
term does not invalidate the finding of a significant difference
in retention between the two groups posited in H 2a.
Referral Programs and Customer Value / 53

TAble 2
Main Results for Differences in Daily Contribution Margin, Churn (i.e., the Converse of Retention),
Observed Customer Value, and Customer lifetime Value
H1b H1a Daily
H2b H3 H3 Daily
Contribution
Contribution
Margin (Time
H2a Churn
(Time
Observed
Customer
Customer
lifetime
Margin Varying) Churn Varying) Value Value
Referral program .076*** — a –.198** .917 49.157*** 39.906***
(.010) — (.059) (1.479) (7.096) 7.152
Age .003*** — .011** .011*** 1.879*** 1.626***
(.000) — (.002) (.002) (.283) (.285)
Female –.009 — –.034 –.034 –4.459 –3.376
(.010) — (.056) (.056) (6.902) (6.958)
Sarried –.078* — –.027 –.028 –52.798* –52.258*
(.033) — (.166) (.166) (22.427) (22.563)
Single –.040 — –.163 –.163 –27.306 –24.035
(.033) — (.167) (.167) (22.573) (22.706)
Divorced –.016 — –.176 –.177 –12.278 –7.656
(.037) — (.183) (.183) (24.776) (24.933)
Widowed .111* — –.470* –.470* 76.085* 87.249**
(.046) — (.212) (.212) (31.128) (31.355)
Acquired January 2006 .172*** — –1.828** –1.833*** 228.228*** 247.960***
(.039) — (.201) (.201) (31.589) (31.666)
Acquired February 2006 .063* — –1.365** –1.369*** 127.706*** 133.591***
(.031) — (.160) (.159) (24.172) (24.411)
Acquired March 2006 .089** — –1.155** –1.157*** 136.393*** 135.755***
(.026) — (.126) (.126) (19.103) (19.280)
Acquired April 2006 .084** — –1.215** –1.208*** 124.793*** 123.153***
(.027) — (.140) (.140) (18.753) (18.895)
Acquired May 2006 .082** — –1.529** –1.524*** 114.302*** 119.426***
(.025) — (.150) (.150) (16.791) (16.909)
Acquired June 2006 .066** — –1.016** –1.013*** 91.090*** 92.643***
(.022) — (.122) (.122) (14.326) (14.475)
Acquired July 2006 .062** — –1.026** –1.023*** 79.574*** 84.200***
(.021) — (.122) (.122) (12.717) (12.839)
Acquired August 2006 .059** — –.841** –.838*** 69.213*** 73.167***
(.020) — (.119) (.119) (12.111) (12.233)
Acquired September 2006 .077** — –.679** –.676*** 72.213*** 76.352***
(.022) — (.126) (.126) (13.199) (13.335)
Acquired October 2006 .037 — –.434** –.432*** 36.602* 39.391***
(.020) — (.108) (.108) (11.133) (11.257)
Acquired November 2006 .021 — –.217* –.215* 19.252 20.551
(.019) — (.105) (.105) (10.497) (10.632)
Year 2007 (dummy) –1.306
(.732)
Year 2008 (dummy) –2.259
(1.258)
Cumulative Days (in thousands)
Cumulative days (in thousands) ¥
3.513
(1.994)
referral program –.231**
(.085)
Log(duration) ¥ referral program –.176
(.232)
Constant .154*** 66.250* 120.949***
(.040) (26.742) (26.937)
Observations 9495 28,353 9495 9495 9495 9495
R² .025 .350 .040 .040
Log-pseudo-likelihood
*p < .05.
**p < .01.
***p < .001.
aCaptured by customer-specific fixed effects.
Notes: Robust standard errors are in parentheses.
–11,715.6 –11,715.4
54 / Journal of Marketing, January 2011

tomer is approximately 25% more valuable to the bank than
a comparable nonreferred customer, within the observation
period. If we take into account the difference in acquisition
costs of approximately 20 euros, the difference in customer
value is nearly 35%. These results strongly support H 3.
Because the margin differential of referred customers
erodes over time even though the loyalty differential does
not, the question arises whether referred customers remain
more valuable beyond the observation period. Repeating the
analysis for the six-year customer lifetime value, we show
that referred customers indeed remain more valuable. The
average customer lifetime value of referred customers is
approximately 6 euros higher than that of other customers
(Mann–Whitney test, p < .001). After we control for differences
in customer demographics and time of acquisition,
the value differential is approximately 40 euros (Column 6
of Table 2; p < .001). Because the average customer lifetime
value of a nonreferred customer is 253 euros, a referred customer
is approximately 16% more valuable to the bank than
a comparable nonreferred customer over a horizon of six
years. If we take into account the difference in acquisition
costs of approximately 20 euros, the difference in customer
lifetime value is approximately 25%. This value differential
is quite considerable.
We also assess the extent to which the differences in
customer value are robust across various subsets of customers.
Table 3 reports the regression coefficients for the
referral program in models of customer value, with the
same controls as in the previous analysis in Columns 5 and
6 of Table 2. Row 1 of Table 3 shows that the results for the
retail customer segment are nearly identical to those for the
entire sample. This is not surprising, because retail customers
make up approximately 90% of our overall sample.
More informative is that the difference in customer value
also exists in the nonretail segment (Row 2 of Table 3).
Rows 3 and 4 of Table 3 show that the positive referral
differential exists among high-margin customers, defined as
those in the top decile based on margin, but not low-margin
customers, defined as those in the bottom decile based on
margin. 8 The remaining rows in Table 3 show that sizable
value differentials between referred and nonreferred customers
exist among both men and women and among all
age ranges, except for those over the age of 55. Overall, the
acquisition through a referral program is associated with
higher customer value for the majority of customer types,
but not for all. These results suggest that using referral programs
might not be beneficial in all customer segments, an
idea we develop further in the “Discussion” section.
Robustness Checks
Table 4 shows that the hypothesis tests are robust to including
retail versus nonretail segment membership as an additional
control variable and allowing the effect of the referral
program to vary as a function of age, sex, marital status, and
retail segment membership. Given the results of Table 3, we
8Low-margin customers and high-margin customers are found
in both the retail and the nonretail segments.
TAble 3
Results for Difference in Customer Value Within
Various Segments
Observed Customer
Customer lifetime
Value Value
(Robust (Robust N (N of
Standard Standard Referred
errors) errors) Customers)
Retail customers 48.620*** 39.082*** 8384
(6.574) (6.633) (4473)
Nonretail customers 77.309** 69.803* 1111
(29.855) (30.023) (536)
High margin customers 80.421** 69.669* 950
(27.768) (28.004) (533)
Low margin customers –1.146 –13.212*** 962
(1.581) (2.087) (247)
Male customers 51.679*** 42.305*** 4371
(10.600) (10.669) (2150)
Female customers 47.437*** 38.274*** 5124
(9.604) (9.690) (2859)
65 years of age –1.577 –8.589 1306
*p < .05.
**p < .01.
***p < .001.
(21.421) (21.409) (536)
Notes: Each row displays the coefficient of referral program in models
with the same control variables as in Table 2 but estimated
for specific segments.
also allowed for a nonlinear effect of age. 9 We mean-center
all variables that interact with referral program, so its coefficient
still reflects the main effect. This coefficient keeps
its sign and significance in each model, so the hypotheses
remain supported. The coefficients are larger than in Table
2, in which we did not control for segment membership and
nonlinear age effects, indicating that our main analysis provides
rather conservative estimates of the referral effects.
As a second robustness check, we repeated the analyses
presented in Table 2 for the sample including all outliers.
The direction and significance of the referral effect remained
the same, but the size of several effects increased. The difference
in daily contribution margin increased from 7.6
cents to 16 cents per day, the margin erosion increased from
23.1 cents to 45.4 cents per thousand days, the churn hazard
reduction remained at 20%, and the difference in customer
lifetime value increased from 40 euros to 66 euros. These
9Because some readers may be interested in how the effect of
referral program is moderated by covariates in the time-varying
contribution model, we estimate the latter using a random coefficients
specification rather than the fixed-effects specification used
in Table 2.
Referral Programs and Customer Value / 55

56 / Journal of Marketing, January 2011
TAble 4
Robustness Checks Allowing for Referral effects to be Moderated
H1b H1a Daily
H2b H3 H3 Daily
Contribution
Contribution
Margin (Time
H2a Churn
(Time
Observed
Customer
Customer
lifetime
Margin Varying) Churn Varying) Value Value
Referral program .133*** .228*** –.270** .791 88.413*** 79.695***
(.017) (.062) (.084) (1.479) (10.870) (10.949)
Age (mean centered) a .310*** .529** 1.421*** 1.421*** 199.247*** 164.893***
(.060) (.191) (.361) (.362) (41.286) (41.647)
Age2 (mean centered) a –.004 –.779 –.014 –.014 –2.276 –2.190
(.003) (.733) (.014) (.014) (1.834) (1.848)
Female (mean centered) –.008 .035 .017 .016 –3.822 –2.828
(.013) (.041) (.076) (.076) (9.393) (9.487)
Married (mean centered) .050 .828*** –.445* –.445* 34.372 38.493
(.039) (.126) (.201) (.202) (27.838) (28.187)
Single (mean centered) .080* .933*** –.472* –.473* 52.010 59.890*
(.040) (.128) (.211) (.211) (27.774) (28.119)
Divorced (mean centered) .111* .916*** –.589* –.590* 77.122* 86.686**
(.044) (.139) (.231) (.231) (31.289) (31.696)
Widowed (mean centered) .340*** 1.104*** –1.011*** –1.012*** 236.446*** 254.709***
(.058) (.154) (.272) (.273) (40.821) (41.259)
Nonretail segment (mean centered) .440*** .551*** –.263* –.263* 310.169*** 311.262***
(.031) (.060) (.124) (.124) (21.972) (22.152)
Age ¥ referral programa,b .151 .081 –.531 –.532 99.676 117.447*
(.086) (.255) (.514) (.514) (57.745) (58.183)
Age2 ¥ referral programa, b –.015*** –.311 .022 .022 –10.212*** –10.392***
(.004) (1.018) (.020) (.020) (2.539) (2.556)
Female ¥ referral programb –.005 –.019 –.104 –.104 –2.980 –2.982
(.020) (.056) (.112) (.112) (13.210) (13.320)
Married ¥ referral programb –.162* –.974*** .953** .951** –108.801* –114.852*
(.068) (.174) (.362) (.362) (45.824) (46.049)
Single ¥ referral programb –.097 –.972*** .743* .743* –62.859 –70.679
(.068) (.175) (.366) (.366) (45.988) (46.207)
Divorced ¥ referral programb –.147* –.982*** .916* .917* –103.598* –112.706*
(.074) (.190) (.394) (.394) (50.159) (50.458)
Widowed ¥ referral programb –.270** –1.181*** 1.246** 1.248** –197.134** –211.207**
(.091) (.222) (.456) (.456) (60.695) (61.046)
Nonretail ¥ referral programb –.025 –.084 .014 .013 –49.754 –49.193
(.046) (.084) (.188) (.188) (30.805) (31.001)
Year 2007 (dummy) –1.487***
(.143)
Year 2008 (dummy) –2.562***
(.236)
Cumulative days (in thousands) 4.004***
(.387)
Cumulative days (in thousands) ¥ –.220*
referral program (.101)
Log(duration) ¥ referral program –.167
(.232)
Constant .211*** .244*** 100.219*** 147.265***
(.016) (.056) (9.809) (9.921)
Observations 9495 28,353 9495 9495 9495 9495
R² .107 .099c .123 .122
Log-pseudo-likelihood
*p < .05.
**p < .01.
***p < .001.
aWe divide age by 100 for better readability.
bThe first variable in this interaction is mean centered.
–11,705.8 –11,705.6
cBecause the model is a random coefficients model estimated with residual maximum likelihood, this value is a pseudo-R-square calculated as
the squared correlation between predicted and actual values.
Notes: Robust standard errors are in parentheses. All models include dummies for month of acquisition.

esults suggest that our main analysis is rather conservative
with regard to the size of the referral differentials.
Although hazard analysis properly accounts for right
censoring, managers are also interested in simply knowing
who is likely to have remained with the firm within a certain
time frame. Therefore, we also assessed the relationship
between referral and the probability of still being with
the bank 21 months after acquisition. This time span is the
longest duration observable without right censoring for
every customer, including those who were acquired last, at
the end of December 2006. Using logistic regression and
controlling for demographics and month of acquisition, we
find that referred customers are approximately 22% less
likely to leave the firm within the first 21 months (p < .01).
Computing the customer lifetime value over three,
rather than six, years after acquisition and repeating the
analysis by controlling for demographics and time of acquisition
yields a value differential between referred and nonreferred
customers of 52 euros (p < .001) rather than 40
euros. Both the size and the statistical significance of the
latter value are rather robust to reestimating the model on
smaller random samples of 90% (39 euros, p < .001), 75%
(42 euros, p < .001), 50% (48 euros, p < .001), and 25% (36
euros, p < .01). We also computed the expected value differential
if there were no difference in retention between referred
and nonreferred customers. The differential in six-year customer
lifetime value decreased from 40 euros to 33 euros.
Finally, we extended the model of margin dynamics and
allowed the effect of time and its interaction with referral to
vary as a function of observed customer demographics,
retail versus nonretail status, and normally distributed unobserved
heterogeneity. This extended random coefficients
model did not fit the data better: The squared correlation
between observed and predicted values (pseudo-R 2)
increased by only .1%, and the Bayesian information criterion
even decreased. Most important, the coefficients of
central interest and the statistical inference were not
affected: Customers acquired through referral had a sizable
initial margin advantage that eroded to zero after approximately
1000 days.
Discussion
Key Findings
Evidence of the economic value of stimulated WOM and of
the customers it helps acquire has been sorely lacking. Our
study addresses this gap in the context of referral programs
and documents the attractiveness of customers acquired
through such a program: Contribution margin, retention, and
customer value all were significantly and sizably higher for
referred customers. In short, referred customers are more
valuable in both the short and the long run. Yet we also find
that the effect is not uniform across all types of customers
and that the referral program was less beneficial when used
to acquire older customers or low-margin customers.
In our application, the value of referred customers in the
six years after acquisition was 40 euros (or 16%) higher
than that of nonreferred customers with similar demograph-
ics and time of acquisition. Considering the initial reward of
25 euros given to the referrer as an investment, this implies
a return on investment of approximately 60% over a sixyear
span. This is a conservative estimate because it does not
take into account that the total acquisition costs of referred
customers are approximately 20 euros lower than those of
other customers.
Implications for Practice
Several scholars have expressed cautious skepticism about
the value of “viral-for-hire” and other stimulated WOM
(e.g., Trusov, Bucklin, and Pauwels 2009; Van den Bulte
2010). Doubts about the benefits of stimulated WOM have
long frustrated managers facing demands to increase their
marketing return on investment. Our findings are important
news for practitioners considering deploying customer
referral programs in their own firm. Assuaging prior skepticism,
we document a positive value differential, in both the
short run and the long run, between customers acquired
through a referral program and other customers. Importantly,
this value differential is larger than the referral fee.
Thus, referral programs can indeed pay off.
The positive differential indicates that abuse by opportunistic
customers and other harmful side effects of referral
programs are much less important than their benefits. The
referral program we analyzed was especially prone to
exploitation because no conditions, such as minimum stay
or assets, applied to the newly acquired customer. Finding a
positive value differential of referred customers in this setting
is especially compelling. Moving beyond referral programs
specifically, our study indicates that a stronger focus
on stimulated rather than organic WOM is worth considering
(Godes and Mayzlin 2009).
However, our results also suggest that firms should
think carefully about what prospects to target with referral
programs and how big of a referral fee to provide. For the
program we analyzed, we found that the customer value differential
is much larger in some segments than in others.
People under the age of 55 and high-margin customers are
more attractive to acquire through a referral program. It is
not necessarily a coincidence that these customers also tend
to be more profitable for banks (and many other consumer
marketers). To the extent that the value differential stems
from better matching and social enrichment, as sociological
theory suggests and as employee referral programs have
documented, referral programs do not “create” higher-value
customers by transforming unattractive prospects into
attractive customers. Rather, they help firms selectively
acquire more valuable prospects and retain them longer at
lower cost. Thus, instead of the currently practiced “all-in”
approach, firms should design and target referral programs
such that attractive customers are more likely to be enticed.
Managers must also make their customers aware of their
referral programs. Bank of America, for example, communicates
its referral program on all its automated teller
machines throughout the United States. Connecting referral
programs with online activities might help further increase
their reach beyond existing customers’ networks of strong
Referral Programs and Customer Value / 57

ties and face-to-face interactions. Managers must also make
it convenient for prospects to actually become a customer. A
possible application is to partner with online communities
and make it easy for people to start a relationship with the
firm online, immediately after they receive a referral from
an existing customer in the same community. Our results
suggest that such awareness and facilitation efforts should
be targeted selectively to customers who offer the highest
value differential.
The referral fee is another issue that requires attention
when designing a referral program. Many programs offer
the same reward to each referrer (Kumar, Petersen, and
Leone 2010). Yet, as we show, the value of referred customers
can vary widely even for one company. Thus, firms
may benefit from offering rewards based on the value of the
referred customer. However, the question then becomes
how to do this without adding too much complexity to the
program. There may be a simple answer: A standard
homophily argument suggests that valuable referrers are more
likely to generate valuable referrals. Thus, firms may want to
make the referral fee a function of the value of the referrer.
A different approach to take advantage of the referral
effect would be to try to generate conditions in which nonreferred
customers become subject to the same mechanisms
that make referred customers more valuable. To the extent
that the differences we have documented stem from better
matching, from social enrichment, or from other mechanisms
that firms can actively foster among all customers,
firms may be able to dramatically “scale up” the beneficial
referral effect beyond dyads of referring and referred customers.
For example, pharmaceutical companies increasingly
involve local opinion leaders in their speaker programs
and other medical education efforts. They do so to
capitalize on these physicians’ relevance and credibility
with practicing physicians.
Firms in the same industry often reward referrers with
the same amount. For example, most German banks offer
25 euros for a referral, as does the bank we studied. Our
results indicate that managers set the referral fee rather low,
allowing the firm to reap attractive returns from its program.
Offering higher rewards might lead to even more customer
acquisitions while still providing positive returns on
investment. Firms should calculate the reward considering
their specific program and the customers it attracts instead
of merely following their competitors.
Further Research
Our study focuses on referred and nonreferred customers of
one particular bank. Although its customer base and referral
program have no unusual characteristics, replications would
nonetheless be welcome. Such studies require rich, firmspecific
data on a large set of customers, with individual
profitability observed over a long period. Therefore, we
expect replications and extensions to come from other
single-firm studies such as ours and those of Godes and
Mayzlin (2009), Haenlein (2010), Iyengar, Van den Bulte, and
Valente (2011), and Nitzan and Libai (2010). Because the
mechanisms of better matching and social enrichment are
likely to be more important for complex products with important
experience attributes, rather than simple products with
58 / Journal of Marketing, January 2011
mostly search attributes (e.g., Coverdill 1998; Kornish and
Li 2010; Rees 1966), studies of multiple products with varying
levels of complexity would be especially informative.
It is likely that the quality of the matches with the firm
deteriorates as existing customers refer more new customers.
It would be of practical interest to know at what rate
the quality of referrals decreases and at what point it tends
not to justify the cost of acquisition anymore. It may also be
useful to know if the motivation of the referrer changes
depending on the reward and whether the size of the reward
affects the quality of the referred customer.
Several of the implications for practice point to the
benefits of better understanding the drivers of the value differential
we documented. Although our results are consistent
with the better matching and social enrichment mechanisms
we used to develop our hypotheses, our analysis
focused on the consequences for contribution margin, retention,
and customer value rather than on the intervening
mechanisms. Research aimed at more directly parsing out
the mechanisms is likely to require information about actual
dyads of referring and referred customers. This would
enable researchers to test, for example, the social enrichment
argument by matching the referred customer with the
respective referrer and analyzing the dependence of their
retention. Additional survey data may help document differences
in product knowledge over time and shed light on the
existence of an informational advantage eroding over time.
Having matched dyad-level data on both referring and
referred customers would also make it possible to check
whether referral dyads exhibit homophily and whether the
value of referred customers varies systematically with that
of their referrer (Haenlein 2010; Nitzan and Libai 2010).
This would yield valuable insights for the design of individual
rewards instead of the currently practiced “one-size-fitsall”
approach.
Conclusion
This study provides the first assessment of economically relevant
differences between customers acquired through a referral
program and customers acquired through other methods.
It documents sizable differences in contribution margin,
retention, and customer value; analyzes whether these differences
erode or persist over time; and investigates differences
across customer segments. The finding that, on average,
referred customers are more valuable than other customers
provides the first direct evidence of the financial attractiveness
of referral programs and also offers much-needed evidence
of the financial appeal of stimulated WOM in general.
Improvements in the targeting, design, and implementation
of such programs will require a better understanding
of the drivers of the value differential. The dyadic interdependence
in the behavior of the referrer and the referred
customer deserves special attention in further research
because it is likely to prove highly relevant to both better
theoretical understanding and more effective program
management.

ReFeReNCeS
Belsley, David A., Edwin Kuh, and Roy E. Welsch (1980), Regression
Diagnostics: Identifying Influential Data and Sources of
Collinearity. New York: John Wiley & Sons.
Biggart, Nicole W. (1989), Charismatic Capitalism: Direct Selling
Organizations in America. Chicago: University of Chicago Press.
Biyalogorsky, Eyal, Eitan Gerstner, and Barak Libai (2001), “Customer
Referral Management: Optimal Reward Programs,”
Marketing Science, 20 (1), 82–95.
Blossfeld, Hans-Peter, Alfred Hamerle, and Karl Ulrich Mayer
(1989), Event History Analysis: Statistical Theory and Application
in the Social Sciences. Hillsdale, NJ: Lawrence Erlbaum
Associates.
Bolton, Patrick, Xavier Freixas, and Joel Shapiro (2007), “Conflicts
of Interest, Information Provision, and Competition in the
Financial Services Industry,” Journal of Financial Economics,
85 (2), 297–330.
Castilla, Emilio J. (2005), “Social Networks and Employee Performance
in a Call Center,” American Journal of Sociology, 110
(5), 1243–83.
Choi, Jeonghye (2009), “Social Influence from Existing to New
Customers,” working paper, the Wharton School, University of
Pennsylvania.
Coverdill, James E. (1998), “Personal Contacts and Post-Hire Outcomes:
Theoretical and Empirical Notes on the Significance of
Matching Methods,” Research in Social Stratification and
Mobility, 16, 247–69.
Edwards, Helen and Derek Day (2005), Creating Passion Brands:
Why It’s Time to Lead from the Heart. London: Kogan Page.
Fernandez, Roberto M., Emilio J. Castilla, and Paul Moore (2000),
“Social Capital at Work: Networks and Employment at a Phone
Center,” American Journal of Sociology, 105 (5), 1288–1356.
Fleming, John H., Curt Coffman, and James K. Harter (2005),
“Manage Your Human Sigma,” Harvard Business Review, 83
(7–8), 106–114.
Godes, David and Dina Mayzlin (2009), “Firm-Created Word-of-
Mouth Communication: Evidence from a Field Test,” Marketing
Science, 28 (4), 721–39.
Gouldner, Alvin W. (1960), “The Norm of Reciprocity: A Preliminary
Statement,” American Sociological Review, 25 (April), 161–78.
Gupta, Sunil and Valarie Zeithaml (2006), “Customer Metrics and
Their Impact on Financial Performance,” Marketing Science,
25 (6), 718–39.
Guseva, Alya (2008), Into the Red: The Birth of the Credit Card
Market in Postcommunist Russia. Palo Alto, CA: Stanford University
Press.
Haenlein, Michael (2010), “A Social Network Analysis of Customer-
Level Revenue Distribution,” Marketing Letters, forthcoming.
Helm, Sabrina (2003), “Calculating the Value of Customers’
Referrals,” Managing Service Quality, 13 (2), 124–33.
Irwin, Julie R. and Gary H. McClelland (2001), “Misleading
Heuristics and Moderated Multiple Regression Models,” Journal
of Marketing Research, 38 (February), 100–109.
Iyengar, Raghuram, Christophe Van den Bulte, and Thomas W.
Valente (2011), “Opinion Leadership and Social Contagion in
New Product Diffusion,” Marketing Science, forthcoming.
Kornish, Laura J. and Quiping Li (2010), “Optimal Referral
Bonuses with Asymmetric Information: Firm-Offered and
Interpersonal Incentives,” Marketing Science, 29 (1), 108–121.
Kumar, V., J. Andrew Petersen, and Robert P. Leone (2007), “How
Valuable Is Word of Mouth?” Harvard Business Review, 85
(10), 139–46.
———, ———, and ——— (2010), “Driving Profitability by
Encouraging Customer Referrals: Who, When, and How,”
Journal of Marketing, 74 (September), 1–17.
——— and Denish Shah (2009), “Expanding the Role of Marketing:
From Customer Equity to Market Capitalization,” Journal
of Marketing, 73 (November), 119–36.
Lewis, Michael (2006), “Customer Acquisition Promotions and
Customer Asset Value,” Journal of Marketing Research, 43
(May), 195–203.
Meyer, Bruce D. (1990), “Unemployment Insurance and Unemployment
Spells,” Econometrica, 58 (July), 757–82.
Montgomery, James D. (1991). “Social Networks and Labor-Market
Outcomes: Toward an Economic Analysis,” American Economic
Review, 81 (December), 1408–1418.
Mummert, Hallie (2000), “The Year’s Best Bells & Whistles,” Target
Marketing, 23 (11), 3–5.
Neckerman, Kathryn M. and Roberto M. Fernandez (2003),
“Keeping a Job: Network Hiring and Turnover in a Retail
Bank,” Research in the Sociology of Organizations, 20,
299–318.
Nitzan, Irit and Barak Libai (2010), “Social Effects on Customer
Retention,” Report No. 10-107. Cambridge, MA: Marketing
Science Institute.
Parasuraman, A., Valarie A. Zeithaml, and Leonard L. Berry
(1985), “A Conceptual Model of Service Quality and Its Implications
for Future Research,” Journal of Marketing, 49 (Fall),
41–50.
Rees, Albert (1966), “Information Networks in Labor Markets,”
American Economic Review, 56 (March), 559–66.
Ryu, Gangseog and Lawrence Feick (2007), “A Penny for Your
Thoughts: Referral Reward Programs and Referral Likelihood,”
Journal of Marketing, 71 (January), 84–94.
Svensson, Lars E. (1994), “Estimating and Interpreting Forward
Interest Rates: Sweden 1992-1994,” NBER Working Paper
No. 4871. Cambridge, MA: National Bureau of Economic
Research.
Thaler, Richard H. and Cass R. Sunstein (2008), Nudge: Improving
Decisions About Health, Wealth, and Happiness. New
Haven, CT: Yale University Press.
Trusov, Michael, Randolph E. Bucklin, and Koen Pauwels (2009),
“Effects of Word-of-Mouth Versus Traditional Marketing:
Findings from an Internet Social Networking Site,” Journal of
Marketing, 73 (September), 90–102.
Van den Bulte, Christophe (2010), “Opportunities and Challenges
in Studying Customer Networks,” in The Connected Customer:
The Changing Nature of Consumer and Business Markets, S.
Wuyts, M.G. Dekimpe, E. Gijsbrechts, and R. Pieters, eds.
London: Routledge, 7–35.
——— and Stefan Wuyts (2007), Social Networks and Marketing.
Cambridge, MA: Marketing Science Institute.
Villanueva, Julian, Shijin Yoo, and Dominique M. Hanssens
(2008), “The Impact of Marketing-Induced Versus Word-of-
Mouth Customer Acquisition on Customer Equity Growth,”
Journal of Marketing Research, 45 (February), 48–59.
Wiesel, Thorsten, Bernd Skiera, and Julian Villanueva (2008),
“Customer Equity: An Integral Part of Financial Reporting,”
Journal of Marketing, 72 (March), 1–14.
Wirtz, Jochen and Patricia Chew (2002), “The Effects of Incentives,
Deal Proneness, Satisfaction and Tie Strength on Wordof-Mouth
Behaviour,” International Journal of Service Industry
Management, 13 (2), 141–62.
Wooldridge, Jeffrey M. (2002), Introductory Econometrics: A
Modern Approach, 2d ed. Cincinnati: South-Western.
Yim, Chi Kin, David K. Tse, and Kimmy Wa Chan (2008),
“Strengthening Customer Loyalty Through Intimacy and Passion:
Roles of Customer–Firm Affection and Customer–Staff
Relationships in Services,” Journal of Marketing Research, 45
(December), 741–56.
ZatzNotFunny.com (2008), “Confirmed: TiVo Rewards Killed,”
(accessed June 28, 2010), [available at http://www.zatznotfunny.
com/2008-02/confirmed-tivo-rewards-killed/].
Referral Programs and Customer Value / 59

Oliver Hinz, Bernd Skiera, Christian Barrot, & Jan U. Becker
Seeding Strategies for Viral
Marketing: An Empirical
Comparison
Seeding strategies have strong influences on the success of viral marketing campaigns, but previous studies
using computer simulations and analytical models have produced conflicting recommendations about the optimal
seeding strategy. This study compares four seeding strategies in two complementary small-scale field experiments,
as well as in one real-life viral marketing campaign involving more than 200,000 customers of a mobile phone
service provider. The empirical results show that the best seeding strategies can be up to eight times more
successful than other seeding strategies. Seeding to well-connected people is the most successful approach
because these attractive seeding points are more likely to participate in viral marketing campaigns. This finding
contradicts a common assumption in other studies. Well-connected people also actively use their greater reach
but do not have more influence on their peers than do less well-connected people.
Keywords: viral marketing, seeding strategy, word of mouth, social contagion, targeting
The future of traditional massmedia advertising is
uncertain in the modern environment of increasingly
prevalent digital video recorders and spam filters.
Marketers must realize that 65% of consumers consider
themselves overwhelmed by too many advertising messages,
and nearly 60% believe advertising is not relevant
to them (Porter and Golan 2006). Such information overload
can cause consumers to defer their purchase altogether
(Iyengar and Lepper 2000), and strong evidence indicates
consumers actively avoid traditional marketing instruments
(Hann et al. 2008).
Other empirical evidence also reveals that consumers
increasingly rely on advice from others in personal or professional
networks when making purchase decisions (Hill,
Provost, and Volinsky 2006; Iyengar, Van den Bulte, and
Valente 2011; Schmitt, Skiera, and Van den Bulte 2011).
In particular, online communication appears increasingly
important as more websites offer user-generated content,
Oliver Hinz is Chaired Professor of Information Systems esp.
Electronic Markets, Technische Universität Darmstadt (e-mail: hinz
@wi.tu-darmstadt.de). Bernd Skiera is Chaired Professor of Electronic
Commerce, Department of Marketing, Goethe-University of
Frankfurt (e-mail: skiera@wiwi.uni-frankfurt.de). Christian Barrot is
Assistant Professor of Marketing and Innovation (e-mail: christian
.barrot@the-klu.org), and Jan U. Becker is Assistant Professor
of Marketing and Service Management (e-mail: jan.becker@theklu.org),
Kühne Logistics University. The authors thank Carsten
Takac, Philipp Schmitt, Martin Spann, Lucas Bremer, Christian
Messerschmidt, Daniele Mahoutchian, Nadine Schmidt, and Katharina
Schnell for helpful input on previous versions of this article. In
addition, three anonymous JM referees provided many helpful suggestions.
This research is supported by the E-Finance Lab Frankfurt.
© 2011, American Marketing Association
ISSN: 0022-2429 (print), 1547-7185 (electronic) 55
such as blogs, video and photo sharing opportunities,
and online social networking platforms (e.g., Facebook,
LinkedIn). Companies have adapted to these trends by
shifting their budgets from above-the-line (mass media) to
below-the-line (e.g., promotions, direct mail, viral) marketing
activities.
Not surprisingly then, viral marketing has become a
hot topic. The term “viral marketing” describes the phenomenon
by which consumers mutually share and spread
marketing-relevant information, initially sent out deliberately
by marketers to stimulate and capitalize on word-ofmouth
(WOM) behaviors (Van der Lans et al. 2010). Such
stimuli, often in the form of e-mails, are usually unsolicited
(De Bruyn and Lilien 2008) but easily forwarded to multiple
recipients. These characteristics parallel the traits of
infectious diseases, such that the name and many conceptual
ideas underlying viral marketing build on findings from
epidemiology (Watts and Peretti 2007).
Because viral marketing campaigns leave the dispersion
of marketing messages up to consumers, they tend to be
more cost efficient than traditional massmedia advertising.
For example, one of the first successful viral campaigns,
conducted by Hotmail, generated 12 million subscribers in
just 18 months with a marketing budget of only $50,000.
Google’s Gmail captured a significant share of the e-mail
provider market, even though the only way to sign up for
the service was through a referral. A recent viral advertisement
by Tipp-Ex (“A hunter shoots a bear!”) triggered
nearly 10 million clicks in just four weeks.
However, to enjoy such results, firms must consider four
critical viral marketing success factors: (1) content, in that
the attractiveness of a message makes it memorable (Berger
and Milkman 2011; Berger and Schwartz 2011; Gladwell
2002; Porter and Golan 2006); (2) the structure of the social
Journal of Marketing
Vol. 75 (November 2011), 55–71

network (Bampo et al. 2008); (3) the behavioral characteristics
of the recipients and their incentives for sharing the
message (Arndt 1967); and (4) the seeding strategy, which
determines the initial set of targeted consumers chosen by
the initiator of the viral marketing campaign (Bampo et al.
2008; Kalish, Mahajan, and Muller 1995; Libai, Muller,
and Peres 2005). This last factor is of particular importance
because it falls entirely under the control of the initiator
and can exploit social characteristics (Toubia, Stephen, and
Freud 2010) or observable network metrics. Unfortunately,
there is a “need for more sophisticated and targeted seeding
experimentation” to gain “a better understanding of the role
of hubs in seeding strategies” (Bampo et al. 2008, p. 289).
The conventional wisdom adopts the influentials hypothesis,
which states that targeting opinion leaders and
strongly connected members of social networks (i.e., hubs)
ensures rapid diffusion (for a summary of arguments, see
Iyengar, Van den Bulte, and Valente 2011). However, recent
findings raise doubts. Van den Bulte and Lilien (2001)
show that social contagion, which occurs when adoption is
a function of exposure to other people’s knowledge, attitudes,
or behaviors (Van den Bulte and Wuyts 2007), does
not necessarily influence diffusion, and yet it remains a
basic premise of viral marketing. Such contagion frequently
arises when people who are close in the social structure use
one another to manage uncertainty in prospective decisions
(Granovetter 1985). However, in a computer simulation,
Watts and Dodds (2007) show that well-connected people
are less important as initiators of large cascades of referrals
or early adopters. Their finding, which Thompson (2008)
provocatively summarizes by implying “the tipping point is
toast,” has stimulated a heated debate about optimal seeding
strategies, though no research offers an extensive empirical
comparison of seeding strategies. Van den Bulte (2010)
thus calls for empirical comparisons of seeding strategies
that use sociometric measures (i.e., metrics that capture the
social position of people).
In response, we undertake an empirical comparison of
the success of different seeding strategies for viral marketing
campaigns and identify reasons for variations in these
levels of success. In doing so, we determine whether companies
should care about the seeding of their viral marketing
campaigns and why. In particular, we study whether
well-connected people really are more difficult to activate,
participate more actively in viral campaigns, and have more
influence on their peers than less well-connected people. In
contrast with previous studies that rely on analytical models
or computer simulations, we derive our results from field
experiments and from a real-life viral marketing campaign.
We begin this article by presenting literature relevant to
viral marketing and social contagion theory. We introduce
our theoretical framework, which disentangles the determinants
of social contagion, and present four seeding strategies.
Next, we empirically compare the success of these
seeding strategies in two complementary field experiments
(Studies 1 and 2) that aim at spreading information and
inducing attitudinal changes. Then, we analyze a real-life
viral marketing campaign designed to increase sales, which
provides an economic measure of success. After we identify
the determinants of success, we conclude with a discussion
of our research contributions, managerial implications,
and limitations.
56 / Journal of Marketing, November 2011
Theoretical Framework
When information about an underlying social network is
available, seeding on the basis of this information, as
typically captured by sociometric data, seems promising
(Van den Bulte 2010). Such a strategy can distinguish three
types of people: “hubs,” who are well-connected people
with a high number of connections to others; “fringes,” who
are poorly connected; and “bridges,” who connect two otherwise
unconnected parts of the network. The sociometric
measure of degree centrality captures connectedness within
the local environment (for details, see the Web Appendix at
http://www.marketingpower.com/jmnov11), such that highdegree-centrality
values characterize hubs, whereas lowdegree-centrality
values mark fringes. In contrast, the
sociometric betweenness centrality measure describes the
extent to which a person acts as a network intermediary,
according to the share of shortest communication paths
that pass through that person (see the Web Appendix).
Thus, bridges earn high values on betweenness centrality
measures.
Determinants of Social Contagion
Following Van der Lans et al. (2010), we propose a fourdeterminant
model of social contagion to determine the
success of viral marketing campaigns. First, individual i
receives a viral message from sender s, who can be either a
friend or the campaign initiator that makes i aware of and
informed by the message with information probability I i.
Individual i then may become active and participate in the
campaign with participation probability P i. Given participation,
individual i passes the message to a set of recipients J i,
where n i is the number of recipients (�J i� = n i), such that it
provides a measure of used reach. The number of expected
referrals R i by individual i, then, is the product of the information
probability (I i), the probability of participating (P i),
and the used reach (n i): R i = I i × P i × n i.
The conversion rate w i�j linearly influences the number of
expected successful referrals SR i of individual i on recipients
j (j ∈ J i), given by
(1)
n �i
SRi = Ii × Pi × ni ×
j = 1
w i� j
If a sender i has the same conversion rate for all recipients,
such that w i�j = w i ∀ j ∈ J i, the number of expected
successful referrals can be rewritten as
(2)
n i
SR i = I i × P i × n i × w i�
All these determinants are a function of i’s social position,
though they also may be influenced by the characteristics
of the sender s and the conversion rate w s. Despite
the lack of empirical comparisons of seeding strategies for
viral marketing campaigns, various studies in marketing,
sociology, and epidemiology have analyzed the influence
of the social position (captured by sociometric measures)
on different determinants, such as whether hubs are more
likely to persuade their peers. In Table 1, we summarize
these findings according to the determinants of information
probability I i, participation probability P i, used reach n i,
and conversion rate w i.
�

TABLE 1
Previous Research
Social Position Has Positive Influence on � � �
Expected Empirically
Expected number of Recommendation Tested
Reason for Participation Used Number of Conversion Successfull for Optimal Seeding
Studies Context Contagion Probability P i Reach n i Referrals R i Rate w i SR i Seeding Strategy Strategy
Coleman, Katz, and Product (low risk) A, BU, NP Hub Hub Hub
Menzel (1966)
Becker (1970) Product (low risk) A, BU, NP Hub Hub Hub
Product (high risk) Fringe Fringe Fringe
Simmel (1950); Porter Messages A Fringe Fringe
and Donthu (2008) Messages A
Watts and Dodds (2007) — — Fringe Hub Fringe Fringe Fringe
Leskovec, Adamic, Product (low risk) A, BU Hub Hub Hub Fringe
and Huberman (2007)
Anderson and May Epidemiology A Hub Hub Hub Hub
(1991); Kemper (1980) Epidemiology A
Granovetter (1973); Messages A Bridge Bridge Bridge
Rayport (1996) Messages A
Iyengar, Van den Bulte, Product (high risk) A, BU Hub Hub Hub Hub
and Valente (2011)
Study 1 Messages A Controlled � Hub, fringe, bridge,
random
Study 2 Messages A � Hub, fringe, bridge,
random
Study 3 Product (low risk) A, BU � � � � � Hub, fringe, random
Notes: A = awareness, BU = belief updating, NP = normative pressure, and i = focal individual. Expected number of referrals: R i = P i × n i; Successful number of referrals: SR i = w i × R i.
Seeding Strategies for Viral Marketing / 57

Effect of social position on information and participation
probability. A viral marketing campaign aims to
inform consumers about the viral marketing message and
to encourage them to participate in the campaign by sending
the message to others. In investigating the impact
of social position on information probability, Goldenberg
et al. (2009) indicate that hubs tend to be better informed
than others because they are exposed to innovations earlier
through their multiple social links. In his reanalysis of
Coleman, Katz, and Menzel’s (1966) medical innovation
study, Burt (1987) also recognizes that some people experience
discomfort when peers whose approval they value
adopt an innovation they have not yet adopted; in this case,
social contagion (reflected in a greater probability to participate)
results from normative pressure and status considerations.
This mechanism could explain why Coleman, Katz,
and Menzel (1966) find that highly integrated people (e.g.,
hubs) are more likely to adopt an innovation early than are
more isolated people.
However, in some cases, hubs do not adopt innovations
first (Becker 1970), such as when the innovation does not
suit the hub’s opinion, which may mean that adoption
occurs first at the fringes of the network (Iyengar, Van den
Bulte, and Valente 2011). Another potential explanation of
hubs’ lower participation probability stems from information
overload effects. Because hubs are exposed to so many
contacts, they possess a wealth of information and thus
might be more difficult to activate (Porter and Donthu 2008;
Simmel 1950) or less likely to participate in viral marketing
campaigns. Overall, information and participation probabilities
remain difficult to disentangle; for the purposes of
this study, we assume that all receivers of viral marketing
messages are aware of them. This assumption is likely to
hold for our three empirical studies, and our main findings
remain unchanged even when it does not. The only difference
is that the participation probability would also capture
the probability that a person is informed (and thus aware)
of the viral marketing campaign.
Effect of social position on used reach. Epidemiology
studies indicate that hub constellations foster the spread of
diseases (Anderson and May 1991; Kemper 1980), which
suggests in parallel that hubs should be more attractive for
seeding viral marketing campaigns. However, it is unclear
whether hubs actively and purposefully make use of their
potential reach. The deliberate use of reach is a common
assumption, and yet only Leskovec, Adamic, and
Huberman (2007) actually confirm that hubs send more
messages. Furthermore, their definition of hubs relies on
messaging behavior, such that it cannot offer generalizable
evidence of the assumption that hubs actively use their
greater reach potential.
In contrast, we anticipate that individual i’s used reach
(first generation), added to the used reach of successive
generations that originate from i’s initial direct reach (second
and further generations), which we call i’s “influence
domain” (Lin 1976), depends on the number of others who
already have received the message. In this setting, bridges
are advantageous because they can forward the message to
different parts of the network (Granovetter 1973) that have
not yet been infected with the viral campaign.
58 / Journal of Marketing, November 2011
Effect of social position on conversion rate. A person’s
social position might also indicate the degree of persuasiveness,
as measured by the conversion rate—namely, the
share of referrals that lead to successful referrals. Iyengar,
Van den Bulte, and Valente (2011) find that hubs are more
likely to be heavy users and that, therefore, their influence
is more effective, because they act in accordance with
their own recommendations (e.g., by making heavy use of
the innovation). Leskovec, Adamic, and Huberman (2007)
find that the success rate per recommendation decreases
with the number of recommendations made, which implies
that people have influence over a limited number of friends
but not over everybody they know. This result indicates
that the conversion rate decreases when hubs use their full
reach potential, though it does not preclude the notion that
hubs instead might select relevant subsets of recipients from
among their peers and thus achieve high conversion rates.
Thus, the effect of social position on the conversion rate
remains unclear. Goldenberg et al. (2009) still make what
they call the conservative assumption that hubs are not more
persuasive than others, though without empirical support.
Seeding Strategies
As our review in Table 1 shows, little consensus exists
regarding recommendations for optimal seeding strategies.
Four studies recommend seeding hubs, three recommend
fringes, and one recommends bridges. We analyze these
discrepant recommendations in turn.
If at least one of the determinants I i, P i, n i, or w i increases
with the connectivity of the sender i, and the
remaining determinants are not correlated with higher connectivity,
hubs should be the targets of initial seeding
efforts, because they spread viral information best—as indicated
in Hanaki et al. (2007), Van den Bulte and Joshi
(2007), and Kiss and Bichler (2008). Using hubs as initial
seeding points implies a “high-degree seeding” strategy.
In contrast, Watts and Dodds’s (2007) computer simulations
of interpersonal influence indicate that targeting
well-connected hubs to maximize the spread of information
works only under certain conditions and may be the exception
rather than the rule. They propose instead that a critical
mass of influenceable people, rather than particularly influential
people, drives cascades of influence. The impact on
triggering critical mass is not even proportional to the number
of people who hubs directly influence; rather, according
to Dodds and Watts (2004), the people most easily influenced
have the highest impact on the diffusion. Moreover, if
hubs suffer from information overload because of their central
position in the social network (Porter and Donthu 2008;
Simmel 1950), they must filter or validate the vast amount
of information they receive, such that they may be less
susceptible to information received from anyone outside
their trusted network. In their analytical model, Galeotti
and Goyal (2009) propose targeting low-degree members
instead, on the fringe of the network, if the probability of
adopting a product increases with the absolute number of
adopting neighbors. Sundararajan (2006) similarly suggests
seeding the fringes rather than the hubs, which we refer to
as a “low-degree seeding” strategy.
When analyses focus on the influence domain, encompassing
referrals beyond the first generation, it also
becomes necessary to consider centrality beyond the local

environment. Bridges who connect otherwise separated
subnetworks have vast influence domains, such that seeding
them might enable information to diffuse throughout parts
of the network and prevent a viral message from simply
circulating in an already infected, highly clustered subnetwork.
Accordingly, Rayport (1996) recommends exploiting
the strength of weak ties (i.e., bridges; Granovetter 1973)
to spread a marketing virus. From an opposite perspective,
Watts (2004) similarly recommends eliminating bridges to
prevent epidemics. We thus refer to the idea of seeding
bridges as a “high-betweenness seeding” strategy.
Finally, if there is no correlation between social position
and the determinants I i, P i, n i, and w i, or if the opposing
influences of the determinants nullify one another, there
should be no differences across the proposed strategies or
a random targeting. We also test this “random seeding”
strategy, which further serves as a benchmark situation in
which no information about the social network is available.
Method
We use three studies to empirically compare the success of
seeding strategies and identify which of the determinants are
influenced by the people’s social position. Our three studies
encompass two types of settings that are particularly relevant
for viral marketing. First, viral marketing campaigns
primarily aim to spread information, create awareness,
and improve brand perceptions, which are noneconomic
goals. Second, other campaigns attempt to increase sales
through mutual information exchanges between adopters
and prospective adopters, to trigger belief updating, such
that we can use an economic measure of success.
These goals map well onto the classification that Van den
Bulte and Wuyts (2007) provide to describe five reasons
for social contagion, the first two of which are especially
relevant for viral marketing campaigns. First, people may
become aware of the existence of an innovation through
WOM provided by previous adopters in a simple information
transfer. Second, people may update their beliefs about
the benefits and costs of a product or service. Third, social
contagion may occur through normative pressures, such
that people experience discomfort when they do not comply
with the expectations of their peer group. Fourth, social
contagion can be based on status considerations and competitive
concerns (i.e., the level of competitiveness between
two people). Fifth, complementary network effects might
cause social contagion, in which the benefit of using a product
or service increases with the number of users.
To examine both types of viral marketing campaigns, we
conduct two experimental studies (Studies 1 and 2) that
simulate viral marketing campaigns in which social contagion
mainly involves simple information transfers and
results in greater awareness as a noneconomic measure of
success. The aim is to compare the success of different
seeding strategies. In Study 3, we examine a viral marketing
campaign in which social contagion relies on belief
updating and results in sales (i.e., economic measure). In
Table 2, we summarize the complementary setup of the
three studies, which helps us overcome some individual
limitations of each study.
Experimental Comparison of Seeding Strategies
In Studies 1 and 2, we compare the success of our four
seeding strategies in different conditions and confirm the
robustness of the results across different settings. Trusov,
Bodapati, and Bucklin (2010) point out the neccessity to
conduct such experiments: In analyzing data from a major
social networking site, they find that only approximately
one-fifth of a user’s friends actually influence that user’s
activity on the site. However, they cannot discern how
responsive the “top influencers” are or whether marketers
should use information about underlying social networks to
seed their viral marketing campaigns. Therefore, they call
for further research that uses straightforward field experiments.
Because such experiments can help identify bestpractice
strategies, we compare the four seeding strategies
in two small-scale field experiments.
Study 1: Comparison of seeding strategies in a controlled
setting. We begin with a controlled setup to ensure internal
validity and control for willingness to actively participate P i
(see Table 1). We recruited 120 students from a German
university. The recruitment and commitment processes
ensured relatively similar participants in terms of communication
activity across treatments because all of them
expressed a willingness to contribute actively. Therefore,
we expect minimal variation in activity levels, compared
with a study in which respondents are unaware of their participation
or do not come into direct contact with the experimenter.
A prerequisite for participation was maintaining
an account on a specified online social networking platform
(similar to Facebook). Using proprietary software, we
automatically gathered each participant’s friends list from
the platform and then applied an event-based approach
to specify boundaries, such that we discarded all links
to friends who did not participate in the experiment.
The software Pajek calculated the sociometric measures
(degree centrality and betweenness centrality; see the Web
Appendix at http://www.marketingpower.com/jmnov11) for
each participant.
The social network thus generated consisted of 120
nodes (i.e., participants) with 270 edges (i.e., friendship
relations). Degree centrality ranged from 1 to 17, with a
mean of 4.463 and a standard deviation of 3.362. In other
words, the participants had slightly more than four friends
each, on average, in the respective, bounded social network.
The correlation (.592, p < �01) between the degree centrality
in this small, bounded network created by the artificial
boundary specification strategy (using the criteria “participation
in experiment”) and the degree centrality of the
entire network hosted by this social networking platform
(6.2 million unique users, November 2009) is striking. It
also supports Costenbader and Valente’s (2003) claim that
some centrality metrics are relatively robust across different
network boundaries. That is, the boundary we applied
does not appear to bias degree centrality, even for a sub
sample that comprises as little as .002% of the entire social
network. The betweenness centrality ranged from 0 to .320,
with a standard deviation of .053.
We used these sociometric measures to implement our
four seeding strategies. The seeding relied on the message
function of the social networking platform, such that
we sent unique tokens of information to a varying subset
of participants (the total population remained unchanged
throughout this experiment) and traced the contagion pro-
Seeding Strategies for Viral Marketing / 59

TABLE 2
Summary of Studies
Study 1 Study 2 Study 3
Seeding strategies Four seeding strategies: Four seeding strategies: Three seeding strategies:
•High degree (HD) •High degree (HD) •High degree (HD)
•Low degree (LD) •Low degree (LD) •Low degree (LD)
•High betweenness (HB) •High betweenness (HB) •Random
•Random (control) •Random (control)
Social contagion through Awareness (advertisement) Awareness (advertisement) Belief updating (service
referral)
Motivation Extrinsic motivation for
sharing (experimental
remuneration)
Intrinsic motivation (funny
video about university)
Extrinsic motivation for
sharing (additional airtime for
referral)
Seeding size 10% of network size 7% of network size Entire network
20% of network size
Seeding timing Sequential Parallel —
Social network 120 nodes (small network),
270 edges
1380 nodes (medium-sized
network), 4052 edges
208,829 nodes (very large
network), 7,786,019 edges
Number of treatments 16 = 4 × 2 × 2 4 —
Number of replications 2 (4 treatments missing) 1 —
Number of experimental
settings 28 4 —
Boundary of network Artificial Natural Natural
Design strengths Test of causality, strong
control due to experimental
setup, identification of
individual behavior due to
specific IDs
Design weaknesses Repeated measures due to
sequential timing, artificial
scenario
Specific finding HD and HB are comparable
and outperform random by
+39%–52% and LD by
factor 7–8
cess. These tokens were to be shared by initial recipients
with friends, who in turn were to spread them further. All
receivers were asked to enter the tokens on a website that
we created for this purpose, along with details about from
whom they received these tokens (called the “referrer”).
Because each participant was provided with unique login
information for this website, we could observe the number
of tokens entered on the website (and thus the number of
successful referrals SR i) by each individual i for each of
the seeding strategies. Furthermore, we could distinguish
whether the recipient received the tokens directly from the
experimenter (“Seeded by Experimenter”) or through viral
spreading from friends. We prohibited and did not observe
the use of forums or mailing lists to spread the tokens.
The experiment used a 4 × 2 × 2 full-factorial design.
Following the strategies we defined previously, we seeded
the tokens every few days to hubs (high-degree seeding),
fringes (low-degree seeding), bridges (high-betweenness
seeding), or a random set of participants. We varied the
number of initial seeds, such that the tokens were sent
to either 12 (10%) or 24 (20%) of the 120 participants.
60 / Journal of Marketing, November 2011
Test of causality, realistic
scenario
Potential interaction between
treatments, activity level of
individuals not controlled for,
individuals cannot be
identified
HD and HB are comparable
and outperform random by
+60% and LD by factor 3
Large real-world network
based on firm data,
identification of determinants
Missing edges between
noncustomers (HB could not
be tested), causality cannot
be tested
HD outperforms random by
factor 2 and LD by factor 8–9
We also varied the payment levels for successful referrals
to account for the potential effects of extrinsic motivation
(incentive for sharing yes/no). When they received
no incentive for sharing, participants earned remuneration
only if they correctly entered the secret token (∼.40 EUR
per token; for detailed instructions, see the Web Appendix
at http://www.marketingpower.com/jmnov11). In contrast,
under the incentives-for-sharing condition, they received
an additional monetary reward when they were named
as a referrer (.25 EUR per correctly entered token and
.20 EUR per referral; for detailed instructions, see the
Web Appendix).
Therefore, we systematically varied the 4×2×2 = 16 different
treatments, with two replications per treatment. The
limitations of the social networking platform’s messaging
system prevented us from replicating four specific treatments,
so we obtained a total of 28 experimental settings
(the potential maximum was 4 × 2 × 2 × 2 = 32 experimental
settings). Although we systematically varied the treatments,
we placed the low-incentive-for-sharing before the
high-incentive-for-sharing settings to avoid confusing participants
with different incentive instructions. We always

seeded one token and then captured all responses two
weeks after the seeding.
Overall, 55% of the participants actively spread or
entered unique tokens, resulting in 1155 responses. The
average number of tokens spread per experimental setting
was 41.25, with a standard deviation of 19.21. To
compare the success of the strategies, we use a randomeffects
logistic regression analysis that accounts for individual
behavioral differences according to each participant’s
responsiveness in each experimental setting. We use the
number of correctly entered tokens as a dependent variable,
which can be 1 if the token was correctly reported and 0 if
otherwise. With 120 participants and 28 experimental settings,
we obtained 3360 observations. As the independent
variables, we included dummy-coded treatment variables
that reflect our full-factorial design, as we detail in Table 3.
The model achieves a pseudo R-square of 15.5%.
The proportion of unexplained variance accounted for by
subject-specific differences due to unobserved influences,
labeled �, is greater than 90%. Compared with random
seeding, the high-degree seeding strategy yields a much
higher likelihood of response (odds ratio = 1�53) that is similar
to the high-betweenness seeding strategy (odds ratio =
1�39). In contrast, the low-degree seeding strategy dramatically
decreases the likelihood of response (odds ratio = �19).
Our treatment variable, high seeding (dummy coded
as 0 = 12 seeds and 1 = 24 seeds), positively influences
response likelihood. Furthermore, the type of incentive
offered drives the high odds ratio estimate, which might
explain why extrinsic motivation in the form of monetary
incentives is popular for viral marketing (e.g., recruit-afriend
campaigns offering rewards such as price discounts
or coupons for successful referrers; Biyalogorsky, Gerstner,
and Libai 2001). Finally, the participants who received the
token from the experimenter (“seeded by experimenter”)
exhibited a higher response likelihood, which is not surprising
because the information probability in this case equals 1.
TABLE 3
Individual Probability to Respond (i.e., Entering
the Correct Token at Experimental Website
[Random Effects Logit Model, Study 1])
Variable Odds Ratio SE
Seeding Strategy
Low degree �19∗∗∗ �04
High betweenness 1�39∗ �28
High degree 1�53∗∗ �31
High seeding 1�89∗∗∗ �28
High incentives 38�11∗∗∗ 26�07
Seeded by experimenter 14�36∗∗∗ Random coefficient: User ID
3�74
ln(� 2 u
� 3�36 �17
� u 5�36 �46
P �90 �02
R 2 (pseudo) .16
N 3360
∗ p < �1.
∗∗ p < �05.
∗∗∗ p < �01.
Notes: Two-tailed significance levels. n�s� = not significant. Reference
category: “random” seeding strategy, “low seeding,”“no
incentives,” and “was not seeded by experimenter.”
To compare the various seeding strategies directly, we
also varied the contrast specifications but left the rest of
the model unchanged, which produced the conditional odds
ratio matrix in Table 4. As Table 4 indicates, both highdegree
and high-betweenness seeding increase response
likelihood, in contrast with the random seeding strategy, by
39%–53%. Compared with the low-degree seeding strategy
(second column of Table 4), all other strategies are five
to eight times more successful. However, a comparison of
the two most successful seeding strategies, high betweenness
and high degree, does not yield significant differences.
This result has key implications for marketing practice, in
that degree centrality as a local measure is much easier to
compute than betweenness centrality, which requires information
about the structure of the entire network.
In summary, we find that the low-degree seeding strategy
is inferior to the other three seeding strategies and that both
high-betweenness and high-degree seeding outperform the
random seeding strategy but yield comparable results. However,
we also acknowledge that this experiment might suffer
from sequential effects, which would limit the validity
of our separate analysis of each experimental setting. The
behavior of a respondent in one experimental setting might
be influenced by his or her experience in prior experimental
settings. This problem is driven by the limited number
of participants in our experiment. Therefore, in Study 2 we
include more participants and avoid sequential effects by
implementing the four seeding strategies simultaneously.
Study 2: Comparison of seeding strategies in a field setting.
In a second field experiment, we focused on the entire
online social network of all students enrolled in the MBA
program at the same university as in Study 1. Thus, the
network boundary is defined by participation in the program.
We collected contact information for 1380 students
(1380 nodes, 4052 edges) by crawling the same social networking
platform to collect information on friendships, and
we then calculated the sociometric measures as in Study 1.
TABLE 4
Conditional Odds Ratios of Seeding Strategies
(Study 1)
Low High High
Degree Random Betweenness Degree
Low degree — �19 ∗∗∗ �13 ∗∗∗ �12 ∗∗∗
Random 5�37 ∗∗∗ — �72 ∗ �65 ∗∗
High
betweenness 7�47 ∗∗∗ 1�39 ∗ — �91 n�s�
High degree 8�19 ∗∗∗ 1�53 ∗∗ 1�10 n�s� —
∗ p < �1.
∗∗ p < �05.
∗∗∗ p < �01.
Notes: Two-tailed significance levels. n�s� = not significant. Read the
second column as follows: The odds that a person reacts
to the strategy of random seeding is 5.37 times as large as
that for low-degree seeding, 7.47 times as large in the strategy
of high-betweenness seeding as for low-degree seeding,
and 8.19 times as large in the strategy of high-degree seeding
as for low-degree seeding. The conditional odds ratio
of the two seeding strategies relate inversely. For example,
the odds ratios of random and low degree relate as follows:
�19 = 1/5�37.
Seeding Strategies for Viral Marketing / 61

The mean degree centrality (standard deviation) is 5.872
(7.318). Study 2 also reveals a high and significant correlation
between degree centrality in the bounded network
(1380 MBA students at the university) and degree centrality
in the entire network of the social networking platform
(6.2 million unique users in November 2009). The Pearson
correlation of .824 (p < .001) thus suggests that the number
of friends reported is also a good indicator of degree
centrality in abounded network.
As a proxy for the level of activity, we also used the
time since the last profile update in Study 2. We acquired
information about 849 update time stamps (we could not
access 531 due to the privacy restrictions set by users). On
average, users updated their profile 25.7 weeks ago (Mdn =
15�0), and we observed a weak but significant correlation
between degree centrality and time (in weeks) since the last
profile update (r = −�192, p < �01). We also observed a correlation
between betweenness centrality and time since the
last profile update (r = −�154, p < �01). These negative correlation
simply that participants who updated their profiles
more recently (and probably update them more frequently)
are also more central in the social network. In other words,
activity correlates with centrality and may be an additional
determinant of the viral spread of information in this setting.
In terms of gender (805 male, 569 female, and 6 missing
observations), male participants were more central, such
that the average female participant had .92 fewer connections
than the average male (p < �05). However, this gender
difference becomes insignificant if we control for activity.
The experimental setup for Study 2 was somewhat different.
First, the four treatment groups (hubs, bridges, fringes,
or random sample) were all seeded on the same day. Second,
we eliminated the incentive variation, such that we
did not use extrinsic monetary incentives to stimulate participation.
Third, we did not vary the seeding size and
sent a reminder out to the initial seeds seven days after
the initial seeding. The seeding included 95 participants in
each of the four treatments (70 on Day 1, 25 on Day 2),
or 7% of the total network (which is in line with Jain,
Mahajan, and Muller 1995). The seed message contained a
60
40
20
0
FIGURE 1
Development of the Number of Unique Visits Over Time (Study 2)
Entry Page: Cumulative Number of Unique Visits
80
0 2 4 6 8 10 12 14 16 18 20
Time (Days)
62 / Journal of Marketing, November 2011
High degree
High betweenness
Random
Low degree
65
63
40
22
unique URL for a website with a funny video that we produced
about the participants’ university (the landing page
and video were identical for all treatments). By producing
a new video specifically for this second field experiment,
we ensured that the viral marketing stimulus (i.e., content)
was unknown to all participants. Furthermore, we predicted
that the link to the video would be distributed preferentially
to fellow students (from which we obtained mutual
online social network relationships), rather than to others
outside the university’s social network. In other words, the
social network for Study 2 should represent a coherent,
self-contained social community. One MBA student served
as the initiator who seeded the message to others, according
to the chosen seeding strategy. In addition to the link
to the particular entry page, the message indicated that the
addressees could find a funny video about the university
that had just been created by the initiator.
We tracked website visits for the entry pages and video
download pages of the four sites (one for each strategy)
for 19 days. Figure 1 compares the success of the seeding
strategies. The rank order with respect to their success,
across both dependent variables, is consistent with
the results from our first experiment. That is, high-degree
and high-betweenness seeding clearly outperform both lowdegree
and random seeding. For example, in terms of
videos watched, the high-degree seeding strategy yielded
more than twice the number of responses than did random
seeding. Information about social position thus made
it possible to more than double the number of responses.
We also estimated two random-effects linear models (one
for the entry page, one for the video page) in which we
treated each of the 19 days as a unit of observation, for
which we have four observations. The dependent variable
is thus the number of unique visits for the entry page and
the number of unique video requests from the video page.
We included the seeding and reseeding (reminder) days as
dummy variables and added another dummy variable to
account for weekends. The seeding strategies also are coded
as dummy variables, and the experimental day is a unit
specific random coefficient. Table 5 illustrates the results.
Video Page: Cumulative Number of Unique Visits
80
60
40
20
Seeding activity
High degree
43
35
High betweenness
Random 17
12
Low degree
0
0 2 4 6 8 10 12 14 16 18 20
Time (Days)

TABLE 5
Number of Visits per Day (Random Effects Model, Study 2)
Entry Page Video Page
Unique Visits Unique Visits
Variable Coefficient SE Coefficient SE
High-degree seeding 2�263 ∗∗∗ �766 1�623 ∗∗∗ �547
High-betweenness seeding 2�158 ∗∗∗ �766 1�211 ∗∗ �547
Random seeding �947 n�s� �766 �263 n�s� �547
Seeding day or reseeding 7�128 ∗∗∗ 1�276 4�005 ∗∗∗ �727
Weekend −1�026 n�s� 1�569 −�636 n�s� �794
Intercept −�127 n�s� �911 −�078 n�s� �524
Random Coefficient: Experimental Day
� e 2.375 1.642
P .519 .290
R 2 (overall) .475 .436
∗ p < �1.
∗∗ p < �05.
∗∗∗ p < �01.
Notes: Two-tailed significance levels. n�s� = not significant. Reference categories: “low-degree seeding” and “weekdays” and “no seeding day.”
The models for both the entry and video pages are highly
significant, with explained overall variances (adjusted R 2 )
of 47.5% and 43.6%. The results in Figure 1 confirm our
previous observations: High-degree and high-betweenness
seeding yield comparable results and are three times more
successful than low-degree seeding and 60% more successful
than random seeding. Days with seeding or reseeding
activities yield more unique visits. Responsiveness declined
on weekends (albeit insignificantly), perhaps due to the
overall higher level of online activity by these students on
weekdays.
In summary, Study 2 supports our findings from Study 1
that seeding to hubs and bridges is preferable to seeding
to fringes. However, we also note the potential for interactions
among the activities associated with the four seeding
strategies in Study 2. For example, a participant might have
watched the video after receiving a message from seeding
strategy A and then receive a nearly identical message from
seeding strategy B, in which case this participant is unlikely
to click the link again to watch the same video. Thus,
seeding strategies that foster faster diffusion may have an
advantage that could bias the result and lead to over estimations
of the success of high-degree and high-betweenness
seeding in contrast with random and low-degree seeding.
However, in Study 1, such crossings were not possible as a
result of the sequential timing, and the results remained the
same. Neither Study 1 nor Study 2 identifies the reasons for
the superiority of specific seeding strategies. In addition,
we cannot distinguish between first- and second-generation
referrals. We address these shortcomings in Study 3.
Comparison of the Effect of Seeding Strategies
on the Determinants of Social Contagion in a
Real-Life Viral Marketing Campaign (Study 3)
For Study 3, a mobile phone service provider stimulated
referrals (through text messages) to attract new customers.
The provider tracked all referrals, so we can compare the
economic success of different seeding strategies and analyze
the influence of the corresponding sociometric measures
on all determinants of social contagion (Table 1) in
a real-life setting. This helps us identify the reasons for
any differences. Thus, Study 3 enables us to decompose
the effect of the different determinants that drive the social
contagion process, including participation probability P i,
the used reach n i, the mean conversion rate w i of all referrals
made by ion the expected number of referrals R i, and
the expected number of successful referrals SR i. The viral
marketing campaign of the mobile phone service provider
featured text messages sent to the entire customer base
(n = 208�829 customers), promising a 50% higher reward
than the regular bonus of E 10 worth of airtime for each
new customer referred in the next month. In total, 4549 customers
participated in the campaign, initiating 6392 firstgeneration
referrals, which was a 50% increase over the
average number of referrals. We anticipate that social contagion
works through belief updating, as prospective customers
talk to adopters about the product. Furthermore, in
Becker’s (1970) terms, we classify this product as a lowrisk
offering (cf. trials of untested drugs).
Our analysis of the social contagion process is based on a
rich data set; each referral activity was logged in the online
referral system of the company, because customers had to
initiate the referral messages to friends online. Successful
referrals were confirmed during the registration process of
the new customers, who had to identify their referrer to trigger
the payment of the referral premium. Thus, we gathered
information about whether customers acted on the stimulus
of the referral, captured by the variable program participation
P i, as well as the number of referrals R i and the
number of successful referrals SR i. The mean conversion
rate per referrer w i can be inferred from a comparison of
R i and SR i. We used individual-level communication data
and the number of text messages to others to calculate the
(external) degree centrality. (In total, we evaluated more
than 100 million connections. 1 ) We assumed that any telephone
call or text message between people (independent of
the direction) reflected social ties. Thus, degree centrality
1 All individual-level data were made anonymous with a multistage
encryption process, undertaken by the firm before the analysis.
At no point was any sensitive customer information, such as
names or telephone numbers, disclosed.
Seeding Strategies for Viral Marketing / 63

equals a count of the total number of unique communication
relationships. However, the service can only be referred
to current non customers, so the degree centrality metric
accounts only for ties that customers had to people outside
the service network at the beginning of the viral marketing
campaign, which makes it a form of external degree
centrality. We lacked information about the relationships
of people who were not customers, so we could not measure
betweenness centrality and test the high-betweenness
seeding strategy in Study 3.
We used the following customer characteristics as covariates:
demographic information including age (in years) and
gender (1 = female, and 0 = male); service-specific characteristics,
such as customer tenure (i.e., length of the relationship
with the company in months); and the tariff plan.
We operationalized the tariff plan with a dichotomous variable
to indicate whether the customer chose a community
tariff (=1, including a reduced per-minute price for
calls within the network) or a one-price tariff (=0). Furthermore,
we used two measures of customers’ trust in the
service: payment type (dichotomous variable: automatic =
1, and manual = 0) and refill policy (dichotomous variable:
automatic = 1, and manual = 0). In the case of automatic
payment and refill, customers provided credit card details to
the service provider. Finally, we included information about
the acquisition channel for each customer (1 = offline/retail,
and 0 = online). As additional controls, we include information
on the individual service usage of the customer,
namely, average monthly airtime (in minutes) and monthly
short message service (SMS)—that is, the average monthly
number of SMS sent by a customer.
Our model reflects the two-stage process for each participant,
who first decides whether to participate (P i) and
then chooses to what extent to participate (n i�. A specific
characteristic of the first stage is the relatively large share
of zeros (i.e., nonparticipants), whereas observed values for
the second stage are count measures and highly skewed.
This data structure requires specific two-stage regression
models: either inflation models, such as the zero-inflated
Poisson (ZIP) regression (Lambert 1992), or hurdle models,
such as the Poisson-logit hurdle regression (PLHR) model
(Mullahy 1986). We use a PLHR model, which combines
a logit model to account for the participation decision and
a zero-truncated Poisson regression to analyze the actual
outcomes of participation (e.g., number of successful referrals).
2 In our PLHR specification, the binary variable P i
indicates whether individual i participates in the referral
program (hurdle or logit model). In addition, Used Reach n i
indicates how many referrals the individual i initiates, conditional
on the decision to participate (P i = 1). As an extension,
Converted Reach CR = �n i × w i� indicates how many
successful referrals individual i initiates, again conditional
on the decision to participate. Note that Used Reach n i
and Converted Reach CR i are equivalent to Referrals R i
and Successful Referrals SR i, respectively, conditional on
a program participation probability P i = 1. These variables
2 We choose PLHR over ZIP because the logit stage of the former
is designed to determine what leads to participation (i.e., identifying
referrers, in which we are interested), whereas the inflation
stage of ZIP tries to detect “sure zeros” (i.e., nonparticipants).
64 / Journal of Marketing, November 2011
provide the dependent variables in the Poisson regression
of our PLHR specification.
Let P∗ i be the latent variable related to Pi, n∗ i be the
censored variable related to ni, and CR∗ = �n∗ i × wi�∗ be
the censored variable related to CR = �n∗ i × wi�. Together
with the explanatory variable of (external) degree centrality
and the covariates (age, gender, payment type, refill policy,
acquisition channel, and customer tenure), the PLHR can
be specified as follows:
�
1 if P
Pi =
∗ i > 0
where P∗
i
0 otherwise� = � (3)
P0i + �Pij × XPij + �Pi� (4)
and
(5)
n i =
�
n ∗ i
if P ∗ i >0
0 otherwise�
CR = �n i × w i� =
where n∗
i = � UR0i +� URij ×X URij +� URi�
�
�ni × wi�∗ if CR∗ i > 0
0 otherwise�
where �n i × w i� ∗ = � CR0i + � CRij × X CRij + � CRi�
Thus, X ij contains the explanatory variables j (i.e., degree
centrality and the covariates) and the error terms � Pi, � URi,
and � CRi, which represent unobserved influences on participation
probability, used reach, and the number of successful
referrals.
Seeding strategies in first-generation models. In a first
step, we restricted our analysis to first-generation models;
we only considered referrals directly initiated by customers
who received the seeding stimulus during the viral marketing
campaign. Table 6 contains the parameter estimates
of the PHLR model. The results can be interpreted in two
stages: first, what drives the participation of seeded customers
in the viral marketing campaign (logit component =
LC) and, second, among these participants, what influences
the number of referrals and successful referrals (Poisson
component = PC). With regard to the covariates’ impact
on program participation, we find significant effects of the
demographic variables gender (�LC 2n∗ = −�2171, p < �01) and
age (�LC 3n∗ = −�0209, p < �01), indicating that male and older
customers are more likely to participate, as are customers
with short customer tenures who have just recently adopted
the service (�LC 8n∗ = −�0016, p < �01). The latter finding aligns
with cognitive dissonance theory, in that these customers
might communicate shortly after their purchase decision
to reduce dissonance (Festinger 1957). Furthermore, we
found that customers acquired online are more strongly
engaged in the (online-based) referral program than customers
acquired through the retail channel (�LC 6n∗ = −�9843,
p < �01). A one-price tariff seems easier to communicate;
seeded customers with that tariff option are more likely to
participate in the referral program (�LC 7n∗ = −�1433� p < �01).
With regard to the usage covariates, we find positive and
significant values for monthly airtime (�LC 9n∗ = �0007� p < �01)
and monthly SMS (�LC 10n∗ = �0010� p < �01). The influences
of most covariates are comparable between the logit and
Poisson regression stages, except the acquisition channel,
in that retail customers are less likely to participate in the
program, but if they do, they exhibit significantly greater

activity than online customers (�PC 6n∗ = �7388� p < �01). The
same applies to the usage covariates: Here, monthly air-
time (�PC 9n∗ = −�0007� p < �1) and monthly SMS (�PC 10n∗ =
−�0018� p < �01) show negative effects on participation.
However, the influence of (external) degree centrality
varies between the stages of the model. In the logit regression
stage, degree centrality has a positive and significant
influence on the likelihood to participate P i in the refer-
ral program (�LC 1n∗ = �0032� p < �01). Confirming the results
of Studies 1 and 2, this finding shows that customers with
high degree centrality are more likely to participate than
those with low degree centrality (average degree centrality
of participants = 45�3 vs. nonparticipants = 36�5). However,
in the Poisson regression stage that analyzes only
the group of active referrers, the effect of degree central-
ity is mixed. We find a significant, positive effect on used
reach (�PC 1n∗ = �0012� p < �01), such that customers with highdegree
centrality are not only more likely to participate but
also more active when participating in the viral marketing
campaign. However, we find no significant effect of degree
TABLE 6
Determinants of Number of Referrals, Number of Successful Referrals, and Influence Domain
(Poisson-Logit Hurdle Regression Models, Study 3)
Converted Reach Conditional Influence
Used Reach n ∗ CR = �n ∗ w� ∗ Domain ID T
i
Variable Coefficient SE Coefficient SE Coefficient SE
Logit Component
Degree centrality � 1 �0022 ∗∗∗ �0003 �0021 ∗∗∗ �0004 �0021 ∗∗∗ �0004
Covariates
Gender � 2 −�2287 ∗∗∗ �0318 −�2527 ∗∗∗ �0337 −�2527 ∗∗∗ �0337
Age � 3 −�0202 ∗∗∗ �0013 −�0190 ∗∗∗ �0014 −�0189 ∗∗∗ �0014
Payment type � 4 �0913 ∗∗ �0389 �0668 n�s� �0412 �0668 n�s� �0412
Refill policy � 5 −�0122 n�s� �0376 �0174 n�s� �0396 �0174 n�s� �0396
Acquisition channel � 6 −�9848 ∗∗∗ �0506 −1�0535 ∗∗∗ �0545 −1�0530 ∗∗∗ �0545
Tariff plan � 7 −�1371 ∗∗∗ �0395 −�1253 ∗∗∗ �0419 −�1253 ∗∗∗ �0420
Customer tenure � 8 −�0016 ∗∗∗ �0001 −�0016 ∗∗∗ �0001 −�0016 ∗∗∗ �0001
Monthly airtime � 9 �0007 ∗∗∗ �0002 �0007 ∗∗∗ �0002 �0007 ∗∗∗ �0002
Monthly SMS � 10 �0010 ∗∗∗ �0002 �0009 ∗∗∗ �0002 �0010 ∗∗∗ �0002
Intercept −2�3825 ∗∗∗ �0727 −2�534 ∗∗∗ �0770 −2�5341 ∗∗∗ �0770
Poisson Component
Degree centrality � 1 �0026 ∗∗∗ �0006 �0001 n�s� �0033 −�0025 ∗∗∗ �0007
Covariates
Gender � 2 −�1539 ∗∗∗ �0519 −�1177 n�s� �2206 −�3323 ∗∗∗ �0496
Age � 3 −�0098 ∗∗∗ �0020 −�0104 n�s� �0087 −�0112 ∗∗∗ �0019
Payment type � 4 −�3908 ∗∗∗ �0581 −�1807 n�s� �2492 −�1379 ∗∗ �0544
Refill policy � 5 −�3417 ∗∗∗ �0761 −�1369 n�s� �2819 �0033 n�s� �0608
Acquisition channel � 6 �7408 ∗∗∗ �0561 �5902 ∗∗ �2592 �6273 ∗∗∗ �0548
Tariff plan � 7 −1�0740 ∗∗∗ �0473 −1�1426 ∗∗∗ �2067 −�5322 ∗∗∗ �0466
Customer tenure � 8 −�0007 ∗∗∗ �0002 −�0007 n�s� �0007 −�0013 ∗∗∗ �0001
Monthly airtime � 9 −�0007 ∗ �0004 −�0000 n�s� �0000 �0000 n�s� �0003
Monthly SMS � 10 −�0018 ∗∗∗ �0005 −�0001 n�s� �0019 �0005 n�s� �0004
Intercept �9811 ∗∗∗ �0941 −1�4633 ∗∗∗ �4181 1�1008 ∗∗∗ �0905
Log-likelihood value −25�163 −19�850 −23�723
BIC 50,596 39,969 47,714
N 208,829
∗ p < �1.
∗∗ p < �05.
∗∗∗ p < �01.
Notes: Two-tailed significance levels. n�s� = not significant.
centrality on the referral success of active referrers (� PC
1CR =
−�0002� n�s�).
This result is further confirmed when we analyze the
mean conversion rate of referrals per referrer w i = CR i/n ∗ i
(see Table 7). Again, we do not find a significant effect
of degree centrality for active referrers (� 1w = �0001� n�s�).
Thus, our results offer no support for the assumption that
participating central customers are more persuasive referrers
or better selectors of potential referral targets.
Next, considering that viral marketing campaigns can
be costly, we attempt to identify customers who are most
likely to participate and generate (successful) referrals. We
use the estimated participation probability calculated from
the results of the selection model (see Table 6, “Logit Component”)
to group the full customer base into cohorts and
then compare these cohorts according to their observed participation,
referral, and conversion rates and degree centrality
(see Table 8). The top 5000 cohort corresponds to
a high-degree and the bottom 5000 cohort to a low-degree
seeding strategy; the results in the “Average” column correspond
to random seeding.
Seeding Strategies for Viral Marketing / 65

TABLE 7
Determinants of Conversion Rates, Active
Referrers (Poisson Regression, Study 3)
Poisson Regression Model
Conversion w = CR/n ∗
Variable Coefficient SE
Degree centrality � 1 −�0001 n�s� �0004
Covariates
Gender � 2 �0788 ∗∗ �0350
Age � 3 �0047 ∗∗∗ �0015
Payment type � 4 �1104 ∗∗∗ �0431
Refill policy � 5 �1079 ∗∗∗ �0409
Acquisition channel � 6 −�4951 ∗∗∗ �0616
Tariff plan � 7 �4638 ∗∗∗ �0472
Customer tenure � 8 �0002 ∗ �0001
Intercept −1�2014 ∗∗∗ �0848
Log-likelihood value −5� 074
N 4,549
∗ p < �1.
∗∗ p < �05.
∗∗∗ p < �01.
Note: Two-tailed significance levels. n�s� = not significant. For the
Poisson regression model, we used ln�n� as offset variable.
The results in Table 8 clearly confirm the positive
correlation between degree centrality and the success of
viral marketing: As the estimated participation probability
increases, observed participation, referral, and conversion
rates (i.e., total number of participants, referrals, or successful
referrals divided by number of seeded customers in
the cohort) and degree centrality increase as well. Thus,
the participation rate of the top 5000 cohort is a multiple
of that of the bottom 5000 cohort (4.4% vs. .5%), with a
much higher average degree centrality (70.8 vs. 18.0). A
high-degree seeding strategy would be nearly nine times
as successful as a low-degree strategy. Compared with the
average value of a random strategy, the top 5000 cohort participation
and degree centrality are twice as high; therefore,
targeting hubs doubles the performance of random seeding
for a sample of the same size. Thus, Study 3 clearly shows
the significant, positive effect of degree centrality on viral
marketing participation and activity, in strong support of a
high-degree seeding strategy. However, the results of the
Poisson regression model do not indicate higher referral
success of hubs within the group of active referrers.
Seeding strategies in multiple-generation models. In the
second step of our empirical analysis, we extended the measure
of success to account for a fuller range of the effects
of seeding efforts by including more than one generation
of referrals. First-generation referrals initiate a viral process
that should continue in further generations. The extent
of this viral branching may differ across seeding strategies,
due to their ability to reach different parts of the social network.
Thus, the optimal seeding strategy might change if
we consider multiple generations.
To capture this form of success, we measured all subsequent
referrals that originated from a first-generation referral
during the campaign. We limited the observation period
to 12 months because the company repeated the referral
campaign 13 months after the initial seeding. During our
observation period, the company did not engage in other
promotions that directly focused on referrals, nor did we
find any anomalies (e.g., drastic increases or decreases) in
company-owned or competitive marketing spending.
In the first year after the campaign, 20.8% of all firstgeneration
referrals became active referrers themselves, and
5.8% did so multiple times. We observed viral referral
chains with a maximum length of 29 generations; on average,
every first-generation referral during the campaign led
to .48 additional referrals.
The dependent variable for this analysis is the influence
domain of all successful referrals of a specific
first-generation customer, which equals the number of
successful first-generation referrals, plus the number of successful
referrals in successive generations during the subsequent
12 months. 3 For example, Figure 2 depicts the
influence domain of a referral customer X that spans 22
additional successful referrals over seven generations.
3 Note that, by definition, there is no overlap of influence
domains between two origins; every referred customer has an indegree
of 1, and only one specific referrer is rewarded for every
new customer.
TABLE 8
Relationship of Conversion Rates and Degree Centrality, Full Sample (Study 3)
Customer Cohort
(According to Estimated Participation Probabilities)
Top 5000 Top 10,000 Top 20,000 Top 50,000 Bottom 5000 Average
Participants
Total participation 220 378 671 1385 24 —
Participation rate 4�4% 3�8% 3�4% 2�8% �5% 2�2%
Referrals
Total referrals 292 489 856 1783 26 —
Referral rate 5�8% 4�9% 4�3% 3�6% �5% 3�0%
Successful referrals
Total conversions 191 330 598 1233 19 —
Conversion rate 3�8% 3�3% 3�0% 2�5% �4% 2�0%
Average degree centrality 70�83 60�21 52�13 45�42 18�01 36�48
Notes: Top (Bottom) 5,000/10,000/� � � refer to the cohort of customers with the highest (lowest) estimated participation probabilities, according
to the coefficient estimates of the logit component reported in Table 6. We calculated rates by dividing the total number of participants,
referrals, or successful referrals by the total number of customers in the cohort.
66 / Journal of Marketing, November 2011

3
FIGURE 2
Influence Domain of a Referral Campaign
Participant (Study 3)
2
1
1
Customer X
Initial campaign (origin)
stimulus
2
2
3 3
3
3
4
3 3 4
4
5
7
6
6
7
Referral
generations
7
Customer Y
The parameter estimates of the PLHR model for this
multiple-generation model appear in the right-hand column
of Table 6. The dependent variable IDT i is conditional
on program participation (Pi = 1). When we compare the
regression model parameters across the different dependent
variables, we find similar results for Influence Domain
IDT i and Used Reach ni (e.g., no significant effect of the
usage covariates) but with one important difference: Our
focal variable, degree centrality, is negative (�PC 1ID = −�00205,
p < .01) in the Poisson regression model. That is, among the
participants, more central customers have a smaller influence
domain. The observed network structure of the referral
processes offers a potential explanation of this surprising
TABLE 9
Determinants of Unconditional Influence Domain (OLS Model, Study 3)
OLS Model Unconditional
Poisson Regression
Model Unconditional
Influence Domain (ID R
i ) Influence Domain �IDR
Variable Standard Coefficient SE Coefficient SE
Degree centrality � 1 �010 ∗∗∗ �000 �002 ∗∗∗ �000
Covariates
Gender � 2 −�015 ∗∗∗ �002 −�358 ∗∗∗ �028
Age � 3 −�024 ∗∗∗ �000 −�024 ∗∗∗ �001
Payment type � 4 �001 n�s� �002 �013 n�s� �033
Refill policy � 5 �000 n�s� �002 �003 n�s� �033
Acquisition channel � 6 −�025 ∗∗∗ �002 −�680 ∗∗∗ �039
Tariff plan � 7 −�016 ∗∗∗ �002 −�433 ∗∗∗ �030
Customer tenure � 8 −�031 ∗∗∗ �004 −�002 ∗∗∗ �000
Intercept �091 ∗∗∗ �004 −1�509 ∗∗∗ �058
R 2 (pseudo) .05 .03
N 208,829 208,829
∗ p < �1.
∗∗ p < �05.
∗∗∗ p < �01.
Notes: Two-tailed significance levels. n�s� = not significant.
result. For hubs, we mostly observe short referral chains (if
at all), whereas fringe customers who participate in the viral
marketing campaign demonstrate significantly longer referral
chains. For example, in Figure 2, the fringe customer X
reacts to the campaign and refers the service. Within two
generations, this referral reaches actor Y, who initiates a
total of 15 additional referrals, which increases the influence
domain of X to 22.
Because we find a positive effect of high degree centrality
in the selection model but a negative effect in the
regression model, the overall effect of degree centrality
remains unclear. For a simple test, we performed both an
ordinary least squares (OLS) and a simple Poisson regres-
sion, with the unconditional Influence Domain ID R i
as the
dependent variable for the complete sample of 208,829 customers
who received the viral marketing campaign stimulus.
According to the results in Table 9, the standardized
beta for degree centrality is positive and significant in both
regressions (�OLS 1ID = �010� �PR
1ID = �002� p < �001), such that the
overall effect of high degree centrality as a selection criterion
for seeding a viral marketing campaign is positive.
Therefore, high-degree seeding remains the more successful
strategy, even if we account for a multiple-generation
viral process.
Robustness checks. To check the robustness of our findings,
we also analyzed our data with a set of alternative
approaches, including a Poisson-based (ZIP regression) and
probit–OLS combinations (e.g., Tobit Type II models). Our
core results pertaining to the influence of degree centrality,
including its positive influence on the selection stage
to determine the likelihood of participation and its negative
effect on the influence domain in the regression stage, hold
across all tested models. The results also remain unchanged
when we incorporate alternative individual-level covariates,
such as monthly mobile charges, that represent the attractiveness
of a customer to the provider.
Unlike Studies 1 and 2, Study 3 does not allow us to
assess the causal effect of the seeding strategy on referral
Seeding Strategies for Viral Marketing / 67
i �

success unambiguously. However, it reflects a real-life marketing
application and is based on detailed firm data, such
that it strikingly illustrates the power of network information
in real-life situations.
Research Contribution
General Discussion
Inspired by conflicting recommendations in previous studies
regarding optimal seeding strategies for viral marketing
campaigns, this article empirically compares the performance
of various proposed strategies, examines the magnitude
of differences, and identifies determinants that are
responsible for the superiority of a particular seeding strategy.
To the best of our knowledge, such an experimental
comparison of seeding strategies is unprecedented; previous
literature is based solely on mathematical models and computer
simulations. Our real-life application provides some
answers to controversies about whether hubs are harder
to convince, whether they make use of their reach, and
whether they are more persuasive.
Marketers can achieve the highest number of referrals,
across various settings, if they seed the message to hubs
(high-degree seeding) or bridges (high-betweenness seeding).
These two strategies yield comparable results and both
outperform the random strategy (+52%) and are up to eight
times more successful than seeding to fringes (low-degree
seeding). The superiority of the high-degree seeding strategy
does not rest on a higher conversion rate due to a higher
persuasiveness of hubs but rather on the increased activity
of hubs, which is in line with previous findings (e.g.,
Iyengar, Van den Bulte, and Valente 2011; Scott 2000).
This finding is persistent even when we control for revenue
of customers and thus demonstrates the importance
of social structure beyond customer revenues and customer
loyalty.
Research that suggests a low-degree seeding strategy is
usually based on the central assumption that highly connected
people are more difficult to influence than less connected
people because highly connected people are subject
to the influence of too many others (e.g., Watts and Dodds
2007). Our results (Studies 1 and 3) reject this assumption
and underline Becker’s (1970) suggestion: Hubs are more
likely to engage because viral marketing works mostly
through awareness caused by information transfer from previous
adopters and through belief updating, especially for
low-risk products. The low perceived risk means hubs do
not hesitate before participating. Furthermore, when social
contagion occurs mostly at the awareness stage, the possible
disproportionate persuasiveness of hubs is irrelevant. As
long as the social contagion occurs at the awareness stage
through simple information transfer, hubs are not more
persuasive than other nodes (Godes and Mayzlin 2009;
Iyengar, Van den Bulte, and Valente 2011).
Our analysis of a mobile service provider’s viral marketing
campaign reveals that hubs make slightly more use of
their reach potential. Furthermore, for the group of participating
customers, we find a negative influence of greater
connectivity on the resulting influence domains. Although
in epidemiology studies infectious diseases spread through
hubs, we find that well-connected people do not use their
68 / Journal of Marketing, November 2011
greater reach potential fully in a marketing setting. Spreading
information is costly in terms of both time invested
and the effort needed to capture peers’ attention. Furthermore,
hubs may be less likely to reach other previously
unaffected central actors, such that they are limited in their
overall influence domain. The compelling findings from
sociology and epidemiology thus appear to have been incorrectly
transferred to targeting strategies in viral marketing
settings.
Nevertheless, the social network remains a crucial determinant
of optimal seeding strategies in practice because
a social structure is much easier to observe and measure
than communication intensity, quality, or frequency. Furthermore,
we find robust results even when we control for
the level of communication activity. Therefore, companies
should use social network information about mutual relationships
to determine their viral marketing strategy.
Managerial Implications
Viral marketing is not necessarily an art rather than a
science; marketers can improve their campaigns by using
sociometric data to seed their viral marketing campaigns.
Our multiple studies show that information about the social
structure is valuable, in that seeding the “right” consumers
yields up to eight times more referrals than seeding the
“wrong” ones. In contrast with random seeding, seeding
hubs and bridges can easily increase the number of successful
referrals by more than half. Thus, it is essential
for marketers to adopt an appropriate seeding strategy and
use sociometric data to increase their profits. We conclude
that adding metrics related to social positions to customer
relationship management databases is likely to improve targeting
models substantially.
Many companies already have implicit information about
social ties that they could use to calculate explicit sociometric
measures. Telecommunication providers can exploit
connection data (as we did in Study 3), banks possess
data about money transfers, e-mail providers might analyze
e-mail exchanges, and companies can evaluate behaviors
in company-owned forums. Many companies also have
indirect access to information on social networks, such as
Microsoft through Skype or Google through its Google
mail services, and they could begin to use information
obtained this way (e.g., Hill, Provost, and Volinsky 2006;
Aral, Muchnik, and Sundararajan 2009). Such network
information is further available in the form of friendship
data obtained from online social communities such as Facebook
or LinkedIn (e.g., Hinz and Spann 2008).
Remarkably, we reveal that to target a particular subnetwork
(e.g., students of a particular university, Study 2) with
a viral marketing message, the use of the respective subnetwork’s
sociometric measures is not absolutely required to
implement the desired seeding strategies. Instead, because
the sociometric measures of subnetworks and their total
network are highly correlated, marketers can use the sociometric
measures of the total network, without undertaking
the complex task of determining exact network boundaries.
Conversely, this appealing result also allows marketers to
feel confident in inferring the connectivity of a person in an
overall network from information about his or her connectivity
in a natural subnetwork. Moreover, because betweenness
centrality requires knowledge about the structure of

the entire network, as well as a complex, time-consuming
computation, degree centrality seems to be the best sociometric
measure for marketing practice (see also Kiss and
Bichler 2008).
According to these insights, marketers should pick highly
connected persons as initial seeds if they hope to generate
awareness or encourage transactions through their viral
marketing campaigns because these hubs promise a wider
spread of the viral message. As long as the social contagion
operates at the awareness stage through information transfer,
we do not observe that hubs are more persuasive. Thus,
the sociometric measure of degree centrality cannot be used
to identify persuasive seeding points. The use of demographics
and product-related characteristics (see Table 7)
seems more promising for this purpose. Study 1 reveals
that monetary incentives for referring strongly increase the
spread of viral marketing messages, which supports the use
of such incentives. However, they would also make viral
marketing more costly than is commonly assumed.
Finally, expertise in the domain of social networks is
valuable for seeding purposes. Thus, online communities
such as Facebook might begin to offer information on
members’ social positions to third-party marketers or provide
the option to seed to a specific target group according
to sociometric measures. Specialized service providers
might adopt a similar idea to tailor their offerings with
respect to optimal seeding. Business models might reflect
social network information collected from communication
relationships or domain expertise in certain subject
domains, such that they target the most highly connected
or intermediary persons with specific marketing information
and thus maximize the success of viral marketing campaigns.
For example, the Procter & Gamble subsidiaries
Vocal point and Tremor already use social network information
to introduce new products, enjoying doubled sales
in some test locations.
Limitations and Directions for Further Research
Although we designed our experiments carefully, some of
their shortcomings might limit the validity of the results. In
Study 1, order effects may exist because the sets of participants
in the different experimental settings are not disjunctive.
We designed Study 2 to avoid such an overlap, but
the parallel timing in that experiment could lead to interrelationships
across the different seeding strategies. However,
as we noted previously, seeding strategies that lead
to faster diffusion might have performed better, which also
reflects reality well: In a world in which people become
REFERENCES
Anderson, Roy M. and Robert M. May (1991), Infectious Diseases
of Humans: Dynamics and Control. New York: Oxford
University Press.
Aral, Sinan, Lev Muchnik, and Arun Sundararajan (2009),
“Distinguishing Influence-Based Contagion from Homophily-
Driven Diffusion in Dynamic Networks,” Proceedings of the
National Academy of Sciences, 106 (51), 21544–49.
, , and (2011), “Engineering Social
Contagions: Optimal Network Seeding and Incentive Strate-
increasingly exposed to multiple viral campaigns competing
for their attention and participation, seeding strategies
that lead to faster diffusion may be advantageous for the
initiator of the particular viral marketing campaign.
The student sample and the artificial information content
of the experiments constitute additional limitations in
our experimental studies, though these shortcomings do not
systematically favor one strategy over another. It would be
useful to conduct similar experiments with different samples
and less artificial information, such as real-life viral
marketing campaigns.
In our real-life application, we could compare only highand
low-degree and random seeding strategies. We did
not differentiate referrals with regard to the profit of the
referred customers (for an analysis of the value of customers
acquired through referral programs, see Schmitt,
Skiera, and Van den Bulte 2011), but our robustness checks
show no significant regression-stage effects of additional
covariates on converted reach or influence domain. Our
results regarding the effects of degree centrality also remain
unchanged across the robustness checks.
Most current research focuses on individual choices and
treats the choices of partners (within the social network)
as exogenous; as we do, these studies assume that the network
remains fixed for the duration of the study and unaffected
by it. However, this strong assumption ignores the
likely effects of dynamics inherent to real-life social networks
(e.g., Steglich, Snijders, and Pearson 2010). Marketing
response models that incorporate these effects would
be a fruitful avenue for further research. Another extension
might incorporate information on the dyadic level, such as
tie strength, which could distinguish the success of referrals
from a specific customer. In addition, from a marketing
perspective, it would be useful to intensify research
regarding the effects of incentives, as we find their impact
to be quite significant in Study 1 (e.g., Schmitt, Skiera,
and van den Bulte 2011; Aral, Munchnik, and Sundararajan
2011).
Our combination of two experimental studies and an
ex post analysis of a real-world viral marketing campaign
provides a strong argument that hubs and bridges are key
to the diffusion of viral marketing campaigns. We cannot
confirm recent findings that question the exposed role of
hubs for the success of viral marketing campaigns, but our
analysis of the different determinants in Study 3 yields
additional explanations for why hubs make more attractive
seeding points. These findings should serve as inputs to
create more realistic computer simulations and analytical
models.
gies,” (accessed July 21, 2011), [available at http://ssrn.com/
abstract=1770982].
Arndt, Johan (1967), “Role of Product-Related Conversations
in the Diffusion of a New Product,” Journal of Marketing
Research, 4 (August), 291–95.
Bampo, Mauro, Michael T. Ewing, Dineli R. Mather, David
Stewart, and Mark Wallace (2008), “The Effects of the
Social Structure of Digital Networks on Viral Marketing
Performance,” Information Systems Research, 19 (3),
273–90.
Seeding Strategies for Viral Marketing / 69

Becker, Marshall H. (1970), “Sociometric Location and Innovativeness:
Reformulation and Extension of the Diffusion Model,”
American Sociological Review, 35 (2), 267–82.
Berger, Jonah and Katy Milkman (2011), “Social Transmission,
Emotion, and the Virality of Online Content,”
(accessed July 21, 2011), [available at http://marketing.wharton
.upenn.edu/documents/research/Virality.pdf].
and Eric Schwartz (2011), “What Do People Talk About?
Drivers of Immediate and Ongoing Word-of-Mouth,” Journal
of Marketing Research, forthcoming.
Biyalogorsky, Eyal, Eitan Gerstner, and Barak Libai (2001), “Customer
Referral Management: Optimal Reward Programs,” Marketing
Science, 20 (1), 82–95.
Burt, Ronald S. (1987), “Social Contagion and Innovation: Cohesion
Versus Structural Equivalence,”American Journal of Sociology,
92 (6), 1287–1335.
Coleman, James S., Elihu Katz, and Herbert Menzel (1966), Medical
Innovation: A Diffusion Study. Indianapolis: Bobbs-Merrill.
Costenbader, Elizabeth and Thomas W. Valente (2003), “The Stability
of Centrality Measures When Networks Are Sampled,”
Social Networks, 25 (4), 283–307.
De Bruyn, Arnaud and Gary L. Lilien (2008), “A Multi-Stage
Model of Word-of-Mouth Influence Through Viral Marketing,”
International Journal of Research in Marketing, 25 (3), 151–63.
Dodds, Peter S. and Duncan J. Watts (2004), “Universal Behavior
in a Generalized Model of Contagion,” Physical Review Letters,
92 (21), 1–5.
Festinger, Leon (1957), A Theory of Cognitive Dissonance.
Stanford, CA: Stanford University Press.
Galeotti, Andrea and Sanjeev Goyal (2009), “Influencing the
Influencers: A Theory of Strategic Diffusion,” RAND Journal
of Economics, 40 (3), 509–532.
Gladwell, Malcolm (2002), The Tipping Point: How Little Things
Can Make a Big Difference. Boston: Back Bay Books.
Godes, David and Dana Mayzlin (2009), “Firm-Created Word-of-
Mouth Communication: Evidence from a Field Study,” Marketing
Science, 28 (4), 721–39.
Goldenberg, Jacob, Sangman Han, Donald R. Lehmann, and Jae
Weon Hong (2009), “The Role of Hubs in the Adoption Process,”
Journal of Marketing, 73 (2), 1–13.
Granovetter, Mark S. (1973), “The Strength of Weak Ties,” American
Journal of Sociology, 78 (6), 1360–80.
(1985), “Economic Action and Social Structure: The
Problem of Embeddedness,” American Journal of Sociology, 91
(3), 481–510.
Hanaki, Nobuyuki, Alexander Peterhansl, Peter S. Dodds, and
Duncan J. Watts (2007), “Cooperation in Evolving Social Networks,”
Management Science, 53 (7), 1036–50.
Hann, Il-Horn, Kai-Lung Hui, Sang-Yong T. Lee, and Ivan P.L.
Png (2008), “Consumer Privacy and Marketing Avoidance:
A Static Model,” Management Science, 54 (6), 1094–1103.
Hill, Shawndra, Foster Provost, and Chris Volinsky (2006),
“Network-Based Marketing: Identifying Likely Adopters via
Consumer Networks,” Statistical Science, 21 (2), 256–76.
Hinz, Oliver and Martin Spann (2008), “The Impact of Information
Diffusion on Bidding Behavior in Secret Reserve Price
Auctions,” Information Systems Research, 19 (3), 351–68.
Iyengar, Raghuram, Christophe Van den Bulte, and Thomas W.
Valente (2011), “Opinion Leadership and Social Contagion in
New Product Diffusion,” Marketing Science, 30 (2), 195–212.
Iyengar, Sheena S. and Mark Lepper (2000), “When Choice
Is Demotivating: Can One Desire Too Much of a Good
Thing?” Journal of Personality and Social Psychology, 76 (6),
995–1006.
70 / Journal of Marketing, November 2011
Jain, Dipak, Vijay Mahajan, and Eitan Muller (1995), “An
Approach for Determining Optimal Product Sampling for the
Diffusion of a New Product,” Journal of Product Innovation
Management, 12 (2), 124–35.
Kalish, Shlomo, Vijay Mahajan, and Eitan Muller (1995), “Waterfall
and Sprinkler New-Product Strategies in Competitive
Global Markets,” International Journal of Research in Marketing,
12 (2), 105–119.
Kemper, John T. (1980), “On the Identification of Superspreaders
for Infectious Disease,” Mathematical Biosciences, 48 (1/2),
111–27.
Kiss, Christine and Martin Bichler (2008), “Identification of
Influencers—Measuring Influence in Customer Networks,”
Decision Support Systems, 46 (1), 233–53.
Lambert, Diane (1992), “Zero-Inflated Poisson Regression, with
an Application to Defects in Manufacturing,” Technometrics,
34 (1), 1–14.
Leskovec, Jure, Lada A. Adamic, and Bernardo A. Huberman
(2007), “The Dynamics of Viral Marketing,” ACM Transactions
on the Web, 1 (1), 228–37.
Libai, Barak, Eitan Muller, and Renana Peres (2005), “The Role
of Seeding in Multi-Market Entry,” International Journal of
Research in Marketing, 22 (4), 375–93.
Lin, Nan (1976), Foundations of Social Research. New York:
McGraw-Hill.
Mullahy, John (1986), “Specification and Testing of Some Modified
Count Data Models,” Journal of Econometrics, 33 (3),
341–65.
Porter, Constance E. and Naveen Donthu (2008), “Cultivating
Trust and Harvesting Value in Virtual Communities,” Management
Science, 54 (1), 113–28.
Porter, Lance and Guy J. Golan (2006), “From Subservient Chickens
to Brawny Men: A Comparison of Viral Advertising to
Television Advertising,” Journal of Interactive Advertising, 6
(2), 30–38.
Rayport, Jeffrey (1996), “The Virus of Marketing,” Fast Company
(accessed March 24, 2011), [available at http://www
.fastcompany.com/magazine/06/virus.html].
Schmitt, Philipp, Bernd Skiera, and Christophe Van den Bulte
(2011), “Referral Programs and Customer Value,” Journal of
Marketing, 75 (January), 46–59.
Scott, John (2000), Social Network Analysis: A Handbook, 2d ed.
London: Sage Publications.
Simmel, Georg (1950), The Sociology of Georg Simmel, compiled
and translated by Kurt Wolff. Glencoe, IL: The Free Press.
Steglich, Christian, Tom A.B. Snijders, and Michael Pearson
(2010), “Dynamic Networks and Behavior: Separating
Selection from Influence,” Sociological Methodology, 40 (1),
329–93.
Sundararajan, Arun (2006), “Network Seeding,” Extended
Abstract for WISE 2006, New York University Stern School of
Business.
Thompson, Clive (2008), “Is the Tipping Point Toast?” Fast
Company, 122, 74–105, (accessed March 24, 2011), [available
at http://www.fastcompany.com/magazine/122/is-the-tippingpoint-toast.html].
Toubia, Olivier, Andrew T. Stephen, and Aliza Freud (2010),
“Identifying Active Members in Viral Marketing Campaigns,”
(accessed July 21, 2011), [available at http://sites.google.com/a/
andrewstephen.net/andrew/researc/papers/Toubia_Stephen_Freud_
viralmarketing_nov2010.pdf].
Trusov, Michael, Anand Bodapati, and Randolph E. Bucklin
(2010), “Determining Influential Users in Internet Social Networks,”
Journal of Marketing Research, 47 (August), 643–58.

Van den Bulte, Christophe (2010), “Opportunities and Challenges
in Studying Consumer Networks,” in The Connected Customer,
S. Wuyts, M.G. Dekimpe, E. Gijsbrechts, and R. Pieters, eds.
London: Routledge, 7–35.
and Yogesh V. Joshi (2007), “New Product Diffusion
with Influentials and Imitators,” Marketing Science, 26 (3),
400–421.
and Gary L. Lilien (2001), “Medical Innovation Revisited:
Social Contagion Versus Marketing Effort,” American
Journal of Sociology, 106 (5), 1409–1435.
and Stefan Wuyts (2007), Social Networks and Marketing.
Cambridge, MA: MSI Relevant Knowledge Series.
Van der Lans, Ralf, Gerrit van Bruggen, Jehoshua Eliashberg,
and Berend Wierenga (2010), “A Viral Branching Model for
Predicting the Spread of Electronic Word-of-Mouth,” Marketing
Science, 29 (2), 348–65.
Watts, Duncan J. (2004), “The ‘New’ Science of Networks,”
Annual Review of Sociology, 30, 243–70.
and Peter S. Dodds (2007), “Influentials, Networks, and
Public Opinion Formation,” Journal of Consumer Research, 34
(4), 441–58.
and Jonah Peretti (2007), “Viral Marketing for the Real
World,” Harvard Business Review, 85 (5), 22–23.
Seeding Strategies for Viral Marketing / 71

© 2011, American Marketing Association
ISSN: 0022-2429 (print), 1547-7185 (electronic)
Bernd Skiera, Manuel Bermes, & Lutz Horn
Customer Equity Sustainability
Ratio: A New Metric for Assessing a
Firm’s Future Orientation
Securitization is a remarkable financial instrument; it enables securitizers to increase their short-term profits at the
expense of the long-term value of their customer base. This ability might be tempting for firms, especially because
it does not need to be disclosed transparently to stakeholders. The authors show how their newly developed
customer equity sustainability ratio (CESR) complements customer equity reporting and creates more transparency
about the consequences of securitization for future earnings and the riskiness of the underlying business model.
Their model compares the future value of an existing customer base with current earnings. In an empirical study
of 38 banks in ten countries, the authors demonstrate the limited transparency of long-term value creation in
financial statements. Next, they outline the adequacy of CESR for creating more transparency in empirical cases
of Countrywide Financial Corporation and nine firms from nonbanking industries. They recommend that marketing
should play a prominent role in providing stakeholders with substantial information about the long-term value of the
customer base.
Keywords: customer equity, financial reporting, sustainability, securitization, financial crisis, customer equity
sustainability ratio
The financial crisis currently disrupting the economic
system and banking worldwide is often attributed to
securitization (Ryan 2008), an instrument that pools
assets with an inherent stream of future earnings to support
debt instruments, called asset-backed securities (ABSs), for
sale to investors (Greenbaum and Thakor 1987). 1 Banks
commonly use securitization to manage their portfolio risk
and funding position by transferring loans and the credit
risk of their loan portfolios to other investors (Santomero
and Babbel 1997). In return, the banks do not hold the loans
until maturity in their own books but instead receive earnings
from them directly at their present value (PV).
Unfortunately, as the global financial crisis makes clear,
the inherent risk of ABSs is difficult to calculate because of
their frequent repackaging; most market participants
grossly underestimated it (Coval, Jurek, and Stafford 2009).
1We use ABS to refer to the whole group of structured products,
which can be classified into asset-backed commercial papers,
mortgage-backed securities, and collateralized debt obligations.
Bernd Skiera is Chair of Electronic Commerce, Department of Marketing,
Faculty of Business and Economics (e-mail: skiera@skiera.de), Manuel
Bermes is a doctoral student, E-Finance Lab, House of Finance (e-mail:
bermes@ wiwi.uni-frankfurt.de), and Lutz Horn is a doctoral student,
Retail Banking Competence Center, House of Finance (e-mail: lutzhorn@
gmx.de), Goethe University Frankfurt. The authors thank Andreas Hackethal,
Clemens Jochum, Wolfgang König, Jan Pieter Krahnen, Christian
Leuz, Steffen Meyer, Anita Mosch, Philipp Schmitt, Christian Schulze,
Thorsten Wiesel, seminar participants at the University of Technology
Sydney and Maastricht University, and three anonymous JM reviewers for
their valuable comments on earlier drafts of the article.
118
As a consequence, prices for securities dropped dramatically,
banks were forced to realize heavy write-downs on
their asset bases, and they suffered huge losses and in some
cases bankruptcies (for more detailed descriptions, see
Demyanyk and Van Hemert 2009; Franke and Krahnen
2008).
These risks are well known, if not fully solved; however,
additional problems result from moves toward short-term
profit realizations that are inherent to the securitization of
loans and come at the expense of long-term value creation.
The transformation of periodic loan payments into one
down payment enables a bank to realize the PV of earnings
immediately rather than doing so over the lifetime of the
loan.
Securitization is not limited to the banking industry.
Airlines and sport clubs can securitize earnings from ticket
sales that will occur over the next few years to realize the
earnings in the present rather than spreading them over
time. Although it remains particularly prevalent in banking,
securitization enjoys great popularity in other industries
such as telecommunications, utilities, and the music business;
it is even used by national governments and other public
institutions (Brinkworth 2004; Downey 1999; Ketkar
and Ratha 2008).
Supported by accounting rules, managers do not need to
make the consequences of securitization for long-term
value creation transparent. They have incentives to adjust
their firm’s earnings stream through securitization and
reach their own goals, such as greater personal wealth
(Dechow and Shakespeare 2009). However, the problems
that arise from such adjustments in customer management
Journal of Marketing
Vol. 75 (May 2011), 118 –131

strategies and earnings streams continue to be largely
ignored in current discussions of the financial crisis (e.g.,
Coval, Jurek, and Stafford 2009; Franke and Krahnen
2008). Not surprisingly, metrics that allow for easily detecting
shifts in earning streams are scarce.
This article outlines the problems associated with this
shift in earnings and proposes customer equity reporting
(CER) along with a new ratio, the customer equity sustainability
ratio (CESR), as means to increase the level of transparency
in financial statements. We emphasize the importance
of reporting forward-looking marketing metrics in
financial statements, thereby extending the role of marketing:
It should supply firm stakeholders with substantial
information about the long-term value of the current customer
base (see also Joshi and Hanssens 2010; Tuli,
Bharadwaj, and Kohli 2010). In line with Wiesel, Skiera, and
Villanueva’s (2008) recent proposal, which postulates that
CER can complement financial statements, we argue that
greater transparency, achieved by reporting more forwardlooking
marketing metrics, might have reduced the devastating
consequences of the current financial crisis for banks
and might lead to a suitable use of securitization in industries
outside banking.
We begin by describing securitization, its advantages,
and its disadvantages. Next, we show how CER and CESR
can capture some of the effects of securitization previously
neglected by first focusing on banking business and its
extensive use of securitization. Then, we empirically analyze
the transparency of long-term value creation in the
financial statements of 38 banks in ten countries and apply
our reporting technique, including CESR, to the former
U.S. market leader in mortgage lending and origination,
Countrywide Financial Corporation. We also explore industries
outside banking and show, for nine securitizations of
firms and institutions in diverse industries, how CESR can
detect differences in the extent of shifts in earnings across
time. We conclude with a discussion of the results.
Securitization
Basic Idea of Securitization
Securitization is best known as the pooling and repackaging
of a group of assets (e.g., loans) and the subsequent sale of
tranches, which are the new divisions of the group of assets
classified by asset quality, of this pool to new investors
(Ryan 2008; Santomero and Babbel 1997). For example,
banks commonly sell a variety of financial products, from
mortgages to student loans to credit cards to leasing claims,
which provide the underlying loans for their securitization
(Ketkar and Ratha 2008; Santomero and Babbel 1997).
They transfer the loan assets for regulatory and accounting
purposes into a separate business unit, called a special purpose
vehicle (SPV), and group them into tranches (Greenbaum
and Thakor 1987). Each tranche receives a rating
from a rating agency, from senior (AAA to A rating) to
equity (no rating/residuum). Then, these tranches can be
priced and sold to new investors (Coval, Jurek, and Stafford
2009; Luo, Tang, and Wang 2009), who receive all earnings
from the tranche they own, which means they also confront
any risks arising from the underlying loans (see Franke and
Krahnen 2008).
Not just banks and other financial institutions but virtually
all firms and even state and national governments can
and do use securitization. The main prerequisite to sell
claims to investors is an underlying asset that currently is
generating earnings or will generate them in the future (for
further details, see Kendall 1996; Kothari 2006). A wide
variety of assets from different industries can be used for
securitization, such as future revenues from cellular phone
contracts or electricity consumption, ticket sales from
future soccer games, airline ticket sales, tax revenue receivables,
and royalties of intellectual properties (Brinkworth
2004; Downey 1999; Ketkar and Ratha 2008). Prominent
recent securitization transactions include musicians such as
David Bowie, who issued the first music royalties future
receivables securitization (Burke Sylva 1999); soccer teams
such as Leeds United (United Kingdom), Tottenham Hotspurs
(United Kingdom), and Schalke 04 (Germany), which
have used securitization to fund their entire businesses
(Brinkworth 2004); and U.S. football teams such as the
Denver Broncos, which financed a new stadium by using
securitization “backed by about 4,000 stadium-related contracts,
such as luxury box seats, club seats, a portion of concession
fees and other cash flows” (Gregory 2002, p. 6).
Some banks helped Greece improve its short-term financial
situation and hide its debt level by providing cash in return
for Greece’s government payments in the future, meaning
that Greece has traded away its revenue from the rights of
airport fees and lotteries (Story, Thomas, and Schwartz
2010).
Numerical Example to Outline the Effects of
Securitization
Despite the widespread popularity of securitization, we
concentrate on banks because loans provide a clear illustration
of its effects. We depict a loan example to show the
effects of securitization on the earnings stream (for comparable
arguments, see Fabozzi, Davis, and Choudry 2006). In
this example, we use a bank that issues one five-year loan
volume of $100,000 to customers at the beginning of each
year, with an annual interest rate of 5% paid at the end of
each year. The customers repay the loan linearly over its
lifetime, so the effective loan volume reduces to $80,000 in
Year 2, $60,000 in Year 3, and so on. During the lifetime of
the loan, the bank receives annual interest income of $5,000
in Year 1, $4,000 in Year 2, $3,000 in Year 3, and so on. We
assume an interest rate of 3.5% for the financial debt to refinance
the loan issuance, resulting in interest expenses of
$3,500 in Year 1, $2,800 in Year 2, $2,100 in Year 3, and so
on. We also deduct loan loss provisions, equal to .5% of the
loan volume, to account for potential default of these customers.
So the bank realizes a margin of 1% (5% – 3.5% –
.5%) of the loan volume. Although specific terms of loans,
such as prepayments, deductions, servicing, and other costs,
might increase the complexity of this illustration in practice,
we avoid additional complexity for ease of exposition.
We also assume that annual redemptions repay the corresponding
financial debt.
Customer Equity Sustainability Ratio (CESR) / 119

Table 1 includes the results for a non-securitizing bank,
and Table 2 depicts the outcome for a securitizing bank. To
show the time-shifting effects of securitization on the earnings
of the bank in the second part of our example, we
assume that from Year 8 onward, the bank does not issue
any new loans and the existing loans expire. Until Year 7,
expiring loans are permanently renewed by new issuances,
and for ease of exposition we assume that the bank has
reached a steady state, so that its total loan volume remains
stable. The earnings of $3,000 in Year 1 comprise $1,000
from loans issued in Year 1, $800 from loans of the prior
year (Year 0), $600 from the next to last year (Year –1),
$400 from the remaining loans of Year –2, and $200 from
those of Year –3. In Year 2, the bank issues new loans generating
earnings of $1,000, earnings of loans from Year –2
to Year 1 decline by $200 each, and loans from Year –3
expire.
The bank can either distribute the obtained earnings as
dividends to its shareholders (distribution case) or keep and
reinvest them (reinvestment case). In the distribution case,
the initial equity of $30,000 remains stable over time.
Because earnings also are constant, the bank realizes a
return on equity (ROE) of 10% each year. In the reinvestment
case, we assume an annual return rate of the reinvestment
of 10%, equivalent to the discount rate. The bank reinvests
its earnings each year for a one-year period,
repeated until the end of the loan contract, and does not distribute
any earnings, so the earnings from previous years are
permanently reinvested. For example, in Year 3, the bank
can reinvest the sum ($6,300) of the earnings of the previous
year ($3,000); the earnings of this year ($3,000), which
are labeled as new reinvestment; as well as the earnings
from the reinvestment in Year 2 (10% ¥ $3,000 = $300). Its
equity increases from $33,000 ($30,000 + $3,000) to
$36,300 because of the new reinvestment ($3,000) and the
earnings from reinvestment ($300). Thus, the bank’s starting
equity of $30,000 grows every year from the retained
earnings from loans and reinvestment. The ROE still
amounts to 10% because the relative increase in earnings
equals the rise in equity.
We also consider a securitizing bank (Table 2) that
decides, in Year 3, to transfer the whole loan volume and
related earnings to new investors. 2 We assume the bank
does so at fair market value, or the PV of earnings at the
end of Year 3. Thus, the bank receives the earnings from the
loan securitization of $6,397. This amount reflects the PV
of new loans issued in Year 3, or $2,660, plus the PV of the
remaining loans of Years –1 to 2, equal to $3,737 ([$800 +
$600 + $400 + $200] + [$600 + $400 + $200]/1.1 + [$400 +
$200]/1.1 2 + $200/1.1 3), that were securitized in Year 3 as
well. However, in considering the PV of future years, the
securitizing bank realizes only $2,660, equal to the PV of
the new loans acquired and securitized each year ($1,000 +
$800/1.1 + $600/1.1 2 + $400/1.1 3 + $200/1.1 4).
2We assume a direct origination and distribution of loans to new
investors. However, an indirect sale through a bank-owned SPV,
which is a separate legal entity, is common as a means to obtain
various accounting reliefs, though it has no influence on the results
in our example.
120 / Journal of Marketing, May 2011
We again distinguish between distributing and reinvesting
the bank’s earnings. In the distribution case, equity
remains constant at $30,000. The securitizing bank reaches
an ROE of 10% in Years 1 and 2, which increases to 21.3%
in Year 3 because of the shift to short-term profit realization
through securitization. However, the increase in ROE in
Year 3 comes at the expense of a lower ROE from Year 4
onward; it falls back to a steady-state level of 8.9%.
In Year 3, with earnings of $6,397, the securitizing bank
surpasses the earnings of the non-securitizing bank by more
than 100%. After Year 3, the securitizing bank earns only
$2,660, whereas the non-securitizing bank continues to earn
$3,000 per year, with more future earnings already contracted.
Comparing ROE leads to a similar result: The nonsecuritizing
bank shows a stable ROE of 10%, so the securitizing
bank outperforms it in Year 3 with a ROE of 21.3%
but then realizes only 8.9% from Year 4 onward.
In the reinvestment case, equity increases by the earnings
realized from loans, securitization, and reinvestment.
When securitizing in Year 3, ROE grows to 19.4% because
of the shift from long-term value creation to short-term
profit realization. However, one year later, ROE falls back
to a lower level of 9.2% and rises only slowly to 9.4% in
Year 7. Compared with the distribution case, the differences
in ROE result from the reinvestment of earnings and the
compound interest effects of this reinvestment.
In this example, both the seller (i.e., securitizer) and the
buyer of the loans have equal assessments of the underlying
(risky) earnings and use the same discount rate, and the
seller does not realize a higher return rate than the discount
rate through reinvestment. Thus, securitization neither creates
nor destroys value but again simply shifts the realization
of value over time (Kothari 2006). Differences in value
accrue only when the seller and buyer differ in (1) their
assessments of the underlying (risky) earnings or (2) their
discount rates, because (1) and (2) yield to different PVs, or
(3) the seller can use the earnings brought forward for reinvestment
to realize a return rate greater than the discount
rate. Still, securitizers increase their short-term profits at the
expense of long-term value creation.
Treatment of Securitization by Reporting
Standards
Reporting standards provide few opportunities to detect the
earnings shift from long-term value creation toward shortterm
profit generation because firms are not required to
report quantitative outlooks of their future earnings. The
U.S. Financial Accounting Standards Board (FASB) introduced
FAS 125 and FAS 140, which obligate firms to publish
more detailed information about their securitization
transactions and valuation principles (FASB 1996, 2000;
Ryan 2008); comparable International Financial Reporting
Standards call for less information about securitization in
IAS 32 and 39 (McCreevy 2008). Therefore, current regulations
require banks and firms, respectively, to publish little
information about their securitization activities; they might
disclose securitization volume in the amendments of their
financial statements, but it is not mandatory. The earnings
derived from securitization may be registered as noninterest
or interest income, which do not need to be separated from

TAblE 1
Numerical Example of loans of a Non-Securitizing bank
Non-Securitizing bank Year 1 Year 2 Year 3 Year 4 Year 5 Year 6 Year 7 Year 8 Year 9 Year 10 Year 11
Noninterest income $0 $0 $0 $0 $0 $0 $0 $0 $0 $0 $0
Net interest income from
Loans issued in Year –3 $200
Loans issued in Year –2 $400 $200
Loans issued in Year –1 $600 $400 $200
Loans issued in Year 0 $800 $600 $400 $200
Loans issued in Year 1 $1,000 $800 $600 $400 $200
Loans issued in Year 2 $1,000 $800 $600 $400 $200
Loans issued in Year 3 $1,000 $800 $600 $400 $200
Loans issued in Year 4 $1,000 $800 $600 $400 $200
Loans issued in Year 5 $1,000 $800 $600 $400 $200
Loans issued in Year 6 $1,000 $800 $600 $400 $200
Loans issued in Year 7 $1,000 $800 $600 $400 $200
Net interest income $3,000 $3,000 $3,000 $3,000 $3,000 $3,000 $3,000 $2,000 $1,200 $600 $200
Earnings $3,000 $3,000 $3,000 $3,000 $3,000 $3,000 $3,000 $2,000 $1,200 $600 $200
Distribution Case
Customer equity (as of 12/31) $6,397 $6,397 $6,397 $6,397 $6,397 $6,397 $6,397
Customer equity sustainability
ratio (as of 12/31) .531 .531 .531 .531 .531 .531 .531
Loan volume to customers
(at beginning of year) $300,000 $300,000 $300,000 $300,000 $300,000 $300,000 $300,000
Equity $30,000 $30,000 $30,000 $30,000 $30,000 $30,000 $30,000
Return on equity 10.0% 10.0% 10.0% 10.0% 10.0% 10.0% 10.0%
Reinvestment Case
Earnings from reinvestment $0 $300 $630 $993 $1,392 $1,832 $2,315 $2,846 $3,331 $3,784 $4,222
New reinvestment –$3,000 –$3,300 –$3,630 –$3,993 –$4,392 –$4,832 –$5,315 –$4,846 –$4,531 –$4,384 $42,222
Total result from reinvestment –$3,000 –$3,000 –$3,000 –$3,000 –$3,000 –$3,000 –$3,000 –$2,000 –$1,200 –$600 $46,445
Customer equity (as of 12/31) $6,670 $7,270 $7,930 $8,656 $9,455 $10,333 $11,299
Customer equity sustainability
ratio (as of 12/31) .550 .546 .542 .539 .535 .532 .530
Loan volume to customers
(at beginning of year) $300,000 $300,000 $300,000 $300,000 $300,000 $300,000 $300,000
Equity $30,000 $33,000 $36,300 $39,930 $43,923 $48,315 $53,147
Return on equity 10.0% 10.0% 10.0% 10.0% 10.0% 10.0% 10.0%
Customer Equity Sustainability Ratio (CESR) / 121

TAblE 2
Numerical Example of loans of a Securitizing bank
Securitizing bank Year 1 Year 2 Year 3 Year 4 Year 5 Year 6 Year 7 Year 8 Year 9 Year 10 Year 11
Noninterest income $0 $0 $6,397 $2,660 $2,660 $2,660 $2,660 $0 $0 $0 $0
Net interest income from
Loans issued in Year –3 $200
Loans issued in Year –2 $400 $200
Loans issued in Year –1 $600 $400
Loans issued in Year 0 $800 $600
Loans issued in Year 1 $1,000 $800
Loans issued in Year 2 $1,000
Loans issued in Year 3
Loans issued in Year 4
Loans issued in Year 5
Loans issued in Year 6
Loans issued in Year 7
Net interest income $3,000 $3,000 $0 $0 $0 $0 $0 $0 $0 $0 $0
Earnings $3,000 $3,000 $6,397 $2,660 $2,660 $2,660 $2,660 $0 $0 $0 $0
122 / Journal of Marketing, May 2011
Distribution Case
Customer equity (as of 12/31) $6,397 $6,397 $6,397 $2,660 $2,660 $2,660 $2,660
Customer equity sustainability
ratio (as of 12/31) .531 .531 .000 .000 .000 .000 .000
Loan volume to customers
(at beginning of year) $300,000 $300,000 $0 $0 $0 $0 $0
Equity $30,000 $30,000 $30,000 $30,000 $30,000 $30,000 $30,000
Return on equity 10.0% 10.0% 21.3% 8.9% 8.9% 8.9% 8.9%
Reinvestment Case
Earnings from reinvestment $0 $300 $630 $1,333 $1,732 $2,171 $2,654 $3,186 $3,504 $3,855 $4,240
New reinvestment –$3,000 –$3,300 –$7,027 –$3,993 –$4,392 –$4,832 –$5,315 –$3,186 –$3,504 –$3,855 $42,404
Total result from reinvestment –$3,000 –$3,000 –$6,397 –$2,660 –$2,660 –$2,660 –$2,660 $0 $0 $0 $46,645
Customer equity (as of 12/31) $6,670 $7,270 $8,239 $5,568 $6,366 $7,245 $8,211
Customer equity sustainability
ratio (as of 12/31) .550 .546 .147 .283 .310 .333 .353
Loan volume to customers
(at beginning of year) $300,000 $300,000 $0 $0 $0 $0 $0
Equity $30,000 $33,000 $36,300 $43,327 $47,320 $51,713 $56,544
Return on equity 10.0% 10.0% 19.4% 9.2% 9.3% 9.3% 9.4%

other noninterest or interest income positions. Therefore, the
shift from creating long-term value to realizing short-term
profits can be hidden well (Dechow and Shakespeare 2009).
Advantages and Disadvantages of Securitization
Securitization offers several major advantages to firms (for an
extended discussion, see Franke and Krahnen 2008). First,
securitization helps banks manage the various risks of underlying
loan portfolios. They can construct well-diversified
investments and lower or even avoid risks by transferring
them to other investors. In risk transfers, banks might also
be able to realize profits if the sellers and buyers of loans
maintain different evaluations of the fair value. Second,
firms can attain stronger capital, funding, and liquidity positions
through securitization because it enables them to turn
illiquid assets, such as mortgages, into more liquid assets.
Third, securitization supports banks in their efforts to fulfill
regulatory requirements such as Basel II, because they no
longer need to hold equity for securitized loans (Calem and
LaCour-Little 2004). Fourth, the lower equity requirements
for securitized loans increase ROE.
However, securitization also produces two major dis -
advantages. First, the frequent repackaging and splitting of
ABSs into tranches make it difficult to determine the appropriate
value of those securities. This topic appears in extensive
discussions elsewhere, so we do not detail it here (see
Coval, Jurek, and Stafford 2009; Luo, Tang, and Wang
2009). Second, the short- and long-term effects of securitization
are not equally transparent. The result of the shortterm
effect, usually an increase in earnings, is clearly
described in financial reports, but the corresponding
decrease in long-term value is less obvious and, as we show
in our empirical study, is not well reported by the vast
majority of firms. The low transparency causes various
problems, including those related to management compensation
when variable payments are linked to short-term
profits (Dechow, Myers, and Shakespeare 2010; Dechow
and Shakespeare 2009).
Customer Equity Reporting and
Securitization
Customer Equity Reporting
Customer equity reporting creates the required transparency
by reporting the value of the customer base (i.e., customer
equity) and its development over time. Blattberg and
Deighton (1996) define “customer equity” (CE) as the customer
lifetime values (CLV) of the firm’s current customers j:
( 1)
CE = CLVj. j = 1
If the total lifespan of a customer j is T j, CLV is simply the
PV of customer j’s earnings (Earn j,t) over time t (discounted
at rate i):
J
∑
Earn j, t
( 2)
CLVj
= .
t ( 1 + i)
T j
∑
t = 0
Blattberg and Deighton (1996), Gupta, Lehmann, and
Stuart (2004), Rust, Lemon, and Zeithaml (2004), and Rust,
Zeithaml, and Lemon (2000) lay a foundation for a deeper
understanding and wider acceptance of the CE approach in
science and practice (see also Srinivasan and Hanssens
2009). As Wiesel, Skiera, and Villanueva (2008) indicate,
CER extends this field of application by connecting CE to
financial reporting. We outline an opportunity to define CE
more narrowly than Wiesel, Skiera, and Villanueva do for
firms in industries with high shares of contracted business,
such as banking, telecommunications, and electricity. In
this contractual setting, in general, customer relationships
span several years, and future earnings can be projected
reliably because they are contractually assured. We consider
only contracted future earnings for our calculation of CLV
and CE (unless noted otherwise).
Customer Equity Sustainability Ratio
We propose the CESR as a new ratio to quantify the intensity
of long-term value creation and establish a connection
between a firm’s financial statements and forward-looking
CER. The CESR contrasts the future value of an existing
customer base with current earnings, such that for an individual
customer j, the future value is the PV of all earnings
after the current year t = 0. Defining CESRj as the ratio of
the PV of all future earnings to the corresponding CLV and
rearranging leads to the following:
( 3)
CESR
j
=
Tj
∑
t = 1
Tj
∑
t = 0
Earn j, t
t ( 1 + i)
CLV − Earn
=
Earn
CLVj
j, t
t
( 1 + i)
Earn j
= 1 − .
CLV
j j,
0 , 0
Therefore, the CESR for all current customers is
( 4)
CESR =
J
Tj
∑∑
j = 1 t = 1
J
Tj
∑∑
j = 1 t = 0
Earn
j, t
t
( 1 + i)
=
Earn
j, t
t
( 1 + i)
= 1 −
J
∑
j = 1
J
∑
j = 1
j
j = 1 j = 1
J
j = 1
If we define current earnings as Earn 0 = S J j = 1Earn j,0, we
can rewrite Equation 4 as follows:
In the case of nonnegative earnings and positive CE, the
CESR falls between 0 and 1. A higher CESR indicates that
the future value of the current customer base is high. A
CESR of 0 implies that all earnings occur in the current
year, as in our securitization example, whereas a CESR of 1
J
∑ ∑
Earn
CLV
CLV − Earn
j,
0
∑
Earn0
( 5) CESR = 1 −
.
CE
j,
0
Customer Equity Sustainability Ratio (CESR) / 123
j
.
J
CLV
j
j

presumes that all earnings will be realized in the future,
with no current earnings.
Numerical CER and CESR Example
To explain the relevance of CER and CESR in more detail,
we adapt the concept to our numerical example in Tables 1
and 2. The non-securitizing bank keeps issuing loans at an
annual volume of $100,000 and realizes earnings of $3,000
from the loans each year. In the distribution case, annual CE
is $6,397, or the PV of all current ($3,000) and future,
already contracted earnings ($3,397 = $2,000/1.1 + $1,200/
1.12 + $600/1.13 + $200/1.14). The CESR is .531 (1 –
$3,000/ $6,397). The bank receives 47% ($3,000/$6,397) of
its contracted earnings in the current year and 53% ($3,397/
$6,397) in the future. In the securitization case, the bank
realizes all future earnings (from Year 3 onward) immediately,
so CE from Year 4 onward decreases to $2,660. Thus,
the CESR drops from .531 to 0 (1 – $2,660/$2,660).
In the reinvestment case, the CE of the non-securitizing
bank is $6,670 in Year 1, which comprises the PV of all current
and future loan-related earnings of $6,397 and the PV
of the current earnings in Year 1 reinvested for one year at a
10% return rate ($273 = [$3,000 ¥ .1]/1.1). The reinvestment
gains raise CESR to .550 compared with the distribution
case. In Year 7, a CESR of .530 derives from the current
earnings from loans ($3,000) and reinvestment
($2,315) and from the PV of future earnings from loans
($3,397) and reinvestment ($2,587), or .530 = 1 – ($3,000 +
$2,315)/ ($3,000 + $2,315 + $3,397 + $2,587).
Again, the values of the securitizing bank differ from
those of the non-securitizing bank. When securitizing in
Year 3, CESR drops to .147 because all future earnings
from securitization ($6,397) are realized immediately along
with the reinvestment gains ($630). Only the reinvestment
generates future earnings ($1,212 = $1,333/1.1). Until Year
7, CESR increases to .353 because gains from reinvestment
rise but earnings from securitization remain constant. Note
that CESR also reflects the investment horizon of the reinvestments,
which is one year in our example. If this
investment horizon was five years, it would correspond to
the investment period of the (five-year) loans; therefore, the
time horizon of both investments would be the same.
The major advantage of CESR is that it creates more
transparency and highlights, in a single metric, whether
firms increase their current earnings at the expense of their
long-term value. Comparing the distribution with the reinvestment
case for the non-securitizing and securitizing
bank reveals an additional strength of CESR: If the bank
chooses to keep and reinvest its earnings from loans and
securitization, the bank concurrently fosters its future earnings
position, and CESR increases.
Empirical Studies
Aim
First, we focus in our empirical studies on banking because
of its extensive use of securitization. We determine whether
banks provide sufficient information about the use of securitization
to make their economic impact transparent. Then,
124 / Journal of Marketing, May 2011
we apply our reporting technique to the former market
leader in mortgage lending and origination in the United
States, Countrywide Financial Corporation, which made
heavy use of securitization, and we show that CER and
CESR could have provided much more transparency. Next,
for nine securitizations of firms and institutions in diverse
industries, we show that CESR can detect differences in the
extent of the shift in earnings across time.
Analysis of Securitization Transparency in
Financial Statements
We analyze 38 banks in the most important banking markets
in the United States and Europe to determine how
much information they provide about securitization in their
financial statements (i.e., annual and quarterly reports). The
sample includes two market leaders (measured by market
capitalization) from ten countries and some randomly
selected smaller banks and public institutions. These banks
offer wide service portfolios to customers to originate sufficient
loan volumes for securitization. As we show in Table
3, nearly all banks are involved in securitization, but especially
in the United States, few report the earnings that they
realize separately. Instead, most banks follow the opportunity
provided by current reporting standards and sum their
“real interest earnings” plus “securitization earnings” as
total interest income. This summary makes it difficult, if not
impossible, to evaluate the extent of the shift toward shortterm
profit realization inherent in securitization. In turn, it
remains unclear what share of profit relates to the ongoing
banking business and which part derives from the one-time
effects of securitization.
Description of Countrywide Financial Corporation
Of this data set, we choose Countrywide for our detailed
analysis because it fulfills our information requirements
about securitization and other interest-bearing assets. In
addition, Countrywide was the U.S. market leader in mortgage
lending and origination between 2004 and 2007
(Countrywide Financial Corp. 2008). Because of its heavy
engagement in subprime mortgage–backed security transactions
and the debt market turbulence of the financial crisis,
Countrywide needed a rescue from Bank of America in
February 2008 (Countrywide Financial Corp. 2008) and no
longer publishes its own financial statements.
Analysis of Countrywide Financial Corporation
To apply CER and CESR to Countrywide, we make a few
assumptions and calculations to project future earnings. The
contracted loans generate net interest income and net loan
servicing fees. We assume that loans are amortized linearly
and that contracted earnings are realized over the loans’
lifetimes. We also assume no further earnings from insurance
premiums and other business after 2007 and adopt the
distribution and reinvestment policy of Countrywide.
In Table 4, we provide the resulting CE and CESR for
1998–2007 for the actual business of Countrywide, which
includes securitization (labeled “Securitization Case”). The
increase in CE from $2.567 billion in 1998 to $20.151 billion
in 2006 may be affiliated with higher loan volumes origi-

nated and sold, longer mortgage lifetimes within the servicing
portfolio, or higher interest margins from the remaining
business. Regardless of the cause, Countrywide created additional
value for the future in 2006. The subsequent intense
drop in earnings appeared in the 2007 CE of $9.647 billion,
when the decreasing U.S. mortgage business reduced both
total earnings and CE by approximately 50% in response to
lower income levels and higher loan loss provisions.
In contrast to the immense fluctuations in CE from 1998
to 2007, CESR remained relatively stable, with values of
.291 in 1998 and .372 in 2007. In other words, the earnings
structure was not affected by the economic downturn, even
when the business model suffered serious harm. Countrywide,
with a CESR of .372 in 2007, still realized 62.8% of
the earnings that it originated in the same year but only
TAblE 3
Information About Securitization from 38 banks
General Securitization
Interest Income
from Other
Securitization Interest-bearing
bank
largest National banks
Country Information Volume Earnings Assets
Erste Bank Austria √ √
Raiffeisen International Bank Austria √ √
BNP Paribas France √ √
Société Générale France √ √
Deutsche Bank Germany √ √
Commerzbank Germany √ √
Allied Irish Bank Ireland √ √
Bank of Ireland Ireland √ √
Unicredit Italy √
Intesa San Paolo Italy √ √
ING Group Netherlands √ √
Fortis Netherlands √
Santander Spain √ √
BBVA Spain √ √
Credit Suisse Switzerland √ √ √ √*
UBS Switzerland √
HSBC United Kingdom √ √
Standard Chartered United Kingdom √ √
JPMorgan Chase & Co. United States √ √ √ √*
Wells Fargo
Random Selection banks
United States √ √ √ √*
Bank Austria Austria √
Comdirect Germany N.A.
DAB Bank Germany N.A.
Deutsche Postbank Germany √ √
Hamburger Sparkasse Germany N.A.
KfW Germany √ √
Landesbank Hessen-Thüringen Germany √ √
ABN Amro Netherlands √ √
Rabobank Netherlands √
Zürcher Kantonalbank Switzerland N.A.
Barclays United Kingdom √ √
HBOS United Kingdom √ √
Lloyds TSB United Kingdom √ √
Northern Rock United Kingdom √ √ (√)
RBS United Kingdom √ √
Bank of America United States √ √ √ √*
Countrywide United States √ √ √ √*
Wachovia United States √ √
Notes: √ = Information is clearly available, √* = Information can be derived directly from the data, (√) = Information can be assumed from data,
and N.A. = not available.
37.2% in subsequent years. Because the average mortgage
lifetime is 6.8 years, we believe it is fair to assert that Countrywide
aimed to realize short-term profits.
Because we want to demonstrate that CER and CESR
could have depicted the realization of short-term profits at
the expense of long-term value creation at Countrywide, we
ran a counterfactual analysis (labeled “Non-securitization
Case” in Table 4) in which we examined what would have
happened had Countrywide not securitized at all. In this
analysis, CE and CESR would have increased strongly. The
most extreme difference would have occurred in 2002, when
79.0% of Countrywide’s CE was immediately realized, but
that value would have been just 32.7% in the securitization
case. (For further details on these calculations, see Web
Appendix A at http://www.marketingpower.com/ jmmay11.)
Customer Equity Sustainability Ratio (CESR) / 125

TAblE 4
Countrywide Earnings Structure
A: Countrywide: Securitization Case (Actual business)
1998 1999 2000 2001 2002 2003 2004 2005 2006 2007
126 / Journal of Marketing, May 2011
Originate and Distribute business
Earnings on sale of mortgage loans
and securities 699,433 557,743 611,092 1,498,812 3,377,065 5,541,539 4,386,536 4,300,579 4,897,771 2,220,164
Earnings on sale of other loans and
securities 623,531 406,458 398,544 103,178 94,153 345,897 455,546 561,201 784,076 214,559
Total earnings, originate and distribute
business 1,322,964 964,201 1,009,636 1,601,990 3,471,218 5,887,436 4,842,082 4,861,780 5,681,847 2,434,723
buy and Hold business
Net interest income 66,764 93,933 10,678 331,877 792,230 1,359,390 1,965,541 2,237,935 2,688,514 587,882
Net loan servicing fees 241,817 551,723 584,024 –2,146 –865,752 –463,050 465,650 1,493,167 1,300,655 909,749
Net insurance premiums earned 12,504 75,786 274,039 316,432 561,681 732,816 782,685 953,647 1,171,433 1,523,534
Other 175,363 202,742 195,462 248,506 358,855 462,050 510,669 470,179 574,679 605,549
Total earnings, buy and hold business 496,448 924,184 1,064,203 894,669 847,014 2,091,206 3,724,545 5,154,928 5,735,281 3,626,714
Total earnings 1,819,412 1,888,385 2,073,839 2,496,659 4,318,232 7,978,642 8,566,627 10,016,708 11,417,128 6,061,437
Customer equity 2,567,495 3,604,128 3,753,248 4,174,750 5,468,867 12,087,915 13,868,776 18,195,139 20,151,106 9,646,673
Customer equity sustainability ratio .291 .476 .447 .402 .210 .340 .382 .449 .433 .372
Securitization ratio 58.9% 65.3% 72.1% 95.5% 95.9% 86.1% 89.4% 82.5% 86.1% 90.4%
Return on equity 15.3% 15.2% 11.6% 12.7% 18.2% 35.8% 23.9% 21.9% 19.7% –4.9%
Nonprime loans as a percentage of total
loan production 3.18% 2.69% 6.23% 4.50% 3.74% 4.56% 10.85% 8.94% 8.67% 4.09%
Delinquent mortgage loans as a percentage
of total loan production 3.91% 3.55% 11.30% 14.42% 14.41% 12.46% 11.29% 15.20% 19.03% 27.29%
b: Countrywide: Non-Securitization Case (Counterfactual Analysis)
1998 1999 2000 2001 2002 2003 2004 2005 2006 2007
Total earnings 1,307,973 1,645,620 1,900,767 1,422,166 2,459,488 5,539,118 8,453,218 10,656,583 11,859,540 8,304,510
Customer equity 2,567,495 4,166,711 4,639,131 5,339,600 7,517,683 16,039,134 20,618,857 25,532,604 27,518,455 17,264,104
Customer equity sustainability ratio .491 .605 .590 .734 .673 .655 .590 .583 .569 .519
Notes: Amounts listed are in thousands of U.S. dollars.

Application of the CESR to Industries Outside
Banking
Although securitization is most commonly used in banking,
it enjoys great popularity in many other industries as well.
Few of these firms provide sufficient information to illustrate
the shift from long-term value creation to short-term
profit realization either. In Table 5, we provide an overview
of nine securitizations from a wide range of industries,
including three state governments and several nonprofit
organizations, which provide enough information to calcu-
Reference
Burke Sylva
(1999)
Campbell and
Hashimoto
(2003)
EDF Group
(2003)
Everton Football
Club (2002)
Firm or
Institution Transaction Type
David Bowie In January 1997, David Bowie raised $55 million
from the issue of ten-year asset-backed bonds,
with collateral consisting of future royalties from
25 albums that he recorded before 1990.
Korean Air Lines Korean Air Lines securitized 2003 future
receivables from the sale of its tickets on routes
between Japan and Korea worth ¥27 billion
($242 million) for a period of 36 months.
EDF (Electricité
de France)
Everton FC
(United Kingdom)
In 1999, EDF transferred receivables of €1.1
billion (US$1.2 billion in 1999) of employees’
housing loans to the Electra mutual securitization
fund. Since the end of 2000, EDF has
transferred future trade receivables of energy
supply contracts to a mutual securitization fund,
reaching an amount of €2.1 billion ($1.8 billion in
2001) by 2001.
In 2002, Everton FC (UK soccer club) signed a
securitization agreement serviced by future season
ticket sales and match-day ticket sales. The
£30 million ($45 million) loan must be repaid with
the interest charge over a period of 25 years.
FIFA (2006) FIFA FIFA, world soccer’s governing body, raised
$420 million through a bond deal in 2001, using
expected receipts of $536 million from the 2002
World Cup in Japan and South Korea and the
2006 World Cup in Germany as collateral.
Armstrong (2009) Keele University
(United Kingdom)
Keele University (UK) securitized future student
housing income worth £55.4 million (US$83.8
million) to finance refurbishments of the student
residences and repay existing debt.
Kasprak (2002) State of Alaska Beginning in July 2001, Alaska used 40% of the
tobacco settlement revenue to get an infusion of
cash to build, rehabilitate, and remodel schools.
The 2001 legislature passed in addition a law to
securitize an additional 40% of the settlement
revenue so that the transactions together securitized
80% of the tobacco settlement revenue.
Kasprak (2002) State of
South Carolina
TAblE 5
Securitization in Industries Outside banking
South Carolina passed legislation to securitize its
MSA settlement dollars through 2018. The state
established four funds to develop its water and
wastewater infrastructure, compensate individuals
for losses in tobacco production (25%), and
fund a variety of health care programs and services
(75%).
Kasprak (2002) State of Wisconsin Wisconsin sold $1.54 billion of bonds backed by
state appropriations to refinance debt issued by
Badger Tobacco Asset Securitization Corp.
late the effect of securitization on changes in CESR. (For
more details on the calculations of securitization cases in
industries outside banking, see Web Appendix B at http://
www.marketingpower.com/jmmay11.)
According to Table 5, the transaction volume ranged
from several million to billions of U.S. dollars, and securitization
appears in diverse settings. The most remarkable
result is that the effects of securitization on changes in
CESR differ across the nine securitizations and hardly correlate
with transaction volume. For example, the largest
securitization of $1.8 billion, for EDF Group, diminished
CESR Without
Securitization
(1997):
.911
(2003):
.833
(1999):
.908
(2001):
.882
(2002):
.989
(2001):
.943
(2000):
.938
(2001):
.906
(2001):
.906
(2001):
.915
CESR with
Securitization
(1997):
.230
(2003):
.825
(1999):
.905
(2001):
.876
(2002):
.958
(2001):
.876
(2000):
.891
(2001):
.181
(2001):
.104
(2001):
0
Customer Equity Sustainability Ratio (CESR) / 127

CESR only from .882 to .876, and one of the smallest securitizations
of $242 million, for Korean Air Lines, led to a
comparable drop in the size of CESR. In contrast, the securitizations
of David Bowie and the State of Wisconsin both
prompted drops of CESR from approximately .900 to less
than .250.
In summary, although securitization is widespread in
banking and prominent in the context of the financial crisis,
it also is used commonly outside banking business. The
diversity of firms and institutions indicate that securitization
is applicable in many settings. Although we acknowledge
that the lack of transparency requires strong assumptions
about the missing information in some of the securitization
cases in Table 5, CESR describes well how strongly the
earnings shifted over time.
Discussion and Conclusion
Discussion of Customer Equity Sustainability
Ratio
Our empirical results illustrate that many banks and firms
fail to provide sufficient transparency about their securitization
activities, which makes it difficult, if not impossible, to
evaluate which earnings come from ongoing business and
which result from the one-time effects of securitization.
Unfortunately, current reporting standards provide limited
means to detect such behavior because firms are not
required to report future earnings. Our empirical studies
show that CER and the newly developed CESR provide
more transparency and emphasize that firms that make
extensive use of securitizations should increase their shortterm
profits at the expense of the long-term value of their
customer base. As forward-looking marketing metrics, CER
and CESR add another perspective to backward-oriented
accounting and reporting practices, which provide mainly
historic information. We argue that marketing researchers
have accumulated enough knowledge in the past decade to
determine future values (e.g., by calculating CLV, CE). Ignorance
of future earnings is no longer justified, and marketing
academics should lead the field in the drive to consider future
earnings in financial reporting. Such reporting techniques
may also hasten marketing’s reentry into the boardroom.
Comparison of Customer Equity Sustainability
Ratio with Other Performance Ratios
As a new key ratio, CESR requires comparisons with wellestablished
performance ratios, such as ROE and return on
risk-adjusted capital; efficiency ratios, such as the cost
income ratio; and productivity ratios, including economic
value added. 3 These ratios all use historic data and consider
3Return on risk-adjusted capital = net income/allocated risk capital;
this can be used to compare projects or investments under
consideration on the basis of their implied risk profile. Cost
income ratio = operating expenses/operating income—that is, the
expenses needed to realize an income of $1. Economic value
added = net operating profit – (capital ¥ cost of capital); economic
value added was developed to measure the real surplus a company
generates in a specific period from an appropriate return on the
capital invested.
128 / Journal of Marketing, May 2011
only the current level of a firm’s profit or earnings, which
means they neglect a possible shift from future (i.e., longterm)
value creation to short-term profit realization. With
securitization, the ratios exhibit better values in the respective
year than they would otherwise because short-term profits
decrease cost–income ratio and increase ROE, return on
risk-adjusted capital, and economic value added. In contrast,
securitization lowers CESR immediately and provides stakeholders
with a reliable indicator that the firm has moved from
long-term value creation to short-term profit realization.
Another contribution of CESR becomes obvious when
we compare it over time with the development of earnings,
CE, ROE, and the securitization ratio (i.e., proportion of the
securitized loan volume to total loan volume in a particular
year) of Countrywide. Except for 2007, earnings and CE
increased constantly, suggesting a successful business and
value-creation process. In this empirical case, ROE would
increase only after the securitization intensity had risen and
then would return to its former levels. Furthermore, ROE
could increase in response to other determinants, such as
higher issued loan volumes, greater efficiency, and better
equity structures. Therefore, a high ROE value can, but does
not necessarily, substantiate heavy securitization engagement
or short-term profit realizations.
To determine the relationship between the securitization
ratio and CESR, we use the results from our empirical study
of Countrywide, which reveal an insignificant correlation
of –.241 (p = .503, n = 10). This outcome indicates the
securitization ratio cannot substitute for CESR. Although
the changes in the securitization ratio point to a modification
in the securitization strategy of Countrywide, it cannot
indicate how much contracted business still gets realized in
upcoming years, because this ratio usually measures the
share of loans securitized in one particular year. Moreover,
securitization ratios do not distinguish the case in which a
firm distributes the earnings from securitization to its shareholders
from one in which it reinvests these earnings and
thus retains them in the firm.
As an early warning system, CESR translates information
into an earnings-related metric and includes earnings
from nonsecuritization businesses. At the moment of securitization,
CESR instantly declines, alerting stakeholders to
the firm’s short-term profit realization. Furthermore, CESR
enables stakeholders to recognize the firm’s distribution or
reinvestment strategy, as we show in our example. Information
about developments in the firm’s shift from long-term
value creation to short-term profit realization and the comparisons
with the values of other firms thus provide strong
advantages. In particular, this ratio condenses important
information without losing content and without limits on its
applicability to a wide range of businesses. Legal and regulatory
restrictions and the effort to keep competitive advantages
have hindered firms from publishing a large pool of
data on its future orientation, supporting the application of
CER and CESR.
Relationship Between Customer Equity and
Customer Equity Sustainability Ratio
To deliver the necessary level of transparency, firms should
report both CE and CESR and grant stakeholders insights

into the value of their current customer base and the extent
of their long-term value creation. As we highlight in Figure
1, a high CE value indicates a strong customer base, which
is always better than a weaker one. In contrast, CESR indicates
whether the firm pursues long-term value creation or
short-term profit realization. We recognize explicitly that
CESR has no per se optimal value and leaves normative
conclusions up to stakeholders. For example, if the firm is
able to attract investors that are willing to pay a price above
the fair value of the underlying loans, the sale is beneficial
for the firm and its stakeholders even though securitization
reduces CESR. The reverse holds if buyers are only willing
to pay a price that is lower than the value for the seller. Still,
if a firm with low values of CE and CESR reports high
earnings for the current year, stakeholders should recognize
its weak customer base and drop in future earnings if the
firm cannot acquire enough valuable new customers or generate
more business with its current customers.
Even if sellers and buyers agree on the fair value of
securities, extreme values of CESR could describe the best
case for a firm under special circumstances. For example, a
CESR of 0 can be the most desirable for a firm with liquidity
needs that securitizes its entire future earnings to bridge
the liquidity gap and likely prevent insolvency. For a project
firm to develop and operate an office building with some
lease contracts already signed at construction start, in the
first years of development, CESR will be 1 because no
earnings occur until the building is fully operational.
Thus, there are no general critical values of CESR per
se because CESR is business specific. However, stakeholders
should be alerted if their firm’s CESR is fundamentally
lower than the CESR in the firm’s peer group and be aware
that their firm is focusing on short-term profit realization. In
addition, if the firm’s CESR continuously declines over
time, stakeholders should seriously consider a firm’s
intended shift from long-term value creation to short-term
profit realization. Moreover, it is necessary to exercise caution
if huge increases in compensation or large dividends go
along with strong declines in CESR.
limitations and Further Research
Our study is subject to some limitations that should prompt
further research. First, the banking industry is complex,
though we focus on simple standard loans in our example.
Therefore, further research should identify the implied
earnings structure of more complex products and business
transactions so that CER and CESR can be implemented in
diverse industries.
Second, the aim of this article is to present CESR as a
new metric and outline its benefits in several empirical studies.
Further research should observe how regulatory authorities,
particularly in the financial industry, can best establish
obligations to publish CER and CESR. Securitizations can
have a dramatic impact on variable payments in management
compensation if these are linked to short-term profits.
Further research could more strongly analyze how CER and
CESR can help develop more appropriate compensation
FIGURE 1
Relationship between Customer Equity (CE) and
Customer Equity Sustainability Ratio (CESR)
CESR
(Low)
Short-term
profit realization
with a strong
customer base
Short-term
profit realization
with a weak
customer base
CE (High)
Long-term
value creation
with a strong
customer base
Long-term
value creation
with a weak
customer base
CE (Low)
CESR
(High)
plans and better disseminate information about managers’
long-term value creation efforts.
Third, the quality of CER and CESR crucially depends
on good forecasts, particularly in industries with high
shares of noncontractual business. Marketing researchers
have developed powerful methods to accommodate the
greater amount of uncertainty in such situations (e.g.,
Reinartz and Kumar 2000), but the influence of such methods
on current reporting standards is still limited. Determining
the value of at least half of the nine securitizations displayed
in Table 5 requires profound knowledge of the firm’s
ability to further market the underlying products; however,
currently, marketing scholars seem to play only a minor role
in determining the value of these securitizations. Thus,
future studies could outline in more detail how current marketing
knowledge could be used to increase the quality of
forecasts and determine the value of these securitizations
more accurately.
Fourth, current reporting standards already apply forwardlooking
information and develop accounting forecasts based
on future earnings, such as the fair values of goodwill and
financial assets. External auditors assess uncertainty and reliability,
limiting the space for managerial arbitrariness and
opening the floor to our reporting technique. Therefore, further
research should examine if CER and CESR require more
detailed rules to reduce the risk of managerial arbitrariness.
Fifth, securitization is likely to affect marketers and
their customers, as we illustrate for a bank, though the
effects hold equally for other firms. The lack of future earnings
from the securitized loans of current customers obliges
a bank to intensify its customer acquisition efforts and sell
more loans to ensure that future earnings do not decrease, in
particular if it distributes earnings to its shareholders. However,
too much pressure on customer acquisition is likely to
also increase the chances of acquiring unprofitable and
high-risk customers (Cao and Gruca 2005; Reinartz,
Thomas, and Kumar 2005), a factor that finance research
Customer Equity Sustainability Ratio (CESR) / 129

also recognizes: “As balance sheets expand, new borrowers
must be found. When all prime borrowers have a mortgage
but balance sheets still need to expand, then banks have to
lower their lending standards in order to lend to subprime
borrowers. The seeds of the subsequent downturn in the
credit cycle are thus sown” (Shin 2009, p. 310). Further
research could provide stronger support for this speculative,
albeit likely, statement.
In summary, the lack of transparency in current financial
reporting about the long-term effects of securitization is
unacceptable. Customer equity reporting and the CESR
REFERENCES
Armstrong, Peter (2009), “Appraising Income Securitisation in the
Public Sector: The Sale of Student Rents at Keele University,”
International Journal of Critical Accounting, 1 (1/2), 3–57.
Blattberg, Robert C. and John Deighton (1996), “Manage Marketing
by the Customer Equity Test,” Harvard Business Review,
74 (4), 136–44.
Brinkworth, Stuart (2004), “Bondholders Cry Foul on Football
Securitization,” International Financial Law Review, 23 (3),
22–24.
Burke Sylva, Jennifer (1999), “Bowie Bonds Sold for Far More
than a Song: The Securitization of Intellectual Property as a
Super-Charged Vehicle for High Technology Financing,” Santa
Clara Computer & High Technology Law Journal, 15 (1),
195–230.
Calem, Paul S. and Michael LaCour-Little (2004), “Risk-Based
Capital Requirements for Mortgage Loans,” Journal of Banking
& Finance, 28 (3), 647–72.
Campbell, Neil and Norifusa Hashimoto (2003), “Korean Air
Lines Securitizes Future Yen Ticket Receivables,” International
Financial Law Review, 22 (11), 16–17.
Cao, Yong and Thomas S. Gruca (2005), “Reducing Adverse
Selection Through Customer Relationship Management,” Journal
of Marketing, 69 (October), 219–29.
Countrywide Financial Corp. (2008), “Form 10-K,” annual report.
Coval, Joshua D., Jakub Jurek, and Erik Stafford (2009), “The
Economics of Structured Finance,” Journal of Economic Perspectives,
23 (1), 3–25.
Dechow, Patricia M., Linda A. Myers, and Catherine Shakespeare
(2010), “Fair Value Accounting and Gains from Asset Securitizations:
A Convenient Earnings Management Tool with Compensation
Side-Benefits,” Journal of Accounting and Economics,
49 (1), 2–25.
——— and Catherine Shakespeare (2009), “Do Managers Time
Securitization Transactions to Obtain Accounting Benefits?”
Accounting Review, 84 (1), 99–132.
Demyanyk, Yuliya and Otto Van Hemert (2009), “Understanding
the Subprime Mortgage Crisis,” Review of Financial Studies,
(published electronically May 4), [DOI: 10.1093/rfs/hhp033].
Downey, Connor (1999), “Whole Business Securitization Comes
of Age,” International Financial Law Review, 18 (September),
8–12.
EDF Group (2003), Annual Report 2003. Paris: EDF Group.
Everton Football Club (2002), Annual Report & Statement of
Accounts 2002. Liverpool, UK: Everton Football Club.
Fabozzi, Frank J., Henry A. Davis, and Moorad Choudry (2006),
Introduction to Structured Finance. Hoboken, NJ: John Wiley
& Sons.
FASB (1996), Statement of Financial Accounting Standards No.
125. Norwalk, CT: Financial Accounting Standards Board.
——— (2000), Statement of Financial Accounting Standards No.
140. Norwalk, CT: Financial Accounting Standards Board.
130 / Journal of Marketing, May 2011
provide more transparency to stakeholders and demonstrate
the financial stability of banks and firms. Perhaps CER and
CESR could have helped avoid some of the challenges of
the current financial crisis; more important, perhaps they
can provide insights into how investors and firms can be
supported in their attempts to prevent the next one. In any
case, marketing should not leave the field to finance and
accounting; instead, it should emphatically assert its claim
and provide stakeholders with substantial information about
the long-term value of the customer base.
Franke, Günther and Jan P. Krahnen (2008), “The Future of Securitization,”
paper presented at Brookings–Tokyo Club–Wharton
Conference, Washington, DC (October 16).
Greenbaum, Stuart I. and Anjan V. Thakor (1987), “Bank Funding
Modes: Securitization Versus Deposits,” Journal of Banking &
Finance, 11 (3), 379–401.
Gregory, Michael (2002), “Broncos Finance Stadium via ABS,”
Asset Securitization Report, 2 (7), 1–6.
Gupta, Sunil, Donald R. Lehmann, and Jennifer Ames Stuart
(2004), “Valuing Customers,” Journal of Marketing Research,
41 (February), 7–18.
International Federation of Association Football (2006), FIFA
Financial Report 2006. Zurich: International Federation of
Association Football.
Joshi, Amit and Dominique M. Hanssens (2010), “The Direct and
Indirect Effects of Advertising Spending on Firm Value,” Journal
of Marketing, 74 (January), 20–33.
Kasprak, John (2002), “Securitization of Tobacco Settlement
Funds,” (accessed November 5, 2010), [available at http://
www. cga.ct.gov/2002/olrdata/ph/rpt/2002-R-0736.htm].
Kendall, Leon T. (1996), “Securitization: A New Era in American
Finance,” in A Primer on Securitization, Leon T. Kendall and
Michael J. Fishman, eds. Cambridge: MIT Press, 1–16.
Ketkar, Suhas and Dilip Ratha (2008), “Development Financing
During a Crisis: Securitization of Future Receivables,” working
paper, The World Bank.
Kothari, Vinod (2006), Securitization. Singapore: John Wiley &
Sons.
Luo, Dan, Dragon Yongjun Tang, and Sarah Qian Wang (2009), “A
Little Knowledge Is a Dangerous Thing: Data History, Model
Uncertainty and CDO (Mis-) Pricing,” working paper, School
of Economics and Finance, University of Hong Kong.
McCreevy, Charlie (2008), “Commission Regulation (EC) No
1126/2008,” Official Journal of the European Union, L 320,
1–481.
Reinartz, Werner J. and V. Kumar (2000), “On the Profitability of
Long-Life Customers in a Noncontractual Setting: An Empirical
Investigation and Implications for Marketing,” Journal of
Marketing, 64 (October), 17–35.
———, Jacquelyn S. Thomas, and V. Kumar (2005), “Balancing
Acquisition and Retention Resources to Maximize Customer
Profitability,” Journal of Marketing, 69 (January), 63–79.
Rust, Roland T., Katherine N. Lemon, and Valarie A. Zeithaml
(2004), “Return on Marketing: Using Customer Equity to
Focus Marketing Strategy,” Journal of Marketing, 68 (January),
109–127.
———, Valarie A. Zeithaml, and Katherine N. Lemon (2000),
Driving Customer Equity: How Customer Lifetime Value Is
Reshaping Corporate Strategy. New York: The Free Press.
Ryan, Stephen G. (2008), “Accounting in and for the Subprime
Crisis,” Accounting Review, 83 (6), 1605–1638.

Santomero, Anthony M. and David F. Babbel (1997), Financial
Markets, Instruments, and Institutions. Chicago: McGraw-Hill.
Shin, Hyun Song (2009), “Securitisation and Financial Stability,”
Economic Journal, 119 (March), 309–332.
Srinivasan, Shuba and Dominique M. Hanssens (2009), “Marketing
and Firm Value: Metrics, Methods, Findings, and Future
Directions,” Journal of Marketing Research, 46 (June),
293–312.
Story, Louise, Landon Thomas Jr., and Nelson D. Schwartz
(2010), “Wall Street Helped Mask Debt Fueling Europe’s Crisis,”
The New York Times, (February 14), A1.
Tuli, Kapil R., Sundar G. Bharadwaj, and Ajay K. Kohli (2010),
“Ties That Bind: The Impact of Multiple Types of Ties with a
Customer on Sales Growth and Sales Volatility,” Journal of
Marketing Research, 47 (February), 36–50.
Wiesel, Thorsten, Bernd Skiera, and Julián Villanueva (2008),
“Customer Equity: An Integral Part of Financial Reporting,”
Journal of Marketing, 72 (April), 1–14.
Customer Equity Sustainability Ratio (CESR) / 131

Für weitere Informationen zum E-Finance Lab wenden Sie sich bitte an:
Prof. Dr. Wolfgang König
Vorstandsvorsitzender des E-Finance Lab,
Geschäftsführender Direktor des House of Finance
Grüneburgplatz 1
60323 Frankfurt
Telefon +49 (0)69 / 798 - 34001
Fax +49 (0)69 / 798 - 33910
E-Mail koenig@efinancelab.com
Prof. Dr. Peter Gomber
Stellvertretender Vorstandsvorsitzender des E-Finance Lab
Grüneburgplatz 1
60323 Frankfurt
Telefon +49 (0)69 / 798 - 34683
Fax +49 (0)69 / 798 - 35007
E-Mail gomber@wiwi.uni-frankfurt.de
oder besuchen Sie unsere Webseite
http://www.efinancelab.com