the value proposition of information-based business models

ALBERT-LUDWIGS-UNIVERSITÄT FREIBURG

WIRTSCHAFTS- UND VERHALTENSWISSENSCHAFTLICHE FAKULTÄT

The Value Proposition of

Information-based Business Models:

Evidence from Online Content and

Energy Markets

Inaugural-Dissertation

zur

Erlangung der Doktorwürde

der Wirtschafts- und Verhaltenswissenschaftlichen Fakultät

der Albert-Ludwigs-Universität Freiburg. i. Br.

vorgelegt von

Philipp Bodenbenner

geboren in Münster

WS 2013/2014

ii

Druckdatum: 10.11.2014

Albert-Ludwigs-Universität Freiburg im Breisgau

Wirtschafts-und Verhaltenswissenschaftliche Fakultät

Kollegiengebäude II

Platz der Alten Synagoge

79085 Freiburg

Dekan: Prof. Dr. Albert Gollhofer

Erstgutachter: Prof. Dr. Dirk Neumann

Zweitgutachter: Prof. Dr. Dr. h.c. Günter Müller

Datum des Promotionsbeschlusses: 27.10.2014

iii

ACKNOWLEDGEMENTS

I would like to express my gratitude to all the people who have supported me in

developing this doctoral thesis.

First and foremost, I would like to thank Prof. Dr. Dirk Neumann for his willingness to

oversee this work as my doctoral supervisor. I have greatly appreciated the many ideas and

suggestions he shared and discussed with me. Not only his professional advice, also his

continuous motivation and positive spirit have had a large influence on the successful

completion of this work. Furthermore, I would like to thank Prof. Dr. Dr. h.c. Günter

Müller for accepting the role of a second reviewer for my thesis.

This work has been developed during my time as an external associate at the chair of

Information Systems Research. I want to thank all fellow colleagues and the assistant Carla

Li-Sai for their support and cooperation, and their efforts to bridge the geographical

distance between Freiburg and Bonn. My special thanks go to Stefan Feuerriegel for an

intense collaboration on the energy topics – from idea generation over content discussions

to a number of successful joint publications.

My employer, the consulting company Detecon International GmbH, has provided me with

the opportunity to dedicate a substantial part of my time to accomplish this work, and has

supported me throughout. I would like to especially point out Dr. Volker Rieger in this

respect. My interest in the research topic has been sparked off already in the early days of

my career at Detecon, through intense and stimulating discussions with my colleague Dr.

Christoph Tempich. Since then, many fruitful exchanges with fellow colleagues on this

topic have further shaped my work. Also my former colleagues Dr. Uwe Alkemper and

Michael Friedmann, the founders of YOOCHOOSE GmbH, contributed to my work by

providing me the raw data for my research on recommendation systems.

Last but not least, I would like to thank my wife Christine and my family for their

unconditional support and continuous encouragement over the past years.

i

TABLE OF CONTENTS

I. Thesis Overview ............................................................................................. 1

1 Introduction..................................................................................................... 2

1.1 Motivation ............................................................................................... 2

1.2 Research Outline...................................................................................... 5

1.3 Structure of Thesis ................................................................................... 7

2 Building Blocks for Taxonomy ...................................................................... 9

2.1 Business Model Canvas ......................................................................... 10

2.2 Information Value Chain ....................................................................... 11

3 Taxonomy of Information-based Business Models ...................................... 14

3.1 Classification Framework ...................................................................... 14

3.2 Class 1: Process Information for Transactional Optimization .............. 16

3.3 Class 2: Apply Analytics to Discover Sweet Spots ............................... 18

3.4 Class 3: Leverage Information to Upgrade Offering ............................. 23

3.5 Class 4: Design a Pure-Play Informational Value Proposition ............. 27

4 Case Studies on Information-based Business Models .................................. 32

4.1 Case Study 1: Personalized Recommendations ..................................... 33

4.2 Case Study 2: Electricity Demand Response ........................................ 35

4.3 Comparison of Case Studies .................................................................. 37

5 Summary and Future Research ..................................................................... 41

6 References..................................................................................................... 46

II. Case Study 1: Personalized Recommendations ........................................ 51

A. Always Mystify, Mislead, and Surprise your Customers? Impact of

Personalized Recommendations on the Performance of an Online Content

Provider .............................................................................................................. 52

1 Introduction................................................................................................... 53

2 Personalization and Recommender Systems ................................................ 55

3 Literature Review ......................................................................................... 57

3.1 From Technical to Transactional Evaluation of Personalization........... 58

3.2 The Role of Personalization for Relationship Marketing ...................... 59

3.3 The Long Tail in e-Commerce .............................................................. 60

3.4 Comparison with Previous Studies ........................................................ 61

4 Research Hypotheses .................................................................................... 63

4.1 Whetting the User Appetite ................................................................... 63

4.2 Boosting Revenue .................................................................................. 64

4.3 Cutting the Long Tail Short ................................................................... 66

ii

5 Design of Empirical Study............................................................................ 67

5.1 Sampling & Data Collection ................................................................. 68

5.2 Sample Characteristics .......................................................................... 71

5.3 Modeling Approach ............................................................................... 73

6 Empirical Results .......................................................................................... 74

6.1 Attraction of Recommended Links (H1) ............................................... 75

6.2 Duration of Website Visits (H2) ............................................................ 78

6.3 Frequency of Website Visits (H3) ......................................................... 83

6.4 Breadth of Accessed Item Base (H4) .................................................... 84

6.5 Measuring the Overall Business Value ................................................. 86

6.6 Managerial Implications ........................................................................ 88

7 Summary and Outlook .................................................................................. 89

8 References..................................................................................................... 93

III. Case Study 2: Electricity Demand Response ............................................. 99

A. An Information System Architecture for Demand Response ................. 100

1 Introduction................................................................................................. 101

2 Demand Response in IS Research .............................................................. 102

3 Applying the Design Science Approach ..................................................... 104

3.1 Defining Requirements for the Information System ........................... 104

3.2 Designing the Information System ...................................................... 108

3.3 Demonstrating Capability of the Information System ......................... 111

3.4 Evaluating the Information System ..................................................... 111

4 Revisiting the Design Science Approach ................................................... 113

4.1 Reviewing the Design Science Guidelines .......................................... 113

4.2 Reflections on Design Science ............................................................ 115

5 Conclusion and Outlook ............................................................................. 117

6 References................................................................................................... 118

B. Optimal Information Granularity in Demand Response ........................ 120

1 Introduction................................................................................................. 121

2 Literature Review ....................................................................................... 124

2.1 Demand Response and Smart Grid in IS Research ............................. 124

2.2 Economic Appraisal of Demand Response ......................................... 125

3 Research Questions ..................................................................................... 126

4 System Model for Demand Response......................................................... 128

4.1 Information Exchanges in the Demand Response System .................. 129

4.2 Reference IS Architecture of the Demand Response System ............. 132

5 Definition of the Cost-Value-Model ........................................................... 136

iii

5.1 Model Assumptions ............................................................................. 136

5.2 Cost Structure of IS Components for Demand Response ................... 137

5.3 Savings through Demand Response .................................................... 141

5.4 Value of Information in Demand Response ........................................ 143

6 Evaluation of the Demand Response System ............................................. 144

6.1 Setup and Dataset of Computational Analysis .................................... 144

6.2 Results ................................................................................................. 146

7 Summary and Conclusions ......................................................................... 158

8 References................................................................................................... 161

C. Scenarios for Integrating Demand Response in Electricity Markets ..... 166

1 Introduction................................................................................................. 167

2 Related Work .............................................................................................. 169

2.1 Designing a Demand Response System .............................................. 170

2.2 Financial Dimension of Demand Response......................................... 171

3 Integrating Demand Response into Electricity Markets ............................. 172

3.1 Electricity Procurement at the Product Market ................................... 173

3.2 Market Design of Control Reserve Exchanges.................................... 173

3.3 Definition of Balancing Energy and Imbalance Penalties ................... 175

4 Research Model: Measuring the Financial Impact of DR .......................... 176

4.1 Research Questions ............................................................................. 176

4.2 Cost Structure of Demand Response Systems ..................................... 178

4.3 Measuring the Revenue Potentials from Demand Response Systems 181

5 Evaluation ................................................................................................... 184

5.1 Datasets ................................................................................................ 185

5.2 Results and Managerial Implications .................................................. 186

5.3 Policy Implications for Electricity Markets Design ............................ 190

6 Conclusion and Outlook ............................................................................. 191

7 References................................................................................................... 193

IV. Appendix ..................................................................................................... 196

1 Curriculum Vitae ........................................................................................ 197

2 Referred Research Publications .................................................................. 199

1

I. THESIS OVERVIEW

PART I THESIS OVERVIEW

2

1 Introduction

1.1 Motivation

Since the information and communication technologies (ICT) have started their triumphal

course, particularly with the rise of the internet and the dot-com-boom at the turn of the

millennium, more and more realms of everyday life and society have become digital. The

digital transformation is unstoppable – we are amidst the information age. By now, the

internet has ultimately acquired the status of a commodity and essential infrastructure for

business and private life. In addition, ubiquitous computing and the internet of things have

set out to become the next concepts to turn into reality.

Together with the increasing digitalization occurs an abundance of produced and available

information. Recent years have seen exponential growth of data in business and society.

User-generated content increasingly dominates internet traffic, mostly in social media such

as Facebook, Instagram and YouTube. This phenomenon is further fueled by the mobile

revolution (Cook, 2008). By using their smart phones, people can be online wherever and

whenever they want (Kwok, 2009). Activities and news are posted in real-time and

instantaneously shared with family and friends. Analysys Mason (2013) estimates the

mobile data traffic volume to grow by at least factor six until 2018.

Recently, the trend towards the internet of things has created an additional boost on the

already staggering large data quantities. Forecasts expect 25-50 billion connected sensors,

which never stop producing data, by the year 2020 (Cisco, 2011; Gartner, 2013). This

amount resembles a more than 30-fold increase compared to today. The bandwidth of

possible use cases for connected sensors reaches into all areas of life: from consumer-

oriented appliances all the way to sensors for remote monitoring of industrial machines.

IDC (2012) predicts a doubling of the digital universe every two years until 2020, which

corresponds to an increase of the data amount from around 5,000 exabytes in 2014 to

40,000 exabytes in 2020 (see Figure 1).


3

Furthermore, the digital transformation carries along societal changes (Castells, 2010). The

way in which humans communicate is changing and new forms are emerging – today most

people do not simply talk or write, instead they blog, twitter or chat. Information exchange

happens quicker than ever; while watching news once a day was deemed sufficient years

ago, the latest happenings are nowadays consumable in real-time, but at the same time

suffer to become obsolete similarly fast. The access to information increasingly becomes a

decisive factor for the social position. Terms like information superiority (Strikwerda,

2011) and information poverty (Norris, 2001), or digital divide, are more often used. In

turn, the explosive growth of information demands new competencies from its users. With

information being anything but scarce, everybody needs to find their way through the

information jungle and extract the relevant aspects – this phenomenon even provokes

speaking about “death by information overload” (Hemp, 2009). Not only individuals, but

also enterprises face a massive organizational transformation challenge in order to keep

pace with these developments.

Dating back to at least 1999, the importance of information goods has been recognized by

researchers (Shapiro and Varian, 1999). This way of thinking is progressively diffusing

into the markets. Information is increasingly perceived as a company’s strategic asset and

not only as a by-product of processes. In a well-received note, Tim O’Reilly has declared

data as “the next Intel Inside“ (O'Reilly, 2005). Marissa Meyer (then Google's vice

president of Search Products & User Experience, today CEO of Yahoo) has attributed a

similar importance to information, by proclaiming on a conference in 2007: “Google is all

about large amounts of data” (Perez, 2007). The perception of information as an important

Figure 1. Forecast of Global Data Growth, Based on IDC (2012).

40,000

30,000

20,000

10,000

0

2010 2012 2014 2016 2018 2020

Global

Digital Data in Exabytes


4

factor for companies is underpinned by various industry studies conducted by the

management consultancy Detecon (see e.g. Rieger and Tempich, 2008).

Just being aware of the importance of information is however not sufficient. Redman

(2008b) remarks that companies in most cases “grossly underutilize their data assets”. A

study by IDC (2012) concludes with similar results: today, only a minor part of the digital

universe is explored for analytical value. In a survey among company executives across all

industries, conducted by The Economist (2010), more than three quarters name the

extraction and usage of relevant data from the tangled mass a key challenge.

Companies thus need to react on the digital transformation and innovate their business

models in order to leverage the potential of data to their advantage (Brynjolfsson and

Saunders, 2010; Amit and Zott, 2012; Veit et al., 2014). Internally intensifying the usage

and analysis of data is however just one aspect. The other, more progressive side of the

coin is making information an (essential) part of the company’s value proposition and

marketing it accordingly (Águila-Obra et al., 2007; Bloching et al., 2012). Novel business

models are emerging that are built upon and spawned around information, for example by

producing, distributing or selling information goods. Both incumbents and new players can

venture into information-based business (Clemons, 2009a). Redman (2008a) points out

that “companies with a cohesive strategy for integrating digital and physical elements can

successfully transform their business models and set new directions for entire industries.”

Gartner (2010) has coined the term hyperdigitization, meaning the economic shift towards

virtual goods and services. Its economic output is estimated at US$ 2.9 trillion, which

surmounts the results belonging to the global manufacturing industry.

The disruptive change in business models induced by new trends in information

technology elevates the topic high on the research agenda. Veit et al. (2014) consider the

analysis of IT-enabled business models as one of the most urgent topics for IS research.

This work picks up this research topic and examines the value creation in information-

based business models.


5

For this purpose, a taxonomy is defined to classify business models according to the

respective role played by information. Subsequently the taxonomy is followed by an in-

depth analysis of two information-based business models, underpinned by case studies

using real-world data. Besides categorizing business models according to the role of

information, the focus of this thesis is to quantify the economic value achievable through

information. The examination reaches from considering the potential revenue streams as

well as the cost structure accompanying an information-driven value proposition.

1.2 Research Outline

This section introduces the two research questions that this thesis addresses. Both deal with

the opportunities around the production factor information, which companies need to seize

in order to innovate and advance their businesses in the era of digital transformation.

The first introduction of the term business model in literature dates back to the 1950s

(Bellman et al., 1957). It was not until the late 1990s though that business models as a

research topic have aroused a broader interest in academia. The increase of attention at that

time mainly stems from the emergence of the internet and electronic businesses. Business

models are largely a product of the dot com era; which might be due to the fact that

traditional, offline revenue models were outdated, and new frameworks matching the

changed business environment were required. From this specific area of applicability the

concept expanded again towards a broader context, spanning the information systems

arena as well as the strategic management domain (Weiner et al., 2010; Burkhart et al.,

2011; Zott et al., 2011; Wirtz, 2013). Prime focus of the research have been taxonomies

and components of a business model (Hedman and Kalling, 2001). Even though the

concept has by now gained high attention, it lacks a common definition among the research

community. Osterwalder et al. (2010), for example, state that “a business model describes

the rationale of how an organization creates, delivers, and captures value”. Among the

many available ones, this definition is well received both in academia and practice

(Gassmann et al., 2013).


6

Recently, Veit et al. (2014) have identified the examination of IT-enabled or digital

business models as one core pillar for IS research related to business models. Several

scholars have attempted to define archetypes for such business models. For example,

Timmers (1998) commences with proposing and delimiting a variety of business models

for electronic markets. Rappa (2001) develops a taxonomy for business models on the web.

Pauwels and Weiss (2008) devote their research to variants of business models for online

content providers. Their particular interest is to explore the underlying pricing model as

well as the difficulty to switch from a “free” to a “fee”-based model. Also Clemons

(2009a) studies business models in the internet. He focuses on models with a non-

advertisement based revenue stream and distinguishes two main groups of business models

– one being the sale of virtual items, another one being the creation and access generation

for customers. In a second paper Clemons (2009b) also propagates the difficulties of

monetizing social network content. Wirtz (2001) and Wirtz et al. (2010) define four types

of business models (content, commerce, context, connection) to describe activities in

online markets in the B2C area.

In summary, a number of different approaches exist to categorize the way how business

models can be defined within the realm of electronic markets. Some of them also consider

the types of products being offered. However, not much light has been shed on the role of

information as such as a resource in business models – regardless of whether in the context

of electronic commerce or offline business environments. As information has been

recognized as a vital strategic asset (O'Reilly, 2005), it is deemed appropriate to take a

closer look at the effects on business models resulting from an increasing availability of

information. This directly leads to the first research question.

Research Question 1: What different roles can information play in business models? How

does the role of information affect the characteristics and components of a business

model?


7

Obviously, the profitability is a pivotal evaluation criterion for any business model. This

naturally accounts as well for business models that base their value proposition on

information. As information constitutes the key resource, its (economic) value and cost

structure heavily influence the overall profitability. With regards to value and cost,

information possesses the following characteristics that clearly distinguishes it from

physical economic goods (Shapiro and Varian, 1999; Stähler, 2002; Picot et al., 2003;

Krcmar, 2005). The determination of the value of information is difficult, since

information objects are experience goods. This means that an inspection prior to purchase

is not possible, as the value would already devolve in doing so. Only the ex-post calculable

value can quantify a user’s willingness to pay from an ex-ante perspective. Moreover, the

value of information objects is closely tied to time. Their value may decrease over time, in

case their value driver is novelty (e.g. stock prices). Only in combination with additional

information, the original information object recovers in value. Information does not have

an absolute value, as various users can draw diverging benefits from it at different points in

time. The cost structure of information is dominated by the problem of the first copy. This

phenomenon describes that the unit costs for the initially created information good are

exorbitantly high, while the costs for each subsequent unit are vanishingly small. In

summary, the value assessment of information is very individual and context-sensitive.

Hence, the determination of the value of information requires taking a close look at the

specifics of a company’s business model, in which the information is utilized.

Research Question 2: What is the financial value of information in specific business

model contexts? How can this value be expressed and quantified?

1.3 Structure of Thesis

The thesis consists of three parts to tackle the two aforementioned research questions. To

answer the first research question, a taxonomy is set up to arrange archetypical roles of

information and describe their impact on the specifics of a business model. The second

research question is addressed through determination of the value of information in two


8

selected real-world business models. Figure 2 provides a view on the overall structure and

flow of the document.

Part I starts with an introduction to the topic of the thesis, its structure and the research

outline (chapter I.1). Subsequently, chapter I.2 recapitulates relevant building blocks,

namely the business model canvas and information value chain, which are used to develop

a taxonomy for the classification of business models. Chapter I.3 outlines the taxonomy

with its four classes differing in the role that information plays in each of them. Two

classes address the usage of information for company-internal optimization, whereas as the

other two classes focus on leveraging information to enhance the company’s value

propositions. Hence, the latter two classes describe so-called information-based business

models. Chapter I.4 ties up to this by summarizing two real-world case studies on

information-based business models, which constitute the nucleus of this thesis. The case

studies examine the role of information and its quantitative value contribution in two

different settings, namely personalized recommendations for online content and demand

response in electricity grids. The value drivers and business model characteristics of the

Figure 2. Structure of Thesis.

Part I: Thesis Overview

i. Taxonomy of Information-based Business Models

ii. Comparison of Case Studies on Information-based

Business Models

Part II: Case Study 1:

Personalized Recommendations

A) Impact of Personalized Recommendations

on the Business Model of an Online Content

Provider

Part III: Case Study 2:

Electricity Demand Response

A) An Information System Architecture for

Demand Response

B) Optimal Information Granularity in Demand

Response

C) Scenarios for Integrating Demand Response

in Electricity Markets

• Controlled experiment on the website of

an online newspaper

• Evaluation of impact of personalization

on observable user behavior and

financial KPIs of content provider

• Quantification of information value for

personalized recommendations

• Design of an information system

architecture for demand response based

on required information flows

• Cost-value-model for demand response

and evaluation in a configuration with

real-world data

• Calculation of optimal information value

and best frequency of meter readouts

• Determination of optimal smart meter

roll-out scenario for an electricity retailer

• Assessment of costs and benefits for

integration of demand response into

different electricity markets

• Development of a framework for

categorization of business models

according to the role of information

• Assessment and comparison of value

drivers and activities of value creation for

two selected case studies


9

two studies are carved out and compared to each other. Chapter I.5 concludes the first part

by providing an overall summary of the findings and an outlook on suggested future

research topics.

Subsequently, part II of the thesis delivers insights into the case study on personalized

recommendations. It describes the design and results of an empirical study conducted with

an online news provider. The study is designed as a controlled online experiment and

examines the impact of personalized recommendations on user behavior and consequently,

on key financial performance metrics of a content provider.

Part III elaborates on the second case study, which exposes and quantifies the value of

information for demand response in electricity grids. Part III.A focuses on the in-depth

design of an information system architecture that supports the required information flows

of a demand response mechanism. Successively, part III.B draws the attention to the

determination of the optimal information granularity in a demand response setup.

Moreover, part III.B provides a sensitivity analysis of major cost drivers and elaborates on

a scenario that maximizes the profit for the provider of a demand response infrastructure.

Ultimately, part III.C further extends the financial assessment and examines saving

potentials for an electricity retailer in three market scenarios, namely product market,

control reserve, and balancing energy.

2 Building Blocks for Taxonomy

This chapter describes the two main concepts that serve as basis for developing a

comprehensive taxonomy to classify information-based business models. First, a structure

to describe the specifics and typical components of a business model is required. For this

purpose, the business model canvas by Osterwalder et al. (2010) is chosen, as it has

reached the status of a de-facto standard both in academia and practice (Gassmann et al.,

2013). Second, deeper insights into the concrete value-added steps in processing

information in a business model are desired. Here, the information value chain concept is

used – a common adaption of the seminal work of Porter and Millar (1985) for IS research.


10

2.1 Business Model Canvas

The business model canvas approach by Osterwalder et al. (2010) supports explaining a

business model and illustrating its components. The business model canvas consists of nine

elements describing the value model (value propositions, customer relations, channels, and

customer segments), the operating model (key activities, key resources, and key partners)

and the financial model (cost structure and revenue streams) of a company (see Figure 3).

The following paragraphs provide a detailed description of each component.

Value propositions. A value proposition illustrates the company’s possibilities to satisfy

the needs of its customers by providing products or services. It highlights the unique

selling point or key differentiator of the offering by the company, and thus shows how the

company distinguishes itself from its competitors.

Customer segments. The company needs to identify the customer groups targeted by their

value proposition. Customer segments categorize these target groups based on their

differing characteristics and needs.

Customer relationships. For each of the customer segments, a company needs to identify

the type of relationship designated for it.

Channels. A variety of marketing and distribution channels is available per target

customer segment for the company to sell and deliver its value proposition. The channels

Figure 3. Business Model Canvas according to Osterwalder et al. (2010).

Key Partners

Key Activities

Key Resources

Value

Propositions

Customer

Relationships

Channels

Customer

Segments

Cost Structure Revenue Streams

Operating Model Value Model

Financial Model


11

serve different purposes during the lifecycle of customer interaction: they are used to

create and raise awareness about products, help customers to evaluate the company’s value

proposition, provide a means to purchase products, serve as fulfillment vehicle and provide

after sales support.

Key partners. In the business model concept, the relationship between a company and its

suppliers or distributors is denoted as a partnership. Partnerships allow for risk reduction

and supply chain optimization through close collaboration.

Key activities. The relevant steps required to create value in order to fulfill the company’s

value proposition are considered key activities.

Key resources. All factors, be they financial, human, physical, intellectual or

technological, which serve as input into the key activities to produce a value proposition

are called key resources.

Revenue streams. Each company needs to earn income from what it is doing. Revenue

streams is the collective term for the multitude of possibilities how to do so.

Cost structure. For the cost structure of a company, both variable costs, which directly

depend on the amount of produced goods or services, and fixed costs, which incur

independently of the production scope, need to be considered.

2.2 Information Value Chain

To describe the value-adding steps within a company, the concept of a value chain is well

established. The first prototype of the value chain has been defined by Porter and Millar

(1985) in the mid-eighties. They describe a firm as „a collection of discrete activities

performed to do business that occur within the scope of the firm”. Hereby the primary

activities, such as inbound logistics, operations, outbound logistics, marketing, sales and

service, are differentiated from supporting activities, for example the management of

infrastructure, human resources, financials, technology and procurement.


12

Rayport and Sviokla (1995) augment this approach and flank the physical value chain with

a virtual one. This virtual value chain is described as an „information underlay“, which is

established as information is captured within the various stages along the value chain. The

information-based activities mirror the physical value chain and provide the ability to

perceive the value chain end-to-end. A real value-add through information is not yet taking

place. During the times of the internet boom, Porter further develops his model and

introduces the internet value chain (Porter, 2001). It encompasses the IT support of the

physical value chain by means of automation as well as data exchange across processes or

even companies, with a focus on online sales channels. Many authors take Porter’s

approach as a starting point to further define information-based value creation (see e.g. Lee

and Yang, 2000; Choo, 2002; Spataro and Crow, 2002; Krcmar, 2005; Schwolow and

Jungfalk, 2009; Stocker and Tochtermann, 2009). Their definitions more or less match,

and can be summarized into four key activity fields, which are data generation and

acquisition, storage and transmission, processing and analytics and, finally, presentation

and distribution (see Figure 4). The four steps are described in detail in the following.

Generation & acquisition. The first step of the information value chain corresponds to

Porter’s inbound logistics. This step contains two activities – the internal data generation

as well as the external data acquisition (Schwolow and Jungfalk, 2009). Data can be

generated manually or be the result of automated data capturing. The acquisition of data

encompasses the transmission of data from suppliers and vendors, the purchase of data

from partners as well as the access to public (open) data or user-generated data

(Choo, 2002).

Figure 4. Information Value Chain (based on Lee and Yang, 2000; Choo, 2002; Spataro and

Crow, 2002; Krcmar, 2005; Schwolow and Jungfalk, 2009; Stocker and Tochtermann, 2009).

Generation &

Acquisition

Storage &

Transmission

Processing &

Analytics

Presentation

& Distribution


13

Storage & transmission. Data can be either granular or complex and can come in various,

mostly heterogeneous formats. Before the data can be stored persistently, transforming the

data through standardization and normalization into a structured scheme and thus into a

machine-readable format, is required. Besides storage, the transmission of data is the

second key component of this value chain step. In times of an increasing internet and

mobile pervasion, highly performing networks are abundant. Consequently, the complexity

of transmission drops.

Processing & analytics. Data processing can relate to a number of activities in which data

are manipulated or synthesized (Schwolow and Jungfalk, 2009). Examples for such

activities are data cleansing to increase quality or aggregation of data from different

sources to decrease the level of detail and consequently the search costs (Madnick and

Siegel, 2002; Spataro and Crow, 2002). Analytics refers to the processing of data involving

operations of higher complexity: analytics can generate predictions, apply collaborative

filtering mechanisms, execute complex mappings between data sources, analyze data for

personalization based on specific behavioral patterns, and produce recommendations. The

scope of operations reaches from single, isolated data to large amounts of them.

Nowadays, both processing and analytics in (near to) real-time are possible, often on large

amounts of data – big data.

Presentation & distribution. In the last step of the value chain the information is prepared

for presentation and distribution towards the user (Spataro and Crow, 2002). For

distribution and marketing of information content, a variety of digital channels can be

used, such as app stores, specialized hardware (e.g. the Apple iPad) and social

communities (Schwolow and Jungfalk, 2009). A positive trait of digital distribution is its

speed, which allows for creating and sending information to the consumer on request,

without a significant delay. To satisfy user demands, it is important for the content to be

formatted, edited in layout and adapted to suit the used output device as well as the user’s

preferences. To leverage the direct user interaction for subsequent value creation activities,


14

feedback elements can be incorporated into the presentation layer. Well-known

mechanisms in this respect are “like” buttons and rating systems. These generate further

data and allow feedback to the company. The distribution of information happens at

virtually no cost. Schwolow and Jungfalk (2009) summarize the contribution of the last

value chain step as follows: “Information distribution adds value based on the principle

‘information is expensive to produce, but cheap to reproduce’”.

3 Taxonomy of Information-based Business Models

Having the business model canvas and the information value chain at hand, this chapter

starts with the introduction of a framework helping to classify the role that information

plays in a business model. Roles and influence of information on the different types of

business models according to the four classes defined within the framework are described

in detail thereafter.

3.1 Classification Framework

The classification scheme to be introduced in this section provides the framework to

answer the first research question on the role of information in different types of business

models. Generally, two groups of companies can be differentiated: those, that use

information solely as a residual within the company, and those, that make information an

integral part of their market offering. Strikwerda (2011) denotes this split as the “economy

of two velocities”. Within the first group, the usage of information is highly linked to using

IT, meaning the technology to process information. Loosely speaking, not all companies in

this group can be painted with the same brush. There are the more traditional companies

drawing value from using information to optimize transactions, referred to as class 1 in the

taxonomy, and the ones being “more than simple number-crunching factories”

(Davenport, 2006). The latter smartly apply analytics to increase the quality of strategic

business decisions and are denoted as class 2. The effect that information has on the

business model of class 1 and 2 companies is more or less limited to the components of the


15

operating model, namely resources, activities and partners. In order to let information play

a vital role in the value model and have it directly influence the revenue streams, it has to

be brought to the marketplace in form of an information product (Redman, 2008b, 2008a).

Having crossed the limits of the company itself, information is now part of the value

model, encompassing the components value propositions, customer relations and channels.

However, the step from a sophisticated, yet internal class 2 usage of information towards

selling a pure-play information product might be too big a leap forward, being associated

with a new risk: the company’s success stands and falls by the quality of its information

and the value creation out of it. A sensible intermediary step thus seems to add information

as a complementarity to the traditional (physical) product and “informationalize” it

(Redman, 2008b). The business models creating such a hybrid offering (Möslein and

Kölling, 2007; Ramaswamy, 2008) are categorized as class 3 in this framework, business

models encompassing standalone information products as class 4. Both classes 3 and 4

exhibit a crucial dependence on information as a key production factor, and can thus truly

be coined with the term information-based business models. All four classes together build

a taxonomy to categorize business models according to the role attributed to information:

Class 1: Process information for transactional optimization

Class 2: Apply analytics to discover sweet spots

Class 3: Leverage information to upgrade offering

Class 4: Design a pure-play informational value proposition

These generic classes already allow for categorization of virtually any business model.

Each class is unique in terms of the role that information plays for the business model as a

whole. However, the specifics of the business model are not taken into account so far. To

proceed accordingly and look beneath the surface of the business model, the business

model canvas by Osterwalder et al. (2010) is applied. The canvas serves as a means to

further elaborate the classes and carve out how information affects different components of


16

the business model. Figure 5 illustrates this by highlighting which components are affected

for each class and how the influence of information on each of these components can be

characterized. From class 1 to class 4 the overall impact of information grows as more and

more components are influenced by information usage.

The following sections explain the four classes and the respective impact of information in

more detail. They explicitly highlight how information is applied to create value in each

case. For this, the information value chain is referred to as the prime tool within the key

activities component of the business model canvas.

3.2 Class 1: Process Information for Transactional Optimization

Class 1 describes a business model, which leverages information as a resource to increase

production efficiency. The usage of information in this context is sometimes not even

driven by internal needs, but rather triggered by external factors, such as regulatory

Figure 5. Taxonomy for Categorizing Business Models According to the Role of Information.

Class 4: Design a pure -play informational value proposition

Impact of

information

on business

Constant repetition

Dynamic data

Own

ecosystem

Purely

information

Prosumer

Digital only

Full

personalization

Tailored pricing

Key partners Key activities

Key resources

Value propositions Customer relationships Customer segments

Channels

Value of information

Class 3: Leverage information to upgrade offering

Class 2: Apply analytics to discover sweet spots

Class 1: Process information for transactional optimization

Information

as part of

the offering

Information

used

internally

Feedback loops

Product usage data

Established

ecosystems

Information

as add-on

Information sphere Partial

personalization

Customer lock-in


Key resources

Value propositions Customer relationships Customer segments


Analytics

“Big” data

Data

provider

Granular

segmentation

Increase of steering quality


Key resources

Customer segments


Aggregation

Transactional data

Operational optimization

Key activities

Key resources



17

requirements (e.g. tax legislation) or demands by business partners (e.g. standardized

automated interfaces for electronic data exchange).

Key resources. Besides the traditional production environment (i.e. personnel,

machinery), information can be regarded as an important resource, even in this most basic

class. The types of information mainly originate from business and production processes,

or from interactions with customers and suppliers, such as order data. The communality of

these information types is their transactional nature.

Key activities. Typically, the information relevant for class 1 is collected from the

operational context. It is structured and stored in information systems, to serve enterprise

resource planning (ERP), supply chain management (SCM), or customer relationship

management (CRM). With the information stored and made accessible in a structured way,

step two of the information value chain makes up the bulk of the value creation for class 1.

Though the focus is not on sophisticated analytics, a basic reporting mechanism is usually

part of the environment and often realized with a management information system (MIS).

The MIS aggregates data extracted and summarized from the connected operational

systems, and thus allows the management to draw conclusions and support operational or

strategic decisions (Haag and Cummings, 2013).

Value of information. Information usage within class 1 aims at improving the planning,

coordination and controlling capabilities of a company. Information is applied to optimize

the efficiency of internal processes, but also external ones to business partners and

customers (Glazer, 1993). Examples for such a gain are improved manufacturing and

inventory management, increased proximity to suppliers, customers and clients (e.g.

through electronic data interchange (EDI)), and reduced operational costs (Lucey, 2005).

In order to derive a value of information in this context, transactions are the unit of

measurement to be considered (Glazer, 1993). Transactions describe the exchange of

goods and services for money between the company and its customers. The information


18

value is visible at both the revenue and the cost side – in form of profit resulting from

higher revenue as well as through reduced production or operation costs.

Certainly, using the resource information comes at a cost – in class 1 these include the

expenses required for production, storage and processing of information, which are usually

considered IT costs, but also staff costs for analyzing the data and defining, populating and

maintaining management report structures.

3.3 Class 2: Apply Analytics to Discover Sweet Spots

The core of class 2 is the ability to connect internally generated and externally acquired

data in a smart way to allow for more informed decisions, resulting in competitive

advantages. A survey conducted by PricewaterhouseCoopers (2009) reveals, for example,

that 94 % of participating CEOs consider information about their customers’ and clients’

preferences and needs important or critical to long-term decision-making. Information is

recognized as a strategic asset, and used accordingly. Businesses are starting to

comprehensively mine their data and conduct in-depth analytics. Already at the turn of the

century, studies have shown that business intelligence can be a key success factor (Choo,

2002). Davenport (2006) states that „organizations are competing on analytics not just

because they can – business today is awash in data and data crunchers – but also because

they should”. Looking at the market for business intelligence and analytics services

confirms this trend, as its growth rate outperforms the overall software market

significantly. It is not surprising that big software companies, such as IBM, SAP,

Microsoft and Oracle, spent several billion US$ over the last years to acquire rising stars in

this market segment.

Key resources. Data relevant for class 2 are manifold – they originate from all stages of a

company’s value chain, from sourcing to distribution. More and more new data sources are

tapped. Especially with the internet of things, a number of areas, so far restricted to the

physical realm, are now deemed relevant for information gathering. Equipping elements


19

along the supply chain with sensors, such as warehouses, vehicles, pallets and items,

allows for real-time analytics and immediate action (IBM Global Business Services, 2011).

Due to specialization and distribution of activities on multiple parties along a value chain,

not all of the desired information is readily found within the boundaries of the company,

but has to be acquired from business partners. Also externally available data with limited

relevance at first glance, such as demographic, meteorological or financial data, are

integrated with the data at hand.

Besides information as such, class 2 also demands for a specific employee skill set. The

profiles of data analysts and scientists will be sought after increasingly in the coming years

(Davenport and Patil, 2012). These specialists are able to develop, configure and populate

analytical models and interpret the outcomes as basis for providing meaningful decision

support on management level.

Moreover, an appropriate IT infrastructure and algorithms are a prerequisite to efficiently

and effectively conduct business data analytics. This applies especially to large data sets,

which are classified as big data. Big data is the term for a collection of data sets so large

and complex that it becomes too difficult to process them using on-hand database

management tools or traditional data processing applications. But they offer almost

unlimited possibilities to discover surprising new, multidimensional insights and

correlations (McAffee and Brynjolfsson, 2012). The criteria volume, velocity and variety

are commonly applied to characterize data sets as big. Fueled by the hype surrounding big

data, new trends and technologies are emerging rapidly, such as in-memory databases.

These are able to process increasingly large data sets in a decreasing amount of time,

sometimes as fast as real-time.

Key activities. It becomes evident, that the analytical part of the information value chain is

the core activity for class 2. Business analytics can be categorized into three classes

(Evans, 2013). The most basic class, descriptive analytics, also referred to as reporting (see

Sharda et al., 2013), borrows methods developed in the area of descriptive statistics. Its


20

outcomes are based on standard aggregate database functions, such as counts,

aggregations, sums or averages, often combined with grouping or filtering. Wu (2013)

claims this type of analytics to be the prevailing one these days acting as basis for

management decisions. The second, more advanced class is denoted as predictive

analytics. It utilizes a number of statistical models and other empirical methods, data

mining and machine learning techniques to study and integrate recent and historical data,

to form the basis for extrapolation of the future (Shmueli and Koppius, 2011). Predictive

analytics helps in creating predictions and detecting patterns in large quantities of data that

are not readily apparent with traditional, i.e. descriptive, analyses. It enables so-called

what-if analyses. In connection with customer data, they offer ways to monetize customer

needs and preferences. An example is the electronics retailer Best Buy, who equips its

sales force with data of past customer behavior to enable suggestions of suitable add-on

purchases (IBM Global Business Services, 2011). The third class of business analytics is

prescriptive analytics. This type of analytics is used to apply deterministic and stochastic

optimization approaches to identify the best solution with the goal of maximizing or

minimizing a certain objective. For this, current trends and forecasts are used to calculate

the expected outcomes for different choices of action. The operations research community

has been pushing this class of analytics for quite some time now (Sharda et al., 2013).

Key partners. As stated earlier, the sources for information in scope for class 2 are often

found outside the company’s boundaries. Third parties, whether specialized providers for

data or public sources, are thus required as key partners for a company operating on a class

2 business model. Partners are manifold: Governments are for instance starting up open

data initiatives, such as the Dutch government offering insight into currently issued

permits. Another well-known example is Google Flu Trends, which uses its search results

to forecast the spread of flu epidemics and make this data publicly available.


21

Customer segments. Analytics generate possibilities for a more granular segmentation of

the customer base. The combination of sales data with predictive methods yields potential

for tailored campaigns or mass customization.

Value of information. Class 2 aims at using analytically enriched information as basis for

an increased steering quality, as business decisions can be based on facts rather than gut

feeling or personal experience. Superior decision-making is, next to decreasing operational

costs, regarded as a valid means to create a competitive advantage (Porter and Millar,

1985; Marchand et al., 2002). Information in this case serves as a lever to partly take away

the uncertainty inherent to any decision. The value of information entirely correlates with

the part of uncertainty it is able to extinguish. Krcmar (2005) proposes three variants to

measure the value of information in the context of uncertainty: normatively, realistically

and subjectively.

Following the principle of opportunity costs, the normative information value compares

the expected profit following a decision before and after information provision. It is as also

known as the expected value of perfect information. An extension of this variant is the

dynamic information value, which considers the delay resulting from the time required to

obtain the information. If the options are evaluated for too long in hope for an improved

informational basis, other market players might make a quicker move. The practical

limitation of the normative information value is its ex-post perspective, as the value of the

decision with information cannot be anticipated.

The realistic information value is defined as the empirically measurable profit arising for

the decision maker when using the information. The information value is calculated from

the value of the actions triggered by the information. Actions are for example quantified by

the resulting gains, response times or precision. The realistic information value should only

be used if relatively easy to determine.

Lastly, the subjective information value tries to quantify the gut feeling of a decision

maker. For this, decision makers are asked for their personal opinion on the value of the


22

information. As the name suggests, the resulting value has to be taken with a grain of salt,

as determined by the statements of one or a limited number of persons only.

There are numerous areas to apply analytics in the business context. Davenport and Harris

(2007), Rosenberger et al. (2009) and IBM Global Business Services (2011) describe

typical scenarios where information is used to support decisions on a set of options: How

are products optimally designed to meet desired quality levels or efficacy targets? What is

the best setup for supply chains with regard to e.g. optimal inventory levels or alternative

shipment routes? Who are the most profitable customers, where is cross- or upselling

promising, where to invest in customer loyalty measures? How high should the real-time

price be set in order to obtain the highest possible yield from each individual customer

transaction?

The listed examples show that the value of information can, in abstract terms, be

summarized as the generation of a competitive advantage. This advantage can materialize

on the cost side as well as on the revenue side. Cost savings can be realized e.g. through

more efficient warehousing, just-in-time production or cost avoidance due to early

mitigated risks. In terms of revenue, the individualization of offerings tailored to

customers’ needs and exhausting their real willingness-to-pay to create additional,

otherwise untapped sales, might be the most promising avenue.

The cost of analytics is two-fold: personnel and technology, analogously to the described

key resources. While technology to unlock value from high volumes of information is

becoming mainstream at acceptable costs, the highly skilled personnel required to attach a

purpose and meaning to this information, is relatively scarce (Davenport and Patil, 2012).

Not all of these resources must necessarily be held available within the company, as

sourcing expertise, algorithms or computing power on-demand from the external market is

possible.


23

3.4 Class 3: Leverage Information to Upgrade Offering

While in class 1 and 2 information is used to optimize within the company itself, it

advances in class 3 to become an explicit part of the customer value proposition and as

such participates in the marketable offering of a company.

Value propositions. Even though information now determines certain elements of the

value proposition of a company, its core is still based on a physical product or service. This

basis is augmented with an information sphere. Zott et al. (2011) describe complementarity

as a key success factor in business models; the extension of a physical base product into a

hybrid offering by means of an information sphere can be regarded as a special form of

complementarity. The important prerequisite is that the customer perceives the information

sphere as value-adding.

Prominent examples of business models based on class 3 are Nikeplus, Fiat eco drive and

DHL Track&Trace. Nike equips their customers with motion sensors, embedded in their

shoes or available as a separate bracelet. The sensors track the movements of its wearer,

including values such as acceleration. Users can then share the data on a website with a

global community, embark on competitions or engage coaches to reach goals they have set

for themselves. A similar model has been started by Fiat. Data collected while driving can

be shared in a community. Also here, gamification is used to motivate users to adjust their

behavior – in this case towards a more economic, fuel-saving driving style reducing

emissions. At the very least, users are repeatedly confronted with the product and are thus

forced to think about it regularly. DHL – one example of many logistics operators with the

same practice – enables its customers via track and trace to find out exactly when a

shipment will be delivered. On top of that, dynamic re-routing to other destinations, such

as a retail outlet or a preferred neighbor, as well as selecting a more preferable delivery

slot, is offered to the customers.

Customer relationships. Personalization using the company’s information sphere can

create a strong lock-in effect with customers, which raises their switching costs in case


24

they want to exit the system in favor for another. Amit and Zott (2012) identify the lock-in

effect as a key value driver for business models. Offering a community to share the usage

of the (physical) core product further intensifies the lock-in and creates a network effect.

This effect implies that the more people join a community, the more value is attributed to

the community and the corresponding core product.

Selling a physical product to a customer is a one-off activity. Usually only warranty cases

trigger a further point of contact, and they are mostly handled via retailers, without

participation of the original equipment manufacturer (OEM). Hybrid models of class 3

manage to mitigate this by creating a continuous relationship between the OEM and the

customer, thus leapfrogging various intermediaries in the distribution chain. In this way,

the OEM receives first-hand information about the usage of its product. By activating the

co-creation paradigm, the OEM can entitle the user to configure and adapt the information

sphere according to own preferences. With this, the company closes the customer feedback

loop and obtains regular signals regarding requirements and desires of its customers

(Loukides, 2010). This reflexive relationship with the markets also offers the possibility to

constantly improve the physical product as such – the cycles of change are getting faster

and are more closely linked to customer demands.

Customer segments. The mentioned personalization also affects the customer

segmentation. The options of personalization are much broader than as described in class

2, as the configurable information sphere allows for very individualized offerings, which

are permanently refined based on the ever-increasing amount of collected product usage

data as well as explicit customer feedback.

Key resources. Obviously, a valid data basis is a prerequisite to create a compelling value

proposition. This data basis contains internal and external data sources, where user-

generated content makes up the bulk. Following the quite narrow definition of the OECD

(2007), data in this category are characterized by being made publicly available through

the internet, boasting a certain level of creativity and being created outside of professional


25

practice. In scope are thus texts, photos, videos and postings in social media channels by

private persons. In an extended scope, also information generated by users through their

behavior can be subsumed under user-generated content. Sensors (e.g. Nike Fuelband) or

mobile phones are common devices making it possible to track and transfer collected data

to the provider. Loukides (2010) calls this a “trail of data exhaust that consumers generate

and that can be mined and put to use”. Further examples of user-generated content are

movement profiles and behavioral patterns, e.g. electricity usage profiles or clickstreams

on websites.

Besides the data, organizational skills are required to embed the mind shift in the company

that information is not solely relevant internally, but also needs to be externalized as part

of the offering. Nolan, Norton & Co. (2014) fault the fact that organizations are usually

spreading information ownership and handling across several entities, and plead for a

“corporate information superiority” where commercial, technical and analytical aspects

are holistically looked upon.

Key activities. Similar to class 2, also in this class data analytics is the prime focus along

the information value chain. However, the purpose for which data are analyzed, aggregated

and distributed is different – direct external commercialization, as opposed to internal

optimization. Actively including the customer in the value creation is an integral activity,

as discussed extensively with respect to the key term user-generated content. To achieve

this participation, a transformation of the traditionally sequential steps of value creation

into a cycle is necessary, such that collected data of product usage and user feedback can

serve as input to create new value-added services in the next iteration of the value chain.

This enables users to experience a feedback loop and recognize their contribution on

products they use (Loukides, 2010).

Key partners. Partnerships on various levels are essential for the provisioning of

information-based business models. First of all, companies should strive to become part of

considerable ecosystems with their information sphere. In the B2C realm, these ecosystems


26

are the large and well-known social media networks such as Facebook, Twitter and

Instagram as well as app markets (e.g. iTunes, Google Play). The most effective form of

interconnection is reached when not only data generated in the ecosystem are used, but

also company-centric data are fed back to the ecosystems, thus creating a bidirectional

information flow. Web mashups are another interesting form of enabling partnerships and

data exchange (Yu et al., 2008). They are characterized by their ability to quickly and

dynamically ramp up and down as required. Mashups utilize open application

programming interfaces (API) and data sources to create value added services. Zhu and

Madnick (2009) highlight that the innovative reuse of data through mashups yields great

opportunities. Yahoo Pipes is an example for the power of mashups – using a simple

“drag’n’drop style”, data feeds from various sources can be easily filtered and customized.

Value of information. The value of the information in class 3 is tapped through

establishing a continuous, long-term customer relationship. This generates a high customer

satisfaction and retention due to the lock-in effect and the perceived personalization

achieved through the information sphere. In the long run, these effects lead to competitive

advantages and thus revenue increases.

A prime example for a successful information sphere is the previously described Nikeplus

offer. Its community counts more than 18 million users by now1 and apparently, Nike was

able to increase its share on the US running shoe market by 14 percentage points in the

subsequent three years after the introduction of their hybrid product2. Ofek and Wathieu

(2010) state that “for a Nikeplus customer, the Nike brand is no longer about just the

product attached to his or her feet; it’s about the total exercise experience, including the

community”.

In order to provide an information sphere, investments are required to build up and

maintain a respective infrastructure. Marketing and ramp-up support is needed to get the

1 http://nikeinc.com/news/nike-evolves-just-do-it-with-new-campaign/, Accessed 2 February 2014 2 http://de.slideshare.net/EnterpriseCoCreation/nike-8829199, Accessed 2 February 2014

http://nikeinc.com/news/nike-evolves-just-do-it-with-new-campaign/

http://de.slideshare.net/EnterpriseCoCreation/nike-8829199


27

community going and reach a critical mass. If users are successfully encouraged to deliver

the majority of content, which comes at virtually no cost to the provider, the run costs of

such a community are largely reduced to IT operations costs.

3.5 Class 4: Design a Pure-Play Informational Value Proposition

In this class, information is at the core of the value proposition. As most information is

permanently renewed, its value is tightly interconnected with its up-to-dateness. Therefore,

the ability to align production and marketing to the speed of information stands in the

foreground of the underlying business model.

Value propositions. Physical products now play no or at the most a very subordinate role.

Thus, one can speak of stand-alone or pure-play information products and services.

Loukides (2010) defines an information product as self-nurturing: “A data application

acquires its value from the data itself, and creates more data as a result.” This extended

business impact of information in comparison to class 3 means that the provided

information as such has to have a value for the customer. Redman (2008b) has developed a

number of strategies that strive to create a substantial value proposition for information-

based products3.

The first strategy is to provide new content. It is supported by the works of Amit and Zott

(2012), who describe novelty as a key value driver for business models. Thus, the goal of

this strategy lies in the pursuit of creating new, richer or more targeted data in order to

distribute them quickly. Both raw and processed information can be marketed (Strikwerda,

2011). The relevance of novelty becomes clear when considering the dynamics of data. If a

value proposition is solely based on a static set of base data, problems are likely to

eventually occur if the data are not owned. Competitors are encouraged to license the base

data as well and offer similar, competing services. MapQuest is such a negative example.

They were the first ones to offer online maps in 1995, but were soon overtaken by Google

3 Remark: Redman´s strategy „informationalize“ classifies as a class 3 element in the taxonomy of this thesis, and is therefore not mentioned here again.


28

and Microsoft who based their offerings on the same underlying data (O'Reilly, 2005). To

avoid this issue, it is advisable to add buzz to the static base data. With this, the data set is

continually renewed and thus remains valuable. For example, Optasports rest their offering

upon basic data of soccer players (e.g. name and age). They permanently overlay these

data in a dynamic way with detailed real-time game statistics per player, such as assists,

goals or successful passes.

The second strategy is repackaging of data. Tim Berners-Lee states it as follows in a 2004

interview: “The exciting thing is serendipitous reuse of data: one person puts data up

there for one thing, and another person uses it another way” (Frauenfelder, 2004).

Repackaging thus means to enrich data sets by using data generated by others and

supplement it with own functionality (Zhu and Madnick, 2009). A prominent example for

this is Google News, which summarizes various news sources into an aggregated news

feed.

Unbundling is the third strategy. Here, information is extracted from existing products and

is sold separately. WebKinz, a manufacturer of stuffed animals, delivers a web code along

with its product and lures children onto its website where they can feed and groom

electronic versions of their pets. Selling aggregated versions of the online collected data on

user behavior serves as an additional source of revenue for WebKinz.

The fourth and fifth strategies utilize information asymmetries. If there is an understanding

on potential differences on perceived value of products or services, these can be explicitly

addressed and leveraged. Examples for this are hedge fund managers who exploit

information superiority to create profit from arbitrage transactions. On the contrary,

information asymmetries can be levelled out by bridging the information gap of the less

informed party in a transaction. For example, Immobilienscout provides an online platform

for others to offer apartments and real estate for sale or rent. The platform operator

analyzes the transactions and supplies its users with e.g. resulting average prices per

geographic area, which enables users to set prices accordingly before offering.


29

The sixth strategy is about providing identifiers. Companies applying this strategy define

unique labels and strive to obtain a monopolistic situation within their niche. Examples are

the identities managed by GS1, which uniquely identify articles on sale worldwide. For

this, GS1 licenses numerical ranges and provides the mapping of identities to products.

The identifiers are mostly converted into barcodes, which are printed on the products.

The seventh strategy is called infomediation. The central aspect of this strategy is playing

the role of a broker, which increases the user’s value perception through better guidance to

relevant information. The German company Creditreform for example takes the role of an

infomediary. It aggregates data on enterprises from a variety of public and non-public

sources and based on those allocates credit ratings to enterprises. These can access the

results against the payment of a fee.

The eighth strategy is about data mining and analytics. The activities largely coincide with

the ones described in class 2, which aim at extracting new, valuable insights from data. In

contrast to class 2 though, where these insights are primarily used to optimize own

production or revenue streams for existing products, class 4 goes further and sells these

insights as a standalone product or service. These insights then help the customers to take

better decisions (Zhu and Madnick, 2009). For example, the search engine of Microsoft

Bing Travel bases its value proposition on the evaluation of 225 billion records of flight

prices, and is thus able to give a recommendation on the optimal time to purchase an

airline ticket.

Independent of which strategy, or a combination of them, is selected, the decisive factor is

the ability to carve out a unique selling point. The unique selling point has to clearly

display the superiority of the own information product over those of competitors. This can

be achieved by the classical generic strategies, namely price or quality leadership. The

quality advantage is obtainable through speed, i.e. up-to-dateness of the information, the

quality of the information itself or the usability of information.


30

Customer segments. Pure-play information products are capable to achieve a very high

level of personalization. This level even surmounts the one obtainable with a class 3

strategy, since it is sufficient to adapt the digital product, which is by far easier than doing

the same for a physical product. With this at hand, the stage is set to approach customers

with completely individualized offerings.

Customer relationships. Similarly to class 3, it is vital to establish and keep a very

intimate link to the customer, motivating the customer to become a prosumer of data. A

prosumer is someone who acts both as producer and as consumer (Ritzer and Jurgenson,

2010).

Channels. Nowadays it is possible to limit the number of distribution channels for

information-based products to a single one, the internet, as it is more or less ubiquitously

available. Physical distribution channels can be completely neglected. Increasing attention

should be given to standardized and easy-to-use interfaces, with which customers and

partners can quickly and simply gain access to the offerings. This relates back to the

concepts of mashups and open APIs described earlier for class 3. Apart from that, data

markets emerge as a platform where providers offer their data for purchase (Balazinska et

al., 2011; Kushal et al., 2011). On Windows Azure Data Marketplace for example, house

price data in Wales, historical weather data and sports statistics can be acquired.

Key resources. The core resource in a class 4 pure-play information value proposition are

up-to-date, dynamic data. Thus it is essential that such data are available within the

company, be it through creation or acquisition via external sources. Besides this raw

material, a suitable technical and organizational infrastructure is required to process the

data, as described in classes 2 and 3. As said, for many pure-play information products the

publicly available internet infrastructure is sufficient for an effective communication and

distribution. However, there are scenarios that require a proprietary information and

communication infrastructure. Examples for such scenarios are services that run over an


31

advanced metering infrastructure, which can only partially be realized with the public

internet as a basis.

Key activities. The information used for a class 4 strategy has to “live”. This means that

the information value chain has to be passed through continually in a cyclic manner, in

order to generate new data, which can be integrated with existing data to maintain the

quality. Krcmar (2005) confirms that by noting that the information value chain should not

be regarded as a one-time sequence of value-adding steps, but rather as a cycle, in which

the steps are iterated. This holds especially true if customers of the information product

themselves create further data in the realm of the base product (e.g. user-generated content,

clickstream paths), which are played back and integrated into the database (Rosenberger et

al., 2009).

Key partners. To increase the buzz on the own data, a high usage of the data by as many

partners and in as many ecosystems as possible has to be enforced. At best, the provider is

able to build up an own ecosystem that attracts additional partners to join and locks them

in. Interfaces (e.g. open APIs) which allow for a simple and fast access from outside both

in technical as well as contractual terms, cater for the wide dissemination. This is the only

valid means to not isolate the data and have them sitting idle for a too long time,

perpetually losing value.

Value of information. While in class 3 the informational component is mostly given as a

free add-on on top of the purchased base product, this is no longer possible in a class 4

environment, where the information product is the only source of revenue. Thus, it is a key

question how to capitalize on the value proposition and price the information product. As

mentioned, information has a number of characteristics influencing its value. Most

importantly, information value usually declines as time progresses, and information is an

experience good. Thus, the value materializes only when starting to use the information

and the perception of value varies largely between individuals.


32

Literature reveals a number of approaches towards pricing information goods (e.g. Bakos

and Brynjolfsson, 2000; Jain and Kannan, 2002; Pauwels and Weiss, 2008; Linde, 2009).

Predominant are the following ones:

Usage-based. Also known as pay-per-use, where the user pays a fee directly linked to

the exhibited usage pattern.

Subscription. The user needs to pay a regular subscription fee to receive access to the

respective product or service. This model becomes increasingly popular with online

newspapers, which introduce so-called paywalls to divide free from premium content.

Advertisement. In this model, the user may access a product or service free of charge.

The provider recovers its costs by showing advertisements sponsored by third parties.

License. The user pays a one-off fee to acquire usage rights for the product or service.

Donation. This is a special model where a provider counts on voluntary financial

support by its users. A prominent user of the donation model is Wikipedia.

Freemium. In this model, a basic service is offered free with the option to upgrade.

The part of the offer, which is free, can be time limited, feature limited, seat limited or

customer type limited.

4 Case Studies on Information-based Business Models

The previously developed taxonomy provides a generic framework to describe the

different roles of information in business models. In order to demonstrate its practical

applicability and carve out different patterns of value creation, the framework is populated

with a sample of two real-world business models. The sample encompasses a class 3 and a

class 4 business model, for which case studies are conducted. The studies aim at

quantifying the value of information in monetary terms by analyzing the cost structure as

well as the revenue streams of the respective business model.


33

The following two sections provide an overview on setup, characteristics and results of the

studies. Finally, the validity of the taxonomy is shown by comparing the specifics of the

two business models with a focus on the drivers of the value proposition and the

underlying value-adding activities.

4.1 Case Study 1: Personalized Recommendations

This empirical case study revolves around the business model of an online content

provider, more precisely, an online newspaper provider. In this area, personalized

recommendations increasingly gain in importance. Meanwhile most of the online content

providers, similarly to online retailers, use recommendations on their pages (e.g. YouTube,

Google News, New York Times). In light of class 3, these recommendations complement

the base product, the news article as such. Recommendations in this scenario are provided

to the user by means of a message along the lines of “if you liked this news article, you

might also be interested in the following one”. In order to personalize such

recommendations, recommender systems observe the user behavior and deduct the user’s

preferences from it. User preferences are derived both from the behavior of a single user in

relation to that of other users (i.e. collaborative filtering), as well as from the visited

content (i.e. content filtering).

The case study examines the impact of personalized recommendations on the revenue

stream of an online newspaper provider. In this case, the revenue stream is generated from

displaying advertisements, which are paid for by the advertising parties. The examined

newspaper provider, a major regional player in Germany, operates a website offering a

broad portfolio from global to local news across various categories. Users can access the

website freely, without a paywall or prior registration.

The clickstreams of around 30,000 website users with more than 100,000 clicks within the

experimental observation period form the basis for the statistical testing of the formulated

research hypotheses. A controlled online experiment is chosen as a suitable scientific


34

method to directly establish a causal context between the recommendations and the user

behavior. The controlled experiment randomly assigns first-time website users to one of

two groups. Depending on their group affiliation, the users then receive either genuinely

personalized or randomly generated links to other news articles as recommendations. The

users do not know that they are part of an experiment – both types of links resemble each

other regarding their layout.

In contrast to many other studies in this field, which focus on the user experience of

recommendations, the clear target of this study is to examine the financial impact of

personalized recommendations from an online content provider perspective, both on the

cost and the revenue side.

The results of the study confirm an increased attractiveness of recommended links, when

truly personalized, over standard links. The higher number of clicks as such would already

lead to higher revenue for the content provider, as each click yields a further display of a

paid advertisement. From a relationship-building perspective, online content providers

moreover strive towards a more long-term bond with users, making them frequent and

long-staying visitors on their website. The study also shows a positive impact of

personalized recommendations on the two associated metrics, which measure frequency

and duration of website visits. With a frequency increased by 15 % and a more than 3 %

higher average duration, the positive value of the information inherent in the

recommendations can be derived. The examined scenario yields a substantial yearly

revenue surplus of around € 0.3 M through personalized recommendations for the online

newspaper provider. This should offset the cost for running and licensing the recommender

software. Besides increasing the revenue, personalized recommendations are shown to also

positively influence the provider’s cost base. Through the recommendations, the provider

can direct the users to specific new articles and thus diminish the breadth of the accessed

item portfolio. In the scenario at hand, genuine recommendations result in a more


35

concentrated clicking pattern on the more up-to-date articles of the portfolio. The provider

could therefore delete outdated articles more quickly to save storage costs.

4.2 Case Study 2: Electricity Demand Response

The second case study analyzes a business model in the booming realm of smart power

grids. With a growth in renewable energies, it becomes increasingly important to balance

electricity production and delivery with efficient electricity markets. Enabled by

information technology, so-called demand response mechanisms recently made their way

into the market. They are used to manage demand side resources, i.e. shifting the

electricity demand according to fluctuating supply. Such an application scenario requires

the setup of large information networks in conjunction with a sophisticated communication

infrastructure. Opportunities for new service providers are emerging due to the immense

investments and operating costs of such environments.

This potential is analyzed in more detail in the study. Key player is an information

intermediary, which acts as an aggregation service provider by reading out smart meter

data and providing them to an electricity retailer in a condensed way. Moreover, the

intermediary also transfers the resulting load shift controls to the connected consumers. As

per the definition on the types of information-based business models in this thesis, the

intermediary acts on a class 4 business model.

The goal of the case study is to determine costs and benefits of information systems adding

demand response to electricity markets. Therefore, a system architecture for a demand

response ICT infrastructure is laid out. Core of the system design is an advanced metering

infrastructure, which comprehends smart meters to read out electricity demand patterns,

network components for transmission of meter data, and backend systems that collect and

aggregate the meter data for further processing. The backend system as well determines the

optimal load shifts and computes the respective signals for load control. These signals are

transferred via the same channels that are used for collection and transmission of meter


36

data. For the smart meters two common types of communication technology are

considered, namely mobile phone networks (GSM) and power line carrier (PLC). The

system design does not incorporate any components on the consumers’ side, which are

required for remote control of any electricity-consuming devices.

From the system architecture, a comprehensive cost model is deducted. The model

comprises both the initially required investments as well as the variable operating costs of

the infrastructure, including the communication costs incurred. The cost model is opposed

by a revenue model, which calculates the optimal load shift based on historical electricity

prices and currently available demand response contingencies. In this setting, the value of

the information, i.e. the smart meter records, is determined by the savings that an

electricity retailer can achieve by optimizing its electricity sourcing. These savings are

defined in dependence of the information granularity resulting from the readout frequency

and the year after installation. On the opposite side, the variable costs and the fixed set-up

costs of the advanced metering infrastructure have to be considered. To come up with the

value of a single smart meter readout, the difference between total revenues and costs is

divided by the number of readouts. The information value thus corresponds to the

maximum price, which the data aggregator can demand from the electricity retailer for its

information-based value proposition.

Using real-world data on the electricity consumption and market prices from 2011, the

selected setting does not generate a positive information value with the basic configuration.

The immense investments completely deplete the revenue potential. If restricting the

consideration to variable costs only, a readout frequency between 21 and 57 minutes yields

a positive information value. The optimal readout frequency is calculable at 41 minutes –

the value of 1,000 smart meter readouts equals approximately € 0.04.

In order to obtain a positive information value also with the full cost consideration, two

approaches are followed. First, the sensitivity of system component costs is examined. The

main cost drivers are the hardware costs for the smart meters and the costs for GSM


37

communication. Only with a dramatic market price decline by around 30 % for these

components, a positive result is obtainable. The second approach refrains from a rollout of

smart meters to the entire market. Instead, customers with the highest electricity

consumption are preferred. Only customers whose consumption exceeds a certain

threshold – which equals the marginal costs – are included in the demand response

ecosystem. With this scenario, a net present value of € 0.98 M over a period of 15 years

can be achieved.

Ultimately, an extension to the basic setting examines the incorporation of demand

response strategies into different electricity markets. Whereas the basic setting focus on the

product market, the markets of tertiary reserve and balancing energy resp. imbalance

penalties are considered here as well. First, in the reserve market, the electricity retailer

offers demand response potential as a replacement for tertiary control reserve. By shifting

load the retailer can provide both positive and negative reserve potentials. The available

capacities have to be announced one day ahead. After activation, the tertiary reserve has to

be fully available within 15 minutes. As compensation, the retailer receives a capacity and

a working price. The working price is only paid, when the capacity is actually requested.

Second, the retailer has to project the electricity demand for the next day. Whenever

imminent deviations occur, the retailer can use demand response potentials to mitigate the

payment of the very expensive imbalance penalties. In summary, the comparison shows

that the product market yields the highest profit for an electricity retailer. For the other two

markets, the annual data volume per smart meters significantly beats the data volume

required for addressing the product market. Hence, the operating costs arising from the

communication effort surmounts the additional revenue potentials. The information value

for demand response is negative in both market settings.

4.3 Comparison of Case Studies

The previous sections introduced two very different information-based business models in

their real-world setting. The business models not only map to different classes of the


38

defined taxonomy, and thus contrast regarding their value proposition, they also differ in

other components of their business model. A one-by-one comparison of the two models is

given in Table 6. The following paragraphs examine two areas of the business model in

more depth, namely the value of information as well as the key activities.

The concept of value drivers, which has been defined by Amit and Zott (2012) over the last

years in connection with business model innovation, serves as basis to analyze the value of

information within the business model. The four value drivers are novelty, lock-in,

complementarities and efficiency.

For the business model based on personalized recommendations for online content, the

value drivers complementarities and lock-in effect are the matching ones.

Complementarity describes the value-enhancing effect, which results from linking various

elements into a new, improved value proposition. In this case, the newspaper offering

increases significantly in value by adding the personalization element. This boosts the

second value driver, the lock-in. The more frequently a user visits the website, the better

the recommender gets to know the user and the personal preferences. This in turn raises the

quality of the recommendations, which leads to a higher satisfaction with the user. The

bottom line for the provider is additional revenue through happy users paying longer and

more frequent visits to the website. A positive feedback loop arises, as the incentives for

users to stay and interact with the website multiply.

The business model of the case study dealing with electricity demand response bases its

value proposition on the other two value drivers, novelty and efficiency. By slightly

extending the definition on novelty by Amit and Zott (2012), not only the degree of

business model innovation, but also the novelty of the underlying data can be taken into

account. The business model requires the usage of highly up-to-date data for the various

forecasts inherent to the model, be it for the development of electricity market prices or the

future electricity usage and demand patterns.


39

Personalized Recommendations

- Case Study 1 -

Electricity Demand Response

- Case Study 2 -

Basic Data

Company Online content provider (regional

newspaper)

Information intermediary and aggregator

in electricity market

Taxonomy

Class

Class 3 – information upgrades offering Class 4 –pure-play information product

Value Model

Value

Propositions

Core product is online content,

specifically news articles - both regional

and (inter-) national across various

categories (e.g. news, sports, weather)

Core product is enhanced with

personalized recommended links that are

generated based on the analysis of the

user behavior

Collection, aggregation and transmission

of (near) real-time usage patterns and

demand response contingencies from

electricity consumers

Transmission and distribution of load

control signals to concerned electricity

consumers

Strategies: repackage and mine data &

analytics

Customer

Segments

Personalized content is offered to online

newspaper readers (B2C)

Data are individually packaged to needs

of electricity retailers (B2B)

Customer

Relationships

Customer intimacy due to information

sphere (i.e. recommendations)

Symbiotic relationship between

infomediary and electricity retailer

Channels ./. Data are provided through defined

digital interfaces

Operating Model

Key Resources Observed and recorded user click

streams on news articles

Recommender system (i.e. algorithm)

Processing power and storage capacities

(on premise or on-demand)

Usage data retrieved from regular

readouts of smart meters

Dedicated ICT infrastructure, from smart

meters to backend systems

Capacity for processing and aggregation

of large data sets

Key Partners Partnering with advertising networks Interaction with electricity markets and

further external data providers

Key Activities Analytics on user click stream data as

basis for generation of personalized

recommendations

Collection, transmission, storage, and

aggregation of meter data

Transmission of load control signals

Financial Model

Value of

Information

Value driver: personalization improves

customer experience by complementing

news services and creating lock-in

Revenue stream: content provider

receives compensation for showing

banner ads

Cost structure: licensing and operations

of recommender system, depending on

provider model; license model is either

fixed monthly payment or pay-per-use

(based on shown recommendations)

Value driver: efficient aggregation and

transmission of high volume data in

(near) real-time

Revenue stream: usage-based pricing

(dependent on number of meters and

frequency) or profit sharing with retailer

Cost structure: high investment and

operational costs for infrastructure;

aiming at economies of scale by

leveraging synergies from reuse of

infrastructure for additional use cases

Table 6. Summary of Business Model Components for the Two Conducted Case Studies.


40

The implementation of demand response mechanisms requires building up a

comprehensive and capital-intensive infrastructure. Efficient operations are thus crucial for

the provider to operate profitably. One way to attain efficiency is the reuse of the

established demand response infrastructure for multiple use cases; another is placing the

demand response mechanism on top of an already existing, self-owned infrastructure. The

latter applies for example to telecommunication providers, who already have access to a

communication infrastructure covering the last mile to customers.

A glance on the information value chain reveals further differences (see Figure 7) and thus

emphasizes the key activities to be performed within each business model. The

personalized recommendations, being the final information product, mainly draw their

added value from steps 3 and 4 of the value chain, through application of elaborate

analytics and intelligent visualization for the user. The key areas of value creation in the

demand response scenario are the collection and transmission of data, corresponding to the

first two steps of the information value chain. Furthermore, processing and aggregation

capabilities are required to generate optimal load profiles, but they are less predominant

compared to the other business model.

A last remark points at the generalizability of these results. It cannot be inferred that a class

3 business model is always based on analytics, whereas a class 4 utilizes storage and

Figure 7. Comparison of Value Creation in the Information-based Business Models Evaluated

within the Two Case Studies. The Major Steps of Value Creation are Highlighted in Bold.

Personalized

Recommendations

- Case Study 1 -

Electricity Demand

Response

- Case Study 2 -

Click streams

(user generated)Scalable Database

for storage

Recommender

algorithm

Personalized

recommended links

Internet for transmission

Generation &

Acquisition

Storage &

Transmission

Processing &

Analytics

Presentation

& Distribution

Usage patterns

(user generated)

Aggregation of

collected meter data

Aggregated

demand pattern

Supply schedule Matching algorithmLoad shifting

controls

upstream

downstream

Advanced Metering

Infrastructure


41

transmission to its benefit. Instead, the following conclusion prevails: information-based

business models can capitalize on different elements and activities across all steps of the

information value chain, and transform the gained benefits into real money. Here, the

developed taxonomy has proven to serve as a prevailing approach for extraction of patterns

in information-based value creation. However, a significantly larger sample of (real-world)

case studies is required to conduct an expedient, in-depth comparison of business models

and derive universally valid propositions and patterns of value creation.

5 Summary and Future Research

The abundance of information in today’s world and its continually increasing importance

for both business and society more than justify the analysis of business models centered

around information.

The contribution of this thesis in the realm of IS research is two-fold. First, it provides a

definition for a taxonomy to classify different types of information-based business models

according to the respective role played by information. The used framework combines the

well-established concepts of the information value chain and the business model canvas.

Second, two case studies are conducted in realistic scenarios in the areas of online content

delivery and electricity markets. The studies prove that information can constitute a key

part or even the essence of the value proposition of a company, and adds a positive

financial contribution to its bottom line. The value add of information is quantified and

depicted as precise, monetary amount in the respective study scenarios.

The two case studies exhibit differences pertaining to the business model components,

especially the means of informational value creation along the steps of the information

value chain and the applied value drivers. This shows that business models based on

information are quite flexible, and allow the provider to choose from a variety of revenue

models. On the downside, there is no universally applicable method to generate value from

information. The combination of business domain, types of data, used distribution channels

and value perception by the customer demands for a tailored approach.


42

The two studies put the stake in the ground for further research around information-based

business models. In order to deduct clear trends and general recommendations, more

examples of real-world studies are required. A myriad of relevant industries and

conceivable value propositions exist, but already the realms of the two presented case

studies offer interesting potential for related research. For the personalized

recommendations for online news, a longer running experiment could showcase intensified

effects through a more educated recommender system yielding even better

recommendations. Modifying certain components of the business model of the online

content provider and comparing results to the base scenario would be another interesting

research path: Regarding the key activities, the provider could select different

recommendation mechanisms or test more progressive ways of presenting the

recommendations on the website. A collection and inclusion of demographic data into the

recommendation engine would allow for more granular customer segmentation. In

addition, the financial model could be subject to change. Basing revenue streams not on

advertisement, but subscription or freemium models, would also require a reconsideration

of the underlying key performance indicators of the website provider in order to evaluate

success. Altering the core product, online news articles, and with this the value proposition

as such, into other forms of online content, such as pictures or videos, might yield

interesting comparisons on the effectiveness of personalized recommendations.

For the other case study, demand response in electricity markets, variations on the key

resources, financial model and value proposition are supposable. Using different

communication media and protocols within the setup of the demand response system might

lead to different results with respect to the optimal information granularity, due to an

altered cost structure. The cost side of the financial model generally awakes further

research interest, especially considering the costs incurred for interaction with the

electricity consumers: Who bears the cost for the required infrastructure on consumers’

premises? What incentives are required to encourage participation in demand response? Do


43

they have to be financial in nature, or would the creation of an information sphere around

electricity consumption (e.g. a portal to share one’s eco-sensitive attitude) be sufficient to

win customers over? Further research should definitively be devoted to elaborate on the

value proposition of the information intermediaries and promising opportunities to reuse

the costly infrastructure for additional revenue-generating services, for example monthly

billing or electricity theft prevention.

Besides creating further evidence on the value of information through studies on

information-based business models, a number of overarching questions with strong

relevance for information-centered companies wait to be answered by IS research. Taking

a step back from the numerous examples of successful business models, there is little

substance on the first entrepreneurial steps to take towards the creation of a promising

business idea based on information: What industries are predestined for success? How to

spot suitable data types? How to best transform the rough data diamonds into real financial

value? How can further innovation be utilized to lift information-based business models to

the next level, once the business is running steadily?

For larger corporations, the increased importance attributed to information gives rise to the

question if and how this development should steer the management of skills and setup of

the organization. The management consultancy McKinsey, for example, recommends

establishing data and their analysis as a key function within the company, even granting

them a CxO position (McKinsey, 2013). Even today, many companies already see their

Chief Information Officer (CIO) actually in charge of managing information, as the name

has always been suggesting, and not just technology. Proving or disproving the validity of

empowerment of data governors with empirical evidence seems an interesting field for IS

research.

With information being a key part of the value proposition, it becomes a strategic asset and

thus target of industrial espionage. Effective data security mechanisms, e.g. encryption, are

required to protect the data from fraud and theft – both from within the company, but also


44

from outside, as demonstrated by the recent revelation around the National Security

Agency (NSA). Here, IS research could clearly contribute on the technological aspects of

data security, but even more by establishing strategies for companies to find tradeoffs

between sophisticated, but costly and cumbersome security mechanisms, and retained, yet

risky agility of their company.

Since many information-based businesses operate in the B2C realm, societal trends should

be observed closely in order to recognize the signs potentially affecting success of

information-based business ideas. Many of these trends are ambiguous though, and have

partially been met by counter movements already. It is currently barely foreseeable which

direction will prevail in the mid- to long-term.

Currently, the willingness to generate and share content, coined by the term shareconomy,

is still on the rise. 74 % of Europeans see disclosing personal information as part of

modern life (European Commission, 2011). The question is, whether this positive attitude

towards data and digital goods will continue, or whether the information overflow will lead

to an all-embracing saturation or even weariness at some point in time. Already today, the

trend towards deceleration is visible in other areas of life, e.g. with the slow food

movement. Moreover, the growing do-it-yourself culture with the underlying desire for

tangible goods might give the digital economy a hard time.

Data privacy is another topic worthwhile looking at both sides of the coin. With more

restrictive data privacy laws appearing on the horizon, operating with privacy data

becomes a high-risk endeavor for companies. For example, the drafted EU data protection

law4 features penalties for data protection violations of up to 2 % of a company’s turnover.

Some industries, such as healthcare, apply specifically strict rules, which in some cases

might even become a showstopper for envisaged business models. On the other hand,

actively addressing the topic of data privacy can be a value driver and selling argument –

4 http://ec.europa.eu/justice/newsroom/data-protection/news/120125_en.htm (Press release), Accessed 15 February 2014

http://ec.europa.eu/justice/newsroom/data-protection/news/120125_en.htm


45

as seen with the German Cloud5, a certification body for companies providing “safe”

ground for outsourcing storage of data, in line with the relatively strict German data

protection laws.

With certainty, the society – both providers and consumers – have acknowledged the value

attributed to personal data. Privacy is already denoted as “the new currency of the web”

(Vandor, 2010): The start-up Datacoup, for example, offers cash in exchange for personal

data from social media accounts and credit-card transactions. A study conducted by

ENISA (2012) found out that if users are offered a digital service with a great deal of

privacy, which costs € 0.5 more than an alternative one without privacy protection, users

choose the cheaper option. Even though such incentives work, an increasing number of

people are skeptical towards the large data krakens such as Google or Facebook, which

grab an ever-larger share of the digital sphere of life – impressively demonstrated by

Facebook’s recent acquisitions of Instagram and WhatsApp.

In summary, further practical research is required to confirm the positive trend for

information-based business models emerging from the results of this thesis. This additional

empirical evidence paired with a vigilant observation of societal streams, might provide a

definitive answer to the question whether information truly is – and will stay – a silver

bullet for business models.

5 http://www.german-cloud.de/, Accessed 15 February 2014

http://www.german-cloud.de/


46

6 References

Águila-Obra, A., Padilla-Meléndez, A., and Serarols-Tarrés, C. (2007). Value Creation and

New Intermediaries on Internet. An Exploratory Analysis of the News Industry and the

Web Content Aggregators. International Journal of Information Management, 27 (3),

pp. 187–199.

Amit, R. and Zott, C. (2012). Creating Value Through Business Model Innovation.

MIT Sloan Management Review, 53 (3), pp. 41–49.

Analysys Mason (2013). Mobile Data Forecast. http://www.analysysmason.com/About-

Us/News/Insight/Mobile-data-Oct2013/ Accessed 15 January 2014.

Bakos, Y. and Brynjolfsson, E. (2000). Bundling and Competition on the Internet.

Marketing Science, 19 (1), pp. 63–82.

Balazinska, M., Howe, B., and Suciu, D. (2011). Data Markets in the Cloud: An

Opportunity for the Database Community. Proceedings of the 37th International

Conference on Very Large Data Bases.

Bellman, R., Clark, C.E., Malcolm, D.G., Craft, C.J., and Ricciardi, F.M. (1957). On the

Construction of a Multi-Stage, Multi-Person Business Game. Operations Research, 5

(4), pp. 469–503.

Bloching, B., Luck, L., and Ramge, T. (2012). Data Unser: Wie Kundendaten die

Wirtschaft revolutionieren. Munich, Germany: Redline.

Brynjolfsson, E. and Saunders, A. (2010). Wired for Innovation: How Information

Technology is Reshaping the Economy. Cambridge, USA: MIT Press.

Burkhart, T., Krumeich, J., Werth, D., and Loos, P. (2011). Analyzing the Business Model

Concept - A Comprehensive Classification of Literature. Proceedings of International

Conference on Information Systems (ICIS).

Castells, M. (2010). The Information Age: Economy, Society and Culture. 2nd ed. Oxford,

Great Britain: Wiley-Blackwell.

Choo, C.W. (2002). Information Management for the Intelligent Organization: The Art of

Scanning the Environment. 3rd ed. Medford, USA: Information Today.

Cisco (2011). The Internet of Things: How the Next Evolution of the Internet is Changing

Everything. White Paper.

Clemons, E.K. (2009a). Business Models for Monetizing Internet Applications and Web

Sites: Experience, Theory, and Predictions. Journal of Management Information

Systems, 26 (2), pp. 15–41.

Clemons, E.K. (2009b). The Complex Problem of Monetizing Virtual Electronic Social

Networks. Decision Support Systems, 48 (1), pp. 46–56.

Cook, S. (2008). The Contribution Revolution: Letting Volunteers Build Your Business.

Harvard Business Review, 86 (10), pp. 60–69.

Davenport, T.H. (2006). Competing on Analytics. Harvard Business Review (1).

Davenport, T.H. and Harris, J.G. (2007). Competing on Analytics: The New Science of

Winning. Boston, USA: Harvard Business School Press.

Davenport, T.H. and Patil, D.J. (2012). Data Scientist: The Sexiest Job of the 21st Century.

Harvard Business Review (10), pp. 70–76.

ENISA (2012). Study on Monetising Privacy: An Economic Model for Pricing Personal

Information.


47

European Commission (2011). Attitudes on Data Protection and Electronic Identity in the

European Union. Special Eurobarometer 359.

Evans, J.R. (2013). Business Analytics: Methods, Models, and Decisions. Upper Saddle

River, USA: Pearson.

Frauenfelder, M. (2004). Sir Tim Berners-Lee: Featured Story.

http://www.technologyreview.com/featuredstory/403095/sir-tim-berners-lee/. Accessed

20 February 2014.

Gartner (2010). Hyperdigitization Creates Major Opportunity for IT Services Providers:

Analyst Report.

Gartner (2013). Forecast: The Internet of Things, Worldwide, 2013. Analyst Report.

Gassmann, O., Frankenberger, K., and Csik, M. (2013). Geschäftsmodelle aktiv

innovieren. D. Grichnik and O. Gassmann, eds. Das unternehmerische Unternehmen.

Wiesbaden: Springer Fachmedien, pp. 23–41.

Glazer, R. (1993). Measuring the value of information: The information-intensive

organization. IBM Systems Journal, 32 (1), pp. 99–110.

Haag, S. and Cummings, M. (2013). Management information systems for the information

age. 9th ed. New York, USA: McGraw-Hill Irwin.

Hedman, J. and Kalling, T. (2001). The Business Model: A Means to Understand the

Business Context of Information and Communication Technology: Working Paper:

School of Economics and Management, Lund University.

Hemp, P. (2009). Death by Information Overload. Harvard Business Review (9), pp. 83–

89.

IBM Global Business Services (2011). Digital transformation: Creating new business

models where digital meets physical. Executive Report.

IDC (2012). The Digital Universe in 2020: Big Data, Bigger Digital Shadows, and Biggest

Growth in the Far East. Analyst Report.

Jain, S. and Kannan, P.K. (2002). Pricing of Information Products on Online Servers:

Issues, Models, and Analysis. Management Science, 48 (9), pp. 1123–1142.

Krcmar, H. (2005). Informationsmanagement. 4th ed. Berlin: Springer.

Kushal, A., Moorthy, S., and Kumar, V. (2011). Pricing for Data Markets: Working Paper:

University of Washington.

Kwok, R. (2009). Personal technology: Phoning in data. Nature, 458 (7241), pp. 959–961.

Lee, C.C. and Yang, J. (2000). Knowledge value chain. Journal of Management

Development, 19 (9), pp. 783–794.

Linde, F. (2009). Pricing information goods. Journal of Product & Brand Management, 18

(5), pp. 379–384.

Loukides, M. (2010). What is Data Science? http://radar.oreilly.com/2010/06/what-is-data-

science.html. Accessed 20 February 2014.

Lucey, T. (2005). Management information systems. 9th ed. London: Thomson Learning.

Madnick, S.E. and Siegel, M. (2002). Seizing the Opportunity: Exploiting Web

Aggregation. MIS Quarterly Executive, 1 (1), pp. 35–46.

Marchand, D.A., Kettinger, W.J., and Rollins, J.D. (2002). Information orientation: The

link to business performance. Oxford, Great Britain: Oxford University Press.

http://www.technologyreview.com/featuredstory/403095/sir-tim-berners-lee/

http://radar.oreilly.com/2010/06/what-is-data-science.html

http://radar.oreilly.com/2010/06/what-is-data-science.html


48

McAffee, A. and Brynjolfsson, E. (2012). Big Data: The Management Revolution.

Spotlight on Big Data. Harvard Business Review (10).

McKinsey (2013). Mobilizing your C-suite for big-data analytics: McKinsey Quarterly

November.

Möslein, K. and Kölling, M. (2007). Interaktive Hybride Wertschöpfung als

Innovationsstrategie. D. Streich and D. Wahl, eds. Innovationsfähigkeit in einer

modernen Arbeitswelt: Personalentwicklung, Organisationsentwicklung,

Kompetenzentwicklung: Beiträge der Tagung des BMBF. Frankfurt/Main, Germany:

Campus.

Nolan, Norton & Co. (2014). Competing on Information.

http://www.competingoninformation.com/. Accessed 20 February 2014.

Norris, P. (2001). Digital Divide: Civic Engagement, Information Poverty, and the Internet

Worldwide. Cambridge, USA: Cambridge University Press.

OECD (2007). Participative Web: User-Created Content: Committee for Information,

Computer, and Communications Policy.

Ofek, E. and Wathieu, L. (2010). Are You Ignoring Trends That Could Shake Up Your

Business? Harvard Business Review (7).

O'Reilly, T. (2005). What Is Web 2.0: Design Patterns and Business Models for the Next

Generation of Software. http://oreilly.com/web2/archive/what-is-web-20.html.

Accessed 15 January 2014.

Osterwalder, A., et al. (2010). Business Model Generation: A Handbook for Visionaries,

Game Changers, and Challengers. 1st ed. New Jersey, USA: John Wiley & Sons.

Pauwels, K. and Weiss, A. (2008). Moving from Free to Fee: How Online Firms Market to

Change Their Business Model Successfully. Journal of Marketing, 72 (3), pp. 14–31.

Perez, J.C. (2007). Google Wants Your Phonemes. http://www.infoworld.com/t/data-

management/google-wants-your-phonemes-539. Accessed 15 January 2014.

Picot, A., Reichwald, R., and Wigand, R.T. (2003). Die grenzenlose Unternehmung:

Information, Organisation und Management. Lehrbuch zur Unternehmensführung im

Informationszeitalter. 5th ed. Wiesbaden, Germany: Betriebswirtschaftlicher Verlag

Gabler.

Porter, M.E. (2001). Strategy and the Internet. Harvard Business Review, 79 (3), pp. 62–

78.

Porter, M.E. and Millar, V.E. (1985). How Information Gives You Competitive

Advantage. Harvard Business Review (7/8).

PricewaterhouseCoopers (2009). 12th Annual Global CEO Survey: Redefining Success.

Ramaswamy, V. (2008). Co-creating value through customers' experiences: the Nike case.

Strategy & Leadership, 36 (5), pp. 9–14.

Rappa, M. (2001). Business Models on the Web. http://www.digitalenterprise.org/models/.

Accessed 15 February 2014.

Rayport, J.F. and Sviokla, J.J. (1995). Exploiting the Virtual Value Chain. Harvard

Business Review.

Redman, T.C. (2008a). Data Driven: Profiting from Your Most Important Business Asset.

1st ed. Boston, USA: Harvard Business Press.

Redman, T.C. (2008b). Put Your Data to Work in the Marketplace. Harvard Business

Review (9), pp. 34.

http://www.competingoninformation.com/

http://oreilly.com/web2/archive/what-is-web-20.html

http://www.infoworld.com/t/data-management/google-wants-your-phonemes-539

http://www.infoworld.com/t/data-management/google-wants-your-phonemes-539

http://www.digitalenterprise.org/models/


49

Rieger, V. and Tempich, C. (2008). Information als Geschäftsmodell. IT Management.

Ritzer, G. and Jurgenson, N. (2010). Production, Consumption, Prosumption: The Nature

of Capitalism in the Age of the Digital "Prosumer". Journal of Consumer Culture, 10

(1), pp. 13–36.

Rosenberger, L., Nash, J., and Graham, A. (2009). The Deciding Factor: The Power of

Analytics to Make Every Decision a Winner. 1st ed. San Francisco, USA: Jossey-Bass.

Schwolow, S. and Jungfalk, M. (2009). The Information Value Chain: Strategic

Information Management for Competitive Advantage: Copenhagen Business School.

Shapiro, C. and Varian, H.R. (1999). Information Rules: A Strategic Guide to the Network

Economy. Boston, USA: Harvard Business School Press.

Sharda, R., Asamoah, D.A., and Ponna, N. (2013). Business Analytics: Research and

Teaching Perspectives. Proceedings of the 35th International Conference on Information

Technology Interfaces.

Shmueli, G. and Koppius, O.R. (2011). Predictive Analytics in Information Systems

Research. Research Essay. MIS Quarterly, 35 (3), pp. 553–572.

Spataro, J. and Crow, B. (2002). A Framework for Understanding the Information

Management Market. The Gilbane Report: Bluebill Advisors.

Stähler, P. (2002). Geschäftsmodelle in der digitalen Ökonomie: Merkmale, Strategien und

Auswirkungen. 2nd ed. Cologne, Germany: Josef Eul Verlag.

Stocker, A. and Tochtermann, K. (2009). Geschäftsmodelle im Semantic Web - die

Wertschöpfungskette der Daten. DOK Magazin.

Strikwerda, J. (2011). Competing on Information - An Exploration of Concepts: Working

Paper, Amsterdam Business School.

The Economist (2010). Business 2010: Embracing the Challenge of Change: Economist

Intelligence Unit.

Timmers, P. (1998). Business Models for Electronic Markets. Electronic Markets, 8 (2),

pp. 3–8.

Vandor, M. (2010). Privacy: Managing the New Currency of the Social Web.

http://mashable.com/2010/01/29/privacy-managing-new-currency/. Accessed 12

February 2014.

Veit, D., Clemons, E.K., Benlian, A., Buxmann, P., Hess, T., Kundisch, D., Leimeister,

J.M., Loos, P., and Spann, M. (2014). Business Models – An Information Systems

Research Agenda. Business & Information Systems Engineering, Forthcoming.

Weiner, N., Renner, T., and Kett, H. (2010). Geschäftsmodelle im "Internet der Dienste":

Aktueller Stand in Forschung und Praxis. Stuttgart, Germany: Fraunhofer-Verlag.

Wirtz, B.W. (2001). Electronic Business. 2nd ed. Wiesbaden, Germany: Gabler.

Wirtz, B.W. (2013). Business Model Management: Design, Instrumente, Erfolgsfaktoren

von Geschäftsmodellen. 3rd ed. Wiesbaden, Germany: Springer Gabler.

Wirtz, B.W., Schilke, O., and Ullrich, S. (2010). Strategic Development of Business

Models: Implications of the Web 2.0 for Creating Value on the Internet. Long Range

Planning, 43 (2-3), pp. 272–290.

Wu, M. (2013). Big Data Reduction. http://lithosphere.lithium.com/t5/science-of-social-

blog/Big-Data-Reduction-1-Descriptive-Analytics/ba-p/77766. Accessed 12 February

2014.

http://mashable.com/2010/01/29/privacy-managing-new-currency/

http://lithosphere.lithium.com/t5/science-of-social-blog/Big-Data-Reduction-1-Descriptive-Analytics/ba-p/77766

http://lithosphere.lithium.com/t5/science-of-social-blog/Big-Data-Reduction-1-Descriptive-Analytics/ba-p/77766


50

Yu, J., Benatallah, B., Casati, F., and Daniel, F. (2008). Understanding Mashup

Development. IEEE Internet Computing, 12 (5), pp. 44–52.

Zhu, H. and Madnick, S.E. (2009). Finding New Uses for Information. MIT Sloan

Management Review, 50 (4), pp. 18–21.

Zott, C., Amit, R., and Massa, L. (2011). The Business Model: Recent Developments and

Future Research. Journal of Management, 37 (4), pp. 1019–1042.

51

II. CASE STUDY 1: PERSONALIZED

RECOMMENDATIONS

PART II CASE STUDY 1: PERSONALIZED RECOMMENDATIONS

52

A. Always Mystify, Mislead, and Surprise your Customers?

Impact of Personalized Recommendations on the Performance

of an Online Content Provider

This chapter is a fundamentally revised and an extended version of the papers

Philipp Bodenbenner, Markus Hedwig, Dirk Neumann, 2011: Impact of Recommendations

on Advertising-based Revenue Models. Pre-ICIS Conference – 10th Workshop on E-

Business (WEB 2011), Shanghai, China, 4 December 2011.

Philipp Bodenbenner, Dirk Neumann, 2012: Are Personalized Recommendations the

Savior for Online Content Providers? Multikonferenz Wirtschaftsinformatik (MKWI

2012), Braunschweig, Germany, 29 February – 2 March 2012.

Abstract

Providers of online content need a loyal user base to establish a profitable business model.

This holds especially true when advertising is the primary revenue stream. In this part, we

show that personalized recommendations are a viable lever to increase website traffic.

Therefore, we conduct a controlled experiment on the website of a large German

newspaper to evaluate the impact of recommenders on observable user behavior. The

evaluation clearly shows a positive effect on the provider’s key performance indicators.

Using personalization, the attractiveness of links and the length of visits increase

significantly. The frequency of visits shows a positive trend as well. Combined, these

enhancements lead to annual additional revenue of € 300,000 in our scenario. Moreover,

the personalized recommendations lead to a higher concentration regarding the breadth of

the visited article portfolio. This in turn helps the provider to further optimize the portfolio

and hence lower operating costs.

Keywords

Online Content Provision, Personalization, Recommender System, Value of Information,

Empirical Study, Controlled Experiment, Observable User Behavior, Web Metrics


53

1 Introduction

The growing influence of information systems (IS) and the internet lead to an increasing

flood of information. Information is ubiquitously available and websites offer several

thousand items to their users. Out of these, only a few items can be featured (for a limited

period) on landing pages. Thus, it is a key challenge for content pages to help users find

the items of their interest. Appropriate filters for information processing and search are

required to get a grip on the information repositories.

A promising approach to this challenge is personalization of websites. To create an

effective personalization, all available data reflecting the user behavior has to be taken into

account. The technological development allows collecting individual users’ preferences

with little effort and in turn providing tailored content matching the demand. 80 % of the

internet users welcome the usage of personalized services (Kobsa 2007). 56 % of online

shoppers buy with higher probability from an online store that employs personalization

features than from one using none (Freedman, 2007). Personalization has consequently

become a common practice in online business (see e.g. Ansari and Mela, 2003, Murthi and

Sarkar, 2003, Wirtz et al., 2010), and most of the major online retailers as well as content

providers today employ personalized recommendations on their websites. In summary,

personalization features are regarded as a key success factor of online services (Delone and

McLean, 2003).

Pathak et al. (2010) summarize the tangible benefits personalization mechanisms imply for

the consumer and consequently for the retailer: an increased quality perception, a

broadened horizon on the overall portfolio including cross-selling opportunities and a

growing intimacy with the customer resulting in their loyalty and trust – a key benefit also

for Li (2009). For the service provider it is essential to build up a loyal, recurring customer

base. This especially applies to content service providers, which receive substantial parts

of their revenue stream from advertising. The total revenue of online newspapers consists

of 81.5 % earnings from advertising; the rest is based on subscription (Clemons et al.,


54

2002). This revenue structure is also expected to coin the future online content market

(Herbert and Thurman, 2007).

In the past, numerous studies have described the application of personalization features.

However, there is only little research regarding the actual impact of personalized

recommendations on the observable behavior of website users. In addition, the perspective

of the website provider has mostly not been the core focus – especially not in an online

content provision setting. This work is thus the first attempt to prove by means of observed

user behavior, that there is a positive impact of personalized recommendations on the

revenue of a content website provider. The contributions are as follows:

We conduct an empirical study in a real-world setting, analyzing users in an

experiment with an online newspaper. With this, we add one of the first content

provider cases to the long list of e-commerce examples.

We are among the first to realize our study as an online controlled experiment, thus

being able to directly compare the outcomes between the two observed groups.

We do not rely on the analysis of expressed user intentions, but measure and evaluate

actual user behavior on the website.

We formalize our evaluation model based on accepted and proven web metrics, and

move the impact of personalization on the revenue of the provider in the focus of our

work.

The remainder of part II.A is structured as follows. Chapter 2 defines the terms

personalization and recommender systems and gives an outline on different types of the

same. Chapter 3 presents a discussion of relevant literature in the realms of evaluation of

personalized recommendations, online consumer behavior and sales concentration. Our

research is based upon fundamental concepts from the domain of relationship marketing,

namely trust and commitment. Chapter 4 presents the theoretical foundations and

establishes our research hypotheses, which set all components of the revenue stream of


55

online content providers into relation. Chapter 5 describes the design of our empirical

study, the sample, and the model specification to analyze our hypotheses empirically.

Chapter 6 outlines and discusses the results of the empirical analysis as well as managerial

implications. Chapter 7 concludes by presenting a summary of findings, a discussion of

limitations as well as ideas for future research.

2 Personalization and Recommender Systems

The concepts of personalization from a marketing perspective as well as recommender

systems from an IT perspective constitute the fundamental building blocks of this work.

We briefly introduce both concepts in this chapter.

The practice of using customer information to deliver a targeted solution is known as

personalization or one-to-one marketing (Peppers et al., 1999). Adomavicius and Tuzhilin

(2005) provide the following definition: “personalization is the use of technology and

customer information to tailor electronic commerce interactions between a business and

each individual customer. Using information either previously obtained or provided in real

time about the customer, the exchange between the parties is altered to fit that customer’s

stated needs, as well as needs perceived by the business based on the available customer

information”.

Within the realm of personalization of digital products, online recommendations are a key

method. Being a form of information retrieval or information filtering, online

recommendations provide the user with a pre-selected list of supposedly relevant items.

Online recommendations are classified into three clusters: best sellers, consumers’

reviews, and personalized recommendations (Shi and Wang, 2008). Best seller refers to the

list of items that have been sold or viewed most. Consumers’ reviews provide customer

opinions based on their experiences having used an item. Finally, personalized


56

recommendations6 describe the approach of suggesting items to users (e.g. content, product

information), which match to their individual needs and preferences at the best possible

rate (Resnick and Varian, 1997). These items are taken from a large information

repository. The algorithm, which derives the user’s preferences and matches the relevant

items, is at the core of personalized recommendations (Burke, 2002). The item selection is

based on the probability that the item fits to the available user preferences and needs. Due

to the complexity of creating valuable user profiles, the proper use of IT is critical for

creating personalized recommendations. In the following, we use the term recommender

system as synonym for an information system generating personalized recommendations.

Recommender systems can be classified using three dimensions: the content of the

recommendations, the preference identification method and the recommendation

algorithm. The first dimension describes the characteristics of the recommended items.

Knijnenburg et al. (2012) distinguish between e-commerce recommender systems used for

products, media recommender systems for recommendations on news, video, music and

other similar content, and social network recommender systems to suggest relevant people.

The second classification dimension of recommender systems relates to the approach that

is used for capturing user preferences. There are generally two methods (Pommeranz et al.,

2012): First, users can explicitly explain their preferences, e.g. by rating an item, assigning

weight to item attributes or answering a questionnaire (see e.g. Rashid et al., 2002, Pu et

al., 2012). Second, the implicit method infers the users’ preferences by analyzing their

observable behavior. Therefore, the system records a user’s browsing behavior (e.g. by

capturing keystrokes and clicked hyperlinks) and extracts the relevant information for

creating a user profile (Hauser et al., 2009). Previous research has shown that the implicit

approach performs equally well as the explicit approach in determining user preferences

(Zhang and Seo, 2001). Based on the captured user preferences and profiles, a

recommender system is able to generate personalized recommendations. Different

6 Personalized recommendations mostly have the following appearance on a website: „people who read this article also read …“ respectively “customers who bought this item also bought …”


57

approaches are deployed for the recommendation algorithm, the third classification

dimension: attribute-based filtering, collaborative filtering, and hybrid combinations

thereof (see e.g. Adomavicius and Tuzhilin, 2005, Knijnenburg et al., 2012). Content-

based filtering is the most popular type of attribute-based filtering. Here, the selection is

determined as similarity measure between the items, i.e. items having similar

characteristics being assumed to be closely related (Maidel et al., 2008). Another type of

attribute-based filtering is the demographic-based filtering aiming to create

recommendations based on users’ personal attributes and demographic classes (Burke,

2002). On the contrary, collaborative recommendations provide users with items that were

used by people with similar preferences and behavior in the past (see e.g. Zhang and Li,

2007, Yang and Li, 2009). As of today, recommender systems based on collaborative

filtering algorithms are most common for recommendations of products in online shops

(Pathak et al., 2010).

3 Literature Review

Our review of relevant literature mainly covers three areas. The first part deals with the

evaluation of recommendation mechanisms, in which we observe an evolution from a pure

technical to a more holistic and impact-oriented way. We will tie the concept applied to

assess the recommendation mechanism in our experiment to the most recent examples of

the IS community. We consider the concept of relationship marketing as a key basis of our

evaluation as well, thus we dedicate the second part in this chapter to relevant theoretical

and practical studies in this field. We conclude our literature review by shedding light on

interesting pieces of work relating to personalized recommendations and their impact on

sales concentration in the online business area. The chapter ends with a structured

comparison of related work.


58

3.1 From Technical to Transactional Evaluation of Personalization

Recommender systems are a rather new field of academic research, only becoming a focus

of interest in the mid-1990s in parallel to the propagation of the internet and e-commerce

(see e.g. Resnick and Varian, 1997). Over the next decade, numerous studies on academic

and practical implementations of recommender systems have followed (Resnick et al.,

1994, Kamba et al., 1997, Mock and Vemuri, 1997, Schafer et al., 1999, Ansari et al.,

2000).

The first evaluations of recommender systems focus on system design, technical efficiency

and predictive accuracy while developing quantitative metrics for assessment (Sarwar et

al., 2000, 2001, Burke, 2002, Herlocker et al., 2004, Adomavicius and Tuzhilin, 2005).

The basic assumption of this research (i.e. the more accurate a recommendation, the

better), has triggered numerous challengers. McNee et al. (2006), for example, points out

that these metrics do not suffice in determining the most successful recommendation

mechanism, but propagates, among others, trust and user experience as possible aspects

that make a difference.

With this new insight, IS and marketing researchers have gradually shifted their evaluation

focus of recommender systems towards a more holistic and output-oriented point of view,

also taking user behavior into account. This transactional perspective examines the impact

that personalization or personalized recommendations have on the purchase decision of

customers, their post-purchase satisfaction and the items bought. Examples for such

evaluations are two studies conducted on the basis of real-world data from Amazon

(Kumar and Benbasat, 2006) and LeShop, a Swiss grocery store (Dias et al., 2008). A

study by Pathak et al. (2010) analyzes the influence on sales and price of individual items

recommended by the system.


59

3.2 The Role of Personalization for Relationship Marketing

Increasingly, another approach for evaluating online consumer behavior is gaining

importance in the marketing arena: the relationship building perspective. In contrast to the

transactional view, “the relationship building perspective provides an affective view and

long-term standpoint to investigate the effect of personalization by the intimacy developing

process” (Li et al., 2006). Marketing literature has already coined the term relationship

marketing in the early 1980’s (Berry, 2002). Relationship marketing emphasizes the

importance of developing a long-term relationship with customers for mutual benefit

(Morgan and Hunt, 1994). Almost a decade later, Treacy and Wiersema (1993) have

published their classical article dealing with the topic of using IT for customer retention

and relationship marketing. They are the first to examine the opportunities for establishing

an intimate interaction between a company and their customers by using IT to easily

determine personal user preferences and imminently satisfying these needs based on the

extended knowledge.

Leonard Berry formulates relationship marketing as a very extensive concept that covers

the full process from attracting new customers over reinforcing the relationship to

transforming indifferent customers into loyal ones (Berry, 2002). Studies show that visitors

perceive a website as representative of the original company (Gefen et al., 2003). Thus, the

concept of relationship marketing is applicable to our online scenario without restrictions.

Attention is a scarce resource in the information-based society (Davenport, 2000). The

internet eases alternating rapidly between different websites and continuously testing

something new. Yet, some users stay on a website and return regularly. This might happen

due to missing motivation as well as inertia (Bhattacherjee, 2001). Relationship marketing

provides an alternative reasoning for this user behavior. It considers trust and commitment

as the core determinants of successful and long-term relations (Morgan and Hunt, 1994).

These two values encourage market participants to resist attractive short-term alternatives

in favor of the expected long-term benefits of staying with existing partners.


60

Thus, building continuous relationship maintenance with customers is crucial for websites

and online services. The effectiveness of recommender systems on building such

relationships has been examined by e.g. Komiak and Benbasat (2006), Liang et al. (2007)

and Liang et al. (2009). Liang et al. (2007) denote personalization as an information

technology tool to build close customer relationships. They analyze how personalization

affects the customers’ willingness of self-disclosure, their loyalty and their attitude towards

the personalized recommendations. In later studies, Liang et al. (2009) show that the

customer could experience increased intimacy when personalization techniques are

employed on a website. The intimacy positively affects the customers’ attitude towards

recommendations as well. Komiak and Benbasat (2006) show that “perceived

personalization significantly increases customers’ intention to adopt by increasing

cognitive trust and emotional trust”.

3.3 The Long Tail in e-Commerce

Besides studying the effects of recommendations on loyalty and trust, the analysis of their

impact on sales concentration and diversity constitutes another major field of current

research. With the expansion of online channels and the fact that online venues are able to

display a much larger number of items than regular stores and media, consumption moves

away from concentration on a small number of popular items only (Elberse and

Oberholzer-Gee, 2008, Zhu and Zhang, 2010). With the help of online search and filtering

tools, customers can also look for and discover the lower-selling niche products off the list

of top sellers, the so-called long tail (Anderson, 2006). The tail can be extremely long,

outplaying the traditional 80/20 Pareto principle of offline channels (Elberse, 2008,

Brynjolfsson et al., 2011). A long tail portfolio has an increased relevance for a broader

audience, since it addresses both mainstream and niche interests. The prevailing opinion of

the research community suggests recommender systems helping consumers to discover

new products and thus increasing sales diversity (e.g. Brynjolfsson et al., 2006, Fleder and

Hosanagar, 2009, Pathak et al., 2010, Brynjolfsson et al., 2011). In other words, previous


61

niche items are recommended and adjacently sold more often. A study evaluating sales

behavior for cigars shows that recommendations stimulate users to buy cigars other than

the well-known ones (Zanker et al., 2006). The extension of a top seller portfolio by a

substantial long tail thus seems to make up an appealing strategy for online service

providers as it provides potential from an up- and cross-selling perspective. Hinz and

Eckert (2010) compare multiple search systems (e.g. hit list, recommender systems) with

respect to additional consumption or substitution effects. Their research shows that online

recommendations lead to substitution, which can be beneficial for retailers when niches

yield higher margins than substituted top-sellers.

3.4 Comparison with Previous Studies

Table 8 provides an overview of how our study compares with other similar studies

conducted. In the compilation, we include studies examining the use of personalized

recommendation in an online content provisioning or e-commerce setting, and compare the

design of the study as well as their research focus. It emerges that many studies have one

or more elements in common with ours, but that only the study conducted by Liu et al.

(2010) coincides with ours in all of the examined dimensions. Thus, we close a research

gap with our controlled experiment analyzing the impact of personalized recommendations

on the revenue of an online content provider based on observed user behavior in a real-

world setting.


62

Study

Business

Context Perspective Environment Method

User

Behavior

On

lin

e

Co

nte

nt

E-C

om

mer

ce

Use

r

Pro

vid

er

Rea

l-W

orl

d

Lab

ora

tory

Co

ntr

oll

ed

Ex

per

imen

t

Oth

er

Ob

serv

ed

Ex

pre

ssed

Our study x x x x x

Adomavicius et al. (2012) x x x x x

Brynjolfsson et al. (2011) x x x x x

Chau et al. (2009) x x x x x

Chau et al. (2011) x x x x x x

Chen et al. (2010) x x x x x

Clemons et al. (2006) x x x x x

Dias et al. (2008) x x x x x

Ehrmann & Schmale (2008) x x x x x

Goel et al. (2010) x x x x x

Herlocker et al. (2004) x x x x x

Hijikata et al. (2012) x x x x

Hinz et al. (2011) x x x x x

Huang et al. (2009) x x x x x

Knijnenburg et al. (2012) x x x x x x x

Komiak & Benbasat (2006) x x x x x

Kumar & Benbasat (2006) x x x x x

Kwon et al. (2010) x x x x x

Liang et al. (2007) x x x x x x

Liang et al. (2009, 2011) x x x x x

Liu et al. (2010) x x x x x x x

Mobasher et al. (2011) x x x x x x x

Pathak et al. (2010) x x x x x

Rashid et al. (2002) x x x x x

Senecal & Nantel (2004) x x x x x

Shi & Wang (2008) x x x x x

Tan (2009) x x x x x

Tsao (2013) x x x x x x

Wang & Benbasat (2005) x x x x x

Zanker et al. (2006) x x x x x

Zhou et al. (2011) x x x x x

Zhu & Zhang (2010) x x x x x

Table 8. Overview of Studies in Literature.


63

4 Research Hypotheses

As outlined in the previous chapter, research has driven recommender system evaluation

mainly from the user perspective, their experience and satisfaction. Even though analyses

on transactions and sales diversity touch upon investigating the provider side, we identify a

research gap, namely linking user behavior induced by recommendations to tangible

benefits for the provider. In summary, we aim at contributing to the following overarching

question: how do personalized recommendations affect the observable user behavior and

consequently performance indicators relevant for website providers? In our considerations,

we put particular emphasis on content-driven websites (e.g. newspapers) basing their

revenue stream on advertising. For deriving our research hypotheses, we combine findings

from the domains of relationship marketing and recommender systems. Figure 9 illustrates

our four research hypotheses, which serve as levers to prove the causal connection between

user behavior and generated value for the provider. In the following, we describe the

rationale for selecting these hypotheses and detail their interrelations.

Figure 9. Research Framework: the Levers of Personalized Recommendations.

4.1 Whetting the User Appetite

The first step in proving a positive impact of recommender systems is to show that

consumers actually use and value recommendations offered to them. A study by Liang et

Can the provider capitalize on theusage of recommendations?

Frequency ofWebsite Visits

H2Revenue SideDuration of

Website VisitsH3

Can personalization help torightsize the provider´s portfolio?

Attractiveness ofRecommended Links

H1

Are users keen on personalizedrecommendations?

The User Perspective The Provider Perspective

CostSide

Breadth ofAccessed Item Base

H4


64

al. (2007) shows by experimental evidence that personalization can increase user

satisfaction through accurate recommendations and thus reduces information overload.

This is particularly true in case the user is looking for specific information - in contrast to

erratically scanning websites. In later studies, Liang et al. (2009) show that customers

experience increased intimacy when personalization techniques are employed on a website,

which in turn positively affects the customers’ attitude towards recommendations. We

wrap these findings into our first research hypothesis.

Attractiveness of Recommended Links (H1): Users associate a value with personalized

recommendations. When exposed to recommended links, they significantly more often

select those with a personalization attached.

4.2 Boosting Revenue

To assess the benefit personalization mechanisms imply for a content provider, we need to

take a closer look at the underlying revenue concept. Since most content websites earn

their money by placing advertisements on their pages, it is crucial to attract users and keep

them on their pages as long as possible. As the number of clicks within a certain period

primarily determines the revenue, each additional click stimulates revenue; the more traffic

a web page creates the better. Online stickiness can thus be deemed the key objective of a

website (Bhat et al., 2002). Stickiness determines the efficacy of a website and its content

in holding the visitor’s attention, i.e. visitors are finding what they expect to find

(Ghandour et al., 2010). Li et al. (2006) defines stickiness from the user perspective: the

user repetitively visits and uses a website because of a deeply held commitment towards

the website, despite external influences, which hold the potential of motivating the user to

switch. Lin (2007) adds to this definition and states that stickiness comprehends the user’s

willingness to prolong the duration of stay on the website.

This duration of stay on the website, or length of the user´s visit, is thus a key determinant

in the revenue function of the provider. Here, a visit denotes usage sequences associated


65

with regard to content and time, and also relatable to a certain user (Peterson, 2006).

Inactivity between two user actions exceeding a certain threshold implies the end of the

previous visit and the start of a new one. Efficient recommenders can help provide users

with more relevant content and thus stick users to websites and motivate them to click on

recommended links, since they fit their current preferences (Davenport and Beck, 2001).

Thus, we propose that, overall, personalized recommendations increase the attractiveness

of a website. This entices users to stay longer on the respective site and perform more

clicks. Consequently, we raise the following hypothesis:

Duration of Website Visits (H2): Personalized recommendations significantly increase

the duration of a user’s visit to a content website.

Whilst duration reflects the user’s behavior during a single visit, the factor frequency

serves as a measure to observe the long-term performance of a relationship. Li et al. (2006)

demonstrate commitment and trust positively affecting the intention of visiting a website

regularly. Recommender systems provide a benefit for the user, create a relationship with

the user, and attract users to return to the website (Schafer et al., 1999). The more often a

user visits a page, the more a recommender system learns about user preferences and,

consequently, the recommendations become more accurate. This in turn increases

stickiness and switching costs for the users (see e.g. Schafer et al., 1999, Pathak et al.,

2010). In an experiment with Google News (Liu et al., 2010), the frequency of website

visits on average increases by more than 14 % for the treatment group, which has been

exposed to a modified website with personalized recommendations, compared to the

control group left with a non-personalized website. We consolidate these findings into the

following statement.

Frequency of Website Visits (H3): Personalized recommendations significantly increase

the frequency with which users visit a content website.

As stated above, the revenue of an online content provider is mainly dependent on

advertisements. With each click, banner advertisements appear on the website. A common


66

model for website providers links charges for advertisements to the number of user clicks

on the website where the advertisement is placed. Thus, the number of clicks determines

the revenue of the provider.

A pure count of clicks, however, fails as the one success criterion for content websites,

which aim at tying their users closely and sustainably to their offering. To show this effect

of online stickiness we have broken down the total number of clicks into our two

hypotheses (H2) and (H3). With the duration of a website visit (H2) and the frequency of

returns to the website (H3), we are able to separately test whether personalized

recommendations can noticeably influence one or both of the two components and thus

serve as a lever for an online content provider to boost revenue. By quantifying the effect,

we can subsequently measure the business value of personalized recommendations in our

online content scenario. For the e-commerce realm, Dias et al. (2008) provide a similar

approach and are able to calculate the revenue increase attributed to recommendations as

0.3 %.

4.3 Cutting the Long Tail Short

The business model of an online content provider is based on one major production factor:

information. As the information value chain model suggests, this involves costs for the

creation, storage and presentation of the resource (Krcmar, 2005). Even though IT costs

are decreasing, they should not to neglected, as online retailers and especially online

content providers have to hold a myriad of content items available, from the top sellers to

the items in the long tail. In the e-commerce realm, the long tail is sometimes perceived as

a gold mine, due to the up- and cross-selling potential worth exploring. Whether this

assessment also holds true from a content provider perspective, is questionable. While

retailers aim at directing customers towards more expensive niche products, as for example

shown in a study by Hinz and Eckert (2010), online content providers have a very balanced

portfolio in terms of value. With advertisement-based revenue models, a click on a top

item is worth the same as a click on a rarely viewed one. In contrast to online providers for


67

videos or contents of special interest, the long tail of an online newspaper is of slightly

different nature. The up-to-dateness of a news article presumably serves as a much better

indicator for its popularity than the category it is placed in. We thus assume that articles

that are more recent make up the top sellers, whereas older articles constitute the long tail.

We examine the possibility to force a more narrow usage of the offering onto the user by

means of personalization. This would help the provider to intelligently rightsize the

breadth of the offered portfolio and hence control costs. In our news scenario, the aim is to

direct users to the more recent pages and away from the older ones in order to shorten the

retention period of articles. This leads to our fourth research hypothesis.

Breadth of Accessed Item Base (H4): In an online content provider setting with news

articles, personalized recommendations lead to a decreased diversity in the viewed item

base and a stronger concentration on the more recently added items.

5 Design of Empirical Study

We design the setup of our empirical study with the major objective to analyze actual user

behavior in a real-world setting. All of the previously mentioned studies on relationship

marketing and online stickiness have used the analysis of users’ intentions as a basis. This

means that the underlying data set consists of intentions regarding the future behavior,

which selected users have expressed in a questionnaire or experimental setting (see e.g.

Liang et al., 2007). To our knowledge, no study evaluates actually observable effects

between personalized recommendations and relationship building.

With our study, we assess the impact of personalized recommendations and test the

validity of the previously stated research hypotheses. In this section, we explain the setup

of the empirical study. First, we provide a description of the real-world environment and

give insights into the econometric methods for retrieving the data sample. Second, we

outline the characteristics and basic metrics of our data sample. Finally, we describe the

overarching elements of the evaluation model used for statistical testing of our hypotheses.


68

5.1 Sampling & Data Collection

Our experimental environment is the website of a major regional German newspaper7. The

website portfolio reaches from regional news to global headlines and provides news

articles across various content categories (e.g. politics, sports and culture). The website is

freely accessible for all users, without a paywall or a prior registration. Figure 10 provides

a schematic overview on the structure of a page featuring an online accessible article.

Figure 10. Schematic View on Website Featuring a News Article.

Our research interest focuses on the possibilities of a user to navigate within the online

news article portfolio. Therefore, we differentiate two types of navigation instruments that

are built into the website: links predefined by the editors of the web page, to be named

standard links, on the one hand, and personalized recommended links on the other hand.

Standard links point to other news articles based on content relation and are embedded

across various sections of the website. They reference from the text body of an article to

another article, and direct the user from the front page or the main page of a category to the

available articles. Besides these standard links, the reader is provided with a set of four

personalized recommended links. This list of links is the output of a hybrid recommender

mechanism8 that comprehends both collaborative and content-based filtering methods to

derive a user driven, personalized recommendation. The four recommended links are

7The company asked to remain anonymous. 8YOOCHOOSE (http://www.yoochoose.com) employs a hybrid recommendation engine, combining both stereotype and collaborative filtering algorithms (Inbar et al., 2008).

Page Header

Search Bar

Breadcrumb Navigation

Navigation Bar with major categories

Weather Forecast

AdvertisementNews ArticleText Body

Advertisements

News Article Headline

Current Category

Editorial Links

Links to Local Categories

Personalized Recommendations1. …2. …

3. …4. …

http://www.yoochoose.com/


69

presented to the user in a highlighted box, overwritten by “You might also want to read

…” directly below the currently visible news article.

The observation of the user behavior with regard to the usage of the provided links

constitutes the core of our experiment. Over a period of seven days, 24 hours a day,

different user activities on the website are recorded in an anonymized way.

The tracing of user clicks, or events, yields a set of data records. Each record comprises

date and type of event (i.e. click on a standard or a recommended link), the associated user

ID, and basic information about the news article, such as category and publication date.

We process the collected raw data to prepare it for evaluation. In particular, we identify

series of related events to determine user visits. A visit is defined as a series of consecutive

events produced by a single user, with the time elapsed between single events not greater

than the predefined threshold. We set this inactivity threshold to 20 minutes, as this value

has proven to be the de-facto standard (Booth and Jansen, 2008).

According to Herlocker et al. (2004), there are four dimensions for categorizing user

evaluations of recommender systems. Applying these dimensions to our setting, we

conduct an implicit study, as users are not asked to participate; the study is not executed in

a lab, but as a field experiment, it is outcome-oriented and has a short-term focus from a

time perspective.

Shani and Gunawardana (2011) differentiate between online and offline experiments.

Online experiments, i.e. with real users in an authentic setting, as in our study, are

considered more trustworthy and better suited to achieve outcomes as true-to-life as

possible (McNee et al., 2002). In addition, online experiments allow for a direct

measurement of overall system goals, such as long-term profit or user retention.

We have conducted a one-time study with a single, unrepeated data collection phase,

which we have set up as a controlled experiment (Oehlert, 2000, Gerber and Green, 2012).

Here, we examine the difference of personalized links, which the treatment group is


70

exposed to, versus randomly selected links offered to the control group. Next, we give a

more detailed description of applying the concept of controlled experiments to our setting.

The first question is the general applicability of the concept for a web environment. The

theoretical aspects of controlled experiments have been studied well in offline

environments. Practical aspects of running them in an online setting, e.g. a website, are still

under development. However, contributions of experience for online environments show,

that a transfer of the concept is generally sensible, if certain particularities are respected

and rules obeyed (Resnick et al., 2006, Kohavi et al., 2007, Crook et al., 2009, Kohavi et

al., 2009). We pick up these challenges and suitable remedial actions in more detail in the

respective context. Next, we describe the setup of the controlled experiment.

The experimental units are the users of the online newspaper. These units are considered

independent, as one unit is not dependent on the treatment of another unit. Users are solely

differentiated based on session cookies that are stored locally on their device. This allows

for recognizing recurring visits of a user. Apart from the time of the users’ first visit and

their click stream within the experiment, the users remain anonymous – no additional

demographic information is available. Because of the described characteristics, we choose

the user base to conduct the randomization procedure for our experiment. For this, we

randomly allocate users to either the treatment group or the control group by utilizing their

session IDs generated randomly by the recording system (e.g. not dependent on time). The

last digit of the ID determines the procedure of random assignment: users with an even

digit receive the treatment, whereas those with an odd digit are part of the control group.

The used mechanism, called randomization, is a known, well-understood probabilistic

scheme. It is important that the user receives a consistent experience throughout the

experiment, and this is commonly achieved through randomization based on user IDs

stored in cookies (Kohavi et al., 2009).

The randomization is also essential to allow for an unbiased inference (Shadish et al.,

2010), or, in other words, the direct observation of effects. In contrast to observational


71

studies or natural experiments, where confounding is a common problem, we can directly

assess the causal inference of changing a parameter in our experimental setting. As

randomization does not need independence, normality and the other assumptions that go

with linear models, an unbiased inference is possible.

The users do not know about the ongoing experiment or their group assignment. This

allows the exclusion of the experimenter´s bias, according to which people who knowingly

participate in a scientific experiment behave differently than normal. This is underpinned

by the fact that the newspaper website has already been featuring personalized

recommendations before the start of the experiment, thus users are familiar with this

concept.

During the experiment, the treatment group is provided with genuinely personalized

recommended links that are generated by the recommendation engine whereas the control

group receives placebo recommended links. The placebo links are arbitrarily selected from

the accessible content assortment; and in no way deducted from the previous behavior of

the user. Both types of links share the same appearance and are indistinguishable for the

user concerning the layout.

5.2 Sample Characteristics

Table 11 summarizes the key facts for our controlled experiment, split between treatment

and control group. The figures show that we nearly achieve a 50-50 split between the

groups with our randomization approach – 13,428 users (49.79 %) receive the treatment

and 13,539 users (50.21 %) the control. Thus, we can say that users are equally likely to

see each variant of our experiment.

In order to obtain a meaningful data basis, we cleanse our data from all single events,

meaning all events generated by users who only produce one page click over the whole

observation period. Table 11 shows the cleansed values, on which we base our

calculations, i.e. about 50,000 click events in each of the groups.


72

Variable Total Treatment Group Control Group

Unique Users 26,967 13,428 13,539

Recorded Events 104,985 53,038 51,497

Thereof Clicks on

Standard Links 99,843 50,018 49,825

Thereof Clicks on

Recommended

Links

5,142 3,020 2,122

Visits 48,081 23,930 24,151

Unique Articles

Accessed 24,730 15,199 15,312

Observational Period: 29/06/2011, 10am – 06/07/2011, 10am

Type of Website: Online newspaper featuring global and regional news

Location: Germany

We assume that users prohibiting cookies, which we use as session identifiers, cause most

of these single events. With the exclusion of these events, we mitigate the most substantial

bias of our study – users prohibiting cookies not being recognized as recurring users.

Another user group counted with a single click only could be users accessing the

newspaper website through search engines and then promptly exiting again.

By choosing an observation period of seven days, seasonal effects, which in our case

would be induced by a different user behavior across various days of the week, are levelled

Table 11. Descriptive Data of Empirical Study.

Figure 12. Number of Page Views per Week Day (left).

Number of Page Views per Hour of Day (right).

10

00

01

50

00

20

00

02

50

00

30

00

0

Weekday

Nu

mb

er

of

Cli

cks

Sun Mon Tue Wed Thu Fri Sat

02

00

04

00

06

00

08

00

01

00

00

12

00

01

40

00

Time of Day

Nu

mb

er

of

Cli

cks

0 2 4 6 8 10 12 14 16 18 20 22


73

out and can thus be excluded as an influential factor on the results. These so-called day of

week effects are a well-known phenomenon in web-based experiments (Kohavi et al.,

2009). They are also visible in the evaluation of our experimental data, as shown in Figure

12. Influential effects created by access from abroad, e.g. producing an overlay of the day

of week effect by time difference, are negligible, as the vast majority of users of the

newspaper web page are from Germany. In addition, no extraordinary national or global

events happened within our observed time span, which therefore gives no reason for a

distorted user behavior regarding the articles accessed.

5.3 Modeling Approach

The controlled experiment serves as a means to examine the impact of true personalization.

In other words, it considers the question whether the user recognizes an advantage from

true, personalized recommendations and thus visits the newspaper web page longer and

more often. Moreover, changes regarding the accessed item base are evaluated.

The setup as controlled experiment allows to compare treatment and control group

regarding key performance indicators by variation of the parameter of true personalization.

Within the prevailing Fisher-Neyman-Rubin framework of causal inference, causal effects

are defined as comparisons of potential outcomes under different treatments (Druckman et

al., 2011). Since it is impossible to observe multiple outcomes (realizations of the variable

of interest under different treatments) for any given unit, we approximate the hypothetical

treatment effect by comparing averages of groups. Hence, measuring the difference in

outcomes of treatment and control provides us with the average treatment effect (ATE)

that is regarded as a valid test statistic in a controlled experiment (Gerber and Green,

2012). We measure the outcomes with an Overall Evaluation Criterion (OEC) (Kohavi et

al., 2009). Any differences in the OEC are inevitably the result of the applied treatment,

establishing causality. The Stable Unit Treatment Value Assumption (SUTVA) is valid as

well, since the outcome observation on one unit, i.e. a user, is unaffected by the particular

assignment of treatments to the other units.


74

Our evaluation model formalizes the defined research hypotheses to allow for an

evaluation based on the empirical study data. For each of our hypotheses, we require

individually developed metrics (Newman, 1997) to measure their respective outcome.

Hence, the evaluation model consists of five parts that leverage appropriate key metrics of

a content-driven website to address our research hypotheses: (H1) attraction of

recommended links, (H2) duration of website visits, (H3) frequency of website visits, and

(H4) breadth of accessed item base.

To set up the model, we utilize the rich toolbox of web analytics. The topic of web

analytics is rather new and scientifically still in its infancy. Recently, a few reference

books dealing with web analytics have been published (e.g. Peterson, 2006). In addition, a

standardization body, the Web Analytics Association, has compiled a compendium that

provides definitions for most web metrics (Web Analytics Association, 2008). Schaupp et

al. (2006) state that numerous web metrics have been defined to enable website operators

to capture the effectiveness of the embedded features in a website, and not only to track the

navigational patterns of website users.

For each of our research hypotheses we define a suitable OEC. We then test the

significance of the difference in means between treatment and control group using

significance tests, such as Student’s t-test and Wilcoxon rank-sum test, depending on the

underlying data characteristics and distribution function.

6 Empirical Results

In this chapter, we present the results of testing our research hypotheses against the data of

our empirical study. For each of the hypotheses, we explain the used metric and the setup

of our evaluation model including selection of the OEC, before providing and discussing

the results. Finally, we present our view on the managerial implications relevant for

practical implementations of personalized recommendations.


75

6.1 Attraction of Recommended Links (H1)

To evaluate the attraction of recommended links, we compare the share of clicks that users

conduct on recommended links with those on standard links. Therefore, we revert to a

well-defined web metric, namely the click-through rate. The click-through rate defines the

degree of attraction that a link exhibits for a user. It is typically used to measure the

success of online advertisement banners by relating the number of clicks on a banner to the

overall number of banner displays (Richardson et al., 2007).. We define our metric

analogously as click-through rate of the recommended links and set the number of clicks

on a recommended link in relation to the total number of clicks. This total corresponds to

the number of displays of recommended links in our experimental setting – with each click

on a news article the user is exposed to four recommendation links. Since we compare

relative values between treatment and control group, we do not include the factor four (for

the four recommended links) in our formula. A key concept for our evaluation is a visit that

denotes a series of related events invoked by a single user. Let 𝑉𝑖𝑗 denote the 𝑖-th visit of

user 𝑗. The click-through rate of recommended links in all visits 𝑉𝑖 of user 𝑗 is defined as

𝑝𝑗 =∑ 𝐶𝑅𝑖𝑖

∑ (𝐶𝑆𝑖+𝐶𝑅𝑖)𝑖. (1)

In the following, we a user 𝑗 resembles a member of the treatment group and 𝑗̂ a member

of the control group respectively.

Figure 13. Bar Plot Comparing the Means of Attraction of a Recommended Link for the

Experimental Groups. Error Bars Indicate the 95 % Confidence Interval of the Standard

Error of the Mean.

0.000

0.005

0.010

0.015

control treatment

Experimental Group

Mean o

f A

ttra

ction o

f Lin

k


76

A look at the bar plots in Figure 13 suggests a higher attraction of the recommended links

in the treatment group. We build upon this indication and use the defined metric 𝑝𝑗 as

evaluation criterion for assessing the treatment effect between treatment and control group.

The average treatment effect is formulated as difference in means between treatment and

control

ATE1 = 𝜇T − 𝜇C =∑ 𝑝𝑗𝑗

𝑚−∑ 𝑝�̂��̂�

𝑁 −𝑚.

(2)

The variance of the ATE and the corresponding confidence interval are defined by

𝜎ATE1 = √𝜎𝑇2

𝑚+

𝜎𝐶2

𝑁−𝑚

(3)

and

𝐶𝐼ATE1 = ATE1 ± 1.96 ∙ 𝜎ATE1. (4)

We apply the threshold value of 1.96 corresponding to a 95 % confidence interval of the

standard normal distribution. To test the hypothesis, a significance level or p-value is the

de-facto analysis measure (Shani and Gunawardana, 2011), which we will deploy as well.

However, the distribution of the users’ conversion rate values is positively skewed, and not

normally distributed. A simple transformation, i.e. taking the logarithm, is not possible, as

many users hold a conversion rate of zero, meaning they have never clicked on a

recommended link. The common analysis tool for simple experimental designs with one

control group (Kirk, 1982), the student´s t-test, can thus not be applied. Instead, we

conduct a one-sided Wilcoxon rank-sum test and test 𝜇C < 𝜇T; (see Shani and

Gunawardana, 2011).

The p-value (p< 2.2e-16) suggests supporting the alternative hypothesis; p=0.05 serves as

threshold (Shani and Gunawardana, 2011). The null hypothesis can thus be rejected.

Hence, users in the treatment group show a significantly higher rate of clicks on

recommendations than users in the control group. The treatment effect as difference in

means computes as 𝐴𝑇𝐸 = 𝜇T − 𝜇C = 0.0033, the standard deviation as 𝜎ATE = 0.0005.

A view on the confidence interval of the absolute treatment effect, i.e. 𝐶𝐼abs(𝐴𝑇𝐸) =


77

[0.0024, 0.0043], confirms the result of the significance tests. Since zero is not part of the

interval, the null hypothesis can be rejected (Kohavi et al., 2009).

The absolute probability for clicking a recommended link in the treatment group is

relatively small. This can be easily explained by the unfavorable positioning of the

recommendations on the very bottom of the web page. Since the absolute effect is so small,

we transform it into a percentage effect, which yields a more intuitive meaning than the

absolute difference (Crook et al., 2009, Kohavi et al., 2009). This transformation is

straightforward. The percentage treatment effect computes as

𝐴𝑇𝐸pct =𝜇T−𝜇C

𝜇C∙ 100. (5)

Calculating the corresponding confidence interval of the ATE requires slightly more effort.

For this, we use the Fieller method to calculate the confidence interval of the quotient of

the two means (Motulsky, 2013). This method requires the two means to be Gaussian

variables, which is true for 𝜇𝑇 and 𝜇𝐶. First, we define the auxiliary variable 𝑔 as

𝑔 = (1.96 ∙�̂�T

𝜇T)2.

(6)

Since 𝑔 takes on a value less than 1.0, we can determine the confidence interval for the

quotient. First, we calculate the standard error of the quotient

�̂�pct =𝜇T

(1−𝑔)∙𝜇C∙ √(1 − 𝑔) ∙

�̂�T2

𝜇T2 +

�̂�C2

𝜇C2 .

(7)

Based on this the confidence interval follows as

𝐶𝐼pct =𝜇T

(1−𝑔)∙𝜇C± 1.96 ∙ �̂�pct. (8)

The confidence interval is not necessarily distributed symmetrically around the quotient.

𝐴𝑇𝐸pct computes for our scenario as 31.85 %. Clicking on a recommended link thus bears

a 30 % higher probability in the treatment group compared to the control group. Table 14

contains all details concerning standard error and confidence interval. In summary, we can

show that through a genuine personalization the attractiveness of recommended links

increases. The share of clicks on such links is significantly higher than on placebo links.


78

The users thus seem to recognize the relation of the recommendations to the currently read

news article and to their personal interests.

Experimental Groups Average Treatment Effect

Variable: 𝒑𝒋 Treatment ΩT1 Control ΩC1 Absolute Percentage

Mean 𝜇 13.84e-03 10.50e-03 3.34e-03 31.85 %

Standard Error

of Mean �̂�

0.35e-03 0.32e-03 0.48e-03 5.24 %

Standard

Deviation 𝜎

73.64e-03 65.58e-03 --- ---

Variance 𝜎2 5.42e-03 4.30e-03 --- ---

Confidence

Interval (95 %)

[13.14e-03,

14.54e-03]

[9.87-03,

11.12e-03]

[2.41e-03,

4.28e-03]

[22.04 %,

42.60 %]

6.2 Duration of Website Visits (H2)

The selection of appropriate performance measures for examining website transactions is

not straightforward (Turban, 2010). However, two measures, namely time spent per visit

and number of page views (Huang et al., 2009), are considered to be most suitable to

determine the duration of a visit to a website (Moe and Fader, 2004, Montgomery et al.,

2004). We restrict our model to the latter, as time spent per visit can only be determined

imperfectly in our study setup. This is because the website does not receive a trigger when

the user concludes a visit and leaves the website. Hence, retention time of the user at the

last page of a visit is not captured.

The number of viewed pages (each featuring one news article) within a single visit of a

user determines the length of a visit. Pages can either be reached by clicking a standard

link or a recommended link. Let 𝐶𝑅𝑖𝑗 and 𝐶𝑆𝑖𝑗 denote the number of recommended and

standard links that a user 𝑗 clicks during visit 𝑉𝑖𝑗.

Hence, the length of a visit is defined as

𝑙𝑒𝑛𝑔𝑡ℎ(𝑉𝑖𝑗) = 𝐶𝑅𝑖𝑗 + 𝐶𝑆𝑖𝑗. (9)

Table 14. Average Treatment Effect for Attraction of Recommended Links with the OEC

“Click-through Rate for Recommended Links” (Hypothesis H1).


79

We model the effect of using recommended links on the distribution of length of visits. To

compare treatment and control group, we use the defined metric 𝑙𝑒𝑛𝑔𝑡ℎ(𝑉𝑖) as evaluation

criterion to assess the average treatment effect. We measure the average treatment effect as

difference in means between treatment and control group

𝐴𝑇𝐸2 = 𝜇T − 𝜇C =∑ 𝑙𝑒𝑛𝑔𝑡ℎ(𝑉𝑖𝑗)𝑖𝑗

𝑚−∑ 𝑙𝑒𝑛𝑔𝑡ℎ(𝑉�̂��̂�)�̂��̂�

𝑁 −𝑚

(10)

Variables 𝑖 ̂ and 𝑗̂ indicate visits and users for the control group. The average treatment

effect 𝐴𝑇𝐸2computes as 0.07. As, once again, the absolute value is quite small, we bring in

the percentage effect. With this, we receive an average increase of the visit length of

3.43 % in the treatment group versus the control group. For significance testing, we use the

non-parametric Wilcoxon rank-sum test with continuity correction, as the data is not

normally distributed, but positively skewed (see Figure 15).

Figure 15. Distribution of Length of Visits in Treatment Group. Data is Positively Skewed.

The null hypothesis for the test is defined as true location shift being equal to zero, and the

alternative hypothesis as true location shift greater than zero. The test yields a p-value that

is considerably smaller than the standard reference p-value of 0.05; thus, the result is

highly significant and the null hypothesis can be rejected. Detailed results regarding the

treatment effect are stated in Table 16. As indicated, the length of a visit is significantly

higher in the treatment group than in the control group; true recommendations positively

influence people to stay longer on the website.

01

00

00

20

00

03

00

00

40

00

0

Length of Visit (in Clicks)

Fre

qu

en

cy

1 6 11 16 21 26 31 36


80


Variable: 𝒍𝒆𝒏𝒈𝒕𝒉(𝑽𝒊𝒋) Treatment Control Absolute Percentage

n 23,930 24,151 --- ---

Mean 𝜇 2.22 2.15 0.07 3.43 %

Standard Error of Mean �̂� 0.01 0.01 0.02 0.80 %

Standard Deviation 𝜎 1.94 1.75 --- ---

Variance 𝜎2 3.78 3.07 --- ---

Confidence Interval (95 %) [2.19, 2.24] [2.13, 2.17] [0.03, 0.10] [1.50 %, 4.61 %]

After analyzing the impact of personalized recommendation on the behavior of treatment

and control group, we now take a closer look at the situation within the treatment group.

Even though we have conducted a single study, it actually contains two experiments in

one. According to Shani and Gunawardana (2011), these two experiments can be

categorized as between subjects (i.e. our controlled experiment comparing treatment and

control group) and within subjects. The comparison within subjects points to the analysis

of the user behavior within the treatment group, namely the impact of recommender

systems as such. We examine the influence of the relative position of the first click on a

recommended link on the overall duration of a visit. Thus, we define the length of a visit

𝑙𝑒𝑛𝑔𝑡ℎ(𝑉𝑖) as our response variable. As independent variables, we introduce dichotomous

dummy variables 𝐺𝑟𝑜𝑢𝑝𝑧, where 𝑧 denotes one of 6 levels, i.e. 𝑧 ∈ [0; 5], 𝑧 ∈ ℕ. Each of

the levels indicates a range within the total length of the visit, with 1 corresponding to the

first 20 % of the visit length, 2 corresponding to the range between 20 % and 40 %, and so

on. The value assigned to 𝑧 depends on the range containing the first click on a

recommended link. With 𝑧 = 0 we refer to the visits where no recommend link appears.

Next, we take a look at generic factors that may affect user behavior in our setting and

hence may serve as a control variable. First, the composition of the visitor base may be

subject to change over time. More experienced users, i.e. visitors that have been active on

the site for a longer time, are more likely to use advanced features of the website. We

Table 16. Average Treatment Effect for Duration of Website Visits with the OEC “Length of

Visit in Number of Clicks per Visit” (Hypothesis H2).


81

capture this by defining the experience factor 𝐸𝑋𝑃𝑖𝑗. The time interval between the user’s

first visit (𝑉0𝑗) to the website and the currently observed visit 𝑉𝑖𝑗 expresses user 𝑗’s

experience with the website:

𝐸𝑋𝑃𝑖𝑗 = 𝑡(𝑉𝑖𝑗) − 𝑡(𝑉0𝑗). (11)

The function 𝑡(𝑉𝑖𝑗) returns the starting time of visit 𝑉𝑖𝑗. A higher value of 𝐸𝑋𝑃𝑖𝑗 indicates

that a user has been active on the website for a longer period already. Moreover, we define

an additional dichotomous dummy variable 𝐹𝑉𝑖𝑗 that indicates whether visit 𝑉𝑖𝑗 is the first

visit of user 𝑗 to the website (i.e. 𝑖 = 0). New users are assumed to be more curious and

thus browse the page for a longer time. To control seasonal effects, we define two

additional variables that resemble the weekday (𝐷𝑎𝑦𝑖𝑗) and time of the day (𝑇𝑖𝑚𝑒𝑖𝑗) when

visit 𝑉𝑖𝑗 is started. The expected linear relationship with all the mentioned components

results in the following formula:

𝑙𝑒𝑛𝑔𝑡ℎ(𝑉𝑖𝑗) = 𝛽0 + 𝛽1 ∙ 𝐺𝑟𝑜𝑢𝑝0 + 𝛽2 ∙ 𝐺𝑟𝑜𝑢𝑝1 + 𝛽3 ∙ 𝐺𝑟𝑜𝑢𝑝2 + 𝛽4

∙ 𝐺𝑟𝑜𝑢𝑝3 + 𝛽5 ∙ 𝐺𝑟𝑜𝑢𝑝4 + 𝛽6 ∙ 𝐺𝑟𝑜𝑢𝑝5 + 𝛽7 ∙ 𝐸𝑋𝑃𝑖𝑗

+ 𝛽8 ∙ 𝐹𝑉𝑖𝑗 + 𝛽9 ∙ 𝑇𝑖𝑚𝑒𝑖𝑗 + 𝛽10 ∙ 𝐷𝑎𝑦𝑖𝑗 + 𝜖𝑖𝑗.

(12)

Since a Breusch-Pagan test indicates heteroskedasticity, we conduct a linear regression

with White’s robust standard errors. We examine the model fit with a forward integration

and check, if each of the added variables delivers a marginally significant contribution.

The corresponding variance analysis is illustrated in an ANOVA table (Table 17).

ANOVA Df Sum Square

Mean Square

F

Group 5 19,221 3844.2 1356.436***

FV 1 3,174 3,174.1 1119.993***

EXP 1 86 85.9 30.325***

Time 1 168 167.9 59.256***

Day 1 5 4.9 0.1879

Residuals 23,920 67,790 2.8 ---

Model H2, response variable 𝑙𝑒𝑛𝑔𝑡ℎ(𝑉𝑖𝑗)

Significance levels (2-tailed): *** p<0.001; ** p<0.01; * p<0.05

Table 17. Results of Analysis of Variances (ANOVA) Used for Coefficient Testing.


82

We exclude variable 𝐷𝑎𝑦𝑖𝑗 from the model, as its contribution is not significant. Hence, it

has no explanatory value in analyzing the length of a visit. The results of the linear

regression are depicted in Table 18, both for the minimal model (H2.1) as well as the

interim steps and the full model (H2). The resulting explanatory value 𝑅2 of 0.25 is only of

medium size (Cohen, 1992). This is, however, no surprise since user behavior is potentially

also highly dependent on demographic factors, which we cannot capture in our setting. For

a real-world study with anonymous users, this value seems nevertheless quite substantial.

Method: OLS with robust standard errors

Model H2.1 Model H2.2 Model H2.3 Full Model H2

𝛽0 Intercept 1.99***

(0.01)

1.83***

(0.02)

1.99***

(0.03)

1.96***

(0.03)

𝛽1 RL in 0-20 % 8.50***

(0.65)

8.61***

(0.65)

8.62***

(0.65)

8.62***

(0.65)

𝛽2 RL in 20-40 % 4.07***

(0.19)

4.11***

(0.18)

4.11***

(0.18)

4.11***

(0.18)

𝛽3 RL in 40-60 % 2.22***

(0.12)

2.27***

(0.12)

2.27***

(0.12)

2.27***

(0.12)

𝛽4 RL in 60-80 % 1.93***

(0.06)

1.91***

(0.06)

1.91***

(0.06)

1.91***

(0.06)

𝛽5 RL in 80-100 % 0.80***

(0.10)

0.79***

(0.10)

0.79***

(0.10)

0.79***

(0.10)

𝛽6 EXP --- -5.04e-04***

(7.58e-05)

-5.29e-04***

(7.61e-05)

-5.28e-04***

(7.61e-05)

𝛽7 FV --- 0.76***

(0.03)

0.76***

(0.03)

0.76***

(0.03)

𝛽8 Time --- --- -0.09***

(0.01)

-0.09***

(0.01)

𝛽9 Day --- --- -8.033e-03

(5.686e-03)

---

Adjusted R square 0.2124 0.2483 0.2502 0.2502

F 384.4*** 395.2*** 309*** 346.9***

Observations: 23,930

Dependent variable: length(𝑉𝑖𝑗)

Significance levels (2-tailed): *** p<0.001; ** p<0.01; * p<0.05

The results of the linear regression show the particularly high impact on the visit’s length

by a click on a recommended link when conducted within the first 20 % of the visit time.

In contrast to a visit without any recommended link (on average around two clicks), such

Table 18. Results of Regression Analysis Regarding the Impact of Personalized

Recommendations on the Length of a Visit (Standard Errors are Stated in Parentheses).


83

an event boosts the length of the visit by 8.6 clicks. This positive effect caused by the

recommender declines towards the end of the visit (see values 𝛽1 to 𝛽5). Another

observable outcome is the decreasing length of visits with growing user experience. This

can be explained with a partial loss of curiosity. Moreover, users are more familiar with the

website and probably tend to arrive at interesting articles in a more direct way, for example

by a targeted click on a standard link.

6.3 Frequency of Website Visits (H3)

For hypothesis 3, we examine the frequency of website visits. Visits per unique visitor has

been established as a prevalent measure (Bhat et al., 2002), determining the absolute

number of visits a single user is conducting. Such an absolute metric does not entirely suit

our purpose, as there is a continuous stream of new visitors to the page, also observed

during the course of our experiment. These new users consequently have less time to

conduct visits than the established users. To account for that, we include a time-dependent

component into our metric. The component 𝜏𝑗 defines the time interval available to user 𝑗

to conduct visits

𝜏𝑗 = {𝑡(𝑠𝑡𝑎𝑟𝑡) − 𝑡(𝑉0𝑗), 𝑡(𝑉0𝑗) > 𝑡(𝑠𝑡𝑎𝑟𝑡)

𝑡(𝑒𝑛𝑑) − 𝑡(𝑠𝑡𝑎𝑟𝑡), 𝑜𝑡ℎ𝑒𝑟 ,

(13)

with 𝑡(𝑠𝑡𝑎𝑟𝑡) denoting the start and 𝑡(𝑒𝑛𝑑) the end of our observational period. The

metric visits per unique visitor per time interval is thus given as

Λ𝑗 =max𝑗𝑖

𝜏𝑗.

(14)

We use this measure as overall evaluation criterion for our research hypothesis H3 and

define the treatment effect as follows

𝐴𝑇𝐸3 = 𝜇T − 𝜇C =∑ Λ𝑗𝑗

𝑚−∑ Λ�̂��̂�

𝑁−𝑚. (15)

We achieve a value of 0.20e-05 for ATE3.


84

A summary of results is given in Table 19.

Experimental Groups Absolute

Treatment Effect Treatment Control

N 13,539 13,428 ---

Mean 𝜇 1.47e-05

0.053*

1.26e-05

0.046*

0.20e-05

0.007*

Standard Error of

Mean �̂�

97.44e-05 88.51e-05 1.32e-03

Standard Deviation 𝜎 0.11 0.10 ---

Variance 𝜎2 0.01 0.01 ---

Confidence Interval

(95 %)

[-1.90e-03,

1.92e-03]

[-1.72e-03,

1.74e-03]

[-2.58e-03,

2.58e-03]

The conducted non-parametric Wilcoxon rank-sum test delivers a p-value which slightly

exceeds the threshold (p=0.07), thus the null hypothesis for a significance level of 5 %

cannot be rejected. Despite the absence of a significant outcome, a positive trend in favor

of the treatment group is visible; thus, users provided with authentic personalized

recommendations tend to return to the website more often.

6.4 Breadth of Accessed Item Base (H4)

Next, we examine the influence of personalized recommendations on the breadth of the

accessed item base. For this, we evaluate the age of articles that users read. We now look

at single clicking events on the news pages in the portfolio. With the access of an article 𝑦

being denoted by the event 𝑥, the age of an article is defined as the difference between its

publishing date 𝑡(𝑦) and the access date 𝑡(𝑥),

𝑎𝑥 = 𝑡(𝑥) − 𝑡(𝑦). (16)

To compare the experimental groups, we start by looking at the maximum age of accessed

articles9 corresponding to the maximum retention time of articles on the news website.

9 Measured with 3,852 days (equivalent to 10.5 years) in both groups.

Table 19. Average Treatment Effect for Frequency of Website Visits with the OEC “Visits per

Unique Visitor per Second”. To Achieve Exact Results, the Metric is Measured in Visits per

Second; the Values Marked with * are the Corresponding Values for Visits per Hour.


85

Analyzing the cumulated percentage of clicks in relation to the particular age of the item

reveals differences. Figure 20 shows a steeper curve for the treatment group. This indicates

a higher concentration regarding the portfolio breadth for this group. In the treatment

group, 58.77 % of articles are younger than a day when accessed, while in the control

group only 56.72 % of clicks are targeted at these latest amendments to the article base.

This trend is even stronger at the 90 % mark. The treatment group reaches this threshold

with articles younger than 42 days, for the control group all articles up to an age of 56 days

have to be taken into account.

Figure 20. Distribution of Cumulated Percentage of Clicks in Relation to Age of Accessed

Items – Comparison of Treatment and Control Group.

We further examine this indication and use the average age as overall evaluation criterion

for the comparison of the difference in means between treatment and control group for our

research hypothesis H4. We define the treatment effect as follows:

𝐴𝑇𝐸4 = 𝜇T − 𝜇C =∑ 𝑎𝑥𝑥

𝑚−∑ 𝑎�̂��̂�

𝑁−𝑚. (17)

Variable 𝑥 denotes the events of the control group. The average treatment effect 𝐴𝑇𝐸4

computes as -3.76 days. The details are shown in Table 21.

Again, we conduct a one-sided Wilcoxon rank-sum test and evaluate if the age of the

accessed news articles is significantly smaller in the treatment group. The test reveals a

high significance with p=7.62e-09. The null hypothesis can thus be rejected and the

alternative hypothesis, i.e. true location shift is less than zero, is supported. To sum up, we


86

have shown the significant impact of personalized recommendations on the accessed item

base. Hence, a website provider can use recommendations as an instrument to optimize the

run costs of the website by providing a leaner item portfolio.


Variable: 𝒂𝒙 Treatment Control Absolut Percentage

N 53,038 51,947 --- ---

Max 3,852 3,852 --- ---

Mean 𝜇 77.56 73.80 -3.76 -5.09 %

Standard Error of Mean �̂� 1.47 1.54 2.14 2.96 %

Standard Deviation 𝜎 339.11 350.98 --- ---

Variance 𝜎2 114,995.11 123,186.96 --- ---

Confidence Interval (95 %) [74.67, 80.45] [70.78, 76.82] [-7.94, 0.42] [-11.09 %, 0.05 %]

6.5 Measuring the Overall Business Value

The results of the empirical study show that personalization can influence the user

behavior in order to achieve a positive impact on the key performance indicators. Which

business value can be derived from these improved performance metrics remains the

question be answered. We address this question by considering two perspectives: the

revenue and the cost perspective. As already described for the development of research

hypotheses H2 and H3, the two components duration and frequency of website visits

constitute the basis for the revenue of a content website relying on advertisements. To

show the effect of personalization on the revenue, we take the treatment effects 𝐴𝑇𝐸2 and

𝐴𝑇𝐸3. Even though the test of 𝐴𝑇𝐸3 does not produce a significant result with the

significance level of p=0.05, we still assume that the indicated tendency is notable. With

the (simplified) assumption of the revenue of a content website being directly linked to the

number of clicks, the following equation defines the revenue stream 𝑟 of a content

provider:

𝑟 = ∑ (𝑙𝑒𝑛𝑔𝑡ℎ(𝑉𝑗) ∙ Λ𝑗)𝑗 ∙ 𝜔 ∙𝐶𝑃𝑀

1000, (18)

Table 21. Average Treatment Effect for Breadth of Accessed Item Base with the OEC

“Average Age of Accessed Items” (Hypothesis H4). The Average Age is Measured in Days.


87

where 𝜔 denotes the number of advertisements shown per news site, and 𝐶𝑃𝑀 the

monetary value a content provider receives per 1,000 displays of the advertisement. In

addition, the results of research hypothesis H1 show an increased attractiveness of

recommended links in the treatment group. We refrain from including this into the overall

business value consideration as the effect potentially overlaps with the two other treatment

effects 𝐴𝑇𝐸2 and 𝐴𝑇𝐸3. The profit achievable with personalized recommendation is

computable as the product of the treatment effects across the number of distinct users 𝑁:

𝑟Δ = 𝑁 ∙ 𝐴𝑇𝐸2 ∙ 𝐴𝑇𝐸3 ∙ 𝜔 ∙𝐶𝑃𝑀

1,000. (19)

For our scenario we have the following configuration: We assume that per news article

three advertisements are shown (i.e. 𝜔 = 3) and a price of 𝐶𝑃𝑀= € 20 for 1,000 displays

can be obtained, a common market value for a regional news site. With N=26,697 users, as

in our experiment, this yields an additional revenue of € 31 per hour for the provider. On a

yearly basis, this corresponds to an achievable profit of € 272,080 through personalized

recommendation.

However, personalization cannot only boost the revenue – it can also serve as a valuable

lever to decrease costs of the portfolio. As the calculations of hypothesis H4 with a

negative 𝐴𝑇𝐸4 show, the genuine recommendations produce a more concentrated clicking

pattern than placebo recommendations. Consequently, a provider would lose fewer clicks

by limiting its offer to a leaner portfolio of articles when applying personalized

recommendations. As a basis for deduction of a real savings potential, a more detailed

knowledge about the cost structure of an online news provider is required, specifically the

costs involved for maintaining the article portfolio as well as storage and presentation, e.g.

by means of a content management system. If in our scenario, the provider would restrict

its portfolio to the news articles accounting for 90 % of the clicks, all articles older than 42

days could be removed without causing too much loss on the user side – instead of keeping

all articles up to an age of 10 years.


88

6.6 Managerial Implications

Overall, our results show clear benefits of applying a personalized recommendation

mechanism for our news website provider. True personalization sparks users’ interest and

is able to influence their behavior to create a positive effect on the key performance

indicators of the content provider. The integration of a recommender mechanism into a

content website is rather simple. A multitude of service providers offer such algorithms as

on-premise or on-demand solutions.

We could verify that if personalized, a recommendation link excels a standard link in terms

of its attractiveness. When considering the unfavorable positioning of the

recommendations at the bottom of the page in our research setting, it seems obvious that a

more prominent placement of the links could even increase the absolute number of clicks

on recommendation links compared to standard links. This correction thus yields untapped

potential for the website provider to capitalize on the effect of recommendation.

In order to provide a recommendation mechanism with the full-fledged picture on the users

including all their history of visits to the page, it is essential that all clicks are

unambiguously matched to the correct user. Only then, a recommendation algorithm can

learn about the user preferences and tailor the recommendations. The usage of cookies as

user identification mechanism unfortunately entails a number of imponderables. First,

users prohibiting persistent storage of cookies cannot be recognized correctly when

returning to the website, resulting in creation of a new user ID for each visit. Second, if

multiple users are surfing the internet from the same device, they are not distinguishable as

separate users and will receive the same user ID. Third, single users that deploy more than

one device (e.g. desktop computer and tablet computer) generate multiple, separated

sessions that are not linked to each other. Internet pages with a mandatory user registration

prior to accessing the content, or web shops, which have a checkout procedure including

registration in place, are less dependent on the unhampered usage of cookies by their users.


89

7 Summary and Outlook

In this part, we examine the impact of personalized recommendations on key performance

metrics of a content website. Our analysis bases on a controlled experiment in a real-world

setting with a German online newspaper. With this, we are among the first studies to

evaluate real user behavior in a setting with personalized recommendations in the online

content realm.

The empirical results show the positive impact from a provider perspective. Users

recognize a higher value in the setting with true personalization than in the placebo setting

– the probability of clicking on a respective link is about 5 % higher. The duration of a

visit can also be increased significantly, on average a visit extends by 3.4 %. An even more

considerable effect is visible when comparing visits containing a click on a recommended

link with those without such a link. Here, the average length of the visit increases by factor

five. In contrast to the findings by Hinz et al. (2011), our results do not indicate a

substituting effect attributable to recommendations, as users in fact view more pages, not

just other ones. In addition, the frequency of website visits of a user tends to slightly

increase (ca. 15 %), however we cannot prove significance here. The breadth of the

accessed item portfolio is also controllable by means of personalized recommendations to

the benefit of the provider. In our setting, the average age of an accessed news article

decreases by more than 5 %, while the concentration across the accessed portfolio

increases. The provider would thus not lose as many clicks as without recommendations, if

the article portfolio was reduced. We condense the mentioned effects into a business value,

which yields substantial financial benefits for the online content provider. The provider is

able to increase revenues by around € 0.3 M annually using personalized

recommendations. Figure 22 summarizes the proven effects of our assessment.


90

Figure 22. Summary of Findings Regarding the Impact of Personalized Recommendations,

Following the Depiction by Hinz et al. (2011)

(*only Trend, no Statistically Significant Result Achieved in Experiment).

The setup of our empirical study exhibits two main limitations, which might affect the

general transferability of our results to a broader context.

The first limitation is the narrow time horizon of our study, which closely ties in with the

nature of our experiment as a real-world study. As mentioned, this authentic setting entails

multiple advantages; on the other hand one can only run this experiment for a limited

amount of time. Scientific interest needs to keep a balance with potential drawbacks for the

web site provider due to dissatisfied users. Users do not (and should not) know about the

ongoing experiment, and thus might perceive the placebo recommendations as a quality

issue on the newspaper web site. Consequently, we can derive conclusions only valid for a

short time horizon from this study, and are not able to investigate time series and long-term

effects. Leaving the commercial considerations aside, a longer running controlled

experiment is perfectly possible to conduct. Without a controlled experiment, results are

subject to regression analysis, which does not yield the same clear inference of causality.

The specific application context of our study shows the other limitation. We deliberately

selected an online content provider, and not an e-commerce provider, as we found very

little research in this area so far. However, not all content providers are similar. In our

setting, we examine the use of personalized recommendations for news articles, where up-

Demand per News Article

Visited News Articles (Descending Order According to Demand)

Increased attractiveness of linksIncreased length of visitsIncreased frequency of visits*

Higher concentration of clicks Reduced average age of visited articles


91

to-date content is fundamental. This characteristic becomes particularly apparent in the

results of our hypothesis 4. In case of provisioning of other text content, such as cooking

recipes, or even other content types, such as videos, music or pictures, there might be other

drivers, which determine the default distribution pattern of user interest across the

portfolio. Hinz et al. (2011) analyze such drivers for a video-on-demand platform, and

identify, among others, the quality of blockbusters as an important factor. Generally, we

assume that for content other than news, categories are much more important than the age

of the content. Even transferring our results to other news content providers could be

difficult, if they apply a revenue model not based on advertisements, but use the

increasingly popular paywall concept or a freemium model for example.

We consider our study as a starting point to further explorations of personalized

recommendations in the world of online content provisioning. Our ideas for further

research partially pick up the issues described as limitations of our study.

In order to obtain a better picture on how the relationship between the recommender

system and the user develops, research conducted over a longer time span would be

beneficial. Thus, potential quality increases of true personalized recommendations and

subsequently an even stronger contrast to placebo recommendations could be verified. We

assume a positive relation between longer observation of user preferences and higher

accuracy of given recommendations. User satisfaction and thus the affinity of using

recommended links increases.

With the described assumed difference for varying types of content provided on a website,

a variation of the application scenarios of personalized recommendations could be an

interesting setup for further academic research: Are differences between true and random

recommendations similarly significant for other content as well? What drives users’

navigational pattern within the portfolio?

However, already the specific context of online news leaves enough room for further

research. In our setting, we could only test one type of a hybrid recommendation


92

algorithm. Comparing results for different types of recommenders, all other parameters left

unchanged, might add further evidence from observed user behavior to the field of

transactional recommender system evaluation. The results could provide valuable practical

advice for news content providers in selecting the best-fitting recommendation algorithm

for their website.

In addition, it seems appealing to evaluate the impact and success of personalized

recommendations on other revenue models besides advertisement, such as paywall or

freemium concepts. Another element giving further insight into the analysis of user

behavior is the inclusion of demographic data about users. Are younger peoples more

inclined to follow recommendations than the elderly? Does familiarity with the internet

stimulate or hinder the success of recommender systems? To acquire such demographic

data, it is though essential to have a mandatory registration before users can access the

content.


93

8 References

Adomavicius, G., Bockstedt, J., Curley, S., and Zhang, J. (2012). Effects of Online

Recommendations on Consumers’ Willingness to Pay. Proceedings of the 6th ACM

Conference on Recommender Systems (RecSys).

Adomavicius, G. and Tuzhilin, A. (2005). Toward the Next Generation of Recommender

Systems: a Survey of the State-of-the-Art and Possible Extensions. IEEE Transactions

on Knowledge and Data Engineering, 17 (6), pp. 734–749.

Anderson, C. (2006). The Long Tail: Why the Future of Business is Selling Less of More.

New York: Hyperion.

Ansari, A., Essegaier, S., and Kohli, R. (2000). Internet Recommendation Systems. Journal

of Marketing Research, 37 (3), pp. 363–375.

Ansari, A. and Mela, C.F. (2003). E-Customization. Journal of Marketing Research, 40 (2),

pp. 131–145.

Berry, L.L. (2002). Relationship Marketing of Services-Perspectives from 1983 and 2000.

Journal of Relationship Marketing, 1 (1), pp. 59–70.

Bhat, S., Bevans, M., and Sengupta, S. (2002). Measuring Users' Web Activity to Evaluate

and Enhance Advertising Effectiveness. Journal of Advertising, 31 (3), pp. 97–106.

Bhattacherjee, A. (2001). Understanding Information Systems Continuance: An

Expectation-Confirmation Model. MIS Quarterly, 25 (3), pp. 351.

Booth, D. and Jansen, B. (2008). A Review of Methodologies for Analyzing Websites.

Handbook of Research on Web Log Analysis: Hershey, USA: IGI, pp. 141–162.

Brynjolfsson, E., Hu, J.Y., and Smith, M.D. (2006). From Niches to Riches: The Anatomy

of the Long Tail. Sloan Management Review, 47 (4), pp. 67–71.

Brynjolfsson, E., Hu, Y., and Simester, D. (2011). Goodbye Pareto Principle, Hello Long

Tail: The Effect of Search Costs on the Concentration of Product Sales. Management

Science, 57 (8), pp. 1373–1386.

Burke, R. (2002). Hybrid Recommender Systems: Survey and Experiments. User

Modeling and User-Adapted Interaction, 12 (4), pp. 331–370.

Chau, P.Y.K., Ho, S.Y., and Yao, Y. (2011). The Effects of Malfunctioning Personalized

Services on Users' Trust and Behaviors. Proceedings of the Pacific Asia Conference on

Information Systems (PACIS).

Chau, P.Y.K., Lyytinen, K., Wei, C.-P., Yang, C.C., Lin, F.-R., Liang, T.-P., Chen, H.-Y.,

and Turban, E. (2009). Effect of Personalization on the Perceived Usefulness of Online

Customer Services. Proceedings of the 11th International Conference on Electronic

Commerce (EC), pp. 279–288.

Chen, D.-N., Hu, P.J.-H., Kuo, Y.-R., and Liang, T.-P. (2010). A Web-based Personalized

Recommendation System for Mobile Phone Selection: Design, Implementation, and

Evaluation. Expert Systems with Applications, 37 (12), pp. 8201–8210.

Clemons, E., Bin Gu, and Lang, K. (2002). Newly Vulnerable Markets in an Age of Pure

Information Products: an Analysis of Online Music and Online News. Proceedings of

the 35th Annual Hawaii International Conference on System Sciences (HICSS), pp.

2949–2958.

Clemons, E., Gao, G., and Hitt, L. (2006). When Online Reviews Meet

Hyperdifferentiation: A Study of the Craft Beer Industry. Journal of Management

Information Systems, 23 (2), pp. 149–171.

Cohen, J. (1992). A Power Primer. Psychological Bulletin, 112 (1), pp. 155–159.


94

Crook, T., Frasca, B., Kohavi, R., and Longbotham, R. (2009). Seven Pitfalls to Avoid

when Running Controlled Experiments on the Web. Proceedings of the 15th ACM

SIGKDD International Conference on Knowledge Discovery and Data Mining. pp.

1105–1114.

Davenport, T.H. (2000). Sticky Business. CIO Magazine, 13 (8), pp. 58–60.

Davenport, T.H. and Beck, J.C. (2001). The Attention Economy: Understanding the New

Currency of Business. Boston, USA: Harvard Business School Press.

Delone, W.H. and McLean, E.R. (2003). The DeLone and McLean Model of Information

Systems Success: A Ten-Year Update. Journal of Management Information Systems,

19 (4), pp. 9–30.

Dias, M.B., Locher, D., Li, M., El-Deredy, W., and Lisboa, P.J. (2008). The Value of

Personalised Recommender Systems to E-Business. Proceedings of the 2008 ACM

Conference on Recommender Systems (RecSys), pp. 291-294.

Druckman, J., et al. (2011). Cambridge Handbook of Experimental Political Science.

Cambridge, New York: Cambridge University Press.

Ehrmann, T. and Schmale, H. (2008). The Hitchhiker's Guide to the Long Tail: The

Influence of Online-Reviews and Product Recommendations on Book Sales - Evidence

from German Online Retailing. Proceedings of the 29th International Conference on

Information Systems (ICIS).

Elberse, A. (2008). Should You Invest in the Long Tail. Harvard Business Review,

86 (7/8), pp. 88–96.

Elberse, A. and Oberholzer-Gee, F. (2008). Superstars and Underdogs: An Examination of

the Long Tail Phenomenon in Video Sales. Working Paper No. 07-015. Harvard

Business School.

Fleder, D. and Hosanagar, K. (2009). Blockbuster Culture's Next Rise or Fall: The Impact

of Recommender Systems on Sales Diversity. Management Science, 55 (5), pp. 697–

712.

Freedman, L. (2007). Listen, Learn and Leverage: Merchant Views of Personalization and

Lasting Customer Relationships. Report.

Gefen, D., Karahanna, E., and Straub, D.W. (2003). Trust and TAM in Online Shopping:

An Integrated Model. MIS Quarterly, 27 (1), pp. 51–90.

Gerber, A.S. and Green, D.P. (2012). Field Experiments: Design, Analysis, and

Interpretation. New York, USA: W. W. Norton & Company.

Ghandour, A., Benwell, G., and Deans, K. (2010). Metrics for eCommerce Websites.

Working Paper. University of Otago.

Goel, S., Broder, A., Gabrilovich, E., and Pang, B. (2010). Anatomy of the Long

Tail: Ordinary People with Extraordinary Tastes. Proceedings of the 3rd

ACM

International Conference on Web Search and Data Mining, pp. 201-210.

Hauser, J.R., Urban, G.L., Liberali, G., and Braun, M. (2009). Website Morphing.

Marketing Science, 28 (2), pp. 202–223.

Herbert, J. and Thurman, N. (2007). Paid Content Strategies for News Websites.

Journalism Practice, 1 (2), pp. 208–226.

Herlocker, J.L., Konstan, J.A., Terveen, L.G., and Riedl, J.T. (2004). Evaluating

Collaborative Filtering Recommender Systems. ACM Transactions on Information

Systems, 22 (1), pp. 5–53.


95

Hijikata, Y., Kai, Y., and Nishida, S. (2012). The Relation Between User Intervention and

User Satisfaction for Information Recommendation. Proceedings of the ACM

Symposium on Applied Computing, pp. 2002–2007.

Hinz, O. and Eckert, J. (2010). The Impact of Search and Recommendation Systems on

Sales in Electronic Commerce. Business & Information Systems Engineering, 2 (2), pp.

67–77.

Hinz, O., Eckert, J., and Skiera, B. (2011). Drivers of the Long Tail Phenomenon: An

Empirical Analysis. Journal of Management Information Systems, 27 (4), pp. 43–70.

Huang, P., Lurie, N.H., and Mitra, S. (2009). Searching for Experience on the Web: An

Empirical Examination of Consumer Behavior for Search and Experience Goods.

Journal of Marketing, 73 (2), pp. 55–69.

Inbar, O., Ben-Asher, N., Porat, T., Mimran, D., Shapira, B., Shoval, P., Meyer, J., and

Tractinsky, N. (2008). All the News that's Fit to E-Ink. Proceedings of the 26th Annual

Conference on Human Factors in Computing Systems (CHI), pp. 3621–3626.

Kamba, T., Sakagami, H., and Koseki, Y. (1997). ANATAGONOMY: a Personalized

Newspaper on the World Wide Web. International Journal of Human-Computer

Studies, 46 (6), pp. 789–803.

Kirk, R.E. (1982). Experimental Design: Part One. Foundations of Research Issues: Study

Design, Data Management, Data Reduction, and Data Synthesis. New Jersey, USA:

John Wiley & Sons, Inc.

Knijnenburg, B.P., Willemsen, M.C., Gantner, Z., Soncu, H., and Newell, C. (2012).

Explaining the User Experience of Recommender Systems. User Modeling and User-

Adapted Interaction, 22 (4-5), pp. 441–504.

Kobsa, A. (2007). Privacy-enhanced Personalization. Communications of the ACM,

50 (8), pp. 24–33.

Kohavi, R., Henne, R.M., and Sommerfield, D. (2007). Practical Guide to Controlled

experiments on the Web: Listen to Your Customers not to the HiPPO. Proceedings of

the 13th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pp.

959–967.

Kohavi, R., Longbotham, R., Sommerfield, D., and Henne, R.M. (2009). Controlled

Experiments on the Web: Survey and Practical Guide. Data Mining and Knowledge

Discovery, 18 (1), pp. 140–181.

Komiak, S.Y.X. and Benbasat, I. (2006). The Effects of Personalization and Familiarity on

Trust and Adoption of Recommendation Agents. MIS Quarterly, 30 (4), pp. 941–960.

Krcmar, H. (2005). Informationsmanagement. 4th ed. Berlin, Germany: Springer.

Kumar, N. and Benbasat, I. (2006). Research Note: The Influence of Recommendations

and Consumer Reviews on Evaluations of Websites. Information Systems Research,

17 (4), pp. 425–439.

Kwon, K., Cho, J., and Park, Y. (2010). How to Best Characterize the Personalization

Construct for E-Services. Expert Systems with Applications, 37 (3), pp. 2232–2240.

Li, D., Browne, G., and Wetherbe, J. (2006). Why Do Internet Users Stick with a Specific

Web Site? A Relationship Perspective. International Journal of Electronic Commerce,

10 (4), pp. 105–141.

Li, Y.-W. (2009). Personalization as a Strategy to Build Customer Relationship: The Role

of Intimacy. Proceedings of the Pacific Asia Conference on Information Systems

(PACIS).


96

Liang, T.-P., Lai, H.-J., and Ku, Y.-C. (2007). Personalized Content Recommendation and

User Satisfaction: Theoretical Synthesis and Empirical Findings. Journal of

Management Information Systems, 23 (3), pp. 45–70.

Liang, T.-P., Li, Y.-W., and Turban, E. (2009). Personalized Services as Empathic

Response: The Role of Intimacy. Proceedings of the Pacific Asia Conference on

Information Systems (PACIS).

Liang, T.-P., Li, Y.-W., and Turban, E. (2011). Customer Intimacy in Online Personalized

Recommendation. Working Paper. National Sun Yat-Sen University.

Lin, J.C.-C. (2007). Online Stickiness: its Antecedents and Effect on Purchasing Intention.

Behaviour & Information Technology, 26 (6), pp. 507–516.

Liu, J., Dolan, P., and Pedersen, E.R. (2010). Personalized News Recommendation Based

on Click Behavior. Proceedings of the 15th International Conference on Intelligent User

Interfaces (IUI), pp. 31-40.

Maidel, V., Shoval, P., Shapira, B., and Taieb-Maimon, M. (2008). Evaluation of an

Ontology-content Based Filtering Method for a Personalized Newspaper. Proceedings

of the 2008 ACM Conference on Recommender Systems (RecSys), pp. 91-98.

McNee, S.M., Albert, I., Cosley, D., Gopalkrishnan, P., Lam, S.K., Rashid, A.M., Konstan,

J.A., and Riedl, J. (2002). On the recommending of Citations for Research Papers.

Proceedings of ACM Conference on Computer Supported Cooperative Work (CSCW),

pp. 116–125.

McNee, S.M., Riedl, J.T., and Konstan, J.A. (2006). Being Accurate is not Enough: How

Accuracy Metrics have Hurt Recommender Systems. Proceedings of the Conference on

Human Factors in Computing Systems (CHI), pp. 1097–1101.

Mobasher, B., Burke, R., Jannach, D., Adomavicius, G., Gorgoglione, M., Panniello, U.,

and Tuzhilin, A. (2011). The Effect of Context-aware Recommendations on Customer

Purchasing Behavior and Trust. Proceedings of the 5th ACM Conference on

Recommender Systems, pp. 85-92.

Mock, K.J. and Vemuri, V. (1997). Information Filtering via Hill Climbing, WordNet, and

Index Patterns. Information Processing & Management, 33 (5), pp. 633–644.

Moe, W.W. and Fader, P.S. (2004). Dynamic Conversion Behavior at E-Commerce Sites.

Management Science, 50 (3), pp. 326–335.

Montgomery, A.L., Li, S., Srinivasan, K., and Liechty, J.C. (2004). Modeling Online

Browsing and Path Analysis Using Clickstream Data. Marketing Science, 23 (4), pp.

579–595.

Morgan, R.M. and Hunt, S.D. (1994). The Commitment-Trust Theory of Relationship

Marketing. Journal of Marketing Research, 58 (3), pp. 20–38.

Motulsky, H. (2013). Intuitive Biostatistics: A Nonmathematical Guide to Statistical

Thinking. Oxford, Great Britain: Oxford University Press.

Murthi, B.P.S. and Sarkar, S. (2003). The Role of the Management Sciences in Research

on Personalization. Management Science, 49 (10), pp. 1344–1362.

Newman, W. Better or Just Different? On the Benefits of Designing Interactive Systems in

terms of Critical Parameters. Proceedings of Designing Interactive Systems, 1997.

Oehlert, G.W. (2000). A First Course in Design and Analysis of Experiments. New York,

USA: W.H. Freeman.

Pathak, B., Garfinkel, R., Gopal, R.D., Venkatesan, R., and Yin, F. (2010). Empirical

Analysis of the Impact of Recommender Systems on Sales. Journal of Management

Information Systems, 27 (2), pp. 159–188.


97

Peppers, D., Rogers, M., and Dorf, B. (1999). Is Your Company Ready for One-to-one

Marketing? Harvard Business Review, 77 (1), pp. 151–160.

Peterson, E.T. (2006). The Big Book of Key Performance Indicators. 1st ed.

Pommeranz, A., Broekens, J., Wiggers, P., Brinkman, W.-P., and Jonker, C.M. (2012).

Designing Interfaces for Explicit Preference Elicitation: a User-centered Investigation

of Preference Representation and Elicitation Process. User Modeling and User-Adapted

Interaction, 22 (4-5), pp. 357–397.

Prager, K.J. (1995). The Psychology of Intimacy. New York, USA: Guilford Press.

Pu, P., Chen, L., and Hu, R. (2012). Evaluating Recommender Systems from the User’s

Perspective: Survey of the State of the Art. User Modeling and User-Adapted

Interaction, 22 (4-5), pp. 317–355.

Rashid, A.M., Albert, I., Cosley, D., Lam, S.K., McNee, S.M., Konstan, J.A., and Riedl,

J.T. (2002). Getting to Know You: Learning New User Preferences in Recommender

Systems. Proceedings of the International Conference on Intelligent User Interfaces

(IUI). pp. 127–134.

Resnick, P., Iacovou, N., Suchak, M., Bergstrom, P., and Riedl, J.T. (1994). GroupLens: an

Oen Achitecture for Collaborative Filtering of Netnews. Proceedings of the 1994 ACM

Conference on Computer Supported Cooperative Work, pp. 175–186.

Resnick, P. and Varian, H.R. (1997). Recommender Systems. Communications of the

ACM, 40 (3), pp. 56–58.

Resnick, P., Zeckhauser, R., Swanson, J., and Lockwood, K. (2006). The value of rep-

utation on eBay: A controlled experiment. Experimental Economics, 9 (2), pp. 79–101.

Richardson, M., Dominowska, E., and Ragno, R. (2007). Predicting Clicks: Estimating the

Click-Through Rate for New Ads. Proceedings of the 16th International Conference on

World Wide Web (WWW), pp. 521–530.

Sarwar, B., Karypis, G., Konstan, J.A., and Riedl, J.T. (2000). Analysis of

Recommendation Algorithms for E-Commerce. Proceedings of the 2nd

ACM

conference on Electronic commerce (EC), pp. 158–167.

Sarwar, B., Karypis, G., Konstan, J.A., and Riedl, J.T. (2001). Item-based Collaborative

Filtering Recommendation Algorithms. Proceedings of the 10th International

Conference on World Wide Web, pp. 285–295.

Schafer, J., Konstan, J.A., and Riedl, J.T. (1999). Recommender Systems in E-Commerce.

Proceedings of the 1st ACM Conference on Electronic Commerce (EC), pp. 158–166.

Schaupp, L., Weiguo Fan, and Belanger, F. (2006). Determining Success for Different

Website Goals. Proceedings of the 39th Annual Hawaii International Conference on

System Sciences (HICSS), pp. 107b.

Senecal, S. and Nantel, J. (2004). The Influence of Online Product Recommendations on

Consumers’ Online Choices. Journal of Retailing, 80 (2), pp. 159–169.

Shadish, W.R., Campbell, D.T., and Cook, T.D. (2010). Experimental and Quasi-

experimental Designs for Generalized Causal Inference. Belmont, USA: Wadsworth,

Cengage Learning.

Shani, G. and Gunawardana, A. (2011). Evaluating Recommendation Systems. F. Ricci et

al. (eds.): Recommender Systems Handbook. Boston, USA: Springer US, pp. 257–297.

Shi, L. and Wang, K. (2008). An Laboratory Experiment for Comparing Effectiveness of

Three Types of Online Recommendations. Tsinghua Science & Technology, 13 (3), pp.

293–299.


98

Tan, T.F. (2009). Is Tom Cruise Threatened? Using Netflix Prize Data to Examine the

Long Tail of Electronic Commerce.

Treacy, M. and Wiersema, F. (1993). Customer Intimacy and Other Value Disciplines.

Harvard Business Review, 71 (1), pp. 84–93.

Tsao, W.-Y. (2013). The Fitness of Product Information: Evidence from Online

Recommendations. International Journal of Information Management, 33 (1), pp. 1–9.

Turban, E. (2010). Electronic Commerce 2010: A Managerial Perspective. Upper Saddle

River, NJ: Prentice-Hall.

Wang, W. and Benbasat, I. (2005). Trust in and Adoption of Online Recommendation

Agents. Journal of the Association for Information Systems, 6 (3), pp. 72–101.

Web Analytics Association (2008). Web Analytics Definitions. Standards Document.

Wirtz, B.W., Schilke, O., and Ullrich, S. (2010) Strategic Development of Business

Models: Implications of the Web 2.0 for Creating Value on the Internet. Long Range

Planning, 43 (2-3), pp. 272–290.

Yang, J.-M. and Li, K.F. (2009). Recommendation Based on Rational Inferences in

Collaborative Filtering. Knowledge-Based Systems, 22 (1), pp. 105–114.

Zanker, M., Bricman, M., Gordea, S., Jannach, D., and Jessenitschnig, M. (2006).

Persuasive Online-Selling in Quality and Taste Domains. E-Commerce and Web

Technologies: Lecture Notes in Computer Science, Berlin, Germany: Springer, pp. 51–

60.

Zhang, B.-T. and Seo, Y.-W. (2001). Personalized Web-Document Filtering Using

Reinforcement Learning. Applied Artificial Intelligence, 15 (7), pp. 665–685.

Zhang, X. and Li, Y. (2007). Use of Collaborative Recommendations for Web Search: an

Exploratory User Study. Journal of Information Science, 34 (2), pp. 145–161.

Zhou, R., Khemmarat, S., Gao, L., and Wang, H. (2011). Boosting Video Popularity

Through Recommendation Systems. Proceedings of the International Conference on

Management of Data (SIGMOD) - Databases and Social Networks, pp. 13-18.

Zhu, F. and Zhang, X. (2010). Impact of Online Consumer Reviews on Sales: The

Moderating Role of Product and Consumer Characteristics. Journal of Marketing,

74 (2), pp. 133–148.

99

III. CASE STUDY 2: ELECTRICITY

DEMAND RESPONSE

PART III CASE STUDY 2: ELECTRICITY DEMAND RESPONSE

100

A. An Information System Architecture for Demand Response

This chapter is a revised version of the paper

Philipp Bodenbenner, Stefan Feuerriegel, Dirk Neumann, 2013: Design Science in

Practice: Designing an Electricity Demand Response System. 8th International Conference

on Design Science Research in Information Systems and Technology (DESRIST 2013),

June 11-12, 2013, Helsinki, Finland, Editor: Jan vom Brocke et al, Proceedings Name:

LNCS 7939, pp. 293-307, Springer-Verlag, Berlin.

Abstract

Information systems play an important role in achieving sustainable solutions for the

global economy. In particular, information systems are inevitable when it comes to the

transition from the “current” to the “smart” power grid. This enables an improved

balancing of both electricity supply and demand, by shifting load – based on the projected

supply gap and electricity prices – on the demand side smartly. As this requires a specific

information system, namely a demand response system, we address the challenge of

designing such a system by utilizing the design science approach: determining general

requirements, deducing the corresponding information requirements, analyzing the

information flow, designing a suitable information system, demonstrating its capability,

and, finally, evaluating the design. The design process is reiterated fully until a viable

solution, i.e. an IS artifact, has been developed. This part describes both the design process

as such and the final IS artifact. Moreover, we summarize our lessons learnt from using

and adopting the design science approach within this practical, bottom-up case study.

Keywords

Design Science Research, Green IS, IS-Architecture, Demand Response, Case Study


101

1 Introduction

For a global economy, information systems (IS) play an important role in achieving

sustainable solutions and, in particular, Green IS will facilitate to shape the power supply,

transmission and consumption systems of the future. As a consequence, the IS research

community has recently given significant attention to Green IS.

Until today, it was only possible to control the supply side in electricity markets, but Green

IS introduces the power demand as an additional dimension subject to optimization. This

degree of freedom is not entirely new, but “affordable global communication

infrastructure and embedded systems make it now relatively easy to add a certain portion

of smart to the loads” (Palensky and Dietrich, 2011). This so-called transition from current

to smart power grids will lead to unexpected problems, unprecedented challenges and new

opportunities – such as matching both power demand and supply. One possible path to

balance power demand and supply is given by the concept of demand response. Demand

response (DR) is defined by the U.S. Department of Energy (2006) and the Federal Energy

Regulatory Commission (FERC) (2006) as: “Changes in electric usage by end-use

customers from their normal consumption patterns in response to changes in the price of

electricity over time, or to incentive payments designed to induce lower electricity use at

times of high wholesale market prices or when system reliability is jeopardized.” In short,

this implies shifting load to times when supply exceeds demand.

The purpose of this part is to design a DR system. We develop an IS artifact that enables

retailers to minimize their expenditures on electricity by demand response. This problem is

addressed by using the well-known design science research approach. Melville (2010)

states that “design research is essential to developing innovative IS-enabled solutions to

environmental problems and evaluating their effectiveness”. This encompasses successive

steps (compare Figure 23) of determining requirements including the required information

demand and flow, designing a suitable information system, demonstrating its capability,


102

and, finally, evaluating the design before starting into the next iteration of the design

process.

As a main contribution to IS research, this part describes not only the design of an

information system for demand response, but it also aims to evaluate the design science

approach in practice. We will also share our experiences from applying the design science

approach.

The remainder of this part is structured as follows. In chapter 2, related work on the IS

perspective of DR systems is reviewed. Subsequently, the DR system (chapter 3) is

designed: First, we introduce general requirements. Then, we deduce the corresponding

information requirements, analyze the information flow, list components matching the

information requirements, and evaluate the design. Finally, in chapter 4, we revisit the

design science approach according to the guidelines by Hevner et al. (2004) and, in

addition to that, we focus on what we have learnt from employing the design science

approach in this case study.

2 Demand Response in IS Research

A recent literature review (Strüker and van Dinther, 2012) shows that there are a small, but

growing number of IS-related research papers on demand response. Our own literature

Figure 23. Iterative Design Science Approach with a Closed Feedback Loop Based on Hevner

et al. (2004) and Simon (1996)..

Demonstrate

Capability(Section 3.3)

Define

Requirements(Section 3.1)

Design

System(Section 3.2)

Evaluate

System(Section 3.4)


103

research on these publications reveals that there are many studies demonstrating the

responsiveness of residential, commercial and industrial customers to incentives and

prices. However, we only came across few contributions dealing with the design of a DR

system. These publications have contributed to IS research:

Corbett (2011) claimed that a DR system would increase the information processing

requirements. As a consequence, the author argued that DR systems will incur

massive amounts of data and, thus, a DR system poses an inherent IS problem.

Several authors (e.g. Watson et al. (2010)) advanced towards an IS-architecture. Tan

et al. (2012) designed an actual decision supporting system for demand side

management, but the authors proposed a high-level structure only. Palensky and

Dietrich (2011) constructed a web-based energy information system and named its

typical components, while Law et al. (2012) focused on tethering the end-consumer.

Feuerriegel et al. (2012) and Feuerriegel et al. (2013) integrated demand response into

the energy informatics framework and listed required components, but the authors

lack a rigorous design science approach.

Since the above research papers have concentrated on partial examinations of either

requirements or design, many questions have been left unanswered. All listed publications

lack (1) a combination of design as well as requirements, and (2) an in-depth evaluation of

the proposed design. Consequently, a DR system originating from a design science process

still seems to be an open research questions. Therefore, we focus on a rigorous design

science approach: First, we derive necessary requirements for a DR system. Based on these

requirements, we develop a system design, and evaluate this system design afterwards.

However, we not only design the DR system, but also give insights into the lessons learnt

from iteratively developing the artifact.


104

3 Applying the Design Science Approach

The starting point for our research is the evaluation of financial benefits, which an

electricity retailer can leverage by executing a DR mechanism. As these benefits obviously

do not come free, the cost perspective needs to be examined as well. Core prerequisite –

and a major cost driver – for establishing DR mechanisms is the setup of a broad

information system infrastructure. Therefore, we needed to gain an overview of both the

required components and their relations. As there is no such system blueprint available

(compare section 2), we employ the design science methodology to conceptualize an

information system, which fits the set of defined requirements. Our approach is based on

the multi-stage design science research methodology for IS research, which has been

proposed by Peffers et al. (2007). According to Peffers et al. (2007) our research intent

requires the use of an objective-centered approach. An objective-centered approach is

triggered by a research need that can be addressed by developing an artifact. Consequently,

we start off with defining the objectives for the DR system and subsequently pursue the

previously described design science process (cp. Figure 23).

3.1 Defining Requirements for the Information System

The objectives for the DR system are deduced by transforming the overall problem

description into functional and non-functional requirements. We also integrate knowledge

about current solutions for DR infrastructures and for other smart energy infrastructures

into the definition of the requirements. This knowledge has been both extracted from

literature (cp. chapter 2) and expert interviews. For structuring and documentation of the

requirements, we revert to the analytical level of the design science framework defined by

Vahidov (2006). Hence, the salient features and properties of the system are described in a

technology-independent and fairly descriptive fashion. The remaining section provides an

excerpt of the overall set of requirements. First, we show the functional and non-functional

requirements corresponding to category structure in Vahidov’s framework. Second, we


105

illustrate the corresponding information demand and flow in a dynamic representation (i.e.

category behavior).

Functional and Non-Functional Requirements

Implementation of a DR mechanism requires substantial extensions of today’s power

networks. So far, power networks – on the distribution network level – have been mainly

equipped with passive components and little to no information systems. Thus, Corbett

(2011) projects a strong future demand for increased information processing capabilities of

the power networks to meet the challenges of the inevitable energy transformation. A core

challenge is the matching process of fluctuating electricity supply and demand. The

fluctuation increases – especially on the supply side – due to increased usage of renewable

energy sources, such as wind and solar. DR mechanisms form a viable solution to this

issue. For determination and execution of an expedient load shifting and reduction scheme

(i.e. the outcome of a DR mechanism), the respective information system has to satisfy a

number of functional (RF) and non-functional (RN) requirements. We present the five

most important requirements at a glance as follows:

Near-time access to consumption usage data (RF1). The determination and execution

of an efficient DR mechanism requires high quality forecasts of the electricity

consumption for the next planning horizon (i.e. mostly the next 24 hours). Such a

forecast quality can only be achieved by processing consumers’ real-time resp. near-

time consumption data (Ziekow et al., 2012). Thus, the correspondent usage data

record needs to be regularly transferred from the consumer to the electricity retailer’s

backend via a secure communication channel (Simmhan et al., 2011).

Centralized remote control of connected devices (RF2). For realization of a DR

mechanism, the connected consumers need to support remote control of their devices

(Simmhan et al., 2011, Sui et al., 2011). Based on the calculated load shifting scheme,


106

the electricity retailer transmits control signals for shifting electricity load of the

devices (cp. detailed DR protocol description in Mohagheghi et al. (2010)).

Easy integration into existing IS environment of retailer and consumer (RN1). The

information system needs to be developed with a brownfield approach as today’s

electricity retailers have a comprehensive IS landscape in place already (Sui et al.,

2011). The different information systems provide functionality for monitoring and

controlling the complex power flow networks and connected components (e.g.

distribution station) across different levels (i.e. transportation and distribution

networks). The newly designed information system, which brings additional

smartness to the distribution networks, must fit into this legacy IS environment

smoothly (Li et al., 2010).

Openness to implementation of other use cases (RN2). The DR system must be

shaped in an open and flexible manner. Additional use cases, apart from

implementation of DR mechanisms, should be enabled to use the deployed

infrastructure and the collected data as much as possible. Faruqui et al. (2010) state

automated (monthly) billing based on usage data and improved customer service as

possible use cases.

Cost effectiveness (RN3). Watson et al. (2010) explain that eco-efficiency is a major

sustainability goal. Thus, the information system design should support both a cost-

effective implementation and operation of the infrastructure.

Information Demand and Flow.

The two functional requirements, namely (RF1) and (RF2), serve as basis for deducing the

information demand and flow of the DR system. For this, a sequence diagram similar to

one of the UML method toolkit is used (Object Management Group (OMG), 2012). First,

involved actors are defined, i.e. in this case the electricity retailer, the connected

consumers, and external service providers. Subsequently, use cases for these actors are


107

developed and converted into a sequence diagram, which is illustrated in Figure 24. The

sequence diagram shows the two-step approach of implementing a DR mechanism: (i)

forecasting of electricity load for the next cycle, and (ii) determination and execution of an

appropriate load shifting scheme.

Load Forecasting. First, the electricity retailer determines the load forecast for the

next optimization interval. This calculation is based on weather forecast data (A),

electricity prices (B), historical (C), and current usage data (D). The electricity retailer

retrieves weather data from specialized external service providers. Just as the weather

data, electricity prices can be retrieved from external service providers, e.g. the EEX

(European Energy Exchange (EEX), 2012). These two datasets are already available

to and used by the electricity retailer for calculation of the load forecast today. In

contrast, the usage data, both historical and current, needs to be more granular and

dense compared to data available today (i.e. standardized load profiles). With the

introduction of advanced communication infrastructures, load profiles can be replaced

by real-time usage data, which is retrieved regularly from the consumer.

Load Shifting. If the load forecast algorithm detects a shortage in supply for the next

optimization interval, a load shifting scheme needs to be calculated. There are three

major variants of DR mechanisms (Mohagheghi et al., 2010): (i) incentive-based, (ii)

rate-based, and (iii) consumer-induced. The different mechanisms entail different

Figure 24. Dynamic View of Information Flow in Incentive-based Demand Response

Mechanisms.

Consumer

External

Service

Providers

Electricity

Retailer

Weather

Forecast

Electricity

Prices

Meter

Data

Sensor

Data

DR

Contingencies

Control

Signals

Determination & execution of load

shifting scheme

D

EElectricity

Demand

A

G

B

H

Forecasting of electricity load

for next time interval

Historical

Usage

C

Network

Model

F


108

information requirements and communication protocols. We focus on incentive-based

mechanisms in which consumer commit a-priori to electricity shifting and receive a

monetary compensation in return as soon as a load shift took place. For calculating the

optimal load shift the electricity retailer requires the projected electricity demand (E),

the consumers’ load contingencies (F) as well as the underlying network model (G).

The electricity demand for the next optimization interval is retrieved as a result of the

computed load forecast. The consumers regularly transmit their load contingencies

and, thereby, communicate their maximum amount of shiftable power and maximum

shift duration to the electricity retailer. The network model describes the detailed

topology of the distribution network including the connected components and

consumers. It is maintained within a repository of the electricity retailer. As soon as

the electricity retailer has computed the optimal load shifting scheme, the required

action is communicated towards the consumer. The consumer receives a series of

control signals (H), which determine the shifting of electricity for the next

optimization interval.

3.2 Designing the Information System

The synthetic level of Vahidov’s framework (Vahidov, 2006) serves as basis for

development the abstract, technology-independent system design. As proposed by Hevner

et al. (2004) the knowledge base of reference disciplines and employ available

frameworks, methods, and tools is used.

The development process starts with an evaluation of the current situation. Hence as a

baseline model, a typical information system environment of today’s electricity retailers on

the distribution network level is outlined (e.g. Cassel, 1993). The target model is being

built on this baseline model and a series of available reference models. The process starts

from a generic view and iteratively refines the design towards a more specific solution.

The starting point is the energy informatics framework (Watson et al., 2010), which

provides the generic base structure of an information system dealing with energy demand


109

and supply. Watson et al. (2010) divide an energy system into several building blocks: a

central information system, sensor networks, sensitized objects, flow networks, and related

external stakeholders (cp. Figure 25).

The proposed architectures for DR systems of Mohagheghi et al. (2010), Simmhan et al.

(2011) and Sui et al. (2011) are the base for refining and reconfiguring the Energy

Informatics Framework. Furthermore, we consult reference architectures for smart

metering infrastructures (e.g. Joint Working Group on Standards for Smart Grids (2011)).

For deriving the design characteristics of the information system from the previously

described requirements, the method of quality function deployment (QFD) is deployed. We

use an extended version of this method, which is adapted to the domain of IS (Haag et al.,

1996). This allows for matching requirements and solution proposals for information

systems in a structured fashion. Moreover, the various design alternatives resp. reference

architectures can be evaluated and compared with each other.

In the following paragraphs, we present an exemplary static view of the target model’s

Information System design, which corresponds to the category structure of Vahidov’s

framework. The overall information system design is illustrated in Figure 25.

The distribution management system (DMS) constitutes the core information system in our

setup. This type of information system is already in operation at electricity retailers today.

Figure 25. System View on a Demand Response System with Advanced Metering

Infrastructures Based on Mohagheghi et al. (2010), Watson et al. (2010), Simmhan et al.

(2011), Sui et al. (2011). Components of the Baseline Model are Marked in Gray.

Sensitized ObjectsSensor NetworksInformation SystemExternal Stakeholders

Flow Networks

Energy

Mgmt.

System

Generation

Info. &

Trading

System

Weather

Forecast

Provider

B

Load Control

Device

Distributed

Energy

Resource

A

A Weather Forecasts B Electricity Prices C Historical Usage Data D Current Usage DataLegend:

Distribution Mgmt.

System

Demand Response

Engine

Load Forecasting

Engine

E

Historical

usage

Network

Model

C G

Energy

Markets

E Electricity Demand F DR Contingencies G Network model H DR Control Signals

Smart

Meter

Meter Data

Mgmt.

System

Load Mgmt.

& Control

System

Consuming

Mgmt.

System

Con-

centrator

Network

Mgmt.

System

H

D F

Con-

centrator

upstream

downstream


110

It features functionality for monitoring and controlling an entire distribution network

(Cassel, 1993). Improving reliability and quality of service (e.g. reduction of outages, and

maintaining frequency) are the key tasks of the DMS. The DMS communicates with

external service providers, such as weather data providers and energy markets, via defined

interfaces. Moreover, the DMS features interfaces for communication and data exchange

with other neighboring parts of the power network (e.g. transportation network). For

implementing DR mechanisms, the DMS needs to be extended with a DR Engine, which is

responsible for computing the optimal load shifting scheme. The DMS is enhanced with a

comprehensive communication infrastructure that needs to be established between the

electricity retailer and the consumers. This setup of a real-time communication channel is

essential for the implementation of a DR mechanism.

Component Required

Adjustments

Category Owner

Distribution Management System (DMS) Update, new functions General Retailer

Meter Data Management System

(MDMS)

New component AMI Retailer

Concentrator New component AMI Retailer

Smart Meter New component AMI Retailer

Network Management System (NMS) New component AMI Retailer

Load Management and Control System New component DR Retailer

Consuming Management System New component DR Retailer

Load Control Device New component DR Consumer

External Interface: Weather Forecast Data None General Retailer

External Interface: Electricity Prices None General Retailer

The communication path consists of two parts, namely an upstream and a downstream

channel. The upstream channel transfers the current usage data and load contingencies of

the consumer. This part contains the components that are being referred to as advanced

metering infrastructures. The smart meter reads out the current usage data of the consumer

and sends it towards the central backend. The meter data management system (MDMS)

receives and aggregates the collected usage data and preprocesses it for the DMS. In

Table 26. Overview on Results of Gap Analysis Between Baseline and Target Model.


111

addition, the downstream channel is used by the electricity retailer to communicate the

control signals for load shifting to the consumer.

Based on the defined baseline and target model, we have conducted a gap analysis to

determine which adjustments and extension are required compared to today’s information

systems. The results are presented in Table 26.

3.3 Demonstrating Capability of the Information System

Based on the abstract information system design, we have executed an exemplary

instantiation of the information system to gain first impressions on its capabilities and

especially the cost structure. For the instantiation, we have defined the following setup that

we used for conducting a computational simulation. As the implementation of demand

response will be driven by electricity retailers, we assume an average German electricity

retailer serving 290,000 residents (i.e. 145,000 households) and 75,000 small industrial

customers. These numbers correspond to around 220,000 smart meters that need to be

connected to the distribution network. The remaining IS infrastructure is scaled

accordingly.

3.4 Evaluating the Information System

In this section, we present a qualitative and quantitative evaluation of our proposed

information system design for implementing DR mechanisms. Our evaluation effort is

based on methods described in Venable et al. (2012) and Hevner et al. (2004). We carry

out both ex-ante (i.e. evaluation of the design) as well as ex-post (i.e. evaluation of an

instantiation) assessments (Venable et al., 2012).

Functional requirements [RF1 and RF2]. The evaluation of compliance and coverage of

the proposed information system design with the defined functional requirements is rather

difficult on the level of abstraction that we have described in this part. Nonetheless, we

employed the QFD method to ensure a preferably high coverage of the defined

requirements by the technical product specifications (Haag et al., 1996). We applied the


112

QFD method iteratively for increasing the level of detail of the resulting IS artifact. For

example, for determination of the optimal transmission medium, we started with

comparing the various network types (cp. Figure 27). In a second iteration, we employed

the QFD method to drill down into the communication network design. Therefore, we

opposed the chosen transmission media, namely GSM and PLC, with a number of

available reference models (e.g. Tan et al., 2012) in a relationship matrix to decide on the

IS design. We conducted this method for the remaining requirements analogously.

Summing up, we can assume that our information system design fulfills the defined

functional requirements.

Easy integration into existing IT environment of retailer and consumer [RN1]. We try

to minimize the integration effort by employing a brownfield approach. Thereby, the

current IS environment is being defined as baseline model and, thus, it is being used as a

basis for development of the target model.

Openness to implementation of other use cases [RN2]. Our information system design is

based on the reference model of advanced metering infrastructures (cp. section 3.2). In

doing so, we ensure that additional use cases, which are built upon an advanced metering

infrastructure, can also be realized by using the proposed information system design. In

addition, the collected usage data is being stored in a repository within the central

information system, namely the DMS, which allows an easy, yet secure, access by other

applications.

Figure 27. Extract of the Relationship Matrix, which is the Result of the First Iteration of the

Quality Function Deployment Method. Here, the Relationship Matrix Illustrates the Fit of

Communication Technologies to a Functional Requirement in Form of a Rating (e.g. “++”).

…

[RF1]

Data transfer consumer

to retailer

…

Relationship

A Requirements B Technical product specifications C Requirement priorities D Technical product specification prioritiesLegend:

A

B

D

C

GSMFixed

LinePLC WiMax

++ o ++ -

Communication network…

…

++ ++ ++ ++ ++


113

Cost effectiveness [RN3]. We show that the designed DR system works cost effectively

(cf. Feuerriegel et al. (2012) for details). Therefore, we calculate both costs and savings

from using a DR mechanism using real-world data (cp. detailed description of simulation

environment in section 3.3). The individual cost components are given in Figure 28. By

varying the frequency of collecting usage data from the meters, we have designed four

scenarios to evaluate our model across different information granularities. In short, we gain

a positive annual surplus in the first scenario (60 minute intervals), but this negates with

higher information granularities where the operating costs exceed the savings.

4 Revisiting the Design Science Approach

We look at the design science paradigm from two angles: First, we review the design

science guidelines and check to what extent we have obeyed them. Second, we reflect how

the design science approach supported us during the design process, and we discuss what

we have learnt from our experiences.

4.1 Reviewing the Design Science Guidelines

The problem we tackled in our research was to develop a DR system using a design

science approach. Hevner et al. (2004) states that “in design science paradigm, knowledge

and understanding of a problem domain and its solution are achieved in the building and

application of the designed artifact”. The successful application of design science research

can be evaluated by examining the seven guidelines for good design science research by

Figure 28. Comparison of Savings and Costs from Using the Demand Response System in

2011 Across Different Information Granularities Based on the Methodology from

Feuerriegel et al. (2012).


114

Hevner et al. (2004). In addition to that, we follow the suggestions by Brocke and Seidel

(2012) to consider the relevance of our research in matters of environmental sustainability

for each guideline:

Problem Relevance. The amount of interest and research (cp. Section 2) regarding

the transition to the smart grid attests to the transition’s relevance. Many countries

face this transition as a major challenge (so-called “Energiewende”) to reduce

electricity consumption and production. Without doubt, the energy transition is related

to environmental sustainability.

Research Rigor. Design science research relies on the application of rigorous

methods for both the construction and for the evaluation of the design artifact. Our

architectural design is based on the energy informatics framework by Watson et al.

(2010); the cost side is constituted in a mathematical model incorporating the

components of the system; the realized savings are modeled through a linear

optimization problem.

Design as a Search Process. Implementation and iteration are central to design

science research. We studied prototypes that, based on previous work from literature,

instantiate posed or newly learnt design prescriptions. Their use and impacts were

observed, problems identified, solutions posed and implemented, and the cycle was

then repeated until the status quo is acquired.

Design as an Artifact. Design science research must produce a viable artifact in the

form of e.g. a model; therefore, we presented an architectural design for a DR system

in section 3.2. According to Brocke and Seidel (2012), this design artifact must

contribute to the implementation of sustainable business processes. In fact, when

implemented by an electricity retailer, this artifact can significantly contribute to

environmental sustainability (Feuerriegel et al., 2012).


115

Design Evaluation. The evaluation (cp. sections 3.3 and 3.4) consists of a thorough

review of the previously stated requirements. Although our model suggests that an

average electricity retailer faces huge initial setup costs, we provide evidence that

savings from load shifting exceed the running costs of the information system

significantly – by more than € 2 million per year. These savings ensure that the design

artifact contributes to environmental sustainability as proposed by Brocke and Seidel

(2012).

Research Contributions. The major contributions of this research are the proposed

architecture for a DR system as the design artifact and the evaluation results in the

form of real-world scenarios to gauge both costs and savings. These contributions

advance our understanding of how demand response can help in the transition to the

smart grid. As stated in Brocke and Seidel (2012), this guarantees that our

contributions are clear and verifiable.

Communication of Research. In addition to the previously described design science

process, Peffers et al. (2007) outline a supplemental, final process step, namely

communication. We address this demand by publishing this part and additional

versions (e.g. Feuerriegel et al. (2012)) that contain useful information for managerial

audience (i.e. how demand response can be integrated as a business model). However,

we must ensure that the results will be more accessible to an audience from energy

research in the future. On top of that, we showed how design science can be applied in

practice in IS research and we presented our experience from using a design science

approach.

4.2 Reflections on Design Science

Our experiences from designing the above DR system have brought forth a number of

practical concerns. The most important findings are listed below:


116

It is difficult to plan design science research in advance. Our primary focus was on the

financial benefits from using DR mechanisms. However, we found out – during a later

iteration – that both acquisition and operation of the corresponding infrastructure is

tied to significant financial costs. Here, design science research with its iterative

solving process allowed us to adjust and refine the focus of our research in an

unconventional manner.

While evaluating the quality of a design, we restricted our analysis not only to the

requirements itself, but we also utilized the general guidelines of Hevner et al. (2004)

for reviewing the design process. As these guidelines are general in nature, we

employed additional criteria (March and Smith, 1995, Brocke and Seidel, 2012,

Sonnenberg and Brocke, 2012) in our internal evaluation process to test each model

against the requirements. In addition to that, we noticed the value of real-world

simulations for evaluating design artifacts.

However, we missed a standardized procedure to approach a design science problem

for developing an information system. Though guidelines and evaluation criteria assist

to identify the necessary design steps, the steps themselves are unclear or ambiguous.

This view is shared by various researchers (e.g. Offermann et al., 2009) who point out

the lack of a toolbox applicable to IS design. For us, this became particularly evident

during translation of the requirements into a concrete system design draft. The UML

toolkit does offer a help, although not completely as it does not provide full

constructive support as required. Current IS research tries to close this gap (e.g.

Dromey (2003)). However, the IS community is still far away from the set of

methods, which is available in e.g. the engineering domain (cp. Pahl and Beitz, 1996).

All in all, we think that we could have benefited from support granted by tailored

tools for representing IS artifacts.


117

Reflecting on our experience in designing the DR system, we appreciate design science

research as it assisted us in the search process to discover an effective solution to our

problem.

5 Conclusion and Outlook

The challenge was to design a DR system. We developed an IS artifact that enables

electricity retailers to shift power demand and addressed this problem by using the well-

known design science research approach. This approach encompasses successive steps of

determining requirements including the required information demand and flow, designing

a suitable information system, demonstrating its capability and, finally, evaluating the

design before advancing the next iteration of the design process.

Our IS artifact illustrates the electricity retailer’s required infrastructure. For making a DR

mechanism work, substantial implementation effort is required on the consumers’ premises

as well. The consumers need to install both a powerful home area network (HAN) and

devices that allow for remote control. In future work, we will focus on this consumer

perspective to create a view of the overall system requirements and costs for implementing

a DR mechanism.


118

6 References

Brocke, J. and Seidel, S. (2012). Environmental Sustainability in Design Science Research:

Direct and Indirect Effects of Design Artifacts. Proceedings of the 7th International

Conference on Design Science Research in Information Systems and Technology

(DESRIST), pp. 294–308.

Cassel, W. (1993). Distribution management systems: functions and payback. IEEE

Transactions on Power Systems, 8 (3), pp. 796–801.

Corbett, J. (2011). Demand Management in the Smart Grid: An Information Processing

Perspective. Proceedings of the 17th Americas Conference on Information Systems

(AMCIS).

Dromey, R. (2003). From Requirements to Design: Formalizing the Key Steps.

Proceedings of the 1st International Conference on Software Engineering and Formal

Methods, pp. 2–11.

European Energy Exchange, EEX (2012). Market Data. http://www.eex.com/, accessed

2 May 2012.

Faruqui, A., Harris, D., and Hledik, R. (2010). Unlocking the €53 Billion Savings from

Smart Meters in the EU: How Increasing the Adoption of Dynamic Tariffs Could Make

or Break the EU’s Smart Grid Investment. Energy Policy, 38 (10), pp. 6222–6231.

Federal Energy Regulatory Commission, FERC (2006). Assessment of Demand Response

and Advanced Metering.

Feuerriegel, S., Bodenbenner, P., and Neumann, D. (2013). Is More Information Better

than Less? Understanding the Impact of Demand Response Mechanisms in Energy

Markets. Proceedings of the 21st European Conference on Information Systems (ECIS).

Feuerriegel, S., Strüker, J., and Neumann, D. (2012). Reducing Price Uncertainty Through

Demand Side Management. Proceedings of the 33rd

International Conference on


Haag, S., Raja, M.K., and Schkade, L.L. (1996). Quality Function Deployment Usage in

Software Development. Communications of the ACM, 39 (1), pp. 41–49.

Hevner, A.R., March, S.T., Park, J., and Ram, S. (2004). Design Science in Information

Systems Research. MIS Quarterly, 28 (1), pp. 75–105.

Joint Working Group on Standards for Smart Grids (2011). Final Report on Standards for

Smart Grids. European Telecommunications Standards Institute (ETSI).

Law, Y.W., Alpcan, T., Lee, V.C., Lo, A., Marusic, S., and Palaniswami, M. (2012).

Demand Response Architectures and Load Management Algorithms for Energy-

Efficient Power Grids: A Survey. Proceedings of the 7th International Conference on

Knowledge, Information and Creativity Support Systems (KICSS), pp. 134–141.

Li, Z., Wang, Z., Tournier, J.-C., Peterson, W., Li, W., and Wang, Y. (2010). A Unified

Solution for Advanced Metering Infrastructure Integration with a Distribution

Management System. Proceedings of the 1st IEEE International Conference on Smart

Grid Communications: IEEE, pp. 501–506.

March, S.T. and Smith, G.F. (1995). Design and natural science research on information

technology. Decision Support Systems, 15 (4), pp. 251–266.

Melville, N.P. (2010). Information systems innovation for environmental sustainability.

MIS Quarterly, 34 (1), pp. 1‐21.

Mohagheghi, S., Stoupis, J., Wang, Z., Li, Z., and Kazemzadeh, H. (2010). Demand

Response Architecture: Integration into the Distribution Management System.

http://www.eex.com/


119

Proceedings of the 1st IEEE International Conference on SmartGrid Communications,

pp. 501–506.

Object Management Group, OMG (2012). OMG Unified Modeling Language (OMG

UML). Standard Document.

Offermann, P., Levina, O., Schönherr, M., and Bub, U. (2009). Outline of a Design

Science Research Process. Proceedings of the 4th International Conference on Design

Science Research in Information Systems and Technology.

Pahl, G. and Beitz, W. (1996). Engineering design: A Systematic Approach. London, Great

Britain: Springer.

Palensky, P. and Dietrich, D. (2011). Demand Side Management: Demand Response,

Intelligent Energy Systems, and Smart Loads. IEEE Transactions on Industrial

Informatics, 7 (3), pp. 381–388.

Peffers, K., Tuunanen, T., Rothenberger, M.A., and Chatterjee, S. (2007). A Design

Science Research Methodology for Information Systems Research. Journal of

Management Information Systems, 24 (3), pp. 45–77.

Simmhan, Y., et al. (2011). An Informatics Approach to Demand Response Optimization

in Smart Grids. Technical Report. University of Southern California.

Simon, H.A. (1996). The Sciences of the Artificial. 3rd

ed. Cambridge, MA: MIT Press.

Sonnenberg, C. and Brocke, J. (2012). Evaluations in the Science of the Artificial –

Reconsidering the Build-Evaluate Pattern in Design Science Research. Proceedings of

the 7th International Conference on Design Science Research in Information Systems

and Technology (DESRIST), pp. 381–397.

Strüker, J. and van Dinther, C. (2012). Demand Response in Smart Grids: Research

Opportunities for the IS Discipline. Proceedings of the 18th Americas Conference on

Information Systems (AMCIS).

Sui, H., Sun, Y., and Lee, W.-J. (2011). A Demand Side Management Model Based on

Advanced Metering Infrastructure. Proceedings of the 4th International Conference on

Electric Utility Deregulation and Restructuring and Power Technologies (DRPT), pp.

1586–1589.

Tan, X., Shan, B., Hu, Z., and Wu, S. (2012). Study on Demand Side Management

Decision Supporting System. Proceedings of the 3rd


Software Engineering and Service Science, pp. 111–114.

U.S. Department of Energy (2006). Benefits of Demand Response in Electricity Markets

and Recommendations for Achieving Them. A Report to the Unit States Congress

Pursuant to Section 1252 of the Energy Policy Act of 2005.

Vahidov, R. (2006). Design Researcher’s IS Artifact: a Representational Framework.

Proceedings of the 1st International Conference on Design Science Research in

Information Systems and Technology (DESRIST).

Venable, J., Pries-Heje, J., and Baskerville, R. (2012). A Comprehensive Framework for

Evaluation in Design Science Research. Proceedings of the 7th International Conference

on Design Science Research in Information Systems and Technology (DESRIST), pp.

423–438.

Watson, R.T., Boudreau, M.-C., and Chen, A.J. (2010). Information Systems and

Environmentally Sustainable Development: Energy Informatics and New Directions for

the IS-Community. MIS Quarterly, 34 (1), 23–36.

Ziekow, H., Strüker, J., and Goebel, C. (2012). The Potential of Smart Home Sensors in

Forecasting Household Electricity Demand. Working Paper. University of Freiburg.


120

B. Optimal Information Granularity in Demand Response

This chapter is a fundamentally revised and extended version of the paper

Stefan Feuerriegel, Philipp Bodenbenner, Dirk Neumann, 2013: Is More Information

Better Than Less? Understanding the Impact of Demand Response Mechanisms in Energy

Markets. 21st European Conference on Information Systems (ECIS 2013), Utrecht,

Netherlands, June 5-8, 2013, Paper 192, Completed Research Paper.

Paper nominated for Claudio Ciborra award (most innovative paper) at ECIS 2013.

Abstract

The large-scale integration of intermittent resources of power generation leads to

unprecedented fluctuations on the supply side. An electricity retailer can tackle these

challenges by pursuing strategies of flexible load shifts – so-called demand response

mechanisms. The newly introduced academic domain of energy informatics claims the

integration of flexible loads into electricity markets and the associated trade-off between

ICT deployment and economic benefits as key research challenges. This work tackles these

areas by conducting a techno-economic study. It first elaborates on the comprehensive ICT

architecture of a demand response system. Subsequently, the system design serves as basis

to derive a cost-value-model, which incorporates all relevant cost components and opposes

them to the expected savings of an electricity retailer. The model is parameterized with

real-world data to determine the optimal read-out frequency of smart meters. Opposed to

common belief, an increased information granularity does not unrestrictedly lead to higher

revenues. Beyond a certain threshold, costs surmount the additional benefits. For our set of

parameters, a positive information value of smart meter readouts is achieved within the

interval of 21 to 57 minutes, considering variable costs only. In order to achieve a

profitable setup also with the full cost scope, an alternative scenario with a smart meter

roll-out restricted to the largest consumers is presented.

Keywords

Energy informatics, Green IT/IS, electricity demand response, optimal information

granularity, business value of IS, electronic markets, business models for information

intermediaries


121

1 Introduction

Environmental sustainability has lately been perceived as one of the priority societal goals

(Melville, 2010). This evolving social awareness of green thinking has led to a

fundamental transformation that is currently taking place in the energy industry. In

particular, the introduction of power generation from renewable sources gives rise to

revolutionary changes in the energy industry. This trend is even reinforced by the

anchoring of environmental sustainability goals into governmental policies. For example,

the European Union has issued the target of renewables to make up 20% of the total

electricity generation by the year 2020. Due to its variability and limited predictability, the

increasing presence of electricity from renewables inevitably leads to unprecedented

fluctuations in electricity supply. Hence, power grid operators are facing major challenges

of stabilizing the grid and preventing imbalances between demand and supply (Lopes et

al., 2007).

As the supply side becomes more volatile as more renewables are fed into the systems,

modern systems will attempt to affect the demand side in a way that the demand has a

tendency to accommodate the shape of the fluctuating supply. Instruments that allow for

managing the demand side by shifting electricity load are basically contracts that reward

electricity usage when plenty is available and vice versa. These so called demand response

(DR) instruments allow electricity providers, typically retailers, to flexibly shift load from

peak times associated with high costs to off-peak time slots, where costs are much lower.

Hence, the retailers are enabled to pursuit a load-shaping strategy instead of load-

following (Ipakchi and Albuyeh, 2009). Altogether, the retailer can realize enormous

savings from optimizing its electricity procurement by implementing an effective demand

response management. Nonetheless, today´s demand side management is in its infancy;

installations mostly use rather simple mechanisms that exert direct load control for

singular use cases. For examples the remote control of heating, ventilation and air


122

conditioning systems allows the grid operator to remotely switch on and off a number of

electricity consuming devices according to the actual supply situation.

However, using demand response in more comprehensive application scenarios, such as

enabling an electricity retailer to perform load shifting for thousands of devices, requires

the setup of large information networks in conjunction with sophisticated communication

infrastructure. Having the intricate information flows and huge amount of data in mind,

demand response unveils to be inherently daunting for information systems (IS) research

(Dedrick, 2010; Katz, 2011). Even if “information and communication technology (ICT) is

not a panacea to address all major challenges in a smart grid […], it allows countries to

manage growing amounts of electricity produced from renewable energies and supports

structural shifts in supply and demand” (Reynolds and Mickoleit, 2011). Watson et al.

(2010) emphasize the necessity for developing integrated electricity systems and describe

their goal as the management of “supply and demand to reduce total demand and maintain

demand below established thresholds.” This pushes the boundaries of IS research with its

challenging requirements on information processing (Watson et al., 2010; Corbett, 2011).

Following this avenue of IS research attempting to contribute to an effective and efficient

DR management, we consider a DR system based on an advanced metering infrastructure

(AMI). This type of a DR system is endorsed by McDonald (2008) as it assumes „a

critical role in the discussion of finding least-cost alternatives to new build or expensive

power purchases“. An AMI establishes a two-way communication and thus higher control

and accuracy compared with other infrastructures that are used for direct load control

programs. Hitherto, the deployment of AMIs has been limited to regional islands, due to

the high upfront investment costs and the uncertain profitability of such a system.

This work contributes to the important topic of demand response by adopting an IS lens,

which aims at evaluating the potential of a DR information system for electricity retailers.


123

The contributions are threefold:

System Design. We formulate the necessary information flows and the IS architecture

for setting up a DR system. The IS architecture utilizes a structured approach for

defining energy-related information systems, which is provided by the energy

informatics framework (Watson et al., 2010).

Cost-Value-Model. Assuming the defined IS architecture we derive a comprehensive

cost-value-model for a DR system that accounts for both the saving potentials for the

retailer as well as the cost structure. With both of the elements of the model at hand,

we determine the profitability of a DR system using publicly available real-world

data.

Information Value. Information on the customers’ usage patterns plays a key role in

the smart grid, as more fine-grained information allows a more precise demand

management. In our case we operationalize information as the granularity at which the

data on user behavior are collected by the smart meters. We define the value of

information in demand response as key measure and quantify the monetary value of

alternative granularities. Interestingly, the results of our analysis contradict common

beliefs in a way that more information does not produce better results in terms of

profitability of the electricity retailer.

The remainder of part III.B is structured as follows. In chapter 2 we review related work

on both the IS perspective of DR systems as well as on their financial dimension. Based on

this review of related literature, we derive our research questions in chapter 3 and provide

evidence that these questions are relevant and important to the IS community. In chapter 4

we describe the setup of an information system for demand response. Subsequently, we

derive a cost-value-model describing the DR system from both a cost and revenue

perspective in chapter 5. In chapter 6, we evaluate the cost-value-model in a configuration

with real-world data. Finally, we discuss the results and derive managerial implications in

chapter 7.


124

2 Literature Review

The study of DR systems is quite challenging, as it is inherently an interdisciplinary

research area. Any effort that is directed towards an economic assessment of DR systems

requires a deep understanding on the technical characteristics of a demand side

management system (Goebel et al., 2014). At the heart of any DR system resides an

information system that automatically assumes the task of coordinating energy demand of

the households with the available energy supply.

Thus, the first part of our literature review targets publications contributing to the IS design

of a DR system. Secondly, we explore literature related to costs and benefits of DR

systems.

2.1 Demand Response and Smart Grid in IS Research

A new research area, energy informatics, is currently emerging to address the transition

towards sustainable economies (Malhotra et al., 2013; Goebel et al., 2014). Energy

informatics covers two major research themes: smart energy-saving systems and smart

grids. We restrain our focus to the latter as this covers the topic of demand response.

Adding intelligence to the electrical power grid to create so-called smart grids receives

highest attention from both academia and industry. ICT is regarded as key enabler

especially for integrating new types of flexible load shifting to the electricity grid

(Callaway and Hiskens, 2011; Appelrath and Kagermann, 2012). In particular, ICT is

expected to unlock the full potential of flexible loads and enable effective demand side

management of large numbers of such loads across multiple time scales (Callaway and

Hiskens, 2011). Correspondingly, the bulk of research in this realm deals with the role of

ICT in demand response (Strüker and van Dinther, 2012).

When restricting the focus to demand response as an instrument to achieve smart grids,

interestingly the range of literature narrows considerably. A recently published literature

review (Strüker and van Dinther, 2012) shows that there is only a small - albeit growing -


125

number of publications available. We come across the following few contributions dealing

with the design of a DR system. Most notably Watson et al. (2010) develop the energy

informatics framework, which appears to be a generic structure that allows for clustering

and describing the components of any energy-related system. Corbett (2011) hypothesizes

that a DR system will lead to an increase of the information processing requirements. As a

consequence, the author argues that DR systems will incur massive amounts of data, which

is an inherent IS problem. Various authors work on the architecture and components of a

DR system: Palensky and Dietrich (2011) construct a web-based energy information

system and name its typical components. Tan et al. (2012) propose a high-level design of a

decision supporting system for demand side management. Sui et al. (2011) provide a high-

level overview on how to use an AMI for setting up a DR system. Mohagheghi et al.

(2010) illustrate the integration of DR architectures into the network control center on

distribution network level; in addition, communication flows are illustrated. Finally, Law

et al. (2012) focus on tethering the end-consumer.

Besides these articles focusing on specifics of DR systems, there is a number of

publications in the IS realm dealing with different aspects of smart grids in general. For

our purpose, especially the topics of smart metering and AMIs are interesting (e.g. Galli et

al., 2011; Gungor et al., 2011).

All of these publications provide a starting point for the design of a DR system by

describing its required components. However, they all lack a design of the underlying

strategy that determines and controls the necessary load shifts. Moreover, they do not

consider the economic context in which the DR system is operated.

2.2 Economic Appraisal of Demand Response

So far, the economic potential of demand response in deregulated markets has only been

examined insufficiently (Aghaei and Alizadeh, 2013). The effects are usually considered

on two levels, the household and the aggregate level. For the household level, a number of


126

studies determine by means of fictitious scenarios the savings potentials for consumers

when using demand response (Gottwalt et al., 2011; Gudi et al., 2012; Lujano-Rojas et al.,

2012; Prüggler, 2013; Vasirani and Ossowski, 2013). Estimations on the savings potentials

reach as high as 20%. On the aggregate level, various publications (e.g. Ridder et al.

(2009)) claim a decrease in retailer profits attributed to demand response. Opposing to this

view, other research suggests demand response not diminishing the amount of consumed

energy, but merely shifting the load (Strbac, 2008; Shaw et al., 2009). To quantify the

retailer’s profits, Feuerriegel et al. (2012), Feuerriegel and Neumann (2014) and Aalami et

al. (2010) perform studies with real-world data. However, these authors neglect both

investment and run costs for using the available load shifting potential. Recent references

provide an overview of the economic costs and benefits of demand response through an

AMI on country-level (NERA Economic Consulting, 2008; Dena, 2010; Faruqui et al.,

2010; PricewaterhouseCoopers, 2010; van Horn, 2012). Paulus and Borggrefe (2011)

perform a cost-benefit-study for energy-intensive industries in Germany that sell their load

shifting potential at an exchange for spinning reserve, but do not consider efficiency gains

from more convenient energy purchases.

All in all, the listed publications lack two aspects: Firstly, an in-depth analysis of the total

cost case of a single electricity retailer using DR mechanisms and secondly, a statement on

the optimal information granularity in DR systems.

3 Research Questions

As demonstrated by literature review, many questions have been left unanswered. To

address these research gaps, we combine both perspectives of a DR system, namely

technology and financials, and set up a techno-economic study. The study is tailored to

answer three research questions.

Goebel et al. (2014) formulate “how to integrate distributed energy resources, such as

flexible loads, into existing electricity markets?” as one of the predominant questions of

research in energy informatics. Specifically, the “trade-off between ICT deployment and


127

the achieved benefits in economic terms” should be examined (Goebel et al., 2014).

Strüker and van Dinther (2012) follow this direction by asking “how large is the economic

value of demand response”. Likewise, Strbac (2008) claims that there is a “lack of

understanding of the benefits of Demand Side Management solutions” and, thus, “there

needs to be a comprehensive analysis of the costs and benefits of installing such a

sophisticated infrastructure”. As part of our evaluation, we will integrate both the cost as

well as the revenue perspective in a combined cost-value-model. This leads to our first

research question.

Existence of a Competitive Advantage for the Retailer (RQ1). Does setting up a DR

system based on an AMI create a competitive advantage for a retailer? (A) Do expected

saving potentials related to optimized electricity procurement exceed the IT related run

costs of a DR system? (B) Do the initial installation and setup costs of a DR system

amortize within its lifetime?

An essential, modifiable variable of a DR system is the interval at which usage data

records are read out from the smart meter. The information granularity directly determines

costs and revenues of the DR system. We will evaluate our models’ behavior across

various information granularities. With this, we especially address the open research

questions that have been raised by Watson et al. (2010) and Jagstaidt et al. (2011)

regarding the optimum level of information granularity in a sensor network when

optimizing a given distribution network. In general, there is a presumption that more

information is better than less, since the retailer is able to better plan its energy

procurement when having a more detailed usage profile at hand (e.g. Ziekow et al., 2012).

We transform this issue into the following research question.

Optimal Information Granularity (RQ2). Does the retailer’s profit rise with an increasing

information granularity, i.e. larger number of read-out events of electricity consumption at

smart meters? Does the revenue exceed the costs starting from a certain read-out interval,

i.e. is there a break-even point?


128

The set-up of a full-blown DR system requires substantial capital expenditures

(PricewaterhouseCoopers, 2010 and Capgemini, 2008). Thus, it might be reasonable for an

electricity retailer to restrict the roll-out of smart meters to certain groups of customers. We

define two scenarios, namely a basic and a restricted scenario. The basic scenario stands

for equipping all consumers with smart meters, whereas in the restricted scenario the roll-

out of smart meters is limited to the most valuable consumers. A recent study, conducted in

Germany, recommends to start rolling out smart meters to customers whose monthly

consumption exceeds 6,000 kWh (Ernst & Young, 2013). Energy consumers with a lower

consumption should only be equipped with smart meters at a later stage, since costs outrun

benefits. This matter forms our third research question.

Optimal Strategy for Smart Meter Roll-Out (RQ3). Does the retailer achieve the highest

profit when installing smart meters at all customers, or when limiting it to a certain group

of customers?

The research framework in Figure 29 illustrates the interplay of the three research

questions.

Figure 29. Research Framework with Research Questions RQ1 – RQ3.

4 System Model for Demand Response

To answer the research questions, we begin with viewing the DR system from a technical

perspective. Based on the information demand and required flow of information, we

Revenues & Costs(in monetary unit)

Readout Frequency(in minutes)

60 153045

revenues

costsRQ1

RQ2

Basic scenario

revenues

costs

Restricted scenario

RQ3

break even

profits


129

determine an appropriate architecture of information systems and communication

networks. In the next section, we derive a cost-value-model from the architecture of the

DR system.

In many studies related to demand response (see e.g. EU-DEEP, 2009, Smart Energy

Demand Coalition (SEDC), 2011, and EU funded project ADDRESS), it is frequently

assumed that demand response will be driven by energy retailers. Consequently, we focus

on a setup where DR activities are being integrated on the distribution network level10

.

4.1 Information Exchanges in the Demand Response System

Even though f a number of different information packages have to be exchanged between

the retailer and the consumer to use DR programs, there is currently, no standard protocol

and message specification established (Arnold, 2011). The ongoing debate of researchers

and professionals deems Data Language Messaging Specification/Companion

Specification for Energy Metering (DLMS/COSEM), Open Automated Demand Response

(OpenADR), and Smart Energy Profile (SEP) to be promising candidates for becoming a

standard (McParland, 2012).

Envisioning the information flow in a DR system and the arising data traffic is essential for

deducting the related communication costs later on. In the following, we develop a generic

exchange of information for executing a DR program that is extracted from the

aforementioned protocols.

The success of DR programs relies upon the commitment and participation of the

consumers. Basically, this can be achieved by one of the following levers (Albadi and El-

Saadany, 2008; Mohagheghi et al., 2010).

An incentive-based DR program resembles a-priori commitments by the consumers to

reduce energy. Consumers regularly communicate their maximal shift duration and

10 In this way, we implicitly incorporate requirements imposed by the power grid structure (e.g. congestion and node voltage limitations) into our modeling efforts (Mohagheghi et al. (2010)).


130

shiftable power amount to the retailer. Based on this information the retailer calculates

the optimal load shifting scheme for each consumer and remotely controls the

consumers’ devices accordingly. Finally, the consumers receive compensation

according to the prior defined incentive scheme.

In rate-based DR programs the consumer receives price signals from the retailer.

Prices can be designed either as time-of-use (TOU) prices, i.e. defined prices for

specific time intervals, or dynamic real-time prices. Based on the pricing information,

the consumer adapts the energy consumption and shifts load to more advantageous

times.

Consumer-induced programs require an initial action from the consumer. The

consumer sends bids towards the retailer offering to shift or curtail load. In return, the

retailer pays the consumer for changing the load scheme.

For the sake of simplicity, but without loss of generality, we assume an incentive-based

program that equips the retailer with direct load control. This approach ensures full

efficiency, i.e. all adjustments on the consumer side are executed as demanded – there is

no necessity of taking into consideration individual consumer valuation and behavior. An

incentive-based DR program runs in two phases (see Figure 30): (i) forecasting the load for

a future time slot, and (ii) determining load curtailments if necessary.

Figure 30. Information Flow in an Incentive-based Demand Response System.

Weather

Forecast

Electricity

Prices

Meter

Data

Sensor

Data

DR

Contingencies

Control

Signals

(ii) Determination & execution of

load shifting scheme

D

EElectricity

Load

A

G

B

H

(i) Forecasting of electricity load

for next time interval

Historical

Usage

C

Network

Model

F

External

Service

Provider

Electricity

Retailer

Consumer


131

In the following, we describe the two phases in detail.

Load forecasting. Initially, the retailer forecasts the energy demand and supply for a

future time window based on the historic and current electricity consumption. The

information concerning the consumption is extracted from the usage data records ,

which the retailer collects from the connected smart meters. Historic consumption

data is compiled and stored in a corresponding database. In addition to the

consumption data, the retailer taps two further data sources, namely weather

forecasts and electricity prices , from external sources. The retailer receives

electricity prices from service providers, e.g. the European Energy Exchange, EEX for

short (European Energy Exchange (EEX), 2012). The retrieved information includes

the historic trend of intraday prices, data concerning future derivatives, and the day-

ahead market. In case the load forecasting algorithm detects a shortage in supply, a

suitable load curtailment scheme has to be computed.

Load curtailment. The projected load forecast serves as input for determining the

load curtailment scheme for all consumers that are connected to the distribution

network. For an incentive-based DR program, the computation follows a three-step

approach. First, the retailer requests the consumers to transmit their available DR

contingencies (i.e. their maximal shift duration and shiftable power amount) for the

next optimization interval . In the second step, the retailer uses the projected energy

demand along with the network model and contingencies to determine an optimal

load shifting schedule. The electricity prices are also taken into consideration for the

calculation. Finally, the retailer informs the consumers about the required actions for

the next interval. Hence, the retailer sends a set of control signals to the smart

meters or load control devices, and induces the required load shifts.


132

It should be noted that the illustrated communicational exchange between the retailer and

consumer – intentionally – includes some simplifications11

.

4.2 Reference IS Architecture of the Demand Response System

The prior described information flows of a DR program can be translated into a

corresponding system design, as shown in Figure 31. Basis for our documentation of the IS

architecture is the Energy Informatics Framework (Watson et al., 2010). The Energy

Informatics Framework defines five relevant categories structuring an energy-related

information system: (i) a central information system, (ii) sensor networks, (iii) sensitized

objects, (iv) external stakeholders, and (v) an interface to the flow networks. The DR

system is grouped into the categories of the Energy Informatics Framework as follows.

Figure 31. System View on a Demand Response Information System with Advanced Metering

Infrastructures based on Fan et al. (2013), Lo and Ansari (2011) Mohagheghi et al. (2010),

Simmhan et al. (2011), Sui et al. (2011) and Watson et al. (2010).

Central Information System

As the DR program is realized on distribution network level, the Distribution Management

System (DMS) is the core information system for the retailer. It features a collection of

components and applications allowing for network monitoring and dynamic decisions to

11 In a concrete, executable design of a communication protocol, additional notification and status messages are presumably required. Moreover, security related to smart grid communications is currently a heavily discussed topic in academia

(McDaniel and McLaughlin (2009); Metke and Ekl (2010)). Certain security requirements might thus induce additional

message transfers. Furthermore, all status and monitoring data (e.g. regarding the power quality), which is being transmitted across the network, is not part of our considerations.

Sensitized ObjectsSensor NetworksInformation SystemExternal Stakeholders

Flow Networks

Energy

Mgmt.

System

Generation

Info. &

Trading

System

Weather

Forecast

Provider

B

Load Control

Device

Distributed

Energy

Resource

A

A Weather Forecasts B Electricity Prices C Historical Usage Data D Current Usage DataLegend:

Distribution

Management System

Demand Response

Engine

Load Forecasting

Engine

E

Historical

usage

Network

Model

C G

Energy

Markets

E Electricity Demand F DR Contingencies G Network model H DR Control Signals

Smart

Meter

Meter Data

Management

System

Load Mgmt.

& Control

System

Consumption

Management

System

Data Con-

centrator

Network

Management

System

H

D F

Data Con-

centrator

upstream

downstream


133

optimize resources and manage demands for the entire distribution network (Simmhan et

al., 2011). The distribution network model is heart and soul of the DMS (Uluski, 2011). It

comprises the network model and the load model of the distribution network. The network

model documents the physical network infrastructure, namely energized components, such

as wires, switches, transformers, and their interconnections. To support the execution of

DR programs, the network model needs to describe the infrastructure down to a single

meter, which is not the case in most of today’s models (Uluski,). The load model provides

insights into the historic usage data of consumers in form of load profiles. In future, theses

load profiles can be enhanced by updating them in regular intervals with collected usage

data from the smart meters. Beyond these models, two application modules within the

DMS are of special interest when executing DR programs: the load forecasting engine, to

predict shortages of supply, and the demand response engine, to determine an optimal load

curtailment scheme.

Sensor networks

With the central information system in place, the retailer has to establish the link towards

the consumers by setting up a sensor network. For this purpose, many electricity retailers

have been employing direct load control programs building on rather simple technology,

such as one-directional communication, for many years. However, “the ability to leverage

two-way communications with greater control and accuracy, and with customer input, can

substantially enhance a retailer’s DR capabilities” (McDonald, 2008). Consequently, we

are focusing on a DR system based on an AMI. An AMI is the only sensor network that

fulfills the requirements of near real-time applications and which is suitable for two-way

communications. An AMI comprises both back-end and monitoring infrastructure of the

retailer and smart meters that are deployed at the consumers’ premises. This is where the

previously mentioned communication standardization plays an important role by ensuring

the interoperability between different system components (Fan et al., 2013). We denote

smart meters as electronic meters that collect energy consumption data of the consumer,


134

so-called usage data records, at defined time intervals and optionally feature a unit for two-

way communication (Brophy Haney et al., 2009).

Data within an AMI can be transferred through various types of communication

technology (see e.g. Khalifa et al., 2011; Lo and Ansari, 2011; Fan et al., 2013). However,

mobile phone networks (GSM and its successors) and power line carrier (PLC) are

regarded as the most suitable technologies for smart grid applications (see e.g. Galli et al.,

2011; Gungor et al., 2011; Sauter and Lobashov, 2011; Llaria et al., 2013). They combine

cost-effectiveness and bandwidth, while not requiring the installation of additional cables.

GSM-enabled meters establish direct point-to-point connections with the AMI back-end

leveraging the widely available wireless GSM network. The communication stream from

PLC-enabled meters is sent through data concentrators that are directly connected to the

back-end (see Figure 32). PLC has traditionally been used by electricity retailers for

remote metering and (low-bandwidth) load control applications. The communication

activities are monitored by a Network Management System (NMS), which supervises and

controls the connected network components.

Figure 32. Detailed View on Differences of Required Infrastructures for Communication

Networks based on GSM (top) and PLC (bottom).

Independent of the chosen communication technology, the collected usage data records

reach the central Meter Data Management System (MDMS). The MDMS receives the

usage data records, preprocesses them and transforms them into information pieces that

can be further refined and used in the retailer’s backend system (e.g. billing). Hundreds of

thousands smart meters produce and transmit usage data records in an average smart

distribution network. Hence, a MDMS needs to be enabled to handle the corresponding

large data amounts. The same accounts for traditional DMS that can be easily overloaded

Meter Data

Management

System

Smart

Meter

PLC

Module

Meter Data

Management

System

Data

Concen-

trator

Neighborhood

Area Network

(NAN)

Wide Area

Network

(WAN)

Sensitized

Objects

Mobile

Network

Home Area

Network (HAN)

Smart

Meter

GSM

Module

Sensitized

Objects

Home Area

Network (HAN)


135

by the data volume. Latest developments in database technology, such as methods and

tools for processing big data, provide viable solutions for this challenge. An example for

upgrading and boosting the performance of a DMS can be found in Li et al. (2010).

Smart meters including communication modules, data concentrators, the NMS, and the

MDMS build the upstream channel of a DR program in an AMI. The opposite direction,

namely the downstream channel, consists of a central load management and control system

(LMCS) and consumption management systems. The LMCS distributes DR signals from

the back-end to the consumption management systems. These systems are the counterpart

of smart meters on the downstream channel forwarding the transmitted signals towards the

sensitized objects at the consumers’ premises. Smart meter and consumption management

system are often integrated into a single component.

Sensitized objects

In a DR environment, the sensitized objects are represented by energy consuming (i.e. load

control devices) and producing (i.e. distributed energy resources) devices. These devices

are operated by the consumers and installed on their premises. Controlling these devices

through a DR system requires an interface with the smart meter and the capability of being

controlled remotely. The residential as well as industrial devices and appliances are

connected via a home area network (HAN) (Fan et al., 2013).

Flow networks

Interfaces to field (e.g. switches) and substation devices on distribution network level as

well as the energy management system at the transmission network level secure

information exchange with neighboring and overlying levels of the power network. These

interfaces are often considered being part of the DMS (Uluski, 2011).


136

External stakeholders

The DR system is completed by interfaces for automated data exchange with various

service providers for acquisition of external data sources. As described in the previous

section, the retailer acquires electricity prices and weather forecasts from external service

providers for executing a DR program.

5 Definition of the Cost-Value-Model

In this section we present the cost-value-model to gauge both costs and saving potentials of

implementing a DR program. First, we investigate the costs which the components of the

previously described DR infrastructure generate. Moreover, we derive an optimization

problem to model optimal load shifts and consequently allowing the retailer to procure

cheaper electricity. The cost-value-model is based on Feuerriegel et al. (2013) and

Feuerriegel and Neumann (2014) – i.e. our previous work, where we use a static model to

determine costs and saving potentials of a single year. Now, we extend the model to cover

a multi-year scope and thus support evaluations in the domain of financial accounting.

Finally, we integrate the cost and saving perspectives in a combined metric, namely the

information value for demand response.

5.1 Model Assumptions

Our cost-value-model is based on a number of assumptions, which we present in the

following.

We do not consider any cross-effects of the DR programs. Sample cross-effects range

from the impact on the electricity prices to a fundamental reduction in electricity

demand.

In order to determine the savings, we only consider load shifting. Additional DR

mechanisms, such as peak clipping, are not part of the model (for details see


137

Feuerriegel et al., 2013). Furthermore, we focus on the day-ahead electricity markets,

while neglecting other markets, e.g. for reserve energy.

The model assumes that the retailer owns the DMS and the sensor network. In reality,

the ownership and operations of the infrastructure is distributed among several parties,

especially in liberalized electricity markets. This may require additional efforts to

secure data transfer between the parties.

The model considers the IS and communication infrastructure which is attributed to

the retailer. The components and communication networks of the end user, namely the

sensitized objects and HAN, are not integrated. There are other studies that

concentrate on this part of the smart grid infrastructure (Gottwalt et al., 2011). We

presume the availability of remotely controllable devices on the premises of all end

users, as this is a necessary condition for the efficient use of DR programs in our

setup.

5.2 Cost Structure of IS Components for Demand Response

This section provides a detailed insight into the relevant cost components of a DR system,

both for initial setup and running operational efforts. The cost components are derived

from our previously defined reference IS architecture of a DR system that bases on an AMI

infrastructure.

Scaling the Demand Response System

The number of deployed smart meters at a certain time 𝑡 is the starting point for all

subsequent calculations of the cost model. The number of residential and industrial

consumers determines the overall number of smart meters 𝑥SM. Without loss of generality,

we assume that all smart meters are rolled out to the consumers in the beginning of the first

year. Hence, the full DR capacity is available throughout the whole observation period.

According to a defined parameter 𝜀, the smart meters are equipped with communication


138

modules. The parameter comprises knowledge about the number of residential buildings

and industrial sites, as well as the ratio of communication units per building. Let 𝛽GSM be

the share of deployed meters with communication modules that use GSM technology. The

number of GSM communication modules 𝑥GSM is given by

𝑥GSM = 𝑥SM ∙ 𝜀 ∙ 𝛽GSM, (20)

and, analogously, the number of PLC communication modules 𝑥PLC equals

𝑥PLC = 𝑥SM ∙ 𝜀 ∙ (1 − 𝛽GSM). (21)

Besides the number of deployed smart meters, the generated and transferred data volume at

each smart meter determines the sizing of the back-end systems. The amount of data

obviously depends on the interval of reading out usage records 𝑓 and defining DR control

signals 𝑓DR.

IS-Related Investments for Demand Response

Before operating, the DR system needs to be rolled out and initialized first. This

installation effort requires substantial capital expenditures. The cost model is designed to

permit a step-wise infrastructure deployment, which extends over several years. The

variable 𝑇 denotes the overall observation period corresponding to the infrastructure

lifetime.

For each of the required components, costs are split into hardware and installation costs.

The hardware investments for a component 𝑦 ∈ {𝑆𝑀, 𝐺𝑆𝑀, 𝑃𝐿𝐶} are indicated by ℎ𝑦

whereas the costs for software and installment are denoted by 𝑖𝑦. We assume, without loss

of generality, that software licenses are paid once in advance by the retailer. The cost

model can be easily adapted to incorporate annual fees as well. In summary, the total

infrastructure investments for the DR system depending on a certain readout frequency 𝑓

are defined as

𝐼(𝑓) = 𝐼SM + 𝐼C + 𝐼N + 𝐼M(𝑓) + 𝐼E. (22)


139

Next, we take a closer look at the investment calculation for the smart meters. This

calculation is composed of the smart meter itself and the attached communication modules

for GSM and PLC networks. As stated earlier, only a fraction of smart meters is equipped

with communication modules, thus giving

𝐼SM = 𝑥SM ∙ (ℎSM + 𝑖SM) + 𝑥GSM ∙ (ℎGSM + 𝑖GSM) + 𝑥PLC ∙

(ℎPLC + 𝑖PLC).

(23)

We assume that the features of a LMCS are integrated into the smart meters.

The dimensioning of further metering infrastructure components, i.e. data concentrators,

NMS and MDMS, depends on the number of deployed smart meters

(PricewaterhouseCoopers, 2010). Moreover, scaling of the MDMS is strongly dependent

on the used read-out interval as this determines the data load that has to be processed and

prepared for further analysis. Hence, the initial investments for data concentrators (𝐼C),

NMS (𝐼N) and MDMS (𝐼M) are defined as

𝐼C = ⌈𝑥PLC ∙ 𝜔C⌉ ∙ 𝑖C. (24)

𝐼N = ⌈𝑥SM ∙ 𝜔N⌉ ∙ 𝑖N. (25)

𝐼M(𝑓) = ⌈𝑥SM ∙ 𝜔M(𝑓)⌉ ∙ 𝑖M. (26)

Installation costs of a single unit of each of these components are denoted by 𝑖C, 𝑖N and 𝑖M.

Variables 𝜔C, 𝜔N and 𝜔M(𝑓) indicate the number of smart meters that can be handled by a

single unit of the component.

The initial costs of a DMS basically consist of system development and integration efforts

(PricewaterhouseCoopers, 2010). We do not further detail these costs and set them as

𝐼E = 𝑘. (27)

The annual depreciation can be derived from the annual investment costs. It is convenient

to use 𝑇 for both the lifetime of the infrastructure as well as the depreciation period. Thus,

the annual depreciation, i.e. the annual fix costs 𝐶FIX for the DR systems, are computed as


140

𝐶FIX(𝑓) =𝐼(𝑓)

𝑇. (28)

IS-Related Operating Costs

The operating costs cover the annual expenditures for e.g. maintenance, personnel and

energy demand of all components of the DR system. Based on our assumption that all

smart meters are rolled out at once, the operating costs are assumed to be independent of

the year 𝑡. As for the investment costs, the operating costs of data concentrators, NMS and

MDMS, depend on the number of deployed smart meters. Hence, the annual operating

costs 𝑐OPS(𝑓) total up to

𝑐OPS(𝑓) = 𝑥SM ∙ 𝑐SM + ⌈𝑥PLC ∙ 𝜔C⌉ ∙ 𝑐C + ⌈𝑥SM ∙ 𝜔N⌉ ∙ 𝑐N +

⌈𝑥SM ∙ 𝜔M(𝑓)⌉ ∙ 𝑐M(𝑓) + 𝑐E.

(29)

Communication Costs

Besides the aforementioned costs, additional costs in DR programs arise from the

necessary data transfers. Let 𝑐GSM(𝑡, 𝑣) be the related price function in year 𝑡 for data

transfer of volume 𝑣 via GSM networks; the price function incorporates an annual decline

in communication prices. Communication via PLC-enabled meters does not produce any

volume-based costs. Here, we need to analyze the following two major communication

streams, namely (i) reading out usage data records, and (ii) executing DR programs.

The usage data records are read out by the smart meter at a certain interval 𝑓 measured in

minutes. The collected records, each of size 𝛼UD, are regularly transferred from the smart

meter to the retailer for further processing; the number of transfers per day is denoted

by 𝑔𝑇𝐹. The total data volume per transfer is composed of the actual payload, i.e. 𝑓

1440∙𝑔TF

and additional overhead 𝛼OH. All in all, this results in an annual data volume of

𝑣UD(𝑓) = 365 ∙ 𝑔TF(𝛼UD𝑓

1440 ∙ 𝑔TF+𝛼OH). (30)


141

The same approach is required for defining and sending out the DR control signals from

the retailer towards the consumer. The interval between definitions of control signals is

labeled with 𝑓𝐷𝑅; without loss of generality we set 𝑓𝐷𝑅 = 𝑓. The two-step communication

for executing the DR program, namely retrieving DR contingencies and sending control

signals (see Figure 30), adds up to a total volume 𝛼DR per single transaction. The number

of transfers is the same as in the prior case of the usage data records. Thus, the annual data

volume for executing DR controls is

𝑣DR(𝑓) = 365 ∙ 𝑔TF(𝛼DR𝑓

1440 ∙ 𝑔TF+𝛼OH). (31)

All in all, the total annual communication costs for all deployed meters in year 𝑡 account

for

𝑐COM(𝑓, 𝑡) = 𝑥GSM ∙ 𝑣(𝑓) ∙ 𝑐GSM(𝑡, 𝑣(𝑓)) (32)

with 𝑣(𝑓) = 𝑣𝑈𝐷(𝑓) + 𝑣𝐷𝑅(𝑓).

Total IS-Related Run Costs

The overall run costs 𝐶𝑉𝐴𝑅 for a DR system in year 𝑡 are computed by adding up operation

and communication costs, i.e.

𝐶VAR(𝑓, 𝑡) = 𝑐OPS(𝑓) + 𝑐COM(𝑓, 𝑡). (33)

To determine the cost components, we do not consider any support costs (e.g. setup of a

customer call center), related process costs (e.g. registration of new meters) and additional

efforts for integrating the DR system into the existing IS landscape. Moreover, the costs

only cover IS-related expenses; expenses concerning operations of the underlying power

grid network are not taken into consideration.

5.3 Savings through Demand Response

The costs of a DR system are opposed by the savings that a retailer can realize with

shifting load. The savings mainly result from shifting load from peak times to times when


142

electricity can be purchased cheaper. The determination of the retailer’s savings is based

on the model of Feuerriegel and Neumann (2014). The problem of optimally harnessing

demand response is formulated as a linear optimization problem with an optimization

horizon of 𝑁 time steps given by

min𝑞F ,𝑞A(1),…,𝑞A(𝑁)

(𝑝F𝑡𝑞F +∑ 𝑝A

𝑡 (𝜏)𝑞A(𝜏)𝑁

𝜏=1). (34)

The price for future options is denoted as 𝑝F𝑡 per watt-hour. In addition, a day-ahead spot

market provides energy at price 𝑝A𝑡 (𝜏) per watt-hour for a specific time 𝑛 of the day. The

parameters 𝑞A(𝜏) and 𝑞𝐹 indicate the demanded quantities in the day-ahead auctions and

the future options respectively. The linear optimization problem minimizes the retailers’

expenditures on electricity derivatives consisting of future options (first summand) and

day-ahead auctions (second summand).

Let 𝐷(𝜏) denote the total amount of energy demanded by the retailer at a given time

interval 𝜏. Furthermore, the retailer has a certain load shifting potential at his disposal

indicated by variables 𝛥𝑗(𝜏), where 𝑗 denotes the maximum length of the shift in hours.

The value of 𝐷𝑅𝑗(𝜏, 𝜏′ − 𝜏) denotes the amount of power that is consumed less at time

step 𝜏, but that is additionally required at time step 𝜏′. The following constraints ensure that

both required energy amounts are available at each time interval and shifted loads come

into use in time

qA(τ) + qF N⁄ = D(τ) for τ = 1,… , N, (35a)

qA(τ) + qF/N = D(τ) − DR1(τ, 0) − DR2(τ, 0) + ⋯ + DR1(τ +1,−1) + DR1(τ − 1,+1),

(35b)

DR1(τ, 0) ≤ Δ1(τ), DR2(τ, 0) ≤ Δ2(τ), etc. , (35c)

𝐷𝑅𝑗(𝜏 + 𝑖, − 𝑖), 𝐷𝑅𝑗(𝜏 − 𝑖, +𝑖) ≥ 0 for all 𝑗 and 𝑖 = 1,… , 𝑗, (36)

∑ 𝐷𝑅𝑗(𝜏 + 𝑖, −𝑖)+𝑗𝑖=−𝑗 = 0 for all 𝑗, (37)

qA(τ) ≥ 0 and qF ≥ 0 for τ = 1,… , N. (38)


143

The given optimization problem determines the retailer’s total energy expenditures when

applying a DR program. We transform the optimization problem by defining the

corresponding Lagrange function. For ease of illustration, 𝑔(𝑓, 𝑡) abbreviates the

minimization problem as defined in equation (34). The side constraints and related

Lagrange multipliers are labeled with ℎ𝑖(𝑓, 𝑡) and 𝜆𝑖 respectively.

For determination of potential savings for the retailer we have to calculate the expenditures

𝐸(𝑓, 𝑡) that arise without demand response in use. Combined, the revenue function 𝑅(𝑓, 𝑡),

which indicates the retailer’s savings, can be formulated as a function of read-out

frequencies 𝑓 and the respective year 𝑡

𝑅(𝑓, 𝑡) = 𝐸(𝑓, 𝑡) − 𝑔(𝑓, 𝑡) +∑ 𝜆𝑖ℎ𝑖(𝑓, 𝑡)𝑁

𝑖=1. (39)

5.4 Value of Information in Demand Response

In order to combine the two previously discussed views – namely costs and savings

potentials of a DR system – we introduce an expedient metric named information value of

demand response. In the smart grid realm, the core information piece is given by the usage

data records. Their value for the retailer is determined by multiple components, which can

be derived from the numerous areas for which these records can be utilized, such as billing

process optimization through automated meter reading and optimization of energy demand

and supply. Here, we focus on the contribution of demand response to the overall value of

usage data records. Consequently, we define the information value of demand response of

a single usage data record in year 𝑡 as

𝐼𝑉VAR(𝑓, 𝑡) =𝑅(𝑓, 𝑡) − 𝐶VAR(𝑓, 𝑡)

𝐹(𝑡) (40)

with 𝐹(𝑡) = 𝑀 𝑓⁄ being the total number of readout operations in year 𝑡 and 𝑀 the number

of minutes per year. Based on 𝐼𝑉VAR, the cumulated information value covering the whole

observation period, thereby incorporating possible price declines for hardware and the like

over the years, is defined as


144

∑𝐼𝑉VAR =1

𝑇∑

𝑅(𝑓, 𝑡) − 𝐶(𝑓, 𝑡)

𝐹(𝑡)

𝑇

𝑡=1

. (41)

We supplement 𝐼𝑉VAR with a second metric that incorporates capital expenditures as well

𝐼𝑉TOT(𝑓, 𝑡) =𝑅(𝑓, 𝑡) − 𝐶VAR(𝑓, 𝑡) − 𝐶FIX(𝑓, 𝑡)

𝐹(𝑡). (42)

6 Evaluation of the Demand Response System

This section evaluates the financial performance of the DR system. We start by defining a

comprehensive analysis environment reflecting a real-world setting. Next, we assess the

derived results along the defined research questions and derive general propositions.

Finally, we deduce managerial and policy implications.

6.1 Setup and Dataset of Computational Analysis

As we do not have access to a physical smart grid test bed, we apply a model-driven

research approach to validate our cost-value model – an effective method according to

(Goebel et al., 2014). In order to stay as close to reality as possible, we populate our

previously set up model with publicly available data derived from existing electricity

markets.

We take the perspective of a typical German electricity retailer distributing an annual

electricity amount of 2,000 GWh into the network (E-Control, 2012). Industrial customers

account for about 50% of the total German electricity consumption (Styczynski, 2011). We

exclude this customer segment from further consideration, as industrial companies often

have special agreements concerning the provisioning of electricity and the affiliated load

management activities. The remainder of 1,000 GWh divides equally between residential

households and commercial customers (Styczynski, 2011). According to figures from E-

Control (2012), the annual average energy consumption amounts to 3,500 kWh per

residential household and 6,500 kWh per commercial customer. Taking this distribution

into account, the retailer’s customer base in our setting thus comprehends 143,000


145

residential households – corresponding to 290,000 residents – and 77,000 commercial

customers. Connecting all customers to the DR infrastructure consequently requires

220,000 smart meters. Not all smart meters need to be equipped with an own

communication module, in larger apartment buildings they can be bundled. With

consideration of the distribution of household and building sizes in Germany, the required

communication modules add up to 130,000. To determine electricity usage per hour, we

use the aggregated average daily load profiles for households and commercial customers in

Germany (E.ON, 2012). These standard load patterns are even often applied by retailers in

practice. The capabilities of shifting load vary strongly among both commercial consumers

and residential households (see Table 33).

Commercial

Consumers

Max. Shift Duration (in h) 1 2 12 16

Average Power Shift (in kW) 16,164 8,850 4,650 5,129

Residential

Households

Max. Shift Duration (in h) 1 2 12 24

Average Power Shift (in kW) 7,353 6,750 35,068 3,616

Table 33. DR Potential Through Load Shifting Scaled for the Electricity Retailer in our

Evaluation Setting based on Klobasa (2007).

We evaluate our model in the two previously defined scenarios, namely the basic scenario

and the restricted scenario. In both scenarios, the smart meters are rolled out stepwise to

the consumers; the full coverage is reached after a period of six years. We have chosen a

total observation period of 15 years, since this resembles the typical lifetime of a DR

infrastructure (PricewaterhouseCoopers, 2010). After the 15 years, new investments are

likely to be made to upgrade the infrastructure.

To calculate potential savings we use historic hourly data of the EEX to configure the cost-

value-model (European Energy Exchange (EEX), 2012). The price for future options is

based on the index prices named ELIX Day Base. All additional configuration parameters

of the cost model are summarized in Table 34.


146

Parameter Variable Value

Share of GSM-enabled smart meters 𝛽GSM 20%

Volume-based communication costs for GSM 𝑐GSM Logarithmic cost function

Hardware costs per meter, including GSM-/PLC-module 𝑐SM € 95 / € 80

Number of PLC-enabled meters per concentrator 1𝜔C⁄ 200

Number of meters per MDMS server 1𝜔M⁄ 30,000 for 𝑓 = 15 minutes

Number of meters per NMS 1𝜔N⁄ 1, i.e. NMS assumed to be

continuously scalable

Size of messages per iteration of DR program 𝛾 300 Byte

Table 34. Parameters of Demand Response System based on Expert Estimations,

PricewaterhouseCoopers (2010) and NERA Economic Consulting (2008).

The given optimization model is very complex and extensive; computing continuous

values for the revenue function is not feasible. This is particularly true for very small

values of 𝑓. Hence, we approximate the revenue function by a quadratic polynomial for a

given optimization model configuration (Judd, 1998). Model specification has been tested

with AIC and BIC - with higher polynomials the fit is decreasing again. Basis for

execution of the polynomial interpolation is a discrete number of revenue values that have

been computed by solving the original optimization problem.

6.2 Results

Existence of a Competitive Advantage for the Retailer (RQ1)

Starting point is a setup with an information granularity of 60 minutes. In this default

setup, the usage data is recorded once every 60 minutes or 24 times per day respectively.

To determine the cost structure of a DR system, we assess the cost-value-model under two

conditions: (i) the prices for GSM communication stay constant throughout the observation

period, (ii) prices decrease annually. The first variant is denoted as ΩCON. For the second

variant we assume two degrees of annual price decline of the GSM communication costs,

namely ΩMIN = 5% and ΩMAX = 15%.


147

Looking at the IT-related run costs (i.e. 𝐼𝑉VAR), the electricity retailer is able to realize a

positive value of information only under certain levels of Ω and after a temporal offset.

Figure 35 provides an overview on the annual information value. In year 15, the DR

program generates an information value 𝐼𝑉VAR per 1,000 meter readouts between € -0.013

for ΩCON (constant communication costs) and € 0.103 for ΩMAX (maximum price decline).

The picture further deteriorates when looking at 𝐼𝑉TOT, i.e. taking the required capital

expenditures into consideration as well. The roll-out requires immense investments,

especially for deploying the smart meters at the consumers’ premises. In our setting, the

infrastructure exhibits a negative overall net present value (NPV) of € -22.37 M for

scenario ΩCON with constant communication costs, a runtime of 15 years and a prevalent

discount rate of 5%. A marginally better result can be achieved when assuming a scenario

with maximal price decline (ΩMAX); the NPV is still negative as accounting for € -

20.91 M.

Figure 35. Development of the Variable Information Value 𝑰𝑽𝐕𝐀𝐑 of the Demand Response

System Depending on the Assumed Degree of Price Declines for GSM-Transactions across the

Observation Period. The Figure Shows the Cumulated Information Value for 1,000 Meter

Readouts and a Meter Readout Interval of 60 Minutes.

In the following, we assess which parameters need to be adjusted to obtain a profitable

system configuration. First, we take a look at the used communication technology and

equipment. As data transfer is at the core of a DR program, it is obviously that these

-0.0

20

.00

0.0

20

.04

0.0

60

.08

0.1

0

Year

Info

rmati

on

Valu

e (

per

1,0

00 r

ead

ou

ts),

in

EU

R

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15

IVVAR in scenario CON

IVVAR in scenario MIN

IVVAR in scenario MAX


148

components have a substantial impact on the costs of the DR infrastructure. In the case of

communication via GSM networks, usage-based charges have to be paid in addition to

base fees. In contrast, PLC communication only generates base fees without any usage-

based premiums. Hence, the number of smart meters with GSM communication modules

considerably determines the annual costs. To assess the exact impact, we conduct a

sensitivity analysis by varying the share of smart meters with GSM modules.

The previously shown results are computed assuming a share of smart meters with GSM

modules of 20 % (see Table 34). An interesting question is at which GSM share the system

yields a positive information value and hence gets profitable. For this, we conduct a

sensitivity analysis and vary the percentage of GSM data communication; the rest of

parameters are kept constant. In order to determine the maximum share of GSM

communication, we set the information value to zero and isolate the GSM share 𝛽GSM in

the equation. Figure 36 illustrates the result of the parameter analysis.

Figure 36. Results of Parameter Analysis Showing Impact of GSM Share on the Variable

Information Value (𝑰𝑽𝐕𝐀𝐑) in a Demand Response System. The Figures are based on a Meter

Readout Interval of 60 Minutes and Show Values in Year t=15.

When only considering the run costs, the share of GSM-based communication has to be

below 19.60 % in constant scenario ΩCON and 24.50 % in scenario ΩMAX to achieve a

profitable system configuration. Taking the capital expenditures into calculation as well,


149

the information value is negative for any given GSM share in the chosen system

configuration; even when all communication is carried out via PLC, a positive surplus

cannot be achieved. The share of GSM transactions, i.e. 𝛽GSM, substantially determines the

profitability of a DR system, as Proposition 1 shows.

Proposition 1. The information value 𝐼𝑉𝑉𝐴𝑅 is positive for all shares of GSM

communication that solve both

𝛽GSM ≥ 0 𝑎𝑛𝑑 𝛽GSM <𝑇(𝑅(𝑓) − 𝑐OPS(𝑓))

𝑥SM ∙ 𝜀 ∙ 𝑣(𝑓)∑ 𝑐GSM𝑡 (𝑣(𝑓))𝑇

𝑡=1

. (43)

Proof. A positive information value solves the equation

∑𝐼𝑉VAR(𝑓)

𝑡

=𝑅(𝑓) − 𝐶VAR(𝑓)

𝐹> 0. (44)

Applying the definitions of revenue and costs leads to

𝑓

525,600𝑇∙ (𝑇(𝑅(𝑓) − 𝑐OPS(𝑓, 𝑥SM)) − 𝑥GSM𝑣(𝑓)∑ 𝑐GSM

𝑡 𝑣(𝑓)𝑇

𝑡=1)

> 0.

(45)

Transformation of this equation results in a range of values for 𝛽GSM, which ensure a

positive information value. ∎

As shown, a positive information value cannot be achieved in full cost calculation (i.e.

taking fix costs into consideration as well as run costs) when only the share of GSM

communication is taken as basis for a parameter analysis. Besides the GSM share, the

installation and operation costs of smart meters are a critical cost driver. Due to their

immense amount smart meters have a significant influence on the total cost of a DR

system. Hence, we broaden our analysis by adding the costs of smart meters to the set of

adaptable parameters. The parameter set contains the share of GSM communication and

the costs of smart meters.

Even for this extended parameter analysis noticeable reductions of parameter values are

required to obtain a profitable system configuration. Compared to the initial system


150

configuration, the parameters under consideration have to be decreased by 30.84 % in

constant scenario ΩCON, by 29.64 % for ΩMIN and by 28.30 % for ΩMAX to achieve a

positive total information value.

Figure 37. Results of Extended Parameter Analysis Combining Variations of GSM Share and

Costs of Smart Meters. The Curves Resemble the Cumulated Information value of 𝑰𝑽𝐓𝐎𝐓. The

Figure is based on a Meter Readout Interval of 60 Minutes.

Figure 37 illustrates the results of the extended parameter analysis. Details concerning the

adjustment of the particular parameters are found in Table 38. The initial hardware prices

of smart meters need to be as low as € 65.70 (GSM-enabled) and € 55.39 (PLC-enabled) in

constant scenario ΩCON. The reduced hardware prices are far below current market prices.

They might be only achievable by the purchasing power of a very large electricity retailer.

Scenario Decrease of

Parameters

GSM Share Cost of GSM-

enabled Smart

Meter

Cost of PLC-

enabled Smart

Meter

Initial System Configuration

Initial --- 20.00% € 95.00 € 80.00

Target Values of Parameters

ΩCON -30.84% 13.83% € 65.70 € 55.39

ΩMIN -29.64% 14.07% € 66.84 € 56.29

ΩMAX -28.30% 14.34% € 68.12 € 57.36

Table 38. Results of Parameter Analysis for Two Scenarios of Price Decline. Target Values of

Parameters are Required to Obtain Profitable System Configuration.

-0.6

-0.4

-0.2

0.0

0.2

0.4

Variation of parameters

Info

rmati

on

Valu

e (

per

1,0

00 r

ead

ou

ts),

in

EU

R

-50% -45% -40% -35% -30% -25% -20% -15% -10% -5% -0%

Cum. IVTOT in scenario CON

Cum. IVTOT in scenario MIN

Cum. IVTOT in scenario MAX

-0.1

0-0

.05

0.0

00

.05

0.1

0


Info

rmati

on

Valu

e (

per

1,0

00 r

ead

ou

ts),

in

EU

R

-35% -34% -33% -32% -31% -30% -29% -28% -27% -26% -25%

-0.1

0-0

.05

0.0

00

.05

0.1

0


Info

rmati

on

Valu

e (

per

1,0

00 r

ead

ou

ts),

in

EU

R

-35% -34% -33% -32% -31% -30% -29% -28% -27% -26% -25%

-0.1

0-0

.05

0.0

00

.05

0.1

0


Info

rmati

on

Valu

e (

per

1,0

00 r

ead

ou

ts),

in

EU

R

-35% -34% -33% -32% -31% -30% -29% -28% -27% -26% -25%


151

In summary, the retailer achieves a comparative advantage by operating a DR system. Due

to his superior knowledge of the current demand, he is enabled to decrease peak demands

and optimize electricity purchases accordingly. As a result, the DR mechanism generates a

positive information value – when considering operational costs only – and hence the

retailer gains a financial benefit.

Optimal Information Granularity (RQ2)

To evaluate the previous Research Question 1, we assume a meter read-out frequency of

60 minutes. In fact, 60-minute-intervals are a frequent delivery period when trading at

energy exchanges (European Energy Exchange (EEX), 2012). In the case of Germany, 60-

minute intervals are further enforced by regulatory issues (Bundesnetzagentur, 2012). In

contrast, reviews and studies of many real-world smart meter installments (e.g. Ernst

& Young, 2013) propagate a 15-minute-interval as a standard for collection of usage data

records. European legislation tends towards a minimum read-out interval of 15 minutes as

well (European Commission, 2012). Hence, we continue by varying the frequency of

collecting usage data from the smart meters in a range of 1 to 120 minutes to evaluate our

model across different information granularities.

Figure 39. Information Value 𝑰𝑽𝐕𝐀𝐑 per 1,000 Smart Meter Readouts across Various

Information Granularities, i.e. Read-out Intervals, in Year t=15.

Smart Meter Read-out Interval, in Minutes

Info

rmati

on

Valu

e (

per

1,0

00 r

ead

ou

ts),

in

EU

R

-1.4

-1.2

-1-0

.8-0

.6-0

.4-0

.20

0.2

120 110 100 90 80 70 60 50 40 30 20 10


152

Based on the approximate revenue function and the run costs, we compute the

corresponding information value curve (see Figure 39). The curve is only positive for a

certain range of read-out intervals 𝑓. Solving the maximization problem with integer

solutions yields an optimal information value 𝐼𝑉VAR at a read-out interval of 41 minutes.

Within the set 𝑓 ∈ {21,57}, the information value takes positive values and thus the DR

system generates a surplus for the retailer.

We summarize the findings in the following two propositions, which strongly contradict

the popular belief that more information is better (e.g. Strüker and van Dinther, 2012):

Proposition 2. The information value for demand response increases up to a certain,

optimal information granularity that is determined by solving the linear optimization

model

max𝑓∈ℕ

1

𝑇∑𝐼𝑉VAR(𝑓, 𝑡)

𝑇

𝑡=1

= max𝑓∈ℕ

∑𝑅(𝑓, 𝑡) − 𝐶VAR(𝑓, 𝑡)

𝑇 ∙ 𝐹(𝑡)

𝑇

𝑡=1

. (46)

Here, the target function ∑𝐼𝑉VAR is accumulated over an optimization horizon of 𝑇 steps to

achieve the cumulated information value. The constraints are given by

𝑅(𝑓, 𝑡) ≥ 0, 𝐶VAR(𝑓, 𝑡) ≥ 0 for 𝑡 ∈ [1, 𝑇] and ∀𝑓. (47)

The optimization model for 𝐼𝑉FIX is set up analogously.

Proof. 𝑅(𝑓, 𝑡) is a second-degree polynomial function, i.e. 𝑅(𝑓, 𝑡) = 𝑎𝑥2 + 𝑏𝑥 + 𝑐.

The revenue function determines the polynomial degree of the overall information value

function since the cost function only features first-degree polynomials. The approximation

of the revenue function yields a negative coefficient 𝑎. Hence, a downward opened

parabola represents the information value function. Consequently, the information value

increases up to a certain, optimal value as any parabola exhibits a unique vertex. ∎


153

Proposition 3. When observing the 𝐼𝑉VAR curve in our evaluation setting, its shape

suggests that – in case the optimal information value is positive – only a certain range of

information granularities yield a positive information value. The respective range is

determined by computing the two real roots of 𝐼𝑉VAR and 𝐼𝑉TOT respectively. Both 𝐼𝑉VAR

and 𝐼𝑉TOT are second-degree polynomial functions.

Proof. A downward opened parabola represents the function of the information value (cp.

proof for proposition 2). Hence, it apparently possesses zero, one or two roots. In case the

optimal information value is positive (i.e. larger than zero), the vertex of the parabola is

above the zero line. Then the information value function coercively has two roots – and a

respective range of positive information values can be computed.∎

Figure 40. Revenue and Variable Cost Curve across Various Information Granularities

(Cumulated View over Observation Period of 15 Years).

With these findings at hand, the answer to our research question yields a surprising result.

Contrary to expectations (and current literature), there is not a break even in information

granularity that serves as starting point for continuously rising profits. For the defined

setup and configuration of the DR system, the retailer exhibits a positive information value

for a limited range of read-out intervals only. Figure 40 illustrates the corresponding

revenue and cost curves, which obviously deviate from the revenue and cost curves

showing the expected characteristics as sketched in Figure 29.

20

25

30

35

40

45

50

55

Smart Meter Read-out Interval, in Minutes

Cu

mu

late

d F

inan

cia

l R

esu

lt,

in m

n E

UR

120 110 100 90 80 70 60 50 40 30 20 10

Cumulated Revenues

Cumulated Variable Costs


154

Optimal Smart Meter Roll-Out (RQ3)

In the prior setting for evaluation of Research Questions 1 and 2, we assume that the

retailer employs smart meters at all customers connected to the distribution network. Next,

we want to examine the extent to which retailers can maximize their profit by reducing the

number of deployed smart meters. Smart meters should therefore only be distributed to

financially lucrative consumers. For this purpose, we first determine the marginal costs of

a single smart meter. Here, we compute the marginal run costs per smart meter for two

different read-out intervals 𝑓, namely 15 and 60 minutes. We have chosen these two read-

out intervals as they represent commonly used values in today’s electricity markets.

The revenue is assumed to be linearly related to the addressed energy consumption and can

therefore be simply converted down to a value per MWh to determine the marginal amount

of electricity. The marginal amount of electricity represents the minimal annual amount of

electricity a customer needs to consume in order to become profitable to the retailer in

terms of employing a DR program. The value of the marginal amount of electricity

increases, in line with the marginal costs per smart meter, with an increasing read-out

interval. This is self-evident, since the run costs – as shown in findings related to previous

research questions – grow with an increasing information granularity. Hence, the

electricity retailer should aim at an implementation, where only customers whose power

consumption exceeds a certain threshold, are connected to the DR system. Setting such a

threshold ensures that the customer’s annual consumption yields enough load shifting

potential for amortizing the run costs and communication effort.

Proposition 4. The electricity retailer revenues are maximized in a setup with a limited

smart meter roll-out. The roll-out scheme is determined by the annual consumption of

consumers – installation of a smart meter only pays off for the retailer when a certain

consumption rate is reached. The threshold ∆kWh(𝑓) is determined based on the marginal


155

variable costs 𝐶VAR′ (𝑓, 𝑡) per smart meter and the revenue per kWh of annual consumption

𝑟kWh(𝑓), i.e.

∆kWh(𝑓) =∑ 𝐶VAR

′ (𝑓, 𝑡)𝑡

𝑇 ∙ 𝑟kWh(𝑓). (48)

Proof. The sum of first derivatives of the variable cost function 𝐶VAR with respect to 𝑥SM

determine the marginal run costs of a single smart meter. The division of these marginal

costs by the revenue per kWh results in the threshold that denotes profitable consumers.

We assume that revenue per kWh has a linear appearance and thus can be scaled down. ∎

Next, we select those customers that meet the standards and excel the threshold of annual

electricity consumption. For a read-out interval of 60 minutes, a customer has to consume

an annual electricity amount of 8,407 kWh to spawn a DR potential that can cover the

system costs. The energy consumption of around 0.98 % of all residential households

overshoots this marginal energy amount12

. In addition, 31.66 % of the commercial

consumers fit into the scheme. This limitation on the number of customers leads to 25,758

smart meters that need to be deployed. The same number of communication modules is

required here. Other than in the basic scenario, we assume that each smart meter has to

carry its own communication module. This is due to the fact that the overall number of

residential smart meters is rather low; hence it can be assumed that the respective

households are located in a substantial distance to each other. The overall number of smart

meters corresponds to a fraction of 11.72 % compared to the basic scenario, whereas the

total energy consumption of 258 GWh corresponds to 25.81 %.

Since only profitable, large-scale consumers are equipped with smart meters, the

cumulated revenues exceed the total costs significantly. The surplus is even large enough

to cover the initial capital expenditures. Within the restricted scenario the DR system

12 Residential households are clustered into households with 1,2,3,4 and 5+ members according to the latest census in

Germany (Statistisches Bundesamt (2012)). Within each of the clusters, the electricity consumption is assumed to be

uniformly distributed. For example, for households with 1 member the mean consumption is given as 2,400 kWh per year (co2online (2012)) – the consumption is assumed to be uniformly distributed within limits [1,200 kWh; 3,600 kWh]; the

corresponding standard deviation is calculated as 692.82 kWh. Similarly, the electricity consumption of commercial

customers is assumed to have be arranged around a mean of 6,500 kWh with uniformly distributed values within a range of [1,300 kWh; 11,700 kWh] (E-Control (2012)).


156

yields a positive net present value, namely € 0.89 M. The results of the restricted scenario

for intervals of both 15 and 60 minutes are summarized in Table 41.

Share of HH/CC Total Annual

Consumption

Number of

Smart Meters

Net Present

Value

Read-out Interval: 15 Minutes

Base Scenario 100% / 100% 1,000 GWh 219,781 € -25.45 M

Restricted Scenario 0.66% / 25.00% 209 GWh

(20.90%)

20,172

(9.18%)

€ 0.98 M

Read-out Interval: 60 Minutes

Base Scenario 100% / 100% 1,000 GWh 219,781 € -22.37 M

Restricted Scenario 0.98% / 31.66% 258 GWh

(25.80%)

25,758

(11.72%)

€ 0.89 M

Restricted Scenario - 15 minutes

Marginal Costs per Smart Meter: € 30.14

Marginal Annual Consumption ∆𝑘𝑊ℎ: 9,100 kWh

Restricted Scenario - 60 minutes

Marginal Costs per Smart Meter: € 24.51

Marginal Annual Consumption ∆𝑘𝑊ℎ: 8,408 kWh

Table 41. Comparison of Base and Restricted Scenario across Various Read-out Intervals

(HH = Households, CC = Commercial Customers).

Managerial and Policy Implications

Based on the evaluation of our proposed IS system architecture for a DR system, we derive

a number of contributions and stimuli to support the ongoing debate concerning the

development of the smart grid and introduction of demand response.

Demand response programs can only be realized in grand style when fully-fledged

HANs are in place at the consumers to control devices remotely. To this day, this is

far from reality. Thus, the retailer cannot fully tap the potential of shifting load.

Besides the substantial investments at the consumers’ premises, it requires their

consent. Concerns related to data privacy and security could be veritable roadblocks

here (McKenna et al., 2012).


157

Our model employs an incentive-based DR program. Thus, the consumers need to be

incentivized to take part in the program. Presumably, this can only be achieved by

paying a financial compensation to the consumer. Klobasa (2007) suggest a payment

of € 2 per MWH that is shifted by the consumer. Such compensation further

deteriorates the financial case.

Looking at the substantial investment costs in the basic scenario, an intervention of

the regulator seems to be inevitable to realize a DR system. However, the intervention

could only be justified by non-financial arguments such as decreasing the number of

expensive spinning reserves. Hence, a better approach seems to be provided by the

restricted scenario. Limiting the deployment to the largest consumers ensures a cost-

effective setup where still a substantial amount of energy is shifted to non-peak times.

All in all, the above mentioned aspects cumulate in a single question: who is setting up and

operating a DR infrastructure? This is particularly important in liberalized energy markets

with separated market roles, such as in Germany and other European countries. The

immense investments and operating costs yield veritable business opportunities for new

service providers, such as intermediaries. For them, setting up and operating a DR

infrastructure might generate a viable business case as they can profit from synergies when

offering the infrastructure to more than one electricity retailer (Strüker et al., 2011). In

addition, they may be able to reuse the infrastructure and collected meter data for

additional valuable applications and services (e.g. monthly billing, energy theft

prevention). Taking the value of these applications into account, the total might yield a

positive financial result. These services create additional benefit and no additional IS-

related costs. First thoughts on using cloud-based infrastructures for data collection and

control, and setting up an energy cloud are already on the way (Fang et al., 2012). In such

a scenario, the security and privacy issues would gain particular importance as the

intermediary constitutes a single source of failure (Strüker and Kerschbaum, 2012).

Telecom providers position themselves in the forefront of taking up the role of the


158

intermediary (Mind Commerce, 2012). They are experienced in building up

communication networks or can even reuse parts of their existing communication networks

to set up capabilities for meter data collection and aggregation.

7 Summary and Conclusions

In the near future, the amount of supplied energy will show unprecedented fluctuations due

to the constantly increasing share of power generation from variable resources. Electricity

retailers can address this challenge by leveraging demand response to create a load-shaping

strategy. The implementation of such a strategy calls for the setup of a comprehensive ICT

infrastructure to handle the vast amount of data collected from the smart meters.

As we have shown in this work, savings from load shifting significantly exceed the

running costs of such a DR system. This reasons an interesting business opportunity for

retailers to capitalize on the savings and generate a considerable competitive advantage

through optimization of electricity purchases. Moreover, this work contributes to the

discussion on the optimal information granularity. This research topic is not only restricted

to demand response, but is also extended to smart grid applications (Heinrich and

Zimmermann, 2012; Dalen et al., 2013) and has been gaining increasing traction within the

IS community. We have shown that collecting more information does not necessarily yield

a higher profit for the retailer; increasing revenues are devoured by a disproportionate

increase in costs. In our scenario the retailer can expect a positive information value per

meter readout for readout frequencies between 21 and 57 minutes, with the peak being

reached at an interval of 41 minutes.

However, the tide is turning when considering the huge upfront investments in the

financial equation. These costs do not amortize for a single retailer over the expected

lifetime of 15 years for a DR infrastructure. Hence, alternative solutions are required. This

can either be governmental subsidies, a substantial decrease of installation costs of smart

meters or the use of a different communication technology (esp. beyond GSM).

Alternatively, and this seems to be the most compelling method of resolution, economies


159

of scale or scope can be leveraged. To apply these, information intermediaries are required.

They could profit from the setup of an infrastructure, when offering their services to

multiple retailers.

Finally, we consider a further approach to operate the DR system economically by

restricting the roll-out of smart meters to large-scale consumers whose annual consumption

exceeds a certain threshold. As shown in our evaluation setting, this leads to an overall

share of around 12% of customers that are equipped with a smart meter. These customers

exhibit a level of electricity consumption, which is high enough to ensure payback of the

investment costs for the installed smart meters. A net present value of € 0.89 M is

obtainable for the retailer.

In summary, we deliver a substantial contribution to one of the most pressing research

questions of energy informatics, according Goebel et al. (2014): how can electricity

markets integrate flexible load shifting, and how can economic benefits of information and

information systems be quantified in this context?

Our study still leaves room for future research. This research could particularly aim at the

detailed technical configuration of a DR system, as this has a substantial effect on

profitability. IS research, both in academia and practice, has already endeavored on this.

Installations of smart metering infrastructures and DR systems are, with few exceptions,

still in the pilot phase. (Technical) standards have not yet gained acceptance. Regulatory

efforts might exert a larger impact – the German Federal Office for Information Security

(BSI) (2013) currently prepares the mandatory usage of certain protection profiles for

smart meters. A non-ignorable side effect of this is a noticeably increased data volume.

This again justifies to further evaluate the economic impact of changes in the topology of

the DR system, such as different communication media and protocols.

Moreover, we deem linking household and aggregate level important. In order to analyze

the performance and the benefits of the entire system, also the consumers need to become

part of the equation. First, consumers somehow need to be incentivized to participate in the


160

DR program at all – most probably, this requires a financial compensation. Second, the

consumers have to be equipped with HANs to enable DR mechanisms. So far, we have

restricted the cost model to IS-related costs within the electricity retailer’s realm.

In order to understand the role and importance of intermediaries for the implementation of

demand response, further case studies need to be conducted on the same infrastructure and

underlying data set, and evaluated accordingly. Only by that, the full business case of an

intermediary, including the consideration of intelligent usage of the smart meter data, can

be established. Besides the financial side, aspects like data privacy and data security need

to be included in the consideration.


161

8 References

Aalami, H., Moghaddam, M.P., and Yousefi, G. (2010). Modeling and Prioritizing

Demand Response Programs in Power Markets. Electric Power Systems Research, 80

(4), pp. 426–435.

Aghaei, J. and Alizadeh, M.-I. (2013). Demand Response in Smart Electricity Grids

Equipped with Renewable Energy Sources: A Review. Renewable and Sustainable

Energy Reviews, 18, pp. 64–72.

Albadi, M. and El-Saadany, E. (2008). A summary of demand response in electricity

markets. Electric Power Systems Research, 78 (11), pp. 1989–1996.

Appelrath, H.-J. and Kagermann, H. (2012). Future Energy Grid: Migration to the Internet

of Energy: National Academy of Science and Engineering.

Arnold, G.W. (2011). Challenges and Opportunities in Smart Grid: A Position Article.

Proceedings of the IEEE, 99 (6), pp. 922–927.

Brophy Haney, A., Jamasb, T., and Pollitt, M. (2009). Smart Metering and Electricity

Demand: Technology, Economics and International Experience. Cambridge Working

Papers in Economics: University of Cambridge.

Bundesnetzagentur (2012). Ausgleichsenergiepreis-Abrechnungssystems: BK6-12-024.

Callaway, D.S. and Hiskens, I.A. (2011). Achieving Controllability of Electric Loads.

Proceedings of the IEEE, 99 (1), pp. 184–199.

Capgemini (2008). Demand Response: A Decisive Breakthrough for Europe.

co2online (2012). Stromverbrauch im 2-Personen-Haushalt. http://www.die-

stromsparinitiative.de/stromkosten/stromverbrauch-pro-haushalt/2-personen-

haushalt/infografik-2-2-personen-haushalt/index.html. Accessed 14 February 2014.


Perspective. Proceedings of the 17th Americas Conference on Information Systems.

Dalen, A., Krämer, J., and Weinhardt, C. (2013). Determining the Optimal Level of

Information Granularity for Efficient Energy Consumption Decisions: Experimental

Evidence. Paper 201. Proceedings of the 21st European Conference on Information

Systems (ECIS).

Dedrick, J. (2010). Green IS: Concepts and Issues for Information Systems Research.

Article 11. Communications of the Association for Information Systems, 27 (1).

Dena (2010). dena Grid Study II: Integration of Renewable Energy Sources in the German

Power Supply System from 2015 – 2020 with an Outlook to 2025. Berlin: German

Energy Agency.

E.ON (2012). Standardized Load Profiles: North Rhine-Westphalia. Kassel, E.ON Mitte

AG. http://www.eon-mitte.com/index.php?parent=9215. Accessed 12 May 2012.

E-Control (2012). Energy Consumers' Structure. Vienna, Austria, Energie-Control Austria.

http://www.e-control.at/de/statistik/strom/marktstatistik/verbraucherstruktur. Accessed

14 December 2012.

Ernst & Young (2013). Cost-benefit analysis for the comprehensive use of smart metering

systems.

EU-DEEP (2009). The birth of a European Distributed Energy Partnership that will help

the large-scale implementation of distributed energy resources in Europe. FP6 -

Sustainable Energy Systems Project Reference. Paris, France: European Commission.

http://www.die-stromsparinitiative.de/stromkosten/stromverbrauch-pro-haushalt/2-personen-haushalt/infografik-2-2-personen-haushalt/index.html



http://www.eon-mitte.com/index.php?parent=9215

http://www.e-control.at/de/statistik/strom/marktstatistik/verbraucherstruktur


162

European Commission (2012). Commission Recommendation on Preparations for the

Roll-out of Smart Metering Systems.

European Energy Exchange (EEX) (2012). Market Data. Leipzig, European Energy

Exchange AG. http://www.eex.com/en/Market%20Data. Accessed 15 May 2012.

Fan, Z., Kulkarni, P., Gormus, S., Efthymiou, C., Kalogridis, G., Sooriyabandara, M., Zhu,

Z., Lambotharan, S., and Chin, W.H. (2013). Smart Grid Communications: Overview

of Research Challenges, Solutions, and Standardization Activities. IEEE

Communications Surveys & Tutorials, 15 (1), pp. 21–38.

Fang, X., Misra, S., Xue, G., and Yang, D. (2012). Smart Grid — The New and Improved

Power Grid: A Survey. IEEE Communications Surveys & Tutorials, 14 (4), pp. 944–

980.

Faruqui, A., Harris, D., and Hledik, R. (2010). Unlocking the €53 Billion Savings from

Smart Meters in the EU: How Increasing the Adoption of Dynamic Tariffs could Make

or Break the EU’s Smart Grid Investment. Energy Policy, 38 (10), pp. 6222–6231.

Federal Office for Information Security (BSI) (2013). Protection Profile for the Gateway of

a Smart Metering System (Smart Meter Gateway PP).



Markets: Forthcoming. Paper 192. Proceedings of the 21st European Conference on

Information Systems (ECIS).

Feuerriegel, S. and Neumann, D. (2014). Measuring the Financial Impact of Demand

Response for Electricity Retailers. Energy Policy, 65, pp. 359–368.

Feuerriegel, S., Strüker, J., and Neumann, D. (2012). Reducing Price Uncertainty Through

Demand Side Management. J.F. George, ed. Proceedings of 33rd

International

Conference on Information Systems (ICIS). Atlanta, USA: Association for Information

Systems.

Galli, S., Scaglione, A., and Wang, Z. (2011). For the Grid and Through the Grid: The

Role of Power Line Communications in the Smart Grid. Proceedings of the IEEE, 99

(6), pp. 998–1027.

Goebel, C., Jacobsen, H.-A., Razo, V., Doblander, C., Rivera, J., Ilg, J., Flath, C.,

Schmeck, H., Weinhardt, C., Pathmaperuma, D., Appelrath, H.-J., Sonnenschein, M.,

Lehnhoff, S., Kramer, O., Staake, T., Fleisch, E., Neumann, D., Strüker, J., Erek, K.,

Zarnekow, R., Ziekow, H., and Lässig, J. (2014). Energy Informatics. Business &

Information Systems Engineering, 6 (1), pp. 25–31.

Gottwalt, S., Ketter, W., Block, C., Collins, J., and Weinhardt, C. (2011). Demand Side

Management—A Simulation of Household Behavior under Variable Prices. Energy

Policy, 39 (12), pp. 8163–8174.

Gudi, N., Wang, L., and Devabhaktuni, V. (2012). A Demand Side Management Based

Simulation Platform Incorporating Heuristic Optimization for Management of

Household Appliances. International Journal of Electrical Power & Energy Systems, 43

(1), pp. 185–193.

Gungor, V.C., Sahin, D., Kocak, T., Ergüt, S., Buccella, C., Cecati, C., Hancke, G.P., and

Ergut, S. (2011). Smart Grid Technologies: Communication Technologies and

Standards. IEEE Transactions on Industrial Informatics, 7 (4), pp. 529–539.

Heinrich, B. and Zimmermann, S. (2012). Granularity Metrics for IT Services. J.F. George,

ed. Proceedings of 33rd

International Conference on Information Systems (ICIS).

Atlanta, USA: Association for Information Systems.

http://www.eex.com/en/Market%20Data


163

Ipakchi, A. and Albuyeh, F. (2009). Grid of the Future: Are We Ready to Transition to a

Smart Grid? IEEE Power and Energy Magazine, 7 (2), pp. 52–62.

Jagstaidt, U.C.C., Kossahl, J., and Kolbe, L.M. (2011). Smart Metering Information

Management. Business & Information Systems Engineering, 3 (5), pp. 323–326.

Judd, K.L. (1998). Numerical Methods in Economics. Cambridge, Mass: MIT Press.

Katz, J. (2011). Data - The Power Behind the Smart Grid. http://smartgrid.ieee.org/august-

2011/76-data-power-behind-the-smart-grid. Accessed 15 February 2014.

Khalifa, T., Naik, K., and Nayak, A. (2011). A Survey of Communication Protocols for

Automatic Meter Reading Applications. IEEE Communications Surveys & Tutorials,

13 (2), pp. 168–182.

Klobasa, M. (2007). Dynamische Simulation eines Lastmanagements und Integration von

Windenergie in ein Elektrizitätsnetz auf Landesebene unter regelungstechnischen und

Kostengesichtspunkten. Doctoral Thesis. Swiss Federal Institute of Technology (ETH),

Zurich.

Law, Y.W., Alpcan, T., Lee, V.C., Lo, A., Marusic, S., and Palaniswami, M. (2012).




Li, Z., Wang, Z., Tournier, J.-C., Peterson, W., Li, W., and Wang, Y. (2010). A Unified

Solution for Advanced Metering Infrastructure Integration with a Distribution

Management System. Proceedings of the 1st IEEE International Conference on Smart

Grid Communications, pp. 501–506.

Llaria, A., Jiménez, J., and Curea, O. (2013). Study on Communication Technologies for

the Optimal Operation of Smart Grids. Transactions on Emerging Telecommunications

Technologies, pp. n/a.

Lo, C.-H. and Ansari, N. (2011). The Progressive Smart Grid System from Both Power

and Communications Aspects. IEEE Communications Surveys & Tutorials.

Lopes, J.P., Hatziargyriou, N., Mutale, J., Djapic, P., and Jenkins, N. (2007). Integrating

Distributed Generation into Electric Power Systems: A Review of Drivers, Challenges

and Opportunities. Electric Power Systems Research, 77 (9), pp. 1189–1203.

Lujano-Rojas, J.M., Monteiro, C., Dufo-López, R., and Bernal-Agustín, J.L. (2012).

Optimum Residential Load Management Strategy for Real Time Pricing (RTP)

Demand Response Programs. Energy Policy, 45, pp. 671–679.

Malhotra, A., Melville, N.P., and Watson, R.T. (2013). Spurring Impactful Research on

Information Systems for Environmental Sustainability. MIS Quarterly, 37 (4), pp.

1265–1274.

McDaniel, P. and McLaughlin, S. (2009). Security and Privacy Challenges in the Smart

Grid. IEEE Security & Privacy Magazine, 7 (3), pp. 75–77.

McDonald, J. (2008). Leader or Follower: Developing the Smart Grid Business Case. The

Business Scene. IEEE Power and Energy Magazine, 6 (6), pp. 18–90.

McKenna, E., Richardson, I., and Thomson, M. (2012). Smart Meter Data: Balancing

Consumer Privacy Concerns with Legitimate Applications. Energy Policy, 41, pp. 807–

814.

McParland, C. (2012). OpenADR Open Source Toolkit: Developing Open Source

Software for the Smart Grid: Lawrence Berkeley National Laboratory.

Melville, N.P. (2010). Information Systems Innovation for Environmental Sustainability.

MIS Quarterly, 34 (1), pp. 1‐21.

http://smartgrid.ieee.org/august-2011/76-data-power-behind-the-smart-grid

http://smartgrid.ieee.org/august-2011/76-data-power-behind-the-smart-grid


164

Metke, A.R. and Ekl, R.L. (2010). Security Technology for Smart Grid Networks. IEEE

Transactions on Smart Grid, 1 (1), pp. 99–107.

Mind Commerce (2012). Telecom and Smartgrid Solutions & Market Opportunities:

Analyst Report.

Mohagheghi, S., Stoupis, J., Wang, Z., Li, Z., and Kazemzadeh, H. (2010). Demand

Response Architecture: Integration into the Distribution Management System.

Proceedings of the 1st IEEE International Conference on Smart Grid Communications,

pp. 501–506.

NERA Economic Consulting (2008). Cost Benefit Analysis of Smart Metering and Direct

Load Control: Overview Report for Consultation.

Palensky, P. and Dietrich, D. (2011). Demand Side Management: Demand Response,


Informatics, 7 (3), pp. 381–388.

Paulus, M. and Borggrefe, F. (2011). The Potential of Demand-Side Management in

Energy-Intensive Industries for Electricity Markets in Germany. Applied Energy, 88

(2), pp. 432–441.

PricewaterhouseCoopers (2010). Studie zur Analyse der Kosten-Nutzen einer

österreichweiten Einführung von Smart Metering.

Prüggler, N. (2013). Economic Potential of Demand Response at Household Level—Are

Central-European Market Conditions Sufficient? Energy Policy, 60, pp. 487–498.

Reynolds, T. and Mickoleit, A. (2011). ICT Application for the Smart Grid. OECD Digital

Economy Papers (190).

Ridder, F. de, Hommelberg, M., and Peeters, E. (2009). Four Potential Business Cases for

Demand Side Integration. Proceedings of the 6th International Conference on the

European Energy Market: IEEE, pp. 1–6.

Sauter, T. and Lobashov, M. (2011). End-to-End Communication Architecture for Smart

Grids. IEEE Transactions on Industrial Electronics, 58 (4), pp. 1218–1228.

Shaw, R., Attree, M., Jackson, T., and Kay, M. (2009). The Value of Reducing

Distribution Losses by Domestic Load-Shifting: A Network Perspective. Energy

Policy, 37 (8), pp. 3159–3167.

Simmhan, Y., et al. (2011). An Informatics Approach to Demand Response Optimization

in Smart Grids: Technical Report: University of Southern California.

Smart Energy Demand Coalition (SEDC) (2011). The Demand Response Snap Shot: The

Reality for Demand Response Providers Working in Europe Today. Report.

Statistisches Bundesamt (2012). Haushalte und Familien: Ergebnisse des Mikrozensus.

https://www.destatis.de/DE/Publikationen/Thematisch/Bevoelkerung/HaushalteMikroz

ensus/HaushalteFamilien2010300117004.pdf?__blob=publicationFile. Accessed 14

February 2014.

Strbac, G. (2008). Demand Side Management: Benefits and Challenges. Energy Policy, 36

(12), pp. 4419–4426.

Strüker, J. and Kerschbaum, F. (2012). From a Barrier to a Bridge: Data-Privacy in

Deregulated Smart Grids. Proceedings of the 33rd



Strüker, J. and van Dinther, C. (2012). Demand Response in Smart Grids: Research



https://www.destatis.de/DE/Publikationen/Thematisch/Bevoelkerung/HaushalteMikrozensus/HaushalteFamilien2010300117004.pdf?__blob=publicationFile

https://www.destatis.de/DE/Publikationen/Thematisch/Bevoelkerung/HaushalteMikrozensus/HaushalteFamilien2010300117004.pdf?__blob=publicationFile


165

Strüker, J., Weppner, H., and Bieser, G. (2011). Intermediaries for the Internet of Energy –

Exchanging Smart Meter Data as a Business Model. Paper 103. Proceedings of the 19th

European Conference on Information Systems (ECIS).

Styczynski, Z. (2011). VDE-Studie Demand Side Management: Presentation. Hannover,

Germany.

Sui, H., Sun, Y., and Lee, W.-J. (2011). A Demand Side Management Model Based on

Advanced Metering Infrastructure. Proceedings of the 4th International Conference on

Electric Utility Deregulation and Restructuring and Power Technologies (DRPT), pp.

1586–1589.





Uluski, B. (2011). Distribution Management Systems: Presentation. CRN Summit 2011.

van Horn, K. (2012). The Comparative Economic Assessment of the Impacts of Energy

Storage and Demand Response Resource Participation in the Day-Ahead Electricity

Markets. Master Thesis. University of Illinois at Urbana-Champaign.

Vasirani, M. and Ossowski, S. (2013). Smart Consumer Load Balancing: State of the Art

and an Empirical Evaluation in the Spanish Electricity Market. Artificial Intelligence

Review, 39 (1), pp. 81–95.

Watson, R.T., Boudreau, M.-C., and Chen, A.J. (2010). Information Systems and


the IS-Community. MIS Quarterly, 34 (1), pp. 23–36.

Ziekow, H., Strüker, J., and Goebel, C. (2012). The Potential of Smart Home Sensors in

Forecasting Household Electricity Demand. Working Paper: University of Freiburg.

166

C. Scenarios for Integrating Demand Response in Electricity

Markets

This chapter is a revised version of the paper

Philipp Bodenbenner, Stefan Feuerriegel, Dirk Neumann, 2014: Integrating Demand

Response Systems in Electricity Markets. 22nd

European Conference on Information

Systems (ECIS 2014), Tel Aviv, Israel, June 9-11, 2014 (conditionally accepted).

Abstract

As a consequence of an increasing share of renewable energies, balancing electricity

production and delivery requires efficient electricity markets. At the heart of the electricity

markets are information systems (IS) that coordinate demand and supply in real-time. IS

has recently opened up an alternative towards increasing the efficiency of electricity

markets by managing demand side resources; i.e. shifting electricity demand according to

fluctuating supply by so-called demand response. This part analyzes information systems

that integrate demand response in electricity markets with respect to both the associated

costs and benefits. Using real-world data from 2011, we compare profits of electricity

retailers across three different usage scenarios to determine that load shifting gives the

highest revenue: IS-related costs account for €2.58M exceeded by savings of €3.36M. To

improve demand response integration, our model suggests reducing bid sizes, delivery

periods and the time-lag between market transactions and delivery dates in electricity

markets.

Keywords

E-business, electronic markets, Green IT/IS, business value of IS/value of IS, information

systems, decision making/makers

167

1 Introduction

Designing and implementing efficient electricity markets is the key for reliable production

and delivery of electricity and, hence, satisfaction of the (ever-growing) demand. Due to its

physical characteristics, in particular transience, electrical power dictates tough

requirements on the related market design. The market environment is further complicated

by regulatory issues and the arising variety of stakeholders that are involved in the value

chain of electricity provisioning. Challenges range from the structure and timing of bidding

in wholesale markets to trading in ancillary services while, concurrently, assuring grid

stability. Addressing these issues by designing appropriate electronic markets for trading

of electrical power is a native mission of information systems (IS) research (Bakos, 1997;

Malone et al., 1987; Watson et al., 2010). With this part, we contribute to this mission by

evaluating the potential of IS and electronic markets in managing the demand side of

electricity and introducing demand side resources into various market settings.

One of the major challenges is to successfully unite electricity market design while

ensuring stability of electricity grids – as one focal concern refers to the system stability.

Electricity grids deliver energy from producers to consumers. Due to highly volatile supply

and demand, electricity grids may become unstable, when large deviations from the

desired power frequency occur. The maintenance of grid stability requires continuous

control of the power frequency. Grid operators (see Figure 42) have to immediately

counteract any imbalances by means of short-term control reserve. While grid operators

execute balancing activities in response to individual deviations in power frequency, the

emerging costs are distributed across the associated electricity retailers. Whenever

electricity retailers face unexpected deviations in demand or supply that might affect grid

stability within their control area, they request the so-called balancing energy which comes

at a varying penalty cost.

168

As a remedy to these costs, both balancing energy and control reserve can be (partially)

replaced by so-called demand response mechanisms. Essentially, demand response (DR)

systems manage the demand side of electricity by flexibly shifting power demand

according to the fluctuating supply side. It is defined by the U.S. Department of Energy

(2006) and the Federal Energy Regulatory Commission (FEDC) (2006) as: “Changes in

electric usage by end-use customers from their normal consumption patterns in response

to changes in the price of electricity over time, or to incentive payments designed to induce

lower electricity use at times of high wholesale market prices or when system reliability is

jeopardized.” Consequently, Madrigal and Porter (2012) recommend shifting electricity

load as a measure for grid stability instead of activating control reserve. Load shifting is

controlled and executed by electricity retailers (see Figure 42) that have access to a certain

DR potential at the customer end.

The effective usage of demand response requires real-time data and large scale information

networks in conjunction with sophisticated communication network structures. Having in

mind the intricate information flows and huge amount of data, demand response unveils to

be inherently daunting for IS research. Hence, the inevitable need for information systems

to match supply and demand in the power grid was stressed by Dedrick (2010). This

pushes the boundaries of IS research with its challenging requirements for information

Figure 42. Overview on Relevant Stakeholders in Electricity Retailer’s Environment Based on

Vasirani and Ossowski (2013); Vasirani and Ossowski (2012).

Electricity

Retailer

Provide Electricity

& Control Demand

Provide DR

Potential

Power

Supplier

Grid

Operator

Pronounce

Balancing Energy

Penalty

Communicate

Demand

Deviation

Optimize

Demand Load

Energy

Exchange

Control

Reserves

Offer Bid

Ask Bid Ask Bid

Ask Bid

Residential

Commercial

Industrial

Customers

Offer Bid

Stabilize Grid

169

processing (Corbett, 2011; Watson et al., 2010). All in all, IS research contributes to

managing DR mechanisms efficiently.

As a main contribution to IS research, this part discusses strategies to integrate demand

response into existing electricity markets from a market design perspective. In particular,

individual contributions are as follows:

We layout an overview of application scenarios for demand response in various

market settings, namely control reserve, balancing energy and electricity procurement

at the product market. For each scenario, we provide a rigorous derivation of financial

models to gauge both IS-related costs and corresponding saving potentials. Moreover,

we define and quantify the value of information in DR systems by setting up a

simulation environment using real-world data.

The results provide decision support in selecting DR implementation strategies for

electricity retailers. Based on our simulation, we deduce required changes in market

design that improve the integration of DR resources and increase market efficiency.

Further, we discuss opportunities for DR aggregators and outline corresponding

business cases.

The remainder of this part is structured as follows. In chapter 2, we review related work on

both the IS perspective of DR systems and their financial dimension. Subsequently, we

discuss strategies to integrate DR activities into existing electricity markets. For each of

them, chapter 3 models optimal decisions and derives costs as well as savings to gauge the

financial return. Finally, in chapter 4, we present the results by comparing DR activities

across different application scenarios and discuss policy implications to improve the

integration of demand response into electricity markets.

2 Related Work

A recent literature review (Strüker and van Dinther, 2012) shows that there is a small, but

growing number of IS-related research papers on smart grids. Examining these

170

publications reveals that there are many studies demonstrating the responsiveness of

residential, commercial and industrial customers to incentives and prices. However, we

came across only a few contributions dealing with demand response itself. In this chapter,

we survey related IS contributions such that recent advances towards demand response

systems are revisited. Based on this literature review, we conclude this chapter by deriving

our research questions and providing evidence that these questions are relevant and

important to the IS community.

2.1 Designing a Demand Response System

When it comes to designing a DR system, several publications have contributed to IS

research:

Corbett (2011) argues that a DR system will increase the information processing

requirements. As a consequence, the author demonstrates that DR systems will incur

massive amounts of data and, thus, a DR system poses an inherent IS problem.

Watson et al. (2010) develops the abstract idea of an energy informatics framework.

This framework represents an integrated approach that incorporates all of the elements

of an energy supply and demand system. However, its foundation remains at a high-

level view where the design of individual components is neglected. Several authors

advance towards an IS-architecture. Tan et al. (2012) design an actual decision

supporting system for demand side management, but the authors propose a high-level

structure only. Palensky and Dietrich (2011) construct a web-based energy

information system and name its typical components, while Law et al. (2012) focus on

tethering the end-consumer. Similarly, (Eto et al., 2007) surveys the necessary

communication infrastructure. However, none of these IS publications investigates the

market integration.

Feuerriegel et al. (2012) integrate demand response into the energy informatics

framework and list required components using a design science approach. However,

171

the authors ignore potential issues regarding market integration. Electric vehicles offer

the possibility to use their internal battery for load shifting purposes. Brandt et al.

(2012) and Wagner (2012) analyze potential usages from an IS perspective, but the

publications lack how electric vehicles are integrated from a market perspective.

While the above IS publications focus on the general design of a DR system, a thorough

analysis how demand response is integrated into electricity markets is missing. A

discussion on integration from a market design-related perspective is strongly needed.

2.2 Financial Dimension of Demand Response

Recent research has undertaken steps to investigate demand response from a financial

perspective. Besides the following advances, knowledge on cost and revenues from

demand response usage is still small.

Gottwalt et al. (2011) model DR decisions mathematically to estimate both costs and

revenues. The authors focus on pure load shifting among household and, thus, other

use cases are neglected.

Ahlert and van Dinther (2009) study a more specific setup where the economic

benefits are studied for energy storages at the consumer-end. Similarly, Brandt et al.

(2012) and Wagner (2012) focus on electric vehicles and investigate trading strategies

at the market for control reserve. However, both approaches can be hardly transferred

to measure the cost-benefit-ratio of demand response activities. Shayesteh et al.

(2010) analyze the effects of an incentivized program to participate in reserve markets

by formulating an optimization problem. However, the model is not calibrated by

real-world data and the financial impacts are not evaluated.

Paulus and Borggrefe (2011) perform a cost-benefit-study to compare different

energy-intensive industries in Germany for their ability to use their DR potential at a

reserve energy market. However, the authors do not consider load shifting and, thus,

172

cannot compare which application for demand response achieves the highest financial

benefit.

Since the above research papers concentrate on partial examinations of either DR systems

or their cost-benefit-ratio, many questions are left unanswered. All listed publications lack

(1) a financial comparison of DR programs across different use cases, and (2) an in-depth

evaluation of market integration. Consequently, an analysis of market integration

originating from electronic markets theory still seems to be an open research question.

Therefore, we pursue a rigorous IS approach: First, we derive mathematical models for

optimal DR decision across different use cases. Based on the use cases, we provide

decision support for DR activations. Then, we not only investigate schemes for market

integration, but also give insights into possible issues where the energy market design

requires adjustments.

3 Integrating Demand Response into Electricity Markets

Generally, electricity markets can be divided into three categories, namely a product

market, a control reserve exchange and balancing energy (cp. Figure 43). This is a valid

classification (Kirby, 2004) in almost all developed countries. In the following sections, we

revisit the landscape of existing electricity markets in detail and elaborate on how to

integrate DR resources.

Figure 43. Overview on Electricity Market Structure in Germany (Based on Liebau (2012)):

Market Components Marked in Dark Grey are Assessed Regarding Demand Response

Integration.

Derivatives and Futures

Day-

ahead

Market

Intra-day

Market

-1 day 0-45 minutes +1 hour +1 day

Day-after

Market

Balancing Energy &

Imbalance Penalties

Gate

ClosureDelivery

Delivery of

Reserve

Energy

Tender for

Primary &

Secondary

Reserve

Tender for

Tertiary

Reserve

-1 week

Product Market

Control Reserve Exchange

t

173

3.1 Electricity Procurement at the Product Market

At the product market, electricity can be procured on basis of standardized contracts as

well as over-the-counter deals. At the contract market, both derivatives, in form of futures

and options, and spot transactions (day-ahead and intraday) can be done. Here, the

electricity retailer can use DR mechanisms to optimize the balance of production and

demand, as well as to and prevent procuring electricity during peak times. Hence, demand

response does not serve as market product itself, but as internal tool for the electricity

retailer.

3.2 Market Design of Control Reserve Exchanges

The operator of an electricity grid requires reserve energy to ensure reliable grid

operations. Exemplary, the controlling system frequency should be maintained to be

around 50 Hz (Riedel and Weigt, 2007). In case of major load fluctuations, such as a

power station outage, reserve energy is activated. Reserve energy appears as both positive

and negative reserves: positive reserves are used to offset lack of energy, whereas negative

reserves withdraw power from the network and, thus, address energy excesses. Here, we

distinguish between three types of reserve energy according to activation time and duration

of activation: primary, secondary and tertiary reserve. Primary reserve energy is activated

immediately to stabilize the system and balance a short power deficit respectively surplus

(up to 30 seconds). Secondary reserve is provided by gas turbines and other fast reacting

power stations. In case the power flow disturbance continues for an extended period,

secondary reserve hands over to the tertiary reserve. The defined period before activation

of the tertiary reserve is usually set to 30 minutes. The tertiary reserve has to be fully

available within 15 minutes after activation. Today, tertiary reserve is used only in approx.

3 % of the cases (Riedel and Weigt, 2007). In fact, the demand for tertiary reserve

increases strongly (Paulus and Borggrefe, 2011).

Here, we focus on the marketplace for trading tertiary reserve energy as it provides the

opportunity for integrating DR potentials. Demand response holds the potential to be

174

activated and provided within the required time frame of 15 minutes. By shifting load, the

retailer could generate both positive and negative reserve potentials that can be offered as

resources in the marketplace. In contrast, the requirements for secondary and primary

reserve markets are too restrictive for DR processes (Paulus and Borggrefe, 2011). With

liberalization of the energy markets, markets for trading control reserve have been

established in many European countries (e.g. Denmark, Switzerland and Germany) and

numerous regions in the US (e.g. New England, Texas and California) (Kirby, 2004). In

Germany, the system operators procure reserve energy in a control power market where

bidding is done through an Internet-based marketplace (Regelleistung, 2012). Primary and

second reserve energy is procured in a monthly cycle.

Tertiary reserve is tendered in combinatorial reverse auctions, whereby positive and

negative reserve is tendered separately in time slices of four hours that divide the day into

six even intervals (Riedel and Weigt, 2007). The auctions that take place each day settle

the tertiary reserves for the following day. The auction bids consist of three values, namely

the amount, the capacity price and the working price. As the amount of required reserve

power cannot be determined precisely in advance, each tenderer specifies the amount of

(positive resp. negative) reserve energy capacity that the tender is able to provide for a

specific time slot. Moreover, the tenderer communicates the capacity price (per MW) for

provisioning the offered capacity. The capacity price has the character of an option fee.

Finally, the working price (per MWh) defines the tenderer’s desired price for actual use of

the capacity. The working prices strongly exceed the prices that are achieved at the regular

day-ahead spot markets for electricity.

In future grid operations, a major increase in control reserve capacities is required

(Madrigal and Porter, 2012; Paulus and Borggrefe, 2011). While today’s electricity

markets already bear control reserve as a major cost driver, the situation will further

worsen in future, as demand and the related costs are assumed to rise significantly. This

development is particularly triggered by the growing share of renewable energy supply

175

and, as the European Union has set its target share to 20 %, renewables will evoke severe

consequences. First, “unpredictability leads to greater fluctuation in the reserve

requirements” (Jacobsen and Zvingilaite, 2010) and, in addition to that, “extra system

balancing reserves will be needed” (Gross et al., 2006). Second, Kladnik et al. (2012)

concludes that prices will increase sharply.

3.3 Definition of Balancing Energy and Imbalance Penalties

As described in the previous section, the retailer has to forecast the energy demand for the

next day. In case of deviations from the forecast, control reserves are activated to offset the

imbalances. The grid operator acquires the control reserves at the corresponding exchange.

Here, balancing energy comes into play. It serves as metric for retrospectively allocating

the costs of actually activated control reserves to the respective originator of the

imbalance. In an ex-post view, the deviations in a given time and control area are

accumulated and analyzed. Based on the ex-post analysis of expenditures for grid

stabilization, a price for balancing energy is determined. This price resp. imbalance penalty

has to be paid by the originators of the imbalance.

Similar to control reserves, balancing energy is a major cost driver in electricity markets.

Again “the presence of wind power in a system increases balancing power requirements,

due to its variability and limited predictability” (Vandezande et al., 2010). Hence, DR

mechanisms could provide benefits to the retailer by reducing expenditures for balancing

energy. Whenever the retailer perceives that an imminent deviation of the actual demand

from his projected forecast may occur, the retailer could shift load to prevent imbalances in

the grid. Furthermore, Vandezande et al. (2010) suppose that the current mechanism for

determination and allocation of the balancing energy price is not sufficient, in particular

when it comes to growing integration of renewable energy sources into the grid. They

propose an enhanced market design for balancing energy to ensure efficient integration of

wind power.

176

4 Research Model: Measuring the Financial Impact of DR

This section presents our research scenarios and their mathematical models for both costs

and savings, in order to gauge financial impacts of DR activities. Costs are determined by

the DR system and its components. As DR usage can only be realized with the help of a

comprehensive IS and communications system, we inspect costs of individual system

components as well as communication. In contrast to that, savings arise from shifting load

efficiently. To estimate savings, we follow the frequently used approach (Doostizadeh and

Ghasemi, 2012; Meng and Zeng, 2012) that assumes single electricity retailer to measure

its savings.

4.1 Research Questions

Recently, Strüker and van Dinther (2012) asked “how large is the economic value of

demand response”. Likewise, Strbac (2008) claims that there is a “lack of understanding

of the benefits of demand side management solutions”. To quantify the benefits, our

research model addresses three scenarios, in which electricity retailers can deploy demand

response to optimize their revenue streams. Each scenario is derived from the retailer’s

tasks concerning maintenance of the flow network and the current electricity market

design. The scenarios are defined as follows:

Scenario A: Trade DR potential at the exchange for reserve energy. In our first

scenario, the retailer uses DR potential as a trading good on the exchange for reserve

energy. Here, the retailer can leverage the available DR potential in his distribution

network to offer it both as positive as well as negative reserve energy on the market.

Scenario B: Use DR to avoid Balancing Energy. Electricity retailers have to

forecast their energy demand one day in advance. In case this forecast deviates from

their actual demand, retailers have to pay a penalty. In our first scenario, the

electricity retailer employs demand response to avoid penalties originating from using

177

balancing energy by shifting load whenever the actual demand deviates from the

forecasted value.

Scenario C: Use load shifting to optimize electricity procurement. In our third

scenario, the retailer employs demand response to improve his procurement strategies

at the regular energy exchanges (i.e. day-ahead auctions, futures, derivatives). Here,

the retailer shifts load to non-peak times to purchase at lower prices.

Based on the financial evaluation and comparison of the three usage scenarios for demand

response, we want to derive managerial recommendations concerning the implementation

strategies for electricity retailers. This leads to our first research question.

Research Question 1: Which usage scenario of demand response provides the highest

benefit to electricity retailers?

The implementation of DR systems is closely connected to substantial capital

expenditures. Hence, it appears necessary to either extend the scope of usage beyond

demand response or leverage synergy effects from operating infrastructures for multiple

retailers. Demand response requires collecting and processing massive loads of

information. This may provide ways for specialized players, such as telecommunications

providers or big data specialists, to establish novel, profitable business cases.

Research Question 2: What are business opportunities for information aggregators in the

field of demand response?

Moreover, we aim at going beyond a purely financial view. In addition to determination of

managerial implications and financial benefits for electricity retailers, this IS research

work seeks for providing insights into necessary changes in electricity market design.

Leveraging the potentials of adjusting the demand side requires changes in the electricity

market design. We address this issue with the following research question.

Research Question 3: What are the policy implications for electricity market design?

178

4.2 Cost Structure of Demand Response Systems

The DR system’s architecture, which is identical for all considered scenarios, is illustrated

in Figure 44.

Core component of the DR infrastructure is the central distribution management system,

which provides the retailer with functionality for monitoring and controlling the

distribution flow network. The load forecasting engine and the DR engine are important

when considering DR programs. They are responsible for detecting mismatches in demand

and supply of electricity in a future timeslot, and for triggering measures accordingly.

Moreover, the sensor network represents the currently ongoing technology trend towards

an internet of things. It provides the retailer with a two-way communication channel

towards his customer resp. the connected devices. This type of broadband communication

only enables the full usage of demand response. There is a great variety of communication

channels. We restrict our considerations to the most popular ones (Gungor et al., 2011),

namely powerline communications (PLC) and mobile communications (GSM and

successors). The smart meters form the interface between the retailer’s infrastructure and

the customer’s home area network (HAN). All energy-consuming devices in a household

resp. at a commercial customer’s premise should be connected to the HAN to allow for

remote control.

Figure 44. System View on a Demand Response System with an Advanced Metering

Infrastructure Based on Bodenbenner et al. (2013)

Weather forecasts Electricity prices Usage data DR contingenciesElectricity demand DR control signals

Information System (IS)

Distribution Mgmt.

System (DMS)

Sensitized Objects

(Home Area Network)

Sensor Networks

(Advanced Metering Infrastructure )

External Stakeholders

Flow Networks

Energy

Mgmt.

System

Generation

Info. &

Trading

System

Weather

Forecast

Provider

Energy

Markets

Distributed

Energy Resource

Load Mgmt. &

Control System

(LMCS)

Consuming

Mgmt. System

Network

Management

System (NMS)

Smart

Meter

Meter Data

Mgmt. System

(MDMS)

Load Control

Device

PLC Con-

centrator

C

E

F

A

D

A B C ED FLegend:

B

Demand Response

Engine

Load Forecasting

Engine

179

The system architecture and information flows are the basis for the derivation of the

following cost components. As the infrastructure in the three scenarios does not differ, the

capital expenditures and operating costs are the same in all scenarios. The capital

expenditures comprise the initial costs of procuring and installing the DR system. In

addition, annual costs 𝑐OP are required to operate the system infrastructure. The costs

cover e.g. the maintenance, personnel, energy, etc. for all components of the DR system.

In addition to the operating costs, the execution of a DR program induces volume-based

communication costs. The volume-based costs are limited to traffic that passes through

GSM-based channels. PLC-based communication does not generate any volume-based

charges.

As shown in Figure 45, the information flow in DR programs consists of two parts: first,

reading out the usage data from the smart meter and, second, broadcasting the DR control

signals.

Looking at the first phase, the usage data needs to be collected from the smart meter and

transferred to the central collection servers. The annual volume for a single smart meter for

this upstream communication is defined as

𝑣UD = ( 𝑔UD 𝑔T⁄⏟ No.of read−out

events per transfer

∙ 𝛼UD⏟Message size

+ 𝛼OH⏟Overhead

size

) ∙ 𝑔T⏟No.of transfers

per day

∙ 365⏟Days p.a.

.

(35)

The second phase, namely broadcasting the DR signal to the connected smart meters,

exhibits a similar cost structure as the first phase. Here, 𝛼DR denotes the accumulated

Figure 45. Information Flow in a Demand Response System

Electricity

Retailer

Consumer

External

Service

Providers

Weather

Forecast

Electricity

Prices

Usage Data

Records

Sensor

Data

DR

Conting-

encies

DR Control

Signals

C

BA

Load Shift.

Scheme

Electricity

Demand

D

Phase 1: Forecasting Electricity Load for

Next Time Interval

Phase 2: Determination & Execution of

Load Shifting Scheme

E F

180

message size for the entire protocol stack, which comprehends three steps: initializing the

DR session, retrieving the available DR potentials from the customer, and, finally, sending

the DR control signals towards the customer. Variable 𝑔DR denotes the number of

communication activities that are induced by demand response per day. The annual volume

for the downstream communication per smart meter is defined as

𝑣DR =

(

𝑔DR 𝑔T⁄⏟ No.of DR signalsper transfer

∙ 𝛼DR⏟Message size

+ 𝛼OH⏟Overhead

size)

∙ ( 𝑔T⏟

No.of transfersper day

∙ 365⏟Days p.a.

+ 𝑔A⏟No.of extra transfers p.a.

).

(36)

The requirements concerning the frequency of reading out the smart meters and sending

DR signals differ for the various scenarios. Consequently, the parameters 𝑔UD (i.e. number

of read-out events per day) and 𝑔T have to be individually assigned for each scenario (cp.

Table 46).

Interval of reading

out usage data

records (𝑔𝑈𝐷)

Interval of

defining DR

control signals

(𝑔𝐷𝑅)

Interval of data

transfers (𝑔𝑇)

Additional readout

& transfer

activities (𝑔𝐴)

Scenario A:

Reserve Energy

15 minutes 15 minutes 24 hours depending on

number of reserve

energy activations

Scenario B:

Balancing

Energy

15 minutes 15 minutes 15 minutes 0

Scenario C:

Load Shifting

60 minutes 60 minutes 24 hours 0

In scenario A, DR potentials are offered at the reserve exchange; the notification of

available resources has to take place on the previous day. Corresponding to the required

activation time, the read-out interval is set to 15 minutes. The data has to be transferred

from the smart meter to the retailer once a day. For balancing energy (i.e. scenario B) the

read-out and transfer intervals are set to 15 minutes. This allows the retailer to

continuously check the actual demand against the projected forecast. If it becomes

apparent that imbalances may occur, the retailer can take appropriate action and shift load

to prevent penalties. For scenario C, we take a look at the current product market. Here,

Table 46. Cost Model Configurations Across Scenarios

181

electricity is traded in blocks of 60 minutes. Hence, we adopt the interval to 60 minutes

accordingly. As the day-ahead market is cleared one day prior to delivery, a daily transfer

of the usage data appears sufficient. In summary, the communication costs add up to

𝑐COM = 𝑥GSM⏟

No. of GSM meters

∙ (𝑣UD ∙ 𝑐GSM(𝑣UD)⏟ Comm. cost GSM (per MByte)

+ 𝑣DR ∙ 𝑐GSM(𝑣DR)). (37)

4.3 Measuring the Revenue Potentials from Demand Response

Systems

While the previous section assessed the (communication) cost perspective of demand

response, we now put the attention towards potential savings. Therefore, we derive a

mathematical model to optimize DR decisions. Now, let us assume that we will be granted

for each time interval a certain DR potential. This potential can now be shifted forward or

backward in time. Here, we distinguish DR potential according to the maximum duration 𝑗

(e.g. one, two, etc. hours) that it can be displaced. For each of these shifts 𝑗, the variables

𝛥𝑗(𝑡) denote the available potential. When we shift DR potential between two hours 𝑡 and

𝑡′, this is indicated, with 𝑗 being the maximum length of the shift, by the parameter

𝐷𝑅𝑗(𝑡, 𝑡′ − 𝑡). The value of 𝐷𝑅𝑗(𝑡, 𝑡

′ − 𝑡) denotes the amount of power that is consumed

less at time step 𝑡, but that is additionally required at time step 𝑡′. Knowing the amount of

granted DR potential, we can continue to model decisions among both balancing energy

usage and reserve energy.

Scenario A: Modeling Trading of DR Potential as Tertiary Control Reserve

Electricity retailers can use their demand response potential as a replacement for tertiary

control reserve. Negative control reserve, if requested, implies that demand must be shifted

away. Let 𝑉neg(𝑡) denote the energy level that needs to be decreased. As regulatory issues

force electricity demand to be fixed day-ahead, this additional volume 𝑉neg(𝑡) can only be

shifted to the following day. The optimal time slots are those that give the highest revenue

and result from the following optimization problem. Let the price for electricity at the day-

182

ahead market be given by 𝑝A(𝑡) and the amount we use by 𝑞A(𝑡). Then, we can split our

volume 𝑉neg(𝑡) across the quantities 𝑞A(𝑡) of the next day. Here, the electricity retailer

maximizes its earnings for the next day and the target function is given by

𝑟neg = max𝑞A(1),…,𝑞A(𝑁)∑ 𝑝A(𝑡) 𝑞A(𝑡)𝑁𝑡=1 . (38)

The necessary constraints are as follows. According to equation (39) the volume of

positive reserve energy must equal the additional energy moved here. The second

constraint specifies the quantities 𝑞A(𝑡) from the day-ahead market. The last constraint

guarantees the conservation of used load shifting potential such that all demand that is

shifted away is finally added somewhere else.

𝑉neg(𝑡) =∑𝐷𝑅𝑗(𝑡, 0)

𝑗

,

𝑞A(𝜏) =∑𝐷𝑅𝑗(𝑡, 𝜏 − 𝑡)

𝑗

,

−𝐷𝑅𝑗(𝑡, 0) + ∑ 𝐷𝑅𝑗(𝑡, +𝑖)+𝑗𝑖=𝑁−𝑡+1 = 0 ∀𝑗.

(39)

Next, we derive constraints on the DR potential. Recall that the variables 𝛥1(𝑡), 𝛥2(𝑡),

𝛥3(𝑡), etc. limit the maximum amount of energy that can be displaced. Thus, we constrain

the maximum DR potential that is shifted away by

𝐷𝑅1(𝑡, 0) ≤ 𝛥1(𝑡), 𝐷𝑅2(𝑡, 0) ≤ 𝛥2(𝑡), etc. (40)

Additionally, we need constraints to guarantee that demand is moved solely away. When

we shift energy from time interval 𝑡 by up to 𝑗 hours, this value cannot be negative. Thus,

𝐷𝑅𝑗(𝑡, +𝑖) ≥ 0 for all 𝑗 and 𝑖 = 1,… , 𝑗. (41)

Positive control reserve, if requested, implies that further energy is needed and demand

response potential must be shifted here. We skip neglect a formal derivation of this model

as models for both positive and negative control reserve are almost identical. As all of the

above constraints are either equality constraints or bounds. Optimization problems both are

linear and can be solved using e.g. a simplex algorithm as a unique solution exists.

183

Scenario B: Modeling DR Usage for Avoiding Balancing Energy

Electricity retailers must pay a penalty whenever their actual electricity demand differs

from the expected one. Let 𝐵𝐸(𝑡) denote the gap in watt-hour at time 𝑡. This imbalance

can be either positive or negative; however, the penalty costs rise linearly with the

deviation |𝐵𝐸(𝑡)|. At the same time, each time slot 𝑡 is associated with a penalty price

𝑝BE(𝑡). Similar to Vasirani and Ossowski (2013), Vasirani and Ossowski (2012), an

electricity retailer pays during 𝑁 time slots costs of

𝑐BE = ∑ 𝑝BE(𝑡) |𝐵𝐸(𝑡)|𝑁𝑡=1 . (42)

To reduce these expenditures 𝑐BE for balancing energy, the retailer can harness demand

response as a remedy. Now, we introduce the notation for DR usage and we derive the

corresponding optimization problem. Its target function 𝑐BE,DR minimizes the retailer’s

expenditures and, thus, sums the penalty costs over a time horizon. Let 𝑞BE(𝑡) denote the

amount of balancing energy that must still be requested after DR usage. Then,

𝑐BE,DR = min𝑞BE(1),… ,𝑞BE(𝑁)∑ 𝑝BE(𝑡) |𝑞BE(𝑡)|𝑁𝑡=1 . (43)

In addition to that, we introduce a set of constraints:

Demand equality. With 𝐵𝐸(𝑡) indicating the demand of balancing energy, this

energy must be replaced by the sum of DR activities and the residual balancing

energy, i.e.

𝐵𝐸(𝑡) = 𝑞BE(𝑡) − 𝐷𝑅1(𝑡, 0) − 𝐷𝑅2(𝑡, 0) + ⋯ + 𝐷𝑅1(𝑡 + 1,−1) + 𝐷𝑅1(𝑡 − 1,+1) + 𝐷𝑅2(𝑡 + 2,−2) + 𝐷𝑅2(𝑡 + 1,−1) + 𝐷𝑅2(𝑡 −1,+1) + 𝐷𝑅2(𝑡 − 2,+2) +⋯.

(44)

Conservation of DR potential. We need to guarantee the conservation of used DR

potential such that all demand that is shifted away is finally added somewhere else

∑ 𝐷𝑅𝑗(𝑡 + 𝑖, −𝑖)+𝑗𝑖=−𝑗 = 0 for all 𝑗. (45)

184

Bounds. We derive additional constraints on the potential of demand response. Recall

that the variables 𝛥1(𝑡), 𝛥2(𝑡), 𝛥3(𝑡), etc. limit the maximum amount of energy that

can be displaced. Therefore, we deduce

𝐷𝑅1(𝑡, 0) ≤ 𝛥1(𝑡), 𝐷𝑅2(𝑡, 0) ≤ 𝛥2(𝑡), etc. (46)

Additionally, we need constraints that limit the flow direction (i.e. that demand is moved

solely away from time interval 𝑡). Therefore, when we shift energy from time interval 𝑡 by

𝑗 hours in either direction, this value cannot be negative,

𝐷𝑅𝑗(𝑡 + 𝑖, − 𝑖), 𝐷𝑅𝑗(𝑡 − 𝑖, +𝑖) ≥ 0 for all 𝑗 and 𝑖 = 1,… , 𝑗. (47)

The above optimization can be easily turned into a linear problem. Depending on the sign

of 𝐵𝐸(𝑡), we use a different price that always yields the desired outcome. Thus, we replace

the price 𝑝BE(𝑡) for balancing energy by a new price 𝑝BE′ (𝑡) which is given by

𝑝BE′ (𝑡) = {

−𝑝BE(𝑡), 𝐵𝐸(𝑡) < 0,𝑝BE(𝑡), otherwise

. (48)

As all of the above constraints are either equality constraints or bounds. The optimization

problem is linear, can be solved using e.g. a simplex algorithm and a unique solution

exists.

Scenario C: Use Load Shifting to Optimize Electricity Procurement

For scenario C, we revert to the linear optimization model of Feuerriegel et al. (2012) that

minimizes the retailer’s accumulated expenditures. The model supports two common

energy derivatives, namely future options and day-ahead auctions.

5 Evaluation

In the following section, we test our mathematical model in a setting using real-world data.

The gained results are used for evaluating the above research questions.

185

5.1 Datasets

For our evaluation setting, we assume a fictitious German retailer delivering electricity to

290,000 residents. The retailer’s overall annual energy demand accounts for 2,000 GWh

(E-Control, 2012). The distribution of retailer’s customers is as follows (Styczynski,

2011): 25 % households, 25 % commercial customers and 50 % industrial customers.

Industrial customers are excluded from calculation of the DR saving potentials as

industrial customers hardly participate in load shifting, but rather reduce their energy

consumption when granted financial incentives. All prices for energy derivatives and spot

auctions are based on the historic hourly data of the European Energy Exchange, EEX for

short (European Energy Exchange, 2012). We use the amount of balancing energy

provided by E.ON Mitte13

for the year 2011. The original values account for 1.5 million

inhabitants, and we scale it down to 290,000. The penalty price of balancing energy is

provided by Transnet BW14

for the year 2011. All volumes and prices for tertiary control

reserve are published on an Internet platform (Regelleistung, 2012).

The capabilities of demand response vary strongly among both industry and households.

Klobasa (2007) analyze the market penetration of demand side management and its overall

potential for Germany – see Table 47.

Commercial

Customers

Max. Shift Duration/h 1 2 12 16

Average Power Shift/kW 16 164 8850 4650 5129

Residential

Households

Max. Shift Duration/h 1 2 12 24

Average Power Shift/kW 7353 6750 35 068 3616

Table 47. Demand Response Potential through Load Shifting Scaled for Retailer

(Klobasa 2007).

Across the time of day, the DR potential in households varies strongly and, hence, the

household values in Table 47 are weighted by time-dependent coefficients (Groiß, 2008).

To utilize this potential, we assume a DR system characterized by the following

13 E.ON Mitte AG (2013). Differenzbilanzierung. Web: http://www.eon-mitte.com/de/netz/veroeffentlichungen/strom_/ver

oeffentlichungen_nach_12_abs_3_stromnzv. Accessed 5 April 2013. 14

Transnet BW GmbH (2013). Bilanzkreisabrechnung. Web: http://www.transnetbw.de/strommarkt/

bilanzkreismanagement-und-bilanzkoordination/bilanzkreisabrechnung/. Accessed 5 April 2013.

http://www.eon-mitte.com/de/netz/veroeffentlichungen/strom_/veroeffentlichungen_nach_12_abs_3_stromnzv

http://www.eon-mitte.com/de/netz/veroeffentlichungen/strom_/veroeffentlichungen_nach_12_abs_3_stromnzv

http://www.transnetbw.de/strommarkt/bilanzkreismanagement-und-bilanzkoordination/bilanzkreisabrechnung/

http://www.transnetbw.de/strommarkt/bilanzkreismanagement-und-bilanzkoordination/bilanzkreisabrechnung/

186

parameters. The volume-based costs for GSM communication are assumed to follow a

logarithmic cost function; a higher data volume per meter leads to a cheaper price per

transferred megabyte. The hardware costs for a single, plain smart meter are defined as

€ 60. A GSM resp. PLC communication module implies additional costs of € 35 and € 20.

The overall number of installed smart meters is computed as 220,000 based on an assumed

annual energy consumption of 3,500 kWh per residential household and 6,500 kWh per

commercial customer. Based on an assumed mean of meters per residential building resp.

commercial customer, the number of communication modules is computed as 130,000.

Thereof a share of 15 % communicates via GSM networks; the remaining 85 % are PLC-

based modules. All meters are rolled out and put into operation at the same time.

5.2 Results and Managerial Implications

The results section begins with an introduction of a metric for measuring the value of

information in DR applications. Based on this metric, we subsequently present the

financial outcomes for each of the defined scenarios and the given dataset. This section

concludes with a comparison of the scenarios and respective managerial recommendations.

Information Value in a Demand Response System

In the smart grid realm, the usage data records that are regularly collected from the smart

meters can be considered the core information piece. The collected data records enable

numerous use cases, such as optimization and automation of the billing process. Here, we

focus on the value that can be tapped by optimization of the electricity demand side with

the help of DR programs. The additional revenue potentials from trading the DR resources

at the energy markets resp. saving potentials when avoiding penalties are opposed by

significant costs for setting up and operating the required infrastructure. Combining these

two perspectives, we introduce an expedient metric: the information value of demand

response. The information value specifies the surplus a retailer is able to generate per

single usage data record in a given scenario. Let 𝑇 denote the total number of smart meter

187

readouts per year. Determination of the revenue varies across the scenario. For scenario A,

let the revenue potential 𝑟 be defined as 𝑟 = 𝑐BE − 𝑐BE,DR. Here, 𝑟 resembles the

difference between, on the one hand, paying the penalties and, on the other hand,

optimizing the situation with the help of demand response. In scenario B, for calculation of

the retailer’s revenue potential the earnings from trading positive and negative reserve

energy have to be added up. Thus, the revenue potential is calculated as 𝑟 = 𝑐pos + 𝑐neg.

Taking all together, this leads to the following definition of the information value.

Definition. The information value for demand response expresses an electricity retailer’s

profit per usage data record that is collected from a smart meter. Based on the annual

revenue 𝑟, related annual costs 𝑐OP and 𝑐COM and the number of readout events per year 𝑇

the information value is calculated as

𝐼𝑉DR =

𝑟 − (𝑐OP + 𝑐COM)

𝑇.

(49)

Scenario A: Trading DR Potential at the Exchange for Reserve Energy

In the presented setting, reserve energy has to be activated 6,895 times within the

observation period (i.e. one year). Out of these, 5,612 activations attribute to negative

reserve energy and the remaining 1,283 activations to positive reserve energy.

Accordingly, 6,895 communication activities are initialized leading to extra annual

communication costs of € 313,000. In addition, the infrastructure generates operating costs

and further communication costs (€ 0.890 M) for daily transfer of usage data records.

Within the 6,895 activations of reserve energy, the retailer is enabled to trade a mean

volume of 0.0849 MW of negative and 0.0332 MW of positive reserve energy per

activation (cp. Table 48). Whereas the mean volume of traded energy is substantially

higher for positive reserve energy, for the actually served volume of energy that of

negative reserve energy exceeds the volume of positive reserve energy. The revenue

potentials from traded and served reserve energy accumulate to € 125,000 per year. This

corresponds to an average earning of € 18 per activation. In summary, the total annual

188

running costs of the DR system significantly exceed the additional revenue potentials,

which the retailer could realize by offering DR potential at the reserve energy market. The

revenues cannot even cover the expenses for the additional communication activities that

are caused by the reserve energy activations; in our setting a loss of € 4,000 is generated.

Negative Reserve

Energy

Positive Reserve

Energy

Total

Max traded volume 0.3598 MW 0.1877 MW –

Earnings € 92,930 € 31,800 € 124,720

Avg. earnings per activation € 16.56 € 24.78 € 18.09

Table 48. Overview on Trading Volumes and Financial Impact of Trading DR Potential at

the Exchange for Reserve Energy for Given Evaluation Setup.

Scenario B: Use Demand Response to Avoid Balancing Energy

Scenario B aims at avoiding penalty costs associated with requested balancing energy. Our

fictions retailer faces total costs of around € 798,000. These costs account for all deviations

in forecasted and actual electricity demand occurring the year 2011. As shown in Table 49,

demand response shows a possible path how these costs can be decreased significantly. By

using DR programs, penalties reduce, as almost all balancing energy can be avoided, down

to staggering € 1,700.

Without DR Using DR Relative

Change

Difference

Total expenditures € 798.4 k € 1.7 k –99.8 % € –796.7 k

Required balancing energy 95.28 MWh 1.92 MWh –98.0 % – 93.37 MWh

Table 49. Comparison of Expenditures for Penalties for Balancing Energy.

Scenario B requires a constant information exchange (i.e. every 15 minutes) between the

retailer and the consumer (cp. Table 46). Compared to scenario C, this means a lot more

communication activities. This again, leads to a much higher communication overhead that

is generated. The communication costs add up to € 0.801 M per annum. In addition, the

DR infrastructure generates annual operating costs in the amount of € 2.540 M. Taken

together, this results in running costs of € 3.341 M.

189

This cost pool strongly exceeds the expected revenue potentials for scenario B. The DR

infrastructure produces an annual loss of € 2.544 M. Here, the huge initial investments (i.e.

€ 24.419 M) have not even been considered. However, the additional revenue potentials of

using DR instead of balancing energy outweigh the costs that are directly related to

communication. Hence, setting up the DR infrastructure and employing it for multiple use

cases, including scenario B, could be a viable approach.

Scenario C: Use Load Shifting to Optimize Electricity Procurement

For scenario C, we fall back on results from our previous work (Feuerriegel et al. 2013).

Here, the retailer’s saving potential is computed as € 3.360 M (cp. Table 50). In contrast,

the DR system generates annual running costs of € 3.117 M. This leads to a slightly

positive information value of € 0.05 per 1,000 smart meter readouts.

Summary of Results

Combining the prior evaluation results (cp. Table 50), a clear answer concerning research

question 1 can be formulated: scenario C, i.e. using load shifts to optimize electricity

procurement, provides the largest benefit for the electricity retailer. It is the only scenario

that generates a surplus at all; the two other scenarios are not profitable and, thus, exhibit a

negative information value. However, it should be noted that also for scenario C the

situation changes as soon as the required capital expenditures are taken into account as

well. Then the revenue potentials implied by the DR system can hardly cover the overall

costs. The capital expenditures in our defined scenarios amount to more than € 24 M each.

Assuming a constant annual profit of € 776,000, as in scenario C, the infrastructure has to

be operated (without afresh investments) for more than 30 years. This is more than double

the value of 15 years, which is frequently assumed as realistic lifetime for such an

Information System (PricewaterhouseCoopers, 2010). However, even if the DR system is

not profitable today, the assumed cost increases for control reserves and balancing energy

could lead to a positive financial case in the medium term.

190

Scenario A:

Reserve

Energy

Scenario B:

Balancing

Energy

Scenario C:

Load Shifting

Revenue potentials € 0.124 M € 0.797 M € 3.360 M

Annual running costs € 3.430 M € 3.341 M € 2.584 M

Capital expenditures € 24.419 M € 24.419 M € 24.012 M

Information value per 1000 meter readouts € –0.40 € –0.26 € 0.40

Annual GSM data volume per smart meter 5.61 megabyte 4.81 megabyte 0.77 megabyte

Table 50. Summary Revenue Potentials, Annual Running Costs and Transferred Data

Volumes in Scenarios (A), (B) and (C).

The results illustrate that operations of a complex advanced metering infrastructure does

not – resp. hardly - pay off for an electricity retailer when solely executing DR programs.

However, getting back to our second research question, this application scenario may open

up the field for new players in the energy domain. Once the usage data records have been

collected, the dataset holds the potential to enable additional services beyond demand

response that yield supplementary revenue for the operator. Applications for the collected

dataset reach from process improvements, such as automated billing, to innovative value-

added services. Moreover, aggregators could create synergies when operating a central

data processing infrastructure. This is particularly true for telecommunications providers

that are able to resort to existing communication network infrastructure (e.g.

concentrators). Only this way the prohibitively high initial investments can be

compensated resp. decreased.

5.3 Policy Implications for Electricity Markets Design

The integration of DR resources into current electricity markets requires various changes,

in particular, to the market design of reserve exchanges. In the following, we discuss, in

detail, the required key adjustments:

The tertiary reserve energy exchanges underlie strong restrictions in terms of quality

and quantity of the traded energy reserve. Whereas DR resources are able to comply

with the quality requirements (e.g. being activated and available within 15 minutes),

191

our evaluation shows that the defined minimum volume per transaction has to be

substantially decreased. For example, a minimum amount of 5 MW of tertiary reserve

energy capacity has to be offered today by the provider per time slot in the German or

US marketplace15

. In our evaluation setting, the electricity retailer trades capacity

volumes that on average are less than 1 MW.

While costs for balancing energy and control reserve are charged for 15-minute

intervals, contracts offered at many energy exchanges – such as the European Energy

Exchange – comprise delivery periods of at least one hour. As a result, the retailer has

no chance to procure missing electricity for any imbalances of 15 minutes. Hence, our

model suggests considering electricity contracts with shorter delivery periods.

Auctions for secondary reserve take place 1 month before delivery and for primary

even 6 months before. According to (Paulus and Borggrefe, 2011), “the requirements

for secondary and primary reserve markets are too restrictive” for integration of DR

resources. (Borggrefe and Neuhoff, 2011) adds to this and states that day-ahead

auctions are required. This is in particular true when looking at our setting. Here, we

assume that a DR mechanism is used for addressing increasing fluctuations on the

supply side that is induced by renewables. The electricity retailer is not able to

generate reasonable estimates and offer respective resources that far in advance. Even

intraday auctions might become necessary to leverage the full potential of demand

response – this would also require changes to the current exchanges for tertiary

reserve that execute day-ahead auctions only.

6 Conclusion and Outlook

Due to the integration of intermittent resources of power generation, the amount of

supplied energy will show unprecedented fluctuations. This challenge can be addressed by

using DR systems for shifting power demand according to the fluctuating supply side and,

15 Source: BK6-10-099 in Germany; e.g. PJM (http://www.pjm.com/~/media/documents/manuals/m11.ashx) as a Regional

Transmission Organization, which belongs to the Eastern Interconnection in the United States.

http://www.pjm.com/~/media/documents/manuals/m11.ashx

192

consequently, integrating Information Systems for DR control into electricity markets. As

we have shown in this part, a scenario where electricity retailer leverages demand response

for optimizing the energy procurement strategies is most profitable compared with

application scenarios of using DR resource as tertiary reserve or to avoid balancing energy

penalties. However, this only represents a snapshot. Prices for reserve and balancing

energy are assumed to be significantly rising in the near future (Kladnik et al., 2012;

Madrigal and Porter, 2012), which consequently leads to increased financial benefits for

the retailer and, hence, could make these scenarios financially rewarding.

Based on our evaluation, we have posed policy implications that need be incorporated into

electricity market design. In short, market transactions have to be conducted in short-term

(at least day-ahead) and allow for smaller volumes to be traded. Moreover, we have taken a

look at business opportunities for aggregators. For them, setting up and operating a DR

infrastructure might generate a viable business case as they can profit from synergies when

offering the infrastructure to more than one electricity retailer. In addition, they may be

able to reuse the infrastructure for additional valuable applications and services.

In future work, we plan to extend the financial perspective towards the connected

customers. First, customers need to be somehow incentivized to participate in the DR

program – most probably, this requires a financial compensation. Second, the customers

have to be equipped with Home Area Networks to enable DR mechanisms. So far, we have

restricted the cost model to IS-related costs within the electricity retailer’s realm.

Furthermore, we intend to further enhance the revenue model by allowing intra-day

allocation and shifting of load. This would better reflect the effects of integrating

renewables into the power grid. Hitherto, all market transactions are fixed day-ahead. All

in all, this contributes to the discussion on the optimal information granularity. This

research topic is not only restricted to DR, but is also extended to smart grid applications

(cp. Dalen et al., 2013; Heinrich and Zimmermann, 2012) and has been gaining increasing

traction within the IS community.

193

7 References

Ahlert, K.-H., and van Dinther, C. (2009). Estimating Economic Benefits of Electricity

Storage at the End Consumer Level. Proceedings of Wirtschaftsinformatik 2009 (Paper

143).

Bakos, J. Y. (1997). Reducing Buyer Search Costs: Implications for Electronic

Marketplaces,” Management Science (43:12), pp. 1676–1692.

Bodenbenner, P., Feuerriegel, S., Neumann, D. (2013). Design Science in Practice:

Designing an Electricity Demand Response System. Proceedings of the 8th International

Conference on Design Science Research in Information Systems and Technology

(DESRIST).

Borggrefe, F., and Neuhoff, K. (2011). Balancing and Intraday Market Design: Options for

Wind Integration. DIW Berlin Discussion Paper No. 1162, SSRN Electronic Journal.

Brandt, T., Wagner, S., and Neumann, D. (2012). Road to 2020: IS-Supported Business

Models for Electric Mobility and Electrical Energy Markets. Proceedings of the 33rd



Perspective. Proceedings of the 17th Americas Conference on Information Systems

(AMCIS).

Dalen, A., Krämer, J., and Weinhardt, C. (2013). Determining the Optimal Level of

Information Granularity for Efficient Energy Consumption Decisions: Experimental

Evidence. Proceedings of the 21st European Conference on Information Systems

(ECIS).

Dedrick, J. (2010). Green IS: Concepts and Issues for Information Systems Research.

Communications of the Association for Information Systems (27:1).

Doostizadeh, M., and Ghasemi, H. (2012). A day-ahead electricity pricing model based on

smart metering and demand-side management. Energy (46:1), pp. 221–230.

E-Control (2012). Energy Consumers’ Structure. http://www.e-control.at/. Accessed

2 December 2012.

Eto, J. H., Nelson-Hoffman, J., Torres, C., Hirth, S., Yinger, B., and Kueck, J. (2007).

Demand Response Spinning Reserve Demonstration. Ernest Orlando Lawrence

Berkeley National Laboratory, Berkeley, USA.

European Energy Exchange, EEX (2012). Market Data. http://www.eex.com/. Accessed

2 May 2012.

Federal Energy Regulatory Commission, FERC (2006). Assessment of Demand Response

and Advanced Metering.

Feuerriegel, S., Strueker, J., and Neumann D. (2012). Reducing Price Uncertainty through

Demand Side Management. Proceedings of the 33rd





Markets: Forthcoming. Paper 192. Proceedings of the 21st European Conference on

Information Systems (ECIS).

Gottwalt, S., Ketter, W., Block, C., Collins, J., and Weinhardt, C. (2011). Demand side

management—A simulation of household behavior under variable prices. Energy

Policy (39:12), pp. 8163–8174.

http://www.e-control.at/

http://www.eex.com/

194

Groiß, C. (2008). Power Demand Side Management: Potentiale und technische

Realisierbarkeit im Haushalt. Diploma Thesis. Faculty of Electrical Engineering and

Information Technology, Vienna University of Technology.

Gross, R., Heptonstall, P., Anderson, D., Green, T., Leach, M., and Skea, J. (2006). The

Costs and Impacts of Intermittency: An Assessment of the Evidence on the Costs and

Impacts of Intermittent Generation on the British Electricity Network. Report. UK

Energy Research Centre.

Gungor, V. C., Sahin, D., Kocak, T., Ergüt, S., Buccella, C., Cecati, C., Hancke, G. P., and

Ergut, S. (2011). Smart Grid Technologies: Communication Technologies and

Standards. IEEE Transactions on Industrial Informatics (7:4), pp. 529–539.

Heinrich, B., and Zimmermann, S. (2012). Granularity Metrics for IT Services.

Proceedings of the 33rd


Jacobsen, H. K., and Zvingilaite, E. (2010). Reducing the Market Impact of Large Shares

of Intermittent Energy in Denmark. Energy Policy (38:7), pp. 3403–3413.

Kirby, B. J. (2004). Frequency Regulation Basics and Trends. Report ORNL/TM-

2004/291. Oak Ridge National Laboratory, Oak Ridge, Tennessee, USA.

Kladnik, B., Artac, G., Stokelj, T., Golob, R., and Gubina, A. F. (2012). Demand-Side

Participation in System Reserve Provision in a Stochastic Market Model with High

Wind Penetration. Proceedings of the IEEE International Conference on Power System

Technology (POWERCON), pp. 1–5.

Klobasa, M. (2007). Dynamische Simulation eines Lastmanagements und Integration von

Windenergie in ein Elektrizitätsnetz auf Landesebene unter regelungstechnischen und

Kostengesichtspunkten. Doctoral Thesis, Swiss Federal Institute of Technology (ETH),

Zurich.

Law, Y. W., Alpcan, T., Lee, V. C., Lo, A., Marusic, S., and Palaniswami, M. (2012).




Liebau, B. (2012). Der deutsche Strommarkt: Marktdesign und Anbieterverhalten.

Doctoral Thesis. University of Muenster, Germany.

Madrigal, M., and Porter, K. (2012). Operating and Planning Electricity Grids with

Variable Renewable Generation: Review of Emerging Lessons from Selected

Operational Experiences and Desktop Studies, Washington DC, USA: World Bank.

Malone, T. W., Yates, J., and Benjamin, R. I. (1987). Electronic markets and electronic

hierarchies. Communications of the ACM (30:6), pp. 484–497.

Meng, F.-L., and Zeng, X.-J. (2012). A Stackelberg Game Approach to maximise

electricity retailer's profit and minimse customers' bills for future smart grid.

Proceedings of the 12th UK Workshop on Computational Intelligence (UKCI).

Palensky, P., and Dietrich, D. (2011). Demand Side Management: Demand Response,


Informatics (7:3), pp. 381–388.

Paulus, M., and Borggrefe, F. (2011). The potential of demand-side management in

energy-intensive industries for electricity markets in Germany. Applied Energy (88:2),

pp. 432–441.

PWC Austria (2010). Studie zur Analyse der Kosten-Nutzen einer österreichweiten

Einführung von Smart Metering. Pricewaterhouse Coopers.

195

Regelleistung (2012). Datencenter der deutschen Übertragungsnetzbetreiber:

Erläuterungen zu den Veröffentlichungen auf der gemeinsamen Internet-plattform

www.regelleistung.net. Manual.

Riedel, S., and Weigt, H. (2007). German Electricity Reserve Markets. Electricity Markets

Working Paper No. EM-20, SSRN Electronic Journal.

Rosen, C., and Madlener, R. (2012). Auction Design for Local Reserve Energy Markets.

E.ON Energy Research Center (ed.).

Shayesteh, E., Yousefi, A., and Parsa Moghaddam, M. (2010). A probabilistic risk-based

approach for spinning reserve provision using day-ahead demand response program.

Energy (35:5), pp. 1908–1915.

Strbac, G. (2008). Demand side management: Benefits and challenges. Energy Policy

(36:12), pp. 4419–4426.

Strüker, J., and van Dinther, C. (2012). Demand Response in Smart Grids: Research



Styczynski, Z. (2011). VDE-Studie Demand Side Management. Report. Association for

Electrical, Electronic & Information Technologies.





U.S. Department of Energy (2006). Benefits of Demand Response in Electricity Markets

and Recommendations for Achieving Them. A Report to the Unit States Congress

Pursuant to Section 1252 of the Energy Policy Act of 2005.

Vandezande, L., Meeus, L., Belmans, R., Saguan, M., and Glachant, J.-M. (2010). Well-

functioning balancing markets: A prerequisite for wind power integration. Energy

Policy (38:7), pp. 3146–3154.

Vasirani, M., and Ossowski, S. (2012). A Collaborative Model for Participatory Load

Management in the Smart Grid. Proceedings of the Workshop on AI Problems and

Approaches for Intelligent Environments, pp. 21–26.

Vasirani, M., and Ossowski, S. (2013). Smart Consumer Load Balancing: State of the Art

and an Empirical Evaluation in the Spanish Electricity Market. Artificial Intelligence

Review (39:1), pp. 81–95.

Wagner, S. (2012). Developing Vehicle-to-Grid Regulation Algorithms to Optimize

Economic Revenues. Master Thesis, University of Freiburg.

Watson, R. T., Boudreau, M.-C., and Chen, A. J. (2010). Information Systems and


the IS-Community. MIS Quarterly (34:1), pp. 23–36.

196

IV. APPENDIX

PART IV APPENDIX

197

1 Curriculum Vitae

The vita includes personal data, which have been removed from the online version.

PART IV APPENDIX

198

The vita includes personal data, which have been removed from the online version.

PART IV APPENDIX

199

2 Referred Research Publications

P. Bodenbenner, S. Feuerriegel, D. Neumann, 2014: Integrating Demand Response

Systems in Electricity Markets.

22nd

European Conference on Information Systems (ECIS 2014), Tel Aviv, Israel, June 9-11,

2014 (Forthcoming).

P. Bodenbenner, S. Feuerriegel, D. Neumann, 2013: Design Science in Practice:

Designing an Electricity Demand Response System.

8th

International Conference on Design Science Research in Information Systems and

Technology (DESRIST 2013), June 11-12, 2013, Helsinki, Finland, Editor: Jan vom Brocke et

al, Proceedings Name: LNCS 7939, pp. 293-307, Springer-Verlag, Berlin/Heidelberg.

S. Feuerriegel, P. Bodenbenner, D. Neumann, 2013: Is More Information Better Than

Less? Understanding the Impact of Demand Response Mechanisms in Energy Markets.

21st European Conference on Information Systems (ECIS 2013), Utrecht, Netherlands, June 5-8,

2013, Paper 192, Completed Research Paper.

Nominated for Claudio Ciborra award (most innovative paper) at ECIS 2013.

P. Bodenbenner, D. Neumann, 2012: Are Personalized Recommendations the Savior for

Online Content Providers?

Multikonferenz Wirtschaftsinformatik (MKWI 2012), Braunschweig, Germany, February 29 -

March 2, 2012.

P. Bodenbenner, M. Hedwig, D. Neumann, 2011: Impact of Recommendations on

Advertising-based Revenue Models.

Pre-ICIS Conference - 10th

Workshop on E-Business (WEB 2011), Shanghai, China, 4

December 2011.

the value proposition of information-based business models

Documents