Forecasting Short Term Demand in Heterogeneous Customer Oriented Demand Management Processes

POLITECNICO DI MILANO Faculty of Engineering Ph.D. in Management Engineering - XV cycle

FORECASTING SHORT TERM DEMAND IN HETEROGENEOUS CUSTOMER ORIENTED DEMAND MANAGEMENT PROCESSES

Matteo Kalchschmidt

Thesis Advisor: Prof. Roberto Verganti

Chair, Ph.D. Program in Management Engineering: Prof. Giuliano Noci

December 2002

Acknowledgments

Even if this work has been written by a single person, it is the result of the different contributions of a much wider number of people.

In first place, I would like to thank the whole research group of Business and Innovation Management within the Department of Industrial Engineering at Politecnico di Milano. In particular a special thank is for Professor Roberto Verganti and Professor Emilio Bartezzaghi that helped me out from this hard work and that made me understand why it was so important. Special thanks to Gianluca, Mariano, Raffaella, Stefano, Tommaso, Federico and Alessio, for letting me participate to their experiences and for participating to mine. A particular remark goes to Giulio Zotteri, which I have to thank and blame for pulling me into this adventure. Special thanks go to my family and my love, that even if not in a scientific perspective, for sure contributed to this work. I would also like to thank all the people belonging to the companies here reported (and to those that are not), that helped me in the development of this work.

He cerrado mi balcón por que no quiero oír el llanto pero por detrás de los grises muros no se oye otra cosa que el llanto. Hay muy pocos ángeles que canten, hay muy pocos perros que ladren, mil violines caben en la palma de mi mano. Pero el llanto es un perro inmenso, el llanto es un ángel inmenso, el llanto es un violín inmenso, las lágrimas amordazan al viento, no se oye otra cosa que el llanto.

Federico García Lorca (1936)

Contents

Introduction 1

PART I: LITERATURE REVIEW

Chapter 1: Customer Orientation 9 1 Introduction to the Problem 9 2 Customer Orientation 13 3 Customer Orientation in Supply Chain Management 19 3.1 Supply Chain Structure 21 3.2 Industrial Networks 22 3.3 Customer Oriented Supply Chains 23 3.3.1 Mass Customization 24 3.3.2 Postponement 26 3.3.3 Customer Relationship Management 27 3.4 Demand Chain Management 28 3.5 Customer Oriented Inventory Management 30 4 Conclusions 34

Contents

Chapter 2: Demand Forecasting and Customer 35 Orientation 1 Demand Management 35 1.1 Demand Management Process 36 1.2 Uncertainty Management: the case of lumpy demand 39 2 Demand Forecasting 42 2.1 Methods for Stable Demand based on Demand Data 45 2.1.1 Smoothing and Average based Techniques 46 2.1.2 Bivariate and Multivariate models 46 2.1.3 ARIMA methods 47 2.2 Methods for Lumpy Demand based on Demand Data 49 2.2.1 Croston Method 49 2.2.2 Syntetos and Boylan Method 52 2.3 Methods for Lumpy Demand based on Information regarding 54 Demand Generation Process 2.3.1 Lost Sales Estimation 56 2.3.2 Analysis of Reliability 57 2.3.3 Early Sales 58 2.3.4 Order Overplanning 64 2.3.5 Multi Level Supply Control 67

PART II: THEORY AND METHODOLOGY

Chapter 1: Research Aims and Methodology 71 1 General Considerations 71 1.1 Factors influencing demand uncertainty and heterogeneity 72 1.2 Effects on Customers’ Structure 78 1.3 Limitations of Actual Forecasting Approaches 82 2 Definition of Research Aims 87 3 Research Methodology 88 3.1 Theoretical Stage 88 3.2 Empirical Stage 90

Chapter 2: Theoretical Framework 93 1 Heterogeneity: a definition 93 1.1 Heterogeneity Dimensions 96 1.2 Heterogeneity Measures 99 2 Impact of Heterogeneity on Demand 103

Contents

2.1 Customer Size 104 2.2 Purchasing Politics 106 2.2.1 Lot Sizing 106 2.2.2 Reorder Interval 110 2.3 Promotional Politics 111 2.3.1 Number of Promotions 112 2.3.2 Promotion Size 113 3 Heterogeneity and Forecasting 115 4 Conclusions 122

PART III: EMPIRICAL ANALYSIS

Chapter 1: Introduction to the Action Research Case 125

Chapter 2: Whirlpool Europe 131 1 Context and Purpose 131 2 Introduction to the company 133 2.1 Whirlpool Europe Spare Parts Centre 136 3 Analysis of the Supply Chain and Demand 138 3.1 The Evolution of the Supply Chain Structure 138 3.2 Impact of the Supply Chain structure on Demand Variability 140 4 Solution Design 146 4.1 Alternative a): Current Solution 146 4.2 Alternative b): Literature model 147 4.2.1 Filtering 147 4.2.2 Forecasting 151 4.3 Alternative c): Ad hoc model 152 4.3.1 Forecasting 152 4.3.2 Inventory Management 157 4.4 Alternative d): improving performances through information 159 5 Performance 161 6 Conclusion 166

Chapter 3: Nestlé Italiana 169 1 Context and Purpose 169 2 Introduction to the company 170 2.1 Nestlè Italiana 172

III

Contents

2.2 The Fresh Food Division 173 2.3 Logistic Structure 174 3 Analysis of the Supply Chain and Demand 178 3.1 Analysis of demand variability 178 3.2 Actual forecasting performances 182 4 Solution Design 186 4.1 Forecasting for the 22 monitored customers 187 4.2 Residual customers forecasting approach 194 4.2.1 Aggregate Solution 194 4.2.2 Cluster Based Solution 197 5 Cluster analysis simulation 205 5.1 Simulation 1 207 5.2 Simulation 2 208 5.3 Simulation 3 209 5.4 Simulation results 209 5.5 Application to Nestlè 212 6 Conclusions 214

Chapter 4: Ahold 219 1 Context and Purpose 219 2 Introduction to the company 223 2.1 Ahold History 223 2.2 Description of the Supply Chain 225 2.3 Actual Forecasting System 227 3 Solution Development 230 3.1 Disaggregate Model 230 3.2 Mixed Model 231 3.3 Aggregate Model 233 3.4 Cluster Model 234 4 Analysis of results 238 5 Conclusions 244

Chapter 5: Conclusions and Future Developments 249 1 Analysis of the Action Research results 249 1.1 Demand management approaches within heterogeneous 250 contexts 1.2 Main elements to be considered 256 2 Generalization of findings 260 3 Final remarks 262 4 Future developments 265

Contents

References 267

Introduction

In many industrial contexts firms are dealing with a demand that is ever more uncertain. The reasons tied to this phenomenon are many, however, a major change that is spreading among different sectors is the ever growing attention towards customers. In fact, companies have identified that customers are critical for their businesses not only because they influence directly the success of a specific product or firm, but also because their role is fundamental within many different internal processes. As a matter of fact, firms are paying a ever greater attention towards customers’ influence on business process. Attention is typically given in the product development field, where literature claims that customers have to be involved in the early stage of the product development. Customer orientation is a relevant issue in the marketing-related theory and practice, where attention by both researchers and practitioners has been paid on the proper interaction with customers, thus leading to the so-called Interactive or Direct Marketing.

Even if the customer role within business process have been deeply analyzed, some research field have not yet exploited this issue completely. Attention in this work is given towards demand forecasting and the role of customer when dealing with heterogeneous contexts. In particular, our main attention is paid towards contexts where companies have to manage different customers influenced by different factors and that react differently towards similar external variables.

Introduction

As a matter of fact, companies’ attention also within the forecasting issue is changing towards the problem of dealing with heterogeneous customers. Many firms are aware that due to the changing environment, customers tend to behave differently and that serving properly market with suited solutions is becoming fundamental to properly compete. However, managing heterogeneous contexts becomes critical since it tends to make more difficult understanding and foretelling future market requirements, thus claiming that contributions towards forecasting matters are required. In particular, this work focuses on short term demand forecasting, thus attention is given towards estimating future market requirements for planning purposes.

Within the described context, this research aims at studying the impact of heterogeneity among customers’ requests and to design both a proper methodology and specific approaches to cope with demand in heterogeneous contexts. In particular, three main research questions are considered in this work.

How does heterogeneity in customers’ purchasing processes influence demand variability and forecasting effectiveness?

How should demand be managed when dealing with heterogeneous contexts?

What are the main elements that should be taken into account when dealing with heterogeneous contexts?

To properly answer these questions, this work has been structured in three stages corresponding to the different parts in which the work is structured.

First of all a wide analysis of current literature has been conducted and contributions have been deeply reviewed and analyzed. In particular attention has been paid towards from one side customer-orientation and its impact within business processes, while, from the other, demand management and forecasting contributions have been considered. This analysis let us define the problem addressed and state some of the research hypothesis this work addresses.

Given this framework the empirical methodological can be applied. In particular, the research is based on a two-stage process; the first stage aims at analyzing heterogeneity and its impact on demand variability and forecasting accuracy. We conducted this analysis

Introduction

by means of simulation, in particularly by simulating different sorts of heterogeneity and measuring the impact on demand variability. The outcome of the first stage is the definition of proper answers towards some of the previously defined research questions and the definition of clear hypothesis that have to be tested in the following stage. These hypothesis synthesize the general structure of the forecasting methodology proposed. After this stage, attention is paid towards the detailed design of a proper methodology to cope with forecasting in heterogeneous contexts and thus to the validation of the defined hypothesis. This part has been developed according to an action research methodology, based on the development of 3 action research cases. This research considered in fact 3 companies facing problems in managing properly the demand they handle, due to a relevant heterogeneity in the ways customers purchase. In the different cases proper forecasting approaches, developed coherently with the forecasting methodology previously defined, were designed and tested using real demand data.

The adoption of the described methodology let us achieve the results of the research that can be summarized as follows. First of all, the work identifies that heterogeneity within demand generation process deeply influences both demand variability and forecasting accuracy. In particular, when heterogeneity rises variability tends to rise too; however, more important is the fact that for high heterogeneity forecasting becomes much more difficult to be conducted since, if attention has to be paid to the demand generation process, high heterogeneity implies that different variables influence differently customers, thus making estimation process very complex. Moreover, this work states that this problem is more critical when significant performances in forecasting accuracy are required since, higher performances, as literature claims too, require that attention has to be given towards customers’ purchasing processes.

A second relevant result is the definition of a proper methodology to cope with variable demand in heterogeneous contexts. In particular, the proposed methodology is essentially based on a three-step process. First of all, in fact, it must be understood at which level forecast has to be developed, so whether attention has to be given to single customers, to total demand or somewhere in the middle between these two extreme conditions. Then, if forecast has to be conducted at some detailed level, customers have to be separated by means of heterogeneity, so to identify homogeneous groups of customers that can be forecasted easier. At last proper forecasting techniques have to applied, so that the different demand patterns referred to the different clusters can be effectively modeled and foreseen.

Introduction

The overall methodology has been adopted and tested within the 3 action research cases thus gaining significant operational improvements.

A third major contribution is tied to the definition of proper techniques to manage different sorts of heterogeneity. In particular, in the Whirlpool case we propose an innovative method to manage demand when heterogeneity is essentially due to the difference in the purchasing politic applied by customers and so to their relative dimension. In the Nestlé case, we propose a general method to manage demand affected by promotional activities applied specifically on particular customers. Moreover, the proposed method manages heterogeneity both in the application of promotional activities and in the different customers sensibility towards these phenomena. In the end the Ahold case proposes a similar application of the methodology applied in the Nestlé case that can be easily applied when heterogeneity on customers’ inner characteristics shows up.

At last a relevant contribution is tied to the identification of some guidelines to the applicability of both the proposed methodology and the specific techniques proposed. In particular we identified structural elements of the supply chain and of the market structure that should be considered when applying such approaches. As a matter of fact we identified that heterogeneity has to be considered when systematic phenomena (e.g., seasonality) explains only a minor part of the total demand variability. Moreover, we identified that when supply chain is complex, attention has to be shifted towards heterogeneity, since this element tends to make heterogeneity relevant. As a matter of fact detailed analyses have been developed to understand under which conditions the techniques proposed in the different cases should be adopted.

As previously stated the structure of this work directly replicates the methodology adopted (see Figure 1). In fact in Part I literature review is provided, in particular in Chapter 1 attention is paid towards customer-orientation and its impact on business processes, while in Chapter 2 forecasting literature is reviewed and discussed. Part II, then, defines the theory and methodology of the research; in particular Chapter 1 defines the research aims and methodology, while Chapter 2 describes the theoretical framework and the research hypothesis analyzed in the empirical part. In the end, Part III focuses on the empirical analysis; the first three chapters describe the action research cases conducted, in particular after a brief overview of the considered cases (Chapter 1), Chapter 2 considers the Whirlpool Europe case, while Chapter 3 the Nestlé

Introduction

Italiana situation and Chapter 4 describes the Ahold case. At last Chapter 5 draws conclusion of the overall dissertation and identifies some future developments.

Chapter 1 Chapter 2 Part I Customer Demand Literature Review Orientation Forecasting

Chapter 1 Research Aims and Methodology Part II Theory and Methodology Chapter 2 Theoretical Framework

Chapter 1 Introduction to the Action Research Cases

Chapter 2 Chapter 3 Chapter 4 Part III Whirlpool Nestlé Ahold Empirical Analysis Europe Italiana

Chapter 5 Conclusions and Future Developments

Figure 1: structure of the research

Introduction

PART I:

LITERATURE REVIEW

Chapter 1 Customer Orientation

1 Introduction to the problem

Many industrial sectors are showing relevant changes due to many context factors. The great rise of competitiveness, the new dimensions of complexity and the ever changing rules of environment, have in fact contributed to the development of new ways of conducting business activities. From one side competitiveness is rising since globalization makes easier to access markets and reduces barrier for newcomers. From the other side technology, and in particular Information and Communication Technology, are influencing the way in which businesses are conducted favoring cooperation and interaction between producers, distributors, consumers, suppliers, and so on. However, many firms show an increasing complexity in the technology management since to properly compete many new technologies have to be adopted. Moreover, many industrial sector, show a constant product life time reduction. These factors have in fact contributed to the evolution of environment leading to higher uncertainty in business activities and so to the need of new models to interpret reality.

Part I Chapter 1: Customer Orientation

In fact, in many industrial contexts firms have to deal with a demand which is more and more uncertain. The increasing market turbulence, the evolution of products and service over time and the harder and harder competition between firms, have deeply influenced the variability of demand and the relevance of this uncertainty over the firm performances. Many firms are gradually becoming aware that major attention has to be paid on demand uncertainty and that flexibility is more and more difficult to be achieved and sustained.

Typically at first firms tend to react by leveraging on inner process performances: they try to gain efficiency by re-engineering business processes so to improve both efficiency and efficacy. This, however, isn’t always enough as economical systems are far more complex thus making management a tough job. In this situation many firms try to react by leveraging on some business partners. This is usually a winning solution when problems need many different competencies to be solved and risk sharing is a relevant matter for gaining success. Firms tend to locate partners among suppliers and customers, or, sometimes, among their own competitors. A few decades have in fact shown a significant movement towards core competences thus rising the number of links between firms. These approaches tend to influence flexibility and reacting capabilities that firms need to manage uncertainty so to gain success. This is only a single part of the total solution, since it does provide more ability in reacting towards relevant changes, but it does not make these changes clearer nor more predictable. In other words, firms have typically reacted by reducing the effects of uncertainty on business performances, but not improving their ability in understanding and foreseeing uncertainty itself. Since markets are evolving and becoming much more complicated, understanding markets and customers is becoming an ever more critical leverage of success. From another side, however, it is ever more difficult to understand which changes are taking place and where they will lead since customers are evolving too. They require better services, more complex products, new solutions to their needs and, as a general rule, more customization. In fact, many sectors have seen a great rise of customer importance, since many firms are becoming more and more aware that customers have strong differences among them and they have to provide them with suited products and services. The global effect is that demand is becoming more uncertain and unpredictable, and so demand foreseeing is ever more a critical issue.

Given this problem, many industrial and academic realities have paid great attention towards the impact of customers on business performances and in particular they have

Part I Chapter 1: Customer Orientation

focused on the value of the information provided by customers to provide them with better solutions. The general idea of these approaches is that since uncertainty is given by single customers’ requirements, a possible solution to better manage this phenomenon is to focus on single customers. This is generally true for many business processes but demand forecasting seems the one that most can exploit from this sort of approach. Since demand is becoming more and more unpredictable, new solutions have to be designed by looking at what customers want and how they aim at obtaining it, so looking deeper towards their specific requirements by means of information regarding their purchasing process. Different studies have provided evidence of how using information produced by customers can deeply influence the demand management process and its performances, thus reducing uncertainty effects. In this way customer orientation becomes a relevant matter.

However, this solution is not free of problems. First of all, a relevant matter of these solutions is tied to the costs and difficulties of achieving information. Customers are becoming more and more aware of the value of the information they can provide and they are not likely to make it available very easily. This makes the information retrieval process rather complex thus deeply impacting on the costs of gaining such information. Besides, distribution channels are evolving rapidly, leading to the existence of many different channels by means of which firms can sell their products. This makes it difficult to keep in touch with customers since many different intermediaries are in the middle. Moreover, having a great amount of information shifts attention towards the problem of managing all this data. This becomes a relevant matter for two main reasons. From one side, information is often distributed in different data bases that are not connected among them, so leading to a non efficient data organization. From the other side there is a reduced awareness of the possible tools and methods that can be adopted to manage this information. Actually firms have reacted to these changes in many different ways and both researchers and practitioners have studied and developed new systems, new ideas and new tools to develop current activities. From one side, firms have developed solutions regarding their access to resources; this is mainly concern of all buyer-supplier literature, from the theories of networks to Supply Chain Management. From another point of view, firms have paid great attention to the development of proper internal processes towards the gain of much wider efficiency, typically by means of Business Process Reengineering methodologies. Moreover, one of the main struggles that actually firms are challenging is for sure the changing role of customers. In fact, since the role of customers within business processes is changing, processes themselves have to change so to leverage on customer-orientation. In particular, this work

Part I Chapter 1: Customer Orientation

is focused on demand forecasting and customer orientation, thus to understand properly the relationship between these two issue, attention has to be paid on what the two constituting topics are. This first chapter is in fact aimed at defining properly customer-orientation and its impact on business processes structures and performances. In particular, paragraph 2 will define customer-orientation while paragraph 3 will go in further detail by analyzing the impact of customer-orientation on supply chain management issues. This will make it possible to understand the impacts of this concept on demand management and forecasting issues, which will be detailed in the next chapter.

Part I Chapter 1: Customer Orientation

2 Customer Orientation

As previously anticipated, the evolution of economic contexts requires a change of perspective in the way market is approached. Typically we can trace different concepts under which organizations conduct their activities.

The Production Concept. The production concept holds that consumers will favor those products that are widely available and low in cost. Managers of production- oriented organizations concentrate on achieving high production efficiency and wide distribution coverage. The Product Concept. The product concept considers that the consumer will favor those products that offer the best quality performance, or innovative features. Managers in these organizations focus their attention on making superior products and improving them over time. The Selling Concept. The selling concept holds that consumers, if left alone, will ordinarily not buy enough of the organization’s products. So organizations have to undertake aggressive selling politics. The Market Concept. The market concepts considers that the key to achieving organizational goals consists in determining the needs and wants of target markets and delivering satisfactions more effectively and efficiently than competitors.

These concepts have developed along time widening the number of areas in which attention has to be paid to gain success. Past literature in fact has devoted attention towards the so-called market orientation. Narver and Slater (1990) define market orientation as:

“…the organization culture that most effectively creates the necessary behaviors for the creation of superior value for buyers and thus continuous superior performance for the business.” (p. 21)

In this way market orientation is considered as a strategy that focuses on providing customers with proper solutions to their needs by means of profitable business models. These authors describe a model of market orientation consisting of three conceptually closely related and equally important aspects: customer orientation, competitor orientation and inter-functional coordination (see Figure 1).

Part I Chapter 1: Customer Orientation

Market Orientation

Customer Competitor Inter-functional Orientation Orientation coordination

Figure 1: Narver and Slater’s market orientation model

They state that a proper market orientation has to be achieved by means of melting together these concepts. No general rule applies in defining the proper balance of these dimensions, but some authors identify environmental variables that influence the relationship among these variables and business performances. In particular, Slater and Narver (1994), building on arguments advanced by Day and Wensley (1988), identified four moderator variables: market growth rate, buyer power, seller concentration and competitive hostility. They have argued that a customer orientation has a stronger impact on profitability in markets characterized by a high growth rate, low buyer power, a high degree of competitive hostility and a low seller concentration. From the other side, a competitive orientation has a stronger impact in the reverse situation. Interestingly, many industrial sectors are showing a shifting towards low buyer power and low seller concentration, thus providing empirical evidence that customer-orientation is becoming an ever relevant approach.

Narver and Slater’s model has been found to be the most robust measure of market orientation, both in terms of its application by other researchers and according to confirmatory factor analysis (Matear et al, 1997). Other modifications have been provided by other authors (i.e. Kohli and Jaworski, 1990) but the general model is similar.

Customer orientation, one particular aspect of market orientation in this model, has been deeply analyzed by different researchers regarding many different aspects. A firm can very well define its market orientation but fail in assuming a proper customer orientation, in terms that a firm must achieve the customer’s viewpoint. In other words assuming a customer orientation means being able to listen properly to the customer and to use this information properly in the operations management. Think, for example, to the expectations of a car buyer, where many different needs are concerned, from absolute

Part I Chapter 1: Customer Orientation

reliability, to low fuel consumption, to comfort, to design, and so on. A car model will not be able to satisfy all needs, so the product manager will have to take decisions. However, it is important to adapt at best to the needs of customers without adopting an inner guiding line. For example 3M states that two among three new ideas come from listening properly to customers claims. In these terms, customer orientation may be defined as the organizational approach that focuses all units within a company towards the customer and his/her needs. In other words, customer-orientation represents situations:

“…in which all functions work together to sense, serve and satisfy the customer.” (Kotler, 1991)

Figure 2 exemplifies this definition.

Figure 2: customer orientation within firms

Even if customer orientation is typically a marketing-based issue, literature has provided contributions in other different areas concerned with operations management. In particular, many researchers have claimed for its relevance since customer orientation seems to affect business performances. Studies have been conducted in the US (Ruekert, 1992; Deshpande et al., 1993; Jaworski and Kohli, 1993; Slater and Narver, 1994; Atuahene-Gima, 1996; Balakrishnan, 1996; Slater and Narver, 2000), in Europe (Diamantopoulos and Hart, 1993; Greenley, 1995; Pitt et al., 1996) and in Australia (Conduit and Mavondo, 2001). Evidence of relationships has been found between customer orientation and different measures of performances. Some authors have in fact studied different performance metrics, tied to various managerial issues (from product

Part I Chapter 1: Customer Orientation

development to global performances, and so on), and customer orientation, to check whether this approach influences positively firms activities. Table 1 summarizes the studies related to this issue and the metrics they considered for evaluating performances.

Performance Measure Authors Cole et al., 1993 Profitability Wong and Saunders, 1993 Pitt et al., 1996 Deshpande et al., 1993 Profitability relative to largest competitor Balakrishnan, 1996 Satisfaction with profit Balakrishnan, 1996 Operating profits, profit-sales ratio, cash flow, return on Pelham and Wilson, 1996 investment Slater and Narver, 1994 Return on assets Pelham and Wilson, 1996 Slater and Narver, 2000 Long run financial performance Ruekert, 1992 Product innovation Atuahene-Gima, 1996 Slater and Narver, 1994 New product success Atuahene-Gima, 1996 Mouritsen, 1997 Quality Webb et al., 2000

Table 1: literature review regarding the impact of customer orientation and firm performances

As it can be seen, different relationships have been analyzed providing evidence that customer orientation can impact significantly on firms’ performances. These benefits can be theoretically supported, as customer orientation provides a unifying focus and clear vision to an organization’s strategy centered on creating superior value for customers (Kohli and Jaworsky, 1990). As the strategy reflects consistent goals, objectives and policies, effective interdepartmental relationships enhance the performance of an organization. It should lead to a better work environment for employees, increased productivity and overall organizational success (Cole et al., 1993).

Some of these provided results were tied to market-orientated companies, that in fact had to pay great attention to these issues since critical for their success, and so they need to learn about their customers and to continue to update that learning (Slater and Narver, 1995). However, some authors argue that the influence of customer orientation is pervasive rather than restricted to specific organizational activities (Kohli and Jaworsky, 1990;

Part I Chapter 1: Customer Orientation

Tufano, 1992; Atuahene-Gima, 1996). For sure in different contexts similar approaches towards customers may have different effects on performances, but the general relationship seems to endure. Some authors have in fact identified and analysed context factors that tend to shift the market orientation-performances relationship (Narver and Slater, 1990; Deshpande et al., 1993; Balakrishnan, 1996; Kotler, 1991).

Moreover, while previous marketing activities were primarily focused on increasing market shares in terms of a mass marketing based on single transactions, the past few years saw a paradigmatic switch towards relationship marketing. There are various reasons for this change: on the one hand, this is due to the growing globalization of the markets in tandem with rapid technological progress causing classical competitive strategies such as cost leadership, differentiation and focusing on a market niche to lose their effect. At the same time, cost cutting and quality marketing programs provide an opportunity for competitors to catch up with a competitive edge in ever shorter periods of time. On the other hand, the change in values that is going on in society and manifests itself in an increased trend towards hedonism and individualism results in more differentiated customer needs and behaviour. As a response to these developments, the emphasis of marketing interest shifts from looking into the market share of a company to focus on its share of customers. While marketing had been characterized for a long time in just striving to win new customers, the new paradigm is based on the finding that corporate success can be improved, in particular, by increasing the profitability of existing customer relations and an extension of the period that such relationships last.

Studies have shown in this context that winning new customers can be up to five times more expensive than maintaining existing customer relations. Furthermore, a negative correlation was found between the portion of regular customers who defected and corporate profit (e.g., Reichheld and Sasser, 1990). Establishment of stable business relationships and networks is thus becoming a key target of marketing efforts. Collection of information on partners and creation of an atmosphere of trust, satisfaction, and commitment are of great importance for the establishment and maintenance of customer relations. This in fact can lead to substantial benefits, since it also favours the ratio of comprehension that a supplier has of his/her own customers (Da Silva et al., 2002).

Even if literature presents interesting benefits in implementing a customer-oriented approach, some authors identify possible problems that tend to make this issue quite complex. First of all, developing a proper customer-oriented approach needs to leverage on

Part I Chapter 1: Customer Orientation

the organisation, so efforts have to be put into reengineering organizations to effectively manage these approaches (Mouritsen, 1997; Kotler, 1991). Moreover a relevant problem is tied to make internal process coherent with the external approach. In fact, a critical issue is structuring organizations so that no internal conflicts arise and processes can be managed efficiently. Among typical enterprises processes, relevant attention should be paid on Supply Chain Management issues.

Part I Chapter 1: Customer Orientation

3 Customer Orientation in Supply Chain Management

In the last few years great attention has been paid from both companies and researchers towards Supply Chain Management. Supply Chain Management (SCM) can be defined in many different ways (see Ronchi, 2001). The Global Supply Chain Forum defines SCM as:

“The integration of business processes from end-user through original suppliers that provide products, services, and information that adds value for the customers.” (Lambert et al., 1998)

Many authors have provided different definitions focused on different aspects; however, as provided by Ronchi (2001), all authors that have given any definition of SCM, agree on three main aspects. SCM, in fact, is a:

Process Oriented Approach: supply chain management is described as a way to integrate all the activities performed across different companies with a process perspective. This entails a strong focus both on the final consumer and on the immediate downstream customer. This is very close to the process approach (Hammer, 1990; Davenport, 1993; Hammer and Champy, 1993). However, supply chain management goes beyond and implies not only the process integration among functions within a company, but the integration of activities realized in different companies is also perceived as a competitive factor. Regarding Products, Information and Funds. Most of the authors see supply chain management dealing with products, information and funds flow. Such aspect is directly related to the previous one. Products mean not only physical materials (raw materials, work in progress and finished products) but also services. All the relevant information needed to pursue efficiency and effectiveness across the supply chain must be shared among participants. Finally, an important issue in transactions is surely money and how it is transferred. Network of companies. The last important factor concerns the actors involved: in order to manage efficiently and effectively all flows across the chain, from the beginning to the end consumption, more companies need to interact. The presence of organizations with different and conflicting objectives makes finding the best

Part I Chapter 1: Customer Orientation

supply chain strategy a significant challenge (Simchi-Levi et al., 2000). Coordination and collaboration across the network open the discussion to concepts like trust and partnership, themes which many authors have coped with.

The basic SCM model suggests the management of activities at inter-organizational level as well as the department level. In fact, instead of focusing on the management of inter- firm inventory and transportation capacities, SCM aims to integrate the activities of an entire set of organizations from procurement of material and product components to deliver completed products to the final customer. These activities refer to areas as new product development, customer relationship management, customer service management, procurement and so on (see Figure 3).

Figure 3: supply chain processes

Consequently, SCM leads to improvements in channel performance among all channel members and not only within the local firm. Prominent examples of these positive effects include the supply chains of Wal-Mart (Gill and Abend, 1997), Digital Equipment Company (Arntzen et al., 1995), Hewlett-Packard Corporation (Hammel and Kopczak, 1993) and Lucent Technologies (Ronchi, 2001).

Literature has provided many contributions regarding all the topics connected within supply chain management. In fact, as Figure 3 shows, different areas can be trace back

Part I Chapter 1: Customer Orientation

towards this issue. However, many authors has also focused on the overall management of the supply chain rather than on the different topics described. Even if literature has provided relevant attention towards this issues, however contributions regarding SCM as a whole can be broadly divided in three main categories: first some authors focus on the supply chain structure and its effects on management, then others analyze industrial networks and the relationships among firms belonging to a network, at last some authors focus on customer orientation in supply chain management.

3.1 Supply Chain Structure

The first category of contributions constitutes of studies regarding the supply chain structure (Forrester, 1958, 1961; Burbidge, 1961; Sharman, 1984; Sterman, 1989; Towill et al., 1992; Lee and Billington, 1992; Lee et al., 1997; Holmström, 1994, 1995; Inger et al, 1995; Fisher, 1997). In this cluster attention is given towards the effects of the supply chain structure on firm performances. As provided by these authors, the structure of relationships influences both demand pattern at the different stages belonging to the chain, and decisional variables as inventory politics, purchasing approaches, and so on. A relevant phenomenon that these authors have dealt with is the so-called bullwhip effect. This phenomenon is represented by the fact that demand variability increases in the upper stages of the supply chain, thus leading to higher costs and lower service levels (see Figure 4). Suppliers Producer Distributor Retailer Customers Sales Sales Sales Sales

Time Time Time Time Figure 4: the Bullwhip Effect

Forrester (1961) illustrates this effect and points out that this is due to time varying behaviors of industrial organizations. Sterman (1989) reports evidence of the bullwhip effect in the “beer distribution game” and he interprets the phenomenon as a consequence of players’ systematic irrational behaviors.

Part I Chapter 1: Customer Orientation

Some authors provide evidence of this phenomenon in many different supply chains, as for example in the food industry (Van der Vorst et al., 1998) and in the European grocery chains (Holström, 1997). Moreover, Holström (1997) reports that the bullwhip effect in European grocery chains can increase order levels by 200% at the upper level. Metters (1997) analyzed the impact of the bullwhip effect on profitability. Interestingly, Lee et al (1997) demonstrate that the bullwhip effect is not a consequence of varying behaviors and irrationality, but it is due to demand forecast updating, order batching, price fluctuations and rationing and shortage gaming. These factors lead to excessive inventories, poor customer service, lost revenues, misguided capacity plans, ineffective transportation, and missed production schedules. The authors also suggest possible remedies based on information sharing, channel alignment, and operational efficiency. With information sharing, demand information is transmitted as timely as possible along all the tiers of the supply chain. Channel alignment is the coordination of pricing, transportation, inventory planning, and ownership between upstream and downstream sites in the supply chain. Finally, operationally efficiency refers to activities that improve performance, such as reduced costs and lead times. Some authors evaluate other possible solutions to these problems (Abel, 1985; Caplin, 1985; Kahn, 1987; Sterman, 1989; Cachon, 1999). Chen et al. (2000) have analyzed the effect of forecasting systems and information sharing through the supply chain. They have shown that providing each stage of the supply chain with complete information on customer demand can reduce the bullwhip effect. Xu et al. (2001) have also analyzed the impact of WMI and CRP programs on the magnitude of the bullwhip effect.

3.2 Industrial Networks

A second group of approaches is primary focused on industrial networks and the relationships between organizations in the chain (Williamson, 1985; Heide and John, 1990; Mohr and Spekman, 1994; Hakansson and Snehota, 1995; Kumar et al., 1995; Dyer, 1996a, b, c, 1997; Monczka et al.,1998). These authors essentially focus on the kinds of relationships that can be developed along a supply chain and try to identify which structures are most effective given certain conditions. In this theory field, relevant contributions have also been given regarding Network of Companies. In literature, different names have been adopted to describe networks of companies, some are here provided:

Part I Chapter 1: Customer Orientation

Supply networks can be defined as sets of supply chains, describing the flow of goods and services from original sources to end consumers (Harland, 1996). The relatively recent incorporation of the term "network" into supply chain management research represents an attempt to make the concept wider and more strategic by harnessing the resource potential of the network in a more effective way. An interesting classification of supply networks based on literature review distinguishes between tiered networks and learning networks (Cagliano et al., 1999). Tiered networks are supply networks organized in "tiers", that is levels of suppliers. Learning networks model has been proposed by Stuart et al. (1998) as opposed to tiered networks. In such form of networks, competencies, know-how, and experience on operational practices are shared among members, which often operate at different levels of the value chain. Extended Manufacturing Enterprises (EME) are defined by Busby and Fan (1993) as a way of combining manufacturing operations that is alternative both to pure market and to pure hierarchical structures. In the EME each company specialized in the production of specific goods or services continues operating divided in different firms, in distinct legal and organizational entities, even if collaborating to obtain a final product. Firms combine their activities for periods that greatly exceed the lead times associated to the specific transactions. Many conceptualizations have been provided concerning the virtual enterprise (e.g. Nakane and Hall, 1991; Davidow and Malone, 1992; Seidel and Mey, 1994; Merli and Saccani, 1994; Dean and Carrie, 1996). These organizational models are based on geographically dispersed modular units, which are specialized on particular competencies that are joined together for a definite period of time to satisfy specific market needs.

3.3 Customer Oriented Supply Chains

A third category of contributions can be identified, considering research focused on the integration between customer-oriented activities (i.e. marketing) and logistics issues. In fact previous theories tend to focus on how to obtain greater efficiency by leveraging on internal processes or on suppliers coordination. However, other authors state that relevant improvements can be obtained by introducing a customer-oriented perspective into supply chain management. Customer orientation is in fact a relevant topic also in Supply Chain Management literature and practice. As a matter of fact, attention regarding supply chain issues is actually shifting

Part I Chapter 1: Customer Orientation

towards new problems tied to the customer orientation of both single firms and overall supply chains. As market is ever more fragmented, it becomes more difficult to achieve significant improvements by means of internal process redesign. So some authors have provided new ways to cope with the ever growing variety of products demanded by customers. In these terms different approaches has been provided: from one side firms that are not able or do not desire to focus on each single customer, have tried to leverage on standardization but without affecting product variety, leading to the Mass Customization and Postponement concepts. From another side firms have tried to get closer to each single customer through the development of Customer Relationships. As a matter of fact, another relevant example of how customer orientation is becoming ever more relevant in SCM issues is Customer Relationship Management. These approaches, closed to marketing and product development issues, have deeply influenced many areas within supply chain management research, in particular they have contributed in changing the approaches of logistics and inventory management theories so leading to new paradigms in demand chain management. The next section describes these different contributions.

3.3.1 Mass Customization

This issue derives from the idea of gaining both the compatibility between the low cost and high efficiency of mass production, and the capacity of fulfilling the needs of the individual customer (Pine, 1993). This concept started its development during the ‘60s and ‘70s, when the idea of pure mass market begun to fade. Since then, more and more firms became aware of the trend towards market micro-segmentation and of strategic potential of offering custom products. This approach involves to design and manufacture products and services with increased options for customization. This situation needs to solve the trade- off between customization and operational efficiency. This typically shows up in the conflict between design, which aims at providing one customer with one product, and production, which tends to reduce product varieties. Typically firms that adopt this strategy face different problems, among which higher inventory and manufacturing costs, longer lead times, longer supplier delivery times, etc. In this perspective, companies search for higher process and logistic flexibility in order to deliver customizable products at low cost and short lead time. In order to face these issues firms rely on a set of approaches and practices that are intended at reducing the disruption induced in their operations by the rise of product differences.

Part I Chapter 1: Customer Orientation

The popularity of Mass Customization in the international business community is significant, since it aims to combine an agile customization of products with lean production efficiency within one supply chain (Anderson and Narus, 1995; Kahn, 1998). It breaks with the dilemma that one has to choose between two options: low volume — high variety and high volume — low variety (Gilmore and Pine, 1997; Kotha, 1995). Mass Customization is applied according to two main approaches: from one side firms can leverage on operations performances to make MC affordable, from another side, they focus on shifting customization along the supply chain. According to this last approach, the different ways in which MC is applied can be distinguished according to the supply chain level in which the customization takes place (Feitzinger and Lee, 1997).

In fact, if MC is applied by focusing on operations, we can identify a few approaches: 1. Material Processing: if the production system is flexible enough to produce the specific product a particular customer orders, the firm does not anticipate market demand but produces (or engineers) the product when order is received. 2. Increase range of stock: if the production system can not respond directly to demand, a stock is provided to guarantee product variety. 3. Assembly of core modules: by combining a set of standard parts a greater number of product variants is obtained. The basic idea is that by embedding modularity in a product architecture, it is possible to build product variants by combining the modules, provided that the way they are combined fulfils given constrains. In this way, with a reduced number of options a great number of products can be developed.

If MC is applied by shifting customisation along the supply chain: 1. Information content modification: if customisation takes place after production, during the selling process, the product is standard and not customised, while the message that describes the product is different for each customer. Rather than a different product or a self-tailored one, the standard offering is wrapped up specifically for each customer. In different situations the product may be promoted differently or its attributes and patterns are advertised in different ways. 2. Retailer provides service around product: customisation is provided by the retailer that provides different services that tend to adapt the single product to the customer needs. 3. Make products self-customizing: customisation may take place after the sale is over. This typically considers a standard product, designed so that users can alter it themselves. This practice is characterised by the fact that the standard product may be tailored or reconfigured directly by the customer to match its specific needs.

Part I Chapter 1: Customer Orientation

3.3.2 Postponement

A similar approach to Mass Customization can be found in the Postponement concept. Postponement is an organizational concept whereby some of the activities in the supply chain are not performed until customer orders are received. Companies can then finalize the output in accordance with customer preferences and even customize their products. Meanwhile, they can avoid building up inventories of finished goods in anticipation of future orders. Moreover, transportation between warehouses and factories can be avoided by shipping products directly to the customer rather than keeping them in stock. Postponement has been applied in many different industrial sectors. Relevant experiences can be traced back to the Fashion industry (i.e. Sports Obermeyer and Benetton), the car sector (i.e. MCC, a DaimlerChrysler car company and Volkswagen), personal computers (i.e. Dell and Hewlett Packard). Examples are also reported in the Wine, Chemical and Pharmaceutical industry (Van Hoek, 2001).

The Postponement concept was initially discussed by Alderson (1950), where it was observed that products tend to become differentiated as they approach the point of purchase, i.e., as it flows down the supply chain. Although this “differentiation” improves the marketability, the “manufacturability” of the products becomes more complex. However, by moving the “point of departure” for the differentiated product-models in the manufacturing process closer toward the point of purchase, benefits of consolidation could be exploited which reduces the complexity in manufacturing. This also helps to substantially truncate uncertainty and leads to a reduction in forecast errors by delaying the date at which one must make a production decision associated with a particular demand. The general notion of postponement has also been studied from different perspectives. For example, Zinn and Levy (1988) analyze the optimal location problem for a postponed inventory in a marketing channel. They employ the concept of speculative inventory introduced by Buklin (1965) to develop a framework of governance structures. Shapiro and Heskett (1985), Zinn and Bowersox (1988), Zinn (1990) and Fisher et al. (1994) present other applications of the concept in the logistics and distribution side of business. More recently, the problem has been analyzed as an integrative supply chain model by Lee and Billington (1992).

Part I Chapter 1: Customer Orientation

3.3.3 Customer Relationship Management

Mass Customization and Postponement tend to prepare the supply structure to an efficient response towards customers’ requirements. From a different perspective, many firms have also identified as relevant and critical the development of proper processes to manage the relationship with their customers.

Customer Relationship Management (CRM) is a business strategy designed to help an enterprise understand and anticipate the needs of its potential and current customers. CRM has its roots in relationship marketing which is based in turn on the formative work by Berry (1983), the IMP Group (e.g., Ford, 1990) and Christopher et al. (1991). CRM is defined by Couldwell (1998) as:

“… a combination of business process and technology that seeks to understand a company’s customers from the perspective of who they are, what they do, and what they’re like.”

CRM focuses on customer retention (Lockard, 1998; Deighton, 1998) and relationship development (Galbreath, 1998). Moreover, according to Kutner and Cripps (1997), CRM is founded on four relationship- based concepts: Customers should be managed as important assets. Customer profitability varies; not all customers are equally desirable. Customers vary in their needs, preferences, buying behavior and price sensitivity. By understanding customer drivers and customer profitability, companies can tailor their offerings to maximize the overall value of their customer portfolio.

From a technical point of view, customer data is captured in several different areas of the enterprise, stored in a central database, analyzed and distributed to key points (called touch points). Touch points can include a mobile sales force, inbound and outbound call centers, Web sites, point-of-sale, direct marketing channels, and any other parts of an enterprise that interact with the customer. The distributed data is intended to help foster effective, individual experiences between the company and the customer (Peppard, 2000).

In a sense, CRM is a natural and predictable extension of the evolution of marketing and sales. The first CRM-enabling technologies included basic contact management software

Part I Chapter 1: Customer Orientation

linked to individual PCs. This primitive form of Sales Force Automation (SFA) soon grew to include account management, opportunity management, mail merge, and forecasting. Client, product, marketing, and competitive information were eventually added to the mix. Now CRM projects strive to provide data to every enterprise department that touches the customer. The convergence of information towards a specific figure that interacts with the customer, provides an opportunity to supply the customers with a very specific service. Enterprises want their customers to see one, friendly, corporate face, as opposed to a collection of disconnected departments trying to work together. Ideally, an effective CRM strategy will enable the enterprise to use all of its resources when interfacing with a customer, including marketing, sales, finance and manufacturing, as well as post-sales services. When carefully and strategically employed, econometric, demographic, lifestyle data and decision-support systems can help promote effective CRM. We will not go into detail regarding this issue, however we refer to the previously cited works.

In the different issues related to Supply Chain Management, customer-orientation is in fact spreading out. Moreover literature regarding Supply Chain Management is showing the development of new paradigms, among which demand chain is one of the most relevant one.

3.4 Demand Chain Management

Demand Chain Management (DCM) can be defined as:

“Extending the view of operations from a single business unit or a company to the whole chain. Essentially, DCM is a set of practices aimed at managing and coordinating the whole demand chain, starting from the end customer and working backward to raw material suppliers. There are two fundamental objectives: (1) to develop synergy along the whole demand chain, and (2) to start with specific customer segments and meet their needs rather than focus on internal optimization”. (Vollmann et al., 2000)

The focus is clearly customer-centric, as defined early on by Brace, in explaining the concept of a demand chain as:

Part I Chapter 1: Customer Orientation

“…the whole manufacturing and distribution process may be seen as a sequence of events with but one end in view: it exists to serve the ultimate consumer”. Brace (1989)

DCM is a concept that draws on a large number of academic areas. Founded in operations management literature, DCM is concerned firstly with logistics. This is not a narrow subject but involves strategy across the whole value chain (Carothers and Adams, 1991; Langley and Holcomb, 1992; Shapiro et al., 1993). Secondly, theory for closely integrating operations between manufacturers, suppliers and customers comes from much of the process re-engineering literature (Frohlich and Westbrook, 2001): DCM may encompass just in time; product postponement; mass customization and the exploitation of third-party logistics. Thirdly, DCM must also be concerned with supply chain studies. As a matter of fact, the concepts of demand chain and DCM are essentially based on supply chain models, however the rapid uptake of technology, and in particular the Internet, has resulted in the evolution of the demand chain concept with a shift away from supply chains towards demand chains and DCM. In fact, DCM requires extensive up- and downstream integration between all business partners in order to succeed and these types of connections are favored by the Internet adoption (Frohlich and Westbrook, 2002). The main stimulus behind this has been the shift in power away from the supplier towards the customer (Soliman and Youssef, 2001). A common focus of most demand chain tools is combining supply chain information with analysis of customer interactions, transactions, and demands for goods/services to make more accurate sales and market demand forecasts (Zellen, 2001). In fact some authors provide evidence that this kind of approach can be effective both in traditional contexts (Hines et al., 2002; Williams et al., 2002) and in fast-growing environments (Heikkilä, 2002). However, DCM is generally associated to manufacturing systems since almost no evidence can actually be found of its application in service industries (Frohlich and Westbrook, 2002).

A relevant aspect that changes deeply when moving from the supply to the demand chain concept is that an inner customer perspective has to be adopted. This can be a tricky task since one of the most critical activity that has to be conducted when applying this shift is to identify properly the classes of customers that have to be managed so to design specific demand chains to provide them with the required solution. Literature has in fact devoted relevant attention to this topic, providing different contributions characterized by classifying products so to apply proper managerial approaches.

Part I Chapter 1: Customer Orientation

Table 2 summarizes these major contributions, however details regarding variables and cluster are not provided here, but we refer to each work.

Authors Classification variables Clusters Position of de-coupling point Make and ship to stock Process constraints Make-to-stock Hoekstra and Romme (1992) Product-market constraints Assemble-to-order Delivery service requirements Make-to-order Inventory cost considerations Purchase- and make-to-order System orders Annual sales revenue Customer inventory Annual unit volume Replenishment Fuller et al. (1993) Co-ordination requirements Rapid response Destination volume Nuts and bolts Handling characteristics Slow movers Customer order fulfillment interval Bulk cable Pure standardisation Segmented standardisation Lampel and Mintzberg (1996) Level of aggregation Customised standardisation Level of individualization Tailored customisation Pure customization Product innovation Demand volume stability Product life cycle duration Functional products Fisher (1997) Make-to-order lead time Innovative products Product variety End-of-sale mark down Product life cycle Product customisation Product variety Product value Full speculation Relative delivery time Logistics postponement Pagh and Cooper (1998) Delivery frequency Manufacturing postponement Uncertainty of demand Full postponement Economies of scale Manufacturing and logistics Lean Naylor et al. (1999) Cost, quality, lead time and service level Agile Stability of demand Leagile Innovative-unique and complex Product innovation Innovative-unique and non-complex Lamming et al. (2000) Product uniqueness Functional and complex Product complexity Functional and non-complex

Table 2: literature review regarding product categorization (adapted from Childerhouse et al., 2002)

The effects of this greater attention towards customer needs and preferences can also be identified by the evolution of many supply chain areas that are typically very far from customer’s perspective. A relevant example of this fact is for sure Inventory Management research.

3.5 Customer-oriented Inventory Management

In fact also the focus of logistics is increasingly turning towards providing better services for customers instead of minimising the total logistics costs or maximizing total profits of

Part I Chapter 1: Customer Orientation

supplier. This has lead to a growing interest on partnership and customer satisfaction issues. Papers regarding inventory management can be classified according to whether information is globally spread or locally managed and whether control is centralised or decentralised. Local information implies that each location sees demand only in the form of orders that arrive from the locations it directly supplies. Also, it has visibility of only its own inventory status and cost structure. Global information implies that the decision- maker has visibility of the demand, costs and inventory status of all the locations in the system. Centralised control implies that attempts are made to jointly optimise the entire system. Decentralised control implies that decisions are made independently by separate locations. This classification is summarised in Table 3.

Centralised Decentralised Control Control Global Echelon Stock Installation Stock Information Traditional Local Doesn’t make Replenishment Information any sense Systems Table 3: inventory management approaches

In particular, besides the traditional replenishment systems, based on simple Economic Order Quantity or Order-up-to policies, that consider the different locations in a supply chain as completely independent, the most discussed approaches are from one side Installation Stock Policies and, from the other, Echelon Stock ones. Both approaches share the idea of global information, in terms of knowledge regarding demand at each level of the supply chain and inventory positions for each location, however different assumptions are considered for what regards control. In particular, installation stock policies consider that control is decentralised so, even if each location has information regarding all the supply chain, they define inventory politics only considering their own objectives (both cost minimisation and profit maximisation). From the other side, echelon stock policies consider situations in which control of the entire supply chain is provided only by a single location that essentially pursue global objectives in terms of overall performances of the total supply chain. For a general description and review of these approaches we refer to Federgruen (1993), Axsäter (1993), Van Houtum et al. (1996), Diks et al. (1996), Silver et al. (1998).

In fact the most traditional quantitative framework for distribution network design is the cost minimisation approach (Lee, 1993; Hansen et al., 1992; Tyagi and Das, 1994).

Part I Chapter 1: Customer Orientation

According to Lee (1993) these models simultaneously locate a set of facilities and satisfy the demands of a given set of customers to minimise the total cost of location, operation and transportation. However, the problem of the cost minimisation framework is that it focuses the problem on the deliverer's point of view, and excludes the profitability of the customers. The focus of a more advanced distribution network design framework is on profit maximisation (e.g., Hakimi and Kuo, 1991). In the profit maximisation framework the costs of the distribution network are deducted from the customer's profits. No attention is paid to the customer's wishes, however, and therefore, they are not satisfied. Even if literature regarding these approaches is widely spread, few of them focus on customer service elements, besides penalty costs for stock-outs. Meshkat and Ballou (1996) and Canel and Khumawala (1996) introduce customer service elements in the distribution design problem, in addition to cost and profit information. Typically this, the so-called service-sensitive framework, includes elements like product availability, delivery time requirements and delivery frequencies. Korpela et al. (2001) propose a customer- service based approach to distribution logistics network design. The methodology proposed, in fact, tries to consider simultaneously the strategic importance of customer for a focal firm considered (achieved by means of AHP1 methodology) and the customer’s preferences for logistics service. In fact other authors have identified the need of major integration between pure supply chain management (in particular the logistics issues) and marketing related topics. Alvarado and Kotzab (2001) show that significant results can be obtained by evolving from a efficiency oriented supply chain towards a effective oriented one, represented, in this work, by Efficient Customer Response policies. Also Lambert and Cooper (2000) state that cross-functional integration is fundamental and marketing has to play a critical role. These authors, moreover, claim that main attention should be paid at first to inter- functional integration at each level of the supply chain and only after that focus should be paid to the vertical integration. Ellinger (2000) shows that there is significant relationship between Marketing/Logistics functional integration and operational performances in the supply chain. These authors provide some evidence that efforts to focus supply chain management towards customer needs can pay off in terms of improvements in performances. Other authors agree with these issues stating that, as customer service is largely provided by marketing and logistics, collaborative integration between these two functions is necessary to fully capitalize on potential service improvements (see Christopher, 1993; Bowersox et al., 1995; Mentzer and Kahn, 1996).

1 Analytic Hierarchy Process is a systematic procedure for representing the elements of any problem in the form of a hierarchy. A detailed description can be found in Saaty, 1990

Part I Chapter 1: Customer Orientation

However, integration can be conducted in different ways; Mentzer and Kahn (1996) claim that within an interdepartmental context a multidimensional perspective for integration must be considered. In fact integration can be conducted as interaction, in which attention is paid on information flows, or as collaboration, in which focus is on teamwork and share resources. These authors provide a framework to choose which kind of focus should be paid according to the specific problem faced. Table 4 summarises these results.

Stable product lines Complex products Stable markets Complex orders High interaction Available time Mission-critical items Lower uncertainty Key customer accounts Interdepartmental interaction Product launches New facility parameters Department-specific activities Low interaction Special customer orders Third-party logistics High uncertainty Short-term episode

Low collaboration High collaboration Interdepartmental collaboration

Table 4: processes influenced by interdepartmental interaction and collaboration

Part I Chapter 1: Customer Orientation

4 Conclusions

All contributions regarding SCM, DCM and customer-orientation show that this can be an effective approach when managing the growing turbulence of business contexts. In fact, all the information provided by more effective interaction with customers can lead to substantial benefits, even if actually few experiences provide evidence of this in planning activities. As a matter of fact, much of the literature has focused on marketing, logistics and R&D issue, and few works have devoted attention to the threats and opportunities in demand management. In this work we claim that attention should be paid to this issue. First the business context is becoming more and more complex and turbulent. Firms have to cope with higher competition, smaller product life cycles, and so on. This makes critical for them to use resources at their most and so being able to allocate efforts in the right directions. Moreover, attention has been primary paid on offering products more customised on consumer needs or providing better logistical services. This is usually developed by means of duplication of resources as, for example, inventories. This, given the rising competitiveness is gradually becoming an inefficient solution and firms are actually looking for alternative solutions to improve their ability in making supply meet demand in an uncertain world (Fisher et al. 1994). This leads to the need of new ways of providing internal market forecasts to increase the ability of foreseeing customer needs by leveraging on the demand generation process.

Chapter 2 Demand Forecasting and Customer Orientation

1 Demand Management

In the previous chapter we have identified that customer orientation is a relevant issue, but theory regarding supply chain management is in fact only starting to deal with this matter. In the supply chain related literature great importance is given to uncertainty management. Since focus is always given to the optimization of supply performances in complex networks, attention is usually paid on the sensitivity of proposed solutions. Adopting a specific solution that is very influenced by little environmental distortions make it completely useless, as often the expected results do not coincide with the effective ones. This generally leads to the problem of managing properly environmental uncertainty so to provide robust solutions.

Part I Chapter 2: Demand Forecasting and Customer Orientation

Generally uncertainty is a difficult issue to deal with, moreover it is complicated by the multidimensionality of the problem, as uncertainty can derive from many different aspects. Typically, when considering a manufacturing organization, it can be originated both from external and internal sources. Internal sources are, for example, production that due to possible failures or similar problems may influence the Total Delivery Lead Time. External ones constitute of actors that don’t belong to the specific firm considered. Among these typically we can consider suppliers, customers, competitors, and so on. In this work attention will be paid to uncertainty due to the demand faced by a specific firm. In particular, at first the demand generation process is introduced, then, in paragraph 2, attention is paid towards demand forecasting and a general overview of the major approaches is provided.

1.1 Demand Management Process

Literature has deeply focused on demand uncertainty and its effects on business performances. In particular, many different definitions have been provided to describe uncertainty (Knight, 1921, Galbraith, 1977). First of all turbulence and complexity make difficult to develop a model to understand reality. This makes much more complicated both identifying future scenarios and evaluating how much they are likely to occur. In the end, even if scenarios are identified, it is very difficult to understand which of these will effectively realize. When dealing with production and operations, this general concept can be defined by identifying where uncertainty arises (Van Donselaar, 1989, Murthy and Ma, 1991, Newman et al., 1993). Typically these sources are suppliers (for uncertainty on supply characteristics), production systems (in terms both of machine breakdowns and management of complex structures as job shops), control systems (when adopted model don’t fit the problem considered) and customers (in terms of future requirements and needs). The main focus of this work is on uncertainty due to customers and in particular we will focus on volume uncertainty. In these terms three main factors contribute to the definition of uncertainty. First of all, markets ask for reduced lead times; this typically creates a break between what markets want and internal throughput time. This has typically moved much production environments from an order driven planning towards a forecast driven one, so dealing with a higher degree of uncertainty (Higgins and Browne, 1992).

Part I Chapter 2: Demand Forecasting and Customer Orientation

Secondarily, demand variability has deeply risen in many industrial sectors. Customers requirements differ more and more and their preferences change very quickly due to mass communication. This makes demand at upper level always more difficult to be managed. Moreover, the supply chain is growing ever more complex; in particular, many firms face supply chains that are more heterogeneous and face very different customers all together, so making it difficult to have relevant clues on future needs for such different clients. All these factors tend to change dramatically demand patterns that, in many different industrial contexts, tend to be more and more variable; in particular often demand shows many periods with low or no demand and others in which demand shows tremendous amplifications. Authors have generally referred to this kind of demand as lumpy. Studies regarding demand uncertainty focus on many different issues. From one side, some authors study the impact of uncertainty on the manufacturing planning and control (MPC) system (e.g., Lee and Everett, 1986; Wemmerlöv, 1986; Lin and Krajewsky, 1993; Kadipasaoglu, 1995; Zhao et al., 1995; Wemmerlöv and Whybark, 1984; Ho, 1995; Ho et al, 2001). Other authors design and compare various methods to manage uncertainty (e.g. Wijngaard and Wortmann, 1985; Guerrero et al., 1986; Bassok and Akella, 1991; Murthy and Ma, 1991; White and White, 1992; Newman et al., 1993; Ebert and Lee, 1995).

When dealing with this kind of uncertainty there are 4 possible approaches that can be developed: first of all, one can redefine the decisional problem so to eliminate the source of uncertainty, for example by making the decision system faster; then, it is possible to influence demand so to control it and better managing it, for example by means of marketing campaigns; moreover, efforts can be paid in trying to reduce the effects of uncertainty by means of flexibility and slack introduction, for example inventories; at last, one can try to reduce uncertainty by exploiting better forecasts.

Figure 1 shows these approaches.

Redefine the decisional problem Since demand variability is rising and since delivery lead times are reducing, often firms have to deal with demand uncertainty at end products level, so encountering huge problems due to very high inventory levels. So typically firms are developing business process reengineering so to reduce overall throughput time. Therefore, one possible solution firms have to manage uncertainty is to redefine the problem, making the system more reactive to demand. Another possible solution is to reduce the number of sub-assemblies by

Part I Chapter 2: Demand Forecasting and Customer Orientation

leveraging on common components. In this way the requests for a single component rise, thus reducing the uncertainty effects (Collier, 1982; Baker et al., 1986; Eynan, 1996).

Problem redefinition - TLT reduction - Similarities adoption

Perceived Uncertainty Turbulence uncertainty Production effects Market Forecast system (forecast error)

Demand influence Forecast Reduce the effects of - Marketing activities - Better information uncertainty - Partnership - Better techniques - Slack - Flexibility

Figure 1: the demand management process (adapted from Bartezzaghi et al., 1999c)

Influence demand Another possible solution is based on considering that demand is, at a certain rate, controllable. Marketing activities and partnerships with customers can partially control demand and so let partially operate in situations characterized by lower uncertainty.

Reduce the effects of uncertainty Another possible solution is to cope better with a given uncertainty, by means of flexibility and slack introduction. Flexibility is the ability of a production system to react quickly and with low costs to variations in requests (Slack, 1983). However, if the system is not able to react significantly to demand variability, a possible alternative solution is to create slack so to partially isolate it. Slacks can be different and their adoption is influenced by the peculiarities of the problem considered: one possible solution is adopting time slacks, as for example safety lead times (Wemmerlöv and Whybark, 1984) or, typically, inventories. Even if these solutions are commonly adopted within firms, they are usually not effective when dealing with extremely variable situations. In these conditions, variability must be anticipated.

Reduce uncertainty If the previously introduced solutions are not effective, forecasts have to be produced; so uncertainty can be reduced by improving forecasts.

Part I Chapter 2: Demand Forecasting and Customer Orientation

1.2 Uncertainty Management: the case of lumpy demand

The focus of many authors has been on demand uncertainty and its effects on business performances, in particular when demand is lumpy. Lumpy demand can be defined as: variable, therefore characterized by relevant fluctuations (Wemmerlöv and Whybark, 1984; Wemmerlöv, 1986; Ho, 1995); sporadic, as demand series is characterized by many days with no demand at all (Ward, 1978; Williams, 1982; Van Donselaar, 1989; Fildes and Beard, 1992; Dar-El and Malmborg, 1991; Vereecke and Verstraeten, 1994); nervous, thus leading to show differences between successive demand observation, so implying that cross time correlation is low (Wemmerlöv and Whybark, 1984; Ho, 1995; Bartezzaghi and Verganti, 1995).

In these terms, lumpy demand is characterised by having many days with low or zero demand, and few with real demand peaks. Figure 2 shows an example of lumpy demand.

90 80 70 60 50 40 Demand 30 20 10 0 Time

Figure 2: example on lumpy demand

Literature has devoted attention to demand lumpiness in relation with different research areas. From one side, literature regarding supply chain has dealt with this issue by identifying sources of this phenomenon, possible organizational solutions and feasible supply chain design to deal with this problem. We refer to the previous chapter for details on contributions in this area and in particular to the works related to the Bullwhip Effect.

Part I Chapter 2: Demand Forecasting and Customer Orientation

From another side contributions have been provided regarding inventory management, in particular, by providing models that try to efficiently allocate inventories to deal with such a sporadic demand. Moreover contributions have also been provided by authors concerned with lumpy demand forecasting.

Table 1 provides some of the major contributions regarding this issue in current literature.

Area Work Year Notes Adoption of controlled partial shipments to Banerjee et al. 2001 Supply Chain reduce the effect of lumpiness Management Causes of demand amplification along the Lee et al. 1997 supply chain Adoption of genetic algorithms to define an Mak et al. 1999 optimal inventory control policy Mak and Lai 1995 Optimal replenishment policies Approximated method to define the reorder Janssen et al. 1998 point in a (R, s, Q) inventory model Inventory model based on both stock and Hollier et al. 1995 special deliveries Inventory and Vereecke and Production 1994 Inventory model based on Package Poisson Verstraeten Management Ward 1978 Regression model to evaluate order points Effect of lumpy demand on lot-sizing Ho 1995 performance Effect of lumpy demand on lot-sizing Ho et al. 2001 performance Wemmerlöv and 1984 Definition of lot-sizing policy Whybark Willemain et al. 1994 Evaluation of Croston’s method Bartezzaghi and Application of Order Overplanning to 1995 Verganti forecast lumpy demand Evaluation of performances of Order Verganti 1997 Overplanning when dealing with lumpy demand Johnston and 1996 Evaluation of Croston’s method Boylan Reduce demand uncertainty by extending Fildes and Beard 1992 the information base of forecasts Forecasting Effects of demand shapes on inventory Bartezzaghi et al. 1999b performance Zotteri and Comparison between decentralised and 2001 Verganti centralised forecasting Comparison among three different Bartezzaghi et al. 1999a forecasting techniques Van Donselaar et Adoption of Advance Demand Information to 2001 al. improve SC performances Syntetos and 2001 Modification of Croston’s technique Boylan Table 1: literature review of demand lumpiness in different areas of Operations Management

Part I Chapter 2: Demand Forecasting and Customer Orientation

Literature concerned with supply chain management regarding this issue has in fact essentially analysed the root causes of demand lumpiness. Essentially they can be due to two main factors. From one side the particular demand generation process considered may have some sporadic patterns. Think for example to the spare parts management problem, where typically demand appears at random, it is distributed among many different SKUs (usually 10,000s of different items) and, especially when products are in the end of their life, requirements are usually for few units. In this context demand typically tends to show a lumpy pattern. Alternatively consider products affected by sudden and quick changes in the final customer’s preferences and taste (e.g., in the fashion industry demand for a given colour can change dramatically from year to year). From another side another relevant source of lumpiness is the supply chain itself. This is typical in multi-echelon supply chains where the bullwhip effect shows up (we refer to paragraph 3.1 for details).

When dealing with inventory models in contexts with lumpy demand, different approaches are provided (see paragraph 3.5 in Chapter 1). Interestingly, however, all these approaches tend to share a common factor: they all consider demand as a known and probabilistic variable. In fact all these approaches consider a specific and known demand generation process, thus modelling demand through a probabilistic event, and essentially derive optimal inventory management politics given certain environmental conditions. Literature has devoted major attention to evaluate which statistical distribution fits better with lumpy demand. For example, regarding spare parts they are typically modelled according to a Poisson or Compound Poisson process1 (Friend, 1960; Hadley and Within, 1963). Other authors have analysed when other distributions may be adopted (Vereecke and Verstraeten, 1994).

However, when demand is both lumpy and uncertain, some kind of forecast has to be provided and so focus has to be paid on demand forecasting.

1 The Poisson or Compound Poisson process assumption considers that interarrival time between orders is distributed according to an Exponential distribution, while the number of parts ordered is distributed according to a different function.

Part I Chapter 2: Demand Forecasting and Customer Orientation

2 Demand Forecasting

“If a man gives no thought about what is distant, he will find sorrow near at hand”

Confucius

Forecasting is a relevant matter in many different management issues since it deeply influences overall performances. Organizations invest enormous amounts based on forecasts for new products, factories, retail outlets and contracts. In fact, poor forecasting can lead to disastrous decisions. Since the ‘60s, both researchers and practitioners have identified that forecasting is a relevant issue and have paid major attention to this topic; as a matter of fact, contributions have been given by many different research areas, from econometrics to operations management.

When dealing with manufacturing contexts, demand forecasting is for sure critical. Demand forecasting evaluates information that is used in the demand management process. According to the particular context these information can be adopted to plan production, to define materials or sub-assemblies procurement and to organize distribution, both in terms of material flows and stocks. As a matter of fact forecasting provides a simple input to the decisional process, but, according to the particular needs, the kind of information adopted changes. Consider, for example, a make-to-stock environment, in which, given a desired service level, inventories are evaluated by means of demand forecast and accuracy of this evaluation. On the contrary, firms adopting a make-to-order structure don’t need to fill finished products inventories, while they need forecasts to define material requirements and, sometimes, to allocate production capacity. Forecast accuracy has severe impacts on systems and, in particular, on manufacturing ones (Biggs and Campion, 1982; Vollman et al., 1992; Ritzman and King, 1993). As a matter of fact, this is a relevant issue also for practitioners that spend relevant efforts in gaining accurate predictions of future market requirements (Dalrymple, 1987; Sanders and Manrodt, 1994; Sanders, 1997; Rice, 1997; Wacker and Sprague, 1998; Klassen and Flores, 2001).

Part I Chapter 2: Demand Forecasting and Customer Orientation

Moreover today, the importance of forecasting is for sure much higher, since high inventory politics are not acceptable any more, both in terms of holding cost and in terms of market evolution. To cope with this ever more relevant issue, literature has provided different contributions that have evolved from very simple and easily implemental systems towards more complex solutions for coping with highly variable demand.

Traditionally two basic solutions have been considered: judgmental and quantitative forecasting. The main difference between the two approaches is tied to the fact that the forecasting model is more or less explicit. In particular, judgmental forecasting is based on the knowledge developed by people over time and so their experience and gut feeling play a primer role. This methodology, which typically bases on an inexplicit model, allows exploiting all types of information available, in particular the unstructured one. At the same time it has a major defect, due to the limitations of human capabilities, such as systematic biases and reduced number of variables that can be managed simultaneously (Sanders and Ritzman, 1992). Some studies (Makridakis, 1988; Webby and O’Connor, 1996) show that, at least when demand is stable, quantitative techniques provide better results than judgemental ones. However, people are able to consider a richer set of information compared to algorithms, in particular when experience and sensibility are relevant for the considered problem (Edmundson et al., 1988). Problems arise when the quantity of items that need to be forecasted grows, and so quantitative techniques tend to be preferable. Managers also agree in considering that quantitative techniques are more precise (Rice, 1997), but firms still use predominantly judgemental techniques at least to complete or correct statistical evaluations (Dalrymple, 1987; Sanders and Manrodt, 1994; Watson, 1996).

The fundamental principle of quantitative techniques consists in identifying some kind of link between past and future behaviour, or any relationship among observable variables and demand itself. As a consequence, when these data are not available and no information regarding demand is provided, these methodologies are not applicable. Any quantitative technique relies on data, so these have to be accurate and provided quickly, otherwise any technique will be inefficient. This is often a problem, in particular when dealing with operations (Fisher et al. 2000). However, sometimes the best solution resides in the application of both quantitative and judgemental techniques together (Carbone and Gorr, 1985; Remus et al., 1996; Sanders, 1997).

Part I Chapter 2: Demand Forecasting and Customer Orientation

For what regards quantitative techniques many different contributions can be found in current literature, however given the objectives of these work, these systems can be grouped according to two main dimensions.

First of all we will consider demand variability; in particular we can separate techniques that need a relatively stable demand from those expressively studied to apply with highly variable or uncertain demand, in particular for lumpy demand. From the other side, we can separate forecasting approaches according to which aspect of the demand forecasting process they focus on. As a matter of fact, each forecasting model is characterized by adopting some kind of data (i.e. past orders) and to apply some kind of technique (i.e. moving average). In fact, to apply properly forecasting, one has first to collect data from market and then use these data properly to produce a forecast. In this way forecast accuracy can be influenced first from how effective the chosen technique is to manage the specific problem, but also from the quality and sort of data provided. This idea is exemplified in Figure 3.

Retrieved Perceived Demand Turbulence information Uncertainty generation Information Forecasting process retrieval algorithm (Forecasting error)

Figure 3: demand forecasting process

Through this framework we can divide forecasting techniques in three main groups (as Figure 4 shows).

First we can consider methods adopted to manage stable demand that rely only on demand data. Among these we can find the “classical” and most known forecasting techniques as Moving Average, Exponential Smoothing, ARIMA, and so on. As further described, these techniques are more concentrated on developing proper forecasting algorithms that rely essentially on statistical evaluations of past demand data. A second major group tries to face the problem of dealing with lumpy demand by leveraging again on past demand data. Here we can find techniques that essentially

Part I Chapter 2: Demand Forecasting and Customer Orientation

try to reduce the problems that traditional techniques face when dealing with highly variable requirements. In the end there is third group that leverages on the information retrieval process regarding how demand is generated. As further illustrated these techniques try to solve the issue of lumpy demand forecasting by focusing more on organizational issues than on statistical and econometrical ones.

2.1

Stable Demand

2.2 2.3

Lumpy

Demand Variability Demand Variability Demand

Demand Demand Data Generation Process Source of Information Figure 4: structure of demand forecasting approaches

In this chapter we will at first take a brief look to methods that rely only on demand data and show techniques applied first when demand is rather stable (cell 2.1) and then when demand becomes lumpy (cell 2.2). Then we will concentrate on other techniques studied for lumpy demand and that focus on information regarding the demand generation process (cell 2.3).

2.1 Methods for Stable Demand based on Demand Data

There are many forecasting methods that belong to this quantitative class. In fact the literature on forecasting methods is so voluminous that methods cannot be all recalled here. So this part is meant to give a general overview of these approaches thus focusing on their common characteristics. Among these techniques we can consider: Smoothing and Averages based techniques. Bivariate and Multivariate models. ARIMA methods.

Part I Chapter 2: Demand Forecasting and Customer Orientation

2.1.1 Smoothing and Average based techniques

Many common forecasting methods come under the general category of Moving Averages. Since most time-series experience variations in demand (called error or noise), better estimates are generally possible by smoothing the history through averaging. These techniques, in fact, average or smooth past data in some way. Among these techniques we can find Simple Averages, Simple Moving Averages and Weighted Moving Averages, that essentially evaluate plain averages of past data so to reduce the effect of demand variability (we refer to Makridakis et al., 1983, for a detailed description of these techniques). One of the best-developed moving-average technique is the X-11 version of the Census Method procedure, that manages different kinds of moving averages to properly estimate trend, seasonal, cyclical components (we refer to Croxton and Cowden, 1955; Dagum, 1980; Makridakis et al., 1983, for a detailed analysis of the approach).

Another highly adopted technique is Exponential Smoothing. The use of this technique in inventory forecasting was introduced by Brown (1959); this method is a moving-average forecast which requires minimal data storage. Essentially, exponential smoothing is a moving average with declining weights for past actual values. The weights are set so that they decrease back in time, thus reducing influence of the oldest demand occurrences. This simple approach has been developed by defining other Multiple Exponential Smoothing methods. In fact, because the simple exponential smoothing model lags when there is trend and seasonality, higher-order models have been developed and used. The most popular approaches of higher-order models have been developed by Brown (1963), Holt and Winters (1960).

2.1.2 Bivariate and Multivariate models

Many series have patterns correlated with time or with some other series. These correlations may yield mathematical curves or relationships that can be used for forecasting. The most commonly used statistical approach to curve fitting involves regression by maximum-likelihood criterion. Two types of least-squares curve fitting are commonly used in forecasting systems. In one type, called autoregression, the predicted value is related to past values of another variable. In the other, the predicted value is related to time. These techniques are widely adopted in many different areas; we refer to

Part I Chapter 2: Demand Forecasting and Customer Orientation

Draper and Smith (1981), Pindyck and Rubinfeld (1991) and Makridakis et al. (1983) for a detailed description of these methods.

2.1.3 ARIMA methods

In general, the techniques described so far have just assumed a model and then analyzed the fit of the time series to the model. Box and Jenkins (1976) have developed a unified forecasting system that includes the empirical determination of not only model parameters but also a model. The Box-Jenkins philosophy is that forecasting models should be based on summary statistics of the historical data rather than unverified assumptions (e.g., pre- selected models). Consequently, these models are derived empirically, that is the models are derived through systematic procedures of identification, estimation and diagnosis. The methods developed by Box and Jenkins are commonly called ARIMA (Autoregressive Integrated Moving Average). We refer to Box and Jenkins (1976), Box et al. (1994) and De Lurgio (1997) for a detailed presentation of the methodology.

The basic approach to quantitative forecasting has been the use of historical demand data with extrapolative techniques, following an evolutionary path from the moving average and the exponential smoothing to the ARIMA techniques. The common factor of these methods is the use of historical data without analyzing the demand generation process and the root causes of its variability. Typically this approach appears suitable for regular patterns of demand or in case of systematic variability, such as trends and seasonality, that can be managed on the basis of historical data. The main difference between exponential smoothing and moving average is that the first considers all past periods, without limiting arbitrarily any time horizon, and that weight associated to each data is smoothed exponentially and attributed arbitrarily, as moving average does. Autoregressive methods, instead, rely on links between observations in a single time series between different periods.

Literature has also provided many possible alternatives of these methods that model many different sources of variability, as, for example, seasonality, trend, and so on (we refer to Gould et al, 1993; Makridakis and Wheelwright, 1989). However, since all these techniques rely on regularities of past observations, they guarantee good performances when demand is stationary. When demand variability rises,

Part I Chapter 2: Demand Forecasting and Customer Orientation

in particular when lumpy demand shows up, these techniques tend to perform poorly (Bartezzaghi et al., 1999a). The inadequacy of all these techniques can be identified when applying these systems on time series with many periods with no demand or with long intervals between orders. Figures 7 and 8 compare how exponential smoothing behaves when applied on regular demand and on lumpy pattern.

Demand Smoothing MAD% = 57% 6

0 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49

Figure 5: exponential smoothing applied to stable demand

Demand Smoothing

90 MAD% = 190% 80 70 60 50 40 30 20 10 0 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49

Figure 6: exponential smoothing applied to lumpy demand

Part I Chapter 2: Demand Forecasting and Customer Orientation

This lumpy pattern implies that, given a certain forecast, when the order arrives the forecast is updated to a higher value, but, given the sporadic behaviour, the forecast is much bigger than actual demand, thus rising forecast error. Moreover, after a few periods of no demand, in which forecast reduces due to the update procedure, a new order arrives, and again the forecast is completely biased. This requires the adoption of different techniques studied specifically for dealing with such a variable demand.

2.2 Methods for Lumpy Demand based on Demand Data

As previously introduced, when dealing with lumpy demand, typical smoothing or autoregressive techniques tend to be inadequate. This phenomenon isn’t new at all, in fact many business realities face this problem, from spare parts to fashion products (Cachon and Fisher, 1997; Fisher et al., 1994; Brandolese and Cigolini, 1999). The rising diffusion of this problem, his economical relevance, and his managerial complexity have forced researchers to focus on this topic. As a matter of fact only few approaches have been proposed to properly forecast lumpy demand. Here we will describe the major approaches developed, that essentially base on Croston’s approach.

2.2.1 Croston Method

Croston (1972) proved the inappropriateness of exponential smoothing as a forecasting method when dealing with intermittent demands and he expressed in a quantitative form the bias associated with the use of this method when demand appears at random with many time periods showing no demand at all. He first assumes deterministic demand of magnitude µ occurring every p review intervals.

Subsequently, the demand Yt is represented by:

µ, t = np +1 Yt =  0, otherwise where n = 0, 1, 2, … and p ≥ 1. Conventional exponential smoothing updates estimates every review period whether or not demand occurs during this period.

Part I Chapter 2: Demand Forecasting and Customer Orientation

' In particular, if we are forecasting one period ahead, Yt , the forecast of demand made in period t, is given by: ' ' ' Yt = Yt−1 +αet = αYt + (1−α)Yt−1

where α is the smoothing constant value used, 0 ≤α ≤1, and et the forecast error in period t. Under these assumptions if demand estimates are updated only when demand occurs, the expected demand estimate per time period is not µ p , but rather:

µ pα µα E(Y ' ) = = t p 1− (1−α) p 1− β p where β = 1 – α.

Croston then refers to a stochastic model of arrival and size of demand assuming that 2 demand sizes zt are normally distributed, N(µ, σ ) and that demand is random and has a Bernoulli probability 1 p of occurring in every review period. Under these conditions the expected demand per unit time period is:

µ E(Y ) = t p

In this case, Croston showed that when demand estimates are updated every period using exponential smoothing, then the mean demand estimate equals the population expected value, and the variance of the demand estimates equals:

2 '  p −1 2 σ  α Var(Yt ) =  2 µ +   p p  2 −α

If we isolate, though, the estimation that are made after demand occurs, Croston shows that these estimates have the biased expected value:

' E(Yi ) = µ(α + β p)

Part I Chapter 2: Demand Forecasting and Customer Orientation

The error, expressed as a percentage of the average demand, is shown to be 100α(p – 1) and reveals an increase in estimation error produced by the Bernoulli arrival of orders as compared with constant inter-arrival intervals.

Croston, assuming the above stochastic model of arrival and size of demand, introduced a new method for characterizing the demand per period by modelling the demand in one period from constituent events. According to his method, separate exponential smoothing estimates of the average size of the demand and the average interval between demand incidences are made after demand occurs. If no demand occurs the estimates remain the same.

If we let: ' pt = the exponentially smoothed inter-demand interval, updated only if demand occurs in period t so that: ' E( pt ) = E( pt ) = pt

' and zt = the exponentially smoothed size of demand, updated only if demand occurs in period t so that: ' E(zt ) = E(zt ) = z

' Then the forecast, Yt for the next time period is given by:

' ' zt Yt = ' pt and so the expected estimate of demand per period in that case would be:

 z '  E(z ' ) µ E(Y ' ) = E t  = t = t  '  '  pt  E( pt ) p

Figure 7 compares demand and Croston’s forecast. It can be noted that in fact forecast, at least when demand shows up, tends to be more accurate, compared to exponential smoothing (see Figure 6).

Part I Chapter 2: Demand Forecasting and Customer Orientation

Domanda Smoothing

90 80 70 60 50 40 30 20 10 0 1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49

Figure 7: Exponential smoothing when update is done only when demand shows up

Given this basic method, other authors have provided other similar approaches that essentially modify this solution to reduce other sources of bias (Sweet, 1980).

Johnston and Boylan (1996) propose a model named Size Interval Method, that, as Croston’s one, decomposes demand in two sources, size and interarrival between two orders. These authors share Croston’s approach, but show that this method evaluates only the average value of demand, without considering the estimation of demand variability. However, planning requires information regarding demand variability so to set-up properly inventories. This forced Johnston and Boylan to provide a system to evaluate both mean and variance under the hypothesis that demand follows a Poisson distribution, which usually seems to be applicable. Even if this method provides a better evaluation of orders size distribution, it is not able to evaluate properly interarrival time variability.

2.2.2 Syntetos and Boylan Method

Syntetos and Boylan (2001) propose a modification to the Croston’s procedure since they prove that Croston’s evaluation is biased. In fact, the empirical evidence suggests that Croston's method's theoretical superiority is not reflected in the forecasting accuracy associated with the use of this method. Subsequently, in an effort to identify the causes of this unexpected forecasting

Part I Chapter 2: Demand Forecasting and Customer Orientation

performance, a mistake was found in Croston’s mathematical derivation of the expected estimate of demand per time period. Assuming that order sizes and intervals are independent it states the following:

 z '   1  E t  = E z ' E   '  ()t  '   pt   pt  but  1  1 E  ≠  '  '  pt  E()pt

It can be proved (we refer to Syntetos and Boylan, 2001 for details) that the expected demand per time period for a smoothing constant α = 1, is:

 z '   1   1  1  E t  = E z ' E  = µ − log  .  '  ()t  '      pt   pt   p −1  p 

So if, for example, the average size of demand when it occurs is µ = 6, and the average inter-demand interval is p = 3, the average estimated demand per time period using Croston's method (for α = 1) is µ p = 6/3 = 2 but it is 6 *0.549 = 3.295 (64.75% bias implicitly incorporated in Croston's estimate). The maximum bias over all possible smoothing parameters is given by:

 1  1  µ µ − log  −  p −1  p  p

This is obtained for α = 1. For realistic values, the magnitude of the error is much smaller. Since Croston’s method fails to produce unbiased demand per period estimates, the authors propose a different estimation that in fact equals µ p . Such an expectation is:

  ' '  1  µ E()Yt = E ()zt E ' =  ' pt −1  p  pt c  where c is a constant.

Part I Chapter 2: Demand Forecasting and Customer Orientation

The described methods are capable of better forecasting performances (Willemain et al. 1994; Sani and Kingsman, 1997), compared to traditional smoothing techniques, when demand shows a lumpy pattern. This is due to the fact that they adapt themselves to the sporadic nature of lumpy demand. However, their criticality is that they don’t provide accurate predictions of when demand will occur, but rather they focus more on the expected quantity. This can be a relevant problem when demand shows a very sporadic pattern or when lumpiness is very high, since these systems push inventory to grow up even if demand will occur very far in time and so rising total holding costs.

These problems can be reduced by adopting different techniques based not on past demand data (or at least not only), but that focus more on the demand generation process and that try to estimate future market requirements by looking at information linked to demand variability.

2.3 Methods for Lumpy Demand based on Information regarding Demand Generation Process

The approaches since here described essentially focus on the forecasting algorithm. In all these methods sales data (rarely demand data) are considered as given. This approach is effective when dealing with demand of products characterized by a long life cycle and relatively stable requests. It is not adequate when dealing with highly irregular demand and products with very short life cycles. These approaches, as a matter of fact, face poor information regarding demand and so cannot apply at all. So, it appears to be more relevant to focus more on the sources of information, so exploit something more than demand data only. This idea has been developed in the literature regarding information in business processes (e.g., Azoury, 1985; Lovejoy, 1990); here information is considered as a control variable and so firms can manage (at least partially) the kind of information they require and its quality. A second relevant factor is considering that information retrieval and its use into a forecasting procedure are very tight dimensions; so the forecasting process can be adapted to use information that will be available in the future, and, from the other side, the information retrieval process can be adapted to effectively extract the necessary information.

Part I Chapter 2: Demand Forecasting and Customer Orientation

We can classify these kinds of techniques according to the perspective they adopt in the information retrieval process. As a matter of fact a first group of methods has a product-oriented focus. These techniques, in fact, pay attention to product characteristics and commonalities, and try to leverage on these sources of information to better foresee future market requirements. Among these techniques we can consider the so-called Analysis of Reliability. Close to this kind of approach we can also consider Lost Sales Estimation, even if these technique deeply leverages on demand data. A different perspective is adopted by other techniques that share with the previous ones the focus on demand generation process; however the methods tend to look deeper in the sources of demand variability by focusing directly on customers. Among these techniques we can consider Order Overplanning and Multi Level Supply Control. Some techniques try also to analyze demand variability by means of looking directly towards customer, but leveraging only on demand data. Among these techniques we can consider Early Sales.

So the main difference between these two groups of methods is that they look at different aggregation levels of information focusing respectively on products and on customers (Figure 8 exemplifies this matter).

Traditional Methods Lost Sales Estimation

Products Croston, Analysis of Syntetos and Reliability Boylan, etc Order Early Overplanning

FOCUS ON Sales

Customers Customers Multi Level Supply Control

Demand Data Demand generation Process

SOURCE OF INFORMATION

Figure 8: forecasting approaches classified according to focus and source of information

Part I Chapter 2: Demand Forecasting and Customer Orientation

Among these approaches it also relevant to consider the so-called Collaborative Planning, Forecasting and Replenishment (CPFR), which in fact relies on exploiting information at customer level and regarding the demand generation process. However, since this kind of approach is not a real technique but rather a particular way to structure the overall forecasting process, we argue that it should not be considered in this framework.

2.3.1 Lost Sales Estimation

In the retail sector final demand is typically characterized by a certain stability. So, the assumption of classic forecasting techniques (e.g., exponential smoothing) is often verified, so stating that these techniques can apply properly. However, the implementation of these techniques has shown some problems due to the kind of data adopted to evaluate forecasts. Forecasting algorithms should be provided with demand data, but, since this is rarely known due to stock-outs, sales data are adopted as proxy. Stock-outs reduce to zero sales for the out-of-stock product, but, on the other side, positively influence similar products demand. These noises bias information on past demand making measurable only sales and shipping. In this way it is necessary to understand demand patterns by cleaning sales data from the stock outs for both the considered product and its similar ones. These problems have stimulated the development of techniques that use sales data to estimate demand (Wecker, 1978; Bell, 1981; Agarwal and Smith, 1996). These contributions, however, have always assumed a periodic review of stock levels. This make the problem very complex, because one ignores the exact moment in which a stock out occurs and so the period along which sales are distributed is unknown. A possible solution to this problem is given by the adoption of data that makes possible to monitor continuously demand and stocks (Raman and Zotteri, 1998). In many retail chains POS data are collected by means of bar codes. These precious data on the single transaction make possible to comprehend exactly when stock finish and so this makes the estimation of demand more precise.

Part I Chapter 2: Demand Forecasting and Customer Orientation

2.3.2 Analysis of Reliability

These techniques rely on the evaluation of causal relationships between technical information and market requirements. This technique is properly studied to deal with spare parts or similar markets, since it is intuitively logical to link failure probability for a component to its requirements. One of the most known models in this area is the Yamashina’s Model (Yamashina, 1989).

Yamashina’s Model

Yamashina proposes a model that focuses of reliability of components in order to derive the number of parts that will be requested during a certain period of time. The main variable the model considers is the failure probability defined as:

t Fi ()t = fi (τ )dτ ∫0 where:

Fi(t) is the probability distribution of the components life fi(t) is the probability density function of the components life

If parts are produced according to a production function p(t) per time unit, total production till time L is equal to: L P = p()t dt ∫0

The number of parts still operating at time t is equal to:

t U ()t = p (τ )R (t −τ )dτ ∫0 where R(t) is the probability that a product is still operating at time t and is defined according to the probability density function of the component m(t), as follows:

t R()t =1− m (τ )dτ ∫0

Part I Chapter 2: Demand Forecasting and Customer Orientation

The number of parts required during a period of time equal to dt, is Di(t)dt and can be calculated as follows. th Consider di,m(t)dt is the probability that a components is working and that the m repair of the component has been done during a short time interval dt. Then:

t d p,m ()t dt = di,m−1 (τ )fi (t −τ ){}R ()t R (τ )dτdt ∫0 where :

di,1(t) = fi (t)R(t) Define: ∞ di ()t = ∑di,m ()t m=1 the probability that part i is repaired during a time interval dt So, t Di ()t dt = p (τ )di (t −τ )dτdt ∫0

The number of spare parts for item I, request till time t is equal to:

t Si ()t = Di ()t dt ∫0

According to the distribution of production, the system is capable of evaluating how many parts will have to be repaired in a certain time interval.

Common characteristic of this sort of approaches is that some kind of technical analysis on parts has to be conducted, to evaluate the variables previously introduced. This operation is often complex and costly, moreover when items are many.

2.3.3 Early Sales

The techniques called Early Sales are characterized by evaluating forecasts of total demand during a certain period (e.g., one year or one season in the fashion industry), by means of a partial observation of demand. This is done by leveraging on the fact that a certain heterogeneity in the Delivery Lead Time (DLT) exists among different customers.

Part I Chapter 2: Demand Forecasting and Customer Orientation

Typically these techniques use data as: pattern of demand in analogous periods during past years; orders that have been already collected.

In Figure 9 is reported the scheme according which forecasts are evaluated. Total demand for period t’ is deducted using historical observations (its shape depends on the relative numerousness of customers with different DLT). Given this information, algorithms are developed to evaluate total demand forecast.

Total quantity Forecast Cumulated demand curve •

Algorithm DLT < k

DLT = k

time t - k t t - k t Figure 9: the Early Sales general approach

Assume to know the cumulative demand curve for a product that is very similar to the one you are trying to forecast, and some orders have already been received. By means of the information on actual orders, the cumulative curve placed on the left of Figure 9 can be constructed. By comparing this curve with the one for the old product, one can estimate the curve for the new product. According to the chosen algorithm, different hypothesis are requested to apply this method. Typically these are: 1) orders acquisition must occur similarly for the different products, so to make consistent using past profile for the new demand 2) high correlation between customers behaviour, so that orders achieved form the first buyers (so-called early buyers) represent purchase intentions of the late buyers. This method is build for markets where there is a high innovation between different seasons, so that the specific firm is never in stationary conditions. Typically fashion industry is a good example of such a situation (for an application of this solution we refer to Fisher et al., 1994).

Part I Chapter 2: Demand Forecasting and Customer Orientation

Here are described two models, developed respectively by Kekre, Morton and Smunt (1990), and Fisher and Raman (1992).

Kekre, Morton and Smunt model

In Kekre et al. (1990) two simple algorithms are presented to adopt Early Sales. These algorithms are applicable when customers can be clustered into those with fixed DLT and those with a DLT that can vary between 0 and L.

We can define: t current period

Dt Total demand for period t

Dt-k, t Orders due for t with a DLT equal to k, so orders that are expected for t-k k Dt Total demand for t knwon k periods in advance k * Dt Demand for t not yet achieved during t-k

St Simple moving average of total demand j St Demand forecast for Dt+j for period t + j available during period t

In these conditions, total demand Dt for period t is given by the sum of the orders by customers with DLT between 0 and L (see Figure 10):

L Dt = ∑ Dt−k,t k=0

t - k t t + j

Figure 10: the time indexes considered

Demand for period t, know at t-k, is given by:

L−k k Dt = ∑ Dt−h,t h=0

The meaning of all these quantities is explained in Figure 11.

Part I Chapter 2: Demand Forecasting and Customer Orientation

quantity

k* Dt

Dt-k,t

k Dt

time t - L t - (k + 1) t - k t

Figure 11: the Kekre, Morton and Smunt model

According to the different situations, the authors propose two possible alternative algorithms.

Multiplier algorithm

As the authors state:

“I've sold 30 and the season is 20% complete; (30/0.2) makes a forecast of 150”

First one evaluates the vector of the so-called current scale up factors, which generic element is given by the ratio of actual total demand Dt and how much of it had been effectively ordered. in t – k. D γ k = t ∀k = 1,2,..., L t D k t

Coefficients are then smoothed considering their homologous values respect to periods before t: k k k ψ t = (1−α)⋅ψ t−1 +α ⋅γ t

Part I Chapter 2: Demand Forecasting and Customer Orientation

j Forecast for period t + j is obtained by multiplying the known part of demand at t (Dt+j ) and the smoothed scale factor for j periods:

j k − j j St =ψ t ⋅ Dt+ j

k The algorithm considers then that the ratio Dt Dt is a relatively stable quantity for the k considered market; the current quantity Dt is amplified by means of coefficients obtained from the past cumulative curve. This means that one can consider the hypothesis that the demand curve is “constant” over time and that correlation between customers is stable.

Additive algorithm

j * This algorithm provides a forecast for Dt+j , the part of unknown part of demand for period t + j. Forecast is evaluated according to the observation of demand Dt till current t. k For each t period, a portion Dt * of demand Dt was not known during period t – k; this part is smoothed on past observations during all t periods. This gains the average part of unknown demand for each possible k between 0 and L. j Demand forecast for t + j is obtained by adding the know part of demand Dt+j and the smoothed part of the unknown one: j k − j* j St = Ft + Dt+ j

In this situation one assumes that the unknown part of demand is constant. This portion is not influenced by the known part of demand, but it simply adds. There is no assumption regarding correlation between clusters of customers in terms of DLT, but a correlation in terms of behaviour on different periods. The authors suggest the use of this algorithm instead of simple exponential smoothing, when demand is irregular, but a part of it is known. As a matter of fact the system can gain very good performances when lumpiness is due to a sudden change but that does not affect the all planning horizon. When causal variations occur, the algorithms show the same problems in identifying illusionary correlations, as pure time series models do. When variability is very high, one can adopt mixed methods where the stable part of demand is estimated by means of Early Sales, while the abnormal part is managed through specific methods, based on a greater interaction between customer and supplier or on judgemental evaluations.

Part I Chapter 2: Demand Forecasting and Customer Orientation

Fisher and Raman Model

Fisher and Raman (1992) provide a complete forecasting and planning model. They consider the fashion market characterized by very long lead times. The actual approach in the fashion industry is the Quick Response: it permits to obtain a partial parallelization of production and sales, which permits to adopt early sales to plan purchasing and production. Given the partial observation of demand, Fisher and Raman propose a model that evaluates demand distribution and evaluates production orders considering capacity bounds and Economic Order Quantity. The model evaluated demand distributions by means of: sales forecasts provided by experts; typical bias introduced by experts; early sales; correlations between early sales and total demand.

The system adopts a similar logic as Kekre and Morton’s multiplier algorithm, however, k k instead of a simple amplification coefficient of Dt , the joint distribution of Dt e Dt is evaluated, so the joint distribution of partial observed and total demand. The distribution is derived as follows: average demand is given by the average forecasts provided by a panel of experts; demand variance is obtained by multiplying the variance of the estimates and a reduction coefficient. The coefficient adopted to reduce the standard deviation of experts’ estimates considers the duplicity of uncertainty: it depends on the intrinsic uncertainty of the demand generation process, but also on that due to the bias introduces by the forecasters in estimation process. The first kind of uncertainty is the one that the system wants to evaluate, while the other provides a larger dispersion that should be eliminated; k correlation coefficient between Dt and Dt is considered constant among different seasons and is estimated by means of past data; k the average Dt is estimated on historical data too;

The method also considers an inventory management model that uses the degree of demand uncertainty to define the production quantity. The hypothesis that this model considered are the existence of a constant correlation among customers and that forecasts error is distributed similarly over time.

Part I Chapter 2: Demand Forecasting and Customer Orientation

This system is interesting since it considers the distinction between intrinsic uncertainty and estimated uncertainty. Moreover, compared to Kekre and Morton’s algorithm, it tends to be closer to reality. However, the model is more complex and requires the evaluation of a larger number of parameters though historical data. However, other authors (e.g., Zhao et al., 2001) show that early order commitments can substantially improve forecasting performances.

2.3.4 Order Overplanning

Another technique that focuses on customer information to deal with lumpy demand in the Order Overplanning (Verganti, 1997; Bartezzaghi et al. 1999). This technique considers to the total demand for an MPS unit as the aggregation of the product unit requests for single customers. In particular, each customer demand may be described as a certain distribution f (D ) (for example Verganti, 1997, considers a binomial distribution), given the number of customers N, one can easily evaluate the distribution of product units. In no additional information is provided on purchasing intentions of customers, the best forecast of total demand is given by the average of the introduced distribution; however, order overplanning is able to collect and exploit extra information on future requirements. By focusing on the purchasing process of each customer, sales tries to capture subjective information on possible requirements that will occur in a given time period t. This early information can be used to separate customers in two groups: those customers that are expected to place an order (likely buyers) and those that seem not intended to place any request (unlikely buyers). Figure 12 exemplifies the forecasting process.

Consider the following notations and assumption: L is the number of likely buyers l is the probability that a likely buyer actually places an order u is the probability that an unlikely buyer places an unexpected order each customer demand follows a Binomial distribution B(d, 1 – d), where d is the probability that a customer places an order the forecasting capabilities of Sales can be model according to two binomial distributions B(l, 1 – l) and B(u, 1- u), given the early information collected at t – (TLT + 1)

Part I Chapter 2: Demand Forecasting and Customer Orientation

1) 2) 0,5 Likely Actual buyers orders L 0 012 unexpected orders Potential customers 0,9 N

0 0123 Customers Unlikely not placing buyers any request U

Information Demand Actual collection forecast demand

Figure 12: the Order Overplanning approach

Since likely buyers are only expected to purchase, some of them will probably not place any order. So, the total number AL of actual orders received form the L likely buyers is a random variable with a probability distribution equal to:

 L  AL L−AL f ()AU = l (1− l )    AL  with an average equal to:

AL = Ll

However, some on the unlikely buyer will in the end, place an unexpected order. So the total number AU of unexpected orders is distributed according to:

N −L−A  N − L AU U   f ()AU = u (1− u )    AU  with an average equal to:

AU = (N − L)u

Part I Chapter 2: Demand Forecasting and Customer Orientation

From this procedure the total number of orders can be simply evaluated as:

AL + AU

Given the particular forecasting procedure order overplanning adopts, two main dimensions influence its applicability. First of the system is suitable when the amount of early information collected at single customer level in considerably higher than the information obtained by looking to total demand. Then a relevant variable for its application is the number of customers N. This is due to two main reasons. First of all, N is a major driver of lumpiness (see Bartezzaghi et al. 1999a) and so as N rises, demand lumpiness reduces making order overplanning less effective. Moreover, as N rises, organizational costs for exploiting early information rise too. This in fact limits the application of this technique to contexts where customers are known and it is possible to have good hints on single customer’s purchasing intentions. Figure 13 shows the effect of these two dimensions on order overplanning domain of application.

complete 1 information 0,9 0,8 Order Overplanning 0,7 0,6 EI 0,5 0,4 Part-oriented methods 0,3 0,2 no extra 0,1 information 0

1 10 20 30 40 50 60 70 80 90 100

potential customers (N )

Figure 13: comparison between Order Overplanning and Part Oriented Methods

Part I Chapter 2: Demand Forecasting and Customer Orientation

2.3.5 Multi Level Supply Control

A similar approach to Order Overplanning is Multi Level Supply Control, since in fact they share the common idea of providing the forecaster with some kind of early information regarding market future requirements. Multi Level Supply Control is based on the idea that a manufacturer is provided in an early stage with Advance Demand Information on an aggregate level, which is higher than the item level. Later the manufacturer is provided with the demand at the item level. Aggregation level may differ based on the situation; examples of this aggregation level are: The amount of capacity from a particular bottleneck (De Kok, 1990). The quantity of products from a particular product family, e.g. all products sharing the same extensive component, or the quantity of a particular product supplied to a group of customers, who will specify later on how to divide this amount among them. (Eppen and Schrage, 1981; Van Donselaar, 1990; Van der Heijden et al., 1997; Lee and Tang, 1997). The quantity of products from a particular product family and from a group of manufacturers with quantities of specific items and choice of manufacturer to be specified later (Van Donselaar et al., 2001).

Part I Chapter 2: Demand Forecasting and Customer Orientation

PART II:

THEORY AND METHODOLOGY

Chapter 1 Research Aims and Methodology

1 General Considerations

The previous part provides a general overview of demand management literature and in particular of demand forecasting approaches that researchers have provided since now. As a matter of fact, literature has shown a growing attention towards highly variable demand (i.e. lumpy) so shifting the problem of managing demand uncertainty from a reactive towards a proactive perspective. In fact, as provided in the previous part, authors have gradually moved their attention from developing efficient inventory politics to manage lumpy demand towards forecasting systems able to reduce the effects of uncertainty. Literature has also provided relevant evidence that, when dealing with such variable demand, particular attention has to be paid on the demand generation process, so emphasis has moved from a product perspective, focused on the development of complex forecasting algorithm, towards an information-oriented point of view, where relevant attention is paid to organizational issues. Even if these solutions provide relevant improvements, however, some problems seem to arise. In fact, the environment appears to be more and more affected by a growing complexity, and this trend seems to be widespread in many different industrial contexts. Moreover, all industrial sectors are showing a ever growing attention towards customer level. It is a fact that providing customers with high service level is critical for

Part II Chapter 1: Research Aims and Methodology

guaranteeing future commercial relationship with those specific customers. Many researcher have in fast evaluated that the cost of keeping a customer in their portfolio is very little compared to the cost of gaining a new one. This is due to many different reasons among which the growing competitiveness, the easy access to information regarding suppliers of similar products, the globalization are playing a major role. In this context firms are feeling more and more the pressure of uncertainty and are trying to react by focusing on customer needs and requirements so to rise or at least maintain the service level. However, from another side, firms are experiencing problems in reacting properly to uncertainty, as, if flexibility is needed, it seems to be more difficult to be achieved as internal complexity rises. So typically firms react by allocating redundant resources as inventories, lead times, production facilities and so on, deeply affecting operational performances. Moreover, the ever growing attention towards customers is making the forecasting issue more and more critical since from one side demand becomes more variable and so it tends to be more difficult to be forecasted and, from the other side, forecasting single customers’ needs is rather a tough matter in many different industrial contexts. In this work we argue and further discuss that in particularly heterogeneity among customers and their needs is an important issue since differences in the way in which customers purchase make demand more nervous and less easily forecastable. In this chapter the concept of heterogeneity is introduced. In particular, in the first part environmental factors affecting demand heterogeneity are identified; in paragraph 2 research questions are defined and in paragraph 3 the overall methodology adopted in the research is described.

1.1 Factors influencing demand uncertainty and heterogeneity

Different factors seem to contribute in making demand more uncertain. In fact, in many industrial sectors firms have to deal with a more unpredictable demand; moreover, it seems a widespread phenomenon that firms have to cope with greater difficulty in understanding demand regularities. In other words, from one side demand appears to be more variable and from the other it seems to be more difficult to catch demand regularities. The reasons of this phenomenon may be traced back to many different factors, however we argue that they can be summarized as follows.

Part II Chapter 1: Research Aims and Methodology

Number and complexity of channels

A relevant factor is tied to the growing number of channels firms have to manage to access markets. Many firms have to manage in fact different logistic channels to deal with retail chains, direct customers and gross distributors. This tends to make management more complex as different characteristics are relevant for the different channels. First of all, demand uncertainty is deeply affected by this situation since different demand patterns overlap (i.e. Bullwhip effect). Figure 1 exemplifies this pattern.

Suppliers Producer Distributor Retailer Customers Sales Sales Sales Sales

Time Time Time Time Figure 1: the Bullwhip Effect

However, since a specific firm manages different customers that belong to different supply chain levels, it has to deal with a heterogeneous structure with different requirements in terms of services, needs, demand patterns and so on. In this way a firm has to deal with a demand that is given by the sum of different patterns with different variability (see Figure 2).

Sales Sales Sales + +

Time Time Time Figure 2: the overlap of different variable demand patterns

Part II Chapter 1: Research Aims and Methodology

This in fact makes demand more unpredictable since regularities are more difficult to be found and evaluated, thus making forecasting rather tricky. Many firms are in fact becoming more and more aware of this issue and so they are applying deep reengineering of their supply structure.

An interesting example of this situation is provided by one of the major European white goods producer, located in Italy1. This firm has started developing a relevant project aimed at improving supply chain performances to rise significantly the value perceived by customers. At the beginning of this project, their actual service level was estimated in 70- 80%, even if this estimation was only partial since they were not able to evaluate all out- of-stocks. So they decided to improve their customer service performances by providing different services to different customers’ segments. Table 1 provides an example of the service elements considered for some of the customers’ segments.

Customer type Desired Service level Lead Time Key Account 95% 4-5 weeks Furniture provider 90% 2-3 weeks Direct Customers 85-90% 2-3 days

Table 1: heterogeneity in the service provided towards customers in Whitegoods

To gain this objective the firm focused on different elements and in particular on developing different service politics. However to gain effectively these objective, the firm had to introduce deep changes. First of all, since different service levels had to be gained for different customers, also different managerial perspectives had to be adopted. For example, a Collaborative Planning Forecasting and Replenishment (CPFR) program had to be developed with the key accounts. From another side a deeper integration with the furniture providers had to be adopted to gain anticipated information regarding the end customer order. Moreover a Vendor Managed Inventory (VMI) program had to be implemented to cope with the direct customers. A similar perspective had to be implemented on the supply side defining different relationships with different groups of suppliers, thus leading to the supply structure represented in Figure 3. This change in the supply chain structure has different relevant impact on demand management issues. First of all, to properly manage each segment demand it is important to focus on the different characteristics of these clusters. As a matter of fact, Whitegoods is actually implementing a relevant change in the organizational structure moving from a product-perspective, where functions focus only on product characteristics, towards a

1 For privacy issues we are not allowed to provide details of this company. So we refer to it as Whitegoods.

Part II Chapter 1: Research Aims and Methodology

customer-perspective, where functions are defined according to which segment they manage.

Source Make Move Serve

Customers Suppliers

Segment 3

Source Make Move Serve Segment 2 Segment 1

Suppliers

Figure 3: the change in the global supply chain in the Whitegoods company

Moreover, demand forecasting becomes more challenging since from one side, different customers focus on different product characteristics and different environmental factors, thus requiring that forecasting has to be different for each segment. From another side, since forecast is applied at product-segment level, it tends to be more variable, since demand for a particular product is distributed among different segments. This tends to make demand more nervous so making forecasting trickier.

Life cycle reduction and customization

In the last few years the products’ life cycle has generally reduced: from one side, firm have been stimulated to introduce new products as this is generally paid off by market; from the other, the technological evolution makes possible to obtain very rapidly products with superior quality, better performances, lower costs and price. Moreover, customers require products ever less standardized and ever more customized. In other words, customers’ requests differentiate and personalize ever more, so, to respond properly to what market wants, firms have to provide a range of products and options ever more wide. These factors are relevant for two main reasons. First of all they contribute in making demand more uncertain since by rising the number of new products per time, past history

Part II Chapter 1: Research Aims and Methodology

becomes less adoptable, and since the number of products arises, market requests per products reduce thus making demand for each SKU less stable. Moreover, it becomes more critical to comprehend properly single customer requests. This makes much more difficult to make supply meet demand, thus stresses attention in the operations management towards customers’ requirements.

Emphasis on logistic costs reduction

Another effective problem is tied to the general competitiveness firms have to deal with. In particular, firms pay more and more attention to the reduction of their cost structures and in particular to the logistic ones. This has pushed many firms in different sectors to redesign their distribution chain so to reduce the number of intermediate buffers between central warehouses or production facilities and end customers. Among efficiency aims, other reasons have pushed firms in this direction, mainly tied to the general attention to customer’s service level and to the reduced control of their distribution chains. This in fact has moved firms towards much flat distribution chains, facing directly end customers.

However many experiences show that in general this solution may provide some counter effects. First of all, when intermediate buffers are eliminated, final demand variability affects directly forecasting procedures. Moreover, these changes are usually not developed completely since, for one reason or another, some of the intermediate levels can not be eliminated, as for example they are not owned by the firm or because their importance is too high to get rid of them. This tends to make demand rather composite since different demand patterns get together, thus making demand management rather difficult. Moreover, this sort of logistic revolution tends to shift firms towards end customers thus reducing the possibility of adopting efficiently redundant resources.

Under this emphasis on logistic cost reduction many firms are also facing other changes. For example, in the retail sector, a general widespread phenomenon is the merger of purchasing structures to gain greater efficiency both in the purchasing process and in the logistic activities. Some experiences of these changes can be seen for example in Italy in the realization of Italia Distribuzione, owned by Coop and Conad, that manages directly all purchase, and in the new aggregation of Auchan, Pam, Bennet and Lombardini. This again makes supply chain far more composite and complex to be managed.

Part II Chapter 1: Research Aims and Methodology

Internet

Another relevant variable that has to be taken into account is the changing role of the Internet in operations management. A diffuse conviction concerns the two main possible scope of the Internet adoption (McKinsey, 2000; PWC, 2000): the technology can be exploited either to create a new emerging business (e.g., Amazon, MP3, Autobytel, Vitaminic, e-steel, Bravo-build, Click4talent) or to support existing ones (e.g., Barnes&Noble, Dell, Carpoint, Cisco, Mitsubishi, Sainsbury's, Wal-Mart, Kmart, Esselunga, Dalmine). Many industrial sector have seen a rising development of Internet based solutions to cope better with their customers. Relevant experiences can be found for example in the retail industry (e.g., Wal-Mart, Esselunga, Kmart, Target, Meijer, ShopKo Stores) and in car sector (e.g., Fast-Buyer, AutoZone, Advance Holding Corp., The Pep Boys). Given the fact that e-business opportunities can be exploited only if strongly related to existing activities, the capability to match existing assets with the use of the Internet is becoming more and more crucial in order to build a sustainable competitive advantage. Digital activities will not cannibalize physical ones, but should complement them and improve the overall performance of the company (Gulati and Garino, 2000; Venkatraman, 2001; Short, 2001; Porter, 2001). As a matter of fact, information transfer, frictionless transactions and the Metcalfe's law2 ("The value of a network is equal to the square of its participants") make the Internet a powerful tool to improve, among the other processes, also supply chain performances (Cross, 2000). A recent study (Lancioni et al., 2000) shows that about 90% of companies adopt the Internet in some part of their SCM program; however, this result should be viewed with much caution due to the self-selection bias related to the data collection. In addition, such questionnaire was more related to internal supply chain activities rather than to relations with suppliers. On another hand, the new information and communication technologies provide the possibility to communicate and interact directly in an efficient and effective way with customers, thus making possible disintermediation.

However the Internet adoption implies that customers have a greater visibility on the internal activities of firms. For example, In Dalmine, customers are able to order pipes directly on-line but they also are provided with the possibility of checking where their specific order are in the supply chain. This makes critical to be able to provide customers with detailed and precise information so guaranteeing effectively what they want. In these terms, Internet may influence deeply the way in which demand is managed since it tends to make forecasting trickier.

2 Named after Robert Metcalfe, inventor of Ethernet and founder of 3COM.

Part II Chapter 1: Research Aims and Methodology

1.2 Effects on Customers’ Structure

All the indicated elements, that for sure are deeply influencing the way in which businesses are conducted, have also a strong impact on the customers’ structure firms have to deal with. In fact, firms have become gradually aware that in many contexts it is ever more critical the role that customers play in the inner processes. All the previously indicated elements tend in fact to shift more and more attention towards customers. The rising number and complexity of channels adopted to deliver products imply that different customers have to be served and managed. For example, firms have to deal from one side with retail customers that tend to purchase frequently but for different quantities since typically these customers apply promotional activities, thus making their orders highly unpredictable. From another side, for example, distributors have to be managed, which, due to their particular characteristics tend to buy rarely but for huge amounts of material, thus again making demand more variable. Moreover, direct customer may be considered, which tend to buy rarely, with very different lead times, thus contributing in the rise of demand uncertainty. An interesting experience within Reckitt Benckiser, a house hold cleaning products manufacturer showed in fact such a problem: customers belonging to different channels purchase differently, however, accessing to different channels plays a critical role in the business success. In these terms, the Internet, which plays a relevant role also in the access towards end market, makes channels heterogeneity more critical.

From another side, the ever growing attention towards customization makes customers more sensible towards both products characteristics and service elements. This tends to rise the number of variables that influence customers’ purchases, thus making relevant to identify properly customer’s cluster so to apply proper commercial activities.

Moreover, the supply chain redesign due to the attention towards logistic costs is also making the customers faced by companies more different. In other words, the supply chain becomes more complex, different customers have to be managed, thus influencing their specific attention towards service elements.

As a general effect, all these elements tend to rise customers’ heterogeneity. In fact, from one side, customers tend to behave differently, they are influenced by different elements and require more suited solutions; in these terms customers’ intrinsic heterogeneity rises. From another side, supply chain complexity, due to the coexistence of several distributional channels, the emphasis on efficiency in logistics and the new interaction opportunities given by the Internet, tends to join together customers that have very different behaviors, thus rising the perceived heterogeneity companies have. Figure 4 summarizes these elements.

Part II Chapter 1: Research Aims and Methodology

Number & Complexity of Channels

Life Cycle Customers’ Customers’ Customers’ Internet Reduction & intrinsic heterogeneity perceived Customization heterogeneity Heterogeneity

Emphasis on logistic costs reduction

Figure 4: environmental factors influencing customers’ heterogeneity

The overall effect is that heterogeneity firms have to cope with is ever more relevant and significant. All these elements tend to focus firms’ attention towards deeper levels in the supply chain, since they have to understand better what happens towards the market. However, this forces them to deal with a greater number of customers with many different peculiarities. In other words the shift of attention towards deeper levels of the supply chain makes wider the spectrum of customers that have to be managed. Consider the example provided in Figure 5. In a simple situation the firm located at the upper level (A) manages demand from the companies located at the closest level (B). When attention shifts deeper in the supply chain (Figure 6), firm A has to deal with customers located at the lowest level (C), where heterogeneity tends to be higher and management becomes more complex since the number of customers rises significantly.

Figure 5: attention is paid towards upper levels of the supply chain

Part II Chapter 1: Research Aims and Methodology

Figure 6: more attention towards end customers influences company’s perceived heterogeneity

Thus, a higher attention towards the supply chain makes the company’s perceived heterogeneity higher.

Moreover, since many different channels are adopted to reach customers, firms have to deal with different intermediaries, with different structures and characteristics, thus leading to a growing heterogeneity in their behaviors (see Figure 7).

Figure 7: an example of heterogeneous supply chain

This tends again to make heterogeneity more relevant and demand management more complex, because from one side demand becomes composite, due to the overlap of different sources of demand variability; from another side, the factors that influence customers’ demand are different as, for example, customers belonging to one chain are influenced by prices variations while others are influenced by service conditions. Moreover, the way in which the different customers are or can be managed is heterogeneous, since different kinds of relationships can be developed, different kinds of agreements may be arranged, and so on.

Thus, the previously introduced factors contribute in rising the complexity of the supply chains with the consequent rise in demand heterogeneity.

Part II Chapter 1: Research Aims and Methodology

This phenomenon is relevant for two main reasons. First of all, it may influence demand variability and so may directly affect the performance of both forecasting and inventory management systems. We argue that, as other authors identify (see Bartezzaghi and Verganti, 1995), the degree of heterogeneity may be one of the causes of demand lumpiness and thus affects forecasting. Moreover, we argue that, besides the problem of dealing with lumpy demand, managing heterogeneous demand may far more complex, in terms of forecasting methods, forecasting organization and forecasting application. In fact, as literature states, if demand is uncertain, forecasting has to shift on the demand generation process. However, it may be difficult to identify properly how to deal with a demand that is partially influenced by some factors and partially by others. Consider, for example, two demand series with similar variability (measured for example by the coefficient of variation), but one is deeply affected by seasonality, while the other doesn’t show systematic patterns since its variability is only due to customers’ nervousness. Even if the two series are equally variable, they can not be forecasted with the same effort, since seasonality can be easily accounted as it is easily estimable, while the different customers’ behavior may be very difficult in being estimated also because they represent many different variables. Figure 8 provides an example with two series characterized by equal variability (CV = 20%) due to different sources of variability.

demand 10

0 1 4 7 1013161922252831343740434649 days

Figure 8: comparison between a seasonal series and a random one

This makes forecasting far more difficult. In fact, as literature suggests, when demand variability is high attention has to be paid towards the demand generation process, thus paying attention towards customers. However, a single customer approach may be inapplicable or too costly, due to the numerousness of customers and to the complexity of managing each customer’s demand differently.

Part II Chapter 1: Research Aims and Methodology

In this way we argue that attention has to be given to the demand generation process, but trying to reduce the number of observations. As detailed in the next section, the existing forecasting approaches, both those based on demand data and those based on information, may face relevant problems when heterogeneity becomes more significant.

After the identification of the major limitations actual forecasting approaches have, we will properly define heterogeneity and then focus on the impact of heterogeneity on demand and on demand forecasting.

1.3 Limitations of actual forecasting approaches

In the previous part we have considered the problem of managing demand forecasting in heterogeneous contexts. In particular we considered some of the main contextual factors that influence demand heterogeneity and showed that this phenomenon may influence at some rate demand variability, thus making demand more complex to be forecasted. However, we also stated that when heterogeneity in demand becomes more relevant forecasting is more complex as different forecasting solutions have to be applied, since different factors influence demand variability. So the main problem is identifying proper forecasting approaches that guarantee the following requirements:

Identify properly variability sources: in other terms forecasting systems have to be capable of properly identifying the root causes of demand variability, so those tied to systematic variability and those due to customers’ particular characteristics. Manage properly variability due to heterogeneity: since contributions regarding systematic variability managements are deeply developed, these systems have to be capable of dealing with demand uncertainty due to customers’ behavior and, in particular, to their heterogeneity. Solve the trade-off between cost of forecasting management, ability to catch variability and ability to manage variability. As it has been stated in the previous chapter, when heterogeneity rises, we argue that different forecasting methods and estimations have to be adopted. However, these solutions imply a relevant cost for gaining and managing information, thus this should force not to adopt too complex solutions. From another side, however, variability has to be identified thus forcing to disseminate forecasting approaches while the ability to manage variability again forces not to develop too many approaches. In these terms, the systems adopted should be able to balance the trade off between these dimensions so to be both effective and applicable.

As a matter of fact actual literature lacks of contributions that may properly deal with this issue. We may consider some examples to understand better the implications of this lack of

Part II Chapter 1: Research Aims and Methodology

knowledge, by taking a brief overview to the main limitations of two of the main approaches developed in current literature: Early Sales and Order Overplanning. This claims for the need of different approaches to manage demand when dealing with highly heterogeneous contexts. After these considerations, the research aims will be defines in paragraph 2, while in paragraph 3 the methodology adopted will be described.

Early Sales

Even if literature reports relevant experiences of the good performances this technique has, its applicability is somehow limited to particular sectors. First of all this method fits situations in which medium term forecasting have to be evaluated for products with very short lives. This tends to limit its application in particular contexts as for example fashion goods and, since its applicability is highly time consuming, it can’t be applied continuously.

Moreover, its attention is paid on the possibility of gaining early information regarding future needs from a reduced number of customers and to use these clues to foretell total future demand. This fact, as attention shifts towards deeper level of the demand chain, makes this technique less applicable, since from one side, it becomes more difficult to develop relationships with customers so to gain early information, from the other, a greater number of customers have to provide this information to identify significant patterns. In fact, it can be empirically demonstrated that its accuracy rises as more demand is observed, so it is relevant to gain enough early information to properly evaluate forecasts. Figure 9, Figure 10 and Figure 11 describe this phenomenon.

4000

3500

3000

2500

2000 Demand 1500

1000

500

0 0 500 1000 1500 2000 2500 3000 3500 4000 Forecast Figure 9: comparison between demand and forecast for different products, when no information is available

Part II Chapter 1: Research Aims and Methodology

4000

3500

3000

2500

2000 Demand 1500

1000

500

0 0 500 1000 1500 2000 2500 3000 3500 4000 Forecast Figure 10: comparison between demand and forecast for different products, when 20% of demand is known

4000

3500

3000

2500

2000 Demand 1500

1000

500

0 0 500 1000 1500 2000 2500 3000 3500 4000 Forecast Figure 11: comparison between demand and forecast for different products, when 80% of demand is known

When dealing with many customers it becomes more difficult to achieve information for a certain demand quota, so making efforts to gain early sales rather costly. Moreover, this technique is applicable when demand for the first customers is correlated with that of the late buyers. However, the more heterogeneous are customers the less this assumption is applicable, thus reducing the possibility to adopt this system efficiently.

However, quite interestingly, this technique leverages on the existence of heterogeneity in the demand generation process, since it considers customers that tend to order in different periods, thus considering early versus late buyers. Nevertheless, in many real contexts this approach may not be applicable.

Part II Chapter 1: Research Aims and Methodology

Order Overplanning

Literature has provided evidence of the effectiveness of this technique, however its structure provides some limitations to its applicability when dealing with a relevant number of customers. In fact the cost of achieving early information is directly correlated to the proximity between firm and customers: when interaction is provided by means of intermediaries it becomes more difficult for sales people to get in touch directly with end customers, thus making the applicability of this technique more complicated. From another point of view, the cost of adopting this technique is correlated to the number of customers that have to be managed. In these terms, since attention is paid towards end customers, or at least towards customers located in deeper levels of the demand chain, it is also more costly to apply this method effectively. Moreover, this technique leverages on the existence of a significant lead time between information retrieval and actual demand. Typically when moving towards deeper levels of the distribution chain, this interval tends to reduce or to be less visible, since less attention is given to planning issues and so the demand generation process appears to be less visible.

All these elements tend to make demand management rather difficult and the cited techniques less effective. In fact the limitation of the forecasting techniques literature provides appear to be significant and tied to the inadequacy of the approaches to deal with the root causes of the rise in demand variability. Since fewer Early Buyer (EB) can be identified, and since it becomes more costly to gain Early Information (EI), both Order Overplanning and Early Sales techniques tend to be less effective and their benefits compared to traditional techniques less significant. Figure 12, in fact, shows this phenomenon.

As it can be noted, the more demand is dispersed the less Order Overplanning is applicable, compared to simple Exponential Smoothing. Again, if correlation among customers is reduced (as when customers request are more heterogeneous and dispersed), Early sales are less applicable. These considerations lead us to the definition of the research scopes of this work.

Part II Chapter 1: Research Aims and Methodology

correlation between customers requests

(ρij ) n and f dispersion (D) ρt lumpiness EWMA

Early Sales ρt EI EI EB Order Overplanning

sporadicness irregularity lumpiness (h and cvord )

Figure 12: the influence of Early Information (EI) and Early Buyers (EB) availability to the effectiveness of Order Overplanning and Early Sales compared to Exponentially Weighted Moving Average (EWMA). Adapted from Bartezzaghi et al. (1999a)

Part II Chapter 1: Research Aims and Methodology

2 Definition of research aims

Given the problem previously described, the research questions of this work can be summarised as follows.

1. How does heterogeneity in customers’ purchasing processes influence demand variability and forecasting effectiveness? A first major issue this work wants to deal with is the impact of heterogeneity in demand and its impact on demand variability and forecasting. The main issue is to identify if and to what extent this concept is relevant and why attention should be paid to this issue. In fact, we argue that in heterogeneous contexts demand variability changes and that forecasting becomes more complex; so at first attention will be given to this issue.

2. How should demand be managed when dealing with heterogeneous contexts? Since it is more difficult to manage demand in complex contexts it is ever truer that the focus is shifting towards customers requirements. Since literature does not provide answers to all the problems customer-orientation introduces, the main objective is to contribute to this issue. In this way we want to provide a forecasting strategy to manage demand in heterogeneous segments, so where different classes of customers are considered. Moreover, we want to define specific forecasting approaches that try to leverage on demand heterogeneity thus providing forecasting strategies to deal with such variable demand.

3. What are the main elements that should be taken into account when dealing with heterogeneous contexts? We argue that focusing on heterogeneity may be significant when dealing with complex forecasting problems. However we believe that it is not the global solution for all the forecasting problems. So a major interest is identifying under which conditions the main focus on customers heterogeneity may be significant, thus providing researchers and practitioners with some guidelines to properly identify forecasting approaches. Moreover, since there is no measure of the intrinsic heterogeneity within demand, this may help practitioners in understanding when approaches focused on heterogeneity should be applied.

In the next chapter the empirical methodology adopted to answer these questions is described.

Part II Chapter 1: Research Aims and Methodology

3 Research Methodology

Given the two main objectives of this work, the focus of the analysis is short-term demand forecasting. In particular, short-term demand forecasting is focused on providing estimation of future demand typically for planning purposes as for example production and purchasing planning. In these terms, short-term demand forecasting influences directly operational performances as inventory and service level.

To choose the proper methodology to deal with the described issue, we took into considerations different aspects concerning the research aims. 1. The main objective is to build theory. Since this research aims at proving new ways to solve forecasting problems, these solutions can’t be found in current knowledge or in current managerial experience. This states that the adopted research methodology can’t emphasise descriptive aspects. 2. Since the problem is empirically important, the solution must be empirically valuable. The scope of this research is to provide real solutions to real problems, so relevant attention is given to the fact that real contexts have to be considered. In these terms quantitative methods have to be adopted and real data has to be considered. 3. As the next chapter will detail heterogeneity relies on many different dimensions that jointly influence demand variability. To properly understand heterogeneity and to provide proper methodologies to cope with its effects on demand, we have to focus on the specific elements that constitute heterogeneity.

Given the research questions approached by this work and the previous remarks we decided to adopt a two stage research process, based on a Theoretical Stage, aimed at defining the role of heterogeneity within demand forecasting, and an Empirical Stage, focused on providing methods to cope with demand in heterogeneous contexts.

Figure 13 describes the overall research structure.

3.1 Theoretical Stage

Since the aim of this part of the research is to study the impact of heterogeneity on demand and forecasting, at first we need to identify what uncertainty is and what are its main elements.

In particular, the literature review provided in the previous part and a detailed analysis on the concept of heterogeneity provides the definition of a first set of hypothesis related to

Part II Chapter 1: Research Aims and Methodology

the effects of heterogeneity on demand variability and forecasting effectiveness and so tied to research question 1. In this stage we adopted a simulation approach, since, as further better described, it was not possible to identify real contexts in which heterogeneity was measurable.

Literature review Heterogeneity definition

st 1 set of Theoretical Hypothesis Stage

Simulation

Validation of 1st 2nd set of Theoretical set of Hypothesis Hypothesis Framework

Action Empirical Research Stage

Validation of 2nd Methodology Forecasting Th. Framework set of Hypothesis Development Approaches Dev. Validation

Figure 13: the methodological approach

Simulation based research can be defined as a rational knowledge generation approach (Meredith et al., 1989). This methodology is based on the assumption that objective models can be built so to explain the behaviour of real-life operational processes or that can understand the decision-making problems that are faced by managers. Model-based research can be divided into two main classes. The first class is primarily driven by the model itself, thus it is defined as axiomatic (Meredith et al., 1989). In this class, the primary concern of the research is to obtain solutions within the defined model and make sure that these solutions provide insights into the structure of the problem as defined within the model. The second class of model-based research is primarily driven by empirical findings and measurements. In this class of research, the primary concern of the researcher is to ensure

Part II Chapter 1: Research Aims and Methodology

that there is a model fit between observations and actions in reality and the model made of that reality. Examples of this type of research is the industrial dynamics research conducted by Forrester (Forrester, 1961). In this work, due to the complexity of the issue addressed we adopted this last approach. In particular, we will consider heterogeneity as the causal variable while variability will be the derived variable. Our aim is to evaluate the shape of the relationship function f that links these two elements and the conditions that tends to change its profile. In these terms, during this stage of the research our attention will be paid to:

V = f (H ) where V is demand variability and H is demand heterogeneity.

This approach let us validate the first set of hypothesis and define a theoretical framework regarding a general methodology to cope with heterogeneous demand. Moreover, this general framework was summarized by the definition of a second set of hypothesis regarding demand forecasting approaches within heterogeneous contexts.

3.2 Empirical Stage

These previous results are used to develop the second stage. In particular, given research questions 2 and 3, and the previous remarks regarding the overall scope of the research, this part had to focus on real contexts. Moreover, since attention is given towards the development of new methodologies, this part of the research is based on an Action Research methodology.

Different elements characterise Action Research (AR); in fact AR can be described by means of different characteristics. First of all AR focuses on research in action, rather than research about action. The central idea is that AR uses a scientific approach to study the resolution of important social or organisational issues together with those who experience these issues directly. Second, AR is participative. The researcher participate actively in the cyclical process of the research. Such participation contrasts with traditional research where members of the system are objects of the study. Third, AR is research concurrent with action. The goal is to make that action more effective while simultaneously building up a body of scientific knowledge. Finally, AR is both a sequence of events and an approach to problem solving. As a sequence of events, it comprises iterative cycles of gathering data, feeding them back to those concerned, analysing data, planning action, taking action and evaluating. It is also an

Part II Chapter 1: Research Aims and Methodology

application of the scientific method of fact finding and experimentation to practical problems requiring action solutions and involving the collaboration and co-operation of the action researchers and members of the organisational system. The AR methodology requires the development of a 6-step cycle, a pre-step of understanding and a meta-step to monitor (Coughlan and Coghlan, 2002). In this work we will adopt this scheme (see Figure 14).

Context & Purpose

Data Gathering

Evaluation Data Feedback

Monitoring

Implementation Data Analysis

Action Planning

Figure 14: the action research cycle (adapted from Coughlan and Coghlan, 2002)

The application of this methodology let us validate the second set of hypothesis and the theoretical framework provided. Moreover, these results let us develop an overall forecasting methodology that can be adopted when dealing with highly heterogeneous contexts. In the end the different studies developed provide current forecasting knowledge with new methods to cope with particular heterogeneity factors.

However, since heterogeneity is multi-dimensional, it was not possible to identify real situations that are influenced only by single specific elements. Thus, the considered contexts show elements of heterogeneity that coexist and for which different solutions have to be designed. Moreover, given again the complexity of the problem we were not able to consider situations that face all the different sources of heterogeneity. So we focused only on some of the dimensions on which heterogeneity may show up, and then analyze the generalization of the methodologies provided. As a matter of fact we considered 3 real cases in which methodologies have been developed and implemented. In all these case the previously described research model is adopted. In fact, at first data are collected from the real situations and feed back regarding data knowledge is given to managers. Then data are analysed by means of statistical tools and actions are provided leading to the implementation of the proposed solution obtaining

Part II Chapter 1: Research Aims and Methodology

certain performances. Typically this situation doesn’t end in one round but, given the experience matured, new data are collected and new analysis conducted, thus starting back the cycle. In the description of the different cases, we will adopt this cyclical process and try to show the design of the solution by means of this scheme. Moreover, we will also try to identify to what extent these solutions are applicable, thus again helping practitioners to understand the applicability of these solutions within their forecasting scenarios.

In this way the methodology adopted aims at: 1. develop a specific solution for a specific real problem within a real scenario; 2. identify to what extent the specific methodology can be generalised; 3. identify under which conditions the methodology should be applied.

In the next chapter the theoretical stage will be described, while the empirical stage will be addressed in part III.

Chapter 2 Theoretical Framework

“All animals are equal but some animals are more equal than others” George Orwell – Animal Farm (1946)

1 Heterogeneity: a definition

Customers’ heterogeneity is not a widely studied subject in the demand management field. In fact, major contributions regarding this issue can be found in the Marketing-related literature. This filed of research in particular refers to heterogeneity as:

“…differences between households in their intrinsic preferences for brands and/or their sensitivities to marketing variables such as price, feature advertisements and special display.” Gupta et al. (1997)

In fact attention of works in this field is tied to the effect of heterogeneity in customers’ characteristics on their brand preference. Contributions in these area consider models representing relationships among marketing mix variables.

Part II Chapter 2: Theoretical Framework

Heterogeneity has been characterized in different ways. In particular, it can be at first classified in whether it is a preference heterogeneity or a structural heterogeneity. Preference heterogeneity refers to individual differences in brand preferences and responses to the marketing mix. These sources can be clustered according to the fact that they are observable or unobservable factors. Observable heterogeneity usually refers to factors that can be directly accounted by firms and that can be easily analyzed to identify customers’ preferences. For example, Guadagni and Little (1983) and Gupta (1988) use purchase history as a means of accounting for heterogeneity. Krishnamurti and Raj (1988) and Chiang (1991) use, in addition, income to capture variations in purchase behavior across customers. However, there are also variations across customers that cannot be observed. Such variations are referred to as unobserved heterogeneity. A customer intrinsic preference for the different brands in a product category is an example of such heterogeneity. Heckman and Singer (1984) show that these unobserved factors, if not considered, result in biased estimates for the effects of the included marketing variables on purchase behavior. The studies of Jones and Landwehr (1988), Colombo and Landwehr (1990), Gönül and Srinivasan (1993) and Jain and Vilcassim (1991) further illustrate the need to account for unobserved heterogeneity in such models.

Other authors also consider heterogeneity as a structural factor. Structural heterogeneity refers to differences in the structure of the choice process, accounting for the fact that some customers for example choose first the kind of product they need and then the specific brand, while other focus directly on the brand choice. For example Currim et al. (1988) consider a classification procedure that infers the decision rules adopted by customers. Similar approaches are considered by Kannan and Wright (1991).

As a matter of fact customer heterogeneity is a widely studied subject even if attention has been only given by marketing research. However, some contributions have also been found in the Value analysis literature. In particular, Desarbo et al. (2001) propose a modification of the most known Customer Value Analysis model to take into account differences in preferences among customer, in terms of perceived quality and price sensibility. They also state that this issue is not studied at all in this research area.

Even if marketing-related literature is very rich in contributions regarding customers’ heterogeneity (see also Wedel and Kamakura, 2000, for a detailed description of the major approaches provided), however minor attention has been paid to this issue in other operations research areas, moreover in demand forecasting. In these terms, our attention focuses on the dimensions of heterogeneity that become relevant when dealing with demand forecasting.

Part II Chapter 2: Theoretical Framework

However, we argue that a different definition may be applied when dealing with heterogeneity in operations management. Consider again the definition previously introduced by Gupta et al. (1997). This definition states:

“…differences between households…” Differences considered are evaluated at household level, so at the final customer level, however in a more general perspective, attention can be paid also to customers, that may be final customers, but also intermediaries in the supply chain. In other words we can extend the definition by integrating a business to business perspective beside the business to consumer attention. “…in their intrinsic preferences for brands and/or their sensitivities…” Differences are tied only to customers’ brands preferences and/or their sensitivities to some variables. This makes that heterogeneity appears as an external and uncontrollable factor, while in an industrial context this may not apply. In fact customers may also behave differently according to the service provided, or to the particular conditions applied. In other words sometimes heterogeneity is influenced directly by the considered firms that applies some specific politic on some customers. The reasons of this behavior are many: from one side there may be a lack of resources that forces to apply some activities only to some groups of customers, as for example promotional activities that are associated to specific costs. From another side there may be lack of access towards some customers, so not providing the possibility of influencing their behavior. Moreover there may be managerial activities that provide different customers with different services due to the supply chain structure. For example, some customers may be managed according to a Collaborative Planning approach, while others according to a traditional relationships. In these terms we argue that a much more general definition of heterogeneity should consider also the fact that heterogeneity may be at some rate influenced by firm’s decisions. “…to marketing variables such as price, feature advertisements and special display.” Differences are tied to marketing variables such as price, feature advertisements and special display. However differences may also be related to other dimensions as for example the way in which customers purchase. In fact, there may be customers that purchase very frequently and other very rarely, customers may purchase for relatively small quantities or for huge amount of materials, and so on. In this way customers may differ also according to many other dimensions.

Part II Chapter 2: Theoretical Framework

Given the previous observations we argue that a much more general definition of heterogeneity may be as follows:

Heterogeneity refers to dissimilarities between customers due to differences in their purchasing processes, in their sensitivities to external factors and in their reactions towards specific firms’ activities.

In this work we will refer to heterogeneity as defined here. However to make more clear the implications of this approach we argue that is important to analyze deeper the different dimensions along which heterogeneity may be found.

1.1 Heterogeneity dimensions

Heterogeneity can show up in many different aspects. In fact, literature regarding heterogeneity in marketing has paid attention towards this issues. Marketers are aware that customers differ, and in particular that their decisions may be influenced both from their characteristics and external influences, such as marketing activities and environmental factors.

Figure 1 presents a detailed model of the factors influencing consumer’s buying behavior as provided by Kotler (1991).

Marketing Other Buyer’s Buyer’s decision Buyer’s Stimuli Stimuli characteristics process decisions Product Economic Cultural Problem Product choice Price Technological Social recognition Brand choice Place Political Personal Information Dealer choice Promotion Cultural Psychological search Purchase timing Evaluation Purchase amount Decision Post-purchase behaviour

Figure 1: model of buyer behaviour

We will not go into detail regarding this model, however, as it can be noted, customer’s behavior is influenced by many different variables.

From on side we can consider inner characteristics of the customer, as, for example, their purchasing process, their size, their loyalty to a particular brand, their utility, and so on. From another perspective, also environmental factors may influence the way in which customers behave, for example socio-cultural elements, macroeconomic factors, weather conditions, etc. At last also firm’s decisional factors may influence customer’s

Part II Chapter 2: Theoretical Framework

characteristics, as for example service level provided, pricing conditions, promotional activities, distributional politics, and so on (see Figure 2).

Inner Characteristics

Firm Decisional Customer factors

Environmental factors

Environment

Figure 2: the different sources of demand heterogeneity

All these elements tend to influence how customers behave in terms of product/brand choice and purchasing process. This means that demand pattern can be significantly influenced by customers’ heterogeneity.

Customers’ inner characteristics

These elements refer to factors that directly belong to the single customer’s characteristics. In other words, in this field we consider all elements that make each customer potentially different from others and that is only referred to him/her. In this group we can consider two other main classes of variables: from one side we can consider customer-specific factors as for example their size (in terms of volumes or sales); from the other side, we can consider the inner response of each customer to external factors, as for example promotional activities, weather conditions, and so on. Examples of variables belonging to the first group are: Customer size: we may have customers of different size (for example some small customers and some big ones). Reorder politic: in terms of whether they apply an Economic Order Quantity purchasing logic or an Order-up-to one, size of the single order they place, interarrival between orders.

From the other side, examples of variable belonging to the second group are: Sensibility to price changes: so we may have customers that react more or less differently to promotional activities, and also customers that do not react at all.

Part II Chapter 2: Theoretical Framework

Sensibility to promotional activities: similarly to the previous elements, customer may differ if they do or don’t apply promotions, and the extent to which promotional activities impact on their purchases. Sensibility to weather conditions: these factors that change among customers due to their localization may influence demand seasonality.

Environmental Factors

A different group of variables are related to the existence of environmental elements that may influence customers’ behaviors. In fact, these may apply differently to different customers; examples of these factors are: Socio-cultural factors: the customer profile is of high relevance for consumer goods, since to different customers (in terms of age, sex, instruction level, and so on) correspond different purchasing behaviors. However, also for industrial products this aspects may be relevant since end-customers demand may also influence that of industrial customers. Geo-political and Macro-Economic effects: differences in the potential expenditures of the single customers (income level, income distribution, and so on), economical conditions influence their way of purchasing. Market Competitiveness: the presence of different complementary products, so the level of competitiveness in the market is capable of influencing customers’ purchases.

Decisional Factors

In the end there can be considered variables that are, at least partially, controllable by a considered factors. Among these we can consider: Product characteristics: we intend both the product itself with its characteristics (functionalities, quality, stile, guarantee, and so on) that make it more or less customizable and the connected services offered to the customer (delivery, post-sales assistance, and so on). Price decisions: considering both price determination criteria and discount politics, prizes, payment conditions. Promotional activities: they comprehend the activities developed to promote the product. The company may apply different promotions both in terms of kind of activity (simple discount, bonus pack, coupon, and so on) and entity, moreover these activities may be applied differently towards different customers.

Part II Chapter 2: Theoretical Framework

Place: in other terms differences may exist because different distributional channels are adopted. Service levels decisions: in terms of different services levels provided to different classes of customers. Communication media

As a matter of fact all these elements are strictly tied together (see Figure 3). In fact, environmental elements tend to influence directly customers’ inner characteristics, since, for example, economical welfare makes customers more or less sensible towards price. Moreover, also, firm’s decision may be affected by environmental elements. However, attention should be paid to the controllable part of the decisional variables.

Environment

Firm Customer

Figure 3: mutual influences among the different sources of demand heterogeneity

As it can be seen, heterogeneity may apply on different dimensions. However, one major problem when dealing with heterogeneity is tied to its identification. Since all these factors typically coexist and, moreover, tend to melt together, it usually very difficult to identify the root causes of heterogeneity. As a matter of fact, even if contributions regarding heterogeneity measures exist they tend to be limited.

1.2 Heterogeneity Measures

Measures regarding heterogeneity have been proposed in literature regarding different topics. In particular, in industrial economics research some authors have provided measures to evaluate heterogeneity in customers’ dimension (for a general review of these measures we refer to Clark, 1985). In this area major attention has been paid on market concentration thus measuring how customers differ in terms of size. The most widely used concentration index is the concentration ratio. It is defined as the proportion of industry output accounted for by the r largest firms, where r is an arbitrary number. Thus, if we consider n firms with output xi, ranked from largest and smallest, the concentration ratio is defined as: r x r C = i = s r ∑ x ∑ i i=1 i=1

Part II Chapter 2: Theoretical Framework

n th x where x = x is the industry output and the market share of the i firm is s = i ∑ i i x i=1

Criticisms of the concentration ratio as a measure of concentration centre on the arbitrary selection of r, since it doesn’t consider all the concentration curve, but only the influence of the first r firms. This has lead to the development of different measures, as for example the Hirschamn- Herfindahl index. This index takes account of all points on the concentration curve, being the sum of squared market shares of the firms in the industry:

n 2 n x 2 H =  i  = s ∑ x  ∑ i i=1   i=1

The index, by squaring market shares, gives most weight to the larger firms in the industry. However, in some situations this particular kind of weight may not be appropriate (see for example Hannah and Kay, 1977, for a discussion of the main problems). This has lead to the definition of other measures as for example the Entropy index. This index weights market shares by ln()1 si , in the form:

n E = ∑si ln()1 si i=1

Beside these concentration measure, that can be adopted to evaluate heterogeneity in firms’ size, other measure are provided that focus on measuring inequality. Such measures ignore firm numbers and can be regarded as summary representations of the Lorenz curve, in the same way as absolute measures summarize the concentration curve. The Lorenz curve plots the cumulative percentage of market output against the cumulative percentage of (rather than number of) firms. Firms are also typically cumulated from the smallest to the largest in constructing a Lorenz curve. Several measures of inequality may be mentioned here. The Gini coefficient which is obtained through a Pareto analysis of the customers’ dimensions (Lambert, 1989). In particular, if Pi is the value of the Pareto curve for customer i (see Figure 4), then: n

∑ Pi i=1 h = n

∑()1− Pi i=1

100

Part II Chapter 2: Theoretical Framework

Cumulated customer size (Pi) 1 0,9 0,8 0,7 Homogeneus market 0,6 (h=1) 0,5 0,4 Heterogeneous market 0,3 (h>1) 0,2 0,1 0 Customers 0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1 Figure 4: The Gini’s index: measure of heterogeneity through a Pareto analysis of the customers’ size

Another possible measure is the demand dispersion (Bartezzaghi and Verganti, 1995). High dispersion occurs when the market consists of a great number of customers with homogeneous dimensions, and therefore the overall demand splits in several requests of similar size. Of course, lumpiness increases as demand dispersion decreases. Dispersion may be measured through several dispersion indexes (Pielou, 1977). For example, if Si is the average expected order of customer i, then:

n 2 ∑Si D = i=1  n 2  ∑Si  i=1 

As it can be noted no attention is paid in literature regarding heterogeneity on factors apart from demand size. In fact we argue that when heterogeneity is due to many different factors, it is too complicated to provide an absolute measure of heterogeneity that evaluates all relevant factors. In this situation we argue that qualitative evaluations of demand heterogeneity may apply better; identifying the root causes of heterogeneity may be a sufficient evaluation of the impact of customers’ behavior on demand. In real applications we suggest to adopt a reverse approach: instead of trying to measure the impact of heterogeneity on demand, one should evaluate the contributions of typical systematic variability factors (seasonality, trend, and so on), then evaluate whether managerial elements exist (as for examples promotional activities or commercial conditions) or context factors are relevant (as macroeconomic phenomena, weather conditions and so on). If in the end demand variability is still not much understood (in other terms a relevant part of variability is still present after cleaning demand from these sources), probably the problem is tied in how customers purchase and so attention should

101

Part II Chapter 2: Theoretical Framework

be paid to their specific demand, to analyze whether correlations exist among customers or if any cause of heterogeneity may be identified. As a matter of fact this approach will be adopted in chapter 5 when specific studies regarding real situations will be taken into account. As it will be cleared, specific and detailed analysis of demand data and supply structures will provide sufficient detail in identifying properly where heterogeneity lies.

102

Part II Chapter 2: Theoretical Framework

2 Impact of Heterogeneity on Demand

As previously defined the main scope of this part of the research is to answer to research question 1:

How does heterogeneity in customers’ purchasing processes influence demand variability and forecasting effectiveness?

When dealing with the impact of heterogeneity on demand, few contributions can be found in current literature. Bartezzaghi and Verganti (1995) study the effects of different causes of lumpy demand on forecasting methods applicability. In their work they focused on different possible causes among which heterogeneity was taken into account. In particular they consider heterogeneity in terms of demand size, thus considering homogeneous versus heterogeneous market (see Figure 5 and Figure 6).

EWMA EaSa OrOv (rich EI) OrOv (poor EI)

MAD MAD 120 120 100 100 80 80 60 60 40 40 20 20 0 0 5 20 50 n 5 20 50 n 2.35 1.9 1.85 lumpiness 2.95 2.25 2.19 lumpiness Figure 5: Effect of numerousness Figure 6: Effect of numerousness in in case of homogeneous market case of heterogeneous market

These authors show that when market is heterogeneous, demand tends to more lumpy and then forecasting performances of simple techniques worsen, while techniques as the Order Overplanning tend to perform quite well. In the study they considered Exponential Weighted Moving Average (EWMA), Early Sales (EaSa) and Order Overplanning (OrOv), both when early information provided regarding demand is poor (poor EI) and when early information is rich (rich EI).

Thus in this part we define a first hypothesis which will be addressed in the present chapter:

H1. Customers’ heterogeneity affects demand variability and forecasting accuracy.

103

Part II Chapter 2: Theoretical Framework

So the main focus of this part is tied to the relationship that exists between customers’ heterogeneity and demand pattern. As a matter of fact the analysis of this relationship is rather tricky due to the complexity of the problem. In particular the main problem in analyzing this relationship is tied to the multi-dimension nature of heterogeneity. Given a certain situation customers may differ along many different dimensions, as previously introduced; they can differ for the lot-size they adopt when purchasing, for their purchasing lead-time, for their total demand, for their sensitivity towards environmental factors, and so on. Typically all these factors jointly influence single customer’s purchasing process thus making demand rather composite. In these terms it is very complex to measure heterogeneity on single dimensions since usually all these characteristics show up jointly.

Moreover, a relevant problem is tied to the kind of data firms have to deal with. In fact some dimensions of heterogeneity can not be evaluated since they are not observable. This makes evaluation of heterogeneity sometimes hard or impossible as not sufficient information is provided to deal with such an issue. Given these factors, in this work we will not consider heterogeneity as a whole, but we will at first study the effect of heterogeneity on single dimensions, where heterogeneity measures can be defined. This will make possible to evaluate the impact of heterogeneity on demand and so identify in which situations it can be a relevant issue. Moreover, since the main issue is tied to demand forecasting we will consider only some heterogeneity dimensions, that typically influence directly demand pattern. In particular, we will consider the existence of heterogeneity on customers’ dimension (in terms of total purchases), on lot-sizing politics, on purchasing lead-times and on purchasing politics, in terms of promotional activities.

These analyses are developed by means of simulation. In particular the general structure of these simulations are as follows. First we simulate demand patterns for different customers with the required peculiarities, in terms of heterogeneity. Then we evaluate heterogeneity on the specific dimension analyzed and study its relationship with demand variability. To properly measure these two dimensions we adopted different measures mainly tied, for what regards heterogeneity, to demand dispersion, and, for variability, to the coefficient of variation. However, analysis were conducted also adopting other heterogeneity measure (in particular the Herfiendal and the Entropy indexes), obtaining similar results. Thus, here analysis are provided only in terms of demand dispersion.

2.1 Customer Size

A first cause of demand heterogeneity may be tied to differences in customers’ size, in terms that their total demand is different. Bartezzaghi et al. (1999a) in fact analyze this

104

Part II Chapter 2: Theoretical Framework

situation and its impact on demand variability, however we extended their analysis. We considered the total demand generated by 20 customers. Each customer is distributed according to a Poisson distribution and is characterized by a different total demand size. In particular, we considered 5 possible sizes of total demand equal to 10, 50, 100, 500 and 1000 units. We simulated demand over a 200 days periods and evaluated the coefficient of variation of demand and the dispersion of customer size. Results are summarized in Figure 7.

0,9

0,8

0,7 CV 0,6

0,5

0,4

0,3 0,2 0,25 0,3 0,35 0,4 0,45 Dispersion of customer size

Figure 7: relationship between demand variability, measured by the coefficient of variation (CV) and the dispersion of customer size

As it can be noted there is a linear correlation among CV and dispersion due to the fact that dispersion tends to be very high when some huge customers show up. This kind of relationship is verified by Table 1 that reports the linear regression analysis between these two dimensions.

Term Estimate Std Error t Ratio Prob>|t| RSquare 0,880736 Intercept 0,0235792 0,012817 1,84 0,0670 RSquare Adj 0,880266 Dispersion 2,120259 0,048956 43,31 <.0001 Table 1: linear regression analysis between CV and dispersion

These results report a rather simple and empirical effect: if we have an homogeneous market (in terms of demand size) this tends to be easily forecasted since single customers’ variability is not particularly significant. As a matter of fact it doesn’t matter whether customers are big or not, since in this situation it is all a matter of total demand size. However, when customers show different dimensions demand becomes more variable

105

Part II Chapter 2: Theoretical Framework

since among day with regular demand (when small customers’ orders are placed) other days show huge peaks of demand (when huge customers order).

2.2 Purchasing Politics

A possible cause of heterogeneity among customers is tied to their purchasing politics. In fact customers may differ according to whether they tend to order with different lot-sizing rules. This can be due to many different reasons: from one side customers may have different purchasing cost structures, thus influencing the size of the Economic Order Quantity. In fact, if reorder costs are higher this pushes the EOQ size up, while high inventory holding costs pushes it down. This typically leads to situations where single customers may be very “regular” in their purchases, as they tend to order for similar quantities over time, while total demand may be variable as different patterns overlap.

A similar problem may arise when intervals between successive orders are different, since regularities in demand are more difficult to be found, while single customers may be very regular in their specific pattern. Here we consider these two situations, in which heterogeneity is at first applied only on their lot sizing rules and then it applied only on their reorder intervals. In particular, we separately simulate the impact of heterogeneity on the two variables introduced so to evaluate their impact on demand.

2.2.1 Lot Sizing

In this analysis we simulated the behavior of 4 distribution centers that face demand from end customers distributed according to a Poisson demand. We analyzed total demand generated by the 4 distributors.

Figure 8 describes the situation considered.

Observation Level Distribution Centers

Customers

Figure 8: structure of the simulation considered

106

Part II Chapter 2: Theoretical Framework

The distributors apply a simple (Q, r) inventory politic based on a reorder size of Q and a reorder level of r. To make comparison fair we considered that the total demand for the 4 distribution centers are equal in terms of total sum, while they may vary according to the specific distribution associated. We simulated the effect of different EOQs by varying each distribution centre Q, while r was fixed for each one. We assumed that Q could vary among 4 different values equal to 50, 100, 200 and 300. We run 40 different combinations of Qs and evaluate heterogeneity and total demand variability. To measure demand variability we considered two dimensions: the Coefficient of Variation of Demand (CVD), evaluated as the ratio of total demand Standard Deviation and its Mean, and the Coefficient of Variation of Order size (CVO). These two dimensions are identically only when demand shows up each day, while, if demand has any sporadic patter the two dimensions differ. We decided to adopt these two different measures since they contribute differently to demand variability understanding: in particular CVD gives an evaluation of global variability, while CVO only focuses on demand size variability and is not influenced by sporadic demand. Figure 9 and Figure 11 show the relationships among CVD and the two measures of heterogeneity.

2,5

2 B 1,5 CVD

1 A

0,5

0 0,4 0,45 0,5 0,55 0,6 0,65 0,7 0,75 Dispersion of demand

Figure 9: relationship between variability of demand and its dispersion

As it can be noted an interesting relationship exists among these two measures, since, for a given value of the heterogeneity index (i.e. 0.5) different values of CVD can be identified. This is due to the fact that the different points evaluated represent different clusters of lot values. For example, point A in Figure 9 refers to the situation when all customer have an EOQ equal to 50, while point B when all customers have an EOQ equal to 100, and so on. This means that these situations, that from a heterogeneity point of view are similar, appear differently in the evaluations of variability since the greater the EOQ value, the greater the

107

Part II Chapter 2: Theoretical Framework

variability. For this reason we also considered the evaluation of CVO, since we argue that similar heterogeneity should lead to similar variability.

However, quite interestingly Figure 10 shows that when heterogeneity rises the coefficient of variation tend to “stabilize” on a particular value. This is due to the fact that, if we are in a homogeneous situation when lot sizes are small (point A) and we rise uncertainty, this may be done only by having some customers order for bigger amounts of materials, thus moving towards point C as provide in Figure 10 (line 1). However, if we consider a situation where market is homogeneous but characterized by large customers (point D) and we try to make it more heterogeneous, we have to reduce each customer order thus moving towards point C as provide in Figure 10 (line 2).

These results indicate that a certain amount of heterogeneity may lead demand towards higher variability, however when heterogeneity is very high, the global “entropy” of the system is so high that a compensation effect on variability takes place. In other terms, if customer’s variability is very sparse, total demand variability may be reduced.

3 D 2,5 2 2 B C 1,5 CVD 1 1 A

0,5

0 0,4 0,45 0,5 0,55 0,6 0,65 0,7 0,75 Dispersion of demand

Figure 10: relationship between variability of demand and its dispersion

This idea is confirmed by Figure 11 that represents the dispersion of EOQ and CVO. As it can be noted the relationship is much less dispersed and tends to fit a polynomial pattern. It can be seen that for low levels of demand dispersion the variability in the orders size is low, while for greater values it tends to become much higher. However this relationship is not completely linear since, for greater values compensation effects tend to be predominant, thus demand tends to be more regular.

108

Part II Chapter 2: Theoretical Framework

In fact, by comparing Table 2 that reports results for the linear regression analysis with Table 3 that reports the polynomial regression results, it can be noted that this second assumption seems to fit better.

1,6

1,4

1,2

0,8 CVO 0,6

0,4

0,2

0 0,40,50,60,70,80,9 Dispersion of demand

Figure 11: relationship between variability of orders’ size and demand dispersion

Term Estimate Std Error t Ratio Prob>|t| RSquare 0,609044 Intercept -1,56424 0,203738 -7,68 <.0001 RSquare Adj 0,604031 Heterog 3,9886185 0,361839 11,02 <.0001

Table 2: linear regression analysis between variability and dispersion

Term Estimate Std Error t Ratio Prob>|t| RSquare 0,724064 Intercept -2,624237 0,254331 -10,32 <.0001 RSquare Adj 0,716897 Heterog 5,999915 0,468662 12,80 <.0001 (Heterog-0,56028)^2 -21,36374 3,77093 -5,67 <.0001 Table 3: polynomial regression analysis between variability and dispersion

This analysis shows that heterogeneity in the lot sizing among customers may influence demand variability. In fact, when heterogeneity rises demand becomes more variable thus making forecasting more difficult, however, even if for higher values of heterogeneity demand tends to become less variable, we argue that this doesn’t make forecasting more simple, since even if demand variability is reduced, demand may tend to be more sporadic thus making forecasting tricky.

109

Part II Chapter 2: Theoretical Framework

2.2.2 Reorder Interval

Another possible source of variability is tied to the existence of different reorder intervals among customers. In fact we considered a situation that faces 20 customers that reorder for similar quantities (i.e. 100 units) but with different reorder intervals. We considered 5 possible reorder intervals equal to 5, 10, 15, 20 and 30 days. In these terms heterogeneity is given by the dispersion among customers of the mean interarrival period between each customers successive orders. We conducted a simulation during 300 days of demand and measure the coefficient of variation of total demand, as reported in Figure 12.

1,2

0,8

0,6 CV

0,4

0,2

0 0 0,02 0,04 0,06 0,08 0,1 0,12 Dispersion of the interrarrival time

Figure 12: relationship between demand variability and dispersion of interarrival time

As it can be noted again when heterogeneity is low demand tends to be quite stable, however when dispersion rises demand tends to become more variable at first, but after a certain level of dispersion it becomes less variable due to compensation effects.

Again we analyzed the polynomial regression of these two variables, thus identifying a significant relationship. Table 4 summarizes these results.

Term Estimate Std Error t Ratio Prob>|t| RSquare 0,250179 Intercept 0,3427295 0,070849 4,84 <.0001 RSquare Adj 0,235896 Heterog 5,8007797 1,095354 5,30 <.0001 Heterog^ 2 -347,0846 63,37483 -5,48 <.0001

Table 4: polynomial regression between variability and heterogeneity

110

Part II Chapter 2: Theoretical Framework

2.3 Promotional Politics

Another possible source of variability is tied to promotional activities. In fact, in many industrial sectors, mostly tied to the retail industry, promotions are always applied to influence short term demand. Some experiences in different contexts have shown that demand during promotional periods may rise up to 1000%. Figure 13 shows an example taken from a company producing household cleaning goods. The example reports the total demand (measured in terms of number of cartons at weekly level) for a specific product. As it can be noted among days with low and very stable demand some huge peaks show up. Detailed analysis showed that demand peaks were due to a single customer’s requests.

9000

8000

7000

6000

5000

N. CT N. 4000

3000

2000

1000

0 200036 200038 200040 200042 200044 200046 200048 200050 200052 200102 200104 200106 200108 200110 200112 200114 200116 200118 200120 200122 200124 200126 200128 200130 200132 200134 Weeks

Figure 13: example of demand series affected by promotions

The reasons for applying promotions are many: from one side they influence sales volume and even if prices are reduced, the total effect on sales is marginally positive; moreover, promotions push people to buy a specific brand instead of competitors’ ones. Even if it is quite well known that promotions may deeply affect demand variability, we will take into consideration a few interesting aspects. In fact we want to analyze what is the impact on demand variability of heterogeneity in how customers react towards promotion. This was done by evaluating heterogeneity on two dimensions: from one side we considered the number of promotions applied by customers, so modeling heterogeneity on the different number of promotions each customer conducts; then we considered also the reaction of customers towards promotions, as different customers may react differently towards similar politics.

111

Part II Chapter 2: Theoretical Framework

2.3.1 Number of Promotions

To model the impact of the heterogeneity of the number of promotions applied by customers, we modeled a situation in which 20 customers are considered. They all tend to reorder each day of the simulation period, however, during some periods they apply promotions towards end-customers, thus making their request towards suppliers higher. Their stable demand is distributed according to a Poisson distribution, while promotional activities are applied according to a Bernoulli distribution. We simulated different conditions by means of changing randomly the number of promotions that each customer placed during the simulation period of 300 days. In particular possible values of the number of promotions are 5, 10, 15 and 20. We measured the coefficient of variation of total demand generated towards the supplier and the dispersion of the number of promotions among customers. Figure 14 shows the relationship between these two variables.

0,9

0,85

0,8

0,75 CV

0,7

0,65

0,6 0,2 0,25 0,3 0,35 0,4 0,45 0,5 0,55 0,6 Dispersion of the number of promotions

Figure 14: relationship between demand variability and dispersion among the number of promotions

As it can be noted the results are similar to those provided by other previous simulations. In particular, for low dispersion variability tends to be low, while as heterogeneity rises also variability rises. Again, however, when heterogeneity is very high, variability tends to reduce since the entropy of promotional activities is very high thus inducing compensation effects.

Table 5 summarizes the results of the polynomial regression analysis.

112

Part II Chapter 2: Theoretical Framework

Term Estimate Std Error t Ratio Prob>|t| RSquare 0,377946 Intercept 0,6610618 0,019376 34,12 <.0001 RSquare Adj 0,361789 Heterog 0,2536418 0,051621 4,91 <.0001 (Heterog-0,39032)^2 -4,156337 0,649501 -6,40 <.0001

Table 5: polynomial regression between variability and heterogeneity

2.3.2 Promotion size

Another relevant dimension on which heterogeneity may be relevant, is tied to the specific reaction that customers may have at particular promotional activities. In fact customer may differ also for their sensibility towards promotional activities, as for example some customers tend to purchase marginally more in promotional periods compared to other customers.

We modeled this situation considering 50 customers, which have a stable demand during non promotional periods distributed according to a Poisson distribution, while promotions over customers are distributed according Bernoulli distributions. When a promotion is applied each customer reacts with a different size of demand, thus influencing differently total demand variability. In particular, we assumed that promotions size may vary among 5 possible values: 20, 40, 60, 80 and 100 units. We assumed that in the 300 days period of simulation customers apply 10 promotional activities. Again we measure the coefficient of variation of total demand and dispersion of the different promotion sizes. The relationship between these two variables is provided in Figure 15.

0,65

0,6

0,55 CV 0,5

0,45

0,4 0,12 0,13 0,14 0,15 0,16 0,17 0,18 0,19 Dispersion on promotion size

Figure 15: relationship between variability and dispersion of promotion size

113

Part II Chapter 2: Theoretical Framework

As it can be noted there seems to be a rather linear relationship which in fact seems to be significant (see Table 6). This may appear in contrast with the analysis provided in section 3.2, since in the actual analysis no compensation effects seem to turn out. However, this is due to the fact that since promotional activities are sporadic, compensation effects don’t show up, thus heterogeneity only influences demand in terms of higher variability.

RSquare 0,420698 Term Estimate Std Error t Ratio Prob>|t| Intercept 0,1910549 0,026449 7,22 <.0001 RSquare Adj 0,418212 Heterog 2,2085202 0,169782 13,01 <.0001

Table 6: linear regression between variability and heterogeneity

114

Part II Chapter 2: Theoretical Framework

3 Heterogeneity and Forecasting

Given the results of the previous analysis some considerations can be drawn. First of all demand variability is influenced by heterogeneity in customers purchasing process. As it can be noted, this impact is influenced by the kind of heterogeneity considered and the way in which heterogeneity is derived. In fact, in some conditions, when heterogeneity is low, variability tends to be relatively low too, however, when heterogeneity rises variability rises too leading to a sort of maximum variability condition, after which greater heterogeneity impact by reducing demand variability. This is due to the fact that when heterogeneity is very high some compensation effects take place thus making demand less variable (see for example paragraphs 2.2.1, 2.2.2 and 2.3.1). In other situations (see paragraphs 2.1 and 2.3.2) the higher heterogeneity, the higher variability. A second relevant consideration is that heterogeneity doesn’t influence demand variability always in the same way; it depends on how heterogeneity is applied to demand. For example, if we have many small customers and few huge ones, heterogeneity can be pretty high and so can be demand variability. On the contrary, if we consider many huge customers and a few ones, heterogeneity can be high again, but demand variability is lower since compensation effects take place. So, a first interesting result is that under some conditions demand variability may be influenced by heterogeneity. However, even if under some conditions demand heterogeneity may influence positively demand variability, when compensation effects occur, however this does not mean that demand forecasting is simple. If variability is low, simple forecasting techniques may be applied, however this is true only if the performances of these systems are acceptable for the particular problem considered. As literature shows, given a certain forecasting problem, in terms of a certain demand variability, statistical techniques may be applied at different levels of sophistication. For example, one may apply a simple exponential smoothing, then take into account seasonality, trend, environmental conditions, and so on, thus making the forecasting system more sophisticated or, in other terms, capable of managing different sources of systematic variability. However, at a certain level, when all systematic variability is managed by the forecasting system attention has to be paid towards the residual variability that is tied to customers’ specific behavior. In this situation, if demand is homogeneous, then probably few factors will have to be considered, on the contrary, if demand is heterogeneous, many different variables will have to be considered. In these terms, heterogeneity may for sure reduce demand variability, but, in some conditions, it doesn’t make demand forecasting simpler.

This is ever truer as attention towards inventory performances rises, since, as provided in Chapter 2, demand forecasting performances directly reflects in inventory costs and service

115

Part II Chapter 2: Theoretical Framework

levels. So, in many industrial contexts a ever stronger attention towards forecasting performances is paid, thus making at a certain point critical to focus on customers’ variability. In this scenario demand heterogeneity becomes a critical matter.

Consider a certain forecasting problem in terms of demand variability (see Figure 16); we can describe a certain forecasting solution in terms of its accuracy and its complexity. Accuracy1 refers to how much the system is capable of identifying demand variability, so it measures its quality; complexity refers to how much the system is difficult to be implemented and thus it can be measured in terms of the forecasting system costs, the organizational efforts needed to apply it, the algorithmic complexity of the solution and so on.

Homogeneous Market C D Forecasting Forecasting Accuracy

B Heterogeneous Market

Systematic Customers’ Forecasting A variability variability Complexity

Figure 16: the impact of market heterogeneity on forecasting complexity and accuracy

When the forecasting system is very simple (i.e. a simple moving average), forecasting accuracy is poor (point A), however it is very simple to perform improvements in forecasting, since improvements are “cheap”. This at first is very effective, thus moving as showed towards higher levels of accuracy. However, the more we move on this curve, the less we can improve performances as we are less and less able to capture relevant sources of variability. At a certain point, when all systematic variability is managed and the more complex system available is developed (point B), to improve performance we have to shift attention towards customers and their influence on variability. However, if heterogeneity is low (homogeneous market) this is a rather simple matter since few variables have to be managed; on the contrary, if heterogeneity is high, it becomes a tricky operation since

1 We will not define in deeper detail the concept of accuracy. We refer to Makridakis (1993) and Armstrong and Collopy (1992) for general review of the concept of forecasting accuracy and its measures.

116

Part II Chapter 2: Theoretical Framework

many different variables have to be considered (heterogeneous market). We have, for example, customers that apply promotions and others that don’t, but promotional customers may react differently towards similar activities; we may have customers that purchase differently, and so. Thus in a heterogeneous context forecasting improvement becomes a rather complex operation. Compare for example point C with point D: they show that to gain a certain forecasting accuracy more complexity has to be managed.

When dealing with heterogeneous demand, forecasting faces two main sorts of problems. From one side different customers are influenced differently from similar factors. In these terms it is more difficult to evaluate properly the influence of a certain variable on a specific customer since they behave differently. For example, different customers react differently to similar price changes, to promotional activities, to weather conditions and so on. This makes the estimation process complex since, even if the forecasting structure may be feasible with all customers considered, the estimation process cannot be conducted for all customers together, otherwise the estimation would be biased by the different behaviors. In fact marketing literature previously introduced regarding heterogeneity faces this problem: when customers react differently (in terms of brand loyalty) to similar factors, this heterogeneity has to be taken into account. This should focus attention more on each single customer, however this solution has two main drawbacks: first of all the cost and complexity of estimating each single customer may be very high, thus making benefits less valuable. Moreover, if estimation is provided at a customer level it may be biased by the lack of sufficient information. Think for example to the problem of estimating the sensitivity towards promotional activities. If we try to make this evaluation at a customer level we may find that many customers perform very few promotional activities thus reducing the reliability of the estimation. From the other side if we try to perform it once for all customers, this solution may be not effective if heterogeneity exists on customers’ reaction towards promotion. The major effect is that not considering heterogeneity in the estimation process produces inaccurate forecasting performances, but this may also apply when attention is paid to each single customer separately. This idea is represented in Figure 17 in terms of forecasting ability, in other words, the ability of the forecasting system of catching and managing variability. In fact, to properly foresee demand, a forecasting system needs from one side to catch variability so to be able to identify its sources. From the other side, caught variability has to be managed, in terms that it has to be properly estimated. In fact, if forecasting is conducted at a aggregate level, so without considering the different characteristics of customers, the forecasting system is not capable of understanding too much variability, while it is effective in managing it since few information is required. From the other side, if forecasting is applied at a customer level, the forecasting system is

117

Part II Chapter 2: Theoretical Framework

capable of identifying variability, while it may be not effective to manage it since the system tends to be too complex.

Forecasting ability

Ability to catch variability

Ability to manage variability

Aggregate Disaggregate Forecasting (all customers (all customers level together) separately) Figure 17: relationship between forecasting ability and forecasting level

We argue that the solution to this dilemma is to identify the proper level of analysis and to use heterogeneity at our own benefit.

From another side different customers may also be influenced by different factors. In fact we may have to deal with customers that tend to use promotional activities, while others don’t apply promotional activities at all. At the same time some customers may purchase almost every day or week for limited amount of materials, since, for example, their reorder costs are low, while others may prefer to purchase rarely for huge amounts. In these terms, customers’ requirements may not be properly forecasted by means of a single technique, as different patterns are associated to them. This may be a relevant problem, as a single forecasting technique does not apply properly to each customers. Moreover applying different forecasting methods for all different customers may be too difficult, and again may encounter problems in the estimation process.

Given these factors we argue that when heterogeneity rises forecasting should rise its heterogeneity too. In other words to deal with heterogeneous demand a heterogeneous forecasting approach must be adopted, in terms of different forecasting techniques. However since again a trade-off exists between estimation reliability and effectiveness, the different forecasting approaches have to be developed according to the specific factors of demand heterogeneity.

118

Part II Chapter 2: Theoretical Framework

As previously anticipated, heterogeneity can be due to three main causes: first of all heterogeneity may be associated directly with customers as different customers may react differently to similar influences. From another side heterogeneity can be due to the existence of particular politics the considered firm applies to particular classes of customers. Think for example to promotional activities, which are conducted with only some customers and typically in different ways according to the particular customer. Moreover, also environmental factor may influence heterogeneity, as for example weather conditions may change the way in which customers purchase. However, we argue that a relevant factor must be also considered when analyzing customers’ heterogeneity: the supply chain structure. In fact having a simple, linear supply chain forces firms to deal essentially with the customers’ inner heterogeneity, since, as the supply chain is rather uniform, customers belonging to the same supply chain level are managed. However, if the supply chain is rather complex, as for example different distribution channels are adopted or as different levels of supply chain are directly managed, this tends to make customers more heterogeneous since their inner characteristics will be very different as they belong to very different realities.

So we argue that the complexity of the supply chain can deeply influence customers’ heterogeneity and that particular attention has to be paid in its management. Complexity of the supply chain can be defined as the heterogeneity of both the supply levels considered and managerial activities adopted. Complex supply chains may show from one side many distributional channels that co-exist simultaneously and, from the other, different managerial politics applied, as for example promotional activities. In fact, all the context factors previously introduced shift supply chain towards more complex structures and so more complex managerial systems.

In these terms possible measures of complexity may be as follows: Number of distributional channels managed: in fact the more different channels are adopted together, the more customers belonging to different chains behave differently and are influenced by different factors. Number of different supply levels controlled: firms may have to manage customers that belong to different levels as for example end customers, retailers, distributors and so on. Number of managerial politics applied: many firms apply promotional activities towards customer; these activities may differ according to the customers considered.

So the general framework we consider pays attention from one side towards inner heterogeneity of customers and from the other towards supply chain complexity.

119

Part II Chapter 2: Theoretical Framework

As Figure 18 shows heterogeneity may be relevant also when complexity is low (point A versus B), since we may have customers behaving differently due to their inner characteristics. However, when complexity rises we argue that heterogeneity may not be below a certain level, since different customers are joined together (point C), and, of course, they may also directly contribute to rise more heterogeneity due to their specific characteristics (point D).

Given this framework we argue that two main approaches can be applied when these two dimensions are relevant. When inner heterogeneity has to be managed, customers tend to be influenced by the same variables but their sensitivity towards these factors may be different from one customer to another. In this way probably a single forecasting structure may be adopted for all customers, but since heterogeneity appears in their reaction towards external phenomena, the proper level of analysis has to be identified. In other words, different evaluations have to be considered for single customers. However again focusing on each single customer may not be most suitable methodology, so the level of analysis has to be properly identified.

D Inner B Heterogeneity Heterogeneity C

Supply Chain Complexity

Figure 18: perceived heterogeneity root causes

However, when supply chain complexity tends to be relevant, differences among customers are not only in their inner characteristics, but also in the kinds of variables that influence their demand patterns. In other words, while previously the main forecasting structure holds among customers, here different customers should be forecasted adopting different methodologies that focus on demand heterogeneity factors. In these terms we argue that in this situation different forecasting approaches have to be adopted, focusing on different variables.

120

Part II Chapter 2: Theoretical Framework

In this way, we argue that when heterogeneity in demand rises, also forecasting has to be applied more heterogeneously. This idea is also consistent with most of the literature regarding the value of information in demand forecasting, that state that attention has to be paid to the demand generation process. However, we consider that relevant consideration has to be given towards the heterogeneity of the different demand generation processes that overlap in heterogeneous contexts.

Figure 19 summarizes this concept thus providing the general framework of the solutions provided.

1 forecasting structure

Inner Different estimation Different forecasting structures Heterogeneity Heterogeneity processes

Supply Chain Complexity

Figure 19: general forecasting framework

121

Part II Chapter 2: Theoretical Framework

4 Conclusions

First of all the provided results claim that Hypothesis H1 can be accepted since the previous analyses show that heterogeneity can deeply influence demand variability and that even when demand is not much variable, demand forecasting may be rather tricky if heterogeneous contexts are faced.

Moreover these analyses let us identify some relevant elements that the overall forecasting methodology should take into account. We can summarize the previous arguments according to the following research hypothesis:

H2. Supply chain complexity deeply influences demand heterogeneity and thus demand forecasting approaches.

H3. When heterogeneity becomes relevant, a heterogeneous forecast has to be adopted.

H4. If heterogeneity is due to the supply chain complexity, then different forecasting approaches have to be adopted.

H5. If heterogeneity is due to customers’ inner characteristics, different estimation processes have to be adopted.

Given this framework, our main issue is now twofold: from one side we want to test our hypotheses empirically, moreover it is fundamental to define proper approaches towards forecasting demand within heterogeneous contexts. These issues are considered in the next part by means of the action research cases developed.

122

PART III:

EMPIRICAL ANALYSIS

Chapter 1 Introduction to the Action Research Cases

As previously anticipated, we focus on three cases belonging to different industrial sectors. Details regarding the different situations considered and the specific problem faced are provided in the different parts. In fact in all the three cases’ descriptions, we try to introduce the problem considered, then when available we report the analysis that led us to identify the root causes of the problems faced by the companies, then solutions are developed and tested according to real demand data provided by the different firms and conclusions are drawn. A relevant remark is that in all the three cases the solutions proposed were effectively implemented thus stating that the methodologies developed were significant for the companies considered, thus claiming for relevance both for researchers and practitioners. In each case, however considerations regarding the generalization of the methodologies are conducted and analysis of the applicability of these techniques are taken into account.

Given the previous methodological remarks, we decided to focus on firms facing the short- term demand forecasting problem that have to deal with great heterogeneity among customers. We argue that this issue isn’t market- or product- specific but rather specific to the nature of the problem described. In these terms, we did not limit our analyses to a

125

Part III Chapter 1: Introduction to the Action Research Cases

particular industrial sector but looked widely to manufacturing and distributing companies, so companies for which demand forecasting is a major issue.

The choice of industries and companies was also influenced by the difficulty in finding firms with such problems and available for solving them. In fact the analysis was conducted among 3 companies belonging to different sectors. Table 1 briefly describes the companies involved, while we refer to chapter 5 for a detailed description.

Ahold Nestlè Italiana Whirlpool Europe Industrial sector Retail Food & Beverage White goods Native country Holland Switzerland USA Country Holland Italy Italy Sales 66 billions € 47 billions € 10 billions € Products Grocery Fresh food Spare parts

Table 1: the cases considered in the research process

The different cases considered refer to different structural situations. In particular, in the different cases the effect of customer’s inner characteristics and supply chain complexity are different. In fact, in the Ahold case, the supply chain is made of different stores owned by the company itself, that effectively belong to the same supply chain level. Similar promotional politics are adopted in the different stores and, moreover, promotions are conducted at a national level. In these terms the supply chain is rather simple, while the main cause of heterogeneity can be associated to the customers’ inner characteristics.

In the Whirlpool Europe case, instead, the main source of heterogeneity can be traced back to the supply chain structure. As it will be detailed in the case description, the firm has a very composite supply chain that serves directly both end customers and other intermediaries distributed on different levels of the supply chain. Thus the difference in the inner characteristics of the different customers can be traced back to the supply chain complexity.

At last the Nestlè situation shows the two elements together, since from one side the supply chain complexity is relevant as different distributional channels are adopted and different managerial politics are applied. Moreover, significant is also the difference among the different customers that from one side react differently towards promotional activities and from the other show different reactions towards environmental elements.

126

Part III Chapter 1: Introduction to the Action Research Cases

Figure 1 maps the different cases according to the supply chain complexity and the relevance of the inner characteristics.

Nestlè Italiana Inner Ahold Heterogeneity Heterogeneity Whirlpool Europe

Supply Chain Complexity

Figure 1: the distribution of the considered cases

In fact, we considered different real contexts, in which heterogeneity shows up in different ways. In particular, the Whirlpool case is considered. This situation shows a relevant heterogeneity in the purchasing process of different customers: from one side we have many small customers that tend to order for small quantities, from the other we face few huge customers that order relevant quantities of materials. The overlap of these demand patterns makes demand lumpy and thus the actual forecasting system the firm adopts rather inefficient. As it will be further described, the solution designed leverages on the characteristics of heterogeneity among customers.

From another side, the Nestlé case focuses on two main sources of problems: from one side the supply chain complexity makes customer heterogeneous due to differences in promotional activities: some customers perform promotions while others don’t. Moreover, significant differences rely also in the reaction that customers have towards both promotional activities and external variables. In this way the main problem is dealing with the different demand patterns these customers have and dealing with the variability they introduce in total demand.

At last, in the Ahold case, the main problem is tied to the different reactions customers have towards external and firm’s decisional factors. As it will be further described,

127

Part III Chapter 1: Introduction to the Action Research Cases

customers in this situation are influenced by similar factors as promotional activities, price, weather conditions and so on, however different customers react differently towards these variables, making estimation rather tricky.

Figure 2 summarises the heterogeneity factors considered in the different situations.

Heterogeneity

Sensitivity towards Promotional Reorder politics environmental activities factors Heterogeneity Dimensions

A N W

Action Researches

Figure 2: the heterogeneity factors considered in the research

Few considerations have to be introduced on these cases: 1. Since the problem of managing demand heterogeneity is not industrial specific, we considered experiences in different sectors: Ahold provides an example in the retail industry, while Nestlè belongs to the fresh food sector and Whirlpool to the white goods spare parts business. This claims for the generalization of the methodology proposed. 2. All the situations considered are real contexts and analyses have been applied by means of real demand data. This claims for the relevance of the problem and of the methodologies developed both for researchers and practitioners. 3. All the solutions described have been effectively developed in the different firms. This again claims for the relevance of the solutions provided in real contexts. 4. Even if the solutions developed are designed in specific situations, their applicability is not industry-specific but rather problem-specific. So we argue that they are applicable also in different contexts and, in the different studies, we discuss to what extent they can be generalized. 5. Since some of the information adopted in the different cases go in deep detail into the operations of the different firms, in some cases we were asked by the firms not to

128

Part III Chapter 1: Introduction to the Action Research Cases

publish some kind of information. In some parts, in fact, we were not able to provide deep details on some elements.

In particular, at first the Whirlpool Europe case will be presented, then the Nestlé Italiana situation will be analyzed and at last the solution developed for Ahold will be studied. In the next chapter, the results of the different studies will be summarized and general considerations will be drawn.

129

Part III Chapter 1: Introduction to the Action Research Cases

130

Chapter 2 Whirlpool Europe

1 Context and Purpose

In this case we consider a firm involved in the spare parts business facing problems with the demand forecasting process. In particular, as further described, the root causes of the problems the firm faces can be traced back to the existence of great heterogeneity within the demand generation process due to the overlap of two kinds of customers: from one side many small customers order low amount of materials, while from the other few huge customers create real demand peaks. The problem of managing spare parts demand is relevant for many reasons: first of all it influences the final products business since it influences post sale service quality. Moreover, it is a relevant business as the market is captive, thus firms pay a relevant attention towards this issue. However, it is a very difficult business to cope with, since requirements are usually very dispersed over time.

Many authors have in fact dealt with the problem of managing demand in the spare parts business. Petrovic et al. (1988) present an algorithm for evaluating lot sizes to minimize total inventory costs. The model assumes that demand in a certain period can be represented as a sequence of single unit requests, so the considered period can be

131

Part III Chapter 2: Whirlpool Europe

considered as a sequence in which demand may or may not show up. The probability of receiving and order in a certain period is considered as distributed according to a Beta distribution. The system evaluates the cumulative probability of having a certain amount of orders in the next planning horizon, so making possible to evaluate the best inventory management politic. Shibuya et al. (1998) compare two politics in a continuous review environment applied to spare parts demand, under the assumption of a Poisson distribution. In the first model, if the system runs out of stock before a predetermined time, an urgent order is emitted at the same time. In the second model, the stock level is controlled during certain periods and if the system is out of stock an urgent order is determined, otherwise a normal order is emitted. Other similar approaches are proposed by Cobbaert and Van Oudheusden (1996), Liu and Shi (1999), David et al. (1997), Teunter and Fortuin (1999), Cohen and Kleindorfer (1989), Dekker et al. (1998).

In this situation we provided the development of a new forecasting solution that leverages on demand heterogeneity and tries to focus on the peculiarities of customers’ demand. It can be anticipated that the main objective of improving inventory performance will be achieved by adopting a heterogeneous solution to manage the heterogeneity of demand. In particular, this part is structured as follows: first the company will be described in terms of products, market and internal structure, then attention will be paid on the supply chain structure and its impact on demand variability. Then the solution proposed will be described and at last comparisons among the possible solutions considered will be analyzed.

132

Part III Chapter 2: Whirlpool Europe

2 Introduction to the company

Even if actually Whirlpool Corporation is a worldwide leader company, it has started (as many of its local divisions) as a small society located in a small town. In fact, in 1911, in St. Joseph (Michigan) Upton Machine Co. is founded, producing washers with electric engine. In 1929, Upton mergers with the Nineteen Hundred Washer Co. In 1916 the first order for washers’ sales in Sears, Roebuck and Co. is signed. In 1919 Gottlob Bauknecht opens in Tailfingen, Germany, a small electric workshop that in 1993 will become its first factory. The Bauknecht brand will be acquired by Philips in 1982. During 1943: Guido Borghi and his three sons found in Comerio (Varese) the Ignis society and its brand, that both will be acquired by Philips in 1972 and afterwards by Whirlpool in 1989. In 1950, Nineteen Hundred Corporation changes its name and becomes Whirlpool Corporation. By the same time the automatic drier lines is activated. In 1957, Whirlpool invests in Brasmotor S.A., for a stock participation in the Brazilian white goods market. It also constitutes Appliance Buyers Credit Corporation that will be the beginning of Whirlpool Financial Corporation. During 1968, the firm concludes the research centre of Benton Harbour; global sales are over 1 billion $. In 1969, Whirlpool buys a stocking participation in Inglis Ltd. (Canada), and it will acquire it all in 1990. In Italy, Ignis and Philips constitute I.R.E., (“Industrie Riunite Eurodomestici”, Union of Eurodomestic Industries) to cope with the rising evolution of the European white goods market. During 1970, Ignis and Philips start a joint venture and constitute a production plant for refrigeration in Trento (Italy). The production of microwaves ovens starts in Norrköping (Sweden), and that of dishwashers starts in Neunkirchen (Germany). In 1974, Philips becomes unique responsible of the joint venture with Ignis, locating its operational base in Comerio (Italy). During 1982, Philips acquires Bauknecht in Stuttgart (Germany), but with a separate management of activities from the Italian headquarters. In 1986, Whirlpool acquires the KitchenAid division of Hobart Corporation and buys from Fiat a major stocking participation in Aspera s.r.l., compressor producer. In 1987, Whirlpool constitutes in India the joint venture TVS Whirlpool Ltd with Sundaram Clayton; afterwards the plant of Pondicherry is built, aimed at producing automatic washers. In 1994 Whirlpool acquires the global control of the joint venture.

133

Part III Chapter 2: Whirlpool Europe

During 1989, a European joint venture is established with Philips that enters at third place in the European white goods producers. In 1991 Whirlpool acquires Philips equity and gains overall property. Sales are actually over 6 billions $. During 1990, Whirlpool Europe launches the Philips-Whirlpool brand program to consolidate the identity of Whirlpool name. A joint venture is created with the Japanese Matsushita Electric Industrial Co., aimed at producing vacuum cleaners for the North- American market. The Estate brand is also created for national customers. In 1991, Whirlpool and Tatramat constitute a joint venture for washers’ production in Poprad. Whirlpool will gain control in 1994. Between 1992 and 1993 new subsidiaries are created in Hungary and Poland. In 1994 Whirlpool becomes an independent brand in Europe. During 1995, Whirlpool consolidates its presence in Asia, by means of the acquisition of major equity participations in two joint ventures: Kelvinator in India (New Delhi) for refrigerators production and Whirlpool Narcissus Co. (Shangai) for automatic washers. Another joint venture starts for microwaves ovens production under the name of Whirlpool SMC Microwave (Shunde) Products Co., Ltd., in the Guangdong region. In 1996, Whirlpool Europe acquires the distributing and producing facilities of Gentrade (South Africa) and opens subsidiaries in Romania e Bulgaria. During 1997, Whirlpool Corporation acquires the major part of Brasmotor S.A. (South America). In 1998, Whirlpool starts a new global plant to produce no-frost refrigerators in Pune (India). In Europe, it becomes major supplier of IKEA. Moreover, all eastern control is shifted to Europe.

Actually Whirlpool gains overall sales for more than 10.5 billions $ (1999), with earnings equal to 347 millions $ (1999). As it can be seen from the previous historical presentation, Whirlpool produces and sells various white-goods products in many different countries (actually over 140). Figure 1 shows the distribution of sales in the different countries.

134

Part III Chapter 2: Whirlpool Europe

Africa and Middle East; North 5% America; South 25% America; 8%

Asia, Australia Eastern and New Europe; 8% Zeland; 28%

Western Europe; 26%

Figure 1: Whirlpool sales distribution

Europe represents one of the most important markets but it also shows a great competitiveness due to the relevant number of competitors (see Figure 2).

Figure 2: white goods market composition

Whirlpool Europe (the European division) sells products with different brands. The most important are: BAUKNECHT: the most expensive brand typically associated with a “German” reliability. WHIRLPOOL: identifies products in a medium/high position with a reduced price compared to Bauknecht products. IGNIS: associated to medium/low position since even if quality is comparable to previous products, fewer accessories are provided.

135

Part III Chapter 2: Whirlpool Europe

The Whirlpool Europe directional centre is located in Comerio (Northern Italy) and here are conducted all operational activities as Marketing, Administrations, Finance and Accounting. Whirlpool Europe controls four plants in Italy (Cassinetta, Siena, Napoli and Trento) and seven in the rest of Europe (Calw, Schorndorf, Neunkichen, Amiens, Poprad, Isithebe and Norrköping). Figure 3 reports the territorial distribution of the plants controlled by Whirlpool Europe.

Norrköping

Amiens Neunkirchen Poprad Schorndorf Calw

Trento Comerio Cassinetta

Siena

Napoli Isithebe

Figure 3: distribution of Whirlpool plants

2.1 Whirlpool Europe Spare Parts Centre

The Spare Parts sector represents a very important activity for different reasons: from one side it influences directly its economical performances since it represents 50 millions €, from the other it influences deeply Whirlpool’s core business, as it influences directly the overall company performances. This sector is managed by the Customer Service & Spare Parts Centre, located in Cassinetta that controls all customer service in all Europe.

136

Part III Chapter 2: Whirlpool Europe

The number of SKUs managed is very high, since more than 85.000 items are usually active for an overall stock value of 24 millions €.

More than 200 people control and direct all the operations connected to the spare parts business, as Figure 4 describes. The most relevant operational activities here conducted are those that support the spare parts logistics.

Director

IT Support Director Assistant

ERP Project

Store Spare Parts Controller SPC Customer WEBV spare Technical Operations Procurement Assistance parts inventory Support M 3 people

12 White collars 120 Blue collars 12 people 2 people 8 people 4 people 2 people 30 Seasonal workers

Figure 4: Whirlpool Europe Spare Parts organizational structure

137

Part III Chapter 2: Whirlpool Europe

3 Analysis of the Supply Chain and Demand

To understand the problems faced by the company we focused on the analysis of the supply chain adopted and on demand variability. In particular, these two analyses were fundamental in understanding from two different points of view, the causes of the demand lumpiness the firms showed.

3.1 The Evolution of the Supply Chain Structure

Since a couple of years, Whirlpool Europe Spare Parts has started developing a relevant change in the distribution systems. In particular, since few years ago, spare parts were distributed by means of regional warehouses. In particular, the global stocking unit (MSH, Main Stock Holder) located in Cassinetta, controlled the supply of spare parts in all Europe, but distributes them to final customers through regional units distributed in all major European countries (see Figure 5). In particular, besides these regional warehouses some external wholesalers buy spare parts from the central warehouse.

Central Warehouse

Regional Warehouses Wholesaler

Retailers

Figure 5: supply chain structure before 1997

In 1997, Whirlpool Europe decided to close most of these regional warehouses and start the so-called Direct Delivery Project. The aim of this project was to reduce costs connected to the great number of stocking units. Their presence, in fact, deeply influenced both handling and stocking costs. The elimination of these stocking points made possible to concentrate activities in a single unit (MSH) with relevant impact on overall costs. However, even if in the initial stages of the project the aim was to eliminate all the regional warehouses, during the project development it was decided to keep some of them alive due

138

Part III Chapter 2: Whirlpool Europe

essentially to problems in reaching customers in the farthest regions of Europe (i.e. East Europe, Greece, Finland, etc.). So, compared to the initial plans, the actual logistic structure shows from one side direct customers served directly from Cassinetta, some Wholesalers that, since they don’t belong to the Whirlpool controlled supply chain, keep on their business, and some of the original regional warehouses. Figure 6 represents this situation.

Central Warehouse

Regional Warehouses Wholesaler

Retailers

Figure 6: actual supply chain structure

As a matter of fact, customers are very numerous and with very different sizes: we have almost 15.000 retailers (small repairers both belonging to Whirlpool and private), 200 wholesalers and 10 residual regional warehouses. The effects of this situation on spare parts demand is relevant, because this is deeply influenced by the heterogeneity of customers, so MSH now has to deal with a very high number of order line (more than 13.000 order lines per day) for very different amounts (for the same SKU orders can be from a few units up to thousands). Intuitively, the root causes of this phenomena can be traced back to the existence at the same level in the supply chain of many heterogeneous customers, from one side many customers requesting very few amounts of products and from the other few customers requesting huge amounts of the same products.

By getting rid of the regional warehouses, Whirlpool has eliminated the supply level that made demand rather stable thus obtaining a highly variable demand. The main effects of this change can be seen in the inventory level that, even if total sales have increased of few percentage points in the last year, the overall inventory level has risen in terms of 10s percentage points. Figure 7 shows this phenomenon.

139

Part III Chapter 2: Whirlpool Europe

Important is to consider that the central warehouse previously introduced has no control of the other warehouses belonging to the supply chain. So, the only variable the firm can control to adjust its inventory performances is its own stock level. Moreover, no information is available to the forecasting managers regarding the number of customers at each echelon, nor their specific reordering policies. In these terms both installation and echelon stock systems (see chapter 2 for a detailed description) are not applicable in this situation, as not enough information is available.

$. 53

$. 51 $. 49 I n $. 47 I I S v S n S n $. 45 h e h v h v i n $. 43 i e i e $ (Millions) p t p n p n $. 41 m o m t m t e r $. 39 e o e o n y n r n r $. 37 t t y t y $. 35 98 99 00 Years Figure 7: shipment and inventory between 1998 and 2000

3.2 Impact of the Supply Chain structure on Demand Variability

To properly improve performances we had to define precisely the region of analysis. In particular, many SKUs show very different demand patterns due to many different reasons, but mainly influenced by the life cycle of both the component and the final products to which they refer. In fact, many SKUs show a very sporadic pattern while others show many days with positive demand. In particular, SKUs were classified according to the number of days with demand in one year as provide in Table 1. This was done to separate very sporadic SKUs from those that are in the normal period of their life.

140

Part III Chapter 2: Whirlpool Europe

Class Demand Days N. of SKUs % Value 1 > 48 4.194 Fast Movers 57% 2 24 < ≤ 48 4.363 3 11< ≤ 24 5.241 Medium Movers 20% 4 4 < ≤ 11 8.230 5 1< ≤ 4 8.835 Slow Movers 23% 6 0 ≤ ≤ 1 33.328 Total 64.191 100%

Table 1: definition and distribution of Whirlpool’s spare parts SKUs

As it can be noted almost 50% of SKUs show a real sporadic pattern and belong to products that have ended their guarantee period (thus typically customers prefer to change the final product instead of repairing it) or are in the final part of their life cycle. Moreover it can be noted that even if fast mover SKUs are fewer compared to other classes they are relevant in terms.

During the analysis the firm provided us with demand data for 1.216 SKUs chosen among different sectors (cooking, washing, and so on), different suppliers and different volumes. This was meant in having the greater number of possible situations the firms has to deal with. In particular, we were provided with daily demand and daily number of orders within a period of 204 days for each SKU. Figure 8 and Figure 9 provide a few examples of how demand goes.

350

300

250

200

quantity 150

100

0 1 12 23 34 45 56 67 78 89 100 111 122 133 144 155 166 177 188 199 days

Figure 8: Daily demand for SKU 481921478267

141

Part III Chapter 2: Whirlpool Europe

quantità 30

0 1 12 23 34 45 56 67 78 89 100 111 122 133 144 155 166 177 188 199 giorni

Figure 9: Daily demand for SKU 481912117005

As it can be noted demand shows quite a lumpy pattern since among days with very low demand others shows real demand peaks with demand amplifying up to 10 times.

Important was to analyze demand variability, using data regarding 200 SKUs. The company stores data at daily level, so we were not able to analyze transactional data. We considered both the demand and the number of order lines to understand if demand variability is due to changes in the number of orders received per day or in the orders size. We analyzed the mean, the standard deviation and the coefficient of variation of these variables.

Table 2 shows the values of these parameters for the SKUs in analysis.

Weekly Daily order Weekly orders Daily demand demand lines lines Mean 0.79 3.89 0.19 0.94 Standard Deviation 3.06 6.92 0.29 0.69 Coefficient of Variation 3.89 1.78 1.54 0.73 (St. dev./mean) Table 2: demand data analysis

Table 2 suggests that daily demand is 2.58 (3.89/1.54) times more variable than the number of order lines. This suggests that changes in order size rather than in the number of orders lead to the very high demand variability the company is facing. Quite interestingly this is closely related to the supply chain structure: indeed, the company receives orders of

142

Part III Chapter 2: Whirlpool Europe

very different sizes because on the one hand it serves small customers that receive direct deliveries and, on the other hand, it serves large customers and national warehouses that place batch orders.

More in-depth graphical analyses confirmed that the root cause of demand variability is the overlap of two phenomena: a “regular” pattern consisting of many small orders by many small customers and an “irregular” pattern generated by few huge orders from the largest customers. Figure 10 shows this phenomenon.

DEMAND

30 QUANTITY 20

0 1 10 19 28 37 46 55 64 73 82 91 100 109 118 127 136 145 154 163 172 181 190 199 208 DAYS

REGULAR SERIES IRREGULAR SERIES

70 70

60 60

50 50

40 40

30 30 QUANTITY QUANTITY 20 20

10 10

0 0 1 1 10 19 28 37 46 55 64 73 82 91 10 19 28 37 46 55 64 73 82 91 100 109 118 127 136 145 154 163 172 181 190 199 208 100 109 118 127 136 145 154 163 172 181 190 199 208 DAYS DAYS

Figure 10: Demand vs. Filtered Series

According to what has bee told till this point, we can make some considerations. 1) From the analysis of the supply system, two classes of customers can be identified with homogeneous characteristics inside them: from one side the residual regional warehouses and the wholesalers, from the other the Whirlpool dealer and the private repairer. 2) From the analysis of the time series on a daily basis, these two customers can very well be identified. Beside the frequent and small orders, generated by minor customers, others appear characterized by huge dimension, emitted by relevant customers. Figure 11 shows this phenomenon.

143

Part III Chapter 2: Whirlpool Europe

Probability

Daily Order Size

Figure 11: example of multi-modal demand distribution

3) Demand multimodality is relevant: for almost all SKUs two main dimensions of orders can be identified, a small and a big one. This fact deeply influences demand variability.

In these terms, the main root of demand variability is due to a great heterogeneity in customers’ orders due to the heterogeneous sizes these customers show. From one side we have many small customers, while from the other we have few relevant customers that purchase large quantities. In fact from the analysis of customers’ distribution it can be noted that few customers are responsible for the major part of demand (see Figure 12 for details). This claims that as previously stated, few customers, corresponding to the wholesalers and the residual regional warehouses, are responsible for the demand peaks previously introduced, while the other customers tend to purchase singularly reduced amounts of materials.

7.000.000

6.000.000 90% 9 customers 5.000.000 75% 4.000.000 23 customers 50%

sales € 3.000.000

2.000.000 54 customers 1.000.000

0 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% customers

Figure 12: Pareto’s curve of customers

144

Part III Chapter 2: Whirlpool Europe

This analysis of demand, together with the analysis of the demand management process the firm uses, let us understand the causes of inefficiency. In particular, the firm uses only one demand management process for the two rather different sources of demand (see Figure 13) thus adopting a solution that, at least partially, does not fit with demand characteristics. This is partially due to the fact that though managers recognized conceptually that demand has two modes, they were not able to analytically separate the two demand series. Moreover, the forecasting and inventory management systems the firm uses are incapable of good performances when dealing with variable and composite demand.

145

Part III Chapter 2: Whirlpool Europe

4 Solution Design

To design a more effective solution we developed different forecasting solutions, based on the causes of demand variability and so on the heterogeneity of customers’ demand. In particular, besides the actual system adopted within the considered company, three other solutions were designed. This section describes the four alternatives considered through a stepwise process. At first, we will describe the forecasting and inventory management systems the firm actually uses. This will lead us to understand the root causes of demand variability. Then we will provide 2 different solutions, one based on actual knowledge presented in current literature to solve such problem and a second one that focuses on demand heterogeneity. At last we will evaluate a fourth solution where asymmetric information is provided regarding the forecasting process.

4.1 Alternative a): Current Solution

The actual system the firm uses is built up of two phases. First, a simple exponential smoothing is applied to forecast future demand on the basis of past data. This provides the planning systems of an overall forecast of future requirements that is used to plan requirements towards suppliers. In fact, orders are emitted using a simple order up-to policy. Figure 13 describes this process.

Total Forecasting Inventory Performance (Whybark method) Management Demand Figure 13: Alternative a) scheme

Although this is a specific solution used in a single firm, it is quite representative of a great number of different solutions applied in real contexts: many firms adopt simple demand management (i.e. forecasting and inventory management policies), so we think that this alternative is rather representative of many real situations. It can be easily understood that this solution isn’t capable of forecasting such intermittent demand, since both the forecasting and planning system are developed for dealing with a low variable demand. Improvement in performances can be gained using the knowledge on demand bi-modality (see alternative b).

146

Part III Chapter 2: Whirlpool Europe

4.2 Alternative b): Literature model

As the previous analysis shows heterogeneity among customer is relevant. Since this heterogeneity is due to the supply chain structure, we argue that a possible solution to improve performances can be found in adopting a heterogeneous forecasting system that leverages on the different variables that influence the different customers. In these terms, the first solution we propose is based on the separation of demand in two series (further called “stable” and “irregular” and respectively built up of many small orders and few huge ones) and on the adoption of two ad hoc forecasting techniques, that better fit the two patterns. Figure 14 shows the scheme of alternative b).

Exponential Stable Smoothing Series Method

Total Inventory Filtering Total Performance Demand Forecast Management

Syntetos & Irregular Boylan Series Method

Figure 14: Alternative b) scheme

This solution is based on a three-step process: a filtering procedure to separate the two demand series, the forecasting of the two components and their aggregation, and finally the inventory management.

4.2.1 Filtering

The general idea of the filtering procedure is to identify, according to the demand size, two “areas” in which demand can be decomposed. From one side, we will have the demand occurrences that belong to the stable series, while, from the other, lays that of the demand peaks. Figure 15 exemplifies this approach.

147

Part III Chapter 2: Whirlpool Europe

quantity

Irregular Series

Threshold value

Stable Series

time Figure 15: separation between stable and irregular series

In the case considered, one of the main factors of demand heterogeneity is customer’s size. So, to reduce the impact of demand variability on the forecasting system, we argue that attention can be paid on this aspect. To properly manage the two series previously described we had to develop a filtering system capable of separating peaks from the more “stable” demand. The idea is to build a threshold that separates what should be considered “stable” from what should be considered “irregular”. Every demand observation lower than the threshold is considered "stable", otherwise it is considered irregular (Figure 16 exemplifies the filtering process). A similar application of this approach, developed in the food industry, can be found in Cachon and Fisher (1997).

To properly estimate the threshold, we have to evaluate precisely the even pattern of the stable series and its natural variability. So the problem of estimating correctly this threshold is split into two sub-problems: estimate the mean behavior of the stable series and its variability. One might think of using the average of past demand to estimate the mean of stable series, but this estimator is deeply influenced and biased by the demand peaks. On the contrary, we have to use an estimator of the mean stable demand that isn’t influenced by the peaks. The median of demand is only shortly influenced by the large peaks, and so seems to be a much better metric since it estimates much more correctly the even behavior of the stable series.

148

Part III Chapter 2: Whirlpool Europe

350

300

250

200

quantity 150

100

0 1 13 25 37 49 61 73 85 97 109 121 133 145 157 169 181 193 205 days

0,07

0,06

0,05

0,04

0,03 probability

0,02

0,01

0 0 50 100 150 200 250 300 350 quantity median filter 350

300

250

200

quantity 150

100 filter

50 median

0 1 13 25 37 49 61 73 85 97 109 121 133 145 157 169 181 193 205 days

Figure 16: filter threshold building

To evaluate precisely the standard deviation of the stable series, we couldn't use standard deviation of the total demand as the peaks made total demand variability very much different from the stable series one. So, we developed a two-phase filtering system, aiming at better estimating the standard deviation and the median of the stable pattern. In the first phase the system eliminates the most variable phenomena using the median and standard deviation of the total demand series. The new series obtained is much less variable and it's closer to the real stable pattern. In particular, its median and standard deviation are closer to that of the stable series. Applying the second filtering system, using the new series as input, evaluates correctly the stable part of demand. This idea is represented in Figure 17.

149

Part III Chapter 2: Whirlpool Europe

Total Demand

400 300 200

quantity 100 0 1 19 37 55 73 91 109 127 145 163 181 199 days

Stable Series after 1st filtering Stable Series after 2nd filtering

400 400 300 300 200 200

quantity 100

quantity 100 0 0 1 22 43 64 85 106 127 148 169 190 1 22 43 64 85 106 127 148 169 190 days days

Irregular Series after 1st filtering Irregular Series after 2nd filtering

400 400 300 300 200 200

quantity 100 quantity 100 0 0 1 22 43 64 85 106 127 148 169 190 1 22 43 64 85 106 127 148 169 190 days days

Figure 17: scheme of the two-phase system for the filter evaluation

The output of the filtering system is made of two demand series which appear more internally consistent, and for which two separated forecasting techniques can be developed. To measure the effective results of the filtering process, we compared the volume- weighted coefficients of variation of total demand and of the stable and irregular series. In particular, for what regards the irregular series we estimated variability by means of the modified coefficient of variation, evaluated as the ratio between standard deviation and demand mean of the non-zero orders. This guarantees that the irregular series variability is not biased by the high number of days with no demand.

150

Part III Chapter 2: Whirlpool Europe

This comparison is summarized by Table 3.

Total Demand Stable Series Irregular Series Weighted Coefficient of 3,73 1,15 0,71 Variation Table 3: weighted coefficients of variation of total demand, stable and irregular series

This comparison shows that the filtering system described is capable of identifying the causes of demand inconsistency quite well, as it reduces stable series variability of 2/3 than total demand one, while the irregular series tends to be more internally consistent.

4.2.2 Forecasting

To properly forecast the two series, we had to find two particular methods that fit with the series they are supposed to manage. In particular the stable series is generally quite even, so it seems reasonable to use an exponential smoothing-based system, which performs well when facing low variability patterns. In particular, the system here adopted is the Whybark Method that is the actual forecasting system the firm uses. On the contrary, as the irregular series has typically a strong lumpy pattern, we adopted a specific method built to manage such demand. In particular, we considered the Syntetos and Boylan method to evaluate future requirements. The Syntetos and Boylan method is based on the exponential smoothing, but it has two main peculiarities that distinguish it from simple smoothing techniques. First of all, it splits the forecasting problem in two sub-problems: it estimates separately the mean demand size per period and the inter-arrival time between orders. The objective is to evaluate demand forecast as product of the mean size of orders and of the number of orders during the planning period. Second, forecasts are updated only when demand shows. We refer to Chapter 2 for details on this method.

This particular procedure allows forecast not to be too much influenced by a great number of days with no demand, as, for example, exponential smoothing does. The criticality of this technique lays in the estimation of “when” the peak will occur. Moreover, after the occurrence of an order, the system pushes inventory management systems to reorder thus accumulating parts right after demand has occurred. This, due to demand sporadicity, generates high inventory level for a long time. We believe, and further show, that a different system to evaluate when the order will show up should guarantee better performances.

151

Part III Chapter 2: Whirlpool Europe

4.3 Alternative c): Ad hoc model

We developed a solution that uses the filtering system introduced in alternative b) to evaluate the two series (stable and irregular). Moreover, solution c) uses two separated forecasting and inventory management systems for each of the two series. Figure 18 shows the scheme of solution c).

Exponential Stable Series Stable Smoothing Inventory Series Method Management

Total Total Filtering Performance Demand Orders

Irregular Irregular Solution Series Series Proposed Inventory Management

Figure 18: Alternative c) scheme

4.3.1 Forecasting

As in cases a) and b), we use the Whybark method for the stable series, while for the irregular series, a deeper analysis is needed. The problem of forecasting a peak of demand can be divided in two sub-problems: first of all one must estimate when the peak will occur, then how many parts will be ordered (see Figure 19 for an example).

The analysis of demand allows to separate the two constituent phenomena and so to evaluate separately the inter-arrival time between peaks and the ordered quantities. By observing the characteristics of the irregular series it is notable that even if the peaks are the major cause of the great variability demand has, they show some regularity. In particular, it is notable that both the inter-arrival time between two successive demand peaks and the number of parts ordered are even. This feature is related to the supply chain structure and reordering processes of large customers: we expect that many customers use reorder systems based on the economic order quantity or fixed reorder period, as the demand they manage appears relatively stable. So they will reorder relatively stable quantities with a relatively even frequency.

152

Part III Chapter 2: Whirlpool Europe

350 4 300 Interarrival time 250 86 158 200

quantity 150 Order size 100

0 1 12 23 34 45 56 67 78 89 100 111 122 133 144 155 166 177 188 199 days

Figure 19: The dual-dimension problem of forecasting

In these terms, heterogeneity shows up again in terms of purchasing processes: from one side we have many customers that order according to the single requests they receive from the end market as they tend not to have stock of parts and simply react to the specific market needs. These are typically small repairers that operate at local level. From the other we can identify the large customers (wholesalers and regional warehouses) that tend to order with almost stable interarrival time for almost stable quantities. The forecasting algorithm can exploit these regularities to achieve good performance.

Literature has devoted a major attention to evaluate which statistical distribution fits better to the problem we are looking at. The major part of the literature considers demand of spare parts modeled according to a Poisson or Compound Poisson process1 (Friend, 1960; Hadley and Within, 1963). Other authors have analyzed when other distribution may be adopted (Vereecke and Verstraeten, 1994). We analyzed the demand data provided to evaluate which distribution fits better and we discovered that the SKUs under consideration could be clustered according to two main distributions. Some SKUs fit very well the Poisson assumption, while for others, the Normal distribution appears to be better. The validity of this approach was supported by the analysis of graphical tests as Q-Q-Plot, which demonstrated that the assumption holds. Figure 20 shows an example of this graphical analysis.

Part III Chapter 2: Whirlpool Europe

1 0,9 0,8 0,7 0,6 0,5 0,4 0,3 0,2 0,1 Theorical CumulativeProbability 0 0 0,2 0,4 0,6 0,8 1 Actual Cumulative Probability

Figure 20: example of Q-Q-Plot

Simply this analysis evaluates the actual cumulative probability and compares it with the theoretical one derived from a particular distribution. The distance between each single observation from the mean line can be adopted as a measure of the quality of the distribution assumed. According to this evaluation we analyzed the different SKUs and defined which assumption holds better.

We argue that this distinction is mainly due to how the demand peaks are generated: when many customers are responsible for these phenomena the Poisson assumption tends to perform better, as the process of peak generation appears to be memory-less. Instead, when a single customer is the source of the demand peaks the process tends to have memory and so the regularities in the customers’ reorder process tend to outperform the Poisson model; in this situation the Normal assumption is likely to perform better. In the development of this solution we adopted these two distributions according to how they fit to the data in analysis. However, in both the two situations we can leverage on this information since we can, first, model interarrival time with specific distributions and, second, in those situations where the normal assumption fits better we can introduce a memory process in the demand evaluation.

The proposed method updates the mean and variance of inter-arrival time daily and considering both the occurrence and the absence of a peak. Indeed a few codes, after showing a few peaks in the first months, have no peaks for the rest of the simulation run. So, the updating procedure keeps on estimating that the peaks will occur with the initial frequency.

154

Part III Chapter 2: Whirlpool Europe

Actually this intuitive solution disregards an important piece of information: the time since last peak has occurred. The solution we adopted was of integrating into the updating system also the information that a peak hasn't shown up in a long time. Thus the interarrival time estimate is updated regardless of whether the peak has occurred or not. In particular, if at time T no peak is observed, the estimate is updated considering that it will occur from T on. The better way to make this estimation would have been to use a survival analysis technique (Kleinbaum, 1996); to simplify the solution we used an approximated estimate, which, in fact, doesn't consider the censored distribution of the last peak. In particular, the forecasting for the inter-arrival mean during the t period is equal to:

µˆ = max µ ⋅(1−α ) +α ⋅δ ; µ ⋅(1−α ) +α ⋅δ ⋅(1−α ) +α ⋅i t {}t 1 1 t [ t 1 1 t ] 2 2 t

where: µt is the estimate of mean till the t period

δt is the inter-arrival time between the last two peaks

it the inter-arrival time since the last peak

α1, α2 are two smoothing constants

The first part of the formula updates the mean whenever a peak is observed at time T, while the second one when a peak is not observed at time T. The two smoothing constants are used to weight how much a peak occurrence and delay influence the actual estimate. A similar updating process is applied to the standard deviation of the inter-arrival times.

The estimated parameters (mean and standard deviation) are used to evaluate the probability distribution of the peak occurrence. To evaluate the probability of occurrence of future orders we have to consider a) the information available regarding the time elapsed since last peak and b) the possibility that more than one peak will occur over the planning period. We will discus the two issues separately. a) Whenever a forecast is produced, we have to consider whether the process of generation of a peak is memory-less or not; so if the Normal assumption is considered for a particular SKU we can add the information that till today a peak hasn't occurred yet. To do so we have to compute the probability of occurrence of a peak given that it hasn't already shown up, using a conditioned probability. Figure 21 exemplifies this process.

155

Part III Chapter 2: Whirlpool Europe

Conditioned Simple Distribution Distribution Probability Probability Probability Probability

time t time t

t Figure 21: evaluation of the probability of peak occurrence b) If more than one peak occurs over the planning period, we have to evaluate the cumulative probability of more events. If on day T a peak occurred, the expected date of the nth peak is T plus n times the expected inter-arrival time. Nevertheless, these distributions will show different variances, as the number of possible events (the days in which a peak may occur) rises. In particular we assume inter-arrivals not to be auto- correlated. Thus the standard deviation of the second peak will be equal to 2 ⋅σ , as the number of possible events is double than the first peak ones, for the third peak will be 3,⋅σ for the nth will be n ⋅σ and so on, where σ is the estimated standard deviation of the inter-arrival period. Figure 22 exemplifies this process.

probability

days t t+ LT

Figure 22: probability distribution of successive peaks occurring over the lead time LT

Given the probability of occurrence of one or more peaks during the planning horizon, it is also necessary to evaluate the mean size of the huge orders, which lets forecast the total number of parts expected to be ordered. We did not assume any distribution for this dimension but we estimated the mean order size using an exponential smoothing to avoid

156

Part III Chapter 2: Whirlpool Europe

distortions caused by contingent events. Each time a peak occurs, the estimate of mean size order is updated as follows: µˆ =µˆ ⋅(1 − α) + D ⋅α q,t q,t−1 t

where: µˆ q,t is the estimate at time t

Dt is the peak size at time t α is the smoothing constant

So the system here described every day of the simulation calculates the probability of occurrence of one or more peaks in the lead-time and, according to the mean order size, forecasts the total number of parts that are expected to be ordered.

4.3.2 Inventory management

The forecasting system feeds probabilistic forecasts (i.e. distributions) into the inventory management system. An inventory management system based on a simple technique as order-up-to policy perfectly fits the stable series patterns. This is not true for the irregular series; furthermore, we think that an inventory management that fully exploits the probabilistic information produced by the forecasting system can deeply raise process performances. This sounds reasonable also for the differences among the two kinds of orders, concerning the inventory management. First of all the two series are associated to different economic value. The huge orders size, very much bigger than the small ones, involves that the economic effect of their lost sale or long time stock keeping is particularly critical. Additionally, the huge orders management has a particular impact on the firm reputation. Arranging a huge order reduces significantly the supply availability, which affects the opportunity of accepting future orders from smaller customers. On the contrary, if the huge order is backlogged, although stock is available to many small customers, this has a critical impact because service isn’t granted to an “important” customer. Although the described trade-off exists also for the stable demand, its relevance is much less critical.

For the irregular series, the order management system we propose tries to balance the backlog and stock holding costs, evaluating when it is convenient to place an order. As the expected cost is a function of the probability of occurrence of a peak, the idea is to bind the emission of an order to a minimum probability of occurrence (MPO), estimated by the forecasting system. If the probability of occurrence of demand peaks over the planning

157

Part III Chapter 2: Whirlpool Europe

horizon is lower than MPO, no orders are placed; otherwise the system releases an order. If the MPO value is low, the system will place orders even when the likelihood that it will actually occur is quite low. On the contrary, if this value is high, the system places purchasing orders only when it is extremely likely that at least a large customer will place a big order.

Whenever an order has to be placed, it is convenient to reorder a number of parts equal to the mean peak dimension, estimated by the forecasting system. Figure 23 shows the Service Level / Inventory Level trade-off curves, built for two different values of MPO.

MPO = 0,1 MPO = 0,95 1 0,99 0,98 0,97 0,96 0,95 0,94

Service Level 0,93 0,92 0,91 0,9 0 2000 4000 6000 8000 10000 12000 14000 mean stock

Figure 23: trade-off curves for different values of minimum probability of occurrence (MPO) required to emit an order

Consider the lines plotted for a high value of MPO (i.e. 0.95) and for a low one (i.e. 0.10). The comparison shows that high MPO values perform better than low values for low SL, while this result is turned over for high SL. This is due to the fact that high MPO values stop not-likely orders reducing stocks, but backordering few orders, while low values of MPO catch more orders while increasing stocks. The global performance curve can be plotted by interpolating the lines that represent the performance trade-off for any given threshold value of MPO, as Figure 24 shows. The particular MPO should be chosen according to the target SL the company wants to achieve.

158

Part III Chapter 2: Whirlpool Europe

0,9

0,8

0,7

Service Level 0,6

0,5

0,4 0 2000 4000 6000 8000 10000 12000 mean stock

Figure 24: Interpolated trade-off curve

4.4 Alternative d): improving performances through information

Solution b and c are based on an internally consistent demand management process that fits demand and supply chain characteristics. They use a given set of information (in this case the number of units requested and the number of orders per day in the past) to improve operational performance. On the other hand, past literature suggests that a company might collect a larger set of information to further improve performance. Such a collection of additional information might require organizational efforts, both from the supplier and the buyer side. In this work we argue that since heterogeneous demand is faced, a heterogeneous solution has to be adopted. Coherently with this assumption we developed a forecasting and planning system (alternative c) that leverages on this aspect. However, we argue that also a heterogeneous collection of information can gain significant benefits.

This work, like others in the past, deliberately does not consider the organizational costs given that they are quite hard to model and firm specific. Thus organizational cost can be better estimated through case studies and gut-feel rather than through management science models. On the other hand, a model can quite precisely show the potential benefits of a given amount of information, in terms of service level/inventory level improvements. In particular, we assume that the company should focus its efforts on the irregular portion of demand. Peaks are generated by a small number of large customers and thus the cost of collecting that information should be relatively low. In addition, usually salespeople by

159

Part III Chapter 2: Whirlpool Europe

default devote a large portion of their time to key accounts and thus can forecast future orders quite precisely with a very limited additional effort. On the other hand, those few customers create a significant variability in demand and thus tend to reduce the service level and increase stocks. This makes the information about their future demand extremely useful for forecasting and planning purposes and relatively cheap.

We built a model that simulates such asymmetric collection of information. The model, like in cases b and c manages the stable and irregular series separately. In particular, for the stable series an exponential smoothing forecasting system is used, while, for the irregular series, a different approach is considered. The simulation model assumes that the large customers provide the company with some information about how many units the single customer is going to order in the future. Obviously, the information the company collects is probabilistic rather than deterministic. Thus we assume that both the timing and the size of forthcoming peaks are normally distributed. To properly use real demand data, the model starts from the real timing and real size of the peak and simulates the forecasted timing and size as the actual figures plus a normal error with a zero deviation. This technique, on the one hand, enables us to describe the effects of information collection and, on the other hand, does not alter the real demand time series. The more company’s information and predictions are reliable, the lower is the variance of the two distributions and the better the system performs. Figure 25 shows the scheme of alternative d).

Stable Series Exponential Inventory Stable Smoothing Management Series Method

Total Total Filtering Performance Demand Orders

Irregular Irregular Solution Series Series Proposed Inventory Management

INFORMATION

Figure 25: Alternative d) scheme.

160

Part III Chapter 2: Whirlpool Europe

5 Performance

The four alternatives presented in the previous section are compared through IL/SL trade- off curves. In particular, for each alternative we run simulations using 1214 codes over 209 days. For a given alternative (e.g. alternative a), a point on the trade off curve represents the IL and the SL the company can gain with a given hedge (i.e. with a given parameter that influences the level of safety stocks). Thus a trade-off curve represents all the mixes of performances a company can gain with a given alternative by changing the amount of inventories kept in the warehouse.

Figure 26 shows, as one would have expected, that the alternative d is the best while a performs poorly.

Alternative a) Alternative b) Alternative c) Alternative d)

0,9

0,8

0,7

0,6 Service Level (Quantity)

0,5

0,4 0 2000 4000 6000 8000 10000 12000 14000 16000 18000 20000 Mean Stocks

Figure 26: trade-off curves for the different models analyzed: Alternative a): the model adopted by the firm we analyzed Alternative b): the model based on techniques existing in literature Alternative c): the model proposed in this work Alternative d): the model based on information exchange between firms

To make the comparison easier we consider the amount of inventories the four solutions require to gain a 83% SL and a 95% SL. Table 4 and Table 5 describe the savings alternative x (in the columns) can gain as compared to alternative y (in the rows). A

161

Part III Chapter 2: Whirlpool Europe

stepwise analysis that compares each solution to the next one might better show the additional benefits of the features of each single solution.

a b c d Inven. 10,000 5,100 2,520 1,810

a b c d a 0.0% 49.0% 74.8% 81.9% b 0.0% 50.6%64.5% c 0.0% 28.2% d 0.0% Table 4: Inventory investments and % savings of solutions a-d as compared to solutions a-d. Case of SL = 83%.

a b c d Inven. 14,750 8,800 4,500 3,600

a b c d a 0.0% 40.3% 69.5% 75.6% b 0.0% 48.9%59.1% c 0.0% 20.0% d 0.0% Table 5: Inventory investments and % savings of solutions a-d as compared to solutions a-d. Case of SL = 95%.

a vs. b. Solution b reduces the inventory investment by 49% as compared to the base case of solution a. Thus solution b guarantees a substantial cost reduction. Hence, the filtering algorithm and adoption of two different forecasting approaches for the two different time series does add value to the company. More in general, this result suggests that when demand and supply chain are composite, companies should identify the different demand management problems they are facing to choose specific solutions.

b vs. c. Solution c performs far better than solution b (50.6% stock reduction). Both a, b and c rely on the same information (i.e. past orders and demand), thus the comparison seems quite fair. Solutions b and c share the same filtering algorithm thus differences in performance are due to the forecasting and inventory management. However, it is quite interesting to notice that the accuracy of this new forecasting approach is only slightly better than the forecasting approach suggested by Syntetos and Boylan, while the operational performance differs substantially. We actually think that this is due to the fact that c is an integrated solution where the forecasting and inventory management are consistent. We believe that what really

162

Part III Chapter 2: Whirlpool Europe

makes the difference is that the solution fits with the features of the supply chain and demand and is internally consistent. This, together with decision making theory, suggests that the forecasting solution is not supposed to reduce forecasting errors but rather to support decision making processes (in this case inventory decisions).

c vs. d. Solution d reduces inventory investment by 81.9% as compared to the base case a and by 28.2% as compared to solution c. Solutions c and d rely on a significantly different set of pieces of information, so we can not conclude that d dominates c, we can just conclude that d performs better, but it might well be too expensive (in terms of effort required to collect the information from the customers and to deliver such information to planners). Nevertheless the results seem to be quite interesting since they suggest that forecasting is not just a matter of developing algorithms to fully exploit past data, but can involve organizational processes even in the spare parts business. It is worth saying that the company we co-operated with has used these quantitative analyses to support an effort to collect information from key customers.

Finally, we can compare Table 4 and Table 5. This comparison tells us the impact of service level on differences in performance. As one would have expected the inventory investment increases for all alternatives. In addition, the absolute differences among the four solutions increase thus suggesting that from a business perspective the importance of choosing the right approach to the management of demand grows as the target service level is more and more challenging and the inventory investment enlarges. Though, quite interestingly, percentage differences among solutions are reduced (for example savings of solution b vs. a drop from 49.0% to 43.0%). At a close look it seems rather intuitive: as the company targets extremely high service levels and faces extremely variable demand, it should be capable of meeting the huge maximum potential demand at all times. Regardless of the solution (a, b c or d) this requires to keep a very high inventory all the time just in case a huge (though very unlikely) peak occurs. Thus for high service levels targets absolute differences increase but percentage differences tend to decrease given that all solutions will always require very high inventories.

At last, however, we argue that the proposed solution, in particular Alternative C, may not be the most effective solution. In fact some SKUs may not be perfectly forecasted by this solution compared to others proposed. Thus we studied under which conditions the proposed solution fits better. To study these conditions we compared performances for

163

Part III Chapter 2: Whirlpool Europe

Alternative C and Alternative A in terms of mean stock for a given service level (96%) and evaluated asymmetry of total demand. Figure 27 shows the relationship between percentage mean stock ratio of the two alternatives and the coefficient of asymmetry.

3,5

2,5

1,5

0,5 Mean Stock Ratio (Alternative C/Alternative A) (Alternative C/Alternative Stock Ratio Mean 0 0 3 6 9 12 15 Demand Asymmetry

Figure 27: relationship between mean stock ratio and demand asymmetry

As it can be noted for greater demand asymmetry, the proposed solution (Alternative C) tends to perform better than the actual system adopted by the company (Alternative A). This is due to the fact that if demand doesn’t show a significant multi modality, then demand peaks are not easy to be identified and so the proposed solution doesn’t provide significant improvements. Table 6 summarizes the linear regression results between these two dimensions.

Term Estimate Std Error t Ratio Prob>|t| RSquare 0,144678 Intercept 1,2482705 0,033687 37,06 <.0001 RSquare Adj 0,143674 Asymmetry -0,046621 0,003883 -12,00 <.0001

Table 6: linear regression between mean stock ratio and asymmetry

164

Part III Chapter 2: Whirlpool Europe

In particular among the 854 SKUs considered in this analysis2, 617 perform better with the proposed solution, while for 237 the actual system performs better. Detailed analysis showed that the SKUs for which the actual system is preferable (those reported in the higher part of Figure 27 are characterized by a single peak thus the system, even if it updates peaks interarrival, keeps on claiming that a peak will occur.

2 We didn’t consider those SKUs for which demand asymmetry was not definable, in other terms, those that have less than three days of demand within the considered period. 165

Part III Chapter 2: Whirlpool Europe

6 Conclusion

This case contributes to knowledge in demand management in rather different ways.

First, the work shows the impact of demand heterogeneity on forecasting systems. In fact the situation faced by this company, that shows a relevant heterogeneity between customers requests deeply affect both forecasting and inventory performances. Moreover, it can be noted that the main cause of demand heterogeneity can be traced back to the supply chain complexity that, due to the direct delivery project, becomes greater. So a first contribution is stating that, as provided in the general framework, when supply chain is simple, demand tends to be homogeneous and thus less variable, so forecasting is not a complex issue, while, when supply chain complexity rises, heterogeneity rises too, demand becomes more variable and forecasting more tricky. In fact, this work discusses the rather unexplored problem of managing supply chains with a various numbers of echelons, multi-modal and extremely variable demand, and with lack of visibility over the distribution channel.

Second, this work shows that when demand heterogeneity is essentially due to the complexity of the supply chain, benefits can be found by adopting different forecasting systems, based on demand heterogeneity causes and that leverages on the different customers’ characteristics. This work provides an algorithmic solution that, through the comprehension of the sources of demand variability and through a probabilistic forecast and inventory management, leads to performance improvement as compared to current solutions adopted by the firm and proposed by the literature. In addition, the work shows that a proper collection of information regarding the purchasing plans of few large customers (that usually significantly contribute to the total variance of demand) can improve the performance of the supply chain substantially. Figure 28 models this effect. Third, the work contributes to the more general problem of how to design a solution to manage uncertain and extremely variable demand. a) We clearly show the close relationship between supply chain structure, demand patterns and the characteristic of the demand management solution. In particular, in the case discussed, a supply chain with different numbers of echelons generates a multi- modal demand that asks for a composite managerial solution (compare solutions b and a in section 5).

166

Part III Chapter 2: Whirlpool Europe

b) Moreover, we show that integrated and consistent solutions can outperform classical solutions based on a stepwise process (compare solutions c and b). c) Finally, the work adds to the literature debate on the value of information in supply chain and inventory management. Clearly the model presented is not the solution itself (which relies on organizational processes and exchange of information about inventory levels etc.) but can estimate its benefits (compare solutions d and c). Such model was adopted by the firm we co-operated with to justify organizational efforts in the area of collaborative planning.

Different forecasting structures Inner Heterogeneity Heterogeneity

Supply Chain Complexity

Figure 28: the impact of supply chain complexity in the Whirlpool case

Even if the solution provided has been developed in the specific context of the firm considered, we argue that the methodology can be generalized. In fact, the main factors that have pushed this company towards a greater supply chain complexity are shared among other companies belonging to different industrial sectors. As discussed in paragraph 3.1, one of the main causes of the rising of demand heterogeneity is tied to the rising attention towards logistic costs. In fact, many firms are redesigning their supply chains to reduce costs tied to inventory holding and transportation, which is exactly the same situation this firm shows. We argue, then, that the solutions provided, in terms of general methodology, can be applied in different industrial contexts that face similar supply conditions.

167

Part III Chapter 2: Whirlpool Europe

168

Chapter 3 Nestlé Italiana

1 Context and Purpose

This case considers a company working in the food business that has to deal with a highly variable demand, due essentially to promotional activities. In particular, promotions are applied both at a customer level, essentially only for the main customers, and at a chain level. Thus it becomes relevant both to consider this element in the evaluation of forecast, moreover it becomes critical to consider how the different customers behave during promotional periods.

As a matter of fact literature regarding demand influenced by promotions is not very developed. For example, Cachon and Fisher (1997) faced a similar problem, as they needed to identify demand occurrences affected by promotions. However, their solution based on the analysis of total demand, faces a few problem when customers tend to have different promotional periods. Thus, this situation needs a different solution.

169

Part III Chapter 3: Nestlé Italiana

2 Introduction to the company

Nestlé S.A. is a multinational group which most relevant business is connected to the production and commercialization of food. Nestlé in fact is the first firm in the food sector and one the most important industrial companies in the world. With the activities conducted in 70 countries, the firm has gained in the 1999 a total sales value of 46.373 millions €, producing in 509 plants in 83 countries with 230.929 people.

In 1866, Henry Nestlé, a chemical from Frankfurt, established in Vevey (Switzerland), the society that will take its name. In 1867, it started producing the famous milk and biscuits flour, a revolutionary product for those times, that substitutes milk in periods when child mortality was very high (15-20%), also due to not proper food. In the same year, in Cham (Switzerland) also the Anglo-Swiss Condensed Milk Company started its activities. The two companies started to compete by producing the same products. Only in 1905 the competition ended with the merge of the two in one big company formally known as Nestlé & Anglo-Swiss Condensed Milk Company. From this point Nestlé began its expansion, both by increasing its product variety, and both by raising the number of production plants. In 1929 the company merged with Peter-Cailler-Kohler Chocolates Suisse S.A.. In 1930 due to overstock of coffee grains, the company starts a study on the conservation of coffee and in 1938 Nescafè is launched, followed by other similar products as Nescorè, Nesquik and Milo. In 1947, Nestlé merges with Alimentana S.A., the company that owns the brand Maggi. Afterwards, the product variety rises again by means of mergers with Cross & Blackwell’s jams (UK), Findus’ frozen goods (Scandinavia), Locatelli’s cheese (Italy), Chamboury’s yogurt (France), Libby’s juices (USA), Mont Blanc’s and A l’Ours’ milk (France), Thomy’s seasonings, and many more.

In 1947, the diversification strategy moves Nestlé also towards participation in other business, as for example, with L’Oreal that will become property-owned in 1977, specialized in cosmetics. In 1985 Nestlé acquires the Hills Brothers Coffee Inc. and the Carnation. Other relevant acquisition are in 1988 Buitoni (so gaining control of Buitoni, Perugina, Vismara, Sasso, King’s, Pezzullo and Berni) and Rowntree Mackintosh; in 1992 Terrier, in 1993 Italgel and Gruppo Dolciario Italiano, in 1994 Alpo, in 1997 San Pellegrino and in 1998 Spillers Petfoods.

170

Part III Chapter 3: Nestlé Italiana

In 1999 activities are consolidate by means of the acquisitions of La Cocinera (Spain), Svitoch (Ukraine), Totole (China), La Universal (Ecuador) and Quilmes (Argentina).

Table 1 reports financial results in the year 1999, while Figure 1 and Figure 2 respectively show the distribution of sales along the different product categories and in the different countries.

1999 1998 Sales 46.373 mln € 44.563 mln € Results 2.934 mln € 2.612 mln € Production Units 509 522 Staff 230.929 231.881

Table 1: major indicators in the 1999 and 1998 (adapted from the annual report “Rapporto annuale 1999 Nestlé Italiana”)

Cooked products; 27% Milk based products and ice creams; 26%

Sweets; 14% Beverages; 28%

Pharmaceutical; 5%

Figure 1: sales during 1999 classified according to product nature (adapted from the annual report “Rapporto annuale 1999 Nestlé Italiana”)

171

Part III Chapter 3: Nestlé Italiana

USA Others 22% 24%

Switzerland 2% France Philippines 10% 2% Canada 2% Australia Germany 2% Spain 9% 3% Mexico UK Brazil 3% 7% 4% Japan Italy 5% 5%

Figure 2: 1999 sales distribution among countries (adapted from the annual report “Rapporto annuale 1999 Nestlé Italiana”)

2.1 Nestlé Italiana

Nestlé has activities in Italy since more than a century. In fact since 1875 the company sells its products within Italy. However it is only in 1913 that the Italian division is founded with the opening in Abbiategrasso (Milano) of the first production plant for condensate milk, while 1916 the chocolate factory opens in Intra. Here are summarized the main step the company had made in Italy. 1948 –Acquisition of Maggi, in Italy since 1912 with a plant in Sesto San Giovanni (Milano). 1961 – Acquisition of Locatelli and La Gragnanese, factory in Gragnano (Piacenza). 1971 – Acquisition of Ursina Franck. 1985 – Acquisition of Carnation, with plants in Udine and Bertiolo (Udine). 1988 – Acquisition of the Buitoni’s Group 1993 – Acquisition of Italgel and Gruppo Dolciario Italiano, in the frozen sector with brands Gelati Motta, Antica Gelateria del Corso, La Valle degli Orti, Surgela, Voglia di Pizza, Mare Fresco, and in the sweets sector with Motta and Alemagna. 1998 – Discharge of cheese activities with brand Locatelli. Acquisition of Spiller that becomes Perfood, active in Italy with a plant in Castiglione delle Stiviere (Mantova) and with its brands Felix, Fido, Vitto, Doko.

172

Part III Chapter 3: Nestlé Italiana

2.2 The Fresh Food Division

In this part attention is paid towards the fresh food division of Nestlé Italiana. This unit manages all products characterized by a limited duration, typically one month, and that need to be maintained in environments with controlled temperature (no more than 8 °C). These characteristics imply that these products are peculiar compared to others, so needing to allocate an entire division towards their management. Different products are here considered, from cheese to yogurt, to fresh pasta, and so on. However they all share the similar logistic problems due to a reduced shelf life and to the need of keeping goods in controlled temperature.

The main units that work under this division are: Logistics: dealing with supply chain management, from forecasting to distribution planning; Sales: managing all customers and minor distributors; Marketing: dealing with communication towards end consumer; Production: manages Italian factories for fresh products; Human resources; Accounting; Product Development.

Logistics is structured in different sub-functions: Forecasting: dealing exclusively with demand forecasting; Planning: managing production planning, separately for the two main brands: Nestlé and Buitoni; Budget & Report; Project Development; Transports; Local Area Manager (LAM): managing partnerships with the three distribution centers located in Italy.

Figure 3 shows the organizational structure and the detail for the Logistics unit.

173

Part III Chapter 3: Nestlé Italiana

Management

Accounting New Products Development

Production Marketing Logistics Sales Human Resource

Forecasting Project Budget & Planning Nestlè Planning Transports LAM North LAM North LAM Development Report Buitoni West East South

Figure 3: organization of the logistic structure

2.3 Logistic Structure

Here a brief description of the production plants and the distributional systems is provided

Production in each plant over Europe is specialized so to gain economies of scale; this, of course, implies a greater complexity for distribution, moreover because we are dealing with fresh products. The different plants are reported in Figure 4.

Transportations of Nestlé fresh products are all done by wheel as their short life require speed and flexibility. In Italy 5 distributional centers are provided in Milan, Modena, Rome, Caserta and Catania. Goods are stored here for no more than three days, than they are sent to customers through specific vectors. Figure 5 reports the location of the distributional centers and the total volumes managed.

The third part of the supply chain is represented by end consumers. They are managed through different channels, as Italian distributional system is very heterogeneous.

The main categories of customers are as follows: Retailers: transportation is provided to the 39 centralized centers of the distribution chains that provide directly to single store distribution. Hypermarkets: delivery is done directly to each of the 305 point of sale. Dealers: delivery is provided to the 50 stores of the different customers that serve both stores and retailers. Distributors: goods are delivered to each of the 80 customers that server 2000 point of sale; in this situation goods are owned by Nestlé as long as they are sold to the single store. These structures in fact simply provide a common interface towards point of sales but don’t purchase anything.

174

Part III Chapter 3: Nestlé Italiana

Figure 6 synthesizes the supply structure of Nestlé Fresh Food.

LUNEBURG (1.600) STAVERTON (800) LC1 BERLIN (750) DESSERT CHILDREN SNACK

CUINCY (1.400) DESSERT

WEIDING (600) LISIEUX (10.350) DESSERT FRUTTOLO

ANDREZIEUX (6,900) YOGHURT

VALANCE (300) LEISI (4.100) MIO FRUIT BASIS PASTA

PIANORO (750) SNACK

MORETTA (750) SNACK

OSTUNI (750) VILLADECANS (1.000) SNACK LC1 GO

Figure 4: geographical distribution of production plants

1998 (tons) MILAN 13.833 MODENA 9.405 ROMA 3.724 CASERTA 9.796

CATANIA 3.942

Figure 5: distribution centers located in Italy

175

Part III Chapter 3: Nestlé Italiana

Plants Retailers

Distributional Centers

Hypermarkets

Dealers

Distributors

Figure 6: the supply chain structure

Figure 7 shows the distribution of sales in the different classes of customers.

Distributors Hypermarkets 23% 19%

Retailers 24% Dealers 34%

Figure 7: sales for each supply channel during 1998 (adapted from the annual report “Rapporto annuale 1999 Nestlé Italiana”)

From this information we can identify a few relevant aspects. First of all Nestlé deals with many different customers, belonging to different supply structures, that contribute to total demand similarly, since the total sales per channel are similar (see Figure 7). In these terms the complexity of the supply chain is rather relevant,

176

Part III Chapter 3: Nestlé Italiana

since Nestlé deals with some hypermarkets directly, while from another side deals with other sorts of intermediaries that behave differently according to their particular cost structures. Moreover, since the aggregate size of these customers, in terms of overall contribution to Nestlé’s sales, is similar, we may state that the complexity of the supply chain is particularly relevant. To understand the implications of this structure on demand and forecasting, the next part will analyze demand variability.

177

Part III Chapter 3: Nestlé Italiana

3 Analysis of the Supply Chain and Demand

The previous part introduces the company and its internal structure; moreover indications on its supply chain structure are provided, showing that the firm deals with a rather composite supply structure. In this part we will consider the analysis of demand variability as a first step towards the development of a proper solution. This analysis, together with the analysis of the supply will make possible to understand the main causes of demand variability and to design a proper forecasting solution.

3.1 Analysis of demand variability

The main problem we are dealing with is demand variability; in fact, market requests for the different products the firm manages show a lumpy pattern in which, among periods of rather stable requirements we have weeks with real demand peaks. Figure 8 shows an example of demand evaluated at a distribution centre level.

Figure 8: demand for a single product between January 1999 and May 2000

The first analysis conducted was aimed at evaluating demand variability. We adopted the coefficient of variation (CV) as a measure of demand variability. The analysis was conducted at a national level on the 49 consolidated products. The time period in analysis was made of the years 1998 and 1999 and data was considered at a weekly level, since this is the detail adopted in the forecasting division.

178

Part III Chapter 3: Nestlé Italiana

Figure 9 shows the distribution of the CV on the different products; in particular the smallest is equal to 12% (formaggino Mio), while the maximum value is 85% (Gramigna), compared to a mean value of 33%.

Demand variability is in fact pretty high, so we tried to understand the root causes of demand variability. In fact demand variability may be tied to many different causes. From one side we considered systematic variability as for example seasonality, in other words variability tied to the specific product due to changes in market requirements. As a matter of fact the fresh food industry shows a relevant impact of this factor. From another side we considered managerial variability, as promotional activities, so variability due to particular politics applied towards customers.

35%

30%

25%

20%

15% Frequency 10%

0% 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% CV

Figure 9: Coefficient of Variation for the consolidated products. Values identify the centre of the range of variability

For what regards seasonality, we evaluated seasonality factors, and classified SKUs according to the mean of the seasonality factors S. Figure 10 summarizes these results, considering that: S > 115%: Strong winter seasonality 105% < S < 115%: Low winter seasonality 95% < S < 105%: No seasonality 85% < S < 105%: Low summer seasonality S < 85%: Strong summer seasonality

179

Part III Chapter 3: Nestlé Italiana

45% 40% 35% 30% 25% 20%

Frequency 15% 10% 5% 0% S>115% 105%

Figure 10: analysis of seasonality for consolidate products during 1998 and 1999

This analysis identifies that demand is affected by seasonality, so forecasting must consider this aspect; however, when studying the deseasonalized demand variability seemed to be still high thus stating that other factors had to be considered. Figure 11 shows the distribution of CV for the deseasonalized series.

70%

60%

50%

40%

30% Frequency 20%

10%

0% 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% CV

Figure 11: coefficient of variation for the deseasonalized series

Interesting hints regarding the causes of demand variability are given by the analysis of the demand shape. This is meant to identify how demand is distributed around its mean; this analysis was done by means of the coefficient of asymmetry that is equal to zero if demand is homogeneously distributed around its mean, while it is positive is demand has a right

180

Part III Chapter 3: Nestlé Italiana

asymmetry and negative if the asymmetry is left. Figure 12 shows two examples of right

(f1) and left asymmetry (f2). For details on the impact of asymmetry on demand variability we refer to Bartezzaghi et al. (1999b).

Figure 12: example of right asymmetry (f1) and left asymmetry (f2)

Figure 13 shows the distribution of asymmetry coefficients among the different SKUs.

40%

35%

30%

25%

20%

Frequency 15%

10%

0% -2 1,5 -1 -0,5 0 0,5 1 1,5 2 2,5 3 3,5 4 Asymmetry

Figure 13: distribution of asymmetry among the different SKUs considered

This analysis shows the relevant asymmetry of demand. In particular it shows that mainly right asymmetry can be identified, thus stating that demand tends to be bimodal: among days with rather low demand, others shows very huge amounts of product requirements, thus making demand lumpy. This in fact is consistent with more in depth graphical analysis that, as previously stated, shows days in which huge peaks of demand shows up (see Figure 8 for an example of demand pattern).

By means of interviews with managers and staff we discovered that the main causes of these demand peaks can be trace back to promotional activities. In fact some customers perform promotional activities towards end-consumer and so ask the producer to apply

181

Part III Chapter 3: Nestlé Italiana

similar promotional activities to them. As a matter of fact these activities deeply influence demand moreover because customers that apply promotions are generally the most relevant in terms of total demand. Since the main factors of variability are tied to seasonality and promotional activities, we may identify a few guide lines in the development of the forecasting solution.

1. Given that these factors contribute significantly towards demand variability they have to be taken into account by the forecasting system. 2. Moreover attention has to be paid on the fact that promotional activities are applied differently on different customers and that customers may react differently on promotional activities. In fact, during promotions, some customers may tend to distribute their purchases differently as, for example, some may tend to buy more in the first weeks and less afterwards, while others may try to speculate on the price reduction by buying more in the last weeks so to increment their inventory with less costly material. 3. Different kinds of promotions are applied; in fact, some customers’ promotions are “personalized” since the company arranges promotion specifically for a single customer. However, since this activity is rather costly, it tends to be applied only with the most important customers, while for the others much wider promotions are conducted, typically at chain level. This implies that for many customers, during certain periods of time, promotional activities are applied and again these customers may react differently towards their development.

These considerations are the main basis for the solution development. However, to better understand both the problems faced by the company and the improvements impact, attention has to be paid to the actual forecasting system.

3.2 Actual forecasting performances

The actual forecasting system adopted by the company is essentially based on a simple moving average. The company uses the SKEP forecasting systems provided by Dynasys. That simply evaluates a 19 weeks ahead moving average to model future requirements. It is not surprising that this kind of systems is not capable of dealing properly with such a variable demand. Table 2 summarizes the performance of this system in Nestlé, separated for distributional centre. In particular the system evaluates a two week-ahead forecast to provide planning with the required information to make planning decisions.

182

Part III Chapter 3: Nestlé Italiana

D.C. %BIAS %MAD Milano +3% 28% Modena -3% 22% Roma +8% 28% Caserta +15% 31% Catania -69% 85% Italy -2% 32% Table 2: performances at distribution centre level of the forecasting system adopted by Nestlé

As it can be noted the system is totally inadequate. In fact this is well known within the company, so the planning staff deeply reviews forecasts to produce much accurate evaluations. In particular this review process is done by analyzing the 22 most relevant customers so to identify the impact of particular events on their demand and on the overall requests. Table 3 shows the performances Nestlé actually obtains after the revision by forecasters and the variation between this and the previous ones.

D.C. %BIAS Variation %MAD Variation Milano +5% +2% 21% -7% Modena -0.5% +2.5% 20% -2% Roma +3% -5% 24% -4% Caserta +2% -13% 17% -14% Catania -6% +63% 25% -60% Italy +2% +4% 21% -11% Table 3: Nestlé final forecasting performances

As a matter of fact the human contribution is rather significant, however this activity is rather time consuming and thus the company desires to provide forecasters with a better initial input. The impact of the forecasting error is pretty relevant. We are considering a fresh food business; fresh food has, by law, a maximum life of 4 weeks, moreover, the company may not sell products to customers that have more than 2 weeks of life. So the inventory management of the company is heavily influenced by forecast errors. To measure this impact we considered the relationship between negative forecast errors (so when forecast underestimates demand) and stock-outs. This analysis showed that a relevant correlation exists among these two dimensions. Table 4 shows the evaluation of Pearson’s correlation coefficients at distribution centre level for the considered products.

183

Part III Chapter 3: Nestlé Italiana

Regression D.C. Correlation Coefficient Milano 0.39 38% Modena 0.40 35% Roma 0.54 42% Caserta 0.48 50% Catania 0.90 97% Italy 0.46 45%

Table 4: correlation between negative forecast errors and stock-outs at distribution centre level

As it can be noted correlation are pretty relevant. Single products regression analysis proved that this relationship is significant. Figure 14 shows an example of the relationship among negative errors and stock-outs for a single product.

Stock-outs Stock-outs

Negative forecast errors

Figure 14: relationship between negative forecast errors and stock-outs for Nesquik Snack

So this analysis proved that since correlation is in average equal to 46% and the regression coefficient is in average equal to 0.45 (see Table 4), that means that 1 Kg. error in forecasts transforms into 0.46 Kg. stock-outs. Of course, as also Figure 14 shows, when errors are small, the inventory management system is capable of reducing their impact by means of its flexibility, however, we errors are much higher, stock-out tend to show up significantly.

This analysis proved that forecasting is a relevant matter for this company and, since forecasting performances appear to be low, this justifies putting efforts into this process.

184

Part III Chapter 3: Nestlé Italiana

This forced us to identify new solutions to forecast demand. In particular, from now on, attention will be paid only towards one of the distribution centre: Milano.

185

Part III Chapter 3: Nestlé Italiana

4 Solution Design

From the previous analysis it can be noted that a relevant source of variability is tied to the existence of promotional activities and the presence of high seasonality. Thus significant benefits can be obtained by considering these factors in the demand forecasting process. In particular attention will be paid also on the heterogeneity that exists among customers that apply promotions since we argue that relevant improvements can be obtained by taking into account this element. However a relevant problem is tied to the definition of how many customers consider in this situation. In fact, we argue that the cost of managing forecasts at a customer level to properly consider promotional activities is tied to the number of customers considered. Actually, in fact, the forecasting staff monitors 22 customers chosen according to different criteria, based on customer relevance, historical factors and so on. We decided to focus only on these 22 customers, so to make fair comparisons between this solution and the actual performances. However we will also consider how changing the customers on which apply this method influences forecasting performances.

Figure 15 shows the Pareto curve of demand for each customer.

Figure 15: Pareto curve of total demand

In fact demand tends to be rather concentrate in few customers (the first 55 customers are responsible for 86% of demand). In fact the 22 chosen customers are responsible for

186

Part III Chapter 3: Nestlé Italiana

almost 52% of total demand, thus even considering their contribution to total variability makes it possible to gain a relevant part of forecast.

4.1 Forecasting for the 22 monitored customers

To improve forecasting accuracy on these customers, we introduce two key elements. 1. First of all, as demand variability is mainly tied to single customers’ variability, we should foresee each customer’s demand so to obtain a global forecast by aggregating the single ones. Even if this approach is probably effective, the main problem is the managerial and computational complexity of handling thousands of forecasts. Moreover, as a relevant source of variability is due to promotional activities, we argue that is worth focusing only on customers adopting these policies, which furthermore represent a relevant part of total demand.

2. A second key element is the merger of historical data with information on past and future promotional activities. Since these phenomena are critical for demand forecasting, we believe that a relevant improvement can be obtained by using this information in a quantitative forecasting technique. In the case considered, only information on the timing of promotions is available, and only for a group of 22 monitored customers. This means that, for each customer, the company knows only if during a certain week a promotion takes place.

These elements lead us to the general framework of the solution based on evaluating single forecasts for each of the 22 monitored customers using information regarding their specific promotional activities. In literature there is no established method for forecasting lumpy demand on the basis of historical data and information on the timing of peaks. Cachon and Fisher (1997) propose a solution based on the separation of demand peaks from total demand so to deal with two more even series (similarly to what has been applied in the Whirlpool Europe case). While their approach is based on the evaluation of the impact of peaks on total demand variability, we argue that exploiting the available information regarding these phenomena may lead to a better operational performance.

The approach here adopted is based on the idea of identifying the various elements that cause demand variability and try to develop proper methods to manage them. In particular,

187

Part III Chapter 3: Nestlé Italiana

we hypothesize that variability can be either systematic or irregular, the former due to phenomena like seasonality, and the latter related to events like promotions. Figure 16 shows the general framework of the suggested method, through a step-wise process.

Forecast of regular demand

Isolation of Depureted Evaluation Historical Demand systematic Historical of irregular Data Forecast variability Data variability

Forecast of irregular fluctuations

Figure 16: the general framework for the suggested method We suggest developing a separate forecasting solution for each kind of variability, and to this purpose each of them should be isolated.

Systematic Variability

The first step is to consider systematic variability, which has been widely discussed in the literature, but presents some additional problems when overlapped to the effect of promotional activities, since it becomes more difficult to identify the two contributions to global variability. In particular, to separate the effects of the different components of variability, we suggest to rely on the information regarding the timing of past promotions. Through this information, demand series can be split in two components: a base series, consisting of the regular demand, i.e. demand in the periods without perturbing events like promotions, and a peak series, made of the irregular demand values.

The example in Figure 17 shows a typical demand pattern and the result of the splitting procedure, highlighting the difference between the two series.

188

Part III Chapter 3: Nestlé Italiana

4000

3500 peaks

3000

2500

2000 4000 demand 1500 3500 demand 1000 3000

500

2500

6 0 4 8 0 2 6 1 5 7 5 9 9 3 2000 02 38 4 50 0 09 21 37 49 01 05 17 80 81 8 9 92 945 0 01 98 981 981 983 984 98 990 9 99 991 99 992 9933 99 9 99 00 000 00 S S9 S S S9 S9822 S9826S S9834 S9 S S98 S S S S S9913S S S9 S S S S9941 S S S0 S S S0 S S0021 weeks

demand 4000 1500

3500 1000 base

3000 500

2500 0

17 29 41 01 13 806 810 818 822 830 834 842 846 901 905 913 925 937 949 009 021 S9802 S9 S9 S9814 S9 S9 S9826 S9 S9 S9838 S9 S9 S9 850 S9 S9 S9909 S9 S99 S9 921 S9 S99 S9933 S9 S99 S9 945 S9 S00 S0 005 S0 S00 S0017 S0 2000 weeks demand 1500

1000

500

6 0 8 2 0 4 8 2 6 0 1 5 9 7 1 5 9 3 1 9 1 9 3 1 0 1 1 3 4 0 0 1 2 2 2 4 0 1 2 84 90 9 93 945 005 9802 98 98 9814 98 98 98 9 99 9 99 9913 99 99 99 99 00 00 0017 00 S S S S S S982 S9826 S S983 S983 S S S985 S S S S S S9 S S S9 S9937 S S9 S994 S S0 S000 S S S weeks

Figure 17: the separation of demand through information regarding promotions

As far as the estimation of systematic variability is concerned, this approach provides a base series that is not affected by promotions, but only by seasonality and trend, thus allowing to measure their contribution. In particular, since this kind of variability is not related to the specific customer, but to the specific product and geographic area, the base series of all the customers should be aggregated, so to obtain a better estimation. In this way those periods that have missing values for some customers, because affected by promotions, do not prevent from evaluating systematic fluctuations of demand. Figure 18 shows the global scheme of this evaluation process.

Single Identification Aggregation Estimation of Depuration Depurated customers of the base of the base systematic of historical historical demand series series variability data data

Figure 18: process of isolation of systematic variability

The estimation seasonality factors has a double purpose: first, it allows to clean historical data from these effects, thus obtaining a series that is affected only by irregular fluctuations. Second, it allows to estimate the future effect of this kind of variability, in order to improve forecast accuracy. In case of lumpy demand, irregular variability is also relevant and needs to be managed with an appropriate method.

189

Part III Chapter 3: Nestlé Italiana

Irregular Variability

The proposed method aims at identifying the specific sources of demand variability, by splitting the demand series in the two components previously described on the basis of the information on the promotional periods; this time, however, the series considered is cleaned by the effect of systematic variability. In this way the base series is almost regular, while the peak one is only affected by irregular variability. This second step allows to perform actual forecast, separately for the two series; since the base series is not affected by relevant variability, a traditional extrapolative technique, like exponential smoothing, is perfectly suitable for forecasting.

The peak series instead, although less variable than the original one, still shows high fluctuations; this is due to the fact that demand is not constant during promotional periods, but is affected by the reorder process that specific customers have during these occurrences. Traditional techniques in this case can provide only an estimation of the average demand during promotions, thus allowing only limited improvement in accuracy. In fact, as previously argued, each customer tends to be heterogeneous in his/her reaction towards promotional activities, since they may tend to distribute their purchases differently during promotional periods. In particular, we argue that this kind of behavior is intrinsic in the customer nature, thus stating that customers will always tend to show this kind of behavior during time.

Observation of data, in fact, shows that demand peaks frequently have an ascending trend at the beginning and a descending one at the end, or anyway a characteristic pattern which is rather constant over time. Figure 19 shows an example of how demand peaks for a certain customers are distributed over time. It can be noted that they tend to shape similarly.

Thus the problem of forecasting a peak of demand can be divided in two sub-problems: first of all one must estimate how many parts will be ordered, then what is the shape of the specific peak. We formulate the hypothesis that a correlation between the peaks’ shape and the buying behavior of customers exists; it’s true indeed that if the pattern remains similar over time it can be estimated and effectively used to forecast demand. This assumption has been verified by graphical analysis on the different customers’ behaviors thus showing that in fact customers tend to purchase similarly (in terms of distribution of purchases during

190

Part III Chapter 3: Nestlé Italiana

promotional activities) since, for example some customers leverage on the price discount to rise their inventory thus buying more in the last weeks, while others prefer not to increase their stock position thus they buy marginally more in the first weeks.

Figure 19: example of peaks shape

In this way each promotional week has been classified as one of the following case: a) Initial transitory b) Intermediate period c) Final transitory

For each kind of week the shaping coefficient is estimated according to the past behavior; in particular during the intermediate period it is equal to 1, while for the other periods it is estimated as ratio between the average demand during the specific period and average demand during the intermediate period. Figure 20 exemplifies this procedure.

To introduce this information within the forecasting systems, we adopted the described profiling coefficients similar to seasonality coefficients aimed at evaluating the percentage distribution of purchases during promotional periods. In this way demand forecasting during promotional periods is achieved by forecasting the average demand during promotions with an exponential smoothing and adjusting it on the base of the pattern estimated for the peak.

191

Part III Chapter 3: Nestlé Italiana

Figure 20: promotion profiling

The actual forecast for the customer’s demand is made according to the information about future periods: if a promotion is expected, the forecast for the peak series is used, otherwise the one for the base series is considered. In the end, this forecast can be adjusted by means of the estimated seasonality. In this way it’s possible to obtain a forecast for each monitored customer’s demand and the total demand of the group can be estimated simply adding the single forecasts.

Figure 21 shows the overall process to forecast the monitored customers. Even if this system is able to capture relevant part of variability an important consideration has to be taken into account regarding the customers considered. In particular, even if these 22 monitored customers are responsible for a relevant part of total demand (almost 52%), other few customers contribute significantly. So we suggested considering also other 8 customers thus monitoring 30 customers responsible for 70% of total demand. This solution, which of course influences managerial costs, however impacts significantly on forecasting performances.

192

Part III Chapter 3: Nestlé Italiana

Exponential Peaks profiling Peaks profiling Future Promo Smoothing Promo Info (peaks) Info

Peak series

Demand Seasonality Seasonality data Filter Join

Stable Exponential series Smoothing Customer (base) Forecasts Figure 21: the overall forecasting process for the 22 monitored customers

In particular, Table 5 compares the performance for the 22 and the marginal improvement for the 30 chosen customers of this solution.

30 Customers Model Product 22 Customers Model Improvement %MAD %BIAS %MAD %BIAS Pasta sfoglia 11.7% 7.4% -1.4% +0.2% Salsa pesto 8.2% 3.2% -1.1% +0.3% Fruttolo misto 20.8% 14.2% -1.3% +1.2% Ravioli al brasato 17.7% 5.3% -3.8% +0.3% Base pizza 12.9% 10.1% -1.5% -0.4% Cappelletti 20.3% 9.4% -4.7% +1.9% Fagottini R/R 15.1% 1.1% -2.2% +0.7% Formaggino Mio 18.3% 12.3% -1.5% -1.5% Salsa 4 formaggi 11.3% 8.0% -1.9% +0.4% Fagottini B/R 15.7% 1.7% -2.8% +0.5% Profiteroles 17.8% 0.7% -1.7% +0.4% Average (class A 14.4% 7.1% -1.9% +0.4% products)

Pasta frolla sottile 15.1% 0.8% -3.0% +1.0% Pizza Margherita 21.4% 9.6% -3.0% +3.3% LC1 bianco 28.4% 23.5% -1.2% +0.5% Pasta frolla 14.7% 10.2% -1.8% -1.6% Dolcemio Nestlé 28.8% 27.2% -0.9% +1.4% Strozzapreti 25.2% -2.2% -2.4% +3.1% Average (class B- 20.7% 10.7% -2.2% +1.1% C products)

Total Average 14.9% 7.4% -1.9% +0.4% Table 5: comparison among the 22 monitored based forecasting solution and the 30 monitored one

193

Part III Chapter 3: Nestlé Italiana

4.2 Residual customers forecasting approach

The previous part of the model focuses on a specific part of demand influenced essentially by promotional activities. In particular, this system focuses on those customers that tend to apply promotions and manages this kind of variability source according to the different reactions customers have. However, a relevant part of variability is still unmanaged, since all the customers that have not been considered right now have to be forecasted. For this kind of customers we propose two different methods: a solution that essentially considers these customers as a whole and forecasts them with a simple exponential smoothing technique (first solution), then a different approach that again tries to leverage of the different behaviors customers show (cluster based solution).

4.2.1 Aggregate solution

To evaluate the forecast regarding total demand we have to focus on the remaining customers’ demand. In case the portion of demand generated by not-monitored customers proves to be almost regular, i.e. it has very little non-systematic variability, it’s possible to use an extrapolative technique with good results. This ideal situation is applicable when a reduced number of customers (the monitored ones in this situation) is responsible of the majority of demand variability. In common practice it’s difficult to obtain such a result, but often what is really important is to monitor large customers, since they represent such a big portion of demand that, even if the remaining part is not regular, the total forecast obtained is accurate enough.

At first, to evaluate the forecast of non-monitored customers we adopt a simple exponential smoothing. Table 6 shows the performances of the global solutions compared with those of the forecasting system the company adopts (SKEP) and those obtained after the revision by the forecasting staff. This comparison shows that significant improvements can be obtained by leveraging on this approach. Moreover, since the main difference between the Nestlé actual forecast and that provided by the system is due to the systematic collection of data regarding promotions and to the evaluation of the effect of promotional activities on each single customer in terms of promotional profile, it can be noted that focusing on the sources of demand variability and heterogeneity can deeply influence forecasting performances.

194

Part III Chapter 3: Nestlé Italiana

A matched-pair analysis on the %MAD of the different solutions identifies that the performances between the proposed solution and those provided by SKEP and Nestlé are significantly better. Table 7 and Table 8 provide the results of this analysis. Quite interestingly even if Nestlé performs better than SKEP, this difference is less statistically significant (see Table 9), stating that considering the effects of heterogeneity on demand may lead to significant improvements.

Nestlé Nestlé Product %MAD BIAS SKEP %MAD SKEP BIAS %MAD BIAS P. Sfoglia 9.1% -0.1% 19.0% -1.2% 23.5% -3.2% Salsa pesto 7.9% -1.8% 11.3% 4.5% 13.9% 0.7% Fruttolo m. 10.1% 0.9% 14.1% 6.8% 15.8% 7.4% Ravioli b. 19.5% -4.1% 24.3% 8.6% 20.5% -6.1% Base pizza 9.0% 1.1% 14.4% -2.5% 15.6% -0.4% Cappelletti 18.8% -4.5% 29.9% 15.5% 23.9% -4.9% Fagottini RR 14.8% -2.4% 17.9% 7.5% 27.9% -3.1% Formag.Mio 8.6% 0.6% 15.4% 6.1% 16.3% 2.1% Salsa 4 for. 7.0% -0.9% 9.9% 6.4% 10.4% -0.1% Fagottini BR 15.2% -1.7% 16.0% 7.2% 26.1% -1.7% Profiteroles 13.4% -1.7% 22.8% -5.8% 26.5% -6.9% Average 11.0% -1.0% 17.0% 3.9% 19.5% -0.8% Class A

P. Frolla s. 13.4% 2.5% 22.7% -8.3% 23.8% -2.9% Pizza m. 16.0% -4.0% 24.4% -7.0% 24.7% -2.2% LC1 bianco 12.9% 0.0% 19.4% -0.9% 22.1% 5.6% P. Frolla 14.2% 2.5% 25.5% -2.9% 27.5% 0.7% Dolcemio 9.5% 1.3% 11.0% 3.3% 14.3% 7.7% Strozzapreti 25.2% -6.8% 34.6% 14.4% 28.4% -1.7%

Average 14.4% 0.0% 22.6% -3.3% 23.7% 0.5% Class B

Average All 11.3% -0.9% 17.5% 3.3% 19.9% -0.7% Products

Table 6: comparison between SKEP, Nestlé and the proposed solution performances

195

Part III Chapter 3: Nestlé Italiana

Nestlé 19,5647 t-Ratio 7,688245 Proposed Solution 13,2118 DF 16 Mean Difference 6,35294 Prob > |t| <.0001 Std Error 0,82632 Prob > t <.0001 Upper95% 8,10466 Prob < t 1,0000 Lower95% 4,60122 N 17 Correlation 0,88846 Table 7: matched-pair analysis between Nestlé and the proposed solution

SKEP 21,2471 t-Ratio 8,213283 Proposed Solution 13,2118 DF 16 Mean Difference 8,03529 Prob > |t| <.0001 Std Error 0,97833 Prob > t <.0001 Upper95% 10,1093 Prob < t 1,0000 Lower95% 5,96133 N 17 Correlation 0,72031

Table 8: matched-pair analysis between SKEP and the proposed solution

SKEP 21,2471 t-Ratio 1,569945 Nestlé 19,5647 DF 16 Mean Difference 1,68235 Prob > |t| 0,1360 Std Error 1,0716 Prob > t 0,0680 Upper95% 3,95404 Prob < t 0,9320 Lower95% -0,5893 N 17 Correlation 0,76953 Table 9: matched-pair analysis between SKEP and Nestlé

However, even if forecasting performances are heavily improved we argue that significant benefits can be obtained by focusing again on customers’ heterogeneity.

As previously introduced among the 22 monitored customers that apply promotional activities, all the other customers left tend to show a residual variability. This is due to many reasons, but mainly again seasonality and promotional activities play a primal role. In fact, also many of these residual customers tend to show a rather relevant variability, so we argue that better performances can be obtained by leveraging on these residual customers performances. In fact, while studying the demand of these residual customers, we identified that some of them show some kind of similarities in their demand pattern. The reasons of this behavior may be tied to different elements. For example, as previously anticipated, promotional activities are conducted also for these customers all together, so they all tend to show

196

Part III Chapter 3: Nestlé Italiana

greater requests during these periods, with some differences due to their sensitivity towards price changes. Moreover, they are probably influenced by other similar factors, while their major difference is again in how they react towards these factors. In this way we adopted a modification of the forecasting system for the non-monitored customers, which leverages on these dimensions of heterogeneity, by means of the application clustering techniques.

4.2.2 Cluster based solution

We applied a clustering approach on the remaining customers based on demand data for each single product. Since we are not able to observe the different factors that may influence their demand, we argue that a proxy that may be adopt to cluster customers is demand. In particular, since we want to cluster customers according to their specific demand pattern, the clustering procedure is applied by normalizing demand data and applying a k- means cluster analysis. The normalization process guarantees that cluster are built according to the demand pattern and not according to the demand size, which may bias the estimation process.

Cluster analysis is a convenient method commonly used in many disciplines to categorize entities into groups that are homogeneous along a range of observed characteristics. Cluster analysis, however, encompasses a relatively wide variety of techniques that attempt to form groups with internal cohesion and external isolation (Gordon, 1980). Two major groups of clustering methods can be distinguished by their use of dependent and independent variables. Methods that distinguish dependent and independent variable are called predictive clustering methods. Such methods form clusters that are homogeneous on the estimated relationship between the two sets of variables. Descriptive clustering methods do not make a distinction between dependent and independent variables forming clusters that are homogeneous along a set of observed variables. Each of those two groups of procedures fits into three major categories of cluster analysis that may be distinguished as: nonoverlapping methods, overlapping methods and fuzzy methods.

In the case of nonoverlapping clusters a consumer belongs to one and only one segment. Two different types of nonoverlapping clustering methods are commonly distinguished: hierarchical and nonhierarchical methods. Hierarchical methods do not identify a set of clusters directly. Rather, they identify hierarchical relations among the N objects on the

197

Part III Chapter 3: Nestlé Italiana

basis of some measure of their similarity. Nonhierarchical methods derive a partitioning of the sample into clusters directly from raw data. Methods providing overlapping and fuzzy partitioning relax the assumption of external isolation of the clusters. In the situation of overlapping clusters, a customer may belong to more than one segment. In the case of fuzzy clusters, subjects have partial membership in more than one segment. Figure 22 summarizes the different clustering approaches.

Hierarchical

Nonoverlapping

Nonhierarchical

Clustering Overlapping Methods

Fuzzy

Figure 22: different clustering methodologies

Since for our objectives each customer should belong to a different group, we focused on nonoverlapping techniques. Moreover, these systems are applied by many different statistical software packages, while the other techniques are not very much adopted in software systems. Thus, as we wanted to make the solution really applicable we again adopted nonoverlapping methods. We will not go into detail regarding the available algorithms, we refer to Wedel and Kamakura (2001) for details.

The application of cluster analysis on demand forecasting is a rather new issue. In fact, very few contributions are provided in current literature regarding cluster analysis applications if operations management. For example Tyagi and Das (1999) adopt a clustering technique to group customers in order to minimize demand variability and its effects on inventories performances.

The reactions towards decisional and external factors may different for each customer; however, we also argue that it may differ between different products. There is no reason to believe that customers buy in the same way different products and, moreover, that they

198

Part III Chapter 3: Nestlé Italiana

respond similarly to different factors applied to different products. Thus, we conducted the cluster estimation at a single product level; in other terms, for each product the following procedure was applied. In particular, for each product, the problems that have to be managed are as follows:

1. Decide on which customers the clustering procedure may be applied A relevant problem is tied to the fact that some small customers purchase very rarely, with a vary sporadic pattern: Since the lack of significant demand data doesn’t make clustering applicable, we separated those customers that have very few demand occurrences.

2. Decide how many clusters apply The main problem in deciding how many clusters have to be considered is rather tough. From one side, in fact, the more clusters are accepted the more they should be homogeneous. From another side, however, many clusters make the forecasting process much more complicated and costly. Moreover, this trade-off is complicated by the fact that there is no statistical rule in deciding how many clusters should be adopted. To perform this operation we adopted a hierarchical clustering technique. In particular, this technique evaluates the so-called agglomeration schedule which reports for each iteration of the clustering process a coefficient that represents the distance between the clusters joined at each stage. The empirical rule states that when this coefficients varies significantly, it represents that very different clusters are forced together, thus stating that the considered clusters should be kept separated. Table 10 provides an example of agglomeration schedule. In particular it can be noted that a significant increase in the coefficients occurs between stage 50 and 51, thus stating that 5 cluster should be considered. For details on this application we refer to Wedel and Kamakura (2001). By the adoption of this approach on the products considered in this analysis, the number of clusters essentially varied between 3 and 5.

3. Define clusters To identify which customers fit in which cluster we applied a k-means cluster analysis based on the normalized demand previously introduced. Each cluster is defined by as many variables as the number of weeks used for the setting of the cluster and a cluster centre is defined, corresponding to the customer that better represents the cluster itself. Figure 23 provides an example of this.

199

Part III Chapter 3: Nestlé Italiana

Each customer is assigned to a cluster according to the Euclidean distance from the cluster centre.

Cluster Combined Coefficients Stage Cluster First Next Stage Appears Stage Cluster 1 Cluster 2 Cluster 1 Cluster 2 1 23 31 4117,763 0 0 5 2 28 32 5097,931 0 0 5 3 1 4 5586,632 0 0 12 4 25 29 6321,074 0 0 11 5 23 28 6734,121 1 2 9 6 24 26 6855,236 0 0 10 ……… ……… ……… ……… ……… ……… ……… ……… ……… ……… ……… ……… ……… ……… ……… ……… ……… ……… ……… ……… ……… 46 12 18 19495,605 44 0 52 47 45 47 20217,303 45 43 48 48 45 50 22128,240 47 0 49 49 45 48 23594,773 48 0 50 50 45 46 24031,760 49 38 54 51 1 23 147368,516 37 14 52 52 1 12 164951,641 51 46 53 53 1 34 169836,031 52 41 54 54 1 45 215648,750 53 50 0 Table 10: agglomeration schedule example

Week t1 Cluster

Normalized demand demand Normalized centre

Cluster centre

Normalized demand Week t2 Figure 23: scheme of the clustering approach and the identification of clusters’ centers

In fact correlation among customers and their centers are pretty high. Table 11 provides the correlation inside the different clusters for the different products considered in this analysis, evaluated by means of Pearson’s Coefficient1. This was evaluated by measuring correlation during two different time ranges.

1 Correlation was evaluated as weighted among the different clusters by means of each cluster total demand 200

Part III Chapter 3: Nestlé Italiana

First we considered the period by which clusters are defined (setting phase); the correlation coefficients are pretty high (0,86) since the clustering algorithm essentially tries to maximize this element. We evaluated correlation also during the time interval adopted for testing the forecasting approach (simulation phase), thus discovering that correlations were still pretty high.

Correlation Setting Simulation Pasta sfoglia 0.88 0.82 Salsa pesto 0.80 0.76 Ravioli brasato 0.86 0.83 Fagottini R/R 0.91 0.86 Formaggino Mio 0.86 0.70 Average 0.86 0.80 Table 11: correlation within cluster during the setting and simulation phases

Figure 24 shows an example of the correlation among centre and its cluster.

Kg.

weeks Figure 24: example of correlation between a cluster and its centre (Cedis)

However to effectively adopt this method we have to guarantee that centre are stable during time, so clusters should be regularly updated during time.

201

Part III Chapter 3: Nestlé Italiana

The identification of the cluster centers makes possible to monitor these customers to collect information regarding their variability, similarly to what has been done with the customers previously considered. The forecast for each cluster is evaluated adopting a linear relationship estimated according to the historical demand data of cluster and of its centre. The adoption of this technique is justified by the correlation previously introduced.

The relationship is as follows:

DCt,i,p = ai,p ⋅ X t,i,p + bi,p where:

DCt,i,p = demand of cluster i during period t

Xt,i,p = demand of the customer that represents the i cluster’s centre during t ai,p, bi,p = coefficients of the regression estimated with the least square minimization

Figure 25 shows the scheme of this solution.

Demand data

Forecasting Customers Regression Clusters method definition Clustering relationship forecast

Figure 25: cluster based forecasting approach

For those customers that are not considered into the clustering procedure it can be noted that their overall contribution to demand variability is much reduced and that when aggregated all together the demand pattern tend to be quite stable. Thus we adopted a simple exponential smoothing to forecast their demand as a whole.

Then comparison will be provided between the previously cited proposed solution and the modification based on the clustering approach. The reasons for this two comparisons is due to the kind of data provided by the firm. In fact, while for all the considered products data regarding promotions was available for all the monitored customers, only for 5 SKUs this sort of information was available for the other customers, thus making the clustering approach applicable only on this smaller set of products.

202

Part III Chapter 3: Nestlé Italiana

Results are provided in Table 12.

Product Proposed model Difference %MAD %BIAS %MAD %BIAS Pasta sfoglia 9.8% 0.0% +0.2% +0.2% Salsa pesto 9.1% -1.4% +1.0% -1.9% Ravioli brasato 17.5% -4.9% -0.8% -1.4% Fagottini R/R 13.4% -0.8% -3.4% -1.8% Formaggino Mio 12.9% 5.2% +4.0% +4.7% Average 11.2% -0.6% +0.3% -0.4% Table 12: performance differences between simple regression and cluster based solution

As the table shows apparently there is no significant improvement by the application of the clustering technique, in particular statistical analysis cannot reject the hypothesis that the two systems perform equally. Quite interestingly the potential improvements of this technique are tied to the stability of the clusters; in fact if a cluster maintains stable over time, this provides potential benefits for the system, on the contrary, the system leverages on relationships that don’t exist any more. Figure 26 shows the relationship between the coefficient of correlation during the simulation period and the percentage improvement of the cluster solution.

5,0%

4,0%

3,0%

2,0%

1,0%

0,0% 0,6 0,65 0,7 0,75 0,8 0,85 0,9 -1,0%

-2,0% Percentage improvementPercentage 2 -3,0% R = 0,9136

-4,0% Correlation during simulation period

Figure 26: regression between percentage improvement and cluster stability

The linearity regression states that the less stable are clusters the less effective is their application, which makes sense. However this doesn’t help in understanding when this solution may be applied.

203

Part III Chapter 3: Nestlé Italiana

Interestingly, we identified a significant relationship between the improvement of this method and the variability of the demand that the system cluster (Figure 27 exemplifies this linkage). This made it possible to understand that this sort of system isn’t capable of providing better performances if the residual variability is very low. However, in those case where variability tends to be still relevant this system is capable of rising performances (for example for Fagottini R/R from 13,4% to 10%, improvement of more than 25%).

5,0%

4,0%

3,0%

2,0%

1,0%

0,0% 0% 5% 10% 15% 20% 25% 30% 35% 40% -1,0%

-2,0% Percentage improvementPercentage -3,0%

-4,0% Demand variability

Figure 27: relationship between percentage improvement and demand variability

As described in the demand analysis, residual demand variability tends to be high when demand is deeply affected by seasonality and promotional activities, so since the cluster system performs better when variability is high, we argue that this system should be applied when systematic or managerial variability is predominant. In fact, variability sources as promotions are very customer-specific, thus if some customers show similar reactions towards promotional activities, then their demand pattern will probably look similar and thus clustering provides to be significant. To verify this assumption we tested this approach in much wider detail.

204

Part III Chapter 3: Nestlé Italiana

5 Cluster analysis simulation

In this part of the work we will describe the simulation framework that has been developed to analyze the impact of demand features on the significance of clusters of customers, in order to define a range of applicability of the forecasting method based on them. In particular we adopted a mixed methodology: first of all we simulated demand according to the particular patterns further explained so to measure identify the relationship between clustering applicability and demand. Then we analyzed the same situation by means of the real demand data provided by Nestlé. Demand has been simulated starting from the hypothesis that a link between the customers’ demand exists, hypothesis that needs to be tested when considering real cases but that is based on a realistic ground. We considered systematic, managerial and random variability, assuming that the latter cannot provide correlation, since it is independent for each customer, while the former two do provide correlation, as depending on variables which are common to all the customers. In particular, systematic variability affects total demand and for this reason it should provide correlation, and this effect should be common to all customers; managerial variability instead is related to groups of customers, hence correlation can be found among them. In our simulation, we considered seasonality as systematic variability and the effect of promotions as managerial variability. This choice is due to the fact that these phenomena are quite common in practice. We believe and further argue that these two phenomena even enforce each other, since often promotions are carried out in high seasonality periods; in particular what can provide correlation between demands is the fact that in certain dates, such as feasts or celebrations, all customers tend to practice promotions. Starting from this point, we simulated the demands of a cluster of customers, among whom one has been considered the centre This is the one that will be monitored, and the other customers will be all related to him. This is a proxy of the output of a clustering technique applied to a wider group of customers and is representative of all the clusters derived through cluster analysis. The significance of the identified cluster is measured as the correlation between the demand of the central customer and the total cluster demand. During the simulation tests we will check this correlation, whose stability is assured by the way in which the simulation is modeled.

Namely, the following features has been considered and simulated:

205

Part III Chapter 3: Nestlé Italiana

1. demand seasonality and demand random fluctuations: since the first has been considered deterministic and the second stochastic, we believe that they will have opposite effects on the evolution of the correlation. The assumption on the characteristics of these features is supported by the fact that products’ seasonality is usually known and what is interesting is to understand in what cases the clustering approach is valid.

2. promotions and demand random fluctuations: we consider that promotions occur in particular dates that are well known periods of simultaneous activities by many customers, assuming that the simulated group has been formed during a setting time window in which these phenomena occur. The occurrence of promotions is supposed known, since the central customer is monitored and this implies that information on the timing of this activities is collected.

3. demand seasonality, promotions and demand random fluctuations: the three features have been considered together, to show how actually they interact.

Figure 28 reports the simulation framework adopted.

Random Simulation 1 Variability

Managerial Simulation 2 Variability

Systematic Simulation 3 Variability

Figure 28: the simulation framework

The result of this simulation is a framework that tries to define under which conditions the clustering approach can be useful, provided that if stability is given, accurate forecasts can be obtained.

206

Part III Chapter 3: Nestlé Italiana

5.1 Simulation 1

The first step of the simulation has been the generation of a random demand, in order to produce a pattern for the central customer of the cluster. This demand has been generated according to a Beta distribution, bounded between 0 and 10, with a mean of 5 and variance of 0.5. The reasons that led us to choose for this distribution assumption are mainly twofold: first of all, this particular distribution can easily be used to simulate interesting demand patterns as asymmetry. Beside this, a practical reason is that this distribution can be bounded, in order to limit it only between positive values. The effect of different parameters for the Beta distribution has been tested, in particular giving it an asymmetric form instead of a symmetric one, but the results obtained have proved to be always coherent with the first one, confirming the validity of our choice. To introduce systematic variability, a constant seasonality pattern has been assigned, through a series of coefficients varying of a fixed step each day of simulation, following a symmetric triangular path with a period of one year of simulation. We assumed that the seasonality effect on demand is multiplicative, as is generally accepted by researchers in this field.

Starting from this series of data, 10 more series have been developed, representing the other customers belonging to the same cluster: each of them is made of the sum of the first series and of a random term that represents the random variability. To simulate random variability, a second random number has been generated according to a uniform distribution bounded between 0 and 1, in order to determine both the direction and the entity of the variation with respect to the centre demand. This coefficient multiplies a variability parameter, which determines the maximum variation magnitude. Some tests with different numbers of customers have been performed, showing no significant changes in the results.

Data has been generated for 1.000 days of simulation, representing 5 years of 200 working days each; as a consequence, we considered 5 cycles of seasonality. The product of this simulation is, in the end, made of 11 series: the cluster centre demand and the 10 remaining customers’ demands. From the sum of these series we derived the total cluster demand, according to which seasonality coefficients have been estimated using a moving average of total demand and the assumption of multiplicative model.

To obtain a measure of the different kinds of variability, the Coefficient of Variation (CV) of total demand has been used as a measure of the total variability.

207

Part III Chapter 3: Nestlé Italiana

To estimate the variability due to the random component, we considered the CV of the deseasonalied total demand. To evaluate what part of variability is due to random phenomena, this last value has been divided by the CV of the total demand, obtaining a rate that tends to 1 when random variability is preponderant on seasonality and to 0 in the opposite case. To measure the correlation between the central customer’s demand and the total demand, the Pearson coefficient has been used. The simulation has been iterated for different values of the random variability parameter, in order to obtain different values of the indicator of variability.

5.2 Simulation 2

In the second simulation, demand has been generated in exactly the same way as before, introducing a series of Boolean coefficients that are non-zero when a promotion occurs. The assumption is that customers are clustered also on the base of their promotional policy and consequently they all practice promotions in the same periods. When a promotion period starts, the demand of the central customer changes to a value defined by a promotion size parameter, giving place to a new series with peaks corresponding to promotions. The promotion size has been set to the value of 10, but tests with different values have been performed, showing no significant change in the results. To obtain the demand of each customer in the cluster, this new series is summed to a random variability, generated in the same way as during the first test, through a random number and a parameter that determines the maximum variation. To measure variability this time both the CV of total demand with and without promotion days have been used, the first to estimate total variability and the second only the random contribution. The rate between this two value has been used as an indicator, that, similarly to the previous situation, tends to 1 when random variability is preponderant and to 0 in the opposite case; correlation has been used again as index of the cluster significance. Again the simulation has been iterated to map the values of correlation at different level of random variability.

208

Part III Chapter 3: Nestlé Italiana

5.3 Simulation 3

Finally a third simulation has been implemented to evaluate the interaction of the three kinds of variability in terms of correlation between the central customer’s demand and the total demand of the cluster. The features considered are the same previously described, i.e. seasonality as systematic variability, promotions as managerial variability and the usual random variability. The parameter that has been changed to evaluate correlation is, as usual, the maximum dimension of the random variation.

5.4 Simulation results

The results of the simulation are here described. In particular, Figure 29 shows the relationship that we measured between the correlation coefficient and the CVs ratio previously described, during the first simulation test, when only seasonality and random variability are considered.

Figure 29: relationship between the CVs ratio and clusters correlation – simulation 1

As we can see, there is a high value of the correlation coefficient when random variability is less relevant than seasonality, while this correlation tends to lower when random variability rises.

209

Part III Chapter 3: Nestlé Italiana

The second simulation test has been used to evaluate the effect of promotions compared to random variability, not considering seasonality to separate the contribution. The result is similar to the previous, but with a different path: correlation is very high until random variability is low, but decreases at a faster rate, as shown in Figure 30.

Figure 30: relationship between the CVs ratio and clusters correlation – simulation 2

The third simulation considers all the three kinds of variability: the results, reported in Figure 31, show that the combined effect of seasonality and promotions is still favorable to correlation, while random variability makes it decrease.

It is not possible to compare the results of the different simulations among them directly with the previous indicators, since this measure of variability is affected by the considered features and consequently assumes different values for the same level of variability in each case. To confirm this fact and to compare directly the three cases, the results of the simulation have been plotted considering only the total demand CV as index of variability, since it is a consistent measure. The graph in Figure 32 shows the results: at any given level of total variability, correlation increases passing from the case that considers only seasonality to that with only promotions and finally considering both of them. In fact, for any given level of total variability, the curves related to simulation 1, 2 and 3 show increasing values of correlation; in the same way, a given correlation is reached in presence of higher variability.

210

Part III Chapter 3: Nestlé Italiana

This is due to the fact that even if total variability arises while passing through these solutions, the most relevant part of this variability is due to systematic or managerial factors that influence all customers’ demand.

Figure 31: relationship between the CVs ratio and clusters correlation – simulation 3

Figure 32: relationship between demand variability and clusters correlation in the different simulations considered

211

Part III Chapter 3: Nestlé Italiana

Considering these results, the simulation performed supports our model for selecting customers with the purpose of collecting critical information to forecast lumpy demand. The general result of this simulation is that when demand is deeply influenced by systematic and managerial variability, the use of a clustering approach can be effective. At this stage an empirical test of this approach would be very interesting to understand whether the described conditions can be found in reality.

5.5 Application to Nestlé

To evaluate whether the simulation performed is consistent with reality, a real case has been considered, namely the Italian fresh product division of a worldwide known food multi-national. Historical data related to the period 1999-2000 were available, as well as information on promotional activities of the main customers. For a limited set of products (9 products belonging to different market sectors and specific brands), the approach described in this paper has been followed, in order to test it in a real contest.

First of all the main customers have been identified, evaluating also their contribution to demand variability and the feasibility of information gathering. Once total demand has been decreased of the portion generated by the monitored customers, residual variability has been measured; since it was still considerable, the clustering method has been used. Clusters have been built using data out of a range of 8 weeks of high seasonality and promotional activity in 1999, considering only customers who order frequently enough for a clustering technique to work.

To measure demand variability, both the total demand CV and deseasonalized demand CV have been calculated, estimating seasonality coefficients through a moving average. Finally correlation has been measured inside the clusters, using data of the following year, considering the same 8 weeks; in this way we investigated the joint effect of seasonality and promotions on cluster significance. Since for each product customers have been grouped in a certain number of clusters, ranging from 3 to 5, the weighted average of the correlation value for each of them has been considered.

212

Part III Chapter 3: Nestlé Italiana

In Figure 33 we mapped the position of the considered products on the same graph previously seen: according to demand features, correlation is different, following a path similar to that simulated.

Figure 33: distribution of Nestlé’s products on the CVs – Correlation map

Of course there’s no perfect correspondence between reality and simulation, also because in the simulation we used a given pattern for both seasonality and promotions that doesn’t exactly match with that of the real products. Anyway results are very interesting, since real data are consistent with the simulation. In fact the higher the CVs ratio, the lower correlation seems to exist. The relevant aspect of this test is that actually, for some products, customers can be grouped in some clusters that are steady enough to be used for forecasting.

213

Part III Chapter 3: Nestlé Italiana

6 Conclusions

In conclusion the overall forecasting method is essentially made of three parts: 1. Customers that are relevant in terms of sales and that apply promotional activities are considered separately. The forecasting systems considers demand at a product/customer level and adopts both information regarding promotions and demand data 2. Customers with correlated demand. The forecasting system is similar to the previous one, however it considers new customers that are identified by means of a cluster analysis. In fact forecast is evaluated at a cluster/product level, by means again of information regarding demand data and promotions is adopted. 3. Residual customer. The customers that remain are aggregated and forecasted by means of a simple exponential smoothing technique, so adopting only demand data.

Figure 34 provides a summary of the forecasting approaches, while Figure 35 provides the detailed structure of the overall forecasting system.

Kind of demand Forecasting approach Adopted information

Big customers Forecasts at Promo with highly customer / Activities variable product level info. demand

Customers Forecasts at Cluster with correlated customers’ analysis demand cluster / product level

Customers Forecasts Demand with regular aggregated at data aggregate product level demand

Figure 34: the overall forecasting approach

214 Part III

Exponential Future Promo Peaks Peaks Profiling smoothing Profiling promo Info (peaks) info

Peaks

Figure 35: the overall forecasting solution. solution. forecasting the overall Figure 35: series

Seasonality Custom. Filter Seasonality Join demand and events and events 215

Base Exponential series Smoothing Customer (base) forecast

Forecasting Customers Regression Cluster method choice Clustering relationship forecast Chapter 3: Nestlé Italiana Chapter 3:Nestlé Seasonality Exponential Seasonality Aggregate Forecast Product and events smoothing and events forecast addition forecast

Part III Chapter 3: Nestlé Italiana

A few considerations can be drawn from the results of this case.

First of all, the situation considered shows two main sources of demand heterogeneity: from one side promotional activities are applied on some customers and many distributional channels are managed together; thus, this makes supply chain rather complex so making demand heterogeneity high. Moreover, the reaction of customers towards promotional activities are different: in fact, from one side we have customers that tend to distribute their purchases during promotional periods differently, from the other some customers react differently in terms of average demand towards promotions. In these terms heterogeneity is influenced both from the supply chain complexity and the inner characteristics of the different customers. Coherently with the general framework defined, the solution proposed adopts different forecasting techniques that leverage both on different data and on different algorithms, moreover, the proposed techniques are applied differently on the different customers: for the monitored customers each forecast is applied separately, while for the non monitored customers that still influence demand variability, a cluster based approach is adopted.

Figure 36 describes the relationship that exists between the forecasting solution proposed and the causes of demand heterogeneity.

Different estimation processes

Inner Heterogeneity Heterogeneity Different forecasting structures

Supply Chain Complexity

Figure 36: the overall forecasting method and its relationship with heterogeneity

Moreover, this work provides evidence that improvements in demand forecasting can be achieved by means of the application of clustering techniques. In fact, this solution can be significantly efficient when a relevant trade-off between forecast accuracy and cost of

216

Part III Chapter 3: Nestlé Italiana

managing information exists. In fact, this kind of solution proves to be efficient when customers have relevant heterogeneity sources that can be traced back to some kinds of groups. One major limitation of this solution or at least in the way we applied it, however, is that it focuses only on demand and by means of it tries to identify homogeneous clusters. We argue that potential improvements can be gained by melting together more sources of information, so to obtain more efficient cluster. However, we argue that this solution, which tends to be more simple, should be applied at first, in those contexts that show many different sources of heterogeneity together. Moreover, we also identified environmental conditions, defined in terms of demand patterns peculiarities and causes of demand variability, that influence the cluster based solution application. These results may help companies in identifying whether this kind of solution should be applied or not. As a matter of fact, this analysis claims also for the generalization of this methodology, since many different industrial contexts show high impact of seasonality and promotional activities. Moreover, even if the solution provided has been designed in the Nestlé situation, its main characteristics are specific to the particular problem considered and not to the specific context analyzed.

217

Part III Chapter 3: Nestlé Italiana

218

Chapter 4 Ahold

1 Context and Purpose

Demand forecasting has been widely studied by researchers over the past decades. Many authors with rather different backgrounds (from inventory management to service operations, from econometrics to financial systems) produced a wide array of contributions. However most papers share a common feature: they either propose a new technique (i.e. algorithm) or evaluate the performance of existing ones. However, when trying to implement forecasting techniques practitioners often find out that forecasting is much more complex than the simple design or selection of an appropriate algorithm and involves the choice of the relevant pieces of information, the design of information systems, the control of data quality, and the definition of managerial processes. This has pushed researchers to study also systems to properly define the forecasting process in terms of information requirements and principles to avoid implementation problems. We refer to Armstrong (2001) for a general review of these approaches and techniques.

219

Part III Chapter 4: Ahold

One critical issue concerning implementation and adoption of forecasting techniques is the choice of the appropriate level at which forecasts should be evaluated. Typically forecasting involves the prediction of future demand for a given “product” over a given time bucket for a given location. However, the definition of these three dimensions is all but trivial. E.g., when forecasting fashion items (e.g., shoes) one should make clear whether a product is a combination of a particular design, material, color and size rather than a group of SKUs that share the same basic model. As to the location one should make clear whether the forecast is at store level rather than at distribution centre level or at chain level. The choice of the appropriate level of aggregation depends on the decision making process the forecast is expected to support. E.g., for short-term production planning probably a very detailed demand forecast is required, while for plant design or budgeting a rather aggregate forecast will be used. What makes the problem more complex in real contexts is that usually various levels of aggregation are adopted within a specific firm as various decision-making processes take place at the same time; thus, usually firms have to evaluate forecasts at different levels of aggregation. Indeed, some authors (Small, 1980) have shown that firms adopt different levels of forecasts and that this aggregation level is not directly tied to the specific firm’s characteristic, as size, industrial sector and so on. This leads to the need to identify a proper forecasting process so that forecasts at different levels of aggregation are consistent and provide the required information to each single decision making process (think for example of production planning vs. budgeting). This problem has stimulated research on the level of aggregation of forecasting process, which is often referred to as Hierarchical Forecasting. Hierarchical Forecasting is typically made up of two separate forecasting processes: 1) the bottom-up forecasting process and 2) the top-down forecasting process (Muir, 1979). In the bottom-up approach, individual forecasts for each demand segment (e.g., single SKU, single day, or single store) are combined to produce a forecast of aggregate demand (e.g., group of products, week or group of stores). This is referred to as the cumulative forecast since it is the sum of individual lower level forecasts. In the top-down process, aggregate demand data are used to forecast aggregate demand, then the aggregate forecast is disaggregated to produce what are known as derived forecasts for each demand segment. Typically, disaggregation is applied by means of historical data regarding the different segments, but some authors also provide other techniques to evaluate properly lower level forecasts; we refer to Gross and Jeffrey (1990) for a general review of these techniques.

Literature regarding Hierarchical Forecasting can be structured in two main areas. On the one hand many authors have focused on identifying whether the top-down approach

220

Part III Chapter 4: Ahold

outperforms the bottom-up one. Some authors (Theil, 1954; Grunfeld and Griliches, 1960; Schwarzkoph et al., 1988; Öller, 1989; Ilmakunnas 1990; Kahn, 1998; Lapide, 1998) argue that the top down approach is superior because of its lower cost and greater accuracy during times of reasonably stable demand. On the contrary, others (Orcutt et al., 1968; Zellner and Tobias, 2000; Weatherford et al., 2001) notice that, although it may be appealing to minimize the number of independent forecasts, individual forecasts are essential when it is important to capture differences among demand patterns (e.g. differences among stores and or among products). E.g., when different products have different seasonality, aggregation might make the estimate of seasonality rather tricky.

On the other hand, some authors take a contingent approach and analyze the conditions that make one approach more effective than the other does. Some authors (Mentzer et al., 1984a; Mentzer et al., 1984b; Weatherford et al., 2001) have shown that the level of data aggregation affects forecasting accuracy and that it is relevant for firms to properly identify at which level they should evaluate predictions. In synthesis, these authors provide evidence that a-priori there is no one best way. Miller et al. (1976), Barnea and Lakonishok (1980) and Fliedner (1999) show that the choice of top-down vs. bottom-up forecasting process depends on the degree of correlation among the sub-aggregate forecast variables and upon the magnitude of the correlation between forecast errors of subaggregate variables.

However, when comparing top-down (i.e. aggregate) with bottom-up (i.e. disaggregate) forecasting processes it is often appropriate to measure forecasting accuracy both at aggregate and disaggregate levels. This approach is consistent with the fact that often companies use both aggregate and disaggregate predictions to support different decision- making processes. In addition, the evaluation of forecasting performance both at aggregate and disaggregate levels makes the comparison between the top-down and bottom-up approaches more fair. Indeed, one might expect that measuring forecast accuracy at disaggregate level might make the bottom-up approach look better, and vice versa. This is not a common approach in past literature, however we argue that, since it can provide relevant contribution to understand performance, more attention should be paid to this issue. Moreover, the choice of the appropriate level of aggregation really depends on the demand generation process. Indeed, to accurately forecast demand one needs to estimate the drivers of demand fluctuations that are likely to depend on the single location and single SKU. In these terms customers heterogeneity (here stores heterogeneity) may be a significant issue to take into account. E.g., different stores might have different seasonality and have clients

221

Part III Chapter 4: Ahold

with different price sensitivity; on the other hand some products (i.e. shoes) are likely to sell very well early in the selling season while others sell only toward the end of it. However, the need to estimate these factors generates a trade off between the ability to capture differences among locations (e.g. stores) and SKUs (e.g., colors of shoes) and the ability to accurately estimate these differences. Indeed, as one tries to be very location and SKU specific (e.g. estimate the price sensitivity at store/item level) the number of parameters increases thus reducing the accuracy of the parameter estimates (given a set of information available). The choice of the proper position on this trade off depends on two factors: the availability of information (e.g. number of years of relevant history; number of transactions per period) and the degree of difference among locations/products. The more information is available the more one can choose to capture differences among locations/SKUs; the more the locations/products differ the more one is open to accept some inaccuracy in the estimate of parameters to capture such difference. Consider for example the Nestlé situation: while retailers in a given area share a rather similar seasonality, they differ significantly as to their ability to increase demand during promotions. Thus the forecasting technique estimates only one seasonality factor for all retailers in a given area, while estimates at single retailer level the effect of trade promotions. At last, the literature usually aggregates demand according to the product and/or supply chain structure. E.g., when aggregating demand over stores, one is tempted to cluster stores that are served by a given distribution centre. Indeed, this is very consistent with the output managers are expecting. However, in the paper we will discuss a second option: clustering demand according to the degree of heterogeneity of demand in terms of heterogeneity in time series. In the above example, one can cluster stores according to whether they have similar demand patterns rather than according to their geographical proximity. We believe that clustering (if appropriately implemented) can on the one hand enable to capture differences among stores (e.g., in terms of price sensitivity) as the clustering procedure groups stores with similar demand patterns (e.g., with similar reaction to price changes); on the other hand clustering enables to have a relatively large sample for each cluster of stores. In these terms, clustering is capable of moving the previously mentioned trade off towards more efficient solutions.

222

Part III Chapter 4: Ahold

2 Introduction to the company

Ahold is a multi-channel food provider with 2001 sales of Euro 66.6 billion. Ahold implements a multi-format strategy focused on meeting the needs of 40 million customers every week in currently 27 countries on four continents. Under their own local brand names, Ahold companies operate approximately 9,000 supermarkets, hypermarkets and convenience stores in the U.S., Europe, Latin America and Asia. Ahold also has significant foodservice activities in the U.S. and in three European countries. Worldwide Ahold employs more than 450,000 people.

2.1 Ahold History

On May 27 1887, 22 year-old Albert Heijn took over his father's small grocery store near Zaandam, West Holland and began offering quality products and services at the lowest prices. In 1887, Albert Heijn takes over his father's small Zaandam grocery store, selling a wide variety of products, from groceries, to dredging nets and tar. In 1897, Albert Heijn stores open in Alkmaar, The Hague and Amsterdam. Within 10 years the store count rises to 23. In 1911, The first Albert Heijn brand name products sold are cookies baked by Heijn himself in the kitchen of a local mansion house. Manufacturing company Marvelo develops from this, with a range of activities including production of tea, coffee, peanut butter, and wine bottling. In 1948, Ahold is listed on the Amsterdam stock exchange. From this moment on Ahold shows a continuous growth. In the 1955 Albert Heijn opens first self-service supermarket in Rotterdam and in 1977 Ahold enters the U.S. market for the first time, acquiring the BI-LO supermarket chain in the Carolinas and Georgia. Actually Ahold operates in the United States under the regional brand names of Stop & Shop, Giant (Carlisle), Giant (Landover), Tops, BI-LO and Bruno's In Latin America, Ahold is active in 10 countries with approximately 750 stores. In Asia, Ahold is developing the TOPS supermarket format in Thailand, Malaysia and Indonesia with almost 100 supermarkets. For what regards Europe, by the end of 2001, Ahold operated over 6,500 stores in 13 countries. Combined 2001 sales amounted to euro 21.8 billion, a rise of 31% over the previous year. Sales in Europe in 2001 accounted for 35% of Ahold's worldwide sales. In Holland, Ahold operates five store chains with sales of euro 9.8 billion and over 2,300

223

Part III Chapter 4: Ahold

outlets, including flagship supermarket company Albert Heijn. Albert is Ahold's joint internet-based home delivery service in The Netherlands. Customers can access all five Dutch subsidiaries in one order. These Ahold outlets include Albert Heijn supermarkets, Gall & Gall wine and liquor stores, Etos health and beauty care stores, De Tuinen natural product stores and Deli XLSHOP. Ahold also holds a 73% interest in Schuitema, a wholesale supplier to independent supermarket operators. Through its Dutch food service operation Deli XL, Ahold is also active in Belgium. Figure 1 shows the distribution of sales among the different countries.

Asia Latin America 2% 8%

Europe USA 35% 55%

Figure 1: worldwide sales distribution

Figure 2 shows the distribution of sales within the major countries in Europe.

Poland Czech Republic 3% 4% Portugal 7% Spain 9% The Netherlands 45%

Northern Europe 32%

Figure 2: European sales distribution

224

Part III Chapter 4: Ahold

In this work we will focus on the Ahold division located in Holland. Even if the analysis here conducted are based on information regarding this particular reality, we argue that findings, at least partially, can be generalized.

2.2 Description of the Supply Chain

Due to privacy issues, we are not allowed to go into the detail of the supply chain structure of this company, however we can draw some interesting elements. The main point of interest on this work is related to the division located in Holland. As a matter of fact each country has its own regional division that manages all local activities, while some global issues are managed by much higher units. The division located in Holland faces a rather simple supply chain, in which a central warehouse distributes to the different stores by means of distribution centers (see Figure 3).

From a logistic point of view the central warehouse uses distribution centers, however decisions regarding stores fulfillment and data are managed at the central unit level, thus form a planning point of view, the distribution centers don’t play a relevant role. Planning decisions are aggregated at the central warehouse level as in fact, the central division is provided with data regarding sales, number of customers purchasing, number of customers entering a specific store, and so on. Thus the forecasting management is entirely attributed to the national division. Central Warehouse

Distribution Centers

Stores

Figure 3: supply chain structure

This structure is essentially due to the particular market considered. Compare for example this situation with the Nestlé case in which a very different supply structure is provided. The retailer manages a network of about 400 stores and owns distribution centers that receive goods from suppliers and distribute them to single stores. In addition, the retailer manages several thousands SKUs as variety is believed to be one of the major drivers of traffic in grocery stores.

225

Part III Chapter 4: Ahold

Even if the supply chain is simple, however, demand appears to be variable. In fact, promotional activities are applied thus rising single store demand variability. Since the size of promotions is significant compared to stable demand, even if they are applied to a reduced number of customers, they tend to influence deeply total demand. Figure 4 and Figure 5 provide two examples on two different products.

Figure 4: demand series for item 529955 on weekly level

Figure 5: demand series for item 8667 on weekly level

226

Part III Chapter 4: Ahold

Since this kind of variability is particularly relevant, the forecasting system considers information regarding promotional activities. The next part describes the model adopted to forecast demand. In our study we were able to analyze demand for a set of 5 products (ranging from fast moving sodas to slow moving diapers) in a network of 38 stores for 117 weeks. The data were split into a fit sample of 111 weeks and a test sample of 6 weeks as 6 weeks is the forecasting horizon used by the retailer.

2.3 Actual Forecasting System

The forecasting technique applied is a logarithmic regression. In particular the conceptual model behind the forecasting approach is the n, p, q model that is often used in marketing literature (e.g., see Kotler, 1991) where n is the number of customers visiting a store, p is the probability that a customer buys a product and q is the quantity bought by a customer that actually purchases the product. One interesting feature of this model is that it follows the demand generation process as it first captures the number of customers visiting a store and then the choice of the customers in the store. Here the system is described.

We define: i is the store index j is the item index w is the week index t is the day of the week index, so that xw;t is variable x in week w and day t

There are 4 major steps in the forecasting process

1. forecast of number of customers niw at the week store level

2. forecast of pqijw (units sold per customer visiting the store) at week level

3. derivation of demand forecast at store item week level (dijw)

4. break down of weekly forecasts at the single day level (dijwt)

In this specific company the number of customers entering the store i in a given week t i ij N t ;the demand per week per store per item Dt are the observable variables.

227

Part III Chapter 4: Ahold

i i Aggregate forecast is derived by evaluating predictions on Nt and on PQt indicated

i i respectively as N t and PQ t .

In this work we will take into consideration only the estimation of the choice of customers, since, given our objectives, we do not need to evaluate completely demand. In fact a similar procedure to the one described is applied to estimate the number of customers visiting a store. In particular we consider the estimation of PQt, which represents the number of units of a specific SKU a generic customer buys during a specific week. This enables us to use log regression as PQ is a positive statistic but it is not bound to be smaller than one1. In these terms the model is as follows:

ln(PQt ) =α 0 +α1 ln( pt ) +α 2 At +α 3Tt +α 4 HDIt +α 5Tt−1 +α 6 HDIt−1 +α 7 NTt +

+α8 NTt−1 +α 9 NTt−2 + ∑ βτ Et;τ + ∑γ τ Et−1;τ + ∑δ r Rtr τ τ r where:

pt is the average price of item j at store i in week t;

At is a dummy variable that captures the presence of a promotion for item j in week t;

Tt is the maximum temperature in week t;

HDIt Human Discomfort Index is an index that captures weather conditions including inches of rain in week t; temperature and HDI do not refer to a specific location as stores are located in an homogeneous area;

NTt is the average maximum temperature recorded in past years during week t;

Et;τ is a dummy that captures whether during week t any particular event occurred (events include school holidays, Easter and Christmas);

Rtr is a dummy variable that captures the nature of the promotion (e.g., the mean used to advertise the presence of a promotion). There are 4 different types of promotion.

Rtr is set to 1 if promotion r is active in week t.

As it can be noted the model considers different elements. From one side are considered external factors such as temperature and weather conditions, from the other, managerial

1 In addition, models where p and q were estimated separately were tested but they did not prove to be more effective than the model proposed. 228

Part III Chapter 4: Ahold

elements as promotional activities are taken into account. Quite interestingly the structure of the forecasting system tries to model the demand generation process since variability is assumed to be influenced from one side by the particular values assumed by the different variables and from the other by the particular sensibility of the different customers towards the specific variables. In other words demand variability may be influenced for example by price variations, but also by how the different stores are sensible towards price variations. In fact different customers may react differently towards similar changes.

229

Part III Chapter 4: Ahold

3 Solution Development

The basic structure of the forecasting model adopted suggests that the buying behavior of the customers is influenced by the price of the good being purchased, by weather (temperature during week t and t-1), events (e.g., Christmas, Easter, summer holidays, etc.) and seasonality, measured by the average temperature in the period. Clearly the model can be applied at different levels of aggregation. First of all aggregation can be made directly on data. Indeed, to estimate the parameters of the equation we can both use detailed data for each single store or aggregate data at chain level. E.g. one can analyze the relationship between the penetration rate at a given store with the price at that store (detailed data), or can try to analyze the relationship between the average penetration for the whole chain of 38 stores with the average price in the chain of 38 stores. In addition, the parameters of the models can be evaluated at different aggregation level. Indeed parameters can both refer to single stores or to the whole set of stores. In particular, if parameters are evaluated at store level we are assuming that each store has its own characteristics and in particular that customers buying at a particular store react differently to changes in causal variables. On the contrary, if parameters are estimated at chain level, we are assuming that roughly all stores react the same way. Obviously, in case aggregate data are used, only aggregate parameters can be derived, while with disaggregate data both aggregate and disaggregate parameters estimation are feasible. The combination of the above degrees of aggregation led us to the definition of 3 basic models, as summarized in Table 1 and here described.

Aggregate Mixed Disaggregate Parameters at chain level at chain level at store level Data at chain level at store level at store level Table 1: the 3 basic models considered

Actually, the forecasting approach adopted by the firm is the aggregate one; so the company evaluates both parameters and data at chain level.

3.1 Disaggregate Model

A first option is to apply the above model at a very detailed level where the parameters are estimated separately for each single SKU-store combination. In other words the detailed model assumes that different stores have a different sensitivity to price changes,

230

Part III Chapter 4: Ahold

promotions, weather etc.. Obviously, the data used to fit the model are at store level in order to capture differences among stores. Clearly this model generates forecasts at store level, while the chain level forecast is obtained by adding disaggregated forecasts. In these terms the model is defined as follows:

ij ij ij ij ij j ij ij ij ij ij ln(PQt ) = α 0 +α1 ln( pt ) +α 2 At +α 3 Tt +α 4 HDIt +α 5 Tt−1 +α 6 HDIt−1 +α 7 NTt +

ij ij ij ij ij j α 8 NTt−1 +α 9 NTt−2 + ∑ βτ Et;τ + ∑γ τ Et−1;τ + ∑δ r Rtr τ τ r where: ik pt is the average price of item j at store i in week t;

j At is a dummy variable that captures the presence of a promotion for item j in week t;

j Rtr is a dummy variable that captures the nature of the promotion (e.g., the mean used

j to advertise the presence of a promotion). There are 4 different types of promotion. Rtr is set to 1 if promotion r is active in week t for item j.

As it can be noted, temperature and HDI are assumed as similar to the different stores, as the region considered is rather homogeneous, while price is assumed to be store specific. This option has obviously the advantage of capturing the differences among stores. On the contrary, this option has the disadvantage of a rather small and noisy sample, as the number of parameters to estimate is rather large. In fact, when for example estimating the impact of a promotion of type r for item j in the store i, we don’t have many points to interpolate, since this specificity is rather high. Thus the main problem is that while variability can be very well identified, it tends to be more difficult to estimate its impact.

3.2 Mixed Model

A second option is to assume that all stores share the same sensitivity to causal factors and thus the parameters of the regression equation are common to all stores. In this model, we introduce a store-dummy variable that allows to capture the differences in average penetration of the SKU in various products. Thus this models assumes that store differ in terms of average penetration rate but react similarly to change in price, weather and other causal factors; this system too uses data at store level to estimate parameters. The model captures only a portion of the differences among stores (namely average penetration) but

231

Part III Chapter 4: Ahold

has the advantage of enjoying more reliable estimates as one large sample is used instead of 38 smaller samples. In fact we define:

ij ij Dt PQt = i is the penetration rate for product j in store i in week t; N t

ij ∑ Dt j i PQt = i is the average penetration rate for the chain of stores. ∑ N t i

In this case too forecasts are generated at store level and the chain prediction is the simple sum of demand of store-level forecasts. So the model formulation is as follows:

ij ik j j ij j j j j j j j ln(PQt ) = ∑ S α 0 +α1 ln( pt ) +α 2 At +α 3 Tt +α 4 HDIt +α 5 Tt−1 +α 6 HDIt−1 +α 7 NTt + k j j j j j j α 8 NTt−1 +α 9 NTt−2 + ∑ βτ Et;τ + ∑γ τ Et−1;τ + ∑δ r Rtr τ τ r where: S ik is a store dummy set to 1 if i=k, and zero otherwise.

For both the Disaggregate and Mixed model we can evaluate total chain forecast by simply adding up all demand forecasts for each store as follows: ij ij ij Dt = N t ⋅ PQt

So for each item, total chain forecast is evaluated as: j ij D t = ∑ D t i

Since the detailed model considers data and parameters at a store level, while the mixed one considers parameters at a chain level, at last the aggregate model consider both data and parameters at store level.

232

Part III Chapter 4: Ahold

3.3 Aggregate Model

A third option is to aggregate data and then estimate the effect of causal variables at chain level. The assumption behind this model is that stores react similarly to causal factors such as Christmas or Easter. This model benefits from more stable inputs, as data at chain level are far less noisy than data at store level. The model clearly generates only a forecast at chain level; thus a process to derive store-level forecasts was to be designed. We adopted a rather trivial rule: the store level forecast is equal to the chain level forecast times a factor that captures the percentage of total demand that occurs in a given store. In formulas:

j j j j j j j j j j j ln(PQt ) = α 0 +α1 ln( pt ) +α 2 At +α 3 Tt +α 4 HDIt +α 5 Tt−1 +α 6 HDIt−1 +α 7 NTt +

j j j j j j α 8 NTt−1 +α 9 NTt−2 + ∑ βτ Et;τ + ∑γ τ Et−1;τ + ∑δ r Rtr τ τ r

This leads to the generation of a chain level demand forecast. The chain level demand forecast is broken down at single store level according to the following procedure. For each store, forecast is disaggregated as follows:

ij j j ij j ij D t = D t ⋅[At ⋅ SF1 + (1− At )⋅ SF0 ] where:

    D ij 1− A j  t  ∑()t   t ij  ∑ Dt  SF0ij =  i  j ∑()1− At t     D ij A j  t  ∑ t   t ij  ∑ Dt  SF1ij =  i  j ∑ At t

This solution, compared to the disaggregate one should be capable of estimating better regression parameters, since enough data are provided, however it misses the relevant differences among stores. 233

Part III Chapter 4: Ahold

So the main problem is due to the fact that from one side the disaggregate model is capable of identifying variability, but is not able to estimate properly its impact on demand, from the other the aggregate model reduce problems in the estimation process, but is not capable of identifying properly variability.

While trying to solve the trade-off between these two elements, we identified a fourth solution that tries to gain both objectives together.

3.4 Cluster Model

As discussed in the introduction, in addition to classical solutions we introduce a model based on clustering. In fact aggregation and disaggregation have been till now conducted according to a “logistical” perspective: aggregation may be at a store and chain level. This tends to be a common practice within firms, as also stated by some works in the current literature. Usually aggregation occurs over the supply chain structure and thus stores served by a given distribution centre are grouped together. We argue that clustering stores according to their heterogeneity can lead to better grouping of stores for forecasting and inventory management purposes. Consider for example two different stores, one located in a big city and another located in a small town frequently visited by tourists during summer vacations. They probably show very different demand patterns since seasonality is structurally different in the two situations, so it appears intuitive that forecasts for these stores should be estimated separately. Similarly, customers may react differently towards promotional activities (as also the Nestlé case shows) or to particular kinds of promotions.

As already discussed one extreme solution is to analyze stores one by one, though this solution leads to significant errors of estimate of parameters that might lead to inaccurate forecasts. However, if we cluster stores that show a similar demand pattern (e.g. seasonality or reaction to promotions) we can capture differences while enjoying a relatively large sample. This approach makes a relevant hypothesis: that heterogeneity may be estimated according to each store demand pattern. However, the analyses provided in chapter 3 should claim that this assumption should hold. This was achieved through the ij variable PQ%t that captures how the penetration rate fluctuates over time. This variable is defined as follows:

234

Part III Chapter 4: Ahold

ij ij PQt PQ%t = ij ∑ PQt t

This metric captures the fluctuation in penetration rate over time at a given store and takes the average penetration out of the picture. Thus this variable captures whether causal factor impact on penetration rate of a given SKU at different stores similarly.

Similarly to what has been done in the Nestlé case, we adopted a k-means algorithm for each SKU. In fact, for each SKU, 3 clusters were defined through the k-means algorithm. In this specific case 3 clusters were chosen to obtain clusters that on the one hand are large enough not to resemble single stores and small enough not to resemble the whole chain. The aggregate approach was used to predict demand for each cluster. In this case the chain level forecast is simply equal to the sum of forecasts for the 3 clusters. The store level forecast is disaggregated from cluster forecast like in the case of aggregate model.

Figure 6, Figure 7 and Figure 8 provide an example of the clusters obtained for a specific product by means of the described system.

Figure 6: demand series for the stores belonging to cluster 1

235

Part III Chapter 4: Ahold

Figure 7: demand series for the stores belonging to cluster 2

Figure 8: demand series for the stores belonging to cluster 3

As it can be noted, the first cluster considers stores that are deeply influenced by seasonality and promotional activities, while the second cluster is characterized by stores that have an almost stable demand except for promotional activities that seem to be highly correlated. In the end the third cluster shows stores that have a negative trend pattern still influenced by different promotional activities.

In these terms the model is defined as follows:

236

Part III Chapter 4: Ahold

cj cj cj cj cj j cj cj cj cj ln(PQt ) = α0 + α1 ln(pt )+ α2 At + α3 Tt + α4 HDIt + α5 Tt−1 + α6 HDIt−1 +

cj cj cj cj cj cj j + α7 NTt + α8 NTt−1 + α9 NTt−2 + ∑ βτ Et;τ + ∑γτ Et−1;τ + ∑δr Rtr τ τ r where: ck pt is the average price of item j at cluster c in week t;

In the remaining part of this work we will now on consider and compare the four levels of aggregation we introduced and here summarized in Table 2.

Aggregate Mixed Disaggregate Cluster Parameters at chain level at chain level at store level at cluster level Data at chain level at store level at store level at cluster level Table 2: Description of the different approaches considered

237

Part III Chapter 4: Ahold

4 Analysis of results

In our analyses we investigated the impact of heterogeneity by means of the aggregation of demand and information over locations as the number of SKUs in our sample was rather limited and both the products and the demand patterns were rather different. According to the objectives of the work we will evaluate the forecast accuracy of the models discussed both at single store level and for the chain of 38 stores at week-item-level. As to the performance metric, we adopted MAPE (Mean Absolute Percentage Error) for the chain level forecast. This metric was chosen because of its wide adoption and because it enables to compare performance of a given model across products with significantly different demand rates. For each product, MAPE is defined as follows:

1 1 AE MAPE = ∑∑ i,t m itn Di,t where:

AEi,t represents the absolute error for week t and store i;

Di,t is total demand for week t and store i.

However, this metric was not directly applicable at store level as for some stores slow moving items have some weeks with zero demand. This can lead to relevant biases when for example in a period when demand is zero this metric is indifferent towards a forecast very close to zero (i.e. 0,1) compared to a second forecast completely biased (i.e. 100). Thus a modified MAPE, indicated as MAPE*, was introduced to avoid the problem of zero demand periods. This metric simply evaluates the MAD and compares it to total demand, so letting comparison among different forecasts possible and reducing the bias the MAPE shows in this particular situation. So we evaluated MAPE* as follows:

∑∑ AEi,t MAPE*= it ∑∑Di,t it

Literature does not consider this as a relevant metric for evaluating performances; we refer to Armstrong (2001) and Vollmann et al. (1992) for a general review of forecast error metrics. However, we argue that in some cases it can be a significant measure.

238

Part III Chapter 4: Ahold

In particular MAPE was adopted at chain level while a modified MAPE (MAPE*) was considered for store level. Table 3 summarizes results for chain level, Table 4 shows the difference in performances of the different approaches compared to the others, while Table 5 shows the 2-tailed significance values of the Paired-Samples t test for the difference of means.

ITEM MIXED AGGREGATE DISAGGR. CLUSTER A 25,97% 8,14% 21,08% 11,42% B 15,82% 7,55% 15,73% 3,15% C 16,47% 11,54% 8,90% 8,06% D 21,64% 11,13% 13,66% 10,41% E 21,05% 13,88% 21,71% 15,16% Average 20,19% 10,45% 16,22% 9,64% Table 3: MAPE for the different approaches at chain level

MIXED AGGREGATE DISAGGR. CLUSTER MIXED - 9,74% 3,97% 10,55% AGGREGATE - - -5,77% 0,81% DISAGGR. - - - 6,58% CLUSTER - - - - Table 4: Differences of performances among the different approaches at chain level

MIXED AGGREGATE DISAGGR. CLUSTER MIXED - 0,011 0,095 0,0024 AGGREGATE - - 0,097 0,6031 DISAGGR. - - - 0,0359 CLUSTER - - - - Table 5: Significance levels for the Paired-Sample t tests on mean difference for the considered approaches

First of all we can observe that, as it might have been expected, the Mixed model tends to perform worst. This result in not really surprising as oddly the mixed model fails to capture differences among stores but on the other hand uses rather noisy store-level data. Interestingly, the Aggregate model performs better than the Disaggregate one (almost 6 percentage points less, equal to more than 30% better). One can argue that this is because the accuracy is measured at chain level. However, this hypothesis can be rejected in the light of results at store level. Indeed, at store level the disaggregate model does not perform better than the aggregate one.

A second hypothesis is that the aggregate model performs better because, on the one hand, the stores in the sample are relatively similar as they belong to the same retail chain where

239

Part III Chapter 4: Ahold

prices are relatively constant across stores, displays are similar etc.; thus in this context the advantage of capturing store specific causal factor does not pay off the disadvantage due to the reduction in the information available to estimate the parameters. In other words, given the relatively high homogeneity of stores and the relative lack of information at store level, the advantage of aggregation in terms of greater accuracy of parameter estimates is greater than the disadvantage due to the inability to capture differences among stores. Finally, on average the Cluster solution performs slightly better than the aggregate approach (on the average, it improves the Aggregate model performances by almost 10%); However a Paired-Sample test showed that the difference is not statistically significant (see table 4 for results). It is however interesting to investigate on which products the aggregate model performs better than the cluster one. As Figure 9 shows the cluster approach outperforms the aggregate model for those items that have a rather high demand per store (average weekly demand per store of items B,C and D are 927, 90 and 126 units). On the contrary for slow moving items (average weekly demand per store of items A and E is 3 and 29 units) the aggregate model outperforms the cluster one.

0 012345678 -1

-2 MAPE(aggr)-MAPE(cluster) -3

-4 Ln(avg.weekly demand per store)

Figure 9: Correlation between cluster approach improvements and average weekly demand per store for the 5 products considered

RSquare 0,832202 Term Estimate Std Error t Ratio Prob>|t| RSquare Adj 0,776269 Intercept -4,929016 1,634665 -3,02 0,0570 Ln(D) 1,3901882 0,360406 3,86 0,0308 Table 6: linear regression between MAPE (aggr.)- MAPE (cluster) and average weekly demand

240

Part III Chapter 4: Ahold

This result sheds lights on the conditions under which clustering might work. First, clustering can only work if there is enough past data to group together stores with similar patterns. In case of very slow moving items (like items A and E) it is really hard to capture similarities among stores, as demand at store level is extremely nervous as the decision of a single customer makes a difference. Thus the clustering technique can effectively group similar stores together only for rather fast moving items.

Second, the clustering technique involves some kind of disaggregation as compared to the aggregate model. Though the cluster solution can capture some kind of difference among stores, the ability to accurately estimate the impact of causal factors is reduced by the lower number of stores per cluster. In other words, the amount of information available is not large enough to estimate accurately the slight differences among the stores that belong to the different clusters.

In fact if we consider the different products separately we see that the trade-off between variability identification and variability management lies at different levels. Figure 10 and Figure 11 show for two different products the performances at different levels of aggregation, thus with different attention to stores heterogeneity.

Product B 20,0%

15,0%

10,0% MAPE

5,0%

0,0% Disaggregate Cluster Aggregate Aggregation Level

Figure 10: performances for product B at different aggregation levels

241

Part III Chapter 4: Ahold

Product A

25,0%

20,0%

15,0%

MAPE 10,0%

5,0%

0,0% Disaggregate Cluster Aggregate Aggregation Level

Figure 11: performances for product A at different aggregation levels

These figures show that for product B, the cluster approach is preferred since it reaches the minimum MAPE compared to the other aggregation levels. On the contrary, for product A, the aggregate approach performs better, since the lack of sufficient data states that a too heavy customer-oriented approach is not recommendable.

As previously anticipated, to properly compare the different approaches we had to evaluate performance also at store level. This is also because for the considered firm it is relevant both the aggregate forecast, since this influences the replenishment process of the items, and transportation to stores. This second analysis was conducted by means of MAPE*. Table 7 summarizes results at store level for the different approaches adopted, Table 8 shows the difference in performances of the different approaches compared to the others, while Table 9 shows the 2-tailed significance values of the Paired-Samples t test for the difference of means.

ITEM MIXED AGGREGATE DISAGGR. CLUSTER A 37,40% 32,60% 36,60% 35,06% B 17,72% 16,50% 20,26% 16,07% C 24,94% 26,55% 19,07% 20,94% D 26,31% 23,90% 22,48% 22,50% E 27,71% 35,60% 43,95% 36,82% Average 26,82% 27,03% 28,47% 26,28% Table 7: MAPE* for the different approaches at store level

242

Part III Chapter 4: Ahold

MIXED AGGREGATE DISAGGR. CLUSTER MIXED - -0,21% -1,66% 0,54% AGGREGATE - - -1,44% 0,75% DISAGGR. - - - 2,19% CLUSTER - - - -

Table 8: Differences of performances among the different approaches at store level

MIXED AGGREGATE DISAGGR. CLUSTER MIXED - 0,926 0,694 0,837 AGGREGATE - - 0,623 0,616 DISAGGR. - - - 0,238 CLUSTER - - - -

Table 9: Significance levels for the Paired-Sample t tests on mean difference for the considered approaches

Again, on average, the cluster approach tends to perform better than the other techniques considered, however such a difference with the aggregate and the disaggregate model is not statistically significant. Table 9, in fact, shows that the Paired Sample test does not reject the hypothesis that the performances are similar, so we cannot conclude that globally one method performs better than another. However it is interesting to notice that the detailed approach (i.e. the one that is consistent with the level at which performance are measured in this case) does not outperform other methods. Thus while the cluster and aggregate solution outperform the disaggregate one in terms of chain level accuracy, the disaggregate model does not outperform the aggregate and the mixed ones. In this respect we would argue that under the specific condition of this company the aggregate and cluster models perform better. In the conclusions we will discuss the generalization of this result. In addition, we considered again the correlation between performances and volumes of the different SKUs; this analysis, again, reported similar results as previously: the cluster approach performs better for relevant SKUs while it does not for slow moving items.

243

Part III Chapter 4: Ahold

5 Conclusions

This case appears to be interesting for rather different reasons. First of all this work contributes in stating that the choice of the appropriate level of aggregation for forecasting is a relevant matter. In fact, it can be noted by the presented results that performances change significantly according to the particular aggregation level adopted. We argue that this issue is in general not much analyzed, in particular by practitioners, while we believe it should be considered in much more detail. Second we have provided evidence, even if in a specific context, that there is no best way in defining the proper aggregation level. Even if the solutions analyzed perform differently, we argue that this is not specific to the particular situation considered. Some authors in the past literature have provided contributions in this direction (see paragraph 1 for a review of these works); in this work we have identified some factors that should shift choice in different directions, in particular according to the specific product characteristics (volumes, fast mover, etc.) and the level of analysis considered. Moreover, coherently with this evidence we have provided some guidelines by means of which the proper level of aggregation should be chosen. In particular this choice depends on the information available and how much heterogeneous market is. As described in Figure 12, given a certain aggregation level at which the process is typically conducted and a certain aggregation of data (so the “base case” for the problem considered), the process of forecasting can be conducted by changing these two levels of aggregation.

Process Aggregation Aggregation Level Aggregate Mixed Aggregate

Base Case Base Case Solution

Disaggregate Not applicable Disaggregate

Disaggregate Base Case Aggregate Data Aggregation

Figure 12: The aggregation/disaggregation approaches considered

244

Part III Chapter 4: Ahold

This leads to the first three solutions analyzed in this case and that in fact appear in the previous figure. However, as previously described, at both chain and store level the Mixed approach does not gain accurate performances. We argue that this kind of aggregation process is non-proper as it introduces bias in the estimation given the reduced coherence with the demand generation process. So the main issue is choosing at which level of aggregation evaluate both data and process. The most important fact is that aggregation or disaggregation should be conducted to solve in the most efficient way the trade off between information quantity and detail on the demand generation process. However typically aggregation and disaggregation are developed according to dimensions that are not coherent with the demand generation process (i.e., distribution centers). Therefore, we propose an alternative method, based on a clustering methodology that tries to shift this trade-off, by introducing new aggregation dimensions.

In particular, we argue that a relevant factor that should be considered to evaluate the proper level of aggregation is heterogeneity consistency of the aggregation level and of the demand generation process. In the present work we adopted a particular solution to improve forecast performances. Our solution provides good results since it is more consistent with a relevant source of variability that is stores heterogeneity. As previously introduced all stores face different problem and so different demands; in this way each store is different from the other, in particular for what regards seasonality. As Figure 13 shows, managing each single store separately is very consistent with their specific characteristics and so it is very consistent with their heterogeneity, however it is also very difficult to extrapolate relevant information from many small elements, so the forecasting system is partially blinded by this great complexity. From the other side the aggregate approach tends to be not consistent with demand heterogeneity, and again performances may not be efficient. In these terms aggregation should not be conducted according to dimensions that are not tied to heterogeneity but it should be evaluated according to the root causes of heterogeneity in the demand generation process. This, as the Cluster approach here proposed, focuses on the heterogeneity of the demand generation process and provides a possible alternative solution for the considered problem. What is relevant, however, is not the solution itself, since the dimensions considered are somehow specific to the particular context analyzed, but the methodology adopted that, we argue, can be generalized.

245

Part III Chapter 4: Ahold

Heterogeneity Consistency

High Total Disaggregation Clusters

Low Total Base Case Aggregation Solution

Disaggregate Aggregate Aggregation Level

Figure 13: Aggregation level and Heterogeneity consistency for the different methods considered

In fact this case identifies that focusing on customer heterogeneity may be tricky since even if variability is caught it cannot be managed. This case claims in fact that customer- orientation has to be mitigated by information availability, by means of intermediate solutions. We propose a cluster-based technique that may in some conditions perform well. Moreover, we show that the degree of customer-orientation that should be adopted is influenced by different elements, so companies adopting this approach should conduct proper analysis to identify the level of heterogeneity at which manage demand. This case shows that even if supply chain is rather simple, focusing on heterogeneity may be a significant issue. In fact, when inner heterogeneity is significant, different forecasting processes have to be applied, while the forecasting structure may be similar for the different customers (see Figure 14).

Inner Different estimation processes Heterogeneity Heterogeneity

Supply Chain Complexity

Figure 14: the overall forecasting method and its relationship with heterogeneity

246

Part III Chapter 4: Ahold

Moreover, this case claims that when choosing at which level forecast has to be conducted, attention should be paid to heterogeneity and its dimension. The clustering approach reveals to be a significant and efficient mean to solve the trade off between ability to catch variability and to manage variability. This result is consistent with what the Nestlé case shows and with the general framework provided in Part II.

In the next chapter overall conclusions are drawn and future development are provided.

247

Part III Chapter 4: Ahold

248

Chapter 5 Conclusions and Future Developments

1 Analysis of the Action Research results

This research focuses its attention on the impact of customer’s heterogeneity on demand forecasting. In particular two main research questions have been addressed in this work, as presented in Part II:

How does heterogeneity in customers’ purchasing processes influence demand variability and forecasting effectiveness?

How should demand be managed when dealing with heterogeneous contexts?

What are the main elements that should be taken into account when dealing with heterogeneous contexts?

249

Part III Chapter 5: Conclusions and Future Developments

In particular we previously defined the following research hypothesis:

H1. Customers’ heterogeneity affects demand variability and forecasting accuracy.

H2. Supply chain complexity deeply influences demand heterogeneity and thus demand forecasting approaches.

H3. When heterogeneity becomes relevant, a heterogeneous forecast has to be adopted.

H4. If heterogeneity is due to the supply chain complexity, then different forecasting approaches have to be adopted.

H5. If heterogeneity is due to customers’ inner characteristics, different estimation processes have to be adopted.

These issues are here summarized.

1.1 Demand management approaches within heterogeneous contexts

In this work we considered three cases that face relevant problems in dealing with demand forecasting due to heterogeneous customers thus leading to heterogeneity in the demand generation process. As previously indicated, heterogeneity may apply on different elements, thus we considered only some of the possible sources of demand heterogeneity, however we will provide comments regarding to which extent results can be generalized. In particular, we identified that heterogeneity may be due to two main factors, from one side it depends on their inner characteristics of the customers’ purchasing behavior and, from the other, it depends on the supply chain complexity faced by the considered company. As discussed in part 2, when these factors become relevant, thus making demand more heterogeneous, forecasting has to be heterogeneous too; however, according to the different characteristics demand shows, this may be applied in two rather different ways. If supply chain complexity is relevant, then different forecasting approaches have to be defined and adopted; when inner heterogeneity is significant, then different forecasting estimations have to be applied.

Figure 1 describes the general methodology.

250

Part III Chapter 5: Conclusions and Future Developments

1 forecasting structure

Inner Different estimation Different forecasting structures Heterogeneity Heterogeneity processes

Supply Chain Complexity

Figure 1: the forecasting methodology adopted

In the development of the different solutions for the considered cases we proved that this kind of approach may be effective. In fact, in the Whirlpool case, the supply chain is rather complex due to the direct management of both retailers and wholesalers. This, due to the rationalization of the supply chain, is the major cause of forecasting inaccuracy and poor inventory performances. In this situation demand is composite and refers to different customers with significantly different purchasing processes. and thus different forecasting solutions have to be adopted, focused on the main characteristics of the different groups of customers.

In the Nestlé case, the supply chain is complex as different customers are managed by means of different distributional channels, thus again different forecasting solutions have to be applied. Moreover, since customers tend to deeply influence demand also according to their inner characteristics, different estimation processes have to be applied by means of clustering techniques and promotional periods profiling. In the end, Ahold shows a rather simple supply chain since considered stores belong to the same supply structure; however inner differences between customers seem to be significant thus claiming that different forecasting estimations have to be conducted, as the case describes by means of clustering application. As a matter of fact, in the Whirlpool case, even if customers may add some heterogeneity to the demand generation process, the provided results show that leveraging on the contribution of the supply chain complexity provides significant benefits. Moreover, we argue that in the considered situation a clustering approach applied to the stable demand would have not provided significant benefits since, first of all, the stable series seems to be sufficiently even thus making its forecasts not very problematic and, moreover, clusters’ demand would have been too sporadic given the particular demand pattern of each single

251

Part III Chapter 5: Conclusions and Future Developments

customer. This claims that the problem can be solved at different levels of detail, thus it is relevant to understand what makes sense to be introduced in the forecasting solution.

Figure 2 describes the different cases considered and the methodology adopted.

Nestlé Italiana Inner Ahold Heterogeneity Heterogeneity Whirlpool Europe

Supply Chain Complexity Different forecasting approaches

One forecasting approach, different estimation processes Figure 2: the methodology adopted in the different cases

In all the cases considered, the problem is defining the proper level of detail at which analysis has to be conducted. In fact, from one side demand cannot be managed at the aggregate level since heterogeneity and variability cannot be identified when demand is considered as a whole. However, from the other side demand cannot be managed at the detailed level since this solution may be too costly and moreover the estimation process is not effective since not enough data is provided. In these terms the main problem is to define properly the appropriate level of detail at which demand should be managed. However, this is not sufficient, since when the proper level of detail is defined, customers appear to behave differently and thus they have to be separated in homogeneous groups, so to identify proper forecasting systems to manage their demand.

In other terms the overall methodology is based on a three step process. First of all the level at which demand has to be managed must be identified, then customers should be separated according to their heterogeneity, at last forecasting systems should be designed. Figure 3 describes the overall methodology further described.

252

Part III Chapter 5: Conclusions and Future Developments

At which level Identification of manage relevant heterogeneity demand factors

Identification of Separate homogeneous customers groups of customers

Define a Identification of proper set of proper forecasting forecasting methods for each solutions group

Figure 3: the overall methodology

A) Understand at which level demand should be managed

A first main issue is to understand at which level demand should be managed. In other words the first step is to identify where the main sources of variability rely, thus the dimensions on which attention has to be paid. As a matter of fact, the three cases described start from this point since attention is at first pointed to the proper comprehension of demand variability. In Whirlpool, for example, the analysis of demand made possible to understand that variability was due to heterogeneity in the demand sizes of customers’ orders, thus claiming that the forecasting solution had to focus on this aspect. In Nestlé, the same kind of analysis made possible to identify that promotions and customers’ reactions towards these phenomena were the main source of variability. At last Ahold shows that customers’ characteristics were the significant dimension that caused heterogeneity in demand and thus variability. The main output of this step is the identification of the relevant variables that should be managed.

In all the cases considered, this analysis showed that systematic factors influence demand variability for a rather reduced part, thus claiming that customers influence demand variability significantly. The analysis of the possible dimensions along which heterogeneity may show up proved to be fundamental to identify the level at which demand should be managed.

253

Part III Chapter 5: Conclusions and Future Developments

As a matter of fact this becomes a relevant issue since, as the cases provided show, the solution is not to decide if demand has to be managed at either an aggregate or disaggregate level. The main issue is that the solution lies in between where heterogeneity is better manageable in terms of managerial costs, estimation accuracy and forecast error.

B) Define how to separate customers according to their heterogeneity

The second relevant step is to define properly how to deal with the heterogeneity factors previously identified. As variability relies in the single customer’s demand generation process and since each single customer cannot be managed, demand has to be separated coherently with heterogeneity and variability. In Whirlpool, for example, since variability is due to heterogeneity in the customers’ orders size, the main problem is to separate the orders belonging to the huge customers from those referred to the smallest ones. This problem has been solved by means of the filtering system previously described. In Nestlé, since heterogeneity resides in the promotional activities and in their reaction towards external factors, customers have been separated by means of their managerial behavior (in this case the fact that a customer applies promotions), their size (the 30 monitored customers are responsible for almost 70% of sales) and according to the demand pattern they show, so tied to how customers react towards external factors. In Ahold, the solution proposed is based on a clustering technique that in fact estimates heterogeneity by means of the demand pattern customers show. Thus, from point A), the relevant dimensions along which heterogeneity relies should be identified, while in this step they have to be separated. However, step A) and B) are very tight in the sense that problems in separating heterogeneity factors may influence the level at which demand should be managed. Thus, coherently with the action research methodology adopted in this work, these two steps should be conducted cyclically. This claims that deep analysis of the specific context considered should be applied to identify at which level the most effective solution resides.

C) Design a proper forecasting solution according to their heterogeneity

When heterogeneity dimensions are identified and customers showing these patterns are defined, then a proper forecasting solution may be adopted. The solutions proposed focus on the heterogeneity dimensions and thus leverage on these elements. As a matter

254

Part III Chapter 5: Conclusions and Future Developments

of fact the different solutions developed adopt heterogeneous forecasting systems, each for a different segment of customers. In Whirlpool, two forecasting systems have been adopted, one for each group of customers that leverage on the regularities that peak orders and stable orders show when separated. In Nestlé at first two main approaches have been designed, one for the promotional customers and one for the remaining ones. Interesting improvements have been identified when attention has been shifted on this last group of customers, by focusing on their different characteristics. So in the end, the solution proposed is based on three different forecasting methods. At last, in Ahold, the solution relies in a single forecasting technique but with different estimation processes based on clustering customers according to their reactions towards external factors.

Again, also this step may be tied to step A) if its complexity does not appear to be justified in terms of trade-off between forecasting accuracy and cost.

The overall methodology tries to reduce the efforts in the estimation processes, while making it more accurate, as the results provided in the different cases show. In fact, in Whirlpool, the solution proposed (alternative c in the case), based on the same retrieval of information from the market as the actual solution the company adopts, improves inventory performances significantly by more than 2/3 in terms of inventory investments without affecting service performances. In Nestlé the overall solution is capable of improving the forecasting accuracy by more than 35% on average. In Ahold, performances of the forecasting process are improved by more than 10%, compared to the Aggregate solution presented.

However, even if performances are improved, more important is the proper identification of factors that influence the applicability of the proposed systems. In the next paragraph these elements are summarized thus defining the main elements that contribute in the applicability of the solution described and so aiming at defining the generalization of the proposed approach.

Moreover this work contributes in the development of the overall methodology proposed by contributing in the differing steps. In particular, different sources of heterogeneity have been identified (see Chapter 3) thus providing a possible line of analysis to identify which dimensions affect variability most. Then different statistical techniques have been provided to properly separate customers within uniform groups (the demand filtering system provided in the Whirlpool case and the application of clustering techniques applied in the Nestlé and Ahold cases). At last, new forecasting models have been provided and their

255

Part III Chapter 5: Conclusions and Future Developments

applicability has been discussed to be general in contexts facing similar problems to the ones considered. Figure 4 summarizes these results.

At which level Customers’ inner characteristics manage Decisional factors demand Environmental conditions

Separate Filtering technique customers Clustering technique

Define a proper set of Alternative c) in Whirlpool case Nestlé’s promotional customers forecasting forecast solutions

Figure 4: the contributions of the methodology developed

1.2 Main elements to be considered

The applicability of the overall methodology derives first from the applicability of the methodologies developed in the different cases. As a matter of fact, in the different cases we showed that the proposed approaches perform better in the average, however we also showed that, for some SKUs, they don’t provide significant improvements. However, more importantly, we identified conditions under which the proposed solutions apply, thus making possible to identify to what extent they can be applied in other similar situations. In particular, the clustering technique has been shown to be an effective solution; however its application should be limited first of all to those situations where enough information is provided. As the Ahold case shows this system is deeply influenced by the size of demand and thus by how many variables can be adopted to create proper clusters. Moreover, this system should be applied only when demand is variable since, when variability is reduced, this system doesn’t outperform the single cluster one.

The analysis developed in the Nestlé and Ahold cases provide that this kind of approach should be applied only when:

256

Part III Chapter 5: Conclusions and Future Developments

1) demand variability of the residual demand on which clusters should be adopted is relevant (see Figure 5);

5,0%

4,0%

3,0%

2,0%

1,0%

0,0% 0% 5% 10% 15% 20% 25% 30% 35% 40% -1,0%

-2,0% Percentage improvement -3,0%

-4,0% Demand variability

Figure 5: relationship between improvement of the clustering and demand variability (see Nestlé case for details)

2) demand is not too much dispersed thus making enough data available to properly estimate significant clusters (see Figure 6).

0 012345678 -1

-2 MAPE(aggr)-MAPE(cluster) -3

-4 Ln(avg.weekly demand per store)

Figure 6: relationship between improvement of the clustering approach and average demand (see Ahold case for details)

257

Part III Chapter 5: Conclusions and Future Developments

From another side the filtering system has been effectively applied and is capable of identifying properly the most variable orders; however two main issues have to be considered. First of all, its applicability is tied to the existence of a strong bi-modality, since if demand is less dispersed this system tends to identify erroneously demand peaks (see Figure 7).

3,5

2,5

1,5

0,5 Mean Stock Ratio (Alternative C/Alternative A) (Alternative C/Alternative Stock Ratio Mean 0 0 3 6 9 12 15 Demand Asymmetry

Figure 7: relationship between mean stock ratio and demand asymmetry

Moreover, an interesting aspect is the number of times that standard deviation should be considered in the filter definition as “normal” variation of the stable series. We developed different efforts in identifying the proper values of this parameter, however, from our results this element doesn’t affect significantly system performances, thus claiming that efforts should not be paid too much towards this issue.

Another relevant factor is tied to the information retrieval process. In fact, when customers deeply influence demand variability, coherently with the actual contributions reported in current literature, attention has to be paid on retrieval of information regarding the demand generation process. For example, in the Whirlpool case we showed that significant benefits can be gained by leveraging on information regarding the demand generation process of the most relevant customers. In fact alternative c) leverages on the knowledge regarding customers’ reorder process so feeding the forecasting system with the information

258

Part III Chapter 5: Conclusions and Future Developments

regarding when a specific customer has purchased. Moreover, alternative d) shows that if early information regarding this process is provided, significant benefits can be gained (see Chapter 2 for details). This claims that organizational efforts to develop collaborative relationships with wholesalers may pay off in terms of forecasting accuracy and operational performances improvements.

In the Nestlé case, we showed that information regarding past and future promotional activities can significantly impact on forecasting performances, due to the contribution of the different customers to total demand variability. In this situation becomes critical the inter-functional integration between forecasting people and the units that develop promotional activities with end customers. As a matter of fact, a relevant aspect of the implementation of this solution within the considered company was defining proper organizational procedures and developing a software system to gain this information, so to guarantee that information is provided as soon as it is available.

Thus attention should be paid also in the exploitation of the proper sources of information regarding the different dimensions of heterogeneity and so on developing proper organizational processes to gain and exploit this relevant information.

259

Part III Chapter 5: Conclusions and Future Developments

2 Generalization of findings

All the solutions provided have been developed within specific contexts; however we argue that results can be at some rate generalized. First of all the application of the clustering technique, applied both in the Nestlé and Ahold case, has proved to be effective in dealing with highly variable demand. We argue that this kind of technique can be adopted in different contexts since its applicability is not based on the specific causes of demand variability. Moreover, the results obtained by its application within these two different contexts provide evidence of its general validity. We also studied under which conditions this kind of solution may be applied, stating that it tends to be applicable when demand variability is mainly not completely random and when enough information is provided. This condition may apply to many different industrial contexts where systematic and managerial variability shows up, as seasonality, promotional activities and so on. This claims that this specific methodology may be generalized. Moreover, we argue that the applicability of this system is also tied to the number of SKUs managed. In fact, if many SKUs are managed, probably demand is rather distributed among these products thus reducing the quantity of information available. The analysis conducted within the Whirlpool case proved that this kind of approach is not applicable. Moreover, the higher the number of SKUs the more complex the system will be.

Similarly the filtering procedure provided doesn’t seem to be specific to the context considered, but rather specific to the problem faced, that, as some contributions in literature provide (e.g., Cachon and Fisher, 1997), is common to different industrial sectors. A particular specificity of the solution provided can be identified in the way the filtering procedure is applied. In particular, the system is capable of separating huge orders from small ones, so it is capable of dealing properly with bi-modal demand. However, the solution provided may have some problems when demand becomes multi-modal, thus when different relevant patterns overlap and different classes of customers (in terms of order size) can be identified. However, we argue that also in this situation the procedure can be adopted if information regarding the number of customers’ classes is known. In fact, by applying the filtering procedure different times the different “levels” of demand peaks can be separated so identifying different demand series, tied to the different customers’ classes. The analysis conducted in the Whirlpool case show in fact that the system performances rise when demand asymmetry is higher. This is due to the fact that the system (as applied in this situation) is effective when the two series (stable and peak series) are easily separable, since this condition makes the two forecasting systems more effective.

260

Part III Chapter 5: Conclusions and Future Developments

Moreover, the specific forecasting methods designed in this work are not specific to the particular contexts considered. In fact, the solution proposed in the Nestlé case may be applied whenever customers’ promotions are applied, since in all this situations the assumptions made by the system hold; as a matter of fact, customers may react differently towards promotional activities, thus making relevant to estimate their peaks profile. This claims that the applicability of this solution may be extended towards different forecasting problems where promotional activities are applied at customer level. As a matter of fact the analysis conducted within this case provide guide lines regarding when this kind of approach is feasible. Moreover, the methodology developed for managing peak series in Whirlpool can be generalized as well. In fact, the specificity of the solution is tied to the particular distribution of the interarrival period. However, this distribution may change according to the particular scenario considered, without affecting the applicability of the proposed system. Moreover, the idea of estimating separately the interarrival period and the demand size is not new at all (see Croston, 1972; Johnston and Boylan, 1996; Syntetos and Boylan, 2001). However, a relevant aspect considered in the solution proposed, differently from what has been done by other authors, is considering also the information that an expected order hasn’t shown up, thus to update the forecasting evaluation. This adds to current knowledge on effective systems to estimate when future orders will be emitted thus reducing estimation problems that similar approaches show (we refer to Part I for details regarding these problems). This may be a general issue in other contexts where customers tend to reorder regularly due to the specific cost structures they have.

261

Part III Chapter 5: Conclusions and Future Developments

3 Final remarks

As previously detailed, we argue that both the overall methodology developed and the specific solutions proposed may be adopted at some rate in different contexts. However, we also state that, given the complexity of the problem addressed, detailed analysis have to be applied to identify properly how to adopt these solutions. These results, moreover, validate the hypothesis previously defined in the methodological definition and provide some other relevant results. In these terms we can summarize the global findings of this work by means of 10 principles, referring to three different issue this work focuses on: from one side some results have been provided regarding heterogeneity and its effects on demand and forecasting, moreover, guide lines have been provided to properly design an effective forecasting solution, in the end, some elements have been considered to identify under which conditions the methodologies applied are more relevant. These principles, which partially recall the 5 research hypothesis considered, are here summarized.

Heterogeneity and its effects

1. Customers’ heterogeneity affects demand variability and forecasting accuracy. As provided by the both the simulation results and the empirical analysis conducted, heterogeneity impacts on both variability and forecasting accuracy, since it makes more complex providing effective forecasting systems given a certain forecasting effort.

2. Heterogeneity is more or less influent according to the specific decisional problem considered, in particular in terms of accuracy required. As previously stated, heterogeneity becomes relevant when relevant attention is paid towards forecasting issues. In fact, if for the company’s needs an aggregate forecast is enough, heterogeneity may not be a relevant matter, while, if particular attention is paid towards developing effective solutions, then attention has to be paid towards single customers, thus making heterogeneity a critical issue.

3. Supply chain complexity deeply influences demand heterogeneity and thus demand forecasting approaches. As provided in particular by the Nestlé and Whirlpool case, when supply chain is complex, heterogeneity rises, thus claiming that this is a structural factor that deeply is concerned in making perceived heterogeneity relevant.

262

Part III Chapter 5: Conclusions and Future Developments

How should demand be managed when heterogeneity shows up

4. When heterogeneity becomes relevant, a heterogeneous forecast has to be adopted. As showed in the different cases, the overall solution relies in adopting different forecasting systems for different parts of total demand. Thus heterogeneity may be used at one’s own benefit to cope properly with highly variable demand.

5. If heterogeneity is due to the supply chain complexity, then different forecasting approaches have to be adopted. As provided, by the Whirlpool and Nestlé case, when heterogeneity is mainly due to supply chain complexity, then customers demand different since there are influenced by different elements. Thus, at some rate, it less important to focus on how they differently react towards similar variables since, at some rate they may be influenced by completely different elements.

6. If heterogeneity is due to customers’ inner characteristics, several estimation processes have to be adopted. As provided by Nestlé and Ahold cases, when customers’ inner characteristics are considered as a relevant part of total heterogeneity, then the forecasting structure fits, while several estimation processes have to be conducted. In fact, in the Ahold case we showed, that in some conditions, the cluster-based approach may be effective compare to other aggregation solutions. Similar results have been provided by the Nestlé experience. However, attention has to be to the problem of defining proper cluster of customers on which apply the forecasting methodology.

7. Demand perceived heterogeneity may be reduce by disaggregating and aggregating customers in groups, so to reduce clusters inner variability. As the applications of the clustering approach has showed, heterogeneity may be reduced and thus forecast be applied by means of statistical clustering techniques. However, also managerial clustering, as that provided for the monitored customers in the Nestlé case may be significant, thus claiming that composite approaches should be adopted, by properly melting together statistical, managerial and judgmental elements.

Elements to be considered

8. The less variability is due to systematic phenomena (i.e., seasonality), the more heterogeneity becomes a relevant matter. This should be considered as a design rule: when systematic variability has been already considered or it is not relevant, then heterogeneity among customers

263

Part III Chapter 5: Conclusions and Future Developments

becomes the major issue that should be taken into account. Example of this fact is provided in the Nestlé case, that in fact “cleans” data from seasonality and then classifies customers.

9. The more demand is variable, the more heterogeneity tends to be relevant. As provided in the Nestlé case, if demand is highly variable and if variability is not completely due to systematic variability then, customers should be considered. This may be considered as a general rule that claims that complexity within the forecasting process due to considering customers should be done only when high variability is faced.

10. The more supply chain is complex, the more heterogeneity should be taken into account. This principles directly reflects on principle 3. In fact, it claims that heterogeneity is for sure relevant when supply chain is complex. So when dealing with many different channels, by means of different intermediaries, applying different politics, a heterogeneous forecasting process is worth being applicable.

264

Part III Chapter 5: Conclusions and Future Developments

4 Future Developments

Even if this work provides interesting results within the field of demand forecasting, we argue that more research and empirical practice should be given to this issue. In fact, from the analysis conducted and the experiences matured, some elements seem to need further development. In particular, as it has been stated, in this work there seems to be a relevant overlap within marketing and forecasting issue. In particular, attention has been paid to the role of customers within internal process within companies, which is a relevant issue in all the direct marketing research area. We argue that this overlap is significant and may be used to gain relevant benefits. In particular, we argue that Customer Relationship Management, one of the most developing paradigms within many industrial companies, may significantly help to improve forecasting capabilities.

This work shows that if we pay attention towards customers and try to exploit relevant information regarding their behavior, significant improvements can be achieved. Customer Relationship Management has been typically associated to direct marketing issues, however many companies are experiencing that it becomes ever more a retention tool to keep customers alive and close to the company. CRM tools and methodologies provide the opportunity to interact more with customers and to gain relevant information regarding their characteristics, thus we argue that joining this process with forecasting may be beneficial. A simple example may help focusing this issue. Consider a firm producing machining centers operating worldwide. This company has introduced since a few years a new service available to their customers consisting in the remote monitoring of the machine to manage malfunctions. The idea of this service, when first introduced, was to give customers wider services as more customized functionalities directly provided on the machine. The system is simply based on a remote center that controls the operating machines through a diagnostic system inserted into each machine that provides communication by means of the Internet. The aim of this system is to provide customers with a better service by means of providing them first of all with hints and suggestions regarding the way in which machines are typically adopted (consider that usually malfunctions are due to erroneous working parameters adopted in the production); moreover, the system may provide the company with information regarding particular operating parameters, malfunction details, and so on. This gives the company the possibility to have quick and detailed information regarding the problems of the sold machines, thus letting them forecast easier both the required service time that should be allocated by means of repairer and requirements of spare parts.

265

Part III Chapter 5: Conclusions and Future Developments

Another relevant example of these possibilities has been found in a multi-national chemical company that has adopted CRM as a way to get closer to customers so to force them to provide the company with early sales data. As a matter of fact the firm had to develop major changes within her organizational structure and processes (both internally and externally towards its customers) and had to deeply analyze the profitability of each single customer. The final result was the definition of an heterogeneous forecasting approach that, for some customers, leverages on the early information provided matched with historical analyses, while for others a more aggregate approach has defined that simply leverage on past sales extrapolation. This experience claims too for the relevant of focusing jointly on customer interaction (e.g., CRM) and forecasting approach. However, we argue that since few experiences of the integration of these systems are available, further research should be provided.

A second relevant issue is tied to the implementation of the proposed approaches within the specific companies considered. As a matter of fact, when the decision of realizing these solution was taken a major issue was identifying proper forecasting software that could handle the systems developed. In fact studies regarding this issue showed that also software houses are focusing more on these issues, since some of them are developing systems that can handle data at different forecasting levels, thus making possible to forecast some customers at a detailed level, while for others more aggregate forecasts can be exploited. In these terms we argue that also research regarding the applicability of the proposed methodologies within software systems may be a relevant topic to focus on. As a matter of fact, the companies considered in this study have effectively implemented the forecasting solution proposed by means of different software solutions. This claims that the empirical results obtained are both relevant and, more important in an industrial perspective, applicable.

266

References

Abel A.B., 1985, Inventories, stock-outs and production smoothing, Review of Economic Studies, 52 (2), 283-293.

Agarwal N., S.A. Smith, 1996, Estimating negative binomial demand for retailer inventory management with unobservable lost sales, Naval Research Logistics, 43, 839-861.

Alderson W., 1950, Marketing efficiency and the principle of postponement, Cost and Profit Outlook, 3.

Alvarado U.Y., H. Kotzab, 2001, Supply Chain Management: The Integration of Logistics in Marketing, Industrial Marketing Management, 30, 183-198.

Anderson J.C., J.A. Narus, 1995, Capturing the value of supplementary services, Harvard Business Review, 73, January-February, 75-83.

Arntzen B., G. Brown, T. Harrison, L. Trafton, 1995, Global supply chain management at Digital Corporation, Interfaces, 25 (1), 69-93.

Armstrong J.S., F. Collopy, 1992, Error measures for generalizing about forecasting methods: Empirical comparisons, International Journal of Forecasting, 8, 69-80.

267

References

Armstrong J.S., 2001, Principles of forecasting – a handbook for researchers and practitioners, Kluwer Academic Publishers, Boston – Dordrecht – London.

Atuahene-Gima K., 1996, Market orientation and innovation, Journal of Business Research, 35, 93- 103.

Axsäter S., 1993, Continuous Review Policies for Multi-Level Inventory Systems with Stochastic Demand, Logistics of Production and Inventory – Handbooks in Operation Research and Management Science, vol.4.

Azoury K.S., 1985, Bayes solution to dynamic inventory models under unknown demand distribution, Management Science, 31 (9), 1150-1160.

Baker K.A., M.J. Magazine, H.L.W. Nuttle, 1986, The effect of commonality on safety stock in a simple inventory model, Management Science, vol. 32, n. 8, 982-988.

Balakrishnan S., 1996, Benefits of customer and competitive orientations in industrial markets, Industrial Marketing Management, 25, 257-69.

Banerjee S., A. Banerjee, J. Burton, W. Bistline, 2001, Controlled partial shipments in two-echelon supply chain network: A simulation study, International Journal of Production Research, 71, 91- 100.

Barnea A., J. Lakonishok, 1980, An analysis of the usefulness of disaggregated accounting data for forecasts of corporate performance, Decision Sciences, 11, 17-26.

Bartezzaghi E., R. Verganti, 1995, Managing demand uncertainty through order overplanning, International Journal of Production Economics, 40, 107-120.

Bartezzaghi E., R. Verganti, G. Zotteri, 1999a, A simulation framework for forecasting uncertain lumpy demand, International Journal of Production Economics, 59, 499-510.

Bartezzaghi E., R. Verganti, G. Zotteri, 1999b, Measuring the impact of asymmetric demand distributions on inventories, International Journal of Production Economics, 60-61, 395-404.

Bartezzaghi E., R. Verganti, G. Zotteri, 1999c, I nuovi approcci alla gestione della domanda: il ruolo dell’informazione, in Campi C., A. D’Angelo, A. La Bella, Le opportunità del cambiamento tecnologico, ed. D’Anselmi Editore – Hoepli.

Bassok Y., R. Akella, 1991, Ordering and production decisions with supply quality and demand uncertainty, Management Science, 37, 1556-1574.

Bell P.C., 1981, Adaptive sales forecasting with many stock-outs, Journal of Operational Research Society, 32, 865-873.

268

References

Berry L.L., 1983, Relationship marketing, In Perspectives in service marketing, eds L.L. Berry, G.L. Shostack, G.D. Upah, American Marketing Association, Chicago, IL.

Biggs J.R., W.M. Campion, 1982, The effect and cost of forecast error bias for multiple-stage production-inventory systems, Decision Sciences, 13 (4), 570-584.

Bowersox D.J., J.T. Mentzer, T.W. Speh, 1995, Logistics Leverage, Journal of Business Strategies, 12 (2), 36-49.

Box G.E.P., G.M. Jenkins G.M., 1976, Time Series Analysis: Forecasting and Control, Holden-Day, San Francisco.

Box G.E.P., G.M. Jenkins, G.C. Reinsel, 1994, Time series analysis forecasting and control, 3rd ed., Prentice-Hall, Englewood Cliffs, N.J.

Brace G., 1989, Market powertrain: an imperative to co-operation, In: Proceedings of Commission of European Communities Partnership between Small and Large Firms Conference, Graham & Trotman, London.

Brandolese A., R. Cigolini, 1999, A new model for the strategic management of inventories subject to peaks in market demand, International Journal of Production Research, 37 (8), 1859-1880.

Brown R.G., 1959, Statistical forecasting for inventory control, McGraw-Hill, New York.

Brown R.G., 1963, Smoothing, forecasting and prediction of discrete time series, Prentice-Hall, Englewood Cliffs, N.J.

Buklin L.P., 1965, Postponement, speculation and the structure of distribution channels, Journal of Marketing Research, 2, 26-31.

Burbidge J.L., 1961, The new approach to production, The Production Engineer 40 (12), 769–784.

Busby Fan, 1993, The extended manufacturing enterprise: its nature and its needs, International Journal of Technology Management, 8 (3-4-5).

Cachon G., M. Fisher, 1997, Campbell soup’s continuous replenishment program: evaluation and enhanced inventory decision rules, Production and Operation Management, 6 (3), 266-276.

Cachon G., 1999, Managing supply chain demand variability with scheduled ordering policies, Management Science, 45, (6), 843-856.

Cagliano R., M. Sassatelli, A. Savoldelli, 1999, Supply network reengineering: some evidence from the field, Proceedings of the conference on SMEs, 7th of May.

269

References

Canel C., B. Khumawala, 1996, A mixed-integer programming approach for the international facilities location problem, International Journal of Operations and Production Management, 16 (4), 49-68.

Caplin A.S., 1985, The variability of aggregate demand with (s,S) inventory policies, Econometrics, 53 (6), 1395-1409.

Carbone R., W.L. Gorr, 1985, Accuracy of judgemental forecasting of time series, Decision Sciences, 16, 153-160.

Carothers G., M. Adams, 1991, Competitive advantage through customer value: the role of value- based strategies, In: Stahl, M.J., Bounds, G.M. (Eds.), Competing Globally Through Customer Value. Quorum Books, New York, 32–66.

Chen F., Z. Drezner, J.K. Ryan, D. Simchi-Levi, 2000, Quantifying the Bullwhip Effect in a Simple Supply Chain: The Impact of Forecasting, Lead Times, and Information, Management Science, 46 (3), 436 – 443.

Chiang J., 1991, A simultaneous approach to the whether, what and how much to buy questions, Marketing Science, 10 (4), 297-315.

Childerhouse P., J. Aitken, D.R. Towill, 2002, Analysis and design of focused demand chains, Journal of Operations Management, 20, 675-689.

Christopher M., A. Payne, D. Ballantyne, 1991, Relationship marketing, Butterworth-Heinemann, Oxford.

Christopher M., 1993, Logistics and Competitive Strategy, European Management Journal, 11 (2), 258-261.

Clark R., 1985, Industrial Economics, Blackwell, Cambridge.

Cobbaert K., D. Van Oudheusden, 1996, Inventory models for fast moving spare parts subject to sudden death obsolescence, International Journal of Production Economics, 44, 239 – 248.

Cohen M.A., P.R. Kleindorfer, 1989, Near – optimal service constrained stocking policies for spare parts, Operations Research, 37 (1).

Cole R., P. Bacdayan, B.J. White, 1993, Quality, participation and competitiveness, Californian Management Review (spring), 35 (3), 68-81.

Collier D.A., 1982, Aggregate safety stock levels and component part commonality, Management Science, 28 (11), 1297-1303.

270

References

Colombo R., J.T. Landwehr, 1990, A non-parametric approach to incorporating heterogeneity into the logit model, working paper

Conduit J., F.T. Mavondo, 2001, How critical is internal customer orientation to market orientation, Journal of Business Research, 51, 11-24.

Coughlan P., D. Coghlan, 2002, Action research for operations management, International Journal of Operations and Production Management, 22, (2), 220-240.

Couldwell C., 1998, A data day battle, Computing, 21, May, 64-66.

Cross J.G., 2000, How e-business is transforming supply chain management, Journal of Business Stretegy, March/April, 36-39.

Croston J.K., 1972, Forecasting and stock control for intermittent demands, Operational Research, 23 (3), 289-304.

Croxton F.E., D.J. Cowden, 1955, Applied general statistics, 2nd ed., Prentice-Hall, Englewood Cliffs, N.J.

Currim J.S., R.J. Meyer, N.T. Lee, 1988, Disaggregate tree-structured modelling of consumer choice data, Journal of Marketing Research, 25, 253-265.

Da Silva RV., G. Davies, P. Naudé, 2002, Assessing customer orientation in the context of buyer/supplier relationships using judgemental modelling, Industrial Marketing Management, 31, 241-252.

Dagum, E.B., 1980, The X-11-ARIMA seasonal adjustment Method, Statistics Canada, Ottawa.

Dalrymple D.J., 1987, Sales forecasting practices: Results of a United States survey, International Journal of Forecasting, 3, 379-392.

Dar-El E.M., C.J. Malmborg, 1991, Improved strategy for service level based safety stock determination, Production Planning and Control, 2 (2), 116-121.

Davenport T. H., 1993, Process Innovation, Boston: Harvard Business School Press.

David I., E. Greenshtein, A. Mehrez, 1997, A dynamic programming approach to continuous review obsolescent inventory problem, Naval Research Logistics, 44 (8).

Davidow W.H., M. Malone, 1992, The virtual corporation, Harper Business, New York.

Day G.S., R. Wensley, 1988, A framework for diagnosing competitive superiority, Journal of Marketing, 52, 1-20.

271

References

De Kok A.G., 1990, Hierarchical production planning for consumer goods. European Journal of Operational Research, 45, 55-69.

De Lurgio S.A., 1997, Forecasting principles and applications, Irwin Publishing, Burr Ridge, Ill.

Dean I., A. Carrie, 1996, The franchised factory: a new paradigm for manufacturing, Proceedings of 3rd International Conference of EUROMA, London, 2-4 June.

Deighton J., 1998, The future of interactive marketing, Harvard Business Review, 74 (6), 151-160.

Dekker R., M.J. Kleijn, P.J. De Rooij, 1998, A spare parts stocking policy based on equipment criticality, International Journal of Production Economics, 56 – 57, 69 – 77.

DeSarbo, W., K. Jedidi, I. Sinha, 2001, An Empirical Investigation of the Structural Antecedents of Perceived Value in a Heterogeneous Population, Strategic Management Journal, 22, 845-857.

Deshpande R., J. Farley, F. Webster, 1993, Corporate culture, customer orientation and innovativeness in Japanese firms: a quadrad analysis, Journal of Marketing, 57 (1), 23-37.

Diamantopoulos A., S. Hart, 1993, Linking market orientation and company performance: evidence on Kohli and Jaworski’s Framework, Journal of Strategic Marketing, 1 (2), 92-122.

Diks E.B., A.G. De Kok, A.G. Lagodimos, 1996, Multi-echelon systems: A service measure perspective, European Journal of Operational Research, 95, 1-23.

Donselaar van K., 1989, Material coordination under uncertainty – towards more flexible concepts, Eindhoven University of Technology, dissertation.

Donselaar van K., 1990, Integral stock norms in divergent systems with lot-sizes, European Journal of Operational Research, 45, 70-84.

Donselaar van K, L.R. Kopczak, M. Wouters, 2001, The use of advance demand information in a project-based supply chain, European Journal of Operational Research, 130, 519-538.

Draper N.P., H. Smith, 1981, Applied regression analysis, 2nd ed., Wiley, New York.

Dyer J.H., 1996a, Specialized supplier networks as a source of competitive advantage: evidence from the auto industry, Strategic Management Journal 17 (4), 271–291.

Dyer J.H., 1996b, Does governance matter? Keiretsu alliances and asset specificity as sources of Japanese competitive advantage, Organization Science 7 (6), 649–666.

Dyer J.H., 1996c, How Chrysler created an American Keiretsu, Harvard Business Review 74 (4), 42–56.

272

References

Dyer J.H., 1997, Effective inter-firm collaboration: how firms minimize transaction costs and maximize transaction value, Strategic Management Journal 18 (7), 535–556.

Ebert R.J., T.S. Lee, 1995, Production loss function and subjective assessments of forecast errors: untapped sources for effective master production scheduling, International Journal of Production Research, 33, 137-159.

Edmundson R., M. Lawrence, M. O’Connor, 1988, The use of nontime series information in sales forecasting: a case study, Journal of Forecasting, 10, 597-617.

Ellinger A.E., 2000, Improving Marketing Logistics Cross-Functional Collaboration in the Supply Chain, Industrial Marketing Management, 29, 85-96.

Eppen G., L. Schrage, 1981, Centralized ordering policies in a multi-warehouse systems with lead times and random demand, In: Schwarz L.B. (Ed.), Multi-level Production-Inventory Control Systems: Theory and Practice, North-Holland, Amsterdam.

Eynan A., 1996, The impact of demand correlation on the effectiveness of component commonality, International Journal of Production Research, 2 (2), 116-121.

Federgruen A., 1993, Centralized Planning Models for Multi-Echelon Inventory Systems under Uncertainty, Logistics of Production and Inventory – Handbooks in Operation Research and Management Science, 4.

Feitzinger E., H.L. Lee, 1997, Mass customization at Hewlett Packard: the power of postponement, Harvard Business Review, 75, January-February, 116-121.

Fildes R., C. Beard, 1992, Forecasting system for production and inventory control, International Journal of Operation and Production Management, 12 (5), 4-27.

Fisher M., J.H. Hammond, W.R. Obermayer, A. Raman, 1994, Making supply meet demand in an uncertain world, Harvard Business Review, 72 (3), 83-93.

Fisher M., A. Raman, A.S. McClelland, 2000, Rocket science retailing is almost here, are you ready?, Harvard Business Review, July-August, 115-124.

Fisher M.L., A. Raman, 1992, Reducing the cost of demand uncertainty through accurate response to early sales, Report n. 92-10-03, Harvard University.

Fisher M.L., 1997, What is the right supply chain for your product? Harvard Business Review 75 (2), 105–116.

Fliedner G., 1999, An investigation of aggregate variable time series forecast strategies with specific subaggregate time series statistical correlation, Computers & Operations Research; 26, 1133-49.

273

References

Ford D., 1990, Understanding Business Markets, Academic Press, London.

Forrester J.W., 1958, Industrial dynamics: a major breakthrough for decision makers. Harvard Business Review 36 (4), 37–66.

Forrester J.W., 1961. Industrial Dynamics. MIT Press, NY.

Friend J.K., 1960, Stock Control with random opportunities for replenishment, Operations Research Quarterly, 11 (3), 130-136.

Frohlich M.T., R. Westbrook, 2001, Arcs of integration: an international study of supply chain strategies, Journal of Operations Management 19, 185–200.

Frohlich M.T., R. Westbrook, 2002, Demand chain management in manufacturing and services: web-based integration, drivers and performance, Journal of Operations Management, 20, 729-745.

Fuller J.B., J. O’Conor, R. Rawlinson, 1993, Tailored logistics: the next advantage, Harvard Business Review, May/June, 87-98.

Galbraith J.R., 1977, Organization Design, Addison Wesley, Reading.

Galbreath J., 1998, Relationship management environments, Credit World, 87 (2), 14-21.

Gill P., J. Abend, 1997, Wal-Mart: the supply chain heavyweight Champ., Supply Chain Management Review, 1 (Spring), 12-21.

Gilmore J.H., B.J. Pine, 1997, The four faces of mass customization, Harvard Business Review, 75, January-February, 91-101.

Gönül F., K. Srinivasan, 1993, Modelling unobserved heterogeneity in multinominal logit models: methodological and managerial implications, Marketing Science, 12 (3), 213-229.

Gordon A.D., 1980, Classification, Chapman and Hall, London.

Gould F. J., G.D. Eppen, C.P. Schmidt, 1993, Introductory Management Science, Prentice-Hall.

Greenley G., 1995, Market orientation and company performance: empirical evidence from UK companies, British Journal of Management, 6 (1), 1-13.

Gross C.W., E.S. Jeffrey, 1990, Disaggregation methods to expedite product line forecasting, Journal of Forecasting, 9, 253-254.

Grunfeld Y. Z. Griliches, 1960, Is aggregation necessarily bad?, Review of Economics and Statistics, 42, 1-13.

274

References

Guadagni P., J. Little, 1983, A logit model of brand choice, Marketing Science, 2, 203-238.

Guerrero H.H., K.R. Baker, M.H. Southard, 1986, The dynamics of hedging the master schedule, International Journal of Production Research, 24, 2579-2599.

Gulati R., J. Garino, 2000, Get the right mix of bricks and clicks, Harvard Business Review, May- June.

Gupta S., 1988, Impact of sales promotions on when, what and how much to buy, Journal of Marketing Research, 25, 342-355.

Gupta S., P.K. Chintagunta, D.R. Wittink, 1997, Household heterogeneity and state dependence in a model of purchase strings: Empirical, results and managerial implications, International Journal of Research in Marketing, 14, 341-357.

Hadley G., T.M. Within, 1963, Analysis of Inventory Systems, Prentice-Hall, Englewood Cliffs, N.J.

Hakansson H., I. Snehota, 1995, Developing Relationships in Business Networks, Routledge, London.

Hakimi S.L., C.C. Kuo, 1991, On a general network location allocation problem, European Journal of Operational Research, 55, 31-45.

Hammel T., L. Kopczak, 1993, Tightening the supply chain, Production and Inventory Management Journal, 34 (2), 63-70.

Hammer M., (1990), Reengineering work: don’t automate – obliterate, Harvard Business Review, July-August.

Hammer M., J. Champy, 1993, Re-engineering the corporation: a manifesto for business revolution, London, Nicholas Brealey Publishing.

Hannah L., J.A. Kay, 1977, Concentration in Modern Industry, London, Macmillan.

Hansen P., E.L. Pedrosa Filho, C.C. Ribeiro, 1992, Location and sizing of offshore platforms for oil exploration, European Journal of Operational Research, 58, 202-214.

Harland C. M., 1996, Supply chain management: relationships, chains and networks, British Journal of Management, 7, 63-80.

Heckman J.J., B. Singer, 1984, A method for minimizing the impact of distributional assumptions in econometric models for duration data, Econometrica, 52 (2), 271-320.

Heide J.B., G. John, 1990, Alliances in industrial purchasing: the determinants of joint action in buyer–supplier relationships, Journal of Marketing Research 27 (1), 24–36.

275

References

Heikkilä J., 2002, From supply to demand chain management: efficiency and customer satisfaction, Journal of Operations Management, 20, 747-767.

Higgins P., J. Browne, 1992, MPS: a concurrent planning approach, Production Planning and Control, 3 (1), 2-18.

Hines P., R. Silvi, M. Bartolini, 2002, Demand chain management: an integrative approach in automotive retailing, Journal of Operations Management, 20, 707-728.

Ho C.J., 1995, Examining the impact of demand lumpiness on the lot-sizing performance in MRP systems, International Journal of Production Research, 33 (9), 2579-99.

Ho C.J., K. S.C. Kim, M. Koo, 2001, MRP system performance under lumpy demand environments, Production Planning and Control, 12, (4), 316-325.

Hoekstra S., J. Romme, 1992, Integral logistics structure, McGraw-Hill, London.

Hollier R.H., K.L. Mak, C.L. Lam, 1995, An inventory model for items with demands satisfied from stock or by special deliveries, International Journal of Production Economics, 42, 229-236.

Holmström J., 1994, The relationship between speed and productivity in industry networks: a study of industrial statistics. International Journal of Production Economics 34 (1), 91–98.

Holmström J., 1995, Speed and efficiency—a statistical enquiry of manufacturing industries, International Journal of Production Economics 39 (3), 185–191.

Holmström J., 1997, Product range study of supply chain operations in the European grocery industry, Supply Chain Management, 2 (3), 107-115.

Inger R., A. Braithwaite, M. Christopher, 1995, Creating a manufacturing environment that is in harmony with the market – the ‘how’ of supply chain management, Production Planning & Control, 6 (3), 246 – 257.

Ilmakunnas P., 1990, Aggregation vs. disaggregation in forecasting construction activity, in: T.S. Barker, H. Pesaran, Disaggregation in Econometric Modelling, Routledge, London, 73-86.

Jain D.C., N.J. Vilcassim, 1991, Investigating household purchase timing decisions: a conditional hazard function approach, Marketing Science, 10 (1), 1-23.

Janssen F., R. Heuts, T. de Kok, 1998, On the (R, s, Q) inventory model when demand is modelled as a compound Bernoulli process, European Journal of Operational Research, 104, 423-436.

Jaworsky B., A. Kohli, 1993, Market orientation: antecedents and consequences, Journal of Marketing, 57, 53-70.

276

References

Johnston F.R., J.E. Boylan, 1996, Forecasting for items with intermittent demand, Journal of the Operational Research Society, 47, 113-121.

Jones J.M., J.T. Landwehr, 1988, Removing heterogeneity bias from logit model estimation, Marketing Science, 7 (1), 41-59.

Kadipasaoglu S.N., 1995, The effect of freezing the master production schedule on cost in multilevel production systems, Production and Inventory Management, Third Quarter, 30-36.

Kahn B.E., 1998, Dynamic relationships with customers: high-variety strategies, Journal of the Academy of Marketing Science, 26, Winter, 45-53.

Kahn J.A., 1987, Inventories and the volatility of production, American Economic Review, 77 (4), 667-679.

Kahn K.B., 1998, Revisiting Top-Down versus Bottom-Up Forecasting, Journal of Business Forecasting, Summer: 14-19.

Kannan P.K., G.P. Wright, 1991, Modelling and testing structured markets: a nested logit approach, Marketing Science, 10, 58-82.

Kekre S., T.E. Morton, T.L. Smunt, 1990, Forecasting using partially known demands, International Journal of Forecasting, 6, 115-125.

Klassen R.D., B.E. Flores, 2001, Forecasting practices of Canadian firms: Survey results and comparisons, International Journal of Production Economics, 70, 163-174.

Kleinbaum D. G., 1996, Survival analysis: a self learning text, Springer, New York.

Knight F.M., 1921, Risk uncertainty and profit, Houghton Mifflin, Boston.

Kohli A.J., B.J. Jaworski, 1990, Market orientation: the construct, research propositions and managerial implications, Journal of Marketing, 54, 1 –18.

Korpela J., A. Lehmusvaara, M. Tuominen, 2001, Customer service based design of the supply chain, International Journal of Production Economics, 69, 193-204.

Kotha S., 1995, Mass customization: implementing the emerging paradigm for competitive advantage, Strategic Management Journal, 16, 21-42.

Kotler P., 1991, Marketing Management: Analysis, Planning, Implementation and Control, 7th ed., Prentice Hall, Englewood Cliffs, NJ.

Krishnamurthi L., S.P. Raj, 1988, A model of brand choice and purchase quantity price sensitivities, Marketing Science, 7 (1), 1-20.

277

References

Kumar N., L.K. Scheer, J.-B.E.M. Steenkamp, 1995, The effects of perceived interdependence on dealer attitudes, Journal of Marketing Research 32 (3), 348–356.

Kutner S., J. Cripps, 1997, Managing the customer portfolio of healthcare enterprises, The Healthcare Forum Journal, 40 (5), 52-54.

Lambert R., 1989, Distribution and Re-distribution of Income, A Mathematical Analysis, New York, Manchester University Press.

Lambert D., J. Stock, L. Ellram, 1998, The Global Supply Chain Forum, in Fundamentals of Logistics Management, Irwin-McGraw Hill, Boston, 504.

Lambert D.M., M.C. Cooper, 2000, Issues in Supply Chain Management, Industrial Marketing Management, 29, 65-83.

Lamming R., T. Johnsen, J. Zheng, C. Harland, 2000, An initial classification of supply networks, International Journal of Operations and Production Management, 20, (6), 675-691.

Lampel J., H. Mintzberg, 1996, Customizing customization, Sloan Management Review, Fall, 21- 30.

Lancioni R.A., M.F. Smith, T.A. Oliva, 2000, The role of the internet in supply chain management, Industrial Marketing Management, 29, 45-56.

Langley C.J., M.C. Holcomb, 1992, Creating logistics customer value, Journal of Business Logistics 13 (2), 1–27.

Lapide L., 1998, New Developments in Business Forecasting, Journal of Business Forecasting, Summer, 28-29.

Lee C, 1993, The multiproduct warehouse location problem: applying a decomposition algorithm, International Journal of Physical Distribution and Logistics Management, 23 (6), 3-13.

Lee H.L., C.S. Tang, 1997, Modelling the costs and benefits of delayed product differentiation, Management Science, 43, (1), 40-54.

Lee H.L., V. Padmanabhan, S. Whang, 1997, Information distortion in a supply chain: The Bullwhip effect, Management Science, 43 (4), 546-559.

Lee T.S., E.A. Everett, 1986, Forecasting error evaluation in material requirement planning (MRP) production-inventory systems, Management Science, 32, 1186-1204.

Lee H.L., Billington, C., 1992, Managing supply chain inventory: pitfalls and opportunities, Sloan Management Review 33 (3), 65–73.

278

References

Lin N.-P, L. Krajewski, 1993, A model for master production scheduling in uncertain environments, Decision Science, 23, 839-861.

Liu L., D.H. Shi, 1999, An (s, S) model for inventory with exponential lifetimes and renewal demands, Naval Research Logistics, 46 (1).

Lockard M, 1998, Test your retention IQ, Target Marketing, 21 (3), 32-41

Lovejoy W.S., 1990, Myopic policies for some inventory models with uncertain demand distributions, Management Science, 36 (6), 724-738.

Mak K.L., K.K. Lai, 1995, The determination of optimal partial fill policy for an inventory system with lumpy demand items, Applied mathematical modelling, 19 (12), 724-737.

Mak K.L., Y.S. Wong, Huang G.Q., 1999, Optimal inventory control of lumpy demand items using genetic algorithms, Computers and Industrial Engineering, 37, 273-276.

Makridakis S., 1988, Metaforecasting, International Journal of Forecasting, 4, 467-491.

Makridakis S., S.C. Wheelwright, V.E. McGee, 1983, Forecasting methods and applications, John Wiley and Sons, New York.

Makridakis S., S.C. Wheelwright, 1989, Forecasting methods for management, Wiley, New York.

Makridakis S., 1993, Accuracy measures: Theoretical and practical concerns, International Journal of Forecasting, 9, 527-529.

Matear S., C. Boshoff, B. Gray, P. Matheson, 1997, Models of Market Orientation: A Comparative Study. 8th World Marketing Congress Proceedings, Kuala Lumpur, Malaysia.

McKinsey in Rangone A., 2000, Dispense del corso di Strategia e Pianificazione aziendale, Politecnico di Milano.

Mentzer J.T., J.E. Jr. Cox, 1984a, Application and performance of sales forecasting techniques, Journal of Forecasting, 3, 27-36.

Mentzer J.T., J.E. Jr. Cox, 1984b, A model of the determinants of achieved forecast accuracy, Journal of Business Logistics, 5, 143-55.

Mentzer J.T., K. Kahn, 1996, Logistics and Interdepartmental Integration, International Journal of Physical Distribution and Logistics Management, 26 (8), 6-14. Meredith J.R., A. Raturi, K. Amoako-Gyampah, B. Kaplan, 1989, Alternative research paradigms in operations, Journal of Operations Management, 8, 297-326.

279

References

Merli G., C. Saccani, 1994, L'azienda Olonico-virtuale. Un'opportunità storica per la piccola e media impresa, Il Sole24Ore.

Meshkat H., R. Ballou, 1996, Warehouse location with uncertain stock availability, Journal of Business Logistics, 17 (2), 197-216.

Metters R., 1997, Quantifying the bullwhip effect in supply chains, Journal of Operations Management, 15 (2), 89-100.

Miller J.G., W. Berry, C.-Y.F. Lai, 1976, A comparison of alternative forecasting strategies for multistage production inventory systems, Decision Sciences, 7, 714-24.

Mohr J., R. Spekman, 1994, Characteristics of partnership success: partnership attributes, communication behaviour, and conflict resolution techniques, Strategic Management Journal 15 (2), 135–152.

Monczka R.M., K.J. Petersen, R.B. Handfield, G.L. Ragatz, 1998, Success factors in strategic supplier alliances: the buying company perspective, Decision Sciences 29 (3), 553–573.

Mouritsen J., 1997, Marginalizing the customer: customer orientation, quality and accounting performance, Scandinavian Journal of Management, 13 (1), 5-18.

Muir J.W., 1979, The pyramid principle, Proceedings of 22nd Annual Conference American Production and Inventory Control Society, 105-107.

Murthy D.N.P., L. Ma, 1991, MRP with uncertainty: a review and some extensions, International Journal of Production Economics, 25, 51-64.

Nakane J., R.W. Hall, 1991, Holonic Manufacturing: flexibility - the competitive battle in the 1990s, Production Planning & Control, 2 (1).

Narver J.C., S.F. Slater, 1990, The effect of market orientation on business profitability, Journal of Marketing, 54, 20 – 35.

Naylor J.B., M.M. Naim, D. Berry, 1999, Leagility: interfacing the lean and agile manufacturing paradigms in the total supply chain, International Journal of Production Economics, 62, 107-118.

Newman W.R., M. Hanna, M.J. Maffei, 1993, Dealing with the uncertainties of manufacturing: flexibility, buffers and integration, International Journal of Operations and Production Management, 13, 19-24.

Orcutt G., H.W. Watts, J.B. Edwards, 1968; Data aggregation and information loss, American Economic Review, 58, 773-87.

280

References

Öller L.E., 1989, Aggregating problems when forecasting industrial production using business survey data, Discussion Paper, No. 20, Ministry of Finance, Economics Department, Helsinky.

Pagh J.D., M.L. Cooper, 1998, Supply chain postponement and speculation strategy: how to choose the right strategy, Journal of Business Logistics, 19 (2), 13-33.

Peppard J., 2000, Customer Relationship Management (CRM) in Financial Services, European Management Journal, 18 (3) 312-327.

Petrović R., A. Šenborn, M Vujošević, 1988, A new adaptive algorithm for determination of stocks in spare parts inventory systems, Engineering Costs and Production Economics, 15, 405 – 410.

Pielou, E.C., 1977, Mathematical Ecology John Wiley, New York.

Pindyck R.S., D.L. Rubinfeld, 1991, Econometric models and economic forecasts, McGraw-Hill, New York.

Pine B.J. 1993, Mass Customization: The New Frontier in Business Competition, Harvard University Press, Cambridge, MA.

Pitt L, A. Caruana, P. Berthon, 1996, Market orientation and business performance: some European evidence, International Marketing Review 13, 5-18.

Porter M. E., 2001, Strategy and the Internet, Harvard Business Review, March.

PWC in Rangone A., 2000, Dispense del corso di Strategia e Pianificazione aziendale, Politecnico di Milano.

Raman A., G. Zotteri, 1998, The estimation of retail demand and lost sales, 10th International Symposium on Inventories, Budapest, 24-28 August.

Reichheld FE, W. Sasser, 1990, Zero defections: quality comes to services, Harvard Business Review, 68(5), 105 – 11.

Remus W., M. O’Connor, K. Griggs, 1996, Does feedback improve the accuracy of recurrent judgemental forecasts?, Organizational Behaviour and Human Decision Processes, 66, 22-30.

Rice G., 1997, Forecasting in US firms: a role for TQM?, International Journal of Operations and Production Management, 17, 2, 221-230.

Ritzman L.P., B.E. King, 1993, The relative significance of forecast errors in multistage manufacturing, Journal of Operations Management, 11, 51-65.

Ronchi S., 2001, The role of the internet in customer-supplier relationships: the emergence of collaborative markets, PhD Thesis, Politecnico di Milano, Milano (Italy).

281

References

Ruekert RW, 1992, Developing a market orientation: an organisational strategy perspective, International Journal of Research in Marketing, 9 (3), 225-45.

Saaty T.L., 1990, The Analytic Hierarchy Process, RWS Publications, USA.

Sanders N.R., K. Manrodt, 1994, Forecasting practices in US corporations: survey results, Interfaces, 24, 92-100.

Sanders N.R., L.P. Ritzman, 1992, The need for contextual and technical knowledge in judgmental forecasting, Journal of Behavioural Decision Making 5, 39-52.

Sanders N.R., 1997, The status of forecasting on manufacturing firms, Production and Inventory Management Journal, Second Quarter, 32-35.

Sani B., B.G. Kingsman, 1997, Selecting the best periodic inventory control and demand forecasting methods for low demand items, Journal of the Operational Research Society, 48, 700- 713.

Schwarzkoph A.B., R.J. Tersine, J.S. Morris, 1988, Top-down versus Bottom-Up Forecasting Strategies, International Journal of Production Research, 26 (11), 1833-1843.

Seidel D., M. Mey, 1994, IMS - Holonic manufacturing systems: strategies, Vol. 1, IFW, University of Hannover, Germany.

Shapiro R.D., J.L. Heskett, 1985, Logistics Strategy: Cases and Concepts, West publishing, St. Paul, MN.

Shapiro J.F., V.M. Singhal, S.N. Wagner, 1993, Optimizing the value chain, Interfaces 23 (2), 102– 117.

Sharman G., 1984, The rediscovery of logistics, Harvard Business Review 62 (5), 65–77.

Shibuya T., T. Dohi, S. Osaki, 1998, Optimal continuous review policies for spare parts provisioning with random lead times, International Journal of Production Economics, 55, 257 – 271.

Short J., 2001, Dot.coms and the second generation of e-business: what's next?, Next e-Business: the future as a business, Seminar ISVOR FIAT, Torino, Italy, 23rd January.

Silver E.A., D.F. Pyke, R. Peterson, 1998, Inventory Management and Production Planning and Scheduling 3rd edition, John Wiley & Sons.

Simchi-Levi D., P. Kaminsky, E. Simchi-Levi, 2000, Designing and managing the supply chain, McGraw-Hill Companies, Inc.

282

References

Slack N., 1983, Flexibility as a manufacturing objective, International Journal of Operations and Production Management, 3 (3), 4-13.

Slater S.F., J.C. Narver, 1994, Does competitive environment moderate the market orientation- performance relationship?, Journal of Marketing, 58 (1), 46-55.

Slater S.F, J.C. Narver, 1995, Market orientation and the learning organisation, Journal of Marketing, 59, 63– 74.

Slater S.F., J.C. Narver, 2000, The positive effect o a market orientation on business profitability: a balanced replication, Journal of Business Research, 48, 69-73.

Small R.L., 1980, Sales forecasting in Canada: a survey of practices, The Conference Board of Canada, Study no. 66.

Soliman F., M. Youssef, 2001, The impact of some recent developments in e-business on the management of next generation manufacturing, International Journal of Operations and Production Management 21 (5/6), 538–564.

Sterman J.D., 1989, Modelling managerial behaviour: misperceptions of feedback in a dynamic decision making experiment, Management Science, 35 (3), 321-339.

Stuart I., P. Deckert, D. McCutcheon, R. Kunst, 1998, A leveraged learning network, Sloan Management Review, Summer.

Sweet A.L., 1980, An ad hoc method for forecasting series with zero values, AIEE, 12 (1), 97-103.

Syntetos A.A., J.E. Boylan, 2001, On the bias of intermittent demand estimates, International Journal of Production Economics, 71, 457-466.

Teunter R.H., L. Fortuin, 1999, End – of - life service, International Journal of Production Economics, 59, 487 – 497.

Theil H., 1954, Linear aggregation of economic relations, Amsterdam, North-Holland Publishing Company.

Towill, D.R., M.M. Naim, J. Wikner, 1992, Industrial dynamics simulation models in the design of supply chains, International Journal of Physical Distribution and Logistics Management 22 (5), 3– 13.

Tufano P., 1992, Financial innovation and first mover advantages, Journal of Applied Corporate Finance, 1, 83-87.

283

References

Tyagi R., C. Das, 1994, Manufacturer and warehouse selection for stable relationship in dynamic wholesaling and location problems, International Journal of Physical Distribution and Logistics Management, 25 (6), 54-72.

Tyagi R., C. Das, 1999, Grouping customers for better allocation of resources to serve correlated demands Computers and Operations Research 26, 1041-1058.

Van der Heijden M.C., E.B., Diks, A.G. de Kok, 1997, Stock allocation in general multi-echelon distribution systems with (R,S) order-up-to policies, International Journal pf Production Economics, 49 (2), 157-174.

Van der Vorst J., A. Beulens, W. De Wit, P. Van Beek, 1998, Supply chain management in food chains: improving performance by reducing uncertainty, International Transactions in Operational Research, 5 (6), 487-499.

Van Hoek R.I., 2001, The rediscovery of postponement: a literature review and directions for research, Journal of Operations Management, 19, 161-184.

Van Houtum G. J., K. Inderfurth, W.H.M. Zijm, 1996, Materials co-ordination in stochastic multi- echelon systems, European Journal of Operational Research, 95, 1-23.

Venkatraman V., 2001, E-Business strategies: the next chapter, Next e-Business: the future as a business, Seminar ISVOR FIAT, Torino, Italy, 23rd January.

Vereecke A., P. Verstraeten, 1994, An inventory model for an inventory consisting of lumpy items, slow movers and fast movers, International Journal of Production Economics 35, 379-389.

Verganti R., 1997, Order overplannning with uncertain lumpy demand: a simplified theory, International Journal of Production Research, 35, 12, 3229-3248.

Vollman T.E., W.L. Berry, D.C. Whybark, 1992, Manufacturing planning and control systems, 3rd ed., Richard D. Irwin, Homewood, IL.

Vollmann T.E., C. Cordon, J. Heikkilä, 2000, Teaching supply chain management to business executives, Production and Operations Management Journal 9 (1), 81–90.

Wacker J.G., L.G. Sprague, 1998, Forecasting accuracy: comparing the relative effectiveness of practices between seven developed countries, Journal of Operations Management, 16, 271-290.

Ward J.B., 1978, Determining reorder points when demand is lumpy, Management Science, 24 (6), 623-632.

Watson M., 1996, Forecasting in the Scottish electronics industry, International Journal of Forecasting, 12, 361-371.

284

References

Weatherford L.R., S.E. Kimes, D.A. Scott, 2001, Forecasting for Hotel Revenue Management: Testing Aggregation Against Disaggregation, Cornell Hotel and Restaurant Administration Quarterly, August, 53-64.

Webb D., C. Webster, A. Krepapa, 2000, An exploration of the meaning and outcomes of a customer-defined market orientation, Journal of Business Research, 48, 101-112.

Webby R., M. O’Connor, 1996, Judgmental and statistical time series forecasting: A review of the literature, International Journal of Forecasting, 12, 91-118.

Wecker W.E., 1978, Predicting demand from sales data in the presence of stock-outs, Management Science, 24 (10), 1043-1054.

Wedel M., W.A. Kamakura, 2000, Market segmentation: conceptual and methodological foundations, Kluwer Acedemic Publishers, Norwell, Massachusetts.

Wemmerlöv U., D.C. Whybark, 1984, Lot-sizing under uncertainty in a rolling schedule environment, International Journal of Production Research, 22 (3), 467-484.

Wemmerlöv U., 1986, A time phased order-point system in environments with and without demand uncertainty: A comparative analysis of no-monetary performance variables, International Journal of Production Research, 24 (2), 343-358.

White C.R., B.R. White, 1992, Emergency inventory ordering assuming Poisson demands, Production Planning and Control, 3, 168-174.

Wijngaard J., J.C. Wortmann, 1985, MRP and inventories, European Journal of Operational Research, 20, 185-189.

Willemain T.R., C.N. Smart, J.H. Shockor, P.A. DeSautels, 1994, Forecasting intermittent demand in manufacturing: a comparative evaluation of Croston's method, International Journal of Forecasting, 10, 529-538.

Williams T.M., 1982, Reorder levels for lumpy demand, Journal of the Operational Research Society, 33, 185-189.

Williams T., R. Maull, B. Ellis, 2002, Demand chain management theory: constrains and development from global aerospace supply webs, Journal of Operations Management, 20, 691- 706.

Williamson O.E., 1985, The Economic Institutions of Capitalism, Free Press, NY.

Winters P.R., 1960, Forecasting sales by exponentially weighted moving averages, Management Science, 6, March, 324-342.

285

References

Wong V, J. Saunders, 1993, Business orientations and corporate success, Journal of Strategic Marketing, 1, 20 –40.

Xu K. Y. Dong, P.T. Evers, 2001, Towards better coordination of the supply chain, Transportation Research Part E, 37, 35-54.

Yamashina H., 1989, The service parts control problem, Engineering Costs and Production Economics, 16, 195 – 208.

Zellen, B., 2001, Helping supply chains meet demand chains: Applix reinvents itself with global strategy and new partners, Available at: http://www.interex.org/hpworldnews/hpw101/03erp.html.

Zellner A., J. Tobias, 2000, A Note on Aggregation, Disaggregation and Forecasting Performance, Journal of Forecasting, 19, 457-69.

Zhao X, J. Xie, R.S.M. Lau, 2001, Improving the supply chain performance: use of forecasting models versus early order commitments, International Journal of Production Research, 39 (17), 3923-3939.

Zhao X., J.C. Goodale, T.S. Lee, 1995, Lot-sizing and freezing the master production schedule in material requirement planning systems under demand uncertainty, International Journal of Production Research, 33, 241-276.

Zinn W., 1990, Should you assemble products before an order is received? Business Horizons, 33, 70-73.

Zinn W., D.J. Bowersox, 1988, Planning physical distribution with the principle of postponement, Journal of Business Logistics, 9, 117-136.

Zinn W., Levy M., 1988, Speculative inventory management: a total channel perspective, International Journal of Physical Distribution and Materials Management, 18, (5), 34-39.

Zotteri G., Verganti R., 2001, Multi-level approaches to demand management in complex environments: An analytical model, International Journal of Production Economics, 71, 221-233.

286

Who knows how long I've loved you? You know I love you still. Will I wait a lonely lifetime? If you want me to, I will.

For if I ever saw you, I didn't catch your name. But it never really mattered; I will always feel the same.

Love you forever and forever, Love you with all of my heart. Love you whenever we're together Love you when we're apart.

And when at last I find you, Your song will fill the air. Sing it loud so I can hear you, Make it easy to be near you, For the things you do endear you to me, Oh you know I will.

I will! Lennon – McCartney (1968)